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(
 355                         dd->pcidev,
 356                         "Failed to lock addr %p, %u pages: errno %d\n",
 357                         (void *) vaddr, cnt, -ret);
 358                 goto done;
 359         }
 360         for (i = 0; i < cnt; i++, vaddr += PAGE_SIZE) {
 361                 for (; ntids--; tid++) {
 362                         if (tid == tidcnt)
 363                                 tid = 0;
 364                         if (!dd->pageshadow[ctxttid + tid])
 365                                 break;
 366                 }
 367                 if (ntids < 0) {
 368                         /*
 369                          * Oops, wrapped all the way through their TIDs,
 370                          * and didn't have enough free; see comments at
 371                          * start of routine
 372                          */
 373                         i--;    /* last tidlist[i] not filled in */
 374                         ret = -ENOMEM;
 375                         break;
 376                 }
 377                 tidlist[i] = tid + tidoff;
 378                 /* we "know" system pages and TID pages are same size */
 379                 dd->pageshadow[ctxttid + tid] = pagep[i];
 380                 dd->physshadow[ctxttid + tid] =
 381                         qib_map_page(dd->pcidev, pagep[i], 0, PAGE_SIZE,
 382                                      PCI_DMA_FROMDEVICE);
 383                 /*
 384                  * don't need atomic or it's overhead
 385                  */
 386                 __set_bit(tid, tidmap);
 387                 physaddr = dd->physshadow[ctxttid + tid];
 388                 /* PERFORMANCE: below should almost certainly be cached */
 389                 dd->f_put_tid(dd, &tidbase[tid],
 390                                   RCVHQ_RCV_TYPE_EXPECTED, physaddr);
 391                 /*
 392                  * don't check this tid in qib_ctxtshadow, since we
 393                  * just filled it in; start with the next one.
 394                  */
 395                 tid++;
 396         }
 397
 398         if (ret) {
 399                 u32 limit;
 400 cleanup:
 401                 /* jump here if copy out of updated info failed... */
 402                 /* same code that's in qib_free_tid() */
 403                 limit = sizeof(tidmap) * BITS_PER_BYTE;
 404                 if (limit > tidcnt)
 405                         /* just in case size changes in future */
 406                         limit = tidcnt;
 407                 tid = find_first_bit((const unsigned long *)tidmap, limit);
 408                 for (; tid < limit; tid++) {
 409                         if (!test_bit(tid, tidmap))
 410                                 continue;
 411                         if (dd->pageshadow[ctxttid + tid]) {
 412                                 dma_addr_t phys;
 413
 414                                 phys = dd->physshadow[ctxttid + tid];
 415                                 dd->physshadow[ctxttid + tid] = dd->tidinvalid;
 416                                 /* PERFORMANCE: below should almost certainly
 417                                  * be cached
 418                                  */
 419                                 dd->f_put_tid(dd, &tidbase[tid],
 420                                               RCVHQ_RCV_TYPE_EXPECTED,
 421                                               dd->tidinvalid);
 422                                 pci_unmap_page(dd->pcidev, phys, PAGE_SIZE,
 423                                                PCI_DMA_FROMDEVICE);
 424                                 dd->pageshadow[ctxttid + tid] = NULL;
 425                         }
 426                 }
 427                 qib_release_user_pages(pagep, cnt);
 428         } else {
 429                 /*
 430                  * Copy the updated array, with qib_tid's filled in, back
 431                  * to user.  Since we did the copy in already, this "should
 432                  * never fail" If it does, we have to clean up...
 433                  */
 434                 if (copy_to_user((void __user *)
 435                                  (unsigned long) ti->tidlist,
 436                                  tidlist, cnt * sizeof(*tidlist))) {
 437                         ret = -EFAULT;
 438                         goto cleanup;
 439                 }
 440                 if (copy_to_user((void __user *) (unsigned long) ti->tidmap,
 441                                  tidmap, sizeof(tidmap))) {
 442                         ret = -EFAULT;
 443                         goto cleanup;
 444                 }
 445                 if (tid == tidcnt)
 446                         tid = 0;
 447                 if (!rcd->subctxt_cnt)
 448                         rcd->tidcursor = tid;
 449                 else
 450                         tidcursor_fp(fp) = tid;
 451         }
 452
 453 done:
 454         return ret;
 455 }
 456
 457 /**
 458  * qib_tid_free - free a context TID
 459  * @rcd: the context
 460  * @subctxt: the subcontext
 461  * @ti: the TID info
 462  *
 463  * right now we are unlocking one page at a time, but since
 464  * the intended use of this routine is for a single group of
 465  * virtually contiguous pages, that should change to improve
 466  * performance.  We check that the TID is in range for this context
 467  * but otherwise don't check validity; if user has an error and
 468  * frees the wrong tid, it's only their own data that can thereby
 469  * be corrupted.  We do check that the TID was in use, for sanity
 470  * We always use our idea of the saved address, not the address that
 471  * they pass in to us.
 472  */
 473 static int qib_tid_free(struct qib_ctxtdata *rcd, unsigned subctxt,
 474                         const struct qib_tid_info *ti)
 475 {
 476         int ret = 0;
 477         u32 tid, ctxttid, cnt, limit, tidcnt;
 478         struct qib_devdata *dd = rcd->dd;
 479         u64 __iomem *tidbase;
 480         unsigned long tidmap[8];
 481
 482         if (!dd->pageshadow) {
 483                 ret = -ENOMEM;
 484                 goto done;
 485         }
 486
 487         if (copy_from_user(tidmap, (void __user *)(unsigned long)ti->tidmap,
 488                            sizeof(tidmap))) {
 489                 ret = -EFAULT;
 490                 goto done;
 491         }
 492
 493         ctxttid = rcd->ctxt * dd->rcvtidcnt;
 494         if (!rcd->subctxt_cnt)
 495                 tidcnt = dd->rcvtidcnt;
 496         else if (!subctxt) {
 497                 tidcnt = (dd->rcvtidcnt / rcd->subctxt_cnt) +
 498                          (dd->rcvtidcnt % rcd->subctxt_cnt);
 499                 ctxttid += dd->rcvtidcnt - tidcnt;
 500         } else {
 501                 tidcnt = dd->rcvtidcnt / rcd->subctxt_cnt;
 502                 ctxttid += tidcnt * (subctxt - 1);
 503         }
 504         tidbase = (u64 __iomem *) ((char __iomem *)(dd->kregbase) +
 505                                    dd->rcvtidbase +
 506                                    ctxttid * sizeof(*tidbase));
 507
 508         limit = sizeof(tidmap) * BITS_PER_BYTE;
 509         if (limit > tidcnt)
 510                 /* just in case size changes in future */
 511                 limit = tidcnt;
 512         tid = find_first_bit(tidmap, limit);
 513         for (cnt = 0; tid < limit; tid++) {
 514                 /*
 515                  * small optimization; if we detect a run of 3 or so without
 516                  * any set, use find_first_bit again.  That's mainly to
 517                  * accelerate the case where we wrapped, so we have some at
 518                  * the beginning, and some at the end, and a big gap
 519                  * in the middle.
 520                  */
 521                 if (!test_bit(tid, tidmap))
 522                         continue;
 523                 cnt++;
 524                 if (dd->pageshadow[ctxttid + tid]) {
 525                         struct page *p;
 526                         dma_addr_t phys;
 527
 528                         p = dd->pageshadow[ctxttid + tid];
 529                         dd->pageshadow[ctxttid + tid] = NULL;
 530                         phys = dd->physshadow[ctxttid + tid];
 531                         dd->physshadow[ctxttid + tid] = dd->tidinvalid;
 532                         /* PERFORMANCE: below should almost certainly be
 533                          * cached
 534                          */
 535                         dd->f_put_tid(dd, &tidbase[tid],
 536                                       RCVHQ_RCV_TYPE_EXPECTED, dd->tidinvalid);
 537                         pci_unmap_page(dd->pcidev, phys, PAGE_SIZE,
 538                                        PCI_DMA_FROMDEVICE);
 539                         qib_release_user_pages(&p, 1);
 540                 }
 541         }
 542 done:
 543         return ret;
 544 }
 545
 546 /**
 547  * qib_set_part_key - set a partition key
 548  * @rcd: the context
 549  * @key: the key
 550  *
 551  * We can have up to 4 active at a time (other than the default, which is
 552  * always allowed).  This is somewhat tricky, since multiple contexts may set
 553  * the same key, so we reference count them, and clean up at exit.  All 4
 554  * partition keys are packed into a single qlogic_ib register.  It's an
 555  * error for a process to set the same pkey multiple times.  We provide no
 556  * mechanism to de-allocate a pkey at this time, we may eventually need to
 557  * do that.  I've used the atomic operations, and no locking, and only make
 558  * a single pass through what's available.  This should be more than
 559  * adequate for some time. I'll think about spinlocks or the like if and as
 560  * it's necessary.
 561  */
 562 static int qib_set_part_key(struct qib_ctxtdata *rcd, u16 key)
 563 {
 564         struct qib_pportdata *ppd = rcd->ppd;
 565         int i, any = 0, pidx = -1;
 566         u16 lkey = key & 0x7FFF;
 567         int ret;
 568
 569         if (lkey == (QIB_DEFAULT_P_KEY & 0x7FFF)) {
 570                 /* nothing to do; this key always valid */
 571                 ret = 0;
 572                 goto bail;
 573         }
 574
 575         if (!lkey) {
 576                 ret = -EINVAL;
 577                 goto bail;
 578         }
 579
 580         /*
 581          * Set the full membership bit, because it has to be
 582          * set in the register or the packet, and it seems
 583          * cleaner to set in the register than to force all
 584          * callers to set it.
