drivers/infiniband/hw/ipath/ipath_file_ops.c

   1 /*
   2  * Copyright (c) 2006 QLogic, Inc. All rights reserved.
   3  * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
   4  *
   5  * This software is available to you under a choice of one of two
   6  * licenses.  You may choose to be licensed under the terms of the GNU
   7  * General Public License (GPL) Version 2, available from the file
   8  * COPYING in the main directory of this source tree, or the
   9  * OpenIB.org BSD license below:
  10  *
  11  *     Redistribution and use in source and binary forms, with or
  12  *     without modification, are permitted provided that the following
  13  *     conditions are met:
  14  *
  15  *      - Redistributions of source code must retain the above
  16  *        copyright notice, this list of conditions and the following
  17  *        disclaimer.
  18  *
  19  *      - Redistributions in binary form must reproduce the above
  20  *        copyright notice, this list of conditions and the following
  21  *        disclaimer in the documentation and/or other materials
  22  *        provided with the distribution.
  23  *
  24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  31  * SOFTWARE.
  32  */
  33
  34 #include <linux/pci.h>
  35 #include <linux/poll.h>
  36 #include <linux/cdev.h>
  37 #include <linux/swap.h>
  38 #include <linux/vmalloc.h>
  39 #include <asm/pgtable.h>
  40
  41 #include "ipath_kernel.h"
  42 #include "ipath_common.h"
  43
  44 /*
  45  * mmap64 doesn't allow all 64 bits for 32-bit applications
  46  * so only use the low 43 bits.
  47  */
  48 #define MMAP64_MASK     0x7FFFFFFFFFFUL
  49
  50 static int ipath_open(struct inode *, struct file *);
  51 static int ipath_close(struct inode *, struct file *);
  52 static ssize_t ipath_write(struct file *, const char __user *, size_t,
  53                            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         .open = ipath_open,
  61         .release = ipath_close,
  62         .poll = ipath_poll,
  63         .mmap = ipath_mmap
  64 };
  65
  66 static int ipath_get_base_info(struct file *fp,
  67                                void __user *ubase, size_t ubase_size)
  68 {
  69         struct ipath_portdata *pd = port_fp(fp);
  70         int ret = 0;
  71         struct ipath_base_info *kinfo = NULL;
  72         struct ipath_devdata *dd = pd->port_dd;
  73         unsigned subport_cnt;
  74         int shared, master;
  75         size_t sz;
  76
  77         subport_cnt = pd->port_subport_cnt;
  78         if (!subport_cnt) {
  79                 shared = 0;
  80                 master = 0;
  81                 subport_cnt = 1;
  82         } else {
  83                 shared = 1;
  84                 master = !subport_fp(fp);
  85         }
  86
  87         sz = sizeof(*kinfo);
  88         /* If port sharing is not requested, allow the old size structure */
  89         if (!shared)
  90                 sz -= 3 * sizeof(u64);
  91         if (ubase_size < sz) {
  92                 ipath_cdbg(PROC,
  93                            "Base size %zu, need %zu (version mismatch?)\n",
  94                            ubase_size, sz);
  95                 ret = -EINVAL;
  96                 goto bail;
  97         }
  98
  99         kinfo = kzalloc(sizeof(*kinfo), GFP_KERNEL);
 100         if (kinfo == NULL) {
 101                 ret = -ENOMEM;
 102                 goto bail;
 103         }
 104
 105         ret = dd->ipath_f_get_base_info(pd, kinfo);
 106         if (ret < 0)
 107                 goto bail;
 108
 109         kinfo->spi_rcvhdr_cnt = dd->ipath_rcvhdrcnt;
 110         kinfo->spi_rcvhdrent_size = dd->ipath_rcvhdrentsize;
 111         kinfo->spi_tidegrcnt = dd->ipath_rcvegrcnt;
 112         kinfo->spi_rcv_egrbufsize = dd->ipath_rcvegrbufsize;
 113         /*
 114          * have to mmap whole thing
 115          */
 116         kinfo->spi_rcv_egrbuftotlen =
 117                 pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
 118         kinfo->spi_rcv_egrperchunk = pd->port_rcvegrbufs_perchunk;
 119         kinfo->spi_rcv_egrchunksize = kinfo->spi_rcv_egrbuftotlen /
 120                 pd->port_rcvegrbuf_chunks;
 121         kinfo->spi_tidcnt = dd->ipath_rcvtidcnt / subport_cnt;
 122         if (master)
 123                 kinfo->spi_tidcnt += dd->ipath_rcvtidcnt % subport_cnt;
 124         /*
 125          * for this use, may be ipath_cfgports summed over all chips that
 126          * are are configured and present
 127          */
 128         kinfo->spi_nports = dd->ipath_cfgports;
 129         /* unit (chip/board) our port is on */
 130         kinfo->spi_unit = dd->ipath_unit;
 131         /* for now, only a single page */
 132         kinfo->spi_tid_maxsize = PAGE_SIZE;
 133
 134         /*
 135          * Doing this per port, and based on the skip value, etc.  This has
 136          * to be the actual buffer size, since the protocol code treats it
 137          * as an array.
 138          *
 139          * These have to be set to user addresses in the user code via mmap.
 140          * These values are used on return to user code for the mmap target
 141          * addresses only.  For 32 bit, same 44 bit address problem, so use
 142          * the physical address, not virtual.  Before 2.6.11, using the
 143          * page_address() macro worked, but in 2.6.11, even that returns the
 144          * full 64 bit address (upper bits all 1's).  So far, using the
 145          * physical addresses (or chip offsets, for chip mapping) works, but
 146          * no doubt some future kernel release will change that, and we'll be
 147          * on to yet another method of dealing with this.
 148          */
 149         kinfo->spi_rcvhdr_base = (u64) pd->port_rcvhdrq_phys;
 150         kinfo->spi_rcvhdr_tailaddr = (u64) pd->port_rcvhdrqtailaddr_phys;
 151         kinfo->spi_rcv_egrbufs = (u64) pd->port_rcvegr_phys;
 152         kinfo->spi_pioavailaddr = (u64) dd->ipath_pioavailregs_phys;
 153         kinfo->spi_status = (u64) kinfo->spi_pioavailaddr +
 154                 (void *) dd->ipath_statusp -
 155                 (void *) dd->ipath_pioavailregs_dma;
 156         if (!shared) {
 157                 kinfo->spi_piocnt = dd->ipath_pbufsport;
 158                 kinfo->spi_piobufbase = (u64) pd->port_piobufs;
 159                 kinfo->__spi_uregbase = (u64) dd->ipath_uregbase +
 160                         dd->ipath_palign * pd->port_port;
 161         } else if (master) {
 162                 kinfo->spi_piocnt = (dd->ipath_pbufsport / subport_cnt) +
 163                                     (dd->ipath_pbufsport % subport_cnt);
 164                 /* Master's PIO buffers are after all the slave's */
 165                 kinfo->spi_piobufbase = (u64) pd->port_piobufs +
 166                         dd->ipath_palign *
 167                         (dd->ipath_pbufsport - kinfo->spi_piocnt);
 168                 kinfo->__spi_uregbase = (u64) dd->ipath_uregbase +
 169                         dd->ipath_palign * pd->port_port;
 170         } else {
 171                 unsigned slave = subport_fp(fp) - 1;
 172
 173                 kinfo->spi_piocnt = dd->ipath_pbufsport / subport_cnt;
 174                 kinfo->spi_piobufbase = (u64) pd->port_piobufs +
 175                         dd->ipath_palign * kinfo->spi_piocnt * slave;
 176                 kinfo->__spi_uregbase = ((u64) pd->subport_uregbase +
 177                         PAGE_SIZE * slave) & MMAP64_MASK;
 178
 179                 kinfo->spi_rcvhdr_base = ((u64) pd->subport_rcvhdr_base +
 180                         pd->port_rcvhdrq_size * slave) & MMAP64_MASK;
 181                 kinfo->spi_rcvhdr_tailaddr =
 182                         (u64) pd->port_rcvhdrqtailaddr_phys & MMAP64_MASK;
 183                 kinfo->spi_rcv_egrbufs = ((u64) pd->subport_rcvegrbuf +
 184                         dd->ipath_rcvegrcnt * dd->ipath_rcvegrbufsize * slave) &
 185                         MMAP64_MASK;
 186         }
 187
 188         kinfo->spi_pioindex = (kinfo->spi_piobufbase - dd->ipath_piobufbase) /
 189                 dd->ipath_palign;
 190         kinfo->spi_pioalign = dd->ipath_palign;
 191
 192         kinfo->spi_qpair = IPATH_KD_QP;
 193         kinfo->spi_piosize = dd->ipath_ibmaxlen;
 194         kinfo->spi_mtu = dd->ipath_ibmaxlen;    /* maxlen, not ibmtu */
 195         kinfo->spi_port = pd->port_port;
 196         kinfo->spi_subport = subport_fp(fp);
 197         kinfo->spi_sw_version = IPATH_KERN_SWVERSION;
 198         kinfo->spi_hw_version = dd->ipath_revision;
 199
 200         if (master) {
 201                 kinfo->spi_runtime_flags |= IPATH_RUNTIME_MASTER;
 202                 kinfo->spi_subport_uregbase =
 203                         (u64) pd->subport_uregbase & MMAP64_MASK;
 204                 kinfo->spi_subport_rcvegrbuf =
 205                         (u64) pd->subport_rcvegrbuf & MMAP64_MASK;
 206                 kinfo->spi_subport_rcvhdr_base =
 207                         (u64) pd->subport_rcvhdr_base & MMAP64_MASK;
 208                 ipath_cdbg(PROC, "port %u flags %x %llx %llx %llx\n",
 209                         kinfo->spi_port, kinfo->spi_runtime_flags,
 210                         (unsigned long long) kinfo->spi_subport_uregbase,
 211                         (unsigned long long) kinfo->spi_subport_rcvegrbuf,
 212                         (unsigned long long) kinfo->spi_subport_rcvhdr_base);
 213         }
 214
 215         if (copy_to_user(ubase, kinfo, sizeof(*kinfo)))
 216                 ret = -EFAULT;
 217
 218 bail:
 219         kfree(kinfo);
 220         return ret;
 221 }
 222
 223 /**
 224  * ipath_tid_update - update a port TID
 225  * @pd: the port
 226  * @fp: the ipath device file
 227  * @ti: the TID information
 228  *
 229  * The new implementation as of Oct 2004 is that the driver assigns
 230  * the tid and returns it to the caller.   To make it easier to
 231  * catch bugs, and to reduce search time, we keep a cursor for
 232  * each port, walking the shadow tid array to find one that's not
 233  * in use.
 234  *
 235  * For now, if we can't allocate the full list, we fail, although
 236  * in the long run, we'll allocate as many as we can, and the
 237  * caller will deal with that by trying the remaining pages later.
 238  * That means that when we fail, we have to mark the tids as not in
 239  * use again, in our shadow copy.
 240  *
 241  * It's up to the caller to free the tids when they are done.
 242  * We'll unlock the pages as they free them.
 243  *
 244  * Also, right now we are locking one page at a time, but since
 245  * the intended use of this routine is for a single group of
 246  * virtually contiguous pages, that should change to improve
 247  * performance.
