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
  47 #include "qib.h"
  48 #include "qib_common.h"
  49 #include "qib_user_sdma.h"
  50
  51 static int qib_open(struct inode *, struct file *);
  52 static int qib_close(struct inode *, struct file *);
  53 static ssize_t qib_write(struct file *, const char __user *, size_t, loff_t *);
  54 static ssize_t qib_aio_write(struct kiocb *, const struct iovec *,
  55                              unsigned long, loff_t);
  56 static unsigned int qib_poll(struct file *, struct poll_table_struct *);
  57 static int qib_mmapf(struct file *, struct vm_area_struct *);
  58
  59 static const struct file_operations qib_file_ops = {
  60         .owner = THIS_MODULE,
  61         .write = qib_write,
  62         .aio_write = qib_aio_write,
  63         .open = qib_open,
  64         .release = qib_close,
  65         .poll = qib_poll,
  66         .mmap = qib_mmapf,
  67         .llseek = noop_llseek,
  68 };
  69
  70 /*
  71  * Convert kernel virtual addresses to physical addresses so they don't
  72  * potentially conflict with the chip addresses used as mmap offsets.
  73  * It doesn't really matter what mmap offset we use as long as we can
  74  * interpret it correctly.
  75  */
  76 static u64 cvt_kvaddr(void *p)
  77 {
  78         struct page *page;
  79         u64 paddr = 0;
  80
  81         page = vmalloc_to_page(p);
  82         if (page)
  83                 paddr = page_to_pfn(page) << PAGE_SHIFT;
  84
  85         return paddr;
  86 }
  87
  88 static int qib_get_base_info(struct file *fp, void __user *ubase,
  89                              size_t ubase_size)
  90 {
  91         struct qib_ctxtdata *rcd = ctxt_fp(fp);
  92         int ret = 0;
  93         struct qib_base_info *kinfo = NULL;
  94         struct qib_devdata *dd = rcd->dd;
  95         struct qib_pportdata *ppd = rcd->ppd;
  96         unsigned subctxt_cnt;
  97         int shared, master;
  98         size_t sz;
  99
 100         subctxt_cnt = rcd->subctxt_cnt;
 101         if (!subctxt_cnt) {
 102                 shared = 0;
 103                 master = 0;
 104                 subctxt_cnt = 1;
 105         } else {
 106                 shared = 1;
 107                 master = !subctxt_fp(fp);
 108         }
 109
 110         sz = sizeof(*kinfo);
 111         /* If context sharing is not requested, allow the old size structure */
 112         if (!shared)
 113                 sz -= 7 * sizeof(u64);
 114         if (ubase_size < sz) {
 115                 ret = -EINVAL;
 116                 goto bail;
 117         }
 118
 119         kinfo = kzalloc(sizeof(*kinfo), GFP_KERNEL);
 120         if (kinfo == NULL) {
 121                 ret = -ENOMEM;
 122                 goto bail;
 123         }
 124
 125         ret = dd->f_get_base_info(rcd, kinfo);
 126         if (ret < 0)
 127                 goto bail;
 128
 129         kinfo->spi_rcvhdr_cnt = dd->rcvhdrcnt;
 130         kinfo->spi_rcvhdrent_size = dd->rcvhdrentsize;
 131         kinfo->spi_tidegrcnt = rcd->rcvegrcnt;
 132         kinfo->spi_rcv_egrbufsize = dd->rcvegrbufsize;
 133         /*
 134          * have to mmap whole thing
 135          */
 136         kinfo->spi_rcv_egrbuftotlen =
 137                 rcd->rcvegrbuf_chunks * rcd->rcvegrbuf_size;
 138         kinfo->spi_rcv_egrperchunk = rcd->rcvegrbufs_perchunk;
 139         kinfo->spi_rcv_egrchunksize = kinfo->spi_rcv_egrbuftotlen /
 140                 rcd->rcvegrbuf_chunks;
 141         kinfo->spi_tidcnt = dd->rcvtidcnt / subctxt_cnt;
 142         if (master)
 143                 kinfo->spi_tidcnt += dd->rcvtidcnt % subctxt_cnt;
 144         /*
 145          * for this use, may be cfgctxts summed over all chips that
 146          * are are configured and present
 147          */
 148         kinfo->spi_nctxts = dd->cfgctxts;
 149         /* unit (chip/board) our context is on */
 150         kinfo->spi_unit = dd->unit;
 151         kinfo->spi_port = ppd->port;
 152         /* for now, only a single page */
 153         kinfo->spi_tid_maxsize = PAGE_SIZE;
 154
 155         /*
 156          * Doing this per context, and based on the skip value, etc.  This has
 157          * to be the actual buffer size, since the protocol code treats it
 158          * as an array.
 159          *
 160          * These have to be set to user addresses in the user code via mmap.
 161          * These values are used on return to user code for the mmap target
 162          * addresses only.  For 32 bit, same 44 bit address problem, so use
 163          * the physical address, not virtual.  Before 2.6.11, using the
 164          * page_address() macro worked, but in 2.6.11, even that returns the
 165          * full 64 bit address (upper bits all 1's).  So far, using the
 166          * physical addresses (or chip offsets, for chip mapping) works, but
 167          * no doubt some future kernel release will change that, and we'll be
 168          * on to yet another method of dealing with this.
 169          * Normally only one of rcvhdr_tailaddr or rhf_offset is useful
 170          * since the chips with non-zero rhf_offset don't normally
 171          * enable tail register updates to host memory, but for testing,
 172          * both can be enabled and used.
 173          */
 174         kinfo->spi_rcvhdr_base = (u64) rcd->rcvhdrq_phys;
 175         kinfo->spi_rcvhdr_tailaddr = (u64) rcd->rcvhdrqtailaddr_phys;
 176         kinfo->spi_rhf_offset = dd->rhf_offset;
 177         kinfo->spi_rcv_egrbufs = (u64) rcd->rcvegr_phys;
 178         kinfo->spi_pioavailaddr = (u64) dd->pioavailregs_phys;
 179         /* setup per-unit (not port) status area for user programs */
 180         kinfo->spi_status = (u64) kinfo->spi_pioavailaddr +
 181                 (char *) ppd->statusp -
 182                 (char *) dd->pioavailregs_dma;
 183         kinfo->spi_uregbase = (u64) dd->uregbase + dd->ureg_align * rcd->ctxt;
 184         if (!shared) {
 185                 kinfo->spi_piocnt = rcd->piocnt;
 186                 kinfo->spi_piobufbase = (u64) rcd->piobufs;
 187                 kinfo->spi_sendbuf_status = cvt_kvaddr(rcd->user_event_mask);
 188         } else if (master) {
 189                 kinfo->spi_piocnt = (rcd->piocnt / subctxt_cnt) +
 190                                     (rcd->piocnt % subctxt_cnt);
 191                 /* Master's PIO buffers are after all the slave's */
 192                 kinfo->spi_piobufbase = (u64) rcd->piobufs +
 193                         dd->palign *
 194                         (rcd->piocnt - kinfo->spi_piocnt);
 195         } else {
 196                 unsigned slave = subctxt_fp(fp) - 1;
 197
 198                 kinfo->spi_piocnt = rcd->piocnt / subctxt_cnt;
 199                 kinfo->spi_piobufbase = (u64) rcd->piobufs +
 200                         dd->palign * kinfo->spi_piocnt * slave;
 201         }
 202
 203         if (shared) {
 204                 kinfo->spi_sendbuf_status =
 205                         cvt_kvaddr(&rcd->user_event_mask[subctxt_fp(fp)]);
 206                 /* only spi_subctxt_* fields should be set in this block! */
 207                 kinfo->spi_subctxt_uregbase = cvt_kvaddr(rcd->subctxt_uregbase);
 208
 209                 kinfo->spi_subctxt_rcvegrbuf =
 210                         cvt_kvaddr(rcd->subctxt_rcvegrbuf);
 211                 kinfo->spi_subctxt_rcvhdr_base =
 212                         cvt_kvaddr(rcd->subctxt_rcvhdr_base);
 213         }
 214
 215         /*
 216          * All user buffers are 2KB buffers.  If we ever support
 217          * giving 4KB buffers to user processes, this will need some
 218          * work.  Can't use piobufbase directly, because it has
 219          * both 2K and 4K buffer base values.
 220          */
 221         kinfo->spi_pioindex = (kinfo->spi_piobufbase - dd->pio2k_bufbase) /
 222                 dd->palign;
 223         kinfo->spi_pioalign = dd->palign;
 224         kinfo->spi_qpair = QIB_KD_QP;
 225         /*
 226          * user mode PIO buffers are always 2KB, even when 4KB can
 227          * be received, and sent via the kernel; this is ibmaxlen
 228          * for 2K MTU.
 229          */
 230         kinfo->spi_piosize = dd->piosize2k - 2 * sizeof(u32);
 231         kinfo->spi_mtu = ppd->ibmaxlen; /* maxlen, not ibmtu */
 232         kinfo->spi_ctxt = rcd->ctxt;
 233         kinfo->spi_subctxt = subctxt_fp(fp);
 234         kinfo->spi_sw_version = QIB_KERN_SWVERSION;
 235         kinfo->spi_sw_version |= 1U << 31; /* QLogic-built, not kernel.org */
 236         kinfo->spi_hw_version = dd->revision;
 237
 238         if (master)
 239                 kinfo->spi_runtime_flags |= QIB_RUNTIME_MASTER;
 240
 241         sz = (ubase_size < sizeof(*kinfo)) ? ubase_size : sizeof(*kinfo);
 242         if (copy_to_user(ubase, kinfo, sz))
 243                 ret = -EFAULT;
 244 bail:
 245         kfree(kinfo);
 246         return ret;
 247 }
 248
 249 /**
 250  * qib_tid_update - update a context TID
 251  * @rcd: the context
 252  * @fp: the qib device file
 253  * @ti: the TID information
 254  *
 255  * The new implementation as of Oct 2004 is that the driver assigns
 256  * the tid and returns it to the caller.   To reduce search time, we
 257  * keep a cursor for each context, walking the shadow tid array to find
 258  * one that's not in use.
 259  *
 260  * For now, if we can't allocate the full list, we fail, although
 261  * in the long run, we'll allocate as many as we can, and the
 262  * caller will deal with that by trying the remaining pages later.
 263  * That means that when we fail, we have to mark the tids as not in
 264  * use again, in our shadow copy.
 265  *
 266  * It's up to the caller to free the tids when they are done.
 267  * We'll unlock the pages as they free them.
 268  *
 269  * Also, right now we are locking one page at a time, but since
 270  * the intended use of this routine is for a single group of
 271  * virtually contiguous pages, that should change to improve
 272  * performance.
