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
 188         /*
 189          * Set the PIO avail update threshold to no larger
 190          * than the number of buffers per process. Note that
 191          * we decrease it here, but won't ever increase it.
 192          */
 193         if (dd->ipath_pioupd_thresh &&
 194             kinfo->spi_piocnt < dd->ipath_pioupd_thresh) {
 195                 unsigned long flags;
 196
 197                 dd->ipath_pioupd_thresh = kinfo->spi_piocnt;
 198                 ipath_dbg("Decreased pio update threshold to %u\n",
 199                         dd->ipath_pioupd_thresh);
 200                 spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
 201                 dd->ipath_sendctrl &= ~(INFINIPATH_S_UPDTHRESH_MASK
 202                         << INFINIPATH_S_UPDTHRESH_SHIFT);
 203                 dd->ipath_sendctrl |= dd->ipath_pioupd_thresh
 204                         << INFINIPATH_S_UPDTHRESH_SHIFT;
 205                 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
 206                         dd->ipath_sendctrl);
 207                 spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
 208         }
 209
 210         if (shared) {
 211                 kinfo->spi_port_uregbase = (u64) dd->ipath_uregbase +
 212                         dd->ipath_ureg_align * pd->port_port;
 213                 kinfo->spi_port_rcvegrbuf = kinfo->spi_rcv_egrbufs;
 214                 kinfo->spi_port_rcvhdr_base = kinfo->spi_rcvhdr_base;
 215                 kinfo->spi_port_rcvhdr_tailaddr = kinfo->spi_rcvhdr_tailaddr;
 216
 217                 kinfo->__spi_uregbase = cvt_kvaddr(pd->subport_uregbase +
 218                         PAGE_SIZE * subport_fp(fp));
 219
 220                 kinfo->spi_rcvhdr_base = cvt_kvaddr(pd->subport_rcvhdr_base +
 221                         pd->port_rcvhdrq_size * subport_fp(fp));
 222                 kinfo->spi_rcvhdr_tailaddr = 0;
 223                 kinfo->spi_rcv_egrbufs = cvt_kvaddr(pd->subport_rcvegrbuf +
 224                         pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size *
 225                         subport_fp(fp));
 226
 227                 kinfo->spi_subport_uregbase =
 228                         cvt_kvaddr(pd->subport_uregbase);
 229                 kinfo->spi_subport_rcvegrbuf =
 230                         cvt_kvaddr(pd->subport_rcvegrbuf);
 231                 kinfo->spi_subport_rcvhdr_base =
 232                         cvt_kvaddr(pd->subport_rcvhdr_base);
 233                 ipath_cdbg(PROC, "port %u flags %x %llx %llx %llx\n",
 234                         kinfo->spi_port, kinfo->spi_runtime_flags,
 235                         (unsigned long long) kinfo->spi_subport_uregbase,
 236                         (unsigned long long) kinfo->spi_subport_rcvegrbuf,
 237                         (unsigned long long) kinfo->spi_subport_rcvhdr_base);
 238         }
 239
 240         kinfo->spi_pioindex = (kinfo->spi_piobufbase - dd->ipath_piobufbase) /
 241                 dd->ipath_palign;
 242         kinfo->spi_pioalign = dd->ipath_palign;
 243
 244         kinfo->spi_qpair = IPATH_KD_QP;
 245         /*
 246          * user mode PIO buffers are always 2KB, even when 4KB can
 247          * be received, and sent via the kernel; this is ibmaxlen
 248          * for 2K MTU.
 249          */
 250         kinfo->spi_piosize = dd->ipath_piosize2k - 2 * sizeof(u32);
 251         kinfo->spi_mtu = dd->ipath_ibmaxlen;    /* maxlen, not ibmtu */
 252         kinfo->spi_port = pd->port_port;
 253         kinfo->spi_subport = subport_fp(fp);
 254         kinfo->spi_sw_version = IPATH_KERN_SWVERSION;
 255         kinfo->spi_hw_version = dd->ipath_revision;
 256
 257         if (master) {
 258                 kinfo->spi_runtime_flags |= IPATH_RUNTIME_MASTER;
 259         }
 260
 261         sz = (ubase_size < sizeof(*kinfo)) ? ubase_size : sizeof(*kinfo);
 262         if (copy_to_user(ubase, kinfo, sz))
 263                 ret = -EFAULT;
 264
 265 bail:
 266         kfree(kinfo);
 267         return ret;
 268 }
 269
 270 /**
 271  * ipath_tid_update - update a port TID
 272  * @pd: the port
 273  * @fp: the ipath device file
 274  * @ti: the TID information
 275  *
 276  * The new implementation as of Oct 2004 is that the driver assigns
 277  * the tid and returns it to the caller.   To make it easier to
 278  * catch bugs, and to reduce search time, we keep a cursor for
 279  * each port, walking the shadow tid array to find one that's not
 280  * in use.
 281  *
 282  * For now, if we can't allocate the full list, we fail, although
 283  * in the long run, we'll allocate as many as we can, and the
 284  * caller will deal with that by trying the remaining pages later.
 285  * That means that when we fail, we have to mark the tids as not in
 286  * use again, in our shadow copy.
 287  *
 288  * It's up to the caller to free the tids when they are done.
 289  * We'll unlock the pages as they free them.
 290  *
 291  * Also, right now we are locking one page at a time, but since
 292  * the intended use of this routine is for a single group of
 293  * virtually contiguous pages, that should change to improve
 294  * performance.
 295  */
 296 static int ipath_tid_update(struct ipath_portdata *pd, struct file *fp,
 297                             const struct ipath_tid_info *ti)
 298 {
 299         int ret = 0, ntids;
 300         u32 tid, porttid, cnt, i, tidcnt, tidoff;
 301         u16 *tidlist;
 302         struct ipath_devdata *dd = pd->port_dd;
 303         u64 physaddr;
 304         unsigned long vaddr;
 305         u64 __iomem *tidbase;
 306         unsigned long tidmap[8];
 307         struct page **pagep = NULL;
 308         unsigned subport = subport_fp(fp);
 309
 310         if (!dd->ipath_pageshadow) {
 311                 ret = -ENOMEM;
 312                 goto done;
 313         }
 314
 315         cnt = ti->tidcnt;
 316         if (!cnt) {
 317                 ipath_dbg("After copyin, tidcnt 0, tidlist %llx\n",
 318                           (unsigned long long) ti->tidlist);
 319                 /*
 320                  * Should we treat as success?  likely a bug
 321                  */
 322                 ret = -EFAULT;
 323                 goto done;
 324         }
 325         porttid = pd->port_port * dd->ipath_rcvtidcnt;
 326         if (!pd->port_subport_cnt) {
 327                 tidcnt = dd->ipath_rcvtidcnt;
 328                 tid = pd->port_tidcursor;
 329                 tidoff = 0;
 330         } else if (!subport) {
 331                 tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
 332                          (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
 333                 tidoff = dd->ipath_rcvtidcnt - tidcnt;
 334                 porttid += tidoff;
 335                 tid = tidcursor_fp(fp);
 336         } else {
 337                 tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
 338                 tidoff = tidcnt * (subport - 1);
 339                 porttid += tidoff;
 340                 tid = tidcursor_fp(fp);
 341         }
 342         if (cnt > tidcnt) {
 343                 /* make sure it all fits in port_tid_pg_list */
 344                 dev_info(&dd->pcidev->dev, "Process tried to allocate %u "
 345                          "TIDs, only trying max (%u)\n", cnt, tidcnt);
 346                 cnt = tidcnt;
 347         }
 348         pagep = &((struct page **) pd->port_tid_pg_list)[tidoff];
 349         tidlist = &((u16 *) &pagep[dd->ipath_rcvtidcnt])[tidoff];
 350
 351         memset(tidmap, 0, sizeof(tidmap));
 352         /* before decrement; chip actual # */
 353         ntids = tidcnt;
 354         tidbase = (u64 __iomem *) (((char __iomem *) dd->ipath_kregbase) +
 355                                    dd->ipath_rcvtidbase +
 356                                    porttid * sizeof(*tidbase));
 357
 358         ipath_cdbg(VERBOSE, "Port%u %u tids, cursor %u, tidbase %p\n",
 359                    pd->port_port, cnt, tid, tidbase);
 360
 361         /* virtual address of first page in transfer */
 362         vaddr = ti->tidvaddr;
 363         if (!access_ok(VERIFY_WRITE, (void __user *) vaddr,
 364                        cnt * PAGE_SIZE)) {
 365                 ipath_dbg("Fail vaddr %p, %u pages, !access_ok\n",
 366                           (void *)vaddr, cnt);
 367                 ret = -EFAULT;
 368                 goto done;
 369         }
 370         ret = ipath_get_user_pages(vaddr, cnt, pagep);
 371         if (ret) {
 372                 if (ret == -EBUSY) {
 373                         ipath_dbg("Failed to lock addr %p, %u pages "
 374                                   "(already locked)\n",
 375                                   (void *) vaddr, cnt);
 376                         /*
 377                          * for now, continue, and see what happens but with
 378                          * the new implementation, this should never happen,
 379                          * unless perhaps the user has mpin'ed the pages
 380                          * themselves (something we need to test)
 381                          */
 382                         ret = 0;
 383                 } else {
 384                         dev_info(&dd->pcidev->dev,
 385                                  "Failed to lock addr %p, %u pages: "
 386                                  "errno %d\n", (void *) vaddr, cnt, -ret);
 387                         goto done;
 388                 }
 389         }
 390         for (i = 0; i < cnt; i++, vaddr += PAGE_SIZE) {
 391                 for (; ntids--; tid++) {
 392                         if (tid == tidcnt)
 393                                 tid = 0;
 394                         if (!dd->ipath_pageshadow[porttid + tid])
 395                                 break;
 396                 }
 397                 if (ntids < 0) {
 398                         /*
 399                          * oops, wrapped all the way through their TIDs,
 400                          * and didn't have enough free; see comments at
 401                          * start of routine
 402                          */
 403                         ipath_dbg("Not enough free TIDs for %u pages "
 404                                   "(index %d), failing\n", cnt, i);
 405                         i--;    /* last tidlist[i] not filled in */
 406                         ret = -ENOMEM;
 407                         break;
 408                 }
 409                 tidlist[i] = tid + tidoff;
 410                 ipath_cdbg(VERBOSE, "Updating idx %u to TID %u, "
 411                            "vaddr %lx\n", i, tid + tidoff, vaddr);
 412                 /* we "know" system pages and TID pages are same size */
 413                 dd->ipath_pageshadow[porttid + tid] = pagep[i];
 414                 dd->ipath_physshadow[porttid + tid] = ipath_map_page(
 415                         dd->pcidev, pagep[i], 0, PAGE_SIZE,
 416                         PCI_DMA_FROMDEVICE);
 417                 /*
 418                  * don't need atomic or it's overhead
 419                  */
 420                 __set_bit(tid, tidmap);
 421                 physaddr = dd->ipath_physshadow[porttid + tid];
 422                 ipath_stats.sps_pagelocks++;
 423                 ipath_cdbg(VERBOSE,
 424                            "TID %u, vaddr %lx, physaddr %llx pgp %p\n",
 425                            tid, vaddr, (unsigned long long) physaddr,
 426                            pagep[i]);
 427                 dd->ipath_f_put_tid(dd, &tidbase[tid], RCVHQ_RCV_TYPE_EXPECTED,
 428                                     physaddr);
 429                 /*
 430                  * don't check this tid in ipath_portshadow, since we
 431                  * just filled it in; start with the next one.
