drivers/infiniband/hw/ipath/ipath_file_ops.c

   1 /*
   2  * Copyright (c) 2006, 2007, 2008 QLogic Corporation. All rights reserved.
   3  * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
   4  *
   5  * This software is available to you under a choice of one of two
   6  * licenses.  You may choose to be licensed under the terms of the GNU
   7  * General Public License (GPL) Version 2, available from the file
   8  * COPYING in the main directory of this source tree, or the
   9  * OpenIB.org BSD license below:
  10  *
  11  *     Redistribution and use in source and binary forms, with or
  12  *     without modification, are permitted provided that the following
  13  *     conditions are met:
  14  *
  15  *      - Redistributions of source code must retain the above
  16  *        copyright notice, this list of conditions and the following
  17  *        disclaimer.
  18  *
  19  *      - Redistributions in binary form must reproduce the above
  20  *        copyright notice, this list of conditions and the following
  21  *        disclaimer in the documentation and/or other materials
  22  *        provided with the distribution.
  23  *
  24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  31  * SOFTWARE.
  32  */
  33
  34 #include <linux/pci.h>
  35 #include <linux/poll.h>
  36 #include <linux/cdev.h>
  37 #include <linux/swap.h>
  38 #include <linux/vmalloc.h>
  39 #include <linux/highmem.h>
  40 #include <linux/io.h>
  41 #include <linux/jiffies.h>
  42 #include <asm/pgtable.h>
  43
  44 #include "ipath_kernel.h"
  45 #include "ipath_common.h"
  46 #include "ipath_user_sdma.h"
  47
  48 static int ipath_open(struct inode *, struct file *);
  49 static int ipath_close(struct inode *, struct file *);
  50 static ssize_t ipath_write(struct file *, const char __user *, size_t,
  51                            loff_t *);
  52 static ssize_t ipath_writev(struct kiocb *, const struct iovec *,
  53                             unsigned long , loff_t);
  54 static unsigned int ipath_poll(struct file *, struct poll_table_struct *);
  55 static int ipath_mmap(struct file *, struct vm_area_struct *);
  56
  57 static const struct file_operations ipath_file_ops = {
  58         .owner = THIS_MODULE,
  59         .write = ipath_write,
  60         .aio_write = ipath_writev,
  61         .open = ipath_open,
  62         .release = ipath_close,
  63         .poll = ipath_poll,
  64         .mmap = ipath_mmap
  65 };
  66
  67 /*
  68  * Convert kernel virtual addresses to physical addresses so they don't
  69  * potentially conflict with the chip addresses used as mmap offsets.
  70  * It doesn't really matter what mmap offset we use as long as we can
  71  * interpret it correctly.
  72  */
  73 static u64 cvt_kvaddr(void *p)
  74 {
  75         struct page *page;
  76         u64 paddr = 0;
  77
  78         page = vmalloc_to_page(p);
  79         if (page)
  80                 paddr = page_to_pfn(page) << PAGE_SHIFT;
  81
  82         return paddr;
  83 }
  84
  85 static int ipath_get_base_info(struct file *fp,
  86                                void __user *ubase, size_t ubase_size)
  87 {
  88         struct ipath_portdata *pd = port_fp(fp);
  89         int ret = 0;
  90         struct ipath_base_info *kinfo = NULL;
  91         struct ipath_devdata *dd = pd->port_dd;
  92         unsigned subport_cnt;
  93         int shared, master;
  94         size_t sz;
  95
  96         subport_cnt = pd->port_subport_cnt;
  97         if (!subport_cnt) {
  98                 shared = 0;
  99                 master = 0;
 100                 subport_cnt = 1;
 101         } else {
 102                 shared = 1;
 103                 master = !subport_fp(fp);
 104         }
 105
 106         sz = sizeof(*kinfo);
 107         /* If port sharing is not requested, allow the old size structure */
 108         if (!shared)
 109                 sz -= 7 * sizeof(u64);
 110         if (ubase_size < sz) {
 111                 ipath_cdbg(PROC,
 112                            "Base size %zu, need %zu (version mismatch?)\n",
 113                            ubase_size, sz);
 114                 ret = -EINVAL;
 115                 goto bail;
 116         }
 117
 118         kinfo = kzalloc(sizeof(*kinfo), GFP_KERNEL);
 119         if (kinfo == NULL) {
 120                 ret = -ENOMEM;
 121                 goto bail;
 122         }
 123
 124         ret = dd->ipath_f_get_base_info(pd, kinfo);
 125         if (ret < 0)
 126                 goto bail;
 127
 128         kinfo->spi_rcvhdr_cnt = dd->ipath_rcvhdrcnt;
 129         kinfo->spi_rcvhdrent_size = dd->ipath_rcvhdrentsize;
 130         kinfo->spi_tidegrcnt = dd->ipath_rcvegrcnt;
 131         kinfo->spi_rcv_egrbufsize = dd->ipath_rcvegrbufsize;
 132         /*
 133          * have to mmap whole thing
 134          */
 135         kinfo->spi_rcv_egrbuftotlen =
 136                 pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
 137         kinfo->spi_rcv_egrperchunk = pd->port_rcvegrbufs_perchunk;
 138         kinfo->spi_rcv_egrchunksize = kinfo->spi_rcv_egrbuftotlen /
 139                 pd->port_rcvegrbuf_chunks;
 140         kinfo->spi_tidcnt = dd->ipath_rcvtidcnt / subport_cnt;
 141         if (master)
 142                 kinfo->spi_tidcnt += dd->ipath_rcvtidcnt % subport_cnt;
 143         /*
 144          * for this use, may be ipath_cfgports summed over all chips that
 145          * are are configured and present
 146          */
 147         kinfo->spi_nports = dd->ipath_cfgports;
 148         /* unit (chip/board) our port is on */
 149         kinfo->spi_unit = dd->ipath_unit;
 150         /* for now, only a single page */
 151         kinfo->spi_tid_maxsize = PAGE_SIZE;
 152
 153         /*
 154          * Doing this per port, and based on the skip value, etc.  This has
 155          * to be the actual buffer size, since the protocol code treats it
 156          * as an array.
 157          *
 158          * These have to be set to user addresses in the user code via mmap.
 159          * These values are used on return to user code for the mmap target
 160          * addresses only.  For 32 bit, same 44 bit address problem, so use
 161          * the physical address, not virtual.  Before 2.6.11, using the
 162          * page_address() macro worked, but in 2.6.11, even that returns the
 163          * full 64 bit address (upper bits all 1's).  So far, using the
 164          * physical addresses (or chip offsets, for chip mapping) works, but
 165          * no doubt some future kernel release will change that, and we'll be
 166          * on to yet another method of dealing with this.
 167          */
 168         kinfo->spi_rcvhdr_base = (u64) pd->port_rcvhdrq_phys;
 169         kinfo->spi_rcvhdr_tailaddr = (u64) pd->port_rcvhdrqtailaddr_phys;
 170         kinfo->spi_rcv_egrbufs = (u64) pd->port_rcvegr_phys;
 171         kinfo->spi_pioavailaddr = (u64) dd->ipath_pioavailregs_phys;
 172         kinfo->spi_status = (u64) kinfo->spi_pioavailaddr +
 173                 (void *) dd->ipath_statusp -
 174                 (void *) dd->ipath_pioavailregs_dma;
 175         if (!shared) {
 176                 kinfo->spi_piocnt = pd->port_piocnt;
 177                 kinfo->spi_piobufbase = (u64) pd->port_piobufs;
 178                 kinfo->__spi_uregbase = (u64) dd->ipath_uregbase +
 179                         dd->ipath_ureg_align * pd->port_port;
 180         } else if (master) {
 181                 kinfo->spi_piocnt = (pd->port_piocnt / subport_cnt) +
 182                                     (pd->port_piocnt % subport_cnt);
 183                 /* Master's PIO buffers are after all the slave's */
 184                 kinfo->spi_piobufbase = (u64) pd->port_piobufs +
 185                         dd->ipath_palign *
 186                         (pd->port_piocnt - kinfo->spi_piocnt);
 187         } else {
 188                 unsigned slave = subport_fp(fp) - 1;
 189
 190                 kinfo->spi_piocnt = pd->port_piocnt / subport_cnt;
 191                 kinfo->spi_piobufbase = (u64) pd->port_piobufs +
 192                         dd->ipath_palign * kinfo->spi_piocnt * slave;
 193         }
 194
 195         if (shared) {
 196                 kinfo->spi_port_uregbase = (u64) dd->ipath_uregbase +
 197                         dd->ipath_ureg_align * pd->port_port;
 198                 kinfo->spi_port_rcvegrbuf = kinfo->spi_rcv_egrbufs;
 199                 kinfo->spi_port_rcvhdr_base = kinfo->spi_rcvhdr_base;
 200                 kinfo->spi_port_rcvhdr_tailaddr = kinfo->spi_rcvhdr_tailaddr;
 201
 202                 kinfo->__spi_uregbase = cvt_kvaddr(pd->subport_uregbase +
 203                         PAGE_SIZE * subport_fp(fp));
 204
 205                 kinfo->spi_rcvhdr_base = cvt_kvaddr(pd->subport_rcvhdr_base +
 206                         pd->port_rcvhdrq_size * subport_fp(fp));
 207                 kinfo->spi_rcvhdr_tailaddr = 0;
 208                 kinfo->spi_rcv_egrbufs = cvt_kvaddr(pd->subport_rcvegrbuf +
 209                         pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size *
 210                         subport_fp(fp));
 211
 212                 kinfo->spi_subport_uregbase =
 213                         cvt_kvaddr(pd->subport_uregbase);
 214                 kinfo->spi_subport_rcvegrbuf =
 215                         cvt_kvaddr(pd->subport_rcvegrbuf);
 216                 kinfo->spi_subport_rcvhdr_base =
 217                         cvt_kvaddr(pd->subport_rcvhdr_base);
 218                 ipath_cdbg(PROC, "port %u flags %x %llx %llx %llx\n",
 219                         kinfo->spi_port, kinfo->spi_runtime_flags,
 220                         (unsigned long long) kinfo->spi_subport_uregbase,
 221                         (unsigned long long) kinfo->spi_subport_rcvegrbuf,
 222                         (unsigned long long) kinfo->spi_subport_rcvhdr_base);
 223         }
 224
 225         kinfo->spi_pioindex = (kinfo->spi_piobufbase - dd->ipath_piobufbase) /
 226                 dd->ipath_palign;
 227         kinfo->spi_pioalign = dd->ipath_palign;
 228
 229         kinfo->spi_qpair = IPATH_KD_QP;
 230         /*
 231          * user mode PIO buffers are always 2KB, even when 4KB can
 232          * be received, and sent via the kernel; this is ibmaxlen
 233          * for 2K MTU.
 234          */
 235         kinfo->spi_piosize = dd->ipath_piosize2k - 2 * sizeof(u32);
 236         kinfo->spi_mtu = dd->ipath_ibmaxlen;    /* maxlen, not ibmtu */
 237         kinfo->spi_port = pd->port_port;
 238         kinfo->spi_subport = subport_fp(fp);
 239         kinfo->spi_sw_version = IPATH_KERN_SWVERSION;
 240         kinfo->spi_hw_version = dd->ipath_revision;
 241
 242         if (master) {
 243                 kinfo->spi_runtime_flags |= IPATH_RUNTIME_MASTER;
 244         }
 245
 246         sz = (ubase_size < sizeof(*kinfo)) ? ubase_size : sizeof(*kinfo);
 247         if (copy_to_user(ubase, kinfo, sz))
 248                 ret = -EFAULT;
 249
 250 bail:
 251         kfree(kinfo);
 252         return ret;
 253 }
 254
 255 /**
 256  * ipath_tid_update - update a port TID
 257  * @pd: the port
 258  * @fp: the ipath device file
 259  * @ti: the TID information
 260  *
 261  * The new implementation as of Oct 2004 is that the driver assigns
 262  * the tid and returns it to the caller.   To make it easier to
 263  * catch bugs, and to reduce search time, we keep a cursor for
 264  * each port, walking the shadow tid array to find one that's not
 265  * in use.
 266  *
 267  * For now, if we can't allocate the full list, we fail, although
 268  * in the long run, we'll allocate as many as we can, and the
 269  * caller will deal with that by trying the remaining pages later.
 270  * That means that when we fail, we have to mark the tids as not in
 271  * use again, in our shadow copy.
 272  *
 273  * It's up to the caller to free the tids when they are done.
 274  * We'll unlock the pages as they free them.
 275  *
 276  * Also, right now we are locking one page at a time, but since
 277  * the intended use of this routine is for a single group of
 278  * virtually contiguous pages, that should change to improve
 279  * performance.
