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_ureg_align * 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_ureg_align * 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         /*
 223          * user mode PIO buffers are always 2KB, even when 4KB can
 224          * be received, and sent via the kernel; this is ibmaxlen
 225          * for 2K MTU.
 226          */
 227         kinfo->spi_piosize = dd->ipath_piosize2k - 2 * sizeof(u32);
 228         kinfo->spi_mtu = dd->ipath_ibmaxlen;    /* maxlen, not ibmtu */
 229         kinfo->spi_port = pd->port_port;
 230         kinfo->spi_subport = subport_fp(fp);
 231         kinfo->spi_sw_version = IPATH_KERN_SWVERSION;
 232         kinfo->spi_hw_version = dd->ipath_revision;
 233
 234         if (master) {
 235                 kinfo->spi_runtime_flags |= IPATH_RUNTIME_MASTER;
 236         }
 237
 238         sz = (ubase_size < sizeof(*kinfo)) ? ubase_size : sizeof(*kinfo);
 239         if (copy_to_user(ubase, kinfo, sz))
 240                 ret = -EFAULT;
 241
 242 bail:
 243         kfree(kinfo);
 244         return ret;
 245 }
 246
 247 /**
 248  * ipath_tid_update - update a port TID
 249  * @pd: the port
 250  * @fp: the ipath device file
 251  * @ti: the TID information
 252  *
 253  * The new implementation as of Oct 2004 is that the driver assigns
 254  * the tid and returns it to the caller.   To make it easier to
 255  * catch bugs, and to reduce search time, we keep a cursor for
 256  * each port, walking the shadow tid array to find one that's not
 257  * in use.
 258  *
 259  * For now, if we can't allocate the full list, we fail, although
 260  * in the long run, we'll allocate as many as we can, and the
 261  * caller will deal with that by trying the remaining pages later.
 262  * That means that when we fail, we have to mark the tids as not in
 263  * use again, in our shadow copy.
 264  *
 265  * It's up to the caller to free the tids when they are done.
 266  * We'll unlock the pages as they free them.
 267  *
 268  * Also, right now we are locking one page at a time, but since
 269  * the intended use of this routine is for a single group of
 270  * virtually contiguous pages, that should change to improve
 271  * performance.
 272  */
 273 static int ipath_tid_update(struct ipath_portdata *pd, struct file *fp,
 274                             const struct ipath_tid_info *ti)
 275 {
 276         int ret = 0, ntids;
 277         u32 tid, porttid, cnt, i, tidcnt, tidoff;
 278         u16 *tidlist;
 279         struct ipath_devdata *dd = pd->port_dd;
 280         u64 physaddr;
 281         unsigned long vaddr;
 282         u64 __iomem *tidbase;
 283         unsigned long tidmap[8];
 284         struct page **pagep = NULL;
 285         unsigned subport = subport_fp(fp);
 286
 287         if (!dd->ipath_pageshadow) {
 288                 ret = -ENOMEM;
 289                 goto done;
 290         }
 291
 292         cnt = ti->tidcnt;
 293         if (!cnt) {
 294                 ipath_dbg("After copyin, tidcnt 0, tidlist %llx\n",
 295                           (unsigned long long) ti->tidlist);
 296                 /*
 297                  * Should we treat as success?  likely a bug
 298                  */
 299                 ret = -EFAULT;
 300                 goto done;
 301         }
 302         porttid = pd->port_port * dd->ipath_rcvtidcnt;
 303         if (!pd->port_subport_cnt) {
 304                 tidcnt = dd->ipath_rcvtidcnt;
 305                 tid = pd->port_tidcursor;
 306                 tidoff = 0;
 307         } else if (!subport) {
 308                 tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
 309                          (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
 310                 tidoff = dd->ipath_rcvtidcnt - tidcnt;
 311                 porttid += tidoff;
 312                 tid = tidcursor_fp(fp);
 313         } else {
 314                 tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
 315                 tidoff = tidcnt * (subport - 1);
 316                 porttid += tidoff;
 317                 tid = tidcursor_fp(fp);
 318         }
 319         if (cnt > tidcnt) {
 320                 /* make sure it all fits in port_tid_pg_list */
 321                 dev_info(&dd->pcidev->dev, "Process tried to allocate %u "
 322                          "TIDs, only trying max (%u)\n", cnt, tidcnt);
 323                 cnt = tidcnt;
 324         }
 325         pagep = &((struct page **) pd->port_tid_pg_list)[tidoff];
 326         tidlist = &((u16 *) &pagep[dd->ipath_rcvtidcnt])[tidoff];
 327
 328         memset(tidmap, 0, sizeof(tidmap));
 329         /* before decrement; chip actual # */
 330         ntids = tidcnt;
 331         tidbase = (u64 __iomem *) (((char __iomem *) dd->ipath_kregbase) +
 332                                    dd->ipath_rcvtidbase +
 333                                    porttid * sizeof(*tidbase));
 334
 335         ipath_cdbg(VERBOSE, "Port%u %u tids, cursor %u, tidbase %p\n",
 336                    pd->port_port, cnt, tid, tidbase);
 337
 338         /* virtual address of first page in transfer */
 339         vaddr = ti->tidvaddr;
 340         if (!access_ok(VERIFY_WRITE, (void __user *) vaddr,
 341                        cnt * PAGE_SIZE)) {
 342                 ipath_dbg("Fail vaddr %p, %u pages, !access_ok\n",
 343                           (void *)vaddr, cnt);
 344                 ret = -EFAULT;
 345                 goto done;
 346         }
 347         ret = ipath_get_user_pages(vaddr, cnt, pagep);
 348         if (ret) {
 349                 if (ret == -EBUSY) {
 350                         ipath_dbg("Failed to lock addr %p, %u pages "
 351                                   "(already locked)\n",
 352                                   (void *) vaddr, cnt);
 353                         /*
 354                          * for now, continue, and see what happens but with
 355                          * the new implementation, this should never happen,
 356                          * unless perhaps the user has mpin'ed the pages
 357                          * themselves (something we need to test)
 358                          */
 359                         ret = 0;
 360                 } else {
 361                         dev_info(&dd->pcidev->dev,
 362                                  "Failed to lock addr %p, %u pages: "
 363                                  "errno %d\n", (void *) vaddr, cnt, -ret);
 364                         goto done;
 365                 }
 366         }
 367         for (i = 0; i < cnt; i++, vaddr += PAGE_SIZE) {
 368                 for (; ntids--; tid++) {
 369                         if (tid == tidcnt)
 370                                 tid = 0;
 371                         if (!dd->ipath_pageshadow[porttid + tid])
 372                                 break;
 373                 }
 374                 if (ntids < 0) {
 375                         /*
 376                          * oops, wrapped all the way through their TIDs,
 377                          * and didn't have enough free; see comments at
 378                          * start of routine
 379                          */
 380                         ipath_dbg("Not enough free TIDs for %u pages "
 381                                   "(index %d), failing\n", cnt, i);
 382                         i--;    /* last tidlist[i] not filled in */
 383                         ret = -ENOMEM;
 384                         break;
 385                 }
 386                 tidlist[i] = tid + tidoff;
 387                 ipath_cdbg(VERBOSE, "Updating idx %u to TID %u, "
 388                            "vaddr %lx\n", i, tid + tidoff, vaddr);
 389                 /* we "know" system pages and TID pages are same size */
 390                 dd->ipath_pageshadow[porttid + tid] = pagep[i];
 391                 dd->ipath_physshadow[porttid + tid] = ipath_map_page(
 392                         dd->pcidev, pagep[i], 0, PAGE_SIZE,
 393                         PCI_DMA_FROMDEVICE);
 394                 /*
 395                  * don't need atomic or it's overhead
 396                  */
 397                 __set_bit(tid, tidmap);
 398                 physaddr = dd->ipath_physshadow[porttid + tid];
 399                 ipath_stats.sps_pagelocks++;
 400                 ipath_cdbg(VERBOSE,
 401                            "TID %u, vaddr %lx, physaddr %llx pgp %p\n",
 402                            tid, vaddr, (unsigned long long) physaddr,
 403                            pagep[i]);
 404                 dd->ipath_f_put_tid(dd, &tidbase[tid], RCVHQ_RCV_TYPE_EXPECTED,
 405                                     physaddr);
 406                 /*
 407                  * don't check this tid in ipath_portshadow, since we
 408                  * just filled it in; start with the next one.
 409                  */
 410                 tid++;
 411         }
 412
 413         if (ret) {
 414                 u32 limit;
 415         cleanup:
 416                 /* jump here if copy out of updated info failed... */
 417                 ipath_dbg("After failure (ret=%d), undo %d of %d entries\n",
 418                           -ret, i, cnt);
 419                 /* same code that's in ipath_free_tid() */
 420                 limit = sizeof(tidmap) * BITS_PER_BYTE;
 421                 if (limit > tidcnt)
 422                         /* just in case size changes in future */
 423                         limit = tidcnt;
 424                 tid = find_first_bit((const unsigned long *)tidmap, limit);
 425                 for (; tid < limit; tid++) {
 426                         if (!test_bit(tid, tidmap))
 427                                 continue;
 428                         if (dd->ipath_pageshadow[porttid + tid]) {
 429                                 ipath_cdbg(VERBOSE, "Freeing TID %u\n",
 430                                            tid);
 431                                 dd->ipath_f_put_tid(dd, &tidbase[tid],
 432                                                     RCVHQ_RCV_TYPE_EXPECTED,
 433                                                     dd->ipath_tidinvalid);
 434                                 pci_unmap_page(dd->pcidev,
 435                                         dd->ipath_physshadow[porttid + tid],
 436                                         PAGE_SIZE, PCI_DMA_FROMDEVICE);
 437                                 dd->ipath_pageshadow[porttid + tid] = NULL;
 438                                 ipath_stats.sps_pageunlocks++;
 439                         }
 440                 }
 441                 ipath_release_user_pages(pagep, cnt);
 442         } else {
 443                 /*
 444                  * Copy the updated array, with ipath_tid's filled in, back
 445                  * to user.  Since we did the copy in already, this "should
 446                  * never fail" If it does, we have to clean up...
 447                  */
 448                 if (copy_to_user((void __user *)
 449                                  (unsigned long) ti->tidlist,
 450                                  tidlist, cnt * sizeof(*tidlist))) {
 451                         ret = -EFAULT;
 452                         goto cleanup;
 453                 }
 454                 if (copy_to_user((void __user *) (unsigned long) ti->tidmap,
 455                                  tidmap, sizeof tidmap)) {
 456                         ret = -EFAULT;
 457                         goto cleanup;
 458                 }
 459                 if (tid == tidcnt)
 460                         tid = 0;
 461                 if (!pd->port_subport_cnt)
 462                         pd->port_tidcursor = tid;
 463                 else
 464                         tidcursor_fp(fp) = tid;
 465         }
 466
 467 done:
 468         if (ret)
 469                 ipath_dbg("Failed to map %u TID pages, failing with %d\n",
 470                           ti->tidcnt, -ret);
 471         return ret;
 472 }
 473
 474 /**
 475  * ipath_tid_free - free a port TID
 476  * @pd: the port
 477  * @subport: the subport
 478  * @ti: the TID info
 479  *
 480  * right now we are unlocking one page at a time, but since
 481  * the intended use of this routine is for a single group of
 482  * virtually contiguous pages, that should change to improve
 483  * performance.  We check that the TID is in range for this port
 484  * but otherwise don't check validity; if user has an error and
 485  * frees the wrong tid, it's only their own data that can thereby
 486  * be corrupted.  We do check that the TID was in use, for sanity
 487  * We always use our idea of the saved address, not the address that
 488  * they pass in to us.
