drivers/infiniband/hw/qib/qib_file_ops.c

   1 /*
   2  * Copyright (c) 2006, 2007, 2008, 2009, 2010 QLogic Corporation.
   3  * All rights reserved.
   4  * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the
  10  * OpenIB.org BSD license below:
  11  *
  12  *     Redistribution and use in source and binary forms, with or
  13  *     without modification, are permitted provided that the following
  14  *     conditions are met:
  15  *
  16  *      - Redistributions of source code must retain the above
  17  *        copyright notice, this list of conditions and the following
  18  *        disclaimer.
  19  *
  20  *      - Redistributions in binary form must reproduce the above
  21  *        copyright notice, this list of conditions and the following
  22  *        disclaimer in the documentation and/or other materials
  23  *        provided with the distribution.
  24  *
  25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  32  * SOFTWARE.
  33  */
  34
  35 #include <linux/pci.h>
  36 #include <linux/poll.h>
  37 #include <linux/cdev.h>
  38 #include <linux/swap.h>
  39 #include <linux/vmalloc.h>
  40 #include <linux/highmem.h>
  41 #include <linux/io.h>
  42 #include <linux/uio.h>
  43 #include <linux/jiffies.h>
  44 #include <asm/pgtable.h>
  45 #include <linux/delay.h>
  46 #include <linux/export.h>
  47
  48 #include "qib.h"
  49 #include "qib_common.h"
  50 #include "qib_user_sdma.h"
  51
  52 static int qib_open(struct inode *, struct file *);
  53 static int qib_close(struct inode *, struct file *);
  54 static ssize_t qib_write(struct file *, const char __user *, size_t, loff_t *);
  55 static ssize_t qib_aio_write(struct kiocb *, const struct iovec *,
  56                              unsigned long, loff_t);
  57 static unsigned int qib_poll(struct file *, struct poll_table_struct *);
  58 static int qib_mmapf(struct file *, struct vm_area_struct *);
  59
  60 static const struct file_operations qib_file_ops = {
  61         .owner = THIS_MODULE,
  62         .write = qib_write,
  63         .aio_write = qib_aio_write,
  64         .open = qib_open,
  65         .release = qib_close,
  66         .poll = qib_poll,
  67         .mmap = qib_mmapf,
  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 qib_get_base_info(struct file *fp, void __user *ubase,
  90                              size_t ubase_size)
  91 {
  92         struct qib_ctxtdata *rcd = ctxt_fp(fp);
  93         int ret = 0;
  94         struct qib_base_info *kinfo = NULL;
  95         struct qib_devdata *dd = rcd->dd;
  96         struct qib_pportdata *ppd = rcd->ppd;
  97         unsigned subctxt_cnt;
  98         int shared, master;
  99         size_t sz;
 100
 101         subctxt_cnt = rcd->subctxt_cnt;
 102         if (!subctxt_cnt) {
 103                 shared = 0;
 104                 master = 0;
 105                 subctxt_cnt = 1;
 106         } else {
 107                 shared = 1;
 108                 master = !subctxt_fp(fp);
 109         }
 110
 111         sz = sizeof(*kinfo);
 112         /* If context sharing is not requested, allow the old size structure */
 113         if (!shared)
 114                 sz -= 7 * sizeof(u64);
 115         if (ubase_size < sz) {
 116                 ret = -EINVAL;
 117                 goto bail;
 118         }
 119
 120         kinfo = kzalloc(sizeof(*kinfo), GFP_KERNEL);
 121         if (kinfo == NULL) {
 122                 ret = -ENOMEM;
 123                 goto bail;
 124         }
 125
 126         ret = dd->f_get_base_info(rcd, kinfo);
 127         if (ret < 0)
 128                 goto bail;
 129
 130         kinfo->spi_rcvhdr_cnt = dd->rcvhdrcnt;
 131         kinfo->spi_rcvhdrent_size = dd->rcvhdrentsize;
 132         kinfo->spi_tidegrcnt = rcd->rcvegrcnt;
 133         kinfo->spi_rcv_egrbufsize = dd->rcvegrbufsize;
 134         /*
 135          * have to mmap whole thing
 136          */
 137         kinfo->spi_rcv_egrbuftotlen =
 138                 rcd->rcvegrbuf_chunks * rcd->rcvegrbuf_size;
 139         kinfo->spi_rcv_egrperchunk = rcd->rcvegrbufs_perchunk;
 140         kinfo->spi_rcv_egrchunksize = kinfo->spi_rcv_egrbuftotlen /
 141                 rcd->rcvegrbuf_chunks;
 142         kinfo->spi_tidcnt = dd->rcvtidcnt / subctxt_cnt;
 143         if (master)
 144                 kinfo->spi_tidcnt += dd->rcvtidcnt % subctxt_cnt;
 145         /*
 146          * for this use, may be cfgctxts summed over all chips that
 147          * are are configured and present
 148          */
 149         kinfo->spi_nctxts = dd->cfgctxts;
 150         /* unit (chip/board) our context is on */
 151         kinfo->spi_unit = dd->unit;
 152         kinfo->spi_port = ppd->port;
 153         /* for now, only a single page */
 154         kinfo->spi_tid_maxsize = PAGE_SIZE;
 155
 156         /*
 157          * Doing this per context, and based on the skip value, etc.  This has
 158          * to be the actual buffer size, since the protocol code treats it
 159          * as an array.
 160          *
 161          * These have to be set to user addresses in the user code via mmap.
 162          * These values are used on return to user code for the mmap target
 163          * addresses only.  For 32 bit, same 44 bit address problem, so use
 164          * the physical address, not virtual.  Before 2.6.11, using the
 165          * page_address() macro worked, but in 2.6.11, even that returns the
 166          * full 64 bit address (upper bits all 1's).  So far, using the
 167          * physical addresses (or chip offsets, for chip mapping) works, but
 168          * no doubt some future kernel release will change that, and we'll be
 169          * on to yet another method of dealing with this.
 170          * Normally only one of rcvhdr_tailaddr or rhf_offset is useful
 171          * since the chips with non-zero rhf_offset don't normally
 172          * enable tail register updates to host memory, but for testing,
 173          * both can be enabled and used.
 174          */
 175         kinfo->spi_rcvhdr_base = (u64) rcd->rcvhdrq_phys;
 176         kinfo->spi_rcvhdr_tailaddr = (u64) rcd->rcvhdrqtailaddr_phys;
 177         kinfo->spi_rhf_offset = dd->rhf_offset;
 178         kinfo->spi_rcv_egrbufs = (u64) rcd->rcvegr_phys;
 179         kinfo->spi_pioavailaddr = (u64) dd->pioavailregs_phys;
 180         /* setup per-unit (not port) status area for user programs */
 181         kinfo->spi_status = (u64) kinfo->spi_pioavailaddr +
 182                 (char *) ppd->statusp -
 183                 (char *) dd->pioavailregs_dma;
 184         kinfo->spi_uregbase = (u64) dd->uregbase + dd->ureg_align * rcd->ctxt;
 185         if (!shared) {
 186                 kinfo->spi_piocnt = rcd->piocnt;
 187                 kinfo->spi_piobufbase = (u64) rcd->piobufs;
 188                 kinfo->spi_sendbuf_status = cvt_kvaddr(rcd->user_event_mask);
 189         } else if (master) {
 190                 kinfo->spi_piocnt = (rcd->piocnt / subctxt_cnt) +
 191                                     (rcd->piocnt % subctxt_cnt);
 192                 /* Master's PIO buffers are after all the slave's */
 193                 kinfo->spi_piobufbase = (u64) rcd->piobufs +
 194                         dd->palign *
 195                         (rcd->piocnt - kinfo->spi_piocnt);
 196         } else {
 197                 unsigned slave = subctxt_fp(fp) - 1;
 198
 199                 kinfo->spi_piocnt = rcd->piocnt / subctxt_cnt;
 200                 kinfo->spi_piobufbase = (u64) rcd->piobufs +
 201                         dd->palign * kinfo->spi_piocnt * slave;
 202         }
 203
 204         if (shared) {
 205                 kinfo->spi_sendbuf_status =
 206                         cvt_kvaddr(&rcd->user_event_mask[subctxt_fp(fp)]);
 207                 /* only spi_subctxt_* fields should be set in this block! */
 208                 kinfo->spi_subctxt_uregbase = cvt_kvaddr(rcd->subctxt_uregbase);
 209
 210                 kinfo->spi_subctxt_rcvegrbuf =
 211                         cvt_kvaddr(rcd->subctxt_rcvegrbuf);
 212                 kinfo->spi_subctxt_rcvhdr_base =
 213                         cvt_kvaddr(rcd->subctxt_rcvhdr_base);
 214         }
 215
 216         /*
 217          * All user buffers are 2KB buffers.  If we ever support
 218          * giving 4KB buffers to user processes, this will need some
 219          * work.  Can't use piobufbase directly, because it has
 220          * both 2K and 4K buffer base values.
 221          */
 222         kinfo->spi_pioindex = (kinfo->spi_piobufbase - dd->pio2k_bufbase) /
 223                 dd->palign;
 224         kinfo->spi_pioalign = dd->palign;
 225         kinfo->spi_qpair = QIB_KD_QP;
 226         /*
 227          * user mode PIO buffers are always 2KB, even when 4KB can
 228          * be received, and sent via the kernel; this is ibmaxlen
 229          * for 2K MTU.
 230          */
 231         kinfo->spi_piosize = dd->piosize2k - 2 * sizeof(u32);
 232         kinfo->spi_mtu = ppd->ibmaxlen; /* maxlen, not ibmtu */
 233         kinfo->spi_ctxt = rcd->ctxt;
 234         kinfo->spi_subctxt = subctxt_fp(fp);
 235         kinfo->spi_sw_version = QIB_KERN_SWVERSION;
 236         kinfo->spi_sw_version |= 1U << 31; /* QLogic-built, not kernel.org */
 237         kinfo->spi_hw_version = dd->revision;
 238
 239         if (master)
 240                 kinfo->spi_runtime_flags |= QIB_RUNTIME_MASTER;
 241
 242         sz = (ubase_size < sizeof(*kinfo)) ? ubase_size : sizeof(*kinfo);
 243         if (copy_to_user(ubase, kinfo, sz))
 244                 ret = -EFAULT;
 245 bail:
 246         kfree(kinfo);
 247         return ret;
 248 }
 249
 250 /**
 251  * qib_tid_update - update a context TID
 252  * @rcd: the context
 253  * @fp: the qib device file
 254  * @ti: the TID information
 255  *
 256  * The new implementation as of Oct 2004 is that the driver assigns
 257  * the tid and returns it to the caller.   To reduce search time, we
 258  * keep a cursor for each context, walking the shadow tid array to find
 259  * one that's not in use.
 260  *
 261  * For now, if we can't allocate the full list, we fail, although
 262  * in the long run, we'll allocate as many as we can, and the
 263  * caller will deal with that by trying the remaining pages later.
 264  * That means that when we fail, we have to mark the tids as not in
 265  * use again, in our shadow copy.
 266  *
 267  * It's up to the caller to free the tids when they are done.
 268  * We'll unlock the pages as they free them.
 269  *
 270  * Also, right now we are locking one page at a time, but since
 271  * the intended use of this routine is for a single group of
 272  * virtually contiguous pages, that should change to improve
 273  * performance.