 280  */
 281 static int ipath_tid_update(struct ipath_portdata *pd, struct file *fp,
 282                             const struct ipath_tid_info *ti)
 283 {
 284         int ret = 0, ntids;
 285         u32 tid, porttid, cnt, i, tidcnt, tidoff;
 286         u16 *tidlist;
 287         struct ipath_devdata *dd = pd->port_dd;
 288         u64 physaddr;
 289         unsigned long vaddr;
 290         u64 __iomem *tidbase;
 291         unsigned long tidmap[8];
 292         struct page **pagep = NULL;
 293         unsigned subport = subport_fp(fp);
 294
 295         if (!dd->ipath_pageshadow) {
 296                 ret = -ENOMEM;
 297                 goto done;
 298         }
 299
 300         cnt = ti->tidcnt;
 301         if (!cnt) {
 302                 ipath_dbg("After copyin, tidcnt 0, tidlist %llx\n",
 303                           (unsigned long long) ti->tidlist);
 304                 /*
 305                  * Should we treat as success?  likely a bug
 306                  */
 307                 ret = -EFAULT;
 308                 goto done;
 309         }
 310         porttid = pd->port_port * dd->ipath_rcvtidcnt;
 311         if (!pd->port_subport_cnt) {
 312                 tidcnt = dd->ipath_rcvtidcnt;
 313                 tid = pd->port_tidcursor;
 314                 tidoff = 0;
 315         } else if (!subport) {
 316                 tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
 317                          (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
 318                 tidoff = dd->ipath_rcvtidcnt - tidcnt;
 319                 porttid += tidoff;
 320                 tid = tidcursor_fp(fp);
 321         } else {
 322                 tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
 323                 tidoff = tidcnt * (subport - 1);
 324                 porttid += tidoff;
 325                 tid = tidcursor_fp(fp);
 326         }
 327         if (cnt > tidcnt) {
 328                 /* make sure it all fits in port_tid_pg_list */
 329                 dev_info(&dd->pcidev->dev, "Process tried to allocate %u "
 330                          "TIDs, only trying max (%u)\n", cnt, tidcnt);
 331                 cnt = tidcnt;
 332         }
 333         pagep = &((struct page **) pd->port_tid_pg_list)[tidoff];
 334         tidlist = &((u16 *) &pagep[dd->ipath_rcvtidcnt])[tidoff];
 335
 336         memset(tidmap, 0, sizeof(tidmap));
 337         /* before decrement; chip actual # */
 338         ntids = tidcnt;
 339         tidbase = (u64 __iomem *) (((char __iomem *) dd->ipath_kregbase) +
 340                                    dd->ipath_rcvtidbase +
 341                                    porttid * sizeof(*tidbase));
 342
 343         ipath_cdbg(VERBOSE, "Port%u %u tids, cursor %u, tidbase %p\n",
 344                    pd->port_port, cnt, tid, tidbase);
 345
 346         /* virtual address of first page in transfer */
 347         vaddr = ti->tidvaddr;
 348         if (!access_ok(VERIFY_WRITE, (void __user *) vaddr,
 349                        cnt * PAGE_SIZE)) {
 350                 ipath_dbg("Fail vaddr %p, %u pages, !access_ok\n",
 351                           (void *)vaddr, cnt);
 352                 ret = -EFAULT;
 353                 goto done;
 354         }
 355         ret = ipath_get_user_pages(vaddr, cnt, pagep);
 356         if (ret) {
 357                 if (ret == -EBUSY) {
 358                         ipath_dbg("Failed to lock addr %p, %u pages "
 359                                   "(already locked)\n",
 360                                   (void *) vaddr, cnt);
 361                         /*
 362                          * for now, continue, and see what happens but with
 363                          * the new implementation, this should never happen,
 364                          * unless perhaps the user has mpin'ed the pages
 365                          * themselves (something we need to test)
 366                          */
 367                         ret = 0;
 368                 } else {
 369                         dev_info(&dd->pcidev->dev,
 370                                  "Failed to lock addr %p, %u pages: "
 371                                  "errno %d\n", (void *) vaddr, cnt, -ret);
 372                         goto done;
 373                 }
 374         }
 375         for (i = 0; i < cnt; i++, vaddr += PAGE_SIZE) {
 376                 for (; ntids--; tid++) {
 377                         if (tid == tidcnt)
 378                                 tid = 0;
 379                         if (!dd->ipath_pageshadow[porttid + tid])
 380                                 break;
 381                 }
 382                 if (ntids < 0) {
 383                         /*
 384                          * oops, wrapped all the way through their TIDs,
 385                          * and didn't have enough free; see comments at
 386                          * start of routine
 387                          */
 388                         ipath_dbg("Not enough free TIDs for %u pages "
 389                                   "(index %d), failing\n", cnt, i);
 390                         i--;    /* last tidlist[i] not filled in */
 391                         ret = -ENOMEM;
 392                         break;
 393                 }
 394                 tidlist[i] = tid + tidoff;
 395                 ipath_cdbg(VERBOSE, "Updating idx %u to TID %u, "
 396                            "vaddr %lx\n", i, tid + tidoff, vaddr);
 397                 /* we "know" system pages and TID pages are same size */
 398                 dd->ipath_pageshadow[porttid + tid] = pagep[i];
 399                 dd->ipath_physshadow[porttid + tid] = ipath_map_page(
 400                         dd->pcidev, pagep[i], 0, PAGE_SIZE,
 401                         PCI_DMA_FROMDEVICE);
 402                 /*
 403                  * don't need atomic or it's overhead
 404                  */
 405                 __set_bit(tid, tidmap);
 406                 physaddr = dd->ipath_physshadow[porttid + tid];
 407                 ipath_stats.sps_pagelocks++;
 408                 ipath_cdbg(VERBOSE,
 409                            "TID %u, vaddr %lx, physaddr %llx pgp %p\n",
 410                            tid, vaddr, (unsigned long long) physaddr,
 411                            pagep[i]);
 412                 dd->ipath_f_put_tid(dd, &tidbase[tid], RCVHQ_RCV_TYPE_EXPECTED,
 413                                     physaddr);
 414                 /*
 415                  * don't check this tid in ipath_portshadow, since we
 416                  * just filled it in; start with the next one.
 417                  */
 418                 tid++;
 419         }
 420
 421         if (ret) {
 422                 u32 limit;
 423         cleanup:
 424                 /* jump here if copy out of updated info failed... */
 425                 ipath_dbg("After failure (ret=%d), undo %d of %d entries\n",
 426                           -ret, i, cnt);
 427                 /* same code that's in ipath_free_tid() */
 428                 limit = sizeof(tidmap) * BITS_PER_BYTE;
 429                 if (limit > tidcnt)
 430                         /* just in case size changes in future */
 431                         limit = tidcnt;
 432                 tid = find_first_bit((const unsigned long *)tidmap, limit);
 433                 for (; tid < limit; tid++) {
 434                         if (!test_bit(tid, tidmap))
 435                                 continue;
 436                         if (dd->ipath_pageshadow[porttid + tid]) {
 437                                 ipath_cdbg(VERBOSE, "Freeing TID %u\n",
 438                                            tid);
 439                                 dd->ipath_f_put_tid(dd, &tidbase[tid],
 440                                                     RCVHQ_RCV_TYPE_EXPECTED,
 441                                                     dd->ipath_tidinvalid);
 442                                 pci_unmap_page(dd->pcidev,
 443                                         dd->ipath_physshadow[porttid + tid],
 444                                         PAGE_SIZE, PCI_DMA_FROMDEVICE);
 445                                 dd->ipath_pageshadow[porttid + tid] = NULL;
 446                                 ipath_stats.sps_pageunlocks++;
 447                         }
 448                 }
 449                 ipath_release_user_pages(pagep, cnt);
 450         } else {
 451                 /*
 452                  * Copy the updated array, with ipath_tid's filled in, back
 453                  * to user.  Since we did the copy in already, this "should
 454                  * never fail" If it does, we have to clean up...
 455                  */
 456                 if (copy_to_user((void __user *)
 457                                  (unsigned long) ti->tidlist,
 458                                  tidlist, cnt * sizeof(*tidlist))) {
 459                         ret = -EFAULT;
 460                         goto cleanup;
 461                 }
 462                 if (copy_to_user((void __user *) (unsigned long) ti->tidmap,
 463                                  tidmap, sizeof tidmap)) {
 464                         ret = -EFAULT;
 465                         goto cleanup;
 466                 }
 467                 if (tid == tidcnt)
 468                         tid = 0;
 469                 if (!pd->port_subport_cnt)
 470                         pd->port_tidcursor = tid;
 471                 else
 472                         tidcursor_fp(fp) = tid;
 473         }
 474
 475 done:
 476         if (ret)
 477                 ipath_dbg("Failed to map %u TID pages, failing with %d\n",
 478                           ti->tidcnt, -ret);
 479         return ret;
 480 }
 481
 482 /**
 483  * ipath_tid_free - free a port TID
 484  * @pd: the port
 485  * @subport: the subport
 486  * @ti: the TID info
 487  *
 488  * right now we are unlocking one page at a time, but since
 489  * the intended use of this routine is for a single group of
 490  * virtually contiguous pages, that should change to improve
 491  * performance.  We check that the TID is in range for this port
 492  * but otherwise don't check validity; if user has an error and
 493  * frees the wrong tid, it's only their own data that can thereby
 494  * be corrupted.  We do check that the TID was in use, for sanity
 495  * We always use our idea of the saved address, not the address that
 496  * they pass in to us.
 497  */
 498
 499 static int ipath_tid_free(struct ipath_portdata *pd, unsigned subport,
 500                           const struct ipath_tid_info *ti)
 501 {
 502         int ret = 0;
 503         u32 tid, porttid, cnt, limit, tidcnt;
 504         struct ipath_devdata *dd = pd->port_dd;
 505         u64 __iomem *tidbase;
 506         unsigned long tidmap[8];
 507
 508         if (!dd->ipath_pageshadow) {
 509                 ret = -ENOMEM;
 510                 goto done;
 511         }
 512
 513         if (copy_from_user(tidmap, (void __user *)(unsigned long)ti->tidmap,
 514                            sizeof tidmap)) {
 515                 ret = -EFAULT;
 516                 goto done;
 517         }
 518
 519         porttid = pd->port_port * dd->ipath_rcvtidcnt;
 520         if (!pd->port_subport_cnt)
 521                 tidcnt = dd->ipath_rcvtidcnt;
 522         else if (!subport) {
 523                 tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
 524                          (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
 525                 porttid += dd->ipath_rcvtidcnt - tidcnt;
 526         } else {
 527                 tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
 528                 porttid += tidcnt * (subport - 1);
 529         }
 530         tidbase = (u64 __iomem *) ((char __iomem *)(dd->ipath_kregbase) +
 531                                    dd->ipath_rcvtidbase +
 532                                    porttid * sizeof(*tidbase));
 533
 534         limit = sizeof(tidmap) * BITS_PER_BYTE;
 535         if (limit > tidcnt)
 536                 /* just in case size changes in future */
 537                 limit = tidcnt;
 538         tid = find_first_bit(tidmap, limit);
 539         ipath_cdbg(VERBOSE, "Port%u free %u tids; first bit (max=%d) "
 540                    "set is %d, porttid %u\n", pd->port_port, ti->tidcnt,
 541                    limit, tid, porttid);
 542         for (cnt = 0; tid < limit; tid++) {
 543                 /*
 544                  * small optimization; if we detect a run of 3 or so without
 545                  * any set, use find_first_bit again.  That's mainly to
 546                  * accelerate the case where we wrapped, so we have some at
 547                  * the beginning, and some at the end, and a big gap
 548                  * in the middle.
 549                  */
 550                 if (!test_bit(tid, tidmap))
 551                         continue;
 552                 cnt++;
 553                 if (dd->ipath_pageshadow[porttid + tid]) {
 554                         struct page *p;
 555                         p = dd->ipath_pageshadow[porttid + tid];
 556                         dd->ipath_pageshadow[porttid + tid] = NULL;
 557                         ipath_cdbg(VERBOSE, "PID %u freeing TID %u\n",
 558                                    pid_nr(pd->port_pid), tid);
 559                         dd->ipath_f_put_tid(dd, &tidbase[tid],
 560                                             RCVHQ_RCV_TYPE_EXPECTED,
 561                                             dd->ipath_tidinvalid);
 562                         pci_unmap_page(dd->pcidev,
 563                                 dd->ipath_physshadow[porttid + tid],
 564                                 PAGE_SIZE, PCI_DMA_FROMDEVICE);
 565                         ipath_release_user_pages(&p, 1);
 566                         ipath_stats.sps_pageunlocks++;
 567                 } else
 568                         ipath_dbg("Unused tid %u, ignoring\n", tid);
 569         }
 570         if (cnt != ti->tidcnt)
 571                 ipath_dbg("passed in tidcnt %d, only %d bits set in map\n",
 572                           ti->tidcnt, cnt);
 573 done:
 574         if (ret)
 575                 ipath_dbg("Failed to unmap %u TID pages, failing with %d\n",
 576                           ti->tidcnt, -ret);
 577         return ret;
 578 }
 579
 580 /**
 581  * ipath_set_part_key - set a partition key
 582  * @pd: the port
 583  * @key: the key
 584  *
 585  * We can have up to 4 active at a time (other than the default, which is
 586  * always allowed).  This is somewhat tricky, since multiple ports may set
 587  * the same key, so we reference count them, and clean up at exit.  All 4
 588  * partition keys are packed into a single infinipath register.  It's an
 589  * error for a process to set the same pkey multiple times.  We provide no
 590  * mechanism to de-allocate a pkey at this time, we may eventually need to
 591  * do that.  I've used the atomic operations, and no locking, and only make
 592  * a single pass through what's available.  This should be more than
 593  * adequate for some time. I'll think about spinlocks or the like if and as
 594  * it's necessary.
