drivers/infiniband/hw/qib/qib_file_ops.c

   1 /*
   2  * Copyright (c) 2012, 2013 Intel Corporation. All rights reserved.
   3  * Copyright (c) 2006 - 2012 QLogic Corporation. All rights reserved.
   4  * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the
  10  * OpenIB.org BSD license below:
  11  *
  12  *     Redistribution and use in source and binary forms, with or
  13  *     without modification, are permitted provided that the following
  14  *     conditions are met:
  15  *
  16  *      - Redistributions of source code must retain the above
  17  *        copyright notice, this list of conditions and the following
  18  *        disclaimer.
  19  *
  20  *      - Redistributions in binary form must reproduce the above
  21  *        copyright notice, this list of conditions and the following
  22  *        disclaimer in the documentation and/or other materials
  23  *        provided with the distribution.
  24  *
  25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  32  * SOFTWARE.
  33  */
  34
  35 #include <linux/pci.h>
  36 #include <linux/poll.h>
  37 #include <linux/cdev.h>
  38 #include <linux/swap.h>
  39 #include <linux/vmalloc.h>
  40 #include <linux/highmem.h>
  41 #include <linux/io.h>
  42 #include <linux/jiffies.h>
  43 #include <asm/pgtable.h>
  44 #include <linux/delay.h>
  45 #include <linux/export.h>
  46 #include <linux/uio.h>
  47
  48 #include <rdma/ib.h>
  49
  50 #include "qib.h"
  51 #include "qib_common.h"
  52 #include "qib_user_sdma.h"
  53
  54 #undef pr_fmt
  55 #define pr_fmt(fmt) QIB_DRV_NAME ": " fmt
  56
  57 static int qib_open(struct inode *, struct file *);
  58 static int qib_close(struct inode *, struct file *);
  59 static ssize_t qib_write(struct file *, const char __user *, size_t, loff_t *);
  60 static ssize_t qib_write_iter(struct kiocb *, struct iov_iter *);
  61 static unsigned int qib_poll(struct file *, struct poll_table_struct *);
  62 static int qib_mmapf(struct file *, struct vm_area_struct *);
  63
  64 /*
  65  * This is really, really weird shit - write() and writev() here
  66  * have completely unrelated semantics.  Sucky userland ABI,
  67  * film at 11.
  68  */
  69 static const struct file_operations qib_file_ops = {
  70         .owner = THIS_MODULE,
  71         .write = qib_write,
  72         .write_iter = qib_write_iter,
  73         .open = qib_open,
  74         .release = qib_close,
  75         .poll = qib_poll,
  76         .mmap = qib_mmapf,
  77         .llseek = noop_llseek,
  78 };
  79
  80 /*
  81  * Convert kernel virtual addresses to physical addresses so they don't
  82  * potentially conflict with the chip addresses used as mmap offsets.
  83  * It doesn't really matter what mmap offset we use as long as we can
  84  * interpret it correctly.
  85  */
  86 static u64 cvt_kvaddr(void *p)
  87 {
  88         struct page *page;
  89         u64 paddr = 0;
  90
  91         page = vmalloc_to_page(p);
  92         if (page)
  93                 paddr = page_to_pfn(page) << PAGE_SHIFT;
  94
  95         return paddr;
  96 }
  97
  98 static int qib_get_base_info(struct file *fp, void __user *ubase,
  99                              size_t ubase_size)
 100 {
 101         struct qib_ctxtdata *rcd = ctxt_fp(fp);
 102         int ret = 0;
 103         struct qib_base_info *kinfo = NULL;
 104         struct qib_devdata *dd = rcd->dd;
 105         struct qib_pportdata *ppd = rcd->ppd;
 106         unsigned subctxt_cnt;
 107         int shared, master;
 108         size_t sz;
 109
 110         subctxt_cnt = rcd->subctxt_cnt;
 111         if (!subctxt_cnt) {
 112                 shared = 0;
 113                 master = 0;
 114                 subctxt_cnt = 1;
 115         } else {
 116                 shared = 1;
 117                 master = !subctxt_fp(fp);
 118         }
 119
 120         sz = sizeof(*kinfo);
 121         /* If context sharing is not requested, allow the old size structure */
 122         if (!shared)
 123                 sz -= 7 * sizeof(u64);
 124         if (ubase_size < sz) {
 125                 ret = -EINVAL;
 126                 goto bail;
 127         }
 128
 129         kinfo = kzalloc(sizeof(*kinfo), GFP_KERNEL);
 130         if (kinfo == NULL) {
 131                 ret = -ENOMEM;
 132                 goto bail;
 133         }
 134
 135         ret = dd->f_get_base_info(rcd, kinfo);
 136         if (ret < 0)
 137                 goto bail;
 138
 139         kinfo->spi_rcvhdr_cnt = dd->rcvhdrcnt;
 140         kinfo->spi_rcvhdrent_size = dd->rcvhdrentsize;
 141         kinfo->spi_tidegrcnt = rcd->rcvegrcnt;
 142         kinfo->spi_rcv_egrbufsize = dd->rcvegrbufsize;
 143         /*
 144          * have to mmap whole thing
 145          */
 146         kinfo->spi_rcv_egrbuftotlen =
 147                 rcd->rcvegrbuf_chunks * rcd->rcvegrbuf_size;
 148         kinfo->spi_rcv_egrperchunk = rcd->rcvegrbufs_perchunk;
 149         kinfo->spi_rcv_egrchunksize = kinfo->spi_rcv_egrbuftotlen /
 150                 rcd->rcvegrbuf_chunks;
 151         kinfo->spi_tidcnt = dd->rcvtidcnt / subctxt_cnt;
 152         if (master)
 153                 kinfo->spi_tidcnt += dd->rcvtidcnt % subctxt_cnt;
 154         /*
 155          * for this use, may be cfgctxts summed over all chips that
 156          * are are configured and present
 157          */
 158         kinfo->spi_nctxts = dd->cfgctxts;
 159         /* unit (chip/board) our context is on */
 160         kinfo->spi_unit = dd->unit;
 161         kinfo->spi_port = ppd->port;
 162         /* for now, only a single page */
 163         kinfo->spi_tid_maxsize = PAGE_SIZE;
 164
 165         /*
 166          * Doing this per context, and based on the skip value, etc.  This has
 167          * to be the actual buffer size, since the protocol code treats it
 168          * as an array.
 169          *
 170          * These have to be set to user addresses in the user code via mmap.
 171          * These values are used on return to user code for the mmap target
 172          * addresses only.  For 32 bit, same 44 bit address problem, so use
 173          * the physical address, not virtual.  Before 2.6.11, using the
 174          * page_address() macro worked, but in 2.6.11, even that returns the
 175          * full 64 bit address (upper bits all 1's).  So far, using the
 176          * physical addresses (or chip offsets, for chip mapping) works, but
 177          * no doubt some future kernel release will change that, and we'll be
 178          * on to yet another method of dealing with this.
 179          * Normally only one of rcvhdr_tailaddr or rhf_offset is useful
 180          * since the chips with non-zero rhf_offset don't normally
 181          * enable tail register updates to host memory, but for testing,
 182          * both can be enabled and used.
 183          */
 184         kinfo->spi_rcvhdr_base = (u64) rcd->rcvhdrq_phys;
 185         kinfo->spi_rcvhdr_tailaddr = (u64) rcd->rcvhdrqtailaddr_phys;
 186         kinfo->spi_rhf_offset = dd->rhf_offset;
 187         kinfo->spi_rcv_egrbufs = (u64) rcd->rcvegr_phys;
 188         kinfo->spi_pioavailaddr = (u64) dd->pioavailregs_phys;
 189         /* setup per-unit (not port) status area for user programs */
 190         kinfo->spi_status = (u64) kinfo->spi_pioavailaddr +
 191                 (char *) ppd->statusp -
 192                 (char *) dd->pioavailregs_dma;
 193         kinfo->spi_uregbase = (u64) dd->uregbase + dd->ureg_align * rcd->ctxt;
 194         if (!shared) {
 195                 kinfo->spi_piocnt = rcd->piocnt;
 196                 kinfo->spi_piobufbase = (u64) rcd->piobufs;
 197                 kinfo->spi_sendbuf_status = cvt_kvaddr(rcd->user_event_mask);
 198         } else if (master) {
 199                 kinfo->spi_piocnt = (rcd->piocnt / subctxt_cnt) +
 200                                     (rcd->piocnt % subctxt_cnt);
 201                 /* Master's PIO buffers are after all the slave's */
 202                 kinfo->spi_piobufbase = (u64) rcd->piobufs +
 203                         dd->palign *
 204                         (rcd->piocnt - kinfo->spi_piocnt);
 205         } else {
 206                 unsigned slave = subctxt_fp(fp) - 1;
 207
 208                 kinfo->spi_piocnt = rcd->piocnt / subctxt_cnt;
 209                 kinfo->spi_piobufbase = (u64) rcd->piobufs +
 210                         dd->palign * kinfo->spi_piocnt * slave;
 211         }
 212
 213         if (shared) {
 214                 kinfo->spi_sendbuf_status =
 215                         cvt_kvaddr(&rcd->user_event_mask[subctxt_fp(fp)]);
 216                 /* only spi_subctxt_* fields should be set in this block! */
 217                 kinfo->spi_subctxt_uregbase = cvt_kvaddr(rcd->subctxt_uregbase);
 218
 219                 kinfo->spi_subctxt_rcvegrbuf =
 220                         cvt_kvaddr(rcd->subctxt_rcvegrbuf);
 221                 kinfo->spi_subctxt_rcvhdr_base =
 222                         cvt_kvaddr(rcd->subctxt_rcvhdr_base);
 223         }
 224
 225         /*
 226          * All user buffers are 2KB buffers.  If we ever support
 227          * giving 4KB buffers to user processes, this will need some
 228          * work.  Can't use piobufbase directly, because it has
 229          * both 2K and 4K buffer base values.
 230          */
 231         kinfo->spi_pioindex = (kinfo->spi_piobufbase - dd->pio2k_bufbase) /
 232                 dd->palign;
 233         kinfo->spi_pioalign = dd->palign;
 234         kinfo->spi_qpair = QIB_KD_QP;
 235         /*
 236          * user mode PIO buffers are always 2KB, even when 4KB can
 237          * be received, and sent via the kernel; this is ibmaxlen
 238          * for 2K MTU.
 239          */
 240         kinfo->spi_piosize = dd->piosize2k - 2 * sizeof(u32);
 241         kinfo->spi_mtu = ppd->ibmaxlen; /* maxlen, not ibmtu */
 242         kinfo->spi_ctxt = rcd->ctxt;
 243         kinfo->spi_subctxt = subctxt_fp(fp);
 244         kinfo->spi_sw_version = QIB_KERN_SWVERSION;
 245         kinfo->spi_sw_version |= 1U << 31; /* QLogic-built, not kernel.org */
 246         kinfo->spi_hw_version = dd->revision;
 247
 248         if (master)
 249                 kinfo->spi_runtime_flags |= QIB_RUNTIME_MASTER;
 250
 251         sz = (ubase_size < sizeof(*kinfo)) ? ubase_size : sizeof(*kinfo);
 252         if (copy_to_user(ubase, kinfo, sz))
 253                 ret = -EFAULT;
 254 bail:
 255         kfree(kinfo);
 256         return ret;
 257 }
 258
 259 /**
 260  * qib_tid_update - update a context TID
 261  * @rcd: the context
 262  * @fp: the qib device file
 263  * @ti: the TID information
 264  *
 265  * The new implementation as of Oct 2004 is that the driver assigns
 266  * the tid and returns it to the caller.   To reduce search time, we
 267  * keep a cursor for each context, walking the shadow tid array to find
 268  * one that's not in use.
 269  *
 270  * For now, if we can't allocate the full list, we fail, although
 271  * in the long run, we'll allocate as many as we can, and the
 272  * caller will deal with that by trying the remaining pages later.
 273  * That means that when we fail, we have to mark the tids as not in
 274  * use again, in our shadow copy.
 275  *
 276  * It's up to the caller to free the tids when they are done.
 277  * We'll unlock the pages as they free them.
