drivers/infiniband/hw/ipath/ipath_file_ops.c

   1 /*
   2  * Copyright (c) 2006, 2007, 2008 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 <linux/highmem.h>
  40 #include <linux/io.h>
  41 #include <linux/jiffies.h>
  42 #include <asm/pgtable.h>
  43
  44 #include "ipath_kernel.h"
  45 #include "ipath_common.h"
  46 #include "ipath_user_sdma.h"
  47
  48 static int ipath_open(struct inode *, struct file *);
  49 static int ipath_close(struct inode *, struct file *);
  50 static ssize_t ipath_write(struct file *, const char __user *, size_t,
  51                            loff_t *);
  52 static ssize_t ipath_writev(struct kiocb *, const struct iovec *,
  53                             unsigned long , loff_t);
  54 static unsigned int ipath_poll(struct file *, struct poll_table_struct *);
  55 static int ipath_mmap(struct file *, struct vm_area_struct *);
  56
  57 static const struct file_operations ipath_file_ops = {
  58         .owner = THIS_MODULE,
  59         .write = ipath_write,
  60         .aio_write = ipath_writev,
  61         .open = ipath_open,
  62         .release = ipath_close,
  63         .poll = ipath_poll,
  64         .mmap = ipath_mmap
  65 };
  66
  67 /*
  68  * Convert kernel virtual addresses to physical addresses so they don't
  69  * potentially conflict with the chip addresses used as mmap offsets.
  70  * It doesn't really matter what mmap offset we use as long as we can
  71  * interpret it correctly.
  72  */
  73 static u64 cvt_kvaddr(void *p)
  74 {
  75         struct page *page;
  76         u64 paddr = 0;
  77
  78         page = vmalloc_to_page(p);
  79         if (page)
  80                 paddr = page_to_pfn(page) << PAGE_SHIFT;
  81
  82         return paddr;
  83 }
  84
  85 static int ipath_get_base_info(struct file *fp,
  86                                void __user *ubase, size_t ubase_size)
  87 {
  88         struct ipath_portdata *pd = port_fp(fp);
  89         int ret = 0;
  90         struct ipath_base_info *kinfo = NULL;
  91         struct ipath_devdata *dd = pd->port_dd;
  92         unsigned subport_cnt;
  93         int shared, master;
  94         size_t sz;
  95
  96         subport_cnt = pd->port_subport_cnt;
  97         if (!subport_cnt) {
  98                 shared = 0;
  99                 master = 0;
 100                 subport_cnt = 1;
 101         } else {
 102                 shared = 1;
 103                 master = !subport_fp(fp);
 104         }
 105
 106         sz = sizeof(*kinfo);
 107         /* If port sharing is not requested, allow the old size structure */
 108         if (!shared)
 109                 sz -= 7 * sizeof(u64);
 110         if (ubase_size < sz) {
 111                 ipath_cdbg(PROC,
 112                            "Base size %zu, need %zu (version mismatch?)\n",
 113                            ubase_size, sz);
 114                 ret = -EINVAL;
 115                 goto bail;
 116         }
 117
 118         kinfo = kzalloc(sizeof(*kinfo), GFP_KERNEL);
 119         if (kinfo == NULL) {
 120                 ret = -ENOMEM;
 121                 goto bail;
 122         }
 123
 124         ret = dd->ipath_f_get_base_info(pd, kinfo);
 125         if (ret < 0)
 126                 goto bail;
 127
 128         kinfo->spi_rcvhdr_cnt = dd->ipath_rcvhdrcnt;
 129         kinfo->spi_rcvhdrent_size = dd->ipath_rcvhdrentsize;
 130         kinfo->spi_tidegrcnt = dd->ipath_rcvegrcnt;
 131         kinfo->spi_rcv_egrbufsize = dd->ipath_rcvegrbufsize;
 132         /*
 133          * have to mmap whole thing
 134          */
 135         kinfo->spi_rcv_egrbuftotlen =
 136                 pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
 137         kinfo->spi_rcv_egrperchunk = pd->port_rcvegrbufs_perchunk;
 138         kinfo->spi_rcv_egrchunksize = kinfo->spi_rcv_egrbuftotlen /
 139                 pd->port_rcvegrbuf_chunks;
 140         kinfo->spi_tidcnt = dd->ipath_rcvtidcnt / subport_cnt;
 141         if (master)
 142                 kinfo->spi_tidcnt += dd->ipath_rcvtidcnt % subport_cnt;
 143         /*
 144          * for this use, may be ipath_cfgports summed over all chips that
 145          * are are configured and present
 146          */
 147         kinfo->spi_nports = dd->ipath_cfgports;
 148         /* unit (chip/board) our port is on */
 149         kinfo->spi_unit = dd->ipath_unit;
 150         /* for now, only a single page */
 151         kinfo->spi_tid_maxsize = PAGE_SIZE;
 152
 153         /*
 154          * Doing this per port, and based on the skip value, etc.  This has
 155          * to be the actual buffer size, since the protocol code treats it
 156          * as an array.
 157          *
 158          * These have to be set to user addresses in the user code via mmap.
 159          * These values are used on return to user code for the mmap target
 160          * addresses only.  For 32 bit, same 44 bit address problem, so use
 161          * the physical address, not virtual.  Before 2.6.11, using the
 162          * page_address() macro worked, but in 2.6.11, even that returns the
 163          * full 64 bit address (upper bits all 1's).  So far, using the
 164          * physical addresses (or chip offsets, for chip mapping) works, but
 165          * no doubt some future kernel release will change that, and we'll be
 166          * on to yet another method of dealing with this.
 167          */
 168         kinfo->spi_rcvhdr_base = (u64) pd->port_rcvhdrq_phys;
 169         kinfo->spi_rcvhdr_tailaddr = (u64) pd->port_rcvhdrqtailaddr_phys;
 170         kinfo->spi_rcv_egrbufs = (u64) pd->port_rcvegr_phys;
 171         kinfo->spi_pioavailaddr = (u64) dd->ipath_pioavailregs_phys;
 172         kinfo->spi_status = (u64) kinfo->spi_pioavailaddr +
 173                 (void *) dd->ipath_statusp -
 174                 (void *) dd->ipath_pioavailregs_dma;
 175         if (!shared) {
 176                 kinfo->spi_piocnt = dd->ipath_pbufsport;
 177                 kinfo->spi_piobufbase = (u64) pd->port_piobufs;
 178                 kinfo->__spi_uregbase = (u64) dd->ipath_uregbase +
 179                         dd->ipath_ureg_align * pd->port_port;
 180         } else if (master) {
 181                 kinfo->spi_piocnt = (dd->ipath_pbufsport / subport_cnt) +
 182                                     (dd->ipath_pbufsport % subport_cnt);
 183                 /* Master's PIO buffers are after all the slave's */
 184                 kinfo->spi_piobufbase = (u64) pd->port_piobufs +
 185                         dd->ipath_palign *
 186                         (dd->ipath_pbufsport - kinfo->spi_piocnt);
 187         } else {
 188                 unsigned slave = subport_fp(fp) - 1;
 189
 190                 kinfo->spi_piocnt = dd->ipath_pbufsport / subport_cnt;
 191                 kinfo->spi_piobufbase = (u64) pd->port_piobufs +
 192                         dd->ipath_palign * kinfo->spi_piocnt * slave;
 193         }
 194
 195         /*
 196          * Set the PIO avail update threshold to no larger
 197          * than the number of buffers per process. Note that
 198          * we decrease it here, but won't ever increase it.
 199          */
 200         if (dd->ipath_pioupd_thresh &&
 201             kinfo->spi_piocnt < dd->ipath_pioupd_thresh) {
 202                 unsigned long flags;
 203
 204                 dd->ipath_pioupd_thresh = kinfo->spi_piocnt;
 205                 ipath_dbg("Decreased pio update threshold to %u\n",
 206                         dd->ipath_pioupd_thresh);
 207                 spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
 208                 dd->ipath_sendctrl &= ~(INFINIPATH_S_UPDTHRESH_MASK
 209                         << INFINIPATH_S_UPDTHRESH_SHIFT);
 210                 dd->ipath_sendctrl |= dd->ipath_pioupd_thresh
 211                         << INFINIPATH_S_UPDTHRESH_SHIFT;
 212                 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
 213                         dd->ipath_sendctrl);
 214                 spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
 215         }
 216
 217         if (shared) {
 218                 kinfo->spi_port_uregbase = (u64) dd->ipath_uregbase +
 219                         dd->ipath_ureg_align * pd->port_port;
 220                 kinfo->spi_port_rcvegrbuf = kinfo->spi_rcv_egrbufs;
 221                 kinfo->spi_port_rcvhdr_base = kinfo->spi_rcvhdr_base;
 222                 kinfo->spi_port_rcvhdr_tailaddr = kinfo->spi_rcvhdr_tailaddr;
 223
 224                 kinfo->__spi_uregbase = cvt_kvaddr(pd->subport_uregbase +
 225                         PAGE_SIZE * subport_fp(fp));
 226
 227                 kinfo->spi_rcvhdr_base = cvt_kvaddr(pd->subport_rcvhdr_base +
 228                         pd->port_rcvhdrq_size * subport_fp(fp));
 229                 kinfo->spi_rcvhdr_tailaddr = 0;
 230                 kinfo->spi_rcv_egrbufs = cvt_kvaddr(pd->subport_rcvegrbuf +
 231                         pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size *
 232                         subport_fp(fp));
 233
 234                 kinfo->spi_subport_uregbase =
 235                         cvt_kvaddr(pd->subport_uregbase);
 236                 kinfo->spi_subport_rcvegrbuf =
 237                         cvt_kvaddr(pd->subport_rcvegrbuf);
 238                 kinfo->spi_subport_rcvhdr_base =
 239                         cvt_kvaddr(pd->subport_rcvhdr_base);
 240                 ipath_cdbg(PROC, "port %u flags %x %llx %llx %llx\n",
 241                         kinfo->spi_port, kinfo->spi_runtime_flags,
 242                         (unsigned long long) kinfo->spi_subport_uregbase,
 243                         (unsigned long long) kinfo->spi_subport_rcvegrbuf,
 244                         (unsigned long long) kinfo->spi_subport_rcvhdr_base);
 245         }
 246
 247         kinfo->spi_pioindex = (kinfo->spi_piobufbase - dd->ipath_piobufbase) /
 248                 dd->ipath_palign;
 249         kinfo->spi_pioalign = dd->ipath_palign;
 250
 251         kinfo->spi_qpair = IPATH_KD_QP;
 252         /*
 253          * user mode PIO buffers are always 2KB, even when 4KB can
 254          * be received, and sent via the kernel; this is ibmaxlen
 255          * for 2K MTU.
 256          */
 257         kinfo->spi_piosize = dd->ipath_piosize2k - 2 * sizeof(u32);
 258         kinfo->spi_mtu = dd->ipath_ibmaxlen;    /* maxlen, not ibmtu */
 259         kinfo->spi_port = pd->port_port;
 260         kinfo->spi_subport = subport_fp(fp);
 261         kinfo->spi_sw_version = IPATH_KERN_SWVERSION;
 262         kinfo->spi_hw_version = dd->ipath_revision;
 263
 264         if (master) {
 265                 kinfo->spi_runtime_flags |= IPATH_RUNTIME_MASTER;
 266         }
 267
 268         sz = (ubase_size < sizeof(*kinfo)) ? ubase_size : sizeof(*kinfo);
 269         if (copy_to_user(ubase, kinfo, sz))
 270                 ret = -EFAULT;
 271
 272 bail:
 273         kfree(kinfo);
 274         return ret;
 275 }
 276
 277 /**
 278  * ipath_tid_update - update a port TID
 279  * @pd: the port
 280  * @fp: the ipath device file
 281  * @ti: the TID information
 282  *
 283  * The new implementation as of Oct 2004 is that the driver assigns
 284  * the tid and returns it to the caller.   To make it easier to
 285  * catch bugs, and to reduce search time, we keep a cursor for
 286  * each port, walking the shadow tid array to find one that's not
 287  * in use.
 288  *
 289  * For now, if we can't allocate the full list, we fail, although
 290  * in the long run, we'll allocate as many as we can, and the
 291  * caller will deal with that by trying the remaining pages later.
 292  * That means that when we fail, we have to mark the tids as not in
 293  * use again, in our shadow copy.
 294  *
 295  * It's up to the caller to free the tids when they are done.
 296  * We'll unlock the pages as they free them.
 297  *
 298  * Also, right now we are locking one page at a time, but since
 299  * the intended use of this routine is for a single group of
 300  * virtually contiguous pages, that should change to improve
 301  * performance.
