drivers/infiniband/hw/ipath/ipath_file_ops.c

   1 /*
   2  * Copyright (c) 2006, 2007 QLogic Corporation. 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 "ipath_common.h"
  43
  44 static int ipath_open(struct inode *, struct file *);
  45 static int ipath_close(struct inode *, struct file *);
  46 static ssize_t ipath_write(struct file *, const char __user *, size_t,
  47                            loff_t *);
  48 static unsigned int ipath_poll(struct file *, struct poll_table_struct *);
  49 static int ipath_mmap(struct file *, struct vm_area_struct *);
  50
  51 static const struct file_operations ipath_file_ops = {
  52         .owner = THIS_MODULE,
  53         .write = ipath_write,
  54         .open = ipath_open,
  55         .release = ipath_close,
  56         .poll = ipath_poll,
  57         .mmap = ipath_mmap
  58 };
  59
  60 /*
  61  * Convert kernel virtual addresses to physical addresses so they don't
  62  * potentially conflict with the chip addresses used as mmap offsets.
  63  * It doesn't really matter what mmap offset we use as long as we can
  64  * interpret it correctly.
  65  */
  66 static u64 cvt_kvaddr(void *p)
  67 {
  68         struct page *page;
  69         u64 paddr = 0;
  70
  71         page = vmalloc_to_page(p);
  72         if (page)
  73                 paddr = page_to_pfn(page) << PAGE_SHIFT;
  74
  75         return paddr;
  76 }
  77
  78 static int ipath_get_base_info(struct file *fp,
  79                                void __user *ubase, size_t ubase_size)
  80 {
  81         struct ipath_portdata *pd = port_fp(fp);
  82         int ret = 0;
  83         struct ipath_base_info *kinfo = NULL;
  84         struct ipath_devdata *dd = pd->port_dd;
  85         unsigned subport_cnt;
  86         int shared, master;
  87         size_t sz;
  88
  89         subport_cnt = pd->port_subport_cnt;
  90         if (!subport_cnt) {
  91                 shared = 0;
  92                 master = 0;
  93                 subport_cnt = 1;
  94         } else {
  95                 shared = 1;
  96                 master = !subport_fp(fp);
  97         }
  98
  99         sz = sizeof(*kinfo);
 100         /* If port sharing is not requested, allow the old size structure */
 101         if (!shared)
 102                 sz -= 7 * sizeof(u64);
 103         if (ubase_size < sz) {
 104                 ipath_cdbg(PROC,
 105                            "Base size %zu, need %zu (version mismatch?)\n",
 106                            ubase_size, sz);
 107                 ret = -EINVAL;
 108                 goto bail;
 109         }
 110
 111         kinfo = kzalloc(sizeof(*kinfo), GFP_KERNEL);
 112         if (kinfo == NULL) {
 113                 ret = -ENOMEM;
 114                 goto bail;
 115         }
 116
 117         ret = dd->ipath_f_get_base_info(pd, kinfo);
 118         if (ret < 0)
 119                 goto bail;
 120
 121         kinfo->spi_rcvhdr_cnt = dd->ipath_rcvhdrcnt;
 122         kinfo->spi_rcvhdrent_size = dd->ipath_rcvhdrentsize;
 123         kinfo->spi_tidegrcnt = dd->ipath_rcvegrcnt;
 124         kinfo->spi_rcv_egrbufsize = dd->ipath_rcvegrbufsize;
 125         /*
 126          * have to mmap whole thing
 127          */
 128         kinfo->spi_rcv_egrbuftotlen =
 129                 pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
 130         kinfo->spi_rcv_egrperchunk = pd->port_rcvegrbufs_perchunk;
 131         kinfo->spi_rcv_egrchunksize = kinfo->spi_rcv_egrbuftotlen /
 132                 pd->port_rcvegrbuf_chunks;
 133         kinfo->spi_tidcnt = dd->ipath_rcvtidcnt / subport_cnt;
 134         if (master)
 135                 kinfo->spi_tidcnt += dd->ipath_rcvtidcnt % subport_cnt;
 136         /*
 137          * for this use, may be ipath_cfgports summed over all chips that
 138          * are are configured and present
 139          */
 140         kinfo->spi_nports = dd->ipath_cfgports;
 141         /* unit (chip/board) our port is on */
 142         kinfo->spi_unit = dd->ipath_unit;
 143         /* for now, only a single page */
 144         kinfo->spi_tid_maxsize = PAGE_SIZE;
 145
 146         /*
 147          * Doing this per port, and based on the skip value, etc.  This has
 148          * to be the actual buffer size, since the protocol code treats it
 149          * as an array.
 150          *
 151          * These have to be set to user addresses in the user code via mmap.
 152          * These values are used on return to user code for the mmap target
 153          * addresses only.  For 32 bit, same 44 bit address problem, so use
 154          * the physical address, not virtual.  Before 2.6.11, using the
 155          * page_address() macro worked, but in 2.6.11, even that returns the
 156          * full 64 bit address (upper bits all 1's).  So far, using the
 157          * physical addresses (or chip offsets, for chip mapping) works, but
 158          * no doubt some future kernel release will change that, and we'll be
 159          * on to yet another method of dealing with this.
 160          */
 161         kinfo->spi_rcvhdr_base = (u64) pd->port_rcvhdrq_phys;
 162         kinfo->spi_rcvhdr_tailaddr = (u64) pd->port_rcvhdrqtailaddr_phys;
 163         kinfo->spi_rcv_egrbufs = (u64) pd->port_rcvegr_phys;
 164         kinfo->spi_pioavailaddr = (u64) dd->ipath_pioavailregs_phys;
 165         kinfo->spi_status = (u64) kinfo->spi_pioavailaddr +
 166                 (void *) dd->ipath_statusp -
 167                 (void *) dd->ipath_pioavailregs_dma;
 168         if (!shared) {
 169                 kinfo->spi_piocnt = dd->ipath_pbufsport;
 170                 kinfo->spi_piobufbase = (u64) pd->port_piobufs;
 171                 kinfo->__spi_uregbase = (u64) dd->ipath_uregbase +
 172                         dd->ipath_palign * pd->port_port;
 173         } else if (master) {
 174                 kinfo->spi_piocnt = (dd->ipath_pbufsport / subport_cnt) +
 175                                     (dd->ipath_pbufsport % subport_cnt);
 176                 /* Master's PIO buffers are after all the slave's */
 177                 kinfo->spi_piobufbase = (u64) pd->port_piobufs +
 178                         dd->ipath_palign *
 179                         (dd->ipath_pbufsport - kinfo->spi_piocnt);
 180         } else {
 181                 unsigned slave = subport_fp(fp) - 1;
 182
 183                 kinfo->spi_piocnt = dd->ipath_pbufsport / subport_cnt;
 184                 kinfo->spi_piobufbase = (u64) pd->port_piobufs +
 185                         dd->ipath_palign * kinfo->spi_piocnt * slave;
 186         }
 187         if (shared) {
 188                 kinfo->spi_port_uregbase = (u64) dd->ipath_uregbase +
 189                         dd->ipath_palign * pd->port_port;
 190                 kinfo->spi_port_rcvegrbuf = kinfo->spi_rcv_egrbufs;
 191                 kinfo->spi_port_rcvhdr_base = kinfo->spi_rcvhdr_base;
 192                 kinfo->spi_port_rcvhdr_tailaddr = kinfo->spi_rcvhdr_tailaddr;
 193
 194                 kinfo->__spi_uregbase = cvt_kvaddr(pd->subport_uregbase +
 195                         PAGE_SIZE * subport_fp(fp));
 196
 197                 kinfo->spi_rcvhdr_base = cvt_kvaddr(pd->subport_rcvhdr_base +
 198                         pd->port_rcvhdrq_size * subport_fp(fp));
 199                 kinfo->spi_rcvhdr_tailaddr = 0;
 200                 kinfo->spi_rcv_egrbufs = cvt_kvaddr(pd->subport_rcvegrbuf +
 201                         pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size *
 202                         subport_fp(fp));
 203
 204                 kinfo->spi_subport_uregbase =
 205                         cvt_kvaddr(pd->subport_uregbase);
 206                 kinfo->spi_subport_rcvegrbuf =
 207                         cvt_kvaddr(pd->subport_rcvegrbuf);
 208                 kinfo->spi_subport_rcvhdr_base =
 209                         cvt_kvaddr(pd->subport_rcvhdr_base);
 210                 ipath_cdbg(PROC, "port %u flags %x %llx %llx %llx\n",
 211                         kinfo->spi_port, kinfo->spi_runtime_flags,
 212                         (unsigned long long) kinfo->spi_subport_uregbase,
 213                         (unsigned long long) kinfo->spi_subport_rcvegrbuf,
 214                         (unsigned long long) kinfo->spi_subport_rcvhdr_base);
 215         }
 216
 217         kinfo->spi_pioindex = (kinfo->spi_piobufbase - dd->ipath_piobufbase) /
 218                 dd->ipath_palign;
 219         kinfo->spi_pioalign = dd->ipath_palign;
 220
 221         kinfo->spi_qpair = IPATH_KD_QP;
 222         kinfo->spi_piosize = dd->ipath_ibmaxlen;
 223         kinfo->spi_mtu = dd->ipath_ibmaxlen;    /* maxlen, not ibmtu */
 224         kinfo->spi_port = pd->port_port;
 225         kinfo->spi_subport = subport_fp(fp);
 226         kinfo->spi_sw_version = IPATH_KERN_SWVERSION;
 227         kinfo->spi_hw_version = dd->ipath_revision;
 228
 229         if (master) {
 230                 kinfo->spi_runtime_flags |= IPATH_RUNTIME_MASTER;
 231         }
 232
 233         sz = (ubase_size < sizeof(*kinfo)) ? ubase_size : sizeof(*kinfo);
 234         if (copy_to_user(ubase, kinfo, sz))
 235                 ret = -EFAULT;
 236
 237 bail:
 238         kfree(kinfo);
 239         return ret;
 240 }
 241
 242 /**
 243  * ipath_tid_update - update a port TID
 244  * @pd: the port
 245  * @fp: the ipath device file
 246  * @ti: the TID information
 247  *
 248  * The new implementation as of Oct 2004 is that the driver assigns
 249  * the tid and returns it to the caller.   To make it easier to
 250  * catch bugs, and to reduce search time, we keep a cursor for
 251  * each port, walking the shadow tid array to find one that's not
 252  * in use.
 253  *
 254  * For now, if we can't allocate the full list, we fail, although
 255  * in the long run, we'll allocate as many as we can, and the
 256  * caller will deal with that by trying the remaining pages later.
 257  * That means that when we fail, we have to mark the tids as not in
 258  * use again, in our shadow copy.
 259  *
 260  * It's up to the caller to free the tids when they are done.
 261  * We'll unlock the pages as they free them.
 262  *
 263  * Also, right now we are locking one page at a time, but since
 264  * the intended use of this routine is for a single group of
 265  * virtually contiguous pages, that should change to improve
 266  * performance.