 278  *
 279  * Also, right now we are locking one page at a time, but since
 280  * the intended use of this routine is for a single group of
 281  * virtually contiguous pages, that should change to improve
 282  * performance.
 283  */
 284 static int qib_tid_update(struct qib_ctxtdata *rcd, struct file *fp,
 285                           const struct qib_tid_info *ti)
 286 {
 287         int ret = 0, ntids;
 288         u32 tid, ctxttid, cnt, i, tidcnt, tidoff;
 289         u16 *tidlist;
 290         struct qib_devdata *dd = rcd->dd;
 291         u64 physaddr;
 292         unsigned long vaddr;
 293         u64 __iomem *tidbase;
 294         unsigned long tidmap[8];
 295         struct page **pagep = NULL;
 296         unsigned subctxt = subctxt_fp(fp);
 297
 298         if (!dd->pageshadow) {
 299                 ret = -ENOMEM;
 300                 goto done;
 301         }
 302
 303         cnt = ti->tidcnt;
 304         if (!cnt) {
 305                 ret = -EFAULT;
 306                 goto done;
 307         }
 308         ctxttid = rcd->ctxt * dd->rcvtidcnt;
 309         if (!rcd->subctxt_cnt) {
 310                 tidcnt = dd->rcvtidcnt;
 311                 tid = rcd->tidcursor;
 312                 tidoff = 0;
 313         } else if (!subctxt) {
 314                 tidcnt = (dd->rcvtidcnt / rcd->subctxt_cnt) +
 315                          (dd->rcvtidcnt % rcd->subctxt_cnt);
 316                 tidoff = dd->rcvtidcnt - tidcnt;
 317                 ctxttid += tidoff;
 318                 tid = tidcursor_fp(fp);
 319         } else {
 320                 tidcnt = dd->rcvtidcnt / rcd->subctxt_cnt;
 321                 tidoff = tidcnt * (subctxt - 1);
 322                 ctxttid += tidoff;
 323                 tid = tidcursor_fp(fp);
 324         }
 325         if (cnt > tidcnt) {
 326                 /* make sure it all fits in tid_pg_list */
 327                 qib_devinfo(dd->pcidev,
 328                         "Process tried to allocate %u TIDs, only trying max (%u)\n",
 329                         cnt, tidcnt);
 330                 cnt = tidcnt;
 331         }
 332         pagep = (struct page **) rcd->tid_pg_list;
 333         tidlist = (u16 *) &pagep[dd->rcvtidcnt];
 334         pagep += tidoff;
 335         tidlist += tidoff;
 336
 337         memset(tidmap, 0, sizeof(tidmap));
 338         /* before decrement; chip actual # */
 339         ntids = tidcnt;
 340         tidbase = (u64 __iomem *) (((char __iomem *) dd->kregbase) +
 341                                    dd->rcvtidbase +
 342                                    ctxttid * sizeof(*tidbase));
 343
 344         /* virtual address of first page in transfer */
 345         vaddr = ti->tidvaddr;
 346         if (!access_ok(VERIFY_WRITE, (void __user *) vaddr,
 347                        cnt * PAGE_SIZE)) {
 348                 ret = -EFAULT;
 349                 goto done;
 350         }
 351         ret = qib_get_user_pages(vaddr, cnt, pagep);
 352         if (ret) {
 353                 /*
 354                  * if (ret == -EBUSY)
 355                  * We can't continue because the pagep array won't be
 356                  * initialized. This should never happen,
 357                  * unless perhaps the user has mpin'ed the pages
 358                  * themselves.
 359                  */
 360                 qib_devinfo(
 361                         dd->pcidev,
 362                         "Failed to lock addr %p, %u pages: errno %d\n",
 363                         (void *) vaddr, cnt, -ret);
 364                 goto done;
 365         }
 366         for (i = 0; i < cnt; i++, vaddr += PAGE_SIZE) {
 367                 dma_addr_t daddr;
 368
 369                 for (; ntids--; tid++) {
 370                         if (tid == tidcnt)
 371                                 tid = 0;
 372                         if (!dd->pageshadow[ctxttid + tid])
 373                                 break;
 374                 }
 375                 if (ntids < 0) {
 376                         /*
 377                          * Oops, wrapped all the way through their TIDs,
 378                          * and didn't have enough free; see comments at
 379                          * start of routine
 380                          */
 381                         i--;    /* last tidlist[i] not filled in */
 382                         ret = -ENOMEM;
 383                         break;
 384                 }
 385                 ret = qib_map_page(dd->pcidev, pagep[i], &daddr);
 386                 if (ret)
 387                         break;
 388
 389                 tidlist[i] = tid + tidoff;
 390                 /* we "know" system pages and TID pages are same size */
 391                 dd->pageshadow[ctxttid + tid] = pagep[i];
 392                 dd->physshadow[ctxttid + tid] = daddr;
 393                 /*
 394                  * don't need atomic or it's overhead
 395                  */
 396                 __set_bit(tid, tidmap);
 397                 physaddr = dd->physshadow[ctxttid + tid];
 398                 /* PERFORMANCE: below should almost certainly be cached */
 399                 dd->f_put_tid(dd, &tidbase[tid],
 400                                   RCVHQ_RCV_TYPE_EXPECTED, physaddr);
 401                 /*
 402                  * don't check this tid in qib_ctxtshadow, since we
 403                  * just filled it in; start with the next one.
 404                  */
 405                 tid++;
 406         }
 407
 408         if (ret) {
 409                 u32 limit;
 410 cleanup:
 411                 /* jump here if copy out of updated info failed... */
 412                 /* same code that's in qib_free_tid() */
 413                 limit = sizeof(tidmap) * BITS_PER_BYTE;
 414                 if (limit > tidcnt)
 415                         /* just in case size changes in future */
 416                         limit = tidcnt;
 417                 tid = find_first_bit((const unsigned long *)tidmap, limit);
 418                 for (; tid < limit; tid++) {
 419                         if (!test_bit(tid, tidmap))
 420                                 continue;
 421                         if (dd->pageshadow[ctxttid + tid]) {
 422                                 dma_addr_t phys;
 423
 424                                 phys = dd->physshadow[ctxttid + tid];
 425                                 dd->physshadow[ctxttid + tid] = dd->tidinvalid;
 426                                 /* PERFORMANCE: below should almost certainly
 427                                  * be cached
 428                                  */
 429                                 dd->f_put_tid(dd, &tidbase[tid],
 430                                               RCVHQ_RCV_TYPE_EXPECTED,
 431                                               dd->tidinvalid);
 432                                 pci_unmap_page(dd->pcidev, phys, PAGE_SIZE,
 433                                                PCI_DMA_FROMDEVICE);
 434                                 dd->pageshadow[ctxttid + tid] = NULL;
 435                         }
 436                 }
 437                 qib_release_user_pages(pagep, cnt);
 438         } else {
 439                 /*
 440                  * Copy the updated array, with qib_tid's filled in, back
 441                  * to user.  Since we did the copy in already, this "should
 442                  * never fail" If it does, we have to clean up...
 443                  */
 444                 if (copy_to_user((void __user *)
 445                                  (unsigned long) ti->tidlist,
 446                                  tidlist, cnt * sizeof(*tidlist))) {
 447                         ret = -EFAULT;
 448                         goto cleanup;
 449                 }
 450                 if (copy_to_user((void __user *) (unsigned long) ti->tidmap,
 451                                  tidmap, sizeof(tidmap))) {
 452                         ret = -EFAULT;
 453                         goto cleanup;
 454                 }
 455                 if (tid == tidcnt)
 456                         tid = 0;
 457                 if (!rcd->subctxt_cnt)
 458                         rcd->tidcursor = tid;
 459                 else
 460                         tidcursor_fp(fp) = tid;
 461         }
 462
 463 done:
 464         return ret;
 465 }
 466
 467 /**
 468  * qib_tid_free - free a context TID
 469  * @rcd: the context
 470  * @subctxt: the subcontext
 471  * @ti: the TID info
 472  *
 473  * right now we are unlocking one page at a time, but since
 474  * the intended use of this routine is for a single group of
 475  * virtually contiguous pages, that should change to improve
 476  * performance.  We check that the TID is in range for this context
 477  * but otherwise don't check validity; if user has an error and
 478  * frees the wrong tid, it's only their own data that can thereby
 479  * be corrupted.  We do check that the TID was in use, for sanity
 480  * We always use our idea of the saved address, not the address that
 481  * they pass in to us.
 482  */
 483 static int qib_tid_free(struct qib_ctxtdata *rcd, unsigned subctxt,
 484                         const struct qib_tid_info *ti)
 485 {
 486         int ret = 0;
 487         u32 tid, ctxttid, cnt, limit, tidcnt;
 488         struct qib_devdata *dd = rcd->dd;
 489         u64 __iomem *tidbase;
 490         unsigned long tidmap[8];
 491
 492         if (!dd->pageshadow) {
 493                 ret = -ENOMEM;
 494                 goto done;
 495         }
 496
 497         if (copy_from_user(tidmap, (void __user *)(unsigned long)ti->tidmap,
 498                            sizeof(tidmap))) {
 499                 ret = -EFAULT;
 500                 goto done;
 501         }
 502
 503         ctxttid = rcd->ctxt * dd->rcvtidcnt;
 504         if (!rcd->subctxt_cnt)
 505                 tidcnt = dd->rcvtidcnt;
 506         else if (!subctxt) {
 507                 tidcnt = (dd->rcvtidcnt / rcd->subctxt_cnt) +
 508                          (dd->rcvtidcnt % rcd->subctxt_cnt);
 509                 ctxttid += dd->rcvtidcnt - tidcnt;
 510         } else {
 511                 tidcnt = dd->rcvtidcnt / rcd->subctxt_cnt;
 512                 ctxttid += tidcnt * (subctxt - 1);
 513         }
 514         tidbase = (u64 __iomem *) ((char __iomem *)(dd->kregbase) +
 515                                    dd->rcvtidbase +
 516                                    ctxttid * sizeof(*tidbase));
 517
 518         limit = sizeof(tidmap) * BITS_PER_BYTE;
 519         if (limit > tidcnt)
 520                 /* just in case size changes in future */
 521                 limit = tidcnt;
 522         tid = find_first_bit(tidmap, limit);
 523         for (cnt = 0; tid < limit; tid++) {
 524                 /*
 525                  * small optimization; if we detect a run of 3 or so without
 526                  * any set, use find_first_bit again.  That's mainly to
 527                  * accelerate the case where we wrapped, so we have some at
 528                  * the beginning, and some at the end, and a big gap
 529                  * in the middle.
 530                  */
 531                 if (!test_bit(tid, tidmap))
 532                         continue;
 533                 cnt++;
 534                 if (dd->pageshadow[ctxttid + tid]) {
 535                         struct page *p;
 536                         dma_addr_t phys;
 537
 538                         p = dd->pageshadow[ctxttid + tid];
 539                         dd->pageshadow[ctxttid + tid] = NULL;
 540                         phys = dd->physshadow[ctxttid + tid];
 541                         dd->physshadow[ctxttid + tid] = dd->tidinvalid;
 542                         /* PERFORMANCE: below should almost certainly be
 543                          * cached
 544                          */
 545                         dd->f_put_tid(dd, &tidbase[tid],
 546                                       RCVHQ_RCV_TYPE_EXPECTED, dd->tidinvalid);
 547                         pci_unmap_page(dd->pcidev, phys, PAGE_SIZE,
 548                                        PCI_DMA_FROMDEVICE);
 549                         qib_release_user_pages(&p, 1);
 550                 }
 551         }
 552 done:
 553         return ret;
 554 }
 555
 556 /**
 557  * qib_set_part_key - set a partition key
 558  * @rcd: the context
 559  * @key: the key
 560  *
 561  * We can have up to 4 active at a time (other than the default, which is
 562  * always allowed).  This is somewhat tricky, since multiple contexts may set
 563  * the same key, so we reference count them, and clean up at exit.  All 4
 564  * partition keys are packed into a single qlogic_ib register.  It's an
 565  * error for a process to set the same pkey multiple times.  We provide no
 566  * mechanism to de-allocate a pkey at this time, we may eventually need to
 567  * do that.  I've used the atomic operations, and no locking, and only make
 568  * a single pass through what's available.  This should be more than
 569  * adequate for some time. I'll think about spinlocks or the like if and as
 570  * it's necessary.