 273  */
 274 static int qib_tid_update(struct qib_ctxtdata *rcd, struct file *fp,
 275                           const struct qib_tid_info *ti)
 276 {
 277         int ret = 0, ntids;
 278         u32 tid, ctxttid, cnt, i, tidcnt, tidoff;
 279         u16 *tidlist;
 280         struct qib_devdata *dd = rcd->dd;
 281         u64 physaddr;
 282         unsigned long vaddr;
 283         u64 __iomem *tidbase;
 284         unsigned long tidmap[8];
 285         struct page **pagep = NULL;
 286         unsigned subctxt = subctxt_fp(fp);
 287
 288         if (!dd->pageshadow) {
 289                 ret = -ENOMEM;
 290                 goto done;
 291         }
 292
 293         cnt = ti->tidcnt;
 294         if (!cnt) {
 295                 ret = -EFAULT;
 296                 goto done;
 297         }
 298         ctxttid = rcd->ctxt * dd->rcvtidcnt;
 299         if (!rcd->subctxt_cnt) {
 300                 tidcnt = dd->rcvtidcnt;
 301                 tid = rcd->tidcursor;
 302                 tidoff = 0;
 303         } else if (!subctxt) {
 304                 tidcnt = (dd->rcvtidcnt / rcd->subctxt_cnt) +
 305                          (dd->rcvtidcnt % rcd->subctxt_cnt);
 306                 tidoff = dd->rcvtidcnt - tidcnt;
 307                 ctxttid += tidoff;
 308                 tid = tidcursor_fp(fp);
 309         } else {
 310                 tidcnt = dd->rcvtidcnt / rcd->subctxt_cnt;
 311                 tidoff = tidcnt * (subctxt - 1);
 312                 ctxttid += tidoff;
 313                 tid = tidcursor_fp(fp);
 314         }
 315         if (cnt > tidcnt) {
 316                 /* make sure it all fits in tid_pg_list */
 317                 qib_devinfo(dd->pcidev, "Process tried to allocate %u "
 318                          "TIDs, only trying max (%u)\n", cnt, tidcnt);
 319                 cnt = tidcnt;
 320         }
 321         pagep = (struct page **) rcd->tid_pg_list;
 322         tidlist = (u16 *) &pagep[dd->rcvtidcnt];
 323         pagep += tidoff;
 324         tidlist += tidoff;
 325
 326         memset(tidmap, 0, sizeof(tidmap));
 327         /* before decrement; chip actual # */
 328         ntids = tidcnt;
 329         tidbase = (u64 __iomem *) (((char __iomem *) dd->kregbase) +
 330                                    dd->rcvtidbase +
 331                                    ctxttid * sizeof(*tidbase));
 332
 333         /* virtual address of first page in transfer */
 334         vaddr = ti->tidvaddr;
 335         if (!access_ok(VERIFY_WRITE, (void __user *) vaddr,
 336                        cnt * PAGE_SIZE)) {
 337                 ret = -EFAULT;
 338                 goto done;
 339         }
 340         ret = qib_get_user_pages(vaddr, cnt, pagep);
 341         if (ret) {
 342                 /*
 343                  * if (ret == -EBUSY)
 344                  * We can't continue because the pagep array won't be
 345                  * initialized. This should never happen,
 346                  * unless perhaps the user has mpin'ed the pages
 347                  * themselves.
 348                  */
 349                 qib_devinfo(dd->pcidev,
 350                          "Failed to lock addr %p, %u pages: "
 351                          "errno %d\n", (void *) vaddr, cnt, -ret);
 352                 goto done;
 353         }
 354         for (i = 0; i < cnt; i++, vaddr += PAGE_SIZE) {
 355                 for (; ntids--; tid++) {
 356                         if (tid == tidcnt)
 357                                 tid = 0;
 358                         if (!dd->pageshadow[ctxttid + tid])
 359                                 break;
 360                 }
 361                 if (ntids < 0) {
 362                         /*
 363                          * Oops, wrapped all the way through their TIDs,
 364                          * and didn't have enough free; see comments at
 365                          * start of routine
 366                          */
 367                         i--;    /* last tidlist[i] not filled in */
 368                         ret = -ENOMEM;
 369                         break;
 370                 }
 371                 tidlist[i] = tid + tidoff;
 372                 /* we "know" system pages and TID pages are same size */
 373                 dd->pageshadow[ctxttid + tid] = pagep[i];
 374                 dd->physshadow[ctxttid + tid] =
 375                         qib_map_page(dd->pcidev, pagep[i], 0, PAGE_SIZE,
 376                                      PCI_DMA_FROMDEVICE);
 377                 /*
 378                  * don't need atomic or it's overhead
 379                  */
 380                 __set_bit(tid, tidmap);
 381                 physaddr = dd->physshadow[ctxttid + tid];
 382                 /* PERFORMANCE: below should almost certainly be cached */
 383                 dd->f_put_tid(dd, &tidbase[tid],
 384                                   RCVHQ_RCV_TYPE_EXPECTED, physaddr);
 385                 /*
 386                  * don't check this tid in qib_ctxtshadow, since we
 387                  * just filled it in; start with the next one.
 388                  */
 389                 tid++;
 390         }
 391
 392         if (ret) {
 393                 u32 limit;
 394 cleanup:
 395                 /* jump here if copy out of updated info failed... */
 396                 /* same code that's in qib_free_tid() */
 397                 limit = sizeof(tidmap) * BITS_PER_BYTE;
 398                 if (limit > tidcnt)
 399                         /* just in case size changes in future */
 400                         limit = tidcnt;
 401                 tid = find_first_bit((const unsigned long *)tidmap, limit);
 402                 for (; tid < limit; tid++) {
 403                         if (!test_bit(tid, tidmap))
 404                                 continue;
 405                         if (dd->pageshadow[ctxttid + tid]) {
 406                                 dma_addr_t phys;
 407
 408                                 phys = dd->physshadow[ctxttid + tid];
 409                                 dd->physshadow[ctxttid + tid] = dd->tidinvalid;
 410                                 /* PERFORMANCE: below should almost certainly
 411                                  * be cached
 412                                  */
 413                                 dd->f_put_tid(dd, &tidbase[tid],
 414                                               RCVHQ_RCV_TYPE_EXPECTED,
 415                                               dd->tidinvalid);
 416                                 pci_unmap_page(dd->pcidev, phys, PAGE_SIZE,
 417                                                PCI_DMA_FROMDEVICE);
 418                                 dd->pageshadow[ctxttid + tid] = NULL;
 419                         }
 420                 }
 421                 qib_release_user_pages(pagep, cnt);
 422         } else {
 423                 /*
 424                  * Copy the updated array, with qib_tid's filled in, back
 425                  * to user.  Since we did the copy in already, this "should
 426                  * never fail" If it does, we have to clean up...
 427                  */
 428                 if (copy_to_user((void __user *)
 429                                  (unsigned long) ti->tidlist,
 430                                  tidlist, cnt * sizeof(*tidlist))) {
 431                         ret = -EFAULT;
 432                         goto cleanup;
 433                 }
 434                 if (copy_to_user((void __user *) (unsigned long) ti->tidmap,
 435                                  tidmap, sizeof tidmap)) {
 436                         ret = -EFAULT;
 437                         goto cleanup;
 438                 }
 439                 if (tid == tidcnt)
 440                         tid = 0;
 441                 if (!rcd->subctxt_cnt)
 442                         rcd->tidcursor = tid;
 443                 else
 444                         tidcursor_fp(fp) = tid;
 445         }
 446
 447 done:
 448         return ret;
 449 }
 450
 451 /**
 452  * qib_tid_free - free a context TID
 453  * @rcd: the context
 454  * @subctxt: the subcontext
 455  * @ti: the TID info
 456  *
 457  * right now we are unlocking one page at a time, but since
 458  * the intended use of this routine is for a single group of
 459  * virtually contiguous pages, that should change to improve
 460  * performance.  We check that the TID is in range for this context
 461  * but otherwise don't check validity; if user has an error and
 462  * frees the wrong tid, it's only their own data that can thereby
 463  * be corrupted.  We do check that the TID was in use, for sanity
 464  * We always use our idea of the saved address, not the address that
 465  * they pass in to us.
 466  */
 467 static int qib_tid_free(struct qib_ctxtdata *rcd, unsigned subctxt,
 468                         const struct qib_tid_info *ti)
 469 {
 470         int ret = 0;
 471         u32 tid, ctxttid, cnt, limit, tidcnt;
 472         struct qib_devdata *dd = rcd->dd;
 473         u64 __iomem *tidbase;
 474         unsigned long tidmap[8];
 475
 476         if (!dd->pageshadow) {
 477                 ret = -ENOMEM;
 478                 goto done;
 479         }
 480
 481         if (copy_from_user(tidmap, (void __user *)(unsigned long)ti->tidmap,
 482                            sizeof tidmap)) {
 483                 ret = -EFAULT;
 484                 goto done;
 485         }
 486
 487         ctxttid = rcd->ctxt * dd->rcvtidcnt;
 488         if (!rcd->subctxt_cnt)
 489                 tidcnt = dd->rcvtidcnt;
 490         else if (!subctxt) {
 491                 tidcnt = (dd->rcvtidcnt / rcd->subctxt_cnt) +
 492                          (dd->rcvtidcnt % rcd->subctxt_cnt);
 493                 ctxttid += dd->rcvtidcnt - tidcnt;
 494         } else {
 495                 tidcnt = dd->rcvtidcnt / rcd->subctxt_cnt;
 496                 ctxttid += tidcnt * (subctxt - 1);
 497         }
 498         tidbase = (u64 __iomem *) ((char __iomem *)(dd->kregbase) +
 499                                    dd->rcvtidbase +
 500                                    ctxttid * sizeof(*tidbase));
 501
 502         limit = sizeof(tidmap) * BITS_PER_BYTE;
 503         if (limit > tidcnt)
 504                 /* just in case size changes in future */
 505                 limit = tidcnt;
 506         tid = find_first_bit(tidmap, limit);
 507         for (cnt = 0; tid < limit; tid++) {
 508                 /*
 509                  * small optimization; if we detect a run of 3 or so without
 510                  * any set, use find_first_bit again.  That's mainly to
 511                  * accelerate the case where we wrapped, so we have some at
 512                  * the beginning, and some at the end, and a big gap
 513                  * in the middle.
 514                  */
 515                 if (!test_bit(tid, tidmap))
 516                         continue;
 517                 cnt++;
 518                 if (dd->pageshadow[ctxttid + tid]) {
 519                         struct page *p;
 520                         dma_addr_t phys;
 521
 522                         p = dd->pageshadow[ctxttid + tid];
 523                         dd->pageshadow[ctxttid + tid] = NULL;
 524                         phys = dd->physshadow[ctxttid + tid];
 525                         dd->physshadow[ctxttid + tid] = dd->tidinvalid;
 526                         /* PERFORMANCE: below should almost certainly be
 527                          * cached
 528                          */
 529                         dd->f_put_tid(dd, &tidbase[tid],
 530                                       RCVHQ_RCV_TYPE_EXPECTED, dd->tidinvalid);
 531                         pci_unmap_page(dd->pcidev, phys, PAGE_SIZE,
 532                                        PCI_DMA_FROMDEVICE);
 533                         qib_release_user_pages(&p, 1);
 534                 }
 535         }
 536 done:
 537         return ret;
 538 }
 539
 540 /**
 541  * qib_set_part_key - set a partition key
 542  * @rcd: the context
 543  * @key: the key
 544  *
 545  * We can have up to 4 active at a time (other than the default, which is
 546  * always allowed).  This is somewhat tricky, since multiple contexts may set
 547  * the same key, so we reference count them, and clean up at exit.  All 4
 548  * partition keys are packed into a single qlogic_ib register.  It's an
 549  * error for a process to set the same pkey multiple times.  We provide no
 550  * mechanism to de-allocate a pkey at this time, we may eventually need to
 551  * do that.  I've used the atomic operations, and no locking, and only make
 552  * a single pass through what's available.  This should be more than
 553  * adequate for some time. I'll think about spinlocks or the like if and as
 554  * it's necessary.
