sys/cam/scsi/scsi_pass.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause
   3  *
   4  * Copyright (c) 1997, 1998, 2000 Justin T. Gibbs.
   5  * Copyright (c) 1997, 1998, 1999 Kenneth D. Merry.
   6  * All rights reserved.
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions, and the following disclaimer,
  13  *    without modification, immediately at the beginning of the file.
  14  * 2. The name of the author may not be used to endorse or promote products
  15  *    derived from this software without specific prior written permission.
  16  *
  17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  20  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR
  21  * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  27  * SUCH DAMAGE.
  28  */
  29
  30 #include <sys/cdefs.h>
  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/sysent.h>
  47 #include <sys/taskqueue.h>
  48 #include <vm/uma.h>
  49 #include <vm/vm.h>
  50 #include <vm/vm_extern.h>
  51
  52 #include <machine/bus.h>
  53
  54 #include <cam/cam.h>
  55 #include <cam/cam_ccb.h>
  56 #include <cam/cam_periph.h>
  57 #include <cam/cam_queue.h>
  58 #include <cam/cam_xpt.h>
  59 #include <cam/cam_xpt_periph.h>
  60 #include <cam/cam_debug.h>
  61 #include <cam/cam_compat.h>
  62 #include <cam/cam_xpt_periph.h>
  63
  64 #include <cam/scsi/scsi_all.h>
  65 #include <cam/scsi/scsi_pass.h>
  66
  67 typedef enum {
  68         PASS_FLAG_OPEN                  = 0x01,
  69         PASS_FLAG_LOCKED                = 0x02,
  70         PASS_FLAG_INVALID               = 0x04,
  71         PASS_FLAG_INITIAL_PHYSPATH      = 0x08,
  72         PASS_FLAG_ZONE_INPROG           = 0x10,
  73         PASS_FLAG_ZONE_VALID            = 0x20,
  74         PASS_FLAG_UNMAPPED_CAPABLE      = 0x40,
  75         PASS_FLAG_ABANDONED_REF_SET     = 0x80
  76 } pass_flags;
  77
  78 typedef enum {
  79         PASS_STATE_NORMAL
  80 } pass_state;
  81
  82 typedef enum {
  83         PASS_CCB_BUFFER_IO,
  84         PASS_CCB_QUEUED_IO
  85 } pass_ccb_types;
  86
  87 #define ccb_type        ppriv_field0
  88 #define ccb_ioreq       ppriv_ptr1
  89
  90 /*
  91  * The maximum number of memory segments we preallocate.
  92  */
  93 #define PASS_MAX_SEGS   16
  94
  95 typedef enum {
  96         PASS_IO_NONE            = 0x00,
  97         PASS_IO_USER_SEG_MALLOC = 0x01,
  98         PASS_IO_KERN_SEG_MALLOC = 0x02,
  99         PASS_IO_ABANDONED       = 0x04
 100 } pass_io_flags;
 101
 102 struct pass_io_req {
 103         union ccb                        ccb;
 104         union ccb                       *alloced_ccb;
 105         union ccb                       *user_ccb_ptr;
 106         camq_entry                       user_periph_links;
 107         ccb_ppriv_area                   user_periph_priv;
 108         struct cam_periph_map_info       mapinfo;
 109         pass_io_flags                    flags;
 110         ccb_flags                        data_flags;
 111         int                              num_user_segs;
 112         bus_dma_segment_t                user_segs[PASS_MAX_SEGS];
 113         int                              num_kern_segs;
 114         bus_dma_segment_t                kern_segs[PASS_MAX_SEGS];
 115         bus_dma_segment_t               *user_segptr;
 116         bus_dma_segment_t               *kern_segptr;
 117         int                              num_bufs;
 118         uint32_t                         dirs[CAM_PERIPH_MAXMAPS];
 119         uint32_t                         lengths[CAM_PERIPH_MAXMAPS];
 120         uint8_t                         *user_bufs[CAM_PERIPH_MAXMAPS];
 121         uint8_t                         *kern_bufs[CAM_PERIPH_MAXMAPS];
 122         struct bintime                   start_time;
 123         TAILQ_ENTRY(pass_io_req)         links;
 124 };
 125
 126 struct pass_softc {
 127         pass_state                state;
 128         pass_flags                flags;
 129         uint8_t           pd_type;
 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, uint32_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, uint32_t cam_flags,
 181                                   uint32_t sense_flags);
 182 static  int             passsendccb(struct cam_periph *periph, union ccb *ccb,
 183                                     union ccb *inccb);
 184 static  void            passflags(union ccb *ccb, uint32_t *cam_flags,
 185                                   uint32_t *sense_flags);
 186
 187 static struct periph_driver passdriver =
 188 {
 189         passinit, "pass",
 190         TAILQ_HEAD_INITIALIZER(passdriver.units), /* generation */ 0
 191 };
 192
 193 PERIPHDRIVER_DECLARE(pass, passdriver);
 194
 195 static struct cdevsw pass_cdevsw = {
 196         .d_version =    D_VERSION,
 197         .d_flags =      D_TRACKCLOSE,
 198         .d_open =       passopen,
 199         .d_close =      passclose,
 200         .d_ioctl =      passioctl,
 201         .d_poll =       passpoll,
 202         .d_kqfilter =   passkqfilter,
 203         .d_name =       "pass",
 204 };
 205
 206 static struct filterops passread_filtops = {
 207         .f_isfd =       1,
 208         .f_detach =     passreadfiltdetach,
 209         .f_event =      passreadfilt
 210 };
 211
 212 static MALLOC_DEFINE(M_SCSIPASS, "scsi_pass", "scsi passthrough buffers");
 213
 214 static void
 215 passinit(void)
 216 {
 217         cam_status status;
 218
 219         /*
 220          * Install a global async callback.  This callback will
 221          * receive async callbacks like "new device found".