 595  */
 596 static int ipath_set_part_key(struct ipath_portdata *pd, u16 key)
 597 {
 598         struct ipath_devdata *dd = pd->port_dd;
 599         int i, any = 0, pidx = -1;
 600         u16 lkey = key & 0x7FFF;
 601         int ret;
 602
 603         if (lkey == (IPATH_DEFAULT_P_KEY & 0x7FFF)) {
 604                 /* nothing to do; this key always valid */
 605                 ret = 0;
 606                 goto bail;
 607         }
 608
 609         ipath_cdbg(VERBOSE, "p%u try to set pkey %hx, current keys "
 610                    "%hx:%x %hx:%x %hx:%x %hx:%x\n",
 611                    pd->port_port, key, dd->ipath_pkeys[0],
 612                    atomic_read(&dd->ipath_pkeyrefs[0]), dd->ipath_pkeys[1],
 613                    atomic_read(&dd->ipath_pkeyrefs[1]), dd->ipath_pkeys[2],
 614                    atomic_read(&dd->ipath_pkeyrefs[2]), dd->ipath_pkeys[3],
 615                    atomic_read(&dd->ipath_pkeyrefs[3]));
 616
 617         if (!lkey) {
 618                 ipath_cdbg(PROC, "p%u tries to set key 0, not allowed\n",
 619                            pd->port_port);
 620                 ret = -EINVAL;
 621                 goto bail;
 622         }
 623
 624         /*
 625          * Set the full membership bit, because it has to be
 626          * set in the register or the packet, and it seems
 627          * cleaner to set in the register than to force all
 628          * callers to set it. (see bug 4331)
 629          */
 630         key |= 0x8000;
 631
 632         for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
 633                 if (!pd->port_pkeys[i] && pidx == -1)
 634                         pidx = i;
 635                 if (pd->port_pkeys[i] == key) {
 636                         ipath_cdbg(VERBOSE, "p%u tries to set same pkey "
 637                                    "(%x) more than once\n",
 638                                    pd->port_port, key);
 639                         ret = -EEXIST;
 640                         goto bail;
 641                 }
 642         }
 643         if (pidx == -1) {
 644                 ipath_dbg("All pkeys for port %u already in use, "
 645                           "can't set %x\n", pd->port_port, key);
 646                 ret = -EBUSY;
 647                 goto bail;
 648         }
 649         for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
 650                 if (!dd->ipath_pkeys[i]) {
 651                         any++;
 652                         continue;
 653                 }
 654                 if (dd->ipath_pkeys[i] == key) {
 655                         atomic_t *pkrefs = &dd->ipath_pkeyrefs[i];
 656
 657                         if (atomic_inc_return(pkrefs) > 1) {
 658                                 pd->port_pkeys[pidx] = key;
 659                                 ipath_cdbg(VERBOSE, "p%u set key %x "
 660                                            "matches #%d, count now %d\n",
 661                                            pd->port_port, key, i,
 662                                            atomic_read(pkrefs));
 663                                 ret = 0;
 664                                 goto bail;
 665                         } else {
 666                                 /*
 667                                  * lost race, decrement count, catch below
 668                                  */
 669                                 atomic_dec(pkrefs);
 670                                 ipath_cdbg(VERBOSE, "Lost race, count was "
 671                                            "0, after dec, it's %d\n",
 672                                            atomic_read(pkrefs));
 673                                 any++;
 674                         }
 675                 }
 676                 if ((dd->ipath_pkeys[i] & 0x7FFF) == lkey) {
 677                         /*
 678                          * It makes no sense to have both the limited and
 679                          * full membership PKEY set at the same time since
 680                          * the unlimited one will disable the limited one.
 681                          */
 682                         ret = -EEXIST;
 683                         goto bail;
 684                 }
 685         }
 686         if (!any) {
 687                 ipath_dbg("port %u, all pkeys already in use, "
 688                           "can't set %x\n", pd->port_port, key);
 689                 ret = -EBUSY;
 690                 goto bail;
 691         }
 692         for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
 693                 if (!dd->ipath_pkeys[i] &&
 694                     atomic_inc_return(&dd->ipath_pkeyrefs[i]) == 1) {
 695                         u64 pkey;
 696
 697                         /* for ipathstats, etc. */
 698                         ipath_stats.sps_pkeys[i] = lkey;
 699                         pd->port_pkeys[pidx] = dd->ipath_pkeys[i] = key;
 700                         pkey =
 701                                 (u64) dd->ipath_pkeys[0] |
 702                                 ((u64) dd->ipath_pkeys[1] << 16) |
 703                                 ((u64) dd->ipath_pkeys[2] << 32) |
 704                                 ((u64) dd->ipath_pkeys[3] << 48);
 705                         ipath_cdbg(PROC, "p%u set key %x in #%d, "
 706                                    "portidx %d, new pkey reg %llx\n",
 707                                    pd->port_port, key, i, pidx,
 708                                    (unsigned long long) pkey);
 709                         ipath_write_kreg(
 710                                 dd, dd->ipath_kregs->kr_partitionkey, pkey);
 711
 712                         ret = 0;
 713                         goto bail;
 714                 }
 715         }
 716         ipath_dbg("port %u, all pkeys already in use 2nd pass, "
 717                   "can't set %x\n", pd->port_port, key);
 718         ret = -EBUSY;
 719
 720 bail:
 721         return ret;
 722 }
 723
 724 /**
 725  * ipath_manage_rcvq - manage a port's receive queue
 726  * @pd: the port
 727  * @subport: the subport
 728  * @start_stop: action to carry out
 729  *
 730  * start_stop == 0 disables receive on the port, for use in queue
 731  * overflow conditions.  start_stop==1 re-enables, to be used to
 732  * re-init the software copy of the head register
 733  */
 734 static int ipath_manage_rcvq(struct ipath_portdata *pd, unsigned subport,
 735                              int start_stop)
 736 {
 737         struct ipath_devdata *dd = pd->port_dd;
 738
 739         ipath_cdbg(PROC, "%sabling rcv for unit %u port %u:%u\n",
 740                    start_stop ? "en" : "dis", dd->ipath_unit,
 741                    pd->port_port, subport);
 742         if (subport)
 743                 goto bail;
 744         /* atomically clear receive enable port. */
 745         if (start_stop) {
 746                 /*
 747                  * On enable, force in-memory copy of the tail register to
 748                  * 0, so that protocol code doesn't have to worry about
 749                  * whether or not the chip has yet updated the in-memory
 750                  * copy or not on return from the system call. The chip
 751                  * always resets it's tail register back to 0 on a
 752                  * transition from disabled to enabled.  This could cause a
 753                  * problem if software was broken, and did the enable w/o
 754                  * the disable, but eventually the in-memory copy will be
 755                  * updated and correct itself, even in the face of software
 756                  * bugs.
 757                  */
 758                 if (pd->port_rcvhdrtail_kvaddr)
 759                         ipath_clear_rcvhdrtail(pd);
 760                 set_bit(dd->ipath_r_portenable_shift + pd->port_port,
 761                         &dd->ipath_rcvctrl);
 762         } else
 763                 clear_bit(dd->ipath_r_portenable_shift + pd->port_port,
 764                           &dd->ipath_rcvctrl);
 765         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
 766                          dd->ipath_rcvctrl);
 767         /* now be sure chip saw it before we return */
 768         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 769         if (start_stop) {
 770                 /*
 771                  * And try to be sure that tail reg update has happened too.
 772                  * This should in theory interlock with the RXE changes to
 773                  * the tail register.  Don't assign it to the tail register
 774                  * in memory copy, since we could overwrite an update by the
 775                  * chip if we did.
 776                  */
 777                 ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
 778         }
 779         /* always; new head should be equal to new tail; see above */
 780 bail:
 781         return 0;
 782 }
 783
 784 static void ipath_clean_part_key(struct ipath_portdata *pd,
 785                                  struct ipath_devdata *dd)
 786 {
 787         int i, j, pchanged = 0;
 788         u64 oldpkey;
 789
 790         /* for debugging only */
 791         oldpkey = (u64) dd->ipath_pkeys[0] |
 792                 ((u64) dd->ipath_pkeys[1] << 16) |
 793                 ((u64) dd->ipath_pkeys[2] << 32) |
 794                 ((u64) dd->ipath_pkeys[3] << 48);
 795
 796         for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
 797                 if (!pd->port_pkeys[i])
 798                         continue;
 799                 ipath_cdbg(VERBOSE, "look for key[%d] %hx in pkeys\n", i,
 800                            pd->port_pkeys[i]);
 801                 for (j = 0; j < ARRAY_SIZE(dd->ipath_pkeys); j++) {
 802                         /* check for match independent of the global bit */
 803                         if ((dd->ipath_pkeys[j] & 0x7fff) !=
 804                             (pd->port_pkeys[i] & 0x7fff))
 805                                 continue;
 806                         if (atomic_dec_and_test(&dd->ipath_pkeyrefs[j])) {
 807                                 ipath_cdbg(VERBOSE, "p%u clear key "
 808                                            "%x matches #%d\n",
 809                                            pd->port_port,
 810                                            pd->port_pkeys[i], j);
 811                                 ipath_stats.sps_pkeys[j] =
 812                                         dd->ipath_pkeys[j] = 0;
 813                                 pchanged++;
 814                         }
 815                         else ipath_cdbg(
 816                                 VERBOSE, "p%u key %x matches #%d, "
 817                                 "but ref still %d\n", pd->port_port,
 818                                 pd->port_pkeys[i], j,
 819                                 atomic_read(&dd->ipath_pkeyrefs[j]));
 820                         break;
 821                 }
 822                 pd->port_pkeys[i] = 0;
 823         }
 824         if (pchanged) {
 825                 u64 pkey = (u64) dd->ipath_pkeys[0] |
 826                         ((u64) dd->ipath_pkeys[1] << 16) |
 827                         ((u64) dd->ipath_pkeys[2] << 32) |
 828                         ((u64) dd->ipath_pkeys[3] << 48);
 829                 ipath_cdbg(VERBOSE, "p%u old pkey reg %llx, "
 830                            "new pkey reg %llx\n", pd->port_port,
 831                            (unsigned long long) oldpkey,
 832                            (unsigned long long) pkey);
 833                 ipath_write_kreg(dd, dd->ipath_kregs->kr_partitionkey,
 834                                  pkey);
 835         }
 836 }
 837
 838 /*
 839  * Initialize the port data with the receive buffer sizes
 840  * so this can be done while the master port is locked.
 841  * Otherwise, there is a race with a slave opening the port
 842  * and seeing these fields uninitialized.
 843  */
 844 static void init_user_egr_sizes(struct ipath_portdata *pd)
 845 {
 846         struct ipath_devdata *dd = pd->port_dd;
 847         unsigned egrperchunk, egrcnt, size;
 848
 849         /*
 850          * to avoid wasting a lot of memory, we allocate 32KB chunks of
 851          * physically contiguous memory, advance through it until used up
 852          * and then allocate more.  Of course, we need memory to store those
 853          * extra pointers, now.  Started out with 256KB, but under heavy
 854          * memory pressure (creating large files and then copying them over
 855          * NFS while doing lots of MPI jobs), we hit some allocation
 856          * failures, even though we can sleep...  (2.6.10) Still get
 857          * failures at 64K.  32K is the lowest we can go without wasting
 858          * additional memory.
 859          */
 860         size = 0x8000;
 861         egrperchunk = size / dd->ipath_rcvegrbufsize;
 862         egrcnt = dd->ipath_rcvegrcnt;
 863         pd->port_rcvegrbuf_chunks = (egrcnt + egrperchunk - 1) / egrperchunk;
 864         pd->port_rcvegrbufs_perchunk = egrperchunk;
 865         pd->port_rcvegrbuf_size = size;
 866 }
 867
 868 /**
 869  * ipath_create_user_egr - allocate eager TID buffers
 870  * @pd: the port to allocate TID buffers for
 871  *
 872  * This routine is now quite different for user and kernel, because
 873  * the kernel uses skb's, for the accelerated network performance
 874  * This is the user port version
 875  *
 876  * Allocate the eager TID buffers and program them into infinipath
 877  * They are no longer completely contiguous, we do multiple allocation
 878  * calls.
 879  */
 880 static int ipath_create_user_egr(struct ipath_portdata *pd)
 881 {
 882         struct ipath_devdata *dd = pd->port_dd;
 883         unsigned e, egrcnt, egrperchunk, chunk, egrsize, egroff;
 884         size_t size;
 885         int ret;
 886         gfp_t gfp_flags;
 887
 888         /*
 889          * GFP_USER, but without GFP_FS, so buffer cache can be
 890          * coalesced (we hope); otherwise, even at order 4,
 891          * heavy filesystem activity makes these fail, and we can
 892          * use compound pages.