 555  */
 556 static int qib_set_part_key(struct qib_ctxtdata *rcd, u16 key)
 557 {
 558         struct qib_pportdata *ppd = rcd->ppd;
 559         int i, any = 0, pidx = -1;
 560         u16 lkey = key & 0x7FFF;
 561         int ret;
 562
 563         if (lkey == (QIB_DEFAULT_P_KEY & 0x7FFF)) {
 564                 /* nothing to do; this key always valid */
 565                 ret = 0;
 566                 goto bail;
 567         }
 568
 569         if (!lkey) {
 570                 ret = -EINVAL;
 571                 goto bail;
 572         }
 573
 574         /*
 575          * Set the full membership bit, because it has to be
 576          * set in the register or the packet, and it seems
 577          * cleaner to set in the register than to force all
 578          * callers to set it.
 579          */
 580         key |= 0x8000;
 581
 582         for (i = 0; i < ARRAY_SIZE(rcd->pkeys); i++) {
 583                 if (!rcd->pkeys[i] && pidx == -1)
 584                         pidx = i;
 585                 if (rcd->pkeys[i] == key) {
 586                         ret = -EEXIST;
 587                         goto bail;
 588                 }
 589         }
 590         if (pidx == -1) {
 591                 ret = -EBUSY;
 592                 goto bail;
 593         }
 594         for (any = i = 0; i < ARRAY_SIZE(ppd->pkeys); i++) {
 595                 if (!ppd->pkeys[i]) {
 596                         any++;
 597                         continue;
 598                 }
 599                 if (ppd->pkeys[i] == key) {
 600                         atomic_t *pkrefs = &ppd->pkeyrefs[i];
 601
 602                         if (atomic_inc_return(pkrefs) > 1) {
 603                                 rcd->pkeys[pidx] = key;
 604                                 ret = 0;
 605                                 goto bail;
 606                         } else {
 607                                 /*
 608                                  * lost race, decrement count, catch below
 609                                  */
 610                                 atomic_dec(pkrefs);
 611                                 any++;
 612                         }
 613                 }
 614                 if ((ppd->pkeys[i] & 0x7FFF) == lkey) {
 615                         /*
 616                          * It makes no sense to have both the limited and
 617                          * full membership PKEY set at the same time since
 618                          * the unlimited one will disable the limited one.
 619                          */
 620                         ret = -EEXIST;
 621                         goto bail;
 622                 }
 623         }
 624         if (!any) {
 625                 ret = -EBUSY;
 626                 goto bail;
 627         }
 628         for (any = i = 0; i < ARRAY_SIZE(ppd->pkeys); i++) {
 629                 if (!ppd->pkeys[i] &&
 630                     atomic_inc_return(&ppd->pkeyrefs[i]) == 1) {
 631                         rcd->pkeys[pidx] = key;
 632                         ppd->pkeys[i] = key;
 633                         (void) ppd->dd->f_set_ib_cfg(ppd, QIB_IB_CFG_PKEYS, 0);
 634                         ret = 0;
 635                         goto bail;
 636                 }
 637         }
 638         ret = -EBUSY;
 639
 640 bail:
 641         return ret;
 642 }
 643
 644 /**
 645  * qib_manage_rcvq - manage a context's receive queue
 646  * @rcd: the context
 647  * @subctxt: the subcontext
 648  * @start_stop: action to carry out
 649  *
 650  * start_stop == 0 disables receive on the context, for use in queue
 651  * overflow conditions.  start_stop==1 re-enables, to be used to
 652  * re-init the software copy of the head register
 653  */
 654 static int qib_manage_rcvq(struct qib_ctxtdata *rcd, unsigned subctxt,
 655                            int start_stop)
 656 {
 657         struct qib_devdata *dd = rcd->dd;
 658         unsigned int rcvctrl_op;
 659
 660         if (subctxt)
 661                 goto bail;
 662         /* atomically clear receive enable ctxt. */
 663         if (start_stop) {
 664                 /*
 665                  * On enable, force in-memory copy of the tail register to
 666                  * 0, so that protocol code doesn't have to worry about
 667                  * whether or not the chip has yet updated the in-memory
 668                  * copy or not on return from the system call. The chip
 669                  * always resets it's tail register back to 0 on a
 670                  * transition from disabled to enabled.
 671                  */
 672                 if (rcd->rcvhdrtail_kvaddr)
 673                         qib_clear_rcvhdrtail(rcd);
 674                 rcvctrl_op = QIB_RCVCTRL_CTXT_ENB;
 675         } else
 676                 rcvctrl_op = QIB_RCVCTRL_CTXT_DIS;
 677         dd->f_rcvctrl(rcd->ppd, rcvctrl_op, rcd->ctxt);
 678         /* always; new head should be equal to new tail; see above */
 679 bail:
 680         return 0;
 681 }
 682
 683 static void qib_clean_part_key(struct qib_ctxtdata *rcd,
 684                                struct qib_devdata *dd)
 685 {
 686         int i, j, pchanged = 0;
 687         u64 oldpkey;
 688         struct qib_pportdata *ppd = rcd->ppd;
 689
 690         /* for debugging only */
 691         oldpkey = (u64) ppd->pkeys[0] |
 692                 ((u64) ppd->pkeys[1] << 16) |
 693                 ((u64) ppd->pkeys[2] << 32) |
 694                 ((u64) ppd->pkeys[3] << 48);
 695
 696         for (i = 0; i < ARRAY_SIZE(rcd->pkeys); i++) {
 697                 if (!rcd->pkeys[i])
 698                         continue;
 699                 for (j = 0; j < ARRAY_SIZE(ppd->pkeys); j++) {
 700                         /* check for match independent of the global bit */
 701                         if ((ppd->pkeys[j] & 0x7fff) !=
 702                             (rcd->pkeys[i] & 0x7fff))
 703                                 continue;
 704                         if (atomic_dec_and_test(&ppd->pkeyrefs[j])) {
 705                                 ppd->pkeys[j] = 0;
 706                                 pchanged++;
 707                         }
 708                         break;
 709                 }
 710                 rcd->pkeys[i] = 0;
 711         }
 712         if (pchanged)
 713                 (void) ppd->dd->f_set_ib_cfg(ppd, QIB_IB_CFG_PKEYS, 0);
 714 }
 715
 716 /* common code for the mappings on dma_alloc_coherent mem */
 717 static int qib_mmap_mem(struct vm_area_struct *vma, struct qib_ctxtdata *rcd,
 718                         unsigned len, void *kvaddr, u32 write_ok, char *what)
 719 {
 720         struct qib_devdata *dd = rcd->dd;
 721         unsigned long pfn;
 722         int ret;
 723
 724         if ((vma->vm_end - vma->vm_start) > len) {
 725                 qib_devinfo(dd->pcidev,
 726                          "FAIL on %s: len %lx > %x\n", what,
 727                          vma->vm_end - vma->vm_start, len);
 728                 ret = -EFAULT;
 729                 goto bail;
 730         }
 731
 732         /*
 733          * shared context user code requires rcvhdrq mapped r/w, others
 734          * only allowed readonly mapping.
 735          */
 736         if (!write_ok) {
 737                 if (vma->vm_flags & VM_WRITE) {
 738                         qib_devinfo(dd->pcidev,
 739                                  "%s must be mapped readonly\n", what);
 740                         ret = -EPERM;
 741                         goto bail;
 742                 }
 743
 744                 /* don't allow them to later change with mprotect */
 745                 vma->vm_flags &= ~VM_MAYWRITE;
 746         }
 747
 748         pfn = virt_to_phys(kvaddr) >> PAGE_SHIFT;
 749         ret = remap_pfn_range(vma, vma->vm_start, pfn,
 750                               len, vma->vm_page_prot);
 751         if (ret)
 752                 qib_devinfo(dd->pcidev, "%s ctxt%u mmap of %lx, %x "
 753                          "bytes failed: %d\n", what, rcd->ctxt,
 754                          pfn, len, ret);
 755 bail:
 756         return ret;
 757 }
 758
 759 static int mmap_ureg(struct vm_area_struct *vma, struct qib_devdata *dd,
 760                      u64 ureg)
 761 {
 762         unsigned long phys;
 763         unsigned long sz;
 764         int ret;
 765
 766         /*
 767          * This is real hardware, so use io_remap.  This is the mechanism
 768          * for the user process to update the head registers for their ctxt
 769          * in the chip.
 770          */
 771         sz = dd->flags & QIB_HAS_HDRSUPP ? 2 * PAGE_SIZE : PAGE_SIZE;
 772         if ((vma->vm_end - vma->vm_start) > sz) {
 773                 qib_devinfo(dd->pcidev, "FAIL mmap userreg: reqlen "
 774                          "%lx > PAGE\n", vma->vm_end - vma->vm_start);
 775                 ret = -EFAULT;
 776         } else {
 777                 phys = dd->physaddr + ureg;
 778                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 779
 780                 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
 781                 ret = io_remap_pfn_range(vma, vma->vm_start,
 782                                          phys >> PAGE_SHIFT,
 783                                          vma->vm_end - vma->vm_start,
 784                                          vma->vm_page_prot);
 785         }
 786         return ret;
 787 }
 788
 789 static int mmap_piobufs(struct vm_area_struct *vma,
 790                         struct qib_devdata *dd,
 791                         struct qib_ctxtdata *rcd,
 792                         unsigned piobufs, unsigned piocnt)
 793 {
 794         unsigned long phys;
 795         int ret;
 796
 797         /*
 798          * When we map the PIO buffers in the chip, we want to map them as
 799          * writeonly, no read possible; unfortunately, x86 doesn't allow
 800          * for this in hardware, but we still prevent users from asking
 801          * for it.
 802          */
 803         if ((vma->vm_end - vma->vm_start) > (piocnt * dd->palign)) {
 804                 qib_devinfo(dd->pcidev, "FAIL mmap piobufs: "
 805                          "reqlen %lx > PAGE\n",
 806                          vma->vm_end - vma->vm_start);
 807                 ret = -EINVAL;
 808                 goto bail;
 809         }
 810
 811         phys = dd->physaddr + piobufs;
 812
 813 #if defined(__powerpc__)
 814         /* There isn't a generic way to specify writethrough mappings */
 815         pgprot_val(vma->vm_page_prot) |= _PAGE_NO_CACHE;
 816         pgprot_val(vma->vm_page_prot) |= _PAGE_WRITETHRU;
 817         pgprot_val(vma->vm_page_prot) &= ~_PAGE_GUARDED;
 818 #endif
 819
 820         /*
 821          * don't allow them to later change to readable with mprotect (for when
 822          * not initially mapped readable, as is normally the case)
 823          */
 824         vma->vm_flags &= ~VM_MAYREAD;
 825         vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
 826
 827         if (qib_wc_pat)
 828                 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
 829
 830         ret = io_remap_pfn_range(vma, vma->vm_start, phys >> PAGE_SHIFT,
 831                                  vma->vm_end - vma->vm_start,
 832                                  vma->vm_page_prot);
 833 bail:
 834         return ret;
 835 }
 836
 837 static int mmap_rcvegrbufs(struct vm_area_struct *vma,
 838                            struct qib_ctxtdata *rcd)
 839 {
 840         struct qib_devdata *dd = rcd->dd;
 841         unsigned long start, size;
 842         size_t total_size, i;
 843         unsigned long pfn;
 844         int ret;
 845
 846         size = rcd->rcvegrbuf_size;
 847         total_size = rcd->rcvegrbuf_chunks * size;
 848         if ((vma->vm_end - vma->vm_start) > total_size) {
 849                 qib_devinfo(dd->pcidev, "FAIL on egr bufs: "
 850                          "reqlen %lx > actual %lx\n",
 851                          vma->vm_end - vma->vm_start,
 852                          (unsigned long) total_size);
 853                 ret = -EINVAL;
 854                 goto bail;
 855         }
 856
 857         if (vma->vm_flags & VM_WRITE) {
 858                 qib_devinfo(dd->pcidev, "Can't map eager buffers as "
 859                          "writable (flags=%lx)\n", vma->vm_flags);
 860                 ret = -EPERM;
 861                 goto bail;
 862         }
 863         /* don't allow them to later change to writeable with mprotect */
 864         vma->vm_flags &= ~VM_MAYWRITE;
 865
 866         start = vma->vm_start;
 867
 868         for (i = 0; i < rcd->rcvegrbuf_chunks; i++, start += size) {
 869                 pfn = virt_to_phys(rcd->rcvegrbuf[i]) >> PAGE_SHIFT;
 870                 ret = remap_pfn_range(vma, start, pfn, size,
 871                                       vma->vm_page_prot);
 872                 if (ret < 0)
 873                         goto bail;
 874         }
 875         ret = 0;
 876
 877 bail:
 878         return ret;
 879 }
 880
 881 /*
 882  * qib_file_vma_fault - handle a VMA page fault.