 222          */
 223         status = xpt_register_async(AC_FOUND_DEVICE, passasync, NULL, NULL);
 224
 225         if (status != CAM_REQ_CMP) {
 226                 printf("pass: Failed to attach master async callback "
 227                        "due to status 0x%x!\n", status);
 228         }
 229
 230 }
 231
 232 static void
 233 passrejectios(struct cam_periph *periph)
 234 {
 235         struct pass_io_req *io_req, *io_req2;
 236         struct pass_softc *softc;
 237
 238         softc = (struct pass_softc *)periph->softc;
 239
 240         /*
 241          * The user can no longer get status for I/O on the done queue, so
 242          * clean up all outstanding I/O on the done queue.
 243          */
 244         TAILQ_FOREACH_SAFE(io_req, &softc->done_queue, links, io_req2) {
 245                 TAILQ_REMOVE(&softc->done_queue, io_req, links);
 246                 passiocleanup(softc, io_req);
 247                 uma_zfree(softc->pass_zone, io_req);
 248         }
 249
 250         /*
 251          * The underlying device is gone, so we can't issue these I/Os.
 252          * The devfs node has been shut down, so we can't return status to
 253          * the user.  Free any I/O left on the incoming queue.
 254          */
 255         TAILQ_FOREACH_SAFE(io_req, &softc->incoming_queue, links, io_req2) {
 256                 TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
 257                 passiocleanup(softc, io_req);
 258                 uma_zfree(softc->pass_zone, io_req);
 259         }
 260
 261         /*
 262          * Normally we would put I/Os on the abandoned queue and acquire a
 263          * reference when we saw the final close.  But, the device went
 264          * away and devfs may have moved everything off to deadfs by the
 265          * time the I/O done callback is called; as a result, we won't see
 266          * any more closes.  So, if we have any active I/Os, we need to put
 267          * them on the abandoned queue.  When the abandoned queue is empty,
 268          * we'll release the remaining reference (see below) to the peripheral.
 269          */
 270         TAILQ_FOREACH_SAFE(io_req, &softc->active_queue, links, io_req2) {
 271                 TAILQ_REMOVE(&softc->active_queue, io_req, links);
 272                 io_req->flags |= PASS_IO_ABANDONED;
 273                 TAILQ_INSERT_TAIL(&softc->abandoned_queue, io_req, links);
 274         }
 275
 276         /*
 277          * If we put any I/O on the abandoned queue, acquire a reference.
 278          */
 279         if ((!TAILQ_EMPTY(&softc->abandoned_queue))
 280          && ((softc->flags & PASS_FLAG_ABANDONED_REF_SET) == 0)) {
 281                 cam_periph_doacquire(periph);
 282                 softc->flags |= PASS_FLAG_ABANDONED_REF_SET;
 283         }
 284 }
 285
 286 static void
 287 passdevgonecb(void *arg)
 288 {
 289         struct cam_periph *periph;
 290         struct mtx *mtx;
 291         struct pass_softc *softc;
 292         int i;
 293
 294         periph = (struct cam_periph *)arg;
 295         mtx = cam_periph_mtx(periph);
 296         mtx_lock(mtx);
 297
 298         softc = (struct pass_softc *)periph->softc;
 299         KASSERT(softc->open_count >= 0, ("Negative open count %d",
 300                 softc->open_count));
 301
 302         /*
 303          * When we get this callback, we will get no more close calls from
 304          * devfs.  So if we have any dangling opens, we need to release the
 305          * reference held for that particular context.
 306          */
 307         for (i = 0; i < softc->open_count; i++)
 308                 cam_periph_release_locked(periph);
 309
 310         softc->open_count = 0;
 311
 312         /*
 313          * Release the reference held for the device node, it is gone now.
 314          * Accordingly, inform all queued I/Os of their fate.
 315          */
 316         cam_periph_release_locked(periph);
 317         passrejectios(periph);
 318
 319         /*
 320          * We reference the SIM lock directly here, instead of using
 321          * cam_periph_unlock().  The reason is that the final call to
 322          * cam_periph_release_locked() above could result in the periph
 323          * getting freed.  If that is the case, dereferencing the periph
 324          * with a cam_periph_unlock() call would cause a page fault.
