drivers/staging/rdma/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/export.h>
  39 #include <linux/vmalloc.h>
  40 #include <linux/slab.h>
  41 #include <linux/highmem.h>
  42 #include <linux/io.h>
  43 #include <linux/jiffies.h>
  44 #include <linux/cpu.h>
  45 #include <linux/uio.h>
  46 #include <asm/pgtable.h>
  47
  48 #include <rdma/ib.h>
  49
  50 #include "ipath_kernel.h"
  51 #include "ipath_common.h"
  52 #include "ipath_user_sdma.h"
  53
  54 static int ipath_open(struct inode *, struct file *);
  55 static int ipath_close(struct inode *, struct file *);
  56 static ssize_t ipath_write(struct file *, const char __user *, size_t,
  57                            loff_t *);
  58 static ssize_t ipath_write_iter(struct kiocb *, struct iov_iter *from);
  59 static unsigned int ipath_poll(struct file *, struct poll_table_struct *);
  60 static int ipath_mmap(struct file *, struct vm_area_struct *);
  61
  62 /*
  63  * This is really, really weird shit - write() and writev() here
  64  * have completely unrelated semantics.  Sucky userland ABI,
  65  * film at 11.
  66  */
  67 static const struct file_operations ipath_file_ops = {
  68         .owner = THIS_MODULE,
  69         .write = ipath_write,
  70         .write_iter = ipath_write_iter,
  71         .open = ipath_open,
  72         .release = ipath_close,
  73         .poll = ipath_poll,
  74         .mmap = ipath_mmap,
  75         .llseek = noop_llseek,
  76 };
  77
  78 /*
  79  * Convert kernel virtual addresses to physical addresses so they don't
  80  * potentially conflict with the chip addresses used as mmap offsets.
  81  * It doesn't really matter what mmap offset we use as long as we can
  82  * interpret it correctly.
  83  */
  84 static u64 cvt_kvaddr(void *p)
  85 {
  86         struct page *page;
  87         u64 paddr = 0;
  88
  89         page = vmalloc_to_page(p);
  90         if (page)
  91                 paddr = page_to_pfn(page) << PAGE_SHIFT;
  92
  93         return paddr;
  94 }
  95
  96 static int ipath_get_base_info(struct file *fp,
  97                                void __user *ubase, size_t ubase_size)
  98 {
  99         struct ipath_portdata *pd = port_fp(fp);
 100         int ret = 0;
 101         struct ipath_base_info *kinfo = NULL;
 102         struct ipath_devdata *dd = pd->port_dd;
 103         unsigned subport_cnt;
 104         int shared, master;
 105         size_t sz;
 106
 107         subport_cnt = pd->port_subport_cnt;
 108         if (!subport_cnt) {
 109                 shared = 0;
 110                 master = 0;
 111                 subport_cnt = 1;
 112         } else {
 113                 shared = 1;
 114                 master = !subport_fp(fp);
 115         }
 116
 117         sz = sizeof(*kinfo);
 118         /* If port sharing is not requested, allow the old size structure */
 119         if (!shared)
 120                 sz -= 7 * sizeof(u64);
 121         if (ubase_size < sz) {
 122                 ipath_cdbg(PROC,
 123                            "Base size %zu, need %zu (version mismatch?)\n",
 124                            ubase_size, sz);
 125                 ret = -EINVAL;
 126                 goto bail;
 127         }
 128
 129         kinfo = kzalloc(sizeof(*kinfo), GFP_KERNEL);
 130         if (kinfo == NULL) {
 131                 ret = -ENOMEM;
 132                 goto bail;
 133         }
 134
 135         ret = dd->ipath_f_get_base_info(pd, kinfo);
 136         if (ret < 0)
 137                 goto bail;
 138
 139         kinfo->spi_rcvhdr_cnt = dd->ipath_rcvhdrcnt;
 140         kinfo->spi_rcvhdrent_size = dd->ipath_rcvhdrentsize;
 141         kinfo->spi_tidegrcnt = dd->ipath_rcvegrcnt;
 142         kinfo->spi_rcv_egrbufsize = dd->ipath_rcvegrbufsize;
 143         /*
 144          * have to mmap whole thing
 145          */
 146         kinfo->spi_rcv_egrbuftotlen =
 147                 pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
 148         kinfo->spi_rcv_egrperchunk = pd->port_rcvegrbufs_perchunk;
 149         kinfo->spi_rcv_egrchunksize = kinfo->spi_rcv_egrbuftotlen /
 150                 pd->port_rcvegrbuf_chunks;
 151         kinfo->spi_tidcnt = dd->ipath_rcvtidcnt / subport_cnt;
 152         if (master)
 153                 kinfo->spi_tidcnt += dd->ipath_rcvtidcnt % subport_cnt;
 154         /*
 155          * for this use, may be ipath_cfgports summed over all chips that
 156          * are are configured and present
 157          */
 158         kinfo->spi_nports = dd->ipath_cfgports;
 159         /* unit (chip/board) our port is on */
 160         kinfo->spi_unit = dd->ipath_unit;
 161         /* for now, only a single page */
 162         kinfo->spi_tid_maxsize = PAGE_SIZE;
 163
 164         /*
 165          * Doing this per port, and based on the skip value, etc.  This has
 166          * to be the actual buffer size, since the protocol code treats it
 167          * as an array.
 168          *
 169          * These have to be set to user addresses in the user code via mmap.
 170          * These values are used on return to user code for the mmap target
 171          * addresses only.  For 32 bit, same 44 bit address problem, so use
 172          * the physical address, not virtual.  Before 2.6.11, using the
 173          * page_address() macro worked, but in 2.6.11, even that returns the
 174          * full 64 bit address (upper bits all 1's).  So far, using the
 175          * physical addresses (or chip offsets, for chip mapping) works, but
 176          * no doubt some future kernel release will change that, and we'll be
 177          * on to yet another method of dealing with this.
 178          */
 179         kinfo->spi_rcvhdr_base = (u64) pd->port_rcvhdrq_phys;
 180         kinfo->spi_rcvhdr_tailaddr = (u64) pd->port_rcvhdrqtailaddr_phys;
 181         kinfo->spi_rcv_egrbufs = (u64) pd->port_rcvegr_phys;
 182         kinfo->spi_pioavailaddr = (u64) dd->ipath_pioavailregs_phys;
 183         kinfo->spi_status = (u64) kinfo->spi_pioavailaddr +
 184                 (void *) dd->ipath_statusp -
 185                 (void *) dd->ipath_pioavailregs_dma;
 186         if (!shared) {
 187                 kinfo->spi_piocnt = pd->port_piocnt;
 188                 kinfo->spi_piobufbase = (u64) pd->port_piobufs;
 189                 kinfo->__spi_uregbase = (u64) dd->ipath_uregbase +
 190                         dd->ipath_ureg_align * pd->port_port;
 191         } else if (master) {
 192                 kinfo->spi_piocnt = (pd->port_piocnt / subport_cnt) +
 193                                     (pd->port_piocnt % subport_cnt);
 194                 /* Master's PIO buffers are after all the slave's */
 195                 kinfo->spi_piobufbase = (u64) pd->port_piobufs +
 196                         dd->ipath_palign *
 197                         (pd->port_piocnt - kinfo->spi_piocnt);
 198         } else {
 199                 unsigned slave = subport_fp(fp) - 1;
 200
 201                 kinfo->spi_piocnt = pd->port_piocnt / subport_cnt;
 202                 kinfo->spi_piobufbase = (u64) pd->port_piobufs +
 203                         dd->ipath_palign * kinfo->spi_piocnt * slave;
 204         }
 205
 206         if (shared) {
 207                 kinfo->spi_port_uregbase = (u64) dd->ipath_uregbase +
 208                         dd->ipath_ureg_align * pd->port_port;
 209                 kinfo->spi_port_rcvegrbuf = kinfo->spi_rcv_egrbufs;
 210                 kinfo->spi_port_rcvhdr_base = kinfo->spi_rcvhdr_base;
 211                 kinfo->spi_port_rcvhdr_tailaddr = kinfo->spi_rcvhdr_tailaddr;
 212
 213                 kinfo->__spi_uregbase = cvt_kvaddr(pd->subport_uregbase +
 214                         PAGE_SIZE * subport_fp(fp));
 215
 216                 kinfo->spi_rcvhdr_base = cvt_kvaddr(pd->subport_rcvhdr_base +
 217                         pd->port_rcvhdrq_size * subport_fp(fp));
 218                 kinfo->spi_rcvhdr_tailaddr = 0;
 219                 kinfo->spi_rcv_egrbufs = cvt_kvaddr(pd->subport_rcvegrbuf +
 220                         pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size *
 221                         subport_fp(fp));
 222
 223                 kinfo->spi_subport_uregbase =
 224                         cvt_kvaddr(pd->subport_uregbase);
 225                 kinfo->spi_subport_rcvegrbuf =
 226                         cvt_kvaddr(pd->subport_rcvegrbuf);
 227                 kinfo->spi_subport_rcvhdr_base =
 228                         cvt_kvaddr(pd->subport_rcvhdr_base);
 229                 ipath_cdbg(PROC, "port %u flags %x %llx %llx %llx\n",
 230                         kinfo->spi_port, kinfo->spi_runtime_flags,
 231                         (unsigned long long) kinfo->spi_subport_uregbase,
 232                         (unsigned long long) kinfo->spi_subport_rcvegrbuf,
 233                         (unsigned long long) kinfo->spi_subport_rcvhdr_base);
 234         }
 235
 236         /*
 237          * All user buffers are 2KB buffers.  If we ever support
 238          * giving 4KB buffers to user processes, this will need some
 239          * work.
 240          */
 241         kinfo->spi_pioindex = (kinfo->spi_piobufbase -
 242                 (dd->ipath_piobufbase & 0xffffffff)) / dd->ipath_palign;
 243         kinfo->spi_pioalign = dd->ipath_palign;
 244
 245         kinfo->spi_qpair = IPATH_KD_QP;
 246         /*
 247          * user mode PIO buffers are always 2KB, even when 4KB can
 248          * be received, and sent via the kernel; this is ibmaxlen
 249          * for 2K MTU.
 250          */
 251         kinfo->spi_piosize = dd->ipath_piosize2k - 2 * sizeof(u32);
 252         kinfo->spi_mtu = dd->ipath_ibmaxlen;    /* maxlen, not ibmtu */
 253         kinfo->spi_port = pd->port_port;
 254         kinfo->spi_subport = subport_fp(fp);
 255         kinfo->spi_sw_version = IPATH_KERN_SWVERSION;
 256         kinfo->spi_hw_version = dd->ipath_revision;
 257
 258         if (master) {
 259                 kinfo->spi_runtime_flags |= IPATH_RUNTIME_MASTER;
 260         }
 261
 262         sz = (ubase_size < sizeof(*kinfo)) ? ubase_size : sizeof(*kinfo);
 263         if (copy_to_user(ubase, kinfo, sz))
 264                 ret = -EFAULT;
 265
 266 bail:
 267         kfree(kinfo);
 268         return ret;
 269 }
 270
 271 /**
 272  * ipath_tid_update - update a port TID
 273  * @pd: the port
 274  * @fp: the ipath device file
 275  * @ti: the TID information
 276  *
 277  * The new implementation as of Oct 2004 is that the driver assigns
 278  * the tid and returns it to the caller.   To make it easier to
 279  * catch bugs, and to reduce search time, we keep a cursor for
 280  * each port, walking the shadow tid array to find one that's not
 281  * in use.
 282  *
 283  * For now, if we can't allocate the full list, we fail, although
 284  * in the long run, we'll allocate as many as we can, and the
 285  * caller will deal with that by trying the remaining pages later.
 286  * That means that when we fail, we have to mark the tids as not in
 287  * use again, in our shadow copy.
 288  *
 289  * It's up to the caller to free the tids when they are done.
 290  * We'll unlock the pages as they free them.
 291  *
 292  * Also, right now we are locking one page at a time, but since
 293  * the intended use of this routine is for a single group of
 294  * virtually contiguous pages, that should change to improve
 295  * performance.