 883  */
 884 static int qib_file_vma_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 885 {
 886         struct page *page;
 887
 888         page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
 889         if (!page)
 890                 return VM_FAULT_SIGBUS;
 891
 892         get_page(page);
 893         vmf->page = page;
 894
 895         return 0;
 896 }
 897
 898 static struct vm_operations_struct qib_file_vm_ops = {
 899         .fault = qib_file_vma_fault,
 900 };
 901
 902 static int mmap_kvaddr(struct vm_area_struct *vma, u64 pgaddr,
 903                        struct qib_ctxtdata *rcd, unsigned subctxt)
 904 {
 905         struct qib_devdata *dd = rcd->dd;
 906         unsigned subctxt_cnt;
 907         unsigned long len;
 908         void *addr;
 909         size_t size;
 910         int ret = 0;
 911
 912         subctxt_cnt = rcd->subctxt_cnt;
 913         size = rcd->rcvegrbuf_chunks * rcd->rcvegrbuf_size;
 914
 915         /*
 916          * Each process has all the subctxt uregbase, rcvhdrq, and
 917          * rcvegrbufs mmapped - as an array for all the processes,
 918          * and also separately for this process.
 919          */
 920         if (pgaddr == cvt_kvaddr(rcd->subctxt_uregbase)) {
 921                 addr = rcd->subctxt_uregbase;
 922                 size = PAGE_SIZE * subctxt_cnt;
 923         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_rcvhdr_base)) {
 924                 addr = rcd->subctxt_rcvhdr_base;
 925                 size = rcd->rcvhdrq_size * subctxt_cnt;
 926         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_rcvegrbuf)) {
 927                 addr = rcd->subctxt_rcvegrbuf;
 928                 size *= subctxt_cnt;
 929         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_uregbase +
 930                                         PAGE_SIZE * subctxt)) {
 931                 addr = rcd->subctxt_uregbase + PAGE_SIZE * subctxt;
 932                 size = PAGE_SIZE;
 933         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_rcvhdr_base +
 934                                         rcd->rcvhdrq_size * subctxt)) {
 935                 addr = rcd->subctxt_rcvhdr_base +
 936                         rcd->rcvhdrq_size * subctxt;
 937                 size = rcd->rcvhdrq_size;
 938         } else if (pgaddr == cvt_kvaddr(&rcd->user_event_mask[subctxt])) {
 939                 addr = rcd->user_event_mask;
 940                 size = PAGE_SIZE;
 941         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_rcvegrbuf +
 942                                         size * subctxt)) {
 943                 addr = rcd->subctxt_rcvegrbuf + size * subctxt;
 944                 /* rcvegrbufs are read-only on the slave */
 945                 if (vma->vm_flags & VM_WRITE) {
 946                         qib_devinfo(dd->pcidev,
 947                                  "Can't map eager buffers as "
 948                                  "writable (flags=%lx)\n", vma->vm_flags);
 949                         ret = -EPERM;
 950                         goto bail;
 951                 }
 952                 /*
 953                  * Don't allow permission to later change to writeable
 954                  * with mprotect.
 955                  */
 956                 vma->vm_flags &= ~VM_MAYWRITE;
 957         } else
 958                 goto bail;
 959         len = vma->vm_end - vma->vm_start;
 960         if (len > size) {
 961                 ret = -EINVAL;
 962                 goto bail;
 963         }
 964
 965         vma->vm_pgoff = (unsigned long) addr >> PAGE_SHIFT;
 966         vma->vm_ops = &qib_file_vm_ops;
 967         vma->vm_flags |= VM_RESERVED | VM_DONTEXPAND;
 968         ret = 1;
 969
 970 bail:
 971         return ret;
 972 }
 973
 974 /**
 975  * qib_mmapf - mmap various structures into user space
 976  * @fp: the file pointer
 977  * @vma: the VM area
 978  *
 979  * We use this to have a shared buffer between the kernel and the user code
 980  * for the rcvhdr queue, egr buffers, and the per-context user regs and pio
 981  * buffers in the chip.  We have the open and close entries so we can bump
 982  * the ref count and keep the driver from being unloaded while still mapped.
 983  */
 984 static int qib_mmapf(struct file *fp, struct vm_area_struct *vma)
 985 {
 986         struct qib_ctxtdata *rcd;
 987         struct qib_devdata *dd;
 988         u64 pgaddr, ureg;
 989         unsigned piobufs, piocnt;
 990         int ret, match = 1;
 991
 992         rcd = ctxt_fp(fp);
 993         if (!rcd || !(vma->vm_flags & VM_SHARED)) {
 994                 ret = -EINVAL;
 995                 goto bail;
 996         }
 997         dd = rcd->dd;
 998
 999         /*
1000          * This is the qib_do_user_init() code, mapping the shared buffers
1001          * and per-context user registers into the user process. The address
1002          * referred to by vm_pgoff is the file offset passed via mmap().
1003          * For shared contexts, this is the kernel vmalloc() address of the
1004          * pages to share with the master.
1005          * For non-shared or master ctxts, this is a physical address.
1006          * We only do one mmap for each space mapped.
1007          */
1008         pgaddr = vma->vm_pgoff << PAGE_SHIFT;
1009
1010         /*
1011          * Check for 0 in case one of the allocations failed, but user
1012          * called mmap anyway.
1013          */
1014         if (!pgaddr)  {
1015                 ret = -EINVAL;
1016                 goto bail;
1017         }
1018
1019         /*
1020          * Physical addresses must fit in 40 bits for our hardware.
1021          * Check for kernel virtual addresses first, anything else must
1022          * match a HW or memory address.
1023          */
1024         ret = mmap_kvaddr(vma, pgaddr, rcd, subctxt_fp(fp));
1025         if (ret) {
1026                 if (ret > 0)
1027                         ret = 0;
1028                 goto bail;
1029         }
1030
1031         ureg = dd->uregbase + dd->ureg_align * rcd->ctxt;
1032         if (!rcd->subctxt_cnt) {
1033                 /* ctxt is not shared */
1034                 piocnt = rcd->piocnt;
1035                 piobufs = rcd->piobufs;
1036         } else if (!subctxt_fp(fp)) {
1037                 /* caller is the master */
1038                 piocnt = (rcd->piocnt / rcd->subctxt_cnt) +
1039                          (rcd->piocnt % rcd->subctxt_cnt);
1040                 piobufs = rcd->piobufs +
1041                         dd->palign * (rcd->piocnt - piocnt);
1042         } else {
1043                 unsigned slave = subctxt_fp(fp) - 1;
1044
1045                 /* caller is a slave */
1046                 piocnt = rcd->piocnt / rcd->subctxt_cnt;
1047                 piobufs = rcd->piobufs + dd->palign * piocnt * slave;
1048         }
1049
1050         if (pgaddr == ureg)
1051                 ret = mmap_ureg(vma, dd, ureg);
1052         else if (pgaddr == piobufs)
1053                 ret = mmap_piobufs(vma, dd, rcd, piobufs, piocnt);
1054         else if (pgaddr == dd->pioavailregs_phys)
1055                 /* in-memory copy of pioavail registers */
1056                 ret = qib_mmap_mem(vma, rcd, PAGE_SIZE,
1057                                    (void *) dd->pioavailregs_dma, 0,
1058                                    "pioavail registers");
1059         else if (pgaddr == rcd->rcvegr_phys)
1060                 ret = mmap_rcvegrbufs(vma, rcd);
1061         else if (pgaddr == (u64) rcd->rcvhdrq_phys)
1062                 /*
1063                  * The rcvhdrq itself; multiple pages, contiguous
1064                  * from an i/o perspective.  Shared contexts need
1065                  * to map r/w, so we allow writing.
1066                  */
1067                 ret = qib_mmap_mem(vma, rcd, rcd->rcvhdrq_size,
1068                                    rcd->rcvhdrq, 1, "rcvhdrq");
1069         else if (pgaddr == (u64) rcd->rcvhdrqtailaddr_phys)
1070                 /* in-memory copy of rcvhdrq tail register */
1071                 ret = qib_mmap_mem(vma, rcd, PAGE_SIZE,
1072                                    rcd->rcvhdrtail_kvaddr, 0,
1073                                    "rcvhdrq tail");
1074         else
1075                 match = 0;
1076         if (!match)
1077                 ret = -EINVAL;
1078
1079         vma->vm_private_data = NULL;
1080
1081         if (ret < 0)
1082                 qib_devinfo(dd->pcidev,
1083                          "mmap Failure %d: off %llx len %lx\n",
1084                          -ret, (unsigned long long)pgaddr,
1085                          vma->vm_end - vma->vm_start);
1086 bail:
1087         return ret;
1088 }
1089
1090 static unsigned int qib_poll_urgent(struct qib_ctxtdata *rcd,
1091                                     struct file *fp,
1092                                     struct poll_table_struct *pt)
1093 {
1094         struct qib_devdata *dd = rcd->dd;
1095         unsigned pollflag;
1096
1097         poll_wait(fp, &rcd->wait, pt);
1098
1099         spin_lock_irq(&dd->uctxt_lock);
1100         if (rcd->urgent != rcd->urgent_poll) {
1101                 pollflag = POLLIN | POLLRDNORM;
1102                 rcd->urgent_poll = rcd->urgent;
1103         } else {
1104                 pollflag = 0;
1105                 set_bit(QIB_CTXT_WAITING_URG, &rcd->flag);
1106         }
1107         spin_unlock_irq(&dd->uctxt_lock);
1108
1109         return pollflag;
1110 }
1111
1112 static unsigned int qib_poll_next(struct qib_ctxtdata *rcd,
1113                                   struct file *fp,
1114                                   struct poll_table_struct *pt)
1115 {
1116         struct qib_devdata *dd = rcd->dd;
1117         unsigned pollflag;
1118
1119         poll_wait(fp, &rcd->wait, pt);
1120
1121         spin_lock_irq(&dd->uctxt_lock);
1122         if (dd->f_hdrqempty(rcd)) {
1123                 set_bit(QIB_CTXT_WAITING_RCV, &rcd->flag);
1124                 dd->f_rcvctrl(rcd->ppd, QIB_RCVCTRL_INTRAVAIL_ENB, rcd->ctxt);
1125                 pollflag = 0;
1126         } else
1127                 pollflag = POLLIN | POLLRDNORM;
1128         spin_unlock_irq(&dd->uctxt_lock);
1129
1130         return pollflag;
1131 }
1132
1133 static unsigned int qib_poll(struct file *fp, struct poll_table_struct *pt)
1134 {
1135         struct qib_ctxtdata *rcd;
1136         unsigned pollflag;
1137
1138         rcd = ctxt_fp(fp);
1139         if (!rcd)
1140                 pollflag = POLLERR;
1141         else if (rcd->poll_type == QIB_POLL_TYPE_URGENT)
1142                 pollflag = qib_poll_urgent(rcd, fp, pt);
1143         else  if (rcd->poll_type == QIB_POLL_TYPE_ANYRCV)
1144                 pollflag = qib_poll_next(rcd, fp, pt);
1145         else /* invalid */
1146                 pollflag = POLLERR;
1147
1148         return pollflag;
1149 }
1150
1151 /*
1152  * Check that userland and driver are compatible for subcontexts.