 325          */
 326         mtx_unlock(mtx);
 327
 328         /*
 329          * We have to remove our kqueue context from a thread because it
 330          * may sleep.  It would be nice if we could get a callback from
 331          * kqueue when it is done cleaning up resources.
 332          */
 333         taskqueue_enqueue(taskqueue_thread, &softc->shutdown_kqueue_task);
 334 }
 335
 336 static void
 337 passoninvalidate(struct cam_periph *periph)
 338 {
 339         struct pass_softc *softc;
 340
 341         softc = (struct pass_softc *)periph->softc;
 342
 343         /*
 344          * De-register any async callbacks.
 345          */
 346         xpt_register_async(0, passasync, periph, periph->path);
 347
 348         softc->flags |= PASS_FLAG_INVALID;
 349
 350         /*
 351          * Tell devfs this device has gone away, and ask for a callback
 352          * when it has cleaned up its state.
 353          */
 354         destroy_dev_sched_cb(softc->dev, passdevgonecb, periph);
 355 }
 356
 357 static void
 358 passcleanup(struct cam_periph *periph)
 359 {
 360         struct pass_softc *softc;
 361
 362         softc = (struct pass_softc *)periph->softc;
 363
 364         cam_periph_assert(periph, MA_OWNED);
 365         KASSERT(TAILQ_EMPTY(&softc->active_queue),
 366                 ("%s called when there are commands on the active queue!\n",
 367                 __func__));
 368         KASSERT(TAILQ_EMPTY(&softc->abandoned_queue),
 369                 ("%s called when there are commands on the abandoned queue!\n",
 370                 __func__));
 371         KASSERT(TAILQ_EMPTY(&softc->incoming_queue),
 372                 ("%s called when there are commands on the incoming queue!\n",
 373                 __func__));
 374         KASSERT(TAILQ_EMPTY(&softc->done_queue),
 375                 ("%s called when there are commands on the done queue!\n",
 376                 __func__));
 377
 378         devstat_remove_entry(softc->device_stats);
 379
 380         cam_periph_unlock(periph);
 381
 382         /*
 383          * We call taskqueue_drain() for the physpath task to make sure it
 384          * is complete.  We drop the lock because this can potentially
 385          * sleep.  XXX KDM that is bad.  Need a way to get a callback when
 386          * a taskqueue is drained.
 387          *
 388          * Note that we don't drain the kqueue shutdown task queue.  This
 389          * is because we hold a reference on the periph for kqueue, and
 390          * release that reference from the kqueue shutdown task queue.  So
 391          * we cannot come into this routine unless we've released that
 392          * reference.  Also, because that could be the last reference, we
 393          * could be called from the cam_periph_release() call in
 394          * pass_shutdown_kqueue().  In that case, the taskqueue_drain()
 395          * would deadlock.  It would be preferable if we had a way to
 396          * get a callback when a taskqueue is done.
 397          */
 398         taskqueue_drain(taskqueue_thread, &softc->add_physpath_task);
 399
 400         /*
 401          * It should be safe to destroy the zones from here, because all
 402          * of the references to this peripheral have been freed, and all
 403          * I/O has been terminated and freed.  We check the zones for NULL
 404          * because they may not have been allocated yet if the device went
 405          * away before any asynchronous I/O has been issued.
 406          */
 407         if (softc->pass_zone != NULL)
 408                 uma_zdestroy(softc->pass_zone);
 409         if (softc->pass_io_zone != NULL)
 410                 uma_zdestroy(softc->pass_io_zone);
 411
 412         cam_periph_lock(periph);
 413
 414         free(softc, M_DEVBUF);
 415 }
 416
 417 static void
 418 pass_shutdown_kqueue(void *context, int pending)
 419 {
 420         struct cam_periph *periph;
 421         struct pass_softc *softc;
 422
 423         periph = context;
 424         softc = periph->softc;
 425
 426         knlist_clear(&softc->read_select.si_note, /*is_locked*/ 0);
 427         knlist_destroy(&softc->read_select.si_note);
 428
 429         /*
 430          * Release the reference we held for kqueue.
 431          */
 432         cam_periph_release(periph);
 433 }
 434
 435 static void
 436 pass_add_physpath(void *context, int pending)
 437 {
 438         struct cam_periph *periph;
 439         struct pass_softc *softc;
 440         struct mtx *mtx;
 441         char *physpath;
 442
 443         /*
 444          * If we have one, create a devfs alias for our
 445          * physical path.
 446          */
 447         periph = context;
 448         softc = periph->softc;
 449         physpath = malloc(MAXPATHLEN, M_DEVBUF, M_WAITOK);
 450         mtx = cam_periph_mtx(periph);
 451         mtx_lock(mtx);
 452
 453         if (periph->flags & CAM_PERIPH_INVALID)
 454                 goto out;
 455
 456         if (xpt_getattr(physpath, MAXPATHLEN,
 457                         "GEOM::physpath", periph->path) == 0
 458          && strlen(physpath) != 0) {
 459                 mtx_unlock(mtx);
 460                 make_dev_physpath_alias(MAKEDEV_WAITOK | MAKEDEV_CHECKNAME,
 461                                 &softc->alias_dev, softc->dev,
 462                                 softc->alias_dev, physpath);
 463                 mtx_lock(mtx);
 464         }
 465
 466 out:
 467         /*
 468          * Now that we've made our alias, we no longer have to have a
 469          * reference to the device.