 274  */
 275 static int qib_tid_update(struct qib_ctxtdata *rcd, struct file *fp,
 276                           const struct qib_tid_info *ti)
 277 {
 278         int ret = 0, ntids;
 279         u32 tid, ctxttid, cnt, i, tidcnt, tidoff;
 280         u16 *tidlist;
 281         struct qib_devdata *dd = rcd->dd;
 282         u64 physaddr;
 283         unsigned long vaddr;
 284         u64 __iomem *tidbase;
 285         unsigned long tidmap[8];
 286         struct page **pagep = NULL;
 287         unsigned subctxt = subctxt_fp(fp);
 288
 289         if (!dd->pageshadow) {
 290                 ret = -ENOMEM;
 291                 goto done;
 292         }
 293
 294         cnt = ti->tidcnt;
 295         if (!cnt) {
 296                 ret = -EFAULT;
 297                 goto done;
 298         }
 299         ctxttid = rcd->ctxt * dd->rcvtidcnt;
 300         if (!rcd->subctxt_cnt) {
 301                 tidcnt = dd->rcvtidcnt;
 302                 tid = rcd->tidcursor;
 303                 tidoff = 0;
 304         } else if (!subctxt) {
 305                 tidcnt = (dd->rcvtidcnt / rcd->subctxt_cnt) +
 306                          (dd->rcvtidcnt % rcd->subctxt_cnt);
 307                 tidoff = dd->rcvtidcnt - tidcnt;
 308                 ctxttid += tidoff;
 309                 tid = tidcursor_fp(fp);
 310         } else {
 311                 tidcnt = dd->rcvtidcnt / rcd->subctxt_cnt;
 312                 tidoff = tidcnt * (subctxt - 1);
 313                 ctxttid += tidoff;
 314                 tid = tidcursor_fp(fp);
 315         }
 316         if (cnt > tidcnt) {
 317                 /* make sure it all fits in tid_pg_list */
 318                 qib_devinfo(dd->pcidev, "Process tried to allocate %u "
 319                          "TIDs, only trying max (%u)\n", cnt, tidcnt);
 320                 cnt = tidcnt;
 321         }
 322         pagep = (struct page **) rcd->tid_pg_list;
 323         tidlist = (u16 *) &pagep[dd->rcvtidcnt];
 324         pagep += tidoff;
 325         tidlist += tidoff;
 326
 327         memset(tidmap, 0, sizeof(tidmap));
 328         /* before decrement; chip actual # */
 329         ntids = tidcnt;
 330         tidbase = (u64 __iomem *) (((char __iomem *) dd->kregbase) +
 331                                    dd->rcvtidbase +
 332                                    ctxttid * sizeof(*tidbase));
 333
 334         /* virtual address of first page in transfer */
 335         vaddr = ti->tidvaddr;
 336         if (!access_ok(VERIFY_WRITE, (void __user *) vaddr,
 337                        cnt * PAGE_SIZE)) {
 338                 ret = -EFAULT;
 339                 goto done;
 340         }
 341         ret = qib_get_user_pages(vaddr, cnt, pagep);
 342         if (ret) {
 343                 /*
 344                  * if (ret == -EBUSY)
 345                  * We can't continue because the pagep array won't be
 346                  * initialized. This should never happen,
 347                  * unless perhaps the user has mpin'ed the pages
 348                  * themselves.
 349                  */
 350                 qib_devinfo(dd->pcidev,
 351                          "Failed to lock addr %p, %u pages: "
 352                          "errno %d\n", (void *) vaddr, cnt, -ret);
 353                 goto done;
 354         }
 355         for (i = 0; i < cnt; i++, vaddr += PAGE_SIZE) {
 356                 for (; ntids--; tid++) {
 357                         if (tid == tidcnt)
 358                                 tid = 0;
 359                         if (!dd->pageshadow[ctxttid + tid])
 360                                 break;
 361                 }
 362                 if (ntids < 0) {
 363                         /*
 364                          * Oops, wrapped all the way through their TIDs,
 365                          * and didn't have enough free; see comments at
 366                          * start of routine
 367                          */
 368                         i--;    /* last tidlist[i] not filled in */
 369                         ret = -ENOMEM;
 370                         break;
 371                 }
 372                 tidlist[i] = tid + tidoff;
 373                 /* we "know" system pages and TID pages are same size */
 374                 dd->pageshadow[ctxttid + tid] = pagep[i];
 375                 dd->physshadow[ctxttid + tid] =
 376                         qib_map_page(dd->pcidev, pagep[i], 0, PAGE_SIZE,
 377                                      PCI_DMA_FROMDEVICE);
 378                 /*
 379                  * don't need atomic or it's overhead
 380                  */
 381                 __set_bit(tid, tidmap);
 382                 physaddr = dd->physshadow[ctxttid + tid];
 383                 /* PERFORMANCE: below should almost certainly be cached */
 384                 dd->f_put_tid(dd, &tidbase[tid],
 385                                   RCVHQ_RCV_TYPE_EXPECTED, physaddr);
 386                 /*
 387                  * don't check this tid in qib_ctxtshadow, since we
 388                  * just filled it in; start with the next one.
 389                  */
 390                 tid++;
 391         }
 392
 393         if (ret) {
 394                 u32 limit;
 395 cleanup:
 396                 /* jump here if copy out of updated info failed... */
 397                 /* same code that's in qib_free_tid() */
 398                 limit = sizeof(tidmap) * BITS_PER_BYTE;
 399                 if (limit > tidcnt)
 400                         /* just in case size changes in future */
 401                         limit = tidcnt;
 402                 tid = find_first_bit((const unsigned long *)tidmap, limit);
 403                 for (; tid < limit; tid++) {
 404                         if (!test_bit(tid, tidmap))
 405                                 continue;
 406                         if (dd->pageshadow[ctxttid + tid]) {
 407                                 dma_addr_t phys;
 408
 409                                 phys = dd->physshadow[ctxttid + tid];
 410                                 dd->physshadow[ctxttid + tid] = dd->tidinvalid;
 411                                 /* PERFORMANCE: below should almost certainly
 412                                  * be cached
 413                                  */
 414                                 dd->f_put_tid(dd, &tidbase[tid],
 415                                               RCVHQ_RCV_TYPE_EXPECTED,
 416                                               dd->tidinvalid);
 417                                 pci_unmap_page(dd->pcidev, phys, PAGE_SIZE,
 418                                                PCI_DMA_FROMDEVICE);
 419                                 dd->pageshadow[ctxttid + tid] = NULL;
 420                         }
 421                 }
 422                 qib_release_user_pages(pagep, cnt);
 423         } else {
 424                 /*
 425                  * Copy the updated array, with qib_tid's filled in, back
 426                  * to user.  Since we did the copy in already, this "should
 427                  * never fail" If it does, we have to clean up...
 428                  */
 429                 if (copy_to_user((void __user *)
 430                                  (unsigned long) ti->tidlist,
 431                                  tidlist, cnt * sizeof(*tidlist))) {
 432                         ret = -EFAULT;
 433                         goto cleanup;
 434                 }
 435                 if (copy_to_user((void __user *) (unsigned long) ti->tidmap,
 436                                  tidmap, sizeof tidmap)) {
 437                         ret = -EFAULT;
 438                         goto cleanup;
 439                 }
 440                 if (tid == tidcnt)
 441                         tid = 0;
 442                 if (!rcd->subctxt_cnt)
 443                         rcd->tidcursor = tid;
 444                 else
 445                         tidcursor_fp(fp) = tid;
 446         }
 447
 448 done:
 449         return ret;
 450 }
 451
 452 /**
 453  * qib_tid_free - free a context TID
 454  * @rcd: the context
 455  * @subctxt: the subcontext
 456  * @ti: the TID info
 457  *
 458  * right now we are unlocking one page at a time, but since
 459  * the intended use of this routine is for a single group of
 460  * virtually contiguous pages, that should change to improve
 461  * performance.  We check that the TID is in range for this context
 462  * but otherwise don't check validity; if user has an error and
 463  * frees the wrong tid, it's only their own data that can thereby
 464  * be corrupted.  We do check that the TID was in use, for sanity
 465  * We always use our idea of the saved address, not the address that
 466  * they pass in to us.
 467  */
 468 static int qib_tid_free(struct qib_ctxtdata *rcd, unsigned subctxt,
 469                         const struct qib_tid_info *ti)
 470 {
 471         int ret = 0;
 472         u32 tid, ctxttid, cnt, limit, tidcnt;
 473         struct qib_devdata *dd = rcd->dd;
 474         u64 __iomem *tidbase;
 475         unsigned long tidmap[8];
 476
 477         if (!dd->pageshadow) {
 478                 ret = -ENOMEM;
 479                 goto done;
 480         }
 481
 482         if (copy_from_user(tidmap, (void __user *)(unsigned long)ti->tidmap,
 483                            sizeof tidmap)) {
 484                 ret = -EFAULT;
 485                 goto done;
 486         }
 487
 488         ctxttid = rcd->ctxt * dd->rcvtidcnt;
 489         if (!rcd->subctxt_cnt)
 490                 tidcnt = dd->rcvtidcnt;
 491         else if (!subctxt) {
 492                 tidcnt = (dd->rcvtidcnt / rcd->subctxt_cnt) +
 493                          (dd->rcvtidcnt % rcd->subctxt_cnt);
 494                 ctxttid += dd->rcvtidcnt - tidcnt;
 495         } else {
 496                 tidcnt = dd->rcvtidcnt / rcd->subctxt_cnt;
 497                 ctxttid += tidcnt * (subctxt - 1);
 498         }
 499         tidbase = (u64 __iomem *) ((char __iomem *)(dd->kregbase) +
 500                                    dd->rcvtidbase +
 501                                    ctxttid * sizeof(*tidbase));
 502
 503         limit = sizeof(tidmap) * BITS_PER_BYTE;
 504         if (limit > tidcnt)
 505                 /* just in case size changes in future */
 506                 limit = tidcnt;
 507         tid = find_first_bit(tidmap, limit);
 508         for (cnt = 0; tid < limit; tid++) {
 509                 /*
 510                  * small optimization; if we detect a run of 3 or so without
 511                  * any set, use find_first_bit again.  That's mainly to
 512                  * accelerate the case where we wrapped, so we have some at
 513                  * the beginning, and some at the end, and a big gap
 514                  * in the middle.
 515                  */
 516                 if (!test_bit(tid, tidmap))
 517                         continue;
 518                 cnt++;
 519                 if (dd->pageshadow[ctxttid + tid]) {
 520                         struct page *p;
 521                         dma_addr_t phys;
 522
 523                         p = dd->pageshadow[ctxttid + tid];
 524                         dd->pageshadow[ctxttid + tid] = NULL;
 525                         phys = dd->physshadow[ctxttid + tid];
 526                         dd->physshadow[ctxttid + tid] = dd->tidinvalid;
 527                         /* PERFORMANCE: below should almost certainly be
 528                          * cached
 529                          */
 530                         dd->f_put_tid(dd, &tidbase[tid],
 531                                       RCVHQ_RCV_TYPE_EXPECTED, dd->tidinvalid);
 532                         pci_unmap_page(dd->pcidev, phys, PAGE_SIZE,
 533                                        PCI_DMA_FROMDEVICE);
 534                         qib_release_user_pages(&p, 1);
 535                 }
 536         }
 537 done:
 538         return ret;
 539 }
 540
 541 /**
 542  * qib_set_part_key - set a partition key
 543  * @rcd: the context
 544  * @key: the key
 545  *
 546  * We can have up to 4 active at a time (other than the default, which is
 547  * always allowed).  This is somewhat tricky, since multiple contexts may set
 548  * the same key, so we reference count them, and clean up at exit.  All 4
 549  * partition keys are packed into a single qlogic_ib register.  It's an
 550  * error for a process to set the same pkey multiple times.  We provide no
 551  * mechanism to de-allocate a pkey at this time, we may eventually need to
 552  * do that.  I've used the atomic operations, and no locking, and only make
 553  * a single pass through what's available.  This should be more than
 554  * adequate for some time. I'll think about spinlocks or the like if and as
 555  * it's necessary.
