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 <rdma/ib.h>
  49
  50 #include "qib.h"
  51 #include "qib_common.h"
  52 #include "qib_user_sdma.h"
  53
  54 #undef pr_fmt
  55 #define pr_fmt(fmt) QIB_DRV_NAME ": " fmt
  56
  57 static int qib_open(struct inode *, struct file *);
  58 static int qib_close(struct inode *, struct file *);
  59 static ssize_t qib_write(struct file *, const char __user *, size_t, loff_t *);
  60 static ssize_t qib_aio_write(struct kiocb *, const struct iovec *,
  61                              unsigned long, loff_t);
  62 static unsigned int qib_poll(struct file *, struct poll_table_struct *);
  63 static int qib_mmapf(struct file *, struct vm_area_struct *);
  64
  65 static const struct file_operations qib_file_ops = {
  66         .owner = THIS_MODULE,
  67         .write = qib_write,
  68         .aio_write = qib_aio_write,
  69         .open = qib_open,
  70         .release = qib_close,
  71         .poll = qib_poll,
  72         .mmap = qib_mmapf,
  73         .llseek = noop_llseek,
  74 };
  75
  76 /*
  77  * Convert kernel virtual addresses to physical addresses so they don't
  78  * potentially conflict with the chip addresses used as mmap offsets.
  79  * It doesn't really matter what mmap offset we use as long as we can
  80  * interpret it correctly.
  81  */
  82 static u64 cvt_kvaddr(void *p)
  83 {
  84         struct page *page;
  85         u64 paddr = 0;
  86
  87         page = vmalloc_to_page(p);
  88         if (page)
  89                 paddr = page_to_pfn(page) << PAGE_SHIFT;
  90
  91         return paddr;
  92 }
  93
  94 static int qib_get_base_info(struct file *fp, void __user *ubase,
  95                              size_t ubase_size)
  96 {
  97         struct qib_ctxtdata *rcd = ctxt_fp(fp);
  98         int ret = 0;
  99         struct qib_base_info *kinfo = NULL;
 100         struct qib_devdata *dd = rcd->dd;
 101         struct qib_pportdata *ppd = rcd->ppd;
 102         unsigned subctxt_cnt;
 103         int shared, master;
 104         size_t sz;
 105
 106         subctxt_cnt = rcd->subctxt_cnt;
 107         if (!subctxt_cnt) {
 108                 shared = 0;
 109                 master = 0;
 110                 subctxt_cnt = 1;
 111         } else {
 112                 shared = 1;
 113                 master = !subctxt_fp(fp);
 114         }
 115
 116         sz = sizeof(*kinfo);
 117         /* If context sharing is not requested, allow the old size structure */
 118         if (!shared)
 119                 sz -= 7 * sizeof(u64);
 120         if (ubase_size < sz) {
 121                 ret = -EINVAL;
 122                 goto bail;
 123         }
 124
 125         kinfo = kzalloc(sizeof(*kinfo), GFP_KERNEL);
 126         if (kinfo == NULL) {
 127                 ret = -ENOMEM;
 128                 goto bail;
 129         }
 130
 131         ret = dd->f_get_base_info(rcd, kinfo);
 132         if (ret < 0)
 133                 goto bail;
 134
 135         kinfo->spi_rcvhdr_cnt = dd->rcvhdrcnt;
 136         kinfo->spi_rcvhdrent_size = dd->rcvhdrentsize;
 137         kinfo->spi_tidegrcnt = rcd->rcvegrcnt;
 138         kinfo->spi_rcv_egrbufsize = dd->rcvegrbufsize;
 139         /*
 140          * have to mmap whole thing
 141          */
 142         kinfo->spi_rcv_egrbuftotlen =
 143                 rcd->rcvegrbuf_chunks * rcd->rcvegrbuf_size;
 144         kinfo->spi_rcv_egrperchunk = rcd->rcvegrbufs_perchunk;
 145         kinfo->spi_rcv_egrchunksize = kinfo->spi_rcv_egrbuftotlen /
 146                 rcd->rcvegrbuf_chunks;
 147         kinfo->spi_tidcnt = dd->rcvtidcnt / subctxt_cnt;
 148         if (master)
 149                 kinfo->spi_tidcnt += dd->rcvtidcnt % subctxt_cnt;
 150         /*
 151          * for this use, may be cfgctxts summed over all chips that
 152          * are are configured and present
 153          */
 154         kinfo->spi_nctxts = dd->cfgctxts;
 155         /* unit (chip/board) our context is on */
 156         kinfo->spi_unit = dd->unit;
 157         kinfo->spi_port = ppd->port;
 158         /* for now, only a single page */
 159         kinfo->spi_tid_maxsize = PAGE_SIZE;
 160
 161         /*
 162          * Doing this per context, and based on the skip value, etc.  This has
 163          * to be the actual buffer size, since the protocol code treats it
 164          * as an array.
 165          *
 166          * These have to be set to user addresses in the user code via mmap.
 167          * These values are used on return to user code for the mmap target
 168          * addresses only.  For 32 bit, same 44 bit address problem, so use
 169          * the physical address, not virtual.  Before 2.6.11, using the
 170          * page_address() macro worked, but in 2.6.11, even that returns the
 171          * full 64 bit address (upper bits all 1's).  So far, using the
 172          * physical addresses (or chip offsets, for chip mapping) works, but
 173          * no doubt some future kernel release will change that, and we'll be
 174          * on to yet another method of dealing with this.
 175          * Normally only one of rcvhdr_tailaddr or rhf_offset is useful
 176          * since the chips with non-zero rhf_offset don't normally
 177          * enable tail register updates to host memory, but for testing,
 178          * both can be enabled and used.
 179          */
 180         kinfo->spi_rcvhdr_base = (u64) rcd->rcvhdrq_phys;
 181         kinfo->spi_rcvhdr_tailaddr = (u64) rcd->rcvhdrqtailaddr_phys;
 182         kinfo->spi_rhf_offset = dd->rhf_offset;
 183         kinfo->spi_rcv_egrbufs = (u64) rcd->rcvegr_phys;
 184         kinfo->spi_pioavailaddr = (u64) dd->pioavailregs_phys;
 185         /* setup per-unit (not port) status area for user programs */
 186         kinfo->spi_status = (u64) kinfo->spi_pioavailaddr +
 187                 (char *) ppd->statusp -
 188                 (char *) dd->pioavailregs_dma;
 189         kinfo->spi_uregbase = (u64) dd->uregbase + dd->ureg_align * rcd->ctxt;
 190         if (!shared) {
 191                 kinfo->spi_piocnt = rcd->piocnt;
 192                 kinfo->spi_piobufbase = (u64) rcd->piobufs;
 193                 kinfo->spi_sendbuf_status = cvt_kvaddr(rcd->user_event_mask);
 194         } else if (master) {
 195                 kinfo->spi_piocnt = (rcd->piocnt / subctxt_cnt) +
 196                                     (rcd->piocnt % subctxt_cnt);
 197                 /* Master's PIO buffers are after all the slave's */
 198                 kinfo->spi_piobufbase = (u64) rcd->piobufs +
 199                         dd->palign *
 200                         (rcd->piocnt - kinfo->spi_piocnt);
 201         } else {
 202                 unsigned slave = subctxt_fp(fp) - 1;
 203
 204                 kinfo->spi_piocnt = rcd->piocnt / subctxt_cnt;
 205                 kinfo->spi_piobufbase = (u64) rcd->piobufs +
 206                         dd->palign * kinfo->spi_piocnt * slave;
 207         }
 208
 209         if (shared) {
 210                 kinfo->spi_sendbuf_status =
 211                         cvt_kvaddr(&rcd->user_event_mask[subctxt_fp(fp)]);
 212                 /* only spi_subctxt_* fields should be set in this block! */
 213                 kinfo->spi_subctxt_uregbase = cvt_kvaddr(rcd->subctxt_uregbase);
 214
 215                 kinfo->spi_subctxt_rcvegrbuf =
 216                         cvt_kvaddr(rcd->subctxt_rcvegrbuf);
 217                 kinfo->spi_subctxt_rcvhdr_base =
 218                         cvt_kvaddr(rcd->subctxt_rcvhdr_base);
 219         }
 220
 221         /*
 222          * All user buffers are 2KB buffers.  If we ever support
 223          * giving 4KB buffers to user processes, this will need some
 224          * work.  Can't use piobufbase directly, because it has
 225          * both 2K and 4K buffer base values.
 226          */
 227         kinfo->spi_pioindex = (kinfo->spi_piobufbase - dd->pio2k_bufbase) /
 228                 dd->palign;
 229         kinfo->spi_pioalign = dd->palign;
 230         kinfo->spi_qpair = QIB_KD_QP;
 231         /*
 232          * user mode PIO buffers are always 2KB, even when 4KB can
 233          * be received, and sent via the kernel; this is ibmaxlen
 234          * for 2K MTU.
 235          */
 236         kinfo->spi_piosize = dd->piosize2k - 2 * sizeof(u32);
 237         kinfo->spi_mtu = ppd->ibmaxlen; /* maxlen, not ibmtu */
 238         kinfo->spi_ctxt = rcd->ctxt;
 239         kinfo->spi_subctxt = subctxt_fp(fp);
 240         kinfo->spi_sw_version = QIB_KERN_SWVERSION;
 241         kinfo->spi_sw_version |= 1U << 31; /* QLogic-built, not kernel.org */
 242         kinfo->spi_hw_version = dd->revision;
 243
 244         if (master)
 245                 kinfo->spi_runtime_flags |= QIB_RUNTIME_MASTER;
 246
 247         sz = (ubase_size < sizeof(*kinfo)) ? ubase_size : sizeof(*kinfo);
 248         if (copy_to_user(ubase, kinfo, sz))
 249                 ret = -EFAULT;
 250 bail:
 251         kfree(kinfo);
 252         return ret;
 253 }
 254
 255 /**
 256  * qib_tid_update - update a context TID
 257  * @rcd: the context
 258  * @fp: the qib device file
 259  * @ti: the TID information
 260  *
 261  * The new implementation as of Oct 2004 is that the driver assigns
 262  * the tid and returns it to the caller.   To reduce search time, we
 263  * keep a cursor for each context, walking the shadow tid array to find
 264  * one that's not in use.
 265  *
 266  * For now, if we can't allocate the full list, we fail, although
 267  * in the long run, we'll allocate as many as we can, and the
 268  * caller will deal with that by trying the remaining pages later.
 269  * That means that when we fail, we have to mark the tids as not in
 270  * use again, in our shadow copy.
 271  *
 272  * It's up to the caller to free the tids when they are done.
 273  * We'll unlock the pages as they free them.
 274  *
 275  * Also, right now we are locking one page at a time, but since
 276  * the intended use of this routine is for a single group of
 277  * virtually contiguous pages, that should change to improve
 278  * performance.
