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
 775                 TAILQ_FOREACH_SAFE(io_req, &softc->done_queue, links, io_req2) {
 776                         TAILQ_REMOVE(&softc->done_queue, io_req, links);
 777                         passiocleanup(softc, io_req);
 778                         uma_zfree(softc->pass_zone, io_req);
 779                 }
 780
 781                 TAILQ_FOREACH_SAFE(io_req, &softc->incoming_queue, links,
 782                                    io_req2) {
 783                         TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
 784                         passiocleanup(softc, io_req);
 785                         uma_zfree(softc->pass_zone, io_req);
 786                 }
 787
 788                 /*
 789                  * If there are any active I/Os, we need to forcibly acquire a
 790                  * reference to the peripheral so that we don't go away
 791                  * before they complete.  We'll release the reference when
 792                  * the abandoned queue is empty.
 793                  */
 794                 io_req = TAILQ_FIRST(&softc->active_queue);
 795                 if ((io_req != NULL)
 796                  && (softc->flags & PASS_FLAG_ABANDONED_REF_SET) == 0) {
 797                         cam_periph_doacquire(periph);
 798                         softc->flags |= PASS_FLAG_ABANDONED_REF_SET;
 799                 }
 800
 801                 /*
 802                  * Since the I/O in the active queue is not under our
 803                  * control, just set a flag so that we can clean it up when
 804                  * it completes and put it on the abandoned queue.  This
 805                  * will prevent our sending spurious completions in the
 806                  * event that the device is opened again before these I/Os
 807                  * complete.
 808                  */
 809                 TAILQ_FOREACH_SAFE(io_req, &softc->active_queue, links,
 810                                    io_req2) {
 811                         TAILQ_REMOVE(&softc->active_queue, io_req, links);
 812                         io_req->flags |= PASS_IO_ABANDONED;
 813                         TAILQ_INSERT_TAIL(&softc->abandoned_queue, io_req,
 814                                           links);
 815                 }
 816         }
 817
 818         cam_periph_release_locked(periph);
 819
 820         /*
 821          * We reference the lock directly here, instead of using
 822          * cam_periph_unlock().  The reason is that the call to
 823          * cam_periph_release_locked() above could result in the periph
 824          * getting freed.  If that is the case, dereferencing the periph
 825          * with a cam_periph_unlock() call would cause a page fault.
 826          *
 827          * cam_periph_release() avoids this problem using the same method,
 828          * but we're manually acquiring and dropping the lock here to
 829          * protect the open count and avoid another lock acquisition and
 830          * release.
 831          */
 832         mtx_unlock(mtx);
 833
 834         return (0);
 835 }
 836
 837
 838 static void
 839 passstart(struct cam_periph *periph, union ccb *start_ccb)
 840 {
 841         struct pass_softc *softc;
 842
 843         softc = (struct pass_softc *)periph->softc;
 844
 845         switch (softc->state) {
 846         case PASS_STATE_NORMAL: {
 847                 struct pass_io_req *io_req;
 848
 849                 /*
 850                  * Check for any queued I/O requests that require an
 851                  * allocated slot.
 852                  */
 853                 io_req = TAILQ_FIRST(&softc->incoming_queue);
 854                 if (io_req == NULL) {
 855                         xpt_release_ccb(start_ccb);
 856                         break;
 857                 }
 858                 TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
 859                 TAILQ_INSERT_TAIL(&softc->active_queue, io_req, links);
 860                 /*
 861                  * Merge the user's CCB into the allocated CCB.
 862                  */
 863                 xpt_merge_ccb(start_ccb, &io_req->ccb);
 864                 start_ccb->ccb_h.ccb_type = PASS_CCB_QUEUED_IO;
 865                 start_ccb->ccb_h.ccb_ioreq = io_req;
 866                 start_ccb->ccb_h.cbfcnp = passdone;
 867                 io_req->alloced_ccb = start_ccb;
 868                 binuptime(&io_req->start_time);
 869                 devstat_start_transaction(softc->device_stats,
 870                                           &io_req->start_time);
 871
 872                 xpt_action(start_ccb);
 873
 874                 /*
 875                  * If we have any more I/O waiting, schedule ourselves again.
 876                  */
 877                 if (!TAILQ_EMPTY(&softc->incoming_queue))
 878                         xpt_schedule(periph, CAM_PRIORITY_NORMAL);
 879                 break;
 880         }
 881         default:
 882                 break;
 883         }
 884 }
 885
 886 static void
 887 passdone(struct cam_periph *periph, union ccb *done_ccb)
 888 {
 889         struct pass_softc *softc;
 890         struct ccb_scsiio *csio;
 891
 892         softc = (struct pass_softc *)periph->softc;
 893
 894         cam_periph_assert(periph, MA_OWNED);
 895
 896         csio = &done_ccb->csio;
 897         switch (csio->ccb_h.ccb_type) {
 898         case PASS_CCB_QUEUED_IO: {
 899                 struct pass_io_req *io_req;
 900
 901                 io_req = done_ccb->ccb_h.ccb_ioreq;
 902 #if 0
 903                 xpt_print(periph->path, "%s: called for user CCB %p\n",
 904                           __func__, io_req->user_ccb_ptr);
 905 #endif
 906                 if (((done_ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_CMP)
 907                  && (done_ccb->ccb_h.flags & CAM_PASS_ERR_RECOVER)
 908                  && ((io_req->flags & PASS_IO_ABANDONED) == 0)) {
 909                         int error;
 910
 911                         error = passerror(done_ccb, CAM_RETRY_SELTO,
 912                                           SF_RETRY_UA | SF_NO_PRINT);
 913
 914                         if (error == ERESTART) {
 915                                 /*
 916                                  * A retry was scheduled, so
 917                                  * just return.
 918                                  */
 919                                 return;
 920                         }
 921                 }
 922
 923                 /*
 924                  * Copy the allocated CCB contents back to the malloced CCB
 925                  * so we can give status back to the user when he requests it.
 926                  */
 927                 bcopy(done_ccb, &io_req->ccb, sizeof(*done_ccb));
 928
 929                 /*
 930                  * Log data/transaction completion with devstat(9).
 931                  */
 932                 switch (done_ccb->ccb_h.func_code) {
 933                 case XPT_SCSI_IO:
 934                         devstat_end_transaction(softc->device_stats,
 935                             done_ccb->csio.dxfer_len - done_ccb->csio.resid,
 936                             done_ccb->csio.tag_action & 0x3,
 937                             ((done_ccb->ccb_h.flags & CAM_DIR_MASK) ==
 938                             CAM_DIR_NONE) ? DEVSTAT_NO_DATA :
 939                             (done_ccb->ccb_h.flags & CAM_DIR_OUT) ?
 940                             DEVSTAT_WRITE : DEVSTAT_READ, NULL,
 941                             &io_req->start_time);
 942                         break;
 943                 case XPT_ATA_IO:
 944                         devstat_end_transaction(softc->device_stats,
 945                             done_ccb->ataio.dxfer_len - done_ccb->ataio.resid,
 946                             0, /* Not used in ATA */
 947                             ((done_ccb->ccb_h.flags & CAM_DIR_MASK) ==
 948                             CAM_DIR_NONE) ? DEVSTAT_NO_DATA :
 949                             (done_ccb->ccb_h.flags & CAM_DIR_OUT) ?