 571  */
 572 static int qib_set_part_key(struct qib_ctxtdata *rcd, u16 key)
 573 {
 574         struct qib_pportdata *ppd = rcd->ppd;
 575         int i, any = 0, pidx = -1;
 576         u16 lkey = key & 0x7FFF;
 577         int ret;
 578
 579         if (lkey == (QIB_DEFAULT_P_KEY & 0x7FFF)) {
 580                 /* nothing to do; this key always valid */
 581                 ret = 0;
 582                 goto bail;
 583         }
 584
 585         if (!lkey) {
 586                 ret = -EINVAL;
 587                 goto bail;
 588         }
 589
 590         /*
 591          * Set the full membership bit, because it has to be
 592          * set in the register or the packet, and it seems
 593          * cleaner to set in the register than to force all
 594          * callers to set it.
 595          */
 596         key |= 0x8000;
 597
 598         for (i = 0; i < ARRAY_SIZE(rcd->pkeys); i++) {
 599                 if (!rcd->pkeys[i] && pidx == -1)
 600                         pidx = i;
 601                 if (rcd->pkeys[i] == key) {
 602                         ret = -EEXIST;
 603                         goto bail;
 604                 }
 605         }
 606         if (pidx == -1) {
 607                 ret = -EBUSY;
 608                 goto bail;
 609         }
 610         for (any = i = 0; i < ARRAY_SIZE(ppd->pkeys); i++) {
 611                 if (!ppd->pkeys[i]) {
 612                         any++;
 613                         continue;
 614                 }
 615                 if (ppd->pkeys[i] == key) {
 616                         atomic_t *pkrefs = &ppd->pkeyrefs[i];
 617
 618                         if (atomic_inc_return(pkrefs) > 1) {
 619                                 rcd->pkeys[pidx] = key;
 620                                 ret = 0;
 621                                 goto bail;
 622                         } else {
 623                                 /*
 624                                  * lost race, decrement count, catch below
 625                                  */
 626                                 atomic_dec(pkrefs);
 627                                 any++;
 628                         }
 629                 }
 630                 if ((ppd->pkeys[i] & 0x7FFF) == lkey) {
 631                         /*
 632                          * It makes no sense to have both the limited and
 633                          * full membership PKEY set at the same time since
 634                          * the unlimited one will disable the limited one.
 635                          */
 636                         ret = -EEXIST;
 637                         goto bail;
 638                 }
 639         }
 640         if (!any) {
 641                 ret = -EBUSY;
 642                 goto bail;
 643         }
 644         for (any = i = 0; i < ARRAY_SIZE(ppd->pkeys); i++) {
 645                 if (!ppd->pkeys[i] &&
 646                     atomic_inc_return(&ppd->pkeyrefs[i]) == 1) {
 647                         rcd->pkeys[pidx] = key;
 648                         ppd->pkeys[i] = key;
 649                         (void) ppd->dd->f_set_ib_cfg(ppd, QIB_IB_CFG_PKEYS, 0);
 650                         ret = 0;
 651                         goto bail;
 652                 }
 653         }
 654         ret = -EBUSY;
 655
 656 bail:
 657         return ret;
 658 }
 659
 660 /**
 661  * qib_manage_rcvq - manage a context's receive queue
 662  * @rcd: the context
 663  * @subctxt: the subcontext
 664  * @start_stop: action to carry out
 665  *
 666  * start_stop == 0 disables receive on the context, for use in queue
 667  * overflow conditions.  start_stop==1 re-enables, to be used to
 668  * re-init the software copy of the head register
 669  */
 670 static int qib_manage_rcvq(struct qib_ctxtdata *rcd, unsigned subctxt,
 671                            int start_stop)
 672 {
 673         struct qib_devdata *dd = rcd->dd;
 674         unsigned int rcvctrl_op;
 675
 676         if (subctxt)
 677                 goto bail;
 678         /* atomically clear receive enable ctxt. */
 679         if (start_stop) {
 680                 /*
 681                  * On enable, force in-memory copy of the tail register to
 682                  * 0, so that protocol code doesn't have to worry about
 683                  * whether or not the chip has yet updated the in-memory
 684                  * copy or not on return from the system call. The chip
 685                  * always resets it's tail register back to 0 on a
 686                  * transition from disabled to enabled.
 687                  */
 688                 if (rcd->rcvhdrtail_kvaddr)
 689                         qib_clear_rcvhdrtail(rcd);
 690                 rcvctrl_op = QIB_RCVCTRL_CTXT_ENB;
 691         } else
 692                 rcvctrl_op = QIB_RCVCTRL_CTXT_DIS;
 693         dd->f_rcvctrl(rcd->ppd, rcvctrl_op, rcd->ctxt);
 694         /* always; new head should be equal to new tail; see above */
 695 bail:
 696         return 0;
 697 }
 698
 699 static void qib_clean_part_key(struct qib_ctxtdata *rcd,
 700                                struct qib_devdata *dd)
 701 {
 702         int i, j, pchanged = 0;
 703         u64 oldpkey;
 704         struct qib_pportdata *ppd = rcd->ppd;
 705
 706         /* for debugging only */
 707         oldpkey = (u64) ppd->pkeys[0] |
 708                 ((u64) ppd->pkeys[1] << 16) |
 709                 ((u64) ppd->pkeys[2] << 32) |
 710                 ((u64) ppd->pkeys[3] << 48);
 711
 712         for (i = 0; i < ARRAY_SIZE(rcd->pkeys); i++) {
 713                 if (!rcd->pkeys[i])
 714                         continue;
 715                 for (j = 0; j < ARRAY_SIZE(ppd->pkeys); j++) {
 716                         /* check for match independent of the global bit */
 717                         if ((ppd->pkeys[j] & 0x7fff) !=
 718                             (rcd->pkeys[i] & 0x7fff))
 719                                 continue;
 720                         if (atomic_dec_and_test(&ppd->pkeyrefs[j])) {
 721                                 ppd->pkeys[j] = 0;
 722                                 pchanged++;
 723                         }
 724                         break;
 725                 }
 726                 rcd->pkeys[i] = 0;
 727         }
 728         if (pchanged)
 729                 (void) ppd->dd->f_set_ib_cfg(ppd, QIB_IB_CFG_PKEYS, 0);
 730 }
 731
 732 /* common code for the mappings on dma_alloc_coherent mem */
 733 static int qib_mmap_mem(struct vm_area_struct *vma, struct qib_ctxtdata *rcd,
 734                         unsigned len, void *kvaddr, u32 write_ok, char *what)
 735 {
 736         struct qib_devdata *dd = rcd->dd;
 737         unsigned long pfn;
 738         int ret;
 739
 740         if ((vma->vm_end - vma->vm_start) > len) {
 741                 qib_devinfo(dd->pcidev,
 742                          "FAIL on %s: len %lx > %x\n", what,
 743                          vma->vm_end - vma->vm_start, len);
 744                 ret = -EFAULT;
 745                 goto bail;
 746         }
 747
 748         /*
 749          * shared context user code requires rcvhdrq mapped r/w, others
 750          * only allowed readonly mapping.
 751          */
 752         if (!write_ok) {
 753                 if (vma->vm_flags & VM_WRITE) {
 754                         qib_devinfo(dd->pcidev,
 755                                  "%s must be mapped readonly\n", what);
 756                         ret = -EPERM;
 757                         goto bail;
 758                 }
 759
 760                 /* don't allow them to later change with mprotect */
 761                 vma->vm_flags &= ~VM_MAYWRITE;
 762         }
 763
 764         pfn = virt_to_phys(kvaddr) >> PAGE_SHIFT;
 765         ret = remap_pfn_range(vma, vma->vm_start, pfn,
 766                               len, vma->vm_page_prot);
 767         if (ret)
 768                 qib_devinfo(dd->pcidev,
 769                         "%s ctxt%u mmap of %lx, %x bytes failed: %d\n",
 770                         what, rcd->ctxt, pfn, len, ret);
 771 bail:
 772         return ret;
 773 }
 774
 775 static int mmap_ureg(struct vm_area_struct *vma, struct qib_devdata *dd,
 776                      u64 ureg)
 777 {
 778         unsigned long phys;
 779         unsigned long sz;
 780         int ret;
 781
 782         /*
 783          * This is real hardware, so use io_remap.  This is the mechanism
 784          * for the user process to update the head registers for their ctxt
 785          * in the chip.
 786          */
 787         sz = dd->flags & QIB_HAS_HDRSUPP ? 2 * PAGE_SIZE : PAGE_SIZE;
 788         if ((vma->vm_end - vma->vm_start) > sz) {
 789                 qib_devinfo(dd->pcidev,
 790                         "FAIL mmap userreg: reqlen %lx > PAGE\n",
 791                         vma->vm_end - vma->vm_start);
 792                 ret = -EFAULT;
 793         } else {
 794                 phys = dd->physaddr + ureg;
 795                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 796
 797                 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
 798                 ret = io_remap_pfn_range(vma, vma->vm_start,
 799                                          phys >> PAGE_SHIFT,
 800                                          vma->vm_end - vma->vm_start,
 801                                          vma->vm_page_prot);
 802         }
 803         return ret;
 804 }
 805
 806 static int mmap_piobufs(struct vm_area_struct *vma,
 807                         struct qib_devdata *dd,
 808                         struct qib_ctxtdata *rcd,
 809                         unsigned piobufs, unsigned piocnt)
 810 {
 811         unsigned long phys;
 812         int ret;
 813
 814         /*
 815          * When we map the PIO buffers in the chip, we want to map them as
 816          * writeonly, no read possible; unfortunately, x86 doesn't allow
 817          * for this in hardware, but we still prevent users from asking
 818          * for it.
 819          */
 820         if ((vma->vm_end - vma->vm_start) > (piocnt * dd->palign)) {
 821                 qib_devinfo(dd->pcidev,
 822                         "FAIL mmap piobufs: reqlen %lx > PAGE\n",
 823                          vma->vm_end - vma->vm_start);
 824                 ret = -EINVAL;
 825                 goto bail;
 826         }
 827
 828         phys = dd->physaddr + piobufs;
 829
 830 #if defined(__powerpc__)
 831         /* There isn't a generic way to specify writethrough mappings */
 832         pgprot_val(vma->vm_page_prot) |= _PAGE_NO_CACHE;
 833         pgprot_val(vma->vm_page_prot) |= _PAGE_WRITETHRU;
 834         pgprot_val(vma->vm_page_prot) &= ~_PAGE_GUARDED;
 835 #endif
 836
 837         /*
 838          * don't allow them to later change to readable with mprotect (for when
 839          * not initially mapped readable, as is normally the case)
 840          */
 841         vma->vm_flags &= ~VM_MAYREAD;
 842         vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
 843
 844         /* We used PAT if wc_cookie == 0 */
 845         if (!dd->wc_cookie)
 846                 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
 847
 848         ret = io_remap_pfn_range(vma, vma->vm_start, phys >> PAGE_SHIFT,
 849                                  vma->vm_end - vma->vm_start,
 850                                  vma->vm_page_prot);
 851 bail:
 852         return ret;
 853 }
 854
 855 static int mmap_rcvegrbufs(struct vm_area_struct *vma,
 856                            struct qib_ctxtdata *rcd)
 857 {
 858         struct qib_devdata *dd = rcd->dd;
 859         unsigned long start, size;
 860         size_t total_size, i;
 861         unsigned long pfn;
 862         int ret;
 863
 864         size = rcd->rcvegrbuf_size;
 865         total_size = rcd->rcvegrbuf_chunks * size;
 866         if ((vma->vm_end - vma->vm_start) > total_size) {
 867                 qib_devinfo(dd->pcidev,
 868                         "FAIL on egr bufs: reqlen %lx > actual %lx\n",
 869                          vma->vm_end - vma->vm_start,
 870                          (unsigned long) total_size);
 871                 ret = -EINVAL;
 872                 goto bail;
 873         }
 874
 875         if (vma->vm_flags & VM_WRITE) {
 876                 qib_devinfo(dd->pcidev,
 877                         "Can't map eager buffers as writable (flags=%lx)\n",
 878                         vma->vm_flags);
 879                 ret = -EPERM;
 880                 goto bail;
 881         }
 882         /* don't allow them to later change to writeable with mprotect */
 883         vma->vm_flags &= ~VM_MAYWRITE;
 884
 885         start = vma->vm_start;
 886
 887         for (i = 0; i < rcd->rcvegrbuf_chunks; i++, start += size) {
 888                 pfn = virt_to_phys(rcd->rcvegrbuf[i]) >> PAGE_SHIFT;
 889                 ret = remap_pfn_range(vma, start, pfn, size,
 890                                       vma->vm_page_prot);
 891                 if (ret < 0)
 892                         goto bail;
 893         }
 894         ret = 0;
 895
 896 bail:
 897         return ret;
 898 }
 899
 900 /*
 901  * qib_file_vma_fault - handle a VMA page fault.