 267  */
 268 static int ipath_tid_update(struct ipath_portdata *pd, struct file *fp,
 269                             const struct ipath_tid_info *ti)
 270 {
 271         int ret = 0, ntids;
 272         u32 tid, porttid, cnt, i, tidcnt, tidoff;
 273         u16 *tidlist;
 274         struct ipath_devdata *dd = pd->port_dd;
 275         u64 physaddr;
 276         unsigned long vaddr;
 277         u64 __iomem *tidbase;
 278         unsigned long tidmap[8];
 279         struct page **pagep = NULL;
 280         unsigned subport = subport_fp(fp);
 281
 282         if (!dd->ipath_pageshadow) {
 283                 ret = -ENOMEM;
 284                 goto done;
 285         }
 286
 287         cnt = ti->tidcnt;
 288         if (!cnt) {
 289                 ipath_dbg("After copyin, tidcnt 0, tidlist %llx\n",
 290                           (unsigned long long) ti->tidlist);
 291                 /*
 292                  * Should we treat as success?  likely a bug
 293                  */
 294                 ret = -EFAULT;
 295                 goto done;
 296         }
 297         porttid = pd->port_port * dd->ipath_rcvtidcnt;
 298         if (!pd->port_subport_cnt) {
 299                 tidcnt = dd->ipath_rcvtidcnt;
 300                 tid = pd->port_tidcursor;
 301                 tidoff = 0;
 302         } else if (!subport) {
 303                 tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
 304                          (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
 305                 tidoff = dd->ipath_rcvtidcnt - tidcnt;
 306                 porttid += tidoff;
 307                 tid = tidcursor_fp(fp);
 308         } else {
 309                 tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
 310                 tidoff = tidcnt * (subport - 1);
 311                 porttid += tidoff;
 312                 tid = tidcursor_fp(fp);
 313         }
 314         if (cnt > tidcnt) {
 315                 /* make sure it all fits in port_tid_pg_list */
 316                 dev_info(&dd->pcidev->dev, "Process tried to allocate %u "
 317                          "TIDs, only trying max (%u)\n", cnt, tidcnt);
 318                 cnt = tidcnt;
 319         }
 320         pagep = &((struct page **) pd->port_tid_pg_list)[tidoff];
 321         tidlist = &((u16 *) &pagep[dd->ipath_rcvtidcnt])[tidoff];
 322
 323         memset(tidmap, 0, sizeof(tidmap));
 324         /* before decrement; chip actual # */
 325         ntids = tidcnt;
 326         tidbase = (u64 __iomem *) (((char __iomem *) dd->ipath_kregbase) +
 327                                    dd->ipath_rcvtidbase +
 328                                    porttid * sizeof(*tidbase));
 329
 330         ipath_cdbg(VERBOSE, "Port%u %u tids, cursor %u, tidbase %p\n",
 331                    pd->port_port, cnt, tid, tidbase);
 332
 333         /* virtual address of first page in transfer */
 334         vaddr = ti->tidvaddr;
 335         if (!access_ok(VERIFY_WRITE, (void __user *) vaddr,
 336                        cnt * PAGE_SIZE)) {
 337                 ipath_dbg("Fail vaddr %p, %u pages, !access_ok\n",
 338                           (void *)vaddr, cnt);
 339                 ret = -EFAULT;
 340                 goto done;
 341         }
 342         ret = ipath_get_user_pages(vaddr, cnt, pagep);
 343         if (ret) {
 344                 if (ret == -EBUSY) {
 345                         ipath_dbg("Failed to lock addr %p, %u pages "
 346                                   "(already locked)\n",
 347                                   (void *) vaddr, cnt);
 348                         /*
 349                          * for now, continue, and see what happens but with
 350                          * the new implementation, this should never happen,
 351                          * unless perhaps the user has mpin'ed the pages
 352                          * themselves (something we need to test)
 353                          */
 354                         ret = 0;
 355                 } else {
 356                         dev_info(&dd->pcidev->dev,
 357                                  "Failed to lock addr %p, %u pages: "
 358                                  "errno %d\n", (void *) vaddr, cnt, -ret);
 359                         goto done;
 360                 }
 361         }
 362         for (i = 0; i < cnt; i++, vaddr += PAGE_SIZE) {
 363                 for (; ntids--; tid++) {
 364                         if (tid == tidcnt)
 365                                 tid = 0;
 366                         if (!dd->ipath_pageshadow[porttid + tid])
 367                                 break;
 368                 }
 369                 if (ntids < 0) {
 370                         /*
 371                          * oops, wrapped all the way through their TIDs,
 372                          * and didn't have enough free; see comments at
 373                          * start of routine
 374                          */
 375                         ipath_dbg("Not enough free TIDs for %u pages "
 376                                   "(index %d), failing\n", cnt, i);
 377                         i--;    /* last tidlist[i] not filled in */
 378                         ret = -ENOMEM;
 379                         break;
 380                 }
 381                 tidlist[i] = tid + tidoff;
 382                 ipath_cdbg(VERBOSE, "Updating idx %u to TID %u, "
 383                            "vaddr %lx\n", i, tid + tidoff, vaddr);
 384                 /* we "know" system pages and TID pages are same size */
 385                 dd->ipath_pageshadow[porttid + tid] = pagep[i];
 386                 dd->ipath_physshadow[porttid + tid] = ipath_map_page(
 387                         dd->pcidev, pagep[i], 0, PAGE_SIZE,
 388                         PCI_DMA_FROMDEVICE);
 389                 /*
 390                  * don't need atomic or it's overhead
 391                  */
 392                 __set_bit(tid, tidmap);
 393                 physaddr = dd->ipath_physshadow[porttid + tid];
 394                 ipath_stats.sps_pagelocks++;
 395                 ipath_cdbg(VERBOSE,
 396                            "TID %u, vaddr %lx, physaddr %llx pgp %p\n",
 397                            tid, vaddr, (unsigned long long) physaddr,
 398                            pagep[i]);
 399                 dd->ipath_f_put_tid(dd, &tidbase[tid], RCVHQ_RCV_TYPE_EXPECTED,
 400                                     physaddr);
 401                 /*
 402                  * don't check this tid in ipath_portshadow, since we
 403                  * just filled it in; start with the next one.
 404                  */
 405                 tid++;
 406         }
 407
 408         if (ret) {
 409                 u32 limit;
 410         cleanup:
 411                 /* jump here if copy out of updated info failed... */
 412                 ipath_dbg("After failure (ret=%d), undo %d of %d entries\n",
 413                           -ret, i, cnt);
 414                 /* same code that's in ipath_free_tid() */
 415                 limit = sizeof(tidmap) * BITS_PER_BYTE;
 416                 if (limit > tidcnt)
 417                         /* just in case size changes in future */
 418                         limit = tidcnt;
 419                 tid = find_first_bit((const unsigned long *)tidmap, limit);
 420                 for (; tid < limit; tid++) {
 421                         if (!test_bit(tid, tidmap))
 422                                 continue;
 423                         if (dd->ipath_pageshadow[porttid + tid]) {
 424                                 ipath_cdbg(VERBOSE, "Freeing TID %u\n",
 425                                            tid);
 426                                 dd->ipath_f_put_tid(dd, &tidbase[tid],
 427                                                     RCVHQ_RCV_TYPE_EXPECTED,
 428                                                     dd->ipath_tidinvalid);
 429                                 pci_unmap_page(dd->pcidev,
 430                                         dd->ipath_physshadow[porttid + tid],
 431                                         PAGE_SIZE, PCI_DMA_FROMDEVICE);
 432                                 dd->ipath_pageshadow[porttid + tid] = NULL;
 433                                 ipath_stats.sps_pageunlocks++;
 434                         }
 435                 }
 436                 ipath_release_user_pages(pagep, cnt);
 437         } else {
 438                 /*
 439                  * Copy the updated array, with ipath_tid's filled in, back
 440                  * to user.  Since we did the copy in already, this "should
 441                  * never fail" If it does, we have to clean up...
 442                  */
 443                 if (copy_to_user((void __user *)
 444                                  (unsigned long) ti->tidlist,
 445                                  tidlist, cnt * sizeof(*tidlist))) {
 446                         ret = -EFAULT;
 447                         goto cleanup;
 448                 }
 449                 if (copy_to_user((void __user *) (unsigned long) ti->tidmap,
 450                                  tidmap, sizeof tidmap)) {
 451                         ret = -EFAULT;
 452                         goto cleanup;
 453                 }
 454                 if (tid == tidcnt)
 455                         tid = 0;
 456                 if (!pd->port_subport_cnt)
 457                         pd->port_tidcursor = tid;
 458                 else
 459                         tidcursor_fp(fp) = tid;
 460         }
 461
 462 done:
 463         if (ret)
 464                 ipath_dbg("Failed to map %u TID pages, failing with %d\n",
 465                           ti->tidcnt, -ret);
 466         return ret;
 467 }
 468
 469 /**
 470  * ipath_tid_free - free a port TID
 471  * @pd: the port
 472  * @subport: the subport
 473  * @ti: the TID info
 474  *
 475  * right now we are unlocking one page at a time, but since
 476  * the intended use of this routine is for a single group of
 477  * virtually contiguous pages, that should change to improve
 478  * performance.  We check that the TID is in range for this port
 479  * but otherwise don't check validity; if user has an error and
 480  * frees the wrong tid, it's only their own data that can thereby
 481  * be corrupted.  We do check that the TID was in use, for sanity
 482  * We always use our idea of the saved address, not the address that
 483  * they pass in to us.
 484  */
 485
 486 static int ipath_tid_free(struct ipath_portdata *pd, unsigned subport,
 487                           const struct ipath_tid_info *ti)
 488 {
 489         int ret = 0;
 490         u32 tid, porttid, cnt, limit, tidcnt;
 491         struct ipath_devdata *dd = pd->port_dd;
 492         u64 __iomem *tidbase;
 493         unsigned long tidmap[8];
 494
 495         if (!dd->ipath_pageshadow) {
 496                 ret = -ENOMEM;
 497                 goto done;
 498         }
 499
 500         if (copy_from_user(tidmap, (void __user *)(unsigned long)ti->tidmap,
 501                            sizeof tidmap)) {
 502                 ret = -EFAULT;
 503                 goto done;
 504         }
 505
 506         porttid = pd->port_port * dd->ipath_rcvtidcnt;
 507         if (!pd->port_subport_cnt)
 508                 tidcnt = dd->ipath_rcvtidcnt;
 509         else if (!subport) {
 510                 tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
 511                          (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
 512                 porttid += dd->ipath_rcvtidcnt - tidcnt;
 513         } else {
 514                 tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
 515                 porttid += tidcnt * (subport - 1);
 516         }
 517         tidbase = (u64 __iomem *) ((char __iomem *)(dd->ipath_kregbase) +
 518                                    dd->ipath_rcvtidbase +
 519                                    porttid * sizeof(*tidbase));
 520
 521         limit = sizeof(tidmap) * BITS_PER_BYTE;
 522         if (limit > tidcnt)
 523                 /* just in case size changes in future */
 524                 limit = tidcnt;
 525         tid = find_first_bit(tidmap, limit);
 526         ipath_cdbg(VERBOSE, "Port%u free %u tids; first bit (max=%d) "
 527                    "set is %d, porttid %u\n", pd->port_port, ti->tidcnt,
 528                    limit, tid, porttid);
 529         for (cnt = 0; tid < limit; tid++) {
 530                 /*
 531                  * small optimization; if we detect a run of 3 or so without
 532                  * any set, use find_first_bit again.  That's mainly to
 533                  * accelerate the case where we wrapped, so we have some at
 534                  * the beginning, and some at the end, and a big gap
 535                  * in the middle.
 536                  */
 537                 if (!test_bit(tid, tidmap))
 538                         continue;
 539                 cnt++;
 540                 if (dd->ipath_pageshadow[porttid + tid]) {
 541                         ipath_cdbg(VERBOSE, "PID %u freeing TID %u\n",
 542                                    pd->port_pid, tid);
 543                         dd->ipath_f_put_tid(dd, &tidbase[tid],
 544                                             RCVHQ_RCV_TYPE_EXPECTED,
 545                                             dd->ipath_tidinvalid);
 546                         pci_unmap_page(dd->pcidev,
 547                                 dd->ipath_physshadow[porttid + tid],
 548                                 PAGE_SIZE, PCI_DMA_FROMDEVICE);
 549                         ipath_release_user_pages(
 550                                 &dd->ipath_pageshadow[porttid + tid], 1);
 551                         dd->ipath_pageshadow[porttid + tid] = NULL;
 552                         ipath_stats.sps_pageunlocks++;
 553                 } else
 554                         ipath_dbg("Unused tid %u, ignoring\n", tid);
 555         }
 556         if (cnt != ti->tidcnt)
 557                 ipath_dbg("passed in tidcnt %d, only %d bits set in map\n",
 558                           ti->tidcnt, cnt);
 559 done:
 560         if (ret)
 561                 ipath_dbg("Failed to unmap %u TID pages, failing with %d\n",
 562                           ti->tidcnt, -ret);
 563         return ret;
 564 }
 565
 566 /**
 567  * ipath_set_part_key - set a partition key
 568  * @pd: the port
 569  * @key: the key
 570  *
 571  * We can have up to 4 active at a time (other than the default, which is
 572  * always allowed).  This is somewhat tricky, since multiple ports may set
 573  * the same key, so we reference count them, and clean up at exit.  All 4
 574  * partition keys are packed into a single infinipath register.  It's an
 575  * error for a process to set the same pkey multiple times.  We provide no
 576  * mechanism to de-allocate a pkey at this time, we may eventually need to
 577  * do that.  I've used the atomic operations, and no locking, and only make
 578  * a single pass through what's available.  This should be more than
 579  * adequate for some time. I'll think about spinlocks or the like if and as
 580  * it's necessary.
