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/jiffies.h>
  43 #include <asm/pgtable.h>
  44 #include <linux/delay.h>
  45 #include <linux/export.h>
  46 #include <linux/uio.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_write_iter(struct kiocb *, struct iov_iter *);
  61 static unsigned int qib_poll(struct file *, struct poll_table_struct *);
  62 static int qib_mmapf(struct file *, struct vm_area_struct *);
  63
  64 /*
  65  * This is really, really weird shit - write() and writev() here
  66  * have completely unrelated semantics.  Sucky userland ABI,
  67  * film at 11.
  68  */
  69 static const struct file_operations qib_file_ops = {
  70         .owner = THIS_MODULE,
  71         .write = qib_write,
  72         .write_iter = qib_write_iter,
  73         .open = qib_open,
  74         .release = qib_close,
  75         .poll = qib_poll,
  76         .mmap = qib_mmapf,
  77         .llseek = noop_llseek,
  78 };
  79
  80 /*
  81  * Convert kernel virtual addresses to physical addresses so they don't
  82  * potentially conflict with the chip addresses used as mmap offsets.
  83  * It doesn't really matter what mmap offset we use as long as we can
  84  * interpret it correctly.
  85  */
  86 static u64 cvt_kvaddr(void *p)
  87 {
  88         struct page *page;
  89         u64 paddr = 0;
  90
  91         page = vmalloc_to_page(p);
  92         if (page)
  93                 paddr = page_to_pfn(page) << PAGE_SHIFT;
  94
  95         return paddr;
  96 }
  97
  98 static int qib_get_base_info(struct file *fp, void __user *ubase,
  99                              size_t ubase_size)
 100 {
 101         struct qib_ctxtdata *rcd = ctxt_fp(fp);
 102         int ret = 0;
 103         struct qib_base_info *kinfo = NULL;
 104         struct qib_devdata *dd = rcd->dd;
 105         struct qib_pportdata *ppd = rcd->ppd;
 106         unsigned subctxt_cnt;
 107         int shared, master;
 108         size_t sz;
 109
 110         subctxt_cnt = rcd->subctxt_cnt;
 111         if (!subctxt_cnt) {
 112                 shared = 0;
 113                 master = 0;
 114                 subctxt_cnt = 1;
 115         } else {
 116                 shared = 1;
 117                 master = !subctxt_fp(fp);
 118         }
 119
 120         sz = sizeof(*kinfo);
 121         /* If context sharing is not requested, allow the old size structure */
 122         if (!shared)
 123                 sz -= 7 * sizeof(u64);
 124         if (ubase_size < sz) {
 125                 ret = -EINVAL;
 126                 goto bail;
 127         }
 128
 129         kinfo = kzalloc(sizeof(*kinfo), GFP_KERNEL);
 130         if (kinfo == NULL) {
 131                 ret = -ENOMEM;
 132                 goto bail;
 133         }
 134
 135         ret = dd->f_get_base_info(rcd, kinfo);
 136         if (ret < 0)
 137                 goto bail;
 138
 139         kinfo->spi_rcvhdr_cnt = dd->rcvhdrcnt;
 140         kinfo->spi_rcvhdrent_size = dd->rcvhdrentsize;
 141         kinfo->spi_tidegrcnt = rcd->rcvegrcnt;
 142         kinfo->spi_rcv_egrbufsize = dd->rcvegrbufsize;
 143         /*
 144          * have to mmap whole thing
 145          */
 146         kinfo->spi_rcv_egrbuftotlen =
 147                 rcd->rcvegrbuf_chunks * rcd->rcvegrbuf_size;
 148         kinfo->spi_rcv_egrperchunk = rcd->rcvegrbufs_perchunk;
 149         kinfo->spi_rcv_egrchunksize = kinfo->spi_rcv_egrbuftotlen /
 150                 rcd->rcvegrbuf_chunks;
 151         kinfo->spi_tidcnt = dd->rcvtidcnt / subctxt_cnt;
 152         if (master)
 153                 kinfo->spi_tidcnt += dd->rcvtidcnt % subctxt_cnt;
 154         /*
 155          * for this use, may be cfgctxts summed over all chips that
 156          * are are configured and present
 157          */
 158         kinfo->spi_nctxts = dd->cfgctxts;
 159         /* unit (chip/board) our context is on */
 160         kinfo->spi_unit = dd->unit;
 161         kinfo->spi_port = ppd->port;
 162         /* for now, only a single page */
 163         kinfo->spi_tid_maxsize = PAGE_SIZE;
 164
 165         /*
 166          * Doing this per context, and based on the skip value, etc.  This has
 167          * to be the actual buffer size, since the protocol code treats it
 168          * as an array.
 169          *
 170          * These have to be set to user addresses in the user code via mmap.
 171          * These values are used on return to user code for the mmap target
 172          * addresses only.  For 32 bit, same 44 bit address problem, so use
 173          * the physical address, not virtual.  Before 2.6.11, using the
 174          * page_address() macro worked, but in 2.6.11, even that returns the
 175          * full 64 bit address (upper bits all 1's).  So far, using the
 176          * physical addresses (or chip offsets, for chip mapping) works, but
 177          * no doubt some future kernel release will change that, and we'll be
 178          * on to yet another method of dealing with this.
 179          * Normally only one of rcvhdr_tailaddr or rhf_offset is useful
 180          * since the chips with non-zero rhf_offset don't normally
 181          * enable tail register updates to host memory, but for testing,
 182          * both can be enabled and used.
 183          */
 184         kinfo->spi_rcvhdr_base = (u64) rcd->rcvhdrq_phys;
 185         kinfo->spi_rcvhdr_tailaddr = (u64) rcd->rcvhdrqtailaddr_phys;
 186         kinfo->spi_rhf_offset = dd->rhf_offset;
 187         kinfo->spi_rcv_egrbufs = (u64) rcd->rcvegr_phys;
 188         kinfo->spi_pioavailaddr = (u64) dd->pioavailregs_phys;
 189         /* setup per-unit (not port) status area for user programs */
 190         kinfo->spi_status = (u64) kinfo->spi_pioavailaddr +
 191                 (char *) ppd->statusp -
 192                 (char *) dd->pioavailregs_dma;
 193         kinfo->spi_uregbase = (u64) dd->uregbase + dd->ureg_align * rcd->ctxt;
 194         if (!shared) {
 195                 kinfo->spi_piocnt = rcd->piocnt;
 196                 kinfo->spi_piobufbase = (u64) rcd->piobufs;
 197                 kinfo->spi_sendbuf_status = cvt_kvaddr(rcd->user_event_mask);
 198         } else if (master) {
 199                 kinfo->spi_piocnt = (rcd->piocnt / subctxt_cnt) +
 200                                     (rcd->piocnt % subctxt_cnt);
 201                 /* Master's PIO buffers are after all the slave's */
 202                 kinfo->spi_piobufbase = (u64) rcd->piobufs +
 203                         dd->palign *
 204                         (rcd->piocnt - kinfo->spi_piocnt);
 205         } else {
 206                 unsigned slave = subctxt_fp(fp) - 1;
 207
 208                 kinfo->spi_piocnt = rcd->piocnt / subctxt_cnt;
 209                 kinfo->spi_piobufbase = (u64) rcd->piobufs +
 210                         dd->palign * kinfo->spi_piocnt * slave;
 211         }
 212
 213         if (shared) {
 214                 kinfo->spi_sendbuf_status =
 215                         cvt_kvaddr(&rcd->user_event_mask[subctxt_fp(fp)]);
 216                 /* only spi_subctxt_* fields should be set in this block! */
 217                 kinfo->spi_subctxt_uregbase = cvt_kvaddr(rcd->subctxt_uregbase);
 218
 219                 kinfo->spi_subctxt_rcvegrbuf =
 220                         cvt_kvaddr(rcd->subctxt_rcvegrbuf);
 221                 kinfo->spi_subctxt_rcvhdr_base =
 222                         cvt_kvaddr(rcd->subctxt_rcvhdr_base);
 223         }
 224
 225         /*
 226          * All user buffers are 2KB buffers.  If we ever support
 227          * giving 4KB buffers to user processes, this will need some
 228          * work.  Can't use piobufbase directly, because it has
 229          * both 2K and 4K buffer base values.
 230          */
 231         kinfo->spi_pioindex = (kinfo->spi_piobufbase - dd->pio2k_bufbase) /
 232                 dd->palign;
 233         kinfo->spi_pioalign = dd->palign;
 234         kinfo->spi_qpair = QIB_KD_QP;
 235         /*
 236          * user mode PIO buffers are always 2KB, even when 4KB can
 237          * be received, and sent via the kernel; this is ibmaxlen
 238          * for 2K MTU.
 239          */
 240         kinfo->spi_piosize = dd->piosize2k - 2 * sizeof(u32);
 241         kinfo->spi_mtu = ppd->ibmaxlen; /* maxlen, not ibmtu */
 242         kinfo->spi_ctxt = rcd->ctxt;
 243         kinfo->spi_subctxt = subctxt_fp(fp);
 244         kinfo->spi_sw_version = QIB_KERN_SWVERSION;
 245         kinfo->spi_sw_version |= 1U << 31; /* QLogic-built, not kernel.org */
 246         kinfo->spi_hw_version = dd->revision;
 247
 248         if (master)
 249                 kinfo->spi_runtime_flags |= QIB_RUNTIME_MASTER;
 250
 251         sz = (ubase_size < sizeof(*kinfo)) ? ubase_size : sizeof(*kinfo);
 252         if (copy_to_user(ubase, kinfo, sz))
 253                 ret = -EFAULT;
 254 bail:
 255         kfree(kinfo);
 256         return ret;
 257 }
 258
 259 /**
 260  * qib_tid_update - update a context TID
 261  * @rcd: the context
 262  * @fp: the qib device file
 263  * @ti: the TID information
 264  *
 265  * The new implementation as of Oct 2004 is that the driver assigns
 266  * the tid and returns it to the caller.   To reduce search time, we
 267  * keep a cursor for each context, walking the shadow tid array to find
 268  * one that's not in use.
 269  *
 270  * For now, if we can't allocate the full list, we fail, although
 271  * in the long run, we'll allocate as many as we can, and the
 272  * caller will deal with that by trying the remaining pages later.
 273  * That means that when we fail, we have to mark the tids as not in
 274  * use again, in our shadow copy.
 275  *
 276  * It's up to the caller to free the tids when they are done.
 277  * We'll unlock the pages as they free them.
 278  *
 279  * Also, right now we are locking one page at a time, but since
 280  * the intended use of this routine is for a single group of
 281  * virtually contiguous pages, that should change to improve
 282  * performance.