 585          */
 586         key |= 0x8000;
 587
 588         for (i = 0; i < ARRAY_SIZE(rcd->pkeys); i++) {
 589                 if (!rcd->pkeys[i] && pidx == -1)
 590                         pidx = i;
 591                 if (rcd->pkeys[i] == key) {
 592                         ret = -EEXIST;
 593                         goto bail;
 594                 }
 595         }
 596         if (pidx == -1) {
 597                 ret = -EBUSY;
 598                 goto bail;
 599         }
 600         for (any = i = 0; i < ARRAY_SIZE(ppd->pkeys); i++) {
 601                 if (!ppd->pkeys[i]) {
 602                         any++;
 603                         continue;
 604                 }
 605                 if (ppd->pkeys[i] == key) {
 606                         atomic_t *pkrefs = &ppd->pkeyrefs[i];
 607
 608                         if (atomic_inc_return(pkrefs) > 1) {
 609                                 rcd->pkeys[pidx] = key;
 610                                 ret = 0;
 611                                 goto bail;
 612                         } else {
 613                                 /*
 614                                  * lost race, decrement count, catch below
 615                                  */
 616                                 atomic_dec(pkrefs);
 617                                 any++;
 618                         }
 619                 }
 620                 if ((ppd->pkeys[i] & 0x7FFF) == lkey) {
 621                         /*
 622                          * It makes no sense to have both the limited and
 623                          * full membership PKEY set at the same time since
 624                          * the unlimited one will disable the limited one.
 625                          */
 626                         ret = -EEXIST;
 627                         goto bail;
 628                 }
 629         }
 630         if (!any) {
 631                 ret = -EBUSY;
 632                 goto bail;
 633         }
 634         for (any = i = 0; i < ARRAY_SIZE(ppd->pkeys); i++) {
 635                 if (!ppd->pkeys[i] &&
 636                     atomic_inc_return(&ppd->pkeyrefs[i]) == 1) {
 637                         rcd->pkeys[pidx] = key;
 638                         ppd->pkeys[i] = key;
 639                         (void) ppd->dd->f_set_ib_cfg(ppd, QIB_IB_CFG_PKEYS, 0);
 640                         ret = 0;
 641                         goto bail;
 642                 }
 643         }
 644         ret = -EBUSY;
 645
 646 bail:
 647         return ret;
 648 }
 649
 650 /**
 651  * qib_manage_rcvq - manage a context's receive queue
 652  * @rcd: the context
 653  * @subctxt: the subcontext
 654  * @start_stop: action to carry out
 655  *
 656  * start_stop == 0 disables receive on the context, for use in queue
 657  * overflow conditions.  start_stop==1 re-enables, to be used to
 658  * re-init the software copy of the head register
 659  */
 660 static int qib_manage_rcvq(struct qib_ctxtdata *rcd, unsigned subctxt,
 661                            int start_stop)
 662 {
 663         struct qib_devdata *dd = rcd->dd;
 664         unsigned int rcvctrl_op;
 665
 666         if (subctxt)
 667                 goto bail;
 668         /* atomically clear receive enable ctxt. */
 669         if (start_stop) {
 670                 /*
 671                  * On enable, force in-memory copy of the tail register to
 672                  * 0, so that protocol code doesn't have to worry about
 673                  * whether or not the chip has yet updated the in-memory
 674                  * copy or not on return from the system call. The chip
 675                  * always resets it's tail register back to 0 on a
 676                  * transition from disabled to enabled.
 677                  */
 678                 if (rcd->rcvhdrtail_kvaddr)
 679                         qib_clear_rcvhdrtail(rcd);
 680                 rcvctrl_op = QIB_RCVCTRL_CTXT_ENB;
 681         } else
 682                 rcvctrl_op = QIB_RCVCTRL_CTXT_DIS;
 683         dd->f_rcvctrl(rcd->ppd, rcvctrl_op, rcd->ctxt);
 684         /* always; new head should be equal to new tail; see above */
 685 bail:
 686         return 0;
 687 }
 688
 689 static void qib_clean_part_key(struct qib_ctxtdata *rcd,
 690                                struct qib_devdata *dd)
 691 {
 692         int i, j, pchanged = 0;
 693         u64 oldpkey;
 694         struct qib_pportdata *ppd = rcd->ppd;
 695
 696         /* for debugging only */
 697         oldpkey = (u64) ppd->pkeys[0] |
 698                 ((u64) ppd->pkeys[1] << 16) |
 699                 ((u64) ppd->pkeys[2] << 32) |
 700                 ((u64) ppd->pkeys[3] << 48);
 701
 702         for (i = 0; i < ARRAY_SIZE(rcd->pkeys); i++) {
 703                 if (!rcd->pkeys[i])
 704                         continue;
 705                 for (j = 0; j < ARRAY_SIZE(ppd->pkeys); j++) {
 706                         /* check for match independent of the global bit */
 707                         if ((ppd->pkeys[j] & 0x7fff) !=
 708                             (rcd->pkeys[i] & 0x7fff))
 709                                 continue;
 710                         if (atomic_dec_and_test(&ppd->pkeyrefs[j])) {
 711                                 ppd->pkeys[j] = 0;
 712                                 pchanged++;
 713                         }
 714                         break;
 715                 }
 716                 rcd->pkeys[i] = 0;
 717         }
 718         if (pchanged)
 719                 (void) ppd->dd->f_set_ib_cfg(ppd, QIB_IB_CFG_PKEYS, 0);
 720 }
 721
 722 /* common code for the mappings on dma_alloc_coherent mem */
 723 static int qib_mmap_mem(struct vm_area_struct *vma, struct qib_ctxtdata *rcd,
 724                         unsigned len, void *kvaddr, u32 write_ok, char *what)
 725 {
 726         struct qib_devdata *dd = rcd->dd;
 727         unsigned long pfn;
 728         int ret;
 729
 730         if ((vma->vm_end - vma->vm_start) > len) {
 731                 qib_devinfo(dd->pcidev,
 732                          "FAIL on %s: len %lx > %x\n", what,
 733                          vma->vm_end - vma->vm_start, len);
 734                 ret = -EFAULT;
 735                 goto bail;
 736         }
 737
 738         /*
 739          * shared context user code requires rcvhdrq mapped r/w, others
 740          * only allowed readonly mapping.
 741          */
 742         if (!write_ok) {
 743                 if (vma->vm_flags & VM_WRITE) {
 744                         qib_devinfo(dd->pcidev,
 745                                  "%s must be mapped readonly\n", what);
 746                         ret = -EPERM;
 747                         goto bail;
 748                 }
 749
 750                 /* don't allow them to later change with mprotect */
 751                 vma->vm_flags &= ~VM_MAYWRITE;
 752         }
 753
 754         pfn = virt_to_phys(kvaddr) >> PAGE_SHIFT;
 755         ret = remap_pfn_range(vma, vma->vm_start, pfn,
 756                               len, vma->vm_page_prot);
 757         if (ret)
 758                 qib_devinfo(dd->pcidev,
 759                         "%s ctxt%u mmap of %lx, %x bytes failed: %d\n",
 760                         what, rcd->ctxt, pfn, len, ret);
 761 bail:
 762         return ret;
 763 }
 764
 765 static int mmap_ureg(struct vm_area_struct *vma, struct qib_devdata *dd,
 766                      u64 ureg)
 767 {
 768         unsigned long phys;
 769         unsigned long sz;
 770         int ret;
 771
 772         /*
 773          * This is real hardware, so use io_remap.  This is the mechanism
 774          * for the user process to update the head registers for their ctxt
 775          * in the chip.
 776          */
 777         sz = dd->flags & QIB_HAS_HDRSUPP ? 2 * PAGE_SIZE : PAGE_SIZE;
 778         if ((vma->vm_end - vma->vm_start) > sz) {
 779                 qib_devinfo(dd->pcidev,
 780                         "FAIL mmap userreg: reqlen %lx > PAGE\n",
 781                         vma->vm_end - vma->vm_start);
 782                 ret = -EFAULT;
 783         } else {
 784                 phys = dd->physaddr + ureg;
 785                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 786
 787                 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
 788                 ret = io_remap_pfn_range(vma, vma->vm_start,
 789                                          phys >> PAGE_SHIFT,
 790                                          vma->vm_end - vma->vm_start,
 791                                          vma->vm_page_prot);
 792         }
 793         return ret;
 794 }
 795
 796 static int mmap_piobufs(struct vm_area_struct *vma,
 797                         struct qib_devdata *dd,
 798                         struct qib_ctxtdata *rcd,
 799                         unsigned piobufs, unsigned piocnt)
 800 {
 801         unsigned long phys;
 802         int ret;
 803
 804         /*
 805          * When we map the PIO buffers in the chip, we want to map them as
 806          * writeonly, no read possible; unfortunately, x86 doesn't allow
 807          * for this in hardware, but we still prevent users from asking
 808          * for it.