 489  */
 490
 491 static int ipath_tid_free(struct ipath_portdata *pd, unsigned subport,
 492                           const struct ipath_tid_info *ti)
 493 {
 494         int ret = 0;
 495         u32 tid, porttid, cnt, limit, tidcnt;
 496         struct ipath_devdata *dd = pd->port_dd;
 497         u64 __iomem *tidbase;
 498         unsigned long tidmap[8];
 499
 500         if (!dd->ipath_pageshadow) {
 501                 ret = -ENOMEM;
 502                 goto done;
 503         }
 504
 505         if (copy_from_user(tidmap, (void __user *)(unsigned long)ti->tidmap,
 506                            sizeof tidmap)) {
 507                 ret = -EFAULT;
 508                 goto done;
 509         }
 510
 511         porttid = pd->port_port * dd->ipath_rcvtidcnt;
 512         if (!pd->port_subport_cnt)
 513                 tidcnt = dd->ipath_rcvtidcnt;
 514         else if (!subport) {
 515                 tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
 516                          (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
 517                 porttid += dd->ipath_rcvtidcnt - tidcnt;
 518         } else {
 519                 tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
 520                 porttid += tidcnt * (subport - 1);
 521         }
 522         tidbase = (u64 __iomem *) ((char __iomem *)(dd->ipath_kregbase) +
 523                                    dd->ipath_rcvtidbase +
 524                                    porttid * sizeof(*tidbase));
 525
 526         limit = sizeof(tidmap) * BITS_PER_BYTE;
 527         if (limit > tidcnt)
 528                 /* just in case size changes in future */
 529                 limit = tidcnt;
 530         tid = find_first_bit(tidmap, limit);
 531         ipath_cdbg(VERBOSE, "Port%u free %u tids; first bit (max=%d) "
 532                    "set is %d, porttid %u\n", pd->port_port, ti->tidcnt,
 533                    limit, tid, porttid);
 534         for (cnt = 0; tid < limit; tid++) {
 535                 /*
 536                  * small optimization; if we detect a run of 3 or so without
 537                  * any set, use find_first_bit again.  That's mainly to
 538                  * accelerate the case where we wrapped, so we have some at
 539                  * the beginning, and some at the end, and a big gap
 540                  * in the middle.
 541                  */
 542                 if (!test_bit(tid, tidmap))
 543                         continue;
 544                 cnt++;
 545                 if (dd->ipath_pageshadow[porttid + tid]) {
 546                         struct page *p;
 547                         p = dd->ipath_pageshadow[porttid + tid];
 548                         dd->ipath_pageshadow[porttid + tid] = NULL;
 549                         ipath_cdbg(VERBOSE, "PID %u freeing TID %u\n",
 550                                    pd->port_pid, tid);
 551                         dd->ipath_f_put_tid(dd, &tidbase[tid],
 552                                             RCVHQ_RCV_TYPE_EXPECTED,
 553                                             dd->ipath_tidinvalid);
 554                         pci_unmap_page(dd->pcidev,
 555                                 dd->ipath_physshadow[porttid + tid],
 556                                 PAGE_SIZE, PCI_DMA_FROMDEVICE);
 557                         ipath_release_user_pages(&p, 1);
 558                         ipath_stats.sps_pageunlocks++;
 559                 } else
 560                         ipath_dbg("Unused tid %u, ignoring\n", tid);
 561         }
 562         if (cnt != ti->tidcnt)
 563                 ipath_dbg("passed in tidcnt %d, only %d bits set in map\n",
 564                           ti->tidcnt, cnt);
 565 done:
 566         if (ret)
 567                 ipath_dbg("Failed to unmap %u TID pages, failing with %d\n",
 568                           ti->tidcnt, -ret);
 569         return ret;
 570 }
 571
 572 /**
 573  * ipath_set_part_key - set a partition key
 574  * @pd: the port
 575  * @key: the key
 576  *
 577  * We can have up to 4 active at a time (other than the default, which is
 578  * always allowed).  This is somewhat tricky, since multiple ports may set
 579  * the same key, so we reference count them, and clean up at exit.  All 4
 580  * partition keys are packed into a single infinipath register.  It's an
 581  * error for a process to set the same pkey multiple times.  We provide no
 582  * mechanism to de-allocate a pkey at this time, we may eventually need to
 583  * do that.  I've used the atomic operations, and no locking, and only make
 584  * a single pass through what's available.  This should be more than
 585  * adequate for some time. I'll think about spinlocks or the like if and as
 586  * it's necessary.
 587  */
 588 static int ipath_set_part_key(struct ipath_portdata *pd, u16 key)
 589 {
 590         struct ipath_devdata *dd = pd->port_dd;
 591         int i, any = 0, pidx = -1;
 592         u16 lkey = key & 0x7FFF;
 593         int ret;
 594
 595         if (lkey == (IPATH_DEFAULT_P_KEY & 0x7FFF)) {
 596                 /* nothing to do; this key always valid */
 597                 ret = 0;
 598                 goto bail;
 599         }
 600
 601         ipath_cdbg(VERBOSE, "p%u try to set pkey %hx, current keys "
 602                    "%hx:%x %hx:%x %hx:%x %hx:%x\n",
 603                    pd->port_port, key, dd->ipath_pkeys[0],
 604                    atomic_read(&dd->ipath_pkeyrefs[0]), dd->ipath_pkeys[1],
 605                    atomic_read(&dd->ipath_pkeyrefs[1]), dd->ipath_pkeys[2],
 606                    atomic_read(&dd->ipath_pkeyrefs[2]), dd->ipath_pkeys[3],
 607                    atomic_read(&dd->ipath_pkeyrefs[3]));
 608
 609         if (!lkey) {
 610                 ipath_cdbg(PROC, "p%u tries to set key 0, not allowed\n",
 611                            pd->port_port);
 612                 ret = -EINVAL;
 613                 goto bail;
 614         }
 615
 616         /*
 617          * Set the full membership bit, because it has to be
 618          * set in the register or the packet, and it seems
 619          * cleaner to set in the register than to force all
 620          * callers to set it. (see bug 4331)
 621          */
 622         key |= 0x8000;
 623
 624         for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
 625                 if (!pd->port_pkeys[i] && pidx == -1)
 626                         pidx = i;
 627                 if (pd->port_pkeys[i] == key) {
 628                         ipath_cdbg(VERBOSE, "p%u tries to set same pkey "
 629                                    "(%x) more than once\n",
 630                                    pd->port_port, key);
 631                         ret = -EEXIST;
 632                         goto bail;
 633                 }
 634         }
 635         if (pidx == -1) {
 636                 ipath_dbg("All pkeys for port %u already in use, "
 637                           "can't set %x\n", pd->port_port, key);
 638                 ret = -EBUSY;
 639                 goto bail;
 640         }
 641         for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
 642                 if (!dd->ipath_pkeys[i]) {
 643                         any++;
 644                         continue;
 645                 }
 646                 if (dd->ipath_pkeys[i] == key) {
 647                         atomic_t *pkrefs = &dd->ipath_pkeyrefs[i];
 648
 649                         if (atomic_inc_return(pkrefs) > 1) {
 650                                 pd->port_pkeys[pidx] = key;
 651                                 ipath_cdbg(VERBOSE, "p%u set key %x "
 652                                            "matches #%d, count now %d\n",
 653                                            pd->port_port, key, i,
 654                                            atomic_read(pkrefs));
 655                                 ret = 0;
 656                                 goto bail;
 657                         } else {
 658                                 /*
 659                                  * lost race, decrement count, catch below
 660                                  */
 661                                 atomic_dec(pkrefs);
 662                                 ipath_cdbg(VERBOSE, "Lost race, count was "
 663                                            "0, after dec, it's %d\n",
 664                                            atomic_read(pkrefs));
 665                                 any++;
 666                         }
 667                 }
 668                 if ((dd->ipath_pkeys[i] & 0x7FFF) == lkey) {
 669                         /*
 670                          * It makes no sense to have both the limited and
 671                          * full membership PKEY set at the same time since
 672                          * the unlimited one will disable the limited one.
 673                          */
 674                         ret = -EEXIST;
 675                         goto bail;
 676                 }
 677         }
 678         if (!any) {
 679                 ipath_dbg("port %u, all pkeys already in use, "
 680                           "can't set %x\n", pd->port_port, key);
 681                 ret = -EBUSY;
 682                 goto bail;
 683         }
 684         for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
 685                 if (!dd->ipath_pkeys[i] &&
 686                     atomic_inc_return(&dd->ipath_pkeyrefs[i]) == 1) {
 687                         u64 pkey;
 688
 689                         /* for ipathstats, etc. */
 690                         ipath_stats.sps_pkeys[i] = lkey;
 691                         pd->port_pkeys[pidx] = dd->ipath_pkeys[i] = key;
 692                         pkey =
 693                                 (u64) dd->ipath_pkeys[0] |
 694                                 ((u64) dd->ipath_pkeys[1] << 16) |
 695                                 ((u64) dd->ipath_pkeys[2] << 32) |
 696                                 ((u64) dd->ipath_pkeys[3] << 48);
 697                         ipath_cdbg(PROC, "p%u set key %x in #%d, "
 698                                    "portidx %d, new pkey reg %llx\n",
 699                                    pd->port_port, key, i, pidx,
 700                                    (unsigned long long) pkey);
 701                         ipath_write_kreg(
 702                                 dd, dd->ipath_kregs->kr_partitionkey, pkey);
 703
 704                         ret = 0;
 705                         goto bail;
 706                 }
 707         }
 708         ipath_dbg("port %u, all pkeys already in use 2nd pass, "
 709                   "can't set %x\n", pd->port_port, key);
 710         ret = -EBUSY;
 711
 712 bail:
 713         return ret;
 714 }
 715
 716 /**
 717  * ipath_manage_rcvq - manage a port's receive queue
 718  * @pd: the port
 719  * @subport: the subport
 720  * @start_stop: action to carry out
 721  *
 722  * start_stop == 0 disables receive on the port, for use in queue
 723  * overflow conditions.  start_stop==1 re-enables, to be used to
 724  * re-init the software copy of the head register
 725  */
 726 static int ipath_manage_rcvq(struct ipath_portdata *pd, unsigned subport,
 727                              int start_stop)
 728 {
 729         struct ipath_devdata *dd = pd->port_dd;
 730
 731         ipath_cdbg(PROC, "%sabling rcv for unit %u port %u:%u\n",
 732                    start_stop ? "en" : "dis", dd->ipath_unit,
 733                    pd->port_port, subport);
 734         if (subport)
 735                 goto bail;
 736         /* atomically clear receive enable port. */
 737         if (start_stop) {
 738                 /*
 739                  * On enable, force in-memory copy of the tail register to
 740                  * 0, so that protocol code doesn't have to worry about
 741                  * whether or not the chip has yet updated the in-memory
 742                  * copy or not on return from the system call. The chip
 743                  * always resets it's tail register back to 0 on a
 744                  * transition from disabled to enabled.  This could cause a
 745                  * problem if software was broken, and did the enable w/o
 746                  * the disable, but eventually the in-memory copy will be
 747                  * updated and correct itself, even in the face of software
 748                  * bugs.