 581  */
 582 static int ipath_set_part_key(struct ipath_portdata *pd, u16 key)
 583 {
 584         struct ipath_devdata *dd = pd->port_dd;
 585         int i, any = 0, pidx = -1;
 586         u16 lkey = key & 0x7FFF;
 587         int ret;
 588
 589         if (lkey == (IPATH_DEFAULT_P_KEY & 0x7FFF)) {
 590                 /* nothing to do; this key always valid */
 591                 ret = 0;
 592                 goto bail;
 593         }
 594
 595         ipath_cdbg(VERBOSE, "p%u try to set pkey %hx, current keys "
 596                    "%hx:%x %hx:%x %hx:%x %hx:%x\n",
 597                    pd->port_port, key, dd->ipath_pkeys[0],
 598                    atomic_read(&dd->ipath_pkeyrefs[0]), dd->ipath_pkeys[1],
 599                    atomic_read(&dd->ipath_pkeyrefs[1]), dd->ipath_pkeys[2],
 600                    atomic_read(&dd->ipath_pkeyrefs[2]), dd->ipath_pkeys[3],
 601                    atomic_read(&dd->ipath_pkeyrefs[3]));
 602
 603         if (!lkey) {
 604                 ipath_cdbg(PROC, "p%u tries to set key 0, not allowed\n",
 605                            pd->port_port);
 606                 ret = -EINVAL;
 607                 goto bail;
 608         }
 609
 610         /*
 611          * Set the full membership bit, because it has to be
 612          * set in the register or the packet, and it seems
 613          * cleaner to set in the register than to force all
 614          * callers to set it. (see bug 4331)
 615          */
 616         key |= 0x8000;
 617
 618         for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
 619                 if (!pd->port_pkeys[i] && pidx == -1)
 620                         pidx = i;
 621                 if (pd->port_pkeys[i] == key) {
 622                         ipath_cdbg(VERBOSE, "p%u tries to set same pkey "
 623                                    "(%x) more than once\n",
 624                                    pd->port_port, key);
 625                         ret = -EEXIST;
 626                         goto bail;
 627                 }
 628         }
 629         if (pidx == -1) {
 630                 ipath_dbg("All pkeys for port %u already in use, "
 631                           "can't set %x\n", pd->port_port, key);
 632                 ret = -EBUSY;
 633                 goto bail;
 634         }
 635         for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
 636                 if (!dd->ipath_pkeys[i]) {
 637                         any++;
 638                         continue;
 639                 }
 640                 if (dd->ipath_pkeys[i] == key) {
 641                         atomic_t *pkrefs = &dd->ipath_pkeyrefs[i];
 642
 643                         if (atomic_inc_return(pkrefs) > 1) {
 644                                 pd->port_pkeys[pidx] = key;
 645                                 ipath_cdbg(VERBOSE, "p%u set key %x "
 646                                            "matches #%d, count now %d\n",
 647                                            pd->port_port, key, i,
 648                                            atomic_read(pkrefs));
 649                                 ret = 0;
 650                                 goto bail;
 651                         } else {
 652                                 /*
 653                                  * lost race, decrement count, catch below
 654                                  */
 655                                 atomic_dec(pkrefs);
 656                                 ipath_cdbg(VERBOSE, "Lost race, count was "
 657                                            "0, after dec, it's %d\n",
 658                                            atomic_read(pkrefs));
 659                                 any++;
 660                         }
 661                 }
 662                 if ((dd->ipath_pkeys[i] & 0x7FFF) == lkey) {
 663                         /*
 664                          * It makes no sense to have both the limited and
 665                          * full membership PKEY set at the same time since
 666                          * the unlimited one will disable the limited one.
 667                          */
 668                         ret = -EEXIST;
 669                         goto bail;
 670                 }
 671         }
 672         if (!any) {
 673                 ipath_dbg("port %u, all pkeys already in use, "
 674                           "can't set %x\n", pd->port_port, key);
 675                 ret = -EBUSY;
 676                 goto bail;
 677         }
 678         for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
 679                 if (!dd->ipath_pkeys[i] &&
 680                     atomic_inc_return(&dd->ipath_pkeyrefs[i]) == 1) {
 681                         u64 pkey;
 682
 683                         /* for ipathstats, etc. */
 684                         ipath_stats.sps_pkeys[i] = lkey;
 685                         pd->port_pkeys[pidx] = dd->ipath_pkeys[i] = key;
 686                         pkey =
 687                                 (u64) dd->ipath_pkeys[0] |
 688                                 ((u64) dd->ipath_pkeys[1] << 16) |
 689                                 ((u64) dd->ipath_pkeys[2] << 32) |
 690                                 ((u64) dd->ipath_pkeys[3] << 48);
 691                         ipath_cdbg(PROC, "p%u set key %x in #%d, "
 692                                    "portidx %d, new pkey reg %llx\n",
 693                                    pd->port_port, key, i, pidx,
 694                                    (unsigned long long) pkey);
 695                         ipath_write_kreg(
 696                                 dd, dd->ipath_kregs->kr_partitionkey, pkey);
 697
 698                         ret = 0;
 699                         goto bail;
 700                 }
 701         }
 702         ipath_dbg("port %u, all pkeys already in use 2nd pass, "
 703                   "can't set %x\n", pd->port_port, key);
 704         ret = -EBUSY;
 705
 706 bail:
 707         return ret;
 708 }
 709
 710 /**
 711  * ipath_manage_rcvq - manage a port's receive queue
 712  * @pd: the port
 713  * @subport: the subport
 714  * @start_stop: action to carry out
 715  *
 716  * start_stop == 0 disables receive on the port, for use in queue
 717  * overflow conditions.  start_stop==1 re-enables, to be used to
 718  * re-init the software copy of the head register
 719  */
 720 static int ipath_manage_rcvq(struct ipath_portdata *pd, unsigned subport,
 721                              int start_stop)
 722 {
 723         struct ipath_devdata *dd = pd->port_dd;
 724
 725         ipath_cdbg(PROC, "%sabling rcv for unit %u port %u:%u\n",
 726                    start_stop ? "en" : "dis", dd->ipath_unit,
 727                    pd->port_port, subport);
 728         if (subport)
 729                 goto bail;
 730         /* atomically clear receive enable port. */
 731         if (start_stop) {
 732                 /*
 733                  * On enable, force in-memory copy of the tail register to
 734                  * 0, so that protocol code doesn't have to worry about
 735                  * whether or not the chip has yet updated the in-memory
 736                  * copy or not on return from the system call. The chip
 737                  * always resets it's tail register back to 0 on a
 738                  * transition from disabled to enabled.  This could cause a
 739                  * problem if software was broken, and did the enable w/o
 740                  * the disable, but eventually the in-memory copy will be
 741                  * updated and correct itself, even in the face of software
 742                  * bugs.
 743                  */
 744                 *(volatile u64 *)pd->port_rcvhdrtail_kvaddr = 0;
 745                 set_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port,
 746                         &dd->ipath_rcvctrl);
 747         } else
 748                 clear_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port,
 749                           &dd->ipath_rcvctrl);
 750         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
 751                          dd->ipath_rcvctrl);
 752         /* now be sure chip saw it before we return */
 753         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 754         if (start_stop) {
 755                 /*
 756                  * And try to be sure that tail reg update has happened too.
 757                  * This should in theory interlock with the RXE changes to
 758                  * the tail register.  Don't assign it to the tail register
 759                  * in memory copy, since we could overwrite an update by the
 760                  * chip if we did.
 761                  */
 762                 ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
 763         }
 764         /* always; new head should be equal to new tail; see above */
 765 bail:
 766         return 0;
 767 }
 768
 769 static void ipath_clean_part_key(struct ipath_portdata *pd,
 770                                  struct ipath_devdata *dd)
 771 {
 772         int i, j, pchanged = 0;
 773         u64 oldpkey;
 774
 775         /* for debugging only */
 776         oldpkey = (u64) dd->ipath_pkeys[0] |
 777                 ((u64) dd->ipath_pkeys[1] << 16) |
 778                 ((u64) dd->ipath_pkeys[2] << 32) |
 779                 ((u64) dd->ipath_pkeys[3] << 48);
 780
 781         for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
 782                 if (!pd->port_pkeys[i])
 783                         continue;
 784                 ipath_cdbg(VERBOSE, "look for key[%d] %hx in pkeys\n", i,
 785                            pd->port_pkeys[i]);
 786                 for (j = 0; j < ARRAY_SIZE(dd->ipath_pkeys); j++) {
 787                         /* check for match independent of the global bit */
 788                         if ((dd->ipath_pkeys[j] & 0x7fff) !=
 789                             (pd->port_pkeys[i] & 0x7fff))
 790                                 continue;
 791                         if (atomic_dec_and_test(&dd->ipath_pkeyrefs[j])) {
 792                                 ipath_cdbg(VERBOSE, "p%u clear key "
 793                                            "%x matches #%d\n",
 794                                            pd->port_port,
 795                                            pd->port_pkeys[i], j);
 796                                 ipath_stats.sps_pkeys[j] =
 797                                         dd->ipath_pkeys[j] = 0;
 798                                 pchanged++;
 799                         }
 800                         else ipath_cdbg(
 801                                 VERBOSE, "p%u key %x matches #%d, "
 802                                 "but ref still %d\n", pd->port_port,
 803                                 pd->port_pkeys[i], j,
 804                                 atomic_read(&dd->ipath_pkeyrefs[j]));
 805                         break;
 806                 }
 807                 pd->port_pkeys[i] = 0;
 808         }
 809         if (pchanged) {
 810                 u64 pkey = (u64) dd->ipath_pkeys[0] |
 811                         ((u64) dd->ipath_pkeys[1] << 16) |
 812                         ((u64) dd->ipath_pkeys[2] << 32) |
 813                         ((u64) dd->ipath_pkeys[3] << 48);
 814                 ipath_cdbg(VERBOSE, "p%u old pkey reg %llx, "
 815                            "new pkey reg %llx\n", pd->port_port,
 816                            (unsigned long long) oldpkey,
 817                            (unsigned long long) pkey);
 818                 ipath_write_kreg(dd, dd->ipath_kregs->kr_partitionkey,
 819                                  pkey);
 820         }
 821 }
 822
 823 /*
 824  * Initialize the port data with the receive buffer sizes
 825  * so this can be done while the master port is locked.
 826  * Otherwise, there is a race with a slave opening the port
 827  * and seeing these fields uninitialized.
 828  */
 829 static void init_user_egr_sizes(struct ipath_portdata *pd)
 830 {
 831         struct ipath_devdata *dd = pd->port_dd;
 832         unsigned egrperchunk, egrcnt, size;
 833
 834         /*
 835          * to avoid wasting a lot of memory, we allocate 32KB chunks of
 836          * physically contiguous memory, advance through it until used up
 837          * and then allocate more.  Of course, we need memory to store those
 838          * extra pointers, now.  Started out with 256KB, but under heavy
 839          * memory pressure (creating large files and then copying them over
 840          * NFS while doing lots of MPI jobs), we hit some allocation
 841          * failures, even though we can sleep...  (2.6.10) Still get
 842          * failures at 64K.  32K is the lowest we can go without wasting
 843          * additional memory.
 844          */
 845         size = 0x8000;
 846         egrperchunk = size / dd->ipath_rcvegrbufsize;
 847         egrcnt = dd->ipath_rcvegrcnt;
 848         pd->port_rcvegrbuf_chunks = (egrcnt + egrperchunk - 1) / egrperchunk;
 849         pd->port_rcvegrbufs_perchunk = egrperchunk;
 850         pd->port_rcvegrbuf_size = size;
 851 }
 852
 853 /**
 854  * ipath_create_user_egr - allocate eager TID buffers
 855  * @pd: the port to allocate TID buffers for
 856  *
 857  * This routine is now quite different for user and kernel, because
 858  * the kernel uses skb's, for the accelerated network performance
 859  * This is the user port version
 860  *
 861  * Allocate the eager TID buffers and program them into infinipath
 862  * They are no longer completely contiguous, we do multiple allocation
 863  * calls.
 864  */
 865 static int ipath_create_user_egr(struct ipath_portdata *pd)
 866 {
 867         struct ipath_devdata *dd = pd->port_dd;
 868         unsigned e, egrcnt, egrperchunk, chunk, egrsize, egroff;
 869         size_t size;
 870         int ret;
 871         gfp_t gfp_flags;
 872
 873         /*
 874          * GFP_USER, but without GFP_FS, so buffer cache can be
 875          * coalesced (we hope); otherwise, even at order 4,
 876          * heavy filesystem activity makes these fail, and we can
 877          * use compound pages.