 556  */
 557 static int qib_set_part_key(struct qib_ctxtdata *rcd, u16 key)
 558 {
 559         struct qib_pportdata *ppd = rcd->ppd;
 560         int i, any = 0, pidx = -1;
 561         u16 lkey = key & 0x7FFF;
 562         int ret;
 563
 564         if (lkey == (QIB_DEFAULT_P_KEY & 0x7FFF)) {
 565                 /* nothing to do; this key always valid */
 566                 ret = 0;
 567                 goto bail;
 568         }
 569
 570         if (!lkey) {
 571                 ret = -EINVAL;
 572                 goto bail;
 573         }
 574
 575         /*
 576          * Set the full membership bit, because it has to be
 577          * set in the register or the packet, and it seems
 578          * cleaner to set in the register than to force all
 579          * callers to set it.
 580          */
 581         key |= 0x8000;
 582
 583         for (i = 0; i < ARRAY_SIZE(rcd->pkeys); i++) {
 584                 if (!rcd->pkeys[i] && pidx == -1)
 585                         pidx = i;
 586                 if (rcd->pkeys[i] == key) {
 587                         ret = -EEXIST;
 588                         goto bail;
 589                 }
 590         }
 591         if (pidx == -1) {
 592                 ret = -EBUSY;
 593                 goto bail;
 594         }
 595         for (any = i = 0; i < ARRAY_SIZE(ppd->pkeys); i++) {
 596                 if (!ppd->pkeys[i]) {
 597                         any++;
 598                         continue;
 599                 }
 600                 if (ppd->pkeys[i] == key) {
 601                         atomic_t *pkrefs = &ppd->pkeyrefs[i];
 602
 603                         if (atomic_inc_return(pkrefs) > 1) {
 604                                 rcd->pkeys[pidx] = key;
 605                                 ret = 0;
 606                                 goto bail;
 607                         } else {
 608                                 /*
 609                                  * lost race, decrement count, catch below
 610                                  */
 611                                 atomic_dec(pkrefs);
 612                                 any++;
 613                         }
 614                 }
 615                 if ((ppd->pkeys[i] & 0x7FFF) == lkey) {
 616                         /*
 617                          * It makes no sense to have both the limited and
 618                          * full membership PKEY set at the same time since
 619                          * the unlimited one will disable the limited one.
 620                          */
 621                         ret = -EEXIST;
 622                         goto bail;
 623                 }
 624         }
 625         if (!any) {
 626                 ret = -EBUSY;
 627                 goto bail;
 628         }
 629         for (any = i = 0; i < ARRAY_SIZE(ppd->pkeys); i++) {
 630                 if (!ppd->pkeys[i] &&
 631                     atomic_inc_return(&ppd->pkeyrefs[i]) == 1) {
 632                         rcd->pkeys[pidx] = key;
 633                         ppd->pkeys[i] = key;
 634                         (void) ppd->dd->f_set_ib_cfg(ppd, QIB_IB_CFG_PKEYS, 0);
 635                         ret = 0;
 636                         goto bail;
 637                 }
 638         }
 639         ret = -EBUSY;
 640
 641 bail:
 642         return ret;
 643 }
 644
 645 /**
 646  * qib_manage_rcvq - manage a context's receive queue
 647  * @rcd: the context
 648  * @subctxt: the subcontext
 649  * @start_stop: action to carry out
 650  *
 651  * start_stop == 0 disables receive on the context, for use in queue
 652  * overflow conditions.  start_stop==1 re-enables, to be used to
 653  * re-init the software copy of the head register
 654  */
 655 static int qib_manage_rcvq(struct qib_ctxtdata *rcd, unsigned subctxt,
 656                            int start_stop)
 657 {
 658         struct qib_devdata *dd = rcd->dd;
 659         unsigned int rcvctrl_op;
 660
 661         if (subctxt)
 662                 goto bail;
 663         /* atomically clear receive enable ctxt. */
 664         if (start_stop) {
 665                 /*
 666                  * On enable, force in-memory copy of the tail register to
 667                  * 0, so that protocol code doesn't have to worry about
 668                  * whether or not the chip has yet updated the in-memory
 669                  * copy or not on return from the system call. The chip
 670                  * always resets it's tail register back to 0 on a
 671                  * transition from disabled to enabled.
 672                  */
 673                 if (rcd->rcvhdrtail_kvaddr)
 674                         qib_clear_rcvhdrtail(rcd);
 675                 rcvctrl_op = QIB_RCVCTRL_CTXT_ENB;
 676         } else
 677                 rcvctrl_op = QIB_RCVCTRL_CTXT_DIS;
 678         dd->f_rcvctrl(rcd->ppd, rcvctrl_op, rcd->ctxt);
 679         /* always; new head should be equal to new tail; see above */
 680 bail:
 681         return 0;
 682 }
 683
 684 static void qib_clean_part_key(struct qib_ctxtdata *rcd,
 685                                struct qib_devdata *dd)
 686 {
 687         int i, j, pchanged = 0;
 688         u64 oldpkey;
 689         struct qib_pportdata *ppd = rcd->ppd;
 690
 691         /* for debugging only */
 692         oldpkey = (u64) ppd->pkeys[0] |
 693                 ((u64) ppd->pkeys[1] << 16) |
 694                 ((u64) ppd->pkeys[2] << 32) |
 695                 ((u64) ppd->pkeys[3] << 48);
 696
 697         for (i = 0; i < ARRAY_SIZE(rcd->pkeys); i++) {
 698                 if (!rcd->pkeys[i])
 699                         continue;
 700                 for (j = 0; j < ARRAY_SIZE(ppd->pkeys); j++) {
 701                         /* check for match independent of the global bit */
 702                         if ((ppd->pkeys[j] & 0x7fff) !=
 703                             (rcd->pkeys[i] & 0x7fff))
 704                                 continue;
 705                         if (atomic_dec_and_test(&ppd->pkeyrefs[j])) {
 706                                 ppd->pkeys[j] = 0;
 707                                 pchanged++;
 708                         }
 709                         break;
 710                 }
 711                 rcd->pkeys[i] = 0;
 712         }
 713         if (pchanged)
 714                 (void) ppd->dd->f_set_ib_cfg(ppd, QIB_IB_CFG_PKEYS, 0);
 715 }
 716
 717 /* common code for the mappings on dma_alloc_coherent mem */
 718 static int qib_mmap_mem(struct vm_area_struct *vma, struct qib_ctxtdata *rcd,
 719                         unsigned len, void *kvaddr, u32 write_ok, char *what)
 720 {
 721         struct qib_devdata *dd = rcd->dd;
 722         unsigned long pfn;
 723         int ret;
 724
 725         if ((vma->vm_end - vma->vm_start) > len) {
 726                 qib_devinfo(dd->pcidev,
 727                          "FAIL on %s: len %lx > %x\n", what,
 728                          vma->vm_end - vma->vm_start, len);
 729                 ret = -EFAULT;
 730                 goto bail;
 731         }
 732
 733         /*
 734          * shared context user code requires rcvhdrq mapped r/w, others
 735          * only allowed readonly mapping.
 736          */
 737         if (!write_ok) {
 738                 if (vma->vm_flags & VM_WRITE) {
 739                         qib_devinfo(dd->pcidev,
 740                                  "%s must be mapped readonly\n", what);
 741                         ret = -EPERM;
 742                         goto bail;
 743                 }
 744
 745                 /* don't allow them to later change with mprotect */
 746                 vma->vm_flags &= ~VM_MAYWRITE;
 747         }
 748
 749         pfn = virt_to_phys(kvaddr) >> PAGE_SHIFT;
 750         ret = remap_pfn_range(vma, vma->vm_start, pfn,
 751                               len, vma->vm_page_prot);
 752         if (ret)
 753                 qib_devinfo(dd->pcidev, "%s ctxt%u mmap of %lx, %x "
 754                          "bytes failed: %d\n", what, rcd->ctxt,
 755                          pfn, len, ret);
 756 bail:
 757         return ret;
 758 }
 759
 760 static int mmap_ureg(struct vm_area_struct *vma, struct qib_devdata *dd,
 761                      u64 ureg)
 762 {
 763         unsigned long phys;
 764         unsigned long sz;
 765         int ret;
 766
 767         /*
 768          * This is real hardware, so use io_remap.  This is the mechanism
 769          * for the user process to update the head registers for their ctxt
 770          * in the chip.
 771          */
 772         sz = dd->flags & QIB_HAS_HDRSUPP ? 2 * PAGE_SIZE : PAGE_SIZE;
 773         if ((vma->vm_end - vma->vm_start) > sz) {
 774                 qib_devinfo(dd->pcidev, "FAIL mmap userreg: reqlen "
 775                          "%lx > PAGE\n", vma->vm_end - vma->vm_start);
 776                 ret = -EFAULT;
 777         } else {
 778                 phys = dd->physaddr + ureg;
 779                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 780
 781                 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
 782                 ret = io_remap_pfn_range(vma, vma->vm_start,
 783                                          phys >> PAGE_SHIFT,
 784                                          vma->vm_end - vma->vm_start,
 785                                          vma->vm_page_prot);
 786         }
 787         return ret;
 788 }
 789
 790 static int mmap_piobufs(struct vm_area_struct *vma,
 791                         struct qib_devdata *dd,
 792                         struct qib_ctxtdata *rcd,
 793                         unsigned piobufs, unsigned piocnt)
 794 {
 795         unsigned long phys;
 796         int ret;
 797
 798         /*
 799          * When we map the PIO buffers in the chip, we want to map them as
 800          * writeonly, no read possible; unfortunately, x86 doesn't allow
 801          * for this in hardware, but we still prevent users from asking
 802          * for it.
 803          */
 804         if ((vma->vm_end - vma->vm_start) > (piocnt * dd->palign)) {
 805                 qib_devinfo(dd->pcidev, "FAIL mmap piobufs: "
 806                          "reqlen %lx > PAGE\n",
 807                          vma->vm_end - vma->vm_start);
 808                 ret = -EINVAL;
 809                 goto bail;
 810         }
 811
 812         phys = dd->physaddr + piobufs;
 813
 814 #if defined(__powerpc__)
 815         /* There isn't a generic way to specify writethrough mappings */
 816         pgprot_val(vma->vm_page_prot) |= _PAGE_NO_CACHE;
 817         pgprot_val(vma->vm_page_prot) |= _PAGE_WRITETHRU;
 818         pgprot_val(vma->vm_page_prot) &= ~_PAGE_GUARDED;
 819 #endif
 820
 821         /*
 822          * don't allow them to later change to readable with mprotect (for when
 823          * not initially mapped readable, as is normally the case)
 824          */
 825         vma->vm_flags &= ~VM_MAYREAD;
 826         vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
 827
 828         if (qib_wc_pat)
 829                 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
 830
 831         ret = io_remap_pfn_range(vma, vma->vm_start, phys >> PAGE_SHIFT,
 832                                  vma->vm_end - vma->vm_start,
 833                                  vma->vm_page_prot);
 834 bail:
 835         return ret;
 836 }
 837
 838 static int mmap_rcvegrbufs(struct vm_area_struct *vma,
 839                            struct qib_ctxtdata *rcd)
 840 {
 841         struct qib_devdata *dd = rcd->dd;
 842         unsigned long start, size;
 843         size_t total_size, i;
 844         unsigned long pfn;
 845         int ret;
 846
 847         size = rcd->rcvegrbuf_size;
 848         total_size = rcd->rcvegrbuf_chunks * size;
 849         if ((vma->vm_end - vma->vm_start) > total_size) {
 850                 qib_devinfo(dd->pcidev, "FAIL on egr bufs: "
 851                          "reqlen %lx > actual %lx\n",
 852                          vma->vm_end - vma->vm_start,
 853                          (unsigned long) total_size);
 854                 ret = -EINVAL;
 855                 goto bail;
 856         }
 857
 858         if (vma->vm_flags & VM_WRITE) {
 859                 qib_devinfo(dd->pcidev, "Can't map eager buffers as "
 860                          "writable (flags=%lx)\n", vma->vm_flags);
 861                 ret = -EPERM;
 862                 goto bail;
 863         }
 864         /* don't allow them to later change to writeable with mprotect */
 865         vma->vm_flags &= ~VM_MAYWRITE;
 866
 867         start = vma->vm_start;
 868
 869         for (i = 0; i < rcd->rcvegrbuf_chunks; i++, start += size) {
 870                 pfn = virt_to_phys(rcd->rcvegrbuf[i]) >> PAGE_SHIFT;
 871                 ret = remap_pfn_range(vma, start, pfn, size,
 872                                       vma->vm_page_prot);
 873                 if (ret < 0)
 874                         goto bail;
 875         }
 876         ret = 0;
 877
 878 bail:
 879         return ret;
 880 }
 881
 882 /*
 883  * qib_file_vma_fault - handle a VMA page fault.