 283  */
 284 static int qib_tid_update(struct qib_ctxtdata *rcd, struct file *fp,
 285                           const struct qib_tid_info *ti)
 286 {
 287         int ret = 0, ntids;
 288         u32 tid, ctxttid, cnt, i, tidcnt, tidoff;
 289         u16 *tidlist;
 290         struct qib_devdata *dd = rcd->dd;
 291         u64 physaddr;
 292         unsigned long vaddr;
 293         u64 __iomem *tidbase;
 294         unsigned long tidmap[8];
 295         struct page **pagep = NULL;
 296         unsigned subctxt = subctxt_fp(fp);
 297
 298         if (!dd->pageshadow) {
 299                 ret = -ENOMEM;
 300                 goto done;
 301         }
 302
 303         cnt = ti->tidcnt;
 304         if (!cnt) {
 305                 ret = -EFAULT;
 306                 goto done;
 307         }
 308         ctxttid = rcd->ctxt * dd->rcvtidcnt;
 309         if (!rcd->subctxt_cnt) {
 310                 tidcnt = dd->rcvtidcnt;
 311                 tid = rcd->tidcursor;
 312                 tidoff = 0;
 313         } else if (!subctxt) {
 314                 tidcnt = (dd->rcvtidcnt / rcd->subctxt_cnt) +
 315                          (dd->rcvtidcnt % rcd->subctxt_cnt);
 316                 tidoff = dd->rcvtidcnt - tidcnt;
 317                 ctxttid += tidoff;
 318                 tid = tidcursor_fp(fp);
 319         } else {
 320                 tidcnt = dd->rcvtidcnt / rcd->subctxt_cnt;
 321                 tidoff = tidcnt * (subctxt - 1);
 322                 ctxttid += tidoff;
 323                 tid = tidcursor_fp(fp);
 324         }
 325         if (cnt > tidcnt) {
 326                 /* make sure it all fits in tid_pg_list */
 327                 qib_devinfo(dd->pcidev,
 328                         "Process tried to allocate %u TIDs, only trying max (%u)\n",
 329                         cnt, tidcnt);
 330                 cnt = tidcnt;
 331         }
 332         pagep = (struct page **) rcd->tid_pg_list;
 333         tidlist = (u16 *) &pagep[dd->rcvtidcnt];
 334         pagep += tidoff;
 335         tidlist += tidoff;
 336
 337         memset(tidmap, 0, sizeof(tidmap));
 338         /* before decrement; chip actual # */
 339         ntids = tidcnt;
 340         tidbase = (u64 __iomem *) (((char __iomem *) dd->kregbase) +
 341                                    dd->rcvtidbase +
 342                                    ctxttid * sizeof(*tidbase));
 343
 344         /* virtual address of first page in transfer */
 345         vaddr = ti->tidvaddr;
 346         if (!access_ok(VERIFY_WRITE, (void __user *) vaddr,
 347                        cnt * PAGE_SIZE)) {
 348                 ret = -EFAULT;
 349                 goto done;
 350         }
 351         ret = qib_get_user_pages(vaddr, cnt, pagep);
 352         if (ret) {
 353                 /*
 354                  * if (ret == -EBUSY)
 355                  * We can't continue because the pagep array won't be
 356                  * initialized. This should never happen,
 357                  * unless perhaps the user has mpin'ed the pages
 358                  * themselves.
 359                  */
 360                 qib_devinfo(
 361                         dd->pcidev,
 362                         "Failed to lock addr %p, %u pages: errno %d\n",
 363                         (void *) vaddr, cnt, -ret);
 364                 goto done;
 365         }
 366         for (i = 0; i < cnt; i++, vaddr += PAGE_SIZE) {
 367                 dma_addr_t daddr;
 368
 369                 for (; ntids--; tid++) {
 370                         if (tid == tidcnt)
 371                                 tid = 0;
 372                         if (!dd->pageshadow[ctxttid + tid])
 373                                 break;
 374                 }
 375                 if (ntids < 0) {
 376                         /*
 377                          * Oops, wrapped all the way through their TIDs,
 378                          * and didn't have enough free; see comments at
 379                          * start of routine
 380                          */
 381                         i--;    /* last tidlist[i] not filled in */
 382                         ret = -ENOMEM;
 383                         break;
 384                 }
 385                 ret = qib_map_page(dd->pcidev, pagep[i], &daddr);
 386                 if (ret)
 387                         break;
 388
 389                 tidlist[i] = tid + tidoff;
 390                 /* we "know" system pages and TID pages are same size */
 391                 dd->pageshadow[ctxttid + tid] = pagep[i];
 392                 dd->physshadow[ctxttid + tid] = daddr;
 393                 /*
 394                  * don't need atomic or it's overhead
 395                  */
 396                 __set_bit(tid, tidmap);
 397                 physaddr = dd->physshadow[ctxttid + tid];
 398                 /* PERFORMANCE: below should almost certainly be cached */
 399                 dd->f_put_tid(dd, &tidbase[tid],
 400                                   RCVHQ_RCV_TYPE_EXPECTED, physaddr);
 401                 /*
 402                  * don't check this tid in qib_ctxtshadow, since we
 403                  * just filled it in; start with the next one.
 404                  */
 405                 tid++;
 406         }
 407
 408         if (ret) {
 409                 u32 limit;
 410 cleanup:
 411                 /* jump here if copy out of updated info failed... */
 412                 /* same code that's in qib_free_tid() */
 413                 limit = sizeof(tidmap) * BITS_PER_BYTE;
 414                 if (limit > tidcnt)
 415                         /* just in case size changes in future */
 416                         limit = tidcnt;
 417                 tid = find_first_bit((const unsigned long *)tidmap, limit);
 418                 for (; tid < limit; tid++) {
 419                         if (!test_bit(tid, tidmap))
 420                                 continue;
 421                         if (dd->pageshadow[ctxttid + tid]) {
 422                                 dma_addr_t phys;
 423
 424                                 phys = dd->physshadow[ctxttid + tid];
 425                                 dd->physshadow[ctxttid + tid] = dd->tidinvalid;
 426                                 /* PERFORMANCE: below should almost certainly
 427                                  * be cached
 428                                  */
 429                                 dd->f_put_tid(dd, &tidbase[tid],
 430                                               RCVHQ_RCV_TYPE_EXPECTED,
 431                                               dd->tidinvalid);
 432                                 pci_unmap_page(dd->pcidev, phys, PAGE_SIZE,
 433                                                PCI_DMA_FROMDEVICE);
 434                                 dd->pageshadow[ctxttid + tid] = NULL;
 435                         }
 436                 }
 437                 qib_release_user_pages(pagep, cnt);
 438         } else {
 439                 /*
 440                  * Copy the updated array, with qib_tid's filled in, back
 441                  * to user.  Since we did the copy in already, this "should
 442                  * never fail" If it does, we have to clean up...
 443                  */
 444                 if (copy_to_user((void __user *)
 445                                  (unsigned long) ti->tidlist,
 446                                  tidlist, cnt * sizeof(*tidlist))) {
 447                         ret = -EFAULT;
 448                         goto cleanup;
 449                 }
 450                 if (copy_to_user((void __user *) (unsigned long) ti->tidmap,
 451                                  tidmap, sizeof(tidmap))) {
 452                         ret = -EFAULT;
 453                         goto cleanup;
 454                 }
 455                 if (tid == tidcnt)
 456                         tid = 0;
 457                 if (!rcd->subctxt_cnt)
 458                         rcd->tidcursor = tid;
 459                 else
 460                         tidcursor_fp(fp) = tid;
 461         }
 462
 463 done:
 464         return ret;
 465 }
 466
 467 /**
 468  * qib_tid_free - free a context TID
 469  * @rcd: the context
 470  * @subctxt: the subcontext
 471  * @ti: the TID info
 472  *
 473  * right now we are unlocking one page at a time, but since
 474  * the intended use of this routine is for a single group of
 475  * virtually contiguous pages, that should change to improve
 476  * performance.  We check that the TID is in range for this context
 477  * but otherwise don't check validity; if user has an error and
 478  * frees the wrong tid, it's only their own data that can thereby
 479  * be corrupted.  We do check that the TID was in use, for sanity
 480  * We always use our idea of the saved address, not the address that
 481  * they pass in to us.
 482  */
 483 static int qib_tid_free(struct qib_ctxtdata *rcd, unsigned subctxt,
 484                         const struct qib_tid_info *ti)
 485 {
 486         int ret = 0;
 487         u32 tid, ctxttid, cnt, limit, tidcnt;
 488         struct qib_devdata *dd = rcd->dd;
 489         u64 __iomem *tidbase;
 490         unsigned long tidmap[8];
 491
 492         if (!dd->pageshadow) {
 493                 ret = -ENOMEM;
 494                 goto done;
 495         }
 496
 497         if (copy_from_user(tidmap, (void __user *)(unsigned long)ti->tidmap,
 498                            sizeof(tidmap))) {
 499                 ret = -EFAULT;
 500                 goto done;
 501         }
 502
 503         ctxttid = rcd->ctxt * dd->rcvtidcnt;
 504         if (!rcd->subctxt_cnt)
 505                 tidcnt = dd->rcvtidcnt;
 506         else if (!subctxt) {
 507                 tidcnt = (dd->rcvtidcnt / rcd->subctxt_cnt) +
 508                          (dd->rcvtidcnt % rcd->subctxt_cnt);
 509                 ctxttid += dd->rcvtidcnt - tidcnt;
 510         } else {
 511                 tidcnt = dd->rcvtidcnt / rcd->subctxt_cnt;
 512                 ctxttid += tidcnt * (subctxt - 1);
 513         }
 514         tidbase = (u64 __iomem *) ((char __iomem *)(dd->kregbase) +
 515                                    dd->rcvtidbase +
 516                                    ctxttid * sizeof(*tidbase));
 517
 518         limit = sizeof(tidmap) * BITS_PER_BYTE;
 519         if (limit > tidcnt)
 520                 /* just in case size changes in future */
 521                 limit = tidcnt;
 522         tid = find_first_bit(tidmap, limit);
 523         for (cnt = 0; tid < limit; tid++) {
 524                 /*
 525                  * small optimization; if we detect a run of 3 or so without
 526                  * any set, use find_first_bit again.  That's mainly to
 527                  * accelerate the case where we wrapped, so we have some at
 528                  * the beginning, and some at the end, and a big gap
 529                  * in the middle.
 530                  */
 531                 if (!test_bit(tid, tidmap))
 532                         continue;
 533                 cnt++;
 534                 if (dd->pageshadow[ctxttid + tid]) {
 535                         struct page *p;
 536                         dma_addr_t phys;
 537
 538                         p = dd->pageshadow[ctxttid + tid];
 539                         dd->pageshadow[ctxttid + tid] = NULL;
 540                         phys = dd->physshadow[ctxttid + tid];
 541                         dd->physshadow[ctxttid + tid] = dd->tidinvalid;
 542                         /* PERFORMANCE: below should almost certainly be
 543                          * cached
 544                          */
 545                         dd->f_put_tid(dd, &tidbase[tid],
 546                                       RCVHQ_RCV_TYPE_EXPECTED, dd->tidinvalid);
 547                         pci_unmap_page(dd->pcidev, phys, PAGE_SIZE,
 548                                        PCI_DMA_FROMDEVICE);
 549                         qib_release_user_pages(&p, 1);
 550                 }
 551         }
 552 done:
 553         return ret;
 554 }
 555
 556 /**
 557  * qib_set_part_key - set a partition key
 558  * @rcd: the context
 559  * @key: the key
 560  *
 561  * We can have up to 4 active at a time (other than the default, which is
 562  * always allowed).  This is somewhat tricky, since multiple contexts may set
 563  * the same key, so we reference count them, and clean up at exit.  All 4
 564  * partition keys are packed into a single qlogic_ib register.  It's an
 565  * error for a process to set the same pkey multiple times.  We provide no
 566  * mechanism to de-allocate a pkey at this time, we may eventually need to
 567  * do that.  I've used the atomic operations, and no locking, and only make
 568  * a single pass through what's available.  This should be more than
 569  * adequate for some time. I'll think about spinlocks or the like if and as
 570  * it's necessary.
 571  */
 572 static int qib_set_part_key(struct qib_ctxtdata *rcd, u16 key)
 573 {
 574         struct qib_pportdata *ppd = rcd->ppd;
 575         int i, any = 0, pidx = -1;
 576         u16 lkey = key & 0x7FFF;
 577         int ret;
 578
 579         if (lkey == (QIB_DEFAULT_P_KEY & 0x7FFF)) {
 580                 /* nothing to do; this key always valid */
 581                 ret = 0;
 582                 goto bail;
 583         }
 584
 585         if (!lkey) {
 586                 ret = -EINVAL;
 587                 goto bail;
 588         }
 589
 590         /*
 591          * Set the full membership bit, because it has to be
 592          * set in the register or the packet, and it seems
 593          * cleaner to set in the register than to force all
 594          * callers to set it.