 950                             DEVSTAT_WRITE : DEVSTAT_READ, NULL,
 951                             &io_req->start_time);
 952                         break;
 953                 case XPT_SMP_IO:
 954                         /*
 955                          * XXX KDM this isn't quite right, but there isn't
 956                          * currently an easy way to represent a bidirectional
 957                          * transfer in devstat.  The only way to do it
 958                          * and have the byte counts come out right would
 959                          * mean that we would have to record two
 960                          * transactions, one for the request and one for the
 961                          * response.  For now, so that we report something,
 962                          * just treat the entire thing as a read.
 963                          */
 964                         devstat_end_transaction(softc->device_stats,
 965                             done_ccb->smpio.smp_request_len +
 966                             done_ccb->smpio.smp_response_len,
 967                             DEVSTAT_TAG_SIMPLE, DEVSTAT_READ, NULL,
 968                             &io_req->start_time);
 969                         break;
 970                 default:
 971                         devstat_end_transaction(softc->device_stats, 0,
 972                             DEVSTAT_TAG_NONE, DEVSTAT_NO_DATA, NULL,
 973                             &io_req->start_time);
 974                         break;
 975                 }
 976
 977                 /*
 978                  * In the normal case, take the completed I/O off of the
 979                  * active queue and put it on the done queue.  Notitfy the
 980                  * user that we have a completed I/O.
 981                  */
 982                 if ((io_req->flags & PASS_IO_ABANDONED) == 0) {
 983                         TAILQ_REMOVE(&softc->active_queue, io_req, links);
 984                         TAILQ_INSERT_TAIL(&softc->done_queue, io_req, links);
 985                         selwakeuppri(&softc->read_select, PRIBIO);
 986                         KNOTE_LOCKED(&softc->read_select.si_note, 0);
 987                 } else {
 988                         /*
 989                          * In the case of an abandoned I/O (final close
 990                          * without fetching the I/O), take it off of the
 991                          * abandoned queue and free it.
 992                          */
 993                         TAILQ_REMOVE(&softc->abandoned_queue, io_req, links);
 994                         passiocleanup(softc, io_req);
 995                         uma_zfree(softc->pass_zone, io_req);
 996
 997                         /*
 998                          * Release the done_ccb here, since we may wind up
 999                          * freeing the peripheral when we decrement the
1000                          * reference count below.
1001                          */
1002                         xpt_release_ccb(done_ccb);
1003
1004                         /*
1005                          * If the abandoned queue is empty, we can release
1006                          * our reference to the periph since we won't have
1007                          * any more completions coming.
1008                          */
1009                         if ((TAILQ_EMPTY(&softc->abandoned_queue))
1010                          && (softc->flags & PASS_FLAG_ABANDONED_REF_SET)) {
1011                                 softc->flags &= ~PASS_FLAG_ABANDONED_REF_SET;
1012                                 cam_periph_release_locked(periph);
1013                         }
1014
1015                         /*
1016                          * We have already released the CCB, so we can
1017                          * return.
1018                          */
1019                         return;
1020                 }
1021                 break;
1022         }
1023         }
1024         xpt_release_ccb(done_ccb);
1025 }
1026
1027 static int
1028 passcreatezone(struct cam_periph *periph)
1029 {
1030         struct pass_softc *softc;
1031         int error;
1032
1033         error = 0;
1034         softc = (struct pass_softc *)periph->softc;
1035
1036         cam_periph_assert(periph, MA_OWNED);
1037         KASSERT(((softc->flags & PASS_FLAG_ZONE_VALID) == 0),
1038                 ("%s called when the pass(4) zone is valid!\n", __func__));
1039         KASSERT((softc->pass_zone == NULL),
1040                 ("%s called when the pass(4) zone is allocated!\n", __func__));
1041
1042         if ((softc->flags & PASS_FLAG_ZONE_INPROG) == 0) {
1043
1044                 /*
1045                  * We're the first context through, so we need to create
1046                  * the pass(4) UMA zone for I/O requests.
1047                  */
1048                 softc->flags |= PASS_FLAG_ZONE_INPROG;
1049
1050                 /*
1051                  * uma_zcreate() does a blocking (M_WAITOK) allocation,
1052                  * so we cannot hold a mutex while we call it.
1053                  */
1054                 cam_periph_unlock(periph);
1055
1056                 softc->pass_zone = uma_zcreate(softc->zone_name,
1057                     sizeof(struct pass_io_req), NULL, NULL, NULL, NULL,
1058                     /*align*/ 0, /*flags*/ 0);
1059
1060                 softc->pass_io_zone = uma_zcreate(softc->io_zone_name,
1061                     softc->io_zone_size, NULL, NULL, NULL, NULL,
1062                     /*align*/ 0, /*flags*/ 0);
1063
1064                 cam_periph_lock(periph);
1065
1066                 if ((softc->pass_zone == NULL)
1067                  || (softc->pass_io_zone == NULL)) {
1068                         if (softc->pass_zone == NULL)
1069                                 xpt_print(periph->path, "unable to allocate "
1070                                     "IO Req UMA zone\n");
1071                         else
1072                                 xpt_print(periph->path, "unable to allocate "
1073                                     "IO UMA zone\n");
1074                         softc->flags &= ~PASS_FLAG_ZONE_INPROG;
1075                         goto bailout;
1076                 }
1077
1078                 /*
1079                  * Set the flags appropriately and notify any other waiters.
1080                  */
1081                 softc->flags &= PASS_FLAG_ZONE_INPROG;
1082                 softc->flags |= PASS_FLAG_ZONE_VALID;
1083                 wakeup(&softc->pass_zone);
1084         } else {
1085                 /*
1086                  * In this case, the UMA zone has not yet been created, but
1087                  * another context is in the process of creating it.  We
1088                  * need to sleep until the creation is either done or has
1089                  * failed.
1090                  */
1091                 while ((softc->flags & PASS_FLAG_ZONE_INPROG)
1092                     && ((softc->flags & PASS_FLAG_ZONE_VALID) == 0)) {
1093                         error = msleep(&softc->pass_zone,
1094                                        cam_periph_mtx(periph), PRIBIO,
1095                                        "paszon", 0);
1096                         if (error != 0)
1097                                 goto bailout;
1098                 }
1099                 /*
1100                  * If the zone creation failed, no luck for the user.