 884  */
 885 static int qib_file_vma_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 886 {
 887         struct page *page;
 888
 889         page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
 890         if (!page)
 891                 return VM_FAULT_SIGBUS;
 892
 893         get_page(page);
 894         vmf->page = page;
 895
 896         return 0;
 897 }
 898
 899 static struct vm_operations_struct qib_file_vm_ops = {
 900         .fault = qib_file_vma_fault,
 901 };
 902
 903 static int mmap_kvaddr(struct vm_area_struct *vma, u64 pgaddr,
 904                        struct qib_ctxtdata *rcd, unsigned subctxt)
 905 {
 906         struct qib_devdata *dd = rcd->dd;
 907         unsigned subctxt_cnt;
 908         unsigned long len;
 909         void *addr;
 910         size_t size;
 911         int ret = 0;
 912
 913         subctxt_cnt = rcd->subctxt_cnt;
 914         size = rcd->rcvegrbuf_chunks * rcd->rcvegrbuf_size;
 915
 916         /*
 917          * Each process has all the subctxt uregbase, rcvhdrq, and
 918          * rcvegrbufs mmapped - as an array for all the processes,
 919          * and also separately for this process.
 920          */
 921         if (pgaddr == cvt_kvaddr(rcd->subctxt_uregbase)) {
 922                 addr = rcd->subctxt_uregbase;
 923                 size = PAGE_SIZE * subctxt_cnt;
 924         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_rcvhdr_base)) {
 925                 addr = rcd->subctxt_rcvhdr_base;
 926                 size = rcd->rcvhdrq_size * subctxt_cnt;
 927         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_rcvegrbuf)) {
 928                 addr = rcd->subctxt_rcvegrbuf;
 929                 size *= subctxt_cnt;
 930         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_uregbase +
 931                                         PAGE_SIZE * subctxt)) {
 932                 addr = rcd->subctxt_uregbase + PAGE_SIZE * subctxt;
 933                 size = PAGE_SIZE;
 934         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_rcvhdr_base +
 935                                         rcd->rcvhdrq_size * subctxt)) {
 936                 addr = rcd->subctxt_rcvhdr_base +
 937                         rcd->rcvhdrq_size * subctxt;
 938                 size = rcd->rcvhdrq_size;
 939         } else if (pgaddr == cvt_kvaddr(&rcd->user_event_mask[subctxt])) {
 940                 addr = rcd->user_event_mask;
 941                 size = PAGE_SIZE;
 942         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_rcvegrbuf +
 943                                         size * subctxt)) {
 944                 addr = rcd->subctxt_rcvegrbuf + size * subctxt;
 945                 /* rcvegrbufs are read-only on the slave */
 946                 if (vma->vm_flags & VM_WRITE) {
 947                         qib_devinfo(dd->pcidev,
 948                                  "Can't map eager buffers as "
 949                                  "writable (flags=%lx)\n", vma->vm_flags);
 950                         ret = -EPERM;
 951                         goto bail;
 952                 }
 953                 /*
 954                  * Don't allow permission to later change to writeable
 955                  * with mprotect.
 956                  */
 957                 vma->vm_flags &= ~VM_MAYWRITE;
 958         } else
 959                 goto bail;
 960         len = vma->vm_end - vma->vm_start;
 961         if (len > size) {
 962                 ret = -EINVAL;
 963                 goto bail;
 964         }
 965
 966         vma->vm_pgoff = (unsigned long) addr >> PAGE_SHIFT;
 967         vma->vm_ops = &qib_file_vm_ops;
 968         vma->vm_flags |= VM_RESERVED | VM_DONTEXPAND;
 969         ret = 1;
 970
 971 bail:
 972         return ret;
 973 }
 974
 975 /**
 976  * qib_mmapf - mmap various structures into user space
 977  * @fp: the file pointer
 978  * @vma: the VM area
 979  *
 980  * We use this to have a shared buffer between the kernel and the user code
 981  * for the rcvhdr queue, egr buffers, and the per-context user regs and pio
 982  * buffers in the chip.  We have the open and close entries so we can bump
 983  * the ref count and keep the driver from being unloaded while still mapped.
 984  */
 985 static int qib_mmapf(struct file *fp, struct vm_area_struct *vma)
 986 {
 987         struct qib_ctxtdata *rcd;
 988         struct qib_devdata *dd;
 989         u64 pgaddr, ureg;
 990         unsigned piobufs, piocnt;
 991         int ret, match = 1;
 992
 993         rcd = ctxt_fp(fp);
 994         if (!rcd || !(vma->vm_flags & VM_SHARED)) {
 995                 ret = -EINVAL;
 996                 goto bail;
 997         }
 998         dd = rcd->dd;
 999
1000         /*
1001          * This is the qib_do_user_init() code, mapping the shared buffers
1002          * and per-context user registers into the user process. The address
1003          * referred to by vm_pgoff is the file offset passed via mmap().
1004          * For shared contexts, this is the kernel vmalloc() address of the
1005          * pages to share with the master.
1006          * For non-shared or master ctxts, this is a physical address.
1007          * We only do one mmap for each space mapped.
1008          */
1009         pgaddr = vma->vm_pgoff << PAGE_SHIFT;
1010
1011         /*
1012          * Check for 0 in case one of the allocations failed, but user
1013          * called mmap anyway.
1014          */
1015         if (!pgaddr)  {
1016                 ret = -EINVAL;
1017                 goto bail;
1018         }
1019
1020         /*
1021          * Physical addresses must fit in 40 bits for our hardware.
1022          * Check for kernel virtual addresses first, anything else must
1023          * match a HW or memory address.
1024          */
1025         ret = mmap_kvaddr(vma, pgaddr, rcd, subctxt_fp(fp));
1026         if (ret) {
1027                 if (ret > 0)
1028                         ret = 0;
1029                 goto bail;
1030         }
1031
1032         ureg = dd->uregbase + dd->ureg_align * rcd->ctxt;
1033         if (!rcd->subctxt_cnt) {
1034                 /* ctxt is not shared */
1035                 piocnt = rcd->piocnt;
1036                 piobufs = rcd->piobufs;
1037         } else if (!subctxt_fp(fp)) {
1038                 /* caller is the master */
1039                 piocnt = (rcd->piocnt / rcd->subctxt_cnt) +
1040                          (rcd->piocnt % rcd->subctxt_cnt);
1041                 piobufs = rcd->piobufs +
1042                         dd->palign * (rcd->piocnt - piocnt);
1043         } else {
1044                 unsigned slave = subctxt_fp(fp) - 1;
1045
1046                 /* caller is a slave */
1047                 piocnt = rcd->piocnt / rcd->subctxt_cnt;
1048                 piobufs = rcd->piobufs + dd->palign * piocnt * slave;
1049         }
1050
1051         if (pgaddr == ureg)
1052                 ret = mmap_ureg(vma, dd, ureg);
1053         else if (pgaddr == piobufs)
1054                 ret = mmap_piobufs(vma, dd, rcd, piobufs, piocnt);
1055         else if (pgaddr == dd->pioavailregs_phys)
1056                 /* in-memory copy of pioavail registers */
1057                 ret = qib_mmap_mem(vma, rcd, PAGE_SIZE,
1058                                    (void *) dd->pioavailregs_dma, 0,
1059                                    "pioavail registers");
1060         else if (pgaddr == rcd->rcvegr_phys)
1061                 ret = mmap_rcvegrbufs(vma, rcd);
1062         else if (pgaddr == (u64) rcd->rcvhdrq_phys)
1063                 /*
1064                  * The rcvhdrq itself; multiple pages, contiguous
1065                  * from an i/o perspective.  Shared contexts need
1066                  * to map r/w, so we allow writing.
1067                  */
1068                 ret = qib_mmap_mem(vma, rcd, rcd->rcvhdrq_size,
1069                                    rcd->rcvhdrq, 1, "rcvhdrq");
1070         else if (pgaddr == (u64) rcd->rcvhdrqtailaddr_phys)
1071                 /* in-memory copy of rcvhdrq tail register */
1072                 ret = qib_mmap_mem(vma, rcd, PAGE_SIZE,
1073                                    rcd->rcvhdrtail_kvaddr, 0,
1074                                    "rcvhdrq tail");
1075         else
1076                 match = 0;
1077         if (!match)
1078                 ret = -EINVAL;
1079
1080         vma->vm_private_data = NULL;
1081
1082         if (ret < 0)
1083                 qib_devinfo(dd->pcidev,
1084                          "mmap Failure %d: off %llx len %lx\n",
1085                          -ret, (unsigned long long)pgaddr,
1086                          vma->vm_end - vma->vm_start);
1087 bail:
1088         return ret;
1089 }
1090
1091 static unsigned int qib_poll_urgent(struct qib_ctxtdata *rcd,
1092                                     struct file *fp,
1093                                     struct poll_table_struct *pt)
1094 {
1095         struct qib_devdata *dd = rcd->dd;
1096         unsigned pollflag;
1097
1098         poll_wait(fp, &rcd->wait, pt);
1099
1100         spin_lock_irq(&dd->uctxt_lock);
1101         if (rcd->urgent != rcd->urgent_poll) {
1102                 pollflag = POLLIN | POLLRDNORM;
1103                 rcd->urgent_poll = rcd->urgent;
1104         } else {
1105                 pollflag = 0;
1106                 set_bit(QIB_CTXT_WAITING_URG, &rcd->flag);
1107         }
1108         spin_unlock_irq(&dd->uctxt_lock);
1109
1110         return pollflag;
1111 }
1112
1113 static unsigned int qib_poll_next(struct qib_ctxtdata *rcd,
1114                                   struct file *fp,
1115                                   struct poll_table_struct *pt)
1116 {
1117         struct qib_devdata *dd = rcd->dd;
1118         unsigned pollflag;
1119
1120         poll_wait(fp, &rcd->wait, pt);
1121
1122         spin_lock_irq(&dd->uctxt_lock);
1123         if (dd->f_hdrqempty(rcd)) {
1124                 set_bit(QIB_CTXT_WAITING_RCV, &rcd->flag);
1125                 dd->f_rcvctrl(rcd->ppd, QIB_RCVCTRL_INTRAVAIL_ENB, rcd->ctxt);
1126                 pollflag = 0;
1127         } else
1128                 pollflag = POLLIN | POLLRDNORM;
1129         spin_unlock_irq(&dd->uctxt_lock);
1130
1131         return pollflag;
1132 }
1133
1134 static unsigned int qib_poll(struct file *fp, struct poll_table_struct *pt)
1135 {
1136         struct qib_ctxtdata *rcd;
1137         unsigned pollflag;
1138
1139         rcd = ctxt_fp(fp);
1140         if (!rcd)
1141                 pollflag = POLLERR;
1142         else if (rcd->poll_type == QIB_POLL_TYPE_URGENT)
1143                 pollflag = qib_poll_urgent(rcd, fp, pt);
1144         else  if (rcd->poll_type == QIB_POLL_TYPE_ANYRCV)
1145                 pollflag = qib_poll_next(rcd, fp, pt);
1146         else /* invalid */
1147                 pollflag = POLLERR;
1148
1149         return pollflag;
1150 }
1151
1152 /*
1153  * Check that userland and driver are compatible for subcontexts.
