sys/cam/scsi/scsi_pass.c

   1 /*-
   2  * Copyright (c) 1997, 1998, 2000 Justin T. Gibbs.
   3  * Copyright (c) 1997, 1998, 1999 Kenneth D. Merry.
   4  * All rights reserved.
   5  *
   6  * Redistribution and use in source and binary forms, with or without
   7  * modification, are permitted provided that the following conditions
   8  * are met:
   9  * 1. Redistributions of source code must retain the above copyright
  10  *    notice, this list of conditions, and the following disclaimer,
  11  *    without modification, immediately at the beginning of the file.
  12  * 2. The name of the author may not be used to endorse or promote products
  13  *    derived from this software without specific prior written permission.
  14  *
  15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  18  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR
  19  * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  25  * SUCH DAMAGE.
  26  */
  27
  28 #include <sys/cdefs.h>
  29 __FBSDID("$FreeBSD$");
  30
  31 #include <sys/param.h>
  32 #include <sys/systm.h>
  33 #include <sys/kernel.h>
  34 #include <sys/conf.h>
  35 #include <sys/types.h>
  36 #include <sys/bio.h>
  37 #include <sys/bus.h>
  38 #include <sys/devicestat.h>
  39 #include <sys/errno.h>
  40 #include <sys/fcntl.h>
  41 #include <sys/malloc.h>
  42 #include <sys/proc.h>
  43 #include <sys/poll.h>
  44 #include <sys/selinfo.h>
  45 #include <sys/sdt.h>
  46 #include <sys/taskqueue.h>
  47 #include <vm/uma.h>
  48 #include <vm/vm.h>
  49 #include <vm/vm_extern.h>
  50
  51 #include <machine/bus.h>
  52
  53 #include <cam/cam.h>
  54 #include <cam/cam_ccb.h>
  55 #include <cam/cam_periph.h>
  56 #include <cam/cam_queue.h>
  57 #include <cam/cam_xpt.h>
  58 #include <cam/cam_xpt_periph.h>
  59 #include <cam/cam_debug.h>
  60 #include <cam/cam_compat.h>
  61 #include <cam/cam_xpt_periph.h>
  62
  63 #include <cam/scsi/scsi_all.h>
  64 #include <cam/scsi/scsi_pass.h>
  65
  66 typedef enum {
  67         PASS_FLAG_OPEN                  = 0x01,
  68         PASS_FLAG_LOCKED                = 0x02,
  69         PASS_FLAG_INVALID               = 0x04,
  70         PASS_FLAG_INITIAL_PHYSPATH      = 0x08,
  71         PASS_FLAG_ZONE_INPROG           = 0x10,
  72         PASS_FLAG_ZONE_VALID            = 0x20,
  73         PASS_FLAG_UNMAPPED_CAPABLE      = 0x40,
  74         PASS_FLAG_ABANDONED_REF_SET     = 0x80
  75 } pass_flags;
  76
  77 typedef enum {
  78         PASS_STATE_NORMAL
  79 } pass_state;
  80
  81 typedef enum {
  82         PASS_CCB_BUFFER_IO,
  83         PASS_CCB_QUEUED_IO
  84 } pass_ccb_types;
  85
  86 #define ccb_type        ppriv_field0
  87 #define ccb_ioreq       ppriv_ptr1
  88
  89 /*
  90  * The maximum number of memory segments we preallocate.
  91  */
  92 #define PASS_MAX_SEGS   16
  93
  94 typedef enum {
  95         PASS_IO_NONE            = 0x00,
  96         PASS_IO_USER_SEG_MALLOC = 0x01,
  97         PASS_IO_KERN_SEG_MALLOC = 0x02,
  98         PASS_IO_ABANDONED       = 0x04
  99 } pass_io_flags;
 100
 101 struct pass_io_req {
 102         union ccb                        ccb;
 103         union ccb                       *alloced_ccb;
 104         union ccb                       *user_ccb_ptr;
 105         camq_entry                       user_periph_links;
 106         ccb_ppriv_area                   user_periph_priv;
 107         struct cam_periph_map_info       mapinfo;
 108         pass_io_flags                    flags;
 109         ccb_flags                        data_flags;
 110         int                              num_user_segs;
 111         bus_dma_segment_t                user_segs[PASS_MAX_SEGS];
 112         int                              num_kern_segs;
 113         bus_dma_segment_t                kern_segs[PASS_MAX_SEGS];
 114         bus_dma_segment_t               *user_segptr;
 115         bus_dma_segment_t               *kern_segptr;
 116         int                              num_bufs;
 117         uint32_t                         dirs[CAM_PERIPH_MAXMAPS];
 118         uint32_t                         lengths[CAM_PERIPH_MAXMAPS];
 119         uint8_t                         *user_bufs[CAM_PERIPH_MAXMAPS];
 120         uint8_t                         *kern_bufs[CAM_PERIPH_MAXMAPS];
 121         struct bintime                   start_time;
 122         TAILQ_ENTRY(pass_io_req)         links;
 123 };
 124
 125 struct pass_softc {
 126         pass_state                state;
 127         pass_flags                flags;
 128         u_int8_t                  pd_type;
 129         union ccb                 saved_ccb;
 130         int                       open_count;
 131         u_int                     maxio;
 132         struct devstat           *device_stats;
 133         struct cdev              *dev;
 134         struct cdev              *alias_dev;
 135         struct task               add_physpath_task;
 136         struct task               shutdown_kqueue_task;
 137         struct selinfo            read_select;
 138         TAILQ_HEAD(, pass_io_req) incoming_queue;
 139         TAILQ_HEAD(, pass_io_req) active_queue;
 140         TAILQ_HEAD(, pass_io_req) abandoned_queue;
 141         TAILQ_HEAD(, pass_io_req) done_queue;
 142         struct cam_periph        *periph;
 143         char                      zone_name[12];
 144         char                      io_zone_name[12];
 145         uma_zone_t                pass_zone;
 146         uma_zone_t                pass_io_zone;
 147         size_t                    io_zone_size;
 148 };
 149
 150 static  d_open_t        passopen;
 151 static  d_close_t       passclose;
 152 static  d_ioctl_t       passioctl;
 153 static  d_ioctl_t       passdoioctl;
 154 static  d_poll_t        passpoll;
 155 static  d_kqfilter_t    passkqfilter;
 156 static  void            passreadfiltdetach(struct knote *kn);
 157 static  int             passreadfilt(struct knote *kn, long hint);
 158
 159 static  periph_init_t   passinit;
 160 static  periph_ctor_t   passregister;
 161 static  periph_oninv_t  passoninvalidate;
 162 static  periph_dtor_t   passcleanup;
 163 static  periph_start_t  passstart;
 164 static  void            pass_shutdown_kqueue(void *context, int pending);
 165 static  void            pass_add_physpath(void *context, int pending);
 166 static  void            passasync(void *callback_arg, u_int32_t code,
 167                                   struct cam_path *path, void *arg);
 168 static  void            passdone(struct cam_periph *periph,
 169                                  union ccb *done_ccb);
 170 static  int             passcreatezone(struct cam_periph *periph);
 171 static  void            passiocleanup(struct pass_softc *softc,
 172                                       struct pass_io_req *io_req);
 173 static  int             passcopysglist(struct cam_periph *periph,
 174                                        struct pass_io_req *io_req,
 175                                        ccb_flags direction);
 176 static  int             passmemsetup(struct cam_periph *periph,
 177                                      struct pass_io_req *io_req);
 178 static  int             passmemdone(struct cam_periph *periph,
 179                                     struct pass_io_req *io_req);
 180 static  int             passerror(union ccb *ccb, u_int32_t cam_flags,
 181                                   u_int32_t sense_flags);
 182 static  int             passsendccb(struct cam_periph *periph, union ccb *ccb,
 183                                     union ccb *inccb);
 184
 185 static struct periph_driver passdriver =
 186 {
 187         passinit, "pass",
 188         TAILQ_HEAD_INITIALIZER(passdriver.units), /* generation */ 0
 189 };
 190
 191 PERIPHDRIVER_DECLARE(pass, passdriver);
 192
 193 static struct cdevsw pass_cdevsw = {
 194         .d_version =    D_VERSION,
 195         .d_flags =      D_TRACKCLOSE,
 196         .d_open =       passopen,
 197         .d_close =      passclose,
 198         .d_ioctl =      passioctl,
 199         .d_poll =       passpoll,
 200         .d_kqfilter =   passkqfilter,
 201         .d_name =       "pass",
 202 };
 203
 204 static struct filterops passread_filtops = {
 205         .f_isfd =       1,
 206         .f_detach =     passreadfiltdetach,
 207         .f_event =      passreadfilt
 208 };
 209
 210 static MALLOC_DEFINE(M_SCSIPASS, "scsi_pass", "scsi passthrough buffers");
 211
 212 static void
 213 passinit(void)
 214 {
 215         cam_status status;
 216
 217         /*
 218          * Install a global async callback.  This callback will
 219          * receive async callbacks like "new device found".
 220          */
 221         status = xpt_register_async(AC_FOUND_DEVICE, passasync, NULL, NULL);
 222
 223         if (status != CAM_REQ_CMP) {
 224                 printf("pass: Failed to attach master async callback "
 225                        "due to status 0x%x!\n", status);
 226         }
 227
 228 }
 229
 230 static void
 231 passrejectios(struct cam_periph *periph)
 232 {
 233         struct pass_io_req *io_req, *io_req2;
 234         struct pass_softc *softc;
 235
 236         softc = (struct pass_softc *)periph->softc;
 237
 238         /*
 239          * The user can no longer get status for I/O on the done queue, so
 240          * clean up all outstanding I/O on the done queue.
 241          */
 242         TAILQ_FOREACH_SAFE(io_req, &softc->done_queue, links, io_req2) {
 243                 TAILQ_REMOVE(&softc->done_queue, io_req, links);
 244                 passiocleanup(softc, io_req);
 245                 uma_zfree(softc->pass_zone, io_req);
 246         }
 247
 248         /*
 249          * The underlying device is gone, so we can't issue these I/Os.
 250          * The devfs node has been shut down, so we can't return status to
 251          * the user.  Free any I/O left on the incoming queue.
 252          */
 253         TAILQ_FOREACH_SAFE(io_req, &softc->incoming_queue, links, io_req2) {
 254                 TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
 255                 passiocleanup(softc, io_req);
 256                 uma_zfree(softc->pass_zone, io_req);
 257         }
 258
 259         /*
 260          * Normally we would put I/Os on the abandoned queue and acquire a
 261          * reference when we saw the final close.  But, the device went
 262          * away and devfs may have moved everything off to deadfs by the
 263          * time the I/O done callback is called; as a result, we won't see
 264          * any more closes.  So, if we have any active I/Os, we need to put
 265          * them on the abandoned queue.  When the abandoned queue is empty,
 266          * we'll release the remaining reference (see below) to the peripheral.