1101                  */
1102                 if ((softc->flags & PASS_FLAG_ZONE_VALID) == 0){
1103                         error = ENOMEM;
1104                         goto bailout;
1105                 }
1106         }
1107 bailout:
1108         return (error);
1109 }
1110
1111 static void
1112 passiocleanup(struct pass_softc *softc, struct pass_io_req *io_req)
1113 {
1114         union ccb *ccb;
1115         u_int8_t **data_ptrs[CAM_PERIPH_MAXMAPS];
1116         int i, numbufs;
1117
1118         ccb = &io_req->ccb;
1119
1120         switch (ccb->ccb_h.func_code) {
1121         case XPT_DEV_MATCH:
1122                 numbufs = min(io_req->num_bufs, 2);
1123
1124                 if (numbufs == 1) {
1125                         data_ptrs[0] = (u_int8_t **)&ccb->cdm.matches;
1126                 } else {
1127                         data_ptrs[0] = (u_int8_t **)&ccb->cdm.patterns;
1128                         data_ptrs[1] = (u_int8_t **)&ccb->cdm.matches;
1129                 }
1130                 break;
1131         case XPT_SCSI_IO:
1132         case XPT_CONT_TARGET_IO:
1133                 data_ptrs[0] = &ccb->csio.data_ptr;
1134                 numbufs = min(io_req->num_bufs, 1);
1135                 break;
1136         case XPT_ATA_IO:
1137                 data_ptrs[0] = &ccb->ataio.data_ptr;
1138                 numbufs = min(io_req->num_bufs, 1);
1139                 break;
1140         case XPT_SMP_IO:
1141                 numbufs = min(io_req->num_bufs, 2);
1142                 data_ptrs[0] = &ccb->smpio.smp_request;
1143                 data_ptrs[1] = &ccb->smpio.smp_response;
1144                 break;
1145         case XPT_DEV_ADVINFO:
1146                 numbufs = min(io_req->num_bufs, 1);
1147                 data_ptrs[0] = (uint8_t **)&ccb->cdai.buf;
1148                 break;
1149         case XPT_NVME_IO:
1150         case XPT_NVME_ADMIN:
1151                 data_ptrs[0] = &ccb->nvmeio.data_ptr;
1152                 numbufs = min(io_req->num_bufs, 1);
1153                 break;
1154         default:
1155                 /* allow ourselves to be swapped once again */
1156                 return;
1157                 break; /* NOTREACHED */
1158         }
1159
1160         if (io_req->flags & PASS_IO_USER_SEG_MALLOC) {
1161                 free(io_req->user_segptr, M_SCSIPASS);
1162                 io_req->user_segptr = NULL;
1163         }
1164
1165         /*
1166          * We only want to free memory we malloced.
1167          */
1168         if (io_req->data_flags == CAM_DATA_VADDR) {
1169                 for (i = 0; i < io_req->num_bufs; i++) {
1170                         if (io_req->kern_bufs[i] == NULL)
1171                                 continue;
1172
1173                         free(io_req->kern_bufs[i], M_SCSIPASS);
1174                         io_req->kern_bufs[i] = NULL;
1175                 }
1176         } else if (io_req->data_flags == CAM_DATA_SG) {
1177                 for (i = 0; i < io_req->num_kern_segs; i++) {
1178                         if ((uint8_t *)(uintptr_t)
1179                             io_req->kern_segptr[i].ds_addr == NULL)
1180                                 continue;
1181
1182                         uma_zfree(softc->pass_io_zone, (uint8_t *)(uintptr_t)
1183                             io_req->kern_segptr[i].ds_addr);
1184                         io_req->kern_segptr[i].ds_addr = 0;
1185                 }
1186         }
1187
1188         if (io_req->flags & PASS_IO_KERN_SEG_MALLOC) {
1189                 free(io_req->kern_segptr, M_SCSIPASS);
1190                 io_req->kern_segptr = NULL;
1191         }
1192
1193         if (io_req->data_flags != CAM_DATA_PADDR) {
1194                 for (i = 0; i < numbufs; i++) {
1195                         /*
1196                          * Restore the user's buffer pointers to their
1197                          * previous values.
1198                          */
1199                         if (io_req->user_bufs[i] != NULL)
1200                                 *data_ptrs[i] = io_req->user_bufs[i];
1201                 }
1202         }
1203
1204 }
1205
1206 static int
1207 passcopysglist(struct cam_periph *periph, struct pass_io_req *io_req,
1208                ccb_flags direction)
1209 {
1210         bus_size_t kern_watermark, user_watermark, len_copied, len_to_copy;
1211         bus_dma_segment_t *user_sglist, *kern_sglist;
1212         int i, j, error;
1213
1214         error = 0;
1215         kern_watermark = 0;
1216         user_watermark = 0;
1217         len_to_copy = 0;
1218         len_copied = 0;
1219         user_sglist = io_req->user_segptr;
1220         kern_sglist = io_req->kern_segptr;
1221
1222         for (i = 0, j = 0; i < io_req->num_user_segs &&
1223              j < io_req->num_kern_segs;) {
1224                 uint8_t *user_ptr, *kern_ptr;
1225
1226                 len_to_copy = min(user_sglist[i].ds_len -user_watermark,
1227                     kern_sglist[j].ds_len - kern_watermark);
1228
1229                 user_ptr = (uint8_t *)(uintptr_t)user_sglist[i].ds_addr;
1230                 user_ptr = user_ptr + user_watermark;
1231                 kern_ptr = (uint8_t *)(uintptr_t)kern_sglist[j].ds_addr;
1232                 kern_ptr = kern_ptr + kern_watermark;
1233
1234                 user_watermark += len_to_copy;
1235                 kern_watermark += len_to_copy;
1236
1237                 if (!useracc(user_ptr, len_to_copy,
1238                     (direction == CAM_DIR_IN) ? VM_PROT_WRITE : VM_PROT_READ)) {
1239                         xpt_print(periph->path, "%s: unable to access user "
1240                                   "S/G list element %p len %zu\n", __func__,
1241                                   user_ptr, len_to_copy);
1242                         error = EFAULT;
1243                         goto bailout;
1244                 }
1245
1246                 if (direction == CAM_DIR_IN) {
1247                         error = copyout(kern_ptr, user_ptr, len_to_copy);
1248                         if (error != 0) {
1249                                 xpt_print(periph->path, "%s: copyout of %u "
1250                                           "bytes from %p to %p failed with "
1251                                           "error %d\n", __func__, len_to_copy,
1252                                           kern_ptr, user_ptr, error);
1253                                 goto bailout;
1254                         }
1255                 } else {
1256                         error = copyin(user_ptr, kern_ptr, len_to_copy);
1257                         if (error != 0) {
1258                                 xpt_print(periph->path, "%s: copyin of %u "
1259                                           "bytes from %p to %p failed with "
1260                                           "error %d\n", __func__, len_to_copy,
1261                                           user_ptr, kern_ptr, error);
1262                                 goto bailout;
1263                         }
1264                 }
1265
1266                 len_copied += len_to_copy;
1267
1268                 if (user_sglist[i].ds_len == user_watermark) {
1269                         i++;
1270                         user_watermark = 0;
1271                 }
1272
1273                 if (kern_sglist[j].ds_len == kern_watermark) {
1274                         j++;
1275                         kern_watermark = 0;
1276                 }
1277         }
1278
1279 bailout:
1280
1281         return (error);
1282 }
1283
1284 static int
1285 passmemsetup(struct cam_periph *periph, struct pass_io_req *io_req)
1286 {
1287         union ccb *ccb;
1288         struct pass_softc *softc;
1289         int numbufs, i;
1290         uint8_t **data_ptrs[CAM_PERIPH_MAXMAPS];
1291         uint32_t lengths[CAM_PERIPH_MAXMAPS];
1292         uint32_t dirs[CAM_PERIPH_MAXMAPS];
1293         uint32_t num_segs;
1294         uint16_t *seg_cnt_ptr;
1295         size_t maxmap;
1296         int error;
1297
1298         cam_periph_assert(periph, MA_NOTOWNED);
1299
1300         softc = periph->softc;
1301
1302         error = 0;
1303         ccb = &io_req->ccb;
1304         maxmap = 0;
1305         num_segs = 0;
1306         seg_cnt_ptr = NULL;
1307
1308         switch(ccb->ccb_h.func_code) {
1309         case XPT_DEV_MATCH:
1310                 if (ccb->cdm.match_buf_len == 0) {
1311                         printf("%s: invalid match buffer length 0\n", __func__);
1312                         return(EINVAL);
1313                 }
1314                 if (ccb->cdm.pattern_buf_len > 0) {
1315                         data_ptrs[0] = (u_int8_t **)&ccb->cdm.patterns;
1316                         lengths[0] = ccb->cdm.pattern_buf_len;
1317                         dirs[0] = CAM_DIR_OUT;
1318                         data_ptrs[1] = (u_int8_t **)&ccb->cdm.matches;
1319                         lengths[1] = ccb->cdm.match_buf_len;
1320                         dirs[1] = CAM_DIR_IN;
1321                         numbufs = 2;
1322                 } else {
1323                         data_ptrs[0] = (u_int8_t **)&ccb->cdm.matches;
1324                         lengths[0] = ccb->cdm.match_buf_len;
1325                         dirs[0] = CAM_DIR_IN;
1326                         numbufs = 1;
1327                 }
1328                 io_req->data_flags = CAM_DATA_VADDR;
1329                 break;
1330         case XPT_SCSI_IO:
1331         case XPT_CONT_TARGET_IO:
1332                 if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
1333                         return(0);
1334
1335                 /*
1336                  * The user shouldn't be able to supply a bio.