 302  */
 303 static int ipath_tid_update(struct ipath_portdata *pd, struct file *fp,
 304                             const struct ipath_tid_info *ti)
 305 {
 306         int ret = 0, ntids;
 307         u32 tid, porttid, cnt, i, tidcnt, tidoff;
 308         u16 *tidlist;
 309         struct ipath_devdata *dd = pd->port_dd;
 310         u64 physaddr;
 311         unsigned long vaddr;
 312         u64 __iomem *tidbase;
 313         unsigned long tidmap[8];
 314         struct page **pagep = NULL;
 315         unsigned subport = subport_fp(fp);
 316
 317         if (!dd->ipath_pageshadow) {
 318                 ret = -ENOMEM;
 319                 goto done;
 320         }
 321
 322         cnt = ti->tidcnt;
 323         if (!cnt) {
 324                 ipath_dbg("After copyin, tidcnt 0, tidlist %llx\n",
 325                           (unsigned long long) ti->tidlist);
 326                 /*
 327                  * Should we treat as success?  likely a bug
 328                  */
 329                 ret = -EFAULT;
 330                 goto done;
 331         }
 332         porttid = pd->port_port * dd->ipath_rcvtidcnt;
 333         if (!pd->port_subport_cnt) {
 334                 tidcnt = dd->ipath_rcvtidcnt;
 335                 tid = pd->port_tidcursor;
 336                 tidoff = 0;
 337         } else if (!subport) {
 338                 tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
 339                          (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
 340                 tidoff = dd->ipath_rcvtidcnt - tidcnt;
 341                 porttid += tidoff;
 342                 tid = tidcursor_fp(fp);
 343         } else {
 344                 tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
 345                 tidoff = tidcnt * (subport - 1);
 346                 porttid += tidoff;
 347                 tid = tidcursor_fp(fp);
 348         }
 349         if (cnt > tidcnt) {
 350                 /* make sure it all fits in port_tid_pg_list */
 351                 dev_info(&dd->pcidev->dev, "Process tried to allocate %u "
 352                          "TIDs, only trying max (%u)\n", cnt, tidcnt);
 353                 cnt = tidcnt;
 354         }
 355         pagep = &((struct page **) pd->port_tid_pg_list)[tidoff];
 356         tidlist = &((u16 *) &pagep[dd->ipath_rcvtidcnt])[tidoff];
 357
 358         memset(tidmap, 0, sizeof(tidmap));
 359         /* before decrement; chip actual # */
 360         ntids = tidcnt;
 361         tidbase = (u64 __iomem *) (((char __iomem *) dd->ipath_kregbase) +
 362                                    dd->ipath_rcvtidbase +
 363                                    porttid * sizeof(*tidbase));
 364
 365         ipath_cdbg(VERBOSE, "Port%u %u tids, cursor %u, tidbase %p\n",
 366                    pd->port_port, cnt, tid, tidbase);
 367
 368         /* virtual address of first page in transfer */
 369         vaddr = ti->tidvaddr;
 370         if (!access_ok(VERIFY_WRITE, (void __user *) vaddr,
 371                        cnt * PAGE_SIZE)) {
 372                 ipath_dbg("Fail vaddr %p, %u pages, !access_ok\n",
 373                           (void *)vaddr, cnt);
 374                 ret = -EFAULT;
 375                 goto done;
 376         }
 377         ret = ipath_get_user_pages(vaddr, cnt, pagep);
 378         if (ret) {
 379                 if (ret == -EBUSY) {
 380                         ipath_dbg("Failed to lock addr %p, %u pages "
 381                                   "(already locked)\n",
 382                                   (void *) vaddr, cnt);
 383                         /*
 384                          * for now, continue, and see what happens but with
 385                          * the new implementation, this should never happen,
 386                          * unless perhaps the user has mpin'ed the pages
 387                          * themselves (something we need to test)
 388                          */
 389                         ret = 0;
 390                 } else {
 391                         dev_info(&dd->pcidev->dev,
 392                                  "Failed to lock addr %p, %u pages: "
 393                                  "errno %d\n", (void *) vaddr, cnt, -ret);
 394                         goto done;
 395                 }
 396         }
 397         for (i = 0; i < cnt; i++, vaddr += PAGE_SIZE) {
 398                 for (; ntids--; tid++) {
 399                         if (tid == tidcnt)
 400                                 tid = 0;
 401                         if (!dd->ipath_pageshadow[porttid + tid])
 402                                 break;
 403                 }
 404                 if (ntids < 0) {
 405                         /*
 406                          * oops, wrapped all the way through their TIDs,
 407                          * and didn't have enough free; see comments at
 408                          * start of routine
 409                          */
 410                         ipath_dbg("Not enough free TIDs for %u pages "
 411                                   "(index %d), failing\n", cnt, i);
 412                         i--;    /* last tidlist[i] not filled in */
 413                         ret = -ENOMEM;
 414                         break;
 415                 }
 416                 tidlist[i] = tid + tidoff;
 417                 ipath_cdbg(VERBOSE, "Updating idx %u to TID %u, "
 418                            "vaddr %lx\n", i, tid + tidoff, vaddr);
 419                 /* we "know" system pages and TID pages are same size */
 420                 dd->ipath_pageshadow[porttid + tid] = pagep[i];
 421                 dd->ipath_physshadow[porttid + tid] = ipath_map_page(
 422                         dd->pcidev, pagep[i], 0, PAGE_SIZE,
 423                         PCI_DMA_FROMDEVICE);
 424                 /*
 425                  * don't need atomic or it's overhead
 426                  */
 427                 __set_bit(tid, tidmap);
 428                 physaddr = dd->ipath_physshadow[porttid + tid];
 429                 ipath_stats.sps_pagelocks++;
 430                 ipath_cdbg(VERBOSE,
 431                            "TID %u, vaddr %lx, physaddr %llx pgp %p\n",
 432                            tid, vaddr, (unsigned long long) physaddr,
 433                            pagep[i]);
 434                 dd->ipath_f_put_tid(dd, &tidbase[tid], RCVHQ_RCV_TYPE_EXPECTED,
 435                                     physaddr);
 436                 /*
 437                  * don't check this tid in ipath_portshadow, since we
 438                  * just filled it in; start with the next one.
 439                  */
 440                 tid++;
 441         }
 442
 443         if (ret) {
 444                 u32 limit;
 445         cleanup:
 446                 /* jump here if copy out of updated info failed... */
 447                 ipath_dbg("After failure (ret=%d), undo %d of %d entries\n",
 448                           -ret, i, cnt);
 449                 /* same code that's in ipath_free_tid() */
 450                 limit = sizeof(tidmap) * BITS_PER_BYTE;
 451                 if (limit > tidcnt)
 452                         /* just in case size changes in future */
 453                         limit = tidcnt;
 454                 tid = find_first_bit((const unsigned long *)tidmap, limit);
 455                 for (; tid < limit; tid++) {
 456                         if (!test_bit(tid, tidmap))
 457                                 continue;
 458                         if (dd->ipath_pageshadow[porttid + tid]) {
 459                                 ipath_cdbg(VERBOSE, "Freeing TID %u\n",
 460                                            tid);
 461                                 dd->ipath_f_put_tid(dd, &tidbase[tid],
 462                                                     RCVHQ_RCV_TYPE_EXPECTED,
 463                                                     dd->ipath_tidinvalid);
 464                                 pci_unmap_page(dd->pcidev,
 465                                         dd->ipath_physshadow[porttid + tid],
 466                                         PAGE_SIZE, PCI_DMA_FROMDEVICE);
 467                                 dd->ipath_pageshadow[porttid + tid] = NULL;
 468                                 ipath_stats.sps_pageunlocks++;
 469                         }
 470                 }
 471                 ipath_release_user_pages(pagep, cnt);
 472         } else {
 473                 /*
 474                  * Copy the updated array, with ipath_tid's filled in, back
 475                  * to user.  Since we did the copy in already, this "should
 476                  * never fail" If it does, we have to clean up...
 477                  */
 478                 if (copy_to_user((void __user *)
 479                                  (unsigned long) ti->tidlist,
 480                                  tidlist, cnt * sizeof(*tidlist))) {
 481                         ret = -EFAULT;
 482                         goto cleanup;
 483                 }
 484                 if (copy_to_user((void __user *) (unsigned long) ti->tidmap,
 485                                  tidmap, sizeof tidmap)) {
 486                         ret = -EFAULT;
 487                         goto cleanup;
 488                 }
 489                 if (tid == tidcnt)
 490                         tid = 0;
 491                 if (!pd->port_subport_cnt)
 492                         pd->port_tidcursor = tid;
 493                 else
 494                         tidcursor_fp(fp) = tid;
 495         }
 496
 497 done:
 498         if (ret)
 499                 ipath_dbg("Failed to map %u TID pages, failing with %d\n",
 500                           ti->tidcnt, -ret);
 501         return ret;
 502 }
 503
 504 /**
 505  * ipath_tid_free - free a port TID
 506  * @pd: the port
 507  * @subport: the subport
 508  * @ti: the TID info
 509  *
 510  * right now we are unlocking one page at a time, but since
 511  * the intended use of this routine is for a single group of
 512  * virtually contiguous pages, that should change to improve
 513  * performance.  We check that the TID is in range for this port
 514  * but otherwise don't check validity; if user has an error and
 515  * frees the wrong tid, it's only their own data that can thereby
 516  * be corrupted.  We do check that the TID was in use, for sanity
 517  * We always use our idea of the saved address, not the address that
 518  * they pass in to us.
 519  */
 520
 521 static int ipath_tid_free(struct ipath_portdata *pd, unsigned subport,
 522                           const struct ipath_tid_info *ti)
 523 {
 524         int ret = 0;
 525         u32 tid, porttid, cnt, limit, tidcnt;
 526         struct ipath_devdata *dd = pd->port_dd;
 527         u64 __iomem *tidbase;
 528         unsigned long tidmap[8];
 529
 530         if (!dd->ipath_pageshadow) {
 531                 ret = -ENOMEM;
 532                 goto done;
 533         }
 534
 535         if (copy_from_user(tidmap, (void __user *)(unsigned long)ti->tidmap,
 536                            sizeof tidmap)) {
 537                 ret = -EFAULT;
 538                 goto done;
 539         }
 540
 541         porttid = pd->port_port * dd->ipath_rcvtidcnt;
 542         if (!pd->port_subport_cnt)
 543                 tidcnt = dd->ipath_rcvtidcnt;
 544         else if (!subport) {
 545                 tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
 546                          (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
 547                 porttid += dd->ipath_rcvtidcnt - tidcnt;
 548         } else {
 549                 tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
 550                 porttid += tidcnt * (subport - 1);
 551         }
 552         tidbase = (u64 __iomem *) ((char __iomem *)(dd->ipath_kregbase) +
 553                                    dd->ipath_rcvtidbase +
 554                                    porttid * sizeof(*tidbase));
 555
 556         limit = sizeof(tidmap) * BITS_PER_BYTE;
 557         if (limit > tidcnt)
 558                 /* just in case size changes in future */
 559                 limit = tidcnt;
 560         tid = find_first_bit(tidmap, limit);
 561         ipath_cdbg(VERBOSE, "Port%u free %u tids; first bit (max=%d) "
 562                    "set is %d, porttid %u\n", pd->port_port, ti->tidcnt,
 563                    limit, tid, porttid);
 564         for (cnt = 0; tid < limit; tid++) {
 565                 /*
 566                  * small optimization; if we detect a run of 3 or so without
 567                  * any set, use find_first_bit again.  That's mainly to
 568                  * accelerate the case where we wrapped, so we have some at
 569                  * the beginning, and some at the end, and a big gap
 570                  * in the middle.
 571                  */
 572                 if (!test_bit(tid, tidmap))
 573                         continue;
 574                 cnt++;
 575                 if (dd->ipath_pageshadow[porttid + tid]) {
 576                         struct page *p;
 577                         p = dd->ipath_pageshadow[porttid + tid];
 578                         dd->ipath_pageshadow[porttid + tid] = NULL;
 579                         ipath_cdbg(VERBOSE, "PID %u freeing TID %u\n",
 580                                    pd->port_pid, tid);
 581                         dd->ipath_f_put_tid(dd, &tidbase[tid],
 582                                             RCVHQ_RCV_TYPE_EXPECTED,
 583                                             dd->ipath_tidinvalid);
 584                         pci_unmap_page(dd->pcidev,
 585                                 dd->ipath_physshadow[porttid + tid],
 586                                 PAGE_SIZE, PCI_DMA_FROMDEVICE);
 587                         ipath_release_user_pages(&p, 1);
 588                         ipath_stats.sps_pageunlocks++;
 589                 } else
 590                         ipath_dbg("Unused tid %u, ignoring\n", tid);
 591         }
 592         if (cnt != ti->tidcnt)
 593                 ipath_dbg("passed in tidcnt %d, only %d bits set in map\n",
 594                           ti->tidcnt, cnt);
 595 done:
 596         if (ret)
 597                 ipath_dbg("Failed to unmap %u TID pages, failing with %d\n",
 598                           ti->tidcnt, -ret);
 599         return ret;
 600 }
 601
 602 /**
 603  * ipath_set_part_key - set a partition key
 604  * @pd: the port
 605  * @key: the key
 606  *
 607  * We can have up to 4 active at a time (other than the default, which is
 608  * always allowed).  This is somewhat tricky, since multiple ports may set
 609  * the same key, so we reference count them, and clean up at exit.  All 4
 610  * partition keys are packed into a single infinipath register.  It's an
 611  * error for a process to set the same pkey multiple times.  We provide no
 612  * mechanism to de-allocate a pkey at this time, we may eventually need to
 613  * do that.  I've used the atomic operations, and no locking, and only make
 614  * a single pass through what's available.  This should be more than
 615  * adequate for some time. I'll think about spinlocks or the like if and as
 616  * it's necessary.