 470          */
 471         if ((softc->flags & PASS_FLAG_INITIAL_PHYSPATH) == 0)
 472                 softc->flags |= PASS_FLAG_INITIAL_PHYSPATH;
 473
 474         /*
 475          * We always acquire a reference to the periph before queueing this
 476          * task queue function, so it won't go away before we run.
 477          */
 478         while (pending-- > 0)
 479                 cam_periph_release_locked(periph);
 480         mtx_unlock(mtx);
 481
 482         free(physpath, M_DEVBUF);
 483 }
 484
 485 static void
 486 passasync(void *callback_arg, uint32_t code,
 487           struct cam_path *path, void *arg)
 488 {
 489         struct cam_periph *periph;
 490
 491         periph = (struct cam_periph *)callback_arg;
 492
 493         switch (code) {
 494         case AC_FOUND_DEVICE:
 495         {
 496                 struct ccb_getdev *cgd;
 497                 cam_status status;
 498
 499                 cgd = (struct ccb_getdev *)arg;
 500                 if (cgd == NULL)
 501                         break;
 502
 503                 /*
 504                  * Allocate a peripheral instance for
 505                  * this device and start the probe
 506                  * process.
 507                  */
 508                 status = cam_periph_alloc(passregister, passoninvalidate,
 509                                           passcleanup, passstart, "pass",
 510                                           CAM_PERIPH_BIO, path,
 511                                           passasync, AC_FOUND_DEVICE, cgd);
 512
 513                 if (status != CAM_REQ_CMP
 514                  && status != CAM_REQ_INPROG) {
 515                         const struct cam_status_entry *entry;
 516
 517                         entry = cam_fetch_status_entry(status);
 518
 519                         printf("passasync: Unable to attach new device "
 520                                "due to status %#x: %s\n", status, entry ?
 521                                entry->status_text : "Unknown");
 522                 }
 523
 524                 break;
 525         }
 526         case AC_ADVINFO_CHANGED:
 527         {
 528                 uintptr_t buftype;
 529
 530                 buftype = (uintptr_t)arg;
 531                 if (buftype == CDAI_TYPE_PHYS_PATH) {
 532                         struct pass_softc *softc;
 533
 534                         softc = (struct pass_softc *)periph->softc;
 535                         /*
 536                          * Acquire a reference to the periph before we
 537                          * start the taskqueue, so that we don't run into
 538                          * a situation where the periph goes away before
 539                          * the task queue has a chance to run.
 540                          */
 541                         if (cam_periph_acquire(periph) != 0)
 542                                 break;
 543
 544                         taskqueue_enqueue(taskqueue_thread,
 545                                           &softc->add_physpath_task);
 546                 }
 547                 break;
 548         }
 549         default:
 550                 cam_periph_async(periph, code, path, arg);
 551                 break;
 552         }
 553 }
 554
 555 static cam_status
 556 passregister(struct cam_periph *periph, void *arg)
 557 {
 558         struct pass_softc *softc;
 559         struct ccb_getdev *cgd;
 560         struct ccb_pathinq cpi;
 561         struct make_dev_args args;
 562         int error, no_tags;
 563
 564         cgd = (struct ccb_getdev *)arg;
 565         if (cgd == NULL) {
 566                 printf("%s: no getdev CCB, can't register device\n", __func__);
 567                 return(CAM_REQ_CMP_ERR);
 568         }
 569
 570         softc = (struct pass_softc *)malloc(sizeof(*softc),
 571                                             M_DEVBUF, M_NOWAIT);
 572
 573         if (softc == NULL) {
 574                 printf("%s: Unable to probe new device. "
 575                        "Unable to allocate softc\n", __func__);
 576                 return(CAM_REQ_CMP_ERR);
 577         }
 578
 579         bzero(softc, sizeof(*softc));
 580         softc->state = PASS_STATE_NORMAL;
 581         if (cgd->protocol == PROTO_SCSI || cgd->protocol == PROTO_ATAPI)
 582                 softc->pd_type = SID_TYPE(&cgd->inq_data);
 583         else if (cgd->protocol == PROTO_SATAPM)
 584                 softc->pd_type = T_ENCLOSURE;
 585         else
 586                 softc->pd_type = T_DIRECT;
 587
 588         periph->softc = softc;
 589         softc->periph = periph;
 590         TAILQ_INIT(&softc->incoming_queue);
 591         TAILQ_INIT(&softc->active_queue);
 592         TAILQ_INIT(&softc->abandoned_queue);
 593         TAILQ_INIT(&softc->done_queue);
 594         snprintf(softc->zone_name, sizeof(softc->zone_name), "%s%d",
 595                  periph->periph_name, periph->unit_number);
 596         snprintf(softc->io_zone_name, sizeof(softc->io_zone_name), "%s%dIO",
 597                  periph->periph_name, periph->unit_number);
 598         softc->io_zone_size = maxphys;
 599         knlist_init_mtx(&softc->read_select.si_note, cam_periph_mtx(periph));
 600
 601         xpt_path_inq(&cpi, periph->path);
 602
 603         if (cpi.maxio == 0)
 604                 softc->maxio = DFLTPHYS;        /* traditional default */
 605         else if (cpi.maxio > maxphys)
 606                 softc->maxio = maxphys;         /* for safety */
 607         else
 608                 softc->maxio = cpi.maxio;       /* real value */
 609
 610         if (cpi.hba_misc & PIM_UNMAPPED)
 611                 softc->flags |= PASS_FLAG_UNMAPPED_CAPABLE;
 612
 613         /*
 614          * We pass in 0 for a blocksize, since we don't
 615          * know what the blocksize of this device is, if
 616          * it even has a blocksize.