 749                  */
 750                 if (pd->port_rcvhdrtail_kvaddr)
 751                         ipath_clear_rcvhdrtail(pd);
 752                 set_bit(dd->ipath_r_portenable_shift + pd->port_port,
 753                         &dd->ipath_rcvctrl);
 754         } else
 755                 clear_bit(dd->ipath_r_portenable_shift + pd->port_port,
 756                           &dd->ipath_rcvctrl);
 757         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
 758                          dd->ipath_rcvctrl);
 759         /* now be sure chip saw it before we return */
 760         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 761         if (start_stop) {
 762                 /*
 763                  * And try to be sure that tail reg update has happened too.
 764                  * This should in theory interlock with the RXE changes to
 765                  * the tail register.  Don't assign it to the tail register
 766                  * in memory copy, since we could overwrite an update by the
 767                  * chip if we did.
 768                  */
 769                 ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
 770         }
 771         /* always; new head should be equal to new tail; see above */
 772 bail:
 773         return 0;
 774 }
 775
 776 static void ipath_clean_part_key(struct ipath_portdata *pd,
 777                                  struct ipath_devdata *dd)
 778 {
 779         int i, j, pchanged = 0;
 780         u64 oldpkey;
 781
 782         /* for debugging only */
 783         oldpkey = (u64) dd->ipath_pkeys[0] |
 784                 ((u64) dd->ipath_pkeys[1] << 16) |
 785                 ((u64) dd->ipath_pkeys[2] << 32) |
 786                 ((u64) dd->ipath_pkeys[3] << 48);
 787
 788         for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
 789                 if (!pd->port_pkeys[i])
 790                         continue;
 791                 ipath_cdbg(VERBOSE, "look for key[%d] %hx in pkeys\n", i,
 792                            pd->port_pkeys[i]);
 793                 for (j = 0; j < ARRAY_SIZE(dd->ipath_pkeys); j++) {
 794                         /* check for match independent of the global bit */
 795                         if ((dd->ipath_pkeys[j] & 0x7fff) !=
 796                             (pd->port_pkeys[i] & 0x7fff))
 797                                 continue;
 798                         if (atomic_dec_and_test(&dd->ipath_pkeyrefs[j])) {
 799                                 ipath_cdbg(VERBOSE, "p%u clear key "
 800                                            "%x matches #%d\n",
 801                                            pd->port_port,
 802                                            pd->port_pkeys[i], j);
 803                                 ipath_stats.sps_pkeys[j] =
 804                                         dd->ipath_pkeys[j] = 0;
 805                                 pchanged++;
 806                         }
 807                         else ipath_cdbg(
 808                                 VERBOSE, "p%u key %x matches #%d, "
 809                                 "but ref still %d\n", pd->port_port,
 810                                 pd->port_pkeys[i], j,
 811                                 atomic_read(&dd->ipath_pkeyrefs[j]));
 812                         break;
 813                 }
 814                 pd->port_pkeys[i] = 0;
 815         }
 816         if (pchanged) {
 817                 u64 pkey = (u64) dd->ipath_pkeys[0] |
 818                         ((u64) dd->ipath_pkeys[1] << 16) |
 819                         ((u64) dd->ipath_pkeys[2] << 32) |
 820                         ((u64) dd->ipath_pkeys[3] << 48);
 821                 ipath_cdbg(VERBOSE, "p%u old pkey reg %llx, "
 822                            "new pkey reg %llx\n", pd->port_port,
 823                            (unsigned long long) oldpkey,
 824                            (unsigned long long) pkey);
 825                 ipath_write_kreg(dd, dd->ipath_kregs->kr_partitionkey,
 826                                  pkey);
 827         }
 828 }
 829
 830 /*
 831  * Initialize the port data with the receive buffer sizes
 832  * so this can be done while the master port is locked.
 833  * Otherwise, there is a race with a slave opening the port
 834  * and seeing these fields uninitialized.
 835  */
 836 static void init_user_egr_sizes(struct ipath_portdata *pd)
 837 {
 838         struct ipath_devdata *dd = pd->port_dd;
 839         unsigned egrperchunk, egrcnt, size;
 840
 841         /*
 842          * to avoid wasting a lot of memory, we allocate 32KB chunks of
 843          * physically contiguous memory, advance through it until used up
 844          * and then allocate more.  Of course, we need memory to store those
 845          * extra pointers, now.  Started out with 256KB, but under heavy
 846          * memory pressure (creating large files and then copying them over
 847          * NFS while doing lots of MPI jobs), we hit some allocation
 848          * failures, even though we can sleep...  (2.6.10) Still get
 849          * failures at 64K.  32K is the lowest we can go without wasting
 850          * additional memory.
 851          */
 852         size = 0x8000;
 853         egrperchunk = size / dd->ipath_rcvegrbufsize;
 854         egrcnt = dd->ipath_rcvegrcnt;
 855         pd->port_rcvegrbuf_chunks = (egrcnt + egrperchunk - 1) / egrperchunk;
 856         pd->port_rcvegrbufs_perchunk = egrperchunk;
 857         pd->port_rcvegrbuf_size = size;
 858 }
 859
 860 /**
 861  * ipath_create_user_egr - allocate eager TID buffers
 862  * @pd: the port to allocate TID buffers for
 863  *
 864  * This routine is now quite different for user and kernel, because
 865  * the kernel uses skb's, for the accelerated network performance
 866  * This is the user port version
 867  *
 868  * Allocate the eager TID buffers and program them into infinipath
 869  * They are no longer completely contiguous, we do multiple allocation
 870  * calls.
 871  */
 872 static int ipath_create_user_egr(struct ipath_portdata *pd)
 873 {
 874         struct ipath_devdata *dd = pd->port_dd;
 875         unsigned e, egrcnt, egrperchunk, chunk, egrsize, egroff;
 876         size_t size;
 877         int ret;
 878         gfp_t gfp_flags;
 879
 880         /*
 881          * GFP_USER, but without GFP_FS, so buffer cache can be
 882          * coalesced (we hope); otherwise, even at order 4,
 883          * heavy filesystem activity makes these fail, and we can
 884          * use compound pages.
 885          */
 886         gfp_flags = __GFP_WAIT | __GFP_IO | __GFP_COMP;
 887
 888         egrcnt = dd->ipath_rcvegrcnt;
 889         /* TID number offset for this port */
 890         egroff = (pd->port_port - 1) * egrcnt + dd->ipath_p0_rcvegrcnt;
 891         egrsize = dd->ipath_rcvegrbufsize;
 892         ipath_cdbg(VERBOSE, "Allocating %d egr buffers, at egrtid "
 893                    "offset %x, egrsize %u\n", egrcnt, egroff, egrsize);
 894
 895         chunk = pd->port_rcvegrbuf_chunks;
 896         egrperchunk = pd->port_rcvegrbufs_perchunk;
 897         size = pd->port_rcvegrbuf_size;
 898         pd->port_rcvegrbuf = kmalloc(chunk * sizeof(pd->port_rcvegrbuf[0]),
 899                                      GFP_KERNEL);
 900         if (!pd->port_rcvegrbuf) {
 901                 ret = -ENOMEM;
 902                 goto bail;
 903         }
 904         pd->port_rcvegrbuf_phys =
 905                 kmalloc(chunk * sizeof(pd->port_rcvegrbuf_phys[0]),
 906                         GFP_KERNEL);
 907         if (!pd->port_rcvegrbuf_phys) {
 908                 ret = -ENOMEM;
 909                 goto bail_rcvegrbuf;
 910         }
 911         for (e = 0; e < pd->port_rcvegrbuf_chunks; e++) {
 912
 913                 pd->port_rcvegrbuf[e] = dma_alloc_coherent(
 914                         &dd->pcidev->dev, size, &pd->port_rcvegrbuf_phys[e],
 915                         gfp_flags);
 916
 917                 if (!pd->port_rcvegrbuf[e]) {
 918                         ret = -ENOMEM;
 919                         goto bail_rcvegrbuf_phys;
 920                 }
 921         }
 922
 923         pd->port_rcvegr_phys = pd->port_rcvegrbuf_phys[0];
 924
 925         for (e = chunk = 0; chunk < pd->port_rcvegrbuf_chunks; chunk++) {
 926                 dma_addr_t pa = pd->port_rcvegrbuf_phys[chunk];
 927                 unsigned i;
 928
 929                 for (i = 0; e < egrcnt && i < egrperchunk; e++, i++) {
 930                         dd->ipath_f_put_tid(dd, e + egroff +
 931                                             (u64 __iomem *)
 932                                             ((char __iomem *)
 933                                              dd->ipath_kregbase +
 934                                              dd->ipath_rcvegrbase),
 935                                             RCVHQ_RCV_TYPE_EAGER, pa);
 936                         pa += egrsize;
 937                 }
 938                 cond_resched(); /* don't hog the cpu */
 939         }
 940
 941         ret = 0;
 942         goto bail;
 943
 944 bail_rcvegrbuf_phys:
 945         for (e = 0; e < pd->port_rcvegrbuf_chunks &&
 946                 pd->port_rcvegrbuf[e]; e++) {
 947                 dma_free_coherent(&dd->pcidev->dev, size,
 948                                   pd->port_rcvegrbuf[e],
 949                                   pd->port_rcvegrbuf_phys[e]);
 950
 951         }
 952         kfree(pd->port_rcvegrbuf_phys);
 953         pd->port_rcvegrbuf_phys = NULL;
 954 bail_rcvegrbuf:
 955         kfree(pd->port_rcvegrbuf);
 956         pd->port_rcvegrbuf = NULL;
 957 bail:
 958         return ret;
 959 }
 960
 961
 962 /* common code for the mappings on dma_alloc_coherent mem */
 963 static int ipath_mmap_mem(struct vm_area_struct *vma,
 964         struct ipath_portdata *pd, unsigned len, int write_ok,
 965         void *kvaddr, char *what)
 966 {
 967         struct ipath_devdata *dd = pd->port_dd;
 968         unsigned long pfn;
 969         int ret;
 970
 971         if ((vma->vm_end - vma->vm_start) > len) {
 972                 dev_info(&dd->pcidev->dev,
 973                          "FAIL on %s: len %lx > %x\n", what,
 974                          vma->vm_end - vma->vm_start, len);
 975                 ret = -EFAULT;
 976                 goto bail;
 977         }
 978
 979         if (!write_ok) {
 980                 if (vma->vm_flags & VM_WRITE) {
 981                         dev_info(&dd->pcidev->dev,
 982                                  "%s must be mapped readonly\n", what);
 983                         ret = -EPERM;
 984                         goto bail;
 985                 }
 986
 987                 /* don't allow them to later change with mprotect */
 988                 vma->vm_flags &= ~VM_MAYWRITE;
 989         }
 990
 991         pfn = virt_to_phys(kvaddr) >> PAGE_SHIFT;
 992         ret = remap_pfn_range(vma, vma->vm_start, pfn,
 993                               len, vma->vm_page_prot);
 994         if (ret)
 995                 dev_info(&dd->pcidev->dev, "%s port%u mmap of %lx, %x "
 996                          "bytes r%c failed: %d\n", what, pd->port_port,
 997                          pfn, len, write_ok?'w':'o', ret);
 998         else
 999                 ipath_cdbg(VERBOSE, "%s port%u mmaped %lx, %x bytes "
1000                            "r%c\n", what, pd->port_port, pfn, len,
1001                            write_ok?'w':'o');
1002 bail:
1003         return ret;
1004 }
1005
1006 static int mmap_ureg(struct vm_area_struct *vma, struct ipath_devdata *dd,
1007                      u64 ureg)
1008 {
1009         unsigned long phys;
1010         int ret;
1011
1012         /*
1013          * This is real hardware, so use io_remap.  This is the mechanism
1014          * for the user process to update the head registers for their port
1015          * in the chip.
