drivers/infiniband/hw/ipath/ipath_file_ops.c

   1 /*
   2  * Copyright (c) 2006, 2007, 2008 QLogic Corporation. All rights reserved.
   3  * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
   4  *
   5  * This software is available to you under a choice of one of two
   6  * licenses.  You may choose to be licensed under the terms of the GNU
   7  * General Public License (GPL) Version 2, available from the file
   8  * COPYING in the main directory of this source tree, or the
   9  * OpenIB.org BSD license below:
  10  *
  11  *     Redistribution and use in source and binary forms, with or
  12  *     without modification, are permitted provided that the following
  13  *     conditions are met:
  14  *
  15  *      - Redistributions of source code must retain the above
  16  *        copyright notice, this list of conditions and the following
  17  *        disclaimer.
  18  *
  19  *      - Redistributions in binary form must reproduce the above
  20  *        copyright notice, this list of conditions and the following
  21  *        disclaimer in the documentation and/or other materials
  22  *        provided with the distribution.
  23  *
  24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  31  * SOFTWARE.
  32  */
  33
  34 #include <linux/pci.h>
  35 #include <linux/poll.h>
  36 #include <linux/cdev.h>
  37 #include <linux/swap.h>
  38 #include <linux/vmalloc.h>
  39 #include <linux/slab.h>
  40 #include <linux/highmem.h>
  41 #include <linux/io.h>
  42 #include <linux/jiffies.h>
  43 #include <linux/smp_lock.h>
  44 #include <asm/pgtable.h>
  45
  46 #include "ipath_kernel.h"
  47 #include "ipath_common.h"
  48 #include "ipath_user_sdma.h"
  49
  50 static int ipath_open(struct inode *, struct file *);
  51 static int ipath_close(struct inode *, struct file *);
  52 static ssize_t ipath_write(struct file *, const char __user *, size_t,
  53                            loff_t *);
  54 static ssize_t ipath_writev(struct kiocb *, const struct iovec *,
  55                             unsigned long , loff_t);
  56 static unsigned int ipath_poll(struct file *, struct poll_table_struct *);
  57 static int ipath_mmap(struct file *, struct vm_area_struct *);
  58
  59 static const struct file_operations ipath_file_ops = {
  60         .owner = THIS_MODULE,
  61         .write = ipath_write,
  62         .aio_write = ipath_writev,
  63         .open = ipath_open,
  64         .release = ipath_close,
  65         .poll = ipath_poll,
  66         .mmap = ipath_mmap,
  67         .llseek = noop_llseek,
  68 };
  69
  70 /*
  71  * Convert kernel virtual addresses to physical addresses so they don't
  72  * potentially conflict with the chip addresses used as mmap offsets.
  73  * It doesn't really matter what mmap offset we use as long as we can
  74  * interpret it correctly.
  75  */
  76 static u64 cvt_kvaddr(void *p)
  77 {
  78         struct page *page;
  79         u64 paddr = 0;
  80
  81         page = vmalloc_to_page(p);
  82         if (page)
  83                 paddr = page_to_pfn(page) << PAGE_SHIFT;
  84
  85         return paddr;
  86 }
  87
  88 static int ipath_get_base_info(struct file *fp,
  89                                void __user *ubase, size_t ubase_size)
  90 {
  91         struct ipath_portdata *pd = port_fp(fp);
  92         int ret = 0;
  93         struct ipath_base_info *kinfo = NULL;
  94         struct ipath_devdata *dd = pd->port_dd;
  95         unsigned subport_cnt;
  96         int shared, master;
  97         size_t sz;
  98
  99         subport_cnt = pd->port_subport_cnt;
 100         if (!subport_cnt) {
 101                 shared = 0;
 102                 master = 0;
 103                 subport_cnt = 1;
 104         } else {
 105                 shared = 1;
 106                 master = !subport_fp(fp);
 107         }
 108
 109         sz = sizeof(*kinfo);
 110         /* If port sharing is not requested, allow the old size structure */
 111         if (!shared)
 112                 sz -= 7 * sizeof(u64);
 113         if (ubase_size < sz) {
 114                 ipath_cdbg(PROC,
 115                            "Base size %zu, need %zu (version mismatch?)\n",
 116                            ubase_size, sz);
 117                 ret = -EINVAL;
 118                 goto bail;
 119         }
 120
 121         kinfo = kzalloc(sizeof(*kinfo), GFP_KERNEL);
 122         if (kinfo == NULL) {
 123                 ret = -ENOMEM;
 124                 goto bail;
 125         }
 126
 127         ret = dd->ipath_f_get_base_info(pd, kinfo);
 128         if (ret < 0)
 129                 goto bail;
 130
 131         kinfo->spi_rcvhdr_cnt = dd->ipath_rcvhdrcnt;
 132         kinfo->spi_rcvhdrent_size = dd->ipath_rcvhdrentsize;
 133         kinfo->spi_tidegrcnt = dd->ipath_rcvegrcnt;
 134         kinfo->spi_rcv_egrbufsize = dd->ipath_rcvegrbufsize;
 135         /*
 136          * have to mmap whole thing
 137          */
 138         kinfo->spi_rcv_egrbuftotlen =
 139                 pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
 140         kinfo->spi_rcv_egrperchunk = pd->port_rcvegrbufs_perchunk;
 141         kinfo->spi_rcv_egrchunksize = kinfo->spi_rcv_egrbuftotlen /
 142                 pd->port_rcvegrbuf_chunks;
 143         kinfo->spi_tidcnt = dd->ipath_rcvtidcnt / subport_cnt;
 144         if (master)
 145                 kinfo->spi_tidcnt += dd->ipath_rcvtidcnt % subport_cnt;
 146         /*
 147          * for this use, may be ipath_cfgports summed over all chips that
 148          * are are configured and present
 149          */
 150         kinfo->spi_nports = dd->ipath_cfgports;
 151         /* unit (chip/board) our port is on */
 152         kinfo->spi_unit = dd->ipath_unit;
 153         /* for now, only a single page */
 154         kinfo->spi_tid_maxsize = PAGE_SIZE;
 155
 156         /*
 157          * Doing this per port, and based on the skip value, etc.  This has
 158          * to be the actual buffer size, since the protocol code treats it
 159          * as an array.
 160          *
 161          * These have to be set to user addresses in the user code via mmap.
 162          * These values are used on return to user code for the mmap target
 163          * addresses only.  For 32 bit, same 44 bit address problem, so use
 164          * the physical address, not virtual.  Before 2.6.11, using the
 165          * page_address() macro worked, but in 2.6.11, even that returns the
 166          * full 64 bit address (upper bits all 1's).  So far, using the
 167          * physical addresses (or chip offsets, for chip mapping) works, but
 168          * no doubt some future kernel release will change that, and we'll be
 169          * on to yet another method of dealing with this.
 170          */
 171         kinfo->spi_rcvhdr_base = (u64) pd->port_rcvhdrq_phys;
 172         kinfo->spi_rcvhdr_tailaddr = (u64) pd->port_rcvhdrqtailaddr_phys;
 173         kinfo->spi_rcv_egrbufs = (u64) pd->port_rcvegr_phys;
 174         kinfo->spi_pioavailaddr = (u64) dd->ipath_pioavailregs_phys;
 175         kinfo->spi_status = (u64) kinfo->spi_pioavailaddr +
 176                 (void *) dd->ipath_statusp -
 177                 (void *) dd->ipath_pioavailregs_dma;
 178         if (!shared) {
 179                 kinfo->spi_piocnt = pd->port_piocnt;
 180                 kinfo->spi_piobufbase = (u64) pd->port_piobufs;
 181                 kinfo->__spi_uregbase = (u64) dd->ipath_uregbase +
 182                         dd->ipath_ureg_align * pd->port_port;
 183         } else if (master) {
 184                 kinfo->spi_piocnt = (pd->port_piocnt / subport_cnt) +
 185                                     (pd->port_piocnt % subport_cnt);
 186                 /* Master's PIO buffers are after all the slave's */
 187                 kinfo->spi_piobufbase = (u64) pd->port_piobufs +
 188                         dd->ipath_palign *
 189                         (pd->port_piocnt - kinfo->spi_piocnt);
 190         } else {
 191                 unsigned slave = subport_fp(fp) - 1;
 192
 193                 kinfo->spi_piocnt = pd->port_piocnt / subport_cnt;
 194                 kinfo->spi_piobufbase = (u64) pd->port_piobufs +
 195                         dd->ipath_palign * kinfo->spi_piocnt * slave;
 196         }
 197
 198         if (shared) {
 199                 kinfo->spi_port_uregbase = (u64) dd->ipath_uregbase +
 200                         dd->ipath_ureg_align * pd->port_port;
 201                 kinfo->spi_port_rcvegrbuf = kinfo->spi_rcv_egrbufs;
 202                 kinfo->spi_port_rcvhdr_base = kinfo->spi_rcvhdr_base;
 203                 kinfo->spi_port_rcvhdr_tailaddr = kinfo->spi_rcvhdr_tailaddr;
 204
 205                 kinfo->__spi_uregbase = cvt_kvaddr(pd->subport_uregbase +
 206                         PAGE_SIZE * subport_fp(fp));
 207
 208                 kinfo->spi_rcvhdr_base = cvt_kvaddr(pd->subport_rcvhdr_base +
 209                         pd->port_rcvhdrq_size * subport_fp(fp));
 210                 kinfo->spi_rcvhdr_tailaddr = 0;
 211                 kinfo->spi_rcv_egrbufs = cvt_kvaddr(pd->subport_rcvegrbuf +
 212                         pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size *
 213                         subport_fp(fp));
 214
 215                 kinfo->spi_subport_uregbase =
 216                         cvt_kvaddr(pd->subport_uregbase);
 217                 kinfo->spi_subport_rcvegrbuf =
 218                         cvt_kvaddr(pd->subport_rcvegrbuf);
 219                 kinfo->spi_subport_rcvhdr_base =
 220                         cvt_kvaddr(pd->subport_rcvhdr_base);
 221                 ipath_cdbg(PROC, "port %u flags %x %llx %llx %llx\n",
 222                         kinfo->spi_port, kinfo->spi_runtime_flags,
 223                         (unsigned long long) kinfo->spi_subport_uregbase,
 224                         (unsigned long long) kinfo->spi_subport_rcvegrbuf,
 225                         (unsigned long long) kinfo->spi_subport_rcvhdr_base);
 226         }
 227
 228         /*
 229          * All user buffers are 2KB buffers.  If we ever support
 230          * giving 4KB buffers to user processes, this will need some
 231          * work.
 232          */
 233         kinfo->spi_pioindex = (kinfo->spi_piobufbase -
 234                 (dd->ipath_piobufbase & 0xffffffff)) / dd->ipath_palign;
 235         kinfo->spi_pioalign = dd->ipath_palign;
 236
 237         kinfo->spi_qpair = IPATH_KD_QP;
 238         /*
 239          * user mode PIO buffers are always 2KB, even when 4KB can
 240          * be received, and sent via the kernel; this is ibmaxlen
 241          * for 2K MTU.
 242          */
 243         kinfo->spi_piosize = dd->ipath_piosize2k - 2 * sizeof(u32);
 244         kinfo->spi_mtu = dd->ipath_ibmaxlen;    /* maxlen, not ibmtu */
 245         kinfo->spi_port = pd->port_port;
 246         kinfo->spi_subport = subport_fp(fp);
 247         kinfo->spi_sw_version = IPATH_KERN_SWVERSION;
 248         kinfo->spi_hw_version = dd->ipath_revision;
 249
 250         if (master) {
 251                 kinfo->spi_runtime_flags |= IPATH_RUNTIME_MASTER;
 252         }
 253
 254         sz = (ubase_size < sizeof(*kinfo)) ? ubase_size : sizeof(*kinfo);
 255         if (copy_to_user(ubase, kinfo, sz))
 256                 ret = -EFAULT;
 257
 258 bail:
 259         kfree(kinfo);
 260         return ret;
 261 }
 262
 263 /**
 264  * ipath_tid_update - update a port TID
 265  * @pd: the port
 266  * @fp: the ipath device file
 267  * @ti: the TID information
 268  *
 269  * The new implementation as of Oct 2004 is that the driver assigns
 270  * the tid and returns it to the caller.   To make it easier to
 271  * catch bugs, and to reduce search time, we keep a cursor for
 272  * each port, walking the shadow tid array to find one that's not
 273  * in use.
 274  *
 275  * For now, if we can't allocate the full list, we fail, although
 276  * in the long run, we'll allocate as many as we can, and the
 277  * caller will deal with that by trying the remaining pages later.
 278  * That means that when we fail, we have to mark the tids as not in
 279  * use again, in our shadow copy.
 280  *
 281  * It's up to the caller to free the tids when they are done.
 282  * We'll unlock the pages as they free them.
 283  *
 284  * Also, right now we are locking one page at a time, but since
 285  * the intended use of this routine is for a single group of
 286  * virtually contiguous pages, that should change to improve
 287  * performance.