 809          */
 810         if ((vma->vm_end - vma->vm_start) > (piocnt * dd->palign)) {
 811                 qib_devinfo(dd->pcidev,
 812                         "FAIL mmap piobufs: reqlen %lx > PAGE\n",
 813                          vma->vm_end - vma->vm_start);
 814                 ret = -EINVAL;
 815                 goto bail;
 816         }
 817
 818         phys = dd->physaddr + piobufs;
 819
 820 #if defined(__powerpc__)
 821         /* There isn't a generic way to specify writethrough mappings */
 822         pgprot_val(vma->vm_page_prot) |= _PAGE_NO_CACHE;
 823         pgprot_val(vma->vm_page_prot) |= _PAGE_WRITETHRU;
 824         pgprot_val(vma->vm_page_prot) &= ~_PAGE_GUARDED;
 825 #endif
 826
 827         /*
 828          * don't allow them to later change to readable with mprotect (for when
 829          * not initially mapped readable, as is normally the case)
 830          */
 831         vma->vm_flags &= ~VM_MAYREAD;
 832         vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
 833
 834         if (qib_wc_pat)
 835                 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
 836
 837         ret = io_remap_pfn_range(vma, vma->vm_start, phys >> PAGE_SHIFT,
 838                                  vma->vm_end - vma->vm_start,
 839                                  vma->vm_page_prot);
 840 bail:
 841         return ret;
 842 }
 843
 844 static int mmap_rcvegrbufs(struct vm_area_struct *vma,
 845                            struct qib_ctxtdata *rcd)
 846 {
 847         struct qib_devdata *dd = rcd->dd;
 848         unsigned long start, size;
 849         size_t total_size, i;
 850         unsigned long pfn;
 851         int ret;
 852
 853         size = rcd->rcvegrbuf_size;
 854         total_size = rcd->rcvegrbuf_chunks * size;
 855         if ((vma->vm_end - vma->vm_start) > total_size) {
 856                 qib_devinfo(dd->pcidev,
 857                         "FAIL on egr bufs: reqlen %lx > actual %lx\n",
 858                          vma->vm_end - vma->vm_start,
 859                          (unsigned long) total_size);
 860                 ret = -EINVAL;
 861                 goto bail;
 862         }
 863
 864         if (vma->vm_flags & VM_WRITE) {
 865                 qib_devinfo(dd->pcidev,
 866                         "Can't map eager buffers as writable (flags=%lx)\n",
 867                         vma->vm_flags);
 868                 ret = -EPERM;
 869                 goto bail;
 870         }
 871         /* don't allow them to later change to writeable with mprotect */
 872         vma->vm_flags &= ~VM_MAYWRITE;
 873
 874         start = vma->vm_start;
 875
 876         for (i = 0; i < rcd->rcvegrbuf_chunks; i++, start += size) {
 877                 pfn = virt_to_phys(rcd->rcvegrbuf[i]) >> PAGE_SHIFT;
 878                 ret = remap_pfn_range(vma, start, pfn, size,
 879                                       vma->vm_page_prot);
 880                 if (ret < 0)
 881                         goto bail;
 882         }
 883         ret = 0;
 884
 885 bail:
 886         return ret;
 887 }
 888
 889 /*
 890  * qib_file_vma_fault - handle a VMA page fault.
 891  */
 892 static int qib_file_vma_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 893 {
 894         struct page *page;
 895
 896         page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
 897         if (!page)
 898                 return VM_FAULT_SIGBUS;
 899
 900         get_page(page);
 901         vmf->page = page;
 902
 903         return 0;
 904 }
 905
 906 static struct vm_operations_struct qib_file_vm_ops = {
 907         .fault = qib_file_vma_fault,
 908 };
 909
 910 static int mmap_kvaddr(struct vm_area_struct *vma, u64 pgaddr,
 911                        struct qib_ctxtdata *rcd, unsigned subctxt)
 912 {
 913         struct qib_devdata *dd = rcd->dd;
 914         unsigned subctxt_cnt;
 915         unsigned long len;
 916         void *addr;
 917         size_t size;
 918         int ret = 0;
 919
 920         subctxt_cnt = rcd->subctxt_cnt;
 921         size = rcd->rcvegrbuf_chunks * rcd->rcvegrbuf_size;
 922
 923         /*
 924          * Each process has all the subctxt uregbase, rcvhdrq, and
 925          * rcvegrbufs mmapped - as an array for all the processes,
 926          * and also separately for this process.
 927          */
 928         if (pgaddr == cvt_kvaddr(rcd->subctxt_uregbase)) {
 929                 addr = rcd->subctxt_uregbase;
 930                 size = PAGE_SIZE * subctxt_cnt;
 931         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_rcvhdr_base)) {
 932                 addr = rcd->subctxt_rcvhdr_base;
 933                 size = rcd->rcvhdrq_size * subctxt_cnt;
 934         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_rcvegrbuf)) {
 935                 addr = rcd->subctxt_rcvegrbuf;
 936                 size *= subctxt_cnt;
 937         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_uregbase +
 938                                         PAGE_SIZE * subctxt)) {
 939                 addr = rcd->subctxt_uregbase + PAGE_SIZE * subctxt;
 940                 size = PAGE_SIZE;
 941         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_rcvhdr_base +
 942                                         rcd->rcvhdrq_size * subctxt)) {
 943                 addr = rcd->subctxt_rcvhdr_base +
 944                         rcd->rcvhdrq_size * subctxt;
 945                 size = rcd->rcvhdrq_size;
 946         } else if (pgaddr == cvt_kvaddr(&rcd->user_event_mask[subctxt])) {
 947                 addr = rcd->user_event_mask;
 948                 size = PAGE_SIZE;
 949         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_rcvegrbuf +
 950                                         size * subctxt)) {
 951                 addr = rcd->subctxt_rcvegrbuf + size * subctxt;
 952                 /* rcvegrbufs are read-only on the slave */
 953                 if (vma->vm_flags & VM_WRITE) {
 954                         qib_devinfo(dd->pcidev,
 955                                  "Can't map eager buffers as writable (flags=%lx)\n",
 956                                  vma->vm_flags);
 957                         ret = -EPERM;
 958                         goto bail;
 959                 }
 960                 /*
 961                  * Don't allow permission to later change to writeable
 962                  * with mprotect.
 963                  */
 964                 vma->vm_flags &= ~VM_MAYWRITE;
 965         } else
 966                 goto bail;
 967         len = vma->vm_end - vma->vm_start;
 968         if (len > size) {
 969                 ret = -EINVAL;
 970                 goto bail;
 971         }
 972
 973         vma->vm_pgoff = (unsigned long) addr >> PAGE_SHIFT;
 974         vma->vm_ops = &qib_file_vm_ops;
 975         vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
 976         ret = 1;
 977
 978 bail:
 979         return ret;
 980 }
 981
 982 /**
 983  * qib_mmapf - mmap various structures into user space
 984  * @fp: the file pointer
 985  * @vma: the VM area
 986  *
 987  * We use this to have a shared buffer between the kernel and the user code
 988  * for the rcvhdr queue, egr buffers, and the per-context user regs and pio
 989  * buffers in the chip.  We have the open and close entries so we can bump
 990  * the ref count and keep the driver from being unloaded while still mapped.
 991  */
 992 static int qib_mmapf(struct file *fp, struct vm_area_struct *vma)
 993 {
 994         struct qib_ctxtdata *rcd;
 995         struct qib_devdata *dd;
 996         u64 pgaddr, ureg;
 997         unsigned piobufs, piocnt;
 998         int ret, match = 1;
 999
1000         rcd = ctxt_fp(fp);
1001         if (!rcd || !(vma->vm_flags & VM_SHARED)) {
1002                 ret = -EINVAL;
1003                 goto bail;
1004         }
1005         dd = rcd->dd;
1006
1007         /*
1008          * This is the qib_do_user_init() code, mapping the shared buffers
1009          * and per-context user registers into the user process. The address
1010          * referred to by vm_pgoff is the file offset passed via mmap().
1011          * For shared contexts, this is the kernel vmalloc() address of the
1012          * pages to share with the master.
1013          * For non-shared or master ctxts, this is a physical address.
1014          * We only do one mmap for each space mapped.
1015          */
1016         pgaddr = vma->vm_pgoff << PAGE_SHIFT;
1017
1018         /*
1019          * Check for 0 in case one of the allocations failed, but user
1020          * called mmap anyway.
1021          */
1022         if (!pgaddr)  {
1023                 ret = -EINVAL;
1024                 goto bail;
1025         }
1026
1027         /*
1028          * Physical addresses must fit in 40 bits for our hardware.
1029          * Check for kernel virtual addresses first, anything else must
1030          * match a HW or memory address.
1031          */
1032         ret = mmap_kvaddr(vma, pgaddr, rcd, subctxt_fp(fp));
1033         if (ret) {
1034                 if (ret > 0)
1035                         ret = 0;
1036                 goto bail;
1037         }
1038
1039         ureg = dd->uregbase + dd->ureg_align * rcd->ctxt;
1040         if (!rcd->subctxt_cnt) {
1041                 /* ctxt is not shared */
1042                 piocnt = rcd->piocnt;
1043                 piobufs = rcd->piobufs;
1044         } else if (!subctxt_fp(fp)) {
1045                 /* caller is the master */
1046                 piocnt = (rcd->piocnt / rcd->subctxt_cnt) +
1047                          (rcd->piocnt % rcd->subctxt_cnt);
1048                 piobufs = rcd->piobufs +
1049                         dd->palign * (rcd->piocnt - piocnt);
1050         } else {
1051                 unsigned slave = subctxt_fp(fp) - 1;
1052
1053                 /* caller is a slave */
1054                 piocnt = rcd->piocnt / rcd->subctxt_cnt;
1055                 piobufs = rcd->piobufs + dd->palign * piocnt * slave;
1056         }
1057
1058         if (pgaddr == ureg)
1059                 ret = mmap_ureg(vma, dd, ureg);
1060         else if (pgaddr == piobufs)
1061                 ret = mmap_piobufs(vma, dd, rcd, piobufs, piocnt);
1062         else if (pgaddr == dd->pioavailregs_phys)
1063                 /* in-memory copy of pioavail registers */
1064                 ret = qib_mmap_mem(vma, rcd, PAGE_SIZE,
1065                                    (void *) dd->pioavailregs_dma, 0,
1066                                    "pioavail registers");
1067         else if (pgaddr == rcd->rcvegr_phys)
1068                 ret = mmap_rcvegrbufs(vma, rcd);
1069         else if (pgaddr == (u64) rcd->rcvhdrq_phys)
1070                 /*
1071                  * The rcvhdrq itself; multiple pages, contiguous
1072                  * from an i/o perspective.  Shared contexts need
1073                  * to map r/w, so we allow writing.