 248  */
 249 static int ipath_tid_update(struct ipath_portdata *pd, struct file *fp,
 250                             const struct ipath_tid_info *ti)
 251 {
 252         int ret = 0, ntids;
 253         u32 tid, porttid, cnt, i, tidcnt, tidoff;
 254         u16 *tidlist;
 255         struct ipath_devdata *dd = pd->port_dd;
 256         u64 physaddr;
 257         unsigned long vaddr;
 258         u64 __iomem *tidbase;
 259         unsigned long tidmap[8];
 260         struct page **pagep = NULL;
 261         unsigned subport = subport_fp(fp);
 262
 263         if (!dd->ipath_pageshadow) {
 264                 ret = -ENOMEM;
 265                 goto done;
 266         }
 267
 268         cnt = ti->tidcnt;
 269         if (!cnt) {
 270                 ipath_dbg("After copyin, tidcnt 0, tidlist %llx\n",
 271                           (unsigned long long) ti->tidlist);
 272                 /*
 273                  * Should we treat as success?  likely a bug
 274                  */
 275                 ret = -EFAULT;
 276                 goto done;
 277         }
 278         porttid = pd->port_port * dd->ipath_rcvtidcnt;
 279         if (!pd->port_subport_cnt) {
 280                 tidcnt = dd->ipath_rcvtidcnt;
 281                 tid = pd->port_tidcursor;
 282                 tidoff = 0;
 283         } else if (!subport) {
 284                 tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
 285                          (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
 286                 tidoff = dd->ipath_rcvtidcnt - tidcnt;
 287                 porttid += tidoff;
 288                 tid = tidcursor_fp(fp);
 289         } else {
 290                 tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
 291                 tidoff = tidcnt * (subport - 1);
 292                 porttid += tidoff;
 293                 tid = tidcursor_fp(fp);
 294         }
 295         if (cnt > tidcnt) {
 296                 /* make sure it all fits in port_tid_pg_list */
 297                 dev_info(&dd->pcidev->dev, "Process tried to allocate %u "
 298                          "TIDs, only trying max (%u)\n", cnt, tidcnt);
 299                 cnt = tidcnt;
 300         }
 301         pagep = &((struct page **) pd->port_tid_pg_list)[tidoff];
 302         tidlist = &((u16 *) &pagep[dd->ipath_rcvtidcnt])[tidoff];
 303
 304         memset(tidmap, 0, sizeof(tidmap));
 305         /* before decrement; chip actual # */
 306         ntids = tidcnt;
 307         tidbase = (u64 __iomem *) (((char __iomem *) dd->ipath_kregbase) +
 308                                    dd->ipath_rcvtidbase +
 309                                    porttid * sizeof(*tidbase));
 310
 311         ipath_cdbg(VERBOSE, "Port%u %u tids, cursor %u, tidbase %p\n",
 312                    pd->port_port, cnt, tid, tidbase);
 313
 314         /* virtual address of first page in transfer */
 315         vaddr = ti->tidvaddr;
 316         if (!access_ok(VERIFY_WRITE, (void __user *) vaddr,
 317                        cnt * PAGE_SIZE)) {
 318                 ipath_dbg("Fail vaddr %p, %u pages, !access_ok\n",
 319                           (void *)vaddr, cnt);
 320                 ret = -EFAULT;
 321                 goto done;
 322         }
 323         ret = ipath_get_user_pages(vaddr, cnt, pagep);
 324         if (ret) {
 325                 if (ret == -EBUSY) {
 326                         ipath_dbg("Failed to lock addr %p, %u pages "
 327                                   "(already locked)\n",
 328                                   (void *) vaddr, cnt);
 329                         /*
 330                          * for now, continue, and see what happens but with
 331                          * the new implementation, this should never happen,
 332                          * unless perhaps the user has mpin'ed the pages
 333                          * themselves (something we need to test)
 334                          */
 335                         ret = 0;
 336                 } else {
 337                         dev_info(&dd->pcidev->dev,
 338                                  "Failed to lock addr %p, %u pages: "
 339                                  "errno %d\n", (void *) vaddr, cnt, -ret);
 340                         goto done;
 341                 }
 342         }
 343         for (i = 0; i < cnt; i++, vaddr += PAGE_SIZE) {
 344                 for (; ntids--; tid++) {
 345                         if (tid == tidcnt)
 346                                 tid = 0;
 347                         if (!dd->ipath_pageshadow[porttid + tid])
 348                                 break;
 349                 }
 350                 if (ntids < 0) {
 351                         /*
 352                          * oops, wrapped all the way through their TIDs,
 353                          * and didn't have enough free; see comments at
 354                          * start of routine
 355                          */
 356                         ipath_dbg("Not enough free TIDs for %u pages "
 357                                   "(index %d), failing\n", cnt, i);
 358                         i--;    /* last tidlist[i] not filled in */
 359                         ret = -ENOMEM;
 360                         break;
 361                 }
 362                 tidlist[i] = tid + tidoff;
 363                 ipath_cdbg(VERBOSE, "Updating idx %u to TID %u, "
 364                            "vaddr %lx\n", i, tid + tidoff, vaddr);
 365                 /* we "know" system pages and TID pages are same size */
 366                 dd->ipath_pageshadow[porttid + tid] = pagep[i];
 367                 dd->ipath_physshadow[porttid + tid] = ipath_map_page(
 368                         dd->pcidev, pagep[i], 0, PAGE_SIZE,
 369                         PCI_DMA_FROMDEVICE);
 370                 /*
 371                  * don't need atomic or it's overhead
 372                  */
 373                 __set_bit(tid, tidmap);
 374                 physaddr = dd->ipath_physshadow[porttid + tid];
 375                 ipath_stats.sps_pagelocks++;
 376                 ipath_cdbg(VERBOSE,
 377                            "TID %u, vaddr %lx, physaddr %llx pgp %p\n",
 378                            tid, vaddr, (unsigned long long) physaddr,
 379                            pagep[i]);
 380                 dd->ipath_f_put_tid(dd, &tidbase[tid], 1, physaddr);
 381                 /*
 382                  * don't check this tid in ipath_portshadow, since we
 383                  * just filled it in; start with the next one.
 384                  */
 385                 tid++;
 386         }
 387
 388         if (ret) {
 389                 u32 limit;
 390         cleanup:
 391                 /* jump here if copy out of updated info failed... */
 392                 ipath_dbg("After failure (ret=%d), undo %d of %d entries\n",
 393                           -ret, i, cnt);
 394                 /* same code that's in ipath_free_tid() */
 395                 limit = sizeof(tidmap) * BITS_PER_BYTE;
 396                 if (limit > tidcnt)
 397                         /* just in case size changes in future */
 398                         limit = tidcnt;
 399                 tid = find_first_bit((const unsigned long *)tidmap, limit);
 400                 for (; tid < limit; tid++) {
 401                         if (!test_bit(tid, tidmap))
 402                                 continue;
 403                         if (dd->ipath_pageshadow[porttid + tid]) {
 404                                 ipath_cdbg(VERBOSE, "Freeing TID %u\n",
 405                                            tid);
 406                                 dd->ipath_f_put_tid(dd, &tidbase[tid], 1,
 407                                                     dd->ipath_tidinvalid);
 408                                 pci_unmap_page(dd->pcidev,
 409                                         dd->ipath_physshadow[porttid + tid],
 410                                         PAGE_SIZE, PCI_DMA_FROMDEVICE);
 411                                 dd->ipath_pageshadow[porttid + tid] = NULL;
 412                                 ipath_stats.sps_pageunlocks++;
 413                         }
 414                 }
 415                 ipath_release_user_pages(pagep, cnt);
 416         } else {
 417                 /*
 418                  * Copy the updated array, with ipath_tid's filled in, back
 419                  * to user.  Since we did the copy in already, this "should
 420                  * never fail" If it does, we have to clean up...
 421                  */
 422                 if (copy_to_user((void __user *)
 423                                  (unsigned long) ti->tidlist,
 424                                  tidlist, cnt * sizeof(*tidlist))) {
 425                         ret = -EFAULT;
 426                         goto cleanup;
 427                 }
 428                 if (copy_to_user((void __user *) (unsigned long) ti->tidmap,
 429                                  tidmap, sizeof tidmap)) {
 430                         ret = -EFAULT;
 431                         goto cleanup;
 432                 }
 433                 if (tid == tidcnt)
 434                         tid = 0;
 435                 if (!pd->port_subport_cnt)
 436                         pd->port_tidcursor = tid;
 437                 else
 438                         tidcursor_fp(fp) = tid;
 439         }
 440
 441 done:
 442         if (ret)
 443                 ipath_dbg("Failed to map %u TID pages, failing with %d\n",
 444                           ti->tidcnt, -ret);
 445         return ret;
 446 }
 447
 448 /**
 449  * ipath_tid_free - free a port TID
 450  * @pd: the port
 451  * @subport: the subport
 452  * @ti: the TID info
 453  *
 454  * right now we are unlocking one page at a time, but since
 455  * the intended use of this routine is for a single group of
 456  * virtually contiguous pages, that should change to improve
 457  * performance.  We check that the TID is in range for this port
 458  * but otherwise don't check validity; if user has an error and
 459  * frees the wrong tid, it's only their own data that can thereby
 460  * be corrupted.  We do check that the TID was in use, for sanity
 461  * We always use our idea of the saved address, not the address that
 462  * they pass in to us.
 463  */
 464
 465 static int ipath_tid_free(struct ipath_portdata *pd, unsigned subport,
 466                           const struct ipath_tid_info *ti)
 467 {
 468         int ret = 0;
 469         u32 tid, porttid, cnt, limit, tidcnt;
 470         struct ipath_devdata *dd = pd->port_dd;
 471         u64 __iomem *tidbase;
 472         unsigned long tidmap[8];
 473
 474         if (!dd->ipath_pageshadow) {
 475                 ret = -ENOMEM;
 476                 goto done;
 477         }
 478
 479         if (copy_from_user(tidmap, (void __user *)(unsigned long)ti->tidmap,
 480                            sizeof tidmap)) {
 481                 ret = -EFAULT;
 482                 goto done;
 483         }
 484
 485         porttid = pd->port_port * dd->ipath_rcvtidcnt;
 486         if (!pd->port_subport_cnt)
 487                 tidcnt = dd->ipath_rcvtidcnt;
 488         else if (!subport) {
 489                 tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
 490                          (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
 491                 porttid += dd->ipath_rcvtidcnt - tidcnt;
 492         } else {
 493                 tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
 494                 porttid += tidcnt * (subport - 1);
 495         }
 496         tidbase = (u64 __iomem *) ((char __iomem *)(dd->ipath_kregbase) +
 497                                    dd->ipath_rcvtidbase +
 498                                    porttid * sizeof(*tidbase));
 499
 500         limit = sizeof(tidmap) * BITS_PER_BYTE;
 501         if (limit > tidcnt)
 502                 /* just in case size changes in future */
 503                 limit = tidcnt;
 504         tid = find_first_bit(tidmap, limit);
 505         ipath_cdbg(VERBOSE, "Port%u free %u tids; first bit (max=%d) "
 506                    "set is %d, porttid %u\n", pd->port_port, ti->tidcnt,
 507                    limit, tid, porttid);
 508         for (cnt = 0; tid < limit; tid++) {
 509                 /*
 510                  * small optimization; if we detect a run of 3 or so without
 511                  * any set, use find_first_bit again.  That's mainly to
 512                  * accelerate the case where we wrapped, so we have some at
 513                  * the beginning, and some at the end, and a big gap
 514                  * in the middle.
 515                  */
 516                 if (!test_bit(tid, tidmap))
 517                         continue;
 518                 cnt++;
 519                 if (dd->ipath_pageshadow[porttid + tid]) {
 520                         ipath_cdbg(VERBOSE, "PID %u freeing TID %u\n",
 521                                    pd->port_pid, tid);
 522                         dd->ipath_f_put_tid(dd, &tidbase[tid], 1,
 523                                             dd->ipath_tidinvalid);
 524                         pci_unmap_page(dd->pcidev,
 525                                 dd->ipath_physshadow[porttid + tid],
 526                                 PAGE_SIZE, PCI_DMA_FROMDEVICE);
 527                         ipath_release_user_pages(
 528                                 &dd->ipath_pageshadow[porttid + tid], 1);
 529                         dd->ipath_pageshadow[porttid + tid] = NULL;
 530                         ipath_stats.sps_pageunlocks++;
 531                 } else
 532                         ipath_dbg("Unused tid %u, ignoring\n", tid);
 533         }
 534         if (cnt != ti->tidcnt)
 535                 ipath_dbg("passed in tidcnt %d, only %d bits set in map\n",
 536                           ti->tidcnt, cnt);
 537 done:
 538         if (ret)
 539                 ipath_dbg("Failed to unmap %u TID pages, failing with %d\n",
 540                           ti->tidcnt, -ret);
 541         return ret;
 542 }
 543
 544 /**
 545  * ipath_set_part_key - set a partition key
 546  * @pd: the port
 547  * @key: the key
 548  *
 549  * We can have up to 4 active at a time (other than the default, which is
 550  * always allowed).  This is somewhat tricky, since multiple ports may set
 551  * the same key, so we reference count them, and clean up at exit.  All 4
 552  * partition keys are packed into a single infinipath register.  It's an
 553  * error for a process to set the same pkey multiple times.  We provide no
 554  * mechanism to de-allocate a pkey at this time, we may eventually need to
 555  * do that.  I've used the atomic operations, and no locking, and only make
 556  * a single pass through what's available.  This should be more than
 557  * adequate for some time. I'll think about spinlocks or the like if and as
 558  * it's necessary.
