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