 902  */
 903 static int qib_file_vma_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 904 {
 905         struct page *page;
 906
 907         page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
 908         if (!page)
 909                 return VM_FAULT_SIGBUS;
 910
 911         get_page(page);
 912         vmf->page = page;
 913
 914         return 0;
 915 }
 916
 917 static const struct vm_operations_struct qib_file_vm_ops = {
 918         .fault = qib_file_vma_fault,
 919 };
 920
 921 static int mmap_kvaddr(struct vm_area_struct *vma, u64 pgaddr,
 922                        struct qib_ctxtdata *rcd, unsigned subctxt)
 923 {
 924         struct qib_devdata *dd = rcd->dd;
 925         unsigned subctxt_cnt;
 926         unsigned long len;
 927         void *addr;
 928         size_t size;
 929         int ret = 0;
 930
 931         subctxt_cnt = rcd->subctxt_cnt;
 932         size = rcd->rcvegrbuf_chunks * rcd->rcvegrbuf_size;
 933
 934         /*
 935          * Each process has all the subctxt uregbase, rcvhdrq, and
 936          * rcvegrbufs mmapped - as an array for all the processes,
 937          * and also separately for this process.
 938          */
 939         if (pgaddr == cvt_kvaddr(rcd->subctxt_uregbase)) {
 940                 addr = rcd->subctxt_uregbase;
 941                 size = PAGE_SIZE * subctxt_cnt;
 942         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_rcvhdr_base)) {
 943                 addr = rcd->subctxt_rcvhdr_base;
 944                 size = rcd->rcvhdrq_size * subctxt_cnt;
 945         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_rcvegrbuf)) {
 946                 addr = rcd->subctxt_rcvegrbuf;
 947                 size *= subctxt_cnt;
 948         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_uregbase +
 949                                         PAGE_SIZE * subctxt)) {
 950                 addr = rcd->subctxt_uregbase + PAGE_SIZE * subctxt;
 951                 size = PAGE_SIZE;
 952         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_rcvhdr_base +
 953                                         rcd->rcvhdrq_size * subctxt)) {
 954                 addr = rcd->subctxt_rcvhdr_base +
 955                         rcd->rcvhdrq_size * subctxt;
 956                 size = rcd->rcvhdrq_size;
 957         } else if (pgaddr == cvt_kvaddr(&rcd->user_event_mask[subctxt])) {
 958                 addr = rcd->user_event_mask;
 959                 size = PAGE_SIZE;
 960         } else if (pgaddr == cvt_kvaddr(rcd->subctxt_rcvegrbuf +
 961                                         size * subctxt)) {
 962                 addr = rcd->subctxt_rcvegrbuf + size * subctxt;
 963                 /* rcvegrbufs are read-only on the slave */
 964                 if (vma->vm_flags & VM_WRITE) {
 965                         qib_devinfo(dd->pcidev,
 966                                  "Can't map eager buffers as writable (flags=%lx)\n",
 967                                  vma->vm_flags);
 968                         ret = -EPERM;
 969                         goto bail;
 970                 }
 971                 /*
 972                  * Don't allow permission to later change to writeable
 973                  * with mprotect.
 974                  */
 975                 vma->vm_flags &= ~VM_MAYWRITE;
 976         } else
 977                 goto bail;
 978         len = vma->vm_end - vma->vm_start;
 979         if (len > size) {
 980                 ret = -EINVAL;
 981                 goto bail;
 982         }
 983
 984         vma->vm_pgoff = (unsigned long) addr >> PAGE_SHIFT;
 985         vma->vm_ops = &qib_file_vm_ops;
 986         vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
 987         ret = 1;
 988
 989 bail:
 990         return ret;
 991 }
 992
 993 /**
 994  * qib_mmapf - mmap various structures into user space
 995  * @fp: the file pointer
 996  * @vma: the VM area
 997  *
 998  * We use this to have a shared buffer between the kernel and the user code
 999  * for the rcvhdr queue, egr buffers, and the per-context user regs and pio
1000  * buffers in the chip.  We have the open and close entries so we can bump
1001  * the ref count and keep the driver from being unloaded while still mapped.
1002  */
1003 static int qib_mmapf(struct file *fp, struct vm_area_struct *vma)
1004 {
1005         struct qib_ctxtdata *rcd;
1006         struct qib_devdata *dd;
1007         u64 pgaddr, ureg;
1008         unsigned piobufs, piocnt;
1009         int ret, match = 1;
1010
1011         rcd = ctxt_fp(fp);
1012         if (!rcd || !(vma->vm_flags & VM_SHARED)) {
1013                 ret = -EINVAL;
1014                 goto bail;
1015         }
1016         dd = rcd->dd;
1017
1018         /*
1019          * This is the qib_do_user_init() code, mapping the shared buffers
1020          * and per-context user registers into the user process. The address
1021          * referred to by vm_pgoff is the file offset passed via mmap().
1022          * For shared contexts, this is the kernel vmalloc() address of the
1023          * pages to share with the master.
1024          * For non-shared or master ctxts, this is a physical address.
1025          * We only do one mmap for each space mapped.
1026          */
1027         pgaddr = vma->vm_pgoff << PAGE_SHIFT;
1028
1029         /*
1030          * Check for 0 in case one of the allocations failed, but user
1031          * called mmap anyway.
1032          */
1033         if (!pgaddr)  {
1034                 ret = -EINVAL;
1035                 goto bail;
1036         }
1037
1038         /*
1039          * Physical addresses must fit in 40 bits for our hardware.
1040          * Check for kernel virtual addresses first, anything else must
1041          * match a HW or memory address.
1042          */
1043         ret = mmap_kvaddr(vma, pgaddr, rcd, subctxt_fp(fp));
1044         if (ret) {
1045                 if (ret > 0)
1046                         ret = 0;
1047                 goto bail;
1048         }
1049
1050         ureg = dd->uregbase + dd->ureg_align * rcd->ctxt;
1051         if (!rcd->subctxt_cnt) {
1052                 /* ctxt is not shared */
1053                 piocnt = rcd->piocnt;
1054                 piobufs = rcd->piobufs;
1055         } else if (!subctxt_fp(fp)) {
1056                 /* caller is the master */
1057                 piocnt = (rcd->piocnt / rcd->subctxt_cnt) +
1058                          (rcd->piocnt % rcd->subctxt_cnt);
1059                 piobufs = rcd->piobufs +
1060                         dd->palign * (rcd->piocnt - piocnt);
1061         } else {
1062                 unsigned slave = subctxt_fp(fp) - 1;
1063
1064                 /* caller is a slave */
1065                 piocnt = rcd->piocnt / rcd->subctxt_cnt;
1066                 piobufs = rcd->piobufs + dd->palign * piocnt * slave;
1067         }
1068
1069         if (pgaddr == ureg)
1070                 ret = mmap_ureg(vma, dd, ureg);
1071         else if (pgaddr == piobufs)
1072                 ret = mmap_piobufs(vma, dd, rcd, piobufs, piocnt);
1073         else if (pgaddr == dd->pioavailregs_phys)
1074                 /* in-memory copy of pioavail registers */
1075                 ret = qib_mmap_mem(vma, rcd, PAGE_SIZE,
1076                                    (void *) dd->pioavailregs_dma, 0,
1077                                    "pioavail registers");
1078         else if (pgaddr == rcd->rcvegr_phys)
1079                 ret = mmap_rcvegrbufs(vma, rcd);
1080         else if (pgaddr == (u64) rcd->rcvhdrq_phys)
1081                 /*
1082                  * The rcvhdrq itself; multiple pages, contiguous
1083                  * from an i/o perspective.  Shared contexts need
1084                  * to map r/w, so we allow writing.
1085                  */
1086                 ret = qib_mmap_mem(vma, rcd, rcd->rcvhdrq_size,
1087                                    rcd->rcvhdrq, 1, "rcvhdrq");
1088         else if (pgaddr == (u64) rcd->rcvhdrqtailaddr_phys)
1089                 /* in-memory copy of rcvhdrq tail register */
1090                 ret = qib_mmap_mem(vma, rcd, PAGE_SIZE,
1091                                    rcd->rcvhdrtail_kvaddr, 0,
1092                                    "rcvhdrq tail");
1093         else
1094                 match = 0;
1095         if (!match)
1096                 ret = -EINVAL;
1097
1098         vma->vm_private_data = NULL;
1099
1100         if (ret < 0)
1101                 qib_devinfo(dd->pcidev,
1102                          "mmap Failure %d: off %llx len %lx\n",
1103                          -ret, (unsigned long long)pgaddr,
1104                          vma->vm_end - vma->vm_start);
1105 bail:
1106         return ret;
1107 }
1108
1109 static unsigned int qib_poll_urgent(struct qib_ctxtdata *rcd,
1110                                     struct file *fp,
1111                                     struct poll_table_struct *pt)
1112 {
1113         struct qib_devdata *dd = rcd->dd;
1114         unsigned pollflag;
1115
1116         poll_wait(fp, &rcd->wait, pt);
1117
1118         spin_lock_irq(&dd->uctxt_lock);
1119         if (rcd->urgent != rcd->urgent_poll) {
1120                 pollflag = POLLIN | POLLRDNORM;
1121                 rcd->urgent_poll = rcd->urgent;
1122         } else {
1123                 pollflag = 0;
1124                 set_bit(QIB_CTXT_WAITING_URG, &rcd->flag);
1125         }
1126         spin_unlock_irq(&dd->uctxt_lock);
1127
1128         return pollflag;
1129 }
1130
1131 static unsigned int qib_poll_next(struct qib_ctxtdata *rcd,
1132                                   struct file *fp,
1133                                   struct poll_table_struct *pt)
1134 {
1135         struct qib_devdata *dd = rcd->dd;
1136         unsigned pollflag;
1137
1138         poll_wait(fp, &rcd->wait, pt);
1139
1140         spin_lock_irq(&dd->uctxt_lock);
1141         if (dd->f_hdrqempty(rcd)) {
1142                 set_bit(QIB_CTXT_WAITING_RCV, &rcd->flag);
1143                 dd->f_rcvctrl(rcd->ppd, QIB_RCVCTRL_INTRAVAIL_ENB, rcd->ctxt);
1144                 pollflag = 0;
1145         } else
1146                 pollflag = POLLIN | POLLRDNORM;
1147         spin_unlock_irq(&dd->uctxt_lock);
1148
1149         return pollflag;
1150 }
1151
1152 static unsigned int qib_poll(struct file *fp, struct poll_table_struct *pt)
1153 {
1154         struct qib_ctxtdata *rcd;
1155         unsigned pollflag;
1156
1157         rcd = ctxt_fp(fp);
1158         if (!rcd)
1159                 pollflag = POLLERR;
1160         else if (rcd->poll_type == QIB_POLL_TYPE_URGENT)
1161                 pollflag = qib_poll_urgent(rcd, fp, pt);
1162         else  if (rcd->poll_type == QIB_POLL_TYPE_ANYRCV)
1163                 pollflag = qib_poll_next(rcd, fp, pt);
1164         else /* invalid */
1165                 pollflag = POLLERR;
1166
1167         return pollflag;
1168 }
1169
1170 static void assign_ctxt_affinity(struct file *fp, struct qib_devdata *dd)
1171 {
1172         struct qib_filedata *fd = fp->private_data;
1173         const unsigned int weight = cpumask_weight(&current->cpus_allowed);
1174         const struct cpumask *local_mask = cpumask_of_pcibus(dd->pcidev->bus);
1175         int local_cpu;
1176
1177         /*
1178          * If process has NOT already set it's affinity, select and
1179          * reserve a processor for it on the local NUMA node.