 559  */
 560 static int ipath_set_part_key(struct ipath_portdata *pd, u16 key)
 561 {
 562         struct ipath_devdata *dd = pd->port_dd;
 563         int i, any = 0, pidx = -1;
 564         u16 lkey = key & 0x7FFF;
 565         int ret;
 566
 567         if (lkey == (IPATH_DEFAULT_P_KEY & 0x7FFF)) {
 568                 /* nothing to do; this key always valid */
 569                 ret = 0;
 570                 goto bail;
 571         }
 572
 573         ipath_cdbg(VERBOSE, "p%u try to set pkey %hx, current keys "
 574                    "%hx:%x %hx:%x %hx:%x %hx:%x\n",
 575                    pd->port_port, key, dd->ipath_pkeys[0],
 576                    atomic_read(&dd->ipath_pkeyrefs[0]), dd->ipath_pkeys[1],
 577                    atomic_read(&dd->ipath_pkeyrefs[1]), dd->ipath_pkeys[2],
 578                    atomic_read(&dd->ipath_pkeyrefs[2]), dd->ipath_pkeys[3],
 579                    atomic_read(&dd->ipath_pkeyrefs[3]));
 580
 581         if (!lkey) {
 582                 ipath_cdbg(PROC, "p%u tries to set key 0, not allowed\n",
 583                            pd->port_port);
 584                 ret = -EINVAL;
 585                 goto bail;
 586         }
 587
 588         /*
 589          * Set the full membership bit, because it has to be
 590          * set in the register or the packet, and it seems
 591          * cleaner to set in the register than to force all
 592          * callers to set it. (see bug 4331)
 593          */
 594         key |= 0x8000;
 595
 596         for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
 597                 if (!pd->port_pkeys[i] && pidx == -1)
 598                         pidx = i;
 599                 if (pd->port_pkeys[i] == key) {
 600                         ipath_cdbg(VERBOSE, "p%u tries to set same pkey "
 601                                    "(%x) more than once\n",
 602                                    pd->port_port, key);
 603                         ret = -EEXIST;
 604                         goto bail;
 605                 }
 606         }
 607         if (pidx == -1) {
 608                 ipath_dbg("All pkeys for port %u already in use, "
 609                           "can't set %x\n", pd->port_port, key);
 610                 ret = -EBUSY;
 611                 goto bail;
 612         }
 613         for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
 614                 if (!dd->ipath_pkeys[i]) {
 615                         any++;
 616                         continue;
 617                 }
 618                 if (dd->ipath_pkeys[i] == key) {
 619                         atomic_t *pkrefs = &dd->ipath_pkeyrefs[i];
 620
 621                         if (atomic_inc_return(pkrefs) > 1) {
 622                                 pd->port_pkeys[pidx] = key;
 623                                 ipath_cdbg(VERBOSE, "p%u set key %x "
 624                                            "matches #%d, count now %d\n",
 625                                            pd->port_port, key, i,
 626                                            atomic_read(pkrefs));
 627                                 ret = 0;
 628                                 goto bail;
 629                         } else {
 630                                 /*
 631                                  * lost race, decrement count, catch below
 632                                  */
 633                                 atomic_dec(pkrefs);
 634                                 ipath_cdbg(VERBOSE, "Lost race, count was "
 635                                            "0, after dec, it's %d\n",
 636                                            atomic_read(pkrefs));
 637                                 any++;
 638                         }
 639                 }
 640                 if ((dd->ipath_pkeys[i] & 0x7FFF) == lkey) {
 641                         /*
 642                          * It makes no sense to have both the limited and
 643                          * full membership PKEY set at the same time since
 644                          * the unlimited one will disable the limited one.
 645                          */
 646                         ret = -EEXIST;
 647                         goto bail;
 648                 }
 649         }
 650         if (!any) {
 651                 ipath_dbg("port %u, all pkeys already in use, "
 652                           "can't set %x\n", pd->port_port, key);
 653                 ret = -EBUSY;
 654                 goto bail;
 655         }
 656         for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
 657                 if (!dd->ipath_pkeys[i] &&
 658                     atomic_inc_return(&dd->ipath_pkeyrefs[i]) == 1) {
 659                         u64 pkey;
 660
 661                         /* for ipathstats, etc. */
 662                         ipath_stats.sps_pkeys[i] = lkey;
 663                         pd->port_pkeys[pidx] = dd->ipath_pkeys[i] = key;
 664                         pkey =
 665                                 (u64) dd->ipath_pkeys[0] |
 666                                 ((u64) dd->ipath_pkeys[1] << 16) |
 667                                 ((u64) dd->ipath_pkeys[2] << 32) |
 668                                 ((u64) dd->ipath_pkeys[3] << 48);
 669                         ipath_cdbg(PROC, "p%u set key %x in #%d, "
 670                                    "portidx %d, new pkey reg %llx\n",
 671                                    pd->port_port, key, i, pidx,
 672                                    (unsigned long long) pkey);
 673                         ipath_write_kreg(
 674                                 dd, dd->ipath_kregs->kr_partitionkey, pkey);
 675
 676                         ret = 0;
 677                         goto bail;
 678                 }
 679         }
 680         ipath_dbg("port %u, all pkeys already in use 2nd pass, "
 681                   "can't set %x\n", pd->port_port, key);
 682         ret = -EBUSY;
 683
 684 bail:
 685         return ret;
 686 }
 687
 688 /**
 689  * ipath_manage_rcvq - manage a port's receive queue
 690  * @pd: the port
 691  * @subport: the subport
 692  * @start_stop: action to carry out
 693  *
 694  * start_stop == 0 disables receive on the port, for use in queue
 695  * overflow conditions.  start_stop==1 re-enables, to be used to
 696  * re-init the software copy of the head register
 697  */
 698 static int ipath_manage_rcvq(struct ipath_portdata *pd, unsigned subport,
 699                              int start_stop)
 700 {
 701         struct ipath_devdata *dd = pd->port_dd;
 702
 703         ipath_cdbg(PROC, "%sabling rcv for unit %u port %u:%u\n",
 704                    start_stop ? "en" : "dis", dd->ipath_unit,
 705                    pd->port_port, subport);
 706         if (subport)
 707                 goto bail;
 708         /* atomically clear receive enable port. */
 709         if (start_stop) {
 710                 /*
 711                  * On enable, force in-memory copy of the tail register to
 712                  * 0, so that protocol code doesn't have to worry about
 713                  * whether or not the chip has yet updated the in-memory
 714                  * copy or not on return from the system call. The chip
 715                  * always resets it's tail register back to 0 on a
 716                  * transition from disabled to enabled.  This could cause a
 717                  * problem if software was broken, and did the enable w/o
 718                  * the disable, but eventually the in-memory copy will be
 719                  * updated and correct itself, even in the face of software
 720                  * bugs.
 721                  */
 722                 *(volatile u64 *)pd->port_rcvhdrtail_kvaddr = 0;
 723                 set_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port,
 724                         &dd->ipath_rcvctrl);
 725         } else
 726                 clear_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port,
 727                           &dd->ipath_rcvctrl);
 728         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
 729                          dd->ipath_rcvctrl);
 730         /* now be sure chip saw it before we return */
 731         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 732         if (start_stop) {
 733                 /*
 734                  * And try to be sure that tail reg update has happened too.
 735                  * This should in theory interlock with the RXE changes to
 736                  * the tail register.  Don't assign it to the tail register
 737                  * in memory copy, since we could overwrite an update by the
 738                  * chip if we did.
 739                  */
 740                 ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
 741         }
 742         /* always; new head should be equal to new tail; see above */
 743 bail:
 744         return 0;
 745 }
 746
 747 static void ipath_clean_part_key(struct ipath_portdata *pd,
 748                                  struct ipath_devdata *dd)
 749 {
 750         int i, j, pchanged = 0;
 751         u64 oldpkey;
 752
 753         /* for debugging only */
 754         oldpkey = (u64) dd->ipath_pkeys[0] |
 755                 ((u64) dd->ipath_pkeys[1] << 16) |
 756                 ((u64) dd->ipath_pkeys[2] << 32) |
 757                 ((u64) dd->ipath_pkeys[3] << 48);
 758
 759         for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
 760                 if (!pd->port_pkeys[i])
 761                         continue;
 762                 ipath_cdbg(VERBOSE, "look for key[%d] %hx in pkeys\n", i,
 763                            pd->port_pkeys[i]);
 764                 for (j = 0; j < ARRAY_SIZE(dd->ipath_pkeys); j++) {
 765                         /* check for match independent of the global bit */
 766                         if ((dd->ipath_pkeys[j] & 0x7fff) !=
 767                             (pd->port_pkeys[i] & 0x7fff))
 768                                 continue;
 769                         if (atomic_dec_and_test(&dd->ipath_pkeyrefs[j])) {
 770                                 ipath_cdbg(VERBOSE, "p%u clear key "
 771                                            "%x matches #%d\n",
 772                                            pd->port_port,
 773                                            pd->port_pkeys[i], j);
 774                                 ipath_stats.sps_pkeys[j] =
 775                                         dd->ipath_pkeys[j] = 0;
 776                                 pchanged++;
 777                         }
 778                         else ipath_cdbg(
 779                                 VERBOSE, "p%u key %x matches #%d, "
 780                                 "but ref still %d\n", pd->port_port,
 781                                 pd->port_pkeys[i], j,
 782                                 atomic_read(&dd->ipath_pkeyrefs[j]));
 783                         break;
 784                 }
 785                 pd->port_pkeys[i] = 0;
 786         }
 787         if (pchanged) {
 788                 u64 pkey = (u64) dd->ipath_pkeys[0] |
 789                         ((u64) dd->ipath_pkeys[1] << 16) |
 790                         ((u64) dd->ipath_pkeys[2] << 32) |
 791                         ((u64) dd->ipath_pkeys[3] << 48);
 792                 ipath_cdbg(VERBOSE, "p%u old pkey reg %llx, "
 793                            "new pkey reg %llx\n", pd->port_port,
 794                            (unsigned long long) oldpkey,
 795                            (unsigned long long) pkey);
 796                 ipath_write_kreg(dd, dd->ipath_kregs->kr_partitionkey,
 797                                  pkey);
 798         }
 799 }
 800
 801 /*
 802  * Initialize the port data with the receive buffer sizes
 803  * so this can be done while the master port is locked.
 804  * Otherwise, there is a race with a slave opening the port
 805  * and seeing these fields uninitialized.
 806  */
 807 static void init_user_egr_sizes(struct ipath_portdata *pd)
 808 {
 809         struct ipath_devdata *dd = pd->port_dd;
 810         unsigned egrperchunk, egrcnt, size;
 811
 812         /*
 813          * to avoid wasting a lot of memory, we allocate 32KB chunks of
 814          * physically contiguous memory, advance through it until used up
 815          * and then allocate more.  Of course, we need memory to store those
 816          * extra pointers, now.  Started out with 256KB, but under heavy
 817          * memory pressure (creating large files and then copying them over
 818          * NFS while doing lots of MPI jobs), we hit some allocation
 819          * failures, even though we can sleep...  (2.6.10) Still get
 820          * failures at 64K.  32K is the lowest we can go without wasting
 821          * additional memory.