 878          */
 879         gfp_flags = __GFP_WAIT | __GFP_IO | __GFP_COMP;
 880
 881         egrcnt = dd->ipath_rcvegrcnt;
 882         /* TID number offset for this port */
 883         egroff = pd->port_port * egrcnt;
 884         egrsize = dd->ipath_rcvegrbufsize;
 885         ipath_cdbg(VERBOSE, "Allocating %d egr buffers, at egrtid "
 886                    "offset %x, egrsize %u\n", egrcnt, egroff, egrsize);
 887
 888         chunk = pd->port_rcvegrbuf_chunks;
 889         egrperchunk = pd->port_rcvegrbufs_perchunk;
 890         size = pd->port_rcvegrbuf_size;
 891         pd->port_rcvegrbuf = kmalloc(chunk * sizeof(pd->port_rcvegrbuf[0]),
 892                                      GFP_KERNEL);
 893         if (!pd->port_rcvegrbuf) {
 894                 ret = -ENOMEM;
 895                 goto bail;
 896         }
 897         pd->port_rcvegrbuf_phys =
 898                 kmalloc(chunk * sizeof(pd->port_rcvegrbuf_phys[0]),
 899                         GFP_KERNEL);
 900         if (!pd->port_rcvegrbuf_phys) {
 901                 ret = -ENOMEM;
 902                 goto bail_rcvegrbuf;
 903         }
 904         for (e = 0; e < pd->port_rcvegrbuf_chunks; e++) {
 905
 906                 pd->port_rcvegrbuf[e] = dma_alloc_coherent(
 907                         &dd->pcidev->dev, size, &pd->port_rcvegrbuf_phys[e],
 908                         gfp_flags);
 909
 910                 if (!pd->port_rcvegrbuf[e]) {
 911                         ret = -ENOMEM;
 912                         goto bail_rcvegrbuf_phys;
 913                 }
 914         }
 915
 916         pd->port_rcvegr_phys = pd->port_rcvegrbuf_phys[0];
 917
 918         for (e = chunk = 0; chunk < pd->port_rcvegrbuf_chunks; chunk++) {
 919                 dma_addr_t pa = pd->port_rcvegrbuf_phys[chunk];
 920                 unsigned i;
 921
 922                 for (i = 0; e < egrcnt && i < egrperchunk; e++, i++) {
 923                         dd->ipath_f_put_tid(dd, e + egroff +
 924                                             (u64 __iomem *)
 925                                             ((char __iomem *)
 926                                              dd->ipath_kregbase +
 927                                              dd->ipath_rcvegrbase),
 928                                             RCVHQ_RCV_TYPE_EAGER, pa);
 929                         pa += egrsize;
 930                 }
 931                 cond_resched(); /* don't hog the cpu */
 932         }
 933
 934         ret = 0;
 935         goto bail;
 936
 937 bail_rcvegrbuf_phys:
 938         for (e = 0; e < pd->port_rcvegrbuf_chunks &&
 939                 pd->port_rcvegrbuf[e]; e++) {
 940                 dma_free_coherent(&dd->pcidev->dev, size,
 941                                   pd->port_rcvegrbuf[e],
 942                                   pd->port_rcvegrbuf_phys[e]);
 943
 944         }
 945         kfree(pd->port_rcvegrbuf_phys);
 946         pd->port_rcvegrbuf_phys = NULL;
 947 bail_rcvegrbuf:
 948         kfree(pd->port_rcvegrbuf);
 949         pd->port_rcvegrbuf = NULL;
 950 bail:
 951         return ret;
 952 }
 953
 954
 955 /* common code for the mappings on dma_alloc_coherent mem */
 956 static int ipath_mmap_mem(struct vm_area_struct *vma,
 957         struct ipath_portdata *pd, unsigned len, int write_ok,
 958         void *kvaddr, char *what)
 959 {
 960         struct ipath_devdata *dd = pd->port_dd;
 961         unsigned long pfn;
 962         int ret;
 963
 964         if ((vma->vm_end - vma->vm_start) > len) {
 965                 dev_info(&dd->pcidev->dev,
 966                          "FAIL on %s: len %lx > %x\n", what,
 967                          vma->vm_end - vma->vm_start, len);
 968                 ret = -EFAULT;
 969                 goto bail;
 970         }
 971
 972         if (!write_ok) {
 973                 if (vma->vm_flags & VM_WRITE) {
 974                         dev_info(&dd->pcidev->dev,
 975                                  "%s must be mapped readonly\n", what);
 976                         ret = -EPERM;
 977                         goto bail;
 978                 }
 979
 980                 /* don't allow them to later change with mprotect */
 981                 vma->vm_flags &= ~VM_MAYWRITE;
 982         }
 983
 984         pfn = virt_to_phys(kvaddr) >> PAGE_SHIFT;
 985         ret = remap_pfn_range(vma, vma->vm_start, pfn,
 986                               len, vma->vm_page_prot);
 987         if (ret)
 988                 dev_info(&dd->pcidev->dev, "%s port%u mmap of %lx, %x "
 989                          "bytes r%c failed: %d\n", what, pd->port_port,
 990                          pfn, len, write_ok?'w':'o', ret);
 991         else
 992                 ipath_cdbg(VERBOSE, "%s port%u mmaped %lx, %x bytes "
 993                            "r%c\n", what, pd->port_port, pfn, len,
 994                            write_ok?'w':'o');
 995 bail:
 996         return ret;
 997 }
 998
 999 static int mmap_ureg(struct vm_area_struct *vma, struct ipath_devdata *dd,
1000                      u64 ureg)
1001 {
1002         unsigned long phys;
1003         int ret;
1004
1005         /*
1006          * This is real hardware, so use io_remap.  This is the mechanism
1007          * for the user process to update the head registers for their port
1008          * in the chip.
1009          */
1010         if ((vma->vm_end - vma->vm_start) > PAGE_SIZE) {
1011                 dev_info(&dd->pcidev->dev, "FAIL mmap userreg: reqlen "
1012                          "%lx > PAGE\n", vma->vm_end - vma->vm_start);
1013                 ret = -EFAULT;
1014         } else {
1015                 phys = dd->ipath_physaddr + ureg;
1016                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1017
1018                 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1019                 ret = io_remap_pfn_range(vma, vma->vm_start,
1020                                          phys >> PAGE_SHIFT,
1021                                          vma->vm_end - vma->vm_start,
1022                                          vma->vm_page_prot);
1023         }
1024         return ret;
1025 }
1026
1027 static int mmap_piobufs(struct vm_area_struct *vma,
1028                         struct ipath_devdata *dd,
1029                         struct ipath_portdata *pd,
1030                         unsigned piobufs, unsigned piocnt)
1031 {
1032         unsigned long phys;
1033         int ret;
1034
1035         /*
1036          * When we map the PIO buffers in the chip, we want to map them as
1037          * writeonly, no read possible.   This prevents access to previous
1038          * process data, and catches users who might try to read the i/o
1039          * space due to a bug.
1040          */
1041         if ((vma->vm_end - vma->vm_start) > (piocnt * dd->ipath_palign)) {
1042                 dev_info(&dd->pcidev->dev, "FAIL mmap piobufs: "
1043                          "reqlen %lx > PAGE\n",
1044                          vma->vm_end - vma->vm_start);
1045                 ret = -EINVAL;
1046                 goto bail;
1047         }
1048
1049         phys = dd->ipath_physaddr + piobufs;
1050
1051         /*
1052          * Don't mark this as non-cached, or we don't get the
1053          * write combining behavior we want on the PIO buffers!
1054          */
1055
1056 #if defined(__powerpc__)
1057         /* There isn't a generic way to specify writethrough mappings */
1058         pgprot_val(vma->vm_page_prot) |= _PAGE_NO_CACHE;
1059         pgprot_val(vma->vm_page_prot) |= _PAGE_WRITETHRU;
1060         pgprot_val(vma->vm_page_prot) &= ~_PAGE_GUARDED;
1061 #endif
1062
1063         /*
1064          * don't allow them to later change to readable with mprotect (for when
1065          * not initially mapped readable, as is normally the case)
1066          */
1067         vma->vm_flags &= ~VM_MAYREAD;
1068         vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1069
1070         ret = io_remap_pfn_range(vma, vma->vm_start, phys >> PAGE_SHIFT,
1071                                  vma->vm_end - vma->vm_start,
1072                                  vma->vm_page_prot);
1073 bail:
1074         return ret;
1075 }
1076
1077 static int mmap_rcvegrbufs(struct vm_area_struct *vma,
1078                            struct ipath_portdata *pd)
1079 {
1080         struct ipath_devdata *dd = pd->port_dd;
1081         unsigned long start, size;
1082         size_t total_size, i;
1083         unsigned long pfn;
1084         int ret;
1085
1086         size = pd->port_rcvegrbuf_size;
1087         total_size = pd->port_rcvegrbuf_chunks * size;
1088         if ((vma->vm_end - vma->vm_start) > total_size) {
1089                 dev_info(&dd->pcidev->dev, "FAIL on egr bufs: "
1090                          "reqlen %lx > actual %lx\n",
1091                          vma->vm_end - vma->vm_start,
1092                          (unsigned long) total_size);
1093                 ret = -EINVAL;
1094                 goto bail;
1095         }
1096
1097         if (vma->vm_flags & VM_WRITE) {
1098                 dev_info(&dd->pcidev->dev, "Can't map eager buffers as "
1099                          "writable (flags=%lx)\n", vma->vm_flags);
1100                 ret = -EPERM;
1101                 goto bail;
1102         }
1103         /* don't allow them to later change to writeable with mprotect */
1104         vma->vm_flags &= ~VM_MAYWRITE;
1105
1106         start = vma->vm_start;
1107
1108         for (i = 0; i < pd->port_rcvegrbuf_chunks; i++, start += size) {
1109                 pfn = virt_to_phys(pd->port_rcvegrbuf[i]) >> PAGE_SHIFT;
1110                 ret = remap_pfn_range(vma, start, pfn, size,
1111                                       vma->vm_page_prot);
1112                 if (ret < 0)
1113                         goto bail;
1114         }
1115         ret = 0;
1116
1117 bail:
1118         return ret;
1119 }
1120
1121 /*
1122  * ipath_file_vma_nopage - handle a VMA page fault.
1123  */
1124 static struct page *ipath_file_vma_nopage(struct vm_area_struct *vma,
1125                                           unsigned long address, int *type)
1126 {
1127         unsigned long offset = address - vma->vm_start;
1128         struct page *page = NOPAGE_SIGBUS;
1129         void *pageptr;
1130
1131         /*
1132          * Convert the vmalloc address into a struct page.
1133          */
1134         pageptr = (void *)(offset + (vma->vm_pgoff << PAGE_SHIFT));
1135         page = vmalloc_to_page(pageptr);
1136         if (!page)
1137                 goto out;
1138
1139         /* Increment the reference count. */
1140         get_page(page);
1141         if (type)
1142                 *type = VM_FAULT_MINOR;
1143 out:
1144         return page;
1145 }
1146
1147 static struct vm_operations_struct ipath_file_vm_ops = {
1148         .nopage = ipath_file_vma_nopage,
1149 };
1150
1151 static int mmap_kvaddr(struct vm_area_struct *vma, u64 pgaddr,
1152                        struct ipath_portdata *pd, unsigned subport)
1153 {
1154         unsigned long len;
1155         struct ipath_devdata *dd;
1156         void *addr;
1157         size_t size;
1158         int ret = 0;
1159
1160         /* If the port is not shared, all addresses should be physical */
1161         if (!pd->port_subport_cnt)
1162                 goto bail;
1163
1164         dd = pd->port_dd;
1165         size = pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
1166
1167         /*
1168          * Each process has all the subport uregbase, rcvhdrq, and
1169          * rcvegrbufs mmapped - as an array for all the processes,
1170          * and also separately for this process.
1171          */
1172         if (pgaddr == cvt_kvaddr(pd->subport_uregbase)) {
1173                 addr = pd->subport_uregbase;
1174                 size = PAGE_SIZE * pd->port_subport_cnt;
1175         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvhdr_base)) {
1176                 addr = pd->subport_rcvhdr_base;
1177                 size = pd->port_rcvhdrq_size * pd->port_subport_cnt;
1178         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvegrbuf)) {
1179                 addr = pd->subport_rcvegrbuf;
1180                 size *= pd->port_subport_cnt;
1181         } else if (pgaddr == cvt_kvaddr(pd->subport_uregbase +
1182                                         PAGE_SIZE * subport)) {
1183                 addr = pd->subport_uregbase + PAGE_SIZE * subport;
1184                 size = PAGE_SIZE;
1185         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvhdr_base +
1186                                 pd->port_rcvhdrq_size * subport)) {
1187                 addr = pd->subport_rcvhdr_base +
1188                         pd->port_rcvhdrq_size * subport;
1189                 size = pd->port_rcvhdrq_size;
1190         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvegrbuf +
1191                                size * subport)) {
1192                 addr = pd->subport_rcvegrbuf + size * subport;
1193                 /* rcvegrbufs are read-only on the slave */
1194                 if (vma->vm_flags & VM_WRITE) {
1195                         dev_info(&dd->pcidev->dev,
1196                                  "Can't map eager buffers as "
1197                                  "writable (flags=%lx)\n", vma->vm_flags);
1198                         ret = -EPERM;
1199                         goto bail;
1200                 }
1201                 /*
1202                  * Don't allow permission to later change to writeable
1203                  * with mprotect.