 617  */
 618 static int ipath_set_part_key(struct ipath_portdata *pd, u16 key)
 619 {
 620         struct ipath_devdata *dd = pd->port_dd;
 621         int i, any = 0, pidx = -1;
 622         u16 lkey = key & 0x7FFF;
 623         int ret;
 624
 625         if (lkey == (IPATH_DEFAULT_P_KEY & 0x7FFF)) {
 626                 /* nothing to do; this key always valid */
 627                 ret = 0;
 628                 goto bail;
 629         }
 630
 631         ipath_cdbg(VERBOSE, "p%u try to set pkey %hx, current keys "
 632                    "%hx:%x %hx:%x %hx:%x %hx:%x\n",
 633                    pd->port_port, key, dd->ipath_pkeys[0],
 634                    atomic_read(&dd->ipath_pkeyrefs[0]), dd->ipath_pkeys[1],
 635                    atomic_read(&dd->ipath_pkeyrefs[1]), dd->ipath_pkeys[2],
 636                    atomic_read(&dd->ipath_pkeyrefs[2]), dd->ipath_pkeys[3],
 637                    atomic_read(&dd->ipath_pkeyrefs[3]));
 638
 639         if (!lkey) {
 640                 ipath_cdbg(PROC, "p%u tries to set key 0, not allowed\n",
 641                            pd->port_port);
 642                 ret = -EINVAL;
 643                 goto bail;
 644         }
 645
 646         /*
 647          * Set the full membership bit, because it has to be
 648          * set in the register or the packet, and it seems
 649          * cleaner to set in the register than to force all
 650          * callers to set it. (see bug 4331)
 651          */
 652         key |= 0x8000;
 653
 654         for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
 655                 if (!pd->port_pkeys[i] && pidx == -1)
 656                         pidx = i;
 657                 if (pd->port_pkeys[i] == key) {
 658                         ipath_cdbg(VERBOSE, "p%u tries to set same pkey "
 659                                    "(%x) more than once\n",
 660                                    pd->port_port, key);
 661                         ret = -EEXIST;
 662                         goto bail;
 663                 }
 664         }
 665         if (pidx == -1) {
 666                 ipath_dbg("All pkeys for port %u already in use, "
 667                           "can't set %x\n", pd->port_port, key);
 668                 ret = -EBUSY;
 669                 goto bail;
 670         }
 671         for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
 672                 if (!dd->ipath_pkeys[i]) {
 673                         any++;
 674                         continue;
 675                 }
 676                 if (dd->ipath_pkeys[i] == key) {
 677                         atomic_t *pkrefs = &dd->ipath_pkeyrefs[i];
 678
 679                         if (atomic_inc_return(pkrefs) > 1) {
 680                                 pd->port_pkeys[pidx] = key;
 681                                 ipath_cdbg(VERBOSE, "p%u set key %x "
 682                                            "matches #%d, count now %d\n",
 683                                            pd->port_port, key, i,
 684                                            atomic_read(pkrefs));
 685                                 ret = 0;
 686                                 goto bail;
 687                         } else {
 688                                 /*
 689                                  * lost race, decrement count, catch below
 690                                  */
 691                                 atomic_dec(pkrefs);
 692                                 ipath_cdbg(VERBOSE, "Lost race, count was "
 693                                            "0, after dec, it's %d\n",
 694                                            atomic_read(pkrefs));
 695                                 any++;
 696                         }
 697                 }
 698                 if ((dd->ipath_pkeys[i] & 0x7FFF) == lkey) {
 699                         /*
 700                          * It makes no sense to have both the limited and
 701                          * full membership PKEY set at the same time since
 702                          * the unlimited one will disable the limited one.
 703                          */
 704                         ret = -EEXIST;
 705                         goto bail;
 706                 }
 707         }
 708         if (!any) {
 709                 ipath_dbg("port %u, all pkeys already in use, "
 710                           "can't set %x\n", pd->port_port, key);
 711                 ret = -EBUSY;
 712                 goto bail;
 713         }
 714         for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
 715                 if (!dd->ipath_pkeys[i] &&
 716                     atomic_inc_return(&dd->ipath_pkeyrefs[i]) == 1) {
 717                         u64 pkey;
 718
 719                         /* for ipathstats, etc. */
 720                         ipath_stats.sps_pkeys[i] = lkey;
 721                         pd->port_pkeys[pidx] = dd->ipath_pkeys[i] = key;
 722                         pkey =
 723                                 (u64) dd->ipath_pkeys[0] |
 724                                 ((u64) dd->ipath_pkeys[1] << 16) |
 725                                 ((u64) dd->ipath_pkeys[2] << 32) |
 726                                 ((u64) dd->ipath_pkeys[3] << 48);
 727                         ipath_cdbg(PROC, "p%u set key %x in #%d, "
 728                                    "portidx %d, new pkey reg %llx\n",
 729                                    pd->port_port, key, i, pidx,
 730                                    (unsigned long long) pkey);
 731                         ipath_write_kreg(
 732                                 dd, dd->ipath_kregs->kr_partitionkey, pkey);
 733
 734                         ret = 0;
 735                         goto bail;
 736                 }
 737         }
 738         ipath_dbg("port %u, all pkeys already in use 2nd pass, "
 739                   "can't set %x\n", pd->port_port, key);
 740         ret = -EBUSY;
 741
 742 bail:
 743         return ret;
 744 }
 745
 746 /**
 747  * ipath_manage_rcvq - manage a port's receive queue
 748  * @pd: the port
 749  * @subport: the subport
 750  * @start_stop: action to carry out
 751  *
 752  * start_stop == 0 disables receive on the port, for use in queue
 753  * overflow conditions.  start_stop==1 re-enables, to be used to
 754  * re-init the software copy of the head register
 755  */
 756 static int ipath_manage_rcvq(struct ipath_portdata *pd, unsigned subport,
 757                              int start_stop)
 758 {
 759         struct ipath_devdata *dd = pd->port_dd;
 760
 761         ipath_cdbg(PROC, "%sabling rcv for unit %u port %u:%u\n",
 762                    start_stop ? "en" : "dis", dd->ipath_unit,
 763                    pd->port_port, subport);
 764         if (subport)
 765                 goto bail;
 766         /* atomically clear receive enable port. */
 767         if (start_stop) {
 768                 /*
 769                  * On enable, force in-memory copy of the tail register to
 770                  * 0, so that protocol code doesn't have to worry about
 771                  * whether or not the chip has yet updated the in-memory
 772                  * copy or not on return from the system call. The chip
 773                  * always resets it's tail register back to 0 on a
 774                  * transition from disabled to enabled.  This could cause a
 775                  * problem if software was broken, and did the enable w/o
 776                  * the disable, but eventually the in-memory copy will be
 777                  * updated and correct itself, even in the face of software
 778                  * bugs.
 779                  */
 780                 if (pd->port_rcvhdrtail_kvaddr)
 781                         ipath_clear_rcvhdrtail(pd);
 782                 set_bit(dd->ipath_r_portenable_shift + pd->port_port,
 783                         &dd->ipath_rcvctrl);
 784         } else
 785                 clear_bit(dd->ipath_r_portenable_shift + pd->port_port,
 786                           &dd->ipath_rcvctrl);
 787         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
 788                          dd->ipath_rcvctrl);
 789         /* now be sure chip saw it before we return */
 790         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 791         if (start_stop) {
 792                 /*
 793                  * And try to be sure that tail reg update has happened too.
 794                  * This should in theory interlock with the RXE changes to
 795                  * the tail register.  Don't assign it to the tail register
 796                  * in memory copy, since we could overwrite an update by the
 797                  * chip if we did.
 798                  */
 799                 ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
 800         }
 801         /* always; new head should be equal to new tail; see above */
 802 bail:
 803         return 0;
 804 }
 805
 806 static void ipath_clean_part_key(struct ipath_portdata *pd,
 807                                  struct ipath_devdata *dd)
 808 {
 809         int i, j, pchanged = 0;
 810         u64 oldpkey;
 811
 812         /* for debugging only */
 813         oldpkey = (u64) dd->ipath_pkeys[0] |
 814                 ((u64) dd->ipath_pkeys[1] << 16) |
 815                 ((u64) dd->ipath_pkeys[2] << 32) |
 816                 ((u64) dd->ipath_pkeys[3] << 48);
 817
 818         for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
 819                 if (!pd->port_pkeys[i])
 820                         continue;
 821                 ipath_cdbg(VERBOSE, "look for key[%d] %hx in pkeys\n", i,
 822                            pd->port_pkeys[i]);
 823                 for (j = 0; j < ARRAY_SIZE(dd->ipath_pkeys); j++) {
 824                         /* check for match independent of the global bit */
 825                         if ((dd->ipath_pkeys[j] & 0x7fff) !=
 826                             (pd->port_pkeys[i] & 0x7fff))
 827                                 continue;
 828                         if (atomic_dec_and_test(&dd->ipath_pkeyrefs[j])) {
 829                                 ipath_cdbg(VERBOSE, "p%u clear key "
 830                                            "%x matches #%d\n",
 831                                            pd->port_port,
 832                                            pd->port_pkeys[i], j);
 833                                 ipath_stats.sps_pkeys[j] =
 834                                         dd->ipath_pkeys[j] = 0;
 835                                 pchanged++;
 836                         }
 837                         else ipath_cdbg(
 838                                 VERBOSE, "p%u key %x matches #%d, "
 839                                 "but ref still %d\n", pd->port_port,
 840                                 pd->port_pkeys[i], j,
 841                                 atomic_read(&dd->ipath_pkeyrefs[j]));
 842                         break;
 843                 }
 844                 pd->port_pkeys[i] = 0;
 845         }
 846         if (pchanged) {
 847                 u64 pkey = (u64) dd->ipath_pkeys[0] |
 848                         ((u64) dd->ipath_pkeys[1] << 16) |
 849                         ((u64) dd->ipath_pkeys[2] << 32) |
 850                         ((u64) dd->ipath_pkeys[3] << 48);
 851                 ipath_cdbg(VERBOSE, "p%u old pkey reg %llx, "
 852                            "new pkey reg %llx\n", pd->port_port,
 853                            (unsigned long long) oldpkey,
 854                            (unsigned long long) pkey);
 855                 ipath_write_kreg(dd, dd->ipath_kregs->kr_partitionkey,
 856                                  pkey);
 857         }
 858 }
 859
 860 /*
 861  * Initialize the port data with the receive buffer sizes
 862  * so this can be done while the master port is locked.
 863  * Otherwise, there is a race with a slave opening the port
 864  * and seeing these fields uninitialized.
 865  */
 866 static void init_user_egr_sizes(struct ipath_portdata *pd)
 867 {
 868         struct ipath_devdata *dd = pd->port_dd;
 869         unsigned egrperchunk, egrcnt, size;
 870
 871         /*
 872          * to avoid wasting a lot of memory, we allocate 32KB chunks of
 873          * physically contiguous memory, advance through it until used up
 874          * and then allocate more.  Of course, we need memory to store those
 875          * extra pointers, now.  Started out with 256KB, but under heavy
 876          * memory pressure (creating large files and then copying them over
 877          * NFS while doing lots of MPI jobs), we hit some allocation
 878          * failures, even though we can sleep...  (2.6.10) Still get
 879          * failures at 64K.  32K is the lowest we can go without wasting
 880          * additional memory.
 881          */
 882         size = 0x8000;
 883         egrperchunk = size / dd->ipath_rcvegrbufsize;
 884         egrcnt = dd->ipath_rcvegrcnt;
 885         pd->port_rcvegrbuf_chunks = (egrcnt + egrperchunk - 1) / egrperchunk;
 886         pd->port_rcvegrbufs_perchunk = egrperchunk;
 887         pd->port_rcvegrbuf_size = size;
 888 }
 889
 890 /**
 891  * ipath_create_user_egr - allocate eager TID buffers
 892  * @pd: the port to allocate TID buffers for
 893  *
 894  * This routine is now quite different for user and kernel, because
 895  * the kernel uses skb's, for the accelerated network performance
 896  * This is the user port version
 897  *
 898  * Allocate the eager TID buffers and program them into infinipath
 899  * They are no longer completely contiguous, we do multiple allocation
 900  * calls.
 901  */
 902 static int ipath_create_user_egr(struct ipath_portdata *pd)
 903 {
 904         struct ipath_devdata *dd = pd->port_dd;
 905         unsigned e, egrcnt, egrperchunk, chunk, egrsize, egroff;
 906         size_t size;
 907         int ret;
 908         gfp_t gfp_flags;
 909
 910         /*
 911          * GFP_USER, but without GFP_FS, so buffer cache can be
 912          * coalesced (we hope); otherwise, even at order 4,
 913          * heavy filesystem activity makes these fail, and we can
 914          * use compound pages.