1074                  */
1075                 ret = qib_mmap_mem(vma, rcd, rcd->rcvhdrq_size,
1076                                    rcd->rcvhdrq, 1, "rcvhdrq");
1077         else if (pgaddr == (u64) rcd->rcvhdrqtailaddr_phys)
1078                 /* in-memory copy of rcvhdrq tail register */
1079                 ret = qib_mmap_mem(vma, rcd, PAGE_SIZE,
1080                                    rcd->rcvhdrtail_kvaddr, 0,
1081                                    "rcvhdrq tail");
1082         else
1083                 match = 0;
1084         if (!match)
1085                 ret = -EINVAL;
1086
1087         vma->vm_private_data = NULL;
1088
1089         if (ret < 0)
1090                 qib_devinfo(dd->pcidev,
1091                          "mmap Failure %d: off %llx len %lx\n",
1092                          -ret, (unsigned long long)pgaddr,
1093                          vma->vm_end - vma->vm_start);
1094 bail:
1095         return ret;
1096 }
1097
1098 static unsigned int qib_poll_urgent(struct qib_ctxtdata *rcd,
1099                                     struct file *fp,
1100                                     struct poll_table_struct *pt)
1101 {
1102         struct qib_devdata *dd = rcd->dd;
1103         unsigned pollflag;
1104
1105         poll_wait(fp, &rcd->wait, pt);
1106
1107         spin_lock_irq(&dd->uctxt_lock);
1108         if (rcd->urgent != rcd->urgent_poll) {
1109                 pollflag = POLLIN | POLLRDNORM;
1110                 rcd->urgent_poll = rcd->urgent;
1111         } else {
1112                 pollflag = 0;
1113                 set_bit(QIB_CTXT_WAITING_URG, &rcd->flag);
1114         }
1115         spin_unlock_irq(&dd->uctxt_lock);
1116
1117         return pollflag;
1118 }
1119
1120 static unsigned int qib_poll_next(struct qib_ctxtdata *rcd,
1121                                   struct file *fp,
1122                                   struct poll_table_struct *pt)
1123 {
1124         struct qib_devdata *dd = rcd->dd;
1125         unsigned pollflag;
1126
1127         poll_wait(fp, &rcd->wait, pt);
1128
1129         spin_lock_irq(&dd->uctxt_lock);
1130         if (dd->f_hdrqempty(rcd)) {
1131                 set_bit(QIB_CTXT_WAITING_RCV, &rcd->flag);
1132                 dd->f_rcvctrl(rcd->ppd, QIB_RCVCTRL_INTRAVAIL_ENB, rcd->ctxt);
1133                 pollflag = 0;
1134         } else
1135                 pollflag = POLLIN | POLLRDNORM;
1136         spin_unlock_irq(&dd->uctxt_lock);
1137
1138         return pollflag;
1139 }
1140
1141 static unsigned int qib_poll(struct file *fp, struct poll_table_struct *pt)
1142 {
1143         struct qib_ctxtdata *rcd;
1144         unsigned pollflag;
1145
1146         rcd = ctxt_fp(fp);
1147         if (!rcd)
1148                 pollflag = POLLERR;
1149         else if (rcd->poll_type == QIB_POLL_TYPE_URGENT)
1150                 pollflag = qib_poll_urgent(rcd, fp, pt);
1151         else  if (rcd->poll_type == QIB_POLL_TYPE_ANYRCV)
1152                 pollflag = qib_poll_next(rcd, fp, pt);
1153         else /* invalid */
1154                 pollflag = POLLERR;
1155
1156         return pollflag;
1157 }
1158
1159 static void assign_ctxt_affinity(struct file *fp, struct qib_devdata *dd)
1160 {
1161         struct qib_filedata *fd = fp->private_data;
1162         const unsigned int weight = cpumask_weight(&current->cpus_allowed);
1163         const struct cpumask *local_mask = cpumask_of_pcibus(dd->pcidev->bus);
1164         int local_cpu;
1165
1166         /*
1167          * If process has NOT already set it's affinity, select and
1168          * reserve a processor for it on the local NUMA node.
1169          */
1170         if ((weight >= qib_cpulist_count) &&
1171                 (cpumask_weight(local_mask) <= qib_cpulist_count)) {
1172                 for_each_cpu(local_cpu, local_mask)
1173                         if (!test_and_set_bit(local_cpu, qib_cpulist)) {
1174                                 fd->rec_cpu_num = local_cpu;
1175                                 return;
1176                         }
1177         }
1178
1179         /*
1180          * If process has NOT already set it's affinity, select and
1181          * reserve a processor for it, as a rendevous for all
1182          * users of the driver.  If they don't actually later
1183          * set affinity to this cpu, or set it to some other cpu,
1184          * it just means that sooner or later we don't recommend
1185          * a cpu, and let the scheduler do it's best.
1186          */
1187         if (weight >= qib_cpulist_count) {
1188                 int cpu;
1189
1190                 cpu = find_first_zero_bit(qib_cpulist,
1191                                           qib_cpulist_count);
1192                 if (cpu == qib_cpulist_count)
1193                         qib_dev_err(dd,
1194                         "no cpus avail for affinity PID %u\n",
1195                         current->pid);
1196                 else {
1197                         __set_bit(cpu, qib_cpulist);
1198                         fd->rec_cpu_num = cpu;
1199                 }
1200         }
1201 }
1202
1203 /*
1204  * Check that userland and driver are compatible for subcontexts.
1205  */
1206 static int qib_compatible_subctxts(int user_swmajor, int user_swminor)
1207 {
1208         /* this code is written long-hand for clarity */
1209         if (QIB_USER_SWMAJOR != user_swmajor) {
1210                 /* no promise of compatibility if major mismatch */
1211                 return 0;
1212         }
1213         if (QIB_USER_SWMAJOR == 1) {
1214                 switch (QIB_USER_SWMINOR) {
1215                 case 0:
1216                 case 1:
1217                 case 2:
1218                         /* no subctxt implementation so cannot be compatible */
1219                         return 0;
1220                 case 3:
1221                         /* 3 is only compatible with itself */
1222                         return user_swminor == 3;
1223                 default:
1224                         /* >= 4 are compatible (or are expected to be) */
1225                         return user_swminor <= QIB_USER_SWMINOR;
1226                 }
1227         }
1228         /* make no promises yet for future major versions */
1229         return 0;
1230 }
1231
1232 static int init_subctxts(struct qib_devdata *dd,
1233                          struct qib_ctxtdata *rcd,
1234                          const struct qib_user_info *uinfo)
1235 {
1236         int ret = 0;
1237         unsigned num_subctxts;
1238         size_t size;
1239
1240         /*
1241          * If the user is requesting zero subctxts,
1242          * skip the subctxt allocation.
1243          */
1244         if (uinfo->spu_subctxt_cnt <= 0)
1245                 goto bail;
1246         num_subctxts = uinfo->spu_subctxt_cnt;
1247
1248         /* Check for subctxt compatibility */
1249         if (!qib_compatible_subctxts(uinfo->spu_userversion >> 16,
1250                 uinfo->spu_userversion & 0xffff)) {
1251                 qib_devinfo(dd->pcidev,
1252                          "Mismatched user version (%d.%d) and driver version (%d.%d) while context sharing. Ensure that driver and library are from the same release.\n",
1253                          (int) (uinfo->spu_userversion >> 16),
1254                          (int) (uinfo->spu_userversion & 0xffff),
1255                          QIB_USER_SWMAJOR, QIB_USER_SWMINOR);
1256                 goto bail;
1257         }
1258         if (num_subctxts > QLOGIC_IB_MAX_SUBCTXT) {
1259                 ret = -EINVAL;
1260                 goto bail;
1261         }
1262
1263         rcd->subctxt_uregbase = vmalloc_user(PAGE_SIZE * num_subctxts);
1264         if (!rcd->subctxt_uregbase) {
1265                 ret = -ENOMEM;
1266                 goto bail;
1267         }
1268         /* Note: rcd->rcvhdrq_size isn't initialized yet. */
1269         size = ALIGN(dd->rcvhdrcnt * dd->rcvhdrentsize *
1270                      sizeof(u32), PAGE_SIZE) * num_subctxts;
1271         rcd->subctxt_rcvhdr_base = vmalloc_user(size);
1272         if (!rcd->subctxt_rcvhdr_base) {
1273                 ret = -ENOMEM;
1274                 goto bail_ureg;
1275         }
1276
1277         rcd->subctxt_rcvegrbuf = vmalloc_user(rcd->rcvegrbuf_chunks *
1278                                               rcd->rcvegrbuf_size *
1279                                               num_subctxts);
1280         if (!rcd->subctxt_rcvegrbuf) {
1281                 ret = -ENOMEM;
1282                 goto bail_rhdr;
1283         }
1284
1285         rcd->subctxt_cnt = uinfo->spu_subctxt_cnt;
1286         rcd->subctxt_id = uinfo->spu_subctxt_id;
1287         rcd->active_slaves = 1;
1288         rcd->redirect_seq_cnt = 1;
1289         set_bit(QIB_CTXT_MASTER_UNINIT, &rcd->flag);
1290         goto bail;
1291
1292 bail_rhdr:
1293         vfree(rcd->subctxt_rcvhdr_base);
1294 bail_ureg:
1295         vfree(rcd->subctxt_uregbase);
1296         rcd->subctxt_uregbase = NULL;
1297 bail:
1298         return ret;
1299 }
1300
1301 static int setup_ctxt(struct qib_pportdata *ppd, int ctxt,
1302                       struct file *fp, const struct qib_user_info *uinfo)
1303 {
1304         struct qib_filedata *fd = fp->private_data;
1305         struct qib_devdata *dd = ppd->dd;
1306         struct qib_ctxtdata *rcd;
1307         void *ptmp = NULL;
1308         int ret;
1309         int numa_id;
1310
1311         assign_ctxt_affinity(fp, dd);
1312
1313         numa_id = qib_numa_aware ? ((fd->rec_cpu_num != -1) ?
