drivers/infiniband/hw/qib/qib_file_ops.c

   1 /*
   2  * Copyright (c) 2012, 2013 Intel Corporation. All rights reserved.
   3  * Copyright (c) 2006 - 2012 QLogic Corporation. 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/aio.h>
  43 #include <linux/jiffies.h>
  44 #include <asm/pgtable.h>
  45 #include <linux/delay.h>
  46 #include <linux/export.h>
  47
  48 #include "qib.h"
  49 #include "qib_common.h"
  50 #include "qib_user_sdma.h"
  51
  52 #undef pr_fmt
  53 #define pr_fmt(fmt) QIB_DRV_NAME ": " fmt
  54
  55 static int qib_open(struct inode *, struct file *);
  56 static int qib_close(struct inode *, struct file *);
  57 static ssize_t qib_write(struct file *, const char __user *, size_t, loff_t *);
  58 static ssize_t qib_aio_write(struct kiocb *, const struct iovec *,
  59                              unsigned long, loff_t);
  60 static unsigned int qib_poll(struct file *, struct poll_table_struct *);
  61 static int qib_mmapf(struct file *, struct vm_area_struct *);
  62
  63 static const struct file_operations qib_file_ops = {
  64         .owner = THIS_MODULE,
  65         .write = qib_write,
  66         .aio_write = qib_aio_write,
  67         .open = qib_open,
  68         .release = qib_close,
  69         .poll = qib_poll,
  70         .mmap = qib_mmapf,
  71         .llseek = noop_llseek,
  72 };
  73
  74 /*
  75  * Convert kernel virtual addresses to physical addresses so they don't
  76  * potentially conflict with the chip addresses used as mmap offsets.
  77  * It doesn't really matter what mmap offset we use as long as we can
  78  * interpret it correctly.
  79  */
  80 static u64 cvt_kvaddr(void *p)
  81 {
  82         struct page *page;
  83         u64 paddr = 0;
  84
  85         page = vmalloc_to_page(p);
  86         if (page)
  87                 paddr = page_to_pfn(page) << PAGE_SHIFT;
  88
  89         return paddr;
  90 }
  91
  92 static int qib_get_base_info(struct file *fp, void __user *ubase,
  93                              size_t ubase_size)
  94 {
  95         struct qib_ctxtdata *rcd = ctxt_fp(fp);
  96         int ret = 0;
  97         struct qib_base_info *kinfo = NULL;
  98         struct qib_devdata *dd = rcd->dd;
  99         struct qib_pportdata *ppd = rcd->ppd;
 100         unsigned subctxt_cnt;
 101         int shared, master;
 102         size_t sz;
 103
 104         subctxt_cnt = rcd->subctxt_cnt;
 105         if (!subctxt_cnt) {
 106                 shared = 0;
 107                 master = 0;
 108                 subctxt_cnt = 1;
 109         } else {
 110                 shared = 1;
 111                 master = !subctxt_fp(fp);
 112         }
 113
 114         sz = sizeof(*kinfo);
 115         /* If context sharing is not requested, allow the old size structure */
 116         if (!shared)
 117                 sz -= 7 * sizeof(u64);
 118         if (ubase_size < sz) {
 119                 ret = -EINVAL;
 120                 goto bail;
 121         }
 122
 123         kinfo = kzalloc(sizeof(*kinfo), GFP_KERNEL);
 124         if (kinfo == NULL) {
 125                 ret = -ENOMEM;
 126                 goto bail;
 127         }
 128
 129         ret = dd->f_get_base_info(rcd, kinfo);
 130         if (ret < 0)
 131                 goto bail;
 132
 133         kinfo->spi_rcvhdr_cnt = dd->rcvhdrcnt;
 134         kinfo->spi_rcvhdrent_size = dd->rcvhdrentsize;
 135         kinfo->spi_tidegrcnt = rcd->rcvegrcnt;
 136         kinfo->spi_rcv_egrbufsize = dd->rcvegrbufsize;
 137         /*
 138          * have to mmap whole thing
 139          */
 140         kinfo->spi_rcv_egrbuftotlen =
 141                 rcd->rcvegrbuf_chunks * rcd->rcvegrbuf_size;
 142         kinfo->spi_rcv_egrperchunk = rcd->rcvegrbufs_perchunk;
 143         kinfo->spi_rcv_egrchunksize = kinfo->spi_rcv_egrbuftotlen /
 144                 rcd->rcvegrbuf_chunks;
 145         kinfo->spi_tidcnt = dd->rcvtidcnt / subctxt_cnt;
 146         if (master)
 147                 kinfo->spi_tidcnt += dd->rcvtidcnt % subctxt_cnt;
 148         /*
 149          * for this use, may be cfgctxts summed over all chips that
 150          * are are configured and present
 151          */
 152         kinfo->spi_nctxts = dd->cfgctxts;
 153         /* unit (chip/board) our context is on */
 154         kinfo->spi_unit = dd->unit;
 155         kinfo->spi_port = ppd->port;
 156         /* for now, only a single page */
 157         kinfo->spi_tid_maxsize = PAGE_SIZE;
 158
 159         /*
 160          * Doing this per context, and based on the skip value, etc.  This has
 161          * to be the actual buffer size, since the protocol code treats it
 162          * as an array.
 163          *
 164          * These have to be set to user addresses in the user code via mmap.
 165          * These values are used on return to user code for the mmap target
 166          * addresses only.  For 32 bit, same 44 bit address problem, so use
 167          * the physical address, not virtual.  Before 2.6.11, using the
 168          * page_address() macro worked, but in 2.6.11, even that returns the
 169          * full 64 bit address (upper bits all 1's).  So far, using the
 170          * physical addresses (or chip offsets, for chip mapping) works, but
 171          * no doubt some future kernel release will change that, and we'll be
 172          * on to yet another method of dealing with this.
 173          * Normally only one of rcvhdr_tailaddr or rhf_offset is useful
 174          * since the chips with non-zero rhf_offset don't normally
 175          * enable tail register updates to host memory, but for testing,
 176          * both can be enabled and used.
 177          */
 178         kinfo->spi_rcvhdr_base = (u64) rcd->rcvhdrq_phys;
 179         kinfo->spi_rcvhdr_tailaddr = (u64) rcd->rcvhdrqtailaddr_phys;
 180         kinfo->spi_rhf_offset = dd->rhf_offset;
 181         kinfo->spi_rcv_egrbufs = (u64) rcd->rcvegr_phys;
 182         kinfo->spi_pioavailaddr = (u64) dd->pioavailregs_phys;
 183         /* setup per-unit (not port) status area for user programs */
 184         kinfo->spi_status = (u64) kinfo->spi_pioavailaddr +
 185                 (char *) ppd->statusp -
 186                 (char *) dd->pioavailregs_dma;
 187         kinfo->spi_uregbase = (u64) dd->uregbase + dd->ureg_align * rcd->ctxt;
 188         if (!shared) {
 189                 kinfo->spi_piocnt = rcd->piocnt;
 190                 kinfo->spi_piobufbase = (u64) rcd->piobufs;
 191                 kinfo->spi_sendbuf_status = cvt_kvaddr(rcd->user_event_mask);
 192         } else if (master) {
 193                 kinfo->spi_piocnt = (rcd->piocnt / subctxt_cnt) +
 194                                     (rcd->piocnt % subctxt_cnt);
 195                 /* Master's PIO buffers are after all the slave's */
 196                 kinfo->spi_piobufbase = (u64) rcd->piobufs +
 197                         dd->palign *
 198                         (rcd->piocnt - kinfo->spi_piocnt);
 199         } else {
 200                 unsigned slave = subctxt_fp(fp) - 1;
 201
 202                 kinfo->spi_piocnt = rcd->piocnt / subctxt_cnt;
 203                 kinfo->spi_piobufbase = (u64) rcd->piobufs +
 204                         dd->palign * kinfo->spi_piocnt * slave;
 205         }
 206
 207         if (shared) {
 208                 kinfo->spi_sendbuf_status =
 209                         cvt_kvaddr(&rcd->user_event_mask[subctxt_fp(fp)]);
 210                 /* only spi_subctxt_* fields should be set in this block! */
 211                 kinfo->spi_subctxt_uregbase = cvt_kvaddr(rcd->subctxt_uregbase);
 212
 213                 kinfo->spi_subctxt_rcvegrbuf =
 214                         cvt_kvaddr(rcd->subctxt_rcvegrbuf);
 215                 kinfo->spi_subctxt_rcvhdr_base =
 216                         cvt_kvaddr(rcd->subctxt_rcvhdr_base);
 217         }
 218
 219         /*
 220          * All user buffers are 2KB buffers.  If we ever support
 221          * giving 4KB buffers to user processes, this will need some
 222          * work.  Can't use piobufbase directly, because it has
 223          * both 2K and 4K buffer base values.
 224          */
 225         kinfo->spi_pioindex = (kinfo->spi_piobufbase - dd->pio2k_bufbase) /
 226                 dd->palign;
 227         kinfo->spi_pioalign = dd->palign;
 228         kinfo->spi_qpair = QIB_KD_QP;
 229         /*
 230          * user mode PIO buffers are always 2KB, even when 4KB can
 231          * be received, and sent via the kernel; this is ibmaxlen
 232          * for 2K MTU.
 233          */
 234         kinfo->spi_piosize = dd->piosize2k - 2 * sizeof(u32);
 235         kinfo->spi_mtu = ppd->ibmaxlen; /* maxlen, not ibmtu */
 236         kinfo->spi_ctxt = rcd->ctxt;
 237         kinfo->spi_subctxt = subctxt_fp(fp);
 238         kinfo->spi_sw_version = QIB_KERN_SWVERSION;
 239         kinfo->spi_sw_version |= 1U << 31; /* QLogic-built, not kernel.org */
 240         kinfo->spi_hw_version = dd->revision;
 241
 242         if (master)
 243                 kinfo->spi_runtime_flags |= QIB_RUNTIME_MASTER;
 244
 245         sz = (ubase_size < sizeof(*kinfo)) ? ubase_size : sizeof(*kinfo);
 246         if (copy_to_user(ubase, kinfo, sz))
 247                 ret = -EFAULT;
 248 bail:
 249         kfree(kinfo);
 250         return ret;
 251 }
 252
 253 /**
 254  * qib_tid_update - update a context TID
 255  * @rcd: the context
 256  * @fp: the qib device file
 257  * @ti: the TID information
 258  *
 259  * The new implementation as of Oct 2004 is that the driver assigns
 260  * the tid and returns it to the caller.   To reduce search time, we
 261  * keep a cursor for each context, walking the shadow tid array to find
 262  * one that's not in use.
 263  *
 264  * For now, if we can't allocate the full list, we fail, although
 265  * in the long run, we'll allocate as many as we can, and the
 266  * caller will deal with that by trying the remaining pages later.
 267  * That means that when we fail, we have to mark the tids as not in
 268  * use again, in our shadow copy.
 269  *
 270  * It's up to the caller to free the tids when they are done.
 271  * We'll unlock the pages as they free them.
 272  *
 273  * Also, right now we are locking one page at a time, but since
 274  * the intended use of this routine is for a single group of
 275  * virtually contiguous pages, that should change to improve
 276  * performance.