1154  */
1155 static int qib_compatible_subctxts(int user_swmajor, int user_swminor)
1156 {
1157         /* this code is written long-hand for clarity */
1158         if (QIB_USER_SWMAJOR != user_swmajor) {
1159                 /* no promise of compatibility if major mismatch */
1160                 return 0;
1161         }
1162         if (QIB_USER_SWMAJOR == 1) {
1163                 switch (QIB_USER_SWMINOR) {
1164                 case 0:
1165                 case 1:
1166                 case 2:
1167                         /* no subctxt implementation so cannot be compatible */
1168                         return 0;
1169                 case 3:
1170                         /* 3 is only compatible with itself */
1171                         return user_swminor == 3;
1172                 default:
1173                         /* >= 4 are compatible (or are expected to be) */
1174                         return user_swminor >= 4;
1175                 }
1176         }
1177         /* make no promises yet for future major versions */
1178         return 0;
1179 }
1180
1181 static int init_subctxts(struct qib_devdata *dd,
1182                          struct qib_ctxtdata *rcd,
1183                          const struct qib_user_info *uinfo)
1184 {
1185         int ret = 0;
1186         unsigned num_subctxts;
1187         size_t size;
1188
1189         /*
1190          * If the user is requesting zero subctxts,
1191          * skip the subctxt allocation.
1192          */
1193         if (uinfo->spu_subctxt_cnt <= 0)
1194                 goto bail;
1195         num_subctxts = uinfo->spu_subctxt_cnt;
1196
1197         /* Check for subctxt compatibility */
1198         if (!qib_compatible_subctxts(uinfo->spu_userversion >> 16,
1199                 uinfo->spu_userversion & 0xffff)) {
1200                 qib_devinfo(dd->pcidev,
1201                          "Mismatched user version (%d.%d) and driver "
1202                          "version (%d.%d) while context sharing. Ensure "
1203                          "that driver and library are from the same "
1204                          "release.\n",
1205                          (int) (uinfo->spu_userversion >> 16),
1206                          (int) (uinfo->spu_userversion & 0xffff),
1207                          QIB_USER_SWMAJOR, QIB_USER_SWMINOR);
1208                 goto bail;
1209         }
1210         if (num_subctxts > QLOGIC_IB_MAX_SUBCTXT) {
1211                 ret = -EINVAL;
1212                 goto bail;
1213         }
1214
1215         rcd->subctxt_uregbase = vmalloc_user(PAGE_SIZE * num_subctxts);
1216         if (!rcd->subctxt_uregbase) {
1217                 ret = -ENOMEM;
1218                 goto bail;
1219         }
1220         /* Note: rcd->rcvhdrq_size isn't initialized yet. */
1221         size = ALIGN(dd->rcvhdrcnt * dd->rcvhdrentsize *
1222                      sizeof(u32), PAGE_SIZE) * num_subctxts;
1223         rcd->subctxt_rcvhdr_base = vmalloc_user(size);
1224         if (!rcd->subctxt_rcvhdr_base) {
1225                 ret = -ENOMEM;
1226                 goto bail_ureg;
1227         }
1228
1229         rcd->subctxt_rcvegrbuf = vmalloc_user(rcd->rcvegrbuf_chunks *
1230                                               rcd->rcvegrbuf_size *
1231                                               num_subctxts);
1232         if (!rcd->subctxt_rcvegrbuf) {
1233                 ret = -ENOMEM;
1234                 goto bail_rhdr;
1235         }
1236
1237         rcd->subctxt_cnt = uinfo->spu_subctxt_cnt;
1238         rcd->subctxt_id = uinfo->spu_subctxt_id;
1239         rcd->active_slaves = 1;
1240         rcd->redirect_seq_cnt = 1;
1241         set_bit(QIB_CTXT_MASTER_UNINIT, &rcd->flag);
1242         goto bail;
1243
1244 bail_rhdr:
1245         vfree(rcd->subctxt_rcvhdr_base);
1246 bail_ureg:
1247         vfree(rcd->subctxt_uregbase);
1248         rcd->subctxt_uregbase = NULL;
1249 bail:
1250         return ret;
1251 }
1252
1253 static int setup_ctxt(struct qib_pportdata *ppd, int ctxt,
1254                       struct file *fp, const struct qib_user_info *uinfo)
1255 {
1256         struct qib_devdata *dd = ppd->dd;
1257         struct qib_ctxtdata *rcd;
1258         void *ptmp = NULL;
1259         int ret;
1260
1261         rcd = qib_create_ctxtdata(ppd, ctxt);
1262
1263         /*
1264          * Allocate memory for use in qib_tid_update() at open to
1265          * reduce cost of expected send setup per message segment
1266          */
1267         if (rcd)
1268                 ptmp = kmalloc(dd->rcvtidcnt * sizeof(u16) +
1269                                dd->rcvtidcnt * sizeof(struct page **),
1270                                GFP_KERNEL);
1271
1272         if (!rcd || !ptmp) {
1273                 qib_dev_err(dd, "Unable to allocate ctxtdata "
1274                             "memory, failing open\n");
1275                 ret = -ENOMEM;
1276                 goto bailerr;
1277         }
1278         rcd->userversion = uinfo->spu_userversion;
1279         ret = init_subctxts(dd, rcd, uinfo);
1280         if (ret)
1281                 goto bailerr;
1282         rcd->tid_pg_list = ptmp;
1283         rcd->pid = current->pid;
1284         init_waitqueue_head(&dd->rcd[ctxt]->wait);
1285         strlcpy(rcd->comm, current->comm, sizeof(rcd->comm));
1286         ctxt_fp(fp) = rcd;
1287         qib_stats.sps_ctxts++;
1288         dd->freectxts--;
1289         ret = 0;
1290         goto bail;
1291
1292 bailerr:
1293         dd->rcd[ctxt] = NULL;
1294         kfree(rcd);
1295         kfree(ptmp);
1296 bail:
1297         return ret;
1298 }
1299
1300 static inline int usable(struct qib_pportdata *ppd)
1301 {
1302         struct qib_devdata *dd = ppd->dd;
1303
1304         return dd && (dd->flags & QIB_PRESENT) && dd->kregbase && ppd->lid &&
1305                 (ppd->lflags & QIBL_LINKACTIVE);
1306 }
1307
1308 /*
1309  * Select a context on the given device, either using a requested port
1310  * or the port based on the context number.
1311  */
1312 static int choose_port_ctxt(struct file *fp, struct qib_devdata *dd, u32 port,
1313                             const struct qib_user_info *uinfo)
1314 {
1315         struct qib_pportdata *ppd = NULL;
1316         int ret, ctxt;
1317
1318         if (port) {
1319                 if (!usable(dd->pport + port - 1)) {
1320                         ret = -ENETDOWN;
1321                         goto done;
1322                 } else
1323                         ppd = dd->pport + port - 1;
1324         }
1325         for (ctxt = dd->first_user_ctxt; ctxt < dd->cfgctxts && dd->rcd[ctxt];
1326              ctxt++)
1327                 ;
1328         if (ctxt == dd->cfgctxts) {
1329                 ret = -EBUSY;
1330                 goto done;
1331         }
1332         if (!ppd) {
1333                 u32 pidx = ctxt % dd->num_pports;
1334                 if (usable(dd->pport + pidx))
1335                         ppd = dd->pport + pidx;
1336                 else {
1337                         for (pidx = 0; pidx < dd->num_pports && !ppd;
1338                              pidx++)
1339                                 if (usable(dd->pport + pidx))
1340                                         ppd = dd->pport + pidx;
1341                 }
1342         }
1343         ret = ppd ? setup_ctxt(ppd, ctxt, fp, uinfo) : -ENETDOWN;
1344 done:
1345         return ret;
1346 }
1347
1348 static int find_free_ctxt(int unit, struct file *fp,
1349                           const struct qib_user_info *uinfo)
1350 {
1351         struct qib_devdata *dd = qib_lookup(unit);
1352         int ret;
1353
1354         if (!dd || (uinfo->spu_port && uinfo->spu_port > dd->num_pports))
1355                 ret = -ENODEV;
1356         else
1357                 ret = choose_port_ctxt(fp, dd, uinfo->spu_port, uinfo);
1358
1359         return ret;
1360 }
1361
1362 static int get_a_ctxt(struct file *fp, const struct qib_user_info *uinfo,
1363                       unsigned alg)
1364 {
1365         struct qib_devdata *udd = NULL;
1366         int ret = 0, devmax, npresent, nup, ndev, dusable = 0, i;
1367         u32 port = uinfo->spu_port, ctxt;
1368
1369         devmax = qib_count_units(&npresent, &nup);
1370         if (!npresent) {
1371                 ret = -ENXIO;
1372                 goto done;
1373         }
1374         if (nup == 0) {
1375                 ret = -ENETDOWN;
1376                 goto done;
1377         }
1378
1379         if (alg == QIB_PORT_ALG_ACROSS) {
1380                 unsigned inuse = ~0U;
1381                 /* find device (with ACTIVE ports) with fewest ctxts in use */
1382                 for (ndev = 0; ndev < devmax; ndev++) {
1383                         struct qib_devdata *dd = qib_lookup(ndev);
1384                         unsigned cused = 0, cfree = 0, pusable = 0;
1385                         if (!dd)
1386                                 continue;
1387                         if (port && port <= dd->num_pports &&
1388                             usable(dd->pport + port - 1))
1389                                 pusable = 1;
1390                         else
1391                                 for (i = 0; i < dd->num_pports; i++)
1392                                         if (usable(dd->pport + i))
1393                                                 pusable++;
1394                         if (!pusable)
1395                                 continue;
1396                         for (ctxt = dd->first_user_ctxt; ctxt < dd->cfgctxts;
1397                              ctxt++)
1398                                 if (dd->rcd[ctxt])
1399                                         cused++;
1400                                 else
1401                                         cfree++;
1402                         if (pusable && cfree && cused < inuse) {
1403                                 udd = dd;
1404                                 inuse = cused;
1405                         }
1406                 }
1407                 if (udd) {
1408                         ret = choose_port_ctxt(fp, udd, port, uinfo);
1409                         goto done;
1410                 }
1411         } else {
1412                 for (ndev = 0; ndev < devmax; ndev++) {
1413                         struct qib_devdata *dd = qib_lookup(ndev);
1414                         if (dd) {
1415                                 ret = choose_port_ctxt(fp, dd, port, uinfo);
1416                                 if (!ret)
1417                                         goto done;
1418                                 if (ret == -EBUSY)
1419                                         dusable++;
1420                         }
1421                 }
1422         }
1423         ret = dusable ? -EBUSY : -ENETDOWN;
1424
1425 done:
1426         return ret;
1427 }
1428
1429 static int find_shared_ctxt(struct file *fp,
1430                             const struct qib_user_info *uinfo)
1431 {
1432         int devmax, ndev, i;
1433         int ret = 0;
1434
1435         devmax = qib_count_units(NULL, NULL);
1436
1437         for (ndev = 0; ndev < devmax; ndev++) {
1438                 struct qib_devdata *dd = qib_lookup(ndev);
1439
1440                 /* device portion of usable() */
1441                 if (!(dd && (dd->flags & QIB_PRESENT) && dd->kregbase))
1442                         continue;
1443                 for (i = dd->first_user_ctxt; i < dd->cfgctxts; i++) {
1444                         struct qib_ctxtdata *rcd = dd->rcd[i];
1445
1446                         /* Skip ctxts which are not yet open */
1447                         if (!rcd || !rcd->cnt)
1448                                 continue;
1449                         /* Skip ctxt if it doesn't match the requested one */
1450                         if (rcd->subctxt_id != uinfo->spu_subctxt_id)
1451                                 continue;
1452                         /* Verify the sharing process matches the master */
1453                         if (rcd->subctxt_cnt != uinfo->spu_subctxt_cnt ||
1454                             rcd->userversion != uinfo->spu_userversion ||
1455                             rcd->cnt >= rcd->subctxt_cnt) {
1456                                 ret = -EINVAL;
1457                                 goto done;
1458                         }
1459                         ctxt_fp(fp) = rcd;
1460                         subctxt_fp(fp) = rcd->cnt++;
1461                         rcd->subpid[subctxt_fp(fp)] = current->pid;
1462                         tidcursor_fp(fp) = 0;
1463                         rcd->active_slaves |= 1 << subctxt_fp(fp);
1464                         ret = 1;
1465                         goto done;
1466                 }
1467         }
1468
1469 done:
1470         return ret;
1471 }
1472
1473 static int qib_open(struct inode *in, struct file *fp)
1474 {
1475         /* The real work is performed later in qib_assign_ctxt() */
1476         fp->private_data = kzalloc(sizeof(struct qib_filedata), GFP_KERNEL);
1477         if (fp->private_data) /* no cpu affinity by default */
1478                 ((struct qib_filedata *)fp->private_data)->rec_cpu_num = -1;
1479         return fp->private_data ? 0 : -ENOMEM;
1480 }
1481
1482 /*
1483  * Get ctxt early, so can set affinity prior to memory allocation.