1314                 cpu_to_node(fd->rec_cpu_num) :
1315                 numa_node_id()) : dd->assigned_node_id;
1316
1317         rcd = qib_create_ctxtdata(ppd, ctxt, numa_id);
1318
1319         /*
1320          * Allocate memory for use in qib_tid_update() at open to
1321          * reduce cost of expected send setup per message segment
1322          */
1323         if (rcd)
1324                 ptmp = kmalloc(dd->rcvtidcnt * sizeof(u16) +
1325                                dd->rcvtidcnt * sizeof(struct page **),
1326                                GFP_KERNEL);
1327
1328         if (!rcd || !ptmp) {
1329                 qib_dev_err(dd,
1330                         "Unable to allocate ctxtdata memory, failing open\n");
1331                 ret = -ENOMEM;
1332                 goto bailerr;
1333         }
1334         rcd->userversion = uinfo->spu_userversion;
1335         ret = init_subctxts(dd, rcd, uinfo);
1336         if (ret)
1337                 goto bailerr;
1338         rcd->tid_pg_list = ptmp;
1339         rcd->pid = current->pid;
1340         init_waitqueue_head(&dd->rcd[ctxt]->wait);
1341         strlcpy(rcd->comm, current->comm, sizeof(rcd->comm));
1342         ctxt_fp(fp) = rcd;
1343         qib_stats.sps_ctxts++;
1344         dd->freectxts--;
1345         ret = 0;
1346         goto bail;
1347
1348 bailerr:
1349         if (fd->rec_cpu_num != -1)
1350                 __clear_bit(fd->rec_cpu_num, qib_cpulist);
1351
1352         dd->rcd[ctxt] = NULL;
1353         kfree(rcd);
1354         kfree(ptmp);
1355 bail:
1356         return ret;
1357 }
1358
1359 static inline int usable(struct qib_pportdata *ppd)
1360 {
1361         struct qib_devdata *dd = ppd->dd;
1362
1363         return dd && (dd->flags & QIB_PRESENT) && dd->kregbase && ppd->lid &&
1364                 (ppd->lflags & QIBL_LINKACTIVE);
1365 }
1366
1367 /*
1368  * Select a context on the given device, either using a requested port
1369  * or the port based on the context number.
1370  */
1371 static int choose_port_ctxt(struct file *fp, struct qib_devdata *dd, u32 port,
1372                             const struct qib_user_info *uinfo)
1373 {
1374         struct qib_pportdata *ppd = NULL;
1375         int ret, ctxt;
1376
1377         if (port) {
1378                 if (!usable(dd->pport + port - 1)) {
1379                         ret = -ENETDOWN;
1380                         goto done;
1381                 } else
1382                         ppd = dd->pport + port - 1;
1383         }
1384         for (ctxt = dd->first_user_ctxt; ctxt < dd->cfgctxts && dd->rcd[ctxt];
1385              ctxt++)
1386                 ;
1387         if (ctxt == dd->cfgctxts) {
1388                 ret = -EBUSY;
1389                 goto done;
1390         }
1391         if (!ppd) {
1392                 u32 pidx = ctxt % dd->num_pports;
1393
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
1442                 /* find device (with ACTIVE ports) with fewest ctxts in use */
1443                 for (ndev = 0; ndev < devmax; ndev++) {
1444                         struct qib_devdata *dd = qib_lookup(ndev);
1445                         unsigned cused = 0, cfree = 0, pusable = 0;
1446
1447                         if (!dd)
1448                                 continue;
1449                         if (port && port <= dd->num_pports &&
1450                             usable(dd->pport + port - 1))
1451                                 pusable = 1;
1452                         else
1453                                 for (i = 0; i < dd->num_pports; i++)
1454                                         if (usable(dd->pport + i))
1455                                                 pusable++;
1456                         if (!pusable)
1457                                 continue;
1458                         for (ctxt = dd->first_user_ctxt; ctxt < dd->cfgctxts;
1459                              ctxt++)
1460                                 if (dd->rcd[ctxt])
1461                                         cused++;
1462                                 else
1463                                         cfree++;
1464                         if (cfree && cused < inuse) {
1465                                 udd = dd;
1466                                 inuse = cused;
1467                         }
1468                 }
1469                 if (udd) {
1470                         ret = choose_port_ctxt(fp, udd, port, uinfo);
1471                         goto done;
1472                 }
1473         } else {
1474                 for (ndev = 0; ndev < devmax; ndev++) {
1475                         struct qib_devdata *dd = qib_lookup(ndev);
1476
1477                         if (dd) {
1478                                 ret = choose_port_ctxt(fp, dd, port, uinfo);
1479                                 if (!ret)
1480                                         goto done;
1481                                 if (ret == -EBUSY)
1482                                         dusable++;
1483                         }
1484                 }
1485         }
1486         ret = dusable ? -EBUSY : -ENETDOWN;
1487
1488 done:
1489         return ret;
1490 }
1491
1492 static int find_shared_ctxt(struct file *fp,
1493                             const struct qib_user_info *uinfo)
1494 {
1495         int devmax, ndev, i;
1496         int ret = 0;
1497
1498         devmax = qib_count_units(NULL, NULL);
1499
1500         for (ndev = 0; ndev < devmax; ndev++) {
1501                 struct qib_devdata *dd = qib_lookup(ndev);
1502
1503                 /* device portion of usable() */
1504                 if (!(dd && (dd->flags & QIB_PRESENT) && dd->kregbase))
1505                         continue;
1506                 for (i = dd->first_user_ctxt; i < dd->cfgctxts; i++) {
1507                         struct qib_ctxtdata *rcd = dd->rcd[i];
1508
1509                         /* Skip ctxts which are not yet open */
1510                         if (!rcd || !rcd->cnt)
1511                                 continue;
1512                         /* Skip ctxt if it doesn't match the requested one */
1513                         if (rcd->subctxt_id != uinfo->spu_subctxt_id)
1514                                 continue;
1515                         /* Verify the sharing process matches the master */
1516                         if (rcd->subctxt_cnt != uinfo->spu_subctxt_cnt ||
1517                             rcd->userversion != uinfo->spu_userversion ||
1518                             rcd->cnt >= rcd->subctxt_cnt) {
1519                                 ret = -EINVAL;
1520                                 goto done;
1521                         }
1522                         ctxt_fp(fp) = rcd;
1523                         subctxt_fp(fp) = rcd->cnt++;
1524                         rcd->subpid[subctxt_fp(fp)] = current->pid;
1525                         tidcursor_fp(fp) = 0;
1526                         rcd->active_slaves |= 1 << subctxt_fp(fp);
1527                         ret = 1;
1528                         goto done;
1529                 }
1530         }
1531
1532 done:
1533         return ret;
1534 }
1535
1536 static int qib_open(struct inode *in, struct file *fp)
1537 {
1538         /* The real work is performed later in qib_assign_ctxt() */
1539         fp->private_data = kzalloc(sizeof(struct qib_filedata), GFP_KERNEL);
1540         if (fp->private_data) /* no cpu affinity by default */
1541                 ((struct qib_filedata *)fp->private_data)->rec_cpu_num = -1;
1542         return fp->private_data ? 0 : -ENOMEM;
1543 }
1544
1545 static int find_hca(unsigned int cpu, int *unit)
1546 {
1547         int ret = 0, devmax, npresent, nup, ndev;
1548
1549         *unit = -1;
1550
1551         devmax = qib_count_units(&npresent, &nup);
1552         if (!npresent) {
1553                 ret = -ENXIO;
1554                 goto done;
1555         }
1556         if (!nup) {
1557                 ret = -ENETDOWN;
1558                 goto done;
1559         }
1560         for (ndev = 0; ndev < devmax; ndev++) {
1561                 struct qib_devdata *dd = qib_lookup(ndev);
1562
1563                 if (dd) {
1564                         if (pcibus_to_node(dd->pcidev->bus) < 0) {
1565                                 ret = -EINVAL;
1566                                 goto done;
1567                         }
1568                         if (cpu_to_node(cpu) ==
1569                                 pcibus_to_node(dd->pcidev->bus)) {
1570                                 *unit = ndev;
1571                                 goto done;
1572                         }
1573                 }
1574         }
1575 done:
1576         return ret;
1577 }
1578
1579 static int do_qib_user_sdma_queue_create(struct file *fp)
1580 {
1581         struct qib_filedata *fd = fp->private_data;
1582         struct qib_ctxtdata *rcd = fd->rcd;
1583         struct qib_devdata *dd = rcd->dd;
1584
1585         if (dd->flags & QIB_HAS_SEND_DMA) {
1586
1587                 fd->pq = qib_user_sdma_queue_create(&dd->pcidev->dev,
1588                                                     dd->unit,
1589                                                     rcd->ctxt,
1590                                                     fd->subctxt);
1591                 if (!fd->pq)
1592                         return -ENOMEM;
1593         }
1594
1595         return 0;
1596 }
1597
1598 /*
1599  * Get ctxt early, so can set affinity prior to memory allocation.