 267          */
 268         TAILQ_FOREACH_SAFE(io_req, &softc->active_queue, links, io_req2) {
 269                 TAILQ_REMOVE(&softc->active_queue, io_req, links);
 270                 io_req->flags |= PASS_IO_ABANDONED;
 271                 TAILQ_INSERT_TAIL(&softc->abandoned_queue, io_req, links);
 272         }
 273
 274         /*
 275          * If we put any I/O on the abandoned queue, acquire a reference.
 276          */
 277         if ((!TAILQ_EMPTY(&softc->abandoned_queue))
 278          && ((softc->flags & PASS_FLAG_ABANDONED_REF_SET) == 0)) {
 279                 cam_periph_doacquire(periph);
 280                 softc->flags |= PASS_FLAG_ABANDONED_REF_SET;
 281         }
 282 }
 283
 284 static void
 285 passdevgonecb(void *arg)
 286 {
 287         struct cam_periph *periph;
 288         struct mtx *mtx;
 289         struct pass_softc *softc;
 290         int i;
 291
 292         periph = (struct cam_periph *)arg;
 293         mtx = cam_periph_mtx(periph);
 294         mtx_lock(mtx);
 295
 296         softc = (struct pass_softc *)periph->softc;
 297         KASSERT(softc->open_count >= 0, ("Negative open count %d",
 298                 softc->open_count));
 299
 300         /*
 301          * When we get this callback, we will get no more close calls from
 302          * devfs.  So if we have any dangling opens, we need to release the
 303          * reference held for that particular context.
 304          */
 305         for (i = 0; i < softc->open_count; i++)
 306                 cam_periph_release_locked(periph);
 307
 308         softc->open_count = 0;
 309
 310         /*
 311          * Release the reference held for the device node, it is gone now.
 312          * Accordingly, inform all queued I/Os of their fate.
 313          */
 314         cam_periph_release_locked(periph);
 315         passrejectios(periph);
 316
 317         /*
 318          * We reference the SIM lock directly here, instead of using
 319          * cam_periph_unlock().  The reason is that the final call to
 320          * cam_periph_release_locked() above could result in the periph
 321          * getting freed.  If that is the case, dereferencing the periph
 322          * with a cam_periph_unlock() call would cause a page fault.
 323          */
 324         mtx_unlock(mtx);
 325
 326         /*
 327          * We have to remove our kqueue context from a thread because it
 328          * may sleep.  It would be nice if we could get a callback from
 329          * kqueue when it is done cleaning up resources.
 330          */
 331         taskqueue_enqueue(taskqueue_thread, &softc->shutdown_kqueue_task);
 332 }
 333
 334 static void
 335 passoninvalidate(struct cam_periph *periph)
 336 {
 337         struct pass_softc *softc;
 338
 339         softc = (struct pass_softc *)periph->softc;
 340
 341         /*
 342          * De-register any async callbacks.
 343          */
 344         xpt_register_async(0, passasync, periph, periph->path);
 345
 346         softc->flags |= PASS_FLAG_INVALID;
 347
 348         /*
 349          * Tell devfs this device has gone away, and ask for a callback
 350          * when it has cleaned up its state.
 351          */
 352         destroy_dev_sched_cb(softc->dev, passdevgonecb, periph);
 353 }
 354
 355 static void
 356 passcleanup(struct cam_periph *periph)
 357 {
 358         struct pass_softc *softc;
 359
 360         softc = (struct pass_softc *)periph->softc;
 361
 362         cam_periph_assert(periph, MA_OWNED);
 363         KASSERT(TAILQ_EMPTY(&softc->active_queue),
 364                 ("%s called when there are commands on the active queue!\n",
 365                 __func__));
 366         KASSERT(TAILQ_EMPTY(&softc->abandoned_queue),
 367                 ("%s called when there are commands on the abandoned queue!\n",
 368                 __func__));
 369         KASSERT(TAILQ_EMPTY(&softc->incoming_queue),
 370                 ("%s called when there are commands on the incoming queue!\n",
 371                 __func__));
 372         KASSERT(TAILQ_EMPTY(&softc->done_queue),
 373                 ("%s called when there are commands on the done queue!\n",
 374                 __func__));
 375
 376         devstat_remove_entry(softc->device_stats);
 377
 378         cam_periph_unlock(periph);
 379
 380         /*
 381          * We call taskqueue_drain() for the physpath task to make sure it
 382          * is complete.  We drop the lock because this can potentially
 383          * sleep.  XXX KDM that is bad.  Need a way to get a callback when
 384          * a taskqueue is drained.
 385          *
 386          * Note that we don't drain the kqueue shutdown task queue.  This
 387          * is because we hold a reference on the periph for kqueue, and
 388          * release that reference from the kqueue shutdown task queue.  So
 389          * we cannot come into this routine unless we've released that
 390          * reference.  Also, because that could be the last reference, we
 391          * could be called from the cam_periph_release() call in
 392          * pass_shutdown_kqueue().  In that case, the taskqueue_drain()
 393          * would deadlock.  It would be preferable if we had a way to
 394          * get a callback when a taskqueue is done.
 395          */
 396         taskqueue_drain(taskqueue_thread, &softc->add_physpath_task);
 397
 398         cam_periph_lock(periph);
 399
 400         free(softc, M_DEVBUF);
 401 }
 402
 403 static void
 404 pass_shutdown_kqueue(void *context, int pending)
 405 {
 406         struct cam_periph *periph;
 407         struct pass_softc *softc;
 408
 409         periph = context;
 410         softc = periph->softc;
 411
 412         knlist_clear(&softc->read_select.si_note, /*is_locked*/ 0);
 413         knlist_destroy(&softc->read_select.si_note);
 414
 415         /*
 416          * Release the reference we held for kqueue.
 417          */
 418         cam_periph_release(periph);
 419 }
 420
 421 static void
 422 pass_add_physpath(void *context, int pending)
 423 {
 424         struct cam_periph *periph;
 425         struct pass_softc *softc;
 426         struct mtx *mtx;
 427         char *physpath;
 428
 429         /*
 430          * If we have one, create a devfs alias for our
 431          * physical path.
 432          */
 433         periph = context;
 434         softc = periph->softc;
 435         physpath = malloc(MAXPATHLEN, M_DEVBUF, M_WAITOK);
 436         mtx = cam_periph_mtx(periph);
 437         mtx_lock(mtx);
 438
 439         if (periph->flags & CAM_PERIPH_INVALID)
 440                 goto out;
 441
 442         if (xpt_getattr(physpath, MAXPATHLEN,
 443                         "GEOM::physpath", periph->path) == 0
 444          && strlen(physpath) != 0) {
 445
 446                 mtx_unlock(mtx);
 447                 make_dev_physpath_alias(MAKEDEV_WAITOK, &softc->alias_dev,
 448                                         softc->dev, softc->alias_dev, physpath);
 449                 mtx_lock(mtx);
 450         }
 451
 452 out:
 453         /*
 454          * Now that we've made our alias, we no longer have to have a
 455          * reference to the device.
 456          */
 457         if ((softc->flags & PASS_FLAG_INITIAL_PHYSPATH) == 0)
 458                 softc->flags |= PASS_FLAG_INITIAL_PHYSPATH;
 459
 460         /*
 461          * We always acquire a reference to the periph before queueing this
 462          * task queue function, so it won't go away before we run.
 463          */
 464         while (pending-- > 0)
 465                 cam_periph_release_locked(periph);
 466         mtx_unlock(mtx);
 467
 468         free(physpath, M_DEVBUF);
 469 }
 470
 471 static void
 472 passasync(void *callback_arg, u_int32_t code,
 473           struct cam_path *path, void *arg)
 474 {
 475         struct cam_periph *periph;
 476
 477         periph = (struct cam_periph *)callback_arg;
 478
 479         switch (code) {
 480         case AC_FOUND_DEVICE:
 481         {
 482                 struct ccb_getdev *cgd;
 483                 cam_status status;
 484
 485                 cgd = (struct ccb_getdev *)arg;
 486                 if (cgd == NULL)
 487                         break;
 488
 489                 /*
 490                  * Allocate a peripheral instance for
 491                  * this device and start the probe
 492                  * process.
 493                  */
 494                 status = cam_periph_alloc(passregister, passoninvalidate,
 495                                           passcleanup, passstart, "pass",
 496                                           CAM_PERIPH_BIO, path,
 497                                           passasync, AC_FOUND_DEVICE, cgd);
 498
 499                 if (status != CAM_REQ_CMP
 500                  && status != CAM_REQ_INPROG) {
 501                         const struct cam_status_entry *entry;
 502
 503                         entry = cam_fetch_status_entry(status);
 504
 505                         printf("passasync: Unable to attach new device "
 506                                "due to status %#x: %s\n", status, entry ?
 507                                entry->status_text : "Unknown");
 508                 }
 509
 510                 break;
 511         }
 512         case AC_ADVINFO_CHANGED:
 513         {
 514                 uintptr_t buftype;
 515
 516                 buftype = (uintptr_t)arg;
 517                 if (buftype == CDAI_TYPE_PHYS_PATH) {
 518                         struct pass_softc *softc;
 519                         cam_status status;
 520
 521                         softc = (struct pass_softc *)periph->softc;
 522                         /*
 523                          * Acquire a reference to the periph before we
 524                          * start the taskqueue, so that we don't run into
 525                          * a situation where the periph goes away before
 526                          * the task queue has a chance to run.