 296  */
 297 static int ipath_tid_update(struct ipath_portdata *pd, struct file *fp,
 298                             const struct ipath_tid_info *ti)
 299 {
 300         int ret = 0, ntids;
 301         u32 tid, porttid, cnt, i, tidcnt, tidoff;
 302         u16 *tidlist;
 303         struct ipath_devdata *dd = pd->port_dd;
 304         u64 physaddr;
 305         unsigned long vaddr;
 306         u64 __iomem *tidbase;
 307         unsigned long tidmap[8];
 308         struct page **pagep = NULL;
 309         unsigned subport = subport_fp(fp);
 310
 311         if (!dd->ipath_pageshadow) {
 312                 ret = -ENOMEM;
 313                 goto done;
 314         }
 315
 316         cnt = ti->tidcnt;
 317         if (!cnt) {
 318                 ipath_dbg("After copyin, tidcnt 0, tidlist %llx\n",
 319                           (unsigned long long) ti->tidlist);
 320                 /*
 321                  * Should we treat as success?  likely a bug
 322                  */
 323                 ret = -EFAULT;
 324                 goto done;
 325         }
 326         porttid = pd->port_port * dd->ipath_rcvtidcnt;
 327         if (!pd->port_subport_cnt) {
 328                 tidcnt = dd->ipath_rcvtidcnt;
 329                 tid = pd->port_tidcursor;
 330                 tidoff = 0;
 331         } else if (!subport) {
 332                 tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
 333                          (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
 334                 tidoff = dd->ipath_rcvtidcnt - tidcnt;
 335                 porttid += tidoff;
 336                 tid = tidcursor_fp(fp);
 337         } else {
 338                 tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
 339                 tidoff = tidcnt * (subport - 1);
 340                 porttid += tidoff;
 341                 tid = tidcursor_fp(fp);
 342         }
 343         if (cnt > tidcnt) {
 344                 /* make sure it all fits in port_tid_pg_list */
 345                 dev_info(&dd->pcidev->dev, "Process tried to allocate %u "
 346                          "TIDs, only trying max (%u)\n", cnt, tidcnt);
 347                 cnt = tidcnt;
 348         }
 349         pagep = &((struct page **) pd->port_tid_pg_list)[tidoff];
 350         tidlist = &((u16 *) &pagep[dd->ipath_rcvtidcnt])[tidoff];
 351
 352         memset(tidmap, 0, sizeof(tidmap));
 353         /* before decrement; chip actual # */
 354         ntids = tidcnt;
 355         tidbase = (u64 __iomem *) (((char __iomem *) dd->ipath_kregbase) +
 356                                    dd->ipath_rcvtidbase +
 357                                    porttid * sizeof(*tidbase));
 358
 359         ipath_cdbg(VERBOSE, "Port%u %u tids, cursor %u, tidbase %p\n",
 360                    pd->port_port, cnt, tid, tidbase);
 361
 362         /* virtual address of first page in transfer */
 363         vaddr = ti->tidvaddr;
 364         if (!access_ok(VERIFY_WRITE, (void __user *) vaddr,
 365                        cnt * PAGE_SIZE)) {
 366                 ipath_dbg("Fail vaddr %p, %u pages, !access_ok\n",
 367                           (void *)vaddr, cnt);
 368                 ret = -EFAULT;
 369                 goto done;
 370         }
 371         ret = ipath_get_user_pages(vaddr, cnt, pagep);
 372         if (ret) {
 373                 if (ret == -EBUSY) {
 374                         ipath_dbg("Failed to lock addr %p, %u pages "
 375                                   "(already locked)\n",
 376                                   (void *) vaddr, cnt);
 377                         /*
 378                          * for now, continue, and see what happens but with
 379                          * the new implementation, this should never happen,
 380                          * unless perhaps the user has mpin'ed the pages
 381                          * themselves (something we need to test)
 382                          */
 383                         ret = 0;
 384                 } else {
 385                         dev_info(&dd->pcidev->dev,
 386                                  "Failed to lock addr %p, %u pages: "
 387                                  "errno %d\n", (void *) vaddr, cnt, -ret);
 388                         goto done;
 389                 }
 390         }
 391         for (i = 0; i < cnt; i++, vaddr += PAGE_SIZE) {
 392                 for (; ntids--; tid++) {
 393                         if (tid == tidcnt)
 394                                 tid = 0;
 395                         if (!dd->ipath_pageshadow[porttid + tid])
 396                                 break;
 397                 }
 398                 if (ntids < 0) {
 399                         /*
 400                          * oops, wrapped all the way through their TIDs,
 401                          * and didn't have enough free; see comments at
 402                          * start of routine
 403                          */
 404                         ipath_dbg("Not enough free TIDs for %u pages "
 405                                   "(index %d), failing\n", cnt, i);
 406                         i--;    /* last tidlist[i] not filled in */
 407                         ret = -ENOMEM;
 408                         break;
 409                 }
 410                 tidlist[i] = tid + tidoff;
 411                 ipath_cdbg(VERBOSE, "Updating idx %u to TID %u, "
 412                            "vaddr %lx\n", i, tid + tidoff, vaddr);
 413                 /* we "know" system pages and TID pages are same size */
 414                 dd->ipath_pageshadow[porttid + tid] = pagep[i];
 415                 dd->ipath_physshadow[porttid + tid] = ipath_map_page(
 416                         dd->pcidev, pagep[i], 0, PAGE_SIZE,
 417                         PCI_DMA_FROMDEVICE);
 418                 /*
 419                  * don't need atomic or it's overhead
 420                  */
 421                 __set_bit(tid, tidmap);
 422                 physaddr = dd->ipath_physshadow[porttid + tid];
 423                 ipath_stats.sps_pagelocks++;
 424                 ipath_cdbg(VERBOSE,
 425                            "TID %u, vaddr %lx, physaddr %llx pgp %p\n",
 426                            tid, vaddr, (unsigned long long) physaddr,
 427                            pagep[i]);
 428                 dd->ipath_f_put_tid(dd, &tidbase[tid], RCVHQ_RCV_TYPE_EXPECTED,
 429                                     physaddr);
 430                 /*
 431                  * don't check this tid in ipath_portshadow, since we
 432                  * just filled it in; start with the next one.
 433                  */
 434                 tid++;
 435         }
 436
 437         if (ret) {
 438                 u32 limit;
 439         cleanup:
 440                 /* jump here if copy out of updated info failed... */
 441                 ipath_dbg("After failure (ret=%d), undo %d of %d entries\n",
 442                           -ret, i, cnt);
 443                 /* same code that's in ipath_free_tid() */
 444                 limit = sizeof(tidmap) * BITS_PER_BYTE;
 445                 if (limit > tidcnt)
 446                         /* just in case size changes in future */
 447                         limit = tidcnt;
 448                 tid = find_first_bit((const unsigned long *)tidmap, limit);
 449                 for (; tid < limit; tid++) {
 450                         if (!test_bit(tid, tidmap))
 451                                 continue;
 452                         if (dd->ipath_pageshadow[porttid + tid]) {
 453                                 ipath_cdbg(VERBOSE, "Freeing TID %u\n",
 454                                            tid);
 455                                 dd->ipath_f_put_tid(dd, &tidbase[tid],
 456                                                     RCVHQ_RCV_TYPE_EXPECTED,
 457                                                     dd->ipath_tidinvalid);
 458                                 pci_unmap_page(dd->pcidev,
 459                                         dd->ipath_physshadow[porttid + tid],
 460                                         PAGE_SIZE, PCI_DMA_FROMDEVICE);
 461                                 dd->ipath_pageshadow[porttid + tid] = NULL;
 462                                 ipath_stats.sps_pageunlocks++;
 463                         }
 464                 }
 465                 ipath_release_user_pages(pagep, cnt);
 466         } else {
 467                 /*
 468                  * Copy the updated array, with ipath_tid's filled in, back
 469                  * to user.  Since we did the copy in already, this "should
 470                  * never fail" If it does, we have to clean up...
 471                  */
 472                 if (copy_to_user((void __user *)
 473                                  (unsigned long) ti->tidlist,
 474                                  tidlist, cnt * sizeof(*tidlist))) {
 475                         ret = -EFAULT;
 476                         goto cleanup;
 477                 }
 478                 if (copy_to_user((void __user *) (unsigned long) ti->tidmap,
 479                                  tidmap, sizeof tidmap)) {
 480                         ret = -EFAULT;
 481                         goto cleanup;
 482                 }
 483                 if (tid == tidcnt)
 484                         tid = 0;
 485                 if (!pd->port_subport_cnt)
 486                         pd->port_tidcursor = tid;
 487                 else
 488                         tidcursor_fp(fp) = tid;
 489         }
 490
 491 done:
 492         if (ret)
 493                 ipath_dbg("Failed to map %u TID pages, failing with %d\n",
 494                           ti->tidcnt, -ret);
 495         return ret;
 496 }
 497
 498 /**
 499  * ipath_tid_free - free a port TID
 500  * @pd: the port
 501  * @subport: the subport
 502  * @ti: the TID info
 503  *
 504  * right now we are unlocking one page at a time, but since
 505  * the intended use of this routine is for a single group of
 506  * virtually contiguous pages, that should change to improve
 507  * performance.  We check that the TID is in range for this port
 508  * but otherwise don't check validity; if user has an error and
 509  * frees the wrong tid, it's only their own data that can thereby
 510  * be corrupted.  We do check that the TID was in use, for sanity
 511  * We always use our idea of the saved address, not the address that
 512  * they pass in to us.
 513  */
 514
 515 static int ipath_tid_free(struct ipath_portdata *pd, unsigned subport,
 516                           const struct ipath_tid_info *ti)
 517 {
 518         int ret = 0;
 519         u32 tid, porttid, cnt, limit, tidcnt;
 520         struct ipath_devdata *dd = pd->port_dd;
 521         u64 __iomem *tidbase;
 522         unsigned long tidmap[8];
 523
 524         if (!dd->ipath_pageshadow) {
 525                 ret = -ENOMEM;
 526                 goto done;
 527         }
 528
 529         if (copy_from_user(tidmap, (void __user *)(unsigned long)ti->tidmap,
 530                            sizeof tidmap)) {
 531                 ret = -EFAULT;
 532                 goto done;
 533         }
 534
 535         porttid = pd->port_port * dd->ipath_rcvtidcnt;
 536         if (!pd->port_subport_cnt)
 537                 tidcnt = dd->ipath_rcvtidcnt;
 538         else if (!subport) {
 539                 tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
 540                          (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
 541                 porttid += dd->ipath_rcvtidcnt - tidcnt;
 542         } else {
 543                 tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
 544                 porttid += tidcnt * (subport - 1);
 545         }
 546         tidbase = (u64 __iomem *) ((char __iomem *)(dd->ipath_kregbase) +
 547                                    dd->ipath_rcvtidbase +
 548                                    porttid * sizeof(*tidbase));
 549
 550         limit = sizeof(tidmap) * BITS_PER_BYTE;
 551         if (limit > tidcnt)
 552                 /* just in case size changes in future */
 553                 limit = tidcnt;
 554         tid = find_first_bit(tidmap, limit);
 555         ipath_cdbg(VERBOSE, "Port%u free %u tids; first bit (max=%d) "
 556                    "set is %d, porttid %u\n", pd->port_port, ti->tidcnt,
 557                    limit, tid, porttid);
 558         for (cnt = 0; tid < limit; tid++) {
 559                 /*
 560                  * small optimization; if we detect a run of 3 or so without
 561                  * any set, use find_first_bit again.  That's mainly to
 562                  * accelerate the case where we wrapped, so we have some at
 563                  * the beginning, and some at the end, and a big gap
 564                  * in the middle.
 565                  */
 566                 if (!test_bit(tid, tidmap))
 567                         continue;
 568                 cnt++;
 569                 if (dd->ipath_pageshadow[porttid + tid]) {
 570                         struct page *p;
 571                         p = dd->ipath_pageshadow[porttid + tid];
 572                         dd->ipath_pageshadow[porttid + tid] = NULL;
 573                         ipath_cdbg(VERBOSE, "PID %u freeing TID %u\n",
 574                                    pid_nr(pd->port_pid), tid);
 575                         dd->ipath_f_put_tid(dd, &tidbase[tid],
 576                                             RCVHQ_RCV_TYPE_EXPECTED,
 577                                             dd->ipath_tidinvalid);
 578                         pci_unmap_page(dd->pcidev,
 579                                 dd->ipath_physshadow[porttid + tid],
 580                                 PAGE_SIZE, PCI_DMA_FROMDEVICE);
 581                         ipath_release_user_pages(&p, 1);
 582                         ipath_stats.sps_pageunlocks++;
 583                 } else
 584                         ipath_dbg("Unused tid %u, ignoring\n", tid);
 585         }
 586         if (cnt != ti->tidcnt)
 587                 ipath_dbg("passed in tidcnt %d, only %d bits set in map\n",
 588                           ti->tidcnt, cnt);
 589 done:
 590         if (ret)
 591                 ipath_dbg("Failed to unmap %u TID pages, failing with %d\n",
 592                           ti->tidcnt, -ret);
 593         return ret;
 594 }
 595
 596 /**
 597  * ipath_set_part_key - set a partition key
 598  * @pd: the port
 599  * @key: the key
 600  *
 601  * We can have up to 4 active at a time (other than the default, which is
 602  * always allowed).  This is somewhat tricky, since multiple ports may set
 603  * the same key, so we reference count them, and clean up at exit.  All 4
 604  * partition keys are packed into a single infinipath register.  It's an
 605  * error for a process to set the same pkey multiple times.  We provide no
 606  * mechanism to de-allocate a pkey at this time, we may eventually need to
 607  * do that.  I've used the atomic operations, and no locking, and only make
 608  * a single pass through what's available.  This should be more than
 609  * adequate for some time. I'll think about spinlocks or the like if and as
 610  * it's necessary.