 893          */
 894         gfp_flags = __GFP_WAIT | __GFP_IO | __GFP_COMP;
 895
 896         egrcnt = dd->ipath_rcvegrcnt;
 897         /* TID number offset for this port */
 898         egroff = (pd->port_port - 1) * egrcnt + dd->ipath_p0_rcvegrcnt;
 899         egrsize = dd->ipath_rcvegrbufsize;
 900         ipath_cdbg(VERBOSE, "Allocating %d egr buffers, at egrtid "
 901                    "offset %x, egrsize %u\n", egrcnt, egroff, egrsize);
 902
 903         chunk = pd->port_rcvegrbuf_chunks;
 904         egrperchunk = pd->port_rcvegrbufs_perchunk;
 905         size = pd->port_rcvegrbuf_size;
 906         pd->port_rcvegrbuf = kmalloc(chunk * sizeof(pd->port_rcvegrbuf[0]),
 907                                      GFP_KERNEL);
 908         if (!pd->port_rcvegrbuf) {
 909                 ret = -ENOMEM;
 910                 goto bail;
 911         }
 912         pd->port_rcvegrbuf_phys =
 913                 kmalloc(chunk * sizeof(pd->port_rcvegrbuf_phys[0]),
 914                         GFP_KERNEL);
 915         if (!pd->port_rcvegrbuf_phys) {
 916                 ret = -ENOMEM;
 917                 goto bail_rcvegrbuf;
 918         }
 919         for (e = 0; e < pd->port_rcvegrbuf_chunks; e++) {
 920
 921                 pd->port_rcvegrbuf[e] = dma_alloc_coherent(
 922                         &dd->pcidev->dev, size, &pd->port_rcvegrbuf_phys[e],
 923                         gfp_flags);
 924
 925                 if (!pd->port_rcvegrbuf[e]) {
 926                         ret = -ENOMEM;
 927                         goto bail_rcvegrbuf_phys;
 928                 }
 929         }
 930
 931         pd->port_rcvegr_phys = pd->port_rcvegrbuf_phys[0];
 932
 933         for (e = chunk = 0; chunk < pd->port_rcvegrbuf_chunks; chunk++) {
 934                 dma_addr_t pa = pd->port_rcvegrbuf_phys[chunk];
 935                 unsigned i;
 936
 937                 for (i = 0; e < egrcnt && i < egrperchunk; e++, i++) {
 938                         dd->ipath_f_put_tid(dd, e + egroff +
 939                                             (u64 __iomem *)
 940                                             ((char __iomem *)
 941                                              dd->ipath_kregbase +
 942                                              dd->ipath_rcvegrbase),
 943                                             RCVHQ_RCV_TYPE_EAGER, pa);
 944                         pa += egrsize;
 945                 }
 946                 cond_resched(); /* don't hog the cpu */
 947         }
 948
 949         ret = 0;
 950         goto bail;
 951
 952 bail_rcvegrbuf_phys:
 953         for (e = 0; e < pd->port_rcvegrbuf_chunks &&
 954                 pd->port_rcvegrbuf[e]; e++) {
 955                 dma_free_coherent(&dd->pcidev->dev, size,
 956                                   pd->port_rcvegrbuf[e],
 957                                   pd->port_rcvegrbuf_phys[e]);
 958
 959         }
 960         kfree(pd->port_rcvegrbuf_phys);
 961         pd->port_rcvegrbuf_phys = NULL;
 962 bail_rcvegrbuf:
 963         kfree(pd->port_rcvegrbuf);
 964         pd->port_rcvegrbuf = NULL;
 965 bail:
 966         return ret;
 967 }
 968
 969
 970 /* common code for the mappings on dma_alloc_coherent mem */
 971 static int ipath_mmap_mem(struct vm_area_struct *vma,
 972         struct ipath_portdata *pd, unsigned len, int write_ok,
 973         void *kvaddr, char *what)
 974 {
 975         struct ipath_devdata *dd = pd->port_dd;
 976         unsigned long pfn;
 977         int ret;
 978
 979         if ((vma->vm_end - vma->vm_start) > len) {
 980                 dev_info(&dd->pcidev->dev,
 981                          "FAIL on %s: len %lx > %x\n", what,
 982                          vma->vm_end - vma->vm_start, len);
 983                 ret = -EFAULT;
 984                 goto bail;
 985         }
 986
 987         if (!write_ok) {
 988                 if (vma->vm_flags & VM_WRITE) {
 989                         dev_info(&dd->pcidev->dev,
 990                                  "%s must be mapped readonly\n", what);
 991                         ret = -EPERM;
 992                         goto bail;
 993                 }
 994
 995                 /* don't allow them to later change with mprotect */
 996                 vma->vm_flags &= ~VM_MAYWRITE;
 997         }
 998
 999         pfn = virt_to_phys(kvaddr) >> PAGE_SHIFT;
1000         ret = remap_pfn_range(vma, vma->vm_start, pfn,
1001                               len, vma->vm_page_prot);
1002         if (ret)
1003                 dev_info(&dd->pcidev->dev, "%s port%u mmap of %lx, %x "
1004                          "bytes r%c failed: %d\n", what, pd->port_port,
1005                          pfn, len, write_ok?'w':'o', ret);
1006         else
1007                 ipath_cdbg(VERBOSE, "%s port%u mmaped %lx, %x bytes "
1008                            "r%c\n", what, pd->port_port, pfn, len,
1009                            write_ok?'w':'o');
1010 bail:
1011         return ret;
1012 }
1013
1014 static int mmap_ureg(struct vm_area_struct *vma, struct ipath_devdata *dd,
1015                      u64 ureg)
1016 {
1017         unsigned long phys;
1018         int ret;
1019
1020         /*
1021          * This is real hardware, so use io_remap.  This is the mechanism
1022          * for the user process to update the head registers for their port
1023          * in the chip.
1024          */
1025         if ((vma->vm_end - vma->vm_start) > PAGE_SIZE) {
1026                 dev_info(&dd->pcidev->dev, "FAIL mmap userreg: reqlen "
1027                          "%lx > PAGE\n", vma->vm_end - vma->vm_start);
1028                 ret = -EFAULT;
1029         } else {
1030                 phys = dd->ipath_physaddr + ureg;
1031                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1032
1033                 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1034                 ret = io_remap_pfn_range(vma, vma->vm_start,
1035                                          phys >> PAGE_SHIFT,
1036                                          vma->vm_end - vma->vm_start,
1037                                          vma->vm_page_prot);
1038         }
1039         return ret;
1040 }
1041
1042 static int mmap_piobufs(struct vm_area_struct *vma,
1043                         struct ipath_devdata *dd,
1044                         struct ipath_portdata *pd,
1045                         unsigned piobufs, unsigned piocnt)
1046 {
1047         unsigned long phys;
1048         int ret;
1049
1050         /*
1051          * When we map the PIO buffers in the chip, we want to map them as
1052          * writeonly, no read possible.   This prevents access to previous
1053          * process data, and catches users who might try to read the i/o
1054          * space due to a bug.
1055          */
1056         if ((vma->vm_end - vma->vm_start) > (piocnt * dd->ipath_palign)) {
1057                 dev_info(&dd->pcidev->dev, "FAIL mmap piobufs: "
1058                          "reqlen %lx > PAGE\n",
1059                          vma->vm_end - vma->vm_start);
1060                 ret = -EINVAL;
1061                 goto bail;
1062         }
1063
1064         phys = dd->ipath_physaddr + piobufs;
1065
1066 #if defined(__powerpc__)
1067         /* There isn't a generic way to specify writethrough mappings */
1068         pgprot_val(vma->vm_page_prot) |= _PAGE_NO_CACHE;
1069         pgprot_val(vma->vm_page_prot) |= _PAGE_WRITETHRU;
1070         pgprot_val(vma->vm_page_prot) &= ~_PAGE_GUARDED;
1071 #endif
1072
1073         /*
1074          * don't allow them to later change to readable with mprotect (for when
1075          * not initially mapped readable, as is normally the case)
1076          */
1077         vma->vm_flags &= ~VM_MAYREAD;
1078         vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1079
1080         ret = io_remap_pfn_range(vma, vma->vm_start, phys >> PAGE_SHIFT,
1081                                  vma->vm_end - vma->vm_start,
1082                                  vma->vm_page_prot);
1083 bail:
1084         return ret;
1085 }
1086
1087 static int mmap_rcvegrbufs(struct vm_area_struct *vma,
1088                            struct ipath_portdata *pd)
1089 {
1090         struct ipath_devdata *dd = pd->port_dd;
1091         unsigned long start, size;
1092         size_t total_size, i;
1093         unsigned long pfn;
1094         int ret;
1095
1096         size = pd->port_rcvegrbuf_size;
1097         total_size = pd->port_rcvegrbuf_chunks * size;
1098         if ((vma->vm_end - vma->vm_start) > total_size) {
1099                 dev_info(&dd->pcidev->dev, "FAIL on egr bufs: "
1100                          "reqlen %lx > actual %lx\n",
1101                          vma->vm_end - vma->vm_start,
1102                          (unsigned long) total_size);
1103                 ret = -EINVAL;
1104                 goto bail;
1105         }
1106
1107         if (vma->vm_flags & VM_WRITE) {
1108                 dev_info(&dd->pcidev->dev, "Can't map eager buffers as "
1109                          "writable (flags=%lx)\n", vma->vm_flags);
1110                 ret = -EPERM;
1111                 goto bail;
1112         }
1113         /* don't allow them to later change to writeable with mprotect */
1114         vma->vm_flags &= ~VM_MAYWRITE;
1115
1116         start = vma->vm_start;
1117
1118         for (i = 0; i < pd->port_rcvegrbuf_chunks; i++, start += size) {
1119                 pfn = virt_to_phys(pd->port_rcvegrbuf[i]) >> PAGE_SHIFT;
1120                 ret = remap_pfn_range(vma, start, pfn, size,
1121                                       vma->vm_page_prot);
1122                 if (ret < 0)
1123                         goto bail;
1124         }
1125         ret = 0;
1126
1127 bail:
1128         return ret;
1129 }
1130
1131 /*
1132  * ipath_file_vma_fault - handle a VMA page fault.
1133  */
1134 static int ipath_file_vma_fault(struct vm_area_struct *vma,
1135                                         struct vm_fault *vmf)
1136 {
1137         struct page *page;
1138
1139         page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
1140         if (!page)
1141                 return VM_FAULT_SIGBUS;
1142         get_page(page);
1143         vmf->page = page;
1144
1145         return 0;
1146 }
1147
1148 static struct vm_operations_struct ipath_file_vm_ops = {
1149         .fault = ipath_file_vma_fault,
1150 };
1151
1152 static int mmap_kvaddr(struct vm_area_struct *vma, u64 pgaddr,
1153                        struct ipath_portdata *pd, unsigned subport)
1154 {
1155         unsigned long len;
1156         struct ipath_devdata *dd;
1157         void *addr;
1158         size_t size;
1159         int ret = 0;
1160
1161         /* If the port is not shared, all addresses should be physical */
1162         if (!pd->port_subport_cnt)
1163                 goto bail;
1164
1165         dd = pd->port_dd;
1166         size = pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
1167
1168         /*
1169          * Each process has all the subport uregbase, rcvhdrq, and
1170          * rcvegrbufs mmapped - as an array for all the processes,
1171          * and also separately for this process.
1172          */
1173         if (pgaddr == cvt_kvaddr(pd->subport_uregbase)) {
1174                 addr = pd->subport_uregbase;
1175                 size = PAGE_SIZE * pd->port_subport_cnt;
1176         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvhdr_base)) {
1177                 addr = pd->subport_rcvhdr_base;
1178                 size = pd->port_rcvhdrq_size * pd->port_subport_cnt;
1179         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvegrbuf)) {
1180                 addr = pd->subport_rcvegrbuf;
1181                 size *= pd->port_subport_cnt;
1182         } else if (pgaddr == cvt_kvaddr(pd->subport_uregbase +
1183                                         PAGE_SIZE * subport)) {
1184                 addr = pd->subport_uregbase + PAGE_SIZE * subport;
1185                 size = PAGE_SIZE;
1186         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvhdr_base +
1187                                 pd->port_rcvhdrq_size * subport)) {
1188                 addr = pd->subport_rcvhdr_base +
1189                         pd->port_rcvhdrq_size * subport;
1190                 size = pd->port_rcvhdrq_size;
1191         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvegrbuf +
1192                                size * subport)) {
1193                 addr = pd->subport_rcvegrbuf + size * subport;
1194                 /* rcvegrbufs are read-only on the slave */
1195                 if (vma->vm_flags & VM_WRITE) {
1196                         dev_info(&dd->pcidev->dev,
1197                                  "Can't map eager buffers as "
1198                                  "writable (flags=%lx)\n", vma->vm_flags);
1199                         ret = -EPERM;
1200                         goto bail;
1201                 }
1202                 /*
1203                  * Don't allow permission to later change to writeable
1204                  * with mprotect.
1205                  */
1206                 vma->vm_flags &= ~VM_MAYWRITE;
1207         } else {
1208                 goto bail;
1209         }
1210         len = vma->vm_end - vma->vm_start;
1211         if (len > size) {
1212                 ipath_cdbg(MM, "FAIL: reqlen %lx > %zx\n", len, size);
1213                 ret = -EINVAL;
1214                 goto bail;
1215         }
1216
1217         vma->vm_pgoff = (unsigned long) addr >> PAGE_SHIFT;
1218         vma->vm_ops = &ipath_file_vm_ops;
1219         vma->vm_flags |= VM_RESERVED | VM_DONTEXPAND;
1220         ret = 1;
1221
1222 bail:
1223         return ret;
1224 }
1225
1226 /**
1227  * ipath_mmap - mmap various structures into user space
1228  * @fp: the file pointer
1229  * @vma: the VM area
1230  *
1231  * We use this to have a shared buffer between the kernel and the user code
1232  * for the rcvhdr queue, egr buffers, and the per-port user regs and pio
1233  * buffers in the chip.  We have the open and close entries so we can bump
1234  * the ref count and keep the driver from being unloaded while still mapped.
1235  */
1236 static int ipath_mmap(struct file *fp, struct vm_area_struct *vma)
1237 {
1238         struct ipath_portdata *pd;
1239         struct ipath_devdata *dd;
1240         u64 pgaddr, ureg;
1241         unsigned piobufs, piocnt;
1242         int ret;
1243
1244         pd = port_fp(fp);
1245         if (!pd) {
1246                 ret = -EINVAL;
1247                 goto bail;
1248         }
1249         dd = pd->port_dd;
1250
1251         /*
1252          * This is the ipath_do_user_init() code, mapping the shared buffers
1253          * into the user process. The address referred to by vm_pgoff is the
1254          * file offset passed via mmap().  For shared ports, this is the
1255          * kernel vmalloc() address of the pages to share with the master.