 822          */
 823         size = 0x8000;
 824         egrperchunk = size / dd->ipath_rcvegrbufsize;
 825         egrcnt = dd->ipath_rcvegrcnt;
 826         pd->port_rcvegrbuf_chunks = (egrcnt + egrperchunk - 1) / egrperchunk;
 827         pd->port_rcvegrbufs_perchunk = egrperchunk;
 828         pd->port_rcvegrbuf_size = size;
 829 }
 830
 831 /**
 832  * ipath_create_user_egr - allocate eager TID buffers
 833  * @pd: the port to allocate TID buffers for
 834  *
 835  * This routine is now quite different for user and kernel, because
 836  * the kernel uses skb's, for the accelerated network performance
 837  * This is the user port version
 838  *
 839  * Allocate the eager TID buffers and program them into infinipath
 840  * They are no longer completely contiguous, we do multiple allocation
 841  * calls.
 842  */
 843 static int ipath_create_user_egr(struct ipath_portdata *pd)
 844 {
 845         struct ipath_devdata *dd = pd->port_dd;
 846         unsigned e, egrcnt, egrperchunk, chunk, egrsize, egroff;
 847         size_t size;
 848         int ret;
 849         gfp_t gfp_flags;
 850
 851         /*
 852          * GFP_USER, but without GFP_FS, so buffer cache can be
 853          * coalesced (we hope); otherwise, even at order 4,
 854          * heavy filesystem activity makes these fail, and we can
 855          * use compound pages.
 856          */
 857         gfp_flags = __GFP_WAIT | __GFP_IO | __GFP_COMP;
 858
 859         egrcnt = dd->ipath_rcvegrcnt;
 860         /* TID number offset for this port */
 861         egroff = pd->port_port * egrcnt;
 862         egrsize = dd->ipath_rcvegrbufsize;
 863         ipath_cdbg(VERBOSE, "Allocating %d egr buffers, at egrtid "
 864                    "offset %x, egrsize %u\n", egrcnt, egroff, egrsize);
 865
 866         chunk = pd->port_rcvegrbuf_chunks;
 867         egrperchunk = pd->port_rcvegrbufs_perchunk;
 868         size = pd->port_rcvegrbuf_size;
 869         pd->port_rcvegrbuf = kmalloc(chunk * sizeof(pd->port_rcvegrbuf[0]),
 870                                      GFP_KERNEL);
 871         if (!pd->port_rcvegrbuf) {
 872                 ret = -ENOMEM;
 873                 goto bail;
 874         }
 875         pd->port_rcvegrbuf_phys =
 876                 kmalloc(chunk * sizeof(pd->port_rcvegrbuf_phys[0]),
 877                         GFP_KERNEL);
 878         if (!pd->port_rcvegrbuf_phys) {
 879                 ret = -ENOMEM;
 880                 goto bail_rcvegrbuf;
 881         }
 882         for (e = 0; e < pd->port_rcvegrbuf_chunks; e++) {
 883
 884                 pd->port_rcvegrbuf[e] = dma_alloc_coherent(
 885                         &dd->pcidev->dev, size, &pd->port_rcvegrbuf_phys[e],
 886                         gfp_flags);
 887
 888                 if (!pd->port_rcvegrbuf[e]) {
 889                         ret = -ENOMEM;
 890                         goto bail_rcvegrbuf_phys;
 891                 }
 892         }
 893
 894         pd->port_rcvegr_phys = pd->port_rcvegrbuf_phys[0];
 895
 896         for (e = chunk = 0; chunk < pd->port_rcvegrbuf_chunks; chunk++) {
 897                 dma_addr_t pa = pd->port_rcvegrbuf_phys[chunk];
 898                 unsigned i;
 899
 900                 for (i = 0; e < egrcnt && i < egrperchunk; e++, i++) {
 901                         dd->ipath_f_put_tid(dd, e + egroff +
 902                                             (u64 __iomem *)
 903                                             ((char __iomem *)
 904                                              dd->ipath_kregbase +
 905                                              dd->ipath_rcvegrbase), 0, pa);
 906                         pa += egrsize;
 907                 }
 908                 cond_resched(); /* don't hog the cpu */
 909         }
 910
 911         ret = 0;
 912         goto bail;
 913
 914 bail_rcvegrbuf_phys:
 915         for (e = 0; e < pd->port_rcvegrbuf_chunks &&
 916                 pd->port_rcvegrbuf[e]; e++) {
 917                 dma_free_coherent(&dd->pcidev->dev, size,
 918                                   pd->port_rcvegrbuf[e],
 919                                   pd->port_rcvegrbuf_phys[e]);
 920
 921         }
 922         kfree(pd->port_rcvegrbuf_phys);
 923         pd->port_rcvegrbuf_phys = NULL;
 924 bail_rcvegrbuf:
 925         kfree(pd->port_rcvegrbuf);
 926         pd->port_rcvegrbuf = NULL;
 927 bail:
 928         return ret;
 929 }
 930
 931
 932 /* common code for the mappings on dma_alloc_coherent mem */
 933 static int ipath_mmap_mem(struct vm_area_struct *vma,
 934         struct ipath_portdata *pd, unsigned len, int write_ok,
 935         void *kvaddr, char *what)
 936 {
 937         struct ipath_devdata *dd = pd->port_dd;
 938         unsigned long pfn;
 939         int ret;
 940
 941         if ((vma->vm_end - vma->vm_start) > len) {
 942                 dev_info(&dd->pcidev->dev,
 943                          "FAIL on %s: len %lx > %x\n", what,
 944                          vma->vm_end - vma->vm_start, len);
 945                 ret = -EFAULT;
 946                 goto bail;
 947         }
 948
 949         if (!write_ok) {
 950                 if (vma->vm_flags & VM_WRITE) {
 951                         dev_info(&dd->pcidev->dev,
 952                                  "%s must be mapped readonly\n", what);
 953                         ret = -EPERM;
 954                         goto bail;
 955                 }
 956
 957                 /* don't allow them to later change with mprotect */
 958                 vma->vm_flags &= ~VM_MAYWRITE;
 959         }
 960
 961         pfn = virt_to_phys(kvaddr) >> PAGE_SHIFT;
 962         ret = remap_pfn_range(vma, vma->vm_start, pfn,
 963                               len, vma->vm_page_prot);
 964         if (ret)
 965                 dev_info(&dd->pcidev->dev, "%s port%u mmap of %lx, %x "
 966                          "bytes r%c failed: %d\n", what, pd->port_port,
 967                          pfn, len, write_ok?'w':'o', ret);
 968         else
 969                 ipath_cdbg(VERBOSE, "%s port%u mmaped %lx, %x bytes "
 970                            "r%c\n", what, pd->port_port, pfn, len,
 971                            write_ok?'w':'o');
 972 bail:
 973         return ret;
 974 }
 975
 976 static int mmap_ureg(struct vm_area_struct *vma, struct ipath_devdata *dd,
 977                      u64 ureg)
 978 {
 979         unsigned long phys;
 980         int ret;
 981
 982         /*
 983          * This is real hardware, so use io_remap.  This is the mechanism
 984          * for the user process to update the head registers for their port
 985          * in the chip.
 986          */
 987         if ((vma->vm_end - vma->vm_start) > PAGE_SIZE) {
 988                 dev_info(&dd->pcidev->dev, "FAIL mmap userreg: reqlen "
 989                          "%lx > PAGE\n", vma->vm_end - vma->vm_start);
 990                 ret = -EFAULT;
 991         } else {
 992                 phys = dd->ipath_physaddr + ureg;
 993                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 994
 995                 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
 996                 ret = io_remap_pfn_range(vma, vma->vm_start,
 997                                          phys >> PAGE_SHIFT,
 998                                          vma->vm_end - vma->vm_start,
 999                                          vma->vm_page_prot);
1000         }
1001         return ret;
1002 }
1003
1004 static int mmap_piobufs(struct vm_area_struct *vma,
1005                         struct ipath_devdata *dd,
1006                         struct ipath_portdata *pd,
1007                         unsigned piobufs, unsigned piocnt)
1008 {
1009         unsigned long phys;
1010         int ret;
1011
1012         /*
1013          * When we map the PIO buffers in the chip, we want to map them as
1014          * writeonly, no read possible.   This prevents access to previous
1015          * process data, and catches users who might try to read the i/o
1016          * space due to a bug.
1017          */
1018         if ((vma->vm_end - vma->vm_start) > (piocnt * dd->ipath_palign)) {
1019                 dev_info(&dd->pcidev->dev, "FAIL mmap piobufs: "
1020                          "reqlen %lx > PAGE\n",
1021                          vma->vm_end - vma->vm_start);
1022                 ret = -EINVAL;
1023                 goto bail;
1024         }
1025
1026         phys = dd->ipath_physaddr + piobufs;
1027
1028         /*
1029          * Don't mark this as non-cached, or we don't get the
1030          * write combining behavior we want on the PIO buffers!
1031          */
1032
1033 #if defined(__powerpc__)
1034         /* There isn't a generic way to specify writethrough mappings */
1035         pgprot_val(vma->vm_page_prot) |= _PAGE_NO_CACHE;
1036         pgprot_val(vma->vm_page_prot) |= _PAGE_WRITETHRU;
1037         pgprot_val(vma->vm_page_prot) &= ~_PAGE_GUARDED;
1038 #endif
1039
1040         /*
1041          * don't allow them to later change to readable with mprotect (for when
1042          * not initially mapped readable, as is normally the case)
1043          */
1044         vma->vm_flags &= ~VM_MAYREAD;
1045         vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1046
1047         ret = io_remap_pfn_range(vma, vma->vm_start, phys >> PAGE_SHIFT,
1048                                  vma->vm_end - vma->vm_start,
1049                                  vma->vm_page_prot);
1050 bail:
1051         return ret;
1052 }
1053
1054 static int mmap_rcvegrbufs(struct vm_area_struct *vma,
1055                            struct ipath_portdata *pd)
1056 {
1057         struct ipath_devdata *dd = pd->port_dd;
1058         unsigned long start, size;
1059         size_t total_size, i;
1060         unsigned long pfn;
1061         int ret;
1062
1063         size = pd->port_rcvegrbuf_size;
1064         total_size = pd->port_rcvegrbuf_chunks * size;
1065         if ((vma->vm_end - vma->vm_start) > total_size) {
1066                 dev_info(&dd->pcidev->dev, "FAIL on egr bufs: "
1067                          "reqlen %lx > actual %lx\n",
1068                          vma->vm_end - vma->vm_start,
1069                          (unsigned long) total_size);
1070                 ret = -EINVAL;
1071                 goto bail;
1072         }
1073
1074         if (vma->vm_flags & VM_WRITE) {
1075                 dev_info(&dd->pcidev->dev, "Can't map eager buffers as "
1076                          "writable (flags=%lx)\n", vma->vm_flags);
1077                 ret = -EPERM;
1078                 goto bail;
1079         }
1080         /* don't allow them to later change to writeable with mprotect */
1081         vma->vm_flags &= ~VM_MAYWRITE;
1082
1083         start = vma->vm_start;
1084
1085         for (i = 0; i < pd->port_rcvegrbuf_chunks; i++, start += size) {
1086                 pfn = virt_to_phys(pd->port_rcvegrbuf[i]) >> PAGE_SHIFT;
1087                 ret = remap_pfn_range(vma, start, pfn, size,
1088                                       vma->vm_page_prot);
1089                 if (ret < 0)
1090                         goto bail;
1091         }
1092         ret = 0;
1093
1094 bail:
1095         return ret;
1096 }
1097
1098 /*
1099  * ipath_file_vma_nopage - handle a VMA page fault.
1100  */
1101 static struct page *ipath_file_vma_nopage(struct vm_area_struct *vma,
1102                                           unsigned long address, int *type)
1103 {
1104         unsigned long offset = address - vma->vm_start;
1105         struct page *page = NOPAGE_SIGBUS;
1106         void *pageptr;
1107
1108         /*
1109          * Convert the vmalloc address into a struct page.