 288  */
 289 static int ipath_tid_update(struct ipath_portdata *pd, struct file *fp,
 290                             const struct ipath_tid_info *ti)
 291 {
 292         int ret = 0, ntids;
 293         u32 tid, porttid, cnt, i, tidcnt, tidoff;
 294         u16 *tidlist;
 295         struct ipath_devdata *dd = pd->port_dd;
 296         u64 physaddr;
 297         unsigned long vaddr;
 298         u64 __iomem *tidbase;
 299         unsigned long tidmap[8];
 300         struct page **pagep = NULL;
 301         unsigned subport = subport_fp(fp);
 302
 303         if (!dd->ipath_pageshadow) {
 304                 ret = -ENOMEM;
 305                 goto done;
 306         }
 307
 308         cnt = ti->tidcnt;
 309         if (!cnt) {
 310                 ipath_dbg("After copyin, tidcnt 0, tidlist %llx\n",
 311                           (unsigned long long) ti->tidlist);
 312                 /*
 313                  * Should we treat as success?  likely a bug
 314                  */
 315                 ret = -EFAULT;
 316                 goto done;
 317         }
 318         porttid = pd->port_port * dd->ipath_rcvtidcnt;
 319         if (!pd->port_subport_cnt) {
 320                 tidcnt = dd->ipath_rcvtidcnt;
 321                 tid = pd->port_tidcursor;
 322                 tidoff = 0;
 323         } else if (!subport) {
 324                 tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
 325                          (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
 326                 tidoff = dd->ipath_rcvtidcnt - tidcnt;
 327                 porttid += tidoff;
 328                 tid = tidcursor_fp(fp);
 329         } else {
 330                 tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
 331                 tidoff = tidcnt * (subport - 1);
 332                 porttid += tidoff;
 333                 tid = tidcursor_fp(fp);
 334         }
 335         if (cnt > tidcnt) {
 336                 /* make sure it all fits in port_tid_pg_list */
 337                 dev_info(&dd->pcidev->dev, "Process tried to allocate %u "
 338                          "TIDs, only trying max (%u)\n", cnt, tidcnt);
 339                 cnt = tidcnt;
 340         }
 341         pagep = &((struct page **) pd->port_tid_pg_list)[tidoff];
 342         tidlist = &((u16 *) &pagep[dd->ipath_rcvtidcnt])[tidoff];
 343
 344         memset(tidmap, 0, sizeof(tidmap));
 345         /* before decrement; chip actual # */
 346         ntids = tidcnt;
 347         tidbase = (u64 __iomem *) (((char __iomem *) dd->ipath_kregbase) +
 348                                    dd->ipath_rcvtidbase +
 349                                    porttid * sizeof(*tidbase));
 350
 351         ipath_cdbg(VERBOSE, "Port%u %u tids, cursor %u, tidbase %p\n",
 352                    pd->port_port, cnt, tid, tidbase);
 353
 354         /* virtual address of first page in transfer */
 355         vaddr = ti->tidvaddr;
 356         if (!access_ok(VERIFY_WRITE, (void __user *) vaddr,
 357                        cnt * PAGE_SIZE)) {
 358                 ipath_dbg("Fail vaddr %p, %u pages, !access_ok\n",
 359                           (void *)vaddr, cnt);
 360                 ret = -EFAULT;
 361                 goto done;
 362         }
 363         ret = ipath_get_user_pages(vaddr, cnt, pagep);
 364         if (ret) {
 365                 if (ret == -EBUSY) {
 366                         ipath_dbg("Failed to lock addr %p, %u pages "
 367                                   "(already locked)\n",
 368                                   (void *) vaddr, cnt);
 369                         /*
 370                          * for now, continue, and see what happens but with
 371                          * the new implementation, this should never happen,
 372                          * unless perhaps the user has mpin'ed the pages
 373                          * themselves (something we need to test)
 374                          */
 375                         ret = 0;
 376                 } else {
 377                         dev_info(&dd->pcidev->dev,
 378                                  "Failed to lock addr %p, %u pages: "
 379                                  "errno %d\n", (void *) vaddr, cnt, -ret);
 380                         goto done;
 381                 }
 382         }
 383         for (i = 0; i < cnt; i++, vaddr += PAGE_SIZE) {
 384                 for (; ntids--; tid++) {
 385                         if (tid == tidcnt)
 386                                 tid = 0;
 387                         if (!dd->ipath_pageshadow[porttid + tid])
 388                                 break;
 389                 }
 390                 if (ntids < 0) {
 391                         /*
 392                          * oops, wrapped all the way through their TIDs,
 393                          * and didn't have enough free; see comments at
 394                          * start of routine
 395                          */
 396                         ipath_dbg("Not enough free TIDs for %u pages "
 397                                   "(index %d), failing\n", cnt, i);
 398                         i--;    /* last tidlist[i] not filled in */
 399                         ret = -ENOMEM;
 400                         break;
 401                 }
 402                 tidlist[i] = tid + tidoff;
 403                 ipath_cdbg(VERBOSE, "Updating idx %u to TID %u, "
 404                            "vaddr %lx\n", i, tid + tidoff, vaddr);
 405                 /* we "know" system pages and TID pages are same size */
 406                 dd->ipath_pageshadow[porttid + tid] = pagep[i];
 407                 dd->ipath_physshadow[porttid + tid] = ipath_map_page(
 408                         dd->pcidev, pagep[i], 0, PAGE_SIZE,
 409                         PCI_DMA_FROMDEVICE);
 410                 /*
 411                  * don't need atomic or it's overhead
 412                  */
 413                 __set_bit(tid, tidmap);
 414                 physaddr = dd->ipath_physshadow[porttid + tid];
 415                 ipath_stats.sps_pagelocks++;
 416                 ipath_cdbg(VERBOSE,
 417                            "TID %u, vaddr %lx, physaddr %llx pgp %p\n",
 418                            tid, vaddr, (unsigned long long) physaddr,
 419                            pagep[i]);
 420                 dd->ipath_f_put_tid(dd, &tidbase[tid], RCVHQ_RCV_TYPE_EXPECTED,
 421                                     physaddr);
 422                 /*
 423                  * don't check this tid in ipath_portshadow, since we
 424                  * just filled it in; start with the next one.
 425                  */
 426                 tid++;
 427         }
 428
 429         if (ret) {
 430                 u32 limit;
 431         cleanup:
 432                 /* jump here if copy out of updated info failed... */
 433                 ipath_dbg("After failure (ret=%d), undo %d of %d entries\n",
 434                           -ret, i, cnt);
 435                 /* same code that's in ipath_free_tid() */
 436                 limit = sizeof(tidmap) * BITS_PER_BYTE;
 437                 if (limit > tidcnt)
 438                         /* just in case size changes in future */
 439                         limit = tidcnt;
 440                 tid = find_first_bit((const unsigned long *)tidmap, limit);
 441                 for (; tid < limit; tid++) {
 442                         if (!test_bit(tid, tidmap))
 443                                 continue;
 444                         if (dd->ipath_pageshadow[porttid + tid]) {
 445                                 ipath_cdbg(VERBOSE, "Freeing TID %u\n",
 446                                            tid);
 447                                 dd->ipath_f_put_tid(dd, &tidbase[tid],
 448                                                     RCVHQ_RCV_TYPE_EXPECTED,
 449                                                     dd->ipath_tidinvalid);
 450                                 pci_unmap_page(dd->pcidev,
 451                                         dd->ipath_physshadow[porttid + tid],
 452                                         PAGE_SIZE, PCI_DMA_FROMDEVICE);
 453                                 dd->ipath_pageshadow[porttid + tid] = NULL;
 454                                 ipath_stats.sps_pageunlocks++;
 455                         }
 456                 }
 457                 ipath_release_user_pages(pagep, cnt);
 458         } else {
 459                 /*
 460                  * Copy the updated array, with ipath_tid's filled in, back
 461                  * to user.  Since we did the copy in already, this "should
 462                  * never fail" If it does, we have to clean up...
 463                  */
 464                 if (copy_to_user((void __user *)
 465                                  (unsigned long) ti->tidlist,
 466                                  tidlist, cnt * sizeof(*tidlist))) {
 467                         ret = -EFAULT;
 468                         goto cleanup;
 469                 }
 470                 if (copy_to_user((void __user *) (unsigned long) ti->tidmap,
 471                                  tidmap, sizeof tidmap)) {
 472                         ret = -EFAULT;
 473                         goto cleanup;
 474                 }
 475                 if (tid == tidcnt)
 476                         tid = 0;
 477                 if (!pd->port_subport_cnt)
 478                         pd->port_tidcursor = tid;
 479                 else
 480                         tidcursor_fp(fp) = tid;
 481         }
 482
 483 done:
 484         if (ret)
 485                 ipath_dbg("Failed to map %u TID pages, failing with %d\n",
 486                           ti->tidcnt, -ret);
 487         return ret;
 488 }
 489
 490 /**
 491  * ipath_tid_free - free a port TID
 492  * @pd: the port
 493  * @subport: the subport
 494  * @ti: the TID info
 495  *
 496  * right now we are unlocking one page at a time, but since
 497  * the intended use of this routine is for a single group of
 498  * virtually contiguous pages, that should change to improve
 499  * performance.  We check that the TID is in range for this port
 500  * but otherwise don't check validity; if user has an error and
 501  * frees the wrong tid, it's only their own data that can thereby
 502  * be corrupted.  We do check that the TID was in use, for sanity
 503  * We always use our idea of the saved address, not the address that
 504  * they pass in to us.
 505  */
 506
 507 static int ipath_tid_free(struct ipath_portdata *pd, unsigned subport,
 508                           const struct ipath_tid_info *ti)
 509 {
 510         int ret = 0;
 511         u32 tid, porttid, cnt, limit, tidcnt;
 512         struct ipath_devdata *dd = pd->port_dd;
 513         u64 __iomem *tidbase;
 514         unsigned long tidmap[8];
 515
 516         if (!dd->ipath_pageshadow) {
 517                 ret = -ENOMEM;
 518                 goto done;
 519         }
 520
 521         if (copy_from_user(tidmap, (void __user *)(unsigned long)ti->tidmap,
 522                            sizeof tidmap)) {
 523                 ret = -EFAULT;
 524                 goto done;
 525         }
 526
 527         porttid = pd->port_port * dd->ipath_rcvtidcnt;
 528         if (!pd->port_subport_cnt)
 529                 tidcnt = dd->ipath_rcvtidcnt;
 530         else if (!subport) {
 531                 tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
 532                          (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
 533                 porttid += dd->ipath_rcvtidcnt - tidcnt;
 534         } else {
 535                 tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
 536                 porttid += tidcnt * (subport - 1);
 537         }
 538         tidbase = (u64 __iomem *) ((char __iomem *)(dd->ipath_kregbase) +
 539                                    dd->ipath_rcvtidbase +
 540                                    porttid * sizeof(*tidbase));
 541
 542         limit = sizeof(tidmap) * BITS_PER_BYTE;
 543         if (limit > tidcnt)
 544                 /* just in case size changes in future */
 545                 limit = tidcnt;
 546         tid = find_first_bit(tidmap, limit);
 547         ipath_cdbg(VERBOSE, "Port%u free %u tids; first bit (max=%d) "
 548                    "set is %d, porttid %u\n", pd->port_port, ti->tidcnt,
 549                    limit, tid, porttid);
 550         for (cnt = 0; tid < limit; tid++) {
 551                 /*
 552                  * small optimization; if we detect a run of 3 or so without
 553                  * any set, use find_first_bit again.  That's mainly to
 554                  * accelerate the case where we wrapped, so we have some at
 555                  * the beginning, and some at the end, and a big gap
 556                  * in the middle.
 557                  */
 558                 if (!test_bit(tid, tidmap))
 559                         continue;
 560                 cnt++;
 561                 if (dd->ipath_pageshadow[porttid + tid]) {
 562                         struct page *p;
 563                         p = dd->ipath_pageshadow[porttid + tid];
 564                         dd->ipath_pageshadow[porttid + tid] = NULL;
 565                         ipath_cdbg(VERBOSE, "PID %u freeing TID %u\n",
 566                                    pid_nr(pd->port_pid), tid);
 567                         dd->ipath_f_put_tid(dd, &tidbase[tid],
 568                                             RCVHQ_RCV_TYPE_EXPECTED,
 569                                             dd->ipath_tidinvalid);
 570                         pci_unmap_page(dd->pcidev,
 571                                 dd->ipath_physshadow[porttid + tid],
 572                                 PAGE_SIZE, PCI_DMA_FROMDEVICE);
 573                         ipath_release_user_pages(&p, 1);
 574                         ipath_stats.sps_pageunlocks++;
 575                 } else
 576                         ipath_dbg("Unused tid %u, ignoring\n", tid);
 577         }
 578         if (cnt != ti->tidcnt)
 579                 ipath_dbg("passed in tidcnt %d, only %d bits set in map\n",
 580                           ti->tidcnt, cnt);
 581 done:
 582         if (ret)
 583                 ipath_dbg("Failed to unmap %u TID pages, failing with %d\n",
 584                           ti->tidcnt, -ret);
 585         return ret;
 586 }
 587
 588 /**
 589  * ipath_set_part_key - set a partition key
 590  * @pd: the port
 591  * @key: the key
 592  *
 593  * We can have up to 4 active at a time (other than the default, which is
 594  * always allowed).  This is somewhat tricky, since multiple ports may set
 595  * the same key, so we reference count them, and clean up at exit.  All 4
 596  * partition keys are packed into a single infinipath register.  It's an
 597  * error for a process to set the same pkey multiple times.  We provide no
 598  * mechanism to de-allocate a pkey at this time, we may eventually need to
 599  * do that.  I've used the atomic operations, and no locking, and only make
 600  * a single pass through what's available.  This should be more than
 601  * adequate for some time. I'll think about spinlocks or the like if and as
 602  * it's necessary.