1153  */
1154 static int qib_compatible_subctxts(int user_swmajor, int user_swminor)
1155 {
1156         /* this code is written long-hand for clarity */
1157         if (QIB_USER_SWMAJOR != user_swmajor) {
1158                 /* no promise of compatibility if major mismatch */
1159                 return 0;
1160         }
1161         if (QIB_USER_SWMAJOR == 1) {
1162                 switch (QIB_USER_SWMINOR) {
1163                 case 0:
1164                 case 1:
1165                 case 2:
1166                         /* no subctxt implementation so cannot be compatible */
1167                         return 0;
1168                 case 3:
1169                         /* 3 is only compatible with itself */
1170                         return user_swminor == 3;
1171                 default:
1172                         /* >= 4 are compatible (or are expected to be) */
1173                         return user_swminor >= 4;
1174                 }
1175         }
1176         /* make no promises yet for future major versions */
1177         return 0;
1178 }
1179
1180 static int init_subctxts(struct qib_devdata *dd,
1181                          struct qib_ctxtdata *rcd,
1182                          const struct qib_user_info *uinfo)
1183 {
1184         int ret = 0;
1185         unsigned num_subctxts;
1186         size_t size;
1187
1188         /*
1189          * If the user is requesting zero subctxts,
1190          * skip the subctxt allocation.
1191          */
1192         if (uinfo->spu_subctxt_cnt <= 0)
1193                 goto bail;
1194         num_subctxts = uinfo->spu_subctxt_cnt;
1195
1196         /* Check for subctxt compatibility */
1197         if (!qib_compatible_subctxts(uinfo->spu_userversion >> 16,
1198                 uinfo->spu_userversion & 0xffff)) {
1199                 qib_devinfo(dd->pcidev,
1200                          "Mismatched user version (%d.%d) and driver "
1201                          "version (%d.%d) while context sharing. Ensure "
1202                          "that driver and library are from the same "
1203                          "release.\n",
1204                          (int) (uinfo->spu_userversion >> 16),
1205                          (int) (uinfo->spu_userversion & 0xffff),
1206                          QIB_USER_SWMAJOR, QIB_USER_SWMINOR);
1207                 goto bail;
1208         }
1209         if (num_subctxts > QLOGIC_IB_MAX_SUBCTXT) {
1210                 ret = -EINVAL;
1211                 goto bail;
1212         }
1213
1214         rcd->subctxt_uregbase = vmalloc_user(PAGE_SIZE * num_subctxts);
1215         if (!rcd->subctxt_uregbase) {
1216                 ret = -ENOMEM;
1217                 goto bail;
1218         }
1219         /* Note: rcd->rcvhdrq_size isn't initialized yet. */
1220         size = ALIGN(dd->rcvhdrcnt * dd->rcvhdrentsize *
1221                      sizeof(u32), PAGE_SIZE) * num_subctxts;
1222         rcd->subctxt_rcvhdr_base = vmalloc_user(size);
1223         if (!rcd->subctxt_rcvhdr_base) {
1224                 ret = -ENOMEM;
1225                 goto bail_ureg;
1226         }
1227
1228         rcd->subctxt_rcvegrbuf = vmalloc_user(rcd->rcvegrbuf_chunks *
1229                                               rcd->rcvegrbuf_size *
1230                                               num_subctxts);
1231         if (!rcd->subctxt_rcvegrbuf) {
1232                 ret = -ENOMEM;
1233                 goto bail_rhdr;
1234         }
1235
1236         rcd->subctxt_cnt = uinfo->spu_subctxt_cnt;
1237         rcd->subctxt_id = uinfo->spu_subctxt_id;
1238         rcd->active_slaves = 1;
1239         rcd->redirect_seq_cnt = 1;
1240         set_bit(QIB_CTXT_MASTER_UNINIT, &rcd->flag);
1241         goto bail;
1242
1243 bail_rhdr:
1244         vfree(rcd->subctxt_rcvhdr_base);
1245 bail_ureg:
1246         vfree(rcd->subctxt_uregbase);
1247         rcd->subctxt_uregbase = NULL;
1248 bail:
1249         return ret;
1250 }
1251
1252 static int setup_ctxt(struct qib_pportdata *ppd, int ctxt,
1253                       struct file *fp, const struct qib_user_info *uinfo)
1254 {
1255         struct qib_devdata *dd = ppd->dd;
1256         struct qib_ctxtdata *rcd;
1257         void *ptmp = NULL;
1258         int ret;
1259
1260         rcd = qib_create_ctxtdata(ppd, ctxt);
1261
1262         /*
1263          * Allocate memory for use in qib_tid_update() at open to
1264          * reduce cost of expected send setup per message segment
1265          */
1266         if (rcd)
1267                 ptmp = kmalloc(dd->rcvtidcnt * sizeof(u16) +
1268                                dd->rcvtidcnt * sizeof(struct page **),
1269                                GFP_KERNEL);
1270
1271         if (!rcd || !ptmp) {
1272                 qib_dev_err(dd, "Unable to allocate ctxtdata "
1273                             "memory, failing open\n");
1274                 ret = -ENOMEM;
1275                 goto bailerr;
1276         }
1277         rcd->userversion = uinfo->spu_userversion;
1278         ret = init_subctxts(dd, rcd, uinfo);
1279         if (ret)
1280                 goto bailerr;
1281         rcd->tid_pg_list = ptmp;
1282         rcd->pid = current->pid;
1283         init_waitqueue_head(&dd->rcd[ctxt]->wait);
1284         strlcpy(rcd->comm, current->comm, sizeof(rcd->comm));
1285         ctxt_fp(fp) = rcd;
1286         qib_stats.sps_ctxts++;
1287         ret = 0;
1288         goto bail;
1289
1290 bailerr:
1291         dd->rcd[ctxt] = NULL;
1292         kfree(rcd);
1293         kfree(ptmp);
1294 bail:
1295         return ret;
1296 }
1297
1298 static inline int usable(struct qib_pportdata *ppd)
1299 {
1300         struct qib_devdata *dd = ppd->dd;
1301
1302         return dd && (dd->flags & QIB_PRESENT) && dd->kregbase && ppd->lid &&
1303                 (ppd->lflags & QIBL_LINKACTIVE);
1304 }
1305
1306 /*
1307  * Select a context on the given device, either using a requested port
1308  * or the port based on the context number.
1309  */
1310 static int choose_port_ctxt(struct file *fp, struct qib_devdata *dd, u32 port,
1311                             const struct qib_user_info *uinfo)
1312 {
1313         struct qib_pportdata *ppd = NULL;
1314         int ret, ctxt;
1315
1316         if (port) {
1317                 if (!usable(dd->pport + port - 1)) {
1318                         ret = -ENETDOWN;
1319                         goto done;
1320                 } else
1321                         ppd = dd->pport + port - 1;
1322         }
1323         for (ctxt = dd->first_user_ctxt; ctxt < dd->cfgctxts && dd->rcd[ctxt];
1324              ctxt++)
1325                 ;
1326         if (ctxt == dd->cfgctxts) {
1327                 ret = -EBUSY;
1328                 goto done;
1329         }
1330         if (!ppd) {
1331                 u32 pidx = ctxt % dd->num_pports;
1332                 if (usable(dd->pport + pidx))
1333                         ppd = dd->pport + pidx;
1334                 else {
1335                         for (pidx = 0; pidx < dd->num_pports && !ppd;
1336                              pidx++)
1337                                 if (usable(dd->pport + pidx))
1338                                         ppd = dd->pport + pidx;
1339                 }
1340         }
1341         ret = ppd ? setup_ctxt(ppd, ctxt, fp, uinfo) : -ENETDOWN;
1342 done:
1343         return ret;
1344 }
1345
1346 static int find_free_ctxt(int unit, struct file *fp,
1347                           const struct qib_user_info *uinfo)
1348 {
1349         struct qib_devdata *dd = qib_lookup(unit);
1350         int ret;
1351
1352         if (!dd || (uinfo->spu_port && uinfo->spu_port > dd->num_pports))
1353                 ret = -ENODEV;
1354         else
1355                 ret = choose_port_ctxt(fp, dd, uinfo->spu_port, uinfo);
1356
1357         return ret;
1358 }
1359
1360 static int get_a_ctxt(struct file *fp, const struct qib_user_info *uinfo,
1361                       unsigned alg)
1362 {
1363         struct qib_devdata *udd = NULL;
1364         int ret = 0, devmax, npresent, nup, ndev, dusable = 0, i;
1365         u32 port = uinfo->spu_port, ctxt;
1366
1367         devmax = qib_count_units(&npresent, &nup);
1368         if (!npresent) {
1369                 ret = -ENXIO;
1370                 goto done;
1371         }
1372         if (nup == 0) {
1373                 ret = -ENETDOWN;
1374                 goto done;
1375         }
1376
1377         if (alg == QIB_PORT_ALG_ACROSS) {
1378                 unsigned inuse = ~0U;
1379                 /* find device (with ACTIVE ports) with fewest ctxts in use */
1380                 for (ndev = 0; ndev < devmax; ndev++) {
1381                         struct qib_devdata *dd = qib_lookup(ndev);
1382                         unsigned cused = 0, cfree = 0, pusable = 0;
1383                         if (!dd)
1384                                 continue;
1385                         if (port && port <= dd->num_pports &&
1386                             usable(dd->pport + port - 1))
1387                                 pusable = 1;
1388                         else
1389                                 for (i = 0; i < dd->num_pports; i++)
1390                                         if (usable(dd->pport + i))
1391                                                 pusable++;
1392                         if (!pusable)
1393                                 continue;
1394                         for (ctxt = dd->first_user_ctxt; ctxt < dd->cfgctxts;
1395                              ctxt++)
1396                                 if (dd->rcd[ctxt])
1397                                         cused++;
1398                                 else
1399                                         cfree++;
1400                         if (pusable && cfree && cused < inuse) {
1401                                 udd = dd;
1402                                 inuse = cused;
1403                         }
1404                 }
1405                 if (udd) {
1406                         ret = choose_port_ctxt(fp, udd, port, uinfo);
1407                         goto done;
1408                 }
1409         } else {
1410                 for (ndev = 0; ndev < devmax; ndev++) {
1411                         struct qib_devdata *dd = qib_lookup(ndev);
1412                         if (dd) {
1413                                 ret = choose_port_ctxt(fp, dd, port, uinfo);
1414                                 if (!ret)
1415                                         goto done;
1416                                 if (ret == -EBUSY)
1417                                         dusable++;
1418                         }
1419                 }
1420         }
1421         ret = dusable ? -EBUSY : -ENETDOWN;
1422
1423 done:
1424         return ret;
1425 }
1426
1427 static int find_shared_ctxt(struct file *fp,
1428                             const struct qib_user_info *uinfo)
1429 {
1430         int devmax, ndev, i;
1431         int ret = 0;
1432
1433         devmax = qib_count_units(NULL, NULL);
1434
1435         for (ndev = 0; ndev < devmax; ndev++) {
1436                 struct qib_devdata *dd = qib_lookup(ndev);
1437
1438                 /* device portion of usable() */
1439                 if (!(dd && (dd->flags & QIB_PRESENT) && dd->kregbase))
1440                         continue;
1441                 for (i = dd->first_user_ctxt; i < dd->cfgctxts; i++) {
1442                         struct qib_ctxtdata *rcd = dd->rcd[i];
1443
1444                         /* Skip ctxts which are not yet open */
1445                         if (!rcd || !rcd->cnt)
1446                                 continue;
1447                         /* Skip ctxt if it doesn't match the requested one */
1448                         if (rcd->subctxt_id != uinfo->spu_subctxt_id)
1449                                 continue;
1450                         /* Verify the sharing process matches the master */
1451                         if (rcd->subctxt_cnt != uinfo->spu_subctxt_cnt ||
1452                             rcd->userversion != uinfo->spu_userversion ||
1453                             rcd->cnt >= rcd->subctxt_cnt) {
1454                                 ret = -EINVAL;
1455                                 goto done;
1456                         }
1457                         ctxt_fp(fp) = rcd;
1458                         subctxt_fp(fp) = rcd->cnt++;
1459                         rcd->subpid[subctxt_fp(fp)] = current->pid;
1460                         tidcursor_fp(fp) = 0;
1461                         rcd->active_slaves |= 1 << subctxt_fp(fp);
1462                         ret = 1;
1463                         goto done;
1464                 }
1465         }
1466
1467 done:
1468         return ret;
1469 }
1470
1471 static int qib_open(struct inode *in, struct file *fp)
1472 {
1473         /* The real work is performed later in qib_assign_ctxt() */
1474         fp->private_data = kzalloc(sizeof(struct qib_filedata), GFP_KERNEL);
1475         if (fp->private_data) /* no cpu affinity by default */
1476                 ((struct qib_filedata *)fp->private_data)->rec_cpu_num = -1;
1477         return fp->private_data ? 0 : -ENOMEM;
1478 }
1479
1480 /*
1481  * Get ctxt early, so can set affinity prior to memory allocation.