1016          */
1017         if ((vma->vm_end - vma->vm_start) > PAGE_SIZE) {
1018                 dev_info(&dd->pcidev->dev, "FAIL mmap userreg: reqlen "
1019                          "%lx > PAGE\n", vma->vm_end - vma->vm_start);
1020                 ret = -EFAULT;
1021         } else {
1022                 phys = dd->ipath_physaddr + ureg;
1023                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1024
1025                 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1026                 ret = io_remap_pfn_range(vma, vma->vm_start,
1027                                          phys >> PAGE_SHIFT,
1028                                          vma->vm_end - vma->vm_start,
1029                                          vma->vm_page_prot);
1030         }
1031         return ret;
1032 }
1033
1034 static int mmap_piobufs(struct vm_area_struct *vma,
1035                         struct ipath_devdata *dd,
1036                         struct ipath_portdata *pd,
1037                         unsigned piobufs, unsigned piocnt)
1038 {
1039         unsigned long phys;
1040         int ret;
1041
1042         /*
1043          * When we map the PIO buffers in the chip, we want to map them as
1044          * writeonly, no read possible.   This prevents access to previous
1045          * process data, and catches users who might try to read the i/o
1046          * space due to a bug.
1047          */
1048         if ((vma->vm_end - vma->vm_start) > (piocnt * dd->ipath_palign)) {
1049                 dev_info(&dd->pcidev->dev, "FAIL mmap piobufs: "
1050                          "reqlen %lx > PAGE\n",
1051                          vma->vm_end - vma->vm_start);
1052                 ret = -EINVAL;
1053                 goto bail;
1054         }
1055
1056         phys = dd->ipath_physaddr + piobufs;
1057
1058 #if defined(__powerpc__)
1059         /* There isn't a generic way to specify writethrough mappings */
1060         pgprot_val(vma->vm_page_prot) |= _PAGE_NO_CACHE;
1061         pgprot_val(vma->vm_page_prot) |= _PAGE_WRITETHRU;
1062         pgprot_val(vma->vm_page_prot) &= ~_PAGE_GUARDED;
1063 #endif
1064
1065         /*
1066          * don't allow them to later change to readable with mprotect (for when
1067          * not initially mapped readable, as is normally the case)
1068          */
1069         vma->vm_flags &= ~VM_MAYREAD;
1070         vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1071
1072         ret = io_remap_pfn_range(vma, vma->vm_start, phys >> PAGE_SHIFT,
1073                                  vma->vm_end - vma->vm_start,
1074                                  vma->vm_page_prot);
1075 bail:
1076         return ret;
1077 }
1078
1079 static int mmap_rcvegrbufs(struct vm_area_struct *vma,
1080                            struct ipath_portdata *pd)
1081 {
1082         struct ipath_devdata *dd = pd->port_dd;
1083         unsigned long start, size;
1084         size_t total_size, i;
1085         unsigned long pfn;
1086         int ret;
1087
1088         size = pd->port_rcvegrbuf_size;
1089         total_size = pd->port_rcvegrbuf_chunks * size;
1090         if ((vma->vm_end - vma->vm_start) > total_size) {
1091                 dev_info(&dd->pcidev->dev, "FAIL on egr bufs: "
1092                          "reqlen %lx > actual %lx\n",
1093                          vma->vm_end - vma->vm_start,
1094                          (unsigned long) total_size);
1095                 ret = -EINVAL;
1096                 goto bail;
1097         }
1098
1099         if (vma->vm_flags & VM_WRITE) {
1100                 dev_info(&dd->pcidev->dev, "Can't map eager buffers as "
1101                          "writable (flags=%lx)\n", vma->vm_flags);
1102                 ret = -EPERM;
1103                 goto bail;
1104         }
1105         /* don't allow them to later change to writeable with mprotect */
1106         vma->vm_flags &= ~VM_MAYWRITE;
1107
1108         start = vma->vm_start;
1109
1110         for (i = 0; i < pd->port_rcvegrbuf_chunks; i++, start += size) {
1111                 pfn = virt_to_phys(pd->port_rcvegrbuf[i]) >> PAGE_SHIFT;
1112                 ret = remap_pfn_range(vma, start, pfn, size,
1113                                       vma->vm_page_prot);
1114                 if (ret < 0)
1115                         goto bail;
1116         }
1117         ret = 0;
1118
1119 bail:
1120         return ret;
1121 }
1122
1123 /*
1124  * ipath_file_vma_fault - handle a VMA page fault.
1125  */
1126 static int ipath_file_vma_fault(struct vm_area_struct *vma,
1127                                         struct vm_fault *vmf)
1128 {
1129         struct page *page;
1130
1131         page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
1132         if (!page)
1133                 return VM_FAULT_SIGBUS;
1134         get_page(page);
1135         vmf->page = page;
1136
1137         return 0;
1138 }
1139
1140 static struct vm_operations_struct ipath_file_vm_ops = {
1141         .fault = ipath_file_vma_fault,
1142 };
1143
1144 static int mmap_kvaddr(struct vm_area_struct *vma, u64 pgaddr,
1145                        struct ipath_portdata *pd, unsigned subport)
1146 {
1147         unsigned long len;
1148         struct ipath_devdata *dd;
1149         void *addr;
1150         size_t size;
1151         int ret = 0;
1152
1153         /* If the port is not shared, all addresses should be physical */
1154         if (!pd->port_subport_cnt)
1155                 goto bail;
1156
1157         dd = pd->port_dd;
1158         size = pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
1159
1160         /*
1161          * Each process has all the subport uregbase, rcvhdrq, and
1162          * rcvegrbufs mmapped - as an array for all the processes,
1163          * and also separately for this process.
1164          */
1165         if (pgaddr == cvt_kvaddr(pd->subport_uregbase)) {
1166                 addr = pd->subport_uregbase;
1167                 size = PAGE_SIZE * pd->port_subport_cnt;
1168         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvhdr_base)) {
1169                 addr = pd->subport_rcvhdr_base;
1170                 size = pd->port_rcvhdrq_size * pd->port_subport_cnt;
1171         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvegrbuf)) {
1172                 addr = pd->subport_rcvegrbuf;
1173                 size *= pd->port_subport_cnt;
1174         } else if (pgaddr == cvt_kvaddr(pd->subport_uregbase +
1175                                         PAGE_SIZE * subport)) {
1176                 addr = pd->subport_uregbase + PAGE_SIZE * subport;
1177                 size = PAGE_SIZE;
1178         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvhdr_base +
1179                                 pd->port_rcvhdrq_size * subport)) {
1180                 addr = pd->subport_rcvhdr_base +
1181                         pd->port_rcvhdrq_size * subport;
1182                 size = pd->port_rcvhdrq_size;
1183         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvegrbuf +
1184                                size * subport)) {
1185                 addr = pd->subport_rcvegrbuf + size * subport;
1186                 /* rcvegrbufs are read-only on the slave */
1187                 if (vma->vm_flags & VM_WRITE) {
1188                         dev_info(&dd->pcidev->dev,
1189                                  "Can't map eager buffers as "
1190                                  "writable (flags=%lx)\n", vma->vm_flags);
1191                         ret = -EPERM;
1192                         goto bail;
1193                 }
1194                 /*
1195                  * Don't allow permission to later change to writeable
1196                  * with mprotect.
1197                  */
1198                 vma->vm_flags &= ~VM_MAYWRITE;
1199         } else {
1200                 goto bail;
1201         }
1202         len = vma->vm_end - vma->vm_start;
1203         if (len > size) {
1204                 ipath_cdbg(MM, "FAIL: reqlen %lx > %zx\n", len, size);
1205                 ret = -EINVAL;
1206                 goto bail;
1207         }
1208
1209         vma->vm_pgoff = (unsigned long) addr >> PAGE_SHIFT;
1210         vma->vm_ops = &ipath_file_vm_ops;
1211         vma->vm_flags |= VM_RESERVED | VM_DONTEXPAND;
1212         ret = 1;
1213
1214 bail:
1215         return ret;
1216 }
1217
1218 /**
1219  * ipath_mmap - mmap various structures into user space
1220  * @fp: the file pointer
1221  * @vma: the VM area
1222  *
1223  * We use this to have a shared buffer between the kernel and the user code
1224  * for the rcvhdr queue, egr buffers, and the per-port user regs and pio
1225  * buffers in the chip.  We have the open and close entries so we can bump
1226  * the ref count and keep the driver from being unloaded while still mapped.
1227  */
1228 static int ipath_mmap(struct file *fp, struct vm_area_struct *vma)
1229 {
1230         struct ipath_portdata *pd;
1231         struct ipath_devdata *dd;
1232         u64 pgaddr, ureg;
1233         unsigned piobufs, piocnt;
1234         int ret;
1235
1236         pd = port_fp(fp);
1237         if (!pd) {
1238                 ret = -EINVAL;
1239                 goto bail;
1240         }
1241         dd = pd->port_dd;
1242
1243         /*
1244          * This is the ipath_do_user_init() code, mapping the shared buffers
1245          * into the user process. The address referred to by vm_pgoff is the
1246          * file offset passed via mmap().  For shared ports, this is the
1247          * kernel vmalloc() address of the pages to share with the master.
1248          * For non-shared or master ports, this is a physical address.
1249          * We only do one mmap for each space mapped.
1250          */
1251         pgaddr = vma->vm_pgoff << PAGE_SHIFT;
1252
1253         /*
1254          * Check for 0 in case one of the allocations failed, but user
1255          * called mmap anyway.