 527                          */
 528                         status = cam_periph_acquire(periph);
 529                         if (status != CAM_REQ_CMP)
 530                                 break;
 531
 532                         taskqueue_enqueue(taskqueue_thread,
 533                                           &softc->add_physpath_task);
 534                 }
 535                 break;
 536         }
 537         default:
 538                 cam_periph_async(periph, code, path, arg);
 539                 break;
 540         }
 541 }
 542
 543 static cam_status
 544 passregister(struct cam_periph *periph, void *arg)
 545 {
 546         struct pass_softc *softc;
 547         struct ccb_getdev *cgd;
 548         struct ccb_pathinq cpi;
 549         struct make_dev_args args;
 550         int error, no_tags;
 551
 552         cgd = (struct ccb_getdev *)arg;
 553         if (cgd == NULL) {
 554                 printf("%s: no getdev CCB, can't register device\n", __func__);
 555                 return(CAM_REQ_CMP_ERR);
 556         }
 557
 558         softc = (struct pass_softc *)malloc(sizeof(*softc),
 559                                             M_DEVBUF, M_NOWAIT);
 560
 561         if (softc == NULL) {
 562                 printf("%s: Unable to probe new device. "
 563                        "Unable to allocate softc\n", __func__);
 564                 return(CAM_REQ_CMP_ERR);
 565         }
 566
 567         bzero(softc, sizeof(*softc));
 568         softc->state = PASS_STATE_NORMAL;
 569         if (cgd->protocol == PROTO_SCSI || cgd->protocol == PROTO_ATAPI)
 570                 softc->pd_type = SID_TYPE(&cgd->inq_data);
 571         else if (cgd->protocol == PROTO_SATAPM)
 572                 softc->pd_type = T_ENCLOSURE;
 573         else
 574                 softc->pd_type = T_DIRECT;
 575
 576         periph->softc = softc;
 577         softc->periph = periph;
 578         TAILQ_INIT(&softc->incoming_queue);
 579         TAILQ_INIT(&softc->active_queue);
 580         TAILQ_INIT(&softc->abandoned_queue);
 581         TAILQ_INIT(&softc->done_queue);
 582         snprintf(softc->zone_name, sizeof(softc->zone_name), "%s%d",
 583                  periph->periph_name, periph->unit_number);
 584         snprintf(softc->io_zone_name, sizeof(softc->io_zone_name), "%s%dIO",
 585                  periph->periph_name, periph->unit_number);
 586         softc->io_zone_size = MAXPHYS;
 587         knlist_init_mtx(&softc->read_select.si_note, cam_periph_mtx(periph));
 588
 589         bzero(&cpi, sizeof(cpi));
 590         xpt_setup_ccb(&cpi.ccb_h, periph->path, CAM_PRIORITY_NORMAL);
 591         cpi.ccb_h.func_code = XPT_PATH_INQ;
 592         xpt_action((union ccb *)&cpi);
 593
 594         if (cpi.maxio == 0)
 595                 softc->maxio = DFLTPHYS;        /* traditional default */
 596         else if (cpi.maxio > MAXPHYS)
 597                 softc->maxio = MAXPHYS;         /* for safety */
 598         else
 599                 softc->maxio = cpi.maxio;       /* real value */
 600
 601         if (cpi.hba_misc & PIM_UNMAPPED)
 602                 softc->flags |= PASS_FLAG_UNMAPPED_CAPABLE;
 603
 604         /*
 605          * We pass in 0 for a blocksize, since we don't
 606          * know what the blocksize of this device is, if
 607          * it even has a blocksize.
 608          */
 609         cam_periph_unlock(periph);
 610         no_tags = (cgd->inq_data.flags & SID_CmdQue) == 0;
 611         softc->device_stats = devstat_new_entry("pass",
 612                           periph->unit_number, 0,
 613                           DEVSTAT_NO_BLOCKSIZE
 614                           | (no_tags ? DEVSTAT_NO_ORDERED_TAGS : 0),
 615                           softc->pd_type |
 616                           XPORT_DEVSTAT_TYPE(cpi.transport) |
 617                           DEVSTAT_TYPE_PASS,
 618                           DEVSTAT_PRIORITY_PASS);
 619
 620         /*
 621          * Initialize the taskqueue handler for shutting down kqueue.
 622          */
 623         TASK_INIT(&softc->shutdown_kqueue_task, /*priority*/ 0,
 624                   pass_shutdown_kqueue, periph);
 625
 626         /*
 627          * Acquire a reference to the periph that we can release once we've
 628          * cleaned up the kqueue.
 629          */
 630         if (cam_periph_acquire(periph) != CAM_REQ_CMP) {
 631                 xpt_print(periph->path, "%s: lost periph during "
 632                           "registration!\n", __func__);
 633                 cam_periph_lock(periph);
 634                 return (CAM_REQ_CMP_ERR);
 635         }
 636
 637         /*
 638          * Acquire a reference to the periph before we create the devfs
 639          * instance for it.  We'll release this reference once the devfs
 640          * instance has been freed.
 641          */
 642         if (cam_periph_acquire(periph) != CAM_REQ_CMP) {
 643                 xpt_print(periph->path, "%s: lost periph during "
 644                           "registration!\n", __func__);
 645                 cam_periph_lock(periph);
 646                 return (CAM_REQ_CMP_ERR);
 647         }
 648
 649         /* Register the device */
 650         make_dev_args_init(&args);
 651         args.mda_devsw = &pass_cdevsw;
 652         args.mda_unit = periph->unit_number;
 653         args.mda_uid = UID_ROOT;
 654         args.mda_gid = GID_OPERATOR;
 655         args.mda_mode = 0600;
 656         args.mda_si_drv1 = periph;
 657         error = make_dev_s(&args, &softc->dev, "%s%d", periph->periph_name,
 658             periph->unit_number);
 659         if (error != 0) {
 660                 cam_periph_lock(periph);
 661                 cam_periph_release_locked(periph);
 662                 return (CAM_REQ_CMP_ERR);
 663         }
 664
 665         /*
 666          * Hold a reference to the periph before we create the physical
 667          * path alias so it can't go away.
 668          */
 669         if (cam_periph_acquire(periph) != CAM_REQ_CMP) {
 670                 xpt_print(periph->path, "%s: lost periph during "
 671                           "registration!\n", __func__);
 672                 cam_periph_lock(periph);
 673                 return (CAM_REQ_CMP_ERR);
 674         }
 675
 676         cam_periph_lock(periph);
 677
 678         TASK_INIT(&softc->add_physpath_task, /*priority*/0,
 679                   pass_add_physpath, periph);
 680
 681         /*
 682          * See if physical path information is already available.
 683          */
 684         taskqueue_enqueue(taskqueue_thread, &softc->add_physpath_task);
 685
 686         /*
 687          * Add an async callback so that we get notified if
 688          * this device goes away or its physical path
 689          * (stored in the advanced info data of the EDT) has
 690          * changed.
 691          */
 692         xpt_register_async(AC_LOST_DEVICE | AC_ADVINFO_CHANGED,
 693                            passasync, periph, periph->path);
 694
 695         if (bootverbose)
 696                 xpt_announce_periph(periph, NULL);
 697
 698         return(CAM_REQ_CMP);
 699 }
 700
 701 static int
 702 passopen(struct cdev *dev, int flags, int fmt, struct thread *td)
 703 {
 704         struct cam_periph *periph;
 705         struct pass_softc *softc;
 706         int error;
 707
 708         periph = (struct cam_periph *)dev->si_drv1;
 709         if (cam_periph_acquire(periph) != CAM_REQ_CMP)
 710                 return (ENXIO);
 711
 712         cam_periph_lock(periph);
 713
 714         softc = (struct pass_softc *)periph->softc;
 715
 716         if (softc->flags & PASS_FLAG_INVALID) {
 717                 cam_periph_release_locked(periph);
 718                 cam_periph_unlock(periph);
 719                 return(ENXIO);
 720         }
 721
 722         /*
 723          * Don't allow access when we're running at a high securelevel.
 724          */
 725         error = securelevel_gt(td->td_ucred, 1);
 726         if (error) {
 727                 cam_periph_release_locked(periph);
 728                 cam_periph_unlock(periph);
 729                 return(error);
 730         }
 731
 732         /*
 733          * Only allow read-write access.
 734          */
 735         if (((flags & FWRITE) == 0) || ((flags & FREAD) == 0)) {
 736                 cam_periph_release_locked(periph);
 737                 cam_periph_unlock(periph);
 738                 return(EPERM);
 739         }
 740
 741         /*
 742          * We don't allow nonblocking access.
 743          */
 744         if ((flags & O_NONBLOCK) != 0) {
 745                 xpt_print(periph->path, "can't do nonblocking access\n");
 746                 cam_periph_release_locked(periph);
 747                 cam_periph_unlock(periph);
 748                 return(EINVAL);
 749         }
 750
 751         softc->open_count++;
 752
 753         cam_periph_unlock(periph);
 754
 755         return (error);
 756 }
 757
 758 static int
 759 passclose(struct cdev *dev, int flag, int fmt, struct thread *td)
 760 {
 761         struct  cam_periph *periph;
 762         struct  pass_softc *softc;
 763         struct mtx *mtx;
 764
 765         periph = (struct cam_periph *)dev->si_drv1;
 766         mtx = cam_periph_mtx(periph);
 767         mtx_lock(mtx);
 768
 769         softc = periph->softc;
 770         softc->open_count--;
 771
 772         if (softc->open_count == 0) {
 773                 struct pass_io_req *io_req, *io_req2;
 774                 int need_unlock;
 775
 776                 need_unlock = 0;
 777
 778                 TAILQ_FOREACH_SAFE(io_req, &softc->done_queue, links, io_req2) {
 779                         TAILQ_REMOVE(&softc->done_queue, io_req, links);
 780                         passiocleanup(softc, io_req);
 781                         uma_zfree(softc->pass_zone, io_req);
 782                 }
 783
 784                 TAILQ_FOREACH_SAFE(io_req, &softc->incoming_queue, links,
 785                                    io_req2) {
 786                         TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
 787                         passiocleanup(softc, io_req);
 788                         uma_zfree(softc->pass_zone, io_req);
 789                 }
 790
 791                 /*
 792                  * If there are any active I/Os, we need to forcibly acquire a
 793                  * reference to the peripheral so that we don't go away
 794                  * before they complete.  We'll release the reference when
 795                  * the abandoned queue is empty.
 796                  */
 797                 io_req = TAILQ_FIRST(&softc->active_queue);
 798                 if ((io_req != NULL)
 799                  && (softc->flags & PASS_FLAG_ABANDONED_REF_SET) == 0) {
 800                         cam_periph_doacquire(periph);
 801                         softc->flags |= PASS_FLAG_ABANDONED_REF_SET;
 802                 }
 803
 804                 /*
 805                  * Since the I/O in the active queue is not under our
 806                  * control, just set a flag so that we can clean it up when
 807                  * it completes and put it on the abandoned queue.  This
 808                  * will prevent our sending spurious completions in the
 809                  * event that the device is opened again before these I/Os
 810                  * complete.
 811                  */
 812                 TAILQ_FOREACH_SAFE(io_req, &softc->active_queue, links,
 813                                    io_req2) {
 814                         TAILQ_REMOVE(&softc->active_queue, io_req, links);
 815                         io_req->flags |= PASS_IO_ABANDONED;
 816                         TAILQ_INSERT_TAIL(&softc->abandoned_queue, io_req,
 817                                           links);
 818                 }
 819         }
 820
 821         cam_periph_release_locked(periph);
 822
 823         /*
 824          * We reference the lock directly here, instead of using
 825          * cam_periph_unlock().  The reason is that the call to
 826          * cam_periph_release_locked() above could result in the periph
 827          * getting freed.  If that is the case, dereferencing the periph
 828          * with a cam_periph_unlock() call would cause a page fault.
 829          *
 830          * cam_periph_release() avoids this problem using the same method,
 831          * but we're manually acquiring and dropping the lock here to
 832          * protect the open count and avoid another lock acquisition and
 833          * release.
 834          */
 835         mtx_unlock(mtx);
 836
 837         return (0);
 838 }
 839
 840
 841 static void
 842 passstart(struct cam_periph *periph, union ccb *start_ccb)
 843 {
 844         struct pass_softc *softc;
 845
 846         softc = (struct pass_softc *)periph->softc;
 847
 848         switch (softc->state) {
 849         case PASS_STATE_NORMAL: {
 850                 struct pass_io_req *io_req;
 851
 852                 /*
 853                  * Check for any queued I/O requests that require an
 854                  * allocated slot.