1204                  */
1205                 vma->vm_flags &= ~VM_MAYWRITE;
1206         } else {
1207                 goto bail;
1208         }
1209         len = vma->vm_end - vma->vm_start;
1210         if (len > size) {
1211                 ipath_cdbg(MM, "FAIL: reqlen %lx > %zx\n", len, size);
1212                 ret = -EINVAL;
1213                 goto bail;
1214         }
1215
1216         vma->vm_pgoff = (unsigned long) addr >> PAGE_SHIFT;
1217         vma->vm_ops = &ipath_file_vm_ops;
1218         vma->vm_flags |= VM_RESERVED | VM_DONTEXPAND;
1219         ret = 1;
1220
1221 bail:
1222         return ret;
1223 }
1224
1225 /**
1226  * ipath_mmap - mmap various structures into user space
1227  * @fp: the file pointer
1228  * @vma: the VM area
1229  *
1230  * We use this to have a shared buffer between the kernel and the user code
1231  * for the rcvhdr queue, egr buffers, and the per-port user regs and pio
1232  * buffers in the chip.  We have the open and close entries so we can bump
1233  * the ref count and keep the driver from being unloaded while still mapped.
1234  */
1235 static int ipath_mmap(struct file *fp, struct vm_area_struct *vma)
1236 {
1237         struct ipath_portdata *pd;
1238         struct ipath_devdata *dd;
1239         u64 pgaddr, ureg;
1240         unsigned piobufs, piocnt;
1241         int ret;
1242
1243         pd = port_fp(fp);
1244         if (!pd) {
1245                 ret = -EINVAL;
1246                 goto bail;
1247         }
1248         dd = pd->port_dd;
1249
1250         /*
1251          * This is the ipath_do_user_init() code, mapping the shared buffers
1252          * into the user process. The address referred to by vm_pgoff is the
1253          * file offset passed via mmap().  For shared ports, this is the
1254          * kernel vmalloc() address of the pages to share with the master.
1255          * For non-shared or master ports, this is a physical address.
1256          * We only do one mmap for each space mapped.
1257          */
1258         pgaddr = vma->vm_pgoff << PAGE_SHIFT;
1259
1260         /*
1261          * Check for 0 in case one of the allocations failed, but user
1262          * called mmap anyway.
1263          */
1264         if (!pgaddr)  {
1265                 ret = -EINVAL;
1266                 goto bail;
1267         }
1268
1269         ipath_cdbg(MM, "pgaddr %llx vm_start=%lx len %lx port %u:%u:%u\n",
1270                    (unsigned long long) pgaddr, vma->vm_start,
1271                    vma->vm_end - vma->vm_start, dd->ipath_unit,
1272                    pd->port_port, subport_fp(fp));
1273
1274         /*
1275          * Physical addresses must fit in 40 bits for our hardware.
1276          * Check for kernel virtual addresses first, anything else must
1277          * match a HW or memory address.
1278          */
1279         ret = mmap_kvaddr(vma, pgaddr, pd, subport_fp(fp));
1280         if (ret) {
1281                 if (ret > 0)
1282                         ret = 0;
1283                 goto bail;
1284         }
1285
1286         ureg = dd->ipath_uregbase + dd->ipath_palign * pd->port_port;
1287         if (!pd->port_subport_cnt) {
1288                 /* port is not shared */
1289                 piocnt = dd->ipath_pbufsport;
1290                 piobufs = pd->port_piobufs;
1291         } else if (!subport_fp(fp)) {
1292                 /* caller is the master */
1293                 piocnt = (dd->ipath_pbufsport / pd->port_subport_cnt) +
1294                          (dd->ipath_pbufsport % pd->port_subport_cnt);
1295                 piobufs = pd->port_piobufs +
1296                         dd->ipath_palign * (dd->ipath_pbufsport - piocnt);
1297         } else {
1298                 unsigned slave = subport_fp(fp) - 1;
1299
1300                 /* caller is a slave */
1301                 piocnt = dd->ipath_pbufsport / pd->port_subport_cnt;
1302                 piobufs = pd->port_piobufs + dd->ipath_palign * piocnt * slave;
1303         }
1304
1305         if (pgaddr == ureg)
1306                 ret = mmap_ureg(vma, dd, ureg);
1307         else if (pgaddr == piobufs)
1308                 ret = mmap_piobufs(vma, dd, pd, piobufs, piocnt);
1309         else if (pgaddr == dd->ipath_pioavailregs_phys)
1310                 /* in-memory copy of pioavail registers */
1311                 ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1312                                      (void *) dd->ipath_pioavailregs_dma,
1313                                      "pioavail registers");
1314         else if (pgaddr == pd->port_rcvegr_phys)
1315                 ret = mmap_rcvegrbufs(vma, pd);
1316         else if (pgaddr == (u64) pd->port_rcvhdrq_phys)
1317                 /*
1318                  * The rcvhdrq itself; readonly except on HT (so have
1319                  * to allow writable mapping), multiple pages, contiguous
1320                  * from an i/o perspective.
1321                  */
1322                 ret = ipath_mmap_mem(vma, pd, pd->port_rcvhdrq_size, 1,
1323                                      pd->port_rcvhdrq,
1324                                      "rcvhdrq");
1325         else if (pgaddr == (u64) pd->port_rcvhdrqtailaddr_phys)
1326                 /* in-memory copy of rcvhdrq tail register */
1327                 ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1328                                      pd->port_rcvhdrtail_kvaddr,
1329                                      "rcvhdrq tail");
1330         else
1331                 ret = -EINVAL;
1332
1333         vma->vm_private_data = NULL;
1334
1335         if (ret < 0)
1336                 dev_info(&dd->pcidev->dev,
1337                          "Failure %d on off %llx len %lx\n",
1338                          -ret, (unsigned long long)pgaddr,
1339                          vma->vm_end - vma->vm_start);
1340 bail:
1341         return ret;
1342 }
1343
1344 static unsigned ipath_poll_hdrqfull(struct ipath_portdata *pd)
1345 {
1346         unsigned pollflag = 0;
1347
1348         if ((pd->poll_type & IPATH_POLL_TYPE_OVERFLOW) &&
1349             pd->port_hdrqfull != pd->port_hdrqfull_poll) {
1350                 pollflag |= POLLIN | POLLRDNORM;
1351                 pd->port_hdrqfull_poll = pd->port_hdrqfull;
1352         }
1353
1354         return pollflag;
1355 }
1356
1357 static unsigned int ipath_poll_urgent(struct ipath_portdata *pd,
1358                                       struct file *fp,
1359                                       struct poll_table_struct *pt)
1360 {
1361         unsigned pollflag = 0;
1362         struct ipath_devdata *dd;
1363
1364         dd = pd->port_dd;
1365
1366         /* variable access in ipath_poll_hdrqfull() needs this */
1367         rmb();
1368         pollflag = ipath_poll_hdrqfull(pd);
1369
1370         if (pd->port_urgent != pd->port_urgent_poll) {
1371                 pollflag |= POLLIN | POLLRDNORM;
1372                 pd->port_urgent_poll = pd->port_urgent;
1373         }
1374
1375         if (!pollflag) {
1376                 /* this saves a spin_lock/unlock in interrupt handler... */
1377                 set_bit(IPATH_PORT_WAITING_URG, &pd->port_flag);
1378                 /* flush waiting flag so don't miss an event... */
1379                 wmb();
1380                 poll_wait(fp, &pd->port_wait, pt);
1381         }
1382
1383         return pollflag;
1384 }
1385
1386 static unsigned int ipath_poll_next(struct ipath_portdata *pd,
1387                                     struct file *fp,
1388                                     struct poll_table_struct *pt)
1389 {
1390         u32 head;
1391         u32 tail;
1392         unsigned pollflag = 0;
1393         struct ipath_devdata *dd;
1394
1395         dd = pd->port_dd;
1396
1397         /* variable access in ipath_poll_hdrqfull() needs this */
1398         rmb();
1399         pollflag = ipath_poll_hdrqfull(pd);
1400
1401         head = ipath_read_ureg32(dd, ur_rcvhdrhead, pd->port_port);
1402         tail = *(volatile u64 *)pd->port_rcvhdrtail_kvaddr;
1403
1404         if (head != tail)
1405                 pollflag |= POLLIN | POLLRDNORM;
1406         else {
1407                 /* this saves a spin_lock/unlock in interrupt handler */
1408                 set_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag);
1409                 /* flush waiting flag so we don't miss an event */
1410                 wmb();
1411
1412                 set_bit(pd->port_port + INFINIPATH_R_INTRAVAIL_SHIFT,
1413                         &dd->ipath_rcvctrl);
1414
1415                 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1416                                  dd->ipath_rcvctrl);
1417
1418                 if (dd->ipath_rhdrhead_intr_off) /* arm rcv interrupt */
1419                         ipath_write_ureg(dd, ur_rcvhdrhead,
1420                                          dd->ipath_rhdrhead_intr_off | head,
1421                                          pd->port_port);
1422
1423                 poll_wait(fp, &pd->port_wait, pt);
1424         }
1425
1426         return pollflag;
1427 }
1428
1429 static unsigned int ipath_poll(struct file *fp,
1430                                struct poll_table_struct *pt)
1431 {
1432         struct ipath_portdata *pd;
1433         unsigned pollflag;
1434
1435         pd = port_fp(fp);
1436         if (!pd)
1437                 pollflag = 0;
1438         else if (pd->poll_type & IPATH_POLL_TYPE_URGENT)
1439                 pollflag = ipath_poll_urgent(pd, fp, pt);
1440         else
1441                 pollflag = ipath_poll_next(pd, fp, pt);
1442
1443         return pollflag;
1444 }
1445
1446 static int ipath_supports_subports(int user_swmajor, int user_swminor)
1447 {
1448         /* no subport implementation prior to software version 1.3 */
1449         return (user_swmajor > 1) || (user_swminor >= 3);
1450 }
1451
1452 static int ipath_compatible_subports(int user_swmajor, int user_swminor)
1453 {
1454         /* this code is written long-hand for clarity */
1455         if (IPATH_USER_SWMAJOR != user_swmajor) {
1456                 /* no promise of compatibility if major mismatch */
1457                 return 0;
1458         }
1459         if (IPATH_USER_SWMAJOR == 1) {
1460                 switch (IPATH_USER_SWMINOR) {
1461                 case 0:
1462                 case 1:
1463                 case 2:
1464                         /* no subport implementation so cannot be compatible */
1465                         return 0;
1466                 case 3:
1467                         /* 3 is only compatible with itself */
1468                         return user_swminor == 3;
1469                 default:
1470                         /* >= 4 are compatible (or are expected to be) */
1471                         return user_swminor >= 4;
1472                 }
1473         }
1474         /* make no promises yet for future major versions */
1475         return 0;
1476 }
1477
1478 static int init_subports(struct ipath_devdata *dd,
1479                          struct ipath_portdata *pd,
1480                          const struct ipath_user_info *uinfo)
1481 {
1482         int ret = 0;
1483         unsigned num_subports;
1484         size_t size;
1485
1486         /*
1487          * If the user is requesting zero subports,
1488          * skip the subport allocation.