1110          */
1111         pageptr = (void *)(offset + (vma->vm_pgoff << PAGE_SHIFT));
1112         page = vmalloc_to_page(pageptr);
1113         if (!page)
1114                 goto out;
1115
1116         /* Increment the reference count. */
1117         get_page(page);
1118         if (type)
1119                 *type = VM_FAULT_MINOR;
1120 out:
1121         return page;
1122 }
1123
1124 static struct vm_operations_struct ipath_file_vm_ops = {
1125         .nopage = ipath_file_vma_nopage,
1126 };
1127
1128 static int mmap_kvaddr(struct vm_area_struct *vma, u64 pgaddr,
1129                        struct ipath_portdata *pd, unsigned subport)
1130 {
1131         unsigned long len;
1132         struct ipath_devdata *dd;
1133         void *addr;
1134         size_t size;
1135         int ret;
1136
1137         /* If the port is not shared, all addresses should be physical */
1138         if (!pd->port_subport_cnt) {
1139                 ret = -EINVAL;
1140                 goto bail;
1141         }
1142
1143         dd = pd->port_dd;
1144         size = pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
1145
1146         /*
1147          * Master has all the slave uregbase, rcvhdrq, and
1148          * rcvegrbufs mmapped.
1149          */
1150         if (subport == 0) {
1151                 unsigned num_slaves = pd->port_subport_cnt - 1;
1152
1153                 if (pgaddr == ((u64) pd->subport_uregbase & MMAP64_MASK)) {
1154                         addr = pd->subport_uregbase;
1155                         size = PAGE_SIZE * num_slaves;
1156                 } else if (pgaddr == ((u64) pd->subport_rcvhdr_base &
1157                                       MMAP64_MASK)) {
1158                         addr = pd->subport_rcvhdr_base;
1159                         size = pd->port_rcvhdrq_size * num_slaves;
1160                 } else if (pgaddr == ((u64) pd->subport_rcvegrbuf &
1161                                       MMAP64_MASK)) {
1162                         addr = pd->subport_rcvegrbuf;
1163                         size *= num_slaves;
1164                 } else {
1165                         ret = -EINVAL;
1166                         goto bail;
1167                 }
1168         } else if (pgaddr == (((u64) pd->subport_uregbase +
1169                                PAGE_SIZE * (subport - 1)) & MMAP64_MASK)) {
1170                 addr = pd->subport_uregbase + PAGE_SIZE * (subport - 1);
1171                 size = PAGE_SIZE;
1172         } else if (pgaddr == (((u64) pd->subport_rcvhdr_base +
1173                                pd->port_rcvhdrq_size * (subport - 1)) &
1174                               MMAP64_MASK)) {
1175                 addr = pd->subport_rcvhdr_base +
1176                         pd->port_rcvhdrq_size * (subport - 1);
1177                 size = pd->port_rcvhdrq_size;
1178         } else if (pgaddr == (((u64) pd->subport_rcvegrbuf +
1179                                size * (subport - 1)) & MMAP64_MASK)) {
1180                 addr = pd->subport_rcvegrbuf + size * (subport - 1);
1181                 /* rcvegrbufs are read-only on the slave */
1182                 if (vma->vm_flags & VM_WRITE) {
1183                         dev_info(&dd->pcidev->dev,
1184                                  "Can't map eager buffers as "
1185                                  "writable (flags=%lx)\n", vma->vm_flags);
1186                         ret = -EPERM;
1187                         goto bail;
1188                 }
1189                 /*
1190                  * Don't allow permission to later change to writeable
1191                  * with mprotect.
1192                  */
1193                 vma->vm_flags &= ~VM_MAYWRITE;
1194         } else {
1195                 ret = -EINVAL;
1196                 goto bail;
1197         }
1198         len = vma->vm_end - vma->vm_start;
1199         if (len > size) {
1200                 ipath_cdbg(MM, "FAIL: reqlen %lx > %zx\n", len, size);
1201                 ret = -EINVAL;
1202                 goto bail;
1203         }
1204
1205         vma->vm_pgoff = (unsigned long) addr >> PAGE_SHIFT;
1206         vma->vm_ops = &ipath_file_vm_ops;
1207         vma->vm_flags |= VM_RESERVED | VM_DONTEXPAND;
1208         ret = 0;
1209
1210 bail:
1211         return ret;
1212 }
1213
1214 /**
1215  * ipath_mmap - mmap various structures into user space
1216  * @fp: the file pointer
1217  * @vma: the VM area
1218  *
1219  * We use this to have a shared buffer between the kernel and the user code
1220  * for the rcvhdr queue, egr buffers, and the per-port user regs and pio
1221  * buffers in the chip.  We have the open and close entries so we can bump
1222  * the ref count and keep the driver from being unloaded while still mapped.
1223  */
1224 static int ipath_mmap(struct file *fp, struct vm_area_struct *vma)
1225 {
1226         struct ipath_portdata *pd;
1227         struct ipath_devdata *dd;
1228         u64 pgaddr, ureg;
1229         unsigned piobufs, piocnt;
1230         int ret;
1231
1232         pd = port_fp(fp);
1233         if (!pd) {
1234                 ret = -EINVAL;
1235                 goto bail;
1236         }
1237         dd = pd->port_dd;
1238
1239         /*
1240          * This is the ipath_do_user_init() code, mapping the shared buffers
1241          * into the user process. The address referred to by vm_pgoff is the
1242          * file offset passed via mmap().  For shared ports, this is the
1243          * kernel vmalloc() address of the pages to share with the master.
1244          * For non-shared or master ports, this is a physical address.
1245          * We only do one mmap for each space mapped.
1246          */
1247         pgaddr = vma->vm_pgoff << PAGE_SHIFT;
1248
1249         /*
1250          * Check for 0 in case one of the allocations failed, but user
1251          * called mmap anyway.
1252          */
1253         if (!pgaddr)  {
1254                 ret = -EINVAL;
1255                 goto bail;
1256         }
1257
1258         ipath_cdbg(MM, "pgaddr %llx vm_start=%lx len %lx port %u:%u:%u\n",
1259                    (unsigned long long) pgaddr, vma->vm_start,
1260                    vma->vm_end - vma->vm_start, dd->ipath_unit,
1261                    pd->port_port, subport_fp(fp));
1262
1263         /*
1264          * Physical addresses must fit in 40 bits for our hardware.
1265          * Check for kernel virtual addresses first, anything else must
1266          * match a HW or memory address.
1267          */
1268         if (pgaddr >= (1ULL<<40)) {
1269                 ret = mmap_kvaddr(vma, pgaddr, pd, subport_fp(fp));
1270                 goto bail;
1271         }
1272
1273         if (!pd->port_subport_cnt) {
1274                 /* port is not shared */
1275                 ureg = dd->ipath_uregbase + dd->ipath_palign * pd->port_port;
1276                 piocnt = dd->ipath_pbufsport;
1277                 piobufs = pd->port_piobufs;
1278         } else if (!subport_fp(fp)) {
1279                 /* caller is the master */
1280                 ureg = dd->ipath_uregbase + dd->ipath_palign * pd->port_port;
1281                 piocnt = (dd->ipath_pbufsport / pd->port_subport_cnt) +
1282                          (dd->ipath_pbufsport % pd->port_subport_cnt);
1283                 piobufs = pd->port_piobufs +
1284                         dd->ipath_palign * (dd->ipath_pbufsport - piocnt);
1285         } else {
1286                 unsigned slave = subport_fp(fp) - 1;
1287
1288                 /* caller is a slave */
1289                 ureg = 0;
1290                 piocnt = dd->ipath_pbufsport / pd->port_subport_cnt;
1291                 piobufs = pd->port_piobufs + dd->ipath_palign * piocnt * slave;
1292         }
1293
1294         if (pgaddr == ureg)
1295                 ret = mmap_ureg(vma, dd, ureg);
1296         else if (pgaddr == piobufs)
1297                 ret = mmap_piobufs(vma, dd, pd, piobufs, piocnt);
1298         else if (pgaddr == dd->ipath_pioavailregs_phys)
1299                 /* in-memory copy of pioavail registers */
1300                 ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1301                                      (void *) dd->ipath_pioavailregs_dma,
1302                                      "pioavail registers");
1303         else if (subport_fp(fp))
1304                 /* Subports don't mmap the physical receive buffers */
1305                 ret = -EINVAL;
1306         else if (pgaddr == pd->port_rcvegr_phys)
1307                 ret = mmap_rcvegrbufs(vma, pd);
1308         else if (pgaddr == (u64) pd->port_rcvhdrq_phys)
1309                 /*
1310                  * The rcvhdrq itself; readonly except on HT (so have
1311                  * to allow writable mapping), multiple pages, contiguous
1312                  * from an i/o perspective.
1313                  */
1314                 ret = ipath_mmap_mem(vma, pd, pd->port_rcvhdrq_size, 1,
1315                                      pd->port_rcvhdrq,
1316                                      "rcvhdrq");
1317         else if (pgaddr == (u64) pd->port_rcvhdrqtailaddr_phys)
1318                 /* in-memory copy of rcvhdrq tail register */
1319                 ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1320                                      pd->port_rcvhdrtail_kvaddr,
1321                                      "rcvhdrq tail");
1322         else
1323                 ret = -EINVAL;
1324
1325         vma->vm_private_data = NULL;
1326
1327         if (ret < 0)
1328                 dev_info(&dd->pcidev->dev,
1329                          "Failure %d on off %llx len %lx\n",
1330                          -ret, (unsigned long long)pgaddr,
1331                          vma->vm_end - vma->vm_start);
1332 bail:
1333         return ret;
1334 }
1335
1336 static unsigned int ipath_poll(struct file *fp,
1337                                struct poll_table_struct *pt)
1338 {
1339         struct ipath_portdata *pd;
1340         u32 head, tail;
1341         int bit;
1342         unsigned pollflag = 0;
1343         struct ipath_devdata *dd;
1344
1345         pd = port_fp(fp);
1346         if (!pd)
1347                 goto bail;
1348         dd = pd->port_dd;
1349
1350         bit = pd->port_port + INFINIPATH_R_INTRAVAIL_SHIFT;
1351         set_bit(bit, &dd->ipath_rcvctrl);
1352
1353         /*
1354          * Before blocking, make sure that head is still == tail,
1355          * reading from the chip, so we can be sure the interrupt
1356          * enable has made it to the chip.  If not equal, disable
1357          * interrupt again and return immediately.  This avoids races,
1358          * and the overhead of the chip read doesn't matter much at
1359          * this point, since we are waiting for something anyway.
1360          */
1361
1362         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1363                          dd->ipath_rcvctrl);
1364
1365         head = ipath_read_ureg32(dd, ur_rcvhdrhead, pd->port_port);
1366         tail = ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
1367
1368         if (tail == head) {
1369                 set_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag);
1370                 if (dd->ipath_rhdrhead_intr_off) /* arm rcv interrupt */
1371                         (void)ipath_write_ureg(dd, ur_rcvhdrhead,
1372                                                dd->ipath_rhdrhead_intr_off
1373                                                | head, pd->port_port);
1374                 poll_wait(fp, &pd->port_wait, pt);
1375
1376                 if (test_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag)) {
1377                         /* timed out, no packets received */
1378                         clear_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag);
1379                         pd->port_rcvwait_to++;
1380                 }
1381                 else
1382                         pollflag = POLLIN | POLLRDNORM;
1383         }
1384         else {
1385                 /* it's already happened; don't do wait_event overhead */
1386                 pollflag = POLLIN | POLLRDNORM;
1387                 pd->port_rcvnowait++;
1388         }
1389
1390         clear_bit(bit, &dd->ipath_rcvctrl);
1391         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1392                          dd->ipath_rcvctrl);
1393
1394 bail:
1395         return pollflag;
1396 }
1397
1398 static int init_subports(struct ipath_devdata *dd,
1399                          struct ipath_portdata *pd,
1400                          const struct ipath_user_info *uinfo)
1401 {
1402         int ret = 0;
1403         unsigned num_slaves;
1404         size_t size;
1405
1406         /* Old user binaries don't know about subports */
1407         if ((uinfo->spu_userversion & 0xffff) != IPATH_USER_SWMINOR)
1408                 goto bail;
1409         /*
1410          * If the user is requesting zero or one port,
1411          * skip the subport allocation.