 432                  */
 433                 tid++;
 434         }
 435
 436         if (ret) {
 437                 u32 limit;
 438         cleanup:
 439                 /* jump here if copy out of updated info failed... */
 440                 ipath_dbg("After failure (ret=%d), undo %d of %d entries\n",
 441                           -ret, i, cnt);
 442                 /* same code that's in ipath_free_tid() */
 443                 limit = sizeof(tidmap) * BITS_PER_BYTE;
 444                 if (limit > tidcnt)
 445                         /* just in case size changes in future */
 446                         limit = tidcnt;
 447                 tid = find_first_bit((const unsigned long *)tidmap, limit);
 448                 for (; tid < limit; tid++) {
 449                         if (!test_bit(tid, tidmap))
 450                                 continue;
 451                         if (dd->ipath_pageshadow[porttid + tid]) {
 452                                 ipath_cdbg(VERBOSE, "Freeing TID %u\n",
 453                                            tid);
 454                                 dd->ipath_f_put_tid(dd, &tidbase[tid],
 455                                                     RCVHQ_RCV_TYPE_EXPECTED,
 456                                                     dd->ipath_tidinvalid);
 457                                 pci_unmap_page(dd->pcidev,
 458                                         dd->ipath_physshadow[porttid + tid],
 459                                         PAGE_SIZE, PCI_DMA_FROMDEVICE);
 460                                 dd->ipath_pageshadow[porttid + tid] = NULL;
 461                                 ipath_stats.sps_pageunlocks++;
 462                         }
 463                 }
 464                 ipath_release_user_pages(pagep, cnt);
 465         } else {
 466                 /*
 467                  * Copy the updated array, with ipath_tid's filled in, back
 468                  * to user.  Since we did the copy in already, this "should
 469                  * never fail" If it does, we have to clean up...
 470                  */
 471                 if (copy_to_user((void __user *)
 472                                  (unsigned long) ti->tidlist,
 473                                  tidlist, cnt * sizeof(*tidlist))) {
 474                         ret = -EFAULT;
 475                         goto cleanup;
 476                 }
 477                 if (copy_to_user((void __user *) (unsigned long) ti->tidmap,
 478                                  tidmap, sizeof tidmap)) {
 479                         ret = -EFAULT;
 480                         goto cleanup;
 481                 }
 482                 if (tid == tidcnt)
 483                         tid = 0;
 484                 if (!pd->port_subport_cnt)
 485                         pd->port_tidcursor = tid;
 486                 else
 487                         tidcursor_fp(fp) = tid;
 488         }
 489
 490 done:
 491         if (ret)
 492                 ipath_dbg("Failed to map %u TID pages, failing with %d\n",
 493                           ti->tidcnt, -ret);
 494         return ret;
 495 }
 496
 497 /**
 498  * ipath_tid_free - free a port TID
 499  * @pd: the port
 500  * @subport: the subport
 501  * @ti: the TID info
 502  *
 503  * right now we are unlocking one page at a time, but since
 504  * the intended use of this routine is for a single group of
 505  * virtually contiguous pages, that should change to improve
 506  * performance.  We check that the TID is in range for this port
 507  * but otherwise don't check validity; if user has an error and
 508  * frees the wrong tid, it's only their own data that can thereby
 509  * be corrupted.  We do check that the TID was in use, for sanity
 510  * We always use our idea of the saved address, not the address that
 511  * they pass in to us.
 512  */
 513
 514 static int ipath_tid_free(struct ipath_portdata *pd, unsigned subport,
 515                           const struct ipath_tid_info *ti)
 516 {
 517         int ret = 0;
 518         u32 tid, porttid, cnt, limit, tidcnt;
 519         struct ipath_devdata *dd = pd->port_dd;
 520         u64 __iomem *tidbase;
 521         unsigned long tidmap[8];
 522
 523         if (!dd->ipath_pageshadow) {
 524                 ret = -ENOMEM;
 525                 goto done;
 526         }
 527
 528         if (copy_from_user(tidmap, (void __user *)(unsigned long)ti->tidmap,
 529                            sizeof tidmap)) {
 530                 ret = -EFAULT;
 531                 goto done;
 532         }
 533
 534         porttid = pd->port_port * dd->ipath_rcvtidcnt;
 535         if (!pd->port_subport_cnt)
 536                 tidcnt = dd->ipath_rcvtidcnt;
 537         else if (!subport) {
 538                 tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
 539                          (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
 540                 porttid += dd->ipath_rcvtidcnt - tidcnt;
 541         } else {
 542                 tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
 543                 porttid += tidcnt * (subport - 1);
 544         }
 545         tidbase = (u64 __iomem *) ((char __iomem *)(dd->ipath_kregbase) +
 546                                    dd->ipath_rcvtidbase +
 547                                    porttid * sizeof(*tidbase));
 548
 549         limit = sizeof(tidmap) * BITS_PER_BYTE;
 550         if (limit > tidcnt)
 551                 /* just in case size changes in future */
 552                 limit = tidcnt;
 553         tid = find_first_bit(tidmap, limit);
 554         ipath_cdbg(VERBOSE, "Port%u free %u tids; first bit (max=%d) "
 555                    "set is %d, porttid %u\n", pd->port_port, ti->tidcnt,
 556                    limit, tid, porttid);
 557         for (cnt = 0; tid < limit; tid++) {
 558                 /*
 559                  * small optimization; if we detect a run of 3 or so without
 560                  * any set, use find_first_bit again.  That's mainly to
 561                  * accelerate the case where we wrapped, so we have some at
 562                  * the beginning, and some at the end, and a big gap
 563                  * in the middle.
 564                  */
 565                 if (!test_bit(tid, tidmap))
 566                         continue;
 567                 cnt++;
 568                 if (dd->ipath_pageshadow[porttid + tid]) {
 569                         struct page *p;
 570                         p = dd->ipath_pageshadow[porttid + tid];
 571                         dd->ipath_pageshadow[porttid + tid] = NULL;
 572                         ipath_cdbg(VERBOSE, "PID %u freeing TID %u\n",
 573                                    pd->port_pid, tid);
 574                         dd->ipath_f_put_tid(dd, &tidbase[tid],
 575                                             RCVHQ_RCV_TYPE_EXPECTED,
 576                                             dd->ipath_tidinvalid);
 577                         pci_unmap_page(dd->pcidev,
 578                                 dd->ipath_physshadow[porttid + tid],
 579                                 PAGE_SIZE, PCI_DMA_FROMDEVICE);
 580                         ipath_release_user_pages(&p, 1);
 581                         ipath_stats.sps_pageunlocks++;
 582                 } else
 583                         ipath_dbg("Unused tid %u, ignoring\n", tid);
 584         }
 585         if (cnt != ti->tidcnt)
 586                 ipath_dbg("passed in tidcnt %d, only %d bits set in map\n",
 587                           ti->tidcnt, cnt);
 588 done:
 589         if (ret)
 590                 ipath_dbg("Failed to unmap %u TID pages, failing with %d\n",
 591                           ti->tidcnt, -ret);
 592         return ret;
 593 }
 594
 595 /**
 596  * ipath_set_part_key - set a partition key
 597  * @pd: the port
 598  * @key: the key
 599  *
 600  * We can have up to 4 active at a time (other than the default, which is
 601  * always allowed).  This is somewhat tricky, since multiple ports may set
 602  * the same key, so we reference count them, and clean up at exit.  All 4
 603  * partition keys are packed into a single infinipath register.  It's an
 604  * error for a process to set the same pkey multiple times.  We provide no
 605  * mechanism to de-allocate a pkey at this time, we may eventually need to
 606  * do that.  I've used the atomic operations, and no locking, and only make
 607  * a single pass through what's available.  This should be more than
 608  * adequate for some time. I'll think about spinlocks or the like if and as
 609  * it's necessary.
 610  */
 611 static int ipath_set_part_key(struct ipath_portdata *pd, u16 key)
 612 {
 613         struct ipath_devdata *dd = pd->port_dd;
 614         int i, any = 0, pidx = -1;
 615         u16 lkey = key & 0x7FFF;
 616         int ret;
 617
 618         if (lkey == (IPATH_DEFAULT_P_KEY & 0x7FFF)) {
 619                 /* nothing to do; this key always valid */
 620                 ret = 0;
 621                 goto bail;
 622         }
 623
 624         ipath_cdbg(VERBOSE, "p%u try to set pkey %hx, current keys "
 625                    "%hx:%x %hx:%x %hx:%x %hx:%x\n",
 626                    pd->port_port, key, dd->ipath_pkeys[0],
 627                    atomic_read(&dd->ipath_pkeyrefs[0]), dd->ipath_pkeys[1],
 628                    atomic_read(&dd->ipath_pkeyrefs[1]), dd->ipath_pkeys[2],
 629                    atomic_read(&dd->ipath_pkeyrefs[2]), dd->ipath_pkeys[3],
 630                    atomic_read(&dd->ipath_pkeyrefs[3]));
 631
 632         if (!lkey) {
 633                 ipath_cdbg(PROC, "p%u tries to set key 0, not allowed\n",
 634                            pd->port_port);
 635                 ret = -EINVAL;
 636                 goto bail;
 637         }
 638
 639         /*
 640          * Set the full membership bit, because it has to be
 641          * set in the register or the packet, and it seems
 642          * cleaner to set in the register than to force all
 643          * callers to set it. (see bug 4331)
 644          */
 645         key |= 0x8000;
 646
 647         for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
 648                 if (!pd->port_pkeys[i] && pidx == -1)
 649                         pidx = i;
 650                 if (pd->port_pkeys[i] == key) {
 651                         ipath_cdbg(VERBOSE, "p%u tries to set same pkey "
 652                                    "(%x) more than once\n",
 653                                    pd->port_port, key);
 654                         ret = -EEXIST;
 655                         goto bail;
 656                 }
 657         }
 658         if (pidx == -1) {
 659                 ipath_dbg("All pkeys for port %u already in use, "
 660                           "can't set %x\n", pd->port_port, key);
 661                 ret = -EBUSY;
 662                 goto bail;
 663         }
 664         for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
 665                 if (!dd->ipath_pkeys[i]) {
 666                         any++;
 667                         continue;
 668                 }
 669                 if (dd->ipath_pkeys[i] == key) {
 670                         atomic_t *pkrefs = &dd->ipath_pkeyrefs[i];
 671
 672                         if (atomic_inc_return(pkrefs) > 1) {
 673                                 pd->port_pkeys[pidx] = key;
 674                                 ipath_cdbg(VERBOSE, "p%u set key %x "
 675                                            "matches #%d, count now %d\n",
 676                                            pd->port_port, key, i,
 677                                            atomic_read(pkrefs));
 678                                 ret = 0;
 679                                 goto bail;
 680                         } else {
 681                                 /*
 682                                  * lost race, decrement count, catch below
 683                                  */
 684                                 atomic_dec(pkrefs);
 685                                 ipath_cdbg(VERBOSE, "Lost race, count was "
 686                                            "0, after dec, it's %d\n",
 687                                            atomic_read(pkrefs));
 688                                 any++;
 689                         }
 690                 }
 691                 if ((dd->ipath_pkeys[i] & 0x7FFF) == lkey) {
 692                         /*
 693                          * It makes no sense to have both the limited and
 694                          * full membership PKEY set at the same time since
 695                          * the unlimited one will disable the limited one.