1489          */
1490         if (uinfo->spu_subport_cnt <= 0)
1491                 goto bail;
1492
1493         /* Self-consistency check for ipath_compatible_subports() */
1494         if (ipath_supports_subports(IPATH_USER_SWMAJOR, IPATH_USER_SWMINOR) &&
1495             !ipath_compatible_subports(IPATH_USER_SWMAJOR,
1496                                        IPATH_USER_SWMINOR)) {
1497                 dev_info(&dd->pcidev->dev,
1498                          "Inconsistent ipath_compatible_subports()\n");
1499                 goto bail;
1500         }
1501
1502         /* Check for subport compatibility */
1503         if (!ipath_compatible_subports(uinfo->spu_userversion >> 16,
1504                                        uinfo->spu_userversion & 0xffff)) {
1505                 dev_info(&dd->pcidev->dev,
1506                          "Mismatched user version (%d.%d) and driver "
1507                          "version (%d.%d) while port sharing. Ensure "
1508                          "that driver and library are from the same "
1509                          "release.\n",
1510                          (int) (uinfo->spu_userversion >> 16),
1511                          (int) (uinfo->spu_userversion & 0xffff),
1512                          IPATH_USER_SWMAJOR,
1513                          IPATH_USER_SWMINOR);
1514                 goto bail;
1515         }
1516         if (uinfo->spu_subport_cnt > INFINIPATH_MAX_SUBPORT) {
1517                 ret = -EINVAL;
1518                 goto bail;
1519         }
1520
1521         num_subports = uinfo->spu_subport_cnt;
1522         pd->subport_uregbase = vmalloc(PAGE_SIZE * num_subports);
1523         if (!pd->subport_uregbase) {
1524                 ret = -ENOMEM;
1525                 goto bail;
1526         }
1527         /* Note: pd->port_rcvhdrq_size isn't initialized yet. */
1528         size = ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize *
1529                      sizeof(u32), PAGE_SIZE) * num_subports;
1530         pd->subport_rcvhdr_base = vmalloc(size);
1531         if (!pd->subport_rcvhdr_base) {
1532                 ret = -ENOMEM;
1533                 goto bail_ureg;
1534         }
1535
1536         pd->subport_rcvegrbuf = vmalloc(pd->port_rcvegrbuf_chunks *
1537                                         pd->port_rcvegrbuf_size *
1538                                         num_subports);
1539         if (!pd->subport_rcvegrbuf) {
1540                 ret = -ENOMEM;
1541                 goto bail_rhdr;
1542         }
1543
1544         pd->port_subport_cnt = uinfo->spu_subport_cnt;
1545         pd->port_subport_id = uinfo->spu_subport_id;
1546         pd->active_slaves = 1;
1547         set_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag);
1548         memset(pd->subport_uregbase, 0, PAGE_SIZE * num_subports);
1549         memset(pd->subport_rcvhdr_base, 0, size);
1550         memset(pd->subport_rcvegrbuf, 0, pd->port_rcvegrbuf_chunks *
1551                                          pd->port_rcvegrbuf_size *
1552                                          num_subports);
1553         goto bail;
1554
1555 bail_rhdr:
1556         vfree(pd->subport_rcvhdr_base);
1557 bail_ureg:
1558         vfree(pd->subport_uregbase);
1559         pd->subport_uregbase = NULL;
1560 bail:
1561         return ret;
1562 }
1563
1564 static int try_alloc_port(struct ipath_devdata *dd, int port,
1565                           struct file *fp,
1566                           const struct ipath_user_info *uinfo)
1567 {
1568         struct ipath_portdata *pd;
1569         int ret;
1570
1571         if (!(pd = dd->ipath_pd[port])) {
1572                 void *ptmp;
1573
1574                 pd = kzalloc(sizeof(struct ipath_portdata), GFP_KERNEL);
1575
1576                 /*
1577                  * Allocate memory for use in ipath_tid_update() just once
1578                  * at open, not per call.  Reduces cost of expected send
1579                  * setup.
1580                  */
1581                 ptmp = kmalloc(dd->ipath_rcvtidcnt * sizeof(u16) +
1582                                dd->ipath_rcvtidcnt * sizeof(struct page **),
1583                                GFP_KERNEL);
1584                 if (!pd || !ptmp) {
1585                         ipath_dev_err(dd, "Unable to allocate portdata "
1586                                       "memory, failing open\n");
1587                         ret = -ENOMEM;
1588                         kfree(pd);
1589                         kfree(ptmp);
1590                         goto bail;
1591                 }
1592                 dd->ipath_pd[port] = pd;
1593                 dd->ipath_pd[port]->port_port = port;
1594                 dd->ipath_pd[port]->port_dd = dd;
1595                 dd->ipath_pd[port]->port_tid_pg_list = ptmp;
1596                 init_waitqueue_head(&dd->ipath_pd[port]->port_wait);
1597         }
1598         if (!pd->port_cnt) {
1599                 pd->userversion = uinfo->spu_userversion;
1600                 init_user_egr_sizes(pd);
1601                 if ((ret = init_subports(dd, pd, uinfo)) != 0)
1602                         goto bail;
1603                 ipath_cdbg(PROC, "%s[%u] opened unit:port %u:%u\n",
1604                            current->comm, current->pid, dd->ipath_unit,
1605                            port);
1606                 pd->port_cnt = 1;
1607                 port_fp(fp) = pd;
1608                 pd->port_pid = current->pid;
1609                 strncpy(pd->port_comm, current->comm, sizeof(pd->port_comm));
1610                 ipath_stats.sps_ports++;
1611                 ret = 0;
1612         } else
1613                 ret = -EBUSY;
1614
1615 bail:
1616         return ret;
1617 }
1618
1619 static inline int usable(struct ipath_devdata *dd)
1620 {
1621         return dd &&
1622                 (dd->ipath_flags & IPATH_PRESENT) &&
1623                 dd->ipath_kregbase &&
1624                 dd->ipath_lid &&
1625                 !(dd->ipath_flags & (IPATH_LINKDOWN | IPATH_DISABLED
1626                                      | IPATH_LINKUNK));
1627 }
1628
1629 static int find_free_port(int unit, struct file *fp,
1630                           const struct ipath_user_info *uinfo)
1631 {
1632         struct ipath_devdata *dd = ipath_lookup(unit);
1633         int ret, i;
1634
1635         if (!dd) {
1636                 ret = -ENODEV;
1637                 goto bail;
1638         }
1639
1640         if (!usable(dd)) {
1641                 ret = -ENETDOWN;
1642                 goto bail;
1643         }
1644
1645         for (i = 1; i < dd->ipath_cfgports; i++) {
1646                 ret = try_alloc_port(dd, i, fp, uinfo);
1647                 if (ret != -EBUSY)
1648                         goto bail;
1649         }
1650         ret = -EBUSY;
1651
1652 bail:
1653         return ret;
1654 }
1655
1656 static int find_best_unit(struct file *fp,
1657                           const struct ipath_user_info *uinfo)
1658 {
1659         int ret = 0, i, prefunit = -1, devmax;
1660         int maxofallports, npresent, nup;
1661         int ndev;
1662
1663         devmax = ipath_count_units(&npresent, &nup, &maxofallports);
1664
1665         /*
1666          * This code is present to allow a knowledgeable person to
1667          * specify the layout of processes to processors before opening
1668          * this driver, and then we'll assign the process to the "closest"
1669          * InfiniPath chip to that processor (we assume reasonable connectivity,
1670          * for now).  This code assumes that if affinity has been set
1671          * before this point, that at most one cpu is set; for now this
1672          * is reasonable.  I check for both cpus_empty() and cpus_full(),
1673          * in case some kernel variant sets none of the bits when no
1674          * affinity is set.  2.6.11 and 12 kernels have all present
1675          * cpus set.  Some day we'll have to fix it up further to handle
1676          * a cpu subset.  This algorithm fails for two HT chips connected
1677          * in tunnel fashion.  Eventually this needs real topology
1678          * information.  There may be some issues with dual core numbering
1679          * as well.  This needs more work prior to release.
1680          */
1681         if (!cpus_empty(current->cpus_allowed) &&
1682             !cpus_full(current->cpus_allowed)) {
1683                 int ncpus = num_online_cpus(), curcpu = -1, nset = 0;
1684                 for (i = 0; i < ncpus; i++)
1685                         if (cpu_isset(i, current->cpus_allowed)) {
1686                                 ipath_cdbg(PROC, "%s[%u] affinity set for "
1687                                            "cpu %d/%d\n", current->comm,
1688                                            current->pid, i, ncpus);
1689                                 curcpu = i;
1690                                 nset++;
1691                         }
1692                 if (curcpu != -1 && nset != ncpus) {
1693                         if (npresent) {
1694                                 prefunit = curcpu / (ncpus / npresent);
1695                                 ipath_cdbg(PROC,"%s[%u] %d chips, %d cpus, "
1696                                           "%d cpus/chip, select unit %d\n",
1697                                           current->comm, current->pid,
1698                                           npresent, ncpus, ncpus / npresent,
1699                                           prefunit);
1700                         }
1701                 }
1702         }
1703
1704         /*
1705          * user ports start at 1, kernel port is 0
1706          * For now, we do round-robin access across all chips
1707          */
1708
1709         if (prefunit != -1)
1710                 devmax = prefunit + 1;
1711 recheck:
1712         for (i = 1; i < maxofallports; i++) {
1713                 for (ndev = prefunit != -1 ? prefunit : 0; ndev < devmax;
1714                      ndev++) {
1715                         struct ipath_devdata *dd = ipath_lookup(ndev);
1716
1717                         if (!usable(dd))
1718                                 continue; /* can't use this unit */
1719                         if (i >= dd->ipath_cfgports)
1720                                 /*
1721                                  * Maxed out on users of this unit. Try
1722                                  * next.
1723                                  */
1724                                 continue;
1725                         ret = try_alloc_port(dd, i, fp, uinfo);
1726                         if (!ret)
1727                                 goto done;
1728                 }
1729         }
1730
1731         if (npresent) {
1732                 if (nup == 0) {
1733                         ret = -ENETDOWN;
1734                         ipath_dbg("No ports available (none initialized "
1735                                   "and ready)\n");
1736                 } else {
1737                         if (prefunit > 0) {
1738                                 /* if started above 0, retry from 0 */
1739                                 ipath_cdbg(PROC,
1740                                            "%s[%u] no ports on prefunit "
1741                                            "%d, clear and re-check\n",
1742                                            current->comm, current->pid,
1743                                            prefunit);
1744                                 devmax = ipath_count_units(NULL, NULL,
1745                                                            NULL);
1746                                 prefunit = -1;
1747                                 goto recheck;
1748                         }
1749                         ret = -EBUSY;
1750                         ipath_dbg("No ports available\n");
1751                 }
1752         } else {
1753                 ret = -ENXIO;
1754                 ipath_dbg("No boards found\n");
1755         }
1756
1757 done:
1758         return ret;
1759 }
1760
1761 static int find_shared_port(struct file *fp,
1762                             const struct ipath_user_info *uinfo)
1763 {
1764         int devmax, ndev, i;
1765         int ret = 0;
1766
1767         devmax = ipath_count_units(NULL, NULL, NULL);
1768
1769         for (ndev = 0; ndev < devmax; ndev++) {
1770                 struct ipath_devdata *dd = ipath_lookup(ndev);
1771
1772                 if (!dd)
1773                         continue;
1774                 for (i = 1; i < dd->ipath_cfgports; i++) {
1775                         struct ipath_portdata *pd = dd->ipath_pd[i];
1776
1777                         /* Skip ports which are not yet open */
1778                         if (!pd || !pd->port_cnt)
1779                                 continue;
1780                         /* Skip port if it doesn't match the requested one */
1781                         if (pd->port_subport_id != uinfo->spu_subport_id)
1782                                 continue;
1783                         /* Verify the sharing process matches the master */
1784                         if (pd->port_subport_cnt != uinfo->spu_subport_cnt ||
1785                             pd->userversion != uinfo->spu_userversion ||
1786                             pd->port_cnt >= pd->port_subport_cnt) {
1787                                 ret = -EINVAL;
1788                                 goto done;
1789                         }
1790                         port_fp(fp) = pd;
1791                         subport_fp(fp) = pd->port_cnt++;
1792                         tidcursor_fp(fp) = 0;
1793                         pd->active_slaves |= 1 << subport_fp(fp);
1794                         ipath_cdbg(PROC,
1795                                    "%s[%u] %u sharing %s[%u] unit:port %u:%u\n",
1796                                    current->comm, current->pid,
1797                                    subport_fp(fp),
1798                                    pd->port_comm, pd->port_pid,
1799                                    dd->ipath_unit, pd->port_port);
1800                         ret = 1;
1801                         goto done;
1802                 }
1803         }
1804
1805 done:
1806         return ret;
1807 }
1808
1809 static int ipath_open(struct inode *in, struct file *fp)
1810 {
1811         /* The real work is performed later in ipath_assign_port() */
1812         fp->private_data = kzalloc(sizeof(struct ipath_filedata), GFP_KERNEL);
1813         return fp->private_data ? 0 : -ENOMEM;
1814 }
1815
1816 /* Get port early, so can set affinity prior to memory allocation */
1817 static int ipath_assign_port(struct file *fp,
1818                               const struct ipath_user_info *uinfo)
1819 {
1820         int ret;
1821         int i_minor;
1822         unsigned swmajor, swminor;
1823
1824         /* Check to be sure we haven't already initialized this file */
1825         if (port_fp(fp)) {
1826                 ret = -EINVAL;
1827                 goto done;
1828         }
1829
1830         /* for now, if major version is different, bail */
1831         swmajor = uinfo->spu_userversion >> 16;
1832         if (swmajor != IPATH_USER_SWMAJOR) {
1833                 ipath_dbg("User major version %d not same as driver "
1834                           "major %d\n", uinfo->spu_userversion >> 16,
1835                           IPATH_USER_SWMAJOR);
1836                 ret = -ENODEV;
1837                 goto done;
1838         }
1839
1840         swminor = uinfo->spu_userversion & 0xffff;
1841         if (swminor != IPATH_USER_SWMINOR)
1842                 ipath_dbg("User minor version %d not same as driver "
1843                           "minor %d\n", swminor, IPATH_USER_SWMINOR);
1844
1845         mutex_lock(&ipath_mutex);
1846
1847         if (ipath_compatible_subports(swmajor, swminor) &&
1848             uinfo->spu_subport_cnt &&
1849             (ret = find_shared_port(fp, uinfo))) {
1850                 mutex_unlock(&ipath_mutex);
1851                 if (ret > 0)
1852                         ret = 0;
1853                 goto done;
1854         }
1855
1856         i_minor = iminor(fp->f_path.dentry->d_inode) - IPATH_USER_MINOR_BASE;
1857         ipath_cdbg(VERBOSE, "open on dev %lx (minor %d)\n",
1858                    (long)fp->f_path.dentry->d_inode->i_rdev, i_minor);
1859
1860         if (i_minor)
1861                 ret = find_free_port(i_minor - 1, fp, uinfo);
1862         else
1863                 ret = find_best_unit(fp, uinfo);
1864
1865         mutex_unlock(&ipath_mutex);
1866
1867 done:
1868         return ret;
1869 }
1870
1871
1872 static int ipath_do_user_init(struct file *fp,
1873                               const struct ipath_user_info *uinfo)
1874 {
1875         int ret;
1876         struct ipath_portdata *pd = port_fp(fp);
1877         struct ipath_devdata *dd;
1878         u32 head32;
1879
1880         /* Subports don't need to initialize anything since master did it. */
1881         if (subport_fp(fp)) {
1882                 ret = wait_event_interruptible(pd->port_wait,
1883                         !test_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag));
1884                 goto done;
1885         }
1886
1887         dd = pd->port_dd;
1888
1889         if (uinfo->spu_rcvhdrsize) {
1890                 ret = ipath_setrcvhdrsize(dd, uinfo->spu_rcvhdrsize);
1891                 if (ret)
1892                         goto done;
1893         }
1894
1895         /* for now we do nothing with rcvhdrcnt: uinfo->spu_rcvhdrcnt */
1896
1897         /* for right now, kernel piobufs are at end, so port 1 is at 0 */
1898         pd->port_piobufs = dd->ipath_piobufbase +
1899                 dd->ipath_pbufsport * (pd->port_port - 1) * dd->ipath_palign;
1900         ipath_cdbg(VERBOSE, "Set base of piobufs for port %u to 0x%x\n",
1901                    pd->port_port, pd->port_piobufs);
1902
1903         /*
1904          * Now allocate the rcvhdr Q and eager TIDs; skip the TID
1905          * array for time being.  If pd->port_port > chip-supported,
1906          * we need to do extra stuff here to handle by handling overflow
1907          * through port 0, someday
1908          */
1909         ret = ipath_create_rcvhdrq(dd, pd);
1910         if (!ret)
1911                 ret = ipath_create_user_egr(pd);
1912         if (ret)
1913                 goto done;
1914
1915         /*
1916          * set the eager head register for this port to the current values
1917          * of the tail pointers, since we don't know if they were
1918          * updated on last use of the port.