1412          */
1413         if (uinfo->spu_subport_cnt <= 1)
1414                 goto bail;
1415         if (uinfo->spu_subport_cnt > 4) {
1416                 ret = -EINVAL;
1417                 goto bail;
1418         }
1419
1420         num_slaves = uinfo->spu_subport_cnt - 1;
1421         pd->subport_uregbase = vmalloc(PAGE_SIZE * num_slaves);
1422         if (!pd->subport_uregbase) {
1423                 ret = -ENOMEM;
1424                 goto bail;
1425         }
1426         /* Note: pd->port_rcvhdrq_size isn't initialized yet. */
1427         size = ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize *
1428                      sizeof(u32), PAGE_SIZE) * num_slaves;
1429         pd->subport_rcvhdr_base = vmalloc(size);
1430         if (!pd->subport_rcvhdr_base) {
1431                 ret = -ENOMEM;
1432                 goto bail_ureg;
1433         }
1434
1435         pd->subport_rcvegrbuf = vmalloc(pd->port_rcvegrbuf_chunks *
1436                                         pd->port_rcvegrbuf_size *
1437                                         num_slaves);
1438         if (!pd->subport_rcvegrbuf) {
1439                 ret = -ENOMEM;
1440                 goto bail_rhdr;
1441         }
1442
1443         pd->port_subport_cnt = uinfo->spu_subport_cnt;
1444         pd->port_subport_id = uinfo->spu_subport_id;
1445         pd->active_slaves = 1;
1446         goto bail;
1447
1448 bail_rhdr:
1449         vfree(pd->subport_rcvhdr_base);
1450 bail_ureg:
1451         vfree(pd->subport_uregbase);
1452         pd->subport_uregbase = NULL;
1453 bail:
1454         return ret;
1455 }
1456
1457 static int try_alloc_port(struct ipath_devdata *dd, int port,
1458                           struct file *fp,
1459                           const struct ipath_user_info *uinfo)
1460 {
1461         struct ipath_portdata *pd;
1462         int ret;
1463
1464         if (!(pd = dd->ipath_pd[port])) {
1465                 void *ptmp;
1466
1467                 pd = kzalloc(sizeof(struct ipath_portdata), GFP_KERNEL);
1468
1469                 /*
1470                  * Allocate memory for use in ipath_tid_update() just once
1471                  * at open, not per call.  Reduces cost of expected send
1472                  * setup.
1473                  */
1474                 ptmp = kmalloc(dd->ipath_rcvtidcnt * sizeof(u16) +
1475                                dd->ipath_rcvtidcnt * sizeof(struct page **),
1476                                GFP_KERNEL);
1477                 if (!pd || !ptmp) {
1478                         ipath_dev_err(dd, "Unable to allocate portdata "
1479                                       "memory, failing open\n");
1480                         ret = -ENOMEM;
1481                         kfree(pd);
1482                         kfree(ptmp);
1483                         goto bail;
1484                 }
1485                 dd->ipath_pd[port] = pd;
1486                 dd->ipath_pd[port]->port_port = port;
1487                 dd->ipath_pd[port]->port_dd = dd;
1488                 dd->ipath_pd[port]->port_tid_pg_list = ptmp;
1489                 init_waitqueue_head(&dd->ipath_pd[port]->port_wait);
1490         }
1491         if (!pd->port_cnt) {
1492                 pd->userversion = uinfo->spu_userversion;
1493                 init_user_egr_sizes(pd);
1494                 if ((ret = init_subports(dd, pd, uinfo)) != 0)
1495                         goto bail;
1496                 ipath_cdbg(PROC, "%s[%u] opened unit:port %u:%u\n",
1497                            current->comm, current->pid, dd->ipath_unit,
1498                            port);
1499                 pd->port_cnt = 1;
1500                 port_fp(fp) = pd;
1501                 pd->port_pid = current->pid;
1502                 strncpy(pd->port_comm, current->comm, sizeof(pd->port_comm));
1503                 ipath_stats.sps_ports++;
1504                 ret = 0;
1505         } else
1506                 ret = -EBUSY;
1507
1508 bail:
1509         return ret;
1510 }
1511
1512 static inline int usable(struct ipath_devdata *dd)
1513 {
1514         return dd &&
1515                 (dd->ipath_flags & IPATH_PRESENT) &&
1516                 dd->ipath_kregbase &&
1517                 dd->ipath_lid &&
1518                 !(dd->ipath_flags & (IPATH_LINKDOWN | IPATH_DISABLED
1519                                      | IPATH_LINKUNK));
1520 }
1521
1522 static int find_free_port(int unit, struct file *fp,
1523                           const struct ipath_user_info *uinfo)
1524 {
1525         struct ipath_devdata *dd = ipath_lookup(unit);
1526         int ret, i;
1527
1528         if (!dd) {
1529                 ret = -ENODEV;
1530                 goto bail;
1531         }
1532
1533         if (!usable(dd)) {
1534                 ret = -ENETDOWN;
1535                 goto bail;
1536         }
1537
1538         for (i = 1; i < dd->ipath_cfgports; i++) {
1539                 ret = try_alloc_port(dd, i, fp, uinfo);
1540                 if (ret != -EBUSY)
1541                         goto bail;
1542         }
1543         ret = -EBUSY;
1544
1545 bail:
1546         return ret;
1547 }
1548
1549 static int find_best_unit(struct file *fp,
1550                           const struct ipath_user_info *uinfo)
1551 {
1552         int ret = 0, i, prefunit = -1, devmax;
1553         int maxofallports, npresent, nup;
1554         int ndev;
1555
1556         devmax = ipath_count_units(&npresent, &nup, &maxofallports);
1557
1558         /*
1559          * This code is present to allow a knowledgeable person to
1560          * specify the layout of processes to processors before opening
1561          * this driver, and then we'll assign the process to the "closest"
1562          * InfiniPath chip to that processor (we assume reasonable connectivity,
1563          * for now).  This code assumes that if affinity has been set
1564          * before this point, that at most one cpu is set; for now this
1565          * is reasonable.  I check for both cpus_empty() and cpus_full(),
1566          * in case some kernel variant sets none of the bits when no
1567          * affinity is set.  2.6.11 and 12 kernels have all present
1568          * cpus set.  Some day we'll have to fix it up further to handle
1569          * a cpu subset.  This algorithm fails for two HT chips connected
1570          * in tunnel fashion.  Eventually this needs real topology
1571          * information.  There may be some issues with dual core numbering
1572          * as well.  This needs more work prior to release.
1573          */
1574         if (!cpus_empty(current->cpus_allowed) &&
1575             !cpus_full(current->cpus_allowed)) {
1576                 int ncpus = num_online_cpus(), curcpu = -1;
1577                 for (i = 0; i < ncpus; i++)
1578                         if (cpu_isset(i, current->cpus_allowed)) {
1579                                 ipath_cdbg(PROC, "%s[%u] affinity set for "
1580                                            "cpu %d\n", current->comm,
1581                                            current->pid, i);
1582                                 curcpu = i;
1583                         }
1584                 if (curcpu != -1) {
1585                         if (npresent) {
1586                                 prefunit = curcpu / (ncpus / npresent);
1587                                 ipath_dbg("%s[%u] %d chips, %d cpus, "
1588                                           "%d cpus/chip, select unit %d\n",
1589                                           current->comm, current->pid,
1590                                           npresent, ncpus, ncpus / npresent,
1591                                           prefunit);
1592                         }
1593                 }
1594         }
1595
1596         /*
1597          * user ports start at 1, kernel port is 0
1598          * For now, we do round-robin access across all chips
1599          */
1600
1601         if (prefunit != -1)
1602                 devmax = prefunit + 1;
1603 recheck:
1604         for (i = 1; i < maxofallports; i++) {
1605                 for (ndev = prefunit != -1 ? prefunit : 0; ndev < devmax;
1606                      ndev++) {
1607                         struct ipath_devdata *dd = ipath_lookup(ndev);
1608
1609                         if (!usable(dd))
1610                                 continue; /* can't use this unit */
1611                         if (i >= dd->ipath_cfgports)
1612                                 /*
1613                                  * Maxed out on users of this unit. Try
1614                                  * next.
1615                                  */
1616                                 continue;
1617                         ret = try_alloc_port(dd, i, fp, uinfo);
1618                         if (!ret)
1619                                 goto done;
1620                 }
1621         }
1622
1623         if (npresent) {
1624                 if (nup == 0) {
1625                         ret = -ENETDOWN;
1626                         ipath_dbg("No ports available (none initialized "
1627                                   "and ready)\n");
1628                 } else {
1629                         if (prefunit > 0) {
1630                                 /* if started above 0, retry from 0 */
1631                                 ipath_cdbg(PROC,
1632                                            "%s[%u] no ports on prefunit "
1633                                            "%d, clear and re-check\n",
1634                                            current->comm, current->pid,
1635                                            prefunit);
1636                                 devmax = ipath_count_units(NULL, NULL,
1637                                                            NULL);
1638                                 prefunit = -1;
1639                                 goto recheck;
1640                         }
1641                         ret = -EBUSY;
1642                         ipath_dbg("No ports available\n");
1643                 }
1644         } else {
1645                 ret = -ENXIO;
1646                 ipath_dbg("No boards found\n");
1647         }
1648
1649 done:
1650         return ret;
1651 }
1652
1653 static int find_shared_port(struct file *fp,
1654                             const struct ipath_user_info *uinfo)
1655 {
1656         int devmax, ndev, i;
1657         int ret = 0;
1658
1659         devmax = ipath_count_units(NULL, NULL, NULL);
1660
1661         for (ndev = 0; ndev < devmax; ndev++) {
1662                 struct ipath_devdata *dd = ipath_lookup(ndev);
1663
1664                 if (!dd)
1665                         continue;
1666                 for (i = 1; i < dd->ipath_cfgports; i++) {
1667                         struct ipath_portdata *pd = dd->ipath_pd[i];
1668
1669                         /* Skip ports which are not yet open */
1670                         if (!pd || !pd->port_cnt)
1671                                 continue;
1672                         /* Skip port if it doesn't match the requested one */
1673                         if (pd->port_subport_id != uinfo->spu_subport_id)
1674                                 continue;
1675                         /* Verify the sharing process matches the master */
1676                         if (pd->port_subport_cnt != uinfo->spu_subport_cnt ||
1677                             pd->userversion != uinfo->spu_userversion ||
1678                             pd->port_cnt >= pd->port_subport_cnt) {
1679                                 ret = -EINVAL;
1680                                 goto done;
1681                         }
1682                         port_fp(fp) = pd;
1683                         subport_fp(fp) = pd->port_cnt++;
1684                         tidcursor_fp(fp) = 0;
1685                         pd->active_slaves |= 1 << subport_fp(fp);
1686                         ipath_cdbg(PROC,
1687                                    "%s[%u] %u sharing %s[%u] unit:port %u:%u\n",
1688                                    current->comm, current->pid,
1689                                    subport_fp(fp),
1690                                    pd->port_comm, pd->port_pid,
1691                                    dd->ipath_unit, pd->port_port);
1692                         ret = 1;
1693                         goto done;
1694                 }
1695         }
1696
1697 done:
1698         return ret;
1699 }
1700
1701 static int ipath_open(struct inode *in, struct file *fp)
1702 {
1703         /* The real work is performed later in ipath_assign_port() */
1704         fp->private_data = kzalloc(sizeof(struct ipath_filedata), GFP_KERNEL);
1705         return fp->private_data ? 0 : -ENOMEM;
1706 }
1707
1708
1709 /* Get port early, so can set affinity prior to memory allocation */
1710 static int ipath_assign_port(struct file *fp,
1711                               const struct ipath_user_info *uinfo)
1712 {
1713         int ret;
1714         int i_minor;
1715         unsigned swminor;
1716
1717         /* Check to be sure we haven't already initialized this file */
1718         if (port_fp(fp)) {
1719                 ret = -EINVAL;
1720                 goto done;
1721         }
1722
1723         /* for now, if major version is different, bail */
1724         if ((uinfo->spu_userversion >> 16) != IPATH_USER_SWMAJOR) {
1725                 ipath_dbg("User major version %d not same as driver "
1726                           "major %d\n", uinfo->spu_userversion >> 16,
1727                           IPATH_USER_SWMAJOR);
1728                 ret = -ENODEV;
1729                 goto done;
1730         }
1731
1732         swminor = uinfo->spu_userversion & 0xffff;
1733         if (swminor != IPATH_USER_SWMINOR)
1734                 ipath_dbg("User minor version %d not same as driver "
1735                           "minor %d\n", swminor, IPATH_USER_SWMINOR);
1736
1737         mutex_lock(&ipath_mutex);
1738
1739         if (swminor == IPATH_USER_SWMINOR && uinfo->spu_subport_cnt &&
1740             (ret = find_shared_port(fp, uinfo))) {
1741                 mutex_unlock(&ipath_mutex);
1742                 if (ret > 0)
1743                         ret = 0;
1744                 goto done;
1745         }
1746
1747         i_minor = iminor(fp->f_path.dentry->d_inode) - IPATH_USER_MINOR_BASE;
1748         ipath_cdbg(VERBOSE, "open on dev %lx (minor %d)\n",
1749                    (long)fp->f_path.dentry->d_inode->i_rdev, i_minor);
1750
1751         if (i_minor)
1752                 ret = find_free_port(i_minor - 1, fp, uinfo);
1753         else
1754                 ret = find_best_unit(fp, uinfo);
1755
1756         mutex_unlock(&ipath_mutex);
1757
1758 done:
1759         return ret;
1760 }
1761
1762
1763 static int ipath_do_user_init(struct file *fp,
1764                               const struct ipath_user_info *uinfo)
1765 {
1766         int ret;
1767         struct ipath_portdata *pd;
1768         struct ipath_devdata *dd;
1769         u32 head32;
1770
1771         pd = port_fp(fp);
1772         dd = pd->port_dd;
1773
1774         if (uinfo->spu_rcvhdrsize) {
1775                 ret = ipath_setrcvhdrsize(dd, uinfo->spu_rcvhdrsize);
1776                 if (ret)
1777                         goto done;
1778         }
1779
1780         /* for now we do nothing with rcvhdrcnt: uinfo->spu_rcvhdrcnt */
1781
1782         /* for right now, kernel piobufs are at end, so port 1 is at 0 */
1783         pd->port_piobufs = dd->ipath_piobufbase +
1784                 dd->ipath_pbufsport * (pd->port_port - 1) * dd->ipath_palign;
1785         ipath_cdbg(VERBOSE, "Set base of piobufs for port %u to 0x%x\n",
1786                    pd->port_port, pd->port_piobufs);
1787
1788         /*
1789          * Now allocate the rcvhdr Q and eager TIDs; skip the TID
1790          * array for time being.  If pd->port_port > chip-supported,
1791          * we need to do extra stuff here to handle by handling overflow
1792          * through port 0, someday
1793          */
1794         ret = ipath_create_rcvhdrq(dd, pd);
1795         if (!ret)
1796                 ret = ipath_create_user_egr(pd);
1797         if (ret)
1798                 goto done;
1799
1800         /*
1801          * set the eager head register for this port to the current values
1802          * of the tail pointers, since we don't know if they were
1803          * updated on last use of the port.