 696                          */
 697                         ret = -EEXIST;
 698                         goto bail;
 699                 }
 700         }
 701         if (!any) {
 702                 ipath_dbg("port %u, all pkeys already in use, "
 703                           "can't set %x\n", pd->port_port, key);
 704                 ret = -EBUSY;
 705                 goto bail;
 706         }
 707         for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
 708                 if (!dd->ipath_pkeys[i] &&
 709                     atomic_inc_return(&dd->ipath_pkeyrefs[i]) == 1) {
 710                         u64 pkey;
 711
 712                         /* for ipathstats, etc. */
 713                         ipath_stats.sps_pkeys[i] = lkey;
 714                         pd->port_pkeys[pidx] = dd->ipath_pkeys[i] = key;
 715                         pkey =
 716                                 (u64) dd->ipath_pkeys[0] |
 717                                 ((u64) dd->ipath_pkeys[1] << 16) |
 718                                 ((u64) dd->ipath_pkeys[2] << 32) |
 719                                 ((u64) dd->ipath_pkeys[3] << 48);
 720                         ipath_cdbg(PROC, "p%u set key %x in #%d, "
 721                                    "portidx %d, new pkey reg %llx\n",
 722                                    pd->port_port, key, i, pidx,
 723                                    (unsigned long long) pkey);
 724                         ipath_write_kreg(
 725                                 dd, dd->ipath_kregs->kr_partitionkey, pkey);
 726
 727                         ret = 0;
 728                         goto bail;
 729                 }
 730         }
 731         ipath_dbg("port %u, all pkeys already in use 2nd pass, "
 732                   "can't set %x\n", pd->port_port, key);
 733         ret = -EBUSY;
 734
 735 bail:
 736         return ret;
 737 }
 738
 739 /**
 740  * ipath_manage_rcvq - manage a port's receive queue
 741  * @pd: the port
 742  * @subport: the subport
 743  * @start_stop: action to carry out
 744  *
 745  * start_stop == 0 disables receive on the port, for use in queue
 746  * overflow conditions.  start_stop==1 re-enables, to be used to
 747  * re-init the software copy of the head register
 748  */
 749 static int ipath_manage_rcvq(struct ipath_portdata *pd, unsigned subport,
 750                              int start_stop)
 751 {
 752         struct ipath_devdata *dd = pd->port_dd;
 753
 754         ipath_cdbg(PROC, "%sabling rcv for unit %u port %u:%u\n",
 755                    start_stop ? "en" : "dis", dd->ipath_unit,
 756                    pd->port_port, subport);
 757         if (subport)
 758                 goto bail;
 759         /* atomically clear receive enable port. */
 760         if (start_stop) {
 761                 /*
 762                  * On enable, force in-memory copy of the tail register to
 763                  * 0, so that protocol code doesn't have to worry about
 764                  * whether or not the chip has yet updated the in-memory
 765                  * copy or not on return from the system call. The chip
 766                  * always resets it's tail register back to 0 on a
 767                  * transition from disabled to enabled.  This could cause a
 768                  * problem if software was broken, and did the enable w/o
 769                  * the disable, but eventually the in-memory copy will be
 770                  * updated and correct itself, even in the face of software
 771                  * bugs.
 772                  */
 773                 if (pd->port_rcvhdrtail_kvaddr)
 774                         ipath_clear_rcvhdrtail(pd);
 775                 set_bit(dd->ipath_r_portenable_shift + pd->port_port,
 776                         &dd->ipath_rcvctrl);
 777         } else
 778                 clear_bit(dd->ipath_r_portenable_shift + pd->port_port,
 779                           &dd->ipath_rcvctrl);
 780         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
 781                          dd->ipath_rcvctrl);
 782         /* now be sure chip saw it before we return */
 783         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 784         if (start_stop) {
 785                 /*
 786                  * And try to be sure that tail reg update has happened too.
 787                  * This should in theory interlock with the RXE changes to
 788                  * the tail register.  Don't assign it to the tail register
 789                  * in memory copy, since we could overwrite an update by the
 790                  * chip if we did.
 791                  */
 792                 ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
 793         }
 794         /* always; new head should be equal to new tail; see above */
 795 bail:
 796         return 0;
 797 }
 798
 799 static void ipath_clean_part_key(struct ipath_portdata *pd,
 800                                  struct ipath_devdata *dd)
 801 {
 802         int i, j, pchanged = 0;
 803         u64 oldpkey;
 804
 805         /* for debugging only */
 806         oldpkey = (u64) dd->ipath_pkeys[0] |
 807                 ((u64) dd->ipath_pkeys[1] << 16) |
 808                 ((u64) dd->ipath_pkeys[2] << 32) |
 809                 ((u64) dd->ipath_pkeys[3] << 48);
 810
 811         for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
 812                 if (!pd->port_pkeys[i])
 813                         continue;
 814                 ipath_cdbg(VERBOSE, "look for key[%d] %hx in pkeys\n", i,
 815                            pd->port_pkeys[i]);
 816                 for (j = 0; j < ARRAY_SIZE(dd->ipath_pkeys); j++) {
 817                         /* check for match independent of the global bit */
 818                         if ((dd->ipath_pkeys[j] & 0x7fff) !=
 819                             (pd->port_pkeys[i] & 0x7fff))
 820                                 continue;
 821                         if (atomic_dec_and_test(&dd->ipath_pkeyrefs[j])) {
 822                                 ipath_cdbg(VERBOSE, "p%u clear key "
 823                                            "%x matches #%d\n",
 824                                            pd->port_port,
 825                                            pd->port_pkeys[i], j);
 826                                 ipath_stats.sps_pkeys[j] =
 827                                         dd->ipath_pkeys[j] = 0;
 828                                 pchanged++;
 829                         }
 830                         else ipath_cdbg(
 831                                 VERBOSE, "p%u key %x matches #%d, "
 832                                 "but ref still %d\n", pd->port_port,
 833                                 pd->port_pkeys[i], j,
 834                                 atomic_read(&dd->ipath_pkeyrefs[j]));
 835                         break;
 836                 }
 837                 pd->port_pkeys[i] = 0;
 838         }
 839         if (pchanged) {
 840                 u64 pkey = (u64) dd->ipath_pkeys[0] |
 841                         ((u64) dd->ipath_pkeys[1] << 16) |
 842                         ((u64) dd->ipath_pkeys[2] << 32) |
 843                         ((u64) dd->ipath_pkeys[3] << 48);
 844                 ipath_cdbg(VERBOSE, "p%u old pkey reg %llx, "
 845                            "new pkey reg %llx\n", pd->port_port,
 846                            (unsigned long long) oldpkey,
 847                            (unsigned long long) pkey);
 848                 ipath_write_kreg(dd, dd->ipath_kregs->kr_partitionkey,
 849                                  pkey);
 850         }
 851 }
 852
 853 /*
 854  * Initialize the port data with the receive buffer sizes
 855  * so this can be done while the master port is locked.
 856  * Otherwise, there is a race with a slave opening the port
 857  * and seeing these fields uninitialized.
 858  */
 859 static void init_user_egr_sizes(struct ipath_portdata *pd)
 860 {
 861         struct ipath_devdata *dd = pd->port_dd;
 862         unsigned egrperchunk, egrcnt, size;
 863
 864         /*
 865          * to avoid wasting a lot of memory, we allocate 32KB chunks of
 866          * physically contiguous memory, advance through it until used up
 867          * and then allocate more.  Of course, we need memory to store those
 868          * extra pointers, now.  Started out with 256KB, but under heavy
 869          * memory pressure (creating large files and then copying them over
 870          * NFS while doing lots of MPI jobs), we hit some allocation
 871          * failures, even though we can sleep...  (2.6.10) Still get
 872          * failures at 64K.  32K is the lowest we can go without wasting
 873          * additional memory.
 874          */
 875         size = 0x8000;
 876         egrperchunk = size / dd->ipath_rcvegrbufsize;
 877         egrcnt = dd->ipath_rcvegrcnt;
 878         pd->port_rcvegrbuf_chunks = (egrcnt + egrperchunk - 1) / egrperchunk;
 879         pd->port_rcvegrbufs_perchunk = egrperchunk;
 880         pd->port_rcvegrbuf_size = size;
 881 }
 882
 883 /**
 884  * ipath_create_user_egr - allocate eager TID buffers
 885  * @pd: the port to allocate TID buffers for
 886  *
 887  * This routine is now quite different for user and kernel, because
 888  * the kernel uses skb's, for the accelerated network performance
 889  * This is the user port version
 890  *
 891  * Allocate the eager TID buffers and program them into infinipath
 892  * They are no longer completely contiguous, we do multiple allocation
 893  * calls.
 894  */
 895 static int ipath_create_user_egr(struct ipath_portdata *pd)
 896 {
 897         struct ipath_devdata *dd = pd->port_dd;
 898         unsigned e, egrcnt, egrperchunk, chunk, egrsize, egroff;
 899         size_t size;
 900         int ret;
 901         gfp_t gfp_flags;
 902
 903         /*
 904          * GFP_USER, but without GFP_FS, so buffer cache can be
 905          * coalesced (we hope); otherwise, even at order 4,
 906          * heavy filesystem activity makes these fail, and we can
 907          * use compound pages.
 908          */
 909         gfp_flags = __GFP_WAIT | __GFP_IO | __GFP_COMP;
 910
 911         egrcnt = dd->ipath_rcvegrcnt;
 912         /* TID number offset for this port */
 913         egroff = (pd->port_port - 1) * egrcnt + dd->ipath_p0_rcvegrcnt;
 914         egrsize = dd->ipath_rcvegrbufsize;
 915         ipath_cdbg(VERBOSE, "Allocating %d egr buffers, at egrtid "
 916                    "offset %x, egrsize %u\n", egrcnt, egroff, egrsize);
 917
 918         chunk = pd->port_rcvegrbuf_chunks;
 919         egrperchunk = pd->port_rcvegrbufs_perchunk;
 920         size = pd->port_rcvegrbuf_size;
 921         pd->port_rcvegrbuf = kmalloc(chunk * sizeof(pd->port_rcvegrbuf[0]),
 922                                      GFP_KERNEL);
 923         if (!pd->port_rcvegrbuf) {
 924                 ret = -ENOMEM;
 925                 goto bail;
 926         }
 927         pd->port_rcvegrbuf_phys =
 928                 kmalloc(chunk * sizeof(pd->port_rcvegrbuf_phys[0]),
 929                         GFP_KERNEL);
 930         if (!pd->port_rcvegrbuf_phys) {
 931                 ret = -ENOMEM;
 932                 goto bail_rcvegrbuf;
 933         }
 934         for (e = 0; e < pd->port_rcvegrbuf_chunks; e++) {
 935
 936                 pd->port_rcvegrbuf[e] = dma_alloc_coherent(
 937                         &dd->pcidev->dev, size, &pd->port_rcvegrbuf_phys[e],
 938                         gfp_flags);
 939
 940                 if (!pd->port_rcvegrbuf[e]) {
 941                         ret = -ENOMEM;
 942                         goto bail_rcvegrbuf_phys;
 943                 }
 944         }
 945
 946         pd->port_rcvegr_phys = pd->port_rcvegrbuf_phys[0];
 947
 948         for (e = chunk = 0; chunk < pd->port_rcvegrbuf_chunks; chunk++) {
 949                 dma_addr_t pa = pd->port_rcvegrbuf_phys[chunk];
 950                 unsigned i;
 951
 952                 for (i = 0; e < egrcnt && i < egrperchunk; e++, i++) {
 953                         dd->ipath_f_put_tid(dd, e + egroff +
 954                                             (u64 __iomem *)
 955                                             ((char __iomem *)
 956                                              dd->ipath_kregbase +
 957                                              dd->ipath_rcvegrbase),
 958                                             RCVHQ_RCV_TYPE_EAGER, pa);
 959                         pa += egrsize;
 960                 }
 961                 cond_resched(); /* don't hog the cpu */
 962         }
 963
 964         ret = 0;
 965         goto bail;
 966
 967 bail_rcvegrbuf_phys:
 968         for (e = 0; e < pd->port_rcvegrbuf_chunks &&
 969                 pd->port_rcvegrbuf[e]; e++) {
 970                 dma_free_coherent(&dd->pcidev->dev, size,
 971                                   pd->port_rcvegrbuf[e],
 972                                   pd->port_rcvegrbuf_phys[e]);
 973
 974         }
 975         kfree(pd->port_rcvegrbuf_phys);
 976         pd->port_rcvegrbuf_phys = NULL;
 977 bail_rcvegrbuf:
 978         kfree(pd->port_rcvegrbuf);
 979         pd->port_rcvegrbuf = NULL;
 980 bail:
 981         return ret;
 982 }
 983
 984
 985 /* common code for the mappings on dma_alloc_coherent mem */
 986 static int ipath_mmap_mem(struct vm_area_struct *vma,
 987         struct ipath_portdata *pd, unsigned len, int write_ok,
 988         void *kvaddr, char *what)
 989 {
 990         struct ipath_devdata *dd = pd->port_dd;
 991         unsigned long pfn;
 992         int ret;
 993
 994         if ((vma->vm_end - vma->vm_start) > len) {
 995                 dev_info(&dd->pcidev->dev,
 996                          "FAIL on %s: len %lx > %x\n", what,
 997                          vma->vm_end - vma->vm_start, len);
 998                 ret = -EFAULT;
 999                 goto bail;
1000         }
1001
1002         if (!write_ok) {
1003                 if (vma->vm_flags & VM_WRITE) {
1004                         dev_info(&dd->pcidev->dev,
1005                                  "%s must be mapped readonly\n", what);
1006                         ret = -EPERM;
1007                         goto bail;
1008                 }
1009
1010                 /* don't allow them to later change with mprotect */
1011                 vma->vm_flags &= ~VM_MAYWRITE;
1012         }
1013
1014         pfn = virt_to_phys(kvaddr) >> PAGE_SHIFT;
1015         ret = remap_pfn_range(vma, vma->vm_start, pfn,
1016                               len, vma->vm_page_prot);
1017         if (ret)
1018                 dev_info(&dd->pcidev->dev, "%s port%u mmap of %lx, %x "
1019                          "bytes r%c failed: %d\n", what, pd->port_port,
1020                          pfn, len, write_ok?'w':'o', ret);
1021         else
1022                 ipath_cdbg(VERBOSE, "%s port%u mmaped %lx, %x bytes "
1023                            "r%c\n", what, pd->port_port, pfn, len,
1024                            write_ok?'w':'o');
1025 bail:
1026         return ret;
1027 }
1028
1029 static int mmap_ureg(struct vm_area_struct *vma, struct ipath_devdata *dd,
1030                      u64 ureg)
1031 {
1032         unsigned long phys;
1033         int ret;
1034
1035         /*
1036          * This is real hardware, so use io_remap.  This is the mechanism
1037          * for the user process to update the head registers for their port
1038          * in the chip.