 603  */
 604 static int ipath_set_part_key(struct ipath_portdata *pd, u16 key)
 605 {
 606         struct ipath_devdata *dd = pd->port_dd;
 607         int i, any = 0, pidx = -1;
 608         u16 lkey = key & 0x7FFF;
 609         int ret;
 610
 611         if (lkey == (IPATH_DEFAULT_P_KEY & 0x7FFF)) {
 612                 /* nothing to do; this key always valid */
 613                 ret = 0;
 614                 goto bail;
 615         }
 616
 617         ipath_cdbg(VERBOSE, "p%u try to set pkey %hx, current keys "
 618                    "%hx:%x %hx:%x %hx:%x %hx:%x\n",
 619                    pd->port_port, key, dd->ipath_pkeys[0],
 620                    atomic_read(&dd->ipath_pkeyrefs[0]), dd->ipath_pkeys[1],
 621                    atomic_read(&dd->ipath_pkeyrefs[1]), dd->ipath_pkeys[2],
 622                    atomic_read(&dd->ipath_pkeyrefs[2]), dd->ipath_pkeys[3],
 623                    atomic_read(&dd->ipath_pkeyrefs[3]));
 624
 625         if (!lkey) {
 626                 ipath_cdbg(PROC, "p%u tries to set key 0, not allowed\n",
 627                            pd->port_port);
 628                 ret = -EINVAL;
 629                 goto bail;
 630         }
 631
 632         /*
 633          * Set the full membership bit, because it has to be
 634          * set in the register or the packet, and it seems
 635          * cleaner to set in the register than to force all
 636          * callers to set it. (see bug 4331)
 637          */
 638         key |= 0x8000;
 639
 640         for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
 641                 if (!pd->port_pkeys[i] && pidx == -1)
 642                         pidx = i;
 643                 if (pd->port_pkeys[i] == key) {
 644                         ipath_cdbg(VERBOSE, "p%u tries to set same pkey "
 645                                    "(%x) more than once\n",
 646                                    pd->port_port, key);
 647                         ret = -EEXIST;
 648                         goto bail;
 649                 }
 650         }
 651         if (pidx == -1) {
 652                 ipath_dbg("All pkeys for port %u already in use, "
 653                           "can't set %x\n", pd->port_port, key);
 654                 ret = -EBUSY;
 655                 goto bail;
 656         }
 657         for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
 658                 if (!dd->ipath_pkeys[i]) {
 659                         any++;
 660                         continue;
 661                 }
 662                 if (dd->ipath_pkeys[i] == key) {
 663                         atomic_t *pkrefs = &dd->ipath_pkeyrefs[i];
 664
 665                         if (atomic_inc_return(pkrefs) > 1) {
 666                                 pd->port_pkeys[pidx] = key;
 667                                 ipath_cdbg(VERBOSE, "p%u set key %x "
 668                                            "matches #%d, count now %d\n",
 669                                            pd->port_port, key, i,
 670                                            atomic_read(pkrefs));
 671                                 ret = 0;
 672                                 goto bail;
 673                         } else {
 674                                 /*
 675                                  * lost race, decrement count, catch below
 676                                  */
 677                                 atomic_dec(pkrefs);
 678                                 ipath_cdbg(VERBOSE, "Lost race, count was "
 679                                            "0, after dec, it's %d\n",
 680                                            atomic_read(pkrefs));
 681                                 any++;
 682                         }
 683                 }
 684                 if ((dd->ipath_pkeys[i] & 0x7FFF) == lkey) {
 685                         /*
 686                          * It makes no sense to have both the limited and
 687                          * full membership PKEY set at the same time since
 688                          * the unlimited one will disable the limited one.
 689                          */
 690                         ret = -EEXIST;
 691                         goto bail;
 692                 }
 693         }
 694         if (!any) {
 695                 ipath_dbg("port %u, all pkeys already in use, "
 696                           "can't set %x\n", pd->port_port, key);
 697                 ret = -EBUSY;
 698                 goto bail;
 699         }
 700         for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
 701                 if (!dd->ipath_pkeys[i] &&
 702                     atomic_inc_return(&dd->ipath_pkeyrefs[i]) == 1) {
 703                         u64 pkey;
 704
 705                         /* for ipathstats, etc. */
 706                         ipath_stats.sps_pkeys[i] = lkey;
 707                         pd->port_pkeys[pidx] = dd->ipath_pkeys[i] = key;
 708                         pkey =
 709                                 (u64) dd->ipath_pkeys[0] |
 710                                 ((u64) dd->ipath_pkeys[1] << 16) |
 711                                 ((u64) dd->ipath_pkeys[2] << 32) |
 712                                 ((u64) dd->ipath_pkeys[3] << 48);
 713                         ipath_cdbg(PROC, "p%u set key %x in #%d, "
 714                                    "portidx %d, new pkey reg %llx\n",
 715                                    pd->port_port, key, i, pidx,
 716                                    (unsigned long long) pkey);
 717                         ipath_write_kreg(
 718                                 dd, dd->ipath_kregs->kr_partitionkey, pkey);
 719
 720                         ret = 0;
 721                         goto bail;
 722                 }
 723         }
 724         ipath_dbg("port %u, all pkeys already in use 2nd pass, "
 725                   "can't set %x\n", pd->port_port, key);
 726         ret = -EBUSY;
 727
 728 bail:
 729         return ret;
 730 }
 731
 732 /**
 733  * ipath_manage_rcvq - manage a port's receive queue
 734  * @pd: the port
 735  * @subport: the subport
 736  * @start_stop: action to carry out
 737  *
 738  * start_stop == 0 disables receive on the port, for use in queue
 739  * overflow conditions.  start_stop==1 re-enables, to be used to
 740  * re-init the software copy of the head register
 741  */
 742 static int ipath_manage_rcvq(struct ipath_portdata *pd, unsigned subport,
 743                              int start_stop)
 744 {
 745         struct ipath_devdata *dd = pd->port_dd;
 746
 747         ipath_cdbg(PROC, "%sabling rcv for unit %u port %u:%u\n",
 748                    start_stop ? "en" : "dis", dd->ipath_unit,
 749                    pd->port_port, subport);
 750         if (subport)
 751                 goto bail;
 752         /* atomically clear receive enable port. */
 753         if (start_stop) {
 754                 /*
 755                  * On enable, force in-memory copy of the tail register to
 756                  * 0, so that protocol code doesn't have to worry about
 757                  * whether or not the chip has yet updated the in-memory
 758                  * copy or not on return from the system call. The chip
 759                  * always resets it's tail register back to 0 on a
 760                  * transition from disabled to enabled.  This could cause a
 761                  * problem if software was broken, and did the enable w/o
 762                  * the disable, but eventually the in-memory copy will be
 763                  * updated and correct itself, even in the face of software
 764                  * bugs.
 765                  */
 766                 if (pd->port_rcvhdrtail_kvaddr)
 767                         ipath_clear_rcvhdrtail(pd);
 768                 set_bit(dd->ipath_r_portenable_shift + pd->port_port,
 769                         &dd->ipath_rcvctrl);
 770         } else
 771                 clear_bit(dd->ipath_r_portenable_shift + pd->port_port,
 772                           &dd->ipath_rcvctrl);
 773         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
 774                          dd->ipath_rcvctrl);
 775         /* now be sure chip saw it before we return */
 776         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 777         if (start_stop) {
 778                 /*
 779                  * And try to be sure that tail reg update has happened too.
 780                  * This should in theory interlock with the RXE changes to
 781                  * the tail register.  Don't assign it to the tail register
 782                  * in memory copy, since we could overwrite an update by the
 783                  * chip if we did.
 784                  */
 785                 ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
 786         }
 787         /* always; new head should be equal to new tail; see above */
 788 bail:
 789         return 0;
 790 }
 791
 792 static void ipath_clean_part_key(struct ipath_portdata *pd,
 793                                  struct ipath_devdata *dd)
 794 {
 795         int i, j, pchanged = 0;
 796         u64 oldpkey;
 797
 798         /* for debugging only */
 799         oldpkey = (u64) dd->ipath_pkeys[0] |
 800                 ((u64) dd->ipath_pkeys[1] << 16) |
 801                 ((u64) dd->ipath_pkeys[2] << 32) |
 802                 ((u64) dd->ipath_pkeys[3] << 48);
 803
 804         for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
 805                 if (!pd->port_pkeys[i])
 806                         continue;
 807                 ipath_cdbg(VERBOSE, "look for key[%d] %hx in pkeys\n", i,
 808                            pd->port_pkeys[i]);
 809                 for (j = 0; j < ARRAY_SIZE(dd->ipath_pkeys); j++) {
 810                         /* check for match independent of the global bit */
 811                         if ((dd->ipath_pkeys[j] & 0x7fff) !=
 812                             (pd->port_pkeys[i] & 0x7fff))
 813                                 continue;
 814                         if (atomic_dec_and_test(&dd->ipath_pkeyrefs[j])) {
 815                                 ipath_cdbg(VERBOSE, "p%u clear key "
 816                                            "%x matches #%d\n",
 817                                            pd->port_port,
 818                                            pd->port_pkeys[i], j);
 819                                 ipath_stats.sps_pkeys[j] =
 820                                         dd->ipath_pkeys[j] = 0;
 821                                 pchanged++;
 822                         }
 823                         else ipath_cdbg(
 824                                 VERBOSE, "p%u key %x matches #%d, "
 825                                 "but ref still %d\n", pd->port_port,
 826                                 pd->port_pkeys[i], j,
 827                                 atomic_read(&dd->ipath_pkeyrefs[j]));
 828                         break;
 829                 }
 830                 pd->port_pkeys[i] = 0;
 831         }
 832         if (pchanged) {
 833                 u64 pkey = (u64) dd->ipath_pkeys[0] |
 834                         ((u64) dd->ipath_pkeys[1] << 16) |
 835                         ((u64) dd->ipath_pkeys[2] << 32) |
 836                         ((u64) dd->ipath_pkeys[3] << 48);
 837                 ipath_cdbg(VERBOSE, "p%u old pkey reg %llx, "
 838                            "new pkey reg %llx\n", pd->port_port,
 839                            (unsigned long long) oldpkey,
 840                            (unsigned long long) pkey);
 841                 ipath_write_kreg(dd, dd->ipath_kregs->kr_partitionkey,
 842                                  pkey);
 843         }
 844 }
 845
 846 /*
 847  * Initialize the port data with the receive buffer sizes
 848  * so this can be done while the master port is locked.
 849  * Otherwise, there is a race with a slave opening the port
 850  * and seeing these fields uninitialized.
 851  */
 852 static void init_user_egr_sizes(struct ipath_portdata *pd)
 853 {
 854         struct ipath_devdata *dd = pd->port_dd;
 855         unsigned egrperchunk, egrcnt, size;
 856
 857         /*
 858          * to avoid wasting a lot of memory, we allocate 32KB chunks of
 859          * physically contiguous memory, advance through it until used up
 860          * and then allocate more.  Of course, we need memory to store those
 861          * extra pointers, now.  Started out with 256KB, but under heavy
 862          * memory pressure (creating large files and then copying them over
 863          * NFS while doing lots of MPI jobs), we hit some allocation
 864          * failures, even though we can sleep...  (2.6.10) Still get
 865          * failures at 64K.  32K is the lowest we can go without wasting
 866          * additional memory.
 867          */
 868         size = 0x8000;
 869         egrperchunk = size / dd->ipath_rcvegrbufsize;
 870         egrcnt = dd->ipath_rcvegrcnt;
 871         pd->port_rcvegrbuf_chunks = (egrcnt + egrperchunk - 1) / egrperchunk;
 872         pd->port_rcvegrbufs_perchunk = egrperchunk;
 873         pd->port_rcvegrbuf_size = size;
 874 }
 875
 876 /**
 877  * ipath_create_user_egr - allocate eager TID buffers
 878  * @pd: the port to allocate TID buffers for
 879  *
 880  * This routine is now quite different for user and kernel, because
 881  * the kernel uses skb's, for the accelerated network performance
 882  * This is the user port version
 883  *
 884  * Allocate the eager TID buffers and program them into infinipath
 885  * They are no longer completely contiguous, we do multiple allocation
 886  * calls.
 887  */
 888 static int ipath_create_user_egr(struct ipath_portdata *pd)
 889 {
 890         struct ipath_devdata *dd = pd->port_dd;
 891         unsigned e, egrcnt, egrperchunk, chunk, egrsize, egroff;
 892         size_t size;
 893         int ret;
 894         gfp_t gfp_flags;
 895
 896         /*
 897          * GFP_USER, but without GFP_FS, so buffer cache can be
 898          * coalesced (we hope); otherwise, even at order 4,
 899          * heavy filesystem activity makes these fail, and we can
 900          * use compound pages.