 277  */
 278 static int qib_tid_update(struct qib_ctxtdata *rcd, struct file *fp,
 279                           const struct qib_tid_info *ti)
 280 {
 281         int ret = 0, ntids;
 282         u32 tid, ctxttid, cnt, i, tidcnt, tidoff;
 283         u16 *tidlist;
 284         struct qib_devdata *dd = rcd->dd;
 285         u64 physaddr;
 286         unsigned long vaddr;
 287         u64 __iomem *tidbase;
 288         unsigned long tidmap[8];
 289         struct page **pagep = NULL;
 290         unsigned subctxt = subctxt_fp(fp);
 291
 292         if (!dd->pageshadow) {
 293                 ret = -ENOMEM;
 294                 goto done;
 295         }
 296
 297         cnt = ti->tidcnt;
 298         if (!cnt) {
 299                 ret = -EFAULT;
 300                 goto done;
 301         }
 302         ctxttid = rcd->ctxt * dd->rcvtidcnt;
 303         if (!rcd->subctxt_cnt) {
 304                 tidcnt = dd->rcvtidcnt;
 305                 tid = rcd->tidcursor;
 306                 tidoff = 0;
 307         } else if (!subctxt) {
 308                 tidcnt = (dd->rcvtidcnt / rcd->subctxt_cnt) +
 309                          (dd->rcvtidcnt % rcd->subctxt_cnt);
 310                 tidoff = dd->rcvtidcnt - tidcnt;
 311                 ctxttid += tidoff;
 312                 tid = tidcursor_fp(fp);
 313         } else {
 314                 tidcnt = dd->rcvtidcnt / rcd->subctxt_cnt;
 315                 tidoff = tidcnt * (subctxt - 1);
 316                 ctxttid += tidoff;
 317                 tid = tidcursor_fp(fp);
 318         }
 319         if (cnt > tidcnt) {
 320                 /* make sure it all fits in tid_pg_list */
 321                 qib_devinfo(dd->pcidev,
 322                         "Process tried to allocate %u TIDs, only trying max (%u)\n",
 323                         cnt, tidcnt);
 324                 cnt = tidcnt;
 325         }
 326         pagep = (struct page **) rcd->tid_pg_list;
 327         tidlist = (u16 *) &pagep[dd->rcvtidcnt];
 328         pagep += tidoff;
 329         tidlist += tidoff;
 330
 331         memset(tidmap, 0, sizeof(tidmap));
 332         /* before decrement; chip actual # */
 333         ntids = tidcnt;
 334         tidbase = (u64 __iomem *) (((char __iomem *) dd->kregbase) +
 335                                    dd->rcvtidbase +
 336                                    ctxttid * sizeof(*tidbase));
 337
 338         /* virtual address of first page in transfer */
 339         vaddr = ti->tidvaddr;
 340         if (!access_ok(VERIFY_WRITE, (void __user *) vaddr,
 341                        cnt * PAGE_SIZE)) {
 342                 ret = -EFAULT;
 343                 goto done;
 344         }
 345         ret = qib_get_user_pages(vaddr, cnt, pagep);
 346         if (ret) {
 347                 /*
 348                  * if (ret == -EBUSY)
 349                  * We can't continue because the pagep array won't be
 350                  * initialized. This should never happen,
 351                  * unless perhaps the user has mpin'ed the pages
 352                  * themselves.
 353                  */
 354                 qib_devinfo(dd->pcidev,
 355                          "Failed to lock addr %p, %u pages: "
 356                          "errno %d\n", (void *) vaddr, cnt, -ret);
 357                 goto done;
 358         }
 359         for (i = 0; i < cnt; i++, vaddr += PAGE_SIZE) {
 360                 for (; ntids--; tid++) {
 361                         if (tid == tidcnt)
 362                                 tid = 0;
 363                         if (!dd->pageshadow[ctxttid + tid])
 364                                 break;
 365                 }
 366                 if (ntids < 0) {
 367                         /*
 368                          * Oops, wrapped all the way through their TIDs,
 369                          * and didn't have enough free; see comments at
 370                          * start of routine
 371                          */
 372                         i--;    /* last tidlist[i] not filled in */
 373                         ret = -ENOMEM;
 374                         break;
 375                 }
 376                 tidlist[i] = tid + tidoff;
 377                 /* we "know" system pages and TID pages are same size */
 378                 dd->pageshadow[ctxttid + tid] = pagep[i];
 379                 dd->physshadow[ctxttid + tid] =
 380                         qib_map_page(dd->pcidev, pagep[i], 0, PAGE_SIZE,
 381                                      PCI_DMA_FROMDEVICE);
 382                 /*
 383                  * don't need atomic or it's overhead
 384                  */
 385                 __set_bit(tid, tidmap);
 386                 physaddr = dd->physshadow[ctxttid + tid];
 387                 /* PERFORMANCE: below should almost certainly be cached */
 388                 dd->f_put_tid(dd, &tidbase[tid],
 389                                   RCVHQ_RCV_TYPE_EXPECTED, physaddr);
 390                 /*
 391                  * don't check this tid in qib_ctxtshadow, since we
 392                  * just filled it in; start with the next one.
 393                  */
 394                 tid++;
 395         }
 396
 397         if (ret) {
 398                 u32 limit;
 399 cleanup:
 400                 /* jump here if copy out of updated info failed... */
 401                 /* same code that's in qib_free_tid() */
 402                 limit = sizeof(tidmap) * BITS_PER_BYTE;
 403                 if (limit > tidcnt)
 404                         /* just in case size changes in future */
 405                         limit = tidcnt;
 406                 tid = find_first_bit((const unsigned long *)tidmap, limit);
 407                 for (; tid < limit; tid++) {
 408                         if (!test_bit(tid, tidmap))
 409                                 continue;
 410                         if (dd->pageshadow[ctxttid + tid]) {
 411                                 dma_addr_t phys;
 412
 413                                 phys = dd->physshadow[ctxttid + tid];
 414                                 dd->physshadow[ctxttid + tid] = dd->tidinvalid;
 415                                 /* PERFORMANCE: below should almost certainly
 416                                  * be cached
 417                                  */
 418                                 dd->f_put_tid(dd, &tidbase[tid],
 419                                               RCVHQ_RCV_TYPE_EXPECTED,
 420                                               dd->tidinvalid);
 421                                 pci_unmap_page(dd->pcidev, phys, PAGE_SIZE,
 422                                                PCI_DMA_FROMDEVICE);
 423                                 dd->pageshadow[ctxttid + tid] = NULL;
 424                         }
 425                 }
 426                 qib_release_user_pages(pagep, cnt);
 427         } else {
 428                 /*
 429                  * Copy the updated array, with qib_tid's filled in, back
 430                  * to user.  Since we did the copy in already, this "should
 431                  * never fail" If it does, we have to clean up...
 432                  */
 433                 if (copy_to_user((void __user *)
 434                                  (unsigned long) ti->tidlist,
 435                                  tidlist, cnt * sizeof(*tidlist))) {
 436                         ret = -EFAULT;
 437                         goto cleanup;
 438                 }
 439                 if (copy_to_user((void __user *) (unsigned long) ti->tidmap,
 440                                  tidmap, sizeof tidmap)) {
 441                         ret = -EFAULT;
 442                         goto cleanup;
 443                 }
 444                 if (tid == tidcnt)
 445                         tid = 0;
 446                 if (!rcd->subctxt_cnt)
 447                         rcd->tidcursor = tid;
 448                 else
 449                         tidcursor_fp(fp) = tid;
 450         }
 451
 452 done:
 453         return ret;
 454 }
 455
 456 /**
 457  * qib_tid_free - free a context TID
 458  * @rcd: the context
 459  * @subctxt: the subcontext
 460  * @ti: the TID info
 461  *
 462  * right now we are unlocking one page at a time, but since
 463  * the intended use of this routine is for a single group of
 464  * virtually contiguous pages, that should change to improve
 465  * performance.  We check that the TID is in range for this context
 466  * but otherwise don't check validity; if user has an error and
 467  * frees the wrong tid, it's only their own data that can thereby
 468  * be corrupted.  We do check that the TID was in use, for sanity
 469  * We always use our idea of the saved address, not the address that
 470  * they pass in to us.
 471  */
 472 static int qib_tid_free(struct qib_ctxtdata *rcd, unsigned subctxt,
 473                         const struct qib_tid_info *ti)
 474 {
 475         int ret = 0;
 476         u32 tid, ctxttid, cnt, limit, tidcnt;
 477         struct qib_devdata *dd = rcd->dd;
 478         u64 __iomem *tidbase;
 479         unsigned long tidmap[8];
 480
 481         if (!dd->pageshadow) {
 482                 ret = -ENOMEM;
 483                 goto done;
 484         }
 485
 486         if (copy_from_user(tidmap, (void __user *)(unsigned long)ti->tidmap,
 487                            sizeof tidmap)) {
 488                 ret = -EFAULT;
 489                 goto done;
 490         }
 491
 492         ctxttid = rcd->ctxt * dd->rcvtidcnt;
 493         if (!rcd->subctxt_cnt)
 494                 tidcnt = dd->rcvtidcnt;
 495         else if (!subctxt) {
 496                 tidcnt = (dd->rcvtidcnt / rcd->subctxt_cnt) +
 497                          (dd->rcvtidcnt % rcd->subctxt_cnt);
 498                 ctxttid += dd->rcvtidcnt - tidcnt;
 499         } else {
 500                 tidcnt = dd->rcvtidcnt / rcd->subctxt_cnt;
 501                 ctxttid += tidcnt * (subctxt - 1);
 502         }
 503         tidbase = (u64 __iomem *) ((char __iomem *)(dd->kregbase) +
 504                                    dd->rcvtidbase +
 505                                    ctxttid * sizeof(*tidbase));
 506
 507         limit = sizeof(tidmap) * BITS_PER_BYTE;
 508         if (limit > tidcnt)
 509                 /* just in case size changes in future */
 510                 limit = tidcnt;
 511         tid = find_first_bit(tidmap, limit);
 512         for (cnt = 0; tid < limit; tid++) {
 513                 /*
 514                  * small optimization; if we detect a run of 3 or so without
 515                  * any set, use find_first_bit again.  That's mainly to
 516                  * accelerate the case where we wrapped, so we have some at
 517                  * the beginning, and some at the end, and a big gap
 518                  * in the middle.
 519                  */
 520                 if (!test_bit(tid, tidmap))
 521                         continue;
 522                 cnt++;
 523                 if (dd->pageshadow[ctxttid + tid]) {
 524                         struct page *p;
 525                         dma_addr_t phys;
 526
 527                         p = dd->pageshadow[ctxttid + tid];
 528                         dd->pageshadow[ctxttid + tid] = NULL;
 529                         phys = dd->physshadow[ctxttid + tid];
 530                         dd->physshadow[ctxttid + tid] = dd->tidinvalid;
 531                         /* PERFORMANCE: below should almost certainly be
 532                          * cached
 533                          */
 534                         dd->f_put_tid(dd, &tidbase[tid],
 535                                       RCVHQ_RCV_TYPE_EXPECTED, dd->tidinvalid);
 536                         pci_unmap_page(dd->pcidev, phys, PAGE_SIZE,
 537                                        PCI_DMA_FROMDEVICE);
 538                         qib_release_user_pages(&p, 1);
 539                 }
 540         }
 541 done:
 542         return ret;
 543 }
 544
 545 /**
 546  * qib_set_part_key - set a partition key
 547  * @rcd: the context
 548  * @key: the key
 549  *
 550  * We can have up to 4 active at a time (other than the default, which is
 551  * always allowed).  This is somewhat tricky, since multiple contexts may set
 552  * the same key, so we reference count them, and clean up at exit.  All 4
 553  * partition keys are packed into a single qlogic_ib register.  It's an
 554  * error for a process to set the same pkey multiple times.  We provide no
 555  * mechanism to de-allocate a pkey at this time, we may eventually need to
 556  * do that.  I've used the atomic operations, and no locking, and only make
 557  * a single pass through what's available.  This should be more than
 558  * adequate for some time. I'll think about spinlocks or the like if and as
 559  * it's necessary.
 560  */
 561 static int qib_set_part_key(struct qib_ctxtdata *rcd, u16 key)
 562 {
 563         struct qib_pportdata *ppd = rcd->ppd;
 564         int i, any = 0, pidx = -1;
 565         u16 lkey = key & 0x7FFF;
 566         int ret;
 567
 568         if (lkey == (QIB_DEFAULT_P_KEY & 0x7FFF)) {
 569                 /* nothing to do; this key always valid */
 570                 ret = 0;
 571                 goto bail;
 572         }
 573
 574         if (!lkey) {
 575                 ret = -EINVAL;
 576                 goto bail;
 577         }
 578
 579         /*
 580          * Set the full membership bit, because it has to be
 581          * set in the register or the packet, and it seems
 582          * cleaner to set in the register than to force all
 583          * callers to set it.