 855                  */
 856                 io_req = TAILQ_FIRST(&softc->incoming_queue);
 857                 if (io_req == NULL) {
 858                         xpt_release_ccb(start_ccb);
 859                         break;
 860                 }
 861                 TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
 862                 TAILQ_INSERT_TAIL(&softc->active_queue, io_req, links);
 863                 /*
 864                  * Merge the user's CCB into the allocated CCB.
 865                  */
 866                 xpt_merge_ccb(start_ccb, &io_req->ccb);
 867                 start_ccb->ccb_h.ccb_type = PASS_CCB_QUEUED_IO;
 868                 start_ccb->ccb_h.ccb_ioreq = io_req;
 869                 start_ccb->ccb_h.cbfcnp = passdone;
 870                 io_req->alloced_ccb = start_ccb;
 871                 binuptime(&io_req->start_time);
 872                 devstat_start_transaction(softc->device_stats,
 873                                           &io_req->start_time);
 874
 875                 xpt_action(start_ccb);
 876
 877                 /*
 878                  * If we have any more I/O waiting, schedule ourselves again.
 879                  */
 880                 if (!TAILQ_EMPTY(&softc->incoming_queue))
 881                         xpt_schedule(periph, CAM_PRIORITY_NORMAL);
 882                 break;
 883         }
 884         default:
 885                 break;
 886         }
 887 }
 888
 889 static void
 890 passdone(struct cam_periph *periph, union ccb *done_ccb)
 891 {
 892         struct pass_softc *softc;
 893         struct ccb_scsiio *csio;
 894
 895         softc = (struct pass_softc *)periph->softc;
 896
 897         cam_periph_assert(periph, MA_OWNED);
 898
 899         csio = &done_ccb->csio;
 900         switch (csio->ccb_h.ccb_type) {
 901         case PASS_CCB_QUEUED_IO: {
 902                 struct pass_io_req *io_req;
 903
 904                 io_req = done_ccb->ccb_h.ccb_ioreq;
 905 #if 0
 906                 xpt_print(periph->path, "%s: called for user CCB %p\n",
 907                           __func__, io_req->user_ccb_ptr);
 908 #endif
 909                 if (((done_ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_CMP)
 910                  && (done_ccb->ccb_h.flags & CAM_PASS_ERR_RECOVER)
 911                  && ((io_req->flags & PASS_IO_ABANDONED) == 0)) {
 912                         int error;
 913
 914                         error = passerror(done_ccb, CAM_RETRY_SELTO,
 915                                           SF_RETRY_UA | SF_NO_PRINT);
 916
 917                         if (error == ERESTART) {
 918                                 /*
 919                                  * A retry was scheduled, so
 920                                  * just return.
 921                                  */
 922                                 return;
 923                         }
 924                 }
 925
 926                 /*
 927                  * Copy the allocated CCB contents back to the malloced CCB
 928                  * so we can give status back to the user when he requests it.
 929                  */
 930                 bcopy(done_ccb, &io_req->ccb, sizeof(*done_ccb));
 931
 932                 /*
 933                  * Log data/transaction completion with devstat(9).
 934                  */
 935                 switch (done_ccb->ccb_h.func_code) {
 936                 case XPT_SCSI_IO:
 937                         devstat_end_transaction(softc->device_stats,
 938                             done_ccb->csio.dxfer_len - done_ccb->csio.resid,
 939                             done_ccb->csio.tag_action & 0x3,
 940                             ((done_ccb->ccb_h.flags & CAM_DIR_MASK) ==
 941                             CAM_DIR_NONE) ? DEVSTAT_NO_DATA :
 942                             (done_ccb->ccb_h.flags & CAM_DIR_OUT) ?
 943                             DEVSTAT_WRITE : DEVSTAT_READ, NULL,
 944                             &io_req->start_time);
 945                         break;
 946                 case XPT_ATA_IO:
 947                         devstat_end_transaction(softc->device_stats,
 948                             done_ccb->ataio.dxfer_len - done_ccb->ataio.resid,
 949                             done_ccb->ataio.tag_action & 0x3,
 950                             ((done_ccb->ccb_h.flags & CAM_DIR_MASK) ==
 951                             CAM_DIR_NONE) ? DEVSTAT_NO_DATA :
 952                             (done_ccb->ccb_h.flags & CAM_DIR_OUT) ?
 953                             DEVSTAT_WRITE : DEVSTAT_READ, NULL,
 954                             &io_req->start_time);
 955                         break;
 956                 case XPT_SMP_IO:
 957                         /*
 958                          * XXX KDM this isn't quite right, but there isn't
 959                          * currently an easy way to represent a bidirectional
 960                          * transfer in devstat.  The only way to do it
 961                          * and have the byte counts come out right would
 962                          * mean that we would have to record two
 963                          * transactions, one for the request and one for the
 964                          * response.  For now, so that we report something,
 965                          * just treat the entire thing as a read.
 966                          */
 967                         devstat_end_transaction(softc->device_stats,
 968                             done_ccb->smpio.smp_request_len +
 969                             done_ccb->smpio.smp_response_len,
 970                             DEVSTAT_TAG_SIMPLE, DEVSTAT_READ, NULL,
 971                             &io_req->start_time);
 972                         break;
 973                 default:
 974                         devstat_end_transaction(softc->device_stats, 0,
 975                             DEVSTAT_TAG_NONE, DEVSTAT_NO_DATA, NULL,
 976                             &io_req->start_time);
 977                         break;
 978                 }
 979
 980                 /*
 981                  * In the normal case, take the completed I/O off of the
 982                  * active queue and put it on the done queue.  Notitfy the
 983                  * user that we have a completed I/O.
 984                  */
 985                 if ((io_req->flags & PASS_IO_ABANDONED) == 0) {
 986                         TAILQ_REMOVE(&softc->active_queue, io_req, links);
 987                         TAILQ_INSERT_TAIL(&softc->done_queue, io_req, links);
 988                         selwakeuppri(&softc->read_select, PRIBIO);
 989                         KNOTE_LOCKED(&softc->read_select.si_note, 0);
 990                 } else {
 991                         /*
 992                          * In the case of an abandoned I/O (final close
 993                          * without fetching the I/O), take it off of the
 994                          * abandoned queue and free it.
 995                          */
 996                         TAILQ_REMOVE(&softc->abandoned_queue, io_req, links);
 997                         passiocleanup(softc, io_req);
 998                         uma_zfree(softc->pass_zone, io_req);
 999
1000                         /*
1001                          * Release the done_ccb here, since we may wind up
1002                          * freeing the peripheral when we decrement the
1003                          * reference count below.
1004                          */
1005                         xpt_release_ccb(done_ccb);
1006
1007                         /*
1008                          * If the abandoned queue is empty, we can release
1009                          * our reference to the periph since we won't have
1010                          * any more completions coming.
1011                          */
1012                         if ((TAILQ_EMPTY(&softc->abandoned_queue))
1013                          && (softc->flags & PASS_FLAG_ABANDONED_REF_SET)) {
1014                                 softc->flags &= ~PASS_FLAG_ABANDONED_REF_SET;
1015                                 cam_periph_release_locked(periph);
1016                         }
1017
1018                         /*
1019                          * We have already released the CCB, so we can
1020                          * return.
1021                          */
1022                         return;
1023                 }
1024                 break;
1025         }
1026         }
1027         xpt_release_ccb(done_ccb);
1028 }
1029
1030 static int
1031 passcreatezone(struct cam_periph *periph)
1032 {
1033         struct pass_softc *softc;
1034         int error;
1035
1036         error = 0;
1037         softc = (struct pass_softc *)periph->softc;
1038
1039         cam_periph_assert(periph, MA_OWNED);
1040         KASSERT(((softc->flags & PASS_FLAG_ZONE_VALID) == 0),
1041                 ("%s called when the pass(4) zone is valid!\n", __func__));
1042         KASSERT((softc->pass_zone == NULL),
1043                 ("%s called when the pass(4) zone is allocated!\n", __func__));
1044
1045         if ((softc->flags & PASS_FLAG_ZONE_INPROG) == 0) {
1046
1047                 /*
1048                  * We're the first context through, so we need to create
1049                  * the pass(4) UMA zone for I/O requests.
1050                  */
1051                 softc->flags |= PASS_FLAG_ZONE_INPROG;
1052
1053                 /*
1054                  * uma_zcreate() does a blocking (M_WAITOK) allocation,
1055                  * so we cannot hold a mutex while we call it.
1056                  */
1057                 cam_periph_unlock(periph);
1058
1059                 softc->pass_zone = uma_zcreate(softc->zone_name,
1060                     sizeof(struct pass_io_req), NULL, NULL, NULL, NULL,
1061                     /*align*/ 0, /*flags*/ 0);
1062
1063                 softc->pass_io_zone = uma_zcreate(softc->io_zone_name,
1064                     softc->io_zone_size, NULL, NULL, NULL, NULL,
1065                     /*align*/ 0, /*flags*/ 0);
1066
1067                 cam_periph_lock(periph);
1068
1069                 if ((softc->pass_zone == NULL)
1070                  || (softc->pass_io_zone == NULL)) {
1071                         if (softc->pass_zone == NULL)
1072                                 xpt_print(periph->path, "unable to allocate "
1073                                     "IO Req UMA zone\n");
1074                         else
1075                                 xpt_print(periph->path, "unable to allocate "
1076                                     "IO UMA zone\n");
1077                         softc->flags &= ~PASS_FLAG_ZONE_INPROG;
1078                         goto bailout;
1079                 }
1080
1081                 /*
1082                  * Set the flags appropriately and notify any other waiters.
1083                  */
1084                 softc->flags &= PASS_FLAG_ZONE_INPROG;
1085                 softc->flags |= PASS_FLAG_ZONE_VALID;
1086                 wakeup(&softc->pass_zone);
1087         } else {
1088                 /*
1089                  * In this case, the UMA zone has not yet been created, but
1090                  * another context is in the process of creating it.  We
1091                  * need to sleep until the creation is either done or has
1092                  * failed.
1093                  */
1094                 while ((softc->flags & PASS_FLAG_ZONE_INPROG)
1095                     && ((softc->flags & PASS_FLAG_ZONE_VALID) == 0)) {
1096                         error = msleep(&softc->pass_zone,
1097                                        cam_periph_mtx(periph), PRIBIO,
1098                                        "paszon", 0);
1099                         if (error != 0)
1100                                 goto bailout;
1101                 }
1102                 /*
1103                  * If the zone creation failed, no luck for the user.