 279  */
 280 static int qib_tid_update(struct qib_ctxtdata *rcd, struct file *fp,
 281                           const struct qib_tid_info *ti)
 282 {
 283         int ret = 0, ntids;
 284         u32 tid, ctxttid, cnt, i, tidcnt, tidoff;
 285         u16 *tidlist;
 286         struct qib_devdata *dd = rcd->dd;
 287         u64 physaddr;
 288         unsigned long vaddr;
 289         u64 __iomem *tidbase;
 290         unsigned long tidmap[8];
 291         struct page **pagep = NULL;
 292         unsigned subctxt = subctxt_fp(fp);
 293
 294         if (!dd->pageshadow) {
 295                 ret = -ENOMEM;
 296                 goto done;
 297         }
 298
 299         cnt = ti->tidcnt;
 300         if (!cnt) {
 301                 ret = -EFAULT;
 302                 goto done;
 303         }
 304         ctxttid = rcd->ctxt * dd->rcvtidcnt;
 305         if (!rcd->subctxt_cnt) {
 306                 tidcnt = dd->rcvtidcnt;
 307                 tid = rcd->tidcursor;
 308                 tidoff = 0;
 309         } else if (!subctxt) {
 310                 tidcnt = (dd->rcvtidcnt / rcd->subctxt_cnt) +
 311                          (dd->rcvtidcnt % rcd->subctxt_cnt);
 312                 tidoff = dd->rcvtidcnt - tidcnt;
 313                 ctxttid += tidoff;
 314                 tid = tidcursor_fp(fp);
 315         } else {
 316                 tidcnt = dd->rcvtidcnt / rcd->subctxt_cnt;
 317                 tidoff = tidcnt * (subctxt - 1);
 318                 ctxttid += tidoff;
 319                 tid = tidcursor_fp(fp);
 320         }
 321         if (cnt > tidcnt) {
 322                 /* make sure it all fits in tid_pg_list */
 323                 qib_devinfo(dd->pcidev,
 324                         "Process tried to allocate %u TIDs, only trying max (%u)\n",
 325                         cnt, tidcnt);
 326                 cnt = tidcnt;
 327         }
 328         pagep = (struct page **) rcd->tid_pg_list;
 329         tidlist = (u16 *) &pagep[dd->rcvtidcnt];
 330         pagep += tidoff;
 331         tidlist += tidoff;
 332
 333         memset(tidmap, 0, sizeof(tidmap));
 334         /* before decrement; chip actual # */
 335         ntids = tidcnt;
 336         tidbase = (u64 __iomem *) (((char __iomem *) dd->kregbase) +
 337                                    dd->rcvtidbase +
 338                                    ctxttid * sizeof(*tidbase));
 339
 340         /* virtual address of first page in transfer */
 341         vaddr = ti->tidvaddr;
 342         if (!access_ok(VERIFY_WRITE, (void __user *) vaddr,
 343                        cnt * PAGE_SIZE)) {
 344                 ret = -EFAULT;
 345                 goto done;
 346         }
 347         ret = qib_get_user_pages(vaddr, cnt, pagep);
 348         if (ret) {
 349                 /*
 350                  * if (ret == -EBUSY)
 351                  * We can't continue because the pagep array won't be
 352                  * initialized. This should never happen,
 353                  * unless perhaps the user has mpin'ed the pages
 354                  * themselves.
 355                  */
 356                 qib_devinfo(dd->pcidev,
 357                          "Failed to lock addr %p, %u pages: "
 358                          "errno %d\n", (void *) vaddr, cnt, -ret);
 359                 goto done;
 360         }
 361         for (i = 0; i < cnt; i++, vaddr += PAGE_SIZE) {
 362                 for (; ntids--; tid++) {
 363                         if (tid == tidcnt)
 364                                 tid = 0;
 365                         if (!dd->pageshadow[ctxttid + tid])
 366                                 break;
 367                 }
 368                 if (ntids < 0) {
 369                         /*
 370                          * Oops, wrapped all the way through their TIDs,
 371                          * and didn't have enough free; see comments at
 372                          * start of routine
 373                          */
 374                         i--;    /* last tidlist[i] not filled in */
 375                         ret = -ENOMEM;
 376                         break;
 377                 }
 378                 tidlist[i] = tid + tidoff;
 379                 /* we "know" system pages and TID pages are same size */
 380                 dd->pageshadow[ctxttid + tid] = pagep[i];
 381                 dd->physshadow[ctxttid + tid] =
 382                         qib_map_page(dd->pcidev, pagep[i], 0, PAGE_SIZE,
 383                                      PCI_DMA_FROMDEVICE);
 384                 /*
 385                  * don't need atomic or it's overhead
 386                  */
 387                 __set_bit(tid, tidmap);
 388                 physaddr = dd->physshadow[ctxttid + tid];
 389                 /* PERFORMANCE: below should almost certainly be cached */
 390                 dd->f_put_tid(dd, &tidbase[tid],
 391                                   RCVHQ_RCV_TYPE_EXPECTED, physaddr);
 392                 /*
 393                  * don't check this tid in qib_ctxtshadow, since we
 394                  * just filled it in; start with the next one.
 395                  */
 396                 tid++;
 397         }
 398
 399         if (ret) {
 400                 u32 limit;
 401 cleanup:
 402                 /* jump here if copy out of updated info failed... */
 403                 /* same code that's in qib_free_tid() */
 404                 limit = sizeof(tidmap) * BITS_PER_BYTE;
 405                 if (limit > tidcnt)
 406                         /* just in case size changes in future */
 407                         limit = tidcnt;
 408                 tid = find_first_bit((const unsigned long *)tidmap, limit);
 409                 for (; tid < limit; tid++) {
 410                         if (!test_bit(tid, tidmap))
 411                                 continue;
 412                         if (dd->pageshadow[ctxttid + tid]) {
 413                                 dma_addr_t phys;
 414
 415                                 phys = dd->physshadow[ctxttid + tid];
 416                                 dd->physshadow[ctxttid + tid] = dd->tidinvalid;
 417                                 /* PERFORMANCE: below should almost certainly
 418                                  * be cached
 419                                  */
 420                                 dd->f_put_tid(dd, &tidbase[tid],
 421                                               RCVHQ_RCV_TYPE_EXPECTED,
 422                                               dd->tidinvalid);
 423                                 pci_unmap_page(dd->pcidev, phys, PAGE_SIZE,
 424                                                PCI_DMA_FROMDEVICE);
 425                                 dd->pageshadow[ctxttid + tid] = NULL;
 426                         }
 427                 }
 428                 qib_release_user_pages(pagep, cnt);
 429         } else {
 430                 /*
 431                  * Copy the updated array, with qib_tid's filled in, back
 432                  * to user.  Since we did the copy in already, this "should
 433                  * never fail" If it does, we have to clean up...
 434                  */
 435                 if (copy_to_user((void __user *)
 436                                  (unsigned long) ti->tidlist,
 437                                  tidlist, cnt * sizeof(*tidlist))) {
 438                         ret = -EFAULT;
 439                         goto cleanup;
 440                 }
 441                 if (copy_to_user((void __user *) (unsigned long) ti->tidmap,
 442                                  tidmap, sizeof tidmap)) {
 443                         ret = -EFAULT;
 444                         goto cleanup;
 445                 }
 446                 if (tid == tidcnt)
 447                         tid = 0;
 448                 if (!rcd->subctxt_cnt)
 449                         rcd->tidcursor = tid;
 450                 else
 451                         tidcursor_fp(fp) = tid;
 452         }
 453
 454 done:
 455         return ret;
 456 }
 457
 458 /**
 459  * qib_tid_free - free a context TID
 460  * @rcd: the context
 461  * @subctxt: the subcontext
 462  * @ti: the TID info
 463  *
 464  * right now we are unlocking one page at a time, but since
 465  * the intended use of this routine is for a single group of
 466  * virtually contiguous pages, that should change to improve
 467  * performance.  We check that the TID is in range for this context
 468  * but otherwise don't check validity; if user has an error and
 469  * frees the wrong tid, it's only their own data that can thereby
 470  * be corrupted.  We do check that the TID was in use, for sanity
 471  * We always use our idea of the saved address, not the address that
 472  * they pass in to us.
 473  */
 474 static int qib_tid_free(struct qib_ctxtdata *rcd, unsigned subctxt,
 475                         const struct qib_tid_info *ti)
 476 {
 477         int ret = 0;
 478         u32 tid, ctxttid, cnt, limit, tidcnt;
 479         struct qib_devdata *dd = rcd->dd;
 480         u64 __iomem *tidbase;
 481         unsigned long tidmap[8];
 482
 483         if (!dd->pageshadow) {
 484                 ret = -ENOMEM;
 485                 goto done;
 486         }
 487
 488         if (copy_from_user(tidmap, (void __user *)(unsigned long)ti->tidmap,
 489                            sizeof tidmap)) {
 490                 ret = -EFAULT;
 491                 goto done;
 492         }
 493
 494         ctxttid = rcd->ctxt * dd->rcvtidcnt;
 495         if (!rcd->subctxt_cnt)
 496                 tidcnt = dd->rcvtidcnt;
 497         else if (!subctxt) {
 498                 tidcnt = (dd->rcvtidcnt / rcd->subctxt_cnt) +
 499                          (dd->rcvtidcnt % rcd->subctxt_cnt);
 500                 ctxttid += dd->rcvtidcnt - tidcnt;
 501         } else {
 502                 tidcnt = dd->rcvtidcnt / rcd->subctxt_cnt;
 503                 ctxttid += tidcnt * (subctxt - 1);
 504         }
 505         tidbase = (u64 __iomem *) ((char __iomem *)(dd->kregbase) +
 506                                    dd->rcvtidbase +
 507                                    ctxttid * sizeof(*tidbase));
 508
 509         limit = sizeof(tidmap) * BITS_PER_BYTE;
 510         if (limit > tidcnt)
 511                 /* just in case size changes in future */
 512                 limit = tidcnt;
 513         tid = find_first_bit(tidmap, limit);
 514         for (cnt = 0; tid < limit; tid++) {
 515                 /*
 516                  * small optimization; if we detect a run of 3 or so without
 517                  * any set, use find_first_bit again.  That's mainly to
 518                  * accelerate the case where we wrapped, so we have some at
 519                  * the beginning, and some at the end, and a big gap
 520                  * in the middle.
 521                  */
 522                 if (!test_bit(tid, tidmap))
 523                         continue;
 524                 cnt++;
 525                 if (dd->pageshadow[ctxttid + tid]) {
 526                         struct page *p;
 527                         dma_addr_t phys;
 528
 529                         p = dd->pageshadow[ctxttid + tid];
 530                         dd->pageshadow[ctxttid + tid] = NULL;
 531                         phys = dd->physshadow[ctxttid + tid];
 532                         dd->physshadow[ctxttid + tid] = dd->tidinvalid;
 533                         /* PERFORMANCE: below should almost certainly be
 534                          * cached
 535                          */
 536                         dd->f_put_tid(dd, &tidbase[tid],
 537                                       RCVHQ_RCV_TYPE_EXPECTED, dd->tidinvalid);
 538                         pci_unmap_page(dd->pcidev, phys, PAGE_SIZE,
 539                                        PCI_DMA_FROMDEVICE);
 540                         qib_release_user_pages(&p, 1);
 541                 }
 542         }
 543 done:
 544         return ret;
 545 }
 546
 547 /**
 548  * qib_set_part_key - set a partition key
 549  * @rcd: the context
 550  * @key: the key
 551  *
 552  * We can have up to 4 active at a time (other than the default, which is
 553  * always allowed).  This is somewhat tricky, since multiple contexts may set
 554  * the same key, so we reference count them, and clean up at exit.  All 4
 555  * partition keys are packed into a single qlogic_ib register.  It's an
 556  * error for a process to set the same pkey multiple times.  We provide no
 557  * mechanism to de-allocate a pkey at this time, we may eventually need to
 558  * do that.  I've used the atomic operations, and no locking, and only make
 559  * a single pass through what's available.  This should be more than
 560  * adequate for some time. I'll think about spinlocks or the like if and as
 561  * it's necessary.
