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