 915          */
 916         gfp_flags = __GFP_WAIT | __GFP_IO | __GFP_COMP;
 917
 918         egrcnt = dd->ipath_rcvegrcnt;
 919         /* TID number offset for this port */
 920         egroff = (pd->port_port - 1) * egrcnt + dd->ipath_p0_rcvegrcnt;
 921         egrsize = dd->ipath_rcvegrbufsize;
 922         ipath_cdbg(VERBOSE, "Allocating %d egr buffers, at egrtid "
 923                    "offset %x, egrsize %u\n", egrcnt, egroff, egrsize);
 924
 925         chunk = pd->port_rcvegrbuf_chunks;
 926         egrperchunk = pd->port_rcvegrbufs_perchunk;
 927         size = pd->port_rcvegrbuf_size;
 928         pd->port_rcvegrbuf = kmalloc(chunk * sizeof(pd->port_rcvegrbuf[0]),
 929                                      GFP_KERNEL);
 930         if (!pd->port_rcvegrbuf) {
 931                 ret = -ENOMEM;
 932                 goto bail;
 933         }
 934         pd->port_rcvegrbuf_phys =
 935                 kmalloc(chunk * sizeof(pd->port_rcvegrbuf_phys[0]),
 936                         GFP_KERNEL);
 937         if (!pd->port_rcvegrbuf_phys) {
 938                 ret = -ENOMEM;
 939                 goto bail_rcvegrbuf;
 940         }
 941         for (e = 0; e < pd->port_rcvegrbuf_chunks; e++) {
 942
 943                 pd->port_rcvegrbuf[e] = dma_alloc_coherent(
 944                         &dd->pcidev->dev, size, &pd->port_rcvegrbuf_phys[e],
 945                         gfp_flags);
 946
 947                 if (!pd->port_rcvegrbuf[e]) {
 948                         ret = -ENOMEM;
 949                         goto bail_rcvegrbuf_phys;
 950                 }
 951         }
 952
 953         pd->port_rcvegr_phys = pd->port_rcvegrbuf_phys[0];
 954
 955         for (e = chunk = 0; chunk < pd->port_rcvegrbuf_chunks; chunk++) {
 956                 dma_addr_t pa = pd->port_rcvegrbuf_phys[chunk];
 957                 unsigned i;
 958
 959                 for (i = 0; e < egrcnt && i < egrperchunk; e++, i++) {
 960                         dd->ipath_f_put_tid(dd, e + egroff +
 961                                             (u64 __iomem *)
 962                                             ((char __iomem *)
 963                                              dd->ipath_kregbase +
 964                                              dd->ipath_rcvegrbase),
 965                                             RCVHQ_RCV_TYPE_EAGER, pa);
 966                         pa += egrsize;
 967                 }
 968                 cond_resched(); /* don't hog the cpu */
 969         }
 970
 971         ret = 0;
 972         goto bail;
 973
 974 bail_rcvegrbuf_phys:
 975         for (e = 0; e < pd->port_rcvegrbuf_chunks &&
 976                 pd->port_rcvegrbuf[e]; e++) {
 977                 dma_free_coherent(&dd->pcidev->dev, size,
 978                                   pd->port_rcvegrbuf[e],
 979                                   pd->port_rcvegrbuf_phys[e]);
 980
 981         }
 982         kfree(pd->port_rcvegrbuf_phys);
 983         pd->port_rcvegrbuf_phys = NULL;
 984 bail_rcvegrbuf:
 985         kfree(pd->port_rcvegrbuf);
 986         pd->port_rcvegrbuf = NULL;
 987 bail:
 988         return ret;
 989 }
 990
 991
 992 /* common code for the mappings on dma_alloc_coherent mem */
 993 static int ipath_mmap_mem(struct vm_area_struct *vma,
 994         struct ipath_portdata *pd, unsigned len, int write_ok,
 995         void *kvaddr, char *what)
 996 {
 997         struct ipath_devdata *dd = pd->port_dd;
 998         unsigned long pfn;
 999         int ret;
1000
1001         if ((vma->vm_end - vma->vm_start) > len) {
1002                 dev_info(&dd->pcidev->dev,
1003                          "FAIL on %s: len %lx > %x\n", what,
1004                          vma->vm_end - vma->vm_start, len);
1005                 ret = -EFAULT;
1006                 goto bail;
1007         }
1008
1009         if (!write_ok) {
1010                 if (vma->vm_flags & VM_WRITE) {
1011                         dev_info(&dd->pcidev->dev,
1012                                  "%s must be mapped readonly\n", what);
1013                         ret = -EPERM;
1014                         goto bail;
1015                 }
1016
1017                 /* don't allow them to later change with mprotect */
1018                 vma->vm_flags &= ~VM_MAYWRITE;
1019         }
1020
1021         pfn = virt_to_phys(kvaddr) >> PAGE_SHIFT;
1022         ret = remap_pfn_range(vma, vma->vm_start, pfn,
1023                               len, vma->vm_page_prot);
1024         if (ret)
1025                 dev_info(&dd->pcidev->dev, "%s port%u mmap of %lx, %x "
1026                          "bytes r%c failed: %d\n", what, pd->port_port,
1027                          pfn, len, write_ok?'w':'o', ret);
1028         else
1029                 ipath_cdbg(VERBOSE, "%s port%u mmaped %lx, %x bytes "
1030                            "r%c\n", what, pd->port_port, pfn, len,
1031                            write_ok?'w':'o');
1032 bail:
1033         return ret;
1034 }
1035
1036 static int mmap_ureg(struct vm_area_struct *vma, struct ipath_devdata *dd,
1037                      u64 ureg)
1038 {
1039         unsigned long phys;
1040         int ret;
1041
1042         /*
1043          * This is real hardware, so use io_remap.  This is the mechanism
1044          * for the user process to update the head registers for their port
1045          * in the chip.
1046          */
1047         if ((vma->vm_end - vma->vm_start) > PAGE_SIZE) {
1048                 dev_info(&dd->pcidev->dev, "FAIL mmap userreg: reqlen "
1049                          "%lx > PAGE\n", vma->vm_end - vma->vm_start);
1050                 ret = -EFAULT;
1051         } else {
1052                 phys = dd->ipath_physaddr + ureg;
1053                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1054
1055                 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1056                 ret = io_remap_pfn_range(vma, vma->vm_start,
1057                                          phys >> PAGE_SHIFT,
1058                                          vma->vm_end - vma->vm_start,
1059                                          vma->vm_page_prot);
1060         }
1061         return ret;
1062 }
1063
1064 static int mmap_piobufs(struct vm_area_struct *vma,
1065                         struct ipath_devdata *dd,
1066                         struct ipath_portdata *pd,
1067                         unsigned piobufs, unsigned piocnt)
1068 {
1069         unsigned long phys;
1070         int ret;
1071
1072         /*
1073          * When we map the PIO buffers in the chip, we want to map them as
1074          * writeonly, no read possible.   This prevents access to previous
1075          * process data, and catches users who might try to read the i/o
1076          * space due to a bug.
1077          */
1078         if ((vma->vm_end - vma->vm_start) > (piocnt * dd->ipath_palign)) {
1079                 dev_info(&dd->pcidev->dev, "FAIL mmap piobufs: "
1080                          "reqlen %lx > PAGE\n",
1081                          vma->vm_end - vma->vm_start);
1082                 ret = -EINVAL;
1083                 goto bail;
1084         }
1085
1086         phys = dd->ipath_physaddr + piobufs;
1087
1088 #if defined(__powerpc__)
1089         /* There isn't a generic way to specify writethrough mappings */
1090         pgprot_val(vma->vm_page_prot) |= _PAGE_NO_CACHE;
1091         pgprot_val(vma->vm_page_prot) |= _PAGE_WRITETHRU;
1092         pgprot_val(vma->vm_page_prot) &= ~_PAGE_GUARDED;
1093 #endif
1094
1095         /*
1096          * don't allow them to later change to readable with mprotect (for when
1097          * not initially mapped readable, as is normally the case)
1098          */
1099         vma->vm_flags &= ~VM_MAYREAD;
1100         vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1101
1102         ret = io_remap_pfn_range(vma, vma->vm_start, phys >> PAGE_SHIFT,
1103                                  vma->vm_end - vma->vm_start,
1104                                  vma->vm_page_prot);
1105 bail:
1106         return ret;
1107 }
1108
1109 static int mmap_rcvegrbufs(struct vm_area_struct *vma,
1110                            struct ipath_portdata *pd)
1111 {
1112         struct ipath_devdata *dd = pd->port_dd;
1113         unsigned long start, size;
1114         size_t total_size, i;
1115         unsigned long pfn;
1116         int ret;
1117
1118         size = pd->port_rcvegrbuf_size;
1119         total_size = pd->port_rcvegrbuf_chunks * size;
1120         if ((vma->vm_end - vma->vm_start) > total_size) {
1121                 dev_info(&dd->pcidev->dev, "FAIL on egr bufs: "
1122                          "reqlen %lx > actual %lx\n",
1123                          vma->vm_end - vma->vm_start,
1124                          (unsigned long) total_size);
1125                 ret = -EINVAL;
1126                 goto bail;
1127         }
1128
1129         if (vma->vm_flags & VM_WRITE) {
1130                 dev_info(&dd->pcidev->dev, "Can't map eager buffers as "
1131                          "writable (flags=%lx)\n", vma->vm_flags);
1132                 ret = -EPERM;
1133                 goto bail;
1134         }
1135         /* don't allow them to later change to writeable with mprotect */
1136         vma->vm_flags &= ~VM_MAYWRITE;
1137
1138         start = vma->vm_start;
1139
1140         for (i = 0; i < pd->port_rcvegrbuf_chunks; i++, start += size) {
1141                 pfn = virt_to_phys(pd->port_rcvegrbuf[i]) >> PAGE_SHIFT;
1142                 ret = remap_pfn_range(vma, start, pfn, size,
1143                                       vma->vm_page_prot);
1144                 if (ret < 0)
1145                         goto bail;
1146         }
1147         ret = 0;
1148
1149 bail:
1150         return ret;
1151 }
1152
1153 /*
1154  * ipath_file_vma_fault - handle a VMA page fault.
1155  */
1156 static int ipath_file_vma_fault(struct vm_area_struct *vma,
1157                                         struct vm_fault *vmf)
1158 {
1159         struct page *page;
1160
1161         page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
1162         if (!page)
1163                 return VM_FAULT_SIGBUS;
1164         get_page(page);
1165         vmf->page = page;
1166
1167         return 0;
1168 }
1169
1170 static struct vm_operations_struct ipath_file_vm_ops = {
1171         .fault = ipath_file_vma_fault,
1172 };
1173
1174 static int mmap_kvaddr(struct vm_area_struct *vma, u64 pgaddr,
1175                        struct ipath_portdata *pd, unsigned subport)
1176 {
1177         unsigned long len;
1178         struct ipath_devdata *dd;
1179         void *addr;
1180         size_t size;
1181         int ret = 0;
1182
1183         /* If the port is not shared, all addresses should be physical */
1184         if (!pd->port_subport_cnt)
1185                 goto bail;
1186
1187         dd = pd->port_dd;
1188         size = pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
1189
1190         /*
1191          * Each process has all the subport uregbase, rcvhdrq, and
1192          * rcvegrbufs mmapped - as an array for all the processes,
1193          * and also separately for this process.
1194          */
1195         if (pgaddr == cvt_kvaddr(pd->subport_uregbase)) {
1196                 addr = pd->subport_uregbase;
1197                 size = PAGE_SIZE * pd->port_subport_cnt;
1198         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvhdr_base)) {
1199                 addr = pd->subport_rcvhdr_base;
1200                 size = pd->port_rcvhdrq_size * pd->port_subport_cnt;
1201         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvegrbuf)) {
1202                 addr = pd->subport_rcvegrbuf;
1203                 size *= pd->port_subport_cnt;
1204         } else if (pgaddr == cvt_kvaddr(pd->subport_uregbase +
1205                                         PAGE_SIZE * subport)) {
1206                 addr = pd->subport_uregbase + PAGE_SIZE * subport;
1207                 size = PAGE_SIZE;
1208         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvhdr_base +
1209                                 pd->port_rcvhdrq_size * subport)) {
1210                 addr = pd->subport_rcvhdr_base +
1211                         pd->port_rcvhdrq_size * subport;
1212                 size = pd->port_rcvhdrq_size;
1213         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvegrbuf +
1214                                size * subport)) {
1215                 addr = pd->subport_rcvegrbuf + size * subport;
1216                 /* rcvegrbufs are read-only on the slave */
1217                 if (vma->vm_flags & VM_WRITE) {
1218                         dev_info(&dd->pcidev->dev,
1219                                  "Can't map eager buffers as "
1220                                  "writable (flags=%lx)\n", vma->vm_flags);
1221                         ret = -EPERM;
1222                         goto bail;
1223                 }
1224                 /*
1225                  * Don't allow permission to later change to writeable
1226                  * with mprotect.
1227                  */
1228                 vma->vm_flags &= ~VM_MAYWRITE;
1229         } else {
1230                 goto bail;
1231         }
1232         len = vma->vm_end - vma->vm_start;
1233         if (len > size) {
1234                 ipath_cdbg(MM, "FAIL: reqlen %lx > %zx\n", len, size);
1235                 ret = -EINVAL;
1236                 goto bail;
1237         }
1238
1239         vma->vm_pgoff = (unsigned long) addr >> PAGE_SHIFT;
1240         vma->vm_ops = &ipath_file_vm_ops;
1241         vma->vm_flags |= VM_RESERVED | VM_DONTEXPAND;
1242         ret = 1;
1243
1244 bail:
1245         return ret;
1246 }
1247
1248 /**
1249  * ipath_mmap - mmap various structures into user space
1250  * @fp: the file pointer
1251  * @vma: the VM area
1252  *
1253  * We use this to have a shared buffer between the kernel and the user code
1254  * for the rcvhdr queue, egr buffers, and the per-port user regs and pio
1255  * buffers in the chip.  We have the open and close entries so we can bump
1256  * the ref count and keep the driver from being unloaded while still mapped.