 611  */
 612 static int ipath_set_part_key(struct ipath_portdata *pd, u16 key)
 613 {
 614         struct ipath_devdata *dd = pd->port_dd;
 615         int i, any = 0, pidx = -1;
 616         u16 lkey = key & 0x7FFF;
 617         int ret;
 618
 619         if (lkey == (IPATH_DEFAULT_P_KEY & 0x7FFF)) {
 620                 /* nothing to do; this key always valid */
 621                 ret = 0;
 622                 goto bail;
 623         }
 624
 625         ipath_cdbg(VERBOSE, "p%u try to set pkey %hx, current keys "
 626                    "%hx:%x %hx:%x %hx:%x %hx:%x\n",
 627                    pd->port_port, key, dd->ipath_pkeys[0],
 628                    atomic_read(&dd->ipath_pkeyrefs[0]), dd->ipath_pkeys[1],
 629                    atomic_read(&dd->ipath_pkeyrefs[1]), dd->ipath_pkeys[2],
 630                    atomic_read(&dd->ipath_pkeyrefs[2]), dd->ipath_pkeys[3],
 631                    atomic_read(&dd->ipath_pkeyrefs[3]));
 632
 633         if (!lkey) {
 634                 ipath_cdbg(PROC, "p%u tries to set key 0, not allowed\n",
 635                            pd->port_port);
 636                 ret = -EINVAL;
 637                 goto bail;
 638         }
 639
 640         /*
 641          * Set the full membership bit, because it has to be
 642          * set in the register or the packet, and it seems
 643          * cleaner to set in the register than to force all
 644          * callers to set it. (see bug 4331)
 645          */
 646         key |= 0x8000;
 647
 648         for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
 649                 if (!pd->port_pkeys[i] && pidx == -1)
 650                         pidx = i;
 651                 if (pd->port_pkeys[i] == key) {
 652                         ipath_cdbg(VERBOSE, "p%u tries to set same pkey "
 653                                    "(%x) more than once\n",
 654                                    pd->port_port, key);
 655                         ret = -EEXIST;
 656                         goto bail;
 657                 }
 658         }
 659         if (pidx == -1) {
 660                 ipath_dbg("All pkeys for port %u already in use, "
 661                           "can't set %x\n", pd->port_port, key);
 662                 ret = -EBUSY;
 663                 goto bail;
 664         }
 665         for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
 666                 if (!dd->ipath_pkeys[i]) {
 667                         any++;
 668                         continue;
 669                 }
 670                 if (dd->ipath_pkeys[i] == key) {
 671                         atomic_t *pkrefs = &dd->ipath_pkeyrefs[i];
 672
 673                         if (atomic_inc_return(pkrefs) > 1) {
 674                                 pd->port_pkeys[pidx] = key;
 675                                 ipath_cdbg(VERBOSE, "p%u set key %x "
 676                                            "matches #%d, count now %d\n",
 677                                            pd->port_port, key, i,
 678                                            atomic_read(pkrefs));
 679                                 ret = 0;
 680                                 goto bail;
 681                         } else {
 682                                 /*
 683                                  * lost race, decrement count, catch below
 684                                  */
 685                                 atomic_dec(pkrefs);
 686                                 ipath_cdbg(VERBOSE, "Lost race, count was "
 687                                            "0, after dec, it's %d\n",
 688                                            atomic_read(pkrefs));
 689                                 any++;
 690                         }
 691                 }
 692                 if ((dd->ipath_pkeys[i] & 0x7FFF) == lkey) {
 693                         /*
 694                          * It makes no sense to have both the limited and
 695                          * full membership PKEY set at the same time since
 696                          * the unlimited one will disable the limited one.
 697                          */
 698                         ret = -EEXIST;
 699                         goto bail;
 700                 }
 701         }
 702         if (!any) {
 703                 ipath_dbg("port %u, all pkeys already in use, "
 704                           "can't set %x\n", pd->port_port, key);
 705                 ret = -EBUSY;
 706                 goto bail;
 707         }
 708         for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
 709                 if (!dd->ipath_pkeys[i] &&
 710                     atomic_inc_return(&dd->ipath_pkeyrefs[i]) == 1) {
 711                         u64 pkey;
 712
 713                         /* for ipathstats, etc. */
 714                         ipath_stats.sps_pkeys[i] = lkey;
 715                         pd->port_pkeys[pidx] = dd->ipath_pkeys[i] = key;
 716                         pkey =
 717                                 (u64) dd->ipath_pkeys[0] |
 718                                 ((u64) dd->ipath_pkeys[1] << 16) |
 719                                 ((u64) dd->ipath_pkeys[2] << 32) |
 720                                 ((u64) dd->ipath_pkeys[3] << 48);
 721                         ipath_cdbg(PROC, "p%u set key %x in #%d, "
 722                                    "portidx %d, new pkey reg %llx\n",
 723                                    pd->port_port, key, i, pidx,
 724                                    (unsigned long long) pkey);
 725                         ipath_write_kreg(
 726                                 dd, dd->ipath_kregs->kr_partitionkey, pkey);
 727
 728                         ret = 0;
 729                         goto bail;
 730                 }
 731         }
 732         ipath_dbg("port %u, all pkeys already in use 2nd pass, "
 733                   "can't set %x\n", pd->port_port, key);
 734         ret = -EBUSY;
 735
 736 bail:
 737         return ret;
 738 }
 739
 740 /**
 741  * ipath_manage_rcvq - manage a port's receive queue
 742  * @pd: the port
 743  * @subport: the subport
 744  * @start_stop: action to carry out
 745  *
 746  * start_stop == 0 disables receive on the port, for use in queue
 747  * overflow conditions.  start_stop==1 re-enables, to be used to
 748  * re-init the software copy of the head register
 749  */
 750 static int ipath_manage_rcvq(struct ipath_portdata *pd, unsigned subport,
 751                              int start_stop)
 752 {
 753         struct ipath_devdata *dd = pd->port_dd;
 754
 755         ipath_cdbg(PROC, "%sabling rcv for unit %u port %u:%u\n",
 756                    start_stop ? "en" : "dis", dd->ipath_unit,
 757                    pd->port_port, subport);
 758         if (subport)
 759                 goto bail;
 760         /* atomically clear receive enable port. */
 761         if (start_stop) {
 762                 /*
 763                  * On enable, force in-memory copy of the tail register to
 764                  * 0, so that protocol code doesn't have to worry about
 765                  * whether or not the chip has yet updated the in-memory
 766                  * copy or not on return from the system call. The chip
 767                  * always resets it's tail register back to 0 on a
 768                  * transition from disabled to enabled.  This could cause a
 769                  * problem if software was broken, and did the enable w/o
 770                  * the disable, but eventually the in-memory copy will be
 771                  * updated and correct itself, even in the face of software
 772                  * bugs.
 773                  */
 774                 if (pd->port_rcvhdrtail_kvaddr)
 775                         ipath_clear_rcvhdrtail(pd);
 776                 set_bit(dd->ipath_r_portenable_shift + pd->port_port,
 777                         &dd->ipath_rcvctrl);
 778         } else
 779                 clear_bit(dd->ipath_r_portenable_shift + pd->port_port,
 780                           &dd->ipath_rcvctrl);
 781         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
 782                          dd->ipath_rcvctrl);
 783         /* now be sure chip saw it before we return */
 784         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 785         if (start_stop) {
 786                 /*
 787                  * And try to be sure that tail reg update has happened too.
 788                  * This should in theory interlock with the RXE changes to
 789                  * the tail register.  Don't assign it to the tail register
 790                  * in memory copy, since we could overwrite an update by the
 791                  * chip if we did.
 792                  */
 793                 ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
 794         }
 795         /* always; new head should be equal to new tail; see above */
 796 bail:
 797         return 0;
 798 }
 799
 800 static void ipath_clean_part_key(struct ipath_portdata *pd,
 801                                  struct ipath_devdata *dd)
 802 {
 803         int i, j, pchanged = 0;
 804         u64 oldpkey;
 805
 806         /* for debugging only */
 807         oldpkey = (u64) dd->ipath_pkeys[0] |
 808                 ((u64) dd->ipath_pkeys[1] << 16) |
 809                 ((u64) dd->ipath_pkeys[2] << 32) |
 810                 ((u64) dd->ipath_pkeys[3] << 48);
 811
 812         for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
 813                 if (!pd->port_pkeys[i])
 814                         continue;
 815                 ipath_cdbg(VERBOSE, "look for key[%d] %hx in pkeys\n", i,
 816                            pd->port_pkeys[i]);
 817                 for (j = 0; j < ARRAY_SIZE(dd->ipath_pkeys); j++) {
 818                         /* check for match independent of the global bit */
 819                         if ((dd->ipath_pkeys[j] & 0x7fff) !=
 820                             (pd->port_pkeys[i] & 0x7fff))
 821                                 continue;
 822                         if (atomic_dec_and_test(&dd->ipath_pkeyrefs[j])) {
 823                                 ipath_cdbg(VERBOSE, "p%u clear key "
 824                                            "%x matches #%d\n",
 825                                            pd->port_port,
 826                                            pd->port_pkeys[i], j);
 827                                 ipath_stats.sps_pkeys[j] =
 828                                         dd->ipath_pkeys[j] = 0;
 829                                 pchanged++;
 830                         } else {
 831                                 ipath_cdbg(VERBOSE, "p%u key %x matches #%d, "
 832                                            "but ref still %d\n", pd->port_port,
 833                                            pd->port_pkeys[i], j,
 834                                            atomic_read(&dd->ipath_pkeyrefs[j]));
 835                                 break;
 836                         }
 837                 }
 838                 pd->port_pkeys[i] = 0;
 839         }
 840         if (pchanged) {
 841                 u64 pkey = (u64) dd->ipath_pkeys[0] |
 842                         ((u64) dd->ipath_pkeys[1] << 16) |
 843                         ((u64) dd->ipath_pkeys[2] << 32) |
 844                         ((u64) dd->ipath_pkeys[3] << 48);
 845                 ipath_cdbg(VERBOSE, "p%u old pkey reg %llx, "
 846                            "new pkey reg %llx\n", pd->port_port,
 847                            (unsigned long long) oldpkey,
 848                            (unsigned long long) pkey);
 849                 ipath_write_kreg(dd, dd->ipath_kregs->kr_partitionkey,
 850                                  pkey);
 851         }
 852 }
 853
 854 /*
 855  * Initialize the port data with the receive buffer sizes
 856  * so this can be done while the master port is locked.
 857  * Otherwise, there is a race with a slave opening the port
 858  * and seeing these fields uninitialized.
 859  */
 860 static void init_user_egr_sizes(struct ipath_portdata *pd)
 861 {
 862         struct ipath_devdata *dd = pd->port_dd;
 863         unsigned egrperchunk, egrcnt, size;
 864
 865         /*
 866          * to avoid wasting a lot of memory, we allocate 32KB chunks of
 867          * physically contiguous memory, advance through it until used up
 868          * and then allocate more.  Of course, we need memory to store those
 869          * extra pointers, now.  Started out with 256KB, but under heavy
 870          * memory pressure (creating large files and then copying them over
 871          * NFS while doing lots of MPI jobs), we hit some allocation
 872          * failures, even though we can sleep...  (2.6.10) Still get
 873          * failures at 64K.  32K is the lowest we can go without wasting
 874          * additional memory.
 875          */
 876         size = 0x8000;
 877         egrperchunk = size / dd->ipath_rcvegrbufsize;
 878         egrcnt = dd->ipath_rcvegrcnt;
 879         pd->port_rcvegrbuf_chunks = (egrcnt + egrperchunk - 1) / egrperchunk;
 880         pd->port_rcvegrbufs_perchunk = egrperchunk;
 881         pd->port_rcvegrbuf_size = size;
 882 }
 883
 884 /**
 885  * ipath_create_user_egr - allocate eager TID buffers
 886  * @pd: the port to allocate TID buffers for
 887  *
 888  * This routine is now quite different for user and kernel, because
 889  * the kernel uses skb's, for the accelerated network performance
 890  * This is the user port version
 891  *
 892  * Allocate the eager TID buffers and program them into infinipath
 893  * They are no longer completely contiguous, we do multiple allocation
 894  * calls.
 895  */
 896 static int ipath_create_user_egr(struct ipath_portdata *pd)
 897 {
 898         struct ipath_devdata *dd = pd->port_dd;
 899         unsigned e, egrcnt, egrperchunk, chunk, egrsize, egroff;
 900         size_t size;
 901         int ret;
 902         gfp_t gfp_flags;
 903
 904         /*
 905          * GFP_USER, but without GFP_FS, so buffer cache can be
 906          * coalesced (we hope); otherwise, even at order 4,
 907          * heavy filesystem activity makes these fail, and we can
 908          * use compound pages.