 584          */
 585         key |= 0x8000;
 586
 587         for (i = 0; i < ARRAY_SIZE(rcd->pkeys); i++) {
 588                 if (!rcd->pkeys[i] && pidx == -1)
 589                         pidx = i;
 590                 if (rcd->pkeys[i] == key) {
 591                         ret = -EEXIST;
 592                         goto bail;
 593                 }
 594         }
 595         if (pidx == -1) {
 596                 ret = -EBUSY;
 597                 goto bail;
 598         }
 599         for (any = i = 0; i < ARRAY_SIZE(ppd->pkeys); i++) {
 600                 if (!ppd->pkeys[i]) {
 601                         any++;
 602                         continue;
 603                 }
 604                 if (ppd->pkeys[i] == key) {
 605                         atomic_t *pkrefs = &ppd->pkeyrefs[i];
 606
 607                         if (atomic_inc_return(pkrefs) > 1) {
 608                                 rcd->pkeys[pidx] = key;
 609                                 ret = 0;
 610                                 goto bail;
 611                         } else {
 612                                 /*
 613                                  * lost race, decrement count, catch below
 614                                  */
 615                                 atomic_dec(pkrefs);
 616                                 any++;
 617                         }
 618                 }
 619                 if ((ppd->pkeys[i] & 0x7FFF) == lkey) {
 620                         /*
 621                          * It makes no sense to have both the limited and
 622                          * full membership PKEY set at the same time since
 623                          * the unlimited one will disable the limited one.
 624                          */
 625                         ret = -EEXIST;
 626                         goto bail;
 627                 }
 628         }
 629         if (!any) {
 630                 ret = -EBUSY;
 631                 goto bail;
 632         }
 633         for (any = i = 0; i < ARRAY_SIZE(ppd->pkeys); i++) {
 634                 if (!ppd->pkeys[i] &&
 635                     atomic_inc_return(&ppd->pkeyrefs[i]) == 1) {
 636                         rcd->pkeys[pidx] = key;
 637                         ppd->pkeys[i] = key;
 638                         (void) ppd->dd->f_set_ib_cfg(ppd, QIB_IB_CFG_PKEYS, 0);
 639                         ret = 0;
 640                         goto bail;
 641                 }
 642         }
 643         ret = -EBUSY;
 644
 645 bail:
 646         return ret;
 647 }
 648
 649 /**
 650  * qib_manage_rcvq - manage a context's receive queue
 651  * @rcd: the context
 652  * @subctxt: the subcontext
 653  * @start_stop: action to carry out
 654  *
 655  * start_stop == 0 disables receive on the context, for use in queue
 656  * overflow conditions.  start_stop==1 re-enables, to be used to
 657  * re-init the software copy of the head register
 658  */
 659 static int qib_manage_rcvq(struct qib_ctxtdata *rcd, unsigned subctxt,
 660                            int start_stop)
 661 {
 662         struct qib_devdata *dd = rcd->dd;
 663         unsigned int rcvctrl_op;
 664
 665         if (subctxt)
 666                 goto bail;
 667         /* atomically clear receive enable ctxt. */
 668         if (start_stop) {
 669                 /*
 670                  * On enable, force in-memory copy of the tail register to
 671                  * 0, so that protocol code doesn't have to worry about
 672                  * whether or not the chip has yet updated the in-memory
 673                  * copy or not on return from the system call. The chip
 674                  * always resets it's tail register back to 0 on a
 675                  * transition from disabled to enabled.
 676                  */
 677                 if (rcd->rcvhdrtail_kvaddr)
 678                         qib_clear_rcvhdrtail(rcd);
 679                 rcvctrl_op = QIB_RCVCTRL_CTXT_ENB;
 680         } else
 681                 rcvctrl_op = QIB_RCVCTRL_CTXT_DIS;
 682         dd->f_rcvctrl(rcd->ppd, rcvctrl_op, rcd->ctxt);
 683         /* always; new head should be equal to new tail; see above */
 684 bail:
 685         return 0;
 686 }
 687
 688 static void qib_clean_part_key(struct qib_ctxtdata *rcd,
 689                                struct qib_devdata *dd)
 690 {
 691         int i, j, pchanged = 0;
 692         u64 oldpkey;
 693         struct qib_pportdata *ppd = rcd->ppd;
 694
 695         /* for debugging only */
 696         oldpkey = (u64) ppd->pkeys[0] |
 697                 ((u64) ppd->pkeys[1] << 16) |
 698                 ((u64) ppd->pkeys[2] << 32) |
 699                 ((u64) ppd->pkeys[3] << 48);
 700
 701         for (i = 0; i < ARRAY_SIZE(rcd->pkeys); i++) {
 702                 if (!rcd->pkeys[i])
 703                         continue;
 704                 for (j = 0; j < ARRAY_SIZE(ppd->pkeys); j++) {
 705                         /* check for match independent of the global bit */
 706                         if ((ppd->pkeys[j] & 0x7fff) !=
 707                             (rcd->pkeys[i] & 0x7fff))
 708                                 continue;
 709                         if (atomic_dec_and_test(&ppd->pkeyrefs[j])) {
 710                                 ppd->pkeys[j] = 0;
 711                                 pchanged++;
 712                         }
 713                         break;
 714                 }
 715                 rcd->pkeys[i] = 0;
 716         }
 717         if (pchanged)
 718                 (void) ppd->dd->f_set_ib_cfg(ppd, QIB_IB_CFG_PKEYS, 0);
 719 }
 720
 721 /* common code for the mappings on dma_alloc_coherent mem */
 722 static int qib_mmap_mem(struct vm_area_struct *vma, struct qib_ctxtdata *rcd,
 723                         unsigned len, void *kvaddr, u32 write_ok, char *what)
 724 {
 725         struct qib_devdata *dd = rcd->dd;
 726         unsigned long pfn;
 727         int ret;
 728
 729         if ((vma->vm_end - vma->vm_start) > len) {
 730                 qib_devinfo(dd->pcidev,
 731                          "FAIL on %s: len %lx > %x\n", what,
 732                          vma->vm_end - vma->vm_start, len);
 733                 ret = -EFAULT;
 734                 goto bail;
 735         }
 736
 737         /*
 738          * shared context user code requires rcvhdrq mapped r/w, others
 739          * only allowed readonly mapping.
 740          */
 741         if (!write_ok) {
 742                 if (vma->vm_flags & VM_WRITE) {
 743                         qib_devinfo(dd->pcidev,
 744                                  "%s must be mapped readonly\n", what);
 745                         ret = -EPERM;
 746                         goto bail;
 747                 }
 748
 749                 /* don't allow them to later change with mprotect */
 750                 vma->vm_flags &= ~VM_MAYWRITE;
 751         }
 752
 753         pfn = virt_to_phys(kvaddr) >> PAGE_SHIFT;
 754         ret = remap_pfn_range(vma, vma->vm_start, pfn,
 755                               len, vma->vm_page_prot);
 756         if (ret)
 757                 qib_devinfo(dd->pcidev,
 758                         "%s ctxt%u mmap of %lx, %x bytes failed: %d\n",
 759                         what, rcd->ctxt, pfn, len, ret);
 760 bail:
 761         return ret;
 762 }
 763
 764 static int mmap_ureg(struct vm_area_struct *vma, struct qib_devdata *dd,
 765                      u64 ureg)
 766 {
 767         unsigned long phys;
 768         unsigned long sz;
 769         int ret;
 770
 771         /*
 772          * This is real hardware, so use io_remap.  This is the mechanism
 773          * for the user process to update the head registers for their ctxt
 774          * in the chip.
 775          */
 776         sz = dd->flags & QIB_HAS_HDRSUPP ? 2 * PAGE_SIZE : PAGE_SIZE;
 777         if ((vma->vm_end - vma->vm_start) > sz) {
 778                 qib_devinfo(dd->pcidev,
 779                         "FAIL mmap userreg: reqlen %lx > PAGE\n",
 780                         vma->vm_end - vma->vm_start);
 781                 ret = -EFAULT;
 782         } else {
 783                 phys = dd->physaddr + ureg;
 784                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 785
 786                 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
 787                 ret = io_remap_pfn_range(vma, vma->vm_start,
 788                                          phys >> PAGE_SHIFT,
 789                                          vma->vm_end - vma->vm_start,
 790                                          vma->vm_page_prot);
 791         }
 792         return ret;
 793 }
 794
 795 static int mmap_piobufs(struct vm_area_struct *vma,
 796                         struct qib_devdata *dd,
 797                         struct qib_ctxtdata *rcd,
 798                         unsigned piobufs, unsigned piocnt)
 799 {
 800         unsigned long phys;
 801         int ret;
 802
 803         /*
 804          * When we map the PIO buffers in the chip, we want to map them as
 805          * writeonly, no read possible; unfortunately, x86 doesn't allow
 806          * for this in hardware, but we still prevent users from asking
 807          * for it.
 808          */
 809         if ((vma->vm_end - vma->vm_start) > (piocnt * dd->palign)) {
 810                 qib_devinfo(dd->pcidev,
 811                         "FAIL mmap piobufs: reqlen %lx > PAGE\n",
 812                          vma->vm_end - vma->vm_start);
 813                 ret = -EINVAL;
 814                 goto bail;
 815         }
 816
 817         phys = dd->physaddr + piobufs;
 818
 819 #if defined(__powerpc__)
 820         /* There isn't a generic way to specify writethrough mappings */
 821         pgprot_val(vma->vm_page_prot) |= _PAGE_NO_CACHE;
 822         pgprot_val(vma->vm_page_prot) |= _PAGE_WRITETHRU;
 823         pgprot_val(vma->vm_page_prot) &= ~_PAGE_GUARDED;
 824 #endif
 825
 826         /*
 827          * don't allow them to later change to readable with mprotect (for when
 828          * not initially mapped readable, as is normally the case)
 829          */
 830         vma->vm_flags &= ~VM_MAYREAD;
 831         vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
 832
 833         if (qib_wc_pat)
 834                 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
 835
 836         ret = io_remap_pfn_range(vma, vma->vm_start, phys >> PAGE_SHIFT,
 837                                  vma->vm_end - vma->vm_start,
 838                                  vma->vm_page_prot);
 839 bail:
 840         return ret;
 841 }
 842
 843 static int mmap_rcvegrbufs(struct vm_area_struct *vma,
 844                            struct qib_ctxtdata *rcd)
 845 {
 846         struct qib_devdata *dd = rcd->dd;
 847         unsigned long start, size;
 848         size_t total_size, i;
 849         unsigned long pfn;
 850         int ret;
 851
 852         size = rcd->rcvegrbuf_size;
 853         total_size = rcd->rcvegrbuf_chunks * size;
 854         if ((vma->vm_end - vma->vm_start) > total_size) {
 855                 qib_devinfo(dd->pcidev,
 856                         "FAIL on egr bufs: reqlen %lx > actual %lx\n",
 857                          vma->vm_end - vma->vm_start,
 858                          (unsigned long) total_size);
 859                 ret = -EINVAL;
 860                 goto bail;
 861         }
 862
 863         if (vma->vm_flags & VM_WRITE) {
 864                 qib_devinfo(dd->pcidev,
 865                         "Can't map eager buffers as writable (flags=%lx)\n",
 866                         vma->vm_flags);
 867                 ret = -EPERM;
 868                 goto bail;
 869         }
 870         /* don't allow them to later change to writeable with mprotect */
 871         vma->vm_flags &= ~VM_MAYWRITE;
 872
 873         start = vma->vm_start;
 874
 875         for (i = 0; i < rcd->rcvegrbuf_chunks; i++, start += size) {
 876                 pfn = virt_to_phys(rcd->rcvegrbuf[i]) >> PAGE_SHIFT;
 877                 ret = remap_pfn_range(vma, start, pfn, size,
 878                                       vma->vm_page_prot);
 879                 if (ret < 0)
 880                         goto bail;
 881         }
 882         ret = 0;
 883
 884 bail:
 885         return ret;
 886 }
 887
 888 /*
 889  * qib_file_vma_fault - handle a VMA page fault.