1257  */
1258 static int ipath_mmap(struct file *fp, struct vm_area_struct *vma)
1259 {
1260         struct ipath_portdata *pd;
1261         struct ipath_devdata *dd;
1262         u64 pgaddr, ureg;
1263         unsigned piobufs, piocnt;
1264         int ret;
1265
1266         pd = port_fp(fp);
1267         if (!pd) {
1268                 ret = -EINVAL;
1269                 goto bail;
1270         }
1271         dd = pd->port_dd;
1272
1273         /*
1274          * This is the ipath_do_user_init() code, mapping the shared buffers
1275          * into the user process. The address referred to by vm_pgoff is the
1276          * file offset passed via mmap().  For shared ports, this is the
1277          * kernel vmalloc() address of the pages to share with the master.
1278          * For non-shared or master ports, this is a physical address.
1279          * We only do one mmap for each space mapped.
1280          */
1281         pgaddr = vma->vm_pgoff << PAGE_SHIFT;
1282
1283         /*
1284          * Check for 0 in case one of the allocations failed, but user
1285          * called mmap anyway.
1286          */
1287         if (!pgaddr)  {
1288                 ret = -EINVAL;
1289                 goto bail;
1290         }
1291
1292         ipath_cdbg(MM, "pgaddr %llx vm_start=%lx len %lx port %u:%u:%u\n",
1293                    (unsigned long long) pgaddr, vma->vm_start,
1294                    vma->vm_end - vma->vm_start, dd->ipath_unit,
1295                    pd->port_port, subport_fp(fp));
1296
1297         /*
1298          * Physical addresses must fit in 40 bits for our hardware.
1299          * Check for kernel virtual addresses first, anything else must
1300          * match a HW or memory address.
1301          */
1302         ret = mmap_kvaddr(vma, pgaddr, pd, subport_fp(fp));
1303         if (ret) {
1304                 if (ret > 0)
1305                         ret = 0;
1306                 goto bail;
1307         }
1308
1309         ureg = dd->ipath_uregbase + dd->ipath_ureg_align * pd->port_port;
1310         if (!pd->port_subport_cnt) {
1311                 /* port is not shared */
1312                 piocnt = dd->ipath_pbufsport;
1313                 piobufs = pd->port_piobufs;
1314         } else if (!subport_fp(fp)) {
1315                 /* caller is the master */
1316                 piocnt = (dd->ipath_pbufsport / pd->port_subport_cnt) +
1317                          (dd->ipath_pbufsport % pd->port_subport_cnt);
1318                 piobufs = pd->port_piobufs +
1319                         dd->ipath_palign * (dd->ipath_pbufsport - piocnt);
1320         } else {
1321                 unsigned slave = subport_fp(fp) - 1;
1322
1323                 /* caller is a slave */
1324                 piocnt = dd->ipath_pbufsport / pd->port_subport_cnt;
1325                 piobufs = pd->port_piobufs + dd->ipath_palign * piocnt * slave;
1326         }
1327
1328         if (pgaddr == ureg)
1329                 ret = mmap_ureg(vma, dd, ureg);
1330         else if (pgaddr == piobufs)
1331                 ret = mmap_piobufs(vma, dd, pd, piobufs, piocnt);
1332         else if (pgaddr == dd->ipath_pioavailregs_phys)
1333                 /* in-memory copy of pioavail registers */
1334                 ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1335                                      (void *) dd->ipath_pioavailregs_dma,
1336                                      "pioavail registers");
1337         else if (pgaddr == pd->port_rcvegr_phys)
1338                 ret = mmap_rcvegrbufs(vma, pd);
1339         else if (pgaddr == (u64) pd->port_rcvhdrq_phys)
1340                 /*
1341                  * The rcvhdrq itself; readonly except on HT (so have
1342                  * to allow writable mapping), multiple pages, contiguous
1343                  * from an i/o perspective.
1344                  */
1345                 ret = ipath_mmap_mem(vma, pd, pd->port_rcvhdrq_size, 1,
1346                                      pd->port_rcvhdrq,
1347                                      "rcvhdrq");
1348         else if (pgaddr == (u64) pd->port_rcvhdrqtailaddr_phys)
1349                 /* in-memory copy of rcvhdrq tail register */
1350                 ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1351                                      pd->port_rcvhdrtail_kvaddr,
1352                                      "rcvhdrq tail");
1353         else
1354                 ret = -EINVAL;
1355
1356         vma->vm_private_data = NULL;
1357
1358         if (ret < 0)
1359                 dev_info(&dd->pcidev->dev,
1360                          "Failure %d on off %llx len %lx\n",
1361                          -ret, (unsigned long long)pgaddr,
1362                          vma->vm_end - vma->vm_start);
1363 bail:
1364         return ret;
1365 }
1366
1367 static unsigned ipath_poll_hdrqfull(struct ipath_portdata *pd)
1368 {
1369         unsigned pollflag = 0;
1370
1371         if ((pd->poll_type & IPATH_POLL_TYPE_OVERFLOW) &&
1372             pd->port_hdrqfull != pd->port_hdrqfull_poll) {
1373                 pollflag |= POLLIN | POLLRDNORM;
1374                 pd->port_hdrqfull_poll = pd->port_hdrqfull;
1375         }
1376
1377         return pollflag;
1378 }
1379
1380 static unsigned int ipath_poll_urgent(struct ipath_portdata *pd,
1381                                       struct file *fp,
1382                                       struct poll_table_struct *pt)
1383 {
1384         unsigned pollflag = 0;
1385         struct ipath_devdata *dd;
1386
1387         dd = pd->port_dd;
1388
1389         /* variable access in ipath_poll_hdrqfull() needs this */
1390         rmb();
1391         pollflag = ipath_poll_hdrqfull(pd);
1392
1393         if (pd->port_urgent != pd->port_urgent_poll) {
1394                 pollflag |= POLLIN | POLLRDNORM;
1395                 pd->port_urgent_poll = pd->port_urgent;
1396         }
1397
1398         if (!pollflag) {
1399                 /* this saves a spin_lock/unlock in interrupt handler... */
1400                 set_bit(IPATH_PORT_WAITING_URG, &pd->port_flag);
1401                 /* flush waiting flag so don't miss an event... */
1402                 wmb();
1403                 poll_wait(fp, &pd->port_wait, pt);
1404         }
1405
1406         return pollflag;
1407 }
1408
1409 static unsigned int ipath_poll_next(struct ipath_portdata *pd,
1410                                     struct file *fp,
1411                                     struct poll_table_struct *pt)
1412 {
1413         u32 head;
1414         u32 tail;
1415         unsigned pollflag = 0;
1416         struct ipath_devdata *dd;
1417
1418         dd = pd->port_dd;
1419
1420         /* variable access in ipath_poll_hdrqfull() needs this */
1421         rmb();
1422         pollflag = ipath_poll_hdrqfull(pd);
1423
1424         head = ipath_read_ureg32(dd, ur_rcvhdrhead, pd->port_port);
1425         if (pd->port_rcvhdrtail_kvaddr)
1426                 tail = ipath_get_rcvhdrtail(pd);
1427         else
1428                 tail = ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
1429
1430         if (head != tail)
1431                 pollflag |= POLLIN | POLLRDNORM;
1432         else {
1433                 /* this saves a spin_lock/unlock in interrupt handler */
1434                 set_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag);
1435                 /* flush waiting flag so we don't miss an event */
1436                 wmb();
1437
1438                 set_bit(pd->port_port + dd->ipath_r_intravail_shift,
1439                         &dd->ipath_rcvctrl);
1440
1441                 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1442                                  dd->ipath_rcvctrl);
1443
1444                 if (dd->ipath_rhdrhead_intr_off) /* arm rcv interrupt */
1445                         ipath_write_ureg(dd, ur_rcvhdrhead,
1446                                          dd->ipath_rhdrhead_intr_off | head,
1447                                          pd->port_port);
1448
1449                 poll_wait(fp, &pd->port_wait, pt);
1450         }
1451
1452         return pollflag;
1453 }
1454
1455 static unsigned int ipath_poll(struct file *fp,
1456                                struct poll_table_struct *pt)
1457 {
1458         struct ipath_portdata *pd;
1459         unsigned pollflag;
1460
1461         pd = port_fp(fp);
1462         if (!pd)
1463                 pollflag = 0;
1464         else if (pd->poll_type & IPATH_POLL_TYPE_URGENT)
1465                 pollflag = ipath_poll_urgent(pd, fp, pt);
1466         else
1467                 pollflag = ipath_poll_next(pd, fp, pt);
1468
1469         return pollflag;
1470 }
1471
1472 static int ipath_supports_subports(int user_swmajor, int user_swminor)
1473 {
1474         /* no subport implementation prior to software version 1.3 */
1475         return (user_swmajor > 1) || (user_swminor >= 3);
1476 }
1477
1478 static int ipath_compatible_subports(int user_swmajor, int user_swminor)
1479 {
1480         /* this code is written long-hand for clarity */
1481         if (IPATH_USER_SWMAJOR != user_swmajor) {
1482                 /* no promise of compatibility if major mismatch */
1483                 return 0;
1484         }
1485         if (IPATH_USER_SWMAJOR == 1) {
1486                 switch (IPATH_USER_SWMINOR) {
1487                 case 0:
1488                 case 1:
1489                 case 2:
1490                         /* no subport implementation so cannot be compatible */
1491                         return 0;
1492                 case 3:
1493                         /* 3 is only compatible with itself */
1494                         return user_swminor == 3;
1495                 default:
1496                         /* >= 4 are compatible (or are expected to be) */
1497                         return user_swminor >= 4;
1498                 }
1499         }
1500         /* make no promises yet for future major versions */
1501         return 0;
1502 }
1503
1504 static int init_subports(struct ipath_devdata *dd,
1505                          struct ipath_portdata *pd,
1506                          const struct ipath_user_info *uinfo)
1507 {
1508         int ret = 0;
1509         unsigned num_subports;
1510         size_t size;
1511
1512         /*
1513          * If the user is requesting zero subports,
1514          * skip the subport allocation.
1515          */
1516         if (uinfo->spu_subport_cnt <= 0)
1517                 goto bail;
1518
1519         /* Self-consistency check for ipath_compatible_subports() */
1520         if (ipath_supports_subports(IPATH_USER_SWMAJOR, IPATH_USER_SWMINOR) &&
1521             !ipath_compatible_subports(IPATH_USER_SWMAJOR,
1522                                        IPATH_USER_SWMINOR)) {
1523                 dev_info(&dd->pcidev->dev,
1524                          "Inconsistent ipath_compatible_subports()\n");
1525                 goto bail;
1526         }
1527
1528         /* Check for subport compatibility */
1529         if (!ipath_compatible_subports(uinfo->spu_userversion >> 16,
1530                                        uinfo->spu_userversion & 0xffff)) {
1531                 dev_info(&dd->pcidev->dev,
1532                          "Mismatched user version (%d.%d) and driver "
1533                          "version (%d.%d) while port sharing. Ensure "
1534                          "that driver and library are from the same "
1535                          "release.\n",
1536                          (int) (uinfo->spu_userversion >> 16),
1537                          (int) (uinfo->spu_userversion & 0xffff),
1538                          IPATH_USER_SWMAJOR,
1539                          IPATH_USER_SWMINOR);
1540                 goto bail;
1541         }
1542         if (uinfo->spu_subport_cnt > INFINIPATH_MAX_SUBPORT) {
1543                 ret = -EINVAL;
1544                 goto bail;
1545         }
1546
1547         num_subports = uinfo->spu_subport_cnt;
1548         pd->subport_uregbase = vmalloc(PAGE_SIZE * num_subports);
1549         if (!pd->subport_uregbase) {
1550                 ret = -ENOMEM;
1551                 goto bail;
1552         }
1553         /* Note: pd->port_rcvhdrq_size isn't initialized yet. */
1554         size = ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize *
1555                      sizeof(u32), PAGE_SIZE) * num_subports;
1556         pd->subport_rcvhdr_base = vmalloc(size);
1557         if (!pd->subport_rcvhdr_base) {
1558                 ret = -ENOMEM;
1559                 goto bail_ureg;
1560         }
1561
1562         pd->subport_rcvegrbuf = vmalloc(pd->port_rcvegrbuf_chunks *
1563                                         pd->port_rcvegrbuf_size *
1564                                         num_subports);
1565         if (!pd->subport_rcvegrbuf) {
1566                 ret = -ENOMEM;
1567                 goto bail_rhdr;
1568         }
1569
1570         pd->port_subport_cnt = uinfo->spu_subport_cnt;
1571         pd->port_subport_id = uinfo->spu_subport_id;
1572         pd->active_slaves = 1;
1573         set_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag);
1574         memset(pd->subport_uregbase, 0, PAGE_SIZE * num_subports);
1575         memset(pd->subport_rcvhdr_base, 0, size);
1576         memset(pd->subport_rcvegrbuf, 0, pd->port_rcvegrbuf_chunks *
1577                                          pd->port_rcvegrbuf_size *
1578                                          num_subports);
1579         goto bail;
1580
1581 bail_rhdr:
1582         vfree(pd->subport_rcvhdr_base);
1583 bail_ureg:
1584         vfree(pd->subport_uregbase);
1585         pd->subport_uregbase = NULL;
1586 bail:
1587         return ret;
1588 }
1589
1590 static int try_alloc_port(struct ipath_devdata *dd, int port,
1591                           struct file *fp,
1592                           const struct ipath_user_info *uinfo)
1593 {
1594         struct ipath_portdata *pd;
1595         int ret;
1596
1597         if (!(pd = dd->ipath_pd[port])) {
1598                 void *ptmp;
1599
1600                 pd = kzalloc(sizeof(struct ipath_portdata), GFP_KERNEL);
1601
1602                 /*
1603                  * Allocate memory for use in ipath_tid_update() just once
1604                  * at open, not per call.  Reduces cost of expected send
1605                  * setup.