 595          */
 596         key |= 0x8000;
 597
 598         for (i = 0; i < ARRAY_SIZE(rcd->pkeys); i++) {
 599                 if (!rcd->pkeys[i] && pidx == -1)
 600                         pidx = i;
 601                 if (rcd->pkeys[i] == key) {
 602                         ret = -EEXIST;
 603                         goto bail;
 604                 }
 605         }
 606         if (pidx == -1) {
 607                 ret = -EBUSY;
 608                 goto bail;
 609         }
 610         for (any = i = 0; i < ARRAY_SIZE(ppd->pkeys); i++) {
 611                 if (!ppd->pkeys[i]) {
 612                         any++;
 613                         continue;
 614                 }
 615                 if (ppd->pkeys[i] == key) {
 616                         atomic_t *pkrefs = &ppd->pkeyrefs[i];
 617
 618                         if (atomic_inc_return(pkrefs) > 1) {
 619                                 rcd->pkeys[pidx] = key;
 620                                 ret = 0;
 621                                 goto bail;
 622                         } else {
 623                                 /*
 624                                  * lost race, decrement count, catch below
 625                                  */
 626                                 atomic_dec(pkrefs);
 627                                 any++;
 628                         }
 629                 }
 630                 if ((ppd->pkeys[i] & 0x7FFF) == lkey) {
 631                         /*
 632                          * It makes no sense to have both the limited and
 633                          * full membership PKEY set at the same time since
 634                          * the unlimited one will disable the limited one.
 635                          */
 636                         ret = -EEXIST;
 637                         goto bail;
 638                 }
 639         }
 640         if (!any) {
 641                 ret = -EBUSY;
 642                 goto bail;
 643         }
 644         for (any = i = 0; i < ARRAY_SIZE(ppd->pkeys); i++) {
 645                 if (!ppd->pkeys[i] &&
 646                     atomic_inc_return(&ppd->pkeyrefs[i]) == 1) {
 647                         rcd->pkeys[pidx] = key;
 648                         ppd->pkeys[i] = key;
 649                         (void) ppd->dd->f_set_ib_cfg(ppd, QIB_IB_CFG_PKEYS, 0);
 650                         ret = 0;
 651                         goto bail;
 652                 }
 653         }
 654         ret = -EBUSY;
 655
 656 bail:
 657         return ret;
 658 }
 659
 660 /**
 661  * qib_manage_rcvq - manage a context's receive queue
 662  * @rcd: the context
 663  * @subctxt: the subcontext
 664  * @start_stop: action to carry out
 665  *
 666  * start_stop == 0 disables receive on the context, for use in queue
 667  * overflow conditions.  start_stop==1 re-enables, to be used to
 668  * re-init the software copy of the head register
 669  */
 670 static int qib_manage_rcvq(struct qib_ctxtdata *rcd, unsigned subctxt,
 671                            int start_stop)
 672 {
 673         struct qib_devdata *dd = rcd->dd;
 674         unsigned int rcvctrl_op;
 675
 676         if (subctxt)
 677                 goto bail;
 678         /* atomically clear receive enable ctxt. */
 679         if (start_stop) {
 680                 /*
 681                  * On enable, force in-memory copy of the tail register to
 682                  * 0, so that protocol code doesn't have to worry about
 683                  * whether or not the chip has yet updated the in-memory
 684                  * copy or not on return from the system call. The chip
 685                  * always resets it's tail register back to 0 on a
 686                  * transition from disabled to enabled.
 687                  */
 688                 if (rcd->rcvhdrtail_kvaddr)
 689                         qib_clear_rcvhdrtail(rcd);
 690                 rcvctrl_op = QIB_RCVCTRL_CTXT_ENB;
 691         } else
 692                 rcvctrl_op = QIB_RCVCTRL_CTXT_DIS;
 693         dd->f_rcvctrl(rcd->ppd, rcvctrl_op, rcd->ctxt);
 694         /* always; new head should be equal to new tail; see above */
 695 bail:
 696         return 0;
 697 }
 698
 699 static void qib_clean_part_key(struct qib_ctxtdata *rcd,
 700                                struct qib_devdata *dd)
 701 {
 702         int i, j, pchanged = 0;
 703         u64 oldpkey;
 704         struct qib_pportdata *ppd = rcd->ppd;
 705
 706         /* for debugging only */
 707         oldpkey = (u64) ppd->pkeys[0] |
 708                 ((u64) ppd->pkeys[1] << 16) |
 709                 ((u64) ppd->pkeys[2] << 32) |
 710                 ((u64) ppd->pkeys[3] << 48);
 711
 712         for (i = 0; i < ARRAY_SIZE(rcd->pkeys); i++) {
 713                 if (!rcd->pkeys[i])
 714                         continue;
 715                 for (j = 0; j < ARRAY_SIZE(ppd->pkeys); j++) {
 716                         /* check for match independent of the global bit */
 717                         if ((ppd->pkeys[j] & 0x7fff) !=
 718                             (rcd->pkeys[i] & 0x7fff))
 719                                 continue;
 720                         if (atomic_dec_and_test(&ppd->pkeyrefs[j])) {
 721                                 ppd->pkeys[j] = 0;
 722                                 pchanged++;
 723                         }
 724                         break;
 725                 }
 726                 rcd->pkeys[i] = 0;
 727         }
 728         if (pchanged)
 729                 (void) ppd->dd->f_set_ib_cfg(ppd, QIB_IB_CFG_PKEYS, 0);
 730 }
 731
 732 /* common code for the mappings on dma_alloc_coherent mem */
 733 static int qib_mmap_mem(struct vm_area_struct *vma, struct qib_ctxtdata *rcd,
 734                         unsigned len, void *kvaddr, u32 write_ok, char *what)
 735 {
 736         struct qib_devdata *dd = rcd->dd;
 737         unsigned long pfn;
 738         int ret;
 739
 740         if ((vma->vm_end - vma->vm_start) > len) {
 741                 qib_devinfo(dd->pcidev,
 742                          "FAIL on %s: len %lx > %x\n", what,
 743                          vma->vm_end - vma->vm_start, len);
 744                 ret = -EFAULT;
 745                 goto bail;
 746         }
 747
 748         /*
 749          * shared context user code requires rcvhdrq mapped r/w, others
 750          * only allowed readonly mapping.
 751          */
 752         if (!write_ok) {
 753                 if (vma->vm_flags & VM_WRITE) {
 754                         qib_devinfo(dd->pcidev,
 755                                  "%s must be mapped readonly\n", what);
 756                         ret = -EPERM;
 757                         goto bail;
 758                 }
 759
 760                 /* don't allow them to later change with mprotect */
 761                 vma->vm_flags &= ~VM_MAYWRITE;
 762         }
 763
 764         pfn = virt_to_phys(kvaddr) >> PAGE_SHIFT;
 765         ret = remap_pfn_range(vma, vma->vm_start, pfn,
 766                               len, vma->vm_page_prot);
 767         if (ret)
 768                 qib_devinfo(dd->pcidev,
 769                         "%s ctxt%u mmap of %lx, %x bytes failed: %d\n",
 770                         what, rcd->ctxt, pfn, len, ret);
 771 bail:
 772         return ret;
 773 }
 774
 775 static int mmap_ureg(struct vm_area_struct *vma, struct qib_devdata *dd,
 776                      u64 ureg)
 777 {
 778         unsigned long phys;
 779         unsigned long sz;
 780         int ret;
 781
 782         /*
 783          * This is real hardware, so use io_remap.  This is the mechanism
 784          * for the user process to update the head registers for their ctxt
 785          * in the chip.
 786          */
 787         sz = dd->flags & QIB_HAS_HDRSUPP ? 2 * PAGE_SIZE : PAGE_SIZE;
 788         if ((vma->vm_end - vma->vm_start) > sz) {
 789                 qib_devinfo(dd->pcidev,
 790                         "FAIL mmap userreg: reqlen %lx > PAGE\n",
 791                         vma->vm_end - vma->vm_start);
 792                 ret = -EFAULT;
 793         } else {
 794                 phys = dd->physaddr + ureg;
 795                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 796
 797                 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
 798                 ret = io_remap_pfn_range(vma, vma->vm_start,
 799                                          phys >> PAGE_SHIFT,
 800                                          vma->vm_end - vma->vm_start,
 801                                          vma->vm_page_prot);
 802         }
 803         return ret;
 804 }
 805
 806 static int mmap_piobufs(struct vm_area_struct *vma,
 807                         struct qib_devdata *dd,
 808                         struct qib_ctxtdata *rcd,
 809                         unsigned piobufs, unsigned piocnt)
 810 {
 811         unsigned long phys;
 812         int ret;
 813
 814         /*
 815          * When we map the PIO buffers in the chip, we want to map them as
 816          * writeonly, no read possible; unfortunately, x86 doesn't allow
 817          * for this in hardware, but we still prevent users from asking
 818          * for it.
 819          */
 820         if ((vma->vm_end - vma->vm_start) > (piocnt * dd->palign)) {
 821                 qib_devinfo(dd->pcidev,
 822                         "FAIL mmap piobufs: reqlen %lx > PAGE\n",
 823                          vma->vm_end - vma->vm_start);
 824                 ret = -EINVAL;
 825                 goto bail;
 826         }
 827
 828         phys = dd->physaddr + piobufs;
 829
 830 #if defined(__powerpc__)
 831         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 832 #endif
 833
 834         /*
 835          * don't allow them to later change to readable with mprotect (for when
 836          * not initially mapped readable, as is normally the case)
 837          */
 838         vma->vm_flags &= ~VM_MAYREAD;
 839         vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
 840
 841         /* We used PAT if wc_cookie == 0 */
 842         if (!dd->wc_cookie)
 843                 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
 844
 845         ret = io_remap_pfn_range(vma, vma->vm_start, phys >> PAGE_SHIFT,
 846                                  vma->vm_end - vma->vm_start,
 847                                  vma->vm_page_prot);
 848 bail:
 849         return ret;
 850 }
 851
 852 static int mmap_rcvegrbufs(struct vm_area_struct *vma,
 853                            struct qib_ctxtdata *rcd)
 854 {
 855         struct qib_devdata *dd = rcd->dd;
 856         unsigned long start, size;
 857         size_t total_size, i;
 858         unsigned long pfn;
 859         int ret;
 860
 861         size = rcd->rcvegrbuf_size;
 862         total_size = rcd->rcvegrbuf_chunks * size;
 863         if ((vma->vm_end - vma->vm_start) > total_size) {
 864                 qib_devinfo(dd->pcidev,
 865                         "FAIL on egr bufs: reqlen %lx > actual %lx\n",
 866                          vma->vm_end - vma->vm_start,
 867                          (unsigned long) total_size);
 868                 ret = -EINVAL;
 869                 goto bail;
 870         }
 871
 872         if (vma->vm_flags & VM_WRITE) {
 873                 qib_devinfo(dd->pcidev,
 874                         "Can't map eager buffers as writable (flags=%lx)\n",
 875                         vma->vm_flags);
 876                 ret = -EPERM;
 877                 goto bail;
 878         }
 879         /* don't allow them to later change to writeable with mprotect */
 880         vma->vm_flags &= ~VM_MAYWRITE;
 881
 882         start = vma->vm_start;
 883
 884         for (i = 0; i < rcd->rcvegrbuf_chunks; i++, start += size) {
 885                 pfn = virt_to_phys(rcd->rcvegrbuf[i]) >> PAGE_SHIFT;
 886                 ret = remap_pfn_range(vma, start, pfn, size,
 887                                       vma->vm_page_prot);
 888                 if (ret < 0)
 889                         goto bail;
 890         }
 891         ret = 0;
 892
 893 bail:
 894         return ret;
 895 }
 896
 897 /*
 898  * qib_file_vma_fault - handle a VMA page fault.