1256          */
1257         if (!pgaddr)  {
1258                 ret = -EINVAL;
1259                 goto bail;
1260         }
1261
1262         ipath_cdbg(MM, "pgaddr %llx vm_start=%lx len %lx port %u:%u:%u\n",
1263                    (unsigned long long) pgaddr, vma->vm_start,
1264                    vma->vm_end - vma->vm_start, dd->ipath_unit,
1265                    pd->port_port, subport_fp(fp));
1266
1267         /*
1268          * Physical addresses must fit in 40 bits for our hardware.
1269          * Check for kernel virtual addresses first, anything else must
1270          * match a HW or memory address.
1271          */
1272         ret = mmap_kvaddr(vma, pgaddr, pd, subport_fp(fp));
1273         if (ret) {
1274                 if (ret > 0)
1275                         ret = 0;
1276                 goto bail;
1277         }
1278
1279         ureg = dd->ipath_uregbase + dd->ipath_ureg_align * pd->port_port;
1280         if (!pd->port_subport_cnt) {
1281                 /* port is not shared */
1282                 piocnt = dd->ipath_pbufsport;
1283                 piobufs = pd->port_piobufs;
1284         } else if (!subport_fp(fp)) {
1285                 /* caller is the master */
1286                 piocnt = (dd->ipath_pbufsport / pd->port_subport_cnt) +
1287                          (dd->ipath_pbufsport % pd->port_subport_cnt);
1288                 piobufs = pd->port_piobufs +
1289                         dd->ipath_palign * (dd->ipath_pbufsport - piocnt);
1290         } else {
1291                 unsigned slave = subport_fp(fp) - 1;
1292
1293                 /* caller is a slave */
1294                 piocnt = dd->ipath_pbufsport / pd->port_subport_cnt;
1295                 piobufs = pd->port_piobufs + dd->ipath_palign * piocnt * slave;
1296         }
1297
1298         if (pgaddr == ureg)
1299                 ret = mmap_ureg(vma, dd, ureg);
1300         else if (pgaddr == piobufs)
1301                 ret = mmap_piobufs(vma, dd, pd, piobufs, piocnt);
1302         else if (pgaddr == dd->ipath_pioavailregs_phys)
1303                 /* in-memory copy of pioavail registers */
1304                 ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1305                                      (void *) dd->ipath_pioavailregs_dma,
1306                                      "pioavail registers");
1307         else if (pgaddr == pd->port_rcvegr_phys)
1308                 ret = mmap_rcvegrbufs(vma, pd);
1309         else if (pgaddr == (u64) pd->port_rcvhdrq_phys)
1310                 /*
1311                  * The rcvhdrq itself; readonly except on HT (so have
1312                  * to allow writable mapping), multiple pages, contiguous
1313                  * from an i/o perspective.
1314                  */
1315                 ret = ipath_mmap_mem(vma, pd, pd->port_rcvhdrq_size, 1,
1316                                      pd->port_rcvhdrq,
1317                                      "rcvhdrq");
1318         else if (pgaddr == (u64) pd->port_rcvhdrqtailaddr_phys)
1319                 /* in-memory copy of rcvhdrq tail register */
1320                 ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1321                                      pd->port_rcvhdrtail_kvaddr,
1322                                      "rcvhdrq tail");
1323         else
1324                 ret = -EINVAL;
1325
1326         vma->vm_private_data = NULL;
1327
1328         if (ret < 0)
1329                 dev_info(&dd->pcidev->dev,
1330                          "Failure %d on off %llx len %lx\n",
1331                          -ret, (unsigned long long)pgaddr,
1332                          vma->vm_end - vma->vm_start);
1333 bail:
1334         return ret;
1335 }
1336
1337 static unsigned ipath_poll_hdrqfull(struct ipath_portdata *pd)
1338 {
1339         unsigned pollflag = 0;
1340
1341         if ((pd->poll_type & IPATH_POLL_TYPE_OVERFLOW) &&
1342             pd->port_hdrqfull != pd->port_hdrqfull_poll) {
1343                 pollflag |= POLLIN | POLLRDNORM;
1344                 pd->port_hdrqfull_poll = pd->port_hdrqfull;
1345         }
1346
1347         return pollflag;
1348 }
1349
1350 static unsigned int ipath_poll_urgent(struct ipath_portdata *pd,
1351                                       struct file *fp,
1352                                       struct poll_table_struct *pt)
1353 {
1354         unsigned pollflag = 0;
1355         struct ipath_devdata *dd;
1356
1357         dd = pd->port_dd;
1358
1359         /* variable access in ipath_poll_hdrqfull() needs this */
1360         rmb();
1361         pollflag = ipath_poll_hdrqfull(pd);
1362
1363         if (pd->port_urgent != pd->port_urgent_poll) {
1364                 pollflag |= POLLIN | POLLRDNORM;
1365                 pd->port_urgent_poll = pd->port_urgent;
1366         }
1367
1368         if (!pollflag) {
1369                 /* this saves a spin_lock/unlock in interrupt handler... */
1370                 set_bit(IPATH_PORT_WAITING_URG, &pd->port_flag);
1371                 /* flush waiting flag so don't miss an event... */
1372                 wmb();
1373                 poll_wait(fp, &pd->port_wait, pt);
1374         }
1375
1376         return pollflag;
1377 }
1378
1379 static unsigned int ipath_poll_next(struct ipath_portdata *pd,
1380                                     struct file *fp,
1381                                     struct poll_table_struct *pt)
1382 {
1383         u32 head;
1384         u32 tail;
1385         unsigned pollflag = 0;
1386         struct ipath_devdata *dd;
1387
1388         dd = pd->port_dd;
1389
1390         /* variable access in ipath_poll_hdrqfull() needs this */
1391         rmb();
1392         pollflag = ipath_poll_hdrqfull(pd);
1393
1394         head = ipath_read_ureg32(dd, ur_rcvhdrhead, pd->port_port);
1395         if (pd->port_rcvhdrtail_kvaddr)
1396                 tail = ipath_get_rcvhdrtail(pd);
1397         else
1398                 tail = ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
1399
1400         if (head != tail)
1401                 pollflag |= POLLIN | POLLRDNORM;
1402         else {
1403                 /* this saves a spin_lock/unlock in interrupt handler */
1404                 set_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag);
1405                 /* flush waiting flag so we don't miss an event */
1406                 wmb();
1407
1408                 set_bit(pd->port_port + dd->ipath_r_intravail_shift,
1409                         &dd->ipath_rcvctrl);
1410
1411                 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1412                                  dd->ipath_rcvctrl);
1413
1414                 if (dd->ipath_rhdrhead_intr_off) /* arm rcv interrupt */
1415                         ipath_write_ureg(dd, ur_rcvhdrhead,
1416                                          dd->ipath_rhdrhead_intr_off | head,
1417                                          pd->port_port);
1418
1419                 poll_wait(fp, &pd->port_wait, pt);
1420         }
1421
1422         return pollflag;
1423 }
1424
1425 static unsigned int ipath_poll(struct file *fp,
1426                                struct poll_table_struct *pt)
1427 {
1428         struct ipath_portdata *pd;
1429         unsigned pollflag;
1430
1431         pd = port_fp(fp);
1432         if (!pd)
1433                 pollflag = 0;
1434         else if (pd->poll_type & IPATH_POLL_TYPE_URGENT)
1435                 pollflag = ipath_poll_urgent(pd, fp, pt);
1436         else
1437                 pollflag = ipath_poll_next(pd, fp, pt);
1438
1439         return pollflag;
1440 }
1441
1442 static int ipath_supports_subports(int user_swmajor, int user_swminor)
1443 {
1444         /* no subport implementation prior to software version 1.3 */
1445         return (user_swmajor > 1) || (user_swminor >= 3);
1446 }
1447
1448 static int ipath_compatible_subports(int user_swmajor, int user_swminor)
1449 {
1450         /* this code is written long-hand for clarity */
1451         if (IPATH_USER_SWMAJOR != user_swmajor) {
1452                 /* no promise of compatibility if major mismatch */
1453                 return 0;
1454         }
1455         if (IPATH_USER_SWMAJOR == 1) {
1456                 switch (IPATH_USER_SWMINOR) {
1457                 case 0:
1458                 case 1:
1459                 case 2:
1460                         /* no subport implementation so cannot be compatible */
1461                         return 0;
1462                 case 3:
1463                         /* 3 is only compatible with itself */
1464                         return user_swminor == 3;
1465                 default:
1466                         /* >= 4 are compatible (or are expected to be) */
1467                         return user_swminor >= 4;
1468                 }
1469         }
1470         /* make no promises yet for future major versions */
1471         return 0;
1472 }
1473
1474 static int init_subports(struct ipath_devdata *dd,
1475                          struct ipath_portdata *pd,
1476                          const struct ipath_user_info *uinfo)
1477 {
1478         int ret = 0;
1479         unsigned num_subports;
1480         size_t size;
1481
1482         /*
1483          * If the user is requesting zero subports,
1484          * skip the subport allocation.
1485          */
1486         if (uinfo->spu_subport_cnt <= 0)
1487                 goto bail;
1488
1489         /* Self-consistency check for ipath_compatible_subports() */
1490         if (ipath_supports_subports(IPATH_USER_SWMAJOR, IPATH_USER_SWMINOR) &&
1491             !ipath_compatible_subports(IPATH_USER_SWMAJOR,
1492                                        IPATH_USER_SWMINOR)) {
1493                 dev_info(&dd->pcidev->dev,
1494                          "Inconsistent ipath_compatible_subports()\n");
1495                 goto bail;
1496         }
1497
1498         /* Check for subport compatibility */
1499         if (!ipath_compatible_subports(uinfo->spu_userversion >> 16,
1500                                        uinfo->spu_userversion & 0xffff)) {
1501                 dev_info(&dd->pcidev->dev,
1502                          "Mismatched user version (%d.%d) and driver "
1503                          "version (%d.%d) while port sharing. Ensure "
1504                          "that driver and library are from the same "
1505                          "release.\n",
1506                          (int) (uinfo->spu_userversion >> 16),
1507                          (int) (uinfo->spu_userversion & 0xffff),
1508                          IPATH_USER_SWMAJOR,
1509                          IPATH_USER_SWMINOR);
1510                 goto bail;
1511         }
1512         if (uinfo->spu_subport_cnt > INFINIPATH_MAX_SUBPORT) {
1513                 ret = -EINVAL;
1514                 goto bail;
1515         }
1516
1517         num_subports = uinfo->spu_subport_cnt;
1518         pd->subport_uregbase = vmalloc(PAGE_SIZE * num_subports);
1519         if (!pd->subport_uregbase) {
1520                 ret = -ENOMEM;
1521                 goto bail;
1522         }
1523         /* Note: pd->port_rcvhdrq_size isn't initialized yet. */
1524         size = ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize *
1525                      sizeof(u32), PAGE_SIZE) * num_subports;
1526         pd->subport_rcvhdr_base = vmalloc(size);
1527         if (!pd->subport_rcvhdr_base) {
1528                 ret = -ENOMEM;
1529                 goto bail_ureg;
1530         }
1531
1532         pd->subport_rcvegrbuf = vmalloc(pd->port_rcvegrbuf_chunks *
1533                                         pd->port_rcvegrbuf_size *
1534                                         num_subports);
1535         if (!pd->subport_rcvegrbuf) {
1536                 ret = -ENOMEM;
1537                 goto bail_rhdr;
1538         }
1539
1540         pd->port_subport_cnt = uinfo->spu_subport_cnt;
1541         pd->port_subport_id = uinfo->spu_subport_id;
1542         pd->active_slaves = 1;
1543         set_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag);
1544         memset(pd->subport_uregbase, 0, PAGE_SIZE * num_subports);
1545         memset(pd->subport_rcvhdr_base, 0, size);
1546         memset(pd->subport_rcvegrbuf, 0, pd->port_rcvegrbuf_chunks *
1547                                          pd->port_rcvegrbuf_size *
1548                                          num_subports);
1549         goto bail;
1550
1551 bail_rhdr:
1552         vfree(pd->subport_rcvhdr_base);
1553 bail_ureg:
1554         vfree(pd->subport_uregbase);
1555         pd->subport_uregbase = NULL;
1556 bail:
1557         return ret;
1558 }
1559
1560 static int try_alloc_port(struct ipath_devdata *dd, int port,
1561                           struct file *fp,
1562                           const struct ipath_user_info *uinfo)
1563 {
1564         struct ipath_portdata *pd;
1565         int ret;
1566
1567         if (!(pd = dd->ipath_pd[port])) {
1568                 void *ptmp;
1569
1570                 pd = kzalloc(sizeof(struct ipath_portdata), GFP_KERNEL);
1571
1572                 /*
1573                  * Allocate memory for use in ipath_tid_update() just once
1574                  * at open, not per call.  Reduces cost of expected send
1575                  * setup.