1482  */
1483 static int qib_assign_ctxt(struct file *fp, const struct qib_user_info *uinfo)
1484 {
1485         int ret;
1486         int i_minor;
1487         unsigned swmajor, swminor, alg = QIB_PORT_ALG_ACROSS;
1488
1489         /* Check to be sure we haven't already initialized this file */
1490         if (ctxt_fp(fp)) {
1491                 ret = -EINVAL;
1492                 goto done;
1493         }
1494
1495         /* for now, if major version is different, bail */
1496         swmajor = uinfo->spu_userversion >> 16;
1497         if (swmajor != QIB_USER_SWMAJOR) {
1498                 ret = -ENODEV;
1499                 goto done;
1500         }
1501
1502         swminor = uinfo->spu_userversion & 0xffff;
1503
1504         if (swminor >= 11 && uinfo->spu_port_alg < QIB_PORT_ALG_COUNT)
1505                 alg = uinfo->spu_port_alg;
1506
1507         mutex_lock(&qib_mutex);
1508
1509         if (qib_compatible_subctxts(swmajor, swminor) &&
1510             uinfo->spu_subctxt_cnt) {
1511                 ret = find_shared_ctxt(fp, uinfo);
1512                 if (ret) {
1513                         if (ret > 0)
1514                                 ret = 0;
1515                         goto done_chk_sdma;
1516                 }
1517         }
1518
1519         i_minor = iminor(fp->f_dentry->d_inode) - QIB_USER_MINOR_BASE;
1520         if (i_minor)
1521                 ret = find_free_ctxt(i_minor - 1, fp, uinfo);
1522         else
1523                 ret = get_a_ctxt(fp, uinfo, alg);
1524
1525 done_chk_sdma:
1526         if (!ret) {
1527                 struct qib_filedata *fd = fp->private_data;
1528                 const struct qib_ctxtdata *rcd = fd->rcd;
1529                 const struct qib_devdata *dd = rcd->dd;
1530                 unsigned int weight;
1531
1532                 if (dd->flags & QIB_HAS_SEND_DMA) {
1533                         fd->pq = qib_user_sdma_queue_create(&dd->pcidev->dev,
1534                                                             dd->unit,
1535                                                             rcd->ctxt,
1536                                                             fd->subctxt);
1537                         if (!fd->pq)
1538                                 ret = -ENOMEM;
1539                 }
1540
1541                 /*
1542                  * If process has NOT already set it's affinity, select and
1543                  * reserve a processor for it, as a rendezvous for all
1544                  * users of the driver.  If they don't actually later
1545                  * set affinity to this cpu, or set it to some other cpu,
1546                  * it just means that sooner or later we don't recommend
1547                  * a cpu, and let the scheduler do it's best.
1548                  */
1549                 weight = cpumask_weight(tsk_cpus_allowed(current));
1550                 if (!ret && weight >= qib_cpulist_count) {
1551                         int cpu;
1552                         cpu = find_first_zero_bit(qib_cpulist,
1553                                                   qib_cpulist_count);
1554                         if (cpu != qib_cpulist_count) {
1555                                 __set_bit(cpu, qib_cpulist);
1556                                 fd->rec_cpu_num = cpu;
1557                         }
1558                 } else if (weight == 1 &&
1559                         test_bit(cpumask_first(tsk_cpus_allowed(current)),
1560                                  qib_cpulist))
1561                         qib_devinfo(dd->pcidev, "%s PID %u affinity "
1562                                     "set to cpu %d; already allocated\n",
1563                                     current->comm, current->pid,
1564                                     cpumask_first(tsk_cpus_allowed(current)));
1565         }
1566
1567         mutex_unlock(&qib_mutex);
1568
1569 done:
1570         return ret;
1571 }
1572
1573
1574 static int qib_do_user_init(struct file *fp,
1575                             const struct qib_user_info *uinfo)
1576 {
1577         int ret;
1578         struct qib_ctxtdata *rcd = ctxt_fp(fp);
1579         struct qib_devdata *dd;
1580         unsigned uctxt;
1581
1582         /* Subctxts don't need to initialize anything since master did it. */
1583         if (subctxt_fp(fp)) {
1584                 ret = wait_event_interruptible(rcd->wait,
1585                         !test_bit(QIB_CTXT_MASTER_UNINIT, &rcd->flag));
1586                 goto bail;
1587         }
1588
1589         dd = rcd->dd;
1590
1591         /* some ctxts may get extra buffers, calculate that here */
1592         uctxt = rcd->ctxt - dd->first_user_ctxt;
1593         if (uctxt < dd->ctxts_extrabuf) {
1594                 rcd->piocnt = dd->pbufsctxt + 1;
1595                 rcd->pio_base = rcd->piocnt * uctxt;
1596         } else {
1597                 rcd->piocnt = dd->pbufsctxt;
1598                 rcd->pio_base = rcd->piocnt * uctxt +
1599                         dd->ctxts_extrabuf;
1600         }
1601
1602         /*
1603          * All user buffers are 2KB buffers.  If we ever support
1604          * giving 4KB buffers to user processes, this will need some
1605          * work.  Can't use piobufbase directly, because it has
1606          * both 2K and 4K buffer base values.  So check and handle.
1607          */
1608         if ((rcd->pio_base + rcd->piocnt) > dd->piobcnt2k) {
1609                 if (rcd->pio_base >= dd->piobcnt2k) {
1610                         qib_dev_err(dd,
1611                                     "%u:ctxt%u: no 2KB buffers available\n",
1612                                     dd->unit, rcd->ctxt);
1613                         ret = -ENOBUFS;
1614                         goto bail;
1615                 }
1616                 rcd->piocnt = dd->piobcnt2k - rcd->pio_base;
1617                 qib_dev_err(dd, "Ctxt%u: would use 4KB bufs, using %u\n",
1618                             rcd->ctxt, rcd->piocnt);
1619         }
1620
1621         rcd->piobufs = dd->pio2k_bufbase + rcd->pio_base * dd->palign;
1622         qib_chg_pioavailkernel(dd, rcd->pio_base, rcd->piocnt,
1623                                TXCHK_CHG_TYPE_USER, rcd);
1624         /*
1625          * try to ensure that processes start up with consistent avail update
1626          * for their own range, at least.   If system very quiet, it might
1627          * have the in-memory copy out of date at startup for this range of
1628          * buffers, when a context gets re-used.  Do after the chg_pioavail
1629          * and before the rest of setup, so it's "almost certain" the dma
1630          * will have occurred (can't 100% guarantee, but should be many
1631          * decimals of 9s, with this ordering), given how much else happens
1632          * after this.
1633          */
1634         dd->f_sendctrl(dd->pport, QIB_SENDCTRL_AVAIL_BLIP);
1635
1636         /*
1637          * Now allocate the rcvhdr Q and eager TIDs; skip the TID
1638          * array for time being.  If rcd->ctxt > chip-supported,
1639          * we need to do extra stuff here to handle by handling overflow
1640          * through ctxt 0, someday
1641          */
1642         ret = qib_create_rcvhdrq(dd, rcd);
1643         if (!ret)
1644                 ret = qib_setup_eagerbufs(rcd);
1645         if (ret)
1646                 goto bail_pio;
1647
1648         rcd->tidcursor = 0; /* start at beginning after open */
1649
1650         /* initialize poll variables... */
1651         rcd->urgent = 0;
1652         rcd->urgent_poll = 0;
1653
1654         /*
1655          * Now enable the ctxt for receive.
1656          * For chips that are set to DMA the tail register to memory
1657          * when they change (and when the update bit transitions from
1658          * 0 to 1.  So for those chips, we turn it off and then back on.
1659          * This will (very briefly) affect any other open ctxts, but the
1660          * duration is very short, and therefore isn't an issue.  We
1661          * explicitly set the in-memory tail copy to 0 beforehand, so we
1662          * don't have to wait to be sure the DMA update has happened
1663          * (chip resets head/tail to 0 on transition to enable).
1664          */
1665         if (rcd->rcvhdrtail_kvaddr)
1666                 qib_clear_rcvhdrtail(rcd);
1667
1668         dd->f_rcvctrl(rcd->ppd, QIB_RCVCTRL_CTXT_ENB | QIB_RCVCTRL_TIDFLOW_ENB,
1669                       rcd->ctxt);
1670
1671         /* Notify any waiting slaves */
1672         if (rcd->subctxt_cnt) {
1673                 clear_bit(QIB_CTXT_MASTER_UNINIT, &rcd->flag);
1674                 wake_up(&rcd->wait);
1675         }
1676         return 0;
1677
1678 bail_pio:
1679         qib_chg_pioavailkernel(dd, rcd->pio_base, rcd->piocnt,
1680                                TXCHK_CHG_TYPE_KERN, rcd);
1681 bail:
1682         return ret;
1683 }
1684
1685 /**
1686  * unlock_exptid - unlock any expected TID entries context still had in use
1687  * @rcd: ctxt
1688  *
1689  * We don't actually update the chip here, because we do a bulk update
1690  * below, using f_clear_tids.