1104                  */
1105                 if ((softc->flags & PASS_FLAG_ZONE_VALID) == 0){
1106                         error = ENOMEM;
1107                         goto bailout;
1108                 }
1109         }
1110 bailout:
1111         return (error);
1112 }
1113
1114 static void
1115 passiocleanup(struct pass_softc *softc, struct pass_io_req *io_req)
1116 {
1117         union ccb *ccb;
1118         u_int8_t **data_ptrs[CAM_PERIPH_MAXMAPS];
1119         int i, numbufs;
1120
1121         ccb = &io_req->ccb;
1122
1123         switch (ccb->ccb_h.func_code) {
1124         case XPT_DEV_MATCH:
1125                 numbufs = min(io_req->num_bufs, 2);
1126
1127                 if (numbufs == 1) {
1128                         data_ptrs[0] = (u_int8_t **)&ccb->cdm.matches;
1129                 } else {
1130                         data_ptrs[0] = (u_int8_t **)&ccb->cdm.patterns;
1131                         data_ptrs[1] = (u_int8_t **)&ccb->cdm.matches;
1132                 }
1133                 break;
1134         case XPT_SCSI_IO:
1135         case XPT_CONT_TARGET_IO:
1136                 data_ptrs[0] = &ccb->csio.data_ptr;
1137                 numbufs = min(io_req->num_bufs, 1);
1138                 break;
1139         case XPT_ATA_IO:
1140                 data_ptrs[0] = &ccb->ataio.data_ptr;
1141                 numbufs = min(io_req->num_bufs, 1);
1142                 break;
1143         case XPT_SMP_IO:
1144                 numbufs = min(io_req->num_bufs, 2);
1145                 data_ptrs[0] = &ccb->smpio.smp_request;
1146                 data_ptrs[1] = &ccb->smpio.smp_response;
1147                 break;
1148         case XPT_DEV_ADVINFO:
1149                 numbufs = min(io_req->num_bufs, 1);
1150                 data_ptrs[0] = (uint8_t **)&ccb->cdai.buf;
1151                 break;
1152         default:
1153                 /* allow ourselves to be swapped once again */
1154                 return;
1155                 break; /* NOTREACHED */
1156         }
1157
1158         if (io_req->flags & PASS_IO_USER_SEG_MALLOC) {
1159                 free(io_req->user_segptr, M_SCSIPASS);
1160                 io_req->user_segptr = NULL;
1161         }
1162
1163         /*
1164          * We only want to free memory we malloced.
1165          */
1166         if (io_req->data_flags == CAM_DATA_VADDR) {
1167                 for (i = 0; i < io_req->num_bufs; i++) {
1168                         if (io_req->kern_bufs[i] == NULL)
1169                                 continue;
1170
1171                         free(io_req->kern_bufs[i], M_SCSIPASS);
1172                         io_req->kern_bufs[i] = NULL;
1173                 }
1174         } else if (io_req->data_flags == CAM_DATA_SG) {
1175                 for (i = 0; i < io_req->num_kern_segs; i++) {
1176                         if ((uint8_t *)(uintptr_t)
1177                             io_req->kern_segptr[i].ds_addr == NULL)
1178                                 continue;
1179
1180                         uma_zfree(softc->pass_io_zone, (uint8_t *)(uintptr_t)
1181                             io_req->kern_segptr[i].ds_addr);
1182                         io_req->kern_segptr[i].ds_addr = 0;
1183                 }
1184         }
1185
1186         if (io_req->flags & PASS_IO_KERN_SEG_MALLOC) {
1187                 free(io_req->kern_segptr, M_SCSIPASS);
1188                 io_req->kern_segptr = NULL;
1189         }
1190
1191         if (io_req->data_flags != CAM_DATA_PADDR) {
1192                 for (i = 0; i < numbufs; i++) {
1193                         /*
1194                          * Restore the user's buffer pointers to their
1195                          * previous values.
1196                          */
1197                         if (io_req->user_bufs[i] != NULL)
1198                                 *data_ptrs[i] = io_req->user_bufs[i];
1199                 }
1200         }
1201
1202 }
1203
1204 static int
1205 passcopysglist(struct cam_periph *periph, struct pass_io_req *io_req,
1206                ccb_flags direction)
1207 {
1208         bus_size_t kern_watermark, user_watermark, len_copied, len_to_copy;
1209         bus_dma_segment_t *user_sglist, *kern_sglist;
1210         int i, j, error;
1211
1212         error = 0;
1213         kern_watermark = 0;
1214         user_watermark = 0;
1215         len_to_copy = 0;
1216         len_copied = 0;
1217         user_sglist = io_req->user_segptr;
1218         kern_sglist = io_req->kern_segptr;
1219
1220         for (i = 0, j = 0; i < io_req->num_user_segs &&
1221              j < io_req->num_kern_segs;) {
1222                 uint8_t *user_ptr, *kern_ptr;
1223
1224                 len_to_copy = min(user_sglist[i].ds_len -user_watermark,
1225                     kern_sglist[j].ds_len - kern_watermark);
1226
1227                 user_ptr = (uint8_t *)(uintptr_t)user_sglist[i].ds_addr;
1228                 user_ptr = user_ptr + user_watermark;
1229                 kern_ptr = (uint8_t *)(uintptr_t)kern_sglist[j].ds_addr;
1230                 kern_ptr = kern_ptr + kern_watermark;
1231
1232                 user_watermark += len_to_copy;
1233                 kern_watermark += len_to_copy;
1234
1235                 if (!useracc(user_ptr, len_to_copy,
1236                     (direction == CAM_DIR_IN) ? VM_PROT_WRITE : VM_PROT_READ)) {
1237                         xpt_print(periph->path, "%s: unable to access user "
1238                                   "S/G list element %p len %zu\n", __func__,
1239                                   user_ptr, len_to_copy);
1240                         error = EFAULT;
1241                         goto bailout;
1242                 }
1243
1244                 if (direction == CAM_DIR_IN) {
1245                         error = copyout(kern_ptr, user_ptr, len_to_copy);
1246                         if (error != 0) {
1247                                 xpt_print(periph->path, "%s: copyout of %u "
1248                                           "bytes from %p to %p failed with "
1249                                           "error %d\n", __func__, len_to_copy,
1250                                           kern_ptr, user_ptr, error);
1251                                 goto bailout;
1252                         }
1253                 } else {
1254                         error = copyin(user_ptr, kern_ptr, len_to_copy);
1255                         if (error != 0) {
1256                                 xpt_print(periph->path, "%s: copyin of %u "
1257                                           "bytes from %p to %p failed with "
1258                                           "error %d\n", __func__, len_to_copy,
1259                                           user_ptr, kern_ptr, error);
1260                                 goto bailout;
1261                         }
1262                 }
1263
1264                 len_copied += len_to_copy;
1265
1266                 if (user_sglist[i].ds_len == user_watermark) {
1267                         i++;
1268                         user_watermark = 0;
1269                 }
1270
1271                 if (kern_sglist[j].ds_len == kern_watermark) {
1272                         j++;
1273                         kern_watermark = 0;
1274                 }
1275         }
1276
1277 bailout:
1278
1279         return (error);
1280 }
1281
1282 static int
1283 passmemsetup(struct cam_periph *periph, struct pass_io_req *io_req)
1284 {
1285         union ccb *ccb;
1286         struct pass_softc *softc;
1287         int numbufs, i;
1288         uint8_t **data_ptrs[CAM_PERIPH_MAXMAPS];
1289         uint32_t lengths[CAM_PERIPH_MAXMAPS];
1290         uint32_t dirs[CAM_PERIPH_MAXMAPS];
1291         uint32_t num_segs;
1292         uint16_t *seg_cnt_ptr;
1293         size_t maxmap;
1294         int error;
1295
1296         cam_periph_assert(periph, MA_NOTOWNED);
1297
1298         softc = periph->softc;
1299
1300         error = 0;
1301         ccb = &io_req->ccb;
1302         maxmap = 0;
1303         num_segs = 0;
1304         seg_cnt_ptr = NULL;
1305
1306         switch(ccb->ccb_h.func_code) {
1307         case XPT_DEV_MATCH:
1308                 if (ccb->cdm.match_buf_len == 0) {
1309                         printf("%s: invalid match buffer length 0\n", __func__);
1310                         return(EINVAL);
1311                 }
1312                 if (ccb->cdm.pattern_buf_len > 0) {
1313                         data_ptrs[0] = (u_int8_t **)&ccb->cdm.patterns;
1314                         lengths[0] = ccb->cdm.pattern_buf_len;
1315                         dirs[0] = CAM_DIR_OUT;
1316                         data_ptrs[1] = (u_int8_t **)&ccb->cdm.matches;
1317                         lengths[1] = ccb->cdm.match_buf_len;
1318                         dirs[1] = CAM_DIR_IN;
1319                         numbufs = 2;
1320                 } else {
1321                         data_ptrs[0] = (u_int8_t **)&ccb->cdm.matches;
1322                         lengths[0] = ccb->cdm.match_buf_len;
1323                         dirs[0] = CAM_DIR_IN;
1324                         numbufs = 1;
1325                 }
1326                 io_req->data_flags = CAM_DATA_VADDR;
1327                 break;
1328         case XPT_SCSI_IO:
1329         case XPT_CONT_TARGET_IO:
1330                 if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
1331                         return(0);
1332
1333                 /*
1334                  * The user shouldn't be able to supply a bio.
1335                  */
1336                 if ((ccb->ccb_h.flags & CAM_DATA_MASK) == CAM_DATA_BIO)
1337                         return (EINVAL);
1338
1339                 io_req->data_flags = ccb->ccb_h.flags & CAM_DATA_MASK;
1340
1341                 data_ptrs[0] = &ccb->csio.data_ptr;
1342                 lengths[0] = ccb->csio.dxfer_len;
1343                 dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1344                 num_segs = ccb->csio.sglist_cnt;
1345                 seg_cnt_ptr = &ccb->csio.sglist_cnt;
1346                 numbufs = 1;
1347                 maxmap = softc->maxio;
1348                 break;
1349         case XPT_ATA_IO:
1350                 if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
1351                         return(0);
1352
1353                 /*
1354                  * We only support a single virtual address for ATA I/O.
1355                  */
1356                 if ((ccb->ccb_h.flags & CAM_DATA_MASK) != CAM_DATA_VADDR)
1357                         return (EINVAL);
1358
1359                 io_req->data_flags = CAM_DATA_VADDR;
1360
1361                 data_ptrs[0] = &ccb->ataio.data_ptr;
1362                 lengths[0] = ccb->ataio.dxfer_len;
1363                 dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1364                 numbufs = 1;
1365                 maxmap = softc->maxio;
1366                 break;
1367         case XPT_SMP_IO:
1368                 io_req->data_flags = CAM_DATA_VADDR;
1369
1370                 data_ptrs[0] = &ccb->smpio.smp_request;
1371                 lengths[0] = ccb->smpio.smp_request_len;
1372                 dirs[0] = CAM_DIR_OUT;
1373                 data_ptrs[1] = &ccb->smpio.smp_response;
1374                 lengths[1] = ccb->smpio.smp_response_len;
1375                 dirs[1] = CAM_DIR_IN;
1376                 numbufs = 2;
1377                 maxmap = softc->maxio;
1378                 break;
1379         case XPT_DEV_ADVINFO:
1380                 if (ccb->cdai.bufsiz == 0)
1381                         return (0);
1382
1383                 io_req->data_flags = CAM_DATA_VADDR;
1384
1385                 data_ptrs[0] = (uint8_t **)&ccb->cdai.buf;
1386                 lengths[0] = ccb->cdai.bufsiz;
1387                 dirs[0] = CAM_DIR_IN;
1388                 numbufs = 1;
1389                 break;
1390         default:
1391                 return(EINVAL);
1392                 break; /* NOTREACHED */
1393         }
1394
1395         io_req->num_bufs = numbufs;
1396
1397         /*
1398          * If there is a maximum, check to make sure that the user's
1399          * request fits within the limit.  In general, we should only have
1400          * a maximum length for requests that go to hardware.  Otherwise it
1401          * is whatever we're able to malloc.