1337                  */
1338                 if ((ccb->ccb_h.flags & CAM_DATA_MASK) == CAM_DATA_BIO)
1339                         return (EINVAL);
1340
1341                 io_req->data_flags = ccb->ccb_h.flags & CAM_DATA_MASK;
1342
1343                 data_ptrs[0] = &ccb->csio.data_ptr;
1344                 lengths[0] = ccb->csio.dxfer_len;
1345                 dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1346                 num_segs = ccb->csio.sglist_cnt;
1347                 seg_cnt_ptr = &ccb->csio.sglist_cnt;
1348                 numbufs = 1;
1349                 maxmap = softc->maxio;
1350                 break;
1351         case XPT_ATA_IO:
1352                 if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
1353                         return(0);
1354
1355                 /*
1356                  * We only support a single virtual address for ATA I/O.
1357                  */
1358                 if ((ccb->ccb_h.flags & CAM_DATA_MASK) != CAM_DATA_VADDR)
1359                         return (EINVAL);
1360
1361                 io_req->data_flags = CAM_DATA_VADDR;
1362
1363                 data_ptrs[0] = &ccb->ataio.data_ptr;
1364                 lengths[0] = ccb->ataio.dxfer_len;
1365                 dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1366                 numbufs = 1;
1367                 maxmap = softc->maxio;
1368                 break;
1369         case XPT_SMP_IO:
1370                 io_req->data_flags = CAM_DATA_VADDR;
1371
1372                 data_ptrs[0] = &ccb->smpio.smp_request;
1373                 lengths[0] = ccb->smpio.smp_request_len;
1374                 dirs[0] = CAM_DIR_OUT;
1375                 data_ptrs[1] = &ccb->smpio.smp_response;
1376                 lengths[1] = ccb->smpio.smp_response_len;
1377                 dirs[1] = CAM_DIR_IN;
1378                 numbufs = 2;
1379                 maxmap = softc->maxio;
1380                 break;
1381         case XPT_DEV_ADVINFO:
1382                 if (ccb->cdai.bufsiz == 0)
1383                         return (0);
1384
1385                 io_req->data_flags = CAM_DATA_VADDR;
1386
1387                 data_ptrs[0] = (uint8_t **)&ccb->cdai.buf;
1388                 lengths[0] = ccb->cdai.bufsiz;
1389                 dirs[0] = CAM_DIR_IN;
1390                 numbufs = 1;
1391                 break;
1392         case XPT_NVME_ADMIN:
1393         case XPT_NVME_IO:
1394                 if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
1395                         return (0);
1396
1397                 io_req->data_flags = ccb->ccb_h.flags & CAM_DATA_MASK;
1398
1399                 data_ptrs[0] = &ccb->nvmeio.data_ptr;
1400                 lengths[0] = ccb->nvmeio.dxfer_len;
1401                 dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1402                 num_segs = ccb->nvmeio.sglist_cnt;
1403                 seg_cnt_ptr = &ccb->nvmeio.sglist_cnt;
1404                 numbufs = 1;
1405                 maxmap = softc->maxio;
1406                 break;
1407         default:
1408                 return(EINVAL);
1409                 break; /* NOTREACHED */
1410         }
1411
1412         io_req->num_bufs = numbufs;
1413
1414         /*
1415          * If there is a maximum, check to make sure that the user's
1416          * request fits within the limit.  In general, we should only have
1417          * a maximum length for requests that go to hardware.  Otherwise it
1418          * is whatever we're able to malloc.
1419          */
1420         for (i = 0; i < numbufs; i++) {
1421                 io_req->user_bufs[i] = *data_ptrs[i];
1422                 io_req->dirs[i] = dirs[i];
1423                 io_req->lengths[i] = lengths[i];
1424
1425                 if (maxmap == 0)
1426                         continue;
1427
1428                 if (lengths[i] <= maxmap)
1429                         continue;
1430
1431                 xpt_print(periph->path, "%s: data length %u > max allowed %u "
1432                           "bytes\n", __func__, lengths[i], maxmap);
1433                 error = EINVAL;
1434                 goto bailout;
1435         }
1436
1437         switch (io_req->data_flags) {
1438         case CAM_DATA_VADDR:
1439                 /* Map or copy the buffer into kernel address space */
1440                 for (i = 0; i < numbufs; i++) {
1441                         uint8_t *tmp_buf;
1442
1443                         /*
1444                          * If for some reason no length is specified, we
1445                          * don't need to allocate anything.
1446                          */
1447                         if (io_req->lengths[i] == 0)
1448                                 continue;
1449
1450                         /*
1451                          * Make sure that the user's buffer is accessible
1452                          * to that process.
1453                          */
1454                         if (!useracc(io_req->user_bufs[i], io_req->lengths[i],
1455                             (io_req->dirs[i] == CAM_DIR_IN) ? VM_PROT_WRITE :
1456                              VM_PROT_READ)) {
1457                                 xpt_print(periph->path, "%s: user address %p "
1458                                     "length %u is not accessible\n", __func__,
1459                                     io_req->user_bufs[i], io_req->lengths[i]);
1460                                 error = EFAULT;
1461                                 goto bailout;
1462                         }
1463
1464                         tmp_buf = malloc(lengths[i], M_SCSIPASS,
1465                                          M_WAITOK | M_ZERO);
1466                         io_req->kern_bufs[i] = tmp_buf;
1467                         *data_ptrs[i] = tmp_buf;
1468
1469 #if 0
1470                         xpt_print(periph->path, "%s: malloced %p len %u, user "
1471                                   "buffer %p, operation: %s\n", __func__,
1472                                   tmp_buf, lengths[i], io_req->user_bufs[i],
1473                                   (dirs[i] == CAM_DIR_IN) ? "read" : "write");
1474 #endif
1475                         /*
1476                          * We only need to copy in if the user is writing.
1477                          */
1478                         if (dirs[i] != CAM_DIR_OUT)
1479                                 continue;
1480
1481                         error = copyin(io_req->user_bufs[i],
1482                                        io_req->kern_bufs[i], lengths[i]);
1483                         if (error != 0) {
1484                                 xpt_print(periph->path, "%s: copy of user "
1485                                           "buffer from %p to %p failed with "
1486                                           "error %d\n", __func__,
1487                                           io_req->user_bufs[i],
1488                                           io_req->kern_bufs[i], error);
1489                                 goto bailout;
1490                         }
1491                 }
1492                 break;
1493         case CAM_DATA_PADDR:
1494                 /* Pass down the pointer as-is */
1495                 break;
1496         case CAM_DATA_SG: {
1497                 size_t sg_length, size_to_go, alloc_size;
1498                 uint32_t num_segs_needed;
1499
1500                 /*
1501                  * Copy the user S/G list in, and then copy in the
1502                  * individual segments.