 617          */
 618         cam_periph_unlock(periph);
 619         no_tags = (cgd->inq_data.flags & SID_CmdQue) == 0;
 620         softc->device_stats = devstat_new_entry("pass",
 621                           periph->unit_number, 0,
 622                           DEVSTAT_NO_BLOCKSIZE
 623                           | (no_tags ? DEVSTAT_NO_ORDERED_TAGS : 0),
 624                           softc->pd_type |
 625                           XPORT_DEVSTAT_TYPE(cpi.transport) |
 626                           DEVSTAT_TYPE_PASS,
 627                           DEVSTAT_PRIORITY_PASS);
 628
 629         /*
 630          * Initialize the taskqueue handler for shutting down kqueue.
 631          */
 632         TASK_INIT(&softc->shutdown_kqueue_task, /*priority*/ 0,
 633                   pass_shutdown_kqueue, periph);
 634
 635         /*
 636          * Acquire a reference to the periph that we can release once we've
 637          * cleaned up the kqueue.
 638          */
 639         if (cam_periph_acquire(periph) != 0) {
 640                 xpt_print(periph->path, "%s: lost periph during "
 641                           "registration!\n", __func__);
 642                 cam_periph_lock(periph);
 643                 return (CAM_REQ_CMP_ERR);
 644         }
 645
 646         /*
 647          * Acquire a reference to the periph before we create the devfs
 648          * instance for it.  We'll release this reference once the devfs
 649          * instance has been freed.
 650          */
 651         if (cam_periph_acquire(periph) != 0) {
 652                 xpt_print(periph->path, "%s: lost periph during "
 653                           "registration!\n", __func__);
 654                 cam_periph_lock(periph);
 655                 return (CAM_REQ_CMP_ERR);
 656         }
 657
 658         /* Register the device */
 659         make_dev_args_init(&args);
 660         args.mda_devsw = &pass_cdevsw;
 661         args.mda_unit = periph->unit_number;
 662         args.mda_uid = UID_ROOT;
 663         args.mda_gid = GID_OPERATOR;
 664         args.mda_mode = 0600;
 665         args.mda_si_drv1 = periph;
 666         args.mda_flags = MAKEDEV_NOWAIT;
 667         error = make_dev_s(&args, &softc->dev, "%s%d", periph->periph_name,
 668             periph->unit_number);
 669         if (error != 0) {
 670                 cam_periph_lock(periph);
 671                 cam_periph_release_locked(periph);
 672                 return (CAM_REQ_CMP_ERR);
 673         }
 674
 675         /*
 676          * Hold a reference to the periph before we create the physical
 677          * path alias so it can't go away.
 678          */
 679         if (cam_periph_acquire(periph) != 0) {
 680                 xpt_print(periph->path, "%s: lost periph during "
 681                           "registration!\n", __func__);
 682                 cam_periph_lock(periph);
 683                 return (CAM_REQ_CMP_ERR);
 684         }
 685
 686         cam_periph_lock(periph);
 687
 688         TASK_INIT(&softc->add_physpath_task, /*priority*/0,
 689                   pass_add_physpath, periph);
 690
 691         /*
 692          * See if physical path information is already available.
 693          */
 694         taskqueue_enqueue(taskqueue_thread, &softc->add_physpath_task);
 695
 696         /*
 697          * Add an async callback so that we get notified if
 698          * this device goes away or its physical path
 699          * (stored in the advanced info data of the EDT) has
 700          * changed.
 701          */
 702         xpt_register_async(AC_LOST_DEVICE | AC_ADVINFO_CHANGED,
 703                            passasync, periph, periph->path);
 704
 705         if (bootverbose)
 706                 xpt_announce_periph(periph, NULL);
 707
 708         return(CAM_REQ_CMP);
 709 }
 710
 711 static int
 712 passopen(struct cdev *dev, int flags, int fmt, struct thread *td)
 713 {
 714         struct cam_periph *periph;
 715         struct pass_softc *softc;
 716         int error;
 717
 718         periph = (struct cam_periph *)dev->si_drv1;
 719         if (cam_periph_acquire(periph) != 0)
 720                 return (ENXIO);
 721
 722         cam_periph_lock(periph);
 723
 724         softc = (struct pass_softc *)periph->softc;
 725
 726         if (softc->flags & PASS_FLAG_INVALID) {
 727                 cam_periph_release_locked(periph);
 728                 cam_periph_unlock(periph);
 729                 return(ENXIO);
 730         }
 731
 732         /*
 733          * Don't allow access when we're running at a high securelevel.