 890  */
 891 static int qib_file_vma_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 892 {
 893         struct page *page;
 894
 895         page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
 896         if (!page)
 897                 return VM_FAULT_SIGBUS;
 898
 899         get_page(page);
 900         vmf->page = page;
 901
 902         return 0;
 903 }
 904
 905 static struct vm_operations_struct qib_file_vm_ops = {
 906         .fault = qib_file_vma_fault,
 907 };
 908
 909 static int mmap_kvaddr(struct vm_area_struct *vma, u64 pgaddr,
 910                        struct qib_ctxtdata *rcd, unsigned subctxt)
 911 {
 912         struct qib_devdata *dd = rcd->dd;
 913         unsigned subctxt_cnt;
 914         unsigned long len;
 915         void *addr;
 916         size_t size;
 917         int ret = 0;
 918
 919         subctxt_cnt = rcd->subctxt_cnt;
 920         size = rcd->rcvegrbuf_chunks * rcd->rcvegrbuf_size;
 921
 922         /*
 923          * Each process has all the subctxt uregbase, rcvhdrq, and
 924          * rcvegrbufs mmapped - as an array for all the processes,
 925          * and also separately for this process.
 926          */
 927         if (pgaddr == cvt_kvaddr(rcd->subctxt_uregbase)) {
 928                 addr = rcd->subctxt_uregbase;
 929                 size = PAGE_SIZE * subctxt_cnt;
 930         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_rcvhdr_base)) {
 931                 addr = rcd->subctxt_rcvhdr_base;
 932                 size = rcd->rcvhdrq_size * subctxt_cnt;
 933         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_rcvegrbuf)) {
 934                 addr = rcd->subctxt_rcvegrbuf;
 935                 size *= subctxt_cnt;
 936         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_uregbase +
 937                                         PAGE_SIZE * subctxt)) {
 938                 addr = rcd->subctxt_uregbase + PAGE_SIZE * subctxt;
 939                 size = PAGE_SIZE;
 940         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_rcvhdr_base +
 941                                         rcd->rcvhdrq_size * subctxt)) {
 942                 addr = rcd->subctxt_rcvhdr_base +
 943                         rcd->rcvhdrq_size * subctxt;
 944                 size = rcd->rcvhdrq_size;
 945         } else if (pgaddr == cvt_kvaddr(&rcd->user_event_mask[subctxt])) {
 946                 addr = rcd->user_event_mask;
 947                 size = PAGE_SIZE;
 948         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_rcvegrbuf +
 949                                         size * subctxt)) {
 950                 addr = rcd->subctxt_rcvegrbuf + size * subctxt;
 951                 /* rcvegrbufs are read-only on the slave */
 952                 if (vma->vm_flags & VM_WRITE) {
 953                         qib_devinfo(dd->pcidev,
 954                                  "Can't map eager buffers as "
 955                                  "writable (flags=%lx)\n", vma->vm_flags);
 956                         ret = -EPERM;
 957                         goto bail;
 958                 }
 959                 /*
 960                  * Don't allow permission to later change to writeable
 961                  * with mprotect.
 962                  */
 963                 vma->vm_flags &= ~VM_MAYWRITE;
 964         } else
 965                 goto bail;
 966         len = vma->vm_end - vma->vm_start;
 967         if (len > size) {
 968                 ret = -EINVAL;
 969                 goto bail;
 970         }
 971
 972         vma->vm_pgoff = (unsigned long) addr >> PAGE_SHIFT;
 973         vma->vm_ops = &qib_file_vm_ops;
 974         vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
 975         ret = 1;
 976
 977 bail:
 978         return ret;
 979 }
 980
 981 /**
 982  * qib_mmapf - mmap various structures into user space
 983  * @fp: the file pointer
 984  * @vma: the VM area
 985  *
 986  * We use this to have a shared buffer between the kernel and the user code
 987  * for the rcvhdr queue, egr buffers, and the per-context user regs and pio
 988  * buffers in the chip.  We have the open and close entries so we can bump
 989  * the ref count and keep the driver from being unloaded while still mapped.
 990  */
 991 static int qib_mmapf(struct file *fp, struct vm_area_struct *vma)
 992 {
 993         struct qib_ctxtdata *rcd;
 994         struct qib_devdata *dd;
 995         u64 pgaddr, ureg;
 996         unsigned piobufs, piocnt;
 997         int ret, match = 1;
 998
 999         rcd = ctxt_fp(fp);
1000         if (!rcd || !(vma->vm_flags & VM_SHARED)) {
1001                 ret = -EINVAL;
1002                 goto bail;
1003         }
1004         dd = rcd->dd;
1005
1006         /*
1007          * This is the qib_do_user_init() code, mapping the shared buffers
1008          * and per-context user registers into the user process. The address
1009          * referred to by vm_pgoff is the file offset passed via mmap().
1010          * For shared contexts, this is the kernel vmalloc() address of the
1011          * pages to share with the master.
1012          * For non-shared or master ctxts, this is a physical address.
1013          * We only do one mmap for each space mapped.
1014          */
1015         pgaddr = vma->vm_pgoff << PAGE_SHIFT;
1016
1017         /*
1018          * Check for 0 in case one of the allocations failed, but user
1019          * called mmap anyway.
1020          */
1021         if (!pgaddr)  {
1022                 ret = -EINVAL;
1023                 goto bail;
1024         }
1025
1026         /*
1027          * Physical addresses must fit in 40 bits for our hardware.
1028          * Check for kernel virtual addresses first, anything else must
1029          * match a HW or memory address.
1030          */
1031         ret = mmap_kvaddr(vma, pgaddr, rcd, subctxt_fp(fp));
1032         if (ret) {
1033                 if (ret > 0)
1034                         ret = 0;
1035                 goto bail;
1036         }
1037
1038         ureg = dd->uregbase + dd->ureg_align * rcd->ctxt;
1039         if (!rcd->subctxt_cnt) {
1040                 /* ctxt is not shared */
1041                 piocnt = rcd->piocnt;
1042                 piobufs = rcd->piobufs;
1043         } else if (!subctxt_fp(fp)) {
1044                 /* caller is the master */
1045                 piocnt = (rcd->piocnt / rcd->subctxt_cnt) +
1046                          (rcd->piocnt % rcd->subctxt_cnt);
1047                 piobufs = rcd->piobufs +
1048                         dd->palign * (rcd->piocnt - piocnt);
1049         } else {
1050                 unsigned slave = subctxt_fp(fp) - 1;
1051
1052                 /* caller is a slave */
1053                 piocnt = rcd->piocnt / rcd->subctxt_cnt;
1054                 piobufs = rcd->piobufs + dd->palign * piocnt * slave;
1055         }
1056
1057         if (pgaddr == ureg)
1058                 ret = mmap_ureg(vma, dd, ureg);
1059         else if (pgaddr == piobufs)
1060                 ret = mmap_piobufs(vma, dd, rcd, piobufs, piocnt);
1061         else if (pgaddr == dd->pioavailregs_phys)
1062                 /* in-memory copy of pioavail registers */
1063                 ret = qib_mmap_mem(vma, rcd, PAGE_SIZE,
1064                                    (void *) dd->pioavailregs_dma, 0,
1065                                    "pioavail registers");
1066         else if (pgaddr == rcd->rcvegr_phys)
1067                 ret = mmap_rcvegrbufs(vma, rcd);
1068         else if (pgaddr == (u64) rcd->rcvhdrq_phys)
1069                 /*
1070                  * The rcvhdrq itself; multiple pages, contiguous
1071                  * from an i/o perspective.  Shared contexts need
1072                  * to map r/w, so we allow writing.
1073                  */
1074                 ret = qib_mmap_mem(vma, rcd, rcd->rcvhdrq_size,
1075                                    rcd->rcvhdrq, 1, "rcvhdrq");
1076         else if (pgaddr == (u64) rcd->rcvhdrqtailaddr_phys)
1077                 /* in-memory copy of rcvhdrq tail register */
1078                 ret = qib_mmap_mem(vma, rcd, PAGE_SIZE,
1079                                    rcd->rcvhdrtail_kvaddr, 0,
1080                                    "rcvhdrq tail");
1081         else
1082                 match = 0;
1083         if (!match)
1084                 ret = -EINVAL;
1085
1086         vma->vm_private_data = NULL;
1087
1088         if (ret < 0)
1089                 qib_devinfo(dd->pcidev,
1090                          "mmap Failure %d: off %llx len %lx\n",
1091                          -ret, (unsigned long long)pgaddr,
1092                          vma->vm_end - vma->vm_start);
1093 bail:
1094         return ret;
1095 }
1096
1097 static unsigned int qib_poll_urgent(struct qib_ctxtdata *rcd,
1098                                     struct file *fp,
1099                                     struct poll_table_struct *pt)
1100 {
1101         struct qib_devdata *dd = rcd->dd;
1102         unsigned pollflag;
1103
1104         poll_wait(fp, &rcd->wait, pt);
1105
1106         spin_lock_irq(&dd->uctxt_lock);
1107         if (rcd->urgent != rcd->urgent_poll) {
1108                 pollflag = POLLIN | POLLRDNORM;
1109                 rcd->urgent_poll = rcd->urgent;
1110         } else {
1111                 pollflag = 0;
1112                 set_bit(QIB_CTXT_WAITING_URG, &rcd->flag);
1113         }
1114         spin_unlock_irq(&dd->uctxt_lock);
1115
1116         return pollflag;
1117 }
1118
1119 static unsigned int qib_poll_next(struct qib_ctxtdata *rcd,
1120                                   struct file *fp,
1121                                   struct poll_table_struct *pt)
1122 {
1123         struct qib_devdata *dd = rcd->dd;
1124         unsigned pollflag;
1125
1126         poll_wait(fp, &rcd->wait, pt);
1127
1128         spin_lock_irq(&dd->uctxt_lock);
1129         if (dd->f_hdrqempty(rcd)) {
1130                 set_bit(QIB_CTXT_WAITING_RCV, &rcd->flag);
1131                 dd->f_rcvctrl(rcd->ppd, QIB_RCVCTRL_INTRAVAIL_ENB, rcd->ctxt);
1132                 pollflag = 0;
1133         } else
1134                 pollflag = POLLIN | POLLRDNORM;
1135         spin_unlock_irq(&dd->uctxt_lock);
1136
1137         return pollflag;
1138 }
1139
1140 static unsigned int qib_poll(struct file *fp, struct poll_table_struct *pt)
1141 {
1142         struct qib_ctxtdata *rcd;
1143         unsigned pollflag;
1144
1145         rcd = ctxt_fp(fp);
1146         if (!rcd)
1147                 pollflag = POLLERR;
1148         else if (rcd->poll_type == QIB_POLL_TYPE_URGENT)
1149                 pollflag = qib_poll_urgent(rcd, fp, pt);
1150         else  if (rcd->poll_type == QIB_POLL_TYPE_ANYRCV)
1151                 pollflag = qib_poll_next(rcd, fp, pt);
1152         else /* invalid */
1153                 pollflag = POLLERR;
1154
1155         return pollflag;
1156 }
1157
1158 static void assign_ctxt_affinity(struct file *fp, struct qib_devdata *dd)
1159 {
1160         struct qib_filedata *fd = fp->private_data;
1161         const unsigned int weight = cpumask_weight(&current->cpus_allowed);
1162         const struct cpumask *local_mask = cpumask_of_pcibus(dd->pcidev->bus);
1163         int local_cpu;
1164
1165         /*
1166          * If process has NOT already set it's affinity, select and
1167          * reserve a processor for it on the local NUMA node.