1039          */
1040         if ((vma->vm_end - vma->vm_start) > PAGE_SIZE) {
1041                 dev_info(&dd->pcidev->dev, "FAIL mmap userreg: reqlen "
1042                          "%lx > PAGE\n", vma->vm_end - vma->vm_start);
1043                 ret = -EFAULT;
1044         } else {
1045                 phys = dd->ipath_physaddr + ureg;
1046                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1047
1048                 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1049                 ret = io_remap_pfn_range(vma, vma->vm_start,
1050                                          phys >> PAGE_SHIFT,
1051                                          vma->vm_end - vma->vm_start,
1052                                          vma->vm_page_prot);
1053         }
1054         return ret;
1055 }
1056
1057 static int mmap_piobufs(struct vm_area_struct *vma,
1058                         struct ipath_devdata *dd,
1059                         struct ipath_portdata *pd,
1060                         unsigned piobufs, unsigned piocnt)
1061 {
1062         unsigned long phys;
1063         int ret;
1064
1065         /*
1066          * When we map the PIO buffers in the chip, we want to map them as
1067          * writeonly, no read possible.   This prevents access to previous
1068          * process data, and catches users who might try to read the i/o
1069          * space due to a bug.
1070          */
1071         if ((vma->vm_end - vma->vm_start) > (piocnt * dd->ipath_palign)) {
1072                 dev_info(&dd->pcidev->dev, "FAIL mmap piobufs: "
1073                          "reqlen %lx > PAGE\n",
1074                          vma->vm_end - vma->vm_start);
1075                 ret = -EINVAL;
1076                 goto bail;
1077         }
1078
1079         phys = dd->ipath_physaddr + piobufs;
1080
1081 #if defined(__powerpc__)
1082         /* There isn't a generic way to specify writethrough mappings */
1083         pgprot_val(vma->vm_page_prot) |= _PAGE_NO_CACHE;
1084         pgprot_val(vma->vm_page_prot) |= _PAGE_WRITETHRU;
1085         pgprot_val(vma->vm_page_prot) &= ~_PAGE_GUARDED;
1086 #endif
1087
1088         /*
1089          * don't allow them to later change to readable with mprotect (for when
1090          * not initially mapped readable, as is normally the case)
1091          */
1092         vma->vm_flags &= ~VM_MAYREAD;
1093         vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1094
1095         ret = io_remap_pfn_range(vma, vma->vm_start, phys >> PAGE_SHIFT,
1096                                  vma->vm_end - vma->vm_start,
1097                                  vma->vm_page_prot);
1098 bail:
1099         return ret;
1100 }
1101
1102 static int mmap_rcvegrbufs(struct vm_area_struct *vma,
1103                            struct ipath_portdata *pd)
1104 {
1105         struct ipath_devdata *dd = pd->port_dd;
1106         unsigned long start, size;
1107         size_t total_size, i;
1108         unsigned long pfn;
1109         int ret;
1110
1111         size = pd->port_rcvegrbuf_size;
1112         total_size = pd->port_rcvegrbuf_chunks * size;
1113         if ((vma->vm_end - vma->vm_start) > total_size) {
1114                 dev_info(&dd->pcidev->dev, "FAIL on egr bufs: "
1115                          "reqlen %lx > actual %lx\n",
1116                          vma->vm_end - vma->vm_start,
1117                          (unsigned long) total_size);
1118                 ret = -EINVAL;
1119                 goto bail;
1120         }
1121
1122         if (vma->vm_flags & VM_WRITE) {
1123                 dev_info(&dd->pcidev->dev, "Can't map eager buffers as "
1124                          "writable (flags=%lx)\n", vma->vm_flags);
1125                 ret = -EPERM;
1126                 goto bail;
1127         }
1128         /* don't allow them to later change to writeable with mprotect */
1129         vma->vm_flags &= ~VM_MAYWRITE;
1130
1131         start = vma->vm_start;
1132
1133         for (i = 0; i < pd->port_rcvegrbuf_chunks; i++, start += size) {
1134                 pfn = virt_to_phys(pd->port_rcvegrbuf[i]) >> PAGE_SHIFT;
1135                 ret = remap_pfn_range(vma, start, pfn, size,
1136                                       vma->vm_page_prot);
1137                 if (ret < 0)
1138                         goto bail;
1139         }
1140         ret = 0;
1141
1142 bail:
1143         return ret;
1144 }
1145
1146 /*
1147  * ipath_file_vma_fault - handle a VMA page fault.
1148  */
1149 static int ipath_file_vma_fault(struct vm_area_struct *vma,
1150                                         struct vm_fault *vmf)
1151 {
1152         struct page *page;
1153
1154         page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
1155         if (!page)
1156                 return VM_FAULT_SIGBUS;
1157         get_page(page);
1158         vmf->page = page;
1159
1160         return 0;
1161 }
1162
1163 static struct vm_operations_struct ipath_file_vm_ops = {
1164         .fault = ipath_file_vma_fault,
1165 };
1166
1167 static int mmap_kvaddr(struct vm_area_struct *vma, u64 pgaddr,
1168                        struct ipath_portdata *pd, unsigned subport)
1169 {
1170         unsigned long len;
1171         struct ipath_devdata *dd;
1172         void *addr;
1173         size_t size;
1174         int ret = 0;
1175
1176         /* If the port is not shared, all addresses should be physical */
1177         if (!pd->port_subport_cnt)
1178                 goto bail;
1179
1180         dd = pd->port_dd;
1181         size = pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
1182
1183         /*
1184          * Each process has all the subport uregbase, rcvhdrq, and
1185          * rcvegrbufs mmapped - as an array for all the processes,
1186          * and also separately for this process.
1187          */
1188         if (pgaddr == cvt_kvaddr(pd->subport_uregbase)) {
1189                 addr = pd->subport_uregbase;
1190                 size = PAGE_SIZE * pd->port_subport_cnt;
1191         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvhdr_base)) {
1192                 addr = pd->subport_rcvhdr_base;
1193                 size = pd->port_rcvhdrq_size * pd->port_subport_cnt;
1194         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvegrbuf)) {
1195                 addr = pd->subport_rcvegrbuf;
1196                 size *= pd->port_subport_cnt;
1197         } else if (pgaddr == cvt_kvaddr(pd->subport_uregbase +
1198                                         PAGE_SIZE * subport)) {
1199                 addr = pd->subport_uregbase + PAGE_SIZE * subport;
1200                 size = PAGE_SIZE;
1201         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvhdr_base +
1202                                 pd->port_rcvhdrq_size * subport)) {
1203                 addr = pd->subport_rcvhdr_base +
1204                         pd->port_rcvhdrq_size * subport;
1205                 size = pd->port_rcvhdrq_size;
1206         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvegrbuf +
1207                                size * subport)) {
1208                 addr = pd->subport_rcvegrbuf + size * subport;
1209                 /* rcvegrbufs are read-only on the slave */
1210                 if (vma->vm_flags & VM_WRITE) {
1211                         dev_info(&dd->pcidev->dev,
1212                                  "Can't map eager buffers as "
1213                                  "writable (flags=%lx)\n", vma->vm_flags);
1214                         ret = -EPERM;
1215                         goto bail;
1216                 }
1217                 /*
1218                  * Don't allow permission to later change to writeable
1219                  * with mprotect.
1220                  */
1221                 vma->vm_flags &= ~VM_MAYWRITE;
1222         } else {
1223                 goto bail;
1224         }
1225         len = vma->vm_end - vma->vm_start;
1226         if (len > size) {
1227                 ipath_cdbg(MM, "FAIL: reqlen %lx > %zx\n", len, size);
1228                 ret = -EINVAL;
1229                 goto bail;
1230         }
1231
1232         vma->vm_pgoff = (unsigned long) addr >> PAGE_SHIFT;
1233         vma->vm_ops = &ipath_file_vm_ops;
1234         vma->vm_flags |= VM_RESERVED | VM_DONTEXPAND;
1235         ret = 1;
1236
1237 bail:
1238         return ret;
1239 }
1240
1241 /**
1242  * ipath_mmap - mmap various structures into user space
1243  * @fp: the file pointer
1244  * @vma: the VM area
1245  *
1246  * We use this to have a shared buffer between the kernel and the user code
1247  * for the rcvhdr queue, egr buffers, and the per-port user regs and pio
1248  * buffers in the chip.  We have the open and close entries so we can bump
1249  * the ref count and keep the driver from being unloaded while still mapped.
1250  */
1251 static int ipath_mmap(struct file *fp, struct vm_area_struct *vma)
1252 {
1253         struct ipath_portdata *pd;
1254         struct ipath_devdata *dd;
1255         u64 pgaddr, ureg;
1256         unsigned piobufs, piocnt;
1257         int ret;
1258
1259         pd = port_fp(fp);
1260         if (!pd) {
1261                 ret = -EINVAL;
1262                 goto bail;
1263         }
1264         dd = pd->port_dd;
1265
1266         /*
1267          * This is the ipath_do_user_init() code, mapping the shared buffers
1268          * into the user process. The address referred to by vm_pgoff is the
1269          * file offset passed via mmap().  For shared ports, this is the
1270          * kernel vmalloc() address of the pages to share with the master.
1271          * For non-shared or master ports, this is a physical address.
1272          * We only do one mmap for each space mapped.
1273          */
1274         pgaddr = vma->vm_pgoff << PAGE_SHIFT;
1275
1276         /*
1277          * Check for 0 in case one of the allocations failed, but user
1278          * called mmap anyway.
1279          */
1280         if (!pgaddr)  {
1281                 ret = -EINVAL;
1282                 goto bail;
1283         }
1284
1285         ipath_cdbg(MM, "pgaddr %llx vm_start=%lx len %lx port %u:%u:%u\n",
1286                    (unsigned long long) pgaddr, vma->vm_start,
1287                    vma->vm_end - vma->vm_start, dd->ipath_unit,
1288                    pd->port_port, subport_fp(fp));
1289
1290         /*
1291          * Physical addresses must fit in 40 bits for our hardware.
1292          * Check for kernel virtual addresses first, anything else must
1293          * match a HW or memory address.