1503                  */
1504                 /*
1505                  * We shouldn't see this, but check just in case.
1506                  */
1507                 if (numbufs != 1) {
1508                         xpt_print(periph->path, "%s: cannot currently handle "
1509                                   "more than one S/G list per CCB\n", __func__);
1510                         error = EINVAL;
1511                         goto bailout;
1512                 }
1513
1514                 /*
1515                  * We have to have at least one segment.
1516                  */
1517                 if (num_segs == 0) {
1518                         xpt_print(periph->path, "%s: CAM_DATA_SG flag set, "
1519                                   "but sglist_cnt=0!\n", __func__);
1520                         error = EINVAL;
1521                         goto bailout;
1522                 }
1523
1524                 /*
1525                  * Make sure the user specified the total length and didn't
1526                  * just leave it to us to decode the S/G list.
1527                  */
1528                 if (lengths[0] == 0) {
1529                         xpt_print(periph->path, "%s: no dxfer_len specified, "
1530                                   "but CAM_DATA_SG flag is set!\n", __func__);
1531                         error = EINVAL;
1532                         goto bailout;
1533                 }
1534
1535                 /*
1536                  * We allocate buffers in io_zone_size increments for an
1537                  * S/G list.  This will generally be MAXPHYS.
1538                  */
1539                 if (lengths[0] <= softc->io_zone_size)
1540                         num_segs_needed = 1;
1541                 else {
1542                         num_segs_needed = lengths[0] / softc->io_zone_size;
1543                         if ((lengths[0] % softc->io_zone_size) != 0)
1544                                 num_segs_needed++;
1545                 }
1546
1547                 /* Figure out the size of the S/G list */
1548                 sg_length = num_segs * sizeof(bus_dma_segment_t);
1549                 io_req->num_user_segs = num_segs;
1550                 io_req->num_kern_segs = num_segs_needed;
1551
1552                 /* Save the user's S/G list pointer for later restoration */
1553                 io_req->user_bufs[0] = *data_ptrs[0];
1554
1555                 /*
1556                  * If we have enough segments allocated by default to handle
1557                  * the length of the user's S/G list,
1558                  */
1559                 if (num_segs > PASS_MAX_SEGS) {
1560                         io_req->user_segptr = malloc(sizeof(bus_dma_segment_t) *
1561                             num_segs, M_SCSIPASS, M_WAITOK | M_ZERO);
1562                         io_req->flags |= PASS_IO_USER_SEG_MALLOC;
1563                 } else
1564                         io_req->user_segptr = io_req->user_segs;
1565
1566                 if (!useracc(*data_ptrs[0], sg_length, VM_PROT_READ)) {
1567                         xpt_print(periph->path, "%s: unable to access user "
1568                                   "S/G list at %p\n", __func__, *data_ptrs[0]);
1569                         error = EFAULT;
1570                         goto bailout;
1571                 }
1572
1573                 error = copyin(*data_ptrs[0], io_req->user_segptr, sg_length);
1574                 if (error != 0) {
1575                         xpt_print(periph->path, "%s: copy of user S/G list "
1576                                   "from %p to %p failed with error %d\n",
1577                                   __func__, *data_ptrs[0], io_req->user_segptr,
1578                                   error);
1579                         goto bailout;
1580                 }
1581
1582                 if (num_segs_needed > PASS_MAX_SEGS) {
1583                         io_req->kern_segptr = malloc(sizeof(bus_dma_segment_t) *
1584                             num_segs_needed, M_SCSIPASS, M_WAITOK | M_ZERO);
1585                         io_req->flags |= PASS_IO_KERN_SEG_MALLOC;
1586                 } else {
1587                         io_req->kern_segptr = io_req->kern_segs;
1588                 }
1589
1590                 /*
1591                  * Allocate the kernel S/G list.
1592                  */
1593                 for (size_to_go = lengths[0], i = 0;
1594                      size_to_go > 0 && i < num_segs_needed;
1595                      i++, size_to_go -= alloc_size) {
1596                         uint8_t *kern_ptr;
1597
1598                         alloc_size = min(size_to_go, softc->io_zone_size);
1599                         kern_ptr = uma_zalloc(softc->pass_io_zone, M_WAITOK);
1600                         io_req->kern_segptr[i].ds_addr =
1601                             (bus_addr_t)(uintptr_t)kern_ptr;
1602                         io_req->kern_segptr[i].ds_len = alloc_size;
1603                 }
1604                 if (size_to_go > 0) {
1605                         printf("%s: size_to_go = %zu, software error!\n",
1606                                __func__, size_to_go);
1607                         error = EINVAL;
1608                         goto bailout;
1609                 }
1610
1611                 *data_ptrs[0] = (uint8_t *)io_req->kern_segptr;
1612                 *seg_cnt_ptr = io_req->num_kern_segs;
1613
1614                 /*
1615                  * We only need to copy data here if the user is writing.
1616                  */
1617                 if (dirs[0] == CAM_DIR_OUT)
1618                         error = passcopysglist(periph, io_req, dirs[0]);
1619                 break;
1620         }
1621         case CAM_DATA_SG_PADDR: {
1622                 size_t sg_length;
1623
1624                 /*
1625                  * We shouldn't see this, but check just in case.
1626                  */
1627                 if (numbufs != 1) {
1628                         printf("%s: cannot currently handle more than one "
1629                                "S/G list per CCB\n", __func__);
1630                         error = EINVAL;
1631                         goto bailout;
1632                 }
1633
1634                 /*
1635                  * We have to have at least one segment.
1636                  */
1637                 if (num_segs == 0) {
1638                         xpt_print(periph->path, "%s: CAM_DATA_SG_PADDR flag "
1639                                   "set, but sglist_cnt=0!\n", __func__);
1640                         error = EINVAL;
1641                         goto bailout;
1642                 }
1643
1644                 /*
1645                  * Make sure the user specified the total length and didn't
1646                  * just leave it to us to decode the S/G list.
1647                  */
1648                 if (lengths[0] == 0) {
1649                         xpt_print(periph->path, "%s: no dxfer_len specified, "
1650                                   "but CAM_DATA_SG flag is set!\n", __func__);
1651                         error = EINVAL;
1652                         goto bailout;
1653                 }
1654
1655                 /* Figure out the size of the S/G list */
1656                 sg_length = num_segs * sizeof(bus_dma_segment_t);
1657                 io_req->num_user_segs = num_segs;
1658                 io_req->num_kern_segs = io_req->num_user_segs;
1659
1660                 /* Save the user's S/G list pointer for later restoration */
1661                 io_req->user_bufs[0] = *data_ptrs[0];
1662
1663                 if (num_segs > PASS_MAX_SEGS) {
1664                         io_req->user_segptr = malloc(sizeof(bus_dma_segment_t) *
1665                             num_segs, M_SCSIPASS, M_WAITOK | M_ZERO);
1666                         io_req->flags |= PASS_IO_USER_SEG_MALLOC;
1667                 } else
1668                         io_req->user_segptr = io_req->user_segs;
1669
1670                 io_req->kern_segptr = io_req->user_segptr;
1671
1672                 error = copyin(*data_ptrs[0], io_req->user_segptr, sg_length);
1673                 if (error != 0) {
1674                         xpt_print(periph->path, "%s: copy of user S/G list "
1675                                   "from %p to %p failed with error %d\n",
1676                                   __func__, *data_ptrs[0], io_req->user_segptr,
1677                                   error);
1678                         goto bailout;
1679                 }
1680                 break;
1681         }
1682         default:
1683         case CAM_DATA_BIO:
1684                 /*
1685                  * A user shouldn't be attaching a bio to the CCB.  It
1686                  * isn't a user-accessible structure.