1600  */
1601 static int qib_assign_ctxt(struct file *fp, const struct qib_user_info *uinfo)
1602 {
1603         int ret;
1604         int i_minor;
1605         unsigned swmajor, swminor, alg = QIB_PORT_ALG_ACROSS;
1606
1607         /* Check to be sure we haven't already initialized this file */
1608         if (ctxt_fp(fp)) {
1609                 ret = -EINVAL;
1610                 goto done;
1611         }
1612
1613         /* for now, if major version is different, bail */
1614         swmajor = uinfo->spu_userversion >> 16;
1615         if (swmajor != QIB_USER_SWMAJOR) {
1616                 ret = -ENODEV;
1617                 goto done;
1618         }
1619
1620         swminor = uinfo->spu_userversion & 0xffff;
1621
1622         if (swminor >= 11 && uinfo->spu_port_alg < QIB_PORT_ALG_COUNT)
1623                 alg = uinfo->spu_port_alg;
1624
1625         mutex_lock(&qib_mutex);
1626
1627         if (qib_compatible_subctxts(swmajor, swminor) &&
1628             uinfo->spu_subctxt_cnt) {
1629                 ret = find_shared_ctxt(fp, uinfo);
1630                 if (ret > 0) {
1631                         ret = do_qib_user_sdma_queue_create(fp);
1632                         if (!ret)
1633                                 assign_ctxt_affinity(fp, (ctxt_fp(fp))->dd);
1634                         goto done_ok;
1635                 }
1636         }
1637
1638         i_minor = iminor(file_inode(fp)) - QIB_USER_MINOR_BASE;
1639         if (i_minor)
1640                 ret = find_free_ctxt(i_minor - 1, fp, uinfo);
1641         else {
1642                 int unit;
1643                 const unsigned int cpu = cpumask_first(&current->cpus_allowed);
1644                 const unsigned int weight =
1645                         cpumask_weight(&current->cpus_allowed);
1646
1647                 if (weight == 1 && !test_bit(cpu, qib_cpulist))
1648                         if (!find_hca(cpu, &unit) && unit >= 0)
1649                                 if (!find_free_ctxt(unit, fp, uinfo)) {
1650                                         ret = 0;
1651                                         goto done_chk_sdma;
1652                                 }
1653                 ret = get_a_ctxt(fp, uinfo, alg);
1654         }
1655
1656 done_chk_sdma:
1657         if (!ret)
1658                 ret = do_qib_user_sdma_queue_create(fp);
1659 done_ok:
1660         mutex_unlock(&qib_mutex);
1661
1662 done:
1663         return ret;
1664 }
1665
1666
1667 static int qib_do_user_init(struct file *fp,
1668                             const struct qib_user_info *uinfo)
1669 {
1670         int ret;
1671         struct qib_ctxtdata *rcd = ctxt_fp(fp);
1672         struct qib_devdata *dd;
1673         unsigned uctxt;
1674
1675         /* Subctxts don't need to initialize anything since master did it. */
1676         if (subctxt_fp(fp)) {
1677                 ret = wait_event_interruptible(rcd->wait,
1678                         !test_bit(QIB_CTXT_MASTER_UNINIT, &rcd->flag));
1679                 goto bail;
1680         }
1681
1682         dd = rcd->dd;
1683
1684         /* some ctxts may get extra buffers, calculate that here */
1685         uctxt = rcd->ctxt - dd->first_user_ctxt;
1686         if (uctxt < dd->ctxts_extrabuf) {
1687                 rcd->piocnt = dd->pbufsctxt + 1;
1688                 rcd->pio_base = rcd->piocnt * uctxt;
1689         } else {
1690                 rcd->piocnt = dd->pbufsctxt;
1691                 rcd->pio_base = rcd->piocnt * uctxt +
1692                         dd->ctxts_extrabuf;
1693         }
1694
1695         /*
1696          * All user buffers are 2KB buffers.  If we ever support
1697          * giving 4KB buffers to user processes, this will need some
1698          * work.  Can't use piobufbase directly, because it has
1699          * both 2K and 4K buffer base values.  So check and handle.
1700          */
1701         if ((rcd->pio_base + rcd->piocnt) > dd->piobcnt2k) {
1702                 if (rcd->pio_base >= dd->piobcnt2k) {
1703                         qib_dev_err(dd,
1704                                     "%u:ctxt%u: no 2KB buffers available\n",
1705                                     dd->unit, rcd->ctxt);
1706                         ret = -ENOBUFS;
1707                         goto bail;
1708                 }
1709                 rcd->piocnt = dd->piobcnt2k - rcd->pio_base;
1710                 qib_dev_err(dd, "Ctxt%u: would use 4KB bufs, using %u\n",
1711                             rcd->ctxt, rcd->piocnt);
1712         }
1713
1714         rcd->piobufs = dd->pio2k_bufbase + rcd->pio_base * dd->palign;
1715         qib_chg_pioavailkernel(dd, rcd->pio_base, rcd->piocnt,
1716                                TXCHK_CHG_TYPE_USER, rcd);
1717         /*
1718          * try to ensure that processes start up with consistent avail update
1719          * for their own range, at least.   If system very quiet, it might
1720          * have the in-memory copy out of date at startup for this range of
1721          * buffers, when a context gets re-used.  Do after the chg_pioavail
1722          * and before the rest of setup, so it's "almost certain" the dma
1723          * will have occurred (can't 100% guarantee, but should be many
1724          * decimals of 9s, with this ordering), given how much else happens
1725          * after this.
1726          */
1727         dd->f_sendctrl(dd->pport, QIB_SENDCTRL_AVAIL_BLIP);
1728
1729         /*
1730          * Now allocate the rcvhdr Q and eager TIDs; skip the TID
1731          * array for time being.  If rcd->ctxt > chip-supported,
1732          * we need to do extra stuff here to handle by handling overflow
1733          * through ctxt 0, someday
1734          */
1735         ret = qib_create_rcvhdrq(dd, rcd);
1736         if (!ret)
1737                 ret = qib_setup_eagerbufs(rcd);
1738         if (ret)
1739                 goto bail_pio;
1740
1741         rcd->tidcursor = 0; /* start at beginning after open */
1742
1743         /* initialize poll variables... */
1744         rcd->urgent = 0;
1745         rcd->urgent_poll = 0;
1746
1747         /*
1748          * Now enable the ctxt for receive.
1749          * For chips that are set to DMA the tail register to memory
1750          * when they change (and when the update bit transitions from
1751          * 0 to 1.  So for those chips, we turn it off and then back on.
1752          * This will (very briefly) affect any other open ctxts, but the
1753          * duration is very short, and therefore isn't an issue.  We
1754          * explicitly set the in-memory tail copy to 0 beforehand, so we
1755          * don't have to wait to be sure the DMA update has happened
1756          * (chip resets head/tail to 0 on transition to enable).
1757          */
1758         if (rcd->rcvhdrtail_kvaddr)
1759                 qib_clear_rcvhdrtail(rcd);
1760
1761         dd->f_rcvctrl(rcd->ppd, QIB_RCVCTRL_CTXT_ENB | QIB_RCVCTRL_TIDFLOW_ENB,
1762                       rcd->ctxt);
1763
1764         /* Notify any waiting slaves */
1765         if (rcd->subctxt_cnt) {
1766                 clear_bit(QIB_CTXT_MASTER_UNINIT, &rcd->flag);
1767                 wake_up(&rcd->wait);
1768         }
1769         return 0;
1770
1771 bail_pio:
1772         qib_chg_pioavailkernel(dd, rcd->pio_base, rcd->piocnt,
1773                                TXCHK_CHG_TYPE_KERN, rcd);
1774 bail:
1775         return ret;
1776 }
1777
1778 /**
1779  * unlock_exptid - unlock any expected TID entries context still had in use
1780  * @rcd: ctxt
1781  *
1782  * We don't actually update the chip here, because we do a bulk update
1783  * below, using f_clear_tids.
1784  */
1785 static void unlock_expected_tids(struct qib_ctxtdata *rcd)
1786 {
1787         struct qib_devdata *dd = rcd->dd;
1788         int ctxt_tidbase = rcd->ctxt * dd->rcvtidcnt;
1789         int i, cnt = 0, maxtid = ctxt_tidbase + dd->rcvtidcnt;
1790
1791         for (i = ctxt_tidbase; i < maxtid; i++) {
1792                 struct page *p = dd->pageshadow[i];
1793                 dma_addr_t phys;
1794
1795                 if (!p)
1796                         continue;
1797
1798                 phys = dd->physshadow[i];
1799                 dd->physshadow[i] = dd->tidinvalid;
1800                 dd->pageshadow[i] = NULL;
1801                 pci_unmap_page(dd->pcidev, phys, PAGE_SIZE,
1802                                PCI_DMA_FROMDEVICE);
1803                 qib_release_user_pages(&p, 1);
1804                 cnt++;
1805         }
1806 }
1807
1808 static int qib_close(struct inode *in, struct file *fp)
1809 {
1810         int ret = 0;
1811         struct qib_filedata *fd;
1812         struct qib_ctxtdata *rcd;
1813         struct qib_devdata *dd;
1814         unsigned long flags;
1815         unsigned ctxt;
1816         pid_t pid;
1817
1818         mutex_lock(&qib_mutex);
1819
1820         fd = fp->private_data;
1821         fp->private_data = NULL;
1822         rcd = fd->rcd;
1823         if (!rcd) {
1824                 mutex_unlock(&qib_mutex);
1825                 goto bail;
1826         }
1827
1828         dd = rcd->dd;
1829
1830         /* ensure all pio buffer writes in progress are flushed */
1831         qib_flush_wc();
1832
1833         /* drain user sdma queue */
1834         if (fd->pq) {
1835                 qib_user_sdma_queue_drain(rcd->ppd, fd->pq);
1836                 qib_user_sdma_queue_destroy(fd->pq);
1837         }
1838
1839         if (fd->rec_cpu_num != -1)
1840                 __clear_bit(fd->rec_cpu_num, qib_cpulist);
1841
1842         if (--rcd->cnt) {
1843                 /*
1844                  * XXX If the master closes the context before the slave(s),
1845                  * revoke the mmap for the eager receive queue so
1846                  * the slave(s) don't wait for receive data forever.