 562  */
 563 static int qib_set_part_key(struct qib_ctxtdata *rcd, u16 key)
 564 {
 565         struct qib_pportdata *ppd = rcd->ppd;
 566         int i, any = 0, pidx = -1;
 567         u16 lkey = key & 0x7FFF;
 568         int ret;
 569
 570         if (lkey == (QIB_DEFAULT_P_KEY & 0x7FFF)) {
 571                 /* nothing to do; this key always valid */
 572                 ret = 0;
 573                 goto bail;
 574         }
 575
 576         if (!lkey) {
 577                 ret = -EINVAL;
 578                 goto bail;
 579         }
 580
 581         /*
 582          * Set the full membership bit, because it has to be
 583          * set in the register or the packet, and it seems
 584          * cleaner to set in the register than to force all
 585          * callers to set it.
 586          */
 587         key |= 0x8000;
 588
 589         for (i = 0; i < ARRAY_SIZE(rcd->pkeys); i++) {
 590                 if (!rcd->pkeys[i] && pidx == -1)
 591                         pidx = i;
 592                 if (rcd->pkeys[i] == key) {
 593                         ret = -EEXIST;
 594                         goto bail;
 595                 }
 596         }
 597         if (pidx == -1) {
 598                 ret = -EBUSY;
 599                 goto bail;
 600         }
 601         for (any = i = 0; i < ARRAY_SIZE(ppd->pkeys); i++) {
 602                 if (!ppd->pkeys[i]) {
 603                         any++;
 604                         continue;
 605                 }
 606                 if (ppd->pkeys[i] == key) {
 607                         atomic_t *pkrefs = &ppd->pkeyrefs[i];
 608
 609                         if (atomic_inc_return(pkrefs) > 1) {
 610                                 rcd->pkeys[pidx] = key;
 611                                 ret = 0;
 612                                 goto bail;
 613                         } else {
 614                                 /*
 615                                  * lost race, decrement count, catch below
 616                                  */
 617                                 atomic_dec(pkrefs);
 618                                 any++;
 619                         }
 620                 }
 621                 if ((ppd->pkeys[i] & 0x7FFF) == lkey) {
 622                         /*
 623                          * It makes no sense to have both the limited and
 624                          * full membership PKEY set at the same time since
 625                          * the unlimited one will disable the limited one.
 626                          */
 627                         ret = -EEXIST;
 628                         goto bail;
 629                 }
 630         }
 631         if (!any) {
 632                 ret = -EBUSY;
 633                 goto bail;
 634         }
 635         for (any = i = 0; i < ARRAY_SIZE(ppd->pkeys); i++) {
 636                 if (!ppd->pkeys[i] &&
 637                     atomic_inc_return(&ppd->pkeyrefs[i]) == 1) {
 638                         rcd->pkeys[pidx] = key;
 639                         ppd->pkeys[i] = key;
 640                         (void) ppd->dd->f_set_ib_cfg(ppd, QIB_IB_CFG_PKEYS, 0);
 641                         ret = 0;
 642                         goto bail;
 643                 }
 644         }
 645         ret = -EBUSY;
 646
 647 bail:
 648         return ret;
 649 }
 650
 651 /**
 652  * qib_manage_rcvq - manage a context's receive queue
 653  * @rcd: the context
 654  * @subctxt: the subcontext
 655  * @start_stop: action to carry out
 656  *
 657  * start_stop == 0 disables receive on the context, for use in queue
 658  * overflow conditions.  start_stop==1 re-enables, to be used to
 659  * re-init the software copy of the head register
 660  */
 661 static int qib_manage_rcvq(struct qib_ctxtdata *rcd, unsigned subctxt,
 662                            int start_stop)
 663 {
 664         struct qib_devdata *dd = rcd->dd;
 665         unsigned int rcvctrl_op;
 666
 667         if (subctxt)
 668                 goto bail;
 669         /* atomically clear receive enable ctxt. */
 670         if (start_stop) {
 671                 /*
 672                  * On enable, force in-memory copy of the tail register to
 673                  * 0, so that protocol code doesn't have to worry about
 674                  * whether or not the chip has yet updated the in-memory
 675                  * copy or not on return from the system call. The chip
 676                  * always resets it's tail register back to 0 on a
 677                  * transition from disabled to enabled.
 678                  */
 679                 if (rcd->rcvhdrtail_kvaddr)
 680                         qib_clear_rcvhdrtail(rcd);
 681                 rcvctrl_op = QIB_RCVCTRL_CTXT_ENB;
 682         } else
 683                 rcvctrl_op = QIB_RCVCTRL_CTXT_DIS;
 684         dd->f_rcvctrl(rcd->ppd, rcvctrl_op, rcd->ctxt);
 685         /* always; new head should be equal to new tail; see above */
 686 bail:
 687         return 0;
 688 }
 689
 690 static void qib_clean_part_key(struct qib_ctxtdata *rcd,
 691                                struct qib_devdata *dd)
 692 {
 693         int i, j, pchanged = 0;
 694         u64 oldpkey;
 695         struct qib_pportdata *ppd = rcd->ppd;
 696
 697         /* for debugging only */
 698         oldpkey = (u64) ppd->pkeys[0] |
 699                 ((u64) ppd->pkeys[1] << 16) |
 700                 ((u64) ppd->pkeys[2] << 32) |
 701                 ((u64) ppd->pkeys[3] << 48);
 702
 703         for (i = 0; i < ARRAY_SIZE(rcd->pkeys); i++) {
 704                 if (!rcd->pkeys[i])
 705                         continue;
 706                 for (j = 0; j < ARRAY_SIZE(ppd->pkeys); j++) {
 707                         /* check for match independent of the global bit */
 708                         if ((ppd->pkeys[j] & 0x7fff) !=
 709                             (rcd->pkeys[i] & 0x7fff))
 710                                 continue;
 711                         if (atomic_dec_and_test(&ppd->pkeyrefs[j])) {
 712                                 ppd->pkeys[j] = 0;
 713                                 pchanged++;
 714                         }
 715                         break;
 716                 }
 717                 rcd->pkeys[i] = 0;
 718         }
 719         if (pchanged)
 720                 (void) ppd->dd->f_set_ib_cfg(ppd, QIB_IB_CFG_PKEYS, 0);
 721 }
 722
 723 /* common code for the mappings on dma_alloc_coherent mem */
 724 static int qib_mmap_mem(struct vm_area_struct *vma, struct qib_ctxtdata *rcd,
 725                         unsigned len, void *kvaddr, u32 write_ok, char *what)
 726 {
 727         struct qib_devdata *dd = rcd->dd;
 728         unsigned long pfn;
 729         int ret;
 730
 731         if ((vma->vm_end - vma->vm_start) > len) {
 732                 qib_devinfo(dd->pcidev,
 733                          "FAIL on %s: len %lx > %x\n", what,
 734                          vma->vm_end - vma->vm_start, len);
 735                 ret = -EFAULT;
 736                 goto bail;
 737         }
 738
 739         /*
 740          * shared context user code requires rcvhdrq mapped r/w, others
 741          * only allowed readonly mapping.
 742          */
 743         if (!write_ok) {
 744                 if (vma->vm_flags & VM_WRITE) {
 745                         qib_devinfo(dd->pcidev,
 746                                  "%s must be mapped readonly\n", what);
 747                         ret = -EPERM;
 748                         goto bail;
 749                 }
 750
 751                 /* don't allow them to later change with mprotect */
 752                 vma->vm_flags &= ~VM_MAYWRITE;
 753         }
 754
 755         pfn = virt_to_phys(kvaddr) >> PAGE_SHIFT;
 756         ret = remap_pfn_range(vma, vma->vm_start, pfn,
 757                               len, vma->vm_page_prot);
 758         if (ret)
 759                 qib_devinfo(dd->pcidev,
 760                         "%s ctxt%u mmap of %lx, %x bytes failed: %d\n",
 761                         what, rcd->ctxt, pfn, len, ret);
 762 bail:
 763         return ret;
 764 }
 765
 766 static int mmap_ureg(struct vm_area_struct *vma, struct qib_devdata *dd,
 767                      u64 ureg)
 768 {
 769         unsigned long phys;
 770         unsigned long sz;
 771         int ret;
 772
 773         /*
 774          * This is real hardware, so use io_remap.  This is the mechanism
 775          * for the user process to update the head registers for their ctxt
 776          * in the chip.
 777          */
 778         sz = dd->flags & QIB_HAS_HDRSUPP ? 2 * PAGE_SIZE : PAGE_SIZE;
 779         if ((vma->vm_end - vma->vm_start) > sz) {
 780                 qib_devinfo(dd->pcidev,
 781                         "FAIL mmap userreg: reqlen %lx > PAGE\n",
 782                         vma->vm_end - vma->vm_start);
 783                 ret = -EFAULT;
 784         } else {
 785                 phys = dd->physaddr + ureg;
 786                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 787
 788                 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
 789                 ret = io_remap_pfn_range(vma, vma->vm_start,
 790                                          phys >> PAGE_SHIFT,
 791                                          vma->vm_end - vma->vm_start,
 792                                          vma->vm_page_prot);
 793         }
 794         return ret;
 795 }
 796
 797 static int mmap_piobufs(struct vm_area_struct *vma,
 798                         struct qib_devdata *dd,
 799                         struct qib_ctxtdata *rcd,
 800                         unsigned piobufs, unsigned piocnt)
 801 {
 802         unsigned long phys;
 803         int ret;
 804
 805         /*
 806          * When we map the PIO buffers in the chip, we want to map them as
 807          * writeonly, no read possible; unfortunately, x86 doesn't allow
 808          * for this in hardware, but we still prevent users from asking
 809          * for it.
 810          */
 811         if ((vma->vm_end - vma->vm_start) > (piocnt * dd->palign)) {
 812                 qib_devinfo(dd->pcidev,
 813                         "FAIL mmap piobufs: reqlen %lx > PAGE\n",
 814                          vma->vm_end - vma->vm_start);
 815                 ret = -EINVAL;
 816                 goto bail;
 817         }
 818
 819         phys = dd->physaddr + piobufs;
 820
 821 #if defined(__powerpc__)
 822         /* There isn't a generic way to specify writethrough mappings */
 823         pgprot_val(vma->vm_page_prot) |= _PAGE_NO_CACHE;
 824         pgprot_val(vma->vm_page_prot) |= _PAGE_WRITETHRU;
 825         pgprot_val(vma->vm_page_prot) &= ~_PAGE_GUARDED;
 826 #endif
 827
 828         /*
 829          * don't allow them to later change to readable with mprotect (for when
 830          * not initially mapped readable, as is normally the case)
 831          */
 832         vma->vm_flags &= ~VM_MAYREAD;
 833         vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
 834
 835         if (qib_wc_pat)
 836                 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
 837
 838         ret = io_remap_pfn_range(vma, vma->vm_start, phys >> PAGE_SHIFT,
 839                                  vma->vm_end - vma->vm_start,
 840                                  vma->vm_page_prot);
 841 bail:
 842         return ret;
 843 }
 844
 845 static int mmap_rcvegrbufs(struct vm_area_struct *vma,
 846                            struct qib_ctxtdata *rcd)
 847 {
 848         struct qib_devdata *dd = rcd->dd;
 849         unsigned long start, size;
 850         size_t total_size, i;
 851         unsigned long pfn;
 852         int ret;
 853
 854         size = rcd->rcvegrbuf_size;
 855         total_size = rcd->rcvegrbuf_chunks * size;
 856         if ((vma->vm_end - vma->vm_start) > total_size) {
 857                 qib_devinfo(dd->pcidev,
 858                         "FAIL on egr bufs: reqlen %lx > actual %lx\n",
 859                          vma->vm_end - vma->vm_start,
 860                          (unsigned long) total_size);
 861                 ret = -EINVAL;
 862                 goto bail;
 863         }
 864
 865         if (vma->vm_flags & VM_WRITE) {
 866                 qib_devinfo(dd->pcidev,
 867                         "Can't map eager buffers as writable (flags=%lx)\n",
 868                         vma->vm_flags);
 869                 ret = -EPERM;
 870                 goto bail;
 871         }
 872         /* don't allow them to later change to writeable with mprotect */
 873         vma->vm_flags &= ~VM_MAYWRITE;
 874
 875         start = vma->vm_start;
 876
 877         for (i = 0; i < rcd->rcvegrbuf_chunks; i++, start += size) {
 878                 pfn = virt_to_phys(rcd->rcvegrbuf[i]) >> PAGE_SHIFT;
 879                 ret = remap_pfn_range(vma, start, pfn, size,
 880                                       vma->vm_page_prot);
 881                 if (ret < 0)
 882                         goto bail;
 883         }
 884         ret = 0;
 885
 886 bail:
 887         return ret;
 888 }
 889
 890 /*
 891  * qib_file_vma_fault - handle a VMA page fault.