1168          */
1169         if ((weight >= qib_cpulist_count) &&
1170                 (cpumask_weight(local_mask) <= qib_cpulist_count)) {
1171                 for_each_cpu(local_cpu, local_mask)
1172                         if (!test_and_set_bit(local_cpu, qib_cpulist)) {
1173                                 fd->rec_cpu_num = local_cpu;
1174                                 return;
1175                         }
1176         }
1177
1178         /*
1179          * If process has NOT already set it's affinity, select and
1180          * reserve a processor for it, as a rendevous for all
1181          * users of the driver.  If they don't actually later
1182          * set affinity to this cpu, or set it to some other cpu,
1183          * it just means that sooner or later we don't recommend
1184          * a cpu, and let the scheduler do it's best.
1185          */
1186         if (weight >= qib_cpulist_count) {
1187                 int cpu;
1188                 cpu = find_first_zero_bit(qib_cpulist,
1189                                           qib_cpulist_count);
1190                 if (cpu == qib_cpulist_count)
1191                         qib_dev_err(dd,
1192                         "no cpus avail for affinity PID %u\n",
1193                         current->pid);
1194                 else {
1195                         __set_bit(cpu, qib_cpulist);
1196                         fd->rec_cpu_num = cpu;
1197                 }
1198         }
1199 }
1200
1201 /*
1202  * Check that userland and driver are compatible for subcontexts.
1203  */
1204 static int qib_compatible_subctxts(int user_swmajor, int user_swminor)
1205 {
1206         /* this code is written long-hand for clarity */
1207         if (QIB_USER_SWMAJOR != user_swmajor) {
1208                 /* no promise of compatibility if major mismatch */
1209                 return 0;
1210         }
1211         if (QIB_USER_SWMAJOR == 1) {
1212                 switch (QIB_USER_SWMINOR) {
1213                 case 0:
1214                 case 1:
1215                 case 2:
1216                         /* no subctxt implementation so cannot be compatible */
1217                         return 0;
1218                 case 3:
1219                         /* 3 is only compatible with itself */
1220                         return user_swminor == 3;
1221                 default:
1222                         /* >= 4 are compatible (or are expected to be) */
1223                         return user_swminor <= QIB_USER_SWMINOR;
1224                 }
1225         }
1226         /* make no promises yet for future major versions */
1227         return 0;
1228 }
1229
1230 static int init_subctxts(struct qib_devdata *dd,
1231                          struct qib_ctxtdata *rcd,
1232                          const struct qib_user_info *uinfo)
1233 {
1234         int ret = 0;
1235         unsigned num_subctxts;
1236         size_t size;
1237
1238         /*
1239          * If the user is requesting zero subctxts,
1240          * skip the subctxt allocation.
1241          */
1242         if (uinfo->spu_subctxt_cnt <= 0)
1243                 goto bail;
1244         num_subctxts = uinfo->spu_subctxt_cnt;
1245
1246         /* Check for subctxt compatibility */
1247         if (!qib_compatible_subctxts(uinfo->spu_userversion >> 16,
1248                 uinfo->spu_userversion & 0xffff)) {
1249                 qib_devinfo(dd->pcidev,
1250                          "Mismatched user version (%d.%d) and driver "
1251                          "version (%d.%d) while context sharing. Ensure "
1252                          "that driver and library are from the same "
1253                          "release.\n",
1254                          (int) (uinfo->spu_userversion >> 16),
1255                          (int) (uinfo->spu_userversion & 0xffff),
1256                          QIB_USER_SWMAJOR, QIB_USER_SWMINOR);
1257                 goto bail;
1258         }
1259         if (num_subctxts > QLOGIC_IB_MAX_SUBCTXT) {
1260                 ret = -EINVAL;
1261                 goto bail;
1262         }
1263
1264         rcd->subctxt_uregbase = vmalloc_user(PAGE_SIZE * num_subctxts);
1265         if (!rcd->subctxt_uregbase) {
1266                 ret = -ENOMEM;
1267                 goto bail;
1268         }
1269         /* Note: rcd->rcvhdrq_size isn't initialized yet. */
1270         size = ALIGN(dd->rcvhdrcnt * dd->rcvhdrentsize *
1271                      sizeof(u32), PAGE_SIZE) * num_subctxts;
1272         rcd->subctxt_rcvhdr_base = vmalloc_user(size);
1273         if (!rcd->subctxt_rcvhdr_base) {
1274                 ret = -ENOMEM;
1275                 goto bail_ureg;
1276         }
1277
1278         rcd->subctxt_rcvegrbuf = vmalloc_user(rcd->rcvegrbuf_chunks *
1279                                               rcd->rcvegrbuf_size *
1280                                               num_subctxts);
1281         if (!rcd->subctxt_rcvegrbuf) {
1282                 ret = -ENOMEM;
1283                 goto bail_rhdr;
1284         }
1285
1286         rcd->subctxt_cnt = uinfo->spu_subctxt_cnt;
1287         rcd->subctxt_id = uinfo->spu_subctxt_id;
1288         rcd->active_slaves = 1;
1289         rcd->redirect_seq_cnt = 1;
1290         set_bit(QIB_CTXT_MASTER_UNINIT, &rcd->flag);
1291         goto bail;
1292
1293 bail_rhdr:
1294         vfree(rcd->subctxt_rcvhdr_base);
1295 bail_ureg:
1296         vfree(rcd->subctxt_uregbase);
1297         rcd->subctxt_uregbase = NULL;
1298 bail:
1299         return ret;
1300 }
1301
1302 static int setup_ctxt(struct qib_pportdata *ppd, int ctxt,
1303                       struct file *fp, const struct qib_user_info *uinfo)
1304 {
1305         struct qib_filedata *fd = fp->private_data;
1306         struct qib_devdata *dd = ppd->dd;
1307         struct qib_ctxtdata *rcd;
1308         void *ptmp = NULL;
1309         int ret;
1310         int numa_id;
1311
1312         assign_ctxt_affinity(fp, dd);
1313
1314         numa_id = qib_numa_aware ? ((fd->rec_cpu_num != -1) ?
1315                 cpu_to_node(fd->rec_cpu_num) :
1316                 numa_node_id()) : dd->assigned_node_id;
1317
1318         rcd = qib_create_ctxtdata(ppd, ctxt, numa_id);
1319
1320         /*
1321          * Allocate memory for use in qib_tid_update() at open to
1322          * reduce cost of expected send setup per message segment
1323          */
1324         if (rcd)
1325                 ptmp = kmalloc(dd->rcvtidcnt * sizeof(u16) +
1326                                dd->rcvtidcnt * sizeof(struct page **),
1327                                GFP_KERNEL);
1328
1329         if (!rcd || !ptmp) {
1330                 qib_dev_err(dd,
1331                         "Unable to allocate ctxtdata memory, failing open\n");
1332                 ret = -ENOMEM;
1333                 goto bailerr;
1334         }
1335         rcd->userversion = uinfo->spu_userversion;
1336         ret = init_subctxts(dd, rcd, uinfo);
1337         if (ret)
1338                 goto bailerr;
1339         rcd->tid_pg_list = ptmp;
1340         rcd->pid = current->pid;
1341         init_waitqueue_head(&dd->rcd[ctxt]->wait);
1342         strlcpy(rcd->comm, current->comm, sizeof(rcd->comm));
1343         ctxt_fp(fp) = rcd;
1344         qib_stats.sps_ctxts++;
1345         dd->freectxts--;
1346         ret = 0;
1347         goto bail;
1348
1349 bailerr:
1350         if (fd->rec_cpu_num != -1)
1351                 __clear_bit(fd->rec_cpu_num, qib_cpulist);
1352
1353         dd->rcd[ctxt] = NULL;
1354         kfree(rcd);
1355         kfree(ptmp);
1356 bail:
1357         return ret;
1358 }
1359
1360 static inline int usable(struct qib_pportdata *ppd)
1361 {
1362         struct qib_devdata *dd = ppd->dd;
1363
1364         return dd && (dd->flags & QIB_PRESENT) && dd->kregbase && ppd->lid &&
1365                 (ppd->lflags & QIBL_LINKACTIVE);
1366 }
1367
1368 /*
1369  * Select a context on the given device, either using a requested port
1370  * or the port based on the context number.
1371  */
1372 static int choose_port_ctxt(struct file *fp, struct qib_devdata *dd, u32 port,
1373                             const struct qib_user_info *uinfo)
1374 {
1375         struct qib_pportdata *ppd = NULL;
1376         int ret, ctxt;
1377
1378         if (port) {
1379                 if (!usable(dd->pport + port - 1)) {
1380                         ret = -ENETDOWN;
1381                         goto done;
1382                 } else
1383                         ppd = dd->pport + port - 1;
1384         }
1385         for (ctxt = dd->first_user_ctxt; ctxt < dd->cfgctxts && dd->rcd[ctxt];
1386              ctxt++)
1387                 ;
1388         if (ctxt == dd->cfgctxts) {
1389                 ret = -EBUSY;
1390                 goto done;
1391         }
1392         if (!ppd) {
1393                 u32 pidx = ctxt % dd->num_pports;
1394                 if (usable(dd->pport + pidx))
1395                         ppd = dd->pport + pidx;
1396                 else {
1397                         for (pidx = 0; pidx < dd->num_pports && !ppd;
1398                              pidx++)
1399                                 if (usable(dd->pport + pidx))
1400                                         ppd = dd->pport + pidx;
1401                 }
1402         }
1403         ret = ppd ? setup_ctxt(ppd, ctxt, fp, uinfo) : -ENETDOWN;
1404 done:
1405         return ret;
1406 }
1407
1408 static int find_free_ctxt(int unit, struct file *fp,
1409                           const struct qib_user_info *uinfo)
1410 {
1411         struct qib_devdata *dd = qib_lookup(unit);
1412         int ret;
1413
1414         if (!dd || (uinfo->spu_port && uinfo->spu_port > dd->num_pports))
1415                 ret = -ENODEV;
1416         else
1417                 ret = choose_port_ctxt(fp, dd, uinfo->spu_port, uinfo);
1418
1419         return ret;
1420 }
1421
1422 static int get_a_ctxt(struct file *fp, const struct qib_user_info *uinfo,
1423                       unsigned alg)
1424 {
1425         struct qib_devdata *udd = NULL;
1426         int ret = 0, devmax, npresent, nup, ndev, dusable = 0, i;
1427         u32 port = uinfo->spu_port, ctxt;
1428
1429         devmax = qib_count_units(&npresent, &nup);
1430         if (!npresent) {
1431                 ret = -ENXIO;
1432                 goto done;
1433         }
1434         if (nup == 0) {
1435                 ret = -ENETDOWN;
1436                 goto done;
1437         }
1438
1439         if (alg == QIB_PORT_ALG_ACROSS) {
1440                 unsigned inuse = ~0U;
1441                 /* find device (with ACTIVE ports) with fewest ctxts in use */
1442                 for (ndev = 0; ndev < devmax; ndev++) {
1443                         struct qib_devdata *dd = qib_lookup(ndev);
1444                         unsigned cused = 0, cfree = 0, pusable = 0;
1445                         if (!dd)
1446                                 continue;
1447                         if (port && port <= dd->num_pports &&
1448                             usable(dd->pport + port - 1))
1449                                 pusable = 1;
1450                         else
1451                                 for (i = 0; i < dd->num_pports; i++)
1452                                         if (usable(dd->pport + i))
1453                                                 pusable++;
1454                         if (!pusable)
1455                                 continue;
1456                         for (ctxt = dd->first_user_ctxt; ctxt < dd->cfgctxts;
1457                              ctxt++)
1458                                 if (dd->rcd[ctxt])
1459                                         cused++;
1460                                 else
1461                                         cfree++;
1462                         if (cfree && cused < inuse) {
1463                                 udd = dd;
1464                                 inuse = cused;
1465                         }
1466                 }
1467                 if (udd) {
1468                         ret = choose_port_ctxt(fp, udd, port, uinfo);
1469                         goto done;
1470                 }
1471         } else {
1472                 for (ndev = 0; ndev < devmax; ndev++) {
1473                         struct qib_devdata *dd = qib_lookup(ndev);
1474                         if (dd) {
1475                                 ret = choose_port_ctxt(fp, dd, port, uinfo);