1691  */
1692 static void unlock_expected_tids(struct qib_ctxtdata *rcd)
1693 {
1694         struct qib_devdata *dd = rcd->dd;
1695         int ctxt_tidbase = rcd->ctxt * dd->rcvtidcnt;
1696         int i, cnt = 0, maxtid = ctxt_tidbase + dd->rcvtidcnt;
1697
1698         for (i = ctxt_tidbase; i < maxtid; i++) {
1699                 struct page *p = dd->pageshadow[i];
1700                 dma_addr_t phys;
1701
1702                 if (!p)
1703                         continue;
1704
1705                 phys = dd->physshadow[i];
1706                 dd->physshadow[i] = dd->tidinvalid;
1707                 dd->pageshadow[i] = NULL;
1708                 pci_unmap_page(dd->pcidev, phys, PAGE_SIZE,
1709                                PCI_DMA_FROMDEVICE);
1710                 qib_release_user_pages(&p, 1);
1711                 cnt++;
1712         }
1713 }
1714
1715 static int qib_close(struct inode *in, struct file *fp)
1716 {
1717         int ret = 0;
1718         struct qib_filedata *fd;
1719         struct qib_ctxtdata *rcd;
1720         struct qib_devdata *dd;
1721         unsigned long flags;
1722         unsigned ctxt;
1723         pid_t pid;
1724
1725         mutex_lock(&qib_mutex);
1726
1727         fd = fp->private_data;
1728         fp->private_data = NULL;
1729         rcd = fd->rcd;
1730         if (!rcd) {
1731                 mutex_unlock(&qib_mutex);
1732                 goto bail;
1733         }
1734
1735         dd = rcd->dd;
1736
1737         /* ensure all pio buffer writes in progress are flushed */
1738         qib_flush_wc();
1739
1740         /* drain user sdma queue */
1741         if (fd->pq) {
1742                 qib_user_sdma_queue_drain(rcd->ppd, fd->pq);
1743                 qib_user_sdma_queue_destroy(fd->pq);
1744         }
1745
1746         if (fd->rec_cpu_num != -1)
1747                 __clear_bit(fd->rec_cpu_num, qib_cpulist);
1748
1749         if (--rcd->cnt) {
1750                 /*
1751                  * XXX If the master closes the context before the slave(s),
1752                  * revoke the mmap for the eager receive queue so
1753                  * the slave(s) don't wait for receive data forever.
1754                  */
1755                 rcd->active_slaves &= ~(1 << fd->subctxt);
1756                 rcd->subpid[fd->subctxt] = 0;
1757                 mutex_unlock(&qib_mutex);
1758                 goto bail;
1759         }
1760
1761         /* early; no interrupt users after this */
1762         spin_lock_irqsave(&dd->uctxt_lock, flags);
1763         ctxt = rcd->ctxt;
1764         dd->rcd[ctxt] = NULL;
1765         pid = rcd->pid;
1766         rcd->pid = 0;
1767         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
1768
1769         if (rcd->rcvwait_to || rcd->piowait_to ||
1770             rcd->rcvnowait || rcd->pionowait) {
1771                 rcd->rcvwait_to = 0;
1772                 rcd->piowait_to = 0;
1773                 rcd->rcvnowait = 0;
1774                 rcd->pionowait = 0;
1775         }
1776         if (rcd->flag)
1777                 rcd->flag = 0;
1778
1779         if (dd->kregbase) {
1780                 /* atomically clear receive enable ctxt and intr avail. */
1781                 dd->f_rcvctrl(rcd->ppd, QIB_RCVCTRL_CTXT_DIS |
1782                                   QIB_RCVCTRL_INTRAVAIL_DIS, ctxt);
1783
1784                 /* clean up the pkeys for this ctxt user */
1785                 qib_clean_part_key(rcd, dd);
1786                 qib_disarm_piobufs(dd, rcd->pio_base, rcd->piocnt);
1787                 qib_chg_pioavailkernel(dd, rcd->pio_base,
1788                                        rcd->piocnt, TXCHK_CHG_TYPE_KERN, NULL);
1789
1790                 dd->f_clear_tids(dd, rcd);
1791
1792                 if (dd->pageshadow)
1793                         unlock_expected_tids(rcd);
1794                 qib_stats.sps_ctxts--;
1795         }
1796
1797         mutex_unlock(&qib_mutex);
1798         qib_free_ctxtdata(dd, rcd); /* after releasing the mutex */
1799
1800 bail:
1801         kfree(fd);
1802         return ret;
1803 }
1804
1805 static int qib_ctxt_info(struct file *fp, struct qib_ctxt_info __user *uinfo)
1806 {
1807         struct qib_ctxt_info info;
1808         int ret;
1809         size_t sz;
1810         struct qib_ctxtdata *rcd = ctxt_fp(fp);
1811         struct qib_filedata *fd;
1812
1813         fd = fp->private_data;
1814
1815         info.num_active = qib_count_active_units();
1816         info.unit = rcd->dd->unit;
1817         info.port = rcd->ppd->port;
1818         info.ctxt = rcd->ctxt;
1819         info.subctxt =  subctxt_fp(fp);
1820         /* Number of user ctxts available for this device. */
1821         info.num_ctxts = rcd->dd->cfgctxts - rcd->dd->first_user_ctxt;
1822         info.num_subctxts = rcd->subctxt_cnt;
1823         info.rec_cpu = fd->rec_cpu_num;
1824         sz = sizeof(info);
1825
1826         if (copy_to_user(uinfo, &info, sz)) {
1827                 ret = -EFAULT;
1828                 goto bail;
1829         }
1830         ret = 0;
1831
1832 bail:
1833         return ret;
1834 }
1835
1836 static int qib_sdma_get_inflight(struct qib_user_sdma_queue *pq,
1837                                  u32 __user *inflightp)
1838 {
1839         const u32 val = qib_user_sdma_inflight_counter(pq);
1840
1841         if (put_user(val, inflightp))
1842                 return -EFAULT;
1843
1844         return 0;
1845 }
1846
1847 static int qib_sdma_get_complete(struct qib_pportdata *ppd,
1848                                  struct qib_user_sdma_queue *pq,
1849                                  u32 __user *completep)
1850 {
1851         u32 val;
1852         int err;
1853
1854         if (!pq)
1855                 return -EINVAL;
1856
1857         err = qib_user_sdma_make_progress(ppd, pq);
1858         if (err < 0)
1859                 return err;
1860
1861         val = qib_user_sdma_complete_counter(pq);
1862         if (put_user(val, completep))
1863                 return -EFAULT;
1864
1865         return 0;
1866 }
1867
1868 static int disarm_req_delay(struct qib_ctxtdata *rcd)
1869 {
1870         int ret = 0;
1871
1872         if (!usable(rcd->ppd)) {
1873                 int i;
1874                 /*
1875                  * if link is down, or otherwise not usable, delay
1876                  * the caller up to 30 seconds, so we don't thrash
1877                  * in trying to get the chip back to ACTIVE, and
1878                  * set flag so they make the call again.
1879                  */
1880                 if (rcd->user_event_mask) {
1881                         /*
1882                          * subctxt_cnt is 0 if not shared, so do base
1883                          * separately, first, then remaining subctxt, if any
1884                          */
1885                         set_bit(_QIB_EVENT_DISARM_BUFS_BIT,
1886                                 &rcd->user_event_mask[0]);
1887                         for (i = 1; i < rcd->subctxt_cnt; i++)
1888                                 set_bit(_QIB_EVENT_DISARM_BUFS_BIT,
1889                                         &rcd->user_event_mask[i]);
1890                 }
1891                 for (i = 0; !usable(rcd->ppd) && i < 300; i++)
1892                         msleep(100);
1893                 ret = -ENETDOWN;
1894         }
1895         return ret;
1896 }
1897
1898 /*
1899  * Find all user contexts in use, and set the specified bit in their
1900  * event mask.
1901  * See also find_ctxt() for a similar use, that is specific to send buffers.
1902  */
1903 int qib_set_uevent_bits(struct qib_pportdata *ppd, const int evtbit)
1904 {
1905         struct qib_ctxtdata *rcd;
1906         unsigned ctxt;
1907         int ret = 0;
1908         unsigned long flags;
1909
1910         spin_lock_irqsave(&ppd->dd->uctxt_lock, flags);
1911         for (ctxt = ppd->dd->first_user_ctxt; ctxt < ppd->dd->cfgctxts;
1912              ctxt++) {
1913                 rcd = ppd->dd->rcd[ctxt];
1914                 if (!rcd)
1915                         continue;
1916                 if (rcd->user_event_mask) {
1917                         int i;
1918                         /*
1919                          * subctxt_cnt is 0 if not shared, so do base
1920                          * separately, first, then remaining subctxt, if any
1921                          */
1922                         set_bit(evtbit, &rcd->user_event_mask[0]);
1923                         for (i = 1; i < rcd->subctxt_cnt; i++)
1924                                 set_bit(evtbit, &rcd->user_event_mask[i]);
1925                 }
1926                 ret = 1;
1927                 break;
1928         }
1929         spin_unlock_irqrestore(&ppd->dd->uctxt_lock, flags);
1930
1931         return ret;
1932 }
1933
1934 /*
1935  * clear the event notifier events for this context.
1936  * For the DISARM_BUFS case, we also take action (this obsoletes
1937  * the older QIB_CMD_DISARM_BUFS, but we keep it for backwards
1938  * compatibility.
1939  * Other bits don't currently require actions, just atomically clear.
1940  * User process then performs actions appropriate to bit having been
1941  * set, if desired, and checks again in future.