 892  */
 893 static int qib_file_vma_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 894 {
 895         struct page *page;
 896
 897         page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
 898         if (!page)
 899                 return VM_FAULT_SIGBUS;
 900
 901         get_page(page);
 902         vmf->page = page;
 903
 904         return 0;
 905 }
 906
 907 static struct vm_operations_struct qib_file_vm_ops = {
 908         .fault = qib_file_vma_fault,
 909 };
 910
 911 static int mmap_kvaddr(struct vm_area_struct *vma, u64 pgaddr,
 912                        struct qib_ctxtdata *rcd, unsigned subctxt)
 913 {
 914         struct qib_devdata *dd = rcd->dd;
 915         unsigned subctxt_cnt;
 916         unsigned long len;
 917         void *addr;
 918         size_t size;
 919         int ret = 0;
 920
 921         subctxt_cnt = rcd->subctxt_cnt;
 922         size = rcd->rcvegrbuf_chunks * rcd->rcvegrbuf_size;
 923
 924         /*
 925          * Each process has all the subctxt uregbase, rcvhdrq, and
 926          * rcvegrbufs mmapped - as an array for all the processes,
 927          * and also separately for this process.
 928          */
 929         if (pgaddr == cvt_kvaddr(rcd->subctxt_uregbase)) {
 930                 addr = rcd->subctxt_uregbase;
 931                 size = PAGE_SIZE * subctxt_cnt;
 932         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_rcvhdr_base)) {
 933                 addr = rcd->subctxt_rcvhdr_base;
 934                 size = rcd->rcvhdrq_size * subctxt_cnt;
 935         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_rcvegrbuf)) {
 936                 addr = rcd->subctxt_rcvegrbuf;
 937                 size *= subctxt_cnt;
 938         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_uregbase +
 939                                         PAGE_SIZE * subctxt)) {
 940                 addr = rcd->subctxt_uregbase + PAGE_SIZE * subctxt;
 941                 size = PAGE_SIZE;
 942         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_rcvhdr_base +
 943                                         rcd->rcvhdrq_size * subctxt)) {
 944                 addr = rcd->subctxt_rcvhdr_base +
 945                         rcd->rcvhdrq_size * subctxt;
 946                 size = rcd->rcvhdrq_size;
 947         } else if (pgaddr == cvt_kvaddr(&rcd->user_event_mask[subctxt])) {
 948                 addr = rcd->user_event_mask;
 949                 size = PAGE_SIZE;
 950         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_rcvegrbuf +
 951                                         size * subctxt)) {
 952                 addr = rcd->subctxt_rcvegrbuf + size * subctxt;
 953                 /* rcvegrbufs are read-only on the slave */
 954                 if (vma->vm_flags & VM_WRITE) {
 955                         qib_devinfo(dd->pcidev,
 956                                  "Can't map eager buffers as "
 957                                  "writable (flags=%lx)\n", vma->vm_flags);
 958                         ret = -EPERM;
 959                         goto bail;
 960                 }
 961                 /*
 962                  * Don't allow permission to later change to writeable
 963                  * with mprotect.
 964                  */
 965                 vma->vm_flags &= ~VM_MAYWRITE;
 966         } else
 967                 goto bail;
 968         len = vma->vm_end - vma->vm_start;
 969         if (len > size) {
 970                 ret = -EINVAL;
 971                 goto bail;
 972         }
 973
 974         vma->vm_pgoff = (unsigned long) addr >> PAGE_SHIFT;
 975         vma->vm_ops = &qib_file_vm_ops;
 976         vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
 977         ret = 1;
 978
 979 bail:
 980         return ret;
 981 }
 982
 983 /**
 984  * qib_mmapf - mmap various structures into user space
 985  * @fp: the file pointer
 986  * @vma: the VM area
 987  *
 988  * We use this to have a shared buffer between the kernel and the user code
 989  * for the rcvhdr queue, egr buffers, and the per-context user regs and pio
 990  * buffers in the chip.  We have the open and close entries so we can bump
 991  * the ref count and keep the driver from being unloaded while still mapped.
 992  */
 993 static int qib_mmapf(struct file *fp, struct vm_area_struct *vma)
 994 {
 995         struct qib_ctxtdata *rcd;
 996         struct qib_devdata *dd;
 997         u64 pgaddr, ureg;
 998         unsigned piobufs, piocnt;
 999         int ret, match = 1;
1000
1001         rcd = ctxt_fp(fp);
1002         if (!rcd || !(vma->vm_flags & VM_SHARED)) {
1003                 ret = -EINVAL;
1004                 goto bail;
1005         }
1006         dd = rcd->dd;
1007
1008         /*
1009          * This is the qib_do_user_init() code, mapping the shared buffers
1010          * and per-context user registers into the user process. The address
1011          * referred to by vm_pgoff is the file offset passed via mmap().
1012          * For shared contexts, this is the kernel vmalloc() address of the
1013          * pages to share with the master.
1014          * For non-shared or master ctxts, this is a physical address.
1015          * We only do one mmap for each space mapped.
1016          */
1017         pgaddr = vma->vm_pgoff << PAGE_SHIFT;
1018
1019         /*
1020          * Check for 0 in case one of the allocations failed, but user
1021          * called mmap anyway.
1022          */
1023         if (!pgaddr)  {
1024                 ret = -EINVAL;
1025                 goto bail;
1026         }
1027
1028         /*
1029          * Physical addresses must fit in 40 bits for our hardware.
1030          * Check for kernel virtual addresses first, anything else must
1031          * match a HW or memory address.
1032          */
1033         ret = mmap_kvaddr(vma, pgaddr, rcd, subctxt_fp(fp));
1034         if (ret) {
1035                 if (ret > 0)
1036                         ret = 0;
1037                 goto bail;
1038         }
1039
1040         ureg = dd->uregbase + dd->ureg_align * rcd->ctxt;
1041         if (!rcd->subctxt_cnt) {
1042                 /* ctxt is not shared */
1043                 piocnt = rcd->piocnt;
1044                 piobufs = rcd->piobufs;
1045         } else if (!subctxt_fp(fp)) {
1046                 /* caller is the master */
1047                 piocnt = (rcd->piocnt / rcd->subctxt_cnt) +
1048                          (rcd->piocnt % rcd->subctxt_cnt);
1049                 piobufs = rcd->piobufs +
1050                         dd->palign * (rcd->piocnt - piocnt);
1051         } else {
1052                 unsigned slave = subctxt_fp(fp) - 1;
1053
1054                 /* caller is a slave */
1055                 piocnt = rcd->piocnt / rcd->subctxt_cnt;
1056                 piobufs = rcd->piobufs + dd->palign * piocnt * slave;
1057         }
1058
1059         if (pgaddr == ureg)
1060                 ret = mmap_ureg(vma, dd, ureg);
1061         else if (pgaddr == piobufs)
1062                 ret = mmap_piobufs(vma, dd, rcd, piobufs, piocnt);
1063         else if (pgaddr == dd->pioavailregs_phys)
1064                 /* in-memory copy of pioavail registers */
1065                 ret = qib_mmap_mem(vma, rcd, PAGE_SIZE,
1066                                    (void *) dd->pioavailregs_dma, 0,
1067                                    "pioavail registers");
1068         else if (pgaddr == rcd->rcvegr_phys)
1069                 ret = mmap_rcvegrbufs(vma, rcd);
1070         else if (pgaddr == (u64) rcd->rcvhdrq_phys)
1071                 /*
1072                  * The rcvhdrq itself; multiple pages, contiguous
1073                  * from an i/o perspective.  Shared contexts need
1074                  * to map r/w, so we allow writing.
1075                  */
1076                 ret = qib_mmap_mem(vma, rcd, rcd->rcvhdrq_size,
1077                                    rcd->rcvhdrq, 1, "rcvhdrq");
1078         else if (pgaddr == (u64) rcd->rcvhdrqtailaddr_phys)
1079                 /* in-memory copy of rcvhdrq tail register */
1080                 ret = qib_mmap_mem(vma, rcd, PAGE_SIZE,
1081                                    rcd->rcvhdrtail_kvaddr, 0,
1082                                    "rcvhdrq tail");
1083         else
1084                 match = 0;
1085         if (!match)
1086                 ret = -EINVAL;
1087
1088         vma->vm_private_data = NULL;
1089
1090         if (ret < 0)
1091                 qib_devinfo(dd->pcidev,
1092                          "mmap Failure %d: off %llx len %lx\n",
1093                          -ret, (unsigned long long)pgaddr,
1094                          vma->vm_end - vma->vm_start);
1095 bail:
1096         return ret;
1097 }
1098
1099 static unsigned int qib_poll_urgent(struct qib_ctxtdata *rcd,
1100                                     struct file *fp,
1101                                     struct poll_table_struct *pt)
1102 {
1103         struct qib_devdata *dd = rcd->dd;
1104         unsigned pollflag;
1105
1106         poll_wait(fp, &rcd->wait, pt);
1107
1108         spin_lock_irq(&dd->uctxt_lock);
1109         if (rcd->urgent != rcd->urgent_poll) {
1110                 pollflag = POLLIN | POLLRDNORM;
1111                 rcd->urgent_poll = rcd->urgent;
1112         } else {
1113                 pollflag = 0;
1114                 set_bit(QIB_CTXT_WAITING_URG, &rcd->flag);
1115         }
1116         spin_unlock_irq(&dd->uctxt_lock);
1117
1118         return pollflag;
1119 }
1120
1121 static unsigned int qib_poll_next(struct qib_ctxtdata *rcd,
1122                                   struct file *fp,
1123                                   struct poll_table_struct *pt)
1124 {
1125         struct qib_devdata *dd = rcd->dd;
1126         unsigned pollflag;
1127
1128         poll_wait(fp, &rcd->wait, pt);
1129
1130         spin_lock_irq(&dd->uctxt_lock);
1131         if (dd->f_hdrqempty(rcd)) {
1132                 set_bit(QIB_CTXT_WAITING_RCV, &rcd->flag);
1133                 dd->f_rcvctrl(rcd->ppd, QIB_RCVCTRL_INTRAVAIL_ENB, rcd->ctxt);
1134                 pollflag = 0;
1135         } else
1136                 pollflag = POLLIN | POLLRDNORM;
1137         spin_unlock_irq(&dd->uctxt_lock);
1138
1139         return pollflag;
1140 }
1141
1142 static unsigned int qib_poll(struct file *fp, struct poll_table_struct *pt)
1143 {
1144         struct qib_ctxtdata *rcd;
1145         unsigned pollflag;
1146
1147         rcd = ctxt_fp(fp);
1148         if (!rcd)
1149                 pollflag = POLLERR;
1150         else if (rcd->poll_type == QIB_POLL_TYPE_URGENT)
1151                 pollflag = qib_poll_urgent(rcd, fp, pt);
1152         else  if (rcd->poll_type == QIB_POLL_TYPE_ANYRCV)
1153                 pollflag = qib_poll_next(rcd, fp, pt);
1154         else /* invalid */
1155                 pollflag = POLLERR;
1156
1157         return pollflag;
1158 }
1159
1160 static void assign_ctxt_affinity(struct file *fp, struct qib_devdata *dd)
1161 {
1162         struct qib_filedata *fd = fp->private_data;
1163         const unsigned int weight = cpumask_weight(&current->cpus_allowed);
1164         const struct cpumask *local_mask = cpumask_of_pcibus(dd->pcidev->bus);
1165         int local_cpu;
1166
1167         /*
1168          * If process has NOT already set it's affinity, select and
1169          * reserve a processor for it on the local NUMA node.