1180          */
1181         if ((weight >= qib_cpulist_count) &&
1182                 (cpumask_weight(local_mask) <= qib_cpulist_count)) {
1183                 for_each_cpu(local_cpu, local_mask)
1184                         if (!test_and_set_bit(local_cpu, qib_cpulist)) {
1185                                 fd->rec_cpu_num = local_cpu;
1186                                 return;
1187                         }
1188         }
1189
1190         /*
1191          * If process has NOT already set it's affinity, select and
1192          * reserve a processor for it, as a rendevous for all
1193          * users of the driver.  If they don't actually later
1194          * set affinity to this cpu, or set it to some other cpu,
1195          * it just means that sooner or later we don't recommend
1196          * a cpu, and let the scheduler do it's best.
1197          */
1198         if (weight >= qib_cpulist_count) {
1199                 int cpu;
1200
1201                 cpu = find_first_zero_bit(qib_cpulist,
1202                                           qib_cpulist_count);
1203                 if (cpu == qib_cpulist_count)
1204                         qib_dev_err(dd,
1205                         "no cpus avail for affinity PID %u\n",
1206                         current->pid);
1207                 else {
1208                         __set_bit(cpu, qib_cpulist);
1209                         fd->rec_cpu_num = cpu;
1210                 }
1211         }
1212 }
1213
1214 /*
1215  * Check that userland and driver are compatible for subcontexts.
1216  */
1217 static int qib_compatible_subctxts(int user_swmajor, int user_swminor)
1218 {
1219         /* this code is written long-hand for clarity */
1220         if (QIB_USER_SWMAJOR != user_swmajor) {
1221                 /* no promise of compatibility if major mismatch */
1222                 return 0;
1223         }
1224         if (QIB_USER_SWMAJOR == 1) {
1225                 switch (QIB_USER_SWMINOR) {
1226                 case 0:
1227                 case 1:
1228                 case 2:
1229                         /* no subctxt implementation so cannot be compatible */
1230                         return 0;
1231                 case 3:
1232                         /* 3 is only compatible with itself */
1233                         return user_swminor == 3;
1234                 default:
1235                         /* >= 4 are compatible (or are expected to be) */
1236                         return user_swminor <= QIB_USER_SWMINOR;
1237                 }
1238         }
1239         /* make no promises yet for future major versions */
1240         return 0;
1241 }
1242
1243 static int init_subctxts(struct qib_devdata *dd,
1244                          struct qib_ctxtdata *rcd,
1245                          const struct qib_user_info *uinfo)
1246 {
1247         int ret = 0;
1248         unsigned num_subctxts;
1249         size_t size;
1250
1251         /*
1252          * If the user is requesting zero subctxts,
1253          * skip the subctxt allocation.
1254          */
1255         if (uinfo->spu_subctxt_cnt <= 0)
1256                 goto bail;
1257         num_subctxts = uinfo->spu_subctxt_cnt;
1258
1259         /* Check for subctxt compatibility */
1260         if (!qib_compatible_subctxts(uinfo->spu_userversion >> 16,
1261                 uinfo->spu_userversion & 0xffff)) {
1262                 qib_devinfo(dd->pcidev,
1263                          "Mismatched user version (%d.%d) and driver version (%d.%d) while context sharing. Ensure that driver and library are from the same release.\n",
1264                          (int) (uinfo->spu_userversion >> 16),
1265                          (int) (uinfo->spu_userversion & 0xffff),
1266                          QIB_USER_SWMAJOR, QIB_USER_SWMINOR);
1267                 goto bail;
1268         }
1269         if (num_subctxts > QLOGIC_IB_MAX_SUBCTXT) {
1270                 ret = -EINVAL;
1271                 goto bail;
1272         }
1273
1274         rcd->subctxt_uregbase = vmalloc_user(PAGE_SIZE * num_subctxts);
1275         if (!rcd->subctxt_uregbase) {
1276                 ret = -ENOMEM;
1277                 goto bail;
1278         }
1279         /* Note: rcd->rcvhdrq_size isn't initialized yet. */
1280         size = ALIGN(dd->rcvhdrcnt * dd->rcvhdrentsize *
1281                      sizeof(u32), PAGE_SIZE) * num_subctxts;
1282         rcd->subctxt_rcvhdr_base = vmalloc_user(size);
1283         if (!rcd->subctxt_rcvhdr_base) {
1284                 ret = -ENOMEM;
1285                 goto bail_ureg;
1286         }
1287
1288         rcd->subctxt_rcvegrbuf = vmalloc_user(rcd->rcvegrbuf_chunks *
1289                                               rcd->rcvegrbuf_size *
1290                                               num_subctxts);
1291         if (!rcd->subctxt_rcvegrbuf) {
1292                 ret = -ENOMEM;
1293                 goto bail_rhdr;
1294         }
1295
1296         rcd->subctxt_cnt = uinfo->spu_subctxt_cnt;
1297         rcd->subctxt_id = uinfo->spu_subctxt_id;
1298         rcd->active_slaves = 1;
1299         rcd->redirect_seq_cnt = 1;
1300         set_bit(QIB_CTXT_MASTER_UNINIT, &rcd->flag);
1301         goto bail;
1302
1303 bail_rhdr:
1304         vfree(rcd->subctxt_rcvhdr_base);
1305 bail_ureg:
1306         vfree(rcd->subctxt_uregbase);
1307         rcd->subctxt_uregbase = NULL;
1308 bail:
1309         return ret;
1310 }
1311
1312 static int setup_ctxt(struct qib_pportdata *ppd, int ctxt,
1313                       struct file *fp, const struct qib_user_info *uinfo)
1314 {
1315         struct qib_filedata *fd = fp->private_data;
1316         struct qib_devdata *dd = ppd->dd;
1317         struct qib_ctxtdata *rcd;
1318         void *ptmp = NULL;
1319         int ret;
1320         int numa_id;
1321
1322         assign_ctxt_affinity(fp, dd);
1323
1324         numa_id = qib_numa_aware ? ((fd->rec_cpu_num != -1) ?
1325                 cpu_to_node(fd->rec_cpu_num) :
1326                 numa_node_id()) : dd->assigned_node_id;
1327
1328         rcd = qib_create_ctxtdata(ppd, ctxt, numa_id);
1329
1330         /*
1331          * Allocate memory for use in qib_tid_update() at open to
1332          * reduce cost of expected send setup per message segment
1333          */
1334         if (rcd)
1335                 ptmp = kmalloc(dd->rcvtidcnt * sizeof(u16) +
1336                                dd->rcvtidcnt * sizeof(struct page **),
1337                                GFP_KERNEL);
1338
1339         if (!rcd || !ptmp) {
1340                 qib_dev_err(dd,
1341                         "Unable to allocate ctxtdata memory, failing open\n");
1342                 ret = -ENOMEM;
1343                 goto bailerr;
1344         }
1345         rcd->userversion = uinfo->spu_userversion;
1346         ret = init_subctxts(dd, rcd, uinfo);
1347         if (ret)
1348                 goto bailerr;
1349         rcd->tid_pg_list = ptmp;
1350         rcd->pid = current->pid;
1351         init_waitqueue_head(&dd->rcd[ctxt]->wait);
1352         strlcpy(rcd->comm, current->comm, sizeof(rcd->comm));
1353         ctxt_fp(fp) = rcd;
1354         qib_stats.sps_ctxts++;
1355         dd->freectxts--;
1356         ret = 0;
1357         goto bail;
1358
1359 bailerr:
1360         if (fd->rec_cpu_num != -1)
1361                 __clear_bit(fd->rec_cpu_num, qib_cpulist);
1362
1363         dd->rcd[ctxt] = NULL;
1364         kfree(rcd);
1365         kfree(ptmp);
1366 bail:
1367         return ret;
1368 }
1369
1370 static inline int usable(struct qib_pportdata *ppd)
1371 {
1372         struct qib_devdata *dd = ppd->dd;
1373
1374         return dd && (dd->flags & QIB_PRESENT) && dd->kregbase && ppd->lid &&
1375                 (ppd->lflags & QIBL_LINKACTIVE);
1376 }
1377
1378 /*
1379  * Select a context on the given device, either using a requested port
1380  * or the port based on the context number.