 909          */
 910         gfp_flags = __GFP_RECLAIM | __GFP_IO | __GFP_COMP;
 911
 912         egrcnt = dd->ipath_rcvegrcnt;
 913         /* TID number offset for this port */
 914         egroff = (pd->port_port - 1) * egrcnt + dd->ipath_p0_rcvegrcnt;
 915         egrsize = dd->ipath_rcvegrbufsize;
 916         ipath_cdbg(VERBOSE, "Allocating %d egr buffers, at egrtid "
 917                    "offset %x, egrsize %u\n", egrcnt, egroff, egrsize);
 918
 919         chunk = pd->port_rcvegrbuf_chunks;
 920         egrperchunk = pd->port_rcvegrbufs_perchunk;
 921         size = pd->port_rcvegrbuf_size;
 922         pd->port_rcvegrbuf = kmalloc(chunk * sizeof(pd->port_rcvegrbuf[0]),
 923                                      GFP_KERNEL);
 924         if (!pd->port_rcvegrbuf) {
 925                 ret = -ENOMEM;
 926                 goto bail;
 927         }
 928         pd->port_rcvegrbuf_phys =
 929                 kmalloc(chunk * sizeof(pd->port_rcvegrbuf_phys[0]),
 930                         GFP_KERNEL);
 931         if (!pd->port_rcvegrbuf_phys) {
 932                 ret = -ENOMEM;
 933                 goto bail_rcvegrbuf;
 934         }
 935         for (e = 0; e < pd->port_rcvegrbuf_chunks; e++) {
 936
 937                 pd->port_rcvegrbuf[e] = dma_alloc_coherent(
 938                         &dd->pcidev->dev, size, &pd->port_rcvegrbuf_phys[e],
 939                         gfp_flags);
 940
 941                 if (!pd->port_rcvegrbuf[e]) {
 942                         ret = -ENOMEM;
 943                         goto bail_rcvegrbuf_phys;
 944                 }
 945         }
 946
 947         pd->port_rcvegr_phys = pd->port_rcvegrbuf_phys[0];
 948
 949         for (e = chunk = 0; chunk < pd->port_rcvegrbuf_chunks; chunk++) {
 950                 dma_addr_t pa = pd->port_rcvegrbuf_phys[chunk];
 951                 unsigned i;
 952
 953                 for (i = 0; e < egrcnt && i < egrperchunk; e++, i++) {
 954                         dd->ipath_f_put_tid(dd, e + egroff +
 955                                             (u64 __iomem *)
 956                                             ((char __iomem *)
 957                                              dd->ipath_kregbase +
 958                                              dd->ipath_rcvegrbase),
 959                                             RCVHQ_RCV_TYPE_EAGER, pa);
 960                         pa += egrsize;
 961                 }
 962                 cond_resched(); /* don't hog the cpu */
 963         }
 964
 965         ret = 0;
 966         goto bail;
 967
 968 bail_rcvegrbuf_phys:
 969         for (e = 0; e < pd->port_rcvegrbuf_chunks &&
 970                 pd->port_rcvegrbuf[e]; e++) {
 971                 dma_free_coherent(&dd->pcidev->dev, size,
 972                                   pd->port_rcvegrbuf[e],
 973                                   pd->port_rcvegrbuf_phys[e]);
 974
 975         }
 976         kfree(pd->port_rcvegrbuf_phys);
 977         pd->port_rcvegrbuf_phys = NULL;
 978 bail_rcvegrbuf:
 979         kfree(pd->port_rcvegrbuf);
 980         pd->port_rcvegrbuf = NULL;
 981 bail:
 982         return ret;
 983 }
 984
 985
 986 /* common code for the mappings on dma_alloc_coherent mem */
 987 static int ipath_mmap_mem(struct vm_area_struct *vma,
 988         struct ipath_portdata *pd, unsigned len, int write_ok,
 989         void *kvaddr, char *what)
 990 {
 991         struct ipath_devdata *dd = pd->port_dd;
 992         unsigned long pfn;
 993         int ret;
 994
 995         if ((vma->vm_end - vma->vm_start) > len) {
 996                 dev_info(&dd->pcidev->dev,
 997                          "FAIL on %s: len %lx > %x\n", what,
 998                          vma->vm_end - vma->vm_start, len);
 999                 ret = -EFAULT;
1000                 goto bail;
1001         }
1002
1003         if (!write_ok) {
1004                 if (vma->vm_flags & VM_WRITE) {
1005                         dev_info(&dd->pcidev->dev,
1006                                  "%s must be mapped readonly\n", what);
1007                         ret = -EPERM;
1008                         goto bail;
1009                 }
1010
1011                 /* don't allow them to later change with mprotect */
1012                 vma->vm_flags &= ~VM_MAYWRITE;
1013         }
1014
1015         pfn = virt_to_phys(kvaddr) >> PAGE_SHIFT;
1016         ret = remap_pfn_range(vma, vma->vm_start, pfn,
1017                               len, vma->vm_page_prot);
1018         if (ret)
1019                 dev_info(&dd->pcidev->dev, "%s port%u mmap of %lx, %x "
1020                          "bytes r%c failed: %d\n", what, pd->port_port,
1021                          pfn, len, write_ok?'w':'o', ret);
1022         else
1023                 ipath_cdbg(VERBOSE, "%s port%u mmaped %lx, %x bytes "
1024                            "r%c\n", what, pd->port_port, pfn, len,
1025                            write_ok?'w':'o');
1026 bail:
1027         return ret;
1028 }
1029
1030 static int mmap_ureg(struct vm_area_struct *vma, struct ipath_devdata *dd,
1031                      u64 ureg)
1032 {
1033         unsigned long phys;
1034         int ret;
1035
1036         /*
1037          * This is real hardware, so use io_remap.  This is the mechanism
1038          * for the user process to update the head registers for their port
1039          * in the chip.
1040          */
1041         if ((vma->vm_end - vma->vm_start) > PAGE_SIZE) {
1042                 dev_info(&dd->pcidev->dev, "FAIL mmap userreg: reqlen "
1043                          "%lx > PAGE\n", vma->vm_end - vma->vm_start);
1044                 ret = -EFAULT;
1045         } else {
1046                 phys = dd->ipath_physaddr + ureg;
1047                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1048
1049                 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1050                 ret = io_remap_pfn_range(vma, vma->vm_start,
1051                                          phys >> PAGE_SHIFT,
1052                                          vma->vm_end - vma->vm_start,
1053                                          vma->vm_page_prot);
1054         }
1055         return ret;
1056 }
1057
1058 static int mmap_piobufs(struct vm_area_struct *vma,
1059                         struct ipath_devdata *dd,
1060                         struct ipath_portdata *pd,
1061                         unsigned piobufs, unsigned piocnt)
1062 {
1063         unsigned long phys;
1064         int ret;
1065
1066         /*
1067          * When we map the PIO buffers in the chip, we want to map them as
1068          * writeonly, no read possible.   This prevents access to previous
1069          * process data, and catches users who might try to read the i/o
1070          * space due to a bug.
1071          */
1072         if ((vma->vm_end - vma->vm_start) > (piocnt * dd->ipath_palign)) {
1073                 dev_info(&dd->pcidev->dev, "FAIL mmap piobufs: "
1074                          "reqlen %lx > PAGE\n",
1075                          vma->vm_end - vma->vm_start);
1076                 ret = -EINVAL;
1077                 goto bail;
1078         }
1079
1080         phys = dd->ipath_physaddr + piobufs;
1081
1082 #if defined(__powerpc__)
1083         /* There isn't a generic way to specify writethrough mappings */
1084         pgprot_val(vma->vm_page_prot) |= _PAGE_NO_CACHE;
1085         pgprot_val(vma->vm_page_prot) |= _PAGE_WRITETHRU;
1086         pgprot_val(vma->vm_page_prot) &= ~_PAGE_GUARDED;
1087 #endif
1088
1089         /*
1090          * don't allow them to later change to readable with mprotect (for when
1091          * not initially mapped readable, as is normally the case)
1092          */
1093         vma->vm_flags &= ~VM_MAYREAD;
1094         vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1095
1096         ret = io_remap_pfn_range(vma, vma->vm_start, phys >> PAGE_SHIFT,
1097                                  vma->vm_end - vma->vm_start,
1098                                  vma->vm_page_prot);
1099 bail:
1100         return ret;
1101 }
1102
1103 static int mmap_rcvegrbufs(struct vm_area_struct *vma,
1104                            struct ipath_portdata *pd)
1105 {
1106         struct ipath_devdata *dd = pd->port_dd;
1107         unsigned long start, size;
1108         size_t total_size, i;
1109         unsigned long pfn;
1110         int ret;
1111
1112         size = pd->port_rcvegrbuf_size;
1113         total_size = pd->port_rcvegrbuf_chunks * size;
1114         if ((vma->vm_end - vma->vm_start) > total_size) {
1115                 dev_info(&dd->pcidev->dev, "FAIL on egr bufs: "
1116                          "reqlen %lx > actual %lx\n",
1117                          vma->vm_end - vma->vm_start,
1118                          (unsigned long) total_size);
1119                 ret = -EINVAL;
1120                 goto bail;
1121         }
1122
1123         if (vma->vm_flags & VM_WRITE) {
1124                 dev_info(&dd->pcidev->dev, "Can't map eager buffers as "
1125                          "writable (flags=%lx)\n", vma->vm_flags);
1126                 ret = -EPERM;
1127                 goto bail;
1128         }
1129         /* don't allow them to later change to writeable with mprotect */
1130         vma->vm_flags &= ~VM_MAYWRITE;
1131
1132         start = vma->vm_start;
1133
1134         for (i = 0; i < pd->port_rcvegrbuf_chunks; i++, start += size) {
1135                 pfn = virt_to_phys(pd->port_rcvegrbuf[i]) >> PAGE_SHIFT;
1136                 ret = remap_pfn_range(vma, start, pfn, size,
1137                                       vma->vm_page_prot);
1138                 if (ret < 0)
1139                         goto bail;
1140         }
1141         ret = 0;
1142
1143 bail:
1144         return ret;
1145 }
1146
1147 /*
1148  * ipath_file_vma_fault - handle a VMA page fault.
1149  */
1150 static int ipath_file_vma_fault(struct vm_area_struct *vma,
1151                                         struct vm_fault *vmf)
1152 {
1153         struct page *page;
1154
1155         page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
1156         if (!page)
1157                 return VM_FAULT_SIGBUS;
1158         get_page(page);
1159         vmf->page = page;
1160
1161         return 0;
1162 }
1163
1164 static const struct vm_operations_struct ipath_file_vm_ops = {
1165         .fault = ipath_file_vma_fault,
1166 };
1167
1168 static int mmap_kvaddr(struct vm_area_struct *vma, u64 pgaddr,
1169                        struct ipath_portdata *pd, unsigned subport)
1170 {
1171         unsigned long len;
1172         struct ipath_devdata *dd;
1173         void *addr;
1174         size_t size;
1175         int ret = 0;
1176
1177         /* If the port is not shared, all addresses should be physical */
1178         if (!pd->port_subport_cnt)
1179                 goto bail;
1180
1181         dd = pd->port_dd;
1182         size = pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
1183
1184         /*
1185          * Each process has all the subport uregbase, rcvhdrq, and
1186          * rcvegrbufs mmapped - as an array for all the processes,
1187          * and also separately for this process.
1188          */
1189         if (pgaddr == cvt_kvaddr(pd->subport_uregbase)) {
1190                 addr = pd->subport_uregbase;
1191                 size = PAGE_SIZE * pd->port_subport_cnt;
1192         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvhdr_base)) {
1193                 addr = pd->subport_rcvhdr_base;
1194                 size = pd->port_rcvhdrq_size * pd->port_subport_cnt;
1195         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvegrbuf)) {
1196                 addr = pd->subport_rcvegrbuf;
1197                 size *= pd->port_subport_cnt;
1198         } else if (pgaddr == cvt_kvaddr(pd->subport_uregbase +
1199                                         PAGE_SIZE * subport)) {
1200                 addr = pd->subport_uregbase + PAGE_SIZE * subport;
1201                 size = PAGE_SIZE;
1202         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvhdr_base +
1203                                 pd->port_rcvhdrq_size * subport)) {
1204                 addr = pd->subport_rcvhdr_base +
1205                         pd->port_rcvhdrq_size * subport;
1206                 size = pd->port_rcvhdrq_size;
1207         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvegrbuf +
1208                                size * subport)) {
1209                 addr = pd->subport_rcvegrbuf + size * subport;
1210                 /* rcvegrbufs are read-only on the slave */
1211                 if (vma->vm_flags & VM_WRITE) {
1212                         dev_info(&dd->pcidev->dev,
1213                                  "Can't map eager buffers as "
1214                                  "writable (flags=%lx)\n", vma->vm_flags);
1215                         ret = -EPERM;
1216                         goto bail;
1217                 }
1218                 /*
1219                  * Don't allow permission to later change to writeable
1220                  * with mprotect.
1221                  */
1222                 vma->vm_flags &= ~VM_MAYWRITE;
1223         } else {
1224                 goto bail;
1225         }
1226         len = vma->vm_end - vma->vm_start;
1227         if (len > size) {
1228                 ipath_cdbg(MM, "FAIL: reqlen %lx > %zx\n", len, size);
1229                 ret = -EINVAL;
1230                 goto bail;
1231         }
1232
1233         vma->vm_pgoff = (unsigned long) addr >> PAGE_SHIFT;
1234         vma->vm_ops = &ipath_file_vm_ops;
1235         vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
1236         ret = 1;
1237
1238 bail:
1239         return ret;
1240 }
1241
1242 /**
1243  * ipath_mmap - mmap various structures into user space
1244  * @fp: the file pointer
1245  * @vma: the VM area
1246  *
1247  * We use this to have a shared buffer between the kernel and the user code
1248  * for the rcvhdr queue, egr buffers, and the per-port user regs and pio
1249  * buffers in the chip.  We have the open and close entries so we can bump
1250  * the ref count and keep the driver from being unloaded while still mapped.
1251  */
1252 static int ipath_mmap(struct file *fp, struct vm_area_struct *vma)
1253 {
1254         struct ipath_portdata *pd;
1255         struct ipath_devdata *dd;
1256         u64 pgaddr, ureg;
1257         unsigned piobufs, piocnt;
1258         int ret;
1259
1260         pd = port_fp(fp);
1261         if (!pd) {
1262                 ret = -EINVAL;
1263                 goto bail;
1264         }
1265         dd = pd->port_dd;
1266
1267         /*
1268          * This is the ipath_do_user_init() code, mapping the shared buffers
1269          * into the user process. The address referred to by vm_pgoff is the
1270          * file offset passed via mmap().  For shared ports, this is the
1271          * kernel vmalloc() address of the pages to share with the master.
1272          * For non-shared or master ports, this is a physical address.