1402          */
1403         for (i = 0; i < numbufs; i++) {
1404                 io_req->user_bufs[i] = *data_ptrs[i];
1405                 io_req->dirs[i] = dirs[i];
1406                 io_req->lengths[i] = lengths[i];
1407
1408                 if (maxmap == 0)
1409                         continue;
1410
1411                 if (lengths[i] <= maxmap)
1412                         continue;
1413
1414                 xpt_print(periph->path, "%s: data length %u > max allowed %u "
1415                           "bytes\n", __func__, lengths[i], maxmap);
1416                 error = EINVAL;
1417                 goto bailout;
1418         }
1419
1420         switch (io_req->data_flags) {
1421         case CAM_DATA_VADDR:
1422                 /* Map or copy the buffer into kernel address space */
1423                 for (i = 0; i < numbufs; i++) {
1424                         uint8_t *tmp_buf;
1425
1426                         /*
1427                          * If for some reason no length is specified, we
1428                          * don't need to allocate anything.
1429                          */
1430                         if (io_req->lengths[i] == 0)
1431                                 continue;
1432
1433                         /*
1434                          * Make sure that the user's buffer is accessible
1435                          * to that process.
1436                          */
1437                         if (!useracc(io_req->user_bufs[i], io_req->lengths[i],
1438                             (io_req->dirs[i] == CAM_DIR_IN) ? VM_PROT_WRITE :
1439                              VM_PROT_READ)) {
1440                                 xpt_print(periph->path, "%s: user address %p "
1441                                     "length %u is not accessible\n", __func__,
1442                                     io_req->user_bufs[i], io_req->lengths[i]);
1443                                 error = EFAULT;
1444                                 goto bailout;
1445                         }
1446
1447                         tmp_buf = malloc(lengths[i], M_SCSIPASS,
1448                                          M_WAITOK | M_ZERO);
1449                         io_req->kern_bufs[i] = tmp_buf;
1450                         *data_ptrs[i] = tmp_buf;
1451
1452 #if 0
1453                         xpt_print(periph->path, "%s: malloced %p len %u, user "
1454                                   "buffer %p, operation: %s\n", __func__,
1455                                   tmp_buf, lengths[i], io_req->user_bufs[i],
1456                                   (dirs[i] == CAM_DIR_IN) ? "read" : "write");
1457 #endif
1458                         /*
1459                          * We only need to copy in if the user is writing.
1460                          */
1461                         if (dirs[i] != CAM_DIR_OUT)
1462                                 continue;
1463
1464                         error = copyin(io_req->user_bufs[i],
1465                                        io_req->kern_bufs[i], lengths[i]);
1466                         if (error != 0) {
1467                                 xpt_print(periph->path, "%s: copy of user "
1468                                           "buffer from %p to %p failed with "
1469                                           "error %d\n", __func__,
1470                                           io_req->user_bufs[i],
1471                                           io_req->kern_bufs[i], error);
1472                                 goto bailout;
1473                         }
1474                 }
1475                 break;
1476         case CAM_DATA_PADDR:
1477                 /* Pass down the pointer as-is */
1478                 break;
1479         case CAM_DATA_SG: {
1480                 size_t sg_length, size_to_go, alloc_size;
1481                 uint32_t num_segs_needed;
1482
1483                 /*
1484                  * Copy the user S/G list in, and then copy in the
1485                  * individual segments.
1486                  */
1487                 /*
1488                  * We shouldn't see this, but check just in case.
1489                  */
1490                 if (numbufs != 1) {
1491                         xpt_print(periph->path, "%s: cannot currently handle "
1492                                   "more than one S/G list per CCB\n", __func__);
1493                         error = EINVAL;
1494                         goto bailout;
1495                 }
1496
1497                 /*
1498                  * We have to have at least one segment.
1499                  */
1500                 if (num_segs == 0) {
1501                         xpt_print(periph->path, "%s: CAM_DATA_SG flag set, "
1502                                   "but sglist_cnt=0!\n", __func__);
1503                         error = EINVAL;
1504                         goto bailout;
1505                 }
1506
1507                 /*
1508                  * Make sure the user specified the total length and didn't
1509                  * just leave it to us to decode the S/G list.
1510                  */
1511                 if (lengths[0] == 0) {
1512                         xpt_print(periph->path, "%s: no dxfer_len specified, "
1513                                   "but CAM_DATA_SG flag is set!\n", __func__);
1514                         error = EINVAL;
1515                         goto bailout;
1516                 }
1517
1518                 /*
1519                  * We allocate buffers in io_zone_size increments for an
1520                  * S/G list.  This will generally be MAXPHYS.
1521                  */
1522                 if (lengths[0] <= softc->io_zone_size)
1523                         num_segs_needed = 1;
1524                 else {
1525                         num_segs_needed = lengths[0] / softc->io_zone_size;
1526                         if ((lengths[0] % softc->io_zone_size) != 0)
1527                                 num_segs_needed++;
1528                 }
1529
1530                 /* Figure out the size of the S/G list */
1531                 sg_length = num_segs * sizeof(bus_dma_segment_t);
1532                 io_req->num_user_segs = num_segs;
1533                 io_req->num_kern_segs = num_segs_needed;
1534
1535                 /* Save the user's S/G list pointer for later restoration */
1536                 io_req->user_bufs[0] = *data_ptrs[0];
1537
1538                 /*
1539                  * If we have enough segments allocated by default to handle
1540                  * the length of the user's S/G list,
1541                  */
1542                 if (num_segs > PASS_MAX_SEGS) {
1543                         io_req->user_segptr = malloc(sizeof(bus_dma_segment_t) *
1544                             num_segs, M_SCSIPASS, M_WAITOK | M_ZERO);
1545                         io_req->flags |= PASS_IO_USER_SEG_MALLOC;
1546                 } else
1547                         io_req->user_segptr = io_req->user_segs;
1548
1549                 if (!useracc(*data_ptrs[0], sg_length, VM_PROT_READ)) {
1550                         xpt_print(periph->path, "%s: unable to access user "
1551                                   "S/G list at %p\n", __func__, *data_ptrs[0]);
1552                         error = EFAULT;
1553                         goto bailout;
1554                 }
1555
1556                 error = copyin(*data_ptrs[0], io_req->user_segptr, sg_length);
1557                 if (error != 0) {
1558                         xpt_print(periph->path, "%s: copy of user S/G list "
1559                                   "from %p to %p failed with error %d\n",
1560                                   __func__, *data_ptrs[0], io_req->user_segptr,
1561                                   error);
1562                         goto bailout;
1563                 }
1564
1565                 if (num_segs_needed > PASS_MAX_SEGS) {
1566                         io_req->kern_segptr = malloc(sizeof(bus_dma_segment_t) *
1567                             num_segs_needed, M_SCSIPASS, M_WAITOK | M_ZERO);
1568                         io_req->flags |= PASS_IO_KERN_SEG_MALLOC;
1569                 } else {
1570                         io_req->kern_segptr = io_req->kern_segs;
1571                 }
1572
1573                 /*
1574                  * Allocate the kernel S/G list.
1575                  */
1576                 for (size_to_go = lengths[0], i = 0;
1577                      size_to_go > 0 && i < num_segs_needed;
1578                      i++, size_to_go -= alloc_size) {
1579                         uint8_t *kern_ptr;
1580
1581                         alloc_size = min(size_to_go, softc->io_zone_size);
1582                         kern_ptr = uma_zalloc(softc->pass_io_zone, M_WAITOK);
1583                         io_req->kern_segptr[i].ds_addr =
1584                             (bus_addr_t)(uintptr_t)kern_ptr;
1585                         io_req->kern_segptr[i].ds_len = alloc_size;
1586                 }
1587                 if (size_to_go > 0) {
1588                         printf("%s: size_to_go = %zu, software error!\n",
1589                                __func__, size_to_go);
1590                         error = EINVAL;
1591                         goto bailout;
1592                 }
1593
1594                 *data_ptrs[0] = (uint8_t *)io_req->kern_segptr;
1595                 *seg_cnt_ptr = io_req->num_kern_segs;
1596
1597                 /*
1598                  * We only need to copy data here if the user is writing.
1599                  */
1600                 if (dirs[0] == CAM_DIR_OUT)
1601                         error = passcopysglist(periph, io_req, dirs[0]);
1602                 break;
1603         }
1604         case CAM_DATA_SG_PADDR: {
1605                 size_t sg_length;
1606
1607                 /*
1608                  * We shouldn't see this, but check just in case.
1609                  */
1610                 if (numbufs != 1) {
1611                         printf("%s: cannot currently handle more than one "
1612                                "S/G list per CCB\n", __func__);
1613                         error = EINVAL;
1614                         goto bailout;
1615                 }
1616
1617                 /*
1618                  * We have to have at least one segment.
1619                  */
1620                 if (num_segs == 0) {
1621                         xpt_print(periph->path, "%s: CAM_DATA_SG_PADDR flag "
1622                                   "set, but sglist_cnt=0!\n", __func__);
1623                         error = EINVAL;
1624                         goto bailout;
1625                 }
1626
1627                 /*
1628                  * Make sure the user specified the total length and didn't
1629                  * just leave it to us to decode the S/G list.