1381  */
1382 static int choose_port_ctxt(struct file *fp, struct qib_devdata *dd, u32 port,
1383                             const struct qib_user_info *uinfo)
1384 {
1385         struct qib_pportdata *ppd = NULL;
1386         int ret, ctxt;
1387
1388         if (port) {
1389                 if (!usable(dd->pport + port - 1)) {
1390                         ret = -ENETDOWN;
1391                         goto done;
1392                 } else
1393                         ppd = dd->pport + port - 1;
1394         }
1395         for (ctxt = dd->first_user_ctxt; ctxt < dd->cfgctxts && dd->rcd[ctxt];
1396              ctxt++)
1397                 ;
1398         if (ctxt == dd->cfgctxts) {
1399                 ret = -EBUSY;
1400                 goto done;
1401         }
1402         if (!ppd) {
1403                 u32 pidx = ctxt % dd->num_pports;
1404
1405                 if (usable(dd->pport + pidx))
1406                         ppd = dd->pport + pidx;
1407                 else {
1408                         for (pidx = 0; pidx < dd->num_pports && !ppd;
1409                              pidx++)
1410                                 if (usable(dd->pport + pidx))
1411                                         ppd = dd->pport + pidx;
1412                 }
1413         }
1414         ret = ppd ? setup_ctxt(ppd, ctxt, fp, uinfo) : -ENETDOWN;
1415 done:
1416         return ret;
1417 }
1418
1419 static int find_free_ctxt(int unit, struct file *fp,
1420                           const struct qib_user_info *uinfo)
1421 {
1422         struct qib_devdata *dd = qib_lookup(unit);
1423         int ret;
1424
1425         if (!dd || (uinfo->spu_port && uinfo->spu_port > dd->num_pports))
1426                 ret = -ENODEV;
1427         else
1428                 ret = choose_port_ctxt(fp, dd, uinfo->spu_port, uinfo);
1429
1430         return ret;
1431 }
1432
1433 static int get_a_ctxt(struct file *fp, const struct qib_user_info *uinfo,
1434                       unsigned alg)
1435 {
1436         struct qib_devdata *udd = NULL;
1437         int ret = 0, devmax, npresent, nup, ndev, dusable = 0, i;
1438         u32 port = uinfo->spu_port, ctxt;
1439
1440         devmax = qib_count_units(&npresent, &nup);
1441         if (!npresent) {
1442                 ret = -ENXIO;
1443                 goto done;
1444         }
1445         if (nup == 0) {
1446                 ret = -ENETDOWN;
1447                 goto done;
1448         }
1449
1450         if (alg == QIB_PORT_ALG_ACROSS) {
1451                 unsigned inuse = ~0U;
1452
1453                 /* find device (with ACTIVE ports) with fewest ctxts in use */
1454                 for (ndev = 0; ndev < devmax; ndev++) {
1455                         struct qib_devdata *dd = qib_lookup(ndev);
1456                         unsigned cused = 0, cfree = 0, pusable = 0;
1457
1458                         if (!dd)
1459                                 continue;
1460                         if (port && port <= dd->num_pports &&
1461                             usable(dd->pport + port - 1))
1462                                 pusable = 1;
1463                         else
1464                                 for (i = 0; i < dd->num_pports; i++)
1465                                         if (usable(dd->pport + i))
1466                                                 pusable++;
1467                         if (!pusable)
1468                                 continue;
1469                         for (ctxt = dd->first_user_ctxt; ctxt < dd->cfgctxts;
1470                              ctxt++)
1471                                 if (dd->rcd[ctxt])
1472                                         cused++;
1473                                 else
1474                                         cfree++;
1475                         if (cfree && cused < inuse) {
1476                                 udd = dd;
1477                                 inuse = cused;
1478                         }
1479                 }
1480                 if (udd) {
1481                         ret = choose_port_ctxt(fp, udd, port, uinfo);
1482                         goto done;
1483                 }
1484         } else {
1485                 for (ndev = 0; ndev < devmax; ndev++) {
1486                         struct qib_devdata *dd = qib_lookup(ndev);
1487
1488                         if (dd) {
1489                                 ret = choose_port_ctxt(fp, dd, port, uinfo);
1490                                 if (!ret)
1491                                         goto done;
1492                                 if (ret == -EBUSY)
1493                                         dusable++;
1494                         }
1495                 }
1496         }
1497         ret = dusable ? -EBUSY : -ENETDOWN;
1498
1499 done:
1500         return ret;
1501 }
1502
1503 static int find_shared_ctxt(struct file *fp,
1504                             const struct qib_user_info *uinfo)
1505 {
1506         int devmax, ndev, i;
1507         int ret = 0;
1508
1509         devmax = qib_count_units(NULL, NULL);
1510
1511         for (ndev = 0; ndev < devmax; ndev++) {
1512                 struct qib_devdata *dd = qib_lookup(ndev);
1513
1514                 /* device portion of usable() */
1515                 if (!(dd && (dd->flags & QIB_PRESENT) && dd->kregbase))
1516                         continue;
1517                 for (i = dd->first_user_ctxt; i < dd->cfgctxts; i++) {
1518                         struct qib_ctxtdata *rcd = dd->rcd[i];
1519
1520                         /* Skip ctxts which are not yet open */
1521                         if (!rcd || !rcd->cnt)
1522                                 continue;
1523                         /* Skip ctxt if it doesn't match the requested one */
1524                         if (rcd->subctxt_id != uinfo->spu_subctxt_id)
1525                                 continue;
1526                         /* Verify the sharing process matches the master */
1527                         if (rcd->subctxt_cnt != uinfo->spu_subctxt_cnt ||
1528                             rcd->userversion != uinfo->spu_userversion ||
1529                             rcd->cnt >= rcd->subctxt_cnt) {
1530                                 ret = -EINVAL;
1531                                 goto done;
1532                         }
1533                         ctxt_fp(fp) = rcd;
1534                         subctxt_fp(fp) = rcd->cnt++;
1535                         rcd->subpid[subctxt_fp(fp)] = current->pid;
1536                         tidcursor_fp(fp) = 0;
1537                         rcd->active_slaves |= 1 << subctxt_fp(fp);
1538                         ret = 1;
1539                         goto done;
1540                 }
1541         }
1542
1543 done:
1544         return ret;
1545 }
1546
1547 static int qib_open(struct inode *in, struct file *fp)
1548 {
1549         /* The real work is performed later in qib_assign_ctxt() */
1550         fp->private_data = kzalloc(sizeof(struct qib_filedata), GFP_KERNEL);
1551         if (fp->private_data) /* no cpu affinity by default */
1552                 ((struct qib_filedata *)fp->private_data)->rec_cpu_num = -1;
1553         return fp->private_data ? 0 : -ENOMEM;
1554 }
1555
1556 static int find_hca(unsigned int cpu, int *unit)
1557 {
1558         int ret = 0, devmax, npresent, nup, ndev;
1559
1560         *unit = -1;
1561
1562         devmax = qib_count_units(&npresent, &nup);
1563         if (!npresent) {
1564                 ret = -ENXIO;
1565                 goto done;
1566         }
1567         if (!nup) {
1568                 ret = -ENETDOWN;
1569                 goto done;
1570         }
1571         for (ndev = 0; ndev < devmax; ndev++) {
1572                 struct qib_devdata *dd = qib_lookup(ndev);
1573
1574                 if (dd) {
1575                         if (pcibus_to_node(dd->pcidev->bus) < 0) {
1576                                 ret = -EINVAL;
1577                                 goto done;
1578                         }
1579                         if (cpu_to_node(cpu) ==
1580                                 pcibus_to_node(dd->pcidev->bus)) {
1581                                 *unit = ndev;
1582                                 goto done;
1583                         }
1584                 }
1585         }
1586 done:
1587         return ret;
1588 }
1589
1590 static int do_qib_user_sdma_queue_create(struct file *fp)
1591 {
1592         struct qib_filedata *fd = fp->private_data;
1593         struct qib_ctxtdata *rcd = fd->rcd;
1594         struct qib_devdata *dd = rcd->dd;
1595
1596         if (dd->flags & QIB_HAS_SEND_DMA) {
1597
1598                 fd->pq = qib_user_sdma_queue_create(&dd->pcidev->dev,
1599                                                     dd->unit,
1600                                                     rcd->ctxt,
1601                                                     fd->subctxt);
1602                 if (!fd->pq)
1603                         return -ENOMEM;
1604         }
1605
1606         return 0;
1607 }
1608
1609 /*
1610  * Get ctxt early, so can set affinity prior to memory allocation.
1611  */
1612 static int qib_assign_ctxt(struct file *fp, const struct qib_user_info *uinfo)
1613 {
1614         int ret;
1615         int i_minor;
1616         unsigned swmajor, swminor, alg = QIB_PORT_ALG_ACROSS;
1617
1618         /* Check to be sure we haven't already initialized this file */
1619         if (ctxt_fp(fp)) {
1620                 ret = -EINVAL;
1621                 goto done;
1622         }
1623
1624         /* for now, if major version is different, bail */
1625         swmajor = uinfo->spu_userversion >> 16;
1626         if (swmajor != QIB_USER_SWMAJOR) {
1627                 ret = -ENODEV;
1628                 goto done;
1629         }
1630
1631         swminor = uinfo->spu_userversion & 0xffff;
1632
1633         if (swminor >= 11 && uinfo->spu_port_alg < QIB_PORT_ALG_COUNT)
1634                 alg = uinfo->spu_port_alg;
1635
1636         mutex_lock(&qib_mutex);
1637
1638         if (qib_compatible_subctxts(swmajor, swminor) &&
1639             uinfo->spu_subctxt_cnt) {
1640                 ret = find_shared_ctxt(fp, uinfo);
1641                 if (ret > 0) {
1642                         ret = do_qib_user_sdma_queue_create(fp);
1643                         if (!ret)
1644                                 assign_ctxt_affinity(fp, (ctxt_fp(fp))->dd);
1645                         goto done_ok;
1646                 }
1647         }
1648
1649         i_minor = iminor(file_inode(fp)) - QIB_USER_MINOR_BASE;
1650         if (i_minor)
1651                 ret = find_free_ctxt(i_minor - 1, fp, uinfo);
1652         else {
1653                 int unit;
1654                 const unsigned int cpu = cpumask_first(&current->cpus_allowed);
1655                 const unsigned int weight =
1656                         cpumask_weight(&current->cpus_allowed);
1657
1658                 if (weight == 1 && !test_bit(cpu, qib_cpulist))
1659                         if (!find_hca(cpu, &unit) && unit >= 0)
1660                                 if (!find_free_ctxt(unit, fp, uinfo)) {
1661                                         ret = 0;
1662                                         goto done_chk_sdma;
1663                                 }
1664                 ret = get_a_ctxt(fp, uinfo, alg);
1665         }
1666
1667 done_chk_sdma:
1668         if (!ret)
1669                 ret = do_qib_user_sdma_queue_create(fp);
1670 done_ok:
1671         mutex_unlock(&qib_mutex);
1672
1673 done:
1674         return ret;
1675 }
1676
1677
1678 static int qib_do_user_init(struct file *fp,
1679                             const struct qib_user_info *uinfo)
1680 {
1681         int ret;
1682         struct qib_ctxtdata *rcd = ctxt_fp(fp);
1683         struct qib_devdata *dd;
1684         unsigned uctxt;
1685
1686         /* Subctxts don't need to initialize anything since master did it. */
1687         if (subctxt_fp(fp)) {
1688                 ret = wait_event_interruptible(rcd->wait,
1689                         !test_bit(QIB_CTXT_MASTER_UNINIT, &rcd->flag));
1690                 goto bail;
1691         }
1692
1693         dd = rcd->dd;
1694
1695         /* some ctxts may get extra buffers, calculate that here */
1696         uctxt = rcd->ctxt - dd->first_user_ctxt;
1697         if (uctxt < dd->ctxts_extrabuf) {
1698                 rcd->piocnt = dd->pbufsctxt + 1;
1699                 rcd->pio_base = rcd->piocnt * uctxt;
1700         } else {
1701                 rcd->piocnt = dd->pbufsctxt;
1702                 rcd->pio_base = rcd->piocnt * uctxt +
1703                         dd->ctxts_extrabuf;
1704         }
1705
1706         /*
1707          * All user buffers are 2KB buffers.  If we ever support
1708          * giving 4KB buffers to user processes, this will need some
1709          * work.  Can't use piobufbase directly, because it has
1710          * both 2K and 4K buffer base values.  So check and handle.
1711          */
1712         if ((rcd->pio_base + rcd->piocnt) > dd->piobcnt2k) {
1713                 if (rcd->pio_base >= dd->piobcnt2k) {
1714                         qib_dev_err(dd,
1715                                     "%u:ctxt%u: no 2KB buffers available\n",
1716                                     dd->unit, rcd->ctxt);
1717                         ret = -ENOBUFS;
1718                         goto bail;
1719                 }
1720                 rcd->piocnt = dd->piobcnt2k - rcd->pio_base;
1721                 qib_dev_err(dd, "Ctxt%u: would use 4KB bufs, using %u\n",
1722                             rcd->ctxt, rcd->piocnt);
1723         }
1724
1725         rcd->piobufs = dd->pio2k_bufbase + rcd->pio_base * dd->palign;
1726         qib_chg_pioavailkernel(dd, rcd->pio_base, rcd->piocnt,
1727                                TXCHK_CHG_TYPE_USER, rcd);
1728         /*
1729          * try to ensure that processes start up with consistent avail update
1730          * for their own range, at least.   If system very quiet, it might
1731          * have the in-memory copy out of date at startup for this range of
1732          * buffers, when a context gets re-used.  Do after the chg_pioavail
1733          * and before the rest of setup, so it's "almost certain" the dma
1734          * will have occurred (can't 100% guarantee, but should be many
1735          * decimals of 9s, with this ordering), given how much else happens
1736          * after this.
1737          */
1738         dd->f_sendctrl(dd->pport, QIB_SENDCTRL_AVAIL_BLIP);
1739
1740         /*
1741          * Now allocate the rcvhdr Q and eager TIDs; skip the TID
1742          * array for time being.  If rcd->ctxt > chip-supported,
1743          * we need to do extra stuff here to handle by handling overflow
1744          * through ctxt 0, someday
1745          */
1746         ret = qib_create_rcvhdrq(dd, rcd);
1747         if (!ret)
1748                 ret = qib_setup_eagerbufs(rcd);
1749         if (ret)
1750                 goto bail_pio;
1751
1752         rcd->tidcursor = 0; /* start at beginning after open */
1753
1754         /* initialize poll variables... */
1755         rcd->urgent = 0;
1756         rcd->urgent_poll = 0;
1757
1758         /*
1759          * Now enable the ctxt for receive.
1760          * For chips that are set to DMA the tail register to memory
1761          * when they change (and when the update bit transitions from
1762          * 0 to 1.  So for those chips, we turn it off and then back on.
1763          * This will (very briefly) affect any other open ctxts, but the
1764          * duration is very short, and therefore isn't an issue.  We
1765          * explicitly set the in-memory tail copy to 0 beforehand, so we
1766          * don't have to wait to be sure the DMA update has happened
1767          * (chip resets head/tail to 0 on transition to enable).