1804          */
1805         head32 = ipath_read_ureg32(dd, ur_rcvegrindextail, pd->port_port);
1806         ipath_write_ureg(dd, ur_rcvegrindexhead, head32, pd->port_port);
1807         dd->ipath_lastegrheads[pd->port_port] = -1;
1808         dd->ipath_lastrcvhdrqtails[pd->port_port] = -1;
1809         ipath_cdbg(VERBOSE, "Wrote port%d egrhead %x from tail regs\n",
1810                 pd->port_port, head32);
1811         pd->port_tidcursor = 0; /* start at beginning after open */
1812         /*
1813          * now enable the port; the tail registers will be written to memory
1814          * by the chip as soon as it sees the write to
1815          * dd->ipath_kregs->kr_rcvctrl.  The update only happens on
1816          * transition from 0 to 1, so clear it first, then set it as part of
1817          * enabling the port.  This will (very briefly) affect any other
1818          * open ports, but it shouldn't be long enough to be an issue.
1819          * We explictly set the in-memory copy to 0 beforehand, so we don't
1820          * have to wait to be sure the DMA update has happened.
1821          */
1822         *(volatile u64 *)pd->port_rcvhdrtail_kvaddr = 0ULL;
1823         set_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port,
1824                 &dd->ipath_rcvctrl);
1825         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1826                          dd->ipath_rcvctrl & ~INFINIPATH_R_TAILUPD);
1827         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1828                          dd->ipath_rcvctrl);
1829 done:
1830         return ret;
1831 }
1832
1833 /**
1834  * unlock_exptid - unlock any expected TID entries port still had in use
1835  * @pd: port
1836  *
1837  * We don't actually update the chip here, because we do a bulk update
1838  * below, using ipath_f_clear_tids.
1839  */
1840 static void unlock_expected_tids(struct ipath_portdata *pd)
1841 {
1842         struct ipath_devdata *dd = pd->port_dd;
1843         int port_tidbase = pd->port_port * dd->ipath_rcvtidcnt;
1844         int i, cnt = 0, maxtid = port_tidbase + dd->ipath_rcvtidcnt;
1845
1846         ipath_cdbg(VERBOSE, "Port %u unlocking any locked expTID pages\n",
1847                    pd->port_port);
1848         for (i = port_tidbase; i < maxtid; i++) {
1849                 if (!dd->ipath_pageshadow[i])
1850                         continue;
1851
1852                 pci_unmap_page(dd->pcidev, dd->ipath_physshadow[i],
1853                         PAGE_SIZE, PCI_DMA_FROMDEVICE);
1854                 ipath_release_user_pages_on_close(&dd->ipath_pageshadow[i],
1855                                                   1);
1856                 dd->ipath_pageshadow[i] = NULL;
1857                 cnt++;
1858                 ipath_stats.sps_pageunlocks++;
1859         }
1860         if (cnt)
1861                 ipath_cdbg(VERBOSE, "Port %u locked %u expTID entries\n",
1862                            pd->port_port, cnt);
1863
1864         if (ipath_stats.sps_pagelocks || ipath_stats.sps_pageunlocks)
1865                 ipath_cdbg(VERBOSE, "%llu pages locked, %llu unlocked\n",
1866                            (unsigned long long) ipath_stats.sps_pagelocks,
1867                            (unsigned long long)
1868                            ipath_stats.sps_pageunlocks);
1869 }
1870
1871 static int ipath_close(struct inode *in, struct file *fp)
1872 {
1873         int ret = 0;
1874         struct ipath_filedata *fd;
1875         struct ipath_portdata *pd;
1876         struct ipath_devdata *dd;
1877         unsigned port;
1878
1879         ipath_cdbg(VERBOSE, "close on dev %lx, private data %p\n",
1880                    (long)in->i_rdev, fp->private_data);
1881
1882         mutex_lock(&ipath_mutex);
1883
1884         fd = (struct ipath_filedata *) fp->private_data;
1885         fp->private_data = NULL;
1886         pd = fd->pd;
1887         if (!pd) {
1888                 mutex_unlock(&ipath_mutex);
1889                 goto bail;
1890         }
1891         if (--pd->port_cnt) {
1892                 /*
1893                  * XXX If the master closes the port before the slave(s),
1894                  * revoke the mmap for the eager receive queue so
1895                  * the slave(s) don't wait for receive data forever.
1896                  */
1897                 pd->active_slaves &= ~(1 << fd->subport);
1898                 mutex_unlock(&ipath_mutex);
1899                 goto bail;
1900         }
1901         port = pd->port_port;
1902         dd = pd->port_dd;
1903
1904         if (pd->port_hdrqfull) {
1905                 ipath_cdbg(PROC, "%s[%u] had %u rcvhdrqfull errors "
1906                            "during run\n", pd->port_comm, pd->port_pid,
1907                            pd->port_hdrqfull);
1908                 pd->port_hdrqfull = 0;
1909         }
1910
1911         if (pd->port_rcvwait_to || pd->port_piowait_to
1912             || pd->port_rcvnowait || pd->port_pionowait) {
1913                 ipath_cdbg(VERBOSE, "port%u, %u rcv, %u pio wait timeo; "
1914                            "%u rcv %u, pio already\n",
1915                            pd->port_port, pd->port_rcvwait_to,
1916                            pd->port_piowait_to, pd->port_rcvnowait,
1917                            pd->port_pionowait);
1918                 pd->port_rcvwait_to = pd->port_piowait_to =
1919                         pd->port_rcvnowait = pd->port_pionowait = 0;
1920         }
1921         if (pd->port_flag) {
1922                 ipath_dbg("port %u port_flag still set to 0x%lx\n",
1923                           pd->port_port, pd->port_flag);
1924                 pd->port_flag = 0;
1925         }
1926
1927         if (dd->ipath_kregbase) {
1928                 int i;
1929                 /* atomically clear receive enable port. */
1930                 clear_bit(INFINIPATH_R_PORTENABLE_SHIFT + port,
1931                           &dd->ipath_rcvctrl);
1932                 ipath_write_kreg( dd, dd->ipath_kregs->kr_rcvctrl,
1933                         dd->ipath_rcvctrl);
1934                 /* and read back from chip to be sure that nothing
1935                  * else is in flight when we do the rest */
1936                 (void)ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
1937
1938                 /* clean up the pkeys for this port user */
1939                 ipath_clean_part_key(pd, dd);
1940                 /*
1941                  * be paranoid, and never write 0's to these, just use an
1942                  * unused part of the port 0 tail page.  Of course,
1943                  * rcvhdraddr points to a large chunk of memory, so this
1944                  * could still trash things, but at least it won't trash
1945                  * page 0, and by disabling the port, it should stop "soon",
1946                  * even if a packet or two is in already in flight after we
1947                  * disabled the port.