1576                  */
1577                 ptmp = kmalloc(dd->ipath_rcvtidcnt * sizeof(u16) +
1578                                dd->ipath_rcvtidcnt * sizeof(struct page **),
1579                                GFP_KERNEL);
1580                 if (!pd || !ptmp) {
1581                         ipath_dev_err(dd, "Unable to allocate portdata "
1582                                       "memory, failing open\n");
1583                         ret = -ENOMEM;
1584                         kfree(pd);
1585                         kfree(ptmp);
1586                         goto bail;
1587                 }
1588                 dd->ipath_pd[port] = pd;
1589                 dd->ipath_pd[port]->port_port = port;
1590                 dd->ipath_pd[port]->port_dd = dd;
1591                 dd->ipath_pd[port]->port_tid_pg_list = ptmp;
1592                 init_waitqueue_head(&dd->ipath_pd[port]->port_wait);
1593         }
1594         if (!pd->port_cnt) {
1595                 pd->userversion = uinfo->spu_userversion;
1596                 init_user_egr_sizes(pd);
1597                 if ((ret = init_subports(dd, pd, uinfo)) != 0)
1598                         goto bail;
1599                 ipath_cdbg(PROC, "%s[%u] opened unit:port %u:%u\n",
1600                            current->comm, current->pid, dd->ipath_unit,
1601                            port);
1602                 pd->port_cnt = 1;
1603                 port_fp(fp) = pd;
1604                 pd->port_pid = current->pid;
1605                 strncpy(pd->port_comm, current->comm, sizeof(pd->port_comm));
1606                 ipath_chg_pioavailkernel(dd,
1607                         dd->ipath_pbufsport * (pd->port_port - 1),
1608                         dd->ipath_pbufsport, 0);
1609                 ipath_stats.sps_ports++;
1610                 ret = 0;
1611         } else
1612                 ret = -EBUSY;
1613
1614 bail:
1615         return ret;
1616 }
1617
1618 static inline int usable(struct ipath_devdata *dd)
1619 {
1620         return dd &&
1621                 (dd->ipath_flags & IPATH_PRESENT) &&
1622                 dd->ipath_kregbase &&
1623                 dd->ipath_lid &&
1624                 !(dd->ipath_flags & (IPATH_LINKDOWN | IPATH_DISABLED
1625                                      | IPATH_LINKUNK));
1626 }
1627
1628 static int find_free_port(int unit, struct file *fp,
1629                           const struct ipath_user_info *uinfo)
1630 {
1631         struct ipath_devdata *dd = ipath_lookup(unit);
1632         int ret, i;
1633
1634         if (!dd) {
1635                 ret = -ENODEV;
1636                 goto bail;
1637         }
1638
1639         if (!usable(dd)) {
1640                 ret = -ENETDOWN;
1641                 goto bail;
1642         }
1643
1644         for (i = 1; i < dd->ipath_cfgports; i++) {
1645                 ret = try_alloc_port(dd, i, fp, uinfo);
1646                 if (ret != -EBUSY)
1647                         goto bail;
1648         }
1649         ret = -EBUSY;
1650
1651 bail:
1652         return ret;
1653 }
1654
1655 static int find_best_unit(struct file *fp,
1656                           const struct ipath_user_info *uinfo)
1657 {
1658         int ret = 0, i, prefunit = -1, devmax;
1659         int maxofallports, npresent, nup;
1660         int ndev;
1661
1662         devmax = ipath_count_units(&npresent, &nup, &maxofallports);
1663
1664         /*
1665          * This code is present to allow a knowledgeable person to
1666          * specify the layout of processes to processors before opening
1667          * this driver, and then we'll assign the process to the "closest"
1668          * InfiniPath chip to that processor (we assume reasonable connectivity,
1669          * for now).  This code assumes that if affinity has been set
1670          * before this point, that at most one cpu is set; for now this
1671          * is reasonable.  I check for both cpus_empty() and cpus_full(),
1672          * in case some kernel variant sets none of the bits when no
1673          * affinity is set.  2.6.11 and 12 kernels have all present
1674          * cpus set.  Some day we'll have to fix it up further to handle
1675          * a cpu subset.  This algorithm fails for two HT chips connected
1676          * in tunnel fashion.  Eventually this needs real topology
1677          * information.  There may be some issues with dual core numbering
1678          * as well.  This needs more work prior to release.
1679          */
1680         if (!cpus_empty(current->cpus_allowed) &&
1681             !cpus_full(current->cpus_allowed)) {
1682                 int ncpus = num_online_cpus(), curcpu = -1, nset = 0;
1683                 for (i = 0; i < ncpus; i++)
1684                         if (cpu_isset(i, current->cpus_allowed)) {
1685                                 ipath_cdbg(PROC, "%s[%u] affinity set for "
1686                                            "cpu %d/%d\n", current->comm,
1687                                            current->pid, i, ncpus);
1688                                 curcpu = i;
1689                                 nset++;
1690                         }
1691                 if (curcpu != -1 && nset != ncpus) {
1692                         if (npresent) {
1693                                 prefunit = curcpu / (ncpus / npresent);
1694                                 ipath_cdbg(PROC,"%s[%u] %d chips, %d cpus, "
1695                                           "%d cpus/chip, select unit %d\n",
1696                                           current->comm, current->pid,
1697                                           npresent, ncpus, ncpus / npresent,
1698                                           prefunit);
1699                         }
1700                 }
1701         }
1702
1703         /*
1704          * user ports start at 1, kernel port is 0
1705          * For now, we do round-robin access across all chips
1706          */
1707
1708         if (prefunit != -1)
1709                 devmax = prefunit + 1;
1710 recheck:
1711         for (i = 1; i < maxofallports; i++) {
1712                 for (ndev = prefunit != -1 ? prefunit : 0; ndev < devmax;
1713                      ndev++) {
1714                         struct ipath_devdata *dd = ipath_lookup(ndev);
1715
1716                         if (!usable(dd))
1717                                 continue; /* can't use this unit */
1718                         if (i >= dd->ipath_cfgports)
1719                                 /*
1720                                  * Maxed out on users of this unit. Try
1721                                  * next.
1722                                  */
1723                                 continue;
1724                         ret = try_alloc_port(dd, i, fp, uinfo);
1725                         if (!ret)
1726                                 goto done;
1727                 }
1728         }
1729
1730         if (npresent) {
1731                 if (nup == 0) {
1732                         ret = -ENETDOWN;
1733                         ipath_dbg("No ports available (none initialized "
1734                                   "and ready)\n");
1735                 } else {
1736                         if (prefunit > 0) {
1737                                 /* if started above 0, retry from 0 */
1738                                 ipath_cdbg(PROC,
1739                                            "%s[%u] no ports on prefunit "
1740                                            "%d, clear and re-check\n",
1741                                            current->comm, current->pid,
1742                                            prefunit);
1743                                 devmax = ipath_count_units(NULL, NULL,
1744                                                            NULL);
1745                                 prefunit = -1;
1746                                 goto recheck;
1747                         }
1748                         ret = -EBUSY;
1749                         ipath_dbg("No ports available\n");
1750                 }
1751         } else {
1752                 ret = -ENXIO;
1753                 ipath_dbg("No boards found\n");
1754         }
1755
1756 done:
1757         return ret;
1758 }
1759
1760 static int find_shared_port(struct file *fp,
1761                             const struct ipath_user_info *uinfo)
1762 {
1763         int devmax, ndev, i;
1764         int ret = 0;
1765
1766         devmax = ipath_count_units(NULL, NULL, NULL);
1767
1768         for (ndev = 0; ndev < devmax; ndev++) {
1769                 struct ipath_devdata *dd = ipath_lookup(ndev);
1770
1771                 if (!usable(dd))
1772                         continue;
1773                 for (i = 1; i < dd->ipath_cfgports; i++) {
1774                         struct ipath_portdata *pd = dd->ipath_pd[i];
1775
1776                         /* Skip ports which are not yet open */
1777                         if (!pd || !pd->port_cnt)
1778                                 continue;
1779                         /* Skip port if it doesn't match the requested one */
1780                         if (pd->port_subport_id != uinfo->spu_subport_id)
1781                                 continue;
1782                         /* Verify the sharing process matches the master */
1783                         if (pd->port_subport_cnt != uinfo->spu_subport_cnt ||
1784                             pd->userversion != uinfo->spu_userversion ||
1785                             pd->port_cnt >= pd->port_subport_cnt) {
1786                                 ret = -EINVAL;
1787                                 goto done;
1788                         }
1789                         port_fp(fp) = pd;
1790                         subport_fp(fp) = pd->port_cnt++;
1791                         pd->port_subpid[subport_fp(fp)] = current->pid;
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         pd->port_lastrcvhdrqtail = -1;
1923         ipath_cdbg(VERBOSE, "Wrote port%d egrhead %x from tail regs\n",
1924                 pd->port_port, head32);
1925         pd->port_tidcursor = 0; /* start at beginning after open */
1926
1927         /* initialize poll variables... */
1928         pd->port_urgent = 0;
1929         pd->port_urgent_poll = 0;
1930         pd->port_hdrqfull_poll = pd->port_hdrqfull;
1931
1932         /*
1933          * now enable the port; the tail registers will be written to memory
1934          * by the chip as soon as it sees the write to
1935          * dd->ipath_kregs->kr_rcvctrl.  The update only happens on
1936          * transition from 0 to 1, so clear it first, then set it as part of
1937          * enabling the port.  This will (very briefly) affect any other
1938          * open ports, but it shouldn't be long enough to be an issue.