1630                  */
1631                 if (lengths[0] == 0) {
1632                         xpt_print(periph->path, "%s: no dxfer_len specified, "
1633                                   "but CAM_DATA_SG flag is set!\n", __func__);
1634                         error = EINVAL;
1635                         goto bailout;
1636                 }
1637
1638                 /* Figure out the size of the S/G list */
1639                 sg_length = num_segs * sizeof(bus_dma_segment_t);
1640                 io_req->num_user_segs = num_segs;
1641                 io_req->num_kern_segs = io_req->num_user_segs;
1642
1643                 /* Save the user's S/G list pointer for later restoration */
1644                 io_req->user_bufs[0] = *data_ptrs[0];
1645
1646                 if (num_segs > PASS_MAX_SEGS) {
1647                         io_req->user_segptr = malloc(sizeof(bus_dma_segment_t) *
1648                             num_segs, M_SCSIPASS, M_WAITOK | M_ZERO);
1649                         io_req->flags |= PASS_IO_USER_SEG_MALLOC;
1650                 } else
1651                         io_req->user_segptr = io_req->user_segs;
1652
1653                 io_req->kern_segptr = io_req->user_segptr;
1654
1655                 error = copyin(*data_ptrs[0], io_req->user_segptr, sg_length);
1656                 if (error != 0) {
1657                         xpt_print(periph->path, "%s: copy of user S/G list "
1658                                   "from %p to %p failed with error %d\n",
1659                                   __func__, *data_ptrs[0], io_req->user_segptr,
1660                                   error);
1661                         goto bailout;
1662                 }
1663                 break;
1664         }
1665         default:
1666         case CAM_DATA_BIO:
1667                 /*
1668                  * A user shouldn't be attaching a bio to the CCB.  It
1669                  * isn't a user-accessible structure.
1670                  */
1671                 error = EINVAL;
1672                 break;
1673         }
1674
1675 bailout:
1676         if (error != 0)
1677                 passiocleanup(softc, io_req);
1678
1679         return (error);
1680 }
1681
1682 static int
1683 passmemdone(struct cam_periph *periph, struct pass_io_req *io_req)
1684 {
1685         struct pass_softc *softc;
1686         union ccb *ccb;
1687         int error;
1688         int i;
1689
1690         error = 0;
1691         softc = (struct pass_softc *)periph->softc;
1692         ccb = &io_req->ccb;
1693
1694         switch (io_req->data_flags) {
1695         case CAM_DATA_VADDR:
1696                 /*
1697                  * Copy back to the user buffer if this was a read.
1698                  */
1699                 for (i = 0; i < io_req->num_bufs; i++) {
1700                         if (io_req->dirs[i] != CAM_DIR_IN)
1701                                 continue;
1702
1703                         error = copyout(io_req->kern_bufs[i],
1704                             io_req->user_bufs[i], io_req->lengths[i]);
1705                         if (error != 0) {
1706                                 xpt_print(periph->path, "Unable to copy %u "
1707                                           "bytes from %p to user address %p\n",
1708                                           io_req->lengths[i],
1709                                           io_req->kern_bufs[i],
1710                                           io_req->user_bufs[i]);
1711                                 goto bailout;
1712                         }
1713
1714                 }
1715                 break;
1716         case CAM_DATA_PADDR:
1717                 /* Do nothing.  The pointer is a physical address already */
1718                 break;
1719         case CAM_DATA_SG:
1720                 /*
1721                  * Copy back to the user buffer if this was a read.
1722                  * Restore the user's S/G list buffer pointer.
1723                  */
1724                 if (io_req->dirs[0] == CAM_DIR_IN)
1725                         error = passcopysglist(periph, io_req, io_req->dirs[0]);
1726                 break;
1727         case CAM_DATA_SG_PADDR:
1728                 /*
1729                  * Restore the user's S/G list buffer pointer.  No need to
1730                  * copy.
1731                  */
1732                 break;
1733         default:
1734         case CAM_DATA_BIO:
1735                 error = EINVAL;
1736                 break;
1737         }
1738
1739 bailout:
1740         /*
1741          * Reset the user's pointers to their original values and free
1742          * allocated memory.
1743          */
1744         passiocleanup(softc, io_req);
1745
1746         return (error);
1747 }
1748
1749 static int
1750 passioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flag, struct thread *td)
1751 {
1752         int error;
1753
1754         if ((error = passdoioctl(dev, cmd, addr, flag, td)) == ENOTTY) {
1755                 error = cam_compat_ioctl(dev, cmd, addr, flag, td, passdoioctl);
1756         }
1757         return (error);
1758 }
1759
1760 static int
1761 passdoioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flag, struct thread *td)
1762 {
1763         struct  cam_periph *periph;
1764         struct  pass_softc *softc;
1765         int     error;
1766         uint32_t priority;
1767
1768         periph = (struct cam_periph *)dev->si_drv1;
1769         cam_periph_lock(periph);
1770         softc = (struct pass_softc *)periph->softc;
1771
1772         error = 0;
1773
1774         switch (cmd) {
1775
1776         case CAMIOCOMMAND:
1777         {
1778                 union ccb *inccb;
1779                 union ccb *ccb;
1780                 int ccb_malloced;
1781
1782                 inccb = (union ccb *)addr;
1783
1784                 /*
1785                  * Some CCB types, like scan bus and scan lun can only go
1786                  * through the transport layer device.
1787                  */
1788                 if (inccb->ccb_h.func_code & XPT_FC_XPT_ONLY) {
1789                         xpt_print(periph->path, "CCB function code %#x is "
1790                             "restricted to the XPT device\n",
1791                             inccb->ccb_h.func_code);
1792                         error = ENODEV;
1793                         break;
1794                 }
1795
1796                 /* Compatibility for RL/priority-unaware code. */
1797                 priority = inccb->ccb_h.pinfo.priority;
1798                 if (priority <= CAM_PRIORITY_OOB)
1799                     priority += CAM_PRIORITY_OOB + 1;
1800
1801                 /*
1802                  * Non-immediate CCBs need a CCB from the per-device pool
1803                  * of CCBs, which is scheduled by the transport layer.
1804                  * Immediate CCBs and user-supplied CCBs should just be
1805                  * malloced.
1806                  */
1807                 if ((inccb->ccb_h.func_code & XPT_FC_QUEUED)
1808                  && ((inccb->ccb_h.func_code & XPT_FC_USER_CCB) == 0)) {
1809                         ccb = cam_periph_getccb(periph, priority);
1810                         ccb_malloced = 0;
1811                 } else {
1812                         ccb = xpt_alloc_ccb_nowait();
1813
1814                         if (ccb != NULL)
1815                                 xpt_setup_ccb(&ccb->ccb_h, periph->path,
1816                                               priority);
1817                         ccb_malloced = 1;
1818                 }
1819
1820                 if (ccb == NULL) {
1821                         xpt_print(periph->path, "unable to allocate CCB\n");
1822                         error = ENOMEM;
1823                         break;
1824                 }
1825
1826                 error = passsendccb(periph, ccb, inccb);
1827
1828                 if (ccb_malloced)
1829                         xpt_free_ccb(ccb);
1830                 else
1831                         xpt_release_ccb(ccb);
1832
1833                 break;
1834         }
1835         case CAMIOQUEUE:
1836         {
1837                 struct pass_io_req *io_req;
1838                 union ccb **user_ccb, *ccb;
1839                 xpt_opcode fc;
1840
1841                 if ((softc->flags & PASS_FLAG_ZONE_VALID) == 0) {
1842                         error = passcreatezone(periph);
1843                         if (error != 0)
1844                                 goto bailout;
1845                 }
1846
1847                 /*
1848                  * We're going to do a blocking allocation for this I/O
1849                  * request, so we have to drop the lock.
1850                  */
1851                 cam_periph_unlock(periph);
1852
1853                 io_req = uma_zalloc(softc->pass_zone, M_WAITOK | M_ZERO);
1854                 ccb = &io_req->ccb;
1855                 user_ccb = (union ccb **)addr;
1856
1857                 /*
1858                  * Unlike the CAMIOCOMMAND ioctl above, we only have a
1859                  * pointer to the user's CCB, so we have to copy the whole
1860                  * thing in to a buffer we have allocated (above) instead
1861                  * of allowing the ioctl code to malloc a buffer and copy
1862                  * it in.
1863                  *
1864                  * This is an advantage for this asynchronous interface,
1865                  * since we don't want the memory to get freed while the
1866                  * CCB is outstanding.
1867                  */
1868 #if 0
1869                 xpt_print(periph->path, "Copying user CCB %p to "
1870                           "kernel address %p\n", *user_ccb, ccb);
1871 #endif
1872                 error = copyin(*user_ccb, ccb, sizeof(*ccb));
1873                 if (error != 0) {
1874                         xpt_print(periph->path, "Copy of user CCB %p to "
1875                                   "kernel address %p failed with error %d\n",
1876                                   *user_ccb, ccb, error);
1877                         uma_zfree(softc->pass_zone, io_req);
1878                         cam_periph_lock(periph);
1879                         break;
1880                 }
1881
1882                 if (ccb->ccb_h.flags & CAM_CDB_POINTER) {
1883                         if (ccb->csio.cdb_len > IOCDBLEN) {
1884                                 error = EINVAL;
1885                                 break;
1886                         }
1887                         error = copyin(ccb->csio.cdb_io.cdb_ptr,
1888                             ccb->csio.cdb_io.cdb_bytes, ccb->csio.cdb_len);
1889                         if (error)
1890                                 break;
1891                         ccb->ccb_h.flags &= ~CAM_CDB_POINTER;
1892                 }
1893
1894                 /*
1895                  * Some CCB types, like scan bus and scan lun can only go
1896                  * through the transport layer device.
1897                  */
1898                 if (ccb->ccb_h.func_code & XPT_FC_XPT_ONLY) {
1899                         xpt_print(periph->path, "CCB function code %#x is "
1900                             "restricted to the XPT device\n",
1901                             ccb->ccb_h.func_code);
1902                         uma_zfree(softc->pass_zone, io_req);
1903                         cam_periph_lock(periph);
1904                         error = ENODEV;
1905                         break;
1906                 }
1907
1908                 /*
1909                  * Save the user's CCB pointer as well as his linked list
1910                  * pointers and peripheral private area so that we can
1911                  * restore these later.
1912                  */
1913                 io_req->user_ccb_ptr = *user_ccb;
1914                 io_req->user_periph_links = ccb->ccb_h.periph_links;
1915                 io_req->user_periph_priv = ccb->ccb_h.periph_priv;
1916
1917                 /*
1918                  * Now that we've saved the user's values, we can set our
1919                  * own peripheral private entry.
1920                  */
1921                 ccb->ccb_h.ccb_ioreq = io_req;
1922
1923                 /* Compatibility for RL/priority-unaware code. */
1924                 priority = ccb->ccb_h.pinfo.priority;
1925                 if (priority <= CAM_PRIORITY_OOB)
1926                     priority += CAM_PRIORITY_OOB + 1;
1927
1928                 /*
1929                  * Setup fields in the CCB like the path and the priority.
1930                  * The path in particular cannot be done in userland, since
1931                  * it is a pointer to a kernel data structure.
1932                  */
1933                 xpt_setup_ccb_flags(&ccb->ccb_h, periph->path, priority,
1934                                     ccb->ccb_h.flags);
1935
1936                 /*
1937                  * Setup our done routine.  There is no way for the user to
1938                  * have a valid pointer here.
1939                  */
1940                 ccb->ccb_h.cbfcnp = passdone;
1941
1942                 fc = ccb->ccb_h.func_code;
1943                 /*
1944                  * If this function code has memory that can be mapped in
1945                  * or out, we need to call passmemsetup().