1256          * For non-shared or master ports, this is a physical address.
1257          * We only do one mmap for each space mapped.
1258          */
1259         pgaddr = vma->vm_pgoff << PAGE_SHIFT;
1260
1261         /*
1262          * Check for 0 in case one of the allocations failed, but user
1263          * called mmap anyway.
1264          */
1265         if (!pgaddr)  {
1266                 ret = -EINVAL;
1267                 goto bail;
1268         }
1269
1270         ipath_cdbg(MM, "pgaddr %llx vm_start=%lx len %lx port %u:%u:%u\n",
1271                    (unsigned long long) pgaddr, vma->vm_start,
1272                    vma->vm_end - vma->vm_start, dd->ipath_unit,
1273                    pd->port_port, subport_fp(fp));
1274
1275         /*
1276          * Physical addresses must fit in 40 bits for our hardware.
1277          * Check for kernel virtual addresses first, anything else must
1278          * match a HW or memory address.
1279          */
1280         ret = mmap_kvaddr(vma, pgaddr, pd, subport_fp(fp));
1281         if (ret) {
1282                 if (ret > 0)
1283                         ret = 0;
1284                 goto bail;
1285         }
1286
1287         ureg = dd->ipath_uregbase + dd->ipath_ureg_align * pd->port_port;
1288         if (!pd->port_subport_cnt) {
1289                 /* port is not shared */
1290                 piocnt = pd->port_piocnt;
1291                 piobufs = pd->port_piobufs;
1292         } else if (!subport_fp(fp)) {
1293                 /* caller is the master */
1294                 piocnt = (pd->port_piocnt / pd->port_subport_cnt) +
1295                          (pd->port_piocnt % pd->port_subport_cnt);
1296                 piobufs = pd->port_piobufs +
1297                         dd->ipath_palign * (pd->port_piocnt - piocnt);
1298         } else {
1299                 unsigned slave = subport_fp(fp) - 1;
1300
1301                 /* caller is a slave */
1302                 piocnt = pd->port_piocnt / pd->port_subport_cnt;
1303                 piobufs = pd->port_piobufs + dd->ipath_palign * piocnt * slave;
1304         }
1305
1306         if (pgaddr == ureg)
1307                 ret = mmap_ureg(vma, dd, ureg);
1308         else if (pgaddr == piobufs)
1309                 ret = mmap_piobufs(vma, dd, pd, piobufs, piocnt);
1310         else if (pgaddr == dd->ipath_pioavailregs_phys)
1311                 /* in-memory copy of pioavail registers */
1312                 ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1313                                      (void *) dd->ipath_pioavailregs_dma,
1314                                      "pioavail registers");
1315         else if (pgaddr == pd->port_rcvegr_phys)
1316                 ret = mmap_rcvegrbufs(vma, pd);
1317         else if (pgaddr == (u64) pd->port_rcvhdrq_phys)
1318                 /*
1319                  * The rcvhdrq itself; readonly except on HT (so have
1320                  * to allow writable mapping), multiple pages, contiguous
1321                  * from an i/o perspective.
1322                  */
1323                 ret = ipath_mmap_mem(vma, pd, pd->port_rcvhdrq_size, 1,
1324                                      pd->port_rcvhdrq,
1325                                      "rcvhdrq");
1326         else if (pgaddr == (u64) pd->port_rcvhdrqtailaddr_phys)
1327                 /* in-memory copy of rcvhdrq tail register */
1328                 ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1329                                      pd->port_rcvhdrtail_kvaddr,
1330                                      "rcvhdrq tail");
1331         else
1332                 ret = -EINVAL;
1333
1334         vma->vm_private_data = NULL;
1335
1336         if (ret < 0)
1337                 dev_info(&dd->pcidev->dev,
1338                          "Failure %d on off %llx len %lx\n",
1339                          -ret, (unsigned long long)pgaddr,
1340                          vma->vm_end - vma->vm_start);
1341 bail:
1342         return ret;
1343 }
1344
1345 static unsigned ipath_poll_hdrqfull(struct ipath_portdata *pd)
1346 {
1347         unsigned pollflag = 0;
1348
1349         if ((pd->poll_type & IPATH_POLL_TYPE_OVERFLOW) &&
1350             pd->port_hdrqfull != pd->port_hdrqfull_poll) {
1351                 pollflag |= POLLIN | POLLRDNORM;
1352                 pd->port_hdrqfull_poll = pd->port_hdrqfull;
1353         }
1354
1355         return pollflag;
1356 }
1357
1358 static unsigned int ipath_poll_urgent(struct ipath_portdata *pd,
1359                                       struct file *fp,
1360                                       struct poll_table_struct *pt)
1361 {
1362         unsigned pollflag = 0;
1363         struct ipath_devdata *dd;
1364
1365         dd = pd->port_dd;
1366
1367         /* variable access in ipath_poll_hdrqfull() needs this */
1368         rmb();
1369         pollflag = ipath_poll_hdrqfull(pd);
1370
1371         if (pd->port_urgent != pd->port_urgent_poll) {
1372                 pollflag |= POLLIN | POLLRDNORM;
1373                 pd->port_urgent_poll = pd->port_urgent;
1374         }
1375
1376         if (!pollflag) {
1377                 /* this saves a spin_lock/unlock in interrupt handler... */
1378                 set_bit(IPATH_PORT_WAITING_URG, &pd->port_flag);
1379                 /* flush waiting flag so don't miss an event... */
1380                 wmb();
1381                 poll_wait(fp, &pd->port_wait, pt);
1382         }
1383
1384         return pollflag;
1385 }
1386
1387 static unsigned int ipath_poll_next(struct ipath_portdata *pd,
1388                                     struct file *fp,
1389                                     struct poll_table_struct *pt)
1390 {
1391         u32 head;
1392         u32 tail;
1393         unsigned pollflag = 0;
1394         struct ipath_devdata *dd;
1395
1396         dd = pd->port_dd;
1397
1398         /* variable access in ipath_poll_hdrqfull() needs this */
1399         rmb();
1400         pollflag = ipath_poll_hdrqfull(pd);
1401
1402         head = ipath_read_ureg32(dd, ur_rcvhdrhead, pd->port_port);
1403         if (pd->port_rcvhdrtail_kvaddr)
1404                 tail = ipath_get_rcvhdrtail(pd);
1405         else
1406                 tail = ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
1407
1408         if (head != tail)
1409                 pollflag |= POLLIN | POLLRDNORM;
1410         else {
1411                 /* this saves a spin_lock/unlock in interrupt handler */
1412                 set_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag);
1413                 /* flush waiting flag so we don't miss an event */
1414                 wmb();
1415
1416                 set_bit(pd->port_port + dd->ipath_r_intravail_shift,
1417                         &dd->ipath_rcvctrl);
1418
1419                 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1420                                  dd->ipath_rcvctrl);
1421
1422                 if (dd->ipath_rhdrhead_intr_off) /* arm rcv interrupt */
1423                         ipath_write_ureg(dd, ur_rcvhdrhead,
1424                                          dd->ipath_rhdrhead_intr_off | head,
1425                                          pd->port_port);
1426
1427                 poll_wait(fp, &pd->port_wait, pt);
1428         }
1429
1430         return pollflag;
1431 }
1432
1433 static unsigned int ipath_poll(struct file *fp,
1434                                struct poll_table_struct *pt)
1435 {
1436         struct ipath_portdata *pd;
1437         unsigned pollflag;
1438
1439         pd = port_fp(fp);
1440         if (!pd)
1441                 pollflag = 0;
1442         else if (pd->poll_type & IPATH_POLL_TYPE_URGENT)
1443                 pollflag = ipath_poll_urgent(pd, fp, pt);
1444         else
1445                 pollflag = ipath_poll_next(pd, fp, pt);
1446
1447         return pollflag;
1448 }
1449
1450 static int ipath_supports_subports(int user_swmajor, int user_swminor)
1451 {
1452         /* no subport implementation prior to software version 1.3 */
1453         return (user_swmajor > 1) || (user_swminor >= 3);
1454 }
1455
1456 static int ipath_compatible_subports(int user_swmajor, int user_swminor)
1457 {
1458         /* this code is written long-hand for clarity */
1459         if (IPATH_USER_SWMAJOR != user_swmajor) {
1460                 /* no promise of compatibility if major mismatch */
1461                 return 0;
1462         }
1463         if (IPATH_USER_SWMAJOR == 1) {
1464                 switch (IPATH_USER_SWMINOR) {
1465                 case 0:
1466                 case 1:
1467                 case 2:
1468                         /* no subport implementation so cannot be compatible */
1469                         return 0;
1470                 case 3:
1471                         /* 3 is only compatible with itself */
1472                         return user_swminor == 3;
1473                 default:
1474                         /* >= 4 are compatible (or are expected to be) */
1475                         return user_swminor >= 4;
1476                 }
1477         }
1478         /* make no promises yet for future major versions */
1479         return 0;
1480 }
1481
1482 static int init_subports(struct ipath_devdata *dd,
1483                          struct ipath_portdata *pd,
1484                          const struct ipath_user_info *uinfo)
1485 {
1486         int ret = 0;
1487         unsigned num_subports;
1488         size_t size;
1489
1490         /*
1491          * If the user is requesting zero subports,
1492          * skip the subport allocation.
1493          */
1494         if (uinfo->spu_subport_cnt <= 0)
1495                 goto bail;
1496
1497         /* Self-consistency check for ipath_compatible_subports() */
1498         if (ipath_supports_subports(IPATH_USER_SWMAJOR, IPATH_USER_SWMINOR) &&
1499             !ipath_compatible_subports(IPATH_USER_SWMAJOR,
1500                                        IPATH_USER_SWMINOR)) {
1501                 dev_info(&dd->pcidev->dev,
1502                          "Inconsistent ipath_compatible_subports()\n");
1503                 goto bail;
1504         }
1505
1506         /* Check for subport compatibility */
1507         if (!ipath_compatible_subports(uinfo->spu_userversion >> 16,
1508                                        uinfo->spu_userversion & 0xffff)) {
1509                 dev_info(&dd->pcidev->dev,
1510                          "Mismatched user version (%d.%d) and driver "
1511                          "version (%d.%d) while port sharing. Ensure "
1512                          "that driver and library are from the same "
1513                          "release.\n",
1514                          (int) (uinfo->spu_userversion >> 16),
1515                          (int) (uinfo->spu_userversion & 0xffff),
1516                          IPATH_USER_SWMAJOR,
1517                          IPATH_USER_SWMINOR);
1518                 goto bail;
1519         }
1520         if (uinfo->spu_subport_cnt > INFINIPATH_MAX_SUBPORT) {
1521                 ret = -EINVAL;
1522                 goto bail;
1523         }
1524
1525         num_subports = uinfo->spu_subport_cnt;
1526         pd->subport_uregbase = vmalloc(PAGE_SIZE * num_subports);
1527         if (!pd->subport_uregbase) {
1528                 ret = -ENOMEM;
1529                 goto bail;
1530         }
1531         /* Note: pd->port_rcvhdrq_size isn't initialized yet. */
1532         size = ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize *
1533                      sizeof(u32), PAGE_SIZE) * num_subports;
1534         pd->subport_rcvhdr_base = vmalloc(size);
1535         if (!pd->subport_rcvhdr_base) {
1536                 ret = -ENOMEM;
1537                 goto bail_ureg;
1538         }
1539
1540         pd->subport_rcvegrbuf = vmalloc(pd->port_rcvegrbuf_chunks *
1541                                         pd->port_rcvegrbuf_size *
1542                                         num_subports);
1543         if (!pd->subport_rcvegrbuf) {
1544                 ret = -ENOMEM;
1545                 goto bail_rhdr;
1546         }
1547
1548         pd->port_subport_cnt = uinfo->spu_subport_cnt;
1549         pd->port_subport_id = uinfo->spu_subport_id;
1550         pd->active_slaves = 1;
1551         set_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag);
1552         memset(pd->subport_uregbase, 0, PAGE_SIZE * num_subports);
1553         memset(pd->subport_rcvhdr_base, 0, size);
1554         memset(pd->subport_rcvegrbuf, 0, pd->port_rcvegrbuf_chunks *
1555                                          pd->port_rcvegrbuf_size *
1556                                          num_subports);
1557         goto bail;
1558
1559 bail_rhdr:
1560         vfree(pd->subport_rcvhdr_base);
1561 bail_ureg:
1562         vfree(pd->subport_uregbase);
1563         pd->subport_uregbase = NULL;
1564 bail:
1565         return ret;
1566 }
1567
1568 static int try_alloc_port(struct ipath_devdata *dd, int port,
1569                           struct file *fp,
1570                           const struct ipath_user_info *uinfo)
1571 {
1572         struct ipath_portdata *pd;
1573         int ret;
1574
1575         if (!(pd = dd->ipath_pd[port])) {
1576                 void *ptmp;
1577
1578                 pd = kzalloc(sizeof(struct ipath_portdata), GFP_KERNEL);
1579
1580                 /*
1581                  * Allocate memory for use in ipath_tid_update() just once
1582                  * at open, not per call.  Reduces cost of expected send
1583                  * setup.