1847                  */
1848                 rcd->active_slaves &= ~(1 << fd->subctxt);
1849                 rcd->subpid[fd->subctxt] = 0;
1850                 mutex_unlock(&qib_mutex);
1851                 goto bail;
1852         }
1853
1854         /* early; no interrupt users after this */
1855         spin_lock_irqsave(&dd->uctxt_lock, flags);
1856         ctxt = rcd->ctxt;
1857         dd->rcd[ctxt] = NULL;
1858         pid = rcd->pid;
1859         rcd->pid = 0;
1860         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
1861
1862         if (rcd->rcvwait_to || rcd->piowait_to ||
1863             rcd->rcvnowait || rcd->pionowait) {
1864                 rcd->rcvwait_to = 0;
1865                 rcd->piowait_to = 0;
1866                 rcd->rcvnowait = 0;
1867                 rcd->pionowait = 0;
1868         }
1869         if (rcd->flag)
1870                 rcd->flag = 0;
1871
1872         if (dd->kregbase) {
1873                 /* atomically clear receive enable ctxt and intr avail. */
1874                 dd->f_rcvctrl(rcd->ppd, QIB_RCVCTRL_CTXT_DIS |
1875                                   QIB_RCVCTRL_INTRAVAIL_DIS, ctxt);
1876
1877                 /* clean up the pkeys for this ctxt user */
1878                 qib_clean_part_key(rcd, dd);
1879                 qib_disarm_piobufs(dd, rcd->pio_base, rcd->piocnt);
1880                 qib_chg_pioavailkernel(dd, rcd->pio_base,
1881                                        rcd->piocnt, TXCHK_CHG_TYPE_KERN, NULL);
1882
1883                 dd->f_clear_tids(dd, rcd);
1884
1885                 if (dd->pageshadow)
1886                         unlock_expected_tids(rcd);
1887                 qib_stats.sps_ctxts--;
1888                 dd->freectxts++;
1889         }
1890
1891         mutex_unlock(&qib_mutex);
1892         qib_free_ctxtdata(dd, rcd); /* after releasing the mutex */
1893
1894 bail:
1895         kfree(fd);
1896         return ret;
1897 }
1898
1899 static int qib_ctxt_info(struct file *fp, struct qib_ctxt_info __user *uinfo)
1900 {
1901         struct qib_ctxt_info info;
1902         int ret;
1903         size_t sz;
1904         struct qib_ctxtdata *rcd = ctxt_fp(fp);
1905         struct qib_filedata *fd;
1906
1907         fd = fp->private_data;
1908
1909         info.num_active = qib_count_active_units();
1910         info.unit = rcd->dd->unit;
1911         info.port = rcd->ppd->port;
1912         info.ctxt = rcd->ctxt;
1913         info.subctxt =  subctxt_fp(fp);
1914         /* Number of user ctxts available for this device. */
1915         info.num_ctxts = rcd->dd->cfgctxts - rcd->dd->first_user_ctxt;
1916         info.num_subctxts = rcd->subctxt_cnt;
1917         info.rec_cpu = fd->rec_cpu_num;
1918         sz = sizeof(info);
1919
1920         if (copy_to_user(uinfo, &info, sz)) {
1921                 ret = -EFAULT;
1922                 goto bail;
1923         }
1924         ret = 0;
1925
1926 bail:
1927         return ret;
1928 }
1929
1930 static int qib_sdma_get_inflight(struct qib_user_sdma_queue *pq,
1931                                  u32 __user *inflightp)
1932 {
1933         const u32 val = qib_user_sdma_inflight_counter(pq);
1934
1935         if (put_user(val, inflightp))
1936                 return -EFAULT;
1937
1938         return 0;
1939 }
1940
1941 static int qib_sdma_get_complete(struct qib_pportdata *ppd,
1942                                  struct qib_user_sdma_queue *pq,
1943                                  u32 __user *completep)
1944 {
1945         u32 val;
1946         int err;
1947
1948         if (!pq)
1949                 return -EINVAL;
1950
1951         err = qib_user_sdma_make_progress(ppd, pq);
1952         if (err < 0)
1953                 return err;
1954
1955         val = qib_user_sdma_complete_counter(pq);
1956         if (put_user(val, completep))
1957                 return -EFAULT;
1958
1959         return 0;
1960 }
1961
1962 static int disarm_req_delay(struct qib_ctxtdata *rcd)
1963 {
1964         int ret = 0;
1965
1966         if (!usable(rcd->ppd)) {
1967                 int i;
1968                 /*
1969                  * if link is down, or otherwise not usable, delay
1970                  * the caller up to 30 seconds, so we don't thrash
1971                  * in trying to get the chip back to ACTIVE, and
1972                  * set flag so they make the call again.
1973                  */
1974                 if (rcd->user_event_mask) {
1975                         /*
1976                          * subctxt_cnt is 0 if not shared, so do base
1977                          * separately, first, then remaining subctxt, if any
1978                          */
1979                         set_bit(_QIB_EVENT_DISARM_BUFS_BIT,
1980                                 &rcd->user_event_mask[0]);
1981                         for (i = 1; i < rcd->subctxt_cnt; i++)
1982                                 set_bit(_QIB_EVENT_DISARM_BUFS_BIT,
1983                                         &rcd->user_event_mask[i]);
1984                 }
1985                 for (i = 0; !usable(rcd->ppd) && i < 300; i++)
1986                         msleep(100);
1987                 ret = -ENETDOWN;
1988         }
1989         return ret;
1990 }
1991
1992 /*
1993  * Find all user contexts in use, and set the specified bit in their
1994  * event mask.
1995  * See also find_ctxt() for a similar use, that is specific to send buffers.
1996  */
1997 int qib_set_uevent_bits(struct qib_pportdata *ppd, const int evtbit)
1998 {
1999         struct qib_ctxtdata *rcd;
2000         unsigned ctxt;
2001         int ret = 0;
2002         unsigned long flags;
2003
2004         spin_lock_irqsave(&ppd->dd->uctxt_lock, flags);
2005         for (ctxt = ppd->dd->first_user_ctxt; ctxt < ppd->dd->cfgctxts;
2006              ctxt++) {
2007                 rcd = ppd->dd->rcd[ctxt];
2008                 if (!rcd)
2009                         continue;
2010                 if (rcd->user_event_mask) {
2011                         int i;
2012                         /*
2013                          * subctxt_cnt is 0 if not shared, so do base
2014                          * separately, first, then remaining subctxt, if any
2015                          */
2016                         set_bit(evtbit, &rcd->user_event_mask[0]);
2017                         for (i = 1; i < rcd->subctxt_cnt; i++)
2018                                 set_bit(evtbit, &rcd->user_event_mask[i]);
2019                 }
2020                 ret = 1;
2021                 break;
2022         }
2023         spin_unlock_irqrestore(&ppd->dd->uctxt_lock, flags);
2024
2025         return ret;
2026 }
2027
2028 /*
2029  * clear the event notifier events for this context.
2030  * For the DISARM_BUFS case, we also take action (this obsoletes
2031  * the older QIB_CMD_DISARM_BUFS, but we keep it for backwards
2032  * compatibility.
2033  * Other bits don't currently require actions, just atomically clear.
2034  * User process then performs actions appropriate to bit having been
2035  * set, if desired, and checks again in future.
2036  */
2037 static int qib_user_event_ack(struct qib_ctxtdata *rcd, int subctxt,
2038                               unsigned long events)
2039 {
2040         int ret = 0, i;
2041
2042         for (i = 0; i <= _QIB_MAX_EVENT_BIT; i++) {
2043                 if (!test_bit(i, &events))
2044                         continue;
2045                 if (i == _QIB_EVENT_DISARM_BUFS_BIT) {
2046                         (void)qib_disarm_piobufs_ifneeded(rcd);
2047                         ret = disarm_req_delay(rcd);
2048                 } else
2049                         clear_bit(i, &rcd->user_event_mask[subctxt]);
2050         }
2051         return ret;
2052 }
2053
2054 static ssize_t qib_write(struct file *fp, const char __user *data,
2055                          size_t count, loff_t *off)
2056 {
2057         const struct qib_cmd __user *ucmd;
2058         struct qib_ctxtdata *rcd;
2059         const void __user *src;
2060         size_t consumed, copy = 0;
2061         struct qib_cmd cmd;
2062         ssize_t ret = 0;
2063         void *dest;
2064
2065         if (count < sizeof(cmd.type)) {
2066                 ret = -EINVAL;
2067                 goto bail;
2068         }
2069
2070         ucmd = (const struct qib_cmd __user *) data;
2071
2072         if (copy_from_user(&cmd.type, &ucmd->type, sizeof(cmd.type))) {
2073                 ret = -EFAULT;
2074                 goto bail;
2075         }
2076
2077         consumed = sizeof(cmd.type);
2078
2079         switch (cmd.type) {
2080         case QIB_CMD_ASSIGN_CTXT:
2081         case QIB_CMD_USER_INIT:
2082                 copy = sizeof(cmd.cmd.user_info);
2083                 dest = &cmd.cmd.user_info;
2084                 src = &ucmd->cmd.user_info;
2085                 break;
2086
2087         case QIB_CMD_RECV_CTRL:
2088                 copy = sizeof(cmd.cmd.recv_ctrl);
2089                 dest = &cmd.cmd.recv_ctrl;
2090                 src = &ucmd->cmd.recv_ctrl;
2091                 break;
2092
2093         case QIB_CMD_CTXT_INFO:
2094                 copy = sizeof(cmd.cmd.ctxt_info);
2095                 dest = &cmd.cmd.ctxt_info;
2096                 src = &ucmd->cmd.ctxt_info;
2097                 break;
2098
2099         case QIB_CMD_TID_UPDATE:
2100         case QIB_CMD_TID_FREE:
2101                 copy = sizeof(cmd.cmd.tid_info);
2102                 dest = &cmd.cmd.tid_info;
2103                 src = &ucmd->cmd.tid_info;
2104                 break;
2105
2106         case QIB_CMD_SET_PART_KEY:
2107                 copy = sizeof(cmd.cmd.part_key);
2108                 dest = &cmd.cmd.part_key;
2109                 src = &ucmd->cmd.part_key;
2110                 break;
2111
2112         case QIB_CMD_DISARM_BUFS:
2113         case QIB_CMD_PIOAVAILUPD: /* force an update of PIOAvail reg */
2114                 copy = 0;
2115                 src = NULL;
2116                 dest = NULL;
2117                 break;
2118