1768          */
1769         if (rcd->rcvhdrtail_kvaddr)
1770                 qib_clear_rcvhdrtail(rcd);
1771
1772         dd->f_rcvctrl(rcd->ppd, QIB_RCVCTRL_CTXT_ENB | QIB_RCVCTRL_TIDFLOW_ENB,
1773                       rcd->ctxt);
1774
1775         /* Notify any waiting slaves */
1776         if (rcd->subctxt_cnt) {
1777                 clear_bit(QIB_CTXT_MASTER_UNINIT, &rcd->flag);
1778                 wake_up(&rcd->wait);
1779         }
1780         return 0;
1781
1782 bail_pio:
1783         qib_chg_pioavailkernel(dd, rcd->pio_base, rcd->piocnt,
1784                                TXCHK_CHG_TYPE_KERN, rcd);
1785 bail:
1786         return ret;
1787 }
1788
1789 /**
1790  * unlock_exptid - unlock any expected TID entries context still had in use
1791  * @rcd: ctxt
1792  *
1793  * We don't actually update the chip here, because we do a bulk update
1794  * below, using f_clear_tids.
1795  */
1796 static void unlock_expected_tids(struct qib_ctxtdata *rcd)
1797 {
1798         struct qib_devdata *dd = rcd->dd;
1799         int ctxt_tidbase = rcd->ctxt * dd->rcvtidcnt;
1800         int i, cnt = 0, maxtid = ctxt_tidbase + dd->rcvtidcnt;
1801
1802         for (i = ctxt_tidbase; i < maxtid; i++) {
1803                 struct page *p = dd->pageshadow[i];
1804                 dma_addr_t phys;
1805
1806                 if (!p)
1807                         continue;
1808
1809                 phys = dd->physshadow[i];
1810                 dd->physshadow[i] = dd->tidinvalid;
1811                 dd->pageshadow[i] = NULL;
1812                 pci_unmap_page(dd->pcidev, phys, PAGE_SIZE,
1813                                PCI_DMA_FROMDEVICE);
1814                 qib_release_user_pages(&p, 1);
1815                 cnt++;
1816         }
1817 }
1818
1819 static int qib_close(struct inode *in, struct file *fp)
1820 {
1821         int ret = 0;
1822         struct qib_filedata *fd;
1823         struct qib_ctxtdata *rcd;
1824         struct qib_devdata *dd;
1825         unsigned long flags;
1826         unsigned ctxt;
1827         pid_t pid;
1828
1829         mutex_lock(&qib_mutex);
1830
1831         fd = fp->private_data;
1832         fp->private_data = NULL;
1833         rcd = fd->rcd;
1834         if (!rcd) {
1835                 mutex_unlock(&qib_mutex);
1836                 goto bail;
1837         }
1838
1839         dd = rcd->dd;
1840
1841         /* ensure all pio buffer writes in progress are flushed */
1842         qib_flush_wc();
1843
1844         /* drain user sdma queue */
1845         if (fd->pq) {
1846                 qib_user_sdma_queue_drain(rcd->ppd, fd->pq);
1847                 qib_user_sdma_queue_destroy(fd->pq);
1848         }
1849
1850         if (fd->rec_cpu_num != -1)
1851                 __clear_bit(fd->rec_cpu_num, qib_cpulist);
1852
1853         if (--rcd->cnt) {
1854                 /*
1855                  * XXX If the master closes the context before the slave(s),
1856                  * revoke the mmap for the eager receive queue so
1857                  * the slave(s) don't wait for receive data forever.
1858                  */
1859                 rcd->active_slaves &= ~(1 << fd->subctxt);
1860                 rcd->subpid[fd->subctxt] = 0;
1861                 mutex_unlock(&qib_mutex);
1862                 goto bail;
1863         }
1864
1865         /* early; no interrupt users after this */
1866         spin_lock_irqsave(&dd->uctxt_lock, flags);
1867         ctxt = rcd->ctxt;
1868         dd->rcd[ctxt] = NULL;
1869         pid = rcd->pid;
1870         rcd->pid = 0;
1871         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
1872
1873         if (rcd->rcvwait_to || rcd->piowait_to ||
1874             rcd->rcvnowait || rcd->pionowait) {
1875                 rcd->rcvwait_to = 0;
1876                 rcd->piowait_to = 0;
1877                 rcd->rcvnowait = 0;
1878                 rcd->pionowait = 0;
1879         }
1880         if (rcd->flag)
1881                 rcd->flag = 0;
1882
1883         if (dd->kregbase) {
1884                 /* atomically clear receive enable ctxt and intr avail. */
1885                 dd->f_rcvctrl(rcd->ppd, QIB_RCVCTRL_CTXT_DIS |
1886                                   QIB_RCVCTRL_INTRAVAIL_DIS, ctxt);
1887
1888                 /* clean up the pkeys for this ctxt user */
1889                 qib_clean_part_key(rcd, dd);
1890                 qib_disarm_piobufs(dd, rcd->pio_base, rcd->piocnt);
1891                 qib_chg_pioavailkernel(dd, rcd->pio_base,
1892                                        rcd->piocnt, TXCHK_CHG_TYPE_KERN, NULL);
1893
1894                 dd->f_clear_tids(dd, rcd);
1895
1896                 if (dd->pageshadow)
1897                         unlock_expected_tids(rcd);
1898                 qib_stats.sps_ctxts--;
1899                 dd->freectxts++;
1900         }
1901
1902         mutex_unlock(&qib_mutex);
1903         qib_free_ctxtdata(dd, rcd); /* after releasing the mutex */
1904
1905 bail:
1906         kfree(fd);
1907         return ret;
1908 }
1909
1910 static int qib_ctxt_info(struct file *fp, struct qib_ctxt_info __user *uinfo)
1911 {
1912         struct qib_ctxt_info info;
1913         int ret;
1914         size_t sz;
1915         struct qib_ctxtdata *rcd = ctxt_fp(fp);
1916         struct qib_filedata *fd;
1917
1918         fd = fp->private_data;
1919
1920         info.num_active = qib_count_active_units();
1921         info.unit = rcd->dd->unit;
1922         info.port = rcd->ppd->port;
1923         info.ctxt = rcd->ctxt;
1924         info.subctxt =  subctxt_fp(fp);
1925         /* Number of user ctxts available for this device. */
1926         info.num_ctxts = rcd->dd->cfgctxts - rcd->dd->first_user_ctxt;
1927         info.num_subctxts = rcd->subctxt_cnt;
1928         info.rec_cpu = fd->rec_cpu_num;
1929         sz = sizeof(info);
1930
1931         if (copy_to_user(uinfo, &info, sz)) {
1932                 ret = -EFAULT;
1933                 goto bail;
1934         }
1935         ret = 0;
1936
1937 bail:
1938         return ret;
1939 }
1940
1941 static int qib_sdma_get_inflight(struct qib_user_sdma_queue *pq,
1942                                  u32 __user *inflightp)
1943 {
1944         const u32 val = qib_user_sdma_inflight_counter(pq);
1945
1946         if (put_user(val, inflightp))
1947                 return -EFAULT;
1948
1949         return 0;
1950 }
1951
1952 static int qib_sdma_get_complete(struct qib_pportdata *ppd,
1953                                  struct qib_user_sdma_queue *pq,
1954                                  u32 __user *completep)
1955 {
1956         u32 val;
1957         int err;
1958
1959         if (!pq)
1960                 return -EINVAL;
1961
1962         err = qib_user_sdma_make_progress(ppd, pq);
1963         if (err < 0)
1964                 return err;
1965
1966         val = qib_user_sdma_complete_counter(pq);
1967         if (put_user(val, completep))
1968                 return -EFAULT;
1969
1970         return 0;
1971 }
1972
1973 static int disarm_req_delay(struct qib_ctxtdata *rcd)
1974 {
1975         int ret = 0;
1976
1977         if (!usable(rcd->ppd)) {
1978                 int i;
1979                 /*
1980                  * if link is down, or otherwise not usable, delay
1981                  * the caller up to 30 seconds, so we don't thrash
1982                  * in trying to get the chip back to ACTIVE, and
1983                  * set flag so they make the call again.
1984                  */
1985                 if (rcd->user_event_mask) {
1986                         /*
1987                          * subctxt_cnt is 0 if not shared, so do base
1988                          * separately, first, then remaining subctxt, if any
1989                          */
1990                         set_bit(_QIB_EVENT_DISARM_BUFS_BIT,
1991                                 &rcd->user_event_mask[0]);
1992                         for (i = 1; i < rcd->subctxt_cnt; i++)
1993                                 set_bit(_QIB_EVENT_DISARM_BUFS_BIT,
1994                                         &rcd->user_event_mask[i]);
1995                 }
1996                 for (i = 0; !usable(rcd->ppd) && i < 300; i++)
1997                         msleep(100);
1998                 ret = -ENETDOWN;
1999         }
2000         return ret;
2001 }
2002
2003 /*
2004  * Find all user contexts in use, and set the specified bit in their
2005  * event mask.
2006  * See also find_ctxt() for a similar use, that is specific to send buffers.
2007  */
2008 int qib_set_uevent_bits(struct qib_pportdata *ppd, const int evtbit)
2009 {
2010         struct qib_ctxtdata *rcd;
2011         unsigned ctxt;
2012         int ret = 0;
2013         unsigned long flags;
2014
2015         spin_lock_irqsave(&ppd->dd->uctxt_lock, flags);
2016         for (ctxt = ppd->dd->first_user_ctxt; ctxt < ppd->dd->cfgctxts;
2017              ctxt++) {
2018                 rcd = ppd->dd->rcd[ctxt];
2019                 if (!rcd)
2020                         continue;
2021                 if (rcd->user_event_mask) {
2022                         int i;
2023                         /*
2024                          * subctxt_cnt is 0 if not shared, so do base
2025                          * separately, first, then remaining subctxt, if any
2026                          */
2027                         set_bit(evtbit, &rcd->user_event_mask[0]);
2028                         for (i = 1; i < rcd->subctxt_cnt; i++)
2029                                 set_bit(evtbit, &rcd->user_event_mask[i]);
2030                 }
2031                 ret = 1;
2032                 break;
2033         }
2034         spin_unlock_irqrestore(&ppd->dd->uctxt_lock, flags);
2035
2036         return ret;
2037 }
2038
2039 /*
2040  * clear the event notifier events for this context.
2041  * For the DISARM_BUFS case, we also take action (this obsoletes
2042  * the older QIB_CMD_DISARM_BUFS, but we keep it for backwards
2043  * compatibility.
2044  * Other bits don't currently require actions, just atomically clear.
2045  * User process then performs actions appropriate to bit having been
2046  * set, if desired, and checks again in future.