1584                  */
1585                 ptmp = kmalloc(dd->ipath_rcvtidcnt * sizeof(u16) +
1586                                dd->ipath_rcvtidcnt * sizeof(struct page **),
1587                                GFP_KERNEL);
1588                 if (!pd || !ptmp) {
1589                         ipath_dev_err(dd, "Unable to allocate portdata "
1590                                       "memory, failing open\n");
1591                         ret = -ENOMEM;
1592                         kfree(pd);
1593                         kfree(ptmp);
1594                         goto bail;
1595                 }
1596                 dd->ipath_pd[port] = pd;
1597                 dd->ipath_pd[port]->port_port = port;
1598                 dd->ipath_pd[port]->port_dd = dd;
1599                 dd->ipath_pd[port]->port_tid_pg_list = ptmp;
1600                 init_waitqueue_head(&dd->ipath_pd[port]->port_wait);
1601         }
1602         if (!pd->port_cnt) {
1603                 pd->userversion = uinfo->spu_userversion;
1604                 init_user_egr_sizes(pd);
1605                 if ((ret = init_subports(dd, pd, uinfo)) != 0)
1606                         goto bail;
1607                 ipath_cdbg(PROC, "%s[%u] opened unit:port %u:%u\n",
1608                            current->comm, current->pid, dd->ipath_unit,
1609                            port);
1610                 pd->port_cnt = 1;
1611                 port_fp(fp) = pd;
1612                 pd->port_pid = get_pid(task_pid(current));
1613                 strncpy(pd->port_comm, current->comm, sizeof(pd->port_comm));
1614                 ipath_stats.sps_ports++;
1615                 ret = 0;
1616         } else
1617                 ret = -EBUSY;
1618
1619 bail:
1620         return ret;
1621 }
1622
1623 static inline int usable(struct ipath_devdata *dd)
1624 {
1625         return dd &&
1626                 (dd->ipath_flags & IPATH_PRESENT) &&
1627                 dd->ipath_kregbase &&
1628                 dd->ipath_lid &&
1629                 !(dd->ipath_flags & (IPATH_LINKDOWN | IPATH_DISABLED
1630                                      | IPATH_LINKUNK));
1631 }
1632
1633 static int find_free_port(int unit, struct file *fp,
1634                           const struct ipath_user_info *uinfo)
1635 {
1636         struct ipath_devdata *dd = ipath_lookup(unit);
1637         int ret, i;
1638
1639         if (!dd) {
1640                 ret = -ENODEV;
1641                 goto bail;
1642         }
1643
1644         if (!usable(dd)) {
1645                 ret = -ENETDOWN;
1646                 goto bail;
1647         }
1648
1649         for (i = 1; i < dd->ipath_cfgports; i++) {
1650                 ret = try_alloc_port(dd, i, fp, uinfo);
1651                 if (ret != -EBUSY)
1652                         goto bail;
1653         }
1654         ret = -EBUSY;
1655
1656 bail:
1657         return ret;
1658 }
1659
1660 static int find_best_unit(struct file *fp,
1661                           const struct ipath_user_info *uinfo)
1662 {
1663         int ret = 0, i, prefunit = -1, devmax;
1664         int maxofallports, npresent, nup;
1665         int ndev;
1666
1667         devmax = ipath_count_units(&npresent, &nup, &maxofallports);
1668
1669         /*
1670          * This code is present to allow a knowledgeable person to
1671          * specify the layout of processes to processors before opening
1672          * this driver, and then we'll assign the process to the "closest"
1673          * InfiniPath chip to that processor (we assume reasonable connectivity,
1674          * for now).  This code assumes that if affinity has been set
1675          * before this point, that at most one cpu is set; for now this
1676          * is reasonable.  I check for both cpus_empty() and cpus_full(),
1677          * in case some kernel variant sets none of the bits when no
1678          * affinity is set.  2.6.11 and 12 kernels have all present
1679          * cpus set.  Some day we'll have to fix it up further to handle
1680          * a cpu subset.  This algorithm fails for two HT chips connected
1681          * in tunnel fashion.  Eventually this needs real topology
1682          * information.  There may be some issues with dual core numbering
1683          * as well.  This needs more work prior to release.
1684          */
1685         if (!cpus_empty(current->cpus_allowed) &&
1686             !cpus_full(current->cpus_allowed)) {
1687                 int ncpus = num_online_cpus(), curcpu = -1, nset = 0;
1688                 for (i = 0; i < ncpus; i++)
1689                         if (cpu_isset(i, current->cpus_allowed)) {
1690                                 ipath_cdbg(PROC, "%s[%u] affinity set for "
1691                                            "cpu %d/%d\n", current->comm,
1692                                            current->pid, i, ncpus);
1693                                 curcpu = i;
1694                                 nset++;
1695                         }
1696                 if (curcpu != -1 && nset != ncpus) {
1697                         if (npresent) {
1698                                 prefunit = curcpu / (ncpus / npresent);
1699                                 ipath_cdbg(PROC,"%s[%u] %d chips, %d cpus, "
1700                                           "%d cpus/chip, select unit %d\n",
1701                                           current->comm, current->pid,
1702                                           npresent, ncpus, ncpus / npresent,
1703                                           prefunit);
1704                         }
1705                 }
1706         }
1707
1708         /*
1709          * user ports start at 1, kernel port is 0
1710          * For now, we do round-robin access across all chips
1711          */
1712
1713         if (prefunit != -1)
1714                 devmax = prefunit + 1;
1715 recheck:
1716         for (i = 1; i < maxofallports; i++) {
1717                 for (ndev = prefunit != -1 ? prefunit : 0; ndev < devmax;
1718                      ndev++) {
1719                         struct ipath_devdata *dd = ipath_lookup(ndev);
1720
1721                         if (!usable(dd))
1722                                 continue; /* can't use this unit */
1723                         if (i >= dd->ipath_cfgports)
1724                                 /*
1725                                  * Maxed out on users of this unit. Try
1726                                  * next.
1727                                  */
1728                                 continue;
1729                         ret = try_alloc_port(dd, i, fp, uinfo);
1730                         if (!ret)
1731                                 goto done;
1732                 }
1733         }
1734
1735         if (npresent) {
1736                 if (nup == 0) {
1737                         ret = -ENETDOWN;
1738                         ipath_dbg("No ports available (none initialized "
1739                                   "and ready)\n");
1740                 } else {
1741                         if (prefunit > 0) {
1742                                 /* if started above 0, retry from 0 */
1743                                 ipath_cdbg(PROC,
1744                                            "%s[%u] no ports on prefunit "
1745                                            "%d, clear and re-check\n",
1746                                            current->comm, current->pid,
1747                                            prefunit);
1748                                 devmax = ipath_count_units(NULL, NULL,
1749                                                            NULL);
1750                                 prefunit = -1;
1751                                 goto recheck;
1752                         }
1753                         ret = -EBUSY;
1754                         ipath_dbg("No ports available\n");
1755                 }
1756         } else {
1757                 ret = -ENXIO;
1758                 ipath_dbg("No boards found\n");
1759         }
1760
1761 done:
1762         return ret;
1763 }
1764
1765 static int find_shared_port(struct file *fp,
1766                             const struct ipath_user_info *uinfo)
1767 {
1768         int devmax, ndev, i;
1769         int ret = 0;
1770
1771         devmax = ipath_count_units(NULL, NULL, NULL);
1772
1773         for (ndev = 0; ndev < devmax; ndev++) {
1774                 struct ipath_devdata *dd = ipath_lookup(ndev);
1775
1776                 if (!usable(dd))
1777                         continue;
1778                 for (i = 1; i < dd->ipath_cfgports; i++) {
1779                         struct ipath_portdata *pd = dd->ipath_pd[i];
1780
1781                         /* Skip ports which are not yet open */
1782                         if (!pd || !pd->port_cnt)
1783                                 continue;
1784                         /* Skip port if it doesn't match the requested one */
1785                         if (pd->port_subport_id != uinfo->spu_subport_id)
1786                                 continue;
1787                         /* Verify the sharing process matches the master */
1788                         if (pd->port_subport_cnt != uinfo->spu_subport_cnt ||
1789                             pd->userversion != uinfo->spu_userversion ||
1790                             pd->port_cnt >= pd->port_subport_cnt) {
1791                                 ret = -EINVAL;
1792                                 goto done;
1793                         }
1794                         port_fp(fp) = pd;
1795                         subport_fp(fp) = pd->port_cnt++;
1796                         pd->port_subpid[subport_fp(fp)] =
1797                                 get_pid(task_pid(current));
1798                         tidcursor_fp(fp) = 0;
1799                         pd->active_slaves |= 1 << subport_fp(fp);
1800                         ipath_cdbg(PROC,
1801                                    "%s[%u] %u sharing %s[%u] unit:port %u:%u\n",
1802                                    current->comm, current->pid,
1803                                    subport_fp(fp),
1804                                    pd->port_comm, pid_nr(pd->port_pid),
1805                                    dd->ipath_unit, pd->port_port);
1806                         ret = 1;
1807                         goto done;
1808                 }
1809         }
1810
1811 done:
1812         return ret;
1813 }
1814
1815 /* No BKL needed here */
1816 static int ipath_open(struct inode *in, struct file *fp)
1817 {
1818         /* The real work is performed later in ipath_assign_port() */
1819         fp->private_data = kzalloc(sizeof(struct ipath_filedata), GFP_KERNEL);
1820         return fp->private_data ? 0 : -ENOMEM;
1821 }
1822
1823 /* Get port early, so can set affinity prior to memory allocation */
1824 static int ipath_assign_port(struct file *fp,
1825                               const struct ipath_user_info *uinfo)
1826 {
1827         int ret;
1828         int i_minor;
1829         unsigned swmajor, swminor;
1830
1831         /* Check to be sure we haven't already initialized this file */
1832         if (port_fp(fp)) {
1833                 ret = -EINVAL;
1834                 goto done;
1835         }
1836
1837         /* for now, if major version is different, bail */
1838         swmajor = uinfo->spu_userversion >> 16;
1839         if (swmajor != IPATH_USER_SWMAJOR) {
1840                 ipath_dbg("User major version %d not same as driver "
1841                           "major %d\n", uinfo->spu_userversion >> 16,
1842                           IPATH_USER_SWMAJOR);
1843                 ret = -ENODEV;
1844                 goto done;
1845         }
1846
1847         swminor = uinfo->spu_userversion & 0xffff;
1848         if (swminor != IPATH_USER_SWMINOR)
1849                 ipath_dbg("User minor version %d not same as driver "
1850                           "minor %d\n", swminor, IPATH_USER_SWMINOR);
1851
1852         mutex_lock(&ipath_mutex);
1853
1854         if (ipath_compatible_subports(swmajor, swminor) &&
1855             uinfo->spu_subport_cnt &&
1856             (ret = find_shared_port(fp, uinfo))) {
1857                 if (ret > 0)
1858                         ret = 0;
1859                 goto done_chk_sdma;
1860         }
1861
1862         i_minor = iminor(fp->f_path.dentry->d_inode) - IPATH_USER_MINOR_BASE;
1863         ipath_cdbg(VERBOSE, "open on dev %lx (minor %d)\n",
1864                    (long)fp->f_path.dentry->d_inode->i_rdev, i_minor);
1865
1866         if (i_minor)
1867                 ret = find_free_port(i_minor - 1, fp, uinfo);
1868         else
1869                 ret = find_best_unit(fp, uinfo);
1870
1871 done_chk_sdma:
1872         if (!ret) {
1873                 struct ipath_filedata *fd = fp->private_data;
1874                 const struct ipath_portdata *pd = fd->pd;
1875                 const struct ipath_devdata *dd = pd->port_dd;
1876
1877                 fd->pq = ipath_user_sdma_queue_create(&dd->pcidev->dev,
1878                                                       dd->ipath_unit,
1879                                                       pd->port_port,
1880                                                       fd->subport);
1881
1882                 if (!fd->pq)
1883                         ret = -ENOMEM;
1884         }
1885
1886         mutex_unlock(&ipath_mutex);
1887
1888 done:
1889         return ret;
1890 }
1891
1892
1893 static int ipath_do_user_init(struct file *fp,
1894                               const struct ipath_user_info *uinfo)
1895 {
1896         int ret;
1897         struct ipath_portdata *pd = port_fp(fp);
1898         struct ipath_devdata *dd;
1899         u32 head32;
1900
1901         /* Subports don't need to initialize anything since master did it. */
1902         if (subport_fp(fp)) {
1903                 ret = wait_event_interruptible(pd->port_wait,
1904                         !test_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag));
1905                 goto done;
1906         }
1907
1908         dd = pd->port_dd;
1909
1910         if (uinfo->spu_rcvhdrsize) {
1911                 ret = ipath_setrcvhdrsize(dd, uinfo->spu_rcvhdrsize);
1912                 if (ret)
1913                         goto done;
1914         }
1915
1916         /* for now we do nothing with rcvhdrcnt: uinfo->spu_rcvhdrcnt */
1917
1918         /* some ports may get extra buffers, calculate that here */
1919         if (pd->port_port <= dd->ipath_ports_extrabuf)
1920                 pd->port_piocnt = dd->ipath_pbufsport + 1;
1921         else
1922                 pd->port_piocnt = dd->ipath_pbufsport;
1923
1924         /* for right now, kernel piobufs are at end, so port 1 is at 0 */
1925         if (pd->port_port <= dd->ipath_ports_extrabuf)
1926                 pd->port_pio_base = (dd->ipath_pbufsport + 1)
1927                         * (pd->port_port - 1);
1928         else
1929                 pd->port_pio_base = dd->ipath_ports_extrabuf +
1930                         dd->ipath_pbufsport * (pd->port_port - 1);
1931         pd->port_piobufs = dd->ipath_piobufbase +
1932                 pd->port_pio_base * dd->ipath_palign;
1933         ipath_cdbg(VERBOSE, "piobuf base for port %u is 0x%x, piocnt %u,"
1934                 " first pio %u\n", pd->port_port, pd->port_piobufs,
1935                 pd->port_piocnt, pd->port_pio_base);
1936         ipath_chg_pioavailkernel(dd, pd->port_pio_base, pd->port_piocnt, 0);
1937
1938         /*
1939          * Now allocate the rcvhdr Q and eager TIDs; skip the TID
1940          * array for time being.  If pd->port_port > chip-supported,
1941          * we need to do extra stuff here to handle by handling overflow
1942          * through port 0, someday
1943          */
1944         ret = ipath_create_rcvhdrq(dd, pd);
1945         if (!ret)
1946                 ret = ipath_create_user_egr(pd);
1947         if (ret)
1948                 goto done;
1949
1950         /*
1951          * set the eager head register for this port to the current values
1952          * of the tail pointers, since we don't know if they were
1953          * updated on last use of the port.