 899  */
 900 static int qib_file_vma_fault(struct vm_fault *vmf)
 901 {
 902         struct page *page;
 903
 904         page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
 905         if (!page)
 906                 return VM_FAULT_SIGBUS;
 907
 908         get_page(page);
 909         vmf->page = page;
 910
 911         return 0;
 912 }
 913
 914 static const struct vm_operations_struct qib_file_vm_ops = {
 915         .fault = qib_file_vma_fault,
 916 };
 917
 918 static int mmap_kvaddr(struct vm_area_struct *vma, u64 pgaddr,
 919                        struct qib_ctxtdata *rcd, unsigned subctxt)
 920 {
 921         struct qib_devdata *dd = rcd->dd;
 922         unsigned subctxt_cnt;
 923         unsigned long len;
 924         void *addr;
 925         size_t size;
 926         int ret = 0;
 927
 928         subctxt_cnt = rcd->subctxt_cnt;
 929         size = rcd->rcvegrbuf_chunks * rcd->rcvegrbuf_size;
 930
 931         /*
 932          * Each process has all the subctxt uregbase, rcvhdrq, and
 933          * rcvegrbufs mmapped - as an array for all the processes,
 934          * and also separately for this process.
 935          */
 936         if (pgaddr == cvt_kvaddr(rcd->subctxt_uregbase)) {
 937                 addr = rcd->subctxt_uregbase;
 938                 size = PAGE_SIZE * subctxt_cnt;
 939         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_rcvhdr_base)) {
 940                 addr = rcd->subctxt_rcvhdr_base;
 941                 size = rcd->rcvhdrq_size * subctxt_cnt;
 942         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_rcvegrbuf)) {
 943                 addr = rcd->subctxt_rcvegrbuf;
 944                 size *= subctxt_cnt;
 945         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_uregbase +
 946                                         PAGE_SIZE * subctxt)) {
 947                 addr = rcd->subctxt_uregbase + PAGE_SIZE * subctxt;
 948                 size = PAGE_SIZE;
 949         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_rcvhdr_base +
 950                                         rcd->rcvhdrq_size * subctxt)) {
 951                 addr = rcd->subctxt_rcvhdr_base +
 952                         rcd->rcvhdrq_size * subctxt;
 953                 size = rcd->rcvhdrq_size;
 954         } else if (pgaddr == cvt_kvaddr(&rcd->user_event_mask[subctxt])) {
 955                 addr = rcd->user_event_mask;
 956                 size = PAGE_SIZE;
 957         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_rcvegrbuf +
 958                                         size * subctxt)) {
 959                 addr = rcd->subctxt_rcvegrbuf + size * subctxt;
 960                 /* rcvegrbufs are read-only on the slave */
 961                 if (vma->vm_flags & VM_WRITE) {
 962                         qib_devinfo(dd->pcidev,
 963                                  "Can't map eager buffers as writable (flags=%lx)\n",
 964                                  vma->vm_flags);
 965                         ret = -EPERM;
 966                         goto bail;
 967                 }
 968                 /*
 969                  * Don't allow permission to later change to writeable
 970                  * with mprotect.
 971                  */
 972                 vma->vm_flags &= ~VM_MAYWRITE;
 973         } else
 974                 goto bail;
 975         len = vma->vm_end - vma->vm_start;
 976         if (len > size) {
 977                 ret = -EINVAL;
 978                 goto bail;
 979         }
 980
 981         vma->vm_pgoff = (unsigned long) addr >> PAGE_SHIFT;
 982         vma->vm_ops = &qib_file_vm_ops;
 983         vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
 984         ret = 1;
 985
 986 bail:
 987         return ret;
 988 }
 989
 990 /**
 991  * qib_mmapf - mmap various structures into user space
 992  * @fp: the file pointer
 993  * @vma: the VM area
 994  *
 995  * We use this to have a shared buffer between the kernel and the user code
 996  * for the rcvhdr queue, egr buffers, and the per-context user regs and pio
 997  * buffers in the chip.  We have the open and close entries so we can bump
 998  * the ref count and keep the driver from being unloaded while still mapped.
 999  */
1000 static int qib_mmapf(struct file *fp, struct vm_area_struct *vma)
1001 {
1002         struct qib_ctxtdata *rcd;
1003         struct qib_devdata *dd;
1004         u64 pgaddr, ureg;
1005         unsigned piobufs, piocnt;
1006         int ret, match = 1;
1007
1008         rcd = ctxt_fp(fp);
1009         if (!rcd || !(vma->vm_flags & VM_SHARED)) {
1010                 ret = -EINVAL;
1011                 goto bail;
1012         }
1013         dd = rcd->dd;
1014
1015         /*
1016          * This is the qib_do_user_init() code, mapping the shared buffers
1017          * and per-context user registers into the user process. The address
1018          * referred to by vm_pgoff is the file offset passed via mmap().
1019          * For shared contexts, this is the kernel vmalloc() address of the
1020          * pages to share with the master.
1021          * For non-shared or master ctxts, this is a physical address.
1022          * We only do one mmap for each space mapped.
1023          */
1024         pgaddr = vma->vm_pgoff << PAGE_SHIFT;
1025
1026         /*
1027          * Check for 0 in case one of the allocations failed, but user
1028          * called mmap anyway.
1029          */
1030         if (!pgaddr)  {
1031                 ret = -EINVAL;
1032                 goto bail;
1033         }
1034
1035         /*
1036          * Physical addresses must fit in 40 bits for our hardware.
1037          * Check for kernel virtual addresses first, anything else must
1038          * match a HW or memory address.
1039          */
1040         ret = mmap_kvaddr(vma, pgaddr, rcd, subctxt_fp(fp));
1041         if (ret) {
1042                 if (ret > 0)
1043                         ret = 0;
1044                 goto bail;
1045         }
1046
1047         ureg = dd->uregbase + dd->ureg_align * rcd->ctxt;
1048         if (!rcd->subctxt_cnt) {
1049                 /* ctxt is not shared */
1050                 piocnt = rcd->piocnt;
1051                 piobufs = rcd->piobufs;
1052         } else if (!subctxt_fp(fp)) {
1053                 /* caller is the master */
1054                 piocnt = (rcd->piocnt / rcd->subctxt_cnt) +
1055                          (rcd->piocnt % rcd->subctxt_cnt);
1056                 piobufs = rcd->piobufs +
1057                         dd->palign * (rcd->piocnt - piocnt);
1058         } else {
1059                 unsigned slave = subctxt_fp(fp) - 1;
1060
1061                 /* caller is a slave */
1062                 piocnt = rcd->piocnt / rcd->subctxt_cnt;
1063                 piobufs = rcd->piobufs + dd->palign * piocnt * slave;
1064         }
1065
1066         if (pgaddr == ureg)
1067                 ret = mmap_ureg(vma, dd, ureg);
1068         else if (pgaddr == piobufs)
1069                 ret = mmap_piobufs(vma, dd, rcd, piobufs, piocnt);
1070         else if (pgaddr == dd->pioavailregs_phys)
1071                 /* in-memory copy of pioavail registers */
1072                 ret = qib_mmap_mem(vma, rcd, PAGE_SIZE,
1073                                    (void *) dd->pioavailregs_dma, 0,
1074                                    "pioavail registers");
1075         else if (pgaddr == rcd->rcvegr_phys)
1076                 ret = mmap_rcvegrbufs(vma, rcd);
1077         else if (pgaddr == (u64) rcd->rcvhdrq_phys)
1078                 /*
1079                  * The rcvhdrq itself; multiple pages, contiguous
1080                  * from an i/o perspective.  Shared contexts need
1081                  * to map r/w, so we allow writing.
1082                  */
1083                 ret = qib_mmap_mem(vma, rcd, rcd->rcvhdrq_size,
1084                                    rcd->rcvhdrq, 1, "rcvhdrq");
1085         else if (pgaddr == (u64) rcd->rcvhdrqtailaddr_phys)
1086                 /* in-memory copy of rcvhdrq tail register */
1087                 ret = qib_mmap_mem(vma, rcd, PAGE_SIZE,
1088                                    rcd->rcvhdrtail_kvaddr, 0,
1089                                    "rcvhdrq tail");
1090         else
1091                 match = 0;
1092         if (!match)
1093                 ret = -EINVAL;
1094
1095         vma->vm_private_data = NULL;
1096
1097         if (ret < 0)
1098                 qib_devinfo(dd->pcidev,
1099                          "mmap Failure %d: off %llx len %lx\n",
1100                          -ret, (unsigned long long)pgaddr,
1101                          vma->vm_end - vma->vm_start);
1102 bail:
1103         return ret;
1104 }
1105
1106 static unsigned int qib_poll_urgent(struct qib_ctxtdata *rcd,
1107                                     struct file *fp,
1108                                     struct poll_table_struct *pt)
1109 {
1110         struct qib_devdata *dd = rcd->dd;
1111         unsigned pollflag;
1112
1113         poll_wait(fp, &rcd->wait, pt);
1114
1115         spin_lock_irq(&dd->uctxt_lock);
1116         if (rcd->urgent != rcd->urgent_poll) {
1117                 pollflag = POLLIN | POLLRDNORM;
1118                 rcd->urgent_poll = rcd->urgent;
1119         } else {
1120                 pollflag = 0;
1121                 set_bit(QIB_CTXT_WAITING_URG, &rcd->flag);
1122         }
1123         spin_unlock_irq(&dd->uctxt_lock);
1124
1125         return pollflag;
1126 }
1127
1128 static unsigned int qib_poll_next(struct qib_ctxtdata *rcd,
1129                                   struct file *fp,
1130                                   struct poll_table_struct *pt)
1131 {
1132         struct qib_devdata *dd = rcd->dd;
1133         unsigned pollflag;
1134
1135         poll_wait(fp, &rcd->wait, pt);
1136
1137         spin_lock_irq(&dd->uctxt_lock);
1138         if (dd->f_hdrqempty(rcd)) {
1139                 set_bit(QIB_CTXT_WAITING_RCV, &rcd->flag);
1140                 dd->f_rcvctrl(rcd->ppd, QIB_RCVCTRL_INTRAVAIL_ENB, rcd->ctxt);
1141                 pollflag = 0;
1142         } else
1143                 pollflag = POLLIN | POLLRDNORM;
1144         spin_unlock_irq(&dd->uctxt_lock);
1145
1146         return pollflag;
1147 }
1148
1149 static unsigned int qib_poll(struct file *fp, struct poll_table_struct *pt)
1150 {
1151         struct qib_ctxtdata *rcd;
1152         unsigned pollflag;
1153
1154         rcd = ctxt_fp(fp);
1155         if (!rcd)
1156                 pollflag = POLLERR;
1157         else if (rcd->poll_type == QIB_POLL_TYPE_URGENT)
1158                 pollflag = qib_poll_urgent(rcd, fp, pt);
1159         else  if (rcd->poll_type == QIB_POLL_TYPE_ANYRCV)
1160                 pollflag = qib_poll_next(rcd, fp, pt);
1161         else /* invalid */
1162                 pollflag = POLLERR;
1163
1164         return pollflag;
1165 }
1166
1167 static void assign_ctxt_affinity(struct file *fp, struct qib_devdata *dd)
1168 {
1169         struct qib_filedata *fd = fp->private_data;
1170         const unsigned int weight = cpumask_weight(&current->cpus_allowed);
1171         const struct cpumask *local_mask = cpumask_of_pcibus(dd->pcidev->bus);
1172         int local_cpu;
1173
1174         /*
1175          * If process has NOT already set it's affinity, select and
1176          * reserve a processor for it on the local NUMA node.
1177          */
1178         if ((weight >= qib_cpulist_count) &&
1179                 (cpumask_weight(local_mask) <= qib_cpulist_count)) {
1180                 for_each_cpu(local_cpu, local_mask)
1181                         if (!test_and_set_bit(local_cpu, qib_cpulist)) {
1182                                 fd->rec_cpu_num = local_cpu;
1183                                 return;
1184                         }
1185         }
1186
1187         /*
1188          * If process has NOT already set it's affinity, select and
1189          * reserve a processor for it, as a rendevous for all
1190          * users of the driver.  If they don't actually later
1191          * set affinity to this cpu, or set it to some other cpu,
1192          * it just means that sooner or later we don't recommend
1193          * a cpu, and let the scheduler do it's best.