1294          */
1295         ret = mmap_kvaddr(vma, pgaddr, pd, subport_fp(fp));
1296         if (ret) {
1297                 if (ret > 0)
1298                         ret = 0;
1299                 goto bail;
1300         }
1301
1302         ureg = dd->ipath_uregbase + dd->ipath_ureg_align * pd->port_port;
1303         if (!pd->port_subport_cnt) {
1304                 /* port is not shared */
1305                 piocnt = dd->ipath_pbufsport;
1306                 piobufs = pd->port_piobufs;
1307         } else if (!subport_fp(fp)) {
1308                 /* caller is the master */
1309                 piocnt = (dd->ipath_pbufsport / pd->port_subport_cnt) +
1310                          (dd->ipath_pbufsport % pd->port_subport_cnt);
1311                 piobufs = pd->port_piobufs +
1312                         dd->ipath_palign * (dd->ipath_pbufsport - piocnt);
1313         } else {
1314                 unsigned slave = subport_fp(fp) - 1;
1315
1316                 /* caller is a slave */
1317                 piocnt = dd->ipath_pbufsport / pd->port_subport_cnt;
1318                 piobufs = pd->port_piobufs + dd->ipath_palign * piocnt * slave;
1319         }
1320
1321         if (pgaddr == ureg)
1322                 ret = mmap_ureg(vma, dd, ureg);
1323         else if (pgaddr == piobufs)
1324                 ret = mmap_piobufs(vma, dd, pd, piobufs, piocnt);
1325         else if (pgaddr == dd->ipath_pioavailregs_phys)
1326                 /* in-memory copy of pioavail registers */
1327                 ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1328                                      (void *) dd->ipath_pioavailregs_dma,
1329                                      "pioavail registers");
1330         else if (pgaddr == pd->port_rcvegr_phys)
1331                 ret = mmap_rcvegrbufs(vma, pd);
1332         else if (pgaddr == (u64) pd->port_rcvhdrq_phys)
1333                 /*
1334                  * The rcvhdrq itself; readonly except on HT (so have
1335                  * to allow writable mapping), multiple pages, contiguous
1336                  * from an i/o perspective.
1337                  */
1338                 ret = ipath_mmap_mem(vma, pd, pd->port_rcvhdrq_size, 1,
1339                                      pd->port_rcvhdrq,
1340                                      "rcvhdrq");
1341         else if (pgaddr == (u64) pd->port_rcvhdrqtailaddr_phys)
1342                 /* in-memory copy of rcvhdrq tail register */
1343                 ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1344                                      pd->port_rcvhdrtail_kvaddr,
1345                                      "rcvhdrq tail");
1346         else
1347                 ret = -EINVAL;
1348
1349         vma->vm_private_data = NULL;
1350
1351         if (ret < 0)
1352                 dev_info(&dd->pcidev->dev,
1353                          "Failure %d on off %llx len %lx\n",
1354                          -ret, (unsigned long long)pgaddr,
1355                          vma->vm_end - vma->vm_start);
1356 bail:
1357         return ret;
1358 }
1359
1360 static unsigned ipath_poll_hdrqfull(struct ipath_portdata *pd)
1361 {
1362         unsigned pollflag = 0;
1363
1364         if ((pd->poll_type & IPATH_POLL_TYPE_OVERFLOW) &&
1365             pd->port_hdrqfull != pd->port_hdrqfull_poll) {
1366                 pollflag |= POLLIN | POLLRDNORM;
1367                 pd->port_hdrqfull_poll = pd->port_hdrqfull;
1368         }
1369
1370         return pollflag;
1371 }
1372
1373 static unsigned int ipath_poll_urgent(struct ipath_portdata *pd,
1374                                       struct file *fp,
1375                                       struct poll_table_struct *pt)
1376 {
1377         unsigned pollflag = 0;
1378         struct ipath_devdata *dd;
1379
1380         dd = pd->port_dd;
1381
1382         /* variable access in ipath_poll_hdrqfull() needs this */
1383         rmb();
1384         pollflag = ipath_poll_hdrqfull(pd);
1385
1386         if (pd->port_urgent != pd->port_urgent_poll) {
1387                 pollflag |= POLLIN | POLLRDNORM;
1388                 pd->port_urgent_poll = pd->port_urgent;
1389         }
1390
1391         if (!pollflag) {
1392                 /* this saves a spin_lock/unlock in interrupt handler... */
1393                 set_bit(IPATH_PORT_WAITING_URG, &pd->port_flag);
1394                 /* flush waiting flag so don't miss an event... */
1395                 wmb();
1396                 poll_wait(fp, &pd->port_wait, pt);
1397         }
1398
1399         return pollflag;
1400 }
1401
1402 static unsigned int ipath_poll_next(struct ipath_portdata *pd,
1403                                     struct file *fp,
1404                                     struct poll_table_struct *pt)
1405 {
1406         u32 head;
1407         u32 tail;
1408         unsigned pollflag = 0;
1409         struct ipath_devdata *dd;
1410
1411         dd = pd->port_dd;
1412
1413         /* variable access in ipath_poll_hdrqfull() needs this */
1414         rmb();
1415         pollflag = ipath_poll_hdrqfull(pd);
1416
1417         head = ipath_read_ureg32(dd, ur_rcvhdrhead, pd->port_port);
1418         if (pd->port_rcvhdrtail_kvaddr)
1419                 tail = ipath_get_rcvhdrtail(pd);
1420         else
1421                 tail = ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
1422
1423         if (head != tail)
1424                 pollflag |= POLLIN | POLLRDNORM;
1425         else {
1426                 /* this saves a spin_lock/unlock in interrupt handler */
1427                 set_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag);
1428                 /* flush waiting flag so we don't miss an event */
1429                 wmb();
1430
1431                 set_bit(pd->port_port + dd->ipath_r_intravail_shift,
1432                         &dd->ipath_rcvctrl);
1433
1434                 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1435                                  dd->ipath_rcvctrl);
1436
1437                 if (dd->ipath_rhdrhead_intr_off) /* arm rcv interrupt */
1438                         ipath_write_ureg(dd, ur_rcvhdrhead,
1439                                          dd->ipath_rhdrhead_intr_off | head,
1440                                          pd->port_port);
1441
1442                 poll_wait(fp, &pd->port_wait, pt);
1443         }
1444
1445         return pollflag;
1446 }
1447
1448 static unsigned int ipath_poll(struct file *fp,
1449                                struct poll_table_struct *pt)
1450 {
1451         struct ipath_portdata *pd;
1452         unsigned pollflag;
1453
1454         pd = port_fp(fp);
1455         if (!pd)
1456                 pollflag = 0;
1457         else if (pd->poll_type & IPATH_POLL_TYPE_URGENT)
1458                 pollflag = ipath_poll_urgent(pd, fp, pt);
1459         else
1460                 pollflag = ipath_poll_next(pd, fp, pt);
1461
1462         return pollflag;
1463 }
1464
1465 static int ipath_supports_subports(int user_swmajor, int user_swminor)
1466 {
1467         /* no subport implementation prior to software version 1.3 */
1468         return (user_swmajor > 1) || (user_swminor >= 3);
1469 }
1470
1471 static int ipath_compatible_subports(int user_swmajor, int user_swminor)
1472 {
1473         /* this code is written long-hand for clarity */
1474         if (IPATH_USER_SWMAJOR != user_swmajor) {
1475                 /* no promise of compatibility if major mismatch */
1476                 return 0;
1477         }
1478         if (IPATH_USER_SWMAJOR == 1) {
1479                 switch (IPATH_USER_SWMINOR) {
1480                 case 0:
1481                 case 1:
1482                 case 2:
1483                         /* no subport implementation so cannot be compatible */
1484                         return 0;
1485                 case 3:
1486                         /* 3 is only compatible with itself */
1487                         return user_swminor == 3;
1488                 default:
1489                         /* >= 4 are compatible (or are expected to be) */
1490                         return user_swminor >= 4;
1491                 }
1492         }
1493         /* make no promises yet for future major versions */
1494         return 0;
1495 }
1496
1497 static int init_subports(struct ipath_devdata *dd,
1498                          struct ipath_portdata *pd,
1499                          const struct ipath_user_info *uinfo)
1500 {
1501         int ret = 0;
1502         unsigned num_subports;
1503         size_t size;
1504
1505         /*
1506          * If the user is requesting zero subports,
1507          * skip the subport allocation.
1508          */
1509         if (uinfo->spu_subport_cnt <= 0)
1510                 goto bail;
1511
1512         /* Self-consistency check for ipath_compatible_subports() */
1513         if (ipath_supports_subports(IPATH_USER_SWMAJOR, IPATH_USER_SWMINOR) &&
1514             !ipath_compatible_subports(IPATH_USER_SWMAJOR,
1515                                        IPATH_USER_SWMINOR)) {
1516                 dev_info(&dd->pcidev->dev,
1517                          "Inconsistent ipath_compatible_subports()\n");
1518                 goto bail;
1519         }
1520
1521         /* Check for subport compatibility */
1522         if (!ipath_compatible_subports(uinfo->spu_userversion >> 16,
1523                                        uinfo->spu_userversion & 0xffff)) {
1524                 dev_info(&dd->pcidev->dev,
1525                          "Mismatched user version (%d.%d) and driver "
1526                          "version (%d.%d) while port sharing. Ensure "
1527                          "that driver and library are from the same "
1528                          "release.\n",
1529                          (int) (uinfo->spu_userversion >> 16),
1530                          (int) (uinfo->spu_userversion & 0xffff),
1531                          IPATH_USER_SWMAJOR,
1532                          IPATH_USER_SWMINOR);
1533                 goto bail;
1534         }
1535         if (uinfo->spu_subport_cnt > INFINIPATH_MAX_SUBPORT) {
1536                 ret = -EINVAL;
1537                 goto bail;
1538         }
1539
1540         num_subports = uinfo->spu_subport_cnt;
1541         pd->subport_uregbase = vmalloc(PAGE_SIZE * num_subports);
1542         if (!pd->subport_uregbase) {
1543                 ret = -ENOMEM;
1544                 goto bail;
1545         }
1546         /* Note: pd->port_rcvhdrq_size isn't initialized yet. */
1547         size = ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize *
1548                      sizeof(u32), PAGE_SIZE) * num_subports;
1549         pd->subport_rcvhdr_base = vmalloc(size);
1550         if (!pd->subport_rcvhdr_base) {
1551                 ret = -ENOMEM;
1552                 goto bail_ureg;
1553         }
1554
1555         pd->subport_rcvegrbuf = vmalloc(pd->port_rcvegrbuf_chunks *
1556                                         pd->port_rcvegrbuf_size *
1557                                         num_subports);
1558         if (!pd->subport_rcvegrbuf) {
1559                 ret = -ENOMEM;
1560                 goto bail_rhdr;
1561         }
1562
1563         pd->port_subport_cnt = uinfo->spu_subport_cnt;
1564         pd->port_subport_id = uinfo->spu_subport_id;
1565         pd->active_slaves = 1;
1566         set_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag);
1567         memset(pd->subport_uregbase, 0, PAGE_SIZE * num_subports);
1568         memset(pd->subport_rcvhdr_base, 0, size);
1569         memset(pd->subport_rcvegrbuf, 0, pd->port_rcvegrbuf_chunks *
1570                                          pd->port_rcvegrbuf_size *
1571                                          num_subports);
1572         goto bail;
1573
1574 bail_rhdr:
1575         vfree(pd->subport_rcvhdr_base);
1576 bail_ureg:
1577         vfree(pd->subport_uregbase);
1578         pd->subport_uregbase = NULL;
1579 bail:
1580         return ret;
1581 }
1582
1583 static int try_alloc_port(struct ipath_devdata *dd, int port,
1584                           struct file *fp,
1585                           const struct ipath_user_info *uinfo)
1586 {
1587         struct ipath_portdata *pd;
1588         int ret;
1589
1590         if (!(pd = dd->ipath_pd[port])) {
1591                 void *ptmp;
1592
1593                 pd = kzalloc(sizeof(struct ipath_portdata), GFP_KERNEL);
1594
1595                 /*
1596                  * Allocate memory for use in ipath_tid_update() just once
1597                  * at open, not per call.  Reduces cost of expected send
1598                  * setup.