1919          */
1920         head32 = ipath_read_ureg32(dd, ur_rcvegrindextail, pd->port_port);
1921         ipath_write_ureg(dd, ur_rcvegrindexhead, head32, pd->port_port);
1922         dd->ipath_lastegrheads[pd->port_port] = -1;
1923         dd->ipath_lastrcvhdrqtails[pd->port_port] = -1;
1924         ipath_cdbg(VERBOSE, "Wrote port%d egrhead %x from tail regs\n",
1925                 pd->port_port, head32);
1926         pd->port_tidcursor = 0; /* start at beginning after open */
1927
1928         /* initialize poll variables... */
1929         pd->port_urgent = 0;
1930         pd->port_urgent_poll = 0;
1931         pd->port_hdrqfull_poll = pd->port_hdrqfull;
1932
1933         /*
1934          * now enable the port; the tail registers will be written to memory
1935          * by the chip as soon as it sees the write to
1936          * dd->ipath_kregs->kr_rcvctrl.  The update only happens on
1937          * transition from 0 to 1, so clear it first, then set it as part of
1938          * enabling the port.  This will (very briefly) affect any other
1939          * open ports, but it shouldn't be long enough to be an issue.
1940          * We explictly set the in-memory copy to 0 beforehand, so we don't
1941          * have to wait to be sure the DMA update has happened.
1942          */
1943         *(volatile u64 *)pd->port_rcvhdrtail_kvaddr = 0ULL;
1944         set_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port,
1945                 &dd->ipath_rcvctrl);
1946         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1947                          dd->ipath_rcvctrl & ~INFINIPATH_R_TAILUPD);
1948         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1949                          dd->ipath_rcvctrl);
1950         /* Notify any waiting slaves */
1951         if (pd->port_subport_cnt) {
1952                 clear_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag);
1953                 wake_up(&pd->port_wait);
1954         }
1955 done:
1956         return ret;
1957 }
1958
1959 /**
1960  * unlock_exptid - unlock any expected TID entries port still had in use
1961  * @pd: port
1962  *
1963  * We don't actually update the chip here, because we do a bulk update
1964  * below, using ipath_f_clear_tids.
1965  */
1966 static void unlock_expected_tids(struct ipath_portdata *pd)
1967 {
1968         struct ipath_devdata *dd = pd->port_dd;
1969         int port_tidbase = pd->port_port * dd->ipath_rcvtidcnt;
1970         int i, cnt = 0, maxtid = port_tidbase + dd->ipath_rcvtidcnt;
1971
1972         ipath_cdbg(VERBOSE, "Port %u unlocking any locked expTID pages\n",
1973                    pd->port_port);
1974         for (i = port_tidbase; i < maxtid; i++) {
1975                 if (!dd->ipath_pageshadow[i])
1976                         continue;
1977
1978                 pci_unmap_page(dd->pcidev, dd->ipath_physshadow[i],
1979                         PAGE_SIZE, PCI_DMA_FROMDEVICE);
1980                 ipath_release_user_pages_on_close(&dd->ipath_pageshadow[i],
1981                                                   1);
1982                 dd->ipath_pageshadow[i] = NULL;
1983                 cnt++;
1984                 ipath_stats.sps_pageunlocks++;
1985         }
1986         if (cnt)
1987                 ipath_cdbg(VERBOSE, "Port %u locked %u expTID entries\n",
1988                            pd->port_port, cnt);
1989
1990         if (ipath_stats.sps_pagelocks || ipath_stats.sps_pageunlocks)
1991                 ipath_cdbg(VERBOSE, "%llu pages locked, %llu unlocked\n",
1992                            (unsigned long long) ipath_stats.sps_pagelocks,
1993                            (unsigned long long)
1994                            ipath_stats.sps_pageunlocks);
1995 }
1996
1997 static int ipath_close(struct inode *in, struct file *fp)
1998 {
1999         int ret = 0;
2000         struct ipath_filedata *fd;
2001         struct ipath_portdata *pd;
2002         struct ipath_devdata *dd;
2003         unsigned port;
2004
2005         ipath_cdbg(VERBOSE, "close on dev %lx, private data %p\n",
2006                    (long)in->i_rdev, fp->private_data);
2007
2008         mutex_lock(&ipath_mutex);
2009
2010         fd = (struct ipath_filedata *) fp->private_data;
2011         fp->private_data = NULL;
2012         pd = fd->pd;
2013         if (!pd) {
2014                 mutex_unlock(&ipath_mutex);
2015                 goto bail;
2016         }
2017         if (--pd->port_cnt) {
2018                 /*
2019                  * XXX If the master closes the port before the slave(s),
2020                  * revoke the mmap for the eager receive queue so
2021                  * the slave(s) don't wait for receive data forever.
2022                  */
2023                 pd->active_slaves &= ~(1 << fd->subport);
2024                 mutex_unlock(&ipath_mutex);
2025                 goto bail;
2026         }
2027         port = pd->port_port;
2028         dd = pd->port_dd;
2029
2030         if (pd->port_hdrqfull) {
2031                 ipath_cdbg(PROC, "%s[%u] had %u rcvhdrqfull errors "
2032                            "during run\n", pd->port_comm, pd->port_pid,
2033                            pd->port_hdrqfull);
2034                 pd->port_hdrqfull = 0;
2035         }
2036
2037         if (pd->port_rcvwait_to || pd->port_piowait_to
2038             || pd->port_rcvnowait || pd->port_pionowait) {
2039                 ipath_cdbg(VERBOSE, "port%u, %u rcv, %u pio wait timeo; "
2040                            "%u rcv %u, pio already\n",
2041                            pd->port_port, pd->port_rcvwait_to,
2042                            pd->port_piowait_to, pd->port_rcvnowait,
2043                            pd->port_pionowait);
2044                 pd->port_rcvwait_to = pd->port_piowait_to =
2045                         pd->port_rcvnowait = pd->port_pionowait = 0;
2046         }
2047         if (pd->port_flag) {
2048                 ipath_dbg("port %u port_flag still set to 0x%lx\n",
2049                           pd->port_port, pd->port_flag);
2050                 pd->port_flag = 0;
2051         }
2052
2053         if (dd->ipath_kregbase) {
2054                 int i;
2055                 /* atomically clear receive enable port and intr avail. */
2056                 clear_bit(INFINIPATH_R_PORTENABLE_SHIFT + port,
2057                           &dd->ipath_rcvctrl);
2058                 clear_bit(pd->port_port + INFINIPATH_R_INTRAVAIL_SHIFT,
2059                           &dd->ipath_rcvctrl);
2060                 ipath_write_kreg( dd, dd->ipath_kregs->kr_rcvctrl,
2061                         dd->ipath_rcvctrl);
2062                 /* and read back from chip to be sure that nothing
2063                  * else is in flight when we do the rest */
2064                 (void)ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
2065
2066                 /* clean up the pkeys for this port user */
2067                 ipath_clean_part_key(pd, dd);
2068                 /*
2069                  * be paranoid, and never write 0's to these, just use an
2070                  * unused part of the port 0 tail page.  Of course,
2071                  * rcvhdraddr points to a large chunk of memory, so this
2072                  * could still trash things, but at least it won't trash
2073                  * page 0, and by disabling the port, it should stop "soon",
2074                  * even if a packet or two is in already in flight after we
2075                  * disabled the port.