1194          */
1195         if (weight >= qib_cpulist_count) {
1196                 int cpu;
1197
1198                 cpu = find_first_zero_bit(qib_cpulist,
1199                                           qib_cpulist_count);
1200                 if (cpu == qib_cpulist_count)
1201                         qib_dev_err(dd,
1202                         "no cpus avail for affinity PID %u\n",
1203                         current->pid);
1204                 else {
1205                         __set_bit(cpu, qib_cpulist);
1206                         fd->rec_cpu_num = cpu;
1207                 }
1208         }
1209 }
1210
1211 /*
1212  * Check that userland and driver are compatible for subcontexts.
1213  */
1214 static int qib_compatible_subctxts(int user_swmajor, int user_swminor)
1215 {
1216         /* this code is written long-hand for clarity */
1217         if (QIB_USER_SWMAJOR != user_swmajor) {
1218                 /* no promise of compatibility if major mismatch */
1219                 return 0;
1220         }
1221         if (QIB_USER_SWMAJOR == 1) {
1222                 switch (QIB_USER_SWMINOR) {
1223                 case 0:
1224                 case 1:
1225                 case 2:
1226                         /* no subctxt implementation so cannot be compatible */
1227                         return 0;
1228                 case 3:
1229                         /* 3 is only compatible with itself */
1230                         return user_swminor == 3;
1231                 default:
1232                         /* >= 4 are compatible (or are expected to be) */
1233                         return user_swminor <= QIB_USER_SWMINOR;
1234                 }
1235         }
1236         /* make no promises yet for future major versions */
1237         return 0;
1238 }
1239
1240 static int init_subctxts(struct qib_devdata *dd,
1241                          struct qib_ctxtdata *rcd,
1242                          const struct qib_user_info *uinfo)
1243 {
1244         int ret = 0;
1245         unsigned num_subctxts;
1246         size_t size;
1247
1248         /*
1249          * If the user is requesting zero subctxts,
1250          * skip the subctxt allocation.
1251          */
1252         if (uinfo->spu_subctxt_cnt <= 0)
1253                 goto bail;
1254         num_subctxts = uinfo->spu_subctxt_cnt;
1255
1256         /* Check for subctxt compatibility */
1257         if (!qib_compatible_subctxts(uinfo->spu_userversion >> 16,
1258                 uinfo->spu_userversion & 0xffff)) {
1259                 qib_devinfo(dd->pcidev,
1260                          "Mismatched user version (%d.%d) and driver version (%d.%d) while context sharing. Ensure that driver and library are from the same release.\n",
1261                          (int) (uinfo->spu_userversion >> 16),
1262                          (int) (uinfo->spu_userversion & 0xffff),
1263                          QIB_USER_SWMAJOR, QIB_USER_SWMINOR);
1264                 goto bail;
1265         }
1266         if (num_subctxts > QLOGIC_IB_MAX_SUBCTXT) {
1267                 ret = -EINVAL;
1268                 goto bail;
1269         }
1270
1271         rcd->subctxt_uregbase = vmalloc_user(PAGE_SIZE * num_subctxts);
1272         if (!rcd->subctxt_uregbase) {
1273                 ret = -ENOMEM;
1274                 goto bail;
1275         }
1276         /* Note: rcd->rcvhdrq_size isn't initialized yet. */
1277         size = ALIGN(dd->rcvhdrcnt * dd->rcvhdrentsize *
1278                      sizeof(u32), PAGE_SIZE) * num_subctxts;
1279         rcd->subctxt_rcvhdr_base = vmalloc_user(size);
1280         if (!rcd->subctxt_rcvhdr_base) {
1281                 ret = -ENOMEM;
1282                 goto bail_ureg;
1283         }
1284
1285         rcd->subctxt_rcvegrbuf = vmalloc_user(rcd->rcvegrbuf_chunks *
1286                                               rcd->rcvegrbuf_size *
1287                                               num_subctxts);
1288         if (!rcd->subctxt_rcvegrbuf) {
1289                 ret = -ENOMEM;
1290                 goto bail_rhdr;
1291         }
1292
1293         rcd->subctxt_cnt = uinfo->spu_subctxt_cnt;
1294         rcd->subctxt_id = uinfo->spu_subctxt_id;
1295         rcd->active_slaves = 1;
1296         rcd->redirect_seq_cnt = 1;
1297         set_bit(QIB_CTXT_MASTER_UNINIT, &rcd->flag);
1298         goto bail;
1299
1300 bail_rhdr:
1301         vfree(rcd->subctxt_rcvhdr_base);
1302 bail_ureg:
1303         vfree(rcd->subctxt_uregbase);
1304         rcd->subctxt_uregbase = NULL;
1305 bail:
1306         return ret;
1307 }
1308
1309 static int setup_ctxt(struct qib_pportdata *ppd, int ctxt,
1310                       struct file *fp, const struct qib_user_info *uinfo)
1311 {
1312         struct qib_filedata *fd = fp->private_data;
1313         struct qib_devdata *dd = ppd->dd;
1314         struct qib_ctxtdata *rcd;
1315         void *ptmp = NULL;
1316         int ret;
1317         int numa_id;
1318
1319         assign_ctxt_affinity(fp, dd);
1320
1321         numa_id = qib_numa_aware ? ((fd->rec_cpu_num != -1) ?
1322                 cpu_to_node(fd->rec_cpu_num) :
1323                 numa_node_id()) : dd->assigned_node_id;
1324
1325         rcd = qib_create_ctxtdata(ppd, ctxt, numa_id);
1326
1327         /*
1328          * Allocate memory for use in qib_tid_update() at open to
1329          * reduce cost of expected send setup per message segment
1330          */
1331         if (rcd)
1332                 ptmp = kmalloc(dd->rcvtidcnt * sizeof(u16) +
1333                                dd->rcvtidcnt * sizeof(struct page **),
1334                                GFP_KERNEL);
1335
1336         if (!rcd || !ptmp) {
1337                 qib_dev_err(dd,
1338                         "Unable to allocate ctxtdata memory, failing open\n");
1339                 ret = -ENOMEM;
1340                 goto bailerr;
1341         }
1342         rcd->userversion = uinfo->spu_userversion;
1343         ret = init_subctxts(dd, rcd, uinfo);
1344         if (ret)
1345                 goto bailerr;
1346         rcd->tid_pg_list = ptmp;
1347         rcd->pid = current->pid;
1348         init_waitqueue_head(&dd->rcd[ctxt]->wait);
1349         strlcpy(rcd->comm, current->comm, sizeof(rcd->comm));
1350         ctxt_fp(fp) = rcd;
1351         qib_stats.sps_ctxts++;
1352         dd->freectxts--;
1353         ret = 0;
1354         goto bail;
1355
1356 bailerr:
1357         if (fd->rec_cpu_num != -1)
1358                 __clear_bit(fd->rec_cpu_num, qib_cpulist);
1359
1360         dd->rcd[ctxt] = NULL;
1361         kfree(rcd);
1362         kfree(ptmp);
1363 bail:
1364         return ret;
1365 }
1366
1367 static inline int usable(struct qib_pportdata *ppd)
1368 {
1369         struct qib_devdata *dd = ppd->dd;
1370
1371         return dd && (dd->flags & QIB_PRESENT) && dd->kregbase && ppd->lid &&
1372                 (ppd->lflags & QIBL_LINKACTIVE);
1373 }
1374
1375 /*
1376  * Select a context on the given device, either using a requested port
1377  * or the port based on the context number.
1378  */
1379 static int choose_port_ctxt(struct file *fp, struct qib_devdata *dd, u32 port,
1380                             const struct qib_user_info *uinfo)
1381 {
1382         struct qib_pportdata *ppd = NULL;
1383         int ret, ctxt;
1384
1385         if (port) {
1386                 if (!usable(dd->pport + port - 1)) {
1387                         ret = -ENETDOWN;
1388                         goto done;
1389                 } else
1390                         ppd = dd->pport + port - 1;
1391         }
1392         for (ctxt = dd->first_user_ctxt; ctxt < dd->cfgctxts && dd->rcd[ctxt];
1393              ctxt++)
1394                 ;
1395         if (ctxt == dd->cfgctxts) {
1396                 ret = -EBUSY;
1397                 goto done;
1398         }
1399         if (!ppd) {
1400                 u32 pidx = ctxt % dd->num_pports;
1401
1402                 if (usable(dd->pport + pidx))
1403                         ppd = dd->pport + pidx;
1404                 else {
1405                         for (pidx = 0; pidx < dd->num_pports && !ppd;
1406                              pidx++)
1407                                 if (usable(dd->pport + pidx))
1408                                         ppd = dd->pport + pidx;
1409                 }
1410         }
1411         ret = ppd ? setup_ctxt(ppd, ctxt, fp, uinfo) : -ENETDOWN;
1412 done:
1413         return ret;
1414 }
1415
1416 static int find_free_ctxt(int unit, struct file *fp,
1417                           const struct qib_user_info *uinfo)
1418 {
1419         struct qib_devdata *dd = qib_lookup(unit);
1420         int ret;
1421
1422         if (!dd || (uinfo->spu_port && uinfo->spu_port > dd->num_pports))
1423                 ret = -ENODEV;
1424         else
1425                 ret = choose_port_ctxt(fp, dd, uinfo->spu_port, uinfo);
1426
1427         return ret;
1428 }
1429
1430 static int get_a_ctxt(struct file *fp, const struct qib_user_info *uinfo,
1431                       unsigned alg)
1432 {
1433         struct qib_devdata *udd = NULL;
1434         int ret = 0, devmax, npresent, nup, ndev, dusable = 0, i;
1435         u32 port = uinfo->spu_port, ctxt;
1436
1437         devmax = qib_count_units(&npresent, &nup);
1438         if (!npresent) {
1439                 ret = -ENXIO;
1440                 goto done;
1441         }
1442         if (nup == 0) {
1443                 ret = -ENETDOWN;
1444                 goto done;
1445         }
1446
1447         if (alg == QIB_PORT_ALG_ACROSS) {
1448                 unsigned inuse = ~0U;
1449
1450                 /* find device (with ACTIVE ports) with fewest ctxts in use */
1451                 for (ndev = 0; ndev < devmax; ndev++) {
1452                         struct qib_devdata *dd = qib_lookup(ndev);
1453                         unsigned cused = 0, cfree = 0, pusable = 0;
1454
1455                         if (!dd)
1456                                 continue;
1457                         if (port && port <= dd->num_pports &&
1458                             usable(dd->pport + port - 1))
1459                                 pusable = 1;
1460                         else
1461                                 for (i = 0; i < dd->num_pports; i++)
1462                                         if (usable(dd->pport + i))
1463                                                 pusable++;
1464                         if (!pusable)
1465                                 continue;
1466                         for (ctxt = dd->first_user_ctxt; ctxt < dd->cfgctxts;
1467                              ctxt++)
1468                                 if (dd->rcd[ctxt])
1469                                         cused++;
1470                                 else
1471                                         cfree++;
1472                         if (cfree && cused < inuse) {
1473                                 udd = dd;
1474                                 inuse = cused;
1475                         }
1476                 }
1477                 if (udd) {
1478                         ret = choose_port_ctxt(fp, udd, port, uinfo);
1479                         goto done;
1480                 }
1481         } else {
1482                 for (ndev = 0; ndev < devmax; ndev++) {
1483                         struct qib_devdata *dd = qib_lookup(ndev);
1484
1485                         if (dd) {
1486                                 ret = choose_port_ctxt(fp, dd, port, uinfo);
1487                                 if (!ret)
1488                                         goto done;
1489                                 if (ret == -EBUSY)
1490                                         dusable++;
1491                         }
1492                 }
1493         }
1494         ret = dusable ? -EBUSY : -ENETDOWN;
1495
1496 done:
1497         return ret;
1498 }
1499
1500 static int find_shared_ctxt(struct file *fp,