1954          */
1955         head32 = ipath_read_ureg32(dd, ur_rcvegrindextail, pd->port_port);
1956         ipath_write_ureg(dd, ur_rcvegrindexhead, head32, pd->port_port);
1957         pd->port_lastrcvhdrqtail = -1;
1958         ipath_cdbg(VERBOSE, "Wrote port%d egrhead %x from tail regs\n",
1959                 pd->port_port, head32);
1960         pd->port_tidcursor = 0; /* start at beginning after open */
1961
1962         /* initialize poll variables... */
1963         pd->port_urgent = 0;
1964         pd->port_urgent_poll = 0;
1965         pd->port_hdrqfull_poll = pd->port_hdrqfull;
1966
1967         /*
1968          * Now enable the port for receive.
1969          * For chips that are set to DMA the tail register to memory
1970          * when they change (and when the update bit transitions from
1971          * 0 to 1.  So for those chips, we turn it off and then back on.
1972          * This will (very briefly) affect any other open ports, but the
1973          * duration is very short, and therefore isn't an issue.  We
1974          * explictly set the in-memory tail copy to 0 beforehand, so we
1975          * don't have to wait to be sure the DMA update has happened
1976          * (chip resets head/tail to 0 on transition to enable).
1977          */
1978         set_bit(dd->ipath_r_portenable_shift + pd->port_port,
1979                 &dd->ipath_rcvctrl);
1980         if (!(dd->ipath_flags & IPATH_NODMA_RTAIL)) {
1981                 if (pd->port_rcvhdrtail_kvaddr)
1982                         ipath_clear_rcvhdrtail(pd);
1983                 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1984                         dd->ipath_rcvctrl &
1985                         ~(1ULL << dd->ipath_r_tailupd_shift));
1986         }
1987         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1988                          dd->ipath_rcvctrl);
1989         /* Notify any waiting slaves */
1990         if (pd->port_subport_cnt) {
1991                 clear_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag);
1992                 wake_up(&pd->port_wait);
1993         }
1994 done:
1995         return ret;
1996 }
1997
1998 /**
1999  * unlock_exptid - unlock any expected TID entries port still had in use
2000  * @pd: port
2001  *
2002  * We don't actually update the chip here, because we do a bulk update
2003  * below, using ipath_f_clear_tids.
2004  */
2005 static void unlock_expected_tids(struct ipath_portdata *pd)
2006 {
2007         struct ipath_devdata *dd = pd->port_dd;
2008         int port_tidbase = pd->port_port * dd->ipath_rcvtidcnt;
2009         int i, cnt = 0, maxtid = port_tidbase + dd->ipath_rcvtidcnt;
2010
2011         ipath_cdbg(VERBOSE, "Port %u unlocking any locked expTID pages\n",
2012                    pd->port_port);
2013         for (i = port_tidbase; i < maxtid; i++) {
2014                 struct page *ps = dd->ipath_pageshadow[i];
2015
2016                 if (!ps)
2017                         continue;
2018
2019                 dd->ipath_pageshadow[i] = NULL;
2020                 pci_unmap_page(dd->pcidev, dd->ipath_physshadow[i],
2021                         PAGE_SIZE, PCI_DMA_FROMDEVICE);
2022                 ipath_release_user_pages_on_close(&ps, 1);
2023                 cnt++;
2024                 ipath_stats.sps_pageunlocks++;
2025         }
2026         if (cnt)
2027                 ipath_cdbg(VERBOSE, "Port %u locked %u expTID entries\n",
2028                            pd->port_port, cnt);
2029
2030         if (ipath_stats.sps_pagelocks || ipath_stats.sps_pageunlocks)
2031                 ipath_cdbg(VERBOSE, "%llu pages locked, %llu unlocked\n",
2032                            (unsigned long long) ipath_stats.sps_pagelocks,
2033                            (unsigned long long)
2034                            ipath_stats.sps_pageunlocks);
2035 }
2036
2037 static int ipath_close(struct inode *in, struct file *fp)
2038 {
2039         int ret = 0;
2040         struct ipath_filedata *fd;
2041         struct ipath_portdata *pd;
2042         struct ipath_devdata *dd;
2043         unsigned port;
2044
2045         ipath_cdbg(VERBOSE, "close on dev %lx, private data %p\n",
2046                    (long)in->i_rdev, fp->private_data);
2047
2048         mutex_lock(&ipath_mutex);
2049
2050         fd = (struct ipath_filedata *) fp->private_data;
2051         fp->private_data = NULL;
2052         pd = fd->pd;
2053         if (!pd) {
2054                 mutex_unlock(&ipath_mutex);
2055                 goto bail;
2056         }
2057
2058         dd = pd->port_dd;
2059
2060         /* drain user sdma queue */
2061         ipath_user_sdma_queue_drain(dd, fd->pq);
2062         ipath_user_sdma_queue_destroy(fd->pq);
2063
2064         if (--pd->port_cnt) {
2065                 /*
2066                  * XXX If the master closes the port before the slave(s),
2067                  * revoke the mmap for the eager receive queue so
2068                  * the slave(s) don't wait for receive data forever.
2069                  */
2070                 pd->active_slaves &= ~(1 << fd->subport);
2071                 put_pid(pd->port_subpid[fd->subport]);
2072                 pd->port_subpid[fd->subport] = NULL;
2073                 mutex_unlock(&ipath_mutex);
2074                 goto bail;
2075         }
2076         port = pd->port_port;
2077
2078         if (pd->port_hdrqfull) {
2079                 ipath_cdbg(PROC, "%s[%u] had %u rcvhdrqfull errors "
2080                            "during run\n", pd->port_comm, pid_nr(pd->port_pid),
2081                            pd->port_hdrqfull);
2082                 pd->port_hdrqfull = 0;
2083         }
2084
2085         if (pd->port_rcvwait_to || pd->port_piowait_to
2086             || pd->port_rcvnowait || pd->port_pionowait) {
2087                 ipath_cdbg(VERBOSE, "port%u, %u rcv, %u pio wait timeo; "
2088                            "%u rcv %u, pio already\n",
2089                            pd->port_port, pd->port_rcvwait_to,
2090                            pd->port_piowait_to, pd->port_rcvnowait,
2091                            pd->port_pionowait);
2092                 pd->port_rcvwait_to = pd->port_piowait_to =
2093                         pd->port_rcvnowait = pd->port_pionowait = 0;
2094         }
2095         if (pd->port_flag) {
2096                 ipath_cdbg(PROC, "port %u port_flag set: 0x%lx\n",
2097                           pd->port_port, pd->port_flag);
2098                 pd->port_flag = 0;
2099         }
2100
2101         if (dd->ipath_kregbase) {
2102                 /* atomically clear receive enable port and intr avail. */
2103                 clear_bit(dd->ipath_r_portenable_shift + port,
2104                           &dd->ipath_rcvctrl);
2105                 clear_bit(pd->port_port + dd->ipath_r_intravail_shift,
2106                           &dd->ipath_rcvctrl);
2107                 ipath_write_kreg( dd, dd->ipath_kregs->kr_rcvctrl,
2108                         dd->ipath_rcvctrl);
2109                 /* and read back from chip to be sure that nothing
2110                  * else is in flight when we do the rest */
2111                 (void)ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
2112
2113                 /* clean up the pkeys for this port user */
2114                 ipath_clean_part_key(pd, dd);
2115                 /*
2116                  * be paranoid, and never write 0's to these, just use an
2117                  * unused part of the port 0 tail page.  Of course,
2118                  * rcvhdraddr points to a large chunk of memory, so this
2119                  * could still trash things, but at least it won't trash
2120                  * page 0, and by disabling the port, it should stop "soon",
2121                  * even if a packet or two is in already in flight after we
2122                  * disabled the port.
2123                  */
2124                 ipath_write_kreg_port(dd,
2125                         dd->ipath_kregs->kr_rcvhdrtailaddr, port,
2126                         dd->ipath_dummy_hdrq_phys);
2127                 ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdraddr,
2128                         pd->port_port, dd->ipath_dummy_hdrq_phys);
2129
2130                 ipath_disarm_piobufs(dd, pd->port_pio_base, pd->port_piocnt);
2131                 ipath_chg_pioavailkernel(dd, pd->port_pio_base,
2132                         pd->port_piocnt, 1);
2133
2134                 dd->ipath_f_clear_tids(dd, pd->port_port);
2135
2136                 if (dd->ipath_pageshadow)
2137                         unlock_expected_tids(pd);
2138                 ipath_stats.sps_ports--;
2139                 ipath_cdbg(PROC, "%s[%u] closed port %u:%u\n",
2140                            pd->port_comm, pid_nr(pd->port_pid),
2141                            dd->ipath_unit, port);
2142         }
2143
2144         put_pid(pd->port_pid);
2145         pd->port_pid = NULL;
2146         dd->ipath_pd[pd->port_port] = NULL; /* before releasing mutex */
2147         mutex_unlock(&ipath_mutex);
2148         ipath_free_pddata(dd, pd); /* after releasing the mutex */
2149
2150 bail:
2151         kfree(fd);
2152         return ret;
2153 }
2154
2155 static int ipath_port_info(struct ipath_portdata *pd, u16 subport,
2156                            struct ipath_port_info __user *uinfo)
2157 {
2158         struct ipath_port_info info;
2159         int nup;
2160         int ret;
2161         size_t sz;
2162
2163         (void) ipath_count_units(NULL, &nup, NULL);
2164         info.num_active = nup;
2165         info.unit = pd->port_dd->ipath_unit;
2166         info.port = pd->port_port;
2167         info.subport = subport;
2168         /* Don't return new fields if old library opened the port. */
2169         if (ipath_supports_subports(pd->userversion >> 16,
2170                                     pd->userversion & 0xffff)) {
2171                 /* Number of user ports available for this device. */
2172                 info.num_ports = pd->port_dd->ipath_cfgports - 1;
2173                 info.num_subports = pd->port_subport_cnt;
2174                 sz = sizeof(info);
2175         } else
2176                 sz = sizeof(info) - 2 * sizeof(u16);
2177
2178         if (copy_to_user(uinfo, &info, sz)) {
2179                 ret = -EFAULT;
2180                 goto bail;
2181         }
2182         ret = 0;
2183
2184 bail:
2185         return ret;
2186 }
2187
2188 static int ipath_get_slave_info(struct ipath_portdata *pd,
2189                                 void __user *slave_mask_addr)
2190 {
2191         int ret = 0;
2192
2193         if (copy_to_user(slave_mask_addr, &pd->active_slaves, sizeof(u32)))
2194                 ret = -EFAULT;
2195         return ret;
2196 }
2197
2198 static int ipath_sdma_get_inflight(struct ipath_user_sdma_queue *pq,
2199                                    u32 __user *inflightp)
2200 {
2201         const u32 val = ipath_user_sdma_inflight_counter(pq);
2202
2203         if (put_user(val, inflightp))
2204                 return -EFAULT;
2205
2206         return 0;
2207 }
2208
2209 static int ipath_sdma_get_complete(struct ipath_devdata *dd,
2210                                    struct ipath_user_sdma_queue *pq,
2211                                    u32 __user *completep)
2212 {
2213         u32 val;
2214         int err;
2215
2216         err = ipath_user_sdma_make_progress(dd, pq);
2217         if (err < 0)
2218                 return err;
2219
2220         val = ipath_user_sdma_complete_counter(pq);
2221         if (put_user(val, completep))
2222                 return -EFAULT;
2223
2224         return 0;
2225 }
2226
2227 static ssize_t ipath_write(struct file *fp, const char __user *data,
2228                            size_t count, loff_t *off)
2229 {
2230         const struct ipath_cmd __user *ucmd;
2231         struct ipath_portdata *pd;
2232         const void __user *src;
2233         size_t consumed, copy;
2234         struct ipath_cmd cmd;
2235         ssize_t ret = 0;
2236         void *dest;
2237
2238         if (count < sizeof(cmd.type)) {
2239                 ret = -EINVAL;
2240                 goto bail;
2241         }
2242
2243         ucmd = (const struct ipath_cmd __user *) data;
2244
2245         if (copy_from_user(&cmd.type, &ucmd->type, sizeof(cmd.type))) {
2246                 ret = -EFAULT;
2247                 goto bail;
2248         }
2249
2250         consumed = sizeof(cmd.type);
2251
2252         switch (cmd.type) {
2253         case IPATH_CMD_ASSIGN_PORT:
2254         case __IPATH_CMD_USER_INIT:
2255         case IPATH_CMD_USER_INIT:
2256                 copy = sizeof(cmd.cmd.user_info);
2257                 dest = &cmd.cmd.user_info;
2258                 src = &ucmd->cmd.user_info;
2259                 break;
2260         case IPATH_CMD_RECV_CTRL:
2261                 copy = sizeof(cmd.cmd.recv_ctrl);