1484  */
1485 static int qib_assign_ctxt(struct file *fp, const struct qib_user_info *uinfo)
1486 {
1487         int ret;
1488         int i_minor;
1489         unsigned swmajor, swminor, alg = QIB_PORT_ALG_ACROSS;
1490
1491         /* Check to be sure we haven't already initialized this file */
1492         if (ctxt_fp(fp)) {
1493                 ret = -EINVAL;
1494                 goto done;
1495         }
1496
1497         /* for now, if major version is different, bail */
1498         swmajor = uinfo->spu_userversion >> 16;
1499         if (swmajor != QIB_USER_SWMAJOR) {
1500                 ret = -ENODEV;
1501                 goto done;
1502         }
1503
1504         swminor = uinfo->spu_userversion & 0xffff;
1505
1506         if (swminor >= 11 && uinfo->spu_port_alg < QIB_PORT_ALG_COUNT)
1507                 alg = uinfo->spu_port_alg;
1508
1509         mutex_lock(&qib_mutex);
1510
1511         if (qib_compatible_subctxts(swmajor, swminor) &&
1512             uinfo->spu_subctxt_cnt) {
1513                 ret = find_shared_ctxt(fp, uinfo);
1514                 if (ret) {
1515                         if (ret > 0)
1516                                 ret = 0;
1517                         goto done_chk_sdma;
1518                 }
1519         }
1520
1521         i_minor = iminor(fp->f_dentry->d_inode) - QIB_USER_MINOR_BASE;
1522         if (i_minor)
1523                 ret = find_free_ctxt(i_minor - 1, fp, uinfo);
1524         else
1525                 ret = get_a_ctxt(fp, uinfo, alg);
1526
1527 done_chk_sdma:
1528         if (!ret) {
1529                 struct qib_filedata *fd = fp->private_data;
1530                 const struct qib_ctxtdata *rcd = fd->rcd;
1531                 const struct qib_devdata *dd = rcd->dd;
1532                 unsigned int weight;
1533
1534                 if (dd->flags & QIB_HAS_SEND_DMA) {
1535                         fd->pq = qib_user_sdma_queue_create(&dd->pcidev->dev,
1536                                                             dd->unit,
1537                                                             rcd->ctxt,
1538                                                             fd->subctxt);
1539                         if (!fd->pq)
1540                                 ret = -ENOMEM;
1541                 }
1542
1543                 /*
1544                  * If process has NOT already set it's affinity, select and
1545                  * reserve a processor for it, as a rendezvous for all
1546                  * users of the driver.  If they don't actually later
1547                  * set affinity to this cpu, or set it to some other cpu,
1548                  * it just means that sooner or later we don't recommend
1549                  * a cpu, and let the scheduler do it's best.
1550                  */
1551                 weight = cpumask_weight(tsk_cpus_allowed(current));
1552                 if (!ret && weight >= qib_cpulist_count) {
1553                         int cpu;
1554                         cpu = find_first_zero_bit(qib_cpulist,
1555                                                   qib_cpulist_count);
1556                         if (cpu != qib_cpulist_count) {
1557                                 __set_bit(cpu, qib_cpulist);
1558                                 fd->rec_cpu_num = cpu;
1559                         }
1560                 } else if (weight == 1 &&
1561                         test_bit(cpumask_first(tsk_cpus_allowed(current)),
1562                                  qib_cpulist))
1563                         qib_devinfo(dd->pcidev, "%s PID %u affinity "
1564                                     "set to cpu %d; already allocated\n",
1565                                     current->comm, current->pid,
1566                                     cpumask_first(tsk_cpus_allowed(current)));
1567         }
1568
1569         mutex_unlock(&qib_mutex);
1570
1571 done:
1572         return ret;
1573 }
1574
1575
1576 static int qib_do_user_init(struct file *fp,
1577                             const struct qib_user_info *uinfo)
1578 {
1579         int ret;
1580         struct qib_ctxtdata *rcd = ctxt_fp(fp);
1581         struct qib_devdata *dd;
1582         unsigned uctxt;
1583
1584         /* Subctxts don't need to initialize anything since master did it. */
1585         if (subctxt_fp(fp)) {
1586                 ret = wait_event_interruptible(rcd->wait,
1587                         !test_bit(QIB_CTXT_MASTER_UNINIT, &rcd->flag));
1588                 goto bail;
1589         }
1590
1591         dd = rcd->dd;
1592
1593         /* some ctxts may get extra buffers, calculate that here */
1594         uctxt = rcd->ctxt - dd->first_user_ctxt;
1595         if (uctxt < dd->ctxts_extrabuf) {
1596                 rcd->piocnt = dd->pbufsctxt + 1;
1597                 rcd->pio_base = rcd->piocnt * uctxt;
1598         } else {
1599                 rcd->piocnt = dd->pbufsctxt;
1600                 rcd->pio_base = rcd->piocnt * uctxt +
1601                         dd->ctxts_extrabuf;
1602         }
1603
1604         /*
1605          * All user buffers are 2KB buffers.  If we ever support
1606          * giving 4KB buffers to user processes, this will need some
1607          * work.  Can't use piobufbase directly, because it has
1608          * both 2K and 4K buffer base values.  So check and handle.
1609          */
1610         if ((rcd->pio_base + rcd->piocnt) > dd->piobcnt2k) {
1611                 if (rcd->pio_base >= dd->piobcnt2k) {
1612                         qib_dev_err(dd,
1613                                     "%u:ctxt%u: no 2KB buffers available\n",
1614                                     dd->unit, rcd->ctxt);
1615                         ret = -ENOBUFS;
1616                         goto bail;
1617                 }
1618                 rcd->piocnt = dd->piobcnt2k - rcd->pio_base;
1619                 qib_dev_err(dd, "Ctxt%u: would use 4KB bufs, using %u\n",
1620                             rcd->ctxt, rcd->piocnt);
1621         }
1622
1623         rcd->piobufs = dd->pio2k_bufbase + rcd->pio_base * dd->palign;
1624         qib_chg_pioavailkernel(dd, rcd->pio_base, rcd->piocnt,
1625                                TXCHK_CHG_TYPE_USER, rcd);
1626         /*
1627          * try to ensure that processes start up with consistent avail update
1628          * for their own range, at least.   If system very quiet, it might
1629          * have the in-memory copy out of date at startup for this range of
1630          * buffers, when a context gets re-used.  Do after the chg_pioavail
1631          * and before the rest of setup, so it's "almost certain" the dma
1632          * will have occurred (can't 100% guarantee, but should be many
1633          * decimals of 9s, with this ordering), given how much else happens
1634          * after this.
1635          */
1636         dd->f_sendctrl(dd->pport, QIB_SENDCTRL_AVAIL_BLIP);
1637
1638         /*
1639          * Now allocate the rcvhdr Q and eager TIDs; skip the TID
1640          * array for time being.  If rcd->ctxt > chip-supported,
1641          * we need to do extra stuff here to handle by handling overflow
1642          * through ctxt 0, someday
1643          */
1644         ret = qib_create_rcvhdrq(dd, rcd);
1645         if (!ret)
1646                 ret = qib_setup_eagerbufs(rcd);
1647         if (ret)
1648                 goto bail_pio;
1649
1650         rcd->tidcursor = 0; /* start at beginning after open */
1651
1652         /* initialize poll variables... */
1653         rcd->urgent = 0;
1654         rcd->urgent_poll = 0;
1655
1656         /*
1657          * Now enable the ctxt for receive.
1658          * For chips that are set to DMA the tail register to memory
1659          * when they change (and when the update bit transitions from
1660          * 0 to 1.  So for those chips, we turn it off and then back on.
1661          * This will (very briefly) affect any other open ctxts, but the
1662          * duration is very short, and therefore isn't an issue.  We
1663          * explicitly set the in-memory tail copy to 0 beforehand, so we
1664          * don't have to wait to be sure the DMA update has happened
1665          * (chip resets head/tail to 0 on transition to enable).
1666          */
1667         if (rcd->rcvhdrtail_kvaddr)
1668                 qib_clear_rcvhdrtail(rcd);
1669
1670         dd->f_rcvctrl(rcd->ppd, QIB_RCVCTRL_CTXT_ENB | QIB_RCVCTRL_TIDFLOW_ENB,
1671                       rcd->ctxt);
1672
1673         /* Notify any waiting slaves */
1674         if (rcd->subctxt_cnt) {
1675                 clear_bit(QIB_CTXT_MASTER_UNINIT, &rcd->flag);
1676                 wake_up(&rcd->wait);
1677         }
1678         return 0;
1679
1680 bail_pio:
1681         qib_chg_pioavailkernel(dd, rcd->pio_base, rcd->piocnt,
1682                                TXCHK_CHG_TYPE_KERN, rcd);
1683 bail:
1684         return ret;
1685 }
1686
1687 /**
1688  * unlock_exptid - unlock any expected TID entries context still had in use
1689  * @rcd: ctxt
1690  *
1691  * We don't actually update the chip here, because we do a bulk update
1692  * below, using f_clear_tids.