1599                  */
1600                 ptmp = kmalloc(dd->ipath_rcvtidcnt * sizeof(u16) +
1601                                dd->ipath_rcvtidcnt * sizeof(struct page **),
1602                                GFP_KERNEL);
1603                 if (!pd || !ptmp) {
1604                         ipath_dev_err(dd, "Unable to allocate portdata "
1605                                       "memory, failing open\n");
1606                         ret = -ENOMEM;
1607                         kfree(pd);
1608                         kfree(ptmp);
1609                         goto bail;
1610                 }
1611                 dd->ipath_pd[port] = pd;
1612                 dd->ipath_pd[port]->port_port = port;
1613                 dd->ipath_pd[port]->port_dd = dd;
1614                 dd->ipath_pd[port]->port_tid_pg_list = ptmp;
1615                 init_waitqueue_head(&dd->ipath_pd[port]->port_wait);
1616         }
1617         if (!pd->port_cnt) {
1618                 pd->userversion = uinfo->spu_userversion;
1619                 init_user_egr_sizes(pd);
1620                 if ((ret = init_subports(dd, pd, uinfo)) != 0)
1621                         goto bail;
1622                 ipath_cdbg(PROC, "%s[%u] opened unit:port %u:%u\n",
1623                            current->comm, current->pid, dd->ipath_unit,
1624                            port);
1625                 pd->port_cnt = 1;
1626                 port_fp(fp) = pd;
1627                 pd->port_pid = current->pid;
1628                 strncpy(pd->port_comm, current->comm, sizeof(pd->port_comm));
1629                 ipath_chg_pioavailkernel(dd,
1630                         dd->ipath_pbufsport * (pd->port_port - 1),
1631                         dd->ipath_pbufsport, 0);
1632                 ipath_stats.sps_ports++;
1633                 ret = 0;
1634         } else
1635                 ret = -EBUSY;
1636
1637 bail:
1638         return ret;
1639 }
1640
1641 static inline int usable(struct ipath_devdata *dd)
1642 {
1643         return dd &&
1644                 (dd->ipath_flags & IPATH_PRESENT) &&
1645                 dd->ipath_kregbase &&
1646                 dd->ipath_lid &&
1647                 !(dd->ipath_flags & (IPATH_LINKDOWN | IPATH_DISABLED
1648                                      | IPATH_LINKUNK));
1649 }
1650
1651 static int find_free_port(int unit, struct file *fp,
1652                           const struct ipath_user_info *uinfo)
1653 {
1654         struct ipath_devdata *dd = ipath_lookup(unit);
1655         int ret, i;
1656
1657         if (!dd) {
1658                 ret = -ENODEV;
1659                 goto bail;
1660         }
1661
1662         if (!usable(dd)) {
1663                 ret = -ENETDOWN;
1664                 goto bail;
1665         }
1666
1667         for (i = 1; i < dd->ipath_cfgports; i++) {
1668                 ret = try_alloc_port(dd, i, fp, uinfo);
1669                 if (ret != -EBUSY)
1670                         goto bail;
1671         }
1672         ret = -EBUSY;
1673
1674 bail:
1675         return ret;
1676 }
1677
1678 static int find_best_unit(struct file *fp,
1679                           const struct ipath_user_info *uinfo)
1680 {
1681         int ret = 0, i, prefunit = -1, devmax;
1682         int maxofallports, npresent, nup;
1683         int ndev;
1684
1685         devmax = ipath_count_units(&npresent, &nup, &maxofallports);
1686
1687         /*
1688          * This code is present to allow a knowledgeable person to
1689          * specify the layout of processes to processors before opening
1690          * this driver, and then we'll assign the process to the "closest"
1691          * InfiniPath chip to that processor (we assume reasonable connectivity,
1692          * for now).  This code assumes that if affinity has been set
1693          * before this point, that at most one cpu is set; for now this
1694          * is reasonable.  I check for both cpus_empty() and cpus_full(),
1695          * in case some kernel variant sets none of the bits when no
1696          * affinity is set.  2.6.11 and 12 kernels have all present
1697          * cpus set.  Some day we'll have to fix it up further to handle
1698          * a cpu subset.  This algorithm fails for two HT chips connected
1699          * in tunnel fashion.  Eventually this needs real topology
1700          * information.  There may be some issues with dual core numbering
1701          * as well.  This needs more work prior to release.
1702          */
1703         if (!cpus_empty(current->cpus_allowed) &&
1704             !cpus_full(current->cpus_allowed)) {
1705                 int ncpus = num_online_cpus(), curcpu = -1, nset = 0;
1706                 for (i = 0; i < ncpus; i++)
1707                         if (cpu_isset(i, current->cpus_allowed)) {
1708                                 ipath_cdbg(PROC, "%s[%u] affinity set for "
1709                                            "cpu %d/%d\n", current->comm,
1710                                            current->pid, i, ncpus);
1711                                 curcpu = i;
1712                                 nset++;
1713                         }
1714                 if (curcpu != -1 && nset != ncpus) {
1715                         if (npresent) {
1716                                 prefunit = curcpu / (ncpus / npresent);
1717                                 ipath_cdbg(PROC,"%s[%u] %d chips, %d cpus, "
1718                                           "%d cpus/chip, select unit %d\n",
1719                                           current->comm, current->pid,
1720                                           npresent, ncpus, ncpus / npresent,
1721                                           prefunit);
1722                         }
1723                 }
1724         }
1725
1726         /*
1727          * user ports start at 1, kernel port is 0
1728          * For now, we do round-robin access across all chips
1729          */
1730
1731         if (prefunit != -1)
1732                 devmax = prefunit + 1;
1733 recheck:
1734         for (i = 1; i < maxofallports; i++) {
1735                 for (ndev = prefunit != -1 ? prefunit : 0; ndev < devmax;
1736                      ndev++) {
1737                         struct ipath_devdata *dd = ipath_lookup(ndev);
1738
1739                         if (!usable(dd))
1740                                 continue; /* can't use this unit */
1741                         if (i >= dd->ipath_cfgports)
1742                                 /*
1743                                  * Maxed out on users of this unit. Try
1744                                  * next.
1745                                  */
1746                                 continue;
1747                         ret = try_alloc_port(dd, i, fp, uinfo);
1748                         if (!ret)
1749                                 goto done;
1750                 }
1751         }
1752
1753         if (npresent) {
1754                 if (nup == 0) {
1755                         ret = -ENETDOWN;
1756                         ipath_dbg("No ports available (none initialized "
1757                                   "and ready)\n");
1758                 } else {
1759                         if (prefunit > 0) {
1760                                 /* if started above 0, retry from 0 */
1761                                 ipath_cdbg(PROC,
1762                                            "%s[%u] no ports on prefunit "
1763                                            "%d, clear and re-check\n",
1764                                            current->comm, current->pid,
1765                                            prefunit);
1766                                 devmax = ipath_count_units(NULL, NULL,
1767                                                            NULL);
1768                                 prefunit = -1;
1769                                 goto recheck;
1770                         }
1771                         ret = -EBUSY;
1772                         ipath_dbg("No ports available\n");
1773                 }
1774         } else {
1775                 ret = -ENXIO;
1776                 ipath_dbg("No boards found\n");
1777         }
1778
1779 done:
1780         return ret;
1781 }
1782
1783 static int find_shared_port(struct file *fp,
1784                             const struct ipath_user_info *uinfo)
1785 {
1786         int devmax, ndev, i;
1787         int ret = 0;
1788
1789         devmax = ipath_count_units(NULL, NULL, NULL);
1790
1791         for (ndev = 0; ndev < devmax; ndev++) {
1792                 struct ipath_devdata *dd = ipath_lookup(ndev);
1793
1794                 if (!usable(dd))
1795                         continue;
1796                 for (i = 1; i < dd->ipath_cfgports; i++) {
1797                         struct ipath_portdata *pd = dd->ipath_pd[i];
1798
1799                         /* Skip ports which are not yet open */
1800                         if (!pd || !pd->port_cnt)
1801                                 continue;
1802                         /* Skip port if it doesn't match the requested one */
1803                         if (pd->port_subport_id != uinfo->spu_subport_id)
1804                                 continue;
1805                         /* Verify the sharing process matches the master */
1806                         if (pd->port_subport_cnt != uinfo->spu_subport_cnt ||
1807                             pd->userversion != uinfo->spu_userversion ||
1808                             pd->port_cnt >= pd->port_subport_cnt) {
1809                                 ret = -EINVAL;
1810                                 goto done;
1811                         }
1812                         port_fp(fp) = pd;
1813                         subport_fp(fp) = pd->port_cnt++;
1814                         pd->port_subpid[subport_fp(fp)] = current->pid;
1815                         tidcursor_fp(fp) = 0;
1816                         pd->active_slaves |= 1 << subport_fp(fp);
1817                         ipath_cdbg(PROC,
1818                                    "%s[%u] %u sharing %s[%u] unit:port %u:%u\n",
1819                                    current->comm, current->pid,
1820                                    subport_fp(fp),
1821                                    pd->port_comm, pd->port_pid,
1822                                    dd->ipath_unit, pd->port_port);
1823                         ret = 1;
1824                         goto done;
1825                 }
1826         }
1827
1828 done:
1829         return ret;
1830 }
1831
1832 static int ipath_open(struct inode *in, struct file *fp)
1833 {
1834         /* The real work is performed later in ipath_assign_port() */
1835         fp->private_data = kzalloc(sizeof(struct ipath_filedata), GFP_KERNEL);
1836         return fp->private_data ? 0 : -ENOMEM;
1837 }
1838
1839 /* Get port early, so can set affinity prior to memory allocation */
1840 static int ipath_assign_port(struct file *fp,
1841                               const struct ipath_user_info *uinfo)
1842 {
1843         int ret;
1844         int i_minor;
1845         unsigned swmajor, swminor;
1846
1847         /* Check to be sure we haven't already initialized this file */
1848         if (port_fp(fp)) {
1849                 ret = -EINVAL;
1850                 goto done;
1851         }
1852
1853         /* for now, if major version is different, bail */
1854         swmajor = uinfo->spu_userversion >> 16;
1855         if (swmajor != IPATH_USER_SWMAJOR) {
1856                 ipath_dbg("User major version %d not same as driver "
1857                           "major %d\n", uinfo->spu_userversion >> 16,
1858                           IPATH_USER_SWMAJOR);
1859                 ret = -ENODEV;
1860                 goto done;
1861         }
1862
1863         swminor = uinfo->spu_userversion & 0xffff;
1864         if (swminor != IPATH_USER_SWMINOR)
1865                 ipath_dbg("User minor version %d not same as driver "
1866                           "minor %d\n", swminor, IPATH_USER_SWMINOR);
1867
1868         mutex_lock(&ipath_mutex);
1869
1870         if (ipath_compatible_subports(swmajor, swminor) &&
1871             uinfo->spu_subport_cnt &&
1872             (ret = find_shared_port(fp, uinfo))) {
1873                 mutex_unlock(&ipath_mutex);
1874                 if (ret > 0)
1875                         ret = 0;
1876                 goto done;
1877         }
1878
1879         i_minor = iminor(fp->f_path.dentry->d_inode) - IPATH_USER_MINOR_BASE;
1880         ipath_cdbg(VERBOSE, "open on dev %lx (minor %d)\n",
1881                    (long)fp->f_path.dentry->d_inode->i_rdev, i_minor);
1882
1883         if (i_minor)
1884                 ret = find_free_port(i_minor - 1, fp, uinfo);
1885         else
1886                 ret = find_best_unit(fp, uinfo);
1887
1888         mutex_unlock(&ipath_mutex);
1889
1890 done:
1891         return ret;
1892 }
1893
1894
1895 static int ipath_do_user_init(struct file *fp,
1896                               const struct ipath_user_info *uinfo)
1897 {
1898         int ret;
1899         struct ipath_portdata *pd = port_fp(fp);
1900         struct ipath_devdata *dd;
1901         u32 head32;
1902
1903         /* Subports don't need to initialize anything since master did it. */
1904         if (subport_fp(fp)) {
1905                 ret = wait_event_interruptible(pd->port_wait,
1906                         !test_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag));
1907                 goto done;
1908         }
1909
1910         dd = pd->port_dd;
1911
1912         if (uinfo->spu_rcvhdrsize) {
1913                 ret = ipath_setrcvhdrsize(dd, uinfo->spu_rcvhdrsize);
1914                 if (ret)
1915                         goto done;
1916         }
1917
1918         /* for now we do nothing with rcvhdrcnt: uinfo->spu_rcvhdrcnt */
1919
1920         /* for right now, kernel piobufs are at end, so port 1 is at 0 */
1921         pd->port_piobufs = dd->ipath_piobufbase +
1922                 dd->ipath_pbufsport * (pd->port_port - 1) * dd->ipath_palign;
1923         ipath_cdbg(VERBOSE, "Set base of piobufs for port %u to 0x%x\n",
1924                    pd->port_port, pd->port_piobufs);
1925
1926         /*
1927          * Now allocate the rcvhdr Q and eager TIDs; skip the TID
1928          * array for time being.  If pd->port_port > chip-supported,
1929          * we need to do extra stuff here to handle by handling overflow
1930          * through port 0, someday
1931          */
1932         ret = ipath_create_rcvhdrq(dd, pd);
1933         if (!ret)
1934                 ret = ipath_create_user_egr(pd);
1935         if (ret)
1936                 goto done;
1937
1938         /*
1939          * set the eager head register for this port to the current values
1940          * of the tail pointers, since we don't know if they were
1941          * updated on last use of the port.