1476                                 if (!ret)
1477                                         goto done;
1478                                 if (ret == -EBUSY)
1479                                         dusable++;
1480                         }
1481                 }
1482         }
1483         ret = dusable ? -EBUSY : -ENETDOWN;
1484
1485 done:
1486         return ret;
1487 }
1488
1489 static int find_shared_ctxt(struct file *fp,
1490                             const struct qib_user_info *uinfo)
1491 {
1492         int devmax, ndev, i;
1493         int ret = 0;
1494
1495         devmax = qib_count_units(NULL, NULL);
1496
1497         for (ndev = 0; ndev < devmax; ndev++) {
1498                 struct qib_devdata *dd = qib_lookup(ndev);
1499
1500                 /* device portion of usable() */
1501                 if (!(dd && (dd->flags & QIB_PRESENT) && dd->kregbase))
1502                         continue;
1503                 for (i = dd->first_user_ctxt; i < dd->cfgctxts; i++) {
1504                         struct qib_ctxtdata *rcd = dd->rcd[i];
1505
1506                         /* Skip ctxts which are not yet open */
1507                         if (!rcd || !rcd->cnt)
1508                                 continue;
1509                         /* Skip ctxt if it doesn't match the requested one */
1510                         if (rcd->subctxt_id != uinfo->spu_subctxt_id)
1511                                 continue;
1512                         /* Verify the sharing process matches the master */
1513                         if (rcd->subctxt_cnt != uinfo->spu_subctxt_cnt ||
1514                             rcd->userversion != uinfo->spu_userversion ||
1515                             rcd->cnt >= rcd->subctxt_cnt) {
1516                                 ret = -EINVAL;
1517                                 goto done;
1518                         }
1519                         ctxt_fp(fp) = rcd;
1520                         subctxt_fp(fp) = rcd->cnt++;
1521                         rcd->subpid[subctxt_fp(fp)] = current->pid;
1522                         tidcursor_fp(fp) = 0;
1523                         rcd->active_slaves |= 1 << subctxt_fp(fp);
1524                         ret = 1;
1525                         goto done;
1526                 }
1527         }
1528
1529 done:
1530         return ret;
1531 }
1532
1533 static int qib_open(struct inode *in, struct file *fp)
1534 {
1535         /* The real work is performed later in qib_assign_ctxt() */
1536         fp->private_data = kzalloc(sizeof(struct qib_filedata), GFP_KERNEL);
1537         if (fp->private_data) /* no cpu affinity by default */
1538                 ((struct qib_filedata *)fp->private_data)->rec_cpu_num = -1;
1539         return fp->private_data ? 0 : -ENOMEM;
1540 }
1541
1542 static int find_hca(unsigned int cpu, int *unit)
1543 {
1544         int ret = 0, devmax, npresent, nup, ndev;
1545
1546         *unit = -1;
1547
1548         devmax = qib_count_units(&npresent, &nup);
1549         if (!npresent) {
1550                 ret = -ENXIO;
1551                 goto done;
1552         }
1553         if (!nup) {
1554                 ret = -ENETDOWN;
1555                 goto done;
1556         }
1557         for (ndev = 0; ndev < devmax; ndev++) {
1558                 struct qib_devdata *dd = qib_lookup(ndev);
1559                 if (dd) {
1560                         if (pcibus_to_node(dd->pcidev->bus) < 0) {
1561                                 ret = -EINVAL;
1562                                 goto done;
1563                         }
1564                         if (cpu_to_node(cpu) ==
1565                                 pcibus_to_node(dd->pcidev->bus)) {
1566                                 *unit = ndev;
1567                                 goto done;
1568                         }
1569                 }
1570         }
1571 done:
1572         return ret;
1573 }
1574
1575 static int do_qib_user_sdma_queue_create(struct file *fp)
1576 {
1577         struct qib_filedata *fd = fp->private_data;
1578         struct qib_ctxtdata *rcd = fd->rcd;
1579         struct qib_devdata *dd = rcd->dd;
1580
1581         if (dd->flags & QIB_HAS_SEND_DMA) {
1582
1583                 fd->pq = qib_user_sdma_queue_create(&dd->pcidev->dev,
1584                                                     dd->unit,
1585                                                     rcd->ctxt,
1586                                                     fd->subctxt);
1587                 if (!fd->pq)
1588                         return -ENOMEM;
1589         }
1590
1591         return 0;
1592 }
1593
1594 /*
1595  * Get ctxt early, so can set affinity prior to memory allocation.
1596  */
1597 static int qib_assign_ctxt(struct file *fp, const struct qib_user_info *uinfo)
1598 {
1599         int ret;
1600         int i_minor;
1601         unsigned swmajor, swminor, alg = QIB_PORT_ALG_ACROSS;
1602
1603         /* Check to be sure we haven't already initialized this file */
1604         if (ctxt_fp(fp)) {
1605                 ret = -EINVAL;
1606                 goto done;
1607         }
1608
1609         /* for now, if major version is different, bail */
1610         swmajor = uinfo->spu_userversion >> 16;
1611         if (swmajor != QIB_USER_SWMAJOR) {
1612                 ret = -ENODEV;
1613                 goto done;
1614         }
1615
1616         swminor = uinfo->spu_userversion & 0xffff;
1617
1618         if (swminor >= 11 && uinfo->spu_port_alg < QIB_PORT_ALG_COUNT)
1619                 alg = uinfo->spu_port_alg;
1620
1621         mutex_lock(&qib_mutex);
1622
1623         if (qib_compatible_subctxts(swmajor, swminor) &&
1624             uinfo->spu_subctxt_cnt) {
1625                 ret = find_shared_ctxt(fp, uinfo);
1626                 if (ret > 0) {
1627                         ret = do_qib_user_sdma_queue_create(fp);
1628                         if (!ret)
1629                                 assign_ctxt_affinity(fp, (ctxt_fp(fp))->dd);
1630                         goto done_ok;
1631                 }
1632         }
1633
1634         i_minor = iminor(file_inode(fp)) - QIB_USER_MINOR_BASE;
1635         if (i_minor)
1636                 ret = find_free_ctxt(i_minor - 1, fp, uinfo);
1637         else {
1638                 int unit;
1639                 const unsigned int cpu = cpumask_first(&current->cpus_allowed);
1640                 const unsigned int weight =
1641                         cpumask_weight(&current->cpus_allowed);
1642
1643                 if (weight == 1 && !test_bit(cpu, qib_cpulist))
1644                         if (!find_hca(cpu, &unit) && unit >= 0)
1645                                 if (!find_free_ctxt(unit, fp, uinfo)) {
1646                                         ret = 0;
1647                                         goto done_chk_sdma;
1648                                 }
1649                 ret = get_a_ctxt(fp, uinfo, alg);
1650         }
1651
1652 done_chk_sdma:
1653         if (!ret)
1654                 ret = do_qib_user_sdma_queue_create(fp);
1655 done_ok:
1656         mutex_unlock(&qib_mutex);
1657
1658 done:
1659         return ret;
1660 }
1661
1662
1663 static int qib_do_user_init(struct file *fp,
1664                             const struct qib_user_info *uinfo)
1665 {
1666         int ret;
1667         struct qib_ctxtdata *rcd = ctxt_fp(fp);
1668         struct qib_devdata *dd;
1669         unsigned uctxt;
1670
1671         /* Subctxts don't need to initialize anything since master did it. */
1672         if (subctxt_fp(fp)) {
1673                 ret = wait_event_interruptible(rcd->wait,
1674                         !test_bit(QIB_CTXT_MASTER_UNINIT, &rcd->flag));
1675                 goto bail;
1676         }
1677
1678         dd = rcd->dd;
1679
1680         /* some ctxts may get extra buffers, calculate that here */
1681         uctxt = rcd->ctxt - dd->first_user_ctxt;
1682         if (uctxt < dd->ctxts_extrabuf) {
1683                 rcd->piocnt = dd->pbufsctxt + 1;
1684                 rcd->pio_base = rcd->piocnt * uctxt;
1685         } else {
1686                 rcd->piocnt = dd->pbufsctxt;
1687                 rcd->pio_base = rcd->piocnt * uctxt +
1688                         dd->ctxts_extrabuf;
1689         }
1690
1691         /*
1692          * All user buffers are 2KB buffers.  If we ever support
1693          * giving 4KB buffers to user processes, this will need some
1694          * work.  Can't use piobufbase directly, because it has
1695          * both 2K and 4K buffer base values.  So check and handle.
1696          */
1697         if ((rcd->pio_base + rcd->piocnt) > dd->piobcnt2k) {
1698                 if (rcd->pio_base >= dd->piobcnt2k) {
1699                         qib_dev_err(dd,
1700                                     "%u:ctxt%u: no 2KB buffers available\n",
1701                                     dd->unit, rcd->ctxt);
1702                         ret = -ENOBUFS;
1703                         goto bail;
1704                 }
1705                 rcd->piocnt = dd->piobcnt2k - rcd->pio_base;
1706                 qib_dev_err(dd, "Ctxt%u: would use 4KB bufs, using %u\n",
1707                             rcd->ctxt, rcd->piocnt);
1708         }
1709
1710         rcd->piobufs = dd->pio2k_bufbase + rcd->pio_base * dd->palign;
1711         qib_chg_pioavailkernel(dd, rcd->pio_base, rcd->piocnt,
1712                                TXCHK_CHG_TYPE_USER, rcd);
1713         /*
1714          * try to ensure that processes start up with consistent avail update
1715          * for their own range, at least.   If system very quiet, it might
1716          * have the in-memory copy out of date at startup for this range of
1717          * buffers, when a context gets re-used.  Do after the chg_pioavail
1718          * and before the rest of setup, so it's "almost certain" the dma
1719          * will have occurred (can't 100% guarantee, but should be many
1720          * decimals of 9s, with this ordering), given how much else happens
1721          * after this.
1722          */
1723         dd->f_sendctrl(dd->pport, QIB_SENDCTRL_AVAIL_BLIP);
1724
1725         /*
1726          * Now allocate the rcvhdr Q and eager TIDs; skip the TID
1727          * array for time being.  If rcd->ctxt > chip-supported,
1728          * we need to do extra stuff here to handle by handling overflow
1729          * through ctxt 0, someday
1730          */
1731         ret = qib_create_rcvhdrq(dd, rcd);
1732         if (!ret)
1733                 ret = qib_setup_eagerbufs(rcd);
1734         if (ret)
1735                 goto bail_pio;
1736
1737         rcd->tidcursor = 0; /* start at beginning after open */
1738
1739         /* initialize poll variables... */
1740         rcd->urgent = 0;
1741         rcd->urgent_poll = 0;
1742
1743         /*
1744          * Now enable the ctxt for receive.
1745          * For chips that are set to DMA the tail register to memory
1746          * when they change (and when the update bit transitions from
1747          * 0 to 1.  So for those chips, we turn it off and then back on.
1748          * This will (very briefly) affect any other open ctxts, but the
1749          * duration is very short, and therefore isn't an issue.  We
1750          * explicitly set the in-memory tail copy to 0 beforehand, so we
1751          * don't have to wait to be sure the DMA update has happened
1752          * (chip resets head/tail to 0 on transition to enable).
1753          */
1754         if (rcd->rcvhdrtail_kvaddr)
1755                 qib_clear_rcvhdrtail(rcd);
1756
1757         dd->f_rcvctrl(rcd->ppd, QIB_RCVCTRL_CTXT_ENB | QIB_RCVCTRL_TIDFLOW_ENB,
1758                       rcd->ctxt);
1759
1760         /* Notify any waiting slaves */
1761         if (rcd->subctxt_cnt) {
1762                 clear_bit(QIB_CTXT_MASTER_UNINIT, &rcd->flag);
1763                 wake_up(&rcd->wait);
1764         }
1765         return 0;
1766
1767 bail_pio:
1768         qib_chg_pioavailkernel(dd, rcd->pio_base, rcd->piocnt,
1769                                TXCHK_CHG_TYPE_KERN, rcd);
1770 bail:
1771         return ret;
1772 }
1773
1774 /**
1775  * unlock_exptid - unlock any expected TID entries context still had in use
1776  * @rcd: ctxt
1777  *
1778  * We don't actually update the chip here, because we do a bulk update
1779  * below, using f_clear_tids.