 734          */
 735         error = securelevel_gt(td->td_ucred, 1);
 736         if (error) {
 737                 cam_periph_release_locked(periph);
 738                 cam_periph_unlock(periph);
 739                 return(error);
 740         }
 741
 742         /*
 743          * Only allow read-write access.
 744          */
 745         if (((flags & FWRITE) == 0) || ((flags & FREAD) == 0)) {
 746                 cam_periph_release_locked(periph);
 747                 cam_periph_unlock(periph);
 748                 return(EPERM);
 749         }
 750
 751         /*
 752          * We don't allow nonblocking access.
 753          */
 754         if ((flags & O_NONBLOCK) != 0) {
 755                 xpt_print(periph->path, "can't do nonblocking access\n");
 756                 cam_periph_release_locked(periph);
 757                 cam_periph_unlock(periph);
 758                 return(EINVAL);
 759         }
 760
 761         softc->open_count++;
 762
 763         cam_periph_unlock(periph);
 764
 765         return (error);
 766 }
 767
 768 static int
 769 passclose(struct cdev *dev, int flag, int fmt, struct thread *td)
 770 {
 771         struct  cam_periph *periph;
 772         struct  pass_softc *softc;
 773         struct mtx *mtx;
 774
 775         periph = (struct cam_periph *)dev->si_drv1;
 776         mtx = cam_periph_mtx(periph);
 777         mtx_lock(mtx);
 778
 779         softc = periph->softc;
 780         softc->open_count--;
 781
 782         if (softc->open_count == 0) {
 783                 struct pass_io_req *io_req, *io_req2;
 784
 785                 TAILQ_FOREACH_SAFE(io_req, &softc->done_queue, links, io_req2) {
 786                         TAILQ_REMOVE(&softc->done_queue, io_req, links);
 787                         passiocleanup(softc, io_req);
 788                         uma_zfree(softc->pass_zone, io_req);
 789                 }
 790
 791                 TAILQ_FOREACH_SAFE(io_req, &softc->incoming_queue, links,
 792                                    io_req2) {
 793                         TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
 794                         passiocleanup(softc, io_req);
 795                         uma_zfree(softc->pass_zone, io_req);
 796                 }
 797
 798                 /*
 799                  * If there are any active I/Os, we need to forcibly acquire a
 800                  * reference to the peripheral so that we don't go away
 801                  * before they complete.  We'll release the reference when
 802                  * the abandoned queue is empty.
 803                  */
 804                 io_req = TAILQ_FIRST(&softc->active_queue);
 805                 if ((io_req != NULL)
 806                  && (softc->flags & PASS_FLAG_ABANDONED_REF_SET) == 0) {
 807                         cam_periph_doacquire(periph);
 808                         softc->flags |= PASS_FLAG_ABANDONED_REF_SET;
 809                 }
 810
 811                 /*
 812                  * Since the I/O in the active queue is not under our
 813                  * control, just set a flag so that we can clean it up when
 814                  * it completes and put it on the abandoned queue.  This
 815                  * will prevent our sending spurious completions in the
 816                  * event that the device is opened again before these I/Os
 817                  * complete.
 818                  */
 819                 TAILQ_FOREACH_SAFE(io_req, &softc->active_queue, links,
 820                                    io_req2) {
 821                         TAILQ_REMOVE(&softc->active_queue, io_req, links);
 822                         io_req->flags |= PASS_IO_ABANDONED;
 823                         TAILQ_INSERT_TAIL(&softc->abandoned_queue, io_req,
 824                                           links);
 825                 }
 826         }
 827
 828         cam_periph_release_locked(periph);
 829
 830         /*
 831          * We reference the lock directly here, instead of using
 832          * cam_periph_unlock().  The reason is that the call to
 833          * cam_periph_release_locked() above could result in the periph
 834          * getting freed.  If that is the case, dereferencing the periph
 835          * with a cam_periph_unlock() call would cause a page fault.
 836          *
 837          * cam_periph_release() avoids this problem using the same method,
 838          * but we're manually acquiring and dropping the lock here to
 839          * protect the open count and avoid another lock acquisition and
 840          * release.
 841          */
 842         mtx_unlock(mtx);
 843
 844         return (0);
 845 }
 846
 847 static void
 848 passstart(struct cam_periph *periph, union ccb *start_ccb)
 849 {
 850         struct pass_softc *softc;
 851
 852         softc = (struct pass_softc *)periph->softc;
 853
 854         switch (softc->state) {
 855         case PASS_STATE_NORMAL: {
 856                 struct pass_io_req *io_req;
 857
 858                 /*
 859                  * Check for any queued I/O requests that require an
 860                  * allocated slot.