1273          * We only do one mmap for each space mapped.
1274          */
1275         pgaddr = vma->vm_pgoff << PAGE_SHIFT;
1276
1277         /*
1278          * Check for 0 in case one of the allocations failed, but user
1279          * called mmap anyway.
1280          */
1281         if (!pgaddr)  {
1282                 ret = -EINVAL;
1283                 goto bail;
1284         }
1285
1286         ipath_cdbg(MM, "pgaddr %llx vm_start=%lx len %lx port %u:%u:%u\n",
1287                    (unsigned long long) pgaddr, vma->vm_start,
1288                    vma->vm_end - vma->vm_start, dd->ipath_unit,
1289                    pd->port_port, subport_fp(fp));
1290
1291         /*
1292          * Physical addresses must fit in 40 bits for our hardware.
1293          * Check for kernel virtual addresses first, anything else must
1294          * match a HW or memory address.
1295          */
1296         ret = mmap_kvaddr(vma, pgaddr, pd, subport_fp(fp));
1297         if (ret) {
1298                 if (ret > 0)
1299                         ret = 0;
1300                 goto bail;
1301         }
1302
1303         ureg = dd->ipath_uregbase + dd->ipath_ureg_align * pd->port_port;
1304         if (!pd->port_subport_cnt) {
1305                 /* port is not shared */
1306                 piocnt = pd->port_piocnt;
1307                 piobufs = pd->port_piobufs;
1308         } else if (!subport_fp(fp)) {
1309                 /* caller is the master */
1310                 piocnt = (pd->port_piocnt / pd->port_subport_cnt) +
1311                          (pd->port_piocnt % pd->port_subport_cnt);
1312                 piobufs = pd->port_piobufs +
1313                         dd->ipath_palign * (pd->port_piocnt - piocnt);
1314         } else {
1315                 unsigned slave = subport_fp(fp) - 1;
1316
1317                 /* caller is a slave */
1318                 piocnt = pd->port_piocnt / pd->port_subport_cnt;
1319                 piobufs = pd->port_piobufs + dd->ipath_palign * piocnt * slave;
1320         }
1321
1322         if (pgaddr == ureg)
1323                 ret = mmap_ureg(vma, dd, ureg);
1324         else if (pgaddr == piobufs)
1325                 ret = mmap_piobufs(vma, dd, pd, piobufs, piocnt);
1326         else if (pgaddr == dd->ipath_pioavailregs_phys)
1327                 /* in-memory copy of pioavail registers */
1328                 ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1329                                      (void *) dd->ipath_pioavailregs_dma,
1330                                      "pioavail registers");
1331         else if (pgaddr == pd->port_rcvegr_phys)
1332                 ret = mmap_rcvegrbufs(vma, pd);
1333         else if (pgaddr == (u64) pd->port_rcvhdrq_phys)
1334                 /*
1335                  * The rcvhdrq itself; readonly except on HT (so have
1336                  * to allow writable mapping), multiple pages, contiguous
1337                  * from an i/o perspective.
1338                  */
1339                 ret = ipath_mmap_mem(vma, pd, pd->port_rcvhdrq_size, 1,
1340                                      pd->port_rcvhdrq,
1341                                      "rcvhdrq");
1342         else if (pgaddr == (u64) pd->port_rcvhdrqtailaddr_phys)
1343                 /* in-memory copy of rcvhdrq tail register */
1344                 ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1345                                      pd->port_rcvhdrtail_kvaddr,
1346                                      "rcvhdrq tail");
1347         else
1348                 ret = -EINVAL;
1349
1350         vma->vm_private_data = NULL;
1351
1352         if (ret < 0)
1353                 dev_info(&dd->pcidev->dev,
1354                          "Failure %d on off %llx len %lx\n",
1355                          -ret, (unsigned long long)pgaddr,
1356                          vma->vm_end - vma->vm_start);
1357 bail:
1358         return ret;
1359 }
1360
1361 static unsigned ipath_poll_hdrqfull(struct ipath_portdata *pd)
1362 {
1363         unsigned pollflag = 0;
1364
1365         if ((pd->poll_type & IPATH_POLL_TYPE_OVERFLOW) &&
1366             pd->port_hdrqfull != pd->port_hdrqfull_poll) {
1367                 pollflag |= POLLIN | POLLRDNORM;
1368                 pd->port_hdrqfull_poll = pd->port_hdrqfull;
1369         }
1370
1371         return pollflag;
1372 }
1373
1374 static unsigned int ipath_poll_urgent(struct ipath_portdata *pd,
1375                                       struct file *fp,
1376                                       struct poll_table_struct *pt)
1377 {
1378         unsigned pollflag = 0;
1379         struct ipath_devdata *dd;
1380
1381         dd = pd->port_dd;
1382
1383         /* variable access in ipath_poll_hdrqfull() needs this */
1384         rmb();
1385         pollflag = ipath_poll_hdrqfull(pd);
1386
1387         if (pd->port_urgent != pd->port_urgent_poll) {
1388                 pollflag |= POLLIN | POLLRDNORM;
1389                 pd->port_urgent_poll = pd->port_urgent;
1390         }
1391
1392         if (!pollflag) {
1393                 /* this saves a spin_lock/unlock in interrupt handler... */
1394                 set_bit(IPATH_PORT_WAITING_URG, &pd->port_flag);
1395                 /* flush waiting flag so don't miss an event... */
1396                 wmb();
1397                 poll_wait(fp, &pd->port_wait, pt);
1398         }
1399
1400         return pollflag;
1401 }
1402
1403 static unsigned int ipath_poll_next(struct ipath_portdata *pd,
1404                                     struct file *fp,
1405                                     struct poll_table_struct *pt)
1406 {
1407         u32 head;
1408         u32 tail;
1409         unsigned pollflag = 0;
1410         struct ipath_devdata *dd;
1411
1412         dd = pd->port_dd;
1413
1414         /* variable access in ipath_poll_hdrqfull() needs this */
1415         rmb();
1416         pollflag = ipath_poll_hdrqfull(pd);
1417
1418         head = ipath_read_ureg32(dd, ur_rcvhdrhead, pd->port_port);
1419         if (pd->port_rcvhdrtail_kvaddr)
1420                 tail = ipath_get_rcvhdrtail(pd);
1421         else
1422                 tail = ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
1423
1424         if (head != tail)
1425                 pollflag |= POLLIN | POLLRDNORM;
1426         else {
1427                 /* this saves a spin_lock/unlock in interrupt handler */
1428                 set_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag);
1429                 /* flush waiting flag so we don't miss an event */
1430                 wmb();
1431
1432                 set_bit(pd->port_port + dd->ipath_r_intravail_shift,
1433                         &dd->ipath_rcvctrl);
1434
1435                 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1436                                  dd->ipath_rcvctrl);
1437
1438                 if (dd->ipath_rhdrhead_intr_off) /* arm rcv interrupt */
1439                         ipath_write_ureg(dd, ur_rcvhdrhead,
1440                                          dd->ipath_rhdrhead_intr_off | head,
1441                                          pd->port_port);
1442
1443                 poll_wait(fp, &pd->port_wait, pt);
1444         }
1445
1446         return pollflag;
1447 }
1448
1449 static unsigned int ipath_poll(struct file *fp,
1450                                struct poll_table_struct *pt)
1451 {
1452         struct ipath_portdata *pd;
1453         unsigned pollflag;
1454
1455         pd = port_fp(fp);
1456         if (!pd)
1457                 pollflag = 0;
1458         else if (pd->poll_type & IPATH_POLL_TYPE_URGENT)
1459                 pollflag = ipath_poll_urgent(pd, fp, pt);
1460         else
1461                 pollflag = ipath_poll_next(pd, fp, pt);
1462
1463         return pollflag;
1464 }
1465
1466 static int ipath_supports_subports(int user_swmajor, int user_swminor)
1467 {
1468         /* no subport implementation prior to software version 1.3 */
1469         return (user_swmajor > 1) || (user_swminor >= 3);
1470 }
1471
1472 static int ipath_compatible_subports(int user_swmajor, int user_swminor)
1473 {
1474         /* this code is written long-hand for clarity */
1475         if (IPATH_USER_SWMAJOR != user_swmajor) {
1476                 /* no promise of compatibility if major mismatch */
1477                 return 0;
1478         }
1479         if (IPATH_USER_SWMAJOR == 1) {
1480                 switch (IPATH_USER_SWMINOR) {
1481                 case 0:
1482                 case 1:
1483                 case 2:
1484                         /* no subport implementation so cannot be compatible */
1485                         return 0;
1486                 case 3:
1487                         /* 3 is only compatible with itself */
1488                         return user_swminor == 3;
1489                 default:
1490                         /* >= 4 are compatible (or are expected to be) */
1491                         return user_swminor >= 4;
1492                 }
1493         }
1494         /* make no promises yet for future major versions */
1495         return 0;
1496 }
1497
1498 static int init_subports(struct ipath_devdata *dd,
1499                          struct ipath_portdata *pd,
1500                          const struct ipath_user_info *uinfo)
1501 {
1502         int ret = 0;
1503         unsigned num_subports;
1504         size_t size;
1505
1506         /*
1507          * If the user is requesting zero subports,
1508          * skip the subport allocation.
1509          */
1510         if (uinfo->spu_subport_cnt <= 0)
1511                 goto bail;
1512
1513         /* Self-consistency check for ipath_compatible_subports() */
1514         if (ipath_supports_subports(IPATH_USER_SWMAJOR, IPATH_USER_SWMINOR) &&
1515             !ipath_compatible_subports(IPATH_USER_SWMAJOR,
1516                                        IPATH_USER_SWMINOR)) {
1517                 dev_info(&dd->pcidev->dev,
1518                          "Inconsistent ipath_compatible_subports()\n");
1519                 goto bail;
1520         }
1521
1522         /* Check for subport compatibility */
1523         if (!ipath_compatible_subports(uinfo->spu_userversion >> 16,
1524                                        uinfo->spu_userversion & 0xffff)) {
1525                 dev_info(&dd->pcidev->dev,
1526                          "Mismatched user version (%d.%d) and driver "
1527                          "version (%d.%d) while port sharing. Ensure "
1528                          "that driver and library are from the same "
1529                          "release.\n",
1530                          (int) (uinfo->spu_userversion >> 16),
1531                          (int) (uinfo->spu_userversion & 0xffff),
1532                          IPATH_USER_SWMAJOR,
1533                          IPATH_USER_SWMINOR);
1534                 goto bail;
1535         }
1536         if (uinfo->spu_subport_cnt > INFINIPATH_MAX_SUBPORT) {
1537                 ret = -EINVAL;
1538                 goto bail;
1539         }
1540
1541         num_subports = uinfo->spu_subport_cnt;
1542         pd->subport_uregbase = vzalloc(PAGE_SIZE * num_subports);
1543         if (!pd->subport_uregbase) {
1544                 ret = -ENOMEM;
1545                 goto bail;
1546         }
1547         /* Note: pd->port_rcvhdrq_size isn't initialized yet. */
1548         size = ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize *
1549                      sizeof(u32), PAGE_SIZE) * num_subports;
1550         pd->subport_rcvhdr_base = vzalloc(size);
1551         if (!pd->subport_rcvhdr_base) {
1552                 ret = -ENOMEM;
1553                 goto bail_ureg;
1554         }
1555
1556         pd->subport_rcvegrbuf = vzalloc(pd->port_rcvegrbuf_chunks *
1557                                         pd->port_rcvegrbuf_size *
1558                                         num_subports);
1559         if (!pd->subport_rcvegrbuf) {
1560                 ret = -ENOMEM;
1561                 goto bail_rhdr;
1562         }
1563
1564         pd->port_subport_cnt = uinfo->spu_subport_cnt;
1565         pd->port_subport_id = uinfo->spu_subport_id;
1566         pd->active_slaves = 1;
1567         set_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag);
1568         goto bail;
1569
1570 bail_rhdr:
1571         vfree(pd->subport_rcvhdr_base);
1572 bail_ureg:
1573         vfree(pd->subport_uregbase);
1574         pd->subport_uregbase = NULL;
1575 bail:
1576         return ret;
1577 }
1578
1579 static int try_alloc_port(struct ipath_devdata *dd, int port,
1580                           struct file *fp,
1581                           const struct ipath_user_info *uinfo)
1582 {
1583         struct ipath_portdata *pd;
1584         int ret;
1585
1586         if (!(pd = dd->ipath_pd[port])) {
1587                 void *ptmp;
1588
1589                 pd = kzalloc(sizeof(struct ipath_portdata), GFP_KERNEL);
1590
1591                 /*
1592                  * Allocate memory for use in ipath_tid_update() just once
1593                  * at open, not per call.  Reduces cost of expected send
1594                  * setup.