 901          */
 902         gfp_flags = __GFP_WAIT | __GFP_IO | __GFP_COMP;
 903
 904         egrcnt = dd->ipath_rcvegrcnt;
 905         /* TID number offset for this port */
 906         egroff = (pd->port_port - 1) * egrcnt + dd->ipath_p0_rcvegrcnt;
 907         egrsize = dd->ipath_rcvegrbufsize;
 908         ipath_cdbg(VERBOSE, "Allocating %d egr buffers, at egrtid "
 909                    "offset %x, egrsize %u\n", egrcnt, egroff, egrsize);
 910
 911         chunk = pd->port_rcvegrbuf_chunks;
 912         egrperchunk = pd->port_rcvegrbufs_perchunk;
 913         size = pd->port_rcvegrbuf_size;
 914         pd->port_rcvegrbuf = kmalloc(chunk * sizeof(pd->port_rcvegrbuf[0]),
 915                                      GFP_KERNEL);
 916         if (!pd->port_rcvegrbuf) {
 917                 ret = -ENOMEM;
 918                 goto bail;
 919         }
 920         pd->port_rcvegrbuf_phys =
 921                 kmalloc(chunk * sizeof(pd->port_rcvegrbuf_phys[0]),
 922                         GFP_KERNEL);
 923         if (!pd->port_rcvegrbuf_phys) {
 924                 ret = -ENOMEM;
 925                 goto bail_rcvegrbuf;
 926         }
 927         for (e = 0; e < pd->port_rcvegrbuf_chunks; e++) {
 928
 929                 pd->port_rcvegrbuf[e] = dma_alloc_coherent(
 930                         &dd->pcidev->dev, size, &pd->port_rcvegrbuf_phys[e],
 931                         gfp_flags);
 932
 933                 if (!pd->port_rcvegrbuf[e]) {
 934                         ret = -ENOMEM;
 935                         goto bail_rcvegrbuf_phys;
 936                 }
 937         }
 938
 939         pd->port_rcvegr_phys = pd->port_rcvegrbuf_phys[0];
 940
 941         for (e = chunk = 0; chunk < pd->port_rcvegrbuf_chunks; chunk++) {
 942                 dma_addr_t pa = pd->port_rcvegrbuf_phys[chunk];
 943                 unsigned i;
 944
 945                 for (i = 0; e < egrcnt && i < egrperchunk; e++, i++) {
 946                         dd->ipath_f_put_tid(dd, e + egroff +
 947                                             (u64 __iomem *)
 948                                             ((char __iomem *)
 949                                              dd->ipath_kregbase +
 950                                              dd->ipath_rcvegrbase),
 951                                             RCVHQ_RCV_TYPE_EAGER, pa);
 952                         pa += egrsize;
 953                 }
 954                 cond_resched(); /* don't hog the cpu */
 955         }
 956
 957         ret = 0;
 958         goto bail;
 959
 960 bail_rcvegrbuf_phys:
 961         for (e = 0; e < pd->port_rcvegrbuf_chunks &&
 962                 pd->port_rcvegrbuf[e]; e++) {
 963                 dma_free_coherent(&dd->pcidev->dev, size,
 964                                   pd->port_rcvegrbuf[e],
 965                                   pd->port_rcvegrbuf_phys[e]);
 966
 967         }
 968         kfree(pd->port_rcvegrbuf_phys);
 969         pd->port_rcvegrbuf_phys = NULL;
 970 bail_rcvegrbuf:
 971         kfree(pd->port_rcvegrbuf);
 972         pd->port_rcvegrbuf = NULL;
 973 bail:
 974         return ret;
 975 }
 976
 977
 978 /* common code for the mappings on dma_alloc_coherent mem */
 979 static int ipath_mmap_mem(struct vm_area_struct *vma,
 980         struct ipath_portdata *pd, unsigned len, int write_ok,
 981         void *kvaddr, char *what)
 982 {
 983         struct ipath_devdata *dd = pd->port_dd;
 984         unsigned long pfn;
 985         int ret;
 986
 987         if ((vma->vm_end - vma->vm_start) > len) {
 988                 dev_info(&dd->pcidev->dev,
 989                          "FAIL on %s: len %lx > %x\n", what,
 990                          vma->vm_end - vma->vm_start, len);
 991                 ret = -EFAULT;
 992                 goto bail;
 993         }
 994
 995         if (!write_ok) {
 996                 if (vma->vm_flags & VM_WRITE) {
 997                         dev_info(&dd->pcidev->dev,
 998                                  "%s must be mapped readonly\n", what);
 999                         ret = -EPERM;
1000                         goto bail;
1001                 }
1002
1003                 /* don't allow them to later change with mprotect */
1004                 vma->vm_flags &= ~VM_MAYWRITE;
1005         }
1006
1007         pfn = virt_to_phys(kvaddr) >> PAGE_SHIFT;
1008         ret = remap_pfn_range(vma, vma->vm_start, pfn,
1009                               len, vma->vm_page_prot);
1010         if (ret)
1011                 dev_info(&dd->pcidev->dev, "%s port%u mmap of %lx, %x "
1012                          "bytes r%c failed: %d\n", what, pd->port_port,
1013                          pfn, len, write_ok?'w':'o', ret);
1014         else
1015                 ipath_cdbg(VERBOSE, "%s port%u mmaped %lx, %x bytes "
1016                            "r%c\n", what, pd->port_port, pfn, len,
1017                            write_ok?'w':'o');
1018 bail:
1019         return ret;
1020 }
1021
1022 static int mmap_ureg(struct vm_area_struct *vma, struct ipath_devdata *dd,
1023                      u64 ureg)
1024 {
1025         unsigned long phys;
1026         int ret;
1027
1028         /*
1029          * This is real hardware, so use io_remap.  This is the mechanism
1030          * for the user process to update the head registers for their port
1031          * in the chip.
1032          */
1033         if ((vma->vm_end - vma->vm_start) > PAGE_SIZE) {
1034                 dev_info(&dd->pcidev->dev, "FAIL mmap userreg: reqlen "
1035                          "%lx > PAGE\n", vma->vm_end - vma->vm_start);
1036                 ret = -EFAULT;
1037         } else {
1038                 phys = dd->ipath_physaddr + ureg;
1039                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1040
1041                 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1042                 ret = io_remap_pfn_range(vma, vma->vm_start,
1043                                          phys >> PAGE_SHIFT,
1044                                          vma->vm_end - vma->vm_start,
1045                                          vma->vm_page_prot);
1046         }
1047         return ret;
1048 }
1049
1050 static int mmap_piobufs(struct vm_area_struct *vma,
1051                         struct ipath_devdata *dd,
1052                         struct ipath_portdata *pd,
1053                         unsigned piobufs, unsigned piocnt)
1054 {
1055         unsigned long phys;
1056         int ret;
1057
1058         /*
1059          * When we map the PIO buffers in the chip, we want to map them as
1060          * writeonly, no read possible.   This prevents access to previous
1061          * process data, and catches users who might try to read the i/o
1062          * space due to a bug.
1063          */
1064         if ((vma->vm_end - vma->vm_start) > (piocnt * dd->ipath_palign)) {
1065                 dev_info(&dd->pcidev->dev, "FAIL mmap piobufs: "
1066                          "reqlen %lx > PAGE\n",
1067                          vma->vm_end - vma->vm_start);
1068                 ret = -EINVAL;
1069                 goto bail;
1070         }
1071
1072         phys = dd->ipath_physaddr + piobufs;
1073
1074 #if defined(__powerpc__)
1075         /* There isn't a generic way to specify writethrough mappings */
1076         pgprot_val(vma->vm_page_prot) |= _PAGE_NO_CACHE;
1077         pgprot_val(vma->vm_page_prot) |= _PAGE_WRITETHRU;
1078         pgprot_val(vma->vm_page_prot) &= ~_PAGE_GUARDED;
1079 #endif
1080
1081         /*
1082          * don't allow them to later change to readable with mprotect (for when
1083          * not initially mapped readable, as is normally the case)
1084          */
1085         vma->vm_flags &= ~VM_MAYREAD;
1086         vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1087
1088         ret = io_remap_pfn_range(vma, vma->vm_start, phys >> PAGE_SHIFT,
1089                                  vma->vm_end - vma->vm_start,
1090                                  vma->vm_page_prot);
1091 bail:
1092         return ret;
1093 }
1094
1095 static int mmap_rcvegrbufs(struct vm_area_struct *vma,
1096                            struct ipath_portdata *pd)
1097 {
1098         struct ipath_devdata *dd = pd->port_dd;
1099         unsigned long start, size;
1100         size_t total_size, i;
1101         unsigned long pfn;
1102         int ret;
1103
1104         size = pd->port_rcvegrbuf_size;
1105         total_size = pd->port_rcvegrbuf_chunks * size;
1106         if ((vma->vm_end - vma->vm_start) > total_size) {
1107                 dev_info(&dd->pcidev->dev, "FAIL on egr bufs: "
1108                          "reqlen %lx > actual %lx\n",
1109                          vma->vm_end - vma->vm_start,
1110                          (unsigned long) total_size);
1111                 ret = -EINVAL;
1112                 goto bail;
1113         }
1114
1115         if (vma->vm_flags & VM_WRITE) {
1116                 dev_info(&dd->pcidev->dev, "Can't map eager buffers as "
1117                          "writable (flags=%lx)\n", vma->vm_flags);
1118                 ret = -EPERM;
1119                 goto bail;
1120         }
1121         /* don't allow them to later change to writeable with mprotect */
1122         vma->vm_flags &= ~VM_MAYWRITE;
1123
1124         start = vma->vm_start;
1125
1126         for (i = 0; i < pd->port_rcvegrbuf_chunks; i++, start += size) {
1127                 pfn = virt_to_phys(pd->port_rcvegrbuf[i]) >> PAGE_SHIFT;
1128                 ret = remap_pfn_range(vma, start, pfn, size,
1129                                       vma->vm_page_prot);
1130                 if (ret < 0)
1131                         goto bail;
1132         }
1133         ret = 0;
1134
1135 bail:
1136         return ret;
1137 }
1138
1139 /*
1140  * ipath_file_vma_fault - handle a VMA page fault.
1141  */
1142 static int ipath_file_vma_fault(struct vm_area_struct *vma,
1143                                         struct vm_fault *vmf)
1144 {
1145         struct page *page;
1146
1147         page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
1148         if (!page)
1149                 return VM_FAULT_SIGBUS;
1150         get_page(page);
1151         vmf->page = page;
1152
1153         return 0;
1154 }
1155
1156 static const struct vm_operations_struct ipath_file_vm_ops = {
1157         .fault = ipath_file_vma_fault,
1158 };
1159
1160 static int mmap_kvaddr(struct vm_area_struct *vma, u64 pgaddr,
1161                        struct ipath_portdata *pd, unsigned subport)
1162 {
1163         unsigned long len;
1164         struct ipath_devdata *dd;
1165         void *addr;
1166         size_t size;
1167         int ret = 0;
1168
1169         /* If the port is not shared, all addresses should be physical */
1170         if (!pd->port_subport_cnt)
1171                 goto bail;
1172
1173         dd = pd->port_dd;
1174         size = pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
1175
1176         /*
1177          * Each process has all the subport uregbase, rcvhdrq, and
1178          * rcvegrbufs mmapped - as an array for all the processes,
1179          * and also separately for this process.
1180          */
1181         if (pgaddr == cvt_kvaddr(pd->subport_uregbase)) {
1182                 addr = pd->subport_uregbase;
1183                 size = PAGE_SIZE * pd->port_subport_cnt;
1184         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvhdr_base)) {
1185                 addr = pd->subport_rcvhdr_base;
1186                 size = pd->port_rcvhdrq_size * pd->port_subport_cnt;
1187         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvegrbuf)) {
1188                 addr = pd->subport_rcvegrbuf;
1189                 size *= pd->port_subport_cnt;
1190         } else if (pgaddr == cvt_kvaddr(pd->subport_uregbase +
1191                                         PAGE_SIZE * subport)) {
1192                 addr = pd->subport_uregbase + PAGE_SIZE * subport;
1193                 size = PAGE_SIZE;
1194         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvhdr_base +
1195                                 pd->port_rcvhdrq_size * subport)) {
1196                 addr = pd->subport_rcvhdr_base +
1197                         pd->port_rcvhdrq_size * subport;
1198                 size = pd->port_rcvhdrq_size;
1199         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvegrbuf +
1200                                size * subport)) {
1201                 addr = pd->subport_rcvegrbuf + size * subport;
1202                 /* rcvegrbufs are read-only on the slave */
1203                 if (vma->vm_flags & VM_WRITE) {
1204                         dev_info(&dd->pcidev->dev,
1205                                  "Can't map eager buffers as "
1206                                  "writable (flags=%lx)\n", vma->vm_flags);
1207                         ret = -EPERM;
1208                         goto bail;
1209                 }
1210                 /*
1211                  * Don't allow permission to later change to writeable
1212                  * with mprotect.
1213                  */
1214                 vma->vm_flags &= ~VM_MAYWRITE;
1215         } else {
1216                 goto bail;
1217         }
1218         len = vma->vm_end - vma->vm_start;
1219         if (len > size) {
1220                 ipath_cdbg(MM, "FAIL: reqlen %lx > %zx\n", len, size);
1221                 ret = -EINVAL;
1222                 goto bail;
1223         }
1224
1225         vma->vm_pgoff = (unsigned long) addr >> PAGE_SHIFT;
1226         vma->vm_ops = &ipath_file_vm_ops;
1227         vma->vm_flags |= VM_RESERVED | VM_DONTEXPAND;
1228         ret = 1;
1229
1230 bail:
1231         return ret;
1232 }
1233
1234 /**
1235  * ipath_mmap - mmap various structures into user space
1236  * @fp: the file pointer
1237  * @vma: the VM area
1238  *
1239  * We use this to have a shared buffer between the kernel and the user code
1240  * for the rcvhdr queue, egr buffers, and the per-port user regs and pio
1241  * buffers in the chip.  We have the open and close entries so we can bump
1242  * the ref count and keep the driver from being unloaded while still mapped.
1243  */
1244 static int ipath_mmap(struct file *fp, struct vm_area_struct *vma)
1245 {
1246         struct ipath_portdata *pd;
1247         struct ipath_devdata *dd;
1248         u64 pgaddr, ureg;
1249         unsigned piobufs, piocnt;
1250         int ret;
1251
1252         pd = port_fp(fp);
1253         if (!pd) {
1254                 ret = -EINVAL;
1255                 goto bail;
1256         }
1257         dd = pd->port_dd;
1258
1259         /*
1260          * This is the ipath_do_user_init() code, mapping the shared buffers
1261          * into the user process. The address referred to by vm_pgoff is the
1262          * file offset passed via mmap().  For shared ports, this is the
1263          * kernel vmalloc() address of the pages to share with the master.