1501                             const struct qib_user_info *uinfo)
1502 {
1503         int devmax, ndev, i;
1504         int ret = 0;
1505
1506         devmax = qib_count_units(NULL, NULL);
1507
1508         for (ndev = 0; ndev < devmax; ndev++) {
1509                 struct qib_devdata *dd = qib_lookup(ndev);
1510
1511                 /* device portion of usable() */
1512                 if (!(dd && (dd->flags & QIB_PRESENT) && dd->kregbase))
1513                         continue;
1514                 for (i = dd->first_user_ctxt; i < dd->cfgctxts; i++) {
1515                         struct qib_ctxtdata *rcd = dd->rcd[i];
1516
1517                         /* Skip ctxts which are not yet open */
1518                         if (!rcd || !rcd->cnt)
1519                                 continue;
1520                         /* Skip ctxt if it doesn't match the requested one */
1521                         if (rcd->subctxt_id != uinfo->spu_subctxt_id)
1522                                 continue;
1523                         /* Verify the sharing process matches the master */
1524                         if (rcd->subctxt_cnt != uinfo->spu_subctxt_cnt ||
1525                             rcd->userversion != uinfo->spu_userversion ||
1526                             rcd->cnt >= rcd->subctxt_cnt) {
1527                                 ret = -EINVAL;
1528                                 goto done;
1529                         }
1530                         ctxt_fp(fp) = rcd;
1531                         subctxt_fp(fp) = rcd->cnt++;
1532                         rcd->subpid[subctxt_fp(fp)] = current->pid;
1533                         tidcursor_fp(fp) = 0;
1534                         rcd->active_slaves |= 1 << subctxt_fp(fp);
1535                         ret = 1;
1536                         goto done;
1537                 }
1538         }
1539
1540 done:
1541         return ret;
1542 }
1543
1544 static int qib_open(struct inode *in, struct file *fp)
1545 {
1546         /* The real work is performed later in qib_assign_ctxt() */
1547         fp->private_data = kzalloc(sizeof(struct qib_filedata), GFP_KERNEL);
1548         if (fp->private_data) /* no cpu affinity by default */
1549                 ((struct qib_filedata *)fp->private_data)->rec_cpu_num = -1;
1550         return fp->private_data ? 0 : -ENOMEM;
1551 }
1552
1553 static int find_hca(unsigned int cpu, int *unit)
1554 {
1555         int ret = 0, devmax, npresent, nup, ndev;
1556
1557         *unit = -1;
1558
1559         devmax = qib_count_units(&npresent, &nup);
1560         if (!npresent) {
1561                 ret = -ENXIO;
1562                 goto done;
1563         }
1564         if (!nup) {
1565                 ret = -ENETDOWN;
1566                 goto done;
1567         }
1568         for (ndev = 0; ndev < devmax; ndev++) {
1569                 struct qib_devdata *dd = qib_lookup(ndev);
1570
1571                 if (dd) {
1572                         if (pcibus_to_node(dd->pcidev->bus) < 0) {
1573                                 ret = -EINVAL;
1574                                 goto done;
1575                         }
1576                         if (cpu_to_node(cpu) ==
1577                                 pcibus_to_node(dd->pcidev->bus)) {
1578                                 *unit = ndev;
1579                                 goto done;
1580                         }
1581                 }
1582         }
1583 done:
1584         return ret;
1585 }
1586
1587 static int do_qib_user_sdma_queue_create(struct file *fp)
1588 {
1589         struct qib_filedata *fd = fp->private_data;
1590         struct qib_ctxtdata *rcd = fd->rcd;
1591         struct qib_devdata *dd = rcd->dd;
1592
1593         if (dd->flags & QIB_HAS_SEND_DMA) {
1594
1595                 fd->pq = qib_user_sdma_queue_create(&dd->pcidev->dev,
1596                                                     dd->unit,
1597                                                     rcd->ctxt,
1598                                                     fd->subctxt);
1599                 if (!fd->pq)
1600                         return -ENOMEM;
1601         }
1602
1603         return 0;
1604 }
1605
1606 /*
1607  * Get ctxt early, so can set affinity prior to memory allocation.
1608  */
1609 static int qib_assign_ctxt(struct file *fp, const struct qib_user_info *uinfo)
1610 {
1611         int ret;
1612         int i_minor;
1613         unsigned swmajor, swminor, alg = QIB_PORT_ALG_ACROSS;
1614
1615         /* Check to be sure we haven't already initialized this file */
1616         if (ctxt_fp(fp)) {
1617                 ret = -EINVAL;
1618                 goto done;
1619         }
1620
1621         /* for now, if major version is different, bail */
1622         swmajor = uinfo->spu_userversion >> 16;
1623         if (swmajor != QIB_USER_SWMAJOR) {
1624                 ret = -ENODEV;
1625                 goto done;
1626         }
1627
1628         swminor = uinfo->spu_userversion & 0xffff;
1629
1630         if (swminor >= 11 && uinfo->spu_port_alg < QIB_PORT_ALG_COUNT)
1631                 alg = uinfo->spu_port_alg;
1632
1633         mutex_lock(&qib_mutex);
1634
1635         if (qib_compatible_subctxts(swmajor, swminor) &&
1636             uinfo->spu_subctxt_cnt) {
1637                 ret = find_shared_ctxt(fp, uinfo);
1638                 if (ret > 0) {
1639                         ret = do_qib_user_sdma_queue_create(fp);
1640                         if (!ret)
1641                                 assign_ctxt_affinity(fp, (ctxt_fp(fp))->dd);
1642                         goto done_ok;
1643                 }
1644         }
1645
1646         i_minor = iminor(file_inode(fp)) - QIB_USER_MINOR_BASE;
1647         if (i_minor)
1648                 ret = find_free_ctxt(i_minor - 1, fp, uinfo);
1649         else {
1650                 int unit;
1651                 const unsigned int cpu = cpumask_first(&current->cpus_allowed);
1652                 const unsigned int weight =
1653                         cpumask_weight(&current->cpus_allowed);
1654
1655                 if (weight == 1 && !test_bit(cpu, qib_cpulist))
1656                         if (!find_hca(cpu, &unit) && unit >= 0)
1657                                 if (!find_free_ctxt(unit, fp, uinfo)) {
1658                                         ret = 0;
1659                                         goto done_chk_sdma;
1660                                 }
1661                 ret = get_a_ctxt(fp, uinfo, alg);
1662         }
1663
1664 done_chk_sdma:
1665         if (!ret)
1666                 ret = do_qib_user_sdma_queue_create(fp);
1667 done_ok:
1668         mutex_unlock(&qib_mutex);
1669
1670 done:
1671         return ret;
1672 }
1673
1674
1675 static int qib_do_user_init(struct file *fp,
1676                             const struct qib_user_info *uinfo)
1677 {
1678         int ret;
1679         struct qib_ctxtdata *rcd = ctxt_fp(fp);
1680         struct qib_devdata *dd;
1681         unsigned uctxt;
1682
1683         /* Subctxts don't need to initialize anything since master did it. */
1684         if (subctxt_fp(fp)) {
1685                 ret = wait_event_interruptible(rcd->wait,
1686                         !test_bit(QIB_CTXT_MASTER_UNINIT, &rcd->flag));
1687                 goto bail;
1688         }
1689
1690         dd = rcd->dd;
1691
1692         /* some ctxts may get extra buffers, calculate that here */
1693         uctxt = rcd->ctxt - dd->first_user_ctxt;
1694         if (uctxt < dd->ctxts_extrabuf) {
1695                 rcd->piocnt = dd->pbufsctxt + 1;
1696                 rcd->pio_base = rcd->piocnt * uctxt;
1697         } else {
1698                 rcd->piocnt = dd->pbufsctxt;
1699                 rcd->pio_base = rcd->piocnt * uctxt +
1700                         dd->ctxts_extrabuf;
1701         }
1702
1703         /*
1704          * All user buffers are 2KB buffers.  If we ever support
1705          * giving 4KB buffers to user processes, this will need some
1706          * work.  Can't use piobufbase directly, because it has
1707          * both 2K and 4K buffer base values.  So check and handle.
1708          */
1709         if ((rcd->pio_base + rcd->piocnt) > dd->piobcnt2k) {
1710                 if (rcd->pio_base >= dd->piobcnt2k) {
1711                         qib_dev_err(dd,
1712                                     "%u:ctxt%u: no 2KB buffers available\n",
1713                                     dd->unit, rcd->ctxt);
1714                         ret = -ENOBUFS;
1715                         goto bail;
1716                 }
1717                 rcd->piocnt = dd->piobcnt2k - rcd->pio_base;
1718                 qib_dev_err(dd, "Ctxt%u: would use 4KB bufs, using %u\n",
1719                             rcd->ctxt, rcd->piocnt);
1720         }
1721
1722         rcd->piobufs = dd->pio2k_bufbase + rcd->pio_base * dd->palign;
1723         qib_chg_pioavailkernel(dd, rcd->pio_base, rcd->piocnt,
1724                                TXCHK_CHG_TYPE_USER, rcd);
1725         /*
1726          * try to ensure that processes start up with consistent avail update
1727          * for their own range, at least.   If system very quiet, it might
1728          * have the in-memory copy out of date at startup for this range of
1729          * buffers, when a context gets re-used.  Do after the chg_pioavail
1730          * and before the rest of setup, so it's "almost certain" the dma
1731          * will have occurred (can't 100% guarantee, but should be many
1732          * decimals of 9s, with this ordering), given how much else happens
1733          * after this.
1734          */
1735         dd->f_sendctrl(dd->pport, QIB_SENDCTRL_AVAIL_BLIP);
1736
1737         /*
1738          * Now allocate the rcvhdr Q and eager TIDs; skip the TID
1739          * array for time being.  If rcd->ctxt > chip-supported,
1740          * we need to do extra stuff here to handle by handling overflow
1741          * through ctxt 0, someday
1742          */
1743         ret = qib_create_rcvhdrq(dd, rcd);
1744         if (!ret)
1745                 ret = qib_setup_eagerbufs(rcd);
1746         if (ret)
1747                 goto bail_pio;
1748
1749         rcd->tidcursor = 0; /* start at beginning after open */
1750
1751         /* initialize poll variables... */
1752         rcd->urgent = 0;
1753         rcd->urgent_poll = 0;
1754
1755         /*
1756          * Now enable the ctxt for receive.
1757          * For chips that are set to DMA the tail register to memory
1758          * when they change (and when the update bit transitions from
1759          * 0 to 1.  So for those chips, we turn it off and then back on.
1760          * This will (very briefly) affect any other open ctxts, but the
1761          * duration is very short, and therefore isn't an issue.  We
1762          * explicitly set the in-memory tail copy to 0 beforehand, so we
1763          * don't have to wait to be sure the DMA update has happened
1764          * (chip resets head/tail to 0 on transition to enable).
1765          */
1766         if (rcd->rcvhdrtail_kvaddr)
1767                 qib_clear_rcvhdrtail(rcd);
1768
1769         dd->f_rcvctrl(rcd->ppd, QIB_RCVCTRL_CTXT_ENB | QIB_RCVCTRL_TIDFLOW_ENB,
1770                       rcd->ctxt);
1771
1772         /* Notify any waiting slaves */
1773         if (rcd->subctxt_cnt) {
1774                 clear_bit(QIB_CTXT_MASTER_UNINIT, &rcd->flag);
1775                 wake_up(&rcd->wait);
1776         }
1777         return 0;
1778
1779 bail_pio:
1780         qib_chg_pioavailkernel(dd, rcd->pio_base, rcd->piocnt,
1781                                TXCHK_CHG_TYPE_KERN, rcd);
1782 bail:
1783         return ret;
1784 }
1785
1786 /**
1787  * unlock_exptid - unlock any expected TID entries context still had in use
1788  * @rcd: ctxt
1789  *
1790  * We don't actually update the chip here, because we do a bulk update
1791  * below, using f_clear_tids.