1170          */
1171         if ((weight >= qib_cpulist_count) &&
1172                 (cpumask_weight(local_mask) <= qib_cpulist_count)) {
1173                 for_each_cpu(local_cpu, local_mask)
1174                         if (!test_and_set_bit(local_cpu, qib_cpulist)) {
1175                                 fd->rec_cpu_num = local_cpu;
1176                                 return;
1177                         }
1178         }
1179
1180         /*
1181          * If process has NOT already set it's affinity, select and
1182          * reserve a processor for it, as a rendevous for all
1183          * users of the driver.  If they don't actually later
1184          * set affinity to this cpu, or set it to some other cpu,
1185          * it just means that sooner or later we don't recommend
1186          * a cpu, and let the scheduler do it's best.
1187          */
1188         if (weight >= qib_cpulist_count) {
1189                 int cpu;
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 "
1253                          "version (%d.%d) while context sharing. Ensure "
1254                          "that driver and library are from the same "
1255                          "release.\n",
1256                          (int) (uinfo->spu_userversion >> 16),
1257                          (int) (uinfo->spu_userversion & 0xffff),
1258                          QIB_USER_SWMAJOR, QIB_USER_SWMINOR);
1259                 goto bail;
1260         }
1261         if (num_subctxts > QLOGIC_IB_MAX_SUBCTXT) {
1262                 ret = -EINVAL;
1263                 goto bail;
1264         }
1265
1266         rcd->subctxt_uregbase = vmalloc_user(PAGE_SIZE * num_subctxts);
1267         if (!rcd->subctxt_uregbase) {
1268                 ret = -ENOMEM;
1269                 goto bail;
1270         }
1271         /* Note: rcd->rcvhdrq_size isn't initialized yet. */
1272         size = ALIGN(dd->rcvhdrcnt * dd->rcvhdrentsize *
1273                      sizeof(u32), PAGE_SIZE) * num_subctxts;
1274         rcd->subctxt_rcvhdr_base = vmalloc_user(size);
1275         if (!rcd->subctxt_rcvhdr_base) {
1276                 ret = -ENOMEM;
1277                 goto bail_ureg;
1278         }
1279
1280         rcd->subctxt_rcvegrbuf = vmalloc_user(rcd->rcvegrbuf_chunks *
1281                                               rcd->rcvegrbuf_size *
1282                                               num_subctxts);
1283         if (!rcd->subctxt_rcvegrbuf) {
1284                 ret = -ENOMEM;
1285                 goto bail_rhdr;
1286         }
1287
1288         rcd->subctxt_cnt = uinfo->spu_subctxt_cnt;
1289         rcd->subctxt_id = uinfo->spu_subctxt_id;
1290         rcd->active_slaves = 1;
1291         rcd->redirect_seq_cnt = 1;
1292         set_bit(QIB_CTXT_MASTER_UNINIT, &rcd->flag);
1293         goto bail;
1294
1295 bail_rhdr:
1296         vfree(rcd->subctxt_rcvhdr_base);
1297 bail_ureg:
1298         vfree(rcd->subctxt_uregbase);
1299         rcd->subctxt_uregbase = NULL;
1300 bail:
1301         return ret;
1302 }
1303
1304 static int setup_ctxt(struct qib_pportdata *ppd, int ctxt,
1305                       struct file *fp, const struct qib_user_info *uinfo)
1306 {
1307         struct qib_filedata *fd = fp->private_data;
1308         struct qib_devdata *dd = ppd->dd;
1309         struct qib_ctxtdata *rcd;
1310         void *ptmp = NULL;
1311         int ret;
1312         int numa_id;
1313
1314         assign_ctxt_affinity(fp, dd);
1315
1316         numa_id = qib_numa_aware ? ((fd->rec_cpu_num != -1) ?
1317                 cpu_to_node(fd->rec_cpu_num) :
1318                 numa_node_id()) : dd->assigned_node_id;
1319
1320         rcd = qib_create_ctxtdata(ppd, ctxt, numa_id);
1321
1322         /*
1323          * Allocate memory for use in qib_tid_update() at open to
1324          * reduce cost of expected send setup per message segment
1325          */
1326         if (rcd)
1327                 ptmp = kmalloc(dd->rcvtidcnt * sizeof(u16) +
1328                                dd->rcvtidcnt * sizeof(struct page **),
1329                                GFP_KERNEL);
1330
1331         if (!rcd || !ptmp) {
1332                 qib_dev_err(dd,
1333                         "Unable to allocate ctxtdata memory, failing open\n");
1334                 ret = -ENOMEM;
1335                 goto bailerr;
1336         }
1337         rcd->userversion = uinfo->spu_userversion;
1338         ret = init_subctxts(dd, rcd, uinfo);
1339         if (ret)
1340                 goto bailerr;
1341         rcd->tid_pg_list = ptmp;
1342         rcd->pid = current->pid;
1343         init_waitqueue_head(&dd->rcd[ctxt]->wait);
1344         strlcpy(rcd->comm, current->comm, sizeof(rcd->comm));
1345         ctxt_fp(fp) = rcd;
1346         qib_stats.sps_ctxts++;
1347         dd->freectxts--;
1348         ret = 0;
1349         goto bail;
1350
1351 bailerr:
1352         if (fd->rec_cpu_num != -1)
1353                 __clear_bit(fd->rec_cpu_num, qib_cpulist);
1354
1355         dd->rcd[ctxt] = NULL;
1356         kfree(rcd);
1357         kfree(ptmp);
1358 bail:
1359         return ret;
1360 }
1361
1362 static inline int usable(struct qib_pportdata *ppd)
1363 {
1364         struct qib_devdata *dd = ppd->dd;
1365
1366         return dd && (dd->flags & QIB_PRESENT) && dd->kregbase && ppd->lid &&
1367                 (ppd->lflags & QIBL_LINKACTIVE);
1368 }
1369
1370 /*
1371  * Select a context on the given device, either using a requested port
1372  * or the port based on the context number.
1373  */
1374 static int choose_port_ctxt(struct file *fp, struct qib_devdata *dd, u32 port,
1375                             const struct qib_user_info *uinfo)
1376 {
1377         struct qib_pportdata *ppd = NULL;
1378         int ret, ctxt;
1379
1380         if (port) {
1381                 if (!usable(dd->pport + port - 1)) {
1382                         ret = -ENETDOWN;
1383                         goto done;
1384                 } else
1385                         ppd = dd->pport + port - 1;
1386         }
1387         for (ctxt = dd->first_user_ctxt; ctxt < dd->cfgctxts && dd->rcd[ctxt];
1388              ctxt++)
1389                 ;
1390         if (ctxt == dd->cfgctxts) {
1391                 ret = -EBUSY;
1392                 goto done;
1393         }
1394         if (!ppd) {
1395                 u32 pidx = ctxt % dd->num_pports;
1396                 if (usable(dd->pport + pidx))
1397                         ppd = dd->pport + pidx;
1398                 else {
1399                         for (pidx = 0; pidx < dd->num_pports && !ppd;
1400                              pidx++)
1401                                 if (usable(dd->pport + pidx))
1402                                         ppd = dd->pport + pidx;
1403                 }
1404         }
1405         ret = ppd ? setup_ctxt(ppd, ctxt, fp, uinfo) : -ENETDOWN;
1406 done:
1407         return ret;
1408 }
1409
1410 static int find_free_ctxt(int unit, struct file *fp,
1411                           const struct qib_user_info *uinfo)
1412 {
1413         struct qib_devdata *dd = qib_lookup(unit);
1414         int ret;
1415
1416         if (!dd || (uinfo->spu_port && uinfo->spu_port > dd->num_pports))
1417                 ret = -ENODEV;
1418         else
1419                 ret = choose_port_ctxt(fp, dd, uinfo->spu_port, uinfo);
1420
1421         return ret;
1422 }
1423
1424 static int get_a_ctxt(struct file *fp, const struct qib_user_info *uinfo,
1425                       unsigned alg)
1426 {
1427         struct qib_devdata *udd = NULL;
1428         int ret = 0, devmax, npresent, nup, ndev, dusable = 0, i;
1429         u32 port = uinfo->spu_port, ctxt;
1430
1431         devmax = qib_count_units(&npresent, &nup);
1432         if (!npresent) {
1433                 ret = -ENXIO;
1434                 goto done;
1435         }
1436         if (nup == 0) {
1437                 ret = -ENETDOWN;
1438                 goto done;
1439         }
1440
1441         if (alg == QIB_PORT_ALG_ACROSS) {
1442                 unsigned inuse = ~0U;
1443                 /* find device (with ACTIVE ports) with fewest ctxts in use */
1444                 for (ndev = 0; ndev < devmax; ndev++) {
1445                         struct qib_devdata *dd = qib_lookup(ndev);
1446                         unsigned cused = 0, cfree = 0, pusable = 0;
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                         if (dd) {
1477                                 ret = choose_port_ctxt(fp, dd, port, uinfo);
1478                                 if (!ret)
1479                                         goto done;
1480                                 if (ret == -EBUSY)
1481                                         dusable++;
1482                         }
1483                 }
1484         }
1485         ret = dusable ? -EBUSY : -ENETDOWN;
1486
1487 done:
1488         return ret;
1489 }
1490
1491 static int find_shared_ctxt(struct file *fp,
1492                             const struct qib_user_info *uinfo)
1493 {
1494         int devmax, ndev, i;
1495         int ret = 0;
1496
1497         devmax = qib_count_units(NULL, NULL);
1498
1499         for (ndev = 0; ndev < devmax; ndev++) {
1500                 struct qib_devdata *dd = qib_lookup(ndev);
1501
1502                 /* device portion of usable() */
1503                 if (!(dd && (dd->flags & QIB_PRESENT) && dd->kregbase))
1504                         continue;
1505                 for (i = dd->first_user_ctxt; i < dd->cfgctxts; i++) {
1506                         struct qib_ctxtdata *rcd = dd->rcd[i];
1507
1508                         /* Skip ctxts which are not yet open */
1509                         if (!rcd || !rcd->cnt)
1510                                 continue;
1511                         /* Skip ctxt if it doesn't match the requested one */
1512                         if (rcd->subctxt_id != uinfo->spu_subctxt_id)
1513                                 continue;
1514                         /* Verify the sharing process matches the master */
1515                         if (rcd->subctxt_cnt != uinfo->spu_subctxt_cnt ||
1516                             rcd->userversion != uinfo->spu_userversion ||
1517                             rcd->cnt >= rcd->subctxt_cnt) {
1518                                 ret = -EINVAL;
1519                                 goto done;
1520                         }
1521                         ctxt_fp(fp) = rcd;
1522                         subctxt_fp(fp) = rcd->cnt++;
1523                         rcd->subpid[subctxt_fp(fp)] = current->pid;
1524                         tidcursor_fp(fp) = 0;
1525                         rcd->active_slaves |= 1 << subctxt_fp(fp);
1526                         ret = 1;
1527                         goto done;
1528                 }
1529         }
1530
1531 done:
1532         return ret;
1533 }
1534
1535 static int qib_open(struct inode *in, struct file *fp)
1536 {
1537         /* The real work is performed later in qib_assign_ctxt() */
1538         fp->private_data = kzalloc(sizeof(struct qib_filedata), GFP_KERNEL);
1539         if (fp->private_data) /* no cpu affinity by default */
1540                 ((struct qib_filedata *)fp->private_data)->rec_cpu_num = -1;
1541         return fp->private_data ? 0 : -ENOMEM;
1542 }
1543
1544 static int find_hca(unsigned int cpu, int *unit)
1545 {
1546         int ret = 0, devmax, npresent, nup, ndev;
1547
1548         *unit = -1;
1549
1550         devmax = qib_count_units(&npresent, &nup);
1551         if (!npresent) {
1552                 ret = -ENXIO;
1553                 goto done;
1554         }
1555         if (!nup) {
1556                 ret = -ENETDOWN;
1557                 goto done;
1558         }
1559         for (ndev = 0; ndev < devmax; ndev++) {
1560                 struct qib_devdata *dd = qib_lookup(ndev);
1561                 if (dd) {
1562                         if (pcibus_to_node(dd->pcidev->bus) < 0) {
1563                                 ret = -EINVAL;
1564                                 goto done;
1565                         }
1566                         if (cpu_to_node(cpu) ==
1567                                 pcibus_to_node(dd->pcidev->bus)) {
1568                                 *unit = ndev;
1569                                 goto done;
1570                         }
1571                 }
1572         }
1573 done:
1574         return ret;
1575 }
1576
1577 static int do_qib_user_sdma_queue_create(struct file *fp)
1578 {
1579         struct qib_filedata *fd = fp->private_data;
1580         struct qib_ctxtdata *rcd = fd->rcd;
1581         struct qib_devdata *dd = rcd->dd;
1582
1583         if (dd->flags & QIB_HAS_SEND_DMA) {
1584
1585                 fd->pq = qib_user_sdma_queue_create(&dd->pcidev->dev,
1586                                                     dd->unit,
1587                                                     rcd->ctxt,
1588                                                     fd->subctxt);
1589                 if (!fd->pq)
1590                         return -ENOMEM;
1591         }
1592
1593         return 0;
1594 }
1595
1596 /*
1597  * Get ctxt early, so can set affinity prior to memory allocation.