1264          * For non-shared or master ports, this is a physical address.
1265          * We only do one mmap for each space mapped.
1266          */
1267         pgaddr = vma->vm_pgoff << PAGE_SHIFT;
1268
1269         /*
1270          * Check for 0 in case one of the allocations failed, but user
1271          * called mmap anyway.
1272          */
1273         if (!pgaddr)  {
1274                 ret = -EINVAL;
1275                 goto bail;
1276         }
1277
1278         ipath_cdbg(MM, "pgaddr %llx vm_start=%lx len %lx port %u:%u:%u\n",
1279                    (unsigned long long) pgaddr, vma->vm_start,
1280                    vma->vm_end - vma->vm_start, dd->ipath_unit,
1281                    pd->port_port, subport_fp(fp));
1282
1283         /*
1284          * Physical addresses must fit in 40 bits for our hardware.
1285          * Check for kernel virtual addresses first, anything else must
1286          * match a HW or memory address.
1287          */
1288         ret = mmap_kvaddr(vma, pgaddr, pd, subport_fp(fp));
1289         if (ret) {
1290                 if (ret > 0)
1291                         ret = 0;
1292                 goto bail;
1293         }
1294
1295         ureg = dd->ipath_uregbase + dd->ipath_ureg_align * pd->port_port;
1296         if (!pd->port_subport_cnt) {
1297                 /* port is not shared */
1298                 piocnt = pd->port_piocnt;
1299                 piobufs = pd->port_piobufs;
1300         } else if (!subport_fp(fp)) {
1301                 /* caller is the master */
1302                 piocnt = (pd->port_piocnt / pd->port_subport_cnt) +
1303                          (pd->port_piocnt % pd->port_subport_cnt);
1304                 piobufs = pd->port_piobufs +
1305                         dd->ipath_palign * (pd->port_piocnt - piocnt);
1306         } else {
1307                 unsigned slave = subport_fp(fp) - 1;
1308
1309                 /* caller is a slave */
1310                 piocnt = pd->port_piocnt / pd->port_subport_cnt;
1311                 piobufs = pd->port_piobufs + dd->ipath_palign * piocnt * slave;
1312         }
1313
1314         if (pgaddr == ureg)
1315                 ret = mmap_ureg(vma, dd, ureg);
1316         else if (pgaddr == piobufs)
1317                 ret = mmap_piobufs(vma, dd, pd, piobufs, piocnt);
1318         else if (pgaddr == dd->ipath_pioavailregs_phys)
1319                 /* in-memory copy of pioavail registers */
1320                 ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1321                                      (void *) dd->ipath_pioavailregs_dma,
1322                                      "pioavail registers");
1323         else if (pgaddr == pd->port_rcvegr_phys)
1324                 ret = mmap_rcvegrbufs(vma, pd);
1325         else if (pgaddr == (u64) pd->port_rcvhdrq_phys)
1326                 /*
1327                  * The rcvhdrq itself; readonly except on HT (so have
1328                  * to allow writable mapping), multiple pages, contiguous
1329                  * from an i/o perspective.
1330                  */
1331                 ret = ipath_mmap_mem(vma, pd, pd->port_rcvhdrq_size, 1,
1332                                      pd->port_rcvhdrq,
1333                                      "rcvhdrq");
1334         else if (pgaddr == (u64) pd->port_rcvhdrqtailaddr_phys)
1335                 /* in-memory copy of rcvhdrq tail register */
1336                 ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1337                                      pd->port_rcvhdrtail_kvaddr,
1338                                      "rcvhdrq tail");
1339         else
1340                 ret = -EINVAL;
1341
1342         vma->vm_private_data = NULL;
1343
1344         if (ret < 0)
1345                 dev_info(&dd->pcidev->dev,
1346                          "Failure %d on off %llx len %lx\n",
1347                          -ret, (unsigned long long)pgaddr,
1348                          vma->vm_end - vma->vm_start);
1349 bail:
1350         return ret;
1351 }
1352
1353 static unsigned ipath_poll_hdrqfull(struct ipath_portdata *pd)
1354 {
1355         unsigned pollflag = 0;
1356
1357         if ((pd->poll_type & IPATH_POLL_TYPE_OVERFLOW) &&
1358             pd->port_hdrqfull != pd->port_hdrqfull_poll) {
1359                 pollflag |= POLLIN | POLLRDNORM;
1360                 pd->port_hdrqfull_poll = pd->port_hdrqfull;
1361         }
1362
1363         return pollflag;
1364 }
1365
1366 static unsigned int ipath_poll_urgent(struct ipath_portdata *pd,
1367                                       struct file *fp,
1368                                       struct poll_table_struct *pt)
1369 {
1370         unsigned pollflag = 0;
1371         struct ipath_devdata *dd;
1372
1373         dd = pd->port_dd;
1374
1375         /* variable access in ipath_poll_hdrqfull() needs this */
1376         rmb();
1377         pollflag = ipath_poll_hdrqfull(pd);
1378
1379         if (pd->port_urgent != pd->port_urgent_poll) {
1380                 pollflag |= POLLIN | POLLRDNORM;
1381                 pd->port_urgent_poll = pd->port_urgent;
1382         }
1383
1384         if (!pollflag) {
1385                 /* this saves a spin_lock/unlock in interrupt handler... */
1386                 set_bit(IPATH_PORT_WAITING_URG, &pd->port_flag);
1387                 /* flush waiting flag so don't miss an event... */
1388                 wmb();
1389                 poll_wait(fp, &pd->port_wait, pt);
1390         }
1391
1392         return pollflag;
1393 }
1394
1395 static unsigned int ipath_poll_next(struct ipath_portdata *pd,
1396                                     struct file *fp,
1397                                     struct poll_table_struct *pt)
1398 {
1399         u32 head;
1400         u32 tail;
1401         unsigned pollflag = 0;
1402         struct ipath_devdata *dd;
1403
1404         dd = pd->port_dd;
1405
1406         /* variable access in ipath_poll_hdrqfull() needs this */
1407         rmb();
1408         pollflag = ipath_poll_hdrqfull(pd);
1409
1410         head = ipath_read_ureg32(dd, ur_rcvhdrhead, pd->port_port);
1411         if (pd->port_rcvhdrtail_kvaddr)
1412                 tail = ipath_get_rcvhdrtail(pd);
1413         else
1414                 tail = ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
1415
1416         if (head != tail)
1417                 pollflag |= POLLIN | POLLRDNORM;
1418         else {
1419                 /* this saves a spin_lock/unlock in interrupt handler */
1420                 set_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag);
1421                 /* flush waiting flag so we don't miss an event */
1422                 wmb();
1423
1424                 set_bit(pd->port_port + dd->ipath_r_intravail_shift,
1425                         &dd->ipath_rcvctrl);
1426
1427                 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1428                                  dd->ipath_rcvctrl);
1429
1430                 if (dd->ipath_rhdrhead_intr_off) /* arm rcv interrupt */
1431                         ipath_write_ureg(dd, ur_rcvhdrhead,
1432                                          dd->ipath_rhdrhead_intr_off | head,
1433                                          pd->port_port);
1434
1435                 poll_wait(fp, &pd->port_wait, pt);
1436         }
1437
1438         return pollflag;
1439 }
1440
1441 static unsigned int ipath_poll(struct file *fp,
1442                                struct poll_table_struct *pt)
1443 {
1444         struct ipath_portdata *pd;
1445         unsigned pollflag;
1446
1447         pd = port_fp(fp);
1448         if (!pd)
1449                 pollflag = 0;
1450         else if (pd->poll_type & IPATH_POLL_TYPE_URGENT)
1451                 pollflag = ipath_poll_urgent(pd, fp, pt);
1452         else
1453                 pollflag = ipath_poll_next(pd, fp, pt);
1454
1455         return pollflag;
1456 }
1457
1458 static int ipath_supports_subports(int user_swmajor, int user_swminor)
1459 {
1460         /* no subport implementation prior to software version 1.3 */
1461         return (user_swmajor > 1) || (user_swminor >= 3);
1462 }
1463
1464 static int ipath_compatible_subports(int user_swmajor, int user_swminor)
1465 {
1466         /* this code is written long-hand for clarity */
1467         if (IPATH_USER_SWMAJOR != user_swmajor) {
1468                 /* no promise of compatibility if major mismatch */
1469                 return 0;
1470         }
1471         if (IPATH_USER_SWMAJOR == 1) {
1472                 switch (IPATH_USER_SWMINOR) {
1473                 case 0:
1474                 case 1:
1475                 case 2:
1476                         /* no subport implementation so cannot be compatible */
1477                         return 0;
1478                 case 3:
1479                         /* 3 is only compatible with itself */
1480                         return user_swminor == 3;
1481                 default:
1482                         /* >= 4 are compatible (or are expected to be) */
1483                         return user_swminor >= 4;
1484                 }
1485         }
1486         /* make no promises yet for future major versions */
1487         return 0;
1488 }
1489
1490 static int init_subports(struct ipath_devdata *dd,
1491                          struct ipath_portdata *pd,
1492                          const struct ipath_user_info *uinfo)
1493 {
1494         int ret = 0;
1495         unsigned num_subports;
1496         size_t size;
1497
1498         /*
1499          * If the user is requesting zero subports,
1500          * skip the subport allocation.
1501          */
1502         if (uinfo->spu_subport_cnt <= 0)
1503                 goto bail;
1504
1505         /* Self-consistency check for ipath_compatible_subports() */
1506         if (ipath_supports_subports(IPATH_USER_SWMAJOR, IPATH_USER_SWMINOR) &&
1507             !ipath_compatible_subports(IPATH_USER_SWMAJOR,
1508                                        IPATH_USER_SWMINOR)) {
1509                 dev_info(&dd->pcidev->dev,
1510                          "Inconsistent ipath_compatible_subports()\n");
1511                 goto bail;
1512         }
1513
1514         /* Check for subport compatibility */
1515         if (!ipath_compatible_subports(uinfo->spu_userversion >> 16,
1516                                        uinfo->spu_userversion & 0xffff)) {
1517                 dev_info(&dd->pcidev->dev,
1518                          "Mismatched user version (%d.%d) and driver "
1519                          "version (%d.%d) while port sharing. Ensure "
1520                          "that driver and library are from the same "
1521                          "release.\n",
1522                          (int) (uinfo->spu_userversion >> 16),
1523                          (int) (uinfo->spu_userversion & 0xffff),
1524                          IPATH_USER_SWMAJOR,
1525                          IPATH_USER_SWMINOR);
1526                 goto bail;
1527         }
1528         if (uinfo->spu_subport_cnt > INFINIPATH_MAX_SUBPORT) {
1529                 ret = -EINVAL;
1530                 goto bail;
1531         }
1532
1533         num_subports = uinfo->spu_subport_cnt;
1534         pd->subport_uregbase = vmalloc(PAGE_SIZE * num_subports);
1535         if (!pd->subport_uregbase) {
1536                 ret = -ENOMEM;
1537                 goto bail;
1538         }
1539         /* Note: pd->port_rcvhdrq_size isn't initialized yet. */
1540         size = ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize *
1541                      sizeof(u32), PAGE_SIZE) * num_subports;
1542         pd->subport_rcvhdr_base = vmalloc(size);
1543         if (!pd->subport_rcvhdr_base) {
1544                 ret = -ENOMEM;
1545                 goto bail_ureg;
1546         }
1547
1548         pd->subport_rcvegrbuf = vmalloc(pd->port_rcvegrbuf_chunks *
1549                                         pd->port_rcvegrbuf_size *
1550                                         num_subports);
1551         if (!pd->subport_rcvegrbuf) {
1552                 ret = -ENOMEM;
1553                 goto bail_rhdr;
1554         }
1555
1556         pd->port_subport_cnt = uinfo->spu_subport_cnt;
1557         pd->port_subport_id = uinfo->spu_subport_id;
1558         pd->active_slaves = 1;
1559         set_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag);
1560         memset(pd->subport_uregbase, 0, PAGE_SIZE * num_subports);
1561         memset(pd->subport_rcvhdr_base, 0, size);
1562         memset(pd->subport_rcvegrbuf, 0, pd->port_rcvegrbuf_chunks *
1563                                          pd->port_rcvegrbuf_size *
1564                                          num_subports);
1565         goto bail;
1566
1567 bail_rhdr:
1568         vfree(pd->subport_rcvhdr_base);
1569 bail_ureg:
1570         vfree(pd->subport_uregbase);
1571         pd->subport_uregbase = NULL;
1572 bail:
1573         return ret;
1574 }
1575
1576 static int try_alloc_port(struct ipath_devdata *dd, int port,
1577                           struct file *fp,
1578                           const struct ipath_user_info *uinfo)
1579 {
1580         struct ipath_portdata *pd;
1581         int ret;
1582
1583         if (!(pd = dd->ipath_pd[port])) {
1584                 void *ptmp;
1585
1586                 pd = kzalloc(sizeof(struct ipath_portdata), GFP_KERNEL);
1587
1588                 /*
1589                  * Allocate memory for use in ipath_tid_update() just once
1590                  * at open, not per call.  Reduces cost of expected send
1591                  * setup.