1792  */
1793 static void unlock_expected_tids(struct qib_ctxtdata *rcd)
1794 {
1795         struct qib_devdata *dd = rcd->dd;
1796         int ctxt_tidbase = rcd->ctxt * dd->rcvtidcnt;
1797         int i, cnt = 0, maxtid = ctxt_tidbase + dd->rcvtidcnt;
1798
1799         for (i = ctxt_tidbase; i < maxtid; i++) {
1800                 struct page *p = dd->pageshadow[i];
1801                 dma_addr_t phys;
1802
1803                 if (!p)
1804                         continue;
1805
1806                 phys = dd->physshadow[i];
1807                 dd->physshadow[i] = dd->tidinvalid;
1808                 dd->pageshadow[i] = NULL;
1809                 pci_unmap_page(dd->pcidev, phys, PAGE_SIZE,
1810                                PCI_DMA_FROMDEVICE);
1811                 qib_release_user_pages(&p, 1);
1812                 cnt++;
1813         }
1814 }
1815
1816 static int qib_close(struct inode *in, struct file *fp)
1817 {
1818         int ret = 0;
1819         struct qib_filedata *fd;
1820         struct qib_ctxtdata *rcd;
1821         struct qib_devdata *dd;
1822         unsigned long flags;
1823         unsigned ctxt;
1824         pid_t pid;
1825
1826         mutex_lock(&qib_mutex);
1827
1828         fd = fp->private_data;
1829         fp->private_data = NULL;
1830         rcd = fd->rcd;
1831         if (!rcd) {
1832                 mutex_unlock(&qib_mutex);
1833                 goto bail;
1834         }
1835
1836         dd = rcd->dd;
1837
1838         /* ensure all pio buffer writes in progress are flushed */
1839         qib_flush_wc();
1840
1841         /* drain user sdma queue */
1842         if (fd->pq) {
1843                 qib_user_sdma_queue_drain(rcd->ppd, fd->pq);
1844                 qib_user_sdma_queue_destroy(fd->pq);
1845         }
1846
1847         if (fd->rec_cpu_num != -1)
1848                 __clear_bit(fd->rec_cpu_num, qib_cpulist);
1849
1850         if (--rcd->cnt) {
1851                 /*
1852                  * XXX If the master closes the context before the slave(s),
1853                  * revoke the mmap for the eager receive queue so
1854                  * the slave(s) don't wait for receive data forever.
1855                  */
1856                 rcd->active_slaves &= ~(1 << fd->subctxt);
1857                 rcd->subpid[fd->subctxt] = 0;
1858                 mutex_unlock(&qib_mutex);
1859                 goto bail;
1860         }
1861
1862         /* early; no interrupt users after this */
1863         spin_lock_irqsave(&dd->uctxt_lock, flags);
1864         ctxt = rcd->ctxt;
1865         dd->rcd[ctxt] = NULL;
1866         pid = rcd->pid;
1867         rcd->pid = 0;
1868         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
1869
1870         if (rcd->rcvwait_to || rcd->piowait_to ||
1871             rcd->rcvnowait || rcd->pionowait) {
1872                 rcd->rcvwait_to = 0;
1873                 rcd->piowait_to = 0;
1874                 rcd->rcvnowait = 0;
1875                 rcd->pionowait = 0;
1876         }
1877         if (rcd->flag)
1878                 rcd->flag = 0;
1879
1880         if (dd->kregbase) {
1881                 /* atomically clear receive enable ctxt and intr avail. */
1882                 dd->f_rcvctrl(rcd->ppd, QIB_RCVCTRL_CTXT_DIS |
1883                                   QIB_RCVCTRL_INTRAVAIL_DIS, ctxt);
1884
1885                 /* clean up the pkeys for this ctxt user */
1886                 qib_clean_part_key(rcd, dd);
1887                 qib_disarm_piobufs(dd, rcd->pio_base, rcd->piocnt);
1888                 qib_chg_pioavailkernel(dd, rcd->pio_base,
1889                                        rcd->piocnt, TXCHK_CHG_TYPE_KERN, NULL);
1890
1891                 dd->f_clear_tids(dd, rcd);
1892
1893                 if (dd->pageshadow)
1894                         unlock_expected_tids(rcd);
1895                 qib_stats.sps_ctxts--;
1896                 dd->freectxts++;
1897         }
1898
1899         mutex_unlock(&qib_mutex);
1900         qib_free_ctxtdata(dd, rcd); /* after releasing the mutex */
1901
1902 bail:
1903         kfree(fd);
1904         return ret;
1905 }
1906
1907 static int qib_ctxt_info(struct file *fp, struct qib_ctxt_info __user *uinfo)
1908 {
1909         struct qib_ctxt_info info;
1910         int ret;
1911         size_t sz;
1912         struct qib_ctxtdata *rcd = ctxt_fp(fp);
1913         struct qib_filedata *fd;
1914
1915         fd = fp->private_data;
1916
1917         info.num_active = qib_count_active_units();
1918         info.unit = rcd->dd->unit;
1919         info.port = rcd->ppd->port;
1920         info.ctxt = rcd->ctxt;
1921         info.subctxt =  subctxt_fp(fp);
1922         /* Number of user ctxts available for this device. */
1923         info.num_ctxts = rcd->dd->cfgctxts - rcd->dd->first_user_ctxt;
1924         info.num_subctxts = rcd->subctxt_cnt;
1925         info.rec_cpu = fd->rec_cpu_num;
1926         sz = sizeof(info);
1927
1928         if (copy_to_user(uinfo, &info, sz)) {
1929                 ret = -EFAULT;
1930                 goto bail;
1931         }
1932         ret = 0;
1933
1934 bail:
1935         return ret;
1936 }
1937
1938 static int qib_sdma_get_inflight(struct qib_user_sdma_queue *pq,
1939                                  u32 __user *inflightp)
1940 {
1941         const u32 val = qib_user_sdma_inflight_counter(pq);
1942
1943         if (put_user(val, inflightp))
1944                 return -EFAULT;
1945
1946         return 0;
1947 }
1948
1949 static int qib_sdma_get_complete(struct qib_pportdata *ppd,
1950                                  struct qib_user_sdma_queue *pq,
1951                                  u32 __user *completep)
1952 {
1953         u32 val;
1954         int err;
1955
1956         if (!pq)
1957                 return -EINVAL;
1958
1959         err = qib_user_sdma_make_progress(ppd, pq);
1960         if (err < 0)
1961                 return err;
1962
1963         val = qib_user_sdma_complete_counter(pq);
1964         if (put_user(val, completep))
1965                 return -EFAULT;
1966
1967         return 0;
1968 }
1969
1970 static int disarm_req_delay(struct qib_ctxtdata *rcd)
1971 {
1972         int ret = 0;
1973
1974         if (!usable(rcd->ppd)) {
1975                 int i;
1976                 /*
1977                  * if link is down, or otherwise not usable, delay
1978                  * the caller up to 30 seconds, so we don't thrash
1979                  * in trying to get the chip back to ACTIVE, and
1980                  * set flag so they make the call again.
1981                  */
1982                 if (rcd->user_event_mask) {
1983                         /*
1984                          * subctxt_cnt is 0 if not shared, so do base
1985                          * separately, first, then remaining subctxt, if any
1986                          */
1987                         set_bit(_QIB_EVENT_DISARM_BUFS_BIT,
1988                                 &rcd->user_event_mask[0]);
1989                         for (i = 1; i < rcd->subctxt_cnt; i++)
1990                                 set_bit(_QIB_EVENT_DISARM_BUFS_BIT,
1991                                         &rcd->user_event_mask[i]);
1992                 }
1993                 for (i = 0; !usable(rcd->ppd) && i < 300; i++)
1994                         msleep(100);
1995                 ret = -ENETDOWN;
1996         }
1997         return ret;
1998 }
1999
2000 /*
2001  * Find all user contexts in use, and set the specified bit in their
2002  * event mask.
2003  * See also find_ctxt() for a similar use, that is specific to send buffers.
2004  */
2005 int qib_set_uevent_bits(struct qib_pportdata *ppd, const int evtbit)
2006 {
2007         struct qib_ctxtdata *rcd;
2008         unsigned ctxt;
2009         int ret = 0;
2010         unsigned long flags;
2011
2012         spin_lock_irqsave(&ppd->dd->uctxt_lock, flags);
2013         for (ctxt = ppd->dd->first_user_ctxt; ctxt < ppd->dd->cfgctxts;
2014              ctxt++) {
2015                 rcd = ppd->dd->rcd[ctxt];
2016                 if (!rcd)
2017                         continue;
2018                 if (rcd->user_event_mask) {
2019                         int i;
2020                         /*
2021                          * subctxt_cnt is 0 if not shared, so do base
2022                          * separately, first, then remaining subctxt, if any
2023                          */
2024                         set_bit(evtbit, &rcd->user_event_mask[0]);
2025                         for (i = 1; i < rcd->subctxt_cnt; i++)
2026                                 set_bit(evtbit, &rcd->user_event_mask[i]);
2027                 }
2028                 ret = 1;
2029                 break;
2030         }
2031         spin_unlock_irqrestore(&ppd->dd->uctxt_lock, flags);
2032
2033         return ret;
2034 }
2035
2036 /*
2037  * clear the event notifier events for this context.
2038  * For the DISARM_BUFS case, we also take action (this obsoletes
2039  * the older QIB_CMD_DISARM_BUFS, but we keep it for backwards
2040  * compatibility.
2041  * Other bits don't currently require actions, just atomically clear.
2042  * User process then performs actions appropriate to bit having been
2043  * set, if desired, and checks again in future.