1598  */
1599 static int qib_assign_ctxt(struct file *fp, const struct qib_user_info *uinfo)
1600 {
1601         int ret;
1602         int i_minor;
1603         unsigned swmajor, swminor, alg = QIB_PORT_ALG_ACROSS;
1604
1605         /* Check to be sure we haven't already initialized this file */
1606         if (ctxt_fp(fp)) {
1607                 ret = -EINVAL;
1608                 goto done;
1609         }
1610
1611         /* for now, if major version is different, bail */
1612         swmajor = uinfo->spu_userversion >> 16;
1613         if (swmajor != QIB_USER_SWMAJOR) {
1614                 ret = -ENODEV;
1615                 goto done;
1616         }
1617
1618         swminor = uinfo->spu_userversion & 0xffff;
1619
1620         if (swminor >= 11 && uinfo->spu_port_alg < QIB_PORT_ALG_COUNT)
1621                 alg = uinfo->spu_port_alg;
1622
1623         mutex_lock(&qib_mutex);
1624
1625         if (qib_compatible_subctxts(swmajor, swminor) &&
1626             uinfo->spu_subctxt_cnt) {
1627                 ret = find_shared_ctxt(fp, uinfo);
1628                 if (ret > 0) {
1629                         ret = do_qib_user_sdma_queue_create(fp);
1630                         if (!ret)
1631                                 assign_ctxt_affinity(fp, (ctxt_fp(fp))->dd);
1632                         goto done_ok;
1633                 }
1634         }
1635
1636         i_minor = iminor(file_inode(fp)) - QIB_USER_MINOR_BASE;
1637         if (i_minor)
1638                 ret = find_free_ctxt(i_minor - 1, fp, uinfo);
1639         else {
1640                 int unit;
1641                 const unsigned int cpu = cpumask_first(&current->cpus_allowed);
1642                 const unsigned int weight =
1643                         cpumask_weight(&current->cpus_allowed);
1644
1645                 if (weight == 1 && !test_bit(cpu, qib_cpulist))
1646                         if (!find_hca(cpu, &unit) && unit >= 0)
1647                                 if (!find_free_ctxt(unit, fp, uinfo)) {
1648                                         ret = 0;
1649                                         goto done_chk_sdma;
1650                                 }
1651                 ret = get_a_ctxt(fp, uinfo, alg);
1652         }
1653
1654 done_chk_sdma:
1655         if (!ret)
1656                 ret = do_qib_user_sdma_queue_create(fp);
1657 done_ok:
1658         mutex_unlock(&qib_mutex);
1659
1660 done:
1661         return ret;
1662 }
1663
1664
1665 static int qib_do_user_init(struct file *fp,
1666                             const struct qib_user_info *uinfo)
1667 {
1668         int ret;
1669         struct qib_ctxtdata *rcd = ctxt_fp(fp);
1670         struct qib_devdata *dd;
1671         unsigned uctxt;
1672
1673         /* Subctxts don't need to initialize anything since master did it. */
1674         if (subctxt_fp(fp)) {
1675                 ret = wait_event_interruptible(rcd->wait,
1676                         !test_bit(QIB_CTXT_MASTER_UNINIT, &rcd->flag));
1677                 goto bail;
1678         }
1679
1680         dd = rcd->dd;
1681
1682         /* some ctxts may get extra buffers, calculate that here */
1683         uctxt = rcd->ctxt - dd->first_user_ctxt;
1684         if (uctxt < dd->ctxts_extrabuf) {
1685                 rcd->piocnt = dd->pbufsctxt + 1;
1686                 rcd->pio_base = rcd->piocnt * uctxt;
1687         } else {
1688                 rcd->piocnt = dd->pbufsctxt;
1689                 rcd->pio_base = rcd->piocnt * uctxt +
1690                         dd->ctxts_extrabuf;
1691         }
1692
1693         /*
1694          * All user buffers are 2KB buffers.  If we ever support
1695          * giving 4KB buffers to user processes, this will need some
1696          * work.  Can't use piobufbase directly, because it has
1697          * both 2K and 4K buffer base values.  So check and handle.
1698          */
1699         if ((rcd->pio_base + rcd->piocnt) > dd->piobcnt2k) {
1700                 if (rcd->pio_base >= dd->piobcnt2k) {
1701                         qib_dev_err(dd,
1702                                     "%u:ctxt%u: no 2KB buffers available\n",
1703                                     dd->unit, rcd->ctxt);
1704                         ret = -ENOBUFS;
1705                         goto bail;
1706                 }
1707                 rcd->piocnt = dd->piobcnt2k - rcd->pio_base;
1708                 qib_dev_err(dd, "Ctxt%u: would use 4KB bufs, using %u\n",
1709                             rcd->ctxt, rcd->piocnt);
1710         }
1711
1712         rcd->piobufs = dd->pio2k_bufbase + rcd->pio_base * dd->palign;
1713         qib_chg_pioavailkernel(dd, rcd->pio_base, rcd->piocnt,
1714                                TXCHK_CHG_TYPE_USER, rcd);
1715         /*
1716          * try to ensure that processes start up with consistent avail update
1717          * for their own range, at least.   If system very quiet, it might
1718          * have the in-memory copy out of date at startup for this range of
1719          * buffers, when a context gets re-used.  Do after the chg_pioavail
1720          * and before the rest of setup, so it's "almost certain" the dma
1721          * will have occurred (can't 100% guarantee, but should be many
1722          * decimals of 9s, with this ordering), given how much else happens
1723          * after this.
1724          */
1725         dd->f_sendctrl(dd->pport, QIB_SENDCTRL_AVAIL_BLIP);
1726
1727         /*
1728          * Now allocate the rcvhdr Q and eager TIDs; skip the TID
1729          * array for time being.  If rcd->ctxt > chip-supported,
1730          * we need to do extra stuff here to handle by handling overflow
1731          * through ctxt 0, someday
1732          */
1733         ret = qib_create_rcvhdrq(dd, rcd);
1734         if (!ret)
1735                 ret = qib_setup_eagerbufs(rcd);
1736         if (ret)
1737                 goto bail_pio;
1738
1739         rcd->tidcursor = 0; /* start at beginning after open */
1740
1741         /* initialize poll variables... */
1742         rcd->urgent = 0;
1743         rcd->urgent_poll = 0;
1744
1745         /*
1746          * Now enable the ctxt for receive.
1747          * For chips that are set to DMA the tail register to memory
1748          * when they change (and when the update bit transitions from
1749          * 0 to 1.  So for those chips, we turn it off and then back on.
1750          * This will (very briefly) affect any other open ctxts, but the
1751          * duration is very short, and therefore isn't an issue.  We
1752          * explicitly set the in-memory tail copy to 0 beforehand, so we
1753          * don't have to wait to be sure the DMA update has happened
1754          * (chip resets head/tail to 0 on transition to enable).
1755          */
1756         if (rcd->rcvhdrtail_kvaddr)
1757                 qib_clear_rcvhdrtail(rcd);
1758
1759         dd->f_rcvctrl(rcd->ppd, QIB_RCVCTRL_CTXT_ENB | QIB_RCVCTRL_TIDFLOW_ENB,
1760                       rcd->ctxt);
1761
1762         /* Notify any waiting slaves */
1763         if (rcd->subctxt_cnt) {
1764                 clear_bit(QIB_CTXT_MASTER_UNINIT, &rcd->flag);
1765                 wake_up(&rcd->wait);
1766         }
1767         return 0;
1768
1769 bail_pio:
1770         qib_chg_pioavailkernel(dd, rcd->pio_base, rcd->piocnt,
1771                                TXCHK_CHG_TYPE_KERN, rcd);
1772 bail:
1773         return ret;
1774 }
1775
1776 /**
1777  * unlock_exptid - unlock any expected TID entries context still had in use
1778  * @rcd: ctxt
1779  *
1780  * We don't actually update the chip here, because we do a bulk update
1781  * below, using f_clear_tids.