1592                  */
1593                 ptmp = kmalloc(dd->ipath_rcvtidcnt * sizeof(u16) +
1594                                dd->ipath_rcvtidcnt * sizeof(struct page **),
1595                                GFP_KERNEL);
1596                 if (!pd || !ptmp) {
1597                         ipath_dev_err(dd, "Unable to allocate portdata "
1598                                       "memory, failing open\n");
1599                         ret = -ENOMEM;
1600                         kfree(pd);
1601                         kfree(ptmp);
1602                         goto bail;
1603                 }
1604                 dd->ipath_pd[port] = pd;
1605                 dd->ipath_pd[port]->port_port = port;
1606                 dd->ipath_pd[port]->port_dd = dd;
1607                 dd->ipath_pd[port]->port_tid_pg_list = ptmp;
1608                 init_waitqueue_head(&dd->ipath_pd[port]->port_wait);
1609         }
1610         if (!pd->port_cnt) {
1611                 pd->userversion = uinfo->spu_userversion;
1612                 init_user_egr_sizes(pd);
1613                 if ((ret = init_subports(dd, pd, uinfo)) != 0)
1614                         goto bail;
1615                 ipath_cdbg(PROC, "%s[%u] opened unit:port %u:%u\n",
1616                            current->comm, current->pid, dd->ipath_unit,
1617                            port);
1618                 pd->port_cnt = 1;
1619                 port_fp(fp) = pd;
1620                 pd->port_pid = get_pid(task_pid(current));
1621                 strlcpy(pd->port_comm, current->comm, sizeof(pd->port_comm));
1622                 ipath_stats.sps_ports++;
1623                 ret = 0;
1624         } else
1625                 ret = -EBUSY;
1626
1627 bail:
1628         return ret;
1629 }
1630
1631 static inline int usable(struct ipath_devdata *dd)
1632 {
1633         return dd &&
1634                 (dd->ipath_flags & IPATH_PRESENT) &&
1635                 dd->ipath_kregbase &&
1636                 dd->ipath_lid &&
1637                 !(dd->ipath_flags & (IPATH_LINKDOWN | IPATH_DISABLED
1638                                      | IPATH_LINKUNK));
1639 }
1640
1641 static int find_free_port(int unit, struct file *fp,
1642                           const struct ipath_user_info *uinfo)
1643 {
1644         struct ipath_devdata *dd = ipath_lookup(unit);
1645         int ret, i;
1646
1647         if (!dd) {
1648                 ret = -ENODEV;
1649                 goto bail;
1650         }
1651
1652         if (!usable(dd)) {
1653                 ret = -ENETDOWN;
1654                 goto bail;
1655         }
1656
1657         for (i = 1; i < dd->ipath_cfgports; i++) {
1658                 ret = try_alloc_port(dd, i, fp, uinfo);
1659                 if (ret != -EBUSY)
1660                         goto bail;
1661         }
1662         ret = -EBUSY;
1663
1664 bail:
1665         return ret;
1666 }
1667
1668 static int find_best_unit(struct file *fp,
1669                           const struct ipath_user_info *uinfo)
1670 {
1671         int ret = 0, i, prefunit = -1, devmax;
1672         int maxofallports, npresent, nup;
1673         int ndev;
1674
1675         devmax = ipath_count_units(&npresent, &nup, &maxofallports);
1676
1677         /*
1678          * This code is present to allow a knowledgeable person to
1679          * specify the layout of processes to processors before opening
1680          * this driver, and then we'll assign the process to the "closest"
1681          * InfiniPath chip to that processor (we assume reasonable connectivity,
1682          * for now).  This code assumes that if affinity has been set
1683          * before this point, that at most one cpu is set; for now this
1684          * is reasonable.  I check for both cpumask_empty() and cpumask_full(),
1685          * in case some kernel variant sets none of the bits when no
1686          * affinity is set.  2.6.11 and 12 kernels have all present
1687          * cpus set.  Some day we'll have to fix it up further to handle
1688          * a cpu subset.  This algorithm fails for two HT chips connected
1689          * in tunnel fashion.  Eventually this needs real topology
1690          * information.  There may be some issues with dual core numbering
1691          * as well.  This needs more work prior to release.
1692          */
1693         if (!cpumask_empty(&current->cpus_allowed) &&
1694             !cpumask_full(&current->cpus_allowed)) {
1695                 int ncpus = num_online_cpus(), curcpu = -1, nset = 0;
1696                 for (i = 0; i < ncpus; i++)
1697                         if (cpumask_test_cpu(i, &current->cpus_allowed)) {
1698                                 ipath_cdbg(PROC, "%s[%u] affinity set for "
1699                                            "cpu %d/%d\n", current->comm,
1700                                            current->pid, i, ncpus);
1701                                 curcpu = i;
1702                                 nset++;
1703                         }
1704                 if (curcpu != -1 && nset != ncpus) {
1705                         if (npresent) {
1706                                 prefunit = curcpu / (ncpus / npresent);
1707                                 ipath_cdbg(PROC,"%s[%u] %d chips, %d cpus, "
1708                                           "%d cpus/chip, select unit %d\n",
1709                                           current->comm, current->pid,
1710                                           npresent, ncpus, ncpus / npresent,
1711                                           prefunit);
1712                         }
1713                 }
1714         }
1715
1716         /*
1717          * user ports start at 1, kernel port is 0
1718          * For now, we do round-robin access across all chips
1719          */
1720
1721         if (prefunit != -1)
1722                 devmax = prefunit + 1;
1723 recheck:
1724         for (i = 1; i < maxofallports; i++) {
1725                 for (ndev = prefunit != -1 ? prefunit : 0; ndev < devmax;
1726                      ndev++) {
1727                         struct ipath_devdata *dd = ipath_lookup(ndev);
1728
1729                         if (!usable(dd))
1730                                 continue; /* can't use this unit */
1731                         if (i >= dd->ipath_cfgports)
1732                                 /*
1733                                  * Maxed out on users of this unit. Try
1734                                  * next.
1735                                  */
1736                                 continue;
1737                         ret = try_alloc_port(dd, i, fp, uinfo);
1738                         if (!ret)
1739                                 goto done;
1740                 }
1741         }
1742
1743         if (npresent) {
1744                 if (nup == 0) {
1745                         ret = -ENETDOWN;
1746                         ipath_dbg("No ports available (none initialized "
1747                                   "and ready)\n");
1748                 } else {
1749                         if (prefunit > 0) {
1750                                 /* if started above 0, retry from 0 */
1751                                 ipath_cdbg(PROC,
1752                                            "%s[%u] no ports on prefunit "
1753                                            "%d, clear and re-check\n",
1754                                            current->comm, current->pid,
1755                                            prefunit);
1756                                 devmax = ipath_count_units(NULL, NULL,
1757                                                            NULL);
1758                                 prefunit = -1;
1759                                 goto recheck;
1760                         }
1761                         ret = -EBUSY;
1762                         ipath_dbg("No ports available\n");
1763                 }
1764         } else {
1765                 ret = -ENXIO;
1766                 ipath_dbg("No boards found\n");
1767         }
1768
1769 done:
1770         return ret;
1771 }
1772
1773 static int find_shared_port(struct file *fp,
1774                             const struct ipath_user_info *uinfo)
1775 {
1776         int devmax, ndev, i;
1777         int ret = 0;
1778
1779         devmax = ipath_count_units(NULL, NULL, NULL);
1780
1781         for (ndev = 0; ndev < devmax; ndev++) {
1782                 struct ipath_devdata *dd = ipath_lookup(ndev);
1783
1784                 if (!usable(dd))
1785                         continue;
1786                 for (i = 1; i < dd->ipath_cfgports; i++) {
1787                         struct ipath_portdata *pd = dd->ipath_pd[i];
1788
1789                         /* Skip ports which are not yet open */
1790                         if (!pd || !pd->port_cnt)
1791                                 continue;
1792                         /* Skip port if it doesn't match the requested one */
1793                         if (pd->port_subport_id != uinfo->spu_subport_id)
1794                                 continue;
1795                         /* Verify the sharing process matches the master */
1796                         if (pd->port_subport_cnt != uinfo->spu_subport_cnt ||
1797                             pd->userversion != uinfo->spu_userversion ||
1798                             pd->port_cnt >= pd->port_subport_cnt) {
1799                                 ret = -EINVAL;
1800                                 goto done;
1801                         }
1802                         port_fp(fp) = pd;
1803                         subport_fp(fp) = pd->port_cnt++;
1804                         pd->port_subpid[subport_fp(fp)] =
1805                                 get_pid(task_pid(current));
1806                         tidcursor_fp(fp) = 0;
1807                         pd->active_slaves |= 1 << subport_fp(fp);
1808                         ipath_cdbg(PROC,
1809                                    "%s[%u] %u sharing %s[%u] unit:port %u:%u\n",
1810                                    current->comm, current->pid,
1811                                    subport_fp(fp),
1812                                    pd->port_comm, pid_nr(pd->port_pid),
1813                                    dd->ipath_unit, pd->port_port);
1814                         ret = 1;
1815                         goto done;
1816                 }
1817         }
1818
1819 done:
1820         return ret;
1821 }
1822
1823 static int ipath_open(struct inode *in, struct file *fp)
1824 {
1825         /* The real work is performed later in ipath_assign_port() */
1826         fp->private_data = kzalloc(sizeof(struct ipath_filedata), GFP_KERNEL);
1827         return fp->private_data ? 0 : -ENOMEM;
1828 }
1829
1830 /* Get port early, so can set affinity prior to memory allocation */
1831 static int ipath_assign_port(struct file *fp,
1832                               const struct ipath_user_info *uinfo)
1833 {
1834         int ret;
1835         int i_minor;
1836         unsigned swmajor, swminor;
1837
1838         /* Check to be sure we haven't already initialized this file */
1839         if (port_fp(fp)) {
1840                 ret = -EINVAL;
1841                 goto done;
1842         }
1843
1844         /* for now, if major version is different, bail */
1845         swmajor = uinfo->spu_userversion >> 16;
1846         if (swmajor != IPATH_USER_SWMAJOR) {
1847                 ipath_dbg("User major version %d not same as driver "
1848                           "major %d\n", uinfo->spu_userversion >> 16,
1849                           IPATH_USER_SWMAJOR);
1850                 ret = -ENODEV;
1851                 goto done;
1852         }
1853
1854         swminor = uinfo->spu_userversion & 0xffff;
1855         if (swminor != IPATH_USER_SWMINOR)
1856                 ipath_dbg("User minor version %d not same as driver "
1857                           "minor %d\n", swminor, IPATH_USER_SWMINOR);
1858
1859         mutex_lock(&ipath_mutex);
1860
1861         if (ipath_compatible_subports(swmajor, swminor) &&
1862             uinfo->spu_subport_cnt &&
1863             (ret = find_shared_port(fp, uinfo))) {
1864                 if (ret > 0)
1865                         ret = 0;
1866                 goto done_chk_sdma;
1867         }
1868
1869         i_minor = iminor(fp->f_path.dentry->d_inode) - IPATH_USER_MINOR_BASE;
1870         ipath_cdbg(VERBOSE, "open on dev %lx (minor %d)\n",
1871                    (long)fp->f_path.dentry->d_inode->i_rdev, i_minor);
1872
1873         if (i_minor)
1874                 ret = find_free_port(i_minor - 1, fp, uinfo);
1875         else
1876                 ret = find_best_unit(fp, uinfo);
1877
1878 done_chk_sdma:
1879         if (!ret) {
1880                 struct ipath_filedata *fd = fp->private_data;
1881                 const struct ipath_portdata *pd = fd->pd;
1882                 const struct ipath_devdata *dd = pd->port_dd;
1883
1884                 fd->pq = ipath_user_sdma_queue_create(&dd->pcidev->dev,
1885                                                       dd->ipath_unit,
1886                                                       pd->port_port,
1887                                                       fd->subport);
1888
1889                 if (!fd->pq)
1890                         ret = -ENOMEM;
1891         }
1892
1893         mutex_unlock(&ipath_mutex);
1894
1895 done:
1896         return ret;
1897 }
1898
1899
1900 static int ipath_do_user_init(struct file *fp,
1901                               const struct ipath_user_info *uinfo)
1902 {
1903         int ret;
1904         struct ipath_portdata *pd = port_fp(fp);
1905         struct ipath_devdata *dd;
1906         u32 head32;
1907
1908         /* Subports don't need to initialize anything since master did it. */
1909         if (subport_fp(fp)) {
1910                 ret = wait_event_interruptible(pd->port_wait,
1911                         !test_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag));
1912                 goto done;
1913         }
1914
1915         dd = pd->port_dd;
1916
1917         if (uinfo->spu_rcvhdrsize) {
1918                 ret = ipath_setrcvhdrsize(dd, uinfo->spu_rcvhdrsize);
1919                 if (ret)
1920                         goto done;
1921         }
1922
1923         /* for now we do nothing with rcvhdrcnt: uinfo->spu_rcvhdrcnt */
1924
1925         /* some ports may get extra buffers, calculate that here */
1926         if (pd->port_port <= dd->ipath_ports_extrabuf)
1927                 pd->port_piocnt = dd->ipath_pbufsport + 1;
1928         else
1929                 pd->port_piocnt = dd->ipath_pbufsport;
1930
1931         /* for right now, kernel piobufs are at end, so port 1 is at 0 */
1932         if (pd->port_port <= dd->ipath_ports_extrabuf)
1933                 pd->port_pio_base = (dd->ipath_pbufsport + 1)
1934                         * (pd->port_port - 1);
1935         else
1936                 pd->port_pio_base = dd->ipath_ports_extrabuf +
1937                         dd->ipath_pbufsport * (pd->port_port - 1);
1938         pd->port_piobufs = dd->ipath_piobufbase +
1939                 pd->port_pio_base * dd->ipath_palign;
1940         ipath_cdbg(VERBOSE, "piobuf base for port %u is 0x%x, piocnt %u,"
1941                 " first pio %u\n", pd->port_port, pd->port_piobufs,
1942                 pd->port_piocnt, pd->port_pio_base);
1943         ipath_chg_pioavailkernel(dd, pd->port_pio_base, pd->port_piocnt, 0);
1944
1945         /*
1946          * Now allocate the rcvhdr Q and eager TIDs; skip the TID
1947          * array for time being.  If pd->port_port > chip-supported,
1948          * we need to do extra stuff here to handle by handling overflow
1949          * through port 0, someday
1950          */
1951         ret = ipath_create_rcvhdrq(dd, pd);
1952         if (!ret)
1953                 ret = ipath_create_user_egr(pd);
1954         if (ret)
1955                 goto done;
1956
1957         /*
1958          * set the eager head register for this port to the current values
1959          * of the tail pointers, since we don't know if they were
1960          * updated on last use of the port.