1693  */
1694 static void unlock_expected_tids(struct qib_ctxtdata *rcd)
1695 {
1696         struct qib_devdata *dd = rcd->dd;
1697         int ctxt_tidbase = rcd->ctxt * dd->rcvtidcnt;
1698         int i, cnt = 0, maxtid = ctxt_tidbase + dd->rcvtidcnt;
1699
1700         for (i = ctxt_tidbase; i < maxtid; i++) {
1701                 struct page *p = dd->pageshadow[i];
1702                 dma_addr_t phys;
1703
1704                 if (!p)
1705                         continue;
1706
1707                 phys = dd->physshadow[i];
1708                 dd->physshadow[i] = dd->tidinvalid;
1709                 dd->pageshadow[i] = NULL;
1710                 pci_unmap_page(dd->pcidev, phys, PAGE_SIZE,
1711                                PCI_DMA_FROMDEVICE);
1712                 qib_release_user_pages(&p, 1);
1713                 cnt++;
1714         }
1715 }
1716
1717 static int qib_close(struct inode *in, struct file *fp)
1718 {
1719         int ret = 0;
1720         struct qib_filedata *fd;
1721         struct qib_ctxtdata *rcd;
1722         struct qib_devdata *dd;
1723         unsigned long flags;
1724         unsigned ctxt;
1725         pid_t pid;
1726
1727         mutex_lock(&qib_mutex);
1728
1729         fd = fp->private_data;
1730         fp->private_data = NULL;
1731         rcd = fd->rcd;
1732         if (!rcd) {
1733                 mutex_unlock(&qib_mutex);
1734                 goto bail;
1735         }
1736
1737         dd = rcd->dd;
1738
1739         /* ensure all pio buffer writes in progress are flushed */
1740         qib_flush_wc();
1741
1742         /* drain user sdma queue */
1743         if (fd->pq) {
1744                 qib_user_sdma_queue_drain(rcd->ppd, fd->pq);
1745                 qib_user_sdma_queue_destroy(fd->pq);
1746         }
1747
1748         if (fd->rec_cpu_num != -1)
1749                 __clear_bit(fd->rec_cpu_num, qib_cpulist);
1750
1751         if (--rcd->cnt) {
1752                 /*
1753                  * XXX If the master closes the context before the slave(s),
1754                  * revoke the mmap for the eager receive queue so
1755                  * the slave(s) don't wait for receive data forever.
1756                  */
1757                 rcd->active_slaves &= ~(1 << fd->subctxt);
1758                 rcd->subpid[fd->subctxt] = 0;
1759                 mutex_unlock(&qib_mutex);
1760                 goto bail;
1761         }
1762
1763         /* early; no interrupt users after this */
1764         spin_lock_irqsave(&dd->uctxt_lock, flags);
1765         ctxt = rcd->ctxt;
1766         dd->rcd[ctxt] = NULL;
1767         pid = rcd->pid;
1768         rcd->pid = 0;
1769         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
1770
1771         if (rcd->rcvwait_to || rcd->piowait_to ||
1772             rcd->rcvnowait || rcd->pionowait) {
1773                 rcd->rcvwait_to = 0;
1774                 rcd->piowait_to = 0;
1775                 rcd->rcvnowait = 0;
1776                 rcd->pionowait = 0;
1777         }
1778         if (rcd->flag)
1779                 rcd->flag = 0;
1780
1781         if (dd->kregbase) {
1782                 /* atomically clear receive enable ctxt and intr avail. */
1783                 dd->f_rcvctrl(rcd->ppd, QIB_RCVCTRL_CTXT_DIS |
1784                                   QIB_RCVCTRL_INTRAVAIL_DIS, ctxt);
1785
1786                 /* clean up the pkeys for this ctxt user */
1787                 qib_clean_part_key(rcd, dd);
1788                 qib_disarm_piobufs(dd, rcd->pio_base, rcd->piocnt);
1789                 qib_chg_pioavailkernel(dd, rcd->pio_base,
1790                                        rcd->piocnt, TXCHK_CHG_TYPE_KERN, NULL);
1791
1792                 dd->f_clear_tids(dd, rcd);
1793
1794                 if (dd->pageshadow)
1795                         unlock_expected_tids(rcd);
1796                 qib_stats.sps_ctxts--;
1797                 dd->freectxts++;
1798         }
1799
1800         mutex_unlock(&qib_mutex);
1801         qib_free_ctxtdata(dd, rcd); /* after releasing the mutex */
1802
1803 bail:
1804         kfree(fd);
1805         return ret;
1806 }
1807
1808 static int qib_ctxt_info(struct file *fp, struct qib_ctxt_info __user *uinfo)
1809 {
1810         struct qib_ctxt_info info;
1811         int ret;
1812         size_t sz;
1813         struct qib_ctxtdata *rcd = ctxt_fp(fp);
1814         struct qib_filedata *fd;
1815
1816         fd = fp->private_data;
1817
1818         info.num_active = qib_count_active_units();
1819         info.unit = rcd->dd->unit;
1820         info.port = rcd->ppd->port;
1821         info.ctxt = rcd->ctxt;
1822         info.subctxt =  subctxt_fp(fp);
1823         /* Number of user ctxts available for this device. */
1824         info.num_ctxts = rcd->dd->cfgctxts - rcd->dd->first_user_ctxt;
1825         info.num_subctxts = rcd->subctxt_cnt;
1826         info.rec_cpu = fd->rec_cpu_num;
1827         sz = sizeof(info);
1828
1829         if (copy_to_user(uinfo, &info, sz)) {
1830                 ret = -EFAULT;
1831                 goto bail;
1832         }
1833         ret = 0;
1834
1835 bail:
1836         return ret;
1837 }
1838
1839 static int qib_sdma_get_inflight(struct qib_user_sdma_queue *pq,
1840                                  u32 __user *inflightp)
1841 {
1842         const u32 val = qib_user_sdma_inflight_counter(pq);
1843
1844         if (put_user(val, inflightp))
1845                 return -EFAULT;
1846
1847         return 0;
1848 }
1849
1850 static int qib_sdma_get_complete(struct qib_pportdata *ppd,
1851                                  struct qib_user_sdma_queue *pq,
1852                                  u32 __user *completep)
1853 {
1854         u32 val;
1855         int err;
1856
1857         if (!pq)
1858                 return -EINVAL;
1859
1860         err = qib_user_sdma_make_progress(ppd, pq);
1861         if (err < 0)
1862                 return err;
1863
1864         val = qib_user_sdma_complete_counter(pq);
1865         if (put_user(val, completep))
1866                 return -EFAULT;
1867
1868         return 0;
1869 }
1870
1871 static int disarm_req_delay(struct qib_ctxtdata *rcd)
1872 {
1873         int ret = 0;
1874
1875         if (!usable(rcd->ppd)) {
1876                 int i;
1877                 /*
1878                  * if link is down, or otherwise not usable, delay
1879                  * the caller up to 30 seconds, so we don't thrash
1880                  * in trying to get the chip back to ACTIVE, and
1881                  * set flag so they make the call again.
1882                  */
1883                 if (rcd->user_event_mask) {
1884                         /*
1885                          * subctxt_cnt is 0 if not shared, so do base
1886                          * separately, first, then remaining subctxt, if any
1887                          */
1888                         set_bit(_QIB_EVENT_DISARM_BUFS_BIT,
1889                                 &rcd->user_event_mask[0]);
1890                         for (i = 1; i < rcd->subctxt_cnt; i++)
1891                                 set_bit(_QIB_EVENT_DISARM_BUFS_BIT,
1892                                         &rcd->user_event_mask[i]);
1893                 }
1894                 for (i = 0; !usable(rcd->ppd) && i < 300; i++)
1895                         msleep(100);
1896                 ret = -ENETDOWN;
1897         }
1898         return ret;
1899 }
1900
1901 /*
1902  * Find all user contexts in use, and set the specified bit in their
1903  * event mask.
1904  * See also find_ctxt() for a similar use, that is specific to send buffers.
1905  */
1906 int qib_set_uevent_bits(struct qib_pportdata *ppd, const int evtbit)
1907 {
1908         struct qib_ctxtdata *rcd;
1909         unsigned ctxt;
1910         int ret = 0;
1911         unsigned long flags;
1912
1913         spin_lock_irqsave(&ppd->dd->uctxt_lock, flags);
1914         for (ctxt = ppd->dd->first_user_ctxt; ctxt < ppd->dd->cfgctxts;
1915              ctxt++) {
1916                 rcd = ppd->dd->rcd[ctxt];
1917                 if (!rcd)
1918                         continue;
1919                 if (rcd->user_event_mask) {
1920                         int i;
1921                         /*
1922                          * subctxt_cnt is 0 if not shared, so do base
1923                          * separately, first, then remaining subctxt, if any
1924                          */
1925                         set_bit(evtbit, &rcd->user_event_mask[0]);
1926                         for (i = 1; i < rcd->subctxt_cnt; i++)
1927                                 set_bit(evtbit, &rcd->user_event_mask[i]);
1928                 }
1929                 ret = 1;
1930                 break;
1931         }
1932         spin_unlock_irqrestore(&ppd->dd->uctxt_lock, flags);
1933
1934         return ret;
1935 }
1936
1937 /*
1938  * clear the event notifier events for this context.
1939  * For the DISARM_BUFS case, we also take action (this obsoletes
1940  * the older QIB_CMD_DISARM_BUFS, but we keep it for backwards
1941  * compatibility.
1942  * Other bits don't currently require actions, just atomically clear.
1943  * User process then performs actions appropriate to bit having been
1944  * set, if desired, and checks again in future.