1687                  */
1688                 error = EINVAL;
1689                 break;
1690         }
1691
1692 bailout:
1693         if (error != 0)
1694                 passiocleanup(softc, io_req);
1695
1696         return (error);
1697 }
1698
1699 static int
1700 passmemdone(struct cam_periph *periph, struct pass_io_req *io_req)
1701 {
1702         struct pass_softc *softc;
1703         union ccb *ccb;
1704         int error;
1705         int i;
1706
1707         error = 0;
1708         softc = (struct pass_softc *)periph->softc;
1709         ccb = &io_req->ccb;
1710
1711         switch (io_req->data_flags) {
1712         case CAM_DATA_VADDR:
1713                 /*
1714                  * Copy back to the user buffer if this was a read.
1715                  */
1716                 for (i = 0; i < io_req->num_bufs; i++) {
1717                         if (io_req->dirs[i] != CAM_DIR_IN)
1718                                 continue;
1719
1720                         error = copyout(io_req->kern_bufs[i],
1721                             io_req->user_bufs[i], io_req->lengths[i]);
1722                         if (error != 0) {
1723                                 xpt_print(periph->path, "Unable to copy %u "
1724                                           "bytes from %p to user address %p\n",
1725                                           io_req->lengths[i],
1726                                           io_req->kern_bufs[i],
1727                                           io_req->user_bufs[i]);
1728                                 goto bailout;
1729                         }
1730
1731                 }
1732                 break;
1733         case CAM_DATA_PADDR:
1734                 /* Do nothing.  The pointer is a physical address already */
1735                 break;
1736         case CAM_DATA_SG:
1737                 /*
1738                  * Copy back to the user buffer if this was a read.
1739                  * Restore the user's S/G list buffer pointer.
1740                  */
1741                 if (io_req->dirs[0] == CAM_DIR_IN)
1742                         error = passcopysglist(periph, io_req, io_req->dirs[0]);
1743                 break;
1744         case CAM_DATA_SG_PADDR:
1745                 /*
1746                  * Restore the user's S/G list buffer pointer.  No need to
1747                  * copy.
1748                  */
1749                 break;
1750         default:
1751         case CAM_DATA_BIO:
1752                 error = EINVAL;
1753                 break;
1754         }
1755
1756 bailout:
1757         /*
1758          * Reset the user's pointers to their original values and free
1759          * allocated memory.
1760          */
1761         passiocleanup(softc, io_req);
1762
1763         return (error);
1764 }
1765
1766 static int
1767 passioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flag, struct thread *td)
1768 {
1769         int error;
1770
1771         if ((error = passdoioctl(dev, cmd, addr, flag, td)) == ENOTTY) {
1772                 error = cam_compat_ioctl(dev, cmd, addr, flag, td, passdoioctl);
1773         }
1774         return (error);
1775 }
1776
1777 static int
1778 passdoioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flag, struct thread *td)
1779 {
1780         struct  cam_periph *periph;
1781         struct  pass_softc *softc;
1782         int     error;
1783         uint32_t priority;
1784
1785         periph = (struct cam_periph *)dev->si_drv1;
1786         cam_periph_lock(periph);
1787         softc = (struct pass_softc *)periph->softc;
1788
1789         error = 0;
1790
1791         switch (cmd) {
1792
1793         case CAMIOCOMMAND:
1794         {
1795                 union ccb *inccb;
1796                 union ccb *ccb;
1797                 int ccb_malloced;
1798
1799                 inccb = (union ccb *)addr;
1800 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
1801                 if (inccb->ccb_h.func_code == XPT_SCSI_IO)
1802                         inccb->csio.bio = NULL;
1803 #endif
1804
1805                 if (inccb->ccb_h.flags & CAM_UNLOCKED) {
1806                         error = EINVAL;
1807                         break;
1808                 }
1809
1810                 /*
1811                  * Some CCB types, like scan bus and scan lun can only go
1812                  * through the transport layer device.
1813                  */
1814                 if (inccb->ccb_h.func_code & XPT_FC_XPT_ONLY) {
1815                         xpt_print(periph->path, "CCB function code %#x is "
1816                             "restricted to the XPT device\n",
1817                             inccb->ccb_h.func_code);
1818                         error = ENODEV;
1819                         break;
1820                 }
1821
1822                 /* Compatibility for RL/priority-unaware code. */
1823                 priority = inccb->ccb_h.pinfo.priority;
1824                 if (priority <= CAM_PRIORITY_OOB)
1825                     priority += CAM_PRIORITY_OOB + 1;
1826
1827                 /*
1828                  * Non-immediate CCBs need a CCB from the per-device pool
1829                  * of CCBs, which is scheduled by the transport layer.
1830                  * Immediate CCBs and user-supplied CCBs should just be
1831                  * malloced.
1832                  */
1833                 if ((inccb->ccb_h.func_code & XPT_FC_QUEUED)
1834                  && ((inccb->ccb_h.func_code & XPT_FC_USER_CCB) == 0)) {
1835                         ccb = cam_periph_getccb(periph, priority);
1836                         ccb_malloced = 0;
1837                 } else {
1838                         ccb = xpt_alloc_ccb_nowait();
1839
1840                         if (ccb != NULL)
1841                                 xpt_setup_ccb(&ccb->ccb_h, periph->path,
1842                                               priority);
1843                         ccb_malloced = 1;
1844                 }
1845
1846                 if (ccb == NULL) {
1847                         xpt_print(periph->path, "unable to allocate CCB\n");
1848                         error = ENOMEM;
1849                         break;
1850                 }
1851
1852                 error = passsendccb(periph, ccb, inccb);
1853
1854                 if (ccb_malloced)
1855                         xpt_free_ccb(ccb);
1856                 else
1857                         xpt_release_ccb(ccb);
1858
1859                 break;
1860         }
1861         case CAMIOQUEUE:
1862         {
1863                 struct pass_io_req *io_req;
1864                 union ccb **user_ccb, *ccb;
1865                 xpt_opcode fc;
1866
1867                 if ((softc->flags & PASS_FLAG_ZONE_VALID) == 0) {
1868                         error = passcreatezone(periph);
1869                         if (error != 0)
1870                                 goto bailout;
1871                 }
1872
1873                 /*
1874                  * We're going to do a blocking allocation for this I/O
1875                  * request, so we have to drop the lock.
1876                  */
1877                 cam_periph_unlock(periph);
1878
1879                 io_req = uma_zalloc(softc->pass_zone, M_WAITOK | M_ZERO);
1880                 ccb = &io_req->ccb;
1881                 user_ccb = (union ccb **)addr;
1882
1883                 /*
1884                  * Unlike the CAMIOCOMMAND ioctl above, we only have a
1885                  * pointer to the user's CCB, so we have to copy the whole
1886                  * thing in to a buffer we have allocated (above) instead
1887                  * of allowing the ioctl code to malloc a buffer and copy
1888                  * it in.
1889                  *
1890                  * This is an advantage for this asynchronous interface,
1891                  * since we don't want the memory to get freed while the
1892                  * CCB is outstanding.