1948                  */
1949                 ipath_write_kreg_port(dd,
1950                         dd->ipath_kregs->kr_rcvhdrtailaddr, port,
1951                         dd->ipath_dummy_hdrq_phys);
1952                 ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdraddr,
1953                         pd->port_port, dd->ipath_dummy_hdrq_phys);
1954
1955                 i = dd->ipath_pbufsport * (port - 1);
1956                 ipath_disarm_piobufs(dd, i, dd->ipath_pbufsport);
1957
1958                 dd->ipath_f_clear_tids(dd, pd->port_port);
1959
1960                 if (dd->ipath_pageshadow)
1961                         unlock_expected_tids(pd);
1962                 ipath_stats.sps_ports--;
1963                 ipath_cdbg(PROC, "%s[%u] closed port %u:%u\n",
1964                            pd->port_comm, pd->port_pid,
1965                            dd->ipath_unit, port);
1966         }
1967
1968         pd->port_pid = 0;
1969         dd->ipath_pd[pd->port_port] = NULL; /* before releasing mutex */
1970         mutex_unlock(&ipath_mutex);
1971         ipath_free_pddata(dd, pd); /* after releasing the mutex */
1972
1973 bail:
1974         kfree(fd);
1975         return ret;
1976 }
1977
1978 static int ipath_port_info(struct ipath_portdata *pd, u16 subport,
1979                            struct ipath_port_info __user *uinfo)
1980 {
1981         struct ipath_port_info info;
1982         int nup;
1983         int ret;
1984         size_t sz;
1985
1986         (void) ipath_count_units(NULL, &nup, NULL);
1987         info.num_active = nup;
1988         info.unit = pd->port_dd->ipath_unit;
1989         info.port = pd->port_port;
1990         info.subport = subport;
1991         /* Don't return new fields if old library opened the port. */
1992         if ((pd->userversion & 0xffff) == IPATH_USER_SWMINOR) {
1993                 /* Number of user ports available for this device. */
1994                 info.num_ports = pd->port_dd->ipath_cfgports - 1;
1995                 info.num_subports = pd->port_subport_cnt;
1996                 sz = sizeof(info);
1997         } else
1998                 sz = sizeof(info) - 2 * sizeof(u16);
1999
2000         if (copy_to_user(uinfo, &info, sz)) {
2001                 ret = -EFAULT;
2002                 goto bail;
2003         }
2004         ret = 0;
2005
2006 bail:
2007         return ret;
2008 }
2009
2010 static int ipath_get_slave_info(struct ipath_portdata *pd,
2011                                 void __user *slave_mask_addr)
2012 {
2013         int ret = 0;
2014
2015         if (copy_to_user(slave_mask_addr, &pd->active_slaves, sizeof(u32)))
2016                 ret = -EFAULT;
2017         return ret;
2018 }
2019
2020 static ssize_t ipath_write(struct file *fp, const char __user *data,
2021                            size_t count, loff_t *off)
2022 {
2023         const struct ipath_cmd __user *ucmd;
2024         struct ipath_portdata *pd;
2025         const void __user *src;
2026         size_t consumed, copy;
2027         struct ipath_cmd cmd;
2028         ssize_t ret = 0;
2029         void *dest;
2030
2031         if (count < sizeof(cmd.type)) {
2032                 ret = -EINVAL;
2033                 goto bail;
2034         }
2035
2036         ucmd = (const struct ipath_cmd __user *) data;
2037
2038         if (copy_from_user(&cmd.type, &ucmd->type, sizeof(cmd.type))) {
2039                 ret = -EFAULT;
2040                 goto bail;
2041         }
2042
2043         consumed = sizeof(cmd.type);
2044
2045         switch (cmd.type) {
2046         case IPATH_CMD_ASSIGN_PORT:
2047         case __IPATH_CMD_USER_INIT:
2048         case IPATH_CMD_USER_INIT:
2049                 copy = sizeof(cmd.cmd.user_info);
2050                 dest = &cmd.cmd.user_info;
2051                 src = &ucmd->cmd.user_info;
2052                 break;
2053         case IPATH_CMD_RECV_CTRL:
2054                 copy = sizeof(cmd.cmd.recv_ctrl);
2055                 dest = &cmd.cmd.recv_ctrl;
2056                 src = &ucmd->cmd.recv_ctrl;
2057                 break;
2058         case IPATH_CMD_PORT_INFO:
2059                 copy = sizeof(cmd.cmd.port_info);
2060                 dest = &cmd.cmd.port_info;
2061                 src = &ucmd->cmd.port_info;
2062                 break;
2063         case IPATH_CMD_TID_UPDATE:
2064         case IPATH_CMD_TID_FREE:
2065                 copy = sizeof(cmd.cmd.tid_info);
2066                 dest = &cmd.cmd.tid_info;
2067                 src = &ucmd->cmd.tid_info;
2068                 break;
2069         case IPATH_CMD_SET_PART_KEY:
2070                 copy = sizeof(cmd.cmd.part_key);
2071                 dest = &cmd.cmd.part_key;
2072                 src = &ucmd->cmd.part_key;
2073                 break;
2074         case IPATH_CMD_SLAVE_INFO:
2075                 copy = sizeof(cmd.cmd.slave_mask_addr);
2076                 dest = &cmd.cmd.slave_mask_addr;
2077                 src = &ucmd->cmd.slave_mask_addr;
2078                 break;
2079         default:
2080                 ret = -EINVAL;
2081                 goto bail;
2082         }
2083
2084         if ((count - consumed) < copy) {
2085                 ret = -EINVAL;
2086                 goto bail;
2087         }
2088
2089         if (copy_from_user(dest, src, copy)) {
2090                 ret = -EFAULT;
2091                 goto bail;
2092         }
2093
2094         consumed += copy;
2095         pd = port_fp(fp);
2096         if (!pd && cmd.type != __IPATH_CMD_USER_INIT &&
2097                 cmd.type != IPATH_CMD_ASSIGN_PORT) {
2098                 ret = -EINVAL;
2099                 goto bail;
2100         }
2101
2102         switch (cmd.type) {
2103         case IPATH_CMD_ASSIGN_PORT:
2104                 ret = ipath_assign_port(fp, &cmd.cmd.user_info);
2105                 if (ret)
2106                         goto bail;
2107                 break;
2108         case __IPATH_CMD_USER_INIT:
2109                 /* backwards compatibility, get port first */
2110                 ret = ipath_assign_port(fp, &cmd.cmd.user_info);
2111                 if (ret)
2112                         goto bail;
2113                 /* and fall through to current version. */
2114         case IPATH_CMD_USER_INIT:
2115                 ret = ipath_do_user_init(fp, &cmd.cmd.user_info);
2116                 if (ret)
2117                         goto bail;
2118                 ret = ipath_get_base_info(
2119                         fp, (void __user *) (unsigned long)
2120                         cmd.cmd.user_info.spu_base_info,
2121                         cmd.cmd.user_info.spu_base_info_size);
2122                 break;
2123         case IPATH_CMD_RECV_CTRL:
2124                 ret = ipath_manage_rcvq(pd, subport_fp(fp), cmd.cmd.recv_ctrl);
2125                 break;
2126         case IPATH_CMD_PORT_INFO:
2127                 ret = ipath_port_info(pd, subport_fp(fp),
2128                                       (struct ipath_port_info __user *)
2129                                       (unsigned long) cmd.cmd.port_info);
2130                 break;
2131         case IPATH_CMD_TID_UPDATE:
2132                 ret = ipath_tid_update(pd, fp, &cmd.cmd.tid_info);
2133                 break;
2134         case IPATH_CMD_TID_FREE:
2135                 ret = ipath_tid_free(pd, subport_fp(fp), &cmd.cmd.tid_info);
2136                 break;
2137         case IPATH_CMD_SET_PART_KEY:
2138                 ret = ipath_set_part_key(pd, cmd.cmd.part_key);
2139                 break;
2140         case IPATH_CMD_SLAVE_INFO:
2141                 ret = ipath_get_slave_info(pd,
2142                                            (void __user *) (unsigned long)
2143                                            cmd.cmd.slave_mask_addr);
2144                 break;
2145         }
2146
2147         if (ret >= 0)
2148                 ret = consumed;
2149
2150 bail:
2151         return ret;
2152 }
2153
2154 static struct class *ipath_class;
2155
2156 static int init_cdev(int minor, char *name, const struct file_operations *fops,
2157                      struct cdev **cdevp, struct class_device **class_devp)
2158 {
2159         const dev_t dev = MKDEV(IPATH_MAJOR, minor);
2160         struct cdev *cdev = NULL;
2161         struct class_device *class_dev = NULL;
2162         int ret;
2163
2164         cdev = cdev_alloc();
2165         if (!cdev) {
2166                 printk(KERN_ERR IPATH_DRV_NAME
2167                        ": Could not allocate cdev for minor %d, %s\n",
2168                        minor, name);
2169                 ret = -ENOMEM;
2170                 goto done;
2171         }
2172
2173         cdev->owner = THIS_MODULE;
2174         cdev->ops = fops;
2175         kobject_set_name(&cdev->kobj, name);
2176
2177         ret = cdev_add(cdev, dev, 1);
2178         if (ret < 0) {
2179                 printk(KERN_ERR IPATH_DRV_NAME
2180                        ": Could not add cdev for minor %d, %s (err %d)\n",
2181                        minor, name, -ret);
2182                 goto err_cdev;
2183         }
2184
2185         class_dev = class_device_create(ipath_class, NULL, dev, NULL, name);
2186
2187         if (IS_ERR(class_dev)) {
2188                 ret = PTR_ERR(class_dev);
2189                 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2190                        "class_dev for minor %d, %s (err %d)\n",
2191                        minor, name, -ret);
2192                 goto err_cdev;
2193         }
2194
2195         goto done;
2196
2197 err_cdev:
2198         cdev_del(cdev);
2199         cdev = NULL;
2200
2201 done:
2202         if (ret >= 0) {
2203                 *cdevp = cdev;
2204                 *class_devp = class_dev;
2205         } else {
2206                 *cdevp = NULL;
2207                 *class_devp = NULL;
2208         }
2209
2210         return ret;
2211 }
2212
2213 int ipath_cdev_init(int minor, char *name, const struct file_operations *fops,
2214                     struct cdev **cdevp, struct class_device **class_devp)
2215 {
2216         return init_cdev(minor, name, fops, cdevp, class_devp);
2217 }
2218
2219 static void cleanup_cdev(struct cdev **cdevp,
2220                          struct class_device **class_devp)
2221 {
2222         struct class_device *class_dev = *class_devp;
2223
2224         if (class_dev) {
2225                 class_device_unregister(class_dev);
2226                 *class_devp = NULL;
2227         }
2228
2229         if (*cdevp) {
2230                 cdev_del(*cdevp);
2231                 *cdevp = NULL;
2232         }
2233 }
2234
2235 void ipath_cdev_cleanup(struct cdev **cdevp,
2236                         struct class_device **class_devp)
2237 {
2238         cleanup_cdev(cdevp, class_devp);
2239 }
2240
2241 static struct cdev *wildcard_cdev;
2242 static struct class_device *wildcard_class_dev;
2243
2244 static const dev_t dev = MKDEV(IPATH_MAJOR, 0);
2245
2246 static int user_init(void)
2247 {
2248         int ret;
2249
2250         ret = register_chrdev_region(dev, IPATH_NMINORS, IPATH_DRV_NAME);
2251         if (ret < 0) {
2252                 printk(KERN_ERR IPATH_DRV_NAME ": Could not register "
2253                        "chrdev region (err %d)\n", -ret);
2254                 goto done;
2255         }
2256
2257         ipath_class = class_create(THIS_MODULE, IPATH_DRV_NAME);
2258
2259         if (IS_ERR(ipath_class)) {
2260                 ret = PTR_ERR(ipath_class);
2261                 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2262                        "device class (err %d)\n", -ret);
2263                 goto bail;
2264         }
2265
2266         goto done;
2267 bail:
2268         unregister_chrdev_region(dev, IPATH_NMINORS);
2269 done:
2270         return ret;
2271 }
2272
2273 static void user_cleanup(void)
2274 {
2275         if (ipath_class) {
2276                 class_destroy(ipath_class);
2277                 ipath_class = NULL;
2278         }
2279
2280         unregister_chrdev_region(dev, IPATH_NMINORS);
2281 }
2282
2283 static atomic_t user_count = ATOMIC_INIT(0);
2284 static atomic_t user_setup = ATOMIC_INIT(0);
2285
2286 int ipath_user_add(struct ipath_devdata *dd)
2287 {
2288         char name[10];
2289         int ret;
2290
2291         if (atomic_inc_return(&user_count) == 1) {
2292                 ret = user_init();
2293                 if (ret < 0) {
2294                         ipath_dev_err(dd, "Unable to set up user support: "
2295                                       "error %d\n", -ret);
2296                         goto bail;
2297                 }
2298                 ret = init_cdev(0, "ipath", &ipath_file_ops, &wildcard_cdev,
2299                                 &wildcard_class_dev);
2300                 if (ret < 0) {
2301                         ipath_dev_err(dd, "Could not create wildcard "
2302                                       "minor: error %d\n", -ret);
2303                         goto bail_user;
2304                 }
2305
2306                 atomic_set(&user_setup, 1);
2307         }
2308
2309         snprintf(name, sizeof(name), "ipath%d", dd->ipath_unit);
2310
2311         ret = init_cdev(dd->ipath_unit + 1, name, &ipath_file_ops,
2312                         &dd->user_cdev, &dd->user_class_dev);
2313         if (ret < 0)
2314                 ipath_dev_err(dd, "Could not create user minor %d, %s\n",
2315                               dd->ipath_unit + 1, name);
2316
2317         goto bail;
2318
2319 bail_user:
2320         user_cleanup();
2321 bail:
2322         return ret;
2323 }
2324
2325 void ipath_user_remove(struct ipath_devdata *dd)
2326 {
2327         cleanup_cdev(&dd->user_cdev, &dd->user_class_dev);
2328
2329         if (atomic_dec_return(&user_count) == 0) {
2330                 if (atomic_read(&user_setup) == 0)
2331                         goto bail;
2332
2333                 cleanup_cdev(&wildcard_cdev, &wildcard_class_dev);
2334                 user_cleanup();
2335
2336                 atomic_set(&user_setup, 0);
2337         }
2338 bail:
2339         return;
2340 }