1595                  */
1596                 ptmp = kmalloc(dd->ipath_rcvtidcnt * sizeof(u16) +
1597                                dd->ipath_rcvtidcnt * sizeof(struct page **),
1598                                GFP_KERNEL);
1599                 if (!pd || !ptmp) {
1600                         ipath_dev_err(dd, "Unable to allocate portdata "
1601                                       "memory, failing open\n");
1602                         ret = -ENOMEM;
1603                         kfree(pd);
1604                         kfree(ptmp);
1605                         goto bail;
1606                 }
1607                 dd->ipath_pd[port] = pd;
1608                 dd->ipath_pd[port]->port_port = port;
1609                 dd->ipath_pd[port]->port_dd = dd;
1610                 dd->ipath_pd[port]->port_tid_pg_list = ptmp;
1611                 init_waitqueue_head(&dd->ipath_pd[port]->port_wait);
1612         }
1613         if (!pd->port_cnt) {
1614                 pd->userversion = uinfo->spu_userversion;
1615                 init_user_egr_sizes(pd);
1616                 if ((ret = init_subports(dd, pd, uinfo)) != 0)
1617                         goto bail;
1618                 ipath_cdbg(PROC, "%s[%u] opened unit:port %u:%u\n",
1619                            current->comm, current->pid, dd->ipath_unit,
1620                            port);
1621                 pd->port_cnt = 1;
1622                 port_fp(fp) = pd;
1623                 pd->port_pid = get_pid(task_pid(current));
1624                 strlcpy(pd->port_comm, current->comm, sizeof(pd->port_comm));
1625                 ipath_stats.sps_ports++;
1626                 ret = 0;
1627         } else
1628                 ret = -EBUSY;
1629
1630 bail:
1631         return ret;
1632 }
1633
1634 static inline int usable(struct ipath_devdata *dd)
1635 {
1636         return dd &&
1637                 (dd->ipath_flags & IPATH_PRESENT) &&
1638                 dd->ipath_kregbase &&
1639                 dd->ipath_lid &&
1640                 !(dd->ipath_flags & (IPATH_LINKDOWN | IPATH_DISABLED
1641                                      | IPATH_LINKUNK));
1642 }
1643
1644 static int find_free_port(int unit, struct file *fp,
1645                           const struct ipath_user_info *uinfo)
1646 {
1647         struct ipath_devdata *dd = ipath_lookup(unit);
1648         int ret, i;
1649
1650         if (!dd) {
1651                 ret = -ENODEV;
1652                 goto bail;
1653         }
1654
1655         if (!usable(dd)) {
1656                 ret = -ENETDOWN;
1657                 goto bail;
1658         }
1659
1660         for (i = 1; i < dd->ipath_cfgports; i++) {
1661                 ret = try_alloc_port(dd, i, fp, uinfo);
1662                 if (ret != -EBUSY)
1663                         goto bail;
1664         }
1665         ret = -EBUSY;
1666
1667 bail:
1668         return ret;
1669 }
1670
1671 static int find_best_unit(struct file *fp,
1672                           const struct ipath_user_info *uinfo)
1673 {
1674         int ret = 0, i, prefunit = -1, devmax;
1675         int maxofallports, npresent, nup;
1676         int ndev;
1677
1678         devmax = ipath_count_units(&npresent, &nup, &maxofallports);
1679
1680         /*
1681          * This code is present to allow a knowledgeable person to
1682          * specify the layout of processes to processors before opening
1683          * this driver, and then we'll assign the process to the "closest"
1684          * InfiniPath chip to that processor (we assume reasonable connectivity,
1685          * for now).  This code assumes that if affinity has been set
1686          * before this point, that at most one cpu is set; for now this
1687          * is reasonable.  I check for both cpumask_empty() and cpumask_full(),
1688          * in case some kernel variant sets none of the bits when no
1689          * affinity is set.  2.6.11 and 12 kernels have all present
1690          * cpus set.  Some day we'll have to fix it up further to handle
1691          * a cpu subset.  This algorithm fails for two HT chips connected
1692          * in tunnel fashion.  Eventually this needs real topology
1693          * information.  There may be some issues with dual core numbering
1694          * as well.  This needs more work prior to release.
1695          */
1696         if (!cpumask_empty(tsk_cpus_allowed(current)) &&
1697             !cpumask_full(tsk_cpus_allowed(current))) {
1698                 int ncpus = num_online_cpus(), curcpu = -1, nset = 0;
1699                 get_online_cpus();
1700                 for_each_online_cpu(i)
1701                         if (cpumask_test_cpu(i, tsk_cpus_allowed(current))) {
1702                                 ipath_cdbg(PROC, "%s[%u] affinity set for "
1703                                            "cpu %d/%d\n", current->comm,
1704                                            current->pid, i, ncpus);
1705                                 curcpu = i;
1706                                 nset++;
1707                         }
1708                 put_online_cpus();
1709                 if (curcpu != -1 && nset != ncpus) {
1710                         if (npresent) {
1711                                 prefunit = curcpu / (ncpus / npresent);
1712                                 ipath_cdbg(PROC,"%s[%u] %d chips, %d cpus, "
1713                                           "%d cpus/chip, select unit %d\n",
1714                                           current->comm, current->pid,
1715                                           npresent, ncpus, ncpus / npresent,
1716                                           prefunit);
1717                         }
1718                 }
1719         }
1720
1721         /*
1722          * user ports start at 1, kernel port is 0
1723          * For now, we do round-robin access across all chips
1724          */
1725
1726         if (prefunit != -1)
1727                 devmax = prefunit + 1;
1728 recheck:
1729         for (i = 1; i < maxofallports; i++) {
1730                 for (ndev = prefunit != -1 ? prefunit : 0; ndev < devmax;
1731                      ndev++) {
1732                         struct ipath_devdata *dd = ipath_lookup(ndev);
1733
1734                         if (!usable(dd))
1735                                 continue; /* can't use this unit */
1736                         if (i >= dd->ipath_cfgports)
1737                                 /*
1738                                  * Maxed out on users of this unit. Try
1739                                  * next.
1740                                  */
1741                                 continue;
1742                         ret = try_alloc_port(dd, i, fp, uinfo);
1743                         if (!ret)
1744                                 goto done;
1745                 }
1746         }
1747
1748         if (npresent) {
1749                 if (nup == 0) {
1750                         ret = -ENETDOWN;
1751                         ipath_dbg("No ports available (none initialized "
1752                                   "and ready)\n");
1753                 } else {
1754                         if (prefunit > 0) {
1755                                 /* if started above 0, retry from 0 */
1756                                 ipath_cdbg(PROC,
1757                                            "%s[%u] no ports on prefunit "
1758                                            "%d, clear and re-check\n",
1759                                            current->comm, current->pid,
1760                                            prefunit);
1761                                 devmax = ipath_count_units(NULL, NULL,
1762                                                            NULL);
1763                                 prefunit = -1;
1764                                 goto recheck;
1765                         }
1766                         ret = -EBUSY;
1767                         ipath_dbg("No ports available\n");
1768                 }
1769         } else {
1770                 ret = -ENXIO;
1771                 ipath_dbg("No boards found\n");
1772         }
1773
1774 done:
1775         return ret;
1776 }
1777
1778 static int find_shared_port(struct file *fp,
1779                             const struct ipath_user_info *uinfo)
1780 {
1781         int devmax, ndev, i;
1782         int ret = 0;
1783
1784         devmax = ipath_count_units(NULL, NULL, NULL);
1785
1786         for (ndev = 0; ndev < devmax; ndev++) {
1787                 struct ipath_devdata *dd = ipath_lookup(ndev);
1788
1789                 if (!usable(dd))
1790                         continue;
1791                 for (i = 1; i < dd->ipath_cfgports; i++) {
1792                         struct ipath_portdata *pd = dd->ipath_pd[i];
1793
1794                         /* Skip ports which are not yet open */
1795                         if (!pd || !pd->port_cnt)
1796                                 continue;
1797                         /* Skip port if it doesn't match the requested one */
1798                         if (pd->port_subport_id != uinfo->spu_subport_id)
1799                                 continue;
1800                         /* Verify the sharing process matches the master */
1801                         if (pd->port_subport_cnt != uinfo->spu_subport_cnt ||
1802                             pd->userversion != uinfo->spu_userversion ||
1803                             pd->port_cnt >= pd->port_subport_cnt) {
1804                                 ret = -EINVAL;
1805                                 goto done;
1806                         }
1807                         port_fp(fp) = pd;
1808                         subport_fp(fp) = pd->port_cnt++;
1809                         pd->port_subpid[subport_fp(fp)] =
1810                                 get_pid(task_pid(current));
1811                         tidcursor_fp(fp) = 0;
1812                         pd->active_slaves |= 1 << subport_fp(fp);
1813                         ipath_cdbg(PROC,
1814                                    "%s[%u] %u sharing %s[%u] unit:port %u:%u\n",
1815                                    current->comm, current->pid,
1816                                    subport_fp(fp),
1817                                    pd->port_comm, pid_nr(pd->port_pid),
1818                                    dd->ipath_unit, pd->port_port);
1819                         ret = 1;
1820                         goto done;
1821                 }
1822         }
1823
1824 done:
1825         return ret;
1826 }
1827
1828 static int ipath_open(struct inode *in, struct file *fp)
1829 {
1830         /* The real work is performed later in ipath_assign_port() */
1831         fp->private_data = kzalloc(sizeof(struct ipath_filedata), GFP_KERNEL);
1832         return fp->private_data ? 0 : -ENOMEM;
1833 }
1834
1835 /* Get port early, so can set affinity prior to memory allocation */
1836 static int ipath_assign_port(struct file *fp,
1837                               const struct ipath_user_info *uinfo)
1838 {
1839         int ret;
1840         int i_minor;
1841         unsigned swmajor, swminor;
1842
1843         /* Check to be sure we haven't already initialized this file */
1844         if (port_fp(fp)) {
1845                 ret = -EINVAL;
1846                 goto done;
1847         }
1848
1849         /* for now, if major version is different, bail */
1850         swmajor = uinfo->spu_userversion >> 16;
1851         if (swmajor != IPATH_USER_SWMAJOR) {
1852                 ipath_dbg("User major version %d not same as driver "
1853                           "major %d\n", uinfo->spu_userversion >> 16,
1854                           IPATH_USER_SWMAJOR);
1855                 ret = -ENODEV;
1856                 goto done;
1857         }
1858
1859         swminor = uinfo->spu_userversion & 0xffff;
1860         if (swminor != IPATH_USER_SWMINOR)
1861                 ipath_dbg("User minor version %d not same as driver "
1862                           "minor %d\n", swminor, IPATH_USER_SWMINOR);
1863
1864         mutex_lock(&ipath_mutex);
1865
1866         if (ipath_compatible_subports(swmajor, swminor) &&
1867             uinfo->spu_subport_cnt &&
1868             (ret = find_shared_port(fp, uinfo))) {
1869                 if (ret > 0)
1870                         ret = 0;
1871                 goto done_chk_sdma;
1872         }
1873
1874         i_minor = iminor(file_inode(fp)) - IPATH_USER_MINOR_BASE;
1875         ipath_cdbg(VERBOSE, "open on dev %lx (minor %d)\n",
1876                    (long)file_inode(fp)->i_rdev, i_minor);
1877
1878         if (i_minor)
1879                 ret = find_free_port(i_minor - 1, fp, uinfo);
1880         else
1881                 ret = find_best_unit(fp, uinfo);
1882
1883 done_chk_sdma:
1884         if (!ret) {
1885                 struct ipath_filedata *fd = fp->private_data;
1886                 const struct ipath_portdata *pd = fd->pd;
1887                 const struct ipath_devdata *dd = pd->port_dd;
1888
1889                 fd->pq = ipath_user_sdma_queue_create(&dd->pcidev->dev,
1890                                                       dd->ipath_unit,
1891                                                       pd->port_port,
1892                                                       fd->subport);
1893
1894                 if (!fd->pq)
1895                         ret = -ENOMEM;
1896         }
1897
1898         mutex_unlock(&ipath_mutex);
1899
1900 done:
1901         return ret;
1902 }
1903
1904
1905 static int ipath_do_user_init(struct file *fp,
1906                               const struct ipath_user_info *uinfo)
1907 {
1908         int ret;
1909         struct ipath_portdata *pd = port_fp(fp);
1910         struct ipath_devdata *dd;
1911         u32 head32;
1912
1913         /* Subports don't need to initialize anything since master did it. */
1914         if (subport_fp(fp)) {
1915                 ret = wait_event_interruptible(pd->port_wait,
1916                         !test_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag));
1917                 goto done;
1918         }
1919
1920         dd = pd->port_dd;
1921
1922         if (uinfo->spu_rcvhdrsize) {
1923                 ret = ipath_setrcvhdrsize(dd, uinfo->spu_rcvhdrsize);
1924                 if (ret)
1925                         goto done;
1926         }
1927
1928         /* for now we do nothing with rcvhdrcnt: uinfo->spu_rcvhdrcnt */
1929
1930         /* some ports may get extra buffers, calculate that here */
1931         if (pd->port_port <= dd->ipath_ports_extrabuf)
1932                 pd->port_piocnt = dd->ipath_pbufsport + 1;
1933         else
1934                 pd->port_piocnt = dd->ipath_pbufsport;
1935
1936         /* for right now, kernel piobufs are at end, so port 1 is at 0 */
1937         if (pd->port_port <= dd->ipath_ports_extrabuf)
1938                 pd->port_pio_base = (dd->ipath_pbufsport + 1)
1939                         * (pd->port_port - 1);
1940         else
1941                 pd->port_pio_base = dd->ipath_ports_extrabuf +
1942                         dd->ipath_pbufsport * (pd->port_port - 1);
1943         pd->port_piobufs = dd->ipath_piobufbase +
1944                 pd->port_pio_base * dd->ipath_palign;
1945         ipath_cdbg(VERBOSE, "piobuf base for port %u is 0x%x, piocnt %u,"
1946                 " first pio %u\n", pd->port_port, pd->port_piobufs,
1947                 pd->port_piocnt, pd->port_pio_base);
1948         ipath_chg_pioavailkernel(dd, pd->port_pio_base, pd->port_piocnt, 0);
1949
1950         /*
1951          * Now allocate the rcvhdr Q and eager TIDs; skip the TID
1952          * array for time being.  If pd->port_port > chip-supported,
1953          * we need to do extra stuff here to handle by handling overflow
1954          * through port 0, someday
1955          */
1956         ret = ipath_create_rcvhdrq(dd, pd);
1957         if (!ret)
1958                 ret = ipath_create_user_egr(pd);
1959         if (ret)
1960                 goto done;
1961
1962         /*
1963          * set the eager head register for this port to the current values
1964          * of the tail pointers, since we don't know if they were
1965          * updated on last use of the port.