2047  */
2048 static int qib_user_event_ack(struct qib_ctxtdata *rcd, int subctxt,
2049                               unsigned long events)
2050 {
2051         int ret = 0, i;
2052
2053         for (i = 0; i <= _QIB_MAX_EVENT_BIT; i++) {
2054                 if (!test_bit(i, &events))
2055                         continue;
2056                 if (i == _QIB_EVENT_DISARM_BUFS_BIT) {
2057                         (void)qib_disarm_piobufs_ifneeded(rcd);
2058                         ret = disarm_req_delay(rcd);
2059                 } else
2060                         clear_bit(i, &rcd->user_event_mask[subctxt]);
2061         }
2062         return ret;
2063 }
2064
2065 static ssize_t qib_write(struct file *fp, const char __user *data,
2066                          size_t count, loff_t *off)
2067 {
2068         const struct qib_cmd __user *ucmd;
2069         struct qib_ctxtdata *rcd;
2070         const void __user *src;
2071         size_t consumed, copy = 0;
2072         struct qib_cmd cmd;
2073         ssize_t ret = 0;
2074         void *dest;
2075
2076         if (WARN_ON_ONCE(!ib_safe_file_access(fp)))
2077                 return -EACCES;
2078
2079         if (count < sizeof(cmd.type)) {
2080                 ret = -EINVAL;
2081                 goto bail;
2082         }
2083
2084         ucmd = (const struct qib_cmd __user *) data;
2085
2086         if (copy_from_user(&cmd.type, &ucmd->type, sizeof(cmd.type))) {
2087                 ret = -EFAULT;
2088                 goto bail;
2089         }
2090
2091         consumed = sizeof(cmd.type);
2092
2093         switch (cmd.type) {
2094         case QIB_CMD_ASSIGN_CTXT:
2095         case QIB_CMD_USER_INIT:
2096                 copy = sizeof(cmd.cmd.user_info);
2097                 dest = &cmd.cmd.user_info;
2098                 src = &ucmd->cmd.user_info;
2099                 break;
2100
2101         case QIB_CMD_RECV_CTRL:
2102                 copy = sizeof(cmd.cmd.recv_ctrl);
2103                 dest = &cmd.cmd.recv_ctrl;
2104                 src = &ucmd->cmd.recv_ctrl;
2105                 break;
2106
2107         case QIB_CMD_CTXT_INFO:
2108                 copy = sizeof(cmd.cmd.ctxt_info);
2109                 dest = &cmd.cmd.ctxt_info;
2110                 src = &ucmd->cmd.ctxt_info;
2111                 break;
2112
2113         case QIB_CMD_TID_UPDATE:
2114         case QIB_CMD_TID_FREE:
2115                 copy = sizeof(cmd.cmd.tid_info);
2116                 dest = &cmd.cmd.tid_info;
2117                 src = &ucmd->cmd.tid_info;
2118                 break;
2119
2120         case QIB_CMD_SET_PART_KEY:
2121                 copy = sizeof(cmd.cmd.part_key);
2122                 dest = &cmd.cmd.part_key;
2123                 src = &ucmd->cmd.part_key;
2124                 break;
2125
2126         case QIB_CMD_DISARM_BUFS:
2127         case QIB_CMD_PIOAVAILUPD: /* force an update of PIOAvail reg */
2128                 copy = 0;
2129                 src = NULL;
2130                 dest = NULL;
2131                 break;
2132
2133         case QIB_CMD_POLL_TYPE:
2134                 copy = sizeof(cmd.cmd.poll_type);
2135                 dest = &cmd.cmd.poll_type;
2136                 src = &ucmd->cmd.poll_type;
2137                 break;
2138
2139         case QIB_CMD_ARMLAUNCH_CTRL:
2140                 copy = sizeof(cmd.cmd.armlaunch_ctrl);
2141                 dest = &cmd.cmd.armlaunch_ctrl;
2142                 src = &ucmd->cmd.armlaunch_ctrl;
2143                 break;
2144
2145         case QIB_CMD_SDMA_INFLIGHT:
2146                 copy = sizeof(cmd.cmd.sdma_inflight);
2147                 dest = &cmd.cmd.sdma_inflight;
2148                 src = &ucmd->cmd.sdma_inflight;
2149                 break;
2150
2151         case QIB_CMD_SDMA_COMPLETE:
2152                 copy = sizeof(cmd.cmd.sdma_complete);
2153                 dest = &cmd.cmd.sdma_complete;
2154                 src = &ucmd->cmd.sdma_complete;
2155                 break;
2156
2157         case QIB_CMD_ACK_EVENT:
2158                 copy = sizeof(cmd.cmd.event_mask);
2159                 dest = &cmd.cmd.event_mask;
2160                 src = &ucmd->cmd.event_mask;
2161                 break;
2162
2163         default:
2164                 ret = -EINVAL;
2165                 goto bail;
2166         }
2167
2168         if (copy) {
2169                 if ((count - consumed) < copy) {
2170                         ret = -EINVAL;
2171                         goto bail;
2172                 }
2173                 if (copy_from_user(dest, src, copy)) {
2174                         ret = -EFAULT;
2175                         goto bail;
2176                 }
2177                 consumed += copy;
2178         }
2179
2180         rcd = ctxt_fp(fp);
2181         if (!rcd && cmd.type != QIB_CMD_ASSIGN_CTXT) {
2182                 ret = -EINVAL;
2183                 goto bail;
2184         }
2185
2186         switch (cmd.type) {
2187         case QIB_CMD_ASSIGN_CTXT:
2188                 ret = qib_assign_ctxt(fp, &cmd.cmd.user_info);
2189                 if (ret)
2190                         goto bail;
2191                 break;
2192
2193         case QIB_CMD_USER_INIT:
2194                 ret = qib_do_user_init(fp, &cmd.cmd.user_info);
2195                 if (ret)
2196                         goto bail;
2197                 ret = qib_get_base_info(fp, (void __user *) (unsigned long)
2198                                         cmd.cmd.user_info.spu_base_info,
2199                                         cmd.cmd.user_info.spu_base_info_size);
2200                 break;
2201
2202         case QIB_CMD_RECV_CTRL:
2203                 ret = qib_manage_rcvq(rcd, subctxt_fp(fp), cmd.cmd.recv_ctrl);
2204                 break;
2205
2206         case QIB_CMD_CTXT_INFO:
2207                 ret = qib_ctxt_info(fp, (struct qib_ctxt_info __user *)
2208                                     (unsigned long) cmd.cmd.ctxt_info);
2209                 break;
2210
2211         case QIB_CMD_TID_UPDATE:
2212                 ret = qib_tid_update(rcd, fp, &cmd.cmd.tid_info);
2213                 break;
2214
2215         case QIB_CMD_TID_FREE:
2216                 ret = qib_tid_free(rcd, subctxt_fp(fp), &cmd.cmd.tid_info);
2217                 break;
2218
2219         case QIB_CMD_SET_PART_KEY:
2220                 ret = qib_set_part_key(rcd, cmd.cmd.part_key);
2221                 break;
2222
2223         case QIB_CMD_DISARM_BUFS:
2224                 (void)qib_disarm_piobufs_ifneeded(rcd);
2225                 ret = disarm_req_delay(rcd);
2226                 break;
2227
2228         case QIB_CMD_PIOAVAILUPD:
2229                 qib_force_pio_avail_update(rcd->dd);
2230                 break;
2231
2232         case QIB_CMD_POLL_TYPE:
2233                 rcd->poll_type = cmd.cmd.poll_type;
2234                 break;
2235
2236         case QIB_CMD_ARMLAUNCH_CTRL:
2237                 rcd->dd->f_set_armlaunch(rcd->dd, cmd.cmd.armlaunch_ctrl);
2238                 break;
2239
2240         case QIB_CMD_SDMA_INFLIGHT:
2241                 ret = qib_sdma_get_inflight(user_sdma_queue_fp(fp),
2242                                             (u32 __user *) (unsigned long)
2243                                             cmd.cmd.sdma_inflight);
2244                 break;
2245
2246         case QIB_CMD_SDMA_COMPLETE:
2247                 ret = qib_sdma_get_complete(rcd->ppd,
2248                                             user_sdma_queue_fp(fp),
2249                                             (u32 __user *) (unsigned long)
2250                                             cmd.cmd.sdma_complete);
2251                 break;
2252
2253         case QIB_CMD_ACK_EVENT:
2254                 ret = qib_user_event_ack(rcd, subctxt_fp(fp),
2255                                          cmd.cmd.event_mask);
2256                 break;
2257         }
2258
2259         if (ret >= 0)
2260                 ret = consumed;
2261
2262 bail:
2263         return ret;
2264 }
2265
2266 static ssize_t qib_write_iter(struct kiocb *iocb, struct iov_iter *from)
2267 {
2268         struct qib_filedata *fp = iocb->ki_filp->private_data;
2269         struct qib_ctxtdata *rcd = ctxt_fp(iocb->ki_filp);
2270         struct qib_user_sdma_queue *pq = fp->pq;
2271
2272         if (!iter_is_iovec(from) || !from->nr_segs || !pq)
2273                 return -EINVAL;
2274
2275         return qib_user_sdma_writev(rcd, pq, from->iov, from->nr_segs);
2276 }
2277
2278 static struct class *qib_class;
2279 static dev_t qib_dev;
2280
2281 int qib_cdev_init(int minor, const char *name,
2282                   const struct file_operations *fops,
2283                   struct cdev **cdevp, struct device **devp)
2284 {
2285         const dev_t dev = MKDEV(MAJOR(qib_dev), minor);
2286         struct cdev *cdev;
2287         struct device *device = NULL;
2288         int ret;
2289
2290         cdev = cdev_alloc();
2291         if (!cdev) {
2292                 pr_err("Could not allocate cdev for minor %d, %s\n",
2293                        minor, name);
2294                 ret = -ENOMEM;
2295                 goto done;
2296         }
2297
2298         cdev->owner = THIS_MODULE;
2299         cdev->ops = fops;
2300         kobject_set_name(&cdev->kobj, name);
2301
2302         ret = cdev_add(cdev, dev, 1);
2303         if (ret < 0) {
2304                 pr_err("Could not add cdev for minor %d, %s (err %d)\n",
2305                        minor, name, -ret);
2306                 goto err_cdev;
2307         }
2308
2309         device = device_create(qib_class, NULL, dev, NULL, "%s", name);
2310         if (!IS_ERR(device))
2311                 goto done;
2312         ret = PTR_ERR(device);
2313         device = NULL;
2314         pr_err("Could not create device for minor %d, %s (err %d)\n",
2315                minor, name, -ret);
2316 err_cdev:
2317         cdev_del(cdev);
2318         cdev = NULL;
2319 done:
2320         *cdevp = cdev;
2321         *devp = device;
2322         return ret;
2323 }
2324
2325 void qib_cdev_cleanup(struct cdev **cdevp, struct device **devp)
2326 {
2327         struct device *device = *devp;
2328
2329         if (device) {
2330                 device_unregister(device);
2331                 *devp = NULL;
2332         }
2333
2334         if (*cdevp) {
2335                 cdev_del(*cdevp);
2336                 *cdevp = NULL;
2337         }
2338 }
2339
2340 static struct cdev *wildcard_cdev;
2341 static struct device *wildcard_device;
2342
2343 int __init qib_dev_init(void)
2344 {
2345         int ret;
2346
2347         ret = alloc_chrdev_region(&qib_dev, 0, QIB_NMINORS, QIB_DRV_NAME);
2348         if (ret < 0) {
2349                 pr_err("Could not allocate chrdev region (err %d)\n", -ret);
2350                 goto done;
2351         }
2352
2353         qib_class = class_create(THIS_MODULE, "ipath");
2354         if (IS_ERR(qib_class)) {
2355                 ret = PTR_ERR(qib_class);
2356                 pr_err("Could not create device class (err %d)\n", -ret);
2357                 unregister_chrdev_region(qib_dev, QIB_NMINORS);
2358         }
2359
2360 done:
2361         return ret;
2362 }
2363
2364 void qib_dev_cleanup(void)
2365 {
2366         if (qib_class) {
2367                 class_destroy(qib_class);
2368                 qib_class = NULL;
2369         }
2370
2371         unregister_chrdev_region(qib_dev, QIB_NMINORS);
2372 }
2373
2374 static atomic_t user_count = ATOMIC_INIT(0);
2375
2376 static void qib_user_remove(struct qib_devdata *dd)
2377 {
2378         if (atomic_dec_return(&user_count) == 0)
2379                 qib_cdev_cleanup(&wildcard_cdev, &wildcard_device);
2380
2381         qib_cdev_cleanup(&dd->user_cdev, &dd->user_device);
2382 }
2383
2384 static int qib_user_add(struct qib_devdata *dd)
2385 {
2386         char name[10];
2387         int ret;
2388
2389         if (atomic_inc_return(&user_count) == 1) {
2390                 ret = qib_cdev_init(0, "ipath", &qib_file_ops,
2391                                     &wildcard_cdev, &wildcard_device);
2392                 if (ret)
2393                         goto done;
2394         }
2395
2396         snprintf(name, sizeof(name), "ipath%d", dd->unit);
2397         ret = qib_cdev_init(dd->unit + 1, name, &qib_file_ops,
2398                             &dd->user_cdev, &dd->user_device);
2399         if (ret)
2400                 qib_user_remove(dd);
2401 done:
2402         return ret;
2403 }
2404
2405 /*
2406  * Create per-unit files in /dev
2407  */
2408 int qib_device_create(struct qib_devdata *dd)
2409 {
2410         int r, ret;
2411
2412         r = qib_user_add(dd);
2413         ret = qib_diag_add(dd);
2414         if (r && !ret)
2415                 ret = r;
2416         return ret;
2417 }
2418
2419 /*
2420  * Remove per-unit files in /dev
2421  * void, core kernel returns no errors for this stuff
2422  */
2423 void qib_device_remove(struct qib_devdata *dd)
2424 {
2425         qib_user_remove(dd);
2426         qib_diag_remove(dd);
2427 }