2262                 dest = &cmd.cmd.recv_ctrl;
2263                 src = &ucmd->cmd.recv_ctrl;
2264                 break;
2265         case IPATH_CMD_PORT_INFO:
2266                 copy = sizeof(cmd.cmd.port_info);
2267                 dest = &cmd.cmd.port_info;
2268                 src = &ucmd->cmd.port_info;
2269                 break;
2270         case IPATH_CMD_TID_UPDATE:
2271         case IPATH_CMD_TID_FREE:
2272                 copy = sizeof(cmd.cmd.tid_info);
2273                 dest = &cmd.cmd.tid_info;
2274                 src = &ucmd->cmd.tid_info;
2275                 break;
2276         case IPATH_CMD_SET_PART_KEY:
2277                 copy = sizeof(cmd.cmd.part_key);
2278                 dest = &cmd.cmd.part_key;
2279                 src = &ucmd->cmd.part_key;
2280                 break;
2281         case __IPATH_CMD_SLAVE_INFO:
2282                 copy = sizeof(cmd.cmd.slave_mask_addr);
2283                 dest = &cmd.cmd.slave_mask_addr;
2284                 src = &ucmd->cmd.slave_mask_addr;
2285                 break;
2286         case IPATH_CMD_PIOAVAILUPD:     // force an update of PIOAvail reg
2287                 copy = 0;
2288                 src = NULL;
2289                 dest = NULL;
2290                 break;
2291         case IPATH_CMD_POLL_TYPE:
2292                 copy = sizeof(cmd.cmd.poll_type);
2293                 dest = &cmd.cmd.poll_type;
2294                 src = &ucmd->cmd.poll_type;
2295                 break;
2296         case IPATH_CMD_ARMLAUNCH_CTRL:
2297                 copy = sizeof(cmd.cmd.armlaunch_ctrl);
2298                 dest = &cmd.cmd.armlaunch_ctrl;
2299                 src = &ucmd->cmd.armlaunch_ctrl;
2300                 break;
2301         case IPATH_CMD_SDMA_INFLIGHT:
2302                 copy = sizeof(cmd.cmd.sdma_inflight);
2303                 dest = &cmd.cmd.sdma_inflight;
2304                 src = &ucmd->cmd.sdma_inflight;
2305                 break;
2306         case IPATH_CMD_SDMA_COMPLETE:
2307                 copy = sizeof(cmd.cmd.sdma_complete);
2308                 dest = &cmd.cmd.sdma_complete;
2309                 src = &ucmd->cmd.sdma_complete;
2310                 break;
2311         default:
2312                 ret = -EINVAL;
2313                 goto bail;
2314         }
2315
2316         if (copy) {
2317                 if ((count - consumed) < copy) {
2318                         ret = -EINVAL;
2319                         goto bail;
2320                 }
2321
2322                 if (copy_from_user(dest, src, copy)) {
2323                         ret = -EFAULT;
2324                         goto bail;
2325                 }
2326
2327                 consumed += copy;
2328         }
2329
2330         pd = port_fp(fp);
2331         if (!pd && cmd.type != __IPATH_CMD_USER_INIT &&
2332                 cmd.type != IPATH_CMD_ASSIGN_PORT) {
2333                 ret = -EINVAL;
2334                 goto bail;
2335         }
2336
2337         switch (cmd.type) {
2338         case IPATH_CMD_ASSIGN_PORT:
2339                 ret = ipath_assign_port(fp, &cmd.cmd.user_info);
2340                 if (ret)
2341                         goto bail;
2342                 break;
2343         case __IPATH_CMD_USER_INIT:
2344                 /* backwards compatibility, get port first */
2345                 ret = ipath_assign_port(fp, &cmd.cmd.user_info);
2346                 if (ret)
2347                         goto bail;
2348                 /* and fall through to current version. */
2349         case IPATH_CMD_USER_INIT:
2350                 ret = ipath_do_user_init(fp, &cmd.cmd.user_info);
2351                 if (ret)
2352                         goto bail;
2353                 ret = ipath_get_base_info(
2354                         fp, (void __user *) (unsigned long)
2355                         cmd.cmd.user_info.spu_base_info,
2356                         cmd.cmd.user_info.spu_base_info_size);
2357                 break;
2358         case IPATH_CMD_RECV_CTRL:
2359                 ret = ipath_manage_rcvq(pd, subport_fp(fp), cmd.cmd.recv_ctrl);
2360                 break;
2361         case IPATH_CMD_PORT_INFO:
2362                 ret = ipath_port_info(pd, subport_fp(fp),
2363                                       (struct ipath_port_info __user *)
2364                                       (unsigned long) cmd.cmd.port_info);
2365                 break;
2366         case IPATH_CMD_TID_UPDATE:
2367                 ret = ipath_tid_update(pd, fp, &cmd.cmd.tid_info);
2368                 break;
2369         case IPATH_CMD_TID_FREE:
2370                 ret = ipath_tid_free(pd, subport_fp(fp), &cmd.cmd.tid_info);
2371                 break;
2372         case IPATH_CMD_SET_PART_KEY:
2373                 ret = ipath_set_part_key(pd, cmd.cmd.part_key);
2374                 break;
2375         case __IPATH_CMD_SLAVE_INFO:
2376                 ret = ipath_get_slave_info(pd,
2377                                            (void __user *) (unsigned long)
2378                                            cmd.cmd.slave_mask_addr);
2379                 break;
2380         case IPATH_CMD_PIOAVAILUPD:
2381                 ipath_force_pio_avail_update(pd->port_dd);
2382                 break;
2383         case IPATH_CMD_POLL_TYPE:
2384                 pd->poll_type = cmd.cmd.poll_type;
2385                 break;
2386         case IPATH_CMD_ARMLAUNCH_CTRL:
2387                 if (cmd.cmd.armlaunch_ctrl)
2388                         ipath_enable_armlaunch(pd->port_dd);
2389                 else
2390                         ipath_disable_armlaunch(pd->port_dd);
2391                 break;
2392         case IPATH_CMD_SDMA_INFLIGHT:
2393                 ret = ipath_sdma_get_inflight(user_sdma_queue_fp(fp),
2394                                               (u32 __user *) (unsigned long)
2395                                               cmd.cmd.sdma_inflight);
2396                 break;
2397         case IPATH_CMD_SDMA_COMPLETE:
2398                 ret = ipath_sdma_get_complete(pd->port_dd,
2399                                               user_sdma_queue_fp(fp),
2400                                               (u32 __user *) (unsigned long)
2401                                               cmd.cmd.sdma_complete);
2402                 break;
2403         }
2404
2405         if (ret >= 0)
2406                 ret = consumed;
2407
2408 bail:
2409         return ret;
2410 }
2411
2412 static ssize_t ipath_writev(struct kiocb *iocb, const struct iovec *iov,
2413                             unsigned long dim, loff_t off)
2414 {
2415         struct file *filp = iocb->ki_filp;
2416         struct ipath_filedata *fp = filp->private_data;
2417         struct ipath_portdata *pd = port_fp(filp);
2418         struct ipath_user_sdma_queue *pq = fp->pq;
2419
2420         if (!dim)
2421                 return -EINVAL;
2422
2423         return ipath_user_sdma_writev(pd->port_dd, pq, iov, dim);
2424 }
2425
2426 static struct class *ipath_class;
2427
2428 static int init_cdev(int minor, char *name, const struct file_operations *fops,
2429                      struct cdev **cdevp, struct device **devp)
2430 {
2431         const dev_t dev = MKDEV(IPATH_MAJOR, minor);
2432         struct cdev *cdev = NULL;
2433         struct device *device = NULL;
2434         int ret;
2435
2436         cdev = cdev_alloc();
2437         if (!cdev) {
2438                 printk(KERN_ERR IPATH_DRV_NAME
2439                        ": Could not allocate cdev for minor %d, %s\n",
2440                        minor, name);
2441                 ret = -ENOMEM;
2442                 goto done;
2443         }
2444
2445         cdev->owner = THIS_MODULE;
2446         cdev->ops = fops;
2447         kobject_set_name(&cdev->kobj, name);
2448
2449         ret = cdev_add(cdev, dev, 1);
2450         if (ret < 0) {
2451                 printk(KERN_ERR IPATH_DRV_NAME
2452                        ": Could not add cdev for minor %d, %s (err %d)\n",
2453                        minor, name, -ret);
2454                 goto err_cdev;
2455         }
2456
2457         device = device_create(ipath_class, NULL, dev, name);
2458
2459         if (IS_ERR(device)) {
2460                 ret = PTR_ERR(device);
2461                 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2462                        "device for minor %d, %s (err %d)\n",
2463                        minor, name, -ret);
2464                 goto err_cdev;
2465         }
2466
2467         goto done;
2468
2469 err_cdev:
2470         cdev_del(cdev);
2471         cdev = NULL;
2472
2473 done:
2474         if (ret >= 0) {
2475                 *cdevp = cdev;
2476                 *devp = device;
2477         } else {
2478                 *cdevp = NULL;
2479                 *devp = NULL;
2480         }
2481
2482         return ret;
2483 }
2484
2485 int ipath_cdev_init(int minor, char *name, const struct file_operations *fops,
2486                     struct cdev **cdevp, struct device **devp)
2487 {
2488         return init_cdev(minor, name, fops, cdevp, devp);
2489 }
2490
2491 static void cleanup_cdev(struct cdev **cdevp,
2492                          struct device **devp)
2493 {
2494         struct device *dev = *devp;
2495
2496         if (dev) {
2497                 device_unregister(dev);
2498                 *devp = NULL;
2499         }
2500
2501         if (*cdevp) {
2502                 cdev_del(*cdevp);
2503                 *cdevp = NULL;
2504         }
2505 }
2506
2507 void ipath_cdev_cleanup(struct cdev **cdevp,
2508                         struct device **devp)
2509 {
2510         cleanup_cdev(cdevp, devp);
2511 }
2512
2513 static struct cdev *wildcard_cdev;
2514 static struct device *wildcard_dev;
2515
2516 static const dev_t dev = MKDEV(IPATH_MAJOR, 0);
2517
2518 static int user_init(void)
2519 {
2520         int ret;
2521
2522         ret = register_chrdev_region(dev, IPATH_NMINORS, IPATH_DRV_NAME);
2523         if (ret < 0) {
2524                 printk(KERN_ERR IPATH_DRV_NAME ": Could not register "
2525                        "chrdev region (err %d)\n", -ret);
2526                 goto done;
2527         }
2528
2529         ipath_class = class_create(THIS_MODULE, IPATH_DRV_NAME);
2530
2531         if (IS_ERR(ipath_class)) {
2532                 ret = PTR_ERR(ipath_class);
2533                 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2534                        "device class (err %d)\n", -ret);
2535                 goto bail;
2536         }
2537
2538         goto done;
2539 bail:
2540         unregister_chrdev_region(dev, IPATH_NMINORS);
2541 done:
2542         return ret;
2543 }
2544
2545 static void user_cleanup(void)
2546 {
2547         if (ipath_class) {
2548                 class_destroy(ipath_class);
2549                 ipath_class = NULL;
2550         }
2551
2552         unregister_chrdev_region(dev, IPATH_NMINORS);
2553 }
2554
2555 static atomic_t user_count = ATOMIC_INIT(0);
2556 static atomic_t user_setup = ATOMIC_INIT(0);
2557
2558 int ipath_user_add(struct ipath_devdata *dd)
2559 {
2560         char name[10];
2561         int ret;
2562
2563         if (atomic_inc_return(&user_count) == 1) {
2564                 ret = user_init();
2565                 if (ret < 0) {
2566                         ipath_dev_err(dd, "Unable to set up user support: "
2567                                       "error %d\n", -ret);
2568                         goto bail;
2569                 }
2570                 ret = init_cdev(0, "ipath", &ipath_file_ops, &wildcard_cdev,
2571                                 &wildcard_dev);
2572                 if (ret < 0) {
2573                         ipath_dev_err(dd, "Could not create wildcard "
2574                                       "minor: error %d\n", -ret);
2575                         goto bail_user;
2576                 }
2577
2578                 atomic_set(&user_setup, 1);
2579         }
2580
2581         snprintf(name, sizeof(name), "ipath%d", dd->ipath_unit);
2582
2583         ret = init_cdev(dd->ipath_unit + 1, name, &ipath_file_ops,
2584                         &dd->user_cdev, &dd->user_dev);
2585         if (ret < 0)
2586                 ipath_dev_err(dd, "Could not create user minor %d, %s\n",
2587                               dd->ipath_unit + 1, name);
2588
2589         goto bail;
2590
2591 bail_user:
2592         user_cleanup();
2593 bail:
2594         return ret;
2595 }
2596
2597 void ipath_user_remove(struct ipath_devdata *dd)
2598 {
2599         cleanup_cdev(&dd->user_cdev, &dd->user_dev);
2600
2601         if (atomic_dec_return(&user_count) == 0) {
2602                 if (atomic_read(&user_setup) == 0)
2603                         goto bail;
2604
2605                 cleanup_cdev(&wildcard_cdev, &wildcard_dev);
2606                 user_cleanup();
2607
2608                 atomic_set(&user_setup, 0);
2609         }
2610 bail:
2611         return;
2612 }