1966          */
1967         head32 = ipath_read_ureg32(dd, ur_rcvegrindextail, pd->port_port);
1968         ipath_write_ureg(dd, ur_rcvegrindexhead, head32, pd->port_port);
1969         pd->port_lastrcvhdrqtail = -1;
1970         ipath_cdbg(VERBOSE, "Wrote port%d egrhead %x from tail regs\n",
1971                 pd->port_port, head32);
1972         pd->port_tidcursor = 0; /* start at beginning after open */
1973
1974         /* initialize poll variables... */
1975         pd->port_urgent = 0;
1976         pd->port_urgent_poll = 0;
1977         pd->port_hdrqfull_poll = pd->port_hdrqfull;
1978
1979         /*
1980          * Now enable the port for receive.
1981          * For chips that are set to DMA the tail register to memory
1982          * when they change (and when the update bit transitions from
1983          * 0 to 1.  So for those chips, we turn it off and then back on.
1984          * This will (very briefly) affect any other open ports, but the
1985          * duration is very short, and therefore isn't an issue.  We
1986          * explicitly set the in-memory tail copy to 0 beforehand, so we
1987          * don't have to wait to be sure the DMA update has happened
1988          * (chip resets head/tail to 0 on transition to enable).
1989          */
1990         set_bit(dd->ipath_r_portenable_shift + pd->port_port,
1991                 &dd->ipath_rcvctrl);
1992         if (!(dd->ipath_flags & IPATH_NODMA_RTAIL)) {
1993                 if (pd->port_rcvhdrtail_kvaddr)
1994                         ipath_clear_rcvhdrtail(pd);
1995                 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1996                         dd->ipath_rcvctrl &
1997                         ~(1ULL << dd->ipath_r_tailupd_shift));
1998         }
1999         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
2000                          dd->ipath_rcvctrl);
2001         /* Notify any waiting slaves */
2002         if (pd->port_subport_cnt) {
2003                 clear_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag);
2004                 wake_up(&pd->port_wait);
2005         }
2006 done:
2007         return ret;
2008 }
2009
2010 /**
2011  * unlock_exptid - unlock any expected TID entries port still had in use
2012  * @pd: port
2013  *
2014  * We don't actually update the chip here, because we do a bulk update
2015  * below, using ipath_f_clear_tids.
2016  */
2017 static void unlock_expected_tids(struct ipath_portdata *pd)
2018 {
2019         struct ipath_devdata *dd = pd->port_dd;
2020         int port_tidbase = pd->port_port * dd->ipath_rcvtidcnt;
2021         int i, cnt = 0, maxtid = port_tidbase + dd->ipath_rcvtidcnt;
2022
2023         ipath_cdbg(VERBOSE, "Port %u unlocking any locked expTID pages\n",
2024                    pd->port_port);
2025         for (i = port_tidbase; i < maxtid; i++) {
2026                 struct page *ps = dd->ipath_pageshadow[i];
2027
2028                 if (!ps)
2029                         continue;
2030
2031                 dd->ipath_pageshadow[i] = NULL;
2032                 pci_unmap_page(dd->pcidev, dd->ipath_physshadow[i],
2033                         PAGE_SIZE, PCI_DMA_FROMDEVICE);
2034                 ipath_release_user_pages_on_close(&ps, 1);
2035                 cnt++;
2036                 ipath_stats.sps_pageunlocks++;
2037         }
2038         if (cnt)
2039                 ipath_cdbg(VERBOSE, "Port %u locked %u expTID entries\n",
2040                            pd->port_port, cnt);
2041
2042         if (ipath_stats.sps_pagelocks || ipath_stats.sps_pageunlocks)
2043                 ipath_cdbg(VERBOSE, "%llu pages locked, %llu unlocked\n",
2044                            (unsigned long long) ipath_stats.sps_pagelocks,
2045                            (unsigned long long)
2046                            ipath_stats.sps_pageunlocks);
2047 }
2048
2049 static int ipath_close(struct inode *in, struct file *fp)
2050 {
2051         struct ipath_filedata *fd;
2052         struct ipath_portdata *pd;
2053         struct ipath_devdata *dd;
2054         unsigned long flags;
2055         unsigned port;
2056         struct pid *pid;
2057
2058         ipath_cdbg(VERBOSE, "close on dev %lx, private data %p\n",
2059                    (long)in->i_rdev, fp->private_data);
2060
2061         mutex_lock(&ipath_mutex);
2062
2063         fd = fp->private_data;
2064         fp->private_data = NULL;
2065         pd = fd->pd;
2066         if (!pd) {
2067                 mutex_unlock(&ipath_mutex);
2068                 goto bail;
2069         }
2070
2071         dd = pd->port_dd;
2072
2073         /* drain user sdma queue */
2074         ipath_user_sdma_queue_drain(dd, fd->pq);
2075         ipath_user_sdma_queue_destroy(fd->pq);
2076
2077         if (--pd->port_cnt) {
2078                 /*
2079                  * XXX If the master closes the port before the slave(s),
2080                  * revoke the mmap for the eager receive queue so
2081                  * the slave(s) don't wait for receive data forever.
2082                  */
2083                 pd->active_slaves &= ~(1 << fd->subport);
2084                 put_pid(pd->port_subpid[fd->subport]);
2085                 pd->port_subpid[fd->subport] = NULL;
2086                 mutex_unlock(&ipath_mutex);
2087                 goto bail;
2088         }
2089         /* early; no interrupt users after this */
2090         spin_lock_irqsave(&dd->ipath_uctxt_lock, flags);
2091         port = pd->port_port;
2092         dd->ipath_pd[port] = NULL;
2093         pid = pd->port_pid;
2094         pd->port_pid = NULL;
2095         spin_unlock_irqrestore(&dd->ipath_uctxt_lock, flags);
2096
2097         if (pd->port_rcvwait_to || pd->port_piowait_to
2098             || pd->port_rcvnowait || pd->port_pionowait) {
2099                 ipath_cdbg(VERBOSE, "port%u, %u rcv, %u pio wait timeo; "
2100                            "%u rcv %u, pio already\n",
2101                            pd->port_port, pd->port_rcvwait_to,
2102                            pd->port_piowait_to, pd->port_rcvnowait,
2103                            pd->port_pionowait);
2104                 pd->port_rcvwait_to = pd->port_piowait_to =
2105                         pd->port_rcvnowait = pd->port_pionowait = 0;
2106         }
2107         if (pd->port_flag) {
2108                 ipath_cdbg(PROC, "port %u port_flag set: 0x%lx\n",
2109                           pd->port_port, pd->port_flag);
2110                 pd->port_flag = 0;
2111         }
2112
2113         if (dd->ipath_kregbase) {
2114                 /* atomically clear receive enable port and intr avail. */
2115                 clear_bit(dd->ipath_r_portenable_shift + port,
2116                           &dd->ipath_rcvctrl);
2117                 clear_bit(pd->port_port + dd->ipath_r_intravail_shift,
2118                           &dd->ipath_rcvctrl);
2119                 ipath_write_kreg( dd, dd->ipath_kregs->kr_rcvctrl,
2120                         dd->ipath_rcvctrl);
2121                 /* and read back from chip to be sure that nothing
2122                  * else is in flight when we do the rest */
2123                 (void)ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
2124
2125                 /* clean up the pkeys for this port user */
2126                 ipath_clean_part_key(pd, dd);
2127                 /*
2128                  * be paranoid, and never write 0's to these, just use an
2129                  * unused part of the port 0 tail page.  Of course,
2130                  * rcvhdraddr points to a large chunk of memory, so this
2131                  * could still trash things, but at least it won't trash
2132                  * page 0, and by disabling the port, it should stop "soon",
2133                  * even if a packet or two is in already in flight after we
2134                  * disabled the port.
2135                  */
2136                 ipath_write_kreg_port(dd,
2137                         dd->ipath_kregs->kr_rcvhdrtailaddr, port,
2138                         dd->ipath_dummy_hdrq_phys);
2139                 ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdraddr,
2140                         pd->port_port, dd->ipath_dummy_hdrq_phys);
2141
2142                 ipath_disarm_piobufs(dd, pd->port_pio_base, pd->port_piocnt);
2143                 ipath_chg_pioavailkernel(dd, pd->port_pio_base,
2144                         pd->port_piocnt, 1);
2145
2146                 dd->ipath_f_clear_tids(dd, pd->port_port);
2147
2148                 if (dd->ipath_pageshadow)
2149                         unlock_expected_tids(pd);
2150                 ipath_stats.sps_ports--;
2151                 ipath_cdbg(PROC, "%s[%u] closed port %u:%u\n",
2152                            pd->port_comm, pid_nr(pid),
2153                            dd->ipath_unit, port);
2154         }
2155
2156         put_pid(pid);
2157         mutex_unlock(&ipath_mutex);
2158         ipath_free_pddata(dd, pd); /* after releasing the mutex */
2159
2160 bail:
2161         kfree(fd);
2162         return 0;
2163 }
2164
2165 static int ipath_port_info(struct ipath_portdata *pd, u16 subport,
2166                            struct ipath_port_info __user *uinfo)
2167 {
2168         struct ipath_port_info info;
2169         int nup;
2170         int ret;
2171         size_t sz;
2172
2173         (void) ipath_count_units(NULL, &nup, NULL);
2174         info.num_active = nup;
2175         info.unit = pd->port_dd->ipath_unit;
2176         info.port = pd->port_port;
2177         info.subport = subport;
2178         /* Don't return new fields if old library opened the port. */
2179         if (ipath_supports_subports(pd->userversion >> 16,
2180                                     pd->userversion & 0xffff)) {
2181                 /* Number of user ports available for this device. */
2182                 info.num_ports = pd->port_dd->ipath_cfgports - 1;
2183                 info.num_subports = pd->port_subport_cnt;
2184                 sz = sizeof(info);
2185         } else
2186                 sz = sizeof(info) - 2 * sizeof(u16);
2187
2188         if (copy_to_user(uinfo, &info, sz)) {
2189                 ret = -EFAULT;
2190                 goto bail;
2191         }
2192         ret = 0;
2193
2194 bail:
2195         return ret;
2196 }
2197
2198 static int ipath_get_slave_info(struct ipath_portdata *pd,
2199                                 void __user *slave_mask_addr)
2200 {
2201         int ret = 0;
2202
2203         if (copy_to_user(slave_mask_addr, &pd->active_slaves, sizeof(u32)))
2204                 ret = -EFAULT;
2205         return ret;
2206 }
2207
2208 static int ipath_sdma_get_inflight(struct ipath_user_sdma_queue *pq,
2209                                    u32 __user *inflightp)
2210 {
2211         const u32 val = ipath_user_sdma_inflight_counter(pq);
2212
2213         if (put_user(val, inflightp))
2214                 return -EFAULT;
2215
2216         return 0;
2217 }
2218
2219 static int ipath_sdma_get_complete(struct ipath_devdata *dd,
2220                                    struct ipath_user_sdma_queue *pq,
2221                                    u32 __user *completep)
2222 {
2223         u32 val;
2224         int err;
2225
2226         err = ipath_user_sdma_make_progress(dd, pq);
2227         if (err < 0)
2228                 return err;
2229
2230         val = ipath_user_sdma_complete_counter(pq);
2231         if (put_user(val, completep))
2232                 return -EFAULT;
2233
2234         return 0;
2235 }
2236
2237 static ssize_t ipath_write(struct file *fp, const char __user *data,
2238                            size_t count, loff_t *off)
2239 {
2240         const struct ipath_cmd __user *ucmd;
2241         struct ipath_portdata *pd;
2242         const void __user *src;
2243         size_t consumed, copy;
2244         struct ipath_cmd cmd;
2245         ssize_t ret = 0;
2246         void *dest;
2247
2248         if (WARN_ON_ONCE(!ib_safe_file_access(fp)))
2249                 return -EACCES;
2250
2251         if (count < sizeof(cmd.type)) {
2252                 ret = -EINVAL;
2253                 goto bail;
2254         }
2255
2256         ucmd = (const struct ipath_cmd __user *) data;
2257
2258         if (copy_from_user(&cmd.type, &ucmd->type, sizeof(cmd.type))) {
2259                 ret = -EFAULT;
2260                 goto bail;
2261         }
2262
2263         consumed = sizeof(cmd.type);
2264
2265         switch (cmd.type) {
2266         case IPATH_CMD_ASSIGN_PORT:
2267         case __IPATH_CMD_USER_INIT:
2268         case IPATH_CMD_USER_INIT:
2269                 copy = sizeof(cmd.cmd.user_info);
2270                 dest = &cmd.cmd.user_info;
2271                 src = &ucmd->cmd.user_info;
2272                 break;
2273         case IPATH_CMD_RECV_CTRL:
2274                 copy = sizeof(cmd.cmd.recv_ctrl);
2275                 dest = &cmd.cmd.recv_ctrl;