2044  */
2045 static int qib_user_event_ack(struct qib_ctxtdata *rcd, int subctxt,
2046                               unsigned long events)
2047 {
2048         int ret = 0, i;
2049
2050         for (i = 0; i <= _QIB_MAX_EVENT_BIT; i++) {
2051                 if (!test_bit(i, &events))
2052                         continue;
2053                 if (i == _QIB_EVENT_DISARM_BUFS_BIT) {
2054                         (void)qib_disarm_piobufs_ifneeded(rcd);
2055                         ret = disarm_req_delay(rcd);
2056                 } else
2057                         clear_bit(i, &rcd->user_event_mask[subctxt]);
2058         }
2059         return ret;
2060 }
2061
2062 static ssize_t qib_write(struct file *fp, const char __user *data,
2063                          size_t count, loff_t *off)
2064 {
2065         const struct qib_cmd __user *ucmd;
2066         struct qib_ctxtdata *rcd;
2067         const void __user *src;
2068         size_t consumed, copy = 0;
2069         struct qib_cmd cmd;
2070         ssize_t ret = 0;
2071         void *dest;
2072
2073         if (!ib_safe_file_access(fp)) {
2074                 pr_err_once("qib_write: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n",
2075                             task_tgid_vnr(current), current->comm);
2076                 return -EACCES;
2077         }
2078
2079         if (count < sizeof(cmd.type)) {
2080                 ret = -EINVAL;
2081                 goto bail;
2082         }
2083
2084         ucmd = (const struct qib_cmd __user *) data;
2085
2086         if (copy_from_user(&cmd.type, &ucmd->type, sizeof(cmd.type))) {
2087                 ret = -EFAULT;
2088                 goto bail;
2089         }
2090
2091         consumed = sizeof(cmd.type);
2092
2093         switch (cmd.type) {
2094         case QIB_CMD_ASSIGN_CTXT:
2095         case QIB_CMD_USER_INIT:
2096                 copy = sizeof(cmd.cmd.user_info);
2097                 dest = &cmd.cmd.user_info;
2098                 src = &ucmd->cmd.user_info;
2099                 break;
2100
2101         case QIB_CMD_RECV_CTRL:
2102                 copy = sizeof(cmd.cmd.recv_ctrl);
2103                 dest = &cmd.cmd.recv_ctrl;
2104                 src = &ucmd->cmd.recv_ctrl;
2105                 break;
2106
2107         case QIB_CMD_CTXT_INFO:
2108                 copy = sizeof(cmd.cmd.ctxt_info);
2109                 dest = &cmd.cmd.ctxt_info;
2110                 src = &ucmd->cmd.ctxt_info;
2111                 break;
2112
2113         case QIB_CMD_TID_UPDATE:
2114         case QIB_CMD_TID_FREE:
2115                 copy = sizeof(cmd.cmd.tid_info);
2116                 dest = &cmd.cmd.tid_info;
2117                 src = &ucmd->cmd.tid_info;
2118                 break;
2119
2120         case QIB_CMD_SET_PART_KEY:
2121                 copy = sizeof(cmd.cmd.part_key);
2122                 dest = &cmd.cmd.part_key;
2123                 src = &ucmd->cmd.part_key;
2124                 break;
2125
2126         case QIB_CMD_DISARM_BUFS:
2127         case QIB_CMD_PIOAVAILUPD: /* force an update of PIOAvail reg */
2128                 copy = 0;
2129                 src = NULL;
2130                 dest = NULL;
2131                 break;
2132
2133         case QIB_CMD_POLL_TYPE:
2134                 copy = sizeof(cmd.cmd.poll_type);
2135                 dest = &cmd.cmd.poll_type;
2136                 src = &ucmd->cmd.poll_type;
2137                 break;
2138
2139         case QIB_CMD_ARMLAUNCH_CTRL:
2140                 copy = sizeof(cmd.cmd.armlaunch_ctrl);
2141                 dest = &cmd.cmd.armlaunch_ctrl;
2142                 src = &ucmd->cmd.armlaunch_ctrl;
2143                 break;
2144
2145         case QIB_CMD_SDMA_INFLIGHT:
2146                 copy = sizeof(cmd.cmd.sdma_inflight);
2147                 dest = &cmd.cmd.sdma_inflight;
2148                 src = &ucmd->cmd.sdma_inflight;
2149                 break;
2150
2151         case QIB_CMD_SDMA_COMPLETE:
2152                 copy = sizeof(cmd.cmd.sdma_complete);
2153                 dest = &cmd.cmd.sdma_complete;
2154                 src = &ucmd->cmd.sdma_complete;
2155                 break;
2156
2157         case QIB_CMD_ACK_EVENT:
2158                 copy = sizeof(cmd.cmd.event_mask);
2159                 dest = &cmd.cmd.event_mask;
2160                 src = &ucmd->cmd.event_mask;
2161                 break;
2162
2163         default:
2164                 ret = -EINVAL;
2165                 goto bail;
2166         }
2167
2168         if (copy) {
2169                 if ((count - consumed) < copy) {
2170                         ret = -EINVAL;
2171                         goto bail;
2172                 }
2173                 if (copy_from_user(dest, src, copy)) {
2174                         ret = -EFAULT;
2175                         goto bail;
2176                 }
2177                 consumed += copy;
2178         }
2179
2180         rcd = ctxt_fp(fp);
2181         if (!rcd && cmd.type != QIB_CMD_ASSIGN_CTXT) {
2182                 ret = -EINVAL;
2183                 goto bail;
2184         }
2185
2186         switch (cmd.type) {
2187         case QIB_CMD_ASSIGN_CTXT:
2188                 if (rcd) {
2189                         ret = -EINVAL;
2190                         goto bail;
2191                 }
2192
2193                 ret = qib_assign_ctxt(fp, &cmd.cmd.user_info);
2194                 if (ret)
2195                         goto bail;
2196                 break;
2197
2198         case QIB_CMD_USER_INIT:
2199                 ret = qib_do_user_init(fp, &cmd.cmd.user_info);
2200                 if (ret)
2201                         goto bail;
2202                 ret = qib_get_base_info(fp, (void __user *) (unsigned long)
2203                                         cmd.cmd.user_info.spu_base_info,
2204                                         cmd.cmd.user_info.spu_base_info_size);
2205                 break;
2206
2207         case QIB_CMD_RECV_CTRL:
2208                 ret = qib_manage_rcvq(rcd, subctxt_fp(fp), cmd.cmd.recv_ctrl);
2209                 break;
2210
2211         case QIB_CMD_CTXT_INFO:
2212                 ret = qib_ctxt_info(fp, (struct qib_ctxt_info __user *)
2213                                     (unsigned long) cmd.cmd.ctxt_info);
2214                 break;
2215
2216         case QIB_CMD_TID_UPDATE:
2217                 ret = qib_tid_update(rcd, fp, &cmd.cmd.tid_info);
2218                 break;
2219
2220         case QIB_CMD_TID_FREE:
2221                 ret = qib_tid_free(rcd, subctxt_fp(fp), &cmd.cmd.tid_info);
2222                 break;
2223
2224         case QIB_CMD_SET_PART_KEY:
2225                 ret = qib_set_part_key(rcd, cmd.cmd.part_key);
2226                 break;
2227
2228         case QIB_CMD_DISARM_BUFS:
2229                 (void)qib_disarm_piobufs_ifneeded(rcd);
2230                 ret = disarm_req_delay(rcd);
2231                 break;
2232
2233         case QIB_CMD_PIOAVAILUPD:
2234                 qib_force_pio_avail_update(rcd->dd);
2235                 break;
2236
2237         case QIB_CMD_POLL_TYPE:
2238                 rcd->poll_type = cmd.cmd.poll_type;
2239                 break;
2240
2241         case QIB_CMD_ARMLAUNCH_CTRL:
2242                 rcd->dd->f_set_armlaunch(rcd->dd, cmd.cmd.armlaunch_ctrl);
2243                 break;
2244
2245         case QIB_CMD_SDMA_INFLIGHT:
2246                 ret = qib_sdma_get_inflight(user_sdma_queue_fp(fp),
2247                                             (u32 __user *) (unsigned long)
2248                                             cmd.cmd.sdma_inflight);
2249                 break;
2250
2251         case QIB_CMD_SDMA_COMPLETE:
2252                 ret = qib_sdma_get_complete(rcd->ppd,
2253                                             user_sdma_queue_fp(fp),
2254                                             (u32 __user *) (unsigned long)
2255                                             cmd.cmd.sdma_complete);
2256                 break;
2257
2258         case QIB_CMD_ACK_EVENT:
2259                 ret = qib_user_event_ack(rcd, subctxt_fp(fp),
2260                                          cmd.cmd.event_mask);
2261                 break;
2262         }
2263
2264         if (ret >= 0)
2265                 ret = consumed;
2266
2267 bail:
2268         return ret;
2269 }
2270
2271 static ssize_t qib_write_iter(struct kiocb *iocb, struct iov_iter *from)
2272 {
2273         struct qib_filedata *fp = iocb->ki_filp->private_data;
2274         struct qib_ctxtdata *rcd = ctxt_fp(iocb->ki_filp);
2275         struct qib_user_sdma_queue *pq = fp->pq;
2276
2277         if (!iter_is_iovec(from) || !from->nr_segs || !pq)
2278                 return -EINVAL;
2279
2280         return qib_user_sdma_writev(rcd, pq, from->iov, from->nr_segs);
2281 }
2282
2283 static struct class *qib_class;
2284 static dev_t qib_dev;
2285
2286 int qib_cdev_init(int minor, const char *name,
2287                   const struct file_operations *fops,
2288                   struct cdev **cdevp, struct device **devp)
2289 {
2290         const dev_t dev = MKDEV(MAJOR(qib_dev), minor);
2291         struct cdev *cdev;
2292         struct device *device = NULL;
2293         int ret;
2294
2295         cdev = cdev_alloc();
2296         if (!cdev) {
2297                 pr_err("Could not allocate cdev for minor %d, %s\n",
2298                        minor, name);
2299                 ret = -ENOMEM;
2300                 goto done;
2301         }
2302
2303         cdev->owner = THIS_MODULE;
2304         cdev->ops = fops;
2305         kobject_set_name(&cdev->kobj, name);
2306
2307         ret = cdev_add(cdev, dev, 1);
2308         if (ret < 0) {
2309                 pr_err("Could not add cdev for minor %d, %s (err %d)\n",
2310                        minor, name, -ret);
2311                 goto err_cdev;
2312         }
2313
2314         device = device_create(qib_class, NULL, dev, NULL, "%s", name);
2315         if (!IS_ERR(device))
2316                 goto done;
2317         ret = PTR_ERR(device);
2318         device = NULL;
2319         pr_err("Could not create device for minor %d, %s (err %d)\n",
2320                minor, name, -ret);
2321 err_cdev:
2322         cdev_del(cdev);
2323         cdev = NULL;
2324 done:
2325         *cdevp = cdev;
2326         *devp = device;
2327         return ret;
2328 }
2329
2330 void qib_cdev_cleanup(struct cdev **cdevp, struct device **devp)
2331 {
2332         struct device *device = *devp;
2333
2334         if (device) {
2335                 device_unregister(device);
2336                 *devp = NULL;
2337         }
2338
2339         if (*cdevp) {
2340                 cdev_del(*cdevp);
2341                 *cdevp = NULL;
2342         }
2343 }
2344
2345 static struct cdev *wildcard_cdev;
2346 static struct device *wildcard_device;
2347
2348 int __init qib_dev_init(void)
2349 {
2350         int ret;
2351
2352         ret = alloc_chrdev_region(&qib_dev, 0, QIB_NMINORS, QIB_DRV_NAME);
2353         if (ret < 0) {
2354                 pr_err("Could not allocate chrdev region (err %d)\n", -ret);
2355                 goto done;
2356         }
2357
2358         qib_class = class_create(THIS_MODULE, "ipath");
2359         if (IS_ERR(qib_class)) {
2360                 ret = PTR_ERR(qib_class);
2361                 pr_err("Could not create device class (err %d)\n", -ret);
2362                 unregister_chrdev_region(qib_dev, QIB_NMINORS);
2363         }
2364
2365 done:
2366         return ret;
2367 }
2368
2369 void qib_dev_cleanup(void)
2370 {
2371         if (qib_class) {
2372                 class_destroy(qib_class);
2373                 qib_class = NULL;
2374         }
2375
2376         unregister_chrdev_region(qib_dev, QIB_NMINORS);
2377 }
2378
2379 static atomic_t user_count = ATOMIC_INIT(0);
2380
2381 static void qib_user_remove(struct qib_devdata *dd)
2382 {
2383         if (atomic_dec_return(&user_count) == 0)
2384                 qib_cdev_cleanup(&wildcard_cdev, &wildcard_device);
2385
2386         qib_cdev_cleanup(&dd->user_cdev, &dd->user_device);
2387 }
2388
2389 static int qib_user_add(struct qib_devdata *dd)
2390 {
2391         char name[10];
2392         int ret;
2393
2394         if (atomic_inc_return(&user_count) == 1) {
2395                 ret = qib_cdev_init(0, "ipath", &qib_file_ops,
2396                                     &wildcard_cdev, &wildcard_device);
2397                 if (ret)
2398                         goto done;
2399         }
2400
2401         snprintf(name, sizeof(name), "ipath%d", dd->unit);
2402         ret = qib_cdev_init(dd->unit + 1, name, &qib_file_ops,
2403                             &dd->user_cdev, &dd->user_device);
2404         if (ret)
2405                 qib_user_remove(dd);
2406 done:
2407         return ret;
2408 }
2409
2410 /*
2411  * Create per-unit files in /dev
2412  */
2413 int qib_device_create(struct qib_devdata *dd)
2414 {
2415         int r, ret;
2416
2417         r = qib_user_add(dd);
2418         ret = qib_diag_add(dd);
2419         if (r && !ret)
2420                 ret = r;
2421         return ret;
2422 }
2423
2424 /*
2425  * Remove per-unit files in /dev
2426  * void, core kernel returns no errors for this stuff
2427  */
2428 void qib_device_remove(struct qib_devdata *dd)
2429 {
2430         qib_user_remove(dd);
2431         qib_diag_remove(dd);
2432 }