1893                  */
1894 #if 0
1895                 xpt_print(periph->path, "Copying user CCB %p to "
1896                           "kernel address %p\n", *user_ccb, ccb);
1897 #endif
1898                 error = copyin(*user_ccb, ccb, sizeof(*ccb));
1899                 if (error != 0) {
1900                         xpt_print(periph->path, "Copy of user CCB %p to "
1901                                   "kernel address %p failed with error %d\n",
1902                                   *user_ccb, ccb, error);
1903                         goto camioqueue_error;
1904                 }
1905 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
1906                 if (ccb->ccb_h.func_code == XPT_SCSI_IO)
1907                         ccb->csio.bio = NULL;
1908 #endif
1909
1910                 if (ccb->ccb_h.flags & CAM_UNLOCKED) {
1911                         error = EINVAL;
1912                         goto camioqueue_error;
1913                 }
1914
1915                 if (ccb->ccb_h.flags & CAM_CDB_POINTER) {
1916                         if (ccb->csio.cdb_len > IOCDBLEN) {
1917                                 error = EINVAL;
1918                                 goto camioqueue_error;
1919                         }
1920                         error = copyin(ccb->csio.cdb_io.cdb_ptr,
1921                             ccb->csio.cdb_io.cdb_bytes, ccb->csio.cdb_len);
1922                         if (error != 0)
1923                                 goto camioqueue_error;
1924                         ccb->ccb_h.flags &= ~CAM_CDB_POINTER;
1925                 }
1926
1927                 /*
1928                  * Some CCB types, like scan bus and scan lun can only go
1929                  * through the transport layer device.
1930                  */
1931                 if (ccb->ccb_h.func_code & XPT_FC_XPT_ONLY) {
1932                         xpt_print(periph->path, "CCB function code %#x is "
1933                             "restricted to the XPT device\n",
1934                             ccb->ccb_h.func_code);
1935                         error = ENODEV;
1936                         goto camioqueue_error;
1937                 }
1938
1939                 /*
1940                  * Save the user's CCB pointer as well as his linked list
1941                  * pointers and peripheral private area so that we can
1942                  * restore these later.
1943                  */
1944                 io_req->user_ccb_ptr = *user_ccb;
1945                 io_req->user_periph_links = ccb->ccb_h.periph_links;
1946                 io_req->user_periph_priv = ccb->ccb_h.periph_priv;
1947
1948                 /*
1949                  * Now that we've saved the user's values, we can set our
1950                  * own peripheral private entry.
1951                  */
1952                 ccb->ccb_h.ccb_ioreq = io_req;
1953
1954                 /* Compatibility for RL/priority-unaware code. */
1955                 priority = ccb->ccb_h.pinfo.priority;
1956                 if (priority <= CAM_PRIORITY_OOB)
1957                     priority += CAM_PRIORITY_OOB + 1;
1958
1959                 /*
1960                  * Setup fields in the CCB like the path and the priority.
1961                  * The path in particular cannot be done in userland, since
1962                  * it is a pointer to a kernel data structure.
1963                  */
1964                 xpt_setup_ccb_flags(&ccb->ccb_h, periph->path, priority,
1965                                     ccb->ccb_h.flags);
1966
1967                 /*
1968                  * Setup our done routine.  There is no way for the user to
1969                  * have a valid pointer here.
1970                  */
1971                 ccb->ccb_h.cbfcnp = passdone;
1972
1973                 fc = ccb->ccb_h.func_code;
1974                 /*
1975                  * If this function code has memory that can be mapped in
1976                  * or out, we need to call passmemsetup().
1977                  */
1978                 if ((fc == XPT_SCSI_IO) || (fc == XPT_ATA_IO)
1979                  || (fc == XPT_SMP_IO) || (fc == XPT_DEV_MATCH)
1980                  || (fc == XPT_DEV_ADVINFO)
1981                  || (fc == XPT_NVME_ADMIN) || (fc == XPT_NVME_IO)) {
1982                         error = passmemsetup(periph, io_req);
1983                         if (error != 0)
1984                                 goto camioqueue_error;
1985                 } else
1986                         io_req->mapinfo.num_bufs_used = 0;
1987
1988                 cam_periph_lock(periph);
1989
1990                 /*
1991                  * Everything goes on the incoming queue initially.
1992                  */
1993                 TAILQ_INSERT_TAIL(&softc->incoming_queue, io_req, links);
1994
1995                 /*
1996                  * If the CCB is queued, and is not a user CCB, then
1997                  * we need to allocate a slot for it.  Call xpt_schedule()
1998                  * so that our start routine will get called when a CCB is
1999                  * available.
2000                  */
2001                 if ((fc & XPT_FC_QUEUED)
2002                  && ((fc & XPT_FC_USER_CCB) == 0)) {
2003                         xpt_schedule(periph, priority);
2004                         break;
2005                 }
2006
2007                 /*
2008                  * At this point, the CCB in question is either an
2009                  * immediate CCB (like XPT_DEV_ADVINFO) or it is a user CCB
2010                  * and therefore should be malloced, not allocated via a slot.
2011                  * Remove the CCB from the incoming queue and add it to the
2012                  * active queue.
2013                  */
2014                 TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
2015                 TAILQ_INSERT_TAIL(&softc->active_queue, io_req, links);
2016
2017                 xpt_action(ccb);
2018
2019                 /*
2020                  * If this is not a queued CCB (i.e. it is an immediate CCB),
2021                  * then it is already done.  We need to put it on the done
2022                  * queue for the user to fetch.
2023                  */
2024                 if ((fc & XPT_FC_QUEUED) == 0) {
2025                         TAILQ_REMOVE(&softc->active_queue, io_req, links);
2026                         TAILQ_INSERT_TAIL(&softc->done_queue, io_req, links);
2027                 }
2028                 break;
2029
2030 camioqueue_error:
2031                 uma_zfree(softc->pass_zone, io_req);
2032                 cam_periph_lock(periph);
2033                 break;
2034         }
2035         case CAMIOGET:
2036         {
2037                 union ccb **user_ccb;
2038                 struct pass_io_req *io_req;
2039                 int old_error;
2040
2041                 user_ccb = (union ccb **)addr;
2042                 old_error = 0;
2043
2044                 io_req = TAILQ_FIRST(&softc->done_queue);
2045                 if (io_req == NULL) {
2046                         error = ENOENT;
2047                         break;
2048                 }
2049
2050                 /*
2051                  * Remove the I/O from the done queue.
2052                  */
2053                 TAILQ_REMOVE(&softc->done_queue, io_req, links);
2054
2055                 /*
2056                  * We have to drop the lock during the copyout because the
2057                  * copyout can result in VM faults that require sleeping.
2058                  */
2059                 cam_periph_unlock(periph);
2060
2061                 /*
2062                  * Do any needed copies (e.g. for reads) and revert the
2063                  * pointers in the CCB back to the user's pointers.
2064                  */
2065                 error = passmemdone(periph, io_req);
2066
2067                 old_error = error;
2068
2069                 io_req->ccb.ccb_h.periph_links = io_req->user_periph_links;
2070                 io_req->ccb.ccb_h.periph_priv = io_req->user_periph_priv;
2071
2072 #if 0
2073                 xpt_print(periph->path, "Copying to user CCB %p from "
2074                           "kernel address %p\n", *user_ccb, &io_req->ccb);
2075 #endif
2076
2077                 error = copyout(&io_req->ccb, *user_ccb, sizeof(union ccb));
2078                 if (error != 0) {
2079                         xpt_print(periph->path, "Copy to user CCB %p from "
2080                                   "kernel address %p failed with error %d\n",
2081                                   *user_ccb, &io_req->ccb, error);
2082                 }
2083
2084                 /*
2085                  * Prefer the first error we got back, and make sure we
2086                  * don't overwrite bad status with good.