1606                  */
1607                 ptmp = kmalloc(dd->ipath_rcvtidcnt * sizeof(u16) +
1608                                dd->ipath_rcvtidcnt * sizeof(struct page **),
1609                                GFP_KERNEL);
1610                 if (!pd || !ptmp) {
1611                         ipath_dev_err(dd, "Unable to allocate portdata "
1612                                       "memory, failing open\n");
1613                         ret = -ENOMEM;
1614                         kfree(pd);
1615                         kfree(ptmp);
1616                         goto bail;
1617                 }
1618                 dd->ipath_pd[port] = pd;
1619                 dd->ipath_pd[port]->port_port = port;
1620                 dd->ipath_pd[port]->port_dd = dd;
1621                 dd->ipath_pd[port]->port_tid_pg_list = ptmp;
1622                 init_waitqueue_head(&dd->ipath_pd[port]->port_wait);
1623         }
1624         if (!pd->port_cnt) {
1625                 pd->userversion = uinfo->spu_userversion;
1626                 init_user_egr_sizes(pd);
1627                 if ((ret = init_subports(dd, pd, uinfo)) != 0)
1628                         goto bail;
1629                 ipath_cdbg(PROC, "%s[%u] opened unit:port %u:%u\n",
1630                            current->comm, current->pid, dd->ipath_unit,
1631                            port);
1632                 pd->port_cnt = 1;
1633                 port_fp(fp) = pd;
1634                 pd->port_pid = current->pid;
1635                 strncpy(pd->port_comm, current->comm, sizeof(pd->port_comm));
1636                 ipath_chg_pioavailkernel(dd,
1637                         dd->ipath_pbufsport * (pd->port_port - 1),
1638                         dd->ipath_pbufsport, 0);
1639                 ipath_stats.sps_ports++;
1640                 ret = 0;
1641         } else
1642                 ret = -EBUSY;
1643
1644 bail:
1645         return ret;
1646 }
1647
1648 static inline int usable(struct ipath_devdata *dd)
1649 {
1650         return dd &&
1651                 (dd->ipath_flags & IPATH_PRESENT) &&
1652                 dd->ipath_kregbase &&
1653                 dd->ipath_lid &&
1654                 !(dd->ipath_flags & (IPATH_LINKDOWN | IPATH_DISABLED
1655                                      | IPATH_LINKUNK));
1656 }
1657
1658 static int find_free_port(int unit, struct file *fp,
1659                           const struct ipath_user_info *uinfo)
1660 {
1661         struct ipath_devdata *dd = ipath_lookup(unit);
1662         int ret, i;
1663
1664         if (!dd) {
1665                 ret = -ENODEV;
1666                 goto bail;
1667         }
1668
1669         if (!usable(dd)) {
1670                 ret = -ENETDOWN;
1671                 goto bail;
1672         }
1673
1674         for (i = 1; i < dd->ipath_cfgports; i++) {
1675                 ret = try_alloc_port(dd, i, fp, uinfo);
1676                 if (ret != -EBUSY)
1677                         goto bail;
1678         }
1679         ret = -EBUSY;
1680
1681 bail:
1682         return ret;
1683 }
1684
1685 static int find_best_unit(struct file *fp,
1686                           const struct ipath_user_info *uinfo)
1687 {
1688         int ret = 0, i, prefunit = -1, devmax;
1689         int maxofallports, npresent, nup;
1690         int ndev;
1691
1692         devmax = ipath_count_units(&npresent, &nup, &maxofallports);
1693
1694         /*
1695          * This code is present to allow a knowledgeable person to
1696          * specify the layout of processes to processors before opening
1697          * this driver, and then we'll assign the process to the "closest"
1698          * InfiniPath chip to that processor (we assume reasonable connectivity,
1699          * for now).  This code assumes that if affinity has been set
1700          * before this point, that at most one cpu is set; for now this
1701          * is reasonable.  I check for both cpus_empty() and cpus_full(),
1702          * in case some kernel variant sets none of the bits when no
1703          * affinity is set.  2.6.11 and 12 kernels have all present
1704          * cpus set.  Some day we'll have to fix it up further to handle
1705          * a cpu subset.  This algorithm fails for two HT chips connected
1706          * in tunnel fashion.  Eventually this needs real topology
1707          * information.  There may be some issues with dual core numbering
1708          * as well.  This needs more work prior to release.
1709          */
1710         if (!cpus_empty(current->cpus_allowed) &&
1711             !cpus_full(current->cpus_allowed)) {
1712                 int ncpus = num_online_cpus(), curcpu = -1, nset = 0;
1713                 for (i = 0; i < ncpus; i++)
1714                         if (cpu_isset(i, current->cpus_allowed)) {
1715                                 ipath_cdbg(PROC, "%s[%u] affinity set for "
1716                                            "cpu %d/%d\n", current->comm,
1717                                            current->pid, i, ncpus);
1718                                 curcpu = i;
1719                                 nset++;
1720                         }
1721                 if (curcpu != -1 && nset != ncpus) {
1722                         if (npresent) {
1723                                 prefunit = curcpu / (ncpus / npresent);
1724                                 ipath_cdbg(PROC,"%s[%u] %d chips, %d cpus, "
1725                                           "%d cpus/chip, select unit %d\n",
1726                                           current->comm, current->pid,
1727                                           npresent, ncpus, ncpus / npresent,
1728                                           prefunit);
1729                         }
1730                 }
1731         }
1732
1733         /*
1734          * user ports start at 1, kernel port is 0
1735          * For now, we do round-robin access across all chips
1736          */
1737
1738         if (prefunit != -1)
1739                 devmax = prefunit + 1;
1740 recheck:
1741         for (i = 1; i < maxofallports; i++) {
1742                 for (ndev = prefunit != -1 ? prefunit : 0; ndev < devmax;
1743                      ndev++) {
1744                         struct ipath_devdata *dd = ipath_lookup(ndev);
1745
1746                         if (!usable(dd))
1747                                 continue; /* can't use this unit */
1748                         if (i >= dd->ipath_cfgports)
1749                                 /*
1750                                  * Maxed out on users of this unit. Try
1751                                  * next.
1752                                  */
1753                                 continue;
1754                         ret = try_alloc_port(dd, i, fp, uinfo);
1755                         if (!ret)
1756                                 goto done;
1757                 }
1758         }
1759
1760         if (npresent) {
1761                 if (nup == 0) {
1762                         ret = -ENETDOWN;
1763                         ipath_dbg("No ports available (none initialized "
1764                                   "and ready)\n");
1765                 } else {
1766                         if (prefunit > 0) {
1767                                 /* if started above 0, retry from 0 */
1768                                 ipath_cdbg(PROC,
1769                                            "%s[%u] no ports on prefunit "
1770                                            "%d, clear and re-check\n",
1771                                            current->comm, current->pid,
1772                                            prefunit);
1773                                 devmax = ipath_count_units(NULL, NULL,
1774                                                            NULL);
1775                                 prefunit = -1;
1776                                 goto recheck;
1777                         }
1778                         ret = -EBUSY;
1779                         ipath_dbg("No ports available\n");
1780                 }
1781         } else {
1782                 ret = -ENXIO;
1783                 ipath_dbg("No boards found\n");
1784         }
1785
1786 done:
1787         return ret;
1788 }
1789
1790 static int find_shared_port(struct file *fp,
1791                             const struct ipath_user_info *uinfo)
1792 {
1793         int devmax, ndev, i;
1794         int ret = 0;
1795
1796         devmax = ipath_count_units(NULL, NULL, NULL);
1797
1798         for (ndev = 0; ndev < devmax; ndev++) {
1799                 struct ipath_devdata *dd = ipath_lookup(ndev);
1800
1801                 if (!usable(dd))
1802                         continue;
1803                 for (i = 1; i < dd->ipath_cfgports; i++) {
1804                         struct ipath_portdata *pd = dd->ipath_pd[i];
1805
1806                         /* Skip ports which are not yet open */
1807                         if (!pd || !pd->port_cnt)
1808                                 continue;
1809                         /* Skip port if it doesn't match the requested one */
1810                         if (pd->port_subport_id != uinfo->spu_subport_id)
1811                                 continue;
1812                         /* Verify the sharing process matches the master */
1813                         if (pd->port_subport_cnt != uinfo->spu_subport_cnt ||
1814                             pd->userversion != uinfo->spu_userversion ||
1815                             pd->port_cnt >= pd->port_subport_cnt) {
1816                                 ret = -EINVAL;
1817                                 goto done;
1818                         }
1819                         port_fp(fp) = pd;
1820                         subport_fp(fp) = pd->port_cnt++;
1821                         pd->port_subpid[subport_fp(fp)] = current->pid;
1822                         tidcursor_fp(fp) = 0;
1823                         pd->active_slaves |= 1 << subport_fp(fp);
1824                         ipath_cdbg(PROC,
1825                                    "%s[%u] %u sharing %s[%u] unit:port %u:%u\n",
1826                                    current->comm, current->pid,
1827                                    subport_fp(fp),
1828                                    pd->port_comm, pd->port_pid,
1829                                    dd->ipath_unit, pd->port_port);
1830                         ret = 1;
1831                         goto done;
1832                 }
1833         }
1834
1835 done:
1836         return ret;
1837 }
1838
1839 static int ipath_open(struct inode *in, struct file *fp)
1840 {
1841         /* The real work is performed later in ipath_assign_port() */
1842         fp->private_data = kzalloc(sizeof(struct ipath_filedata), GFP_KERNEL);
1843         return fp->private_data ? 0 : -ENOMEM;
1844 }
1845
1846 /* Get port early, so can set affinity prior to memory allocation */
1847 static int ipath_assign_port(struct file *fp,
1848                               const struct ipath_user_info *uinfo)
1849 {
1850         int ret;
1851         int i_minor;
1852         unsigned swmajor, swminor;
1853
1854         /* Check to be sure we haven't already initialized this file */
1855         if (port_fp(fp)) {
1856                 ret = -EINVAL;
1857                 goto done;
1858         }
1859
1860         /* for now, if major version is different, bail */
1861         swmajor = uinfo->spu_userversion >> 16;
1862         if (swmajor != IPATH_USER_SWMAJOR) {
1863                 ipath_dbg("User major version %d not same as driver "
1864                           "major %d\n", uinfo->spu_userversion >> 16,
1865                           IPATH_USER_SWMAJOR);
1866                 ret = -ENODEV;
1867                 goto done;
1868         }
1869
1870         swminor = uinfo->spu_userversion & 0xffff;
1871         if (swminor != IPATH_USER_SWMINOR)
1872                 ipath_dbg("User minor version %d not same as driver "
1873                           "minor %d\n", swminor, IPATH_USER_SWMINOR);
1874
1875         mutex_lock(&ipath_mutex);
1876
1877         if (ipath_compatible_subports(swmajor, swminor) &&
1878             uinfo->spu_subport_cnt &&
1879             (ret = find_shared_port(fp, uinfo))) {
1880                 if (ret > 0)
1881                         ret = 0;
1882                 goto done_chk_sdma;
1883         }
1884
1885         i_minor = iminor(fp->f_path.dentry->d_inode) - IPATH_USER_MINOR_BASE;
1886         ipath_cdbg(VERBOSE, "open on dev %lx (minor %d)\n",
1887                    (long)fp->f_path.dentry->d_inode->i_rdev, i_minor);
1888
1889         if (i_minor)
1890                 ret = find_free_port(i_minor - 1, fp, uinfo);
1891         else
1892                 ret = find_best_unit(fp, uinfo);
1893
1894 done_chk_sdma:
1895         if (!ret) {
1896                 struct ipath_filedata *fd = fp->private_data;
1897                 const struct ipath_portdata *pd = fd->pd;
1898                 const struct ipath_devdata *dd = pd->port_dd;
1899
1900                 fd->pq = ipath_user_sdma_queue_create(&dd->pcidev->dev,
1901                                                       dd->ipath_unit,
1902                                                       pd->port_port,
1903                                                       fd->subport);
1904
1905                 if (!fd->pq)
1906                         ret = -ENOMEM;
1907         }
1908
1909         mutex_unlock(&ipath_mutex);
1910
1911 done:
1912         return ret;
1913 }
1914
1915
1916 static int ipath_do_user_init(struct file *fp,
1917                               const struct ipath_user_info *uinfo)
1918 {
1919         int ret;
1920         struct ipath_portdata *pd = port_fp(fp);
1921         struct ipath_devdata *dd;
1922         u32 head32;
1923
1924         /* Subports don't need to initialize anything since master did it. */
1925         if (subport_fp(fp)) {
1926                 ret = wait_event_interruptible(pd->port_wait,
1927                         !test_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag));
1928                 goto done;
1929         }
1930
1931         dd = pd->port_dd;
1932
1933         if (uinfo->spu_rcvhdrsize) {
1934                 ret = ipath_setrcvhdrsize(dd, uinfo->spu_rcvhdrsize);
1935                 if (ret)
1936                         goto done;
1937         }
1938
1939         /* for now we do nothing with rcvhdrcnt: uinfo->spu_rcvhdrcnt */
1940
1941         /* for right now, kernel piobufs are at end, so port 1 is at 0 */
1942         pd->port_piobufs = dd->ipath_piobufbase +
1943                 dd->ipath_pbufsport * (pd->port_port - 1) * dd->ipath_palign;
1944         ipath_cdbg(VERBOSE, "Set base of piobufs for port %u to 0x%x\n",
1945                    pd->port_port, pd->port_piobufs);
1946
1947         /*
1948          * Now allocate the rcvhdr Q and eager TIDs; skip the TID
1949          * array for time being.  If pd->port_port > chip-supported,
1950          * we need to do extra stuff here to handle by handling overflow
1951          * through port 0, someday
1952          */
1953         ret = ipath_create_rcvhdrq(dd, pd);
1954         if (!ret)
1955                 ret = ipath_create_user_egr(pd);
1956         if (ret)
1957                 goto done;
1958
1959         /*
1960          * set the eager head register for this port to the current values
1961          * of the tail pointers, since we don't know if they were
1962          * updated on last use of the port.