1945  */
1946 static int qib_user_event_ack(struct qib_ctxtdata *rcd, int subctxt,
1947                               unsigned long events)
1948 {
1949         int ret = 0, i;
1950
1951         for (i = 0; i <= _QIB_MAX_EVENT_BIT; i++) {
1952                 if (!test_bit(i, &events))
1953                         continue;
1954                 if (i == _QIB_EVENT_DISARM_BUFS_BIT) {
1955                         (void)qib_disarm_piobufs_ifneeded(rcd);
1956                         ret = disarm_req_delay(rcd);
1957                 } else
1958                         clear_bit(i, &rcd->user_event_mask[subctxt]);
1959         }
1960         return ret;
1961 }
1962
1963 static ssize_t qib_write(struct file *fp, const char __user *data,
1964                          size_t count, loff_t *off)
1965 {
1966         const struct qib_cmd __user *ucmd;
1967         struct qib_ctxtdata *rcd;
1968         const void __user *src;
1969         size_t consumed, copy = 0;
1970         struct qib_cmd cmd;
1971         ssize_t ret = 0;
1972         void *dest;
1973
1974         if (count < sizeof(cmd.type)) {
1975                 ret = -EINVAL;
1976                 goto bail;
1977         }
1978
1979         ucmd = (const struct qib_cmd __user *) data;
1980
1981         if (copy_from_user(&cmd.type, &ucmd->type, sizeof(cmd.type))) {
1982                 ret = -EFAULT;
1983                 goto bail;
1984         }
1985
1986         consumed = sizeof(cmd.type);
1987
1988         switch (cmd.type) {
1989         case QIB_CMD_ASSIGN_CTXT:
1990         case QIB_CMD_USER_INIT:
1991                 copy = sizeof(cmd.cmd.user_info);
1992                 dest = &cmd.cmd.user_info;
1993                 src = &ucmd->cmd.user_info;
1994                 break;
1995
1996         case QIB_CMD_RECV_CTRL:
1997                 copy = sizeof(cmd.cmd.recv_ctrl);
1998                 dest = &cmd.cmd.recv_ctrl;
1999                 src = &ucmd->cmd.recv_ctrl;
2000                 break;
2001
2002         case QIB_CMD_CTXT_INFO:
2003                 copy = sizeof(cmd.cmd.ctxt_info);
2004                 dest = &cmd.cmd.ctxt_info;
2005                 src = &ucmd->cmd.ctxt_info;
2006                 break;
2007
2008         case QIB_CMD_TID_UPDATE:
2009         case QIB_CMD_TID_FREE:
2010                 copy = sizeof(cmd.cmd.tid_info);
2011                 dest = &cmd.cmd.tid_info;
2012                 src = &ucmd->cmd.tid_info;
2013                 break;
2014
2015         case QIB_CMD_SET_PART_KEY:
2016                 copy = sizeof(cmd.cmd.part_key);
2017                 dest = &cmd.cmd.part_key;
2018                 src = &ucmd->cmd.part_key;
2019                 break;
2020
2021         case QIB_CMD_DISARM_BUFS:
2022         case QIB_CMD_PIOAVAILUPD: /* force an update of PIOAvail reg */
2023                 copy = 0;
2024                 src = NULL;
2025                 dest = NULL;
2026                 break;
2027
2028         case QIB_CMD_POLL_TYPE:
2029                 copy = sizeof(cmd.cmd.poll_type);
2030                 dest = &cmd.cmd.poll_type;
2031                 src = &ucmd->cmd.poll_type;
2032                 break;
2033
2034         case QIB_CMD_ARMLAUNCH_CTRL:
2035                 copy = sizeof(cmd.cmd.armlaunch_ctrl);
2036                 dest = &cmd.cmd.armlaunch_ctrl;
2037                 src = &ucmd->cmd.armlaunch_ctrl;
2038                 break;
2039
2040         case QIB_CMD_SDMA_INFLIGHT:
2041                 copy = sizeof(cmd.cmd.sdma_inflight);
2042                 dest = &cmd.cmd.sdma_inflight;
2043                 src = &ucmd->cmd.sdma_inflight;
2044                 break;
2045
2046         case QIB_CMD_SDMA_COMPLETE:
2047                 copy = sizeof(cmd.cmd.sdma_complete);
2048                 dest = &cmd.cmd.sdma_complete;
2049                 src = &ucmd->cmd.sdma_complete;
2050                 break;
2051
2052         case QIB_CMD_ACK_EVENT:
2053                 copy = sizeof(cmd.cmd.event_mask);
2054                 dest = &cmd.cmd.event_mask;
2055                 src = &ucmd->cmd.event_mask;
2056                 break;
2057
2058         default:
2059                 ret = -EINVAL;
2060                 goto bail;
2061         }
2062
2063         if (copy) {
2064                 if ((count - consumed) < copy) {
2065                         ret = -EINVAL;
2066                         goto bail;
2067                 }
2068                 if (copy_from_user(dest, src, copy)) {
2069                         ret = -EFAULT;
2070                         goto bail;
2071                 }
2072                 consumed += copy;
2073         }
2074
2075         rcd = ctxt_fp(fp);
2076         if (!rcd && cmd.type != QIB_CMD_ASSIGN_CTXT) {
2077                 ret = -EINVAL;
2078                 goto bail;
2079         }
2080
2081         switch (cmd.type) {
2082         case QIB_CMD_ASSIGN_CTXT:
2083                 ret = qib_assign_ctxt(fp, &cmd.cmd.user_info);
2084                 if (ret)
2085                         goto bail;
2086                 break;
2087
2088         case QIB_CMD_USER_INIT:
2089                 ret = qib_do_user_init(fp, &cmd.cmd.user_info);
2090                 if (ret)
2091                         goto bail;
2092                 ret = qib_get_base_info(fp, (void __user *) (unsigned long)
2093                                         cmd.cmd.user_info.spu_base_info,
2094                                         cmd.cmd.user_info.spu_base_info_size);
2095                 break;
2096
2097         case QIB_CMD_RECV_CTRL:
2098                 ret = qib_manage_rcvq(rcd, subctxt_fp(fp), cmd.cmd.recv_ctrl);
2099                 break;
2100
2101         case QIB_CMD_CTXT_INFO:
2102                 ret = qib_ctxt_info(fp, (struct qib_ctxt_info __user *)
2103                                     (unsigned long) cmd.cmd.ctxt_info);
2104                 break;
2105
2106         case QIB_CMD_TID_UPDATE:
2107                 ret = qib_tid_update(rcd, fp, &cmd.cmd.tid_info);
2108                 break;
2109
2110         case QIB_CMD_TID_FREE:
2111                 ret = qib_tid_free(rcd, subctxt_fp(fp), &cmd.cmd.tid_info);
2112                 break;
2113
2114         case QIB_CMD_SET_PART_KEY:
2115                 ret = qib_set_part_key(rcd, cmd.cmd.part_key);
2116                 break;
2117
2118         case QIB_CMD_DISARM_BUFS:
2119                 (void)qib_disarm_piobufs_ifneeded(rcd);
2120                 ret = disarm_req_delay(rcd);
2121                 break;
2122
2123         case QIB_CMD_PIOAVAILUPD:
2124                 qib_force_pio_avail_update(rcd->dd);
2125                 break;
2126
2127         case QIB_CMD_POLL_TYPE:
2128                 rcd->poll_type = cmd.cmd.poll_type;
2129                 break;
2130
2131         case QIB_CMD_ARMLAUNCH_CTRL:
2132                 rcd->dd->f_set_armlaunch(rcd->dd, cmd.cmd.armlaunch_ctrl);
2133                 break;
2134
2135         case QIB_CMD_SDMA_INFLIGHT:
2136                 ret = qib_sdma_get_inflight(user_sdma_queue_fp(fp),
2137                                             (u32 __user *) (unsigned long)
2138                                             cmd.cmd.sdma_inflight);
2139                 break;
2140
2141         case QIB_CMD_SDMA_COMPLETE:
2142                 ret = qib_sdma_get_complete(rcd->ppd,
2143                                             user_sdma_queue_fp(fp),
2144                                             (u32 __user *) (unsigned long)
2145                                             cmd.cmd.sdma_complete);
2146                 break;
2147
2148         case QIB_CMD_ACK_EVENT:
2149                 ret = qib_user_event_ack(rcd, subctxt_fp(fp),
2150                                          cmd.cmd.event_mask);
2151                 break;
2152         }
2153
2154         if (ret >= 0)
2155                 ret = consumed;
2156
2157 bail:
2158         return ret;
2159 }
2160
2161 static ssize_t qib_aio_write(struct kiocb *iocb, const struct iovec *iov,
2162                              unsigned long dim, loff_t off)
2163 {
2164         struct qib_filedata *fp = iocb->ki_filp->private_data;
2165         struct qib_ctxtdata *rcd = ctxt_fp(iocb->ki_filp);
2166         struct qib_user_sdma_queue *pq = fp->pq;
2167
2168         if (!dim || !pq)
2169                 return -EINVAL;
2170
2171         return qib_user_sdma_writev(rcd, pq, iov, dim);
2172 }
2173
2174 static struct class *qib_class;
2175 static dev_t qib_dev;
2176
2177 int qib_cdev_init(int minor, const char *name,
2178                   const struct file_operations *fops,
2179                   struct cdev **cdevp, struct device **devp)
2180 {
2181         const dev_t dev = MKDEV(MAJOR(qib_dev), minor);
2182         struct cdev *cdev;
2183         struct device *device = NULL;
2184         int ret;
2185
2186         cdev = cdev_alloc();
2187         if (!cdev) {
2188                 printk(KERN_ERR QIB_DRV_NAME
2189                        ": Could not allocate cdev for minor %d, %s\n",
2190                        minor, name);
2191                 ret = -ENOMEM;
2192                 goto done;
2193         }
2194
2195         cdev->owner = THIS_MODULE;
2196         cdev->ops = fops;
2197         kobject_set_name(&cdev->kobj, name);
2198
2199         ret = cdev_add(cdev, dev, 1);
2200         if (ret < 0) {
2201                 printk(KERN_ERR QIB_DRV_NAME
2202                        ": Could not add cdev for minor %d, %s (err %d)\n",
2203                        minor, name, -ret);
2204                 goto err_cdev;
2205         }
2206
2207         device = device_create(qib_class, NULL, dev, NULL, name);
2208         if (!IS_ERR(device))
2209                 goto done;
2210         ret = PTR_ERR(device);
2211         device = NULL;
2212         printk(KERN_ERR QIB_DRV_NAME ": Could not create "
2213                "device for minor %d, %s (err %d)\n",
2214                minor, name, -ret);
2215 err_cdev:
2216         cdev_del(cdev);
2217         cdev = NULL;
2218 done:
2219         *cdevp = cdev;
2220         *devp = device;
2221         return ret;
2222 }
2223
2224 void qib_cdev_cleanup(struct cdev **cdevp, struct device **devp)
2225 {
2226         struct device *device = *devp;
2227
2228         if (device) {
2229                 device_unregister(device);
2230                 *devp = NULL;
2231         }
2232
2233         if (*cdevp) {
2234                 cdev_del(*cdevp);
2235                 *cdevp = NULL;
2236         }
2237 }
2238
2239 static struct cdev *wildcard_cdev;
2240 static struct device *wildcard_device;
2241
2242 int __init qib_dev_init(void)
2243 {
2244         int ret;
2245
2246         ret = alloc_chrdev_region(&qib_dev, 0, QIB_NMINORS, QIB_DRV_NAME);
2247         if (ret < 0) {
2248                 printk(KERN_ERR QIB_DRV_NAME ": Could not allocate "
2249                        "chrdev region (err %d)\n", -ret);
2250                 goto done;
2251         }
2252
2253         qib_class = class_create(THIS_MODULE, "ipath");
2254         if (IS_ERR(qib_class)) {
2255                 ret = PTR_ERR(qib_class);
2256                 printk(KERN_ERR QIB_DRV_NAME ": Could not create "
2257                        "device class (err %d)\n", -ret);
2258                 unregister_chrdev_region(qib_dev, QIB_NMINORS);
2259         }
2260
2261 done:
2262         return ret;
2263 }
2264
2265 void qib_dev_cleanup(void)
2266 {
2267         if (qib_class) {
2268                 class_destroy(qib_class);
2269                 qib_class = NULL;
2270         }
2271
2272         unregister_chrdev_region(qib_dev, QIB_NMINORS);
2273 }
2274
2275 static atomic_t user_count = ATOMIC_INIT(0);
2276
2277 static void qib_user_remove(struct qib_devdata *dd)
2278 {
2279         if (atomic_dec_return(&user_count) == 0)
2280                 qib_cdev_cleanup(&wildcard_cdev, &wildcard_device);
2281
2282         qib_cdev_cleanup(&dd->user_cdev, &dd->user_device);
2283 }
2284
2285 static int qib_user_add(struct qib_devdata *dd)
2286 {
2287         char name[10];
2288         int ret;
2289
2290         if (atomic_inc_return(&user_count) == 1) {
2291                 ret = qib_cdev_init(0, "ipath", &qib_file_ops,
2292                                     &wildcard_cdev, &wildcard_device);
2293                 if (ret)
2294                         goto done;
2295         }
2296
2297         snprintf(name, sizeof(name), "ipath%d", dd->unit);
2298         ret = qib_cdev_init(dd->unit + 1, name, &qib_file_ops,
2299                             &dd->user_cdev, &dd->user_device);
2300         if (ret)
2301                 qib_user_remove(dd);
2302 done:
2303         return ret;
2304 }
2305
2306 /*
2307  * Create per-unit files in /dev
2308  */
2309 int qib_device_create(struct qib_devdata *dd)
2310 {
2311         int r, ret;
2312
2313         r = qib_user_add(dd);
2314         ret = qib_diag_add(dd);
2315         if (r && !ret)
2316                 ret = r;
2317         return ret;
2318 }
2319
2320 /*
2321  * Remove per-unit files in /dev
2322  * void, core kernel returns no errors for this stuff
2323  */
2324 void qib_device_remove(struct qib_devdata *dd)
2325 {
2326         qib_user_remove(dd);
2327         qib_diag_remove(dd);
2328 }