1961          */
1962         head32 = ipath_read_ureg32(dd, ur_rcvegrindextail, pd->port_port);
1963         ipath_write_ureg(dd, ur_rcvegrindexhead, head32, pd->port_port);
1964         pd->port_lastrcvhdrqtail = -1;
1965         ipath_cdbg(VERBOSE, "Wrote port%d egrhead %x from tail regs\n",
1966                 pd->port_port, head32);
1967         pd->port_tidcursor = 0; /* start at beginning after open */
1968
1969         /* initialize poll variables... */
1970         pd->port_urgent = 0;
1971         pd->port_urgent_poll = 0;
1972         pd->port_hdrqfull_poll = pd->port_hdrqfull;
1973
1974         /*
1975          * Now enable the port for receive.
1976          * For chips that are set to DMA the tail register to memory
1977          * when they change (and when the update bit transitions from
1978          * 0 to 1.  So for those chips, we turn it off and then back on.
1979          * This will (very briefly) affect any other open ports, but the
1980          * duration is very short, and therefore isn't an issue.  We
1981          * explictly set the in-memory tail copy to 0 beforehand, so we
1982          * don't have to wait to be sure the DMA update has happened
1983          * (chip resets head/tail to 0 on transition to enable).
1984          */
1985         set_bit(dd->ipath_r_portenable_shift + pd->port_port,
1986                 &dd->ipath_rcvctrl);
1987         if (!(dd->ipath_flags & IPATH_NODMA_RTAIL)) {
1988                 if (pd->port_rcvhdrtail_kvaddr)
1989                         ipath_clear_rcvhdrtail(pd);
1990                 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1991                         dd->ipath_rcvctrl &
1992                         ~(1ULL << dd->ipath_r_tailupd_shift));
1993         }
1994         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1995                          dd->ipath_rcvctrl);
1996         /* Notify any waiting slaves */
1997         if (pd->port_subport_cnt) {
1998                 clear_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag);
1999                 wake_up(&pd->port_wait);
2000         }
2001 done:
2002         return ret;
2003 }
2004
2005 /**
2006  * unlock_exptid - unlock any expected TID entries port still had in use
2007  * @pd: port
2008  *
2009  * We don't actually update the chip here, because we do a bulk update
2010  * below, using ipath_f_clear_tids.
2011  */
2012 static void unlock_expected_tids(struct ipath_portdata *pd)
2013 {
2014         struct ipath_devdata *dd = pd->port_dd;
2015         int port_tidbase = pd->port_port * dd->ipath_rcvtidcnt;
2016         int i, cnt = 0, maxtid = port_tidbase + dd->ipath_rcvtidcnt;
2017
2018         ipath_cdbg(VERBOSE, "Port %u unlocking any locked expTID pages\n",
2019                    pd->port_port);
2020         for (i = port_tidbase; i < maxtid; i++) {
2021                 struct page *ps = dd->ipath_pageshadow[i];
2022
2023                 if (!ps)
2024                         continue;
2025
2026                 dd->ipath_pageshadow[i] = NULL;
2027                 pci_unmap_page(dd->pcidev, dd->ipath_physshadow[i],
2028                         PAGE_SIZE, PCI_DMA_FROMDEVICE);
2029                 ipath_release_user_pages_on_close(&ps, 1);
2030                 cnt++;
2031                 ipath_stats.sps_pageunlocks++;
2032         }
2033         if (cnt)
2034                 ipath_cdbg(VERBOSE, "Port %u locked %u expTID entries\n",
2035                            pd->port_port, cnt);
2036
2037         if (ipath_stats.sps_pagelocks || ipath_stats.sps_pageunlocks)
2038                 ipath_cdbg(VERBOSE, "%llu pages locked, %llu unlocked\n",
2039                            (unsigned long long) ipath_stats.sps_pagelocks,
2040                            (unsigned long long)
2041                            ipath_stats.sps_pageunlocks);
2042 }
2043
2044 static int ipath_close(struct inode *in, struct file *fp)
2045 {
2046         int ret = 0;
2047         struct ipath_filedata *fd;
2048         struct ipath_portdata *pd;
2049         struct ipath_devdata *dd;
2050         unsigned long flags;
2051         unsigned port;
2052         struct pid *pid;
2053
2054         ipath_cdbg(VERBOSE, "close on dev %lx, private data %p\n",
2055                    (long)in->i_rdev, fp->private_data);
2056
2057         mutex_lock(&ipath_mutex);
2058
2059         fd = fp->private_data;
2060         fp->private_data = NULL;
2061         pd = fd->pd;
2062         if (!pd) {
2063                 mutex_unlock(&ipath_mutex);
2064                 goto bail;
2065         }
2066
2067         dd = pd->port_dd;
2068
2069         /* drain user sdma queue */
2070         ipath_user_sdma_queue_drain(dd, fd->pq);
2071         ipath_user_sdma_queue_destroy(fd->pq);
2072
2073         if (--pd->port_cnt) {
2074                 /*
2075                  * XXX If the master closes the port before the slave(s),
2076                  * revoke the mmap for the eager receive queue so
2077                  * the slave(s) don't wait for receive data forever.
2078                  */
2079                 pd->active_slaves &= ~(1 << fd->subport);
2080                 put_pid(pd->port_subpid[fd->subport]);
2081                 pd->port_subpid[fd->subport] = NULL;
2082                 mutex_unlock(&ipath_mutex);
2083                 goto bail;
2084         }
2085         /* early; no interrupt users after this */
2086         spin_lock_irqsave(&dd->ipath_uctxt_lock, flags);
2087         port = pd->port_port;
2088         dd->ipath_pd[port] = NULL;
2089         pid = pd->port_pid;
2090         pd->port_pid = NULL;
2091         spin_unlock_irqrestore(&dd->ipath_uctxt_lock, flags);
2092
2093         if (pd->port_rcvwait_to || pd->port_piowait_to
2094             || pd->port_rcvnowait || pd->port_pionowait) {
2095                 ipath_cdbg(VERBOSE, "port%u, %u rcv, %u pio wait timeo; "
2096                            "%u rcv %u, pio already\n",
2097                            pd->port_port, pd->port_rcvwait_to,
2098                            pd->port_piowait_to, pd->port_rcvnowait,
2099                            pd->port_pionowait);
2100                 pd->port_rcvwait_to = pd->port_piowait_to =
2101                         pd->port_rcvnowait = pd->port_pionowait = 0;
2102         }
2103         if (pd->port_flag) {
2104                 ipath_cdbg(PROC, "port %u port_flag set: 0x%lx\n",
2105                           pd->port_port, pd->port_flag);
2106                 pd->port_flag = 0;
2107         }
2108
2109         if (dd->ipath_kregbase) {
2110                 /* atomically clear receive enable port and intr avail. */
2111                 clear_bit(dd->ipath_r_portenable_shift + port,
2112                           &dd->ipath_rcvctrl);
2113                 clear_bit(pd->port_port + dd->ipath_r_intravail_shift,
2114                           &dd->ipath_rcvctrl);
2115                 ipath_write_kreg( dd, dd->ipath_kregs->kr_rcvctrl,
2116                         dd->ipath_rcvctrl);
2117                 /* and read back from chip to be sure that nothing
2118                  * else is in flight when we do the rest */
2119                 (void)ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
2120
2121                 /* clean up the pkeys for this port user */
2122                 ipath_clean_part_key(pd, dd);
2123                 /*
2124                  * be paranoid, and never write 0's to these, just use an
2125                  * unused part of the port 0 tail page.  Of course,
2126                  * rcvhdraddr points to a large chunk of memory, so this
2127                  * could still trash things, but at least it won't trash
2128                  * page 0, and by disabling the port, it should stop "soon",
2129                  * even if a packet or two is in already in flight after we
2130                  * disabled the port.
2131                  */
2132                 ipath_write_kreg_port(dd,
2133                         dd->ipath_kregs->kr_rcvhdrtailaddr, port,
2134                         dd->ipath_dummy_hdrq_phys);
2135                 ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdraddr,
2136                         pd->port_port, dd->ipath_dummy_hdrq_phys);
2137
2138                 ipath_disarm_piobufs(dd, pd->port_pio_base, pd->port_piocnt);
2139                 ipath_chg_pioavailkernel(dd, pd->port_pio_base,
2140                         pd->port_piocnt, 1);
2141
2142                 dd->ipath_f_clear_tids(dd, pd->port_port);
2143
2144                 if (dd->ipath_pageshadow)
2145                         unlock_expected_tids(pd);
2146                 ipath_stats.sps_ports--;
2147                 ipath_cdbg(PROC, "%s[%u] closed port %u:%u\n",
2148                            pd->port_comm, pid_nr(pid),
2149                            dd->ipath_unit, port);
2150         }
2151
2152         put_pid(pid);
2153         mutex_unlock(&ipath_mutex);
2154         ipath_free_pddata(dd, pd); /* after releasing the mutex */
2155
2156 bail:
2157         kfree(fd);
2158         return ret;
2159 }
2160
2161 static int ipath_port_info(struct ipath_portdata *pd, u16 subport,
2162                            struct ipath_port_info __user *uinfo)
2163 {
2164         struct ipath_port_info info;
2165         int nup;
2166         int ret;
2167         size_t sz;
2168
2169         (void) ipath_count_units(NULL, &nup, NULL);
2170         info.num_active = nup;
2171         info.unit = pd->port_dd->ipath_unit;
2172         info.port = pd->port_port;
2173         info.subport = subport;
2174         /* Don't return new fields if old library opened the port. */
2175         if (ipath_supports_subports(pd->userversion >> 16,
2176                                     pd->userversion & 0xffff)) {
2177                 /* Number of user ports available for this device. */
2178                 info.num_ports = pd->port_dd->ipath_cfgports - 1;
2179                 info.num_subports = pd->port_subport_cnt;
2180                 sz = sizeof(info);
2181         } else
2182                 sz = sizeof(info) - 2 * sizeof(u16);
2183
2184         if (copy_to_user(uinfo, &info, sz)) {
2185                 ret = -EFAULT;
2186                 goto bail;
2187         }
2188         ret = 0;
2189
2190 bail:
2191         return ret;
2192 }
2193
2194 static int ipath_get_slave_info(struct ipath_portdata *pd,
2195                                 void __user *slave_mask_addr)
2196 {
2197         int ret = 0;
2198
2199         if (copy_to_user(slave_mask_addr, &pd->active_slaves, sizeof(u32)))
2200                 ret = -EFAULT;
2201         return ret;
2202 }
2203
2204 static int ipath_sdma_get_inflight(struct ipath_user_sdma_queue *pq,
2205                                    u32 __user *inflightp)
2206 {
2207         const u32 val = ipath_user_sdma_inflight_counter(pq);
2208
2209         if (put_user(val, inflightp))
2210                 return -EFAULT;
2211
2212         return 0;
2213 }
2214
2215 static int ipath_sdma_get_complete(struct ipath_devdata *dd,
2216                                    struct ipath_user_sdma_queue *pq,
2217                                    u32 __user *completep)
2218 {
2219         u32 val;
2220         int err;
2221
2222         err = ipath_user_sdma_make_progress(dd, pq);
2223         if (err < 0)
2224                 return err;
2225
2226         val = ipath_user_sdma_complete_counter(pq);
2227         if (put_user(val, completep))
2228                 return -EFAULT;
2229
2230         return 0;
2231 }
2232
2233 static ssize_t ipath_write(struct file *fp, const char __user *data,
2234                            size_t count, loff_t *off)
2235 {
2236         const struct ipath_cmd __user *ucmd;
2237         struct ipath_portdata *pd;
2238         const void __user *src;
2239         size_t consumed, copy;
2240         struct ipath_cmd cmd;
2241         ssize_t ret = 0;
2242         void *dest;
2243
2244         if (count < sizeof(cmd.type)) {
2245                 ret = -EINVAL;
2246                 goto bail;
2247         }
2248
2249         ucmd = (const struct ipath_cmd __user *) data;
2250
2251         if (copy_from_user(&cmd.type, &ucmd->type, sizeof(cmd.type))) {
2252                 ret = -EFAULT;
2253                 goto bail;
2254         }
2255
2256         consumed = sizeof(cmd.type);
2257
2258         switch (cmd.type) {
2259         case IPATH_CMD_ASSIGN_PORT:
2260         case __IPATH_CMD_USER_INIT:
2261         case IPATH_CMD_USER_INIT:
2262                 copy = sizeof(cmd.cmd.user_info);
2263                 dest = &cmd.cmd.user_info;
2264                 src = &ucmd->cmd.user_info;
2265                 break;
2266         case IPATH_CMD_RECV_CTRL:
2267                 copy = sizeof(cmd.cmd.recv_ctrl);
2268                 dest = &cmd.cmd.recv_ctrl;
2269                 src = &ucmd->cmd.recv_ctrl;