1780  */
1781 static void unlock_expected_tids(struct qib_ctxtdata *rcd)
1782 {
1783         struct qib_devdata *dd = rcd->dd;
1784         int ctxt_tidbase = rcd->ctxt * dd->rcvtidcnt;
1785         int i, cnt = 0, maxtid = ctxt_tidbase + dd->rcvtidcnt;
1786
1787         for (i = ctxt_tidbase; i < maxtid; i++) {
1788                 struct page *p = dd->pageshadow[i];
1789                 dma_addr_t phys;
1790
1791                 if (!p)
1792                         continue;
1793
1794                 phys = dd->physshadow[i];
1795                 dd->physshadow[i] = dd->tidinvalid;
1796                 dd->pageshadow[i] = NULL;
1797                 pci_unmap_page(dd->pcidev, phys, PAGE_SIZE,
1798                                PCI_DMA_FROMDEVICE);
1799                 qib_release_user_pages(&p, 1);
1800                 cnt++;
1801         }
1802 }
1803
1804 static int qib_close(struct inode *in, struct file *fp)
1805 {
1806         int ret = 0;
1807         struct qib_filedata *fd;
1808         struct qib_ctxtdata *rcd;
1809         struct qib_devdata *dd;
1810         unsigned long flags;
1811         unsigned ctxt;
1812         pid_t pid;
1813
1814         mutex_lock(&qib_mutex);
1815
1816         fd = fp->private_data;
1817         fp->private_data = NULL;
1818         rcd = fd->rcd;
1819         if (!rcd) {
1820                 mutex_unlock(&qib_mutex);
1821                 goto bail;
1822         }
1823
1824         dd = rcd->dd;
1825
1826         /* ensure all pio buffer writes in progress are flushed */
1827         qib_flush_wc();
1828
1829         /* drain user sdma queue */
1830         if (fd->pq) {
1831                 qib_user_sdma_queue_drain(rcd->ppd, fd->pq);
1832                 qib_user_sdma_queue_destroy(fd->pq);
1833         }
1834
1835         if (fd->rec_cpu_num != -1)
1836                 __clear_bit(fd->rec_cpu_num, qib_cpulist);
1837
1838         if (--rcd->cnt) {
1839                 /*
1840                  * XXX If the master closes the context before the slave(s),
1841                  * revoke the mmap for the eager receive queue so
1842                  * the slave(s) don't wait for receive data forever.
1843                  */
1844                 rcd->active_slaves &= ~(1 << fd->subctxt);
1845                 rcd->subpid[fd->subctxt] = 0;
1846                 mutex_unlock(&qib_mutex);
1847                 goto bail;
1848         }
1849
1850         /* early; no interrupt users after this */
1851         spin_lock_irqsave(&dd->uctxt_lock, flags);
1852         ctxt = rcd->ctxt;
1853         dd->rcd[ctxt] = NULL;
1854         pid = rcd->pid;
1855         rcd->pid = 0;
1856         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
1857
1858         if (rcd->rcvwait_to || rcd->piowait_to ||
1859             rcd->rcvnowait || rcd->pionowait) {
1860                 rcd->rcvwait_to = 0;
1861                 rcd->piowait_to = 0;
1862                 rcd->rcvnowait = 0;
1863                 rcd->pionowait = 0;
1864         }
1865         if (rcd->flag)
1866                 rcd->flag = 0;
1867
1868         if (dd->kregbase) {
1869                 /* atomically clear receive enable ctxt and intr avail. */
1870                 dd->f_rcvctrl(rcd->ppd, QIB_RCVCTRL_CTXT_DIS |
1871                                   QIB_RCVCTRL_INTRAVAIL_DIS, ctxt);
1872
1873                 /* clean up the pkeys for this ctxt user */
1874                 qib_clean_part_key(rcd, dd);
1875                 qib_disarm_piobufs(dd, rcd->pio_base, rcd->piocnt);
1876                 qib_chg_pioavailkernel(dd, rcd->pio_base,
1877                                        rcd->piocnt, TXCHK_CHG_TYPE_KERN, NULL);
1878
1879                 dd->f_clear_tids(dd, rcd);
1880
1881                 if (dd->pageshadow)
1882                         unlock_expected_tids(rcd);
1883                 qib_stats.sps_ctxts--;
1884                 dd->freectxts++;
1885         }
1886
1887         mutex_unlock(&qib_mutex);
1888         qib_free_ctxtdata(dd, rcd); /* after releasing the mutex */
1889
1890 bail:
1891         kfree(fd);
1892         return ret;
1893 }
1894
1895 static int qib_ctxt_info(struct file *fp, struct qib_ctxt_info __user *uinfo)
1896 {
1897         struct qib_ctxt_info info;
1898         int ret;
1899         size_t sz;
1900         struct qib_ctxtdata *rcd = ctxt_fp(fp);
1901         struct qib_filedata *fd;
1902
1903         fd = fp->private_data;
1904
1905         info.num_active = qib_count_active_units();
1906         info.unit = rcd->dd->unit;
1907         info.port = rcd->ppd->port;
1908         info.ctxt = rcd->ctxt;
1909         info.subctxt =  subctxt_fp(fp);
1910         /* Number of user ctxts available for this device. */
1911         info.num_ctxts = rcd->dd->cfgctxts - rcd->dd->first_user_ctxt;
1912         info.num_subctxts = rcd->subctxt_cnt;
1913         info.rec_cpu = fd->rec_cpu_num;
1914         sz = sizeof(info);
1915
1916         if (copy_to_user(uinfo, &info, sz)) {
1917                 ret = -EFAULT;
1918                 goto bail;
1919         }
1920         ret = 0;
1921
1922 bail:
1923         return ret;
1924 }
1925
1926 static int qib_sdma_get_inflight(struct qib_user_sdma_queue *pq,
1927                                  u32 __user *inflightp)
1928 {
1929         const u32 val = qib_user_sdma_inflight_counter(pq);
1930
1931         if (put_user(val, inflightp))
1932                 return -EFAULT;
1933
1934         return 0;
1935 }
1936
1937 static int qib_sdma_get_complete(struct qib_pportdata *ppd,
1938                                  struct qib_user_sdma_queue *pq,
1939                                  u32 __user *completep)
1940 {
1941         u32 val;
1942         int err;
1943
1944         if (!pq)
1945                 return -EINVAL;
1946
1947         err = qib_user_sdma_make_progress(ppd, pq);
1948         if (err < 0)
1949                 return err;
1950
1951         val = qib_user_sdma_complete_counter(pq);
1952         if (put_user(val, completep))
1953                 return -EFAULT;
1954
1955         return 0;
1956 }
1957
1958 static int disarm_req_delay(struct qib_ctxtdata *rcd)
1959 {
1960         int ret = 0;
1961
1962         if (!usable(rcd->ppd)) {
1963                 int i;
1964                 /*
1965                  * if link is down, or otherwise not usable, delay
1966                  * the caller up to 30 seconds, so we don't thrash
1967                  * in trying to get the chip back to ACTIVE, and
1968                  * set flag so they make the call again.
1969                  */
1970                 if (rcd->user_event_mask) {
1971                         /*
1972                          * subctxt_cnt is 0 if not shared, so do base
1973                          * separately, first, then remaining subctxt, if any
1974                          */
1975                         set_bit(_QIB_EVENT_DISARM_BUFS_BIT,
1976                                 &rcd->user_event_mask[0]);
1977                         for (i = 1; i < rcd->subctxt_cnt; i++)
1978                                 set_bit(_QIB_EVENT_DISARM_BUFS_BIT,
1979                                         &rcd->user_event_mask[i]);
1980                 }
1981                 for (i = 0; !usable(rcd->ppd) && i < 300; i++)
1982                         msleep(100);
1983                 ret = -ENETDOWN;
1984         }
1985         return ret;
1986 }
1987
1988 /*
1989  * Find all user contexts in use, and set the specified bit in their
1990  * event mask.
1991  * See also find_ctxt() for a similar use, that is specific to send buffers.
1992  */
1993 int qib_set_uevent_bits(struct qib_pportdata *ppd, const int evtbit)
1994 {
1995         struct qib_ctxtdata *rcd;
1996         unsigned ctxt;
1997         int ret = 0;
1998         unsigned long flags;
1999
2000         spin_lock_irqsave(&ppd->dd->uctxt_lock, flags);
2001         for (ctxt = ppd->dd->first_user_ctxt; ctxt < ppd->dd->cfgctxts;
2002              ctxt++) {
2003                 rcd = ppd->dd->rcd[ctxt];
2004                 if (!rcd)
2005                         continue;
2006                 if (rcd->user_event_mask) {
2007                         int i;
2008                         /*
2009                          * subctxt_cnt is 0 if not shared, so do base
2010                          * separately, first, then remaining subctxt, if any
2011                          */
2012                         set_bit(evtbit, &rcd->user_event_mask[0]);
2013                         for (i = 1; i < rcd->subctxt_cnt; i++)
2014                                 set_bit(evtbit, &rcd->user_event_mask[i]);
2015                 }
2016                 ret = 1;
2017                 break;
2018         }
2019         spin_unlock_irqrestore(&ppd->dd->uctxt_lock, flags);
2020
2021         return ret;
2022 }
2023
2024 /*
2025  * clear the event notifier events for this context.
2026  * For the DISARM_BUFS case, we also take action (this obsoletes
2027  * the older QIB_CMD_DISARM_BUFS, but we keep it for backwards
2028  * compatibility.
2029  * Other bits don't currently require actions, just atomically clear.
2030  * User process then performs actions appropriate to bit having been
2031  * set, if desired, and checks again in future.