1942  */
1943 static int qib_user_event_ack(struct qib_ctxtdata *rcd, int subctxt,
1944                               unsigned long events)
1945 {
1946         int ret = 0, i;
1947
1948         for (i = 0; i <= _QIB_MAX_EVENT_BIT; i++) {
1949                 if (!test_bit(i, &events))
1950                         continue;
1951                 if (i == _QIB_EVENT_DISARM_BUFS_BIT) {
1952                         (void)qib_disarm_piobufs_ifneeded(rcd);
1953                         ret = disarm_req_delay(rcd);
1954                 } else
1955                         clear_bit(i, &rcd->user_event_mask[subctxt]);
1956         }
1957         return ret;
1958 }
1959
1960 static ssize_t qib_write(struct file *fp, const char __user *data,
1961                          size_t count, loff_t *off)
1962 {
1963         const struct qib_cmd __user *ucmd;
1964         struct qib_ctxtdata *rcd;
1965         const void __user *src;
1966         size_t consumed, copy = 0;
1967         struct qib_cmd cmd;
1968         ssize_t ret = 0;
1969         void *dest;
1970
1971         if (count < sizeof(cmd.type)) {
1972                 ret = -EINVAL;
1973                 goto bail;
1974         }
1975
1976         ucmd = (const struct qib_cmd __user *) data;
1977
1978         if (copy_from_user(&cmd.type, &ucmd->type, sizeof(cmd.type))) {
1979                 ret = -EFAULT;
1980                 goto bail;
1981         }
1982
1983         consumed = sizeof(cmd.type);
1984
1985         switch (cmd.type) {
1986         case QIB_CMD_ASSIGN_CTXT:
1987         case QIB_CMD_USER_INIT:
1988                 copy = sizeof(cmd.cmd.user_info);
1989                 dest = &cmd.cmd.user_info;
1990                 src = &ucmd->cmd.user_info;
1991                 break;
1992
1993         case QIB_CMD_RECV_CTRL:
1994                 copy = sizeof(cmd.cmd.recv_ctrl);
1995                 dest = &cmd.cmd.recv_ctrl;
1996                 src = &ucmd->cmd.recv_ctrl;
1997                 break;
1998
1999         case QIB_CMD_CTXT_INFO:
2000                 copy = sizeof(cmd.cmd.ctxt_info);
2001                 dest = &cmd.cmd.ctxt_info;
2002                 src = &ucmd->cmd.ctxt_info;
2003                 break;
2004
2005         case QIB_CMD_TID_UPDATE:
2006         case QIB_CMD_TID_FREE:
2007                 copy = sizeof(cmd.cmd.tid_info);
2008                 dest = &cmd.cmd.tid_info;
2009                 src = &ucmd->cmd.tid_info;
2010                 break;
2011
2012         case QIB_CMD_SET_PART_KEY:
2013                 copy = sizeof(cmd.cmd.part_key);
2014                 dest = &cmd.cmd.part_key;
2015                 src = &ucmd->cmd.part_key;
2016                 break;
2017
2018         case QIB_CMD_DISARM_BUFS:
2019         case QIB_CMD_PIOAVAILUPD: /* force an update of PIOAvail reg */
2020                 copy = 0;
2021                 src = NULL;
2022                 dest = NULL;
2023                 break;
2024
2025         case QIB_CMD_POLL_TYPE:
2026                 copy = sizeof(cmd.cmd.poll_type);
2027                 dest = &cmd.cmd.poll_type;
2028                 src = &ucmd->cmd.poll_type;
2029                 break;
2030
2031         case QIB_CMD_ARMLAUNCH_CTRL:
2032                 copy = sizeof(cmd.cmd.armlaunch_ctrl);
2033                 dest = &cmd.cmd.armlaunch_ctrl;
2034                 src = &ucmd->cmd.armlaunch_ctrl;
2035                 break;
2036
2037         case QIB_CMD_SDMA_INFLIGHT:
2038                 copy = sizeof(cmd.cmd.sdma_inflight);
2039                 dest = &cmd.cmd.sdma_inflight;
2040                 src = &ucmd->cmd.sdma_inflight;
2041                 break;
2042
2043         case QIB_CMD_SDMA_COMPLETE:
2044                 copy = sizeof(cmd.cmd.sdma_complete);
2045                 dest = &cmd.cmd.sdma_complete;
2046                 src = &ucmd->cmd.sdma_complete;
2047                 break;
2048
2049         case QIB_CMD_ACK_EVENT:
2050                 copy = sizeof(cmd.cmd.event_mask);
2051                 dest = &cmd.cmd.event_mask;
2052                 src = &ucmd->cmd.event_mask;
2053                 break;
2054
2055         default:
2056                 ret = -EINVAL;
2057                 goto bail;
2058         }
2059
2060         if (copy) {
2061                 if ((count - consumed) < copy) {
2062                         ret = -EINVAL;
2063                         goto bail;
2064                 }
2065                 if (copy_from_user(dest, src, copy)) {
2066                         ret = -EFAULT;
2067                         goto bail;
2068                 }
2069                 consumed += copy;
2070         }
2071
2072         rcd = ctxt_fp(fp);
2073         if (!rcd && cmd.type != QIB_CMD_ASSIGN_CTXT) {
2074                 ret = -EINVAL;
2075                 goto bail;
2076         }
2077
2078         switch (cmd.type) {
2079         case QIB_CMD_ASSIGN_CTXT:
2080                 ret = qib_assign_ctxt(fp, &cmd.cmd.user_info);
2081                 if (ret)
2082                         goto bail;
2083                 break;
2084
2085         case QIB_CMD_USER_INIT:
2086                 ret = qib_do_user_init(fp, &cmd.cmd.user_info);
2087                 if (ret)
2088                         goto bail;
2089                 ret = qib_get_base_info(fp, (void __user *) (unsigned long)
2090                                         cmd.cmd.user_info.spu_base_info,
2091                                         cmd.cmd.user_info.spu_base_info_size);
2092                 break;
2093
2094         case QIB_CMD_RECV_CTRL:
2095                 ret = qib_manage_rcvq(rcd, subctxt_fp(fp), cmd.cmd.recv_ctrl);
2096                 break;
2097
2098         case QIB_CMD_CTXT_INFO:
2099                 ret = qib_ctxt_info(fp, (struct qib_ctxt_info __user *)
2100                                     (unsigned long) cmd.cmd.ctxt_info);
2101                 break;
2102
2103         case QIB_CMD_TID_UPDATE:
2104                 ret = qib_tid_update(rcd, fp, &cmd.cmd.tid_info);
2105                 break;
2106
2107         case QIB_CMD_TID_FREE:
2108                 ret = qib_tid_free(rcd, subctxt_fp(fp), &cmd.cmd.tid_info);
2109                 break;
2110
2111         case QIB_CMD_SET_PART_KEY:
2112                 ret = qib_set_part_key(rcd, cmd.cmd.part_key);
2113                 break;
2114
2115         case QIB_CMD_DISARM_BUFS:
2116                 (void)qib_disarm_piobufs_ifneeded(rcd);
2117                 ret = disarm_req_delay(rcd);
2118                 break;
2119
2120         case QIB_CMD_PIOAVAILUPD:
2121                 qib_force_pio_avail_update(rcd->dd);
2122                 break;
2123
2124         case QIB_CMD_POLL_TYPE:
2125                 rcd->poll_type = cmd.cmd.poll_type;
2126                 break;
2127
2128         case QIB_CMD_ARMLAUNCH_CTRL:
2129                 rcd->dd->f_set_armlaunch(rcd->dd, cmd.cmd.armlaunch_ctrl);
2130                 break;
2131
2132         case QIB_CMD_SDMA_INFLIGHT:
2133                 ret = qib_sdma_get_inflight(user_sdma_queue_fp(fp),
2134                                             (u32 __user *) (unsigned long)
2135                                             cmd.cmd.sdma_inflight);
2136                 break;
2137
2138         case QIB_CMD_SDMA_COMPLETE:
2139                 ret = qib_sdma_get_complete(rcd->ppd,
2140                                             user_sdma_queue_fp(fp),
2141                                             (u32 __user *) (unsigned long)
2142                                             cmd.cmd.sdma_complete);
2143                 break;
2144
2145         case QIB_CMD_ACK_EVENT:
2146                 ret = qib_user_event_ack(rcd, subctxt_fp(fp),
2147                                          cmd.cmd.event_mask);
2148                 break;
2149         }
2150
2151         if (ret >= 0)
2152                 ret = consumed;
2153
2154 bail:
2155         return ret;
2156 }
2157
2158 static ssize_t qib_aio_write(struct kiocb *iocb, const struct iovec *iov,
2159                              unsigned long dim, loff_t off)
2160 {
2161         struct qib_filedata *fp = iocb->ki_filp->private_data;
2162         struct qib_ctxtdata *rcd = ctxt_fp(iocb->ki_filp);
2163         struct qib_user_sdma_queue *pq = fp->pq;
2164
2165         if (!dim || !pq)
2166                 return -EINVAL;
2167
2168         return qib_user_sdma_writev(rcd, pq, iov, dim);
2169 }
2170
2171 static struct class *qib_class;
2172 static dev_t qib_dev;
2173
2174 int qib_cdev_init(int minor, const char *name,
2175                   const struct file_operations *fops,
2176                   struct cdev **cdevp, struct device **devp)
2177 {
2178         const dev_t dev = MKDEV(MAJOR(qib_dev), minor);
2179         struct cdev *cdev;
2180         struct device *device = NULL;
2181         int ret;
2182
2183         cdev = cdev_alloc();
2184         if (!cdev) {
2185                 printk(KERN_ERR QIB_DRV_NAME
2186                        ": Could not allocate cdev for minor %d, %s\n",
2187                        minor, name);
2188                 ret = -ENOMEM;
2189                 goto done;
2190         }
2191
2192         cdev->owner = THIS_MODULE;
2193         cdev->ops = fops;
2194         kobject_set_name(&cdev->kobj, name);
2195
2196         ret = cdev_add(cdev, dev, 1);
2197         if (ret < 0) {
2198                 printk(KERN_ERR QIB_DRV_NAME
2199                        ": Could not add cdev for minor %d, %s (err %d)\n",
2200                        minor, name, -ret);
2201                 goto err_cdev;
2202         }
2203
2204         device = device_create(qib_class, NULL, dev, NULL, name);
2205         if (!IS_ERR(device))
2206                 goto done;
2207         ret = PTR_ERR(device);
2208         device = NULL;
2209         printk(KERN_ERR QIB_DRV_NAME ": Could not create "
2210                "device for minor %d, %s (err %d)\n",
2211                minor, name, -ret);
2212 err_cdev:
2213         cdev_del(cdev);
2214         cdev = NULL;
2215 done:
2216         *cdevp = cdev;
2217         *devp = device;
2218         return ret;
2219 }
2220
2221 void qib_cdev_cleanup(struct cdev **cdevp, struct device **devp)
2222 {
2223         struct device *device = *devp;
2224
2225         if (device) {
2226                 device_unregister(device);
2227                 *devp = NULL;
2228         }
2229
2230         if (*cdevp) {
2231                 cdev_del(*cdevp);
2232                 *cdevp = NULL;
2233         }
2234 }
2235
2236 static struct cdev *wildcard_cdev;
2237 static struct device *wildcard_device;
2238
2239 int __init qib_dev_init(void)
2240 {
2241         int ret;
2242
2243         ret = alloc_chrdev_region(&qib_dev, 0, QIB_NMINORS, QIB_DRV_NAME);
2244         if (ret < 0) {
2245                 printk(KERN_ERR QIB_DRV_NAME ": Could not allocate "
2246                        "chrdev region (err %d)\n", -ret);
2247                 goto done;
2248         }
2249
2250         qib_class = class_create(THIS_MODULE, "ipath");
2251         if (IS_ERR(qib_class)) {
2252                 ret = PTR_ERR(qib_class);
2253                 printk(KERN_ERR QIB_DRV_NAME ": Could not create "
2254                        "device class (err %d)\n", -ret);
2255                 unregister_chrdev_region(qib_dev, QIB_NMINORS);
2256         }
2257
2258 done:
2259         return ret;
2260 }
2261
2262 void qib_dev_cleanup(void)
2263 {
2264         if (qib_class) {
2265                 class_destroy(qib_class);
2266                 qib_class = NULL;
2267         }
2268
2269         unregister_chrdev_region(qib_dev, QIB_NMINORS);
2270 }
2271
2272 static atomic_t user_count = ATOMIC_INIT(0);
2273
2274 static void qib_user_remove(struct qib_devdata *dd)
2275 {
2276         if (atomic_dec_return(&user_count) == 0)
2277                 qib_cdev_cleanup(&wildcard_cdev, &wildcard_device);
2278
2279         qib_cdev_cleanup(&dd->user_cdev, &dd->user_device);
2280 }
2281
2282 static int qib_user_add(struct qib_devdata *dd)
2283 {
2284         char name[10];
2285         int ret;
2286
2287         if (atomic_inc_return(&user_count) == 1) {
2288                 ret = qib_cdev_init(0, "ipath", &qib_file_ops,
2289                                     &wildcard_cdev, &wildcard_device);
2290                 if (ret)
2291                         goto done;
2292         }
2293
2294         snprintf(name, sizeof(name), "ipath%d", dd->unit);
2295         ret = qib_cdev_init(dd->unit + 1, name, &qib_file_ops,
2296                             &dd->user_cdev, &dd->user_device);
2297         if (ret)
2298                 qib_user_remove(dd);
2299 done:
2300         return ret;
2301 }
2302
2303 /*
2304  * Create per-unit files in /dev
2305  */
2306 int qib_device_create(struct qib_devdata *dd)
2307 {
2308         int r, ret;
2309
2310         r = qib_user_add(dd);
2311         ret = qib_diag_add(dd);
2312         if (r && !ret)
2313                 ret = r;
2314         return ret;
2315 }
2316
2317 /*
2318  * Remove per-unit files in /dev
2319  * void, core kernel returns no errors for this stuff
2320  */
2321 void qib_device_remove(struct qib_devdata *dd)
2322 {
2323         qib_user_remove(dd);
2324         qib_diag_remove(dd);
2325 }