2270                 break;
2271         case IPATH_CMD_PORT_INFO:
2272                 copy = sizeof(cmd.cmd.port_info);
2273                 dest = &cmd.cmd.port_info;
2274                 src = &ucmd->cmd.port_info;
2275                 break;
2276         case IPATH_CMD_TID_UPDATE:
2277         case IPATH_CMD_TID_FREE:
2278                 copy = sizeof(cmd.cmd.tid_info);
2279                 dest = &cmd.cmd.tid_info;
2280                 src = &ucmd->cmd.tid_info;
2281                 break;
2282         case IPATH_CMD_SET_PART_KEY:
2283                 copy = sizeof(cmd.cmd.part_key);
2284                 dest = &cmd.cmd.part_key;
2285                 src = &ucmd->cmd.part_key;
2286                 break;
2287         case __IPATH_CMD_SLAVE_INFO:
2288                 copy = sizeof(cmd.cmd.slave_mask_addr);
2289                 dest = &cmd.cmd.slave_mask_addr;
2290                 src = &ucmd->cmd.slave_mask_addr;
2291                 break;
2292         case IPATH_CMD_PIOAVAILUPD:     // force an update of PIOAvail reg
2293                 copy = 0;
2294                 src = NULL;
2295                 dest = NULL;
2296                 break;
2297         case IPATH_CMD_POLL_TYPE:
2298                 copy = sizeof(cmd.cmd.poll_type);
2299                 dest = &cmd.cmd.poll_type;
2300                 src = &ucmd->cmd.poll_type;
2301                 break;
2302         case IPATH_CMD_ARMLAUNCH_CTRL:
2303                 copy = sizeof(cmd.cmd.armlaunch_ctrl);
2304                 dest = &cmd.cmd.armlaunch_ctrl;
2305                 src = &ucmd->cmd.armlaunch_ctrl;
2306                 break;
2307         case IPATH_CMD_SDMA_INFLIGHT:
2308                 copy = sizeof(cmd.cmd.sdma_inflight);
2309                 dest = &cmd.cmd.sdma_inflight;
2310                 src = &ucmd->cmd.sdma_inflight;
2311                 break;
2312         case IPATH_CMD_SDMA_COMPLETE:
2313                 copy = sizeof(cmd.cmd.sdma_complete);
2314                 dest = &cmd.cmd.sdma_complete;
2315                 src = &ucmd->cmd.sdma_complete;
2316                 break;
2317         default:
2318                 ret = -EINVAL;
2319                 goto bail;
2320         }
2321
2322         if (copy) {
2323                 if ((count - consumed) < copy) {
2324                         ret = -EINVAL;
2325                         goto bail;
2326                 }
2327
2328                 if (copy_from_user(dest, src, copy)) {
2329                         ret = -EFAULT;
2330                         goto bail;
2331                 }
2332
2333                 consumed += copy;
2334         }
2335
2336         pd = port_fp(fp);
2337         if (!pd && cmd.type != __IPATH_CMD_USER_INIT &&
2338                 cmd.type != IPATH_CMD_ASSIGN_PORT) {
2339                 ret = -EINVAL;
2340                 goto bail;
2341         }
2342
2343         switch (cmd.type) {
2344         case IPATH_CMD_ASSIGN_PORT:
2345                 ret = ipath_assign_port(fp, &cmd.cmd.user_info);
2346                 if (ret)
2347                         goto bail;
2348                 break;
2349         case __IPATH_CMD_USER_INIT:
2350                 /* backwards compatibility, get port first */
2351                 ret = ipath_assign_port(fp, &cmd.cmd.user_info);
2352                 if (ret)
2353                         goto bail;
2354                 /* and fall through to current version. */
2355         case IPATH_CMD_USER_INIT:
2356                 ret = ipath_do_user_init(fp, &cmd.cmd.user_info);
2357                 if (ret)
2358                         goto bail;
2359                 ret = ipath_get_base_info(
2360                         fp, (void __user *) (unsigned long)
2361                         cmd.cmd.user_info.spu_base_info,
2362                         cmd.cmd.user_info.spu_base_info_size);
2363                 break;
2364         case IPATH_CMD_RECV_CTRL:
2365                 ret = ipath_manage_rcvq(pd, subport_fp(fp), cmd.cmd.recv_ctrl);
2366                 break;
2367         case IPATH_CMD_PORT_INFO:
2368                 ret = ipath_port_info(pd, subport_fp(fp),
2369                                       (struct ipath_port_info __user *)
2370                                       (unsigned long) cmd.cmd.port_info);
2371                 break;
2372         case IPATH_CMD_TID_UPDATE:
2373                 ret = ipath_tid_update(pd, fp, &cmd.cmd.tid_info);
2374                 break;
2375         case IPATH_CMD_TID_FREE:
2376                 ret = ipath_tid_free(pd, subport_fp(fp), &cmd.cmd.tid_info);
2377                 break;
2378         case IPATH_CMD_SET_PART_KEY:
2379                 ret = ipath_set_part_key(pd, cmd.cmd.part_key);
2380                 break;
2381         case __IPATH_CMD_SLAVE_INFO:
2382                 ret = ipath_get_slave_info(pd,
2383                                            (void __user *) (unsigned long)
2384                                            cmd.cmd.slave_mask_addr);
2385                 break;
2386         case IPATH_CMD_PIOAVAILUPD:
2387                 ipath_force_pio_avail_update(pd->port_dd);
2388                 break;
2389         case IPATH_CMD_POLL_TYPE:
2390                 pd->poll_type = cmd.cmd.poll_type;
2391                 break;
2392         case IPATH_CMD_ARMLAUNCH_CTRL:
2393                 if (cmd.cmd.armlaunch_ctrl)
2394                         ipath_enable_armlaunch(pd->port_dd);
2395                 else
2396                         ipath_disable_armlaunch(pd->port_dd);
2397                 break;
2398         case IPATH_CMD_SDMA_INFLIGHT:
2399                 ret = ipath_sdma_get_inflight(user_sdma_queue_fp(fp),
2400                                               (u32 __user *) (unsigned long)
2401                                               cmd.cmd.sdma_inflight);
2402                 break;
2403         case IPATH_CMD_SDMA_COMPLETE:
2404                 ret = ipath_sdma_get_complete(pd->port_dd,
2405                                               user_sdma_queue_fp(fp),
2406                                               (u32 __user *) (unsigned long)
2407                                               cmd.cmd.sdma_complete);
2408                 break;
2409         }
2410
2411         if (ret >= 0)
2412                 ret = consumed;
2413
2414 bail:
2415         return ret;
2416 }
2417
2418 static ssize_t ipath_writev(struct kiocb *iocb, const struct iovec *iov,
2419                             unsigned long dim, loff_t off)
2420 {
2421         struct file *filp = iocb->ki_filp;
2422         struct ipath_filedata *fp = filp->private_data;
2423         struct ipath_portdata *pd = port_fp(filp);
2424         struct ipath_user_sdma_queue *pq = fp->pq;
2425
2426         if (!dim)
2427                 return -EINVAL;
2428
2429         return ipath_user_sdma_writev(pd->port_dd, pq, iov, dim);
2430 }
2431
2432 static struct class *ipath_class;
2433
2434 static int init_cdev(int minor, char *name, const struct file_operations *fops,
2435                      struct cdev **cdevp, struct device **devp)
2436 {
2437         const dev_t dev = MKDEV(IPATH_MAJOR, minor);
2438         struct cdev *cdev = NULL;
2439         struct device *device = NULL;
2440         int ret;
2441
2442         cdev = cdev_alloc();
2443         if (!cdev) {
2444                 printk(KERN_ERR IPATH_DRV_NAME
2445                        ": Could not allocate cdev for minor %d, %s\n",
2446                        minor, name);
2447                 ret = -ENOMEM;
2448                 goto done;
2449         }
2450
2451         cdev->owner = THIS_MODULE;
2452         cdev->ops = fops;
2453         kobject_set_name(&cdev->kobj, name);
2454
2455         ret = cdev_add(cdev, dev, 1);
2456         if (ret < 0) {
2457                 printk(KERN_ERR IPATH_DRV_NAME
2458                        ": Could not add cdev for minor %d, %s (err %d)\n",
2459                        minor, name, -ret);
2460                 goto err_cdev;
2461         }
2462
2463         device = device_create(ipath_class, NULL, dev, NULL, name);
2464
2465         if (IS_ERR(device)) {
2466                 ret = PTR_ERR(device);
2467                 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2468                        "device for minor %d, %s (err %d)\n",
2469                        minor, name, -ret);
2470                 goto err_cdev;
2471         }
2472
2473         goto done;
2474
2475 err_cdev:
2476         cdev_del(cdev);
2477         cdev = NULL;
2478
2479 done:
2480         if (ret >= 0) {
2481                 *cdevp = cdev;
2482                 *devp = device;
2483         } else {
2484                 *cdevp = NULL;
2485                 *devp = NULL;
2486         }
2487
2488         return ret;
2489 }
2490
2491 int ipath_cdev_init(int minor, char *name, const struct file_operations *fops,
2492                     struct cdev **cdevp, struct device **devp)
2493 {
2494         return init_cdev(minor, name, fops, cdevp, devp);
2495 }
2496
2497 static void cleanup_cdev(struct cdev **cdevp,
2498                          struct device **devp)
2499 {
2500         struct device *dev = *devp;
2501
2502         if (dev) {
2503                 device_unregister(dev);
2504                 *devp = NULL;
2505         }
2506
2507         if (*cdevp) {
2508                 cdev_del(*cdevp);
2509                 *cdevp = NULL;
2510         }
2511 }
2512
2513 void ipath_cdev_cleanup(struct cdev **cdevp,
2514                         struct device **devp)
2515 {
2516         cleanup_cdev(cdevp, devp);
2517 }
2518
2519 static struct cdev *wildcard_cdev;
2520 static struct device *wildcard_dev;
2521
2522 static const dev_t dev = MKDEV(IPATH_MAJOR, 0);
2523
2524 static int user_init(void)
2525 {
2526         int ret;
2527
2528         ret = register_chrdev_region(dev, IPATH_NMINORS, IPATH_DRV_NAME);
2529         if (ret < 0) {
2530                 printk(KERN_ERR IPATH_DRV_NAME ": Could not register "
2531                        "chrdev region (err %d)\n", -ret);
2532                 goto done;
2533         }
2534
2535         ipath_class = class_create(THIS_MODULE, IPATH_DRV_NAME);
2536
2537         if (IS_ERR(ipath_class)) {
2538                 ret = PTR_ERR(ipath_class);
2539                 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2540                        "device class (err %d)\n", -ret);
2541                 goto bail;
2542         }
2543
2544         goto done;
2545 bail:
2546         unregister_chrdev_region(dev, IPATH_NMINORS);
2547 done:
2548         return ret;
2549 }
2550
2551 static void user_cleanup(void)
2552 {
2553         if (ipath_class) {
2554                 class_destroy(ipath_class);
2555                 ipath_class = NULL;
2556         }
2557
2558         unregister_chrdev_region(dev, IPATH_NMINORS);
2559 }
2560
2561 static atomic_t user_count = ATOMIC_INIT(0);
2562 static atomic_t user_setup = ATOMIC_INIT(0);
2563
2564 int ipath_user_add(struct ipath_devdata *dd)
2565 {
2566         char name[10];
2567         int ret;
2568
2569         if (atomic_inc_return(&user_count) == 1) {
2570                 ret = user_init();
2571                 if (ret < 0) {
2572                         ipath_dev_err(dd, "Unable to set up user support: "
2573                                       "error %d\n", -ret);
2574                         goto bail;
2575                 }
2576                 ret = init_cdev(0, "ipath", &ipath_file_ops, &wildcard_cdev,
2577                                 &wildcard_dev);
2578                 if (ret < 0) {
2579                         ipath_dev_err(dd, "Could not create wildcard "
2580                                       "minor: error %d\n", -ret);
2581                         goto bail_user;
2582                 }
2583
2584                 atomic_set(&user_setup, 1);
2585         }
2586
2587         snprintf(name, sizeof(name), "ipath%d", dd->ipath_unit);
2588
2589         ret = init_cdev(dd->ipath_unit + 1, name, &ipath_file_ops,
2590                         &dd->user_cdev, &dd->user_dev);
2591         if (ret < 0)
2592                 ipath_dev_err(dd, "Could not create user minor %d, %s\n",
2593                               dd->ipath_unit + 1, name);
2594
2595         goto bail;
2596
2597 bail_user:
2598         user_cleanup();
2599 bail:
2600         return ret;
2601 }
2602
2603 void ipath_user_remove(struct ipath_devdata *dd)
2604 {
2605         cleanup_cdev(&dd->user_cdev, &dd->user_dev);
2606
2607         if (atomic_dec_return(&user_count) == 0) {
2608                 if (atomic_read(&user_setup) == 0)
2609                         goto bail;
2610
2611                 cleanup_cdev(&wildcard_cdev, &wildcard_dev);
2612                 user_cleanup();
2613
2614                 atomic_set(&user_setup, 0);
2615         }
2616 bail:
2617         return;
2618 }