1782  */
1783 static void unlock_expected_tids(struct qib_ctxtdata *rcd)
1784 {
1785         struct qib_devdata *dd = rcd->dd;
1786         int ctxt_tidbase = rcd->ctxt * dd->rcvtidcnt;
1787         int i, cnt = 0, maxtid = ctxt_tidbase + dd->rcvtidcnt;
1788
1789         for (i = ctxt_tidbase; i < maxtid; i++) {
1790                 struct page *p = dd->pageshadow[i];
1791                 dma_addr_t phys;
1792
1793                 if (!p)
1794                         continue;
1795
1796                 phys = dd->physshadow[i];
1797                 dd->physshadow[i] = dd->tidinvalid;
1798                 dd->pageshadow[i] = NULL;
1799                 pci_unmap_page(dd->pcidev, phys, PAGE_SIZE,
1800                                PCI_DMA_FROMDEVICE);
1801                 qib_release_user_pages(&p, 1);
1802                 cnt++;
1803         }
1804 }
1805
1806 static int qib_close(struct inode *in, struct file *fp)
1807 {
1808         int ret = 0;
1809         struct qib_filedata *fd;
1810         struct qib_ctxtdata *rcd;
1811         struct qib_devdata *dd;
1812         unsigned long flags;
1813         unsigned ctxt;
1814         pid_t pid;
1815
1816         mutex_lock(&qib_mutex);
1817
1818         fd = fp->private_data;
1819         fp->private_data = NULL;
1820         rcd = fd->rcd;
1821         if (!rcd) {
1822                 mutex_unlock(&qib_mutex);
1823                 goto bail;
1824         }
1825
1826         dd = rcd->dd;
1827
1828         /* ensure all pio buffer writes in progress are flushed */
1829         qib_flush_wc();
1830
1831         /* drain user sdma queue */
1832         if (fd->pq) {
1833                 qib_user_sdma_queue_drain(rcd->ppd, fd->pq);
1834                 qib_user_sdma_queue_destroy(fd->pq);
1835         }
1836
1837         if (fd->rec_cpu_num != -1)
1838                 __clear_bit(fd->rec_cpu_num, qib_cpulist);
1839
1840         if (--rcd->cnt) {
1841                 /*
1842                  * XXX If the master closes the context before the slave(s),
1843                  * revoke the mmap for the eager receive queue so
1844                  * the slave(s) don't wait for receive data forever.
1845                  */
1846                 rcd->active_slaves &= ~(1 << fd->subctxt);
1847                 rcd->subpid[fd->subctxt] = 0;
1848                 mutex_unlock(&qib_mutex);
1849                 goto bail;
1850         }
1851
1852         /* early; no interrupt users after this */
1853         spin_lock_irqsave(&dd->uctxt_lock, flags);
1854         ctxt = rcd->ctxt;
1855         dd->rcd[ctxt] = NULL;
1856         pid = rcd->pid;
1857         rcd->pid = 0;
1858         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
1859
1860         if (rcd->rcvwait_to || rcd->piowait_to ||
1861             rcd->rcvnowait || rcd->pionowait) {
1862                 rcd->rcvwait_to = 0;
1863                 rcd->piowait_to = 0;
1864                 rcd->rcvnowait = 0;
1865                 rcd->pionowait = 0;
1866         }
1867         if (rcd->flag)
1868                 rcd->flag = 0;
1869
1870         if (dd->kregbase) {
1871                 /* atomically clear receive enable ctxt and intr avail. */
1872                 dd->f_rcvctrl(rcd->ppd, QIB_RCVCTRL_CTXT_DIS |
1873                                   QIB_RCVCTRL_INTRAVAIL_DIS, ctxt);
1874
1875                 /* clean up the pkeys for this ctxt user */
1876                 qib_clean_part_key(rcd, dd);
1877                 qib_disarm_piobufs(dd, rcd->pio_base, rcd->piocnt);
1878                 qib_chg_pioavailkernel(dd, rcd->pio_base,
1879                                        rcd->piocnt, TXCHK_CHG_TYPE_KERN, NULL);
1880
1881                 dd->f_clear_tids(dd, rcd);
1882
1883                 if (dd->pageshadow)
1884                         unlock_expected_tids(rcd);
1885                 qib_stats.sps_ctxts--;
1886                 dd->freectxts++;
1887         }
1888
1889         mutex_unlock(&qib_mutex);
1890         qib_free_ctxtdata(dd, rcd); /* after releasing the mutex */
1891
1892 bail:
1893         kfree(fd);
1894         return ret;
1895 }
1896
1897 static int qib_ctxt_info(struct file *fp, struct qib_ctxt_info __user *uinfo)
1898 {
1899         struct qib_ctxt_info info;
1900         int ret;
1901         size_t sz;
1902         struct qib_ctxtdata *rcd = ctxt_fp(fp);
1903         struct qib_filedata *fd;
1904
1905         fd = fp->private_data;
1906
1907         info.num_active = qib_count_active_units();
1908         info.unit = rcd->dd->unit;
1909         info.port = rcd->ppd->port;
1910         info.ctxt = rcd->ctxt;
1911         info.subctxt =  subctxt_fp(fp);
1912         /* Number of user ctxts available for this device. */
1913         info.num_ctxts = rcd->dd->cfgctxts - rcd->dd->first_user_ctxt;
1914         info.num_subctxts = rcd->subctxt_cnt;
1915         info.rec_cpu = fd->rec_cpu_num;
1916         sz = sizeof(info);
1917
1918         if (copy_to_user(uinfo, &info, sz)) {
1919                 ret = -EFAULT;
1920                 goto bail;
1921         }
1922         ret = 0;
1923
1924 bail:
1925         return ret;
1926 }
1927
1928 static int qib_sdma_get_inflight(struct qib_user_sdma_queue *pq,
1929                                  u32 __user *inflightp)
1930 {
1931         const u32 val = qib_user_sdma_inflight_counter(pq);
1932
1933         if (put_user(val, inflightp))
1934                 return -EFAULT;
1935
1936         return 0;
1937 }
1938
1939 static int qib_sdma_get_complete(struct qib_pportdata *ppd,
1940                                  struct qib_user_sdma_queue *pq,
1941                                  u32 __user *completep)
1942 {
1943         u32 val;
1944         int err;
1945
1946         if (!pq)
1947                 return -EINVAL;
1948
1949         err = qib_user_sdma_make_progress(ppd, pq);
1950         if (err < 0)
1951                 return err;
1952
1953         val = qib_user_sdma_complete_counter(pq);
1954         if (put_user(val, completep))
1955                 return -EFAULT;
1956
1957         return 0;
1958 }
1959
1960 static int disarm_req_delay(struct qib_ctxtdata *rcd)
1961 {
1962         int ret = 0;
1963
1964         if (!usable(rcd->ppd)) {
1965                 int i;
1966                 /*
1967                  * if link is down, or otherwise not usable, delay
1968                  * the caller up to 30 seconds, so we don't thrash
1969                  * in trying to get the chip back to ACTIVE, and
1970                  * set flag so they make the call again.
1971                  */
1972                 if (rcd->user_event_mask) {
1973                         /*
1974                          * subctxt_cnt is 0 if not shared, so do base
1975                          * separately, first, then remaining subctxt, if any
1976                          */
1977                         set_bit(_QIB_EVENT_DISARM_BUFS_BIT,
1978                                 &rcd->user_event_mask[0]);
1979                         for (i = 1; i < rcd->subctxt_cnt; i++)
1980                                 set_bit(_QIB_EVENT_DISARM_BUFS_BIT,
1981                                         &rcd->user_event_mask[i]);
1982                 }
1983                 for (i = 0; !usable(rcd->ppd) && i < 300; i++)
1984                         msleep(100);
1985                 ret = -ENETDOWN;
1986         }
1987         return ret;
1988 }
1989
1990 /*
1991  * Find all user contexts in use, and set the specified bit in their
1992  * event mask.
1993  * See also find_ctxt() for a similar use, that is specific to send buffers.
1994  */
1995 int qib_set_uevent_bits(struct qib_pportdata *ppd, const int evtbit)
1996 {
1997         struct qib_ctxtdata *rcd;
1998         unsigned ctxt;
1999         int ret = 0;
2000         unsigned long flags;
2001
2002         spin_lock_irqsave(&ppd->dd->uctxt_lock, flags);
2003         for (ctxt = ppd->dd->first_user_ctxt; ctxt < ppd->dd->cfgctxts;
2004              ctxt++) {
2005                 rcd = ppd->dd->rcd[ctxt];
2006                 if (!rcd)
2007                         continue;
2008                 if (rcd->user_event_mask) {
2009                         int i;
2010                         /*
2011                          * subctxt_cnt is 0 if not shared, so do base
2012                          * separately, first, then remaining subctxt, if any
2013                          */
2014                         set_bit(evtbit, &rcd->user_event_mask[0]);
2015                         for (i = 1; i < rcd->subctxt_cnt; i++)
2016                                 set_bit(evtbit, &rcd->user_event_mask[i]);
2017                 }
2018                 ret = 1;
2019                 break;
2020         }
2021         spin_unlock_irqrestore(&ppd->dd->uctxt_lock, flags);
2022
2023         return ret;
2024 }
2025
2026 /*
2027  * clear the event notifier events for this context.
2028  * For the DISARM_BUFS case, we also take action (this obsoletes
2029  * the older QIB_CMD_DISARM_BUFS, but we keep it for backwards
2030  * compatibility.
2031  * Other bits don't currently require actions, just atomically clear.
2032  * User process then performs actions appropriate to bit having been
2033  * set, if desired, and checks again in future.