2032  */
2033 static int qib_user_event_ack(struct qib_ctxtdata *rcd, int subctxt,
2034                               unsigned long events)
2035 {
2036         int ret = 0, i;
2037
2038         for (i = 0; i <= _QIB_MAX_EVENT_BIT; i++) {
2039                 if (!test_bit(i, &events))
2040                         continue;
2041                 if (i == _QIB_EVENT_DISARM_BUFS_BIT) {
2042                         (void)qib_disarm_piobufs_ifneeded(rcd);
2043                         ret = disarm_req_delay(rcd);
2044                 } else
2045                         clear_bit(i, &rcd->user_event_mask[subctxt]);
2046         }
2047         return ret;
2048 }
2049
2050 static ssize_t qib_write(struct file *fp, const char __user *data,
2051                          size_t count, loff_t *off)
2052 {
2053         const struct qib_cmd __user *ucmd;
2054         struct qib_ctxtdata *rcd;
2055         const void __user *src;
2056         size_t consumed, copy = 0;
2057         struct qib_cmd cmd;
2058         ssize_t ret = 0;
2059         void *dest;
2060
2061         if (count < sizeof(cmd.type)) {
2062                 ret = -EINVAL;
2063                 goto bail;
2064         }
2065
2066         ucmd = (const struct qib_cmd __user *) data;
2067
2068         if (copy_from_user(&cmd.type, &ucmd->type, sizeof(cmd.type))) {
2069                 ret = -EFAULT;
2070                 goto bail;
2071         }
2072
2073         consumed = sizeof(cmd.type);
2074
2075         switch (cmd.type) {
2076         case QIB_CMD_ASSIGN_CTXT:
2077         case QIB_CMD_USER_INIT:
2078                 copy = sizeof(cmd.cmd.user_info);
2079                 dest = &cmd.cmd.user_info;
2080                 src = &ucmd->cmd.user_info;
2081                 break;
2082
2083         case QIB_CMD_RECV_CTRL:
2084                 copy = sizeof(cmd.cmd.recv_ctrl);
2085                 dest = &cmd.cmd.recv_ctrl;
2086                 src = &ucmd->cmd.recv_ctrl;
2087                 break;
2088
2089         case QIB_CMD_CTXT_INFO:
2090                 copy = sizeof(cmd.cmd.ctxt_info);
2091                 dest = &cmd.cmd.ctxt_info;
2092                 src = &ucmd->cmd.ctxt_info;
2093                 break;
2094
2095         case QIB_CMD_TID_UPDATE:
2096         case QIB_CMD_TID_FREE:
2097                 copy = sizeof(cmd.cmd.tid_info);
2098                 dest = &cmd.cmd.tid_info;
2099                 src = &ucmd->cmd.tid_info;
2100                 break;
2101
2102         case QIB_CMD_SET_PART_KEY:
2103                 copy = sizeof(cmd.cmd.part_key);
2104                 dest = &cmd.cmd.part_key;
2105                 src = &ucmd->cmd.part_key;
2106                 break;
2107
2108         case QIB_CMD_DISARM_BUFS:
2109         case QIB_CMD_PIOAVAILUPD: /* force an update of PIOAvail reg */
2110                 copy = 0;
2111                 src = NULL;
2112                 dest = NULL;
2113                 break;
2114
2115         case QIB_CMD_POLL_TYPE:
2116                 copy = sizeof(cmd.cmd.poll_type);
2117                 dest = &cmd.cmd.poll_type;
2118                 src = &ucmd->cmd.poll_type;
2119                 break;
2120
2121         case QIB_CMD_ARMLAUNCH_CTRL:
2122                 copy = sizeof(cmd.cmd.armlaunch_ctrl);
2123                 dest = &cmd.cmd.armlaunch_ctrl;
2124                 src = &ucmd->cmd.armlaunch_ctrl;
2125                 break;
2126
2127         case QIB_CMD_SDMA_INFLIGHT:
2128                 copy = sizeof(cmd.cmd.sdma_inflight);
2129                 dest = &cmd.cmd.sdma_inflight;
2130                 src = &ucmd->cmd.sdma_inflight;
2131                 break;
2132
2133         case QIB_CMD_SDMA_COMPLETE:
2134                 copy = sizeof(cmd.cmd.sdma_complete);
2135                 dest = &cmd.cmd.sdma_complete;
2136                 src = &ucmd->cmd.sdma_complete;
2137                 break;
2138
2139         case QIB_CMD_ACK_EVENT:
2140                 copy = sizeof(cmd.cmd.event_mask);
2141                 dest = &cmd.cmd.event_mask;
2142                 src = &ucmd->cmd.event_mask;
2143                 break;
2144
2145         default:
2146                 ret = -EINVAL;
2147                 goto bail;
2148         }
2149
2150         if (copy) {
2151                 if ((count - consumed) < copy) {
2152                         ret = -EINVAL;
2153                         goto bail;
2154                 }
2155                 if (copy_from_user(dest, src, copy)) {
2156                         ret = -EFAULT;
2157                         goto bail;
2158                 }
2159                 consumed += copy;
2160         }
2161
2162         rcd = ctxt_fp(fp);
2163         if (!rcd && cmd.type != QIB_CMD_ASSIGN_CTXT) {
2164                 ret = -EINVAL;
2165                 goto bail;
2166         }
2167
2168         switch (cmd.type) {
2169         case QIB_CMD_ASSIGN_CTXT:
2170                 ret = qib_assign_ctxt(fp, &cmd.cmd.user_info);
2171                 if (ret)
2172                         goto bail;
2173                 break;
2174
2175         case QIB_CMD_USER_INIT:
2176                 ret = qib_do_user_init(fp, &cmd.cmd.user_info);
2177                 if (ret)
2178                         goto bail;
2179                 ret = qib_get_base_info(fp, (void __user *) (unsigned long)
2180                                         cmd.cmd.user_info.spu_base_info,
2181                                         cmd.cmd.user_info.spu_base_info_size);
2182                 break;
2183
2184         case QIB_CMD_RECV_CTRL:
2185                 ret = qib_manage_rcvq(rcd, subctxt_fp(fp), cmd.cmd.recv_ctrl);
2186                 break;
2187
2188         case QIB_CMD_CTXT_INFO:
2189                 ret = qib_ctxt_info(fp, (struct qib_ctxt_info __user *)
2190                                     (unsigned long) cmd.cmd.ctxt_info);
2191                 break;
2192
2193         case QIB_CMD_TID_UPDATE:
2194                 ret = qib_tid_update(rcd, fp, &cmd.cmd.tid_info);
2195                 break;
2196
2197         case QIB_CMD_TID_FREE:
2198                 ret = qib_tid_free(rcd, subctxt_fp(fp), &cmd.cmd.tid_info);
2199                 break;
2200
2201         case QIB_CMD_SET_PART_KEY:
2202                 ret = qib_set_part_key(rcd, cmd.cmd.part_key);
2203                 break;
2204
2205         case QIB_CMD_DISARM_BUFS:
2206                 (void)qib_disarm_piobufs_ifneeded(rcd);
2207                 ret = disarm_req_delay(rcd);
2208                 break;
2209
2210         case QIB_CMD_PIOAVAILUPD:
2211                 qib_force_pio_avail_update(rcd->dd);
2212                 break;
2213
2214         case QIB_CMD_POLL_TYPE:
2215                 rcd->poll_type = cmd.cmd.poll_type;
2216                 break;
2217
2218         case QIB_CMD_ARMLAUNCH_CTRL:
2219                 rcd->dd->f_set_armlaunch(rcd->dd, cmd.cmd.armlaunch_ctrl);
2220                 break;
2221
2222         case QIB_CMD_SDMA_INFLIGHT:
2223                 ret = qib_sdma_get_inflight(user_sdma_queue_fp(fp),
2224                                             (u32 __user *) (unsigned long)
2225                                             cmd.cmd.sdma_inflight);
2226                 break;
2227
2228         case QIB_CMD_SDMA_COMPLETE:
2229                 ret = qib_sdma_get_complete(rcd->ppd,
2230                                             user_sdma_queue_fp(fp),
2231                                             (u32 __user *) (unsigned long)
2232                                             cmd.cmd.sdma_complete);
2233                 break;
2234
2235         case QIB_CMD_ACK_EVENT:
2236                 ret = qib_user_event_ack(rcd, subctxt_fp(fp),
2237                                          cmd.cmd.event_mask);
2238                 break;
2239         }
2240
2241         if (ret >= 0)
2242                 ret = consumed;
2243
2244 bail:
2245         return ret;
2246 }
2247
2248 static ssize_t qib_aio_write(struct kiocb *iocb, const struct iovec *iov,
2249                              unsigned long dim, loff_t off)
2250 {
2251         struct qib_filedata *fp = iocb->ki_filp->private_data;
2252         struct qib_ctxtdata *rcd = ctxt_fp(iocb->ki_filp);
2253         struct qib_user_sdma_queue *pq = fp->pq;
2254
2255         if (!dim || !pq)
2256                 return -EINVAL;
2257
2258         return qib_user_sdma_writev(rcd, pq, iov, dim);
2259 }
2260
2261 static struct class *qib_class;
2262 static dev_t qib_dev;
2263
2264 int qib_cdev_init(int minor, const char *name,
2265                   const struct file_operations *fops,
2266                   struct cdev **cdevp, struct device **devp)
2267 {
2268         const dev_t dev = MKDEV(MAJOR(qib_dev), minor);
2269         struct cdev *cdev;
2270         struct device *device = NULL;
2271         int ret;
2272
2273         cdev = cdev_alloc();
2274         if (!cdev) {
2275                 pr_err("Could not allocate cdev for minor %d, %s\n",
2276                        minor, name);
2277                 ret = -ENOMEM;
2278                 goto done;
2279         }
2280
2281         cdev->owner = THIS_MODULE;
2282         cdev->ops = fops;
2283         kobject_set_name(&cdev->kobj, name);
2284
2285         ret = cdev_add(cdev, dev, 1);
2286         if (ret < 0) {
2287                 pr_err("Could not add cdev for minor %d, %s (err %d)\n",
2288                        minor, name, -ret);
2289                 goto err_cdev;
2290         }
2291
2292         device = device_create(qib_class, NULL, dev, NULL, "%s", name);
2293         if (!IS_ERR(device))
2294                 goto done;
2295         ret = PTR_ERR(device);
2296         device = NULL;
2297         pr_err("Could not create device for minor %d, %s (err %d)\n",
2298                minor, name, -ret);
2299 err_cdev:
2300         cdev_del(cdev);
2301         cdev = NULL;
2302 done:
2303         *cdevp = cdev;
2304         *devp = device;
2305         return ret;
2306 }
2307
2308 void qib_cdev_cleanup(struct cdev **cdevp, struct device **devp)
2309 {
2310         struct device *device = *devp;
2311
2312         if (device) {
2313                 device_unregister(device);
2314                 *devp = NULL;
2315         }
2316
2317         if (*cdevp) {
2318                 cdev_del(*cdevp);
2319                 *cdevp = NULL;
2320         }
2321 }
2322
2323 static struct cdev *wildcard_cdev;
2324 static struct device *wildcard_device;
2325
2326 int __init qib_dev_init(void)
2327 {
2328         int ret;
2329
2330         ret = alloc_chrdev_region(&qib_dev, 0, QIB_NMINORS, QIB_DRV_NAME);
2331         if (ret < 0) {
2332                 pr_err("Could not allocate chrdev region (err %d)\n", -ret);
2333                 goto done;
2334         }
2335
2336         qib_class = class_create(THIS_MODULE, "ipath");
2337         if (IS_ERR(qib_class)) {
2338                 ret = PTR_ERR(qib_class);
2339                 pr_err("Could not create device class (err %d)\n", -ret);
2340                 unregister_chrdev_region(qib_dev, QIB_NMINORS);
2341         }
2342
2343 done:
2344         return ret;
2345 }
2346
2347 void qib_dev_cleanup(void)
2348 {
2349         if (qib_class) {
2350                 class_destroy(qib_class);
2351                 qib_class = NULL;
2352         }
2353
2354         unregister_chrdev_region(qib_dev, QIB_NMINORS);
2355 }
2356
2357 static atomic_t user_count = ATOMIC_INIT(0);
2358
2359 static void qib_user_remove(struct qib_devdata *dd)
2360 {
2361         if (atomic_dec_return(&user_count) == 0)
2362                 qib_cdev_cleanup(&wildcard_cdev, &wildcard_device);
2363
2364         qib_cdev_cleanup(&dd->user_cdev, &dd->user_device);
2365 }
2366
2367 static int qib_user_add(struct qib_devdata *dd)
2368 {
2369         char name[10];
2370         int ret;
2371
2372         if (atomic_inc_return(&user_count) == 1) {
2373                 ret = qib_cdev_init(0, "ipath", &qib_file_ops,
2374                                     &wildcard_cdev, &wildcard_device);
2375                 if (ret)
2376                         goto done;
2377         }
2378
2379         snprintf(name, sizeof(name), "ipath%d", dd->unit);
2380         ret = qib_cdev_init(dd->unit + 1, name, &qib_file_ops,
2381                             &dd->user_cdev, &dd->user_device);
2382         if (ret)
2383                 qib_user_remove(dd);
2384 done:
2385         return ret;
2386 }
2387
2388 /*
2389  * Create per-unit files in /dev
2390  */
2391 int qib_device_create(struct qib_devdata *dd)
2392 {
2393         int r, ret;
2394
2395         r = qib_user_add(dd);
2396         ret = qib_diag_add(dd);
2397         if (r && !ret)
2398                 ret = r;
2399         return ret;
2400 }
2401
2402 /*
2403  * Remove per-unit files in /dev
2404  * void, core kernel returns no errors for this stuff
2405  */
2406 void qib_device_remove(struct qib_devdata *dd)
2407 {
2408         qib_user_remove(dd);
2409         qib_diag_remove(dd);
2410 }