2087                  */
2088                 if (old_error != 0)
2089                         error = old_error;
2090
2091                 cam_periph_lock(periph);
2092
2093                 /*
2094                  * At this point, if there was an error, we could potentially
2095                  * re-queue the I/O and try again.  But why?  The error
2096                  * would almost certainly happen again.  We might as well
2097                  * not leak memory.
2098                  */
2099                 uma_zfree(softc->pass_zone, io_req);
2100                 break;
2101         }
2102         default:
2103                 error = cam_periph_ioctl(periph, cmd, addr, passerror);
2104                 break;
2105         }
2106
2107 bailout:
2108         cam_periph_unlock(periph);
2109
2110         return(error);
2111 }
2112
2113 static int
2114 passpoll(struct cdev *dev, int poll_events, struct thread *td)
2115 {
2116         struct cam_periph *periph;
2117         struct pass_softc *softc;
2118         int revents;
2119
2120         periph = (struct cam_periph *)dev->si_drv1;
2121         softc = (struct pass_softc *)periph->softc;
2122
2123         revents = poll_events & (POLLOUT | POLLWRNORM);
2124         if ((poll_events & (POLLIN | POLLRDNORM)) != 0) {
2125                 cam_periph_lock(periph);
2126
2127                 if (!TAILQ_EMPTY(&softc->done_queue)) {
2128                         revents |= poll_events & (POLLIN | POLLRDNORM);
2129                 }
2130                 cam_periph_unlock(periph);
2131                 if (revents == 0)
2132                         selrecord(td, &softc->read_select);
2133         }
2134
2135         return (revents);
2136 }
2137
2138 static int
2139 passkqfilter(struct cdev *dev, struct knote *kn)
2140 {
2141         struct cam_periph *periph;
2142         struct pass_softc *softc;
2143
2144         periph = (struct cam_periph *)dev->si_drv1;
2145         softc = (struct pass_softc *)periph->softc;
2146
2147         kn->kn_hook = (caddr_t)periph;
2148         kn->kn_fop = &passread_filtops;
2149         knlist_add(&softc->read_select.si_note, kn, 0);
2150
2151         return (0);
2152 }
2153
2154 static void
2155 passreadfiltdetach(struct knote *kn)
2156 {
2157         struct cam_periph *periph;
2158         struct pass_softc *softc;
2159
2160         periph = (struct cam_periph *)kn->kn_hook;
2161         softc = (struct pass_softc *)periph->softc;
2162
2163         knlist_remove(&softc->read_select.si_note, kn, 0);
2164 }
2165
2166 static int
2167 passreadfilt(struct knote *kn, long hint)
2168 {
2169         struct cam_periph *periph;
2170         struct pass_softc *softc;
2171         int retval;
2172
2173         periph = (struct cam_periph *)kn->kn_hook;
2174         softc = (struct pass_softc *)periph->softc;
2175
2176         cam_periph_assert(periph, MA_OWNED);
2177
2178         if (TAILQ_EMPTY(&softc->done_queue))
2179                 retval = 0;
2180         else
2181                 retval = 1;
2182
2183         return (retval);
2184 }
2185
2186 /*
2187  * Generally, "ccb" should be the CCB supplied by the kernel.  "inccb"
2188  * should be the CCB that is copied in from the user.
2189  */
2190 static int
2191 passsendccb(struct cam_periph *periph, union ccb *ccb, union ccb *inccb)
2192 {
2193         struct pass_softc *softc;
2194         struct cam_periph_map_info mapinfo;
2195         uint8_t *cmd;
2196         xpt_opcode fc;
2197         int error;
2198
2199         softc = (struct pass_softc *)periph->softc;
2200
2201         /*
2202          * There are some fields in the CCB header that need to be
2203          * preserved, the rest we get from the user.
2204          */
2205         xpt_merge_ccb(ccb, inccb);
2206
2207         if (ccb->ccb_h.flags & CAM_CDB_POINTER) {
2208                 cmd = __builtin_alloca(ccb->csio.cdb_len);
2209                 error = copyin(ccb->csio.cdb_io.cdb_ptr, cmd, ccb->csio.cdb_len);
2210                 if (error)
2211                         return (error);
2212                 ccb->csio.cdb_io.cdb_ptr = cmd;
2213         }
2214
2215         /*
2216          */
2217         ccb->ccb_h.cbfcnp = passdone;
2218
2219         /*
2220          * Let cam_periph_mapmem do a sanity check on the data pointer format.
2221          * Even if no data transfer is needed, it's a cheap check and it
2222          * simplifies the code.
2223          */
2224         fc = ccb->ccb_h.func_code;
2225         if ((fc == XPT_SCSI_IO) || (fc == XPT_ATA_IO) || (fc == XPT_SMP_IO)
2226             || (fc == XPT_DEV_MATCH) || (fc == XPT_DEV_ADVINFO) || (fc == XPT_MMC_IO)
2227             || (fc == XPT_NVME_ADMIN) || (fc == XPT_NVME_IO)) {
2228
2229                 bzero(&mapinfo, sizeof(mapinfo));
2230
2231                 /*
2232                  * cam_periph_mapmem calls into proc and vm functions that can
2233                  * sleep as well as trigger I/O, so we can't hold the lock.
2234                  * Dropping it here is reasonably safe.
2235                  */
2236                 cam_periph_unlock(periph);
2237                 error = cam_periph_mapmem(ccb, &mapinfo, softc->maxio);
2238                 cam_periph_lock(periph);
2239
2240                 /*
2241                  * cam_periph_mapmem returned an error, we can't continue.
2242                  * Return the error to the user.
2243                  */
2244                 if (error)
2245                         return(error);
2246         } else
2247                 /* Ensure that the unmap call later on is a no-op. */
2248                 mapinfo.num_bufs_used = 0;
2249
2250         /*
2251          * If the user wants us to perform any error recovery, then honor
2252          * that request.  Otherwise, it's up to the user to perform any
2253          * error recovery.
2254          */
2255         cam_periph_runccb(ccb, (ccb->ccb_h.flags & CAM_PASS_ERR_RECOVER) ?
2256             passerror : NULL, /* cam_flags */ CAM_RETRY_SELTO,
2257             /* sense_flags */ SF_RETRY_UA | SF_NO_PRINT,
2258             softc->device_stats);
2259
2260         cam_periph_unmapmem(ccb, &mapinfo);
2261
2262         ccb->ccb_h.cbfcnp = NULL;
2263         ccb->ccb_h.periph_priv = inccb->ccb_h.periph_priv;
2264         bcopy(ccb, inccb, sizeof(union ccb));
2265
2266         return(0);
2267 }
2268
2269 static int
2270 passerror(union ccb *ccb, u_int32_t cam_flags, u_int32_t sense_flags)
2271 {
2272         struct cam_periph *periph;
2273         struct pass_softc *softc;
2274
2275         periph = xpt_path_periph(ccb->ccb_h.path);
2276         softc = (struct pass_softc *)periph->softc;
2277
2278         return(cam_periph_error(ccb, cam_flags, sense_flags,
2279                                  &softc->saved_ccb));
2280 }