2076                  */
2077                 ipath_write_kreg_port(dd,
2078                         dd->ipath_kregs->kr_rcvhdrtailaddr, port,
2079                         dd->ipath_dummy_hdrq_phys);
2080                 ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdraddr,
2081                         pd->port_port, dd->ipath_dummy_hdrq_phys);
2082
2083                 i = dd->ipath_pbufsport * (port - 1);
2084                 ipath_disarm_piobufs(dd, i, dd->ipath_pbufsport);
2085
2086                 dd->ipath_f_clear_tids(dd, pd->port_port);
2087
2088                 if (dd->ipath_pageshadow)
2089                         unlock_expected_tids(pd);
2090                 ipath_stats.sps_ports--;
2091                 ipath_cdbg(PROC, "%s[%u] closed port %u:%u\n",
2092                            pd->port_comm, pd->port_pid,
2093                            dd->ipath_unit, port);
2094         }
2095
2096         pd->port_pid = 0;
2097         dd->ipath_pd[pd->port_port] = NULL; /* before releasing mutex */
2098         mutex_unlock(&ipath_mutex);
2099         ipath_free_pddata(dd, pd); /* after releasing the mutex */
2100
2101 bail:
2102         kfree(fd);
2103         return ret;
2104 }
2105
2106 static int ipath_port_info(struct ipath_portdata *pd, u16 subport,
2107                            struct ipath_port_info __user *uinfo)
2108 {
2109         struct ipath_port_info info;
2110         int nup;
2111         int ret;
2112         size_t sz;
2113
2114         (void) ipath_count_units(NULL, &nup, NULL);
2115         info.num_active = nup;
2116         info.unit = pd->port_dd->ipath_unit;
2117         info.port = pd->port_port;
2118         info.subport = subport;
2119         /* Don't return new fields if old library opened the port. */
2120         if (ipath_supports_subports(pd->userversion >> 16,
2121                                     pd->userversion & 0xffff)) {
2122                 /* Number of user ports available for this device. */
2123                 info.num_ports = pd->port_dd->ipath_cfgports - 1;
2124                 info.num_subports = pd->port_subport_cnt;
2125                 sz = sizeof(info);
2126         } else
2127                 sz = sizeof(info) - 2 * sizeof(u16);
2128
2129         if (copy_to_user(uinfo, &info, sz)) {
2130                 ret = -EFAULT;
2131                 goto bail;
2132         }
2133         ret = 0;
2134
2135 bail:
2136         return ret;
2137 }
2138
2139 static int ipath_get_slave_info(struct ipath_portdata *pd,
2140                                 void __user *slave_mask_addr)
2141 {
2142         int ret = 0;
2143
2144         if (copy_to_user(slave_mask_addr, &pd->active_slaves, sizeof(u32)))
2145                 ret = -EFAULT;
2146         return ret;
2147 }
2148
2149 static int ipath_force_pio_avail_update(struct ipath_devdata *dd)
2150 {
2151         u64 reg = dd->ipath_sendctrl;
2152
2153         clear_bit(IPATH_S_PIOBUFAVAILUPD, &reg);
2154         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, reg);
2155         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, dd->ipath_sendctrl);
2156
2157         return 0;
2158 }
2159
2160 static ssize_t ipath_write(struct file *fp, const char __user *data,
2161                            size_t count, loff_t *off)
2162 {
2163         const struct ipath_cmd __user *ucmd;
2164         struct ipath_portdata *pd;
2165         const void __user *src;
2166         size_t consumed, copy;
2167         struct ipath_cmd cmd;
2168         ssize_t ret = 0;
2169         void *dest;
2170
2171         if (count < sizeof(cmd.type)) {
2172                 ret = -EINVAL;
2173                 goto bail;
2174         }
2175
2176         ucmd = (const struct ipath_cmd __user *) data;
2177
2178         if (copy_from_user(&cmd.type, &ucmd->type, sizeof(cmd.type))) {
2179                 ret = -EFAULT;
2180                 goto bail;
2181         }
2182
2183         consumed = sizeof(cmd.type);
2184
2185         switch (cmd.type) {
2186         case IPATH_CMD_ASSIGN_PORT:
2187         case __IPATH_CMD_USER_INIT:
2188         case IPATH_CMD_USER_INIT:
2189                 copy = sizeof(cmd.cmd.user_info);
2190                 dest = &cmd.cmd.user_info;
2191                 src = &ucmd->cmd.user_info;
2192                 break;
2193         case IPATH_CMD_RECV_CTRL:
2194                 copy = sizeof(cmd.cmd.recv_ctrl);
2195                 dest = &cmd.cmd.recv_ctrl;
2196                 src = &ucmd->cmd.recv_ctrl;
2197                 break;
2198         case IPATH_CMD_PORT_INFO:
2199                 copy = sizeof(cmd.cmd.port_info);
2200                 dest = &cmd.cmd.port_info;
2201                 src = &ucmd->cmd.port_info;
2202                 break;
2203         case IPATH_CMD_TID_UPDATE:
2204         case IPATH_CMD_TID_FREE:
2205                 copy = sizeof(cmd.cmd.tid_info);
2206                 dest = &cmd.cmd.tid_info;
2207                 src = &ucmd->cmd.tid_info;
2208                 break;
2209         case IPATH_CMD_SET_PART_KEY:
2210                 copy = sizeof(cmd.cmd.part_key);
2211                 dest = &cmd.cmd.part_key;
2212                 src = &ucmd->cmd.part_key;
2213                 break;
2214         case __IPATH_CMD_SLAVE_INFO:
2215                 copy = sizeof(cmd.cmd.slave_mask_addr);
2216                 dest = &cmd.cmd.slave_mask_addr;
2217                 src = &ucmd->cmd.slave_mask_addr;
2218                 break;
2219         case IPATH_CMD_PIOAVAILUPD:     // force an update of PIOAvail reg
2220                 copy = 0;
2221                 src = NULL;
2222                 dest = NULL;
2223                 break;
2224         case IPATH_CMD_POLL_TYPE:
2225                 copy = sizeof(cmd.cmd.poll_type);
2226                 dest = &cmd.cmd.poll_type;
2227                 src = &ucmd->cmd.poll_type;
2228                 break;
2229         default:
2230                 ret = -EINVAL;
2231                 goto bail;
2232         }
2233
2234         if (copy) {
2235                 if ((count - consumed) < copy) {
2236                         ret = -EINVAL;
2237                         goto bail;
2238                 }
2239
2240                 if (copy_from_user(dest, src, copy)) {
2241                         ret = -EFAULT;
2242                         goto bail;
2243                 }
2244
2245                 consumed += copy;
2246         }
2247
2248         pd = port_fp(fp);
2249         if (!pd && cmd.type != __IPATH_CMD_USER_INIT &&
2250                 cmd.type != IPATH_CMD_ASSIGN_PORT) {
2251                 ret = -EINVAL;
2252                 goto bail;
2253         }
2254
2255         switch (cmd.type) {
2256         case IPATH_CMD_ASSIGN_PORT:
2257                 ret = ipath_assign_port(fp, &cmd.cmd.user_info);
2258                 if (ret)
2259                         goto bail;
2260                 break;
2261         case __IPATH_CMD_USER_INIT:
2262                 /* backwards compatibility, get port first */
2263                 ret = ipath_assign_port(fp, &cmd.cmd.user_info);
2264                 if (ret)
2265                         goto bail;
2266                 /* and fall through to current version. */
2267         case IPATH_CMD_USER_INIT:
2268                 ret = ipath_do_user_init(fp, &cmd.cmd.user_info);
2269                 if (ret)
2270                         goto bail;
2271                 ret = ipath_get_base_info(
2272                         fp, (void __user *) (unsigned long)
2273                         cmd.cmd.user_info.spu_base_info,
2274                         cmd.cmd.user_info.spu_base_info_size);
2275                 break;
2276         case IPATH_CMD_RECV_CTRL:
2277                 ret = ipath_manage_rcvq(pd, subport_fp(fp), cmd.cmd.recv_ctrl);
2278                 break;
2279         case IPATH_CMD_PORT_INFO:
2280                 ret = ipath_port_info(pd, subport_fp(fp),
2281                                       (struct ipath_port_info __user *)
2282                                       (unsigned long) cmd.cmd.port_info);
2283                 break;
2284         case IPATH_CMD_TID_UPDATE:
2285                 ret = ipath_tid_update(pd, fp, &cmd.cmd.tid_info);
2286                 break;
2287         case IPATH_CMD_TID_FREE:
2288                 ret = ipath_tid_free(pd, subport_fp(fp), &cmd.cmd.tid_info);
2289                 break;
2290         case IPATH_CMD_SET_PART_KEY:
2291                 ret = ipath_set_part_key(pd, cmd.cmd.part_key);
2292                 break;
2293         case __IPATH_CMD_SLAVE_INFO:
2294                 ret = ipath_get_slave_info(pd,
2295                                            (void __user *) (unsigned long)
2296                                            cmd.cmd.slave_mask_addr);
2297                 break;
2298         case IPATH_CMD_PIOAVAILUPD:
2299                 ret = ipath_force_pio_avail_update(pd->port_dd);
2300                 break;
2301         case IPATH_CMD_POLL_TYPE:
2302                 pd->poll_type = cmd.cmd.poll_type;
2303                 break;
2304         }
2305
2306         if (ret >= 0)
2307                 ret = consumed;
2308
2309 bail:
2310         return ret;
2311 }
2312
2313 static struct class *ipath_class;
2314
2315 static int init_cdev(int minor, char *name, const struct file_operations *fops,
2316                      struct cdev **cdevp, struct class_device **class_devp)
2317 {
2318         const dev_t dev = MKDEV(IPATH_MAJOR, minor);
2319         struct cdev *cdev = NULL;
2320         struct class_device *class_dev = NULL;
2321         int ret;
2322
2323         cdev = cdev_alloc();
2324         if (!cdev) {
2325                 printk(KERN_ERR IPATH_DRV_NAME
2326                        ": Could not allocate cdev for minor %d, %s\n",
2327                        minor, name);
2328                 ret = -ENOMEM;
2329                 goto done;
2330         }
2331
2332         cdev->owner = THIS_MODULE;
2333         cdev->ops = fops;
2334         kobject_set_name(&cdev->kobj, name);
2335
2336         ret = cdev_add(cdev, dev, 1);
2337         if (ret < 0) {
2338                 printk(KERN_ERR IPATH_DRV_NAME
2339                        ": Could not add cdev for minor %d, %s (err %d)\n",
2340                        minor, name, -ret);
2341                 goto err_cdev;
2342         }
2343
2344         class_dev = class_device_create(ipath_class, NULL, dev, NULL, name);
2345
2346         if (IS_ERR(class_dev)) {
2347                 ret = PTR_ERR(class_dev);
2348                 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2349                        "class_dev for minor %d, %s (err %d)\n",
2350                        minor, name, -ret);
2351                 goto err_cdev;
2352         }
2353
2354         goto done;
2355
2356 err_cdev:
2357         cdev_del(cdev);
2358         cdev = NULL;
2359
2360 done:
2361         if (ret >= 0) {
2362                 *cdevp = cdev;
2363                 *class_devp = class_dev;
2364         } else {
2365                 *cdevp = NULL;
2366                 *class_devp = NULL;
2367         }
2368
2369         return ret;
2370 }
2371
2372 int ipath_cdev_init(int minor, char *name, const struct file_operations *fops,
2373                     struct cdev **cdevp, struct class_device **class_devp)
2374 {
2375         return init_cdev(minor, name, fops, cdevp, class_devp);
2376 }
2377
2378 static void cleanup_cdev(struct cdev **cdevp,
2379                          struct class_device **class_devp)
2380 {
2381         struct class_device *class_dev = *class_devp;
2382
2383         if (class_dev) {
2384                 class_device_unregister(class_dev);
2385                 *class_devp = NULL;
2386         }
2387
2388         if (*cdevp) {
2389                 cdev_del(*cdevp);
2390                 *cdevp = NULL;
2391         }
2392 }
2393
2394 void ipath_cdev_cleanup(struct cdev **cdevp,
2395                         struct class_device **class_devp)
2396 {
2397         cleanup_cdev(cdevp, class_devp);
2398 }
2399
2400 static struct cdev *wildcard_cdev;
2401 static struct class_device *wildcard_class_dev;
2402
2403 static const dev_t dev = MKDEV(IPATH_MAJOR, 0);
2404
2405 static int user_init(void)
2406 {
2407         int ret;
2408
2409         ret = register_chrdev_region(dev, IPATH_NMINORS, IPATH_DRV_NAME);
2410         if (ret < 0) {
2411                 printk(KERN_ERR IPATH_DRV_NAME ": Could not register "
2412                        "chrdev region (err %d)\n", -ret);
2413                 goto done;
2414         }
2415
2416         ipath_class = class_create(THIS_MODULE, IPATH_DRV_NAME);
2417
2418         if (IS_ERR(ipath_class)) {
2419                 ret = PTR_ERR(ipath_class);
2420                 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2421                        "device class (err %d)\n", -ret);
2422                 goto bail;
2423         }
2424
2425         goto done;
2426 bail:
2427         unregister_chrdev_region(dev, IPATH_NMINORS);
2428 done:
2429         return ret;
2430 }
2431
2432 static void user_cleanup(void)
2433 {
2434         if (ipath_class) {
2435                 class_destroy(ipath_class);
2436                 ipath_class = NULL;
2437         }
2438
2439         unregister_chrdev_region(dev, IPATH_NMINORS);
2440 }
2441
2442 static atomic_t user_count = ATOMIC_INIT(0);
2443 static atomic_t user_setup = ATOMIC_INIT(0);
2444
2445 int ipath_user_add(struct ipath_devdata *dd)
2446 {
2447         char name[10];
2448         int ret;
2449
2450         if (atomic_inc_return(&user_count) == 1) {
2451                 ret = user_init();
2452                 if (ret < 0) {
2453                         ipath_dev_err(dd, "Unable to set up user support: "
2454                                       "error %d\n", -ret);
2455                         goto bail;
2456                 }
2457                 ret = init_cdev(0, "ipath", &ipath_file_ops, &wildcard_cdev,
2458                                 &wildcard_class_dev);
2459                 if (ret < 0) {
2460                         ipath_dev_err(dd, "Could not create wildcard "
2461                                       "minor: error %d\n", -ret);
2462                         goto bail_user;
2463                 }
2464
2465                 atomic_set(&user_setup, 1);
2466         }
2467
2468         snprintf(name, sizeof(name), "ipath%d", dd->ipath_unit);
2469
2470         ret = init_cdev(dd->ipath_unit + 1, name, &ipath_file_ops,
2471                         &dd->user_cdev, &dd->user_class_dev);
2472         if (ret < 0)
2473                 ipath_dev_err(dd, "Could not create user minor %d, %s\n",
2474                               dd->ipath_unit + 1, name);
2475
2476         goto bail;
2477
2478 bail_user:
2479         user_cleanup();
2480 bail:
2481         return ret;
2482 }
2483
2484 void ipath_user_remove(struct ipath_devdata *dd)
2485 {
2486         cleanup_cdev(&dd->user_cdev, &dd->user_class_dev);
2487
2488         if (atomic_dec_return(&user_count) == 0) {
2489                 if (atomic_read(&user_setup) == 0)
2490                         goto bail;
2491
2492                 cleanup_cdev(&wildcard_cdev, &wildcard_class_dev);
2493                 user_cleanup();
2494
2495                 atomic_set(&user_setup, 0);
2496         }
2497 bail:
2498         return;
2499 }