1963          */
1964         head32 = ipath_read_ureg32(dd, ur_rcvegrindextail, pd->port_port);
1965         ipath_write_ureg(dd, ur_rcvegrindexhead, head32, pd->port_port);
1966         pd->port_lastrcvhdrqtail = -1;
1967         ipath_cdbg(VERBOSE, "Wrote port%d egrhead %x from tail regs\n",
1968                 pd->port_port, head32);
1969         pd->port_tidcursor = 0; /* start at beginning after open */
1970
1971         /* initialize poll variables... */
1972         pd->port_urgent = 0;
1973         pd->port_urgent_poll = 0;
1974         pd->port_hdrqfull_poll = pd->port_hdrqfull;
1975
1976         /*
1977          * Now enable the port for receive.
1978          * For chips that are set to DMA the tail register to memory
1979          * when they change (and when the update bit transitions from
1980          * 0 to 1.  So for those chips, we turn it off and then back on.
1981          * This will (very briefly) affect any other open ports, but the
1982          * duration is very short, and therefore isn't an issue.  We
1983          * explictly set the in-memory tail copy to 0 beforehand, so we
1984          * don't have to wait to be sure the DMA update has happened
1985          * (chip resets head/tail to 0 on transition to enable).
1986          */
1987         set_bit(dd->ipath_r_portenable_shift + pd->port_port,
1988                 &dd->ipath_rcvctrl);
1989         if (!(dd->ipath_flags & IPATH_NODMA_RTAIL)) {
1990                 if (pd->port_rcvhdrtail_kvaddr)
1991                         ipath_clear_rcvhdrtail(pd);
1992                 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1993                         dd->ipath_rcvctrl &
1994                         ~(1ULL << dd->ipath_r_tailupd_shift));
1995         }
1996         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1997                          dd->ipath_rcvctrl);
1998         /* Notify any waiting slaves */
1999         if (pd->port_subport_cnt) {
2000                 clear_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag);
2001                 wake_up(&pd->port_wait);
2002         }
2003 done:
2004         return ret;
2005 }
2006
2007 /**
2008  * unlock_exptid - unlock any expected TID entries port still had in use
2009  * @pd: port
2010  *
2011  * We don't actually update the chip here, because we do a bulk update
2012  * below, using ipath_f_clear_tids.
2013  */
2014 static void unlock_expected_tids(struct ipath_portdata *pd)
2015 {
2016         struct ipath_devdata *dd = pd->port_dd;
2017         int port_tidbase = pd->port_port * dd->ipath_rcvtidcnt;
2018         int i, cnt = 0, maxtid = port_tidbase + dd->ipath_rcvtidcnt;
2019
2020         ipath_cdbg(VERBOSE, "Port %u unlocking any locked expTID pages\n",
2021                    pd->port_port);
2022         for (i = port_tidbase; i < maxtid; i++) {
2023                 struct page *ps = dd->ipath_pageshadow[i];
2024
2025                 if (!ps)
2026                         continue;
2027
2028                 dd->ipath_pageshadow[i] = NULL;
2029                 pci_unmap_page(dd->pcidev, dd->ipath_physshadow[i],
2030                         PAGE_SIZE, PCI_DMA_FROMDEVICE);
2031                 ipath_release_user_pages_on_close(&ps, 1);
2032                 cnt++;
2033                 ipath_stats.sps_pageunlocks++;
2034         }
2035         if (cnt)
2036                 ipath_cdbg(VERBOSE, "Port %u locked %u expTID entries\n",
2037                            pd->port_port, cnt);
2038
2039         if (ipath_stats.sps_pagelocks || ipath_stats.sps_pageunlocks)
2040                 ipath_cdbg(VERBOSE, "%llu pages locked, %llu unlocked\n",
2041                            (unsigned long long) ipath_stats.sps_pagelocks,
2042                            (unsigned long long)
2043                            ipath_stats.sps_pageunlocks);
2044 }
2045
2046 static int ipath_close(struct inode *in, struct file *fp)
2047 {
2048         int ret = 0;
2049         struct ipath_filedata *fd;
2050         struct ipath_portdata *pd;
2051         struct ipath_devdata *dd;
2052         unsigned port;
2053
2054         ipath_cdbg(VERBOSE, "close on dev %lx, private data %p\n",
2055                    (long)in->i_rdev, fp->private_data);
2056
2057         mutex_lock(&ipath_mutex);
2058
2059         fd = (struct ipath_filedata *) fp->private_data;
2060         fp->private_data = NULL;
2061         pd = fd->pd;
2062         if (!pd) {
2063                 mutex_unlock(&ipath_mutex);
2064                 goto bail;
2065         }
2066
2067         dd = pd->port_dd;
2068
2069         /* drain user sdma queue */
2070         ipath_user_sdma_queue_drain(dd, fd->pq);
2071         ipath_user_sdma_queue_destroy(fd->pq);
2072
2073         if (--pd->port_cnt) {
2074                 /*
2075                  * XXX If the master closes the port before the slave(s),
2076                  * revoke the mmap for the eager receive queue so
2077                  * the slave(s) don't wait for receive data forever.
2078                  */
2079                 pd->active_slaves &= ~(1 << fd->subport);
2080                 pd->port_subpid[fd->subport] = 0;
2081                 mutex_unlock(&ipath_mutex);
2082                 goto bail;
2083         }
2084         port = pd->port_port;
2085
2086         if (pd->port_hdrqfull) {
2087                 ipath_cdbg(PROC, "%s[%u] had %u rcvhdrqfull errors "
2088                            "during run\n", pd->port_comm, pd->port_pid,
2089                            pd->port_hdrqfull);
2090                 pd->port_hdrqfull = 0;
2091         }
2092
2093         if (pd->port_rcvwait_to || pd->port_piowait_to
2094             || pd->port_rcvnowait || pd->port_pionowait) {
2095                 ipath_cdbg(VERBOSE, "port%u, %u rcv, %u pio wait timeo; "
2096                            "%u rcv %u, pio already\n",
2097                            pd->port_port, pd->port_rcvwait_to,
2098                            pd->port_piowait_to, pd->port_rcvnowait,
2099                            pd->port_pionowait);
2100                 pd->port_rcvwait_to = pd->port_piowait_to =
2101                         pd->port_rcvnowait = pd->port_pionowait = 0;
2102         }
2103         if (pd->port_flag) {
2104                 ipath_cdbg(PROC, "port %u port_flag set: 0x%lx\n",
2105                           pd->port_port, pd->port_flag);
2106                 pd->port_flag = 0;
2107         }
2108
2109         if (dd->ipath_kregbase) {
2110                 int i;
2111                 /* atomically clear receive enable port and intr avail. */
2112                 clear_bit(dd->ipath_r_portenable_shift + port,
2113                           &dd->ipath_rcvctrl);
2114                 clear_bit(pd->port_port + dd->ipath_r_intravail_shift,
2115                           &dd->ipath_rcvctrl);
2116                 ipath_write_kreg( dd, dd->ipath_kregs->kr_rcvctrl,
2117                         dd->ipath_rcvctrl);
2118                 /* and read back from chip to be sure that nothing
2119                  * else is in flight when we do the rest */
2120                 (void)ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
2121
2122                 /* clean up the pkeys for this port user */
2123                 ipath_clean_part_key(pd, dd);
2124                 /*
2125                  * be paranoid, and never write 0's to these, just use an
2126                  * unused part of the port 0 tail page.  Of course,
2127                  * rcvhdraddr points to a large chunk of memory, so this
2128                  * could still trash things, but at least it won't trash
2129                  * page 0, and by disabling the port, it should stop "soon",
2130                  * even if a packet or two is in already in flight after we
2131                  * disabled the port.
2132                  */
2133                 ipath_write_kreg_port(dd,
2134                         dd->ipath_kregs->kr_rcvhdrtailaddr, port,
2135                         dd->ipath_dummy_hdrq_phys);
2136                 ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdraddr,
2137                         pd->port_port, dd->ipath_dummy_hdrq_phys);
2138
2139                 i = dd->ipath_pbufsport * (port - 1);
2140                 ipath_disarm_piobufs(dd, i, dd->ipath_pbufsport);
2141                 ipath_chg_pioavailkernel(dd, i, dd->ipath_pbufsport, 1);
2142
2143                 dd->ipath_f_clear_tids(dd, pd->port_port);
2144
2145                 if (dd->ipath_pageshadow)
2146                         unlock_expected_tids(pd);
2147                 ipath_stats.sps_ports--;
2148                 ipath_cdbg(PROC, "%s[%u] closed port %u:%u\n",
2149                            pd->port_comm, pd->port_pid,
2150                            dd->ipath_unit, port);
2151         }
2152
2153         pd->port_pid = 0;
2154         dd->ipath_pd[pd->port_port] = NULL; /* before releasing mutex */
2155         mutex_unlock(&ipath_mutex);
2156         ipath_free_pddata(dd, pd); /* after releasing the mutex */
2157
2158 bail:
2159         kfree(fd);
2160         return ret;
2161 }
2162
2163 static int ipath_port_info(struct ipath_portdata *pd, u16 subport,
2164                            struct ipath_port_info __user *uinfo)
2165 {
2166         struct ipath_port_info info;
2167         int nup;
2168         int ret;
2169         size_t sz;
2170
2171         (void) ipath_count_units(NULL, &nup, NULL);
2172         info.num_active = nup;
2173         info.unit = pd->port_dd->ipath_unit;
2174         info.port = pd->port_port;
2175         info.subport = subport;
2176         /* Don't return new fields if old library opened the port. */
2177         if (ipath_supports_subports(pd->userversion >> 16,
2178                                     pd->userversion & 0xffff)) {
2179                 /* Number of user ports available for this device. */
2180                 info.num_ports = pd->port_dd->ipath_cfgports - 1;
2181                 info.num_subports = pd->port_subport_cnt;
2182                 sz = sizeof(info);
2183         } else
2184                 sz = sizeof(info) - 2 * sizeof(u16);
2185
2186         if (copy_to_user(uinfo, &info, sz)) {
2187                 ret = -EFAULT;
2188                 goto bail;
2189         }
2190         ret = 0;
2191
2192 bail:
2193         return ret;
2194 }
2195
2196 static int ipath_get_slave_info(struct ipath_portdata *pd,
2197                                 void __user *slave_mask_addr)
2198 {
2199         int ret = 0;
2200
2201         if (copy_to_user(slave_mask_addr, &pd->active_slaves, sizeof(u32)))
2202                 ret = -EFAULT;
2203         return ret;
2204 }
2205
2206 static int ipath_sdma_get_inflight(struct ipath_user_sdma_queue *pq,
2207                                    u32 __user *inflightp)
2208 {
2209         const u32 val = ipath_user_sdma_inflight_counter(pq);
2210
2211         if (put_user(val, inflightp))
2212                 return -EFAULT;
2213
2214         return 0;
2215 }
2216
2217 static int ipath_sdma_get_complete(struct ipath_devdata *dd,
2218                                    struct ipath_user_sdma_queue *pq,
2219                                    u32 __user *completep)
2220 {
2221         u32 val;
2222         int err;
2223
2224         err = ipath_user_sdma_make_progress(dd, pq);
2225         if (err < 0)
2226                 return err;
2227
2228         val = ipath_user_sdma_complete_counter(pq);
2229         if (put_user(val, completep))
2230                 return -EFAULT;
2231
2232         return 0;
2233 }
2234
2235 static ssize_t ipath_write(struct file *fp, const char __user *data,
2236                            size_t count, loff_t *off)
2237 {
2238         const struct ipath_cmd __user *ucmd;
2239         struct ipath_portdata *pd;
2240         const void __user *src;
2241         size_t consumed, copy;
2242         struct ipath_cmd cmd;
2243         ssize_t ret = 0;
2244         void *dest;
2245
2246         if (count < sizeof(cmd.type)) {
2247                 ret = -EINVAL;
2248                 goto bail;
2249         }
2250
2251         ucmd = (const struct ipath_cmd __user *) data;
2252
2253         if (copy_from_user(&cmd.type, &ucmd->type, sizeof(cmd.type))) {
2254                 ret = -EFAULT;
2255                 goto bail;
2256         }
2257
2258         consumed = sizeof(cmd.type);
2259
2260         switch (cmd.type) {
2261         case IPATH_CMD_ASSIGN_PORT:
2262         case __IPATH_CMD_USER_INIT:
2263         case IPATH_CMD_USER_INIT:
2264                 copy = sizeof(cmd.cmd.user_info);
2265                 dest = &cmd.cmd.user_info;
2266                 src = &ucmd->cmd.user_info;
2267                 break;
2268         case IPATH_CMD_RECV_CTRL:
2269                 copy = sizeof(cmd.cmd.recv_ctrl);
2270                 dest = &cmd.cmd.recv_ctrl;