1942          */
1943         head32 = ipath_read_ureg32(dd, ur_rcvegrindextail, pd->port_port);
1944         ipath_write_ureg(dd, ur_rcvegrindexhead, head32, pd->port_port);
1945         pd->port_lastrcvhdrqtail = -1;
1946         ipath_cdbg(VERBOSE, "Wrote port%d egrhead %x from tail regs\n",
1947                 pd->port_port, head32);
1948         pd->port_tidcursor = 0; /* start at beginning after open */
1949
1950         /* initialize poll variables... */
1951         pd->port_urgent = 0;
1952         pd->port_urgent_poll = 0;
1953         pd->port_hdrqfull_poll = pd->port_hdrqfull;
1954
1955         /*
1956          * Now enable the port for receive.
1957          * For chips that are set to DMA the tail register to memory
1958          * when they change (and when the update bit transitions from
1959          * 0 to 1.  So for those chips, we turn it off and then back on.
1960          * This will (very briefly) affect any other open ports, but the
1961          * duration is very short, and therefore isn't an issue.  We
1962          * explictly set the in-memory tail copy to 0 beforehand, so we
1963          * don't have to wait to be sure the DMA update has happened
1964          * (chip resets head/tail to 0 on transition to enable).
1965          */
1966         set_bit(dd->ipath_r_portenable_shift + pd->port_port,
1967                 &dd->ipath_rcvctrl);
1968         if (!(dd->ipath_flags & IPATH_NODMA_RTAIL)) {
1969                 if (pd->port_rcvhdrtail_kvaddr)
1970                         ipath_clear_rcvhdrtail(pd);
1971                 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1972                         dd->ipath_rcvctrl &
1973                         ~(1ULL << dd->ipath_r_tailupd_shift));
1974         }
1975         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1976                          dd->ipath_rcvctrl);
1977         /* Notify any waiting slaves */
1978         if (pd->port_subport_cnt) {
1979                 clear_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag);
1980                 wake_up(&pd->port_wait);
1981         }
1982 done:
1983         return ret;
1984 }
1985
1986 /**
1987  * unlock_exptid - unlock any expected TID entries port still had in use
1988  * @pd: port
1989  *
1990  * We don't actually update the chip here, because we do a bulk update
1991  * below, using ipath_f_clear_tids.
1992  */
1993 static void unlock_expected_tids(struct ipath_portdata *pd)
1994 {
1995         struct ipath_devdata *dd = pd->port_dd;
1996         int port_tidbase = pd->port_port * dd->ipath_rcvtidcnt;
1997         int i, cnt = 0, maxtid = port_tidbase + dd->ipath_rcvtidcnt;
1998
1999         ipath_cdbg(VERBOSE, "Port %u unlocking any locked expTID pages\n",
2000                    pd->port_port);
2001         for (i = port_tidbase; i < maxtid; i++) {
2002                 struct page *ps = dd->ipath_pageshadow[i];
2003
2004                 if (!ps)
2005                         continue;
2006
2007                 dd->ipath_pageshadow[i] = NULL;
2008                 pci_unmap_page(dd->pcidev, dd->ipath_physshadow[i],
2009                         PAGE_SIZE, PCI_DMA_FROMDEVICE);
2010                 ipath_release_user_pages_on_close(&ps, 1);
2011                 cnt++;
2012                 ipath_stats.sps_pageunlocks++;
2013         }
2014         if (cnt)
2015                 ipath_cdbg(VERBOSE, "Port %u locked %u expTID entries\n",
2016                            pd->port_port, cnt);
2017
2018         if (ipath_stats.sps_pagelocks || ipath_stats.sps_pageunlocks)
2019                 ipath_cdbg(VERBOSE, "%llu pages locked, %llu unlocked\n",
2020                            (unsigned long long) ipath_stats.sps_pagelocks,
2021                            (unsigned long long)
2022                            ipath_stats.sps_pageunlocks);
2023 }
2024
2025 static int ipath_close(struct inode *in, struct file *fp)
2026 {
2027         int ret = 0;
2028         struct ipath_filedata *fd;
2029         struct ipath_portdata *pd;
2030         struct ipath_devdata *dd;
2031         unsigned port;
2032
2033         ipath_cdbg(VERBOSE, "close on dev %lx, private data %p\n",
2034                    (long)in->i_rdev, fp->private_data);
2035
2036         mutex_lock(&ipath_mutex);
2037
2038         fd = (struct ipath_filedata *) fp->private_data;
2039         fp->private_data = NULL;
2040         pd = fd->pd;
2041         if (!pd) {
2042                 mutex_unlock(&ipath_mutex);
2043                 goto bail;
2044         }
2045         if (--pd->port_cnt) {
2046                 /*
2047                  * XXX If the master closes the port before the slave(s),
2048                  * revoke the mmap for the eager receive queue so
2049                  * the slave(s) don't wait for receive data forever.
2050                  */
2051                 pd->active_slaves &= ~(1 << fd->subport);
2052                 pd->port_subpid[fd->subport] = 0;
2053                 mutex_unlock(&ipath_mutex);
2054                 goto bail;
2055         }
2056         port = pd->port_port;
2057         dd = pd->port_dd;
2058
2059         if (pd->port_hdrqfull) {
2060                 ipath_cdbg(PROC, "%s[%u] had %u rcvhdrqfull errors "
2061                            "during run\n", pd->port_comm, pd->port_pid,
2062                            pd->port_hdrqfull);
2063                 pd->port_hdrqfull = 0;
2064         }
2065
2066         if (pd->port_rcvwait_to || pd->port_piowait_to
2067             || pd->port_rcvnowait || pd->port_pionowait) {
2068                 ipath_cdbg(VERBOSE, "port%u, %u rcv, %u pio wait timeo; "
2069                            "%u rcv %u, pio already\n",
2070                            pd->port_port, pd->port_rcvwait_to,
2071                            pd->port_piowait_to, pd->port_rcvnowait,
2072                            pd->port_pionowait);
2073                 pd->port_rcvwait_to = pd->port_piowait_to =
2074                         pd->port_rcvnowait = pd->port_pionowait = 0;
2075         }
2076         if (pd->port_flag) {
2077                 ipath_cdbg(PROC, "port %u port_flag set: 0x%lx\n",
2078                           pd->port_port, pd->port_flag);
2079                 pd->port_flag = 0;
2080         }
2081
2082         if (dd->ipath_kregbase) {
2083                 int i;
2084                 /* atomically clear receive enable port and intr avail. */
2085                 clear_bit(dd->ipath_r_portenable_shift + port,
2086                           &dd->ipath_rcvctrl);
2087                 clear_bit(pd->port_port + dd->ipath_r_intravail_shift,
2088                           &dd->ipath_rcvctrl);
2089                 ipath_write_kreg( dd, dd->ipath_kregs->kr_rcvctrl,
2090                         dd->ipath_rcvctrl);
2091                 /* and read back from chip to be sure that nothing
2092                  * else is in flight when we do the rest */
2093                 (void)ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
2094
2095                 /* clean up the pkeys for this port user */
2096                 ipath_clean_part_key(pd, dd);
2097                 /*
2098                  * be paranoid, and never write 0's to these, just use an
2099                  * unused part of the port 0 tail page.  Of course,
2100                  * rcvhdraddr points to a large chunk of memory, so this
2101                  * could still trash things, but at least it won't trash
2102                  * page 0, and by disabling the port, it should stop "soon",
2103                  * even if a packet or two is in already in flight after we
2104                  * disabled the port.
2105                  */
2106                 ipath_write_kreg_port(dd,
2107                         dd->ipath_kregs->kr_rcvhdrtailaddr, port,
2108                         dd->ipath_dummy_hdrq_phys);
2109                 ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdraddr,
2110                         pd->port_port, dd->ipath_dummy_hdrq_phys);
2111
2112                 i = dd->ipath_pbufsport * (port - 1);
2113                 ipath_disarm_piobufs(dd, i, dd->ipath_pbufsport);
2114                 ipath_chg_pioavailkernel(dd, i, dd->ipath_pbufsport, 1);
2115
2116                 dd->ipath_f_clear_tids(dd, pd->port_port);
2117
2118                 if (dd->ipath_pageshadow)
2119                         unlock_expected_tids(pd);
2120                 ipath_stats.sps_ports--;
2121                 ipath_cdbg(PROC, "%s[%u] closed port %u:%u\n",
2122                            pd->port_comm, pd->port_pid,
2123                            dd->ipath_unit, port);
2124         }
2125
2126         pd->port_pid = 0;
2127         dd->ipath_pd[pd->port_port] = NULL; /* before releasing mutex */
2128         mutex_unlock(&ipath_mutex);
2129         ipath_free_pddata(dd, pd); /* after releasing the mutex */
2130
2131 bail:
2132         kfree(fd);
2133         return ret;
2134 }
2135
2136 static int ipath_port_info(struct ipath_portdata *pd, u16 subport,
2137                            struct ipath_port_info __user *uinfo)
2138 {
2139         struct ipath_port_info info;
2140         int nup;
2141         int ret;
2142         size_t sz;
2143
2144         (void) ipath_count_units(NULL, &nup, NULL);
2145         info.num_active = nup;
2146         info.unit = pd->port_dd->ipath_unit;
2147         info.port = pd->port_port;
2148         info.subport = subport;
2149         /* Don't return new fields if old library opened the port. */
2150         if (ipath_supports_subports(pd->userversion >> 16,
2151                                     pd->userversion & 0xffff)) {
2152                 /* Number of user ports available for this device. */
2153                 info.num_ports = pd->port_dd->ipath_cfgports - 1;
2154                 info.num_subports = pd->port_subport_cnt;
2155                 sz = sizeof(info);
2156         } else
2157                 sz = sizeof(info) - 2 * sizeof(u16);
2158
2159         if (copy_to_user(uinfo, &info, sz)) {
2160                 ret = -EFAULT;
2161                 goto bail;
2162         }
2163         ret = 0;
2164
2165 bail:
2166         return ret;
2167 }
2168
2169 static int ipath_get_slave_info(struct ipath_portdata *pd,
2170                                 void __user *slave_mask_addr)
2171 {
2172         int ret = 0;
2173
2174         if (copy_to_user(slave_mask_addr, &pd->active_slaves, sizeof(u32)))
2175                 ret = -EFAULT;
2176         return ret;
2177 }
2178
2179 static ssize_t ipath_write(struct file *fp, const char __user *data,
2180                            size_t count, loff_t *off)
2181 {
2182         const struct ipath_cmd __user *ucmd;
2183         struct ipath_portdata *pd;
2184         const void __user *src;
2185         size_t consumed, copy;
2186         struct ipath_cmd cmd;
2187         ssize_t ret = 0;
2188         void *dest;
2189
2190         if (count < sizeof(cmd.type)) {
2191                 ret = -EINVAL;
2192                 goto bail;
2193         }
2194
2195         ucmd = (const struct ipath_cmd __user *) data;
2196
2197         if (copy_from_user(&cmd.type, &ucmd->type, sizeof(cmd.type))) {
2198                 ret = -EFAULT;
2199                 goto bail;
2200         }
2201
2202         consumed = sizeof(cmd.type);
2203
2204         switch (cmd.type) {
2205         case IPATH_CMD_ASSIGN_PORT:
2206         case __IPATH_CMD_USER_INIT:
2207         case IPATH_CMD_USER_INIT:
2208                 copy = sizeof(cmd.cmd.user_info);
2209                 dest = &cmd.cmd.user_info;