2119         case QIB_CMD_POLL_TYPE:
2120                 copy = sizeof(cmd.cmd.poll_type);
2121                 dest = &cmd.cmd.poll_type;
2122                 src = &ucmd->cmd.poll_type;
2123                 break;
2124
2125         case QIB_CMD_ARMLAUNCH_CTRL:
2126                 copy = sizeof(cmd.cmd.armlaunch_ctrl);
2127                 dest = &cmd.cmd.armlaunch_ctrl;
2128                 src = &ucmd->cmd.armlaunch_ctrl;
2129                 break;
2130
2131         case QIB_CMD_SDMA_INFLIGHT:
2132                 copy = sizeof(cmd.cmd.sdma_inflight);
2133                 dest = &cmd.cmd.sdma_inflight;
2134                 src = &ucmd->cmd.sdma_inflight;
2135                 break;
2136
2137         case QIB_CMD_SDMA_COMPLETE:
2138                 copy = sizeof(cmd.cmd.sdma_complete);
2139                 dest = &cmd.cmd.sdma_complete;
2140                 src = &ucmd->cmd.sdma_complete;
2141                 break;
2142
2143         case QIB_CMD_ACK_EVENT:
2144                 copy = sizeof(cmd.cmd.event_mask);
2145                 dest = &cmd.cmd.event_mask;
2146                 src = &ucmd->cmd.event_mask;
2147                 break;
2148
2149         default:
2150                 ret = -EINVAL;
2151                 goto bail;
2152         }
2153
2154         if (copy) {
2155                 if ((count - consumed) < copy) {
2156                         ret = -EINVAL;
2157                         goto bail;
2158                 }
2159                 if (copy_from_user(dest, src, copy)) {
2160                         ret = -EFAULT;
2161                         goto bail;
2162                 }
2163                 consumed += copy;
2164         }
2165
2166         rcd = ctxt_fp(fp);
2167         if (!rcd && cmd.type != QIB_CMD_ASSIGN_CTXT) {
2168                 ret = -EINVAL;
2169                 goto bail;
2170         }
2171
2172         switch (cmd.type) {
2173         case QIB_CMD_ASSIGN_CTXT:
2174                 ret = qib_assign_ctxt(fp, &cmd.cmd.user_info);
2175                 if (ret)
2176                         goto bail;
2177                 break;
2178
2179         case QIB_CMD_USER_INIT:
2180                 ret = qib_do_user_init(fp, &cmd.cmd.user_info);
2181                 if (ret)
2182                         goto bail;
2183                 ret = qib_get_base_info(fp, (void __user *) (unsigned long)
2184                                         cmd.cmd.user_info.spu_base_info,
2185                                         cmd.cmd.user_info.spu_base_info_size);
2186                 break;
2187
2188         case QIB_CMD_RECV_CTRL:
2189                 ret = qib_manage_rcvq(rcd, subctxt_fp(fp), cmd.cmd.recv_ctrl);
2190                 break;
2191
2192         case QIB_CMD_CTXT_INFO:
2193                 ret = qib_ctxt_info(fp, (struct qib_ctxt_info __user *)
2194                                     (unsigned long) cmd.cmd.ctxt_info);
2195                 break;
2196
2197         case QIB_CMD_TID_UPDATE:
2198                 ret = qib_tid_update(rcd, fp, &cmd.cmd.tid_info);
2199                 break;
2200
2201         case QIB_CMD_TID_FREE:
2202                 ret = qib_tid_free(rcd, subctxt_fp(fp), &cmd.cmd.tid_info);
2203                 break;
2204
2205         case QIB_CMD_SET_PART_KEY:
2206                 ret = qib_set_part_key(rcd, cmd.cmd.part_key);
2207                 break;
2208
2209         case QIB_CMD_DISARM_BUFS:
2210                 (void)qib_disarm_piobufs_ifneeded(rcd);
2211                 ret = disarm_req_delay(rcd);
2212                 break;
2213
2214         case QIB_CMD_PIOAVAILUPD:
2215                 qib_force_pio_avail_update(rcd->dd);
2216                 break;
2217
2218         case QIB_CMD_POLL_TYPE:
2219                 rcd->poll_type = cmd.cmd.poll_type;
2220                 break;
2221
2222         case QIB_CMD_ARMLAUNCH_CTRL:
2223                 rcd->dd->f_set_armlaunch(rcd->dd, cmd.cmd.armlaunch_ctrl);
2224                 break;
2225
2226         case QIB_CMD_SDMA_INFLIGHT:
2227                 ret = qib_sdma_get_inflight(user_sdma_queue_fp(fp),
2228                                             (u32 __user *) (unsigned long)
2229                                             cmd.cmd.sdma_inflight);
2230                 break;
2231
2232         case QIB_CMD_SDMA_COMPLETE:
2233                 ret = qib_sdma_get_complete(rcd->ppd,
2234                                             user_sdma_queue_fp(fp),
2235                                             (u32 __user *) (unsigned long)
2236                                             cmd.cmd.sdma_complete);
2237                 break;
2238
2239         case QIB_CMD_ACK_EVENT:
2240                 ret = qib_user_event_ack(rcd, subctxt_fp(fp),
2241                                          cmd.cmd.event_mask);
2242                 break;
2243         }
2244
2245         if (ret >= 0)
2246                 ret = consumed;
2247
2248 bail:
2249         return ret;
2250 }
2251
2252 static ssize_t qib_aio_write(struct kiocb *iocb, const struct iovec *iov,
2253                              unsigned long dim, loff_t off)
2254 {
2255         struct qib_filedata *fp = iocb->ki_filp->private_data;
2256         struct qib_ctxtdata *rcd = ctxt_fp(iocb->ki_filp);
2257         struct qib_user_sdma_queue *pq = fp->pq;
2258
2259         if (!dim || !pq)
2260                 return -EINVAL;
2261
2262         return qib_user_sdma_writev(rcd, pq, iov, dim);
2263 }
2264
2265 static struct class *qib_class;
2266 static dev_t qib_dev;
2267
2268 int qib_cdev_init(int minor, const char *name,
2269                   const struct file_operations *fops,
2270                   struct cdev **cdevp, struct device **devp)
2271 {
2272         const dev_t dev = MKDEV(MAJOR(qib_dev), minor);
2273         struct cdev *cdev;
2274         struct device *device = NULL;
2275         int ret;
2276
2277         cdev = cdev_alloc();
2278         if (!cdev) {
2279                 pr_err("Could not allocate cdev for minor %d, %s\n",
2280                        minor, name);
2281                 ret = -ENOMEM;
2282                 goto done;
2283         }
2284
2285         cdev->owner = THIS_MODULE;
2286         cdev->ops = fops;
2287         kobject_set_name(&cdev->kobj, name);
2288
2289         ret = cdev_add(cdev, dev, 1);
2290         if (ret < 0) {
2291                 pr_err("Could not add cdev for minor %d, %s (err %d)\n",
2292                        minor, name, -ret);
2293                 goto err_cdev;
2294         }
2295
2296         device = device_create(qib_class, NULL, dev, NULL, "%s", name);
2297         if (!IS_ERR(device))
2298                 goto done;
2299         ret = PTR_ERR(device);
2300         device = NULL;
2301         pr_err("Could not create device for minor %d, %s (err %d)\n",
2302                minor, name, -ret);
2303 err_cdev:
2304         cdev_del(cdev);
2305         cdev = NULL;
2306 done:
2307         *cdevp = cdev;
2308         *devp = device;
2309         return ret;
2310 }
2311
2312 void qib_cdev_cleanup(struct cdev **cdevp, struct device **devp)
2313 {
2314         struct device *device = *devp;
2315
2316         if (device) {
2317                 device_unregister(device);
2318                 *devp = NULL;
2319         }
2320
2321         if (*cdevp) {
2322                 cdev_del(*cdevp);
2323                 *cdevp = NULL;
2324         }
2325 }
2326
2327 static struct cdev *wildcard_cdev;
2328 static struct device *wildcard_device;
2329
2330 int __init qib_dev_init(void)
2331 {
2332         int ret;
2333
2334         ret = alloc_chrdev_region(&qib_dev, 0, QIB_NMINORS, QIB_DRV_NAME);
2335         if (ret < 0) {
2336                 pr_err("Could not allocate chrdev region (err %d)\n", -ret);
2337                 goto done;
2338         }
2339
2340         qib_class = class_create(THIS_MODULE, "ipath");
2341         if (IS_ERR(qib_class)) {
2342                 ret = PTR_ERR(qib_class);
2343                 pr_err("Could not create device class (err %d)\n", -ret);
2344                 unregister_chrdev_region(qib_dev, QIB_NMINORS);
2345         }
2346
2347 done:
2348         return ret;
2349 }
2350
2351 void qib_dev_cleanup(void)
2352 {
2353         if (qib_class) {
2354                 class_destroy(qib_class);
2355                 qib_class = NULL;
2356         }
2357
2358         unregister_chrdev_region(qib_dev, QIB_NMINORS);
2359 }
2360
2361 static atomic_t user_count = ATOMIC_INIT(0);
2362
2363 static void qib_user_remove(struct qib_devdata *dd)
2364 {
2365         if (atomic_dec_return(&user_count) == 0)
2366                 qib_cdev_cleanup(&wildcard_cdev, &wildcard_device);
2367
2368         qib_cdev_cleanup(&dd->user_cdev, &dd->user_device);
2369 }
2370
2371 static int qib_user_add(struct qib_devdata *dd)
2372 {
2373         char name[10];
2374         int ret;
2375
2376         if (atomic_inc_return(&user_count) == 1) {
2377                 ret = qib_cdev_init(0, "ipath", &qib_file_ops,
2378                                     &wildcard_cdev, &wildcard_device);
2379                 if (ret)
2380                         goto done;
2381         }
2382
2383         snprintf(name, sizeof(name), "ipath%d", dd->unit);
2384         ret = qib_cdev_init(dd->unit + 1, name, &qib_file_ops,
2385                             &dd->user_cdev, &dd->user_device);
2386         if (ret)
2387                 qib_user_remove(dd);
2388 done:
2389         return ret;
2390 }
2391
2392 /*
2393  * Create per-unit files in /dev
2394  */
2395 int qib_device_create(struct qib_devdata *dd)
2396 {
2397         int r, ret;
2398
2399         r = qib_user_add(dd);
2400         ret = qib_diag_add(dd);
2401         if (r && !ret)
2402                 ret = r;
2403         return ret;
2404 }
2405
2406 /*
2407  * Remove per-unit files in /dev
2408  * void, core kernel returns no errors for this stuff
2409  */
2410 void qib_device_remove(struct qib_devdata *dd)
2411 {
2412         qib_user_remove(dd);
2413         qib_diag_remove(dd);
2414 }