1946                  */
1947                 if ((fc == XPT_SCSI_IO) || (fc == XPT_ATA_IO)
1948                  || (fc == XPT_SMP_IO) || (fc == XPT_DEV_MATCH)
1949                  || (fc == XPT_DEV_ADVINFO)) {
1950                         error = passmemsetup(periph, io_req);
1951                         if (error != 0) {
1952                                 uma_zfree(softc->pass_zone, io_req);
1953                                 cam_periph_lock(periph);
1954                                 break;
1955                         }
1956                 } else
1957                         io_req->mapinfo.num_bufs_used = 0;
1958
1959                 cam_periph_lock(periph);
1960
1961                 /*
1962                  * Everything goes on the incoming queue initially.
1963                  */
1964                 TAILQ_INSERT_TAIL(&softc->incoming_queue, io_req, links);
1965
1966                 /*
1967                  * If the CCB is queued, and is not a user CCB, then
1968                  * we need to allocate a slot for it.  Call xpt_schedule()
1969                  * so that our start routine will get called when a CCB is
1970                  * available.
1971                  */
1972                 if ((fc & XPT_FC_QUEUED)
1973                  && ((fc & XPT_FC_USER_CCB) == 0)) {
1974                         xpt_schedule(periph, priority);
1975                         break;
1976                 }
1977
1978                 /*
1979                  * At this point, the CCB in question is either an
1980                  * immediate CCB (like XPT_DEV_ADVINFO) or it is a user CCB
1981                  * and therefore should be malloced, not allocated via a slot.
1982                  * Remove the CCB from the incoming queue and add it to the
1983                  * active queue.
1984                  */
1985                 TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
1986                 TAILQ_INSERT_TAIL(&softc->active_queue, io_req, links);
1987
1988                 xpt_action(ccb);
1989
1990                 /*
1991                  * If this is not a queued CCB (i.e. it is an immediate CCB),
1992                  * then it is already done.  We need to put it on the done
1993                  * queue for the user to fetch.
1994                  */
1995                 if ((fc & XPT_FC_QUEUED) == 0) {
1996                         TAILQ_REMOVE(&softc->active_queue, io_req, links);
1997                         TAILQ_INSERT_TAIL(&softc->done_queue, io_req, links);
1998                 }
1999                 break;
2000         }
2001         case CAMIOGET:
2002         {
2003                 union ccb **user_ccb;
2004                 struct pass_io_req *io_req;
2005                 int old_error;
2006
2007                 user_ccb = (union ccb **)addr;
2008                 old_error = 0;
2009
2010                 io_req = TAILQ_FIRST(&softc->done_queue);
2011                 if (io_req == NULL) {
2012                         error = ENOENT;
2013                         break;
2014                 }
2015
2016                 /*
2017                  * Remove the I/O from the done queue.
2018                  */
2019                 TAILQ_REMOVE(&softc->done_queue, io_req, links);
2020
2021                 /*
2022                  * We have to drop the lock during the copyout because the
2023                  * copyout can result in VM faults that require sleeping.
2024                  */
2025                 cam_periph_unlock(periph);
2026
2027                 /*
2028                  * Do any needed copies (e.g. for reads) and revert the
2029                  * pointers in the CCB back to the user's pointers.
2030                  */
2031                 error = passmemdone(periph, io_req);
2032
2033                 old_error = error;
2034
2035                 io_req->ccb.ccb_h.periph_links = io_req->user_periph_links;
2036                 io_req->ccb.ccb_h.periph_priv = io_req->user_periph_priv;
2037
2038 #if 0
2039                 xpt_print(periph->path, "Copying to user CCB %p from "
2040                           "kernel address %p\n", *user_ccb, &io_req->ccb);
2041 #endif
2042
2043                 error = copyout(&io_req->ccb, *user_ccb, sizeof(union ccb));
2044                 if (error != 0) {
2045                         xpt_print(periph->path, "Copy to user CCB %p from "
2046                                   "kernel address %p failed with error %d\n",
2047                                   *user_ccb, &io_req->ccb, error);
2048                 }
2049
2050                 /*
2051                  * Prefer the first error we got back, and make sure we
2052                  * don't overwrite bad status with good.
2053                  */
2054                 if (old_error != 0)
2055                         error = old_error;
2056
2057                 cam_periph_lock(periph);
2058
2059                 /*
2060                  * At this point, if there was an error, we could potentially
2061                  * re-queue the I/O and try again.  But why?  The error
2062                  * would almost certainly happen again.  We might as well
2063                  * not leak memory.
2064                  */
2065                 uma_zfree(softc->pass_zone, io_req);
2066                 break;
2067         }
2068         default:
2069                 error = cam_periph_ioctl(periph, cmd, addr, passerror);
2070                 break;
2071         }
2072
2073 bailout:
2074         cam_periph_unlock(periph);
2075
2076         return(error);
2077 }
2078
2079 static int
2080 passpoll(struct cdev *dev, int poll_events, struct thread *td)
2081 {
2082         struct cam_periph *periph;
2083         struct pass_softc *softc;
2084         int revents;
2085
2086         periph = (struct cam_periph *)dev->si_drv1;
2087         softc = (struct pass_softc *)periph->softc;
2088
2089         revents = poll_events & (POLLOUT | POLLWRNORM);
2090         if ((poll_events & (POLLIN | POLLRDNORM)) != 0) {
2091                 cam_periph_lock(periph);
2092
2093                 if (!TAILQ_EMPTY(&softc->done_queue)) {
2094                         revents |= poll_events & (POLLIN | POLLRDNORM);
2095                 }
2096                 cam_periph_unlock(periph);
2097                 if (revents == 0)
2098                         selrecord(td, &softc->read_select);
2099         }
2100
2101         return (revents);
2102 }
2103
2104 static int
2105 passkqfilter(struct cdev *dev, struct knote *kn)
2106 {
2107         struct cam_periph *periph;
2108         struct pass_softc *softc;
2109
2110         periph = (struct cam_periph *)dev->si_drv1;
2111         softc = (struct pass_softc *)periph->softc;
2112
2113         kn->kn_hook = (caddr_t)periph;
2114         kn->kn_fop = &passread_filtops;
2115         knlist_add(&softc->read_select.si_note, kn, 0);
2116
2117         return (0);
2118 }
2119
2120 static void
2121 passreadfiltdetach(struct knote *kn)
2122 {
2123         struct cam_periph *periph;
2124         struct pass_softc *softc;
2125
2126         periph = (struct cam_periph *)kn->kn_hook;
2127         softc = (struct pass_softc *)periph->softc;
2128
2129         knlist_remove(&softc->read_select.si_note, kn, 0);
2130 }
2131
2132 static int
2133 passreadfilt(struct knote *kn, long hint)
2134 {
2135         struct cam_periph *periph;
2136         struct pass_softc *softc;
2137         int retval;
2138
2139         periph = (struct cam_periph *)kn->kn_hook;
2140         softc = (struct pass_softc *)periph->softc;
2141
2142         cam_periph_assert(periph, MA_OWNED);
2143
2144         if (TAILQ_EMPTY(&softc->done_queue))
2145                 retval = 0;
2146         else
2147                 retval = 1;
2148
2149         return (retval);
2150 }
2151
2152 /*
2153  * Generally, "ccb" should be the CCB supplied by the kernel.  "inccb"
2154  * should be the CCB that is copied in from the user.
2155  */
2156 static int
2157 passsendccb(struct cam_periph *periph, union ccb *ccb, union ccb *inccb)
2158 {
2159         struct pass_softc *softc;
2160         struct cam_periph_map_info mapinfo;
2161         uint8_t *cmd;
2162         xpt_opcode fc;
2163         int error;
2164
2165         softc = (struct pass_softc *)periph->softc;
2166
2167         /*
2168          * There are some fields in the CCB header that need to be
2169          * preserved, the rest we get from the user.
2170          */
2171         xpt_merge_ccb(ccb, inccb);
2172
2173         if (ccb->ccb_h.flags & CAM_CDB_POINTER) {
2174                 cmd = __builtin_alloca(ccb->csio.cdb_len);
2175                 error = copyin(ccb->csio.cdb_io.cdb_ptr, cmd, ccb->csio.cdb_len);
2176                 if (error)
2177                         return (error);
2178                 ccb->csio.cdb_io.cdb_ptr = cmd;
2179         }
2180
2181         /*
2182          */
2183         ccb->ccb_h.cbfcnp = passdone;
2184
2185         /*
2186          * Let cam_periph_mapmem do a sanity check on the data pointer format.
2187          * Even if no data transfer is needed, it's a cheap check and it
2188          * simplifies the code.
2189          */
2190         fc = ccb->ccb_h.func_code;
2191         if ((fc == XPT_SCSI_IO) || (fc == XPT_ATA_IO) || (fc == XPT_SMP_IO)
2192          || (fc == XPT_DEV_MATCH) || (fc == XPT_DEV_ADVINFO)) {
2193                 bzero(&mapinfo, sizeof(mapinfo));
2194
2195                 /*
2196                  * cam_periph_mapmem calls into proc and vm functions that can
2197                  * sleep as well as trigger I/O, so we can't hold the lock.
2198                  * Dropping it here is reasonably safe.
2199                  */
2200                 cam_periph_unlock(periph);
2201                 error = cam_periph_mapmem(ccb, &mapinfo, softc->maxio);
2202                 cam_periph_lock(periph);
2203
2204                 /*
2205                  * cam_periph_mapmem returned an error, we can't continue.
2206                  * Return the error to the user.
2207                  */
2208                 if (error)
2209                         return(error);
2210         } else
2211                 /* Ensure that the unmap call later on is a no-op. */
2212                 mapinfo.num_bufs_used = 0;
2213
2214         /*
2215          * If the user wants us to perform any error recovery, then honor
2216          * that request.  Otherwise, it's up to the user to perform any
2217          * error recovery.
2218          */
2219         cam_periph_runccb(ccb, (ccb->ccb_h.flags & CAM_PASS_ERR_RECOVER) ?
2220             passerror : NULL, /* cam_flags */ CAM_RETRY_SELTO,
2221             /* sense_flags */ SF_RETRY_UA | SF_NO_PRINT,
2222             softc->device_stats);
2223
2224         cam_periph_unmapmem(ccb, &mapinfo);
2225
2226         ccb->ccb_h.cbfcnp = NULL;
2227         ccb->ccb_h.periph_priv = inccb->ccb_h.periph_priv;
2228         bcopy(ccb, inccb, sizeof(union ccb));
2229
2230         return(0);
2231 }
2232
2233 static int
2234 passerror(union ccb *ccb, u_int32_t cam_flags, u_int32_t sense_flags)
2235 {
2236         struct cam_periph *periph;
2237         struct pass_softc *softc;
2238
2239         periph = xpt_path_periph(ccb->ccb_h.path);
2240         softc = (struct pass_softc *)periph->softc;
2241
2242         return(cam_periph_error(ccb, cam_flags, sense_flags,
2243                                  &softc->saved_ccb));
2244 }