2276                 src = &ucmd->cmd.recv_ctrl;
2277                 break;
2278         case IPATH_CMD_PORT_INFO:
2279                 copy = sizeof(cmd.cmd.port_info);
2280                 dest = &cmd.cmd.port_info;
2281                 src = &ucmd->cmd.port_info;
2282                 break;
2283         case IPATH_CMD_TID_UPDATE:
2284         case IPATH_CMD_TID_FREE:
2285                 copy = sizeof(cmd.cmd.tid_info);
2286                 dest = &cmd.cmd.tid_info;
2287                 src = &ucmd->cmd.tid_info;
2288                 break;
2289         case IPATH_CMD_SET_PART_KEY:
2290                 copy = sizeof(cmd.cmd.part_key);
2291                 dest = &cmd.cmd.part_key;
2292                 src = &ucmd->cmd.part_key;
2293                 break;
2294         case __IPATH_CMD_SLAVE_INFO:
2295                 copy = sizeof(cmd.cmd.slave_mask_addr);
2296                 dest = &cmd.cmd.slave_mask_addr;
2297                 src = &ucmd->cmd.slave_mask_addr;
2298                 break;
2299         case IPATH_CMD_PIOAVAILUPD:     // force an update of PIOAvail reg
2300                 copy = 0;
2301                 src = NULL;
2302                 dest = NULL;
2303                 break;
2304         case IPATH_CMD_POLL_TYPE:
2305                 copy = sizeof(cmd.cmd.poll_type);
2306                 dest = &cmd.cmd.poll_type;
2307                 src = &ucmd->cmd.poll_type;
2308                 break;
2309         case IPATH_CMD_ARMLAUNCH_CTRL:
2310                 copy = sizeof(cmd.cmd.armlaunch_ctrl);
2311                 dest = &cmd.cmd.armlaunch_ctrl;
2312                 src = &ucmd->cmd.armlaunch_ctrl;
2313                 break;
2314         case IPATH_CMD_SDMA_INFLIGHT:
2315                 copy = sizeof(cmd.cmd.sdma_inflight);
2316                 dest = &cmd.cmd.sdma_inflight;
2317                 src = &ucmd->cmd.sdma_inflight;
2318                 break;
2319         case IPATH_CMD_SDMA_COMPLETE:
2320                 copy = sizeof(cmd.cmd.sdma_complete);
2321                 dest = &cmd.cmd.sdma_complete;
2322                 src = &ucmd->cmd.sdma_complete;
2323                 break;
2324         default:
2325                 ret = -EINVAL;
2326                 goto bail;
2327         }
2328
2329         if (copy) {
2330                 if ((count - consumed) < copy) {
2331                         ret = -EINVAL;
2332                         goto bail;
2333                 }
2334
2335                 if (copy_from_user(dest, src, copy)) {
2336                         ret = -EFAULT;
2337                         goto bail;
2338                 }
2339
2340                 consumed += copy;
2341         }
2342
2343         pd = port_fp(fp);
2344         if (!pd && cmd.type != __IPATH_CMD_USER_INIT &&
2345                 cmd.type != IPATH_CMD_ASSIGN_PORT) {
2346                 ret = -EINVAL;
2347                 goto bail;
2348         }
2349
2350         switch (cmd.type) {
2351         case IPATH_CMD_ASSIGN_PORT:
2352                 ret = ipath_assign_port(fp, &cmd.cmd.user_info);
2353                 if (ret)
2354                         goto bail;
2355                 break;
2356         case __IPATH_CMD_USER_INIT:
2357                 /* backwards compatibility, get port first */
2358                 ret = ipath_assign_port(fp, &cmd.cmd.user_info);
2359                 if (ret)
2360                         goto bail;
2361                 /* and fall through to current version. */
2362         case IPATH_CMD_USER_INIT:
2363                 ret = ipath_do_user_init(fp, &cmd.cmd.user_info);
2364                 if (ret)
2365                         goto bail;
2366                 ret = ipath_get_base_info(
2367                         fp, (void __user *) (unsigned long)
2368                         cmd.cmd.user_info.spu_base_info,
2369                         cmd.cmd.user_info.spu_base_info_size);
2370                 break;
2371         case IPATH_CMD_RECV_CTRL:
2372                 ret = ipath_manage_rcvq(pd, subport_fp(fp), cmd.cmd.recv_ctrl);
2373                 break;
2374         case IPATH_CMD_PORT_INFO:
2375                 ret = ipath_port_info(pd, subport_fp(fp),
2376                                       (struct ipath_port_info __user *)
2377                                       (unsigned long) cmd.cmd.port_info);
2378                 break;
2379         case IPATH_CMD_TID_UPDATE:
2380                 ret = ipath_tid_update(pd, fp, &cmd.cmd.tid_info);
2381                 break;
2382         case IPATH_CMD_TID_FREE:
2383                 ret = ipath_tid_free(pd, subport_fp(fp), &cmd.cmd.tid_info);
2384                 break;
2385         case IPATH_CMD_SET_PART_KEY:
2386                 ret = ipath_set_part_key(pd, cmd.cmd.part_key);
2387                 break;
2388         case __IPATH_CMD_SLAVE_INFO:
2389                 ret = ipath_get_slave_info(pd,
2390                                            (void __user *) (unsigned long)
2391                                            cmd.cmd.slave_mask_addr);
2392                 break;
2393         case IPATH_CMD_PIOAVAILUPD:
2394                 ipath_force_pio_avail_update(pd->port_dd);
2395                 break;
2396         case IPATH_CMD_POLL_TYPE:
2397                 pd->poll_type = cmd.cmd.poll_type;
2398                 break;
2399         case IPATH_CMD_ARMLAUNCH_CTRL:
2400                 if (cmd.cmd.armlaunch_ctrl)
2401                         ipath_enable_armlaunch(pd->port_dd);
2402                 else
2403                         ipath_disable_armlaunch(pd->port_dd);
2404                 break;
2405         case IPATH_CMD_SDMA_INFLIGHT:
2406                 ret = ipath_sdma_get_inflight(user_sdma_queue_fp(fp),
2407                                               (u32 __user *) (unsigned long)
2408                                               cmd.cmd.sdma_inflight);
2409                 break;
2410         case IPATH_CMD_SDMA_COMPLETE:
2411                 ret = ipath_sdma_get_complete(pd->port_dd,
2412                                               user_sdma_queue_fp(fp),
2413                                               (u32 __user *) (unsigned long)
2414                                               cmd.cmd.sdma_complete);
2415                 break;
2416         }
2417
2418         if (ret >= 0)
2419                 ret = consumed;
2420
2421 bail:
2422         return ret;
2423 }
2424
2425 static ssize_t ipath_write_iter(struct kiocb *iocb, struct iov_iter *from)
2426 {
2427         struct file *filp = iocb->ki_filp;
2428         struct ipath_filedata *fp = filp->private_data;
2429         struct ipath_portdata *pd = port_fp(filp);
2430         struct ipath_user_sdma_queue *pq = fp->pq;
2431
2432         if (!iter_is_iovec(from) || !from->nr_segs)
2433                 return -EINVAL;
2434
2435         return ipath_user_sdma_writev(pd->port_dd, pq, from->iov, from->nr_segs);
2436 }
2437
2438 static struct class *ipath_class;
2439
2440 static int init_cdev(int minor, char *name, const struct file_operations *fops,
2441                      struct cdev **cdevp, struct device **devp)
2442 {
2443         const dev_t dev = MKDEV(IPATH_MAJOR, minor);
2444         struct cdev *cdev = NULL;
2445         struct device *device = NULL;
2446         int ret;
2447
2448         cdev = cdev_alloc();
2449         if (!cdev) {
2450                 printk(KERN_ERR IPATH_DRV_NAME
2451                        ": Could not allocate cdev for minor %d, %s\n",
2452                        minor, name);
2453                 ret = -ENOMEM;
2454                 goto done;
2455         }
2456
2457         cdev->owner = THIS_MODULE;
2458         cdev->ops = fops;
2459         kobject_set_name(&cdev->kobj, name);
2460
2461         ret = cdev_add(cdev, dev, 1);
2462         if (ret < 0) {
2463                 printk(KERN_ERR IPATH_DRV_NAME
2464                        ": Could not add cdev for minor %d, %s (err %d)\n",
2465                        minor, name, -ret);
2466                 goto err_cdev;
2467         }
2468
2469         device = device_create(ipath_class, NULL, dev, NULL, name);
2470
2471         if (IS_ERR(device)) {
2472                 ret = PTR_ERR(device);
2473                 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2474                        "device for minor %d, %s (err %d)\n",
2475                        minor, name, -ret);
2476                 goto err_cdev;
2477         }
2478
2479         goto done;
2480
2481 err_cdev:
2482         cdev_del(cdev);
2483         cdev = NULL;
2484
2485 done:
2486         if (ret >= 0) {
2487                 *cdevp = cdev;
2488                 *devp = device;
2489         } else {
2490                 *cdevp = NULL;
2491                 *devp = NULL;
2492         }
2493
2494         return ret;
2495 }
2496
2497 int ipath_cdev_init(int minor, char *name, const struct file_operations *fops,
2498                     struct cdev **cdevp, struct device **devp)
2499 {
2500         return init_cdev(minor, name, fops, cdevp, devp);
2501 }
2502
2503 static void cleanup_cdev(struct cdev **cdevp,
2504                          struct device **devp)
2505 {
2506         struct device *dev = *devp;
2507
2508         if (dev) {
2509                 device_unregister(dev);
2510                 *devp = NULL;
2511         }
2512
2513         if (*cdevp) {
2514                 cdev_del(*cdevp);
2515                 *cdevp = NULL;
2516         }
2517 }
2518
2519 void ipath_cdev_cleanup(struct cdev **cdevp,
2520                         struct device **devp)
2521 {
2522         cleanup_cdev(cdevp, devp);
2523 }
2524
2525 static struct cdev *wildcard_cdev;
2526 static struct device *wildcard_dev;
2527
2528 static const dev_t dev = MKDEV(IPATH_MAJOR, 0);
2529
2530 static int user_init(void)
2531 {
2532         int ret;
2533
2534         ret = register_chrdev_region(dev, IPATH_NMINORS, IPATH_DRV_NAME);
2535         if (ret < 0) {
2536                 printk(KERN_ERR IPATH_DRV_NAME ": Could not register "
2537                        "chrdev region (err %d)\n", -ret);
2538                 goto done;
2539         }
2540
2541         ipath_class = class_create(THIS_MODULE, IPATH_DRV_NAME);
2542
2543         if (IS_ERR(ipath_class)) {
2544                 ret = PTR_ERR(ipath_class);
2545                 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2546                        "device class (err %d)\n", -ret);
2547                 goto bail;
2548         }
2549
2550         goto done;
2551 bail:
2552         unregister_chrdev_region(dev, IPATH_NMINORS);
2553 done:
2554         return ret;
2555 }
2556
2557 static void user_cleanup(void)
2558 {
2559         if (ipath_class) {
2560                 class_destroy(ipath_class);
2561                 ipath_class = NULL;
2562         }
2563
2564         unregister_chrdev_region(dev, IPATH_NMINORS);
2565 }
2566
2567 static atomic_t user_count = ATOMIC_INIT(0);
2568 static atomic_t user_setup = ATOMIC_INIT(0);
2569
2570 int ipath_user_add(struct ipath_devdata *dd)
2571 {
2572         char name[10];
2573         int ret;
2574
2575         if (atomic_inc_return(&user_count) == 1) {
2576                 ret = user_init();
2577                 if (ret < 0) {
2578                         ipath_dev_err(dd, "Unable to set up user support: "
2579                                       "error %d\n", -ret);
2580                         goto bail;
2581                 }
2582                 ret = init_cdev(0, "ipath", &ipath_file_ops, &wildcard_cdev,
2583                                 &wildcard_dev);
2584                 if (ret < 0) {
2585                         ipath_dev_err(dd, "Could not create wildcard "
2586                                       "minor: error %d\n", -ret);
2587                         goto bail_user;
2588                 }
2589
2590                 atomic_set(&user_setup, 1);
2591         }
2592
2593         snprintf(name, sizeof(name), "ipath%d", dd->ipath_unit);
2594
2595         ret = init_cdev(dd->ipath_unit + 1, name, &ipath_file_ops,
2596                         &dd->user_cdev, &dd->user_dev);
2597         if (ret < 0)
2598                 ipath_dev_err(dd, "Could not create user minor %d, %s\n",
2599                               dd->ipath_unit + 1, name);
2600
2601         goto bail;
2602
2603 bail_user:
2604         user_cleanup();
2605 bail:
2606         return ret;
2607 }
2608
2609 void ipath_user_remove(struct ipath_devdata *dd)
2610 {
2611         cleanup_cdev(&dd->user_cdev, &dd->user_dev);
2612
2613         if (atomic_dec_return(&user_count) == 0) {
2614                 if (atomic_read(&user_setup) == 0)
2615                         goto bail;
2616
2617                 cleanup_cdev(&wildcard_cdev, &wildcard_dev);
2618                 user_cleanup();
2619
2620                 atomic_set(&user_setup, 0);
2621         }
2622 bail:
2623         return;
2624 }