 861                  */
 862                 io_req = TAILQ_FIRST(&softc->incoming_queue);
 863                 if (io_req == NULL) {
 864                         xpt_release_ccb(start_ccb);
 865                         break;
 866                 }
 867                 TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
 868                 TAILQ_INSERT_TAIL(&softc->active_queue, io_req, links);
 869                 /*
 870                  * Merge the user's CCB into the allocated CCB.
 871                  */
 872                 xpt_merge_ccb(start_ccb, &io_req->ccb);
 873                 start_ccb->ccb_h.ccb_type = PASS_CCB_QUEUED_IO;
 874                 start_ccb->ccb_h.ccb_ioreq = io_req;
 875                 start_ccb->ccb_h.cbfcnp = passdone;
 876                 io_req->alloced_ccb = start_ccb;
 877                 binuptime(&io_req->start_time);
 878                 devstat_start_transaction(softc->device_stats,
 879                                           &io_req->start_time);
 880
 881                 xpt_action(start_ccb);
 882
 883                 /*
 884                  * If we have any more I/O waiting, schedule ourselves again.
 885                  */
 886                 if (!TAILQ_EMPTY(&softc->incoming_queue))
 887                         xpt_schedule(periph, CAM_PRIORITY_NORMAL);
 888                 break;
 889         }
 890         default:
 891                 break;
 892         }
 893 }
 894
 895 static void
 896 passdone(struct cam_periph *periph, union ccb *done_ccb)
 897 {
 898         struct pass_softc *softc;
 899         struct ccb_scsiio *csio;
 900
 901         softc = (struct pass_softc *)periph->softc;
 902
 903         cam_periph_assert(periph, MA_OWNED);
 904
 905         csio = &done_ccb->csio;
 906         switch (csio->ccb_h.ccb_type) {
 907         case PASS_CCB_QUEUED_IO: {
 908                 struct pass_io_req *io_req;
 909
 910                 io_req = done_ccb->ccb_h.ccb_ioreq;
 911 #if 0
 912                 xpt_print(periph->path, "%s: called for user CCB %p\n",
 913                           __func__, io_req->user_ccb_ptr);
 914 #endif
 915                 if (((done_ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_CMP) &&
 916                     ((io_req->flags & PASS_IO_ABANDONED) == 0)) {
 917                         int error;
 918                         uint32_t cam_flags, sense_flags;
 919
 920                         passflags(done_ccb, &cam_flags, &sense_flags);
 921                         error = passerror(done_ccb, cam_flags, sense_flags);
 922
 923                         if (error == ERESTART) {
 924                                 KASSERT(((sense_flags & SF_NO_RETRY) == 0),
 925                                     ("passerror returned ERESTART with no retry requested\n"));
 926                                 return;
 927                         }
 928                 }
 929
 930                 /*
 931                  * Copy the allocated CCB contents back to the malloced CCB
 932                  * so we can give status back to the user when he requests it.
 933                  */
 934                 bcopy(done_ccb, &io_req->ccb, sizeof(*done_ccb));
 935
 936                 /*
 937                  * Log data/transaction completion with devstat(9).
 938                  */
 939                 switch (done_ccb->ccb_h.func_code) {
 940                 case XPT_SCSI_IO:
 941                         devstat_end_transaction(softc->device_stats,
 942                             done_ccb->csio.dxfer_len - done_ccb->csio.resid,
 943                             done_ccb->csio.tag_action & 0x3,
 944                             ((done_ccb->ccb_h.flags & CAM_DIR_MASK) ==
 945                             CAM_DIR_NONE) ? DEVSTAT_NO_DATA :
 946                             (done_ccb->ccb_h.flags & CAM_DIR_OUT) ?
 947                             DEVSTAT_WRITE : DEVSTAT_READ, NULL,
 948                             &io_req->start_time);
 949                         break;
 950                 case XPT_ATA_IO:
 951                         devstat_end_transaction(softc->device_stats,
 952                             done_ccb->ataio.dxfer_len - done_ccb->ataio.resid,
 953                             0, /* Not used in ATA */
 954                             ((done_ccb->ccb_h.flags & CAM_DIR_MASK) ==
 955                             CAM_DIR_NONE) ? DEVSTAT_NO_DATA :
 956                             (done_ccb->ccb_h.flags & CAM_DIR_OUT) ?
 957                             DEVSTAT_WRITE : DEVSTAT_READ, NULL,
 958                             &io_req->start_time);
 959                         break;
 960                 case XPT_SMP_IO:
 961                         /*
 962                          * XXX KDM this isn't quite right, but there isn't
 963                          * currently an easy way to represent a bidirectional
 964                          * transfer in devstat.  The only way to do it
 965                          * and have the byte counts come out right would
 966                          * mean that we would have to record two
 967                          * transactions, one for the request and one for the
 968                          * response.  For now, so that we report something,
 969                          * just treat the entire thing as a read.