1939          * We explictly set the in-memory copy to 0 beforehand, so we don't
1940          * have to wait to be sure the DMA update has happened.
1941          */
1942         if (pd->port_rcvhdrtail_kvaddr)
1943                 ipath_clear_rcvhdrtail(pd);
1944         set_bit(dd->ipath_r_portenable_shift + pd->port_port,
1945                 &dd->ipath_rcvctrl);
1946         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1947                         dd->ipath_rcvctrl &
1948                         ~(1ULL << dd->ipath_r_tailupd_shift));
1949         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1950                          dd->ipath_rcvctrl);
1951         /* Notify any waiting slaves */
1952         if (pd->port_subport_cnt) {
1953                 clear_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag);
1954                 wake_up(&pd->port_wait);
1955         }
1956 done:
1957         return ret;
1958 }
1959
1960 /**
1961  * unlock_exptid - unlock any expected TID entries port still had in use
1962  * @pd: port
1963  *
1964  * We don't actually update the chip here, because we do a bulk update
1965  * below, using ipath_f_clear_tids.
1966  */
1967 static void unlock_expected_tids(struct ipath_portdata *pd)
1968 {
1969         struct ipath_devdata *dd = pd->port_dd;
1970         int port_tidbase = pd->port_port * dd->ipath_rcvtidcnt;
1971         int i, cnt = 0, maxtid = port_tidbase + dd->ipath_rcvtidcnt;
1972
1973         ipath_cdbg(VERBOSE, "Port %u unlocking any locked expTID pages\n",
1974                    pd->port_port);
1975         for (i = port_tidbase; i < maxtid; i++) {
1976                 if (!dd->ipath_pageshadow[i])
1977                         continue;
1978
1979                 pci_unmap_page(dd->pcidev, dd->ipath_physshadow[i],
1980                         PAGE_SIZE, PCI_DMA_FROMDEVICE);
1981                 ipath_release_user_pages_on_close(&dd->ipath_pageshadow[i],
1982                                                   1);
1983                 dd->ipath_pageshadow[i] = NULL;
1984                 cnt++;
1985                 ipath_stats.sps_pageunlocks++;
1986         }
1987         if (cnt)
1988                 ipath_cdbg(VERBOSE, "Port %u locked %u expTID entries\n",
1989                            pd->port_port, cnt);
1990
1991         if (ipath_stats.sps_pagelocks || ipath_stats.sps_pageunlocks)
1992                 ipath_cdbg(VERBOSE, "%llu pages locked, %llu unlocked\n",
1993                            (unsigned long long) ipath_stats.sps_pagelocks,
1994                            (unsigned long long)
1995                            ipath_stats.sps_pageunlocks);
1996 }
1997
1998 static int ipath_close(struct inode *in, struct file *fp)
1999 {
2000         int ret = 0;
2001         struct ipath_filedata *fd;
2002         struct ipath_portdata *pd;
2003         struct ipath_devdata *dd;
2004         unsigned port;
2005
2006         ipath_cdbg(VERBOSE, "close on dev %lx, private data %p\n",
2007                    (long)in->i_rdev, fp->private_data);
2008
2009         mutex_lock(&ipath_mutex);
2010
2011         fd = (struct ipath_filedata *) fp->private_data;
2012         fp->private_data = NULL;
2013         pd = fd->pd;
2014         if (!pd) {
2015                 mutex_unlock(&ipath_mutex);
2016                 goto bail;
2017         }
2018         if (--pd->port_cnt) {
2019                 /*
2020                  * XXX If the master closes the port before the slave(s),
2021                  * revoke the mmap for the eager receive queue so
2022                  * the slave(s) don't wait for receive data forever.
2023                  */
2024                 pd->active_slaves &= ~(1 << fd->subport);
2025                 pd->port_subpid[fd->subport] = 0;
2026                 mutex_unlock(&ipath_mutex);
2027                 goto bail;
2028         }
2029         port = pd->port_port;
2030         dd = pd->port_dd;
2031
2032         if (pd->port_hdrqfull) {
2033                 ipath_cdbg(PROC, "%s[%u] had %u rcvhdrqfull errors "
2034                            "during run\n", pd->port_comm, pd->port_pid,
2035                            pd->port_hdrqfull);
2036                 pd->port_hdrqfull = 0;
2037         }
2038
2039         if (pd->port_rcvwait_to || pd->port_piowait_to
2040             || pd->port_rcvnowait || pd->port_pionowait) {
2041                 ipath_cdbg(VERBOSE, "port%u, %u rcv, %u pio wait timeo; "
2042                            "%u rcv %u, pio already\n",
2043                            pd->port_port, pd->port_rcvwait_to,
2044                            pd->port_piowait_to, pd->port_rcvnowait,
2045                            pd->port_pionowait);
2046                 pd->port_rcvwait_to = pd->port_piowait_to =
2047                         pd->port_rcvnowait = pd->port_pionowait = 0;
2048         }
2049         if (pd->port_flag) {
2050                 ipath_dbg("port %u port_flag still set to 0x%lx\n",
2051                           pd->port_port, pd->port_flag);
2052                 pd->port_flag = 0;
2053         }
2054
2055         if (dd->ipath_kregbase) {
2056                 int i;
2057                 /* atomically clear receive enable port and intr avail. */
2058                 clear_bit(dd->ipath_r_portenable_shift + port,
2059                           &dd->ipath_rcvctrl);
2060                 clear_bit(pd->port_port + dd->ipath_r_intravail_shift,
2061                           &dd->ipath_rcvctrl);
2062                 ipath_write_kreg( dd, dd->ipath_kregs->kr_rcvctrl,
2063                         dd->ipath_rcvctrl);
2064                 /* and read back from chip to be sure that nothing
2065                  * else is in flight when we do the rest */
2066                 (void)ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
2067
2068                 /* clean up the pkeys for this port user */
2069                 ipath_clean_part_key(pd, dd);
2070                 /*
2071                  * be paranoid, and never write 0's to these, just use an
2072                  * unused part of the port 0 tail page.  Of course,
2073                  * rcvhdraddr points to a large chunk of memory, so this
2074                  * could still trash things, but at least it won't trash
2075                  * page 0, and by disabling the port, it should stop "soon",
2076                  * even if a packet or two is in already in flight after we
2077                  * disabled the port.
2078                  */
2079                 ipath_write_kreg_port(dd,
2080                         dd->ipath_kregs->kr_rcvhdrtailaddr, port,
2081                         dd->ipath_dummy_hdrq_phys);
2082                 ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdraddr,
2083                         pd->port_port, dd->ipath_dummy_hdrq_phys);
2084
2085                 i = dd->ipath_pbufsport * (port - 1);
2086                 ipath_disarm_piobufs(dd, i, dd->ipath_pbufsport);
2087                 ipath_chg_pioavailkernel(dd, i, dd->ipath_pbufsport, 1);
2088
2089                 dd->ipath_f_clear_tids(dd, pd->port_port);
2090
2091                 if (dd->ipath_pageshadow)
2092                         unlock_expected_tids(pd);
2093                 ipath_stats.sps_ports--;
2094                 ipath_cdbg(PROC, "%s[%u] closed port %u:%u\n",
2095                            pd->port_comm, pd->port_pid,
2096                            dd->ipath_unit, port);
2097         }
2098
2099         pd->port_pid = 0;
2100         dd->ipath_pd[pd->port_port] = NULL; /* before releasing mutex */
2101         mutex_unlock(&ipath_mutex);
2102         ipath_free_pddata(dd, pd); /* after releasing the mutex */
2103
2104 bail:
2105         kfree(fd);
2106         return ret;
2107 }
2108
2109 static int ipath_port_info(struct ipath_portdata *pd, u16 subport,
2110                            struct ipath_port_info __user *uinfo)
2111 {
2112         struct ipath_port_info info;
2113         int nup;
2114         int ret;
2115         size_t sz;
2116
2117         (void) ipath_count_units(NULL, &nup, NULL);
2118         info.num_active = nup;
2119         info.unit = pd->port_dd->ipath_unit;
2120         info.port = pd->port_port;
2121         info.subport = subport;
2122         /* Don't return new fields if old library opened the port. */
2123         if (ipath_supports_subports(pd->userversion >> 16,
2124                                     pd->userversion & 0xffff)) {
2125                 /* Number of user ports available for this device. */
2126                 info.num_ports = pd->port_dd->ipath_cfgports - 1;
2127                 info.num_subports = pd->port_subport_cnt;
2128                 sz = sizeof(info);
2129         } else
2130                 sz = sizeof(info) - 2 * sizeof(u16);
2131
2132         if (copy_to_user(uinfo, &info, sz)) {
2133                 ret = -EFAULT;
2134                 goto bail;
2135         }
2136         ret = 0;
2137
2138 bail:
2139         return ret;
2140 }
2141
2142 static int ipath_get_slave_info(struct ipath_portdata *pd,
2143                                 void __user *slave_mask_addr)
2144 {
2145         int ret = 0;
2146
2147         if (copy_to_user(slave_mask_addr, &pd->active_slaves, sizeof(u32)))
2148                 ret = -EFAULT;
2149         return ret;
2150 }
2151
2152 static ssize_t ipath_write(struct file *fp, const char __user *data,
2153                            size_t count, loff_t *off)
2154 {
2155         const struct ipath_cmd __user *ucmd;
2156         struct ipath_portdata *pd;
2157         const void __user *src;
2158         size_t consumed, copy;
2159         struct ipath_cmd cmd;
2160         ssize_t ret = 0;
2161         void *dest;
2162
2163         if (count < sizeof(cmd.type)) {
2164                 ret = -EINVAL;
2165                 goto bail;
2166         }
2167
2168         ucmd = (const struct ipath_cmd __user *) data;
2169
2170         if (copy_from_user(&cmd.type, &ucmd->type, sizeof(cmd.type))) {
2171                 ret = -EFAULT;
2172                 goto bail;
2173         }
2174
2175         consumed = sizeof(cmd.type);
2176
2177         switch (cmd.type) {
2178         case IPATH_CMD_ASSIGN_PORT:
2179         case __IPATH_CMD_USER_INIT:
2180         case IPATH_CMD_USER_INIT:
2181                 copy = sizeof(cmd.cmd.user_info);
2182                 dest = &cmd.cmd.user_info;
2183                 src = &ucmd->cmd.user_info;
2184                 break;
2185         case IPATH_CMD_RECV_CTRL:
2186                 copy = sizeof(cmd.cmd.recv_ctrl);
2187                 dest = &cmd.cmd.recv_ctrl;
2188                 src = &ucmd->cmd.recv_ctrl;
2189                 break;
2190         case IPATH_CMD_PORT_INFO:
2191                 copy = sizeof(cmd.cmd.port_info);
2192                 dest = &cmd.cmd.port_info;
2193                 src = &ucmd->cmd.port_info;
2194                 break;
2195         case IPATH_CMD_TID_UPDATE:
2196         case IPATH_CMD_TID_FREE:
2197                 copy = sizeof(cmd.cmd.tid_info);
2198                 dest = &cmd.cmd.tid_info;
2199                 src = &ucmd->cmd.tid_info;
2200                 break;
2201         case IPATH_CMD_SET_PART_KEY:
2202                 copy = sizeof(cmd.cmd.part_key);
2203                 dest = &cmd.cmd.part_key;
2204                 src = &ucmd->cmd.part_key;
2205                 break;
2206         case __IPATH_CMD_SLAVE_INFO:
2207                 copy = sizeof(cmd.cmd.slave_mask_addr);
2208                 dest = &cmd.cmd.slave_mask_addr;
2209                 src = &ucmd->cmd.slave_mask_addr;
2210                 break;
2211         case IPATH_CMD_PIOAVAILUPD:     // force an update of PIOAvail reg
2212                 copy = 0;
2213                 src = NULL;
2214                 dest = NULL;
2215                 break;
2216         case IPATH_CMD_POLL_TYPE:
2217                 copy = sizeof(cmd.cmd.poll_type);
2218                 dest = &cmd.cmd.poll_type;
2219                 src = &ucmd->cmd.poll_type;
2220                 break;
2221         case IPATH_CMD_ARMLAUNCH_CTRL:
2222                 copy = sizeof(cmd.cmd.armlaunch_ctrl);
2223                 dest = &cmd.cmd.armlaunch_ctrl;
2224                 src = &ucmd->cmd.armlaunch_ctrl;
2225                 break;
2226         default:
2227                 ret = -EINVAL;
2228                 goto bail;
2229         }
2230
2231         if (copy) {
2232                 if ((count - consumed) < copy) {
2233                         ret = -EINVAL;
2234                         goto bail;
2235                 }
2236
2237                 if (copy_from_user(dest, src, copy)) {
2238                         ret = -EFAULT;
2239                         goto bail;
2240                 }
2241
2242                 consumed += copy;
2243         }
2244
2245         pd = port_fp(fp);
2246         if (!pd && cmd.type != __IPATH_CMD_USER_INIT &&
2247                 cmd.type != IPATH_CMD_ASSIGN_PORT) {
2248                 ret = -EINVAL;
2249                 goto bail;
2250         }
2251
2252         switch (cmd.type) {
2253         case IPATH_CMD_ASSIGN_PORT:
2254                 ret = ipath_assign_port(fp, &cmd.cmd.user_info);
2255                 if (ret)
2256                         goto bail;
2257                 break;
2258         case __IPATH_CMD_USER_INIT:
2259                 /* backwards compatibility, get port first */
2260                 ret = ipath_assign_port(fp, &cmd.cmd.user_info);
2261                 if (ret)
2262                         goto bail;
2263                 /* and fall through to current version. */
2264         case IPATH_CMD_USER_INIT:
2265                 ret = ipath_do_user_init(fp, &cmd.cmd.user_info);
2266                 if (ret)
2267                         goto bail;
2268                 ret = ipath_get_base_info(
2269                         fp, (void __user *) (unsigned long)
2270                         cmd.cmd.user_info.spu_base_info,
2271                         cmd.cmd.user_info.spu_base_info_size);
2272                 break;
2273         case IPATH_CMD_RECV_CTRL:
2274                 ret = ipath_manage_rcvq(pd, subport_fp(fp), cmd.cmd.recv_ctrl);
2275                 break;
2276         case IPATH_CMD_PORT_INFO:
2277                 ret = ipath_port_info(pd, subport_fp(fp),
2278                                       (struct ipath_port_info __user *)
2279                                       (unsigned long) cmd.cmd.port_info);
2280                 break;
2281         case IPATH_CMD_TID_UPDATE:
2282                 ret = ipath_tid_update(pd, fp, &cmd.cmd.tid_info);
2283                 break;
2284         case IPATH_CMD_TID_FREE:
2285                 ret = ipath_tid_free(pd, subport_fp(fp), &cmd.cmd.tid_info);
2286                 break;
2287         case IPATH_CMD_SET_PART_KEY:
2288                 ret = ipath_set_part_key(pd, cmd.cmd.part_key);
2289                 break;
2290         case __IPATH_CMD_SLAVE_INFO:
2291                 ret = ipath_get_slave_info(pd,
2292                                            (void __user *) (unsigned long)
2293                                            cmd.cmd.slave_mask_addr);
2294                 break;
2295         case IPATH_CMD_PIOAVAILUPD:
2296                 ipath_force_pio_avail_update(pd->port_dd);
2297                 break;
2298         case IPATH_CMD_POLL_TYPE:
2299                 pd->poll_type = cmd.cmd.poll_type;
2300                 break;
2301         case IPATH_CMD_ARMLAUNCH_CTRL:
2302                 if (cmd.cmd.armlaunch_ctrl)
2303                         ipath_enable_armlaunch(pd->port_dd);
2304                 else
2305                         ipath_disable_armlaunch(pd->port_dd);
2306                 break;
2307         }
2308
2309         if (ret >= 0)
2310                 ret = consumed;
2311
2312 bail:
2313         return ret;
2314 }
2315
2316 static struct class *ipath_class;
2317
2318 static int init_cdev(int minor, char *name, const struct file_operations *fops,
2319                      struct cdev **cdevp, struct class_device **class_devp)
2320 {
2321         const dev_t dev = MKDEV(IPATH_MAJOR, minor);
2322         struct cdev *cdev = NULL;
2323         struct class_device *class_dev = NULL;
2324         int ret;
2325
2326         cdev = cdev_alloc();
2327         if (!cdev) {
2328                 printk(KERN_ERR IPATH_DRV_NAME
2329                        ": Could not allocate cdev for minor %d, %s\n",
2330                        minor, name);
2331                 ret = -ENOMEM;
2332                 goto done;
2333         }
2334
2335         cdev->owner = THIS_MODULE;
2336         cdev->ops = fops;
2337         kobject_set_name(&cdev->kobj, name);
2338
2339         ret = cdev_add(cdev, dev, 1);
2340         if (ret < 0) {
2341                 printk(KERN_ERR IPATH_DRV_NAME
2342                        ": Could not add cdev for minor %d, %s (err %d)\n",
2343                        minor, name, -ret);
2344                 goto err_cdev;
2345         }
2346
2347         class_dev = class_device_create(ipath_class, NULL, dev, NULL, name);
2348
2349         if (IS_ERR(class_dev)) {
2350                 ret = PTR_ERR(class_dev);
2351                 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2352                        "class_dev for minor %d, %s (err %d)\n",
2353                        minor, name, -ret);
2354                 goto err_cdev;
2355         }
2356
2357         goto done;
2358
2359 err_cdev:
2360         cdev_del(cdev);
2361         cdev = NULL;
2362
2363 done:
2364         if (ret >= 0) {
2365                 *cdevp = cdev;
2366                 *class_devp = class_dev;
2367         } else {
2368                 *cdevp = NULL;
2369                 *class_devp = NULL;
2370         }
2371
2372         return ret;
2373 }
2374
2375 int ipath_cdev_init(int minor, char *name, const struct file_operations *fops,
2376                     struct cdev **cdevp, struct class_device **class_devp)
2377 {
2378         return init_cdev(minor, name, fops, cdevp, class_devp);
2379 }
2380
2381 static void cleanup_cdev(struct cdev **cdevp,
2382                          struct class_device **class_devp)
2383 {
2384         struct class_device *class_dev = *class_devp;
2385
2386         if (class_dev) {
2387                 class_device_unregister(class_dev);
2388                 *class_devp = NULL;
2389         }
2390
2391         if (*cdevp) {
2392                 cdev_del(*cdevp);
2393                 *cdevp = NULL;
2394         }
2395 }
2396
2397 void ipath_cdev_cleanup(struct cdev **cdevp,
2398                         struct class_device **class_devp)
2399 {
2400         cleanup_cdev(cdevp, class_devp);
2401 }
2402
2403 static struct cdev *wildcard_cdev;
2404 static struct class_device *wildcard_class_dev;
2405
2406 static const dev_t dev = MKDEV(IPATH_MAJOR, 0);
2407
2408 static int user_init(void)
2409 {
2410         int ret;
2411
2412         ret = register_chrdev_region(dev, IPATH_NMINORS, IPATH_DRV_NAME);
2413         if (ret < 0) {
2414                 printk(KERN_ERR IPATH_DRV_NAME ": Could not register "
2415                        "chrdev region (err %d)\n", -ret);
2416                 goto done;
2417         }
2418
2419         ipath_class = class_create(THIS_MODULE, IPATH_DRV_NAME);
2420
2421         if (IS_ERR(ipath_class)) {
2422                 ret = PTR_ERR(ipath_class);
2423                 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2424                        "device class (err %d)\n", -ret);
2425                 goto bail;
2426         }
2427
2428         goto done;
2429 bail:
2430         unregister_chrdev_region(dev, IPATH_NMINORS);
2431 done:
2432         return ret;
2433 }
2434
2435 static void user_cleanup(void)
2436 {
2437         if (ipath_class) {
2438                 class_destroy(ipath_class);
2439                 ipath_class = NULL;
2440         }
2441
2442         unregister_chrdev_region(dev, IPATH_NMINORS);
2443 }
2444
2445 static atomic_t user_count = ATOMIC_INIT(0);
2446 static atomic_t user_setup = ATOMIC_INIT(0);
2447
2448 int ipath_user_add(struct ipath_devdata *dd)
2449 {
2450         char name[10];
2451         int ret;
2452
2453         if (atomic_inc_return(&user_count) == 1) {
2454                 ret = user_init();
2455                 if (ret < 0) {
2456                         ipath_dev_err(dd, "Unable to set up user support: "
2457                                       "error %d\n", -ret);
2458                         goto bail;
2459                 }
2460                 ret = init_cdev(0, "ipath", &ipath_file_ops, &wildcard_cdev,
2461                                 &wildcard_class_dev);
2462                 if (ret < 0) {
2463                         ipath_dev_err(dd, "Could not create wildcard "
2464                                       "minor: error %d\n", -ret);
2465                         goto bail_user;
2466                 }
2467
2468                 atomic_set(&user_setup, 1);
2469         }
2470
2471         snprintf(name, sizeof(name), "ipath%d", dd->ipath_unit);
2472
2473         ret = init_cdev(dd->ipath_unit + 1, name, &ipath_file_ops,
2474                         &dd->user_cdev, &dd->user_class_dev);
2475         if (ret < 0)
2476                 ipath_dev_err(dd, "Could not create user minor %d, %s\n",
2477                               dd->ipath_unit + 1, name);
2478
2479         goto bail;
2480
2481 bail_user:
2482         user_cleanup();
2483 bail:
2484         return ret;
2485 }
2486
2487 void ipath_user_remove(struct ipath_devdata *dd)
2488 {
2489         cleanup_cdev(&dd->user_cdev, &dd->user_class_dev);
2490
2491         if (atomic_dec_return(&user_count) == 0) {
2492                 if (atomic_read(&user_setup) == 0)
2493                         goto bail;
2494
2495                 cleanup_cdev(&wildcard_cdev, &wildcard_class_dev);
2496                 user_cleanup();
2497
2498                 atomic_set(&user_setup, 0);
2499         }
2500 bail:
2501         return;
2502 }