2034  */
2035 static int qib_user_event_ack(struct qib_ctxtdata *rcd, int subctxt,
2036                               unsigned long events)
2037 {
2038         int ret = 0, i;
2039
2040         for (i = 0; i <= _QIB_MAX_EVENT_BIT; i++) {
2041                 if (!test_bit(i, &events))
2042                         continue;
2043                 if (i == _QIB_EVENT_DISARM_BUFS_BIT) {
2044                         (void)qib_disarm_piobufs_ifneeded(rcd);
2045                         ret = disarm_req_delay(rcd);
2046                 } else
2047                         clear_bit(i, &rcd->user_event_mask[subctxt]);
2048         }
2049         return ret;
2050 }
2051
2052 static ssize_t qib_write(struct file *fp, const char __user *data,
2053                          size_t count, loff_t *off)
2054 {
2055         const struct qib_cmd __user *ucmd;
2056         struct qib_ctxtdata *rcd;
2057         const void __user *src;
2058         size_t consumed, copy = 0;
2059         struct qib_cmd cmd;
2060         ssize_t ret = 0;
2061         void *dest;
2062
2063         if (WARN_ON_ONCE(!ib_safe_file_access(fp)))
2064                 return -EACCES;
2065
2066         if (count < sizeof(cmd.type)) {
2067                 ret = -EINVAL;
2068                 goto bail;
2069         }
2070
2071         ucmd = (const struct qib_cmd __user *) data;
2072
2073         if (copy_from_user(&cmd.type, &ucmd->type, sizeof(cmd.type))) {
2074                 ret = -EFAULT;
2075                 goto bail;
2076         }
2077
2078         consumed = sizeof(cmd.type);
2079
2080         switch (cmd.type) {
2081         case QIB_CMD_ASSIGN_CTXT:
2082         case QIB_CMD_USER_INIT:
2083                 copy = sizeof(cmd.cmd.user_info);
2084                 dest = &cmd.cmd.user_info;
2085                 src = &ucmd->cmd.user_info;
2086                 break;
2087
2088         case QIB_CMD_RECV_CTRL:
2089                 copy = sizeof(cmd.cmd.recv_ctrl);
2090                 dest = &cmd.cmd.recv_ctrl;
2091                 src = &ucmd->cmd.recv_ctrl;
2092                 break;
2093
2094         case QIB_CMD_CTXT_INFO:
2095                 copy = sizeof(cmd.cmd.ctxt_info);
2096                 dest = &cmd.cmd.ctxt_info;
2097                 src = &ucmd->cmd.ctxt_info;
2098                 break;
2099
2100         case QIB_CMD_TID_UPDATE:
2101         case QIB_CMD_TID_FREE:
2102                 copy = sizeof(cmd.cmd.tid_info);
2103                 dest = &cmd.cmd.tid_info;
2104                 src = &ucmd->cmd.tid_info;
2105                 break;
2106
2107         case QIB_CMD_SET_PART_KEY:
2108                 copy = sizeof(cmd.cmd.part_key);
2109                 dest = &cmd.cmd.part_key;
2110                 src = &ucmd->cmd.part_key;
2111                 break;
2112
2113         case QIB_CMD_DISARM_BUFS:
2114         case QIB_CMD_PIOAVAILUPD: /* force an update of PIOAvail reg */
2115                 copy = 0;
2116                 src = NULL;
2117                 dest = NULL;
2118                 break;
2119
2120         case QIB_CMD_POLL_TYPE:
2121                 copy = sizeof(cmd.cmd.poll_type);
2122                 dest = &cmd.cmd.poll_type;
2123                 src = &ucmd->cmd.poll_type;
2124                 break;
2125
2126         case QIB_CMD_ARMLAUNCH_CTRL:
2127                 copy = sizeof(cmd.cmd.armlaunch_ctrl);
2128                 dest = &cmd.cmd.armlaunch_ctrl;
2129                 src = &ucmd->cmd.armlaunch_ctrl;
2130                 break;
2131
2132         case QIB_CMD_SDMA_INFLIGHT:
2133                 copy = sizeof(cmd.cmd.sdma_inflight);
2134                 dest = &cmd.cmd.sdma_inflight;
2135                 src = &ucmd->cmd.sdma_inflight;
2136                 break;
2137
2138         case QIB_CMD_SDMA_COMPLETE:
2139                 copy = sizeof(cmd.cmd.sdma_complete);
2140                 dest = &cmd.cmd.sdma_complete;
2141                 src = &ucmd->cmd.sdma_complete;
2142                 break;
2143
2144         case QIB_CMD_ACK_EVENT:
2145                 copy = sizeof(cmd.cmd.event_mask);
2146                 dest = &cmd.cmd.event_mask;
2147                 src = &ucmd->cmd.event_mask;
2148                 break;
2149
2150         default:
2151                 ret = -EINVAL;
2152                 goto bail;
2153         }
2154
2155         if (copy) {
2156                 if ((count - consumed) < copy) {
2157                         ret = -EINVAL;
2158                         goto bail;
2159                 }
2160                 if (copy_from_user(dest, src, copy)) {
2161                         ret = -EFAULT;
2162                         goto bail;
2163                 }
2164                 consumed += copy;
2165         }
2166
2167         rcd = ctxt_fp(fp);
2168         if (!rcd && cmd.type != QIB_CMD_ASSIGN_CTXT) {
2169                 ret = -EINVAL;
2170                 goto bail;
2171         }
2172
2173         switch (cmd.type) {
2174         case QIB_CMD_ASSIGN_CTXT:
2175                 ret = qib_assign_ctxt(fp, &cmd.cmd.user_info);
2176                 if (ret)
2177                         goto bail;
2178                 break;
2179
2180         case QIB_CMD_USER_INIT:
2181                 ret = qib_do_user_init(fp, &cmd.cmd.user_info);
2182                 if (ret)
2183                         goto bail;
2184                 ret = qib_get_base_info(fp, (void __user *) (unsigned long)
2185                                         cmd.cmd.user_info.spu_base_info,
2186                                         cmd.cmd.user_info.spu_base_info_size);
2187                 break;
2188
2189         case QIB_CMD_RECV_CTRL:
2190                 ret = qib_manage_rcvq(rcd, subctxt_fp(fp), cmd.cmd.recv_ctrl);
2191                 break;
2192
2193         case QIB_CMD_CTXT_INFO:
2194                 ret = qib_ctxt_info(fp, (struct qib_ctxt_info __user *)
2195                                     (unsigned long) cmd.cmd.ctxt_info);
2196                 break;
2197
2198         case QIB_CMD_TID_UPDATE:
2199                 ret = qib_tid_update(rcd, fp, &cmd.cmd.tid_info);
2200                 break;
2201
2202         case QIB_CMD_TID_FREE:
2203                 ret = qib_tid_free(rcd, subctxt_fp(fp), &cmd.cmd.tid_info);
2204                 break;
2205
2206         case QIB_CMD_SET_PART_KEY:
2207                 ret = qib_set_part_key(rcd, cmd.cmd.part_key);
2208                 break;
2209
2210         case QIB_CMD_DISARM_BUFS:
2211                 (void)qib_disarm_piobufs_ifneeded(rcd);
2212                 ret = disarm_req_delay(rcd);
2213                 break;
2214
2215         case QIB_CMD_PIOAVAILUPD:
2216                 qib_force_pio_avail_update(rcd->dd);
2217                 break;
2218
2219         case QIB_CMD_POLL_TYPE:
2220                 rcd->poll_type = cmd.cmd.poll_type;
2221                 break;
2222
2223         case QIB_CMD_ARMLAUNCH_CTRL:
2224                 rcd->dd->f_set_armlaunch(rcd->dd, cmd.cmd.armlaunch_ctrl);
2225                 break;
2226
2227         case QIB_CMD_SDMA_INFLIGHT:
2228                 ret = qib_sdma_get_inflight(user_sdma_queue_fp(fp),
2229                                             (u32 __user *) (unsigned long)
2230                                             cmd.cmd.sdma_inflight);
2231                 break;
2232
2233         case QIB_CMD_SDMA_COMPLETE:
2234                 ret = qib_sdma_get_complete(rcd->ppd,
2235                                             user_sdma_queue_fp(fp),
2236                                             (u32 __user *) (unsigned long)
2237                                             cmd.cmd.sdma_complete);
2238                 break;
2239
2240         case QIB_CMD_ACK_EVENT:
2241                 ret = qib_user_event_ack(rcd, subctxt_fp(fp),
2242                                          cmd.cmd.event_mask);
2243                 break;
2244         }
2245
2246         if (ret >= 0)
2247                 ret = consumed;
2248
2249 bail:
2250         return ret;
2251 }
2252
2253 static ssize_t qib_aio_write(struct kiocb *iocb, const struct iovec *iov,
2254                              unsigned long dim, loff_t off)
2255 {
2256         struct qib_filedata *fp = iocb->ki_filp->private_data;
2257         struct qib_ctxtdata *rcd = ctxt_fp(iocb->ki_filp);
2258         struct qib_user_sdma_queue *pq = fp->pq;
2259
2260         if (!dim || !pq)
2261                 return -EINVAL;
2262
2263         return qib_user_sdma_writev(rcd, pq, iov, dim);
2264 }
2265
2266 static struct class *qib_class;
2267 static dev_t qib_dev;
2268
2269 int qib_cdev_init(int minor, const char *name,
2270                   const struct file_operations *fops,
2271                   struct cdev **cdevp, struct device **devp)
2272 {
2273         const dev_t dev = MKDEV(MAJOR(qib_dev), minor);
2274         struct cdev *cdev;
2275         struct device *device = NULL;
2276         int ret;
2277
2278         cdev = cdev_alloc();
2279         if (!cdev) {
2280                 pr_err("Could not allocate cdev for minor %d, %s\n",
2281                        minor, name);
2282                 ret = -ENOMEM;
2283                 goto done;
2284         }
2285
2286         cdev->owner = THIS_MODULE;
2287         cdev->ops = fops;
2288         kobject_set_name(&cdev->kobj, name);
2289
2290         ret = cdev_add(cdev, dev, 1);
2291         if (ret < 0) {
2292                 pr_err("Could not add cdev for minor %d, %s (err %d)\n",
2293                        minor, name, -ret);
2294                 goto err_cdev;
2295         }
2296
2297         device = device_create(qib_class, NULL, dev, NULL, "%s", name);
2298         if (!IS_ERR(device))
2299                 goto done;
2300         ret = PTR_ERR(device);
2301         device = NULL;
2302         pr_err("Could not create device for minor %d, %s (err %d)\n",
2303                minor, name, -ret);
2304 err_cdev:
2305         cdev_del(cdev);
2306         cdev = NULL;
2307 done:
2308         *cdevp = cdev;
2309         *devp = device;
2310         return ret;
2311 }
2312
2313 void qib_cdev_cleanup(struct cdev **cdevp, struct device **devp)
2314 {
2315         struct device *device = *devp;
2316
2317         if (device) {
2318                 device_unregister(device);
2319                 *devp = NULL;
2320         }
2321
2322         if (*cdevp) {
2323                 cdev_del(*cdevp);
2324                 *cdevp = NULL;
2325         }
2326 }
2327
2328 static struct cdev *wildcard_cdev;
2329 static struct device *wildcard_device;
2330
2331 int __init qib_dev_init(void)
2332 {
2333         int ret;
2334
2335         ret = alloc_chrdev_region(&qib_dev, 0, QIB_NMINORS, QIB_DRV_NAME);
2336         if (ret < 0) {
2337                 pr_err("Could not allocate chrdev region (err %d)\n", -ret);
2338                 goto done;
2339         }
2340
2341         qib_class = class_create(THIS_MODULE, "ipath");
2342         if (IS_ERR(qib_class)) {
2343                 ret = PTR_ERR(qib_class);
2344                 pr_err("Could not create device class (err %d)\n", -ret);
2345                 unregister_chrdev_region(qib_dev, QIB_NMINORS);
2346         }
2347
2348 done:
2349         return ret;
2350 }
2351
2352 void qib_dev_cleanup(void)
2353 {
2354         if (qib_class) {
2355                 class_destroy(qib_class);
2356                 qib_class = NULL;
2357         }
2358
2359         unregister_chrdev_region(qib_dev, QIB_NMINORS);
2360 }
2361
2362 static atomic_t user_count = ATOMIC_INIT(0);
2363
2364 static void qib_user_remove(struct qib_devdata *dd)
2365 {
2366         if (atomic_dec_return(&user_count) == 0)
2367                 qib_cdev_cleanup(&wildcard_cdev, &wildcard_device);
2368
2369         qib_cdev_cleanup(&dd->user_cdev, &dd->user_device);
2370 }
2371
2372 static int qib_user_add(struct qib_devdata *dd)
2373 {
2374         char name[10];
2375         int ret;
2376
2377         if (atomic_inc_return(&user_count) == 1) {
2378                 ret = qib_cdev_init(0, "ipath", &qib_file_ops,
2379                                     &wildcard_cdev, &wildcard_device);
2380                 if (ret)
2381                         goto done;
2382         }
2383
2384         snprintf(name, sizeof(name), "ipath%d", dd->unit);
2385         ret = qib_cdev_init(dd->unit + 1, name, &qib_file_ops,
2386                             &dd->user_cdev, &dd->user_device);
2387         if (ret)
2388                 qib_user_remove(dd);
2389 done:
2390         return ret;
2391 }
2392
2393 /*
2394  * Create per-unit files in /dev
2395  */
2396 int qib_device_create(struct qib_devdata *dd)
2397 {
2398         int r, ret;
2399
2400         r = qib_user_add(dd);
2401         ret = qib_diag_add(dd);
2402         if (r && !ret)
2403                 ret = r;
2404         return ret;
2405 }
2406
2407 /*
2408  * Remove per-unit files in /dev
2409  * void, core kernel returns no errors for this stuff
2410  */
2411 void qib_device_remove(struct qib_devdata *dd)
2412 {
2413         qib_user_remove(dd);
2414         qib_diag_remove(dd);
2415 }