2210                 src = &ucmd->cmd.user_info;
2211                 break;
2212         case IPATH_CMD_RECV_CTRL:
2213                 copy = sizeof(cmd.cmd.recv_ctrl);
2214                 dest = &cmd.cmd.recv_ctrl;
2215                 src = &ucmd->cmd.recv_ctrl;
2216                 break;
2217         case IPATH_CMD_PORT_INFO:
2218                 copy = sizeof(cmd.cmd.port_info);
2219                 dest = &cmd.cmd.port_info;
2220                 src = &ucmd->cmd.port_info;
2221                 break;
2222         case IPATH_CMD_TID_UPDATE:
2223         case IPATH_CMD_TID_FREE:
2224                 copy = sizeof(cmd.cmd.tid_info);
2225                 dest = &cmd.cmd.tid_info;
2226                 src = &ucmd->cmd.tid_info;
2227                 break;
2228         case IPATH_CMD_SET_PART_KEY:
2229                 copy = sizeof(cmd.cmd.part_key);
2230                 dest = &cmd.cmd.part_key;
2231                 src = &ucmd->cmd.part_key;
2232                 break;
2233         case __IPATH_CMD_SLAVE_INFO:
2234                 copy = sizeof(cmd.cmd.slave_mask_addr);
2235                 dest = &cmd.cmd.slave_mask_addr;
2236                 src = &ucmd->cmd.slave_mask_addr;
2237                 break;
2238         case IPATH_CMD_PIOAVAILUPD:     // force an update of PIOAvail reg
2239                 copy = 0;
2240                 src = NULL;
2241                 dest = NULL;
2242                 break;
2243         case IPATH_CMD_POLL_TYPE:
2244                 copy = sizeof(cmd.cmd.poll_type);
2245                 dest = &cmd.cmd.poll_type;
2246                 src = &ucmd->cmd.poll_type;
2247                 break;
2248         case IPATH_CMD_ARMLAUNCH_CTRL:
2249                 copy = sizeof(cmd.cmd.armlaunch_ctrl);
2250                 dest = &cmd.cmd.armlaunch_ctrl;
2251                 src = &ucmd->cmd.armlaunch_ctrl;
2252                 break;
2253         default:
2254                 ret = -EINVAL;
2255                 goto bail;
2256         }
2257
2258         if (copy) {
2259                 if ((count - consumed) < copy) {
2260                         ret = -EINVAL;
2261                         goto bail;
2262                 }
2263
2264                 if (copy_from_user(dest, src, copy)) {
2265                         ret = -EFAULT;
2266                         goto bail;
2267                 }
2268
2269                 consumed += copy;
2270         }
2271
2272         pd = port_fp(fp);
2273         if (!pd && cmd.type != __IPATH_CMD_USER_INIT &&
2274                 cmd.type != IPATH_CMD_ASSIGN_PORT) {
2275                 ret = -EINVAL;
2276                 goto bail;
2277         }
2278
2279         switch (cmd.type) {
2280         case IPATH_CMD_ASSIGN_PORT:
2281                 ret = ipath_assign_port(fp, &cmd.cmd.user_info);
2282                 if (ret)
2283                         goto bail;
2284                 break;
2285         case __IPATH_CMD_USER_INIT:
2286                 /* backwards compatibility, get port first */
2287                 ret = ipath_assign_port(fp, &cmd.cmd.user_info);
2288                 if (ret)
2289                         goto bail;
2290                 /* and fall through to current version. */
2291         case IPATH_CMD_USER_INIT:
2292                 ret = ipath_do_user_init(fp, &cmd.cmd.user_info);
2293                 if (ret)
2294                         goto bail;
2295                 ret = ipath_get_base_info(
2296                         fp, (void __user *) (unsigned long)
2297                         cmd.cmd.user_info.spu_base_info,
2298                         cmd.cmd.user_info.spu_base_info_size);
2299                 break;
2300         case IPATH_CMD_RECV_CTRL:
2301                 ret = ipath_manage_rcvq(pd, subport_fp(fp), cmd.cmd.recv_ctrl);
2302                 break;
2303         case IPATH_CMD_PORT_INFO:
2304                 ret = ipath_port_info(pd, subport_fp(fp),
2305                                       (struct ipath_port_info __user *)
2306                                       (unsigned long) cmd.cmd.port_info);
2307                 break;
2308         case IPATH_CMD_TID_UPDATE:
2309                 ret = ipath_tid_update(pd, fp, &cmd.cmd.tid_info);
2310                 break;
2311         case IPATH_CMD_TID_FREE:
2312                 ret = ipath_tid_free(pd, subport_fp(fp), &cmd.cmd.tid_info);
2313                 break;
2314         case IPATH_CMD_SET_PART_KEY:
2315                 ret = ipath_set_part_key(pd, cmd.cmd.part_key);
2316                 break;
2317         case __IPATH_CMD_SLAVE_INFO:
2318                 ret = ipath_get_slave_info(pd,
2319                                            (void __user *) (unsigned long)
2320                                            cmd.cmd.slave_mask_addr);
2321                 break;
2322         case IPATH_CMD_PIOAVAILUPD:
2323                 ipath_force_pio_avail_update(pd->port_dd);
2324                 break;
2325         case IPATH_CMD_POLL_TYPE:
2326                 pd->poll_type = cmd.cmd.poll_type;
2327                 break;
2328         case IPATH_CMD_ARMLAUNCH_CTRL:
2329                 if (cmd.cmd.armlaunch_ctrl)
2330                         ipath_enable_armlaunch(pd->port_dd);
2331                 else
2332                         ipath_disable_armlaunch(pd->port_dd);
2333                 break;
2334         }
2335
2336         if (ret >= 0)
2337                 ret = consumed;
2338
2339 bail:
2340         return ret;
2341 }
2342
2343 static struct class *ipath_class;
2344
2345 static int init_cdev(int minor, char *name, const struct file_operations *fops,
2346                      struct cdev **cdevp, struct class_device **class_devp)
2347 {
2348         const dev_t dev = MKDEV(IPATH_MAJOR, minor);
2349         struct cdev *cdev = NULL;
2350         struct class_device *class_dev = NULL;
2351         int ret;
2352
2353         cdev = cdev_alloc();
2354         if (!cdev) {
2355                 printk(KERN_ERR IPATH_DRV_NAME
2356                        ": Could not allocate cdev for minor %d, %s\n",
2357                        minor, name);
2358                 ret = -ENOMEM;
2359                 goto done;
2360         }
2361
2362         cdev->owner = THIS_MODULE;
2363         cdev->ops = fops;
2364         kobject_set_name(&cdev->kobj, name);
2365
2366         ret = cdev_add(cdev, dev, 1);
2367         if (ret < 0) {
2368                 printk(KERN_ERR IPATH_DRV_NAME
2369                        ": Could not add cdev for minor %d, %s (err %d)\n",
2370                        minor, name, -ret);
2371                 goto err_cdev;
2372         }
2373
2374         class_dev = class_device_create(ipath_class, NULL, dev, NULL, name);
2375
2376         if (IS_ERR(class_dev)) {
2377                 ret = PTR_ERR(class_dev);
2378                 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2379                        "class_dev for minor %d, %s (err %d)\n",
2380                        minor, name, -ret);
2381                 goto err_cdev;
2382         }
2383
2384         goto done;
2385
2386 err_cdev:
2387         cdev_del(cdev);
2388         cdev = NULL;
2389
2390 done:
2391         if (ret >= 0) {
2392                 *cdevp = cdev;
2393                 *class_devp = class_dev;
2394         } else {
2395                 *cdevp = NULL;
2396                 *class_devp = NULL;
2397         }
2398
2399         return ret;
2400 }
2401
2402 int ipath_cdev_init(int minor, char *name, const struct file_operations *fops,
2403                     struct cdev **cdevp, struct class_device **class_devp)
2404 {
2405         return init_cdev(minor, name, fops, cdevp, class_devp);
2406 }
2407
2408 static void cleanup_cdev(struct cdev **cdevp,
2409                          struct class_device **class_devp)
2410 {
2411         struct class_device *class_dev = *class_devp;
2412
2413         if (class_dev) {
2414                 class_device_unregister(class_dev);
2415                 *class_devp = NULL;
2416         }
2417
2418         if (*cdevp) {
2419                 cdev_del(*cdevp);
2420                 *cdevp = NULL;
2421         }
2422 }
2423
2424 void ipath_cdev_cleanup(struct cdev **cdevp,
2425                         struct class_device **class_devp)
2426 {
2427         cleanup_cdev(cdevp, class_devp);
2428 }
2429
2430 static struct cdev *wildcard_cdev;
2431 static struct class_device *wildcard_class_dev;
2432
2433 static const dev_t dev = MKDEV(IPATH_MAJOR, 0);
2434
2435 static int user_init(void)
2436 {
2437         int ret;
2438
2439         ret = register_chrdev_region(dev, IPATH_NMINORS, IPATH_DRV_NAME);
2440         if (ret < 0) {
2441                 printk(KERN_ERR IPATH_DRV_NAME ": Could not register "
2442                        "chrdev region (err %d)\n", -ret);
2443                 goto done;
2444         }
2445
2446         ipath_class = class_create(THIS_MODULE, IPATH_DRV_NAME);
2447
2448         if (IS_ERR(ipath_class)) {
2449                 ret = PTR_ERR(ipath_class);
2450                 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2451                        "device class (err %d)\n", -ret);
2452                 goto bail;
2453         }
2454
2455         goto done;
2456 bail:
2457         unregister_chrdev_region(dev, IPATH_NMINORS);
2458 done:
2459         return ret;
2460 }
2461
2462 static void user_cleanup(void)
2463 {
2464         if (ipath_class) {
2465                 class_destroy(ipath_class);
2466                 ipath_class = NULL;
2467         }
2468
2469         unregister_chrdev_region(dev, IPATH_NMINORS);
2470 }
2471
2472 static atomic_t user_count = ATOMIC_INIT(0);
2473 static atomic_t user_setup = ATOMIC_INIT(0);
2474
2475 int ipath_user_add(struct ipath_devdata *dd)
2476 {
2477         char name[10];
2478         int ret;
2479
2480         if (atomic_inc_return(&user_count) == 1) {
2481                 ret = user_init();
2482                 if (ret < 0) {
2483                         ipath_dev_err(dd, "Unable to set up user support: "
2484                                       "error %d\n", -ret);
2485                         goto bail;
2486                 }
2487                 ret = init_cdev(0, "ipath", &ipath_file_ops, &wildcard_cdev,
2488                                 &wildcard_class_dev);
2489                 if (ret < 0) {
2490                         ipath_dev_err(dd, "Could not create wildcard "
2491                                       "minor: error %d\n", -ret);
2492                         goto bail_user;
2493                 }
2494
2495                 atomic_set(&user_setup, 1);
2496         }
2497
2498         snprintf(name, sizeof(name), "ipath%d", dd->ipath_unit);
2499
2500         ret = init_cdev(dd->ipath_unit + 1, name, &ipath_file_ops,
2501                         &dd->user_cdev, &dd->user_class_dev);
2502         if (ret < 0)
2503                 ipath_dev_err(dd, "Could not create user minor %d, %s\n",
2504                               dd->ipath_unit + 1, name);
2505
2506         goto bail;
2507
2508 bail_user:
2509         user_cleanup();
2510 bail:
2511         return ret;
2512 }
2513
2514 void ipath_user_remove(struct ipath_devdata *dd)
2515 {
2516         cleanup_cdev(&dd->user_cdev, &dd->user_class_dev);
2517
2518         if (atomic_dec_return(&user_count) == 0) {
2519                 if (atomic_read(&user_setup) == 0)
2520                         goto bail;
2521
2522                 cleanup_cdev(&wildcard_cdev, &wildcard_class_dev);
2523                 user_cleanup();
2524
2525                 atomic_set(&user_setup, 0);
2526         }
2527 bail:
2528         return;
2529 }