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