2271                 src = &ucmd->cmd.recv_ctrl;
2272                 break;
2273         case IPATH_CMD_PORT_INFO:
2274                 copy = sizeof(cmd.cmd.port_info);
2275                 dest = &cmd.cmd.port_info;
2276                 src = &ucmd->cmd.port_info;
2277                 break;
2278         case IPATH_CMD_TID_UPDATE:
2279         case IPATH_CMD_TID_FREE:
2280                 copy = sizeof(cmd.cmd.tid_info);
2281                 dest = &cmd.cmd.tid_info;
2282                 src = &ucmd->cmd.tid_info;
2283                 break;
2284         case IPATH_CMD_SET_PART_KEY:
2285                 copy = sizeof(cmd.cmd.part_key);
2286                 dest = &cmd.cmd.part_key;
2287                 src = &ucmd->cmd.part_key;
2288                 break;
2289         case __IPATH_CMD_SLAVE_INFO:
2290                 copy = sizeof(cmd.cmd.slave_mask_addr);
2291                 dest = &cmd.cmd.slave_mask_addr;
2292                 src = &ucmd->cmd.slave_mask_addr;
2293                 break;
2294         case IPATH_CMD_PIOAVAILUPD:     // force an update of PIOAvail reg
2295                 copy = 0;
2296                 src = NULL;
2297                 dest = NULL;
2298                 break;
2299         case IPATH_CMD_POLL_TYPE:
2300                 copy = sizeof(cmd.cmd.poll_type);
2301                 dest = &cmd.cmd.poll_type;
2302                 src = &ucmd->cmd.poll_type;
2303                 break;
2304         case IPATH_CMD_ARMLAUNCH_CTRL:
2305                 copy = sizeof(cmd.cmd.armlaunch_ctrl);
2306                 dest = &cmd.cmd.armlaunch_ctrl;
2307                 src = &ucmd->cmd.armlaunch_ctrl;
2308                 break;
2309         case IPATH_CMD_SDMA_INFLIGHT:
2310                 copy = sizeof(cmd.cmd.sdma_inflight);
2311                 dest = &cmd.cmd.sdma_inflight;
2312                 src = &ucmd->cmd.sdma_inflight;
2313                 break;
2314         case IPATH_CMD_SDMA_COMPLETE:
2315                 copy = sizeof(cmd.cmd.sdma_complete);
2316                 dest = &cmd.cmd.sdma_complete;
2317                 src = &ucmd->cmd.sdma_complete;
2318                 break;
2319         default:
2320                 ret = -EINVAL;
2321                 goto bail;
2322         }
2323
2324         if (copy) {
2325                 if ((count - consumed) < copy) {
2326                         ret = -EINVAL;
2327                         goto bail;
2328                 }
2329
2330                 if (copy_from_user(dest, src, copy)) {
2331                         ret = -EFAULT;
2332                         goto bail;
2333                 }
2334
2335                 consumed += copy;
2336         }
2337
2338         pd = port_fp(fp);
2339         if (!pd && cmd.type != __IPATH_CMD_USER_INIT &&
2340                 cmd.type != IPATH_CMD_ASSIGN_PORT) {
2341                 ret = -EINVAL;
2342                 goto bail;
2343         }
2344
2345         switch (cmd.type) {
2346         case IPATH_CMD_ASSIGN_PORT:
2347                 ret = ipath_assign_port(fp, &cmd.cmd.user_info);
2348                 if (ret)
2349                         goto bail;
2350                 break;
2351         case __IPATH_CMD_USER_INIT:
2352                 /* backwards compatibility, get port first */
2353                 ret = ipath_assign_port(fp, &cmd.cmd.user_info);
2354                 if (ret)
2355                         goto bail;
2356                 /* and fall through to current version. */
2357         case IPATH_CMD_USER_INIT:
2358                 ret = ipath_do_user_init(fp, &cmd.cmd.user_info);
2359                 if (ret)
2360                         goto bail;
2361                 ret = ipath_get_base_info(
2362                         fp, (void __user *) (unsigned long)
2363                         cmd.cmd.user_info.spu_base_info,
2364                         cmd.cmd.user_info.spu_base_info_size);
2365                 break;
2366         case IPATH_CMD_RECV_CTRL:
2367                 ret = ipath_manage_rcvq(pd, subport_fp(fp), cmd.cmd.recv_ctrl);
2368                 break;
2369         case IPATH_CMD_PORT_INFO:
2370                 ret = ipath_port_info(pd, subport_fp(fp),
2371                                       (struct ipath_port_info __user *)
2372                                       (unsigned long) cmd.cmd.port_info);
2373                 break;
2374         case IPATH_CMD_TID_UPDATE:
2375                 ret = ipath_tid_update(pd, fp, &cmd.cmd.tid_info);
2376                 break;
2377         case IPATH_CMD_TID_FREE:
2378                 ret = ipath_tid_free(pd, subport_fp(fp), &cmd.cmd.tid_info);
2379                 break;
2380         case IPATH_CMD_SET_PART_KEY:
2381                 ret = ipath_set_part_key(pd, cmd.cmd.part_key);
2382                 break;
2383         case __IPATH_CMD_SLAVE_INFO:
2384                 ret = ipath_get_slave_info(pd,
2385                                            (void __user *) (unsigned long)
2386                                            cmd.cmd.slave_mask_addr);
2387                 break;
2388         case IPATH_CMD_PIOAVAILUPD:
2389                 ipath_force_pio_avail_update(pd->port_dd);
2390                 break;
2391         case IPATH_CMD_POLL_TYPE:
2392                 pd->poll_type = cmd.cmd.poll_type;
2393                 break;
2394         case IPATH_CMD_ARMLAUNCH_CTRL:
2395                 if (cmd.cmd.armlaunch_ctrl)
2396                         ipath_enable_armlaunch(pd->port_dd);
2397                 else
2398                         ipath_disable_armlaunch(pd->port_dd);
2399                 break;
2400         case IPATH_CMD_SDMA_INFLIGHT:
2401                 ret = ipath_sdma_get_inflight(user_sdma_queue_fp(fp),
2402                                               (u32 __user *) (unsigned long)
2403                                               cmd.cmd.sdma_inflight);
2404                 break;
2405         case IPATH_CMD_SDMA_COMPLETE:
2406                 ret = ipath_sdma_get_complete(pd->port_dd,
2407                                               user_sdma_queue_fp(fp),
2408                                               (u32 __user *) (unsigned long)
2409                                               cmd.cmd.sdma_complete);
2410                 break;
2411         }
2412
2413         if (ret >= 0)
2414                 ret = consumed;
2415
2416 bail:
2417         return ret;
2418 }
2419
2420 static ssize_t ipath_writev(struct kiocb *iocb, const struct iovec *iov,
2421                             unsigned long dim, loff_t off)
2422 {
2423         struct file *filp = iocb->ki_filp;
2424         struct ipath_filedata *fp = filp->private_data;
2425         struct ipath_portdata *pd = port_fp(filp);
2426         struct ipath_user_sdma_queue *pq = fp->pq;
2427
2428         if (!dim)
2429                 return -EINVAL;
2430
2431         return ipath_user_sdma_writev(pd->port_dd, pq, iov, dim);
2432 }
2433
2434 static struct class *ipath_class;
2435
2436 static int init_cdev(int minor, char *name, const struct file_operations *fops,
2437                      struct cdev **cdevp, struct device **devp)
2438 {
2439         const dev_t dev = MKDEV(IPATH_MAJOR, minor);
2440         struct cdev *cdev = NULL;
2441         struct device *device = NULL;
2442         int ret;
2443
2444         cdev = cdev_alloc();
2445         if (!cdev) {
2446                 printk(KERN_ERR IPATH_DRV_NAME
2447                        ": Could not allocate cdev for minor %d, %s\n",
2448                        minor, name);
2449                 ret = -ENOMEM;
2450                 goto done;
2451         }
2452
2453         cdev->owner = THIS_MODULE;
2454         cdev->ops = fops;
2455         kobject_set_name(&cdev->kobj, name);
2456
2457         ret = cdev_add(cdev, dev, 1);
2458         if (ret < 0) {
2459                 printk(KERN_ERR IPATH_DRV_NAME
2460                        ": Could not add cdev for minor %d, %s (err %d)\n",
2461                        minor, name, -ret);
2462                 goto err_cdev;
2463         }
2464
2465         device = device_create(ipath_class, NULL, dev, name);
2466
2467         if (IS_ERR(device)) {
2468                 ret = PTR_ERR(device);
2469                 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2470                        "device for minor %d, %s (err %d)\n",
2471                        minor, name, -ret);
2472                 goto err_cdev;
2473         }
2474
2475         goto done;
2476
2477 err_cdev:
2478         cdev_del(cdev);
2479         cdev = NULL;
2480
2481 done:
2482         if (ret >= 0) {
2483                 *cdevp = cdev;
2484                 *devp = device;
2485         } else {
2486                 *cdevp = NULL;
2487                 *devp = NULL;
2488         }
2489
2490         return ret;
2491 }
2492
2493 int ipath_cdev_init(int minor, char *name, const struct file_operations *fops,
2494                     struct cdev **cdevp, struct device **devp)
2495 {
2496         return init_cdev(minor, name, fops, cdevp, devp);
2497 }
2498
2499 static void cleanup_cdev(struct cdev **cdevp,
2500                          struct device **devp)
2501 {
2502         struct device *dev = *devp;
2503
2504         if (dev) {
2505                 device_unregister(dev);
2506                 *devp = NULL;
2507         }
2508
2509         if (*cdevp) {
2510                 cdev_del(*cdevp);
2511                 *cdevp = NULL;
2512         }
2513 }
2514
2515 void ipath_cdev_cleanup(struct cdev **cdevp,
2516                         struct device **devp)
2517 {
2518         cleanup_cdev(cdevp, devp);
2519 }
2520
2521 static struct cdev *wildcard_cdev;
2522 static struct device *wildcard_dev;
2523
2524 static const dev_t dev = MKDEV(IPATH_MAJOR, 0);
2525
2526 static int user_init(void)
2527 {
2528         int ret;
2529
2530         ret = register_chrdev_region(dev, IPATH_NMINORS, IPATH_DRV_NAME);
2531         if (ret < 0) {
2532                 printk(KERN_ERR IPATH_DRV_NAME ": Could not register "
2533                        "chrdev region (err %d)\n", -ret);
2534                 goto done;
2535         }
2536
2537         ipath_class = class_create(THIS_MODULE, IPATH_DRV_NAME);
2538
2539         if (IS_ERR(ipath_class)) {
2540                 ret = PTR_ERR(ipath_class);
2541                 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2542                        "device class (err %d)\n", -ret);
2543                 goto bail;
2544         }
2545
2546         goto done;
2547 bail:
2548         unregister_chrdev_region(dev, IPATH_NMINORS);
2549 done:
2550         return ret;
2551 }
2552
2553 static void user_cleanup(void)
2554 {
2555         if (ipath_class) {
2556                 class_destroy(ipath_class);
2557                 ipath_class = NULL;
2558         }
2559
2560         unregister_chrdev_region(dev, IPATH_NMINORS);
2561 }
2562
2563 static atomic_t user_count = ATOMIC_INIT(0);
2564 static atomic_t user_setup = ATOMIC_INIT(0);
2565
2566 int ipath_user_add(struct ipath_devdata *dd)
2567 {
2568         char name[10];
2569         int ret;
2570
2571         if (atomic_inc_return(&user_count) == 1) {
2572                 ret = user_init();
2573                 if (ret < 0) {
2574                         ipath_dev_err(dd, "Unable to set up user support: "
2575                                       "error %d\n", -ret);
2576                         goto bail;
2577                 }
2578                 ret = init_cdev(0, "ipath", &ipath_file_ops, &wildcard_cdev,
2579                                 &wildcard_dev);
2580                 if (ret < 0) {
2581                         ipath_dev_err(dd, "Could not create wildcard "
2582                                       "minor: error %d\n", -ret);
2583                         goto bail_user;
2584                 }
2585
2586                 atomic_set(&user_setup, 1);
2587         }
2588
2589         snprintf(name, sizeof(name), "ipath%d", dd->ipath_unit);
2590
2591         ret = init_cdev(dd->ipath_unit + 1, name, &ipath_file_ops,
2592                         &dd->user_cdev, &dd->user_dev);
2593         if (ret < 0)
2594                 ipath_dev_err(dd, "Could not create user minor %d, %s\n",
2595                               dd->ipath_unit + 1, name);
2596
2597         goto bail;
2598
2599 bail_user:
2600         user_cleanup();
2601 bail:
2602         return ret;
2603 }
2604
2605 void ipath_user_remove(struct ipath_devdata *dd)
2606 {
2607         cleanup_cdev(&dd->user_cdev, &dd->user_dev);
2608
2609         if (atomic_dec_return(&user_count) == 0) {
2610                 if (atomic_read(&user_setup) == 0)
2611                         goto bail;
2612
2613                 cleanup_cdev(&wildcard_cdev, &wildcard_dev);
2614                 user_cleanup();
2615
2616                 atomic_set(&user_setup, 0);
2617         }
2618 bail:
2619         return;
2620 }