 970                          */
 971                         devstat_end_transaction(softc->device_stats,
 972                             done_ccb->smpio.smp_request_len +
 973                             done_ccb->smpio.smp_response_len,
 974                             DEVSTAT_TAG_SIMPLE, DEVSTAT_READ, NULL,
 975                             &io_req->start_time);
 976                         break;
 977                 default:
 978                         devstat_end_transaction(softc->device_stats, 0,
 979                             DEVSTAT_TAG_NONE, DEVSTAT_NO_DATA, NULL,
 980                             &io_req->start_time);
 981                         break;
 982                 }
 983
 984                 /*
 985                  * In the normal case, take the completed I/O off of the
 986                  * active queue and put it on the done queue.  Notitfy the
 987                  * user that we have a completed I/O.
 988                  */
 989                 if ((io_req->flags & PASS_IO_ABANDONED) == 0) {
 990                         TAILQ_REMOVE(&softc->active_queue, io_req, links);
 991                         TAILQ_INSERT_TAIL(&softc->done_queue, io_req, links);
 992                         selwakeuppri(&softc->read_select, PRIBIO);
 993                         KNOTE_LOCKED(&softc->read_select.si_note, 0);
 994                 } else {
 995                         /*
 996                          * In the case of an abandoned I/O (final close
 997                          * without fetching the I/O), take it off of the
 998                          * abandoned queue and free it.
 999                          */
1000                         TAILQ_REMOVE(&softc->abandoned_queue, io_req, links);
1001                         passiocleanup(softc, io_req);
1002                         uma_zfree(softc->pass_zone, io_req);
1003
1004                         /*
1005                          * Release the done_ccb here, since we may wind up
1006                          * freeing the peripheral when we decrement the
1007                          * reference count below.
1008                          */
1009                         xpt_release_ccb(done_ccb);
1010
1011                         /*
1012                          * If the abandoned queue is empty, we can release
1013                          * our reference to the periph since we won't have
1014                          * any more completions coming.
1015                          */
1016                         if ((TAILQ_EMPTY(&softc->abandoned_queue))
1017                          && (softc->flags & PASS_FLAG_ABANDONED_REF_SET)) {
1018                                 softc->flags &= ~PASS_FLAG_ABANDONED_REF_SET;
1019                                 cam_periph_release_locked(periph);
1020                         }
1021
1022                         /*
1023                          * We have already released the CCB, so we can
1024                          * return.
1025                          */
1026                         return;
1027                 }
1028                 break;
1029         }
1030         }
1031         xpt_release_ccb(done_ccb);
1032 }
1033
1034 static int
1035 passcreatezone(struct cam_periph *periph)
1036 {
1037         struct pass_softc *softc;
1038         int error;
1039
1040         error = 0;
1041         softc = (struct pass_softc *)periph->softc;
1042
1043         cam_periph_assert(periph, MA_OWNED);
1044         KASSERT(((softc->flags & PASS_FLAG_ZONE_VALID) == 0),
1045                 ("%s called when the pass(4) zone is valid!\n", __func__));
1046         KASSERT((softc->pass_zone == NULL),
1047                 ("%s called when the pass(4) zone is allocated!\n", __func__));
1048
1049         if ((softc->flags & PASS_FLAG_ZONE_INPROG) == 0) {
1050                 /*
1051                  * We're the first context through, so we need to create
1052                  * the pass(4) UMA zone for I/O requests.
1053                  */
1054                 softc->flags |= PASS_FLAG_ZONE_INPROG;
1055
1056                 /*
1057                  * uma_zcreate() does a blocking (M_WAITOK) allocation,
1058                  * so we cannot hold a mutex while we call it.
1059                  */
1060                 cam_periph_unlock(periph);
1061
1062                 softc->pass_zone = uma_zcreate(softc->zone_name,
1063                     sizeof(struct pass_io_req), NULL, NULL, NULL, NULL,
1064                     /*align*/ 0, /*flags*/ 0);
1065
1066                 softc->pass_io_zone = uma_zcreate(softc->io_zone_name,
1067                     softc->io_zone_size, NULL, NULL, NULL, NULL,
1068                     /*align*/ 0, /*flags*/ 0);
1069
1070                 cam_periph_lock(periph);
1071
1072                 if ((softc->pass_zone == NULL)
1073                  || (softc->pass_io_zone == NULL)) {
1074                         if (softc->pass_zone == NULL)
1075                                 xpt_print(periph->path, "unable to allocate "
1076                                     "IO Req UMA zone\n");
1077                         else
1078                                 xpt_print(periph->path, "unable to allocate "
1079                                     "IO UMA zone\n");
1080                         softc->flags &= ~PASS_FLAG_ZONE_INPROG;
1081                         goto bailout;
1082                 }
1083
1084                 /*
1085                  * Set the flags appropriately and notify any other waiters.
1086                  */
1087                 softc->flags &= ~PASS_FLAG_ZONE_INPROG;
1088                 softc->flags |= PASS_FLAG_ZONE_VALID;
1089                 wakeup(&softc->pass_zone);
1090         } else {
1091                 /*
1092                  * In this case, the UMA zone has not yet been created, but
1093                  * another context is in the process of creating it.  We
1094                  * need to sleep until the creation is either done or has
1095                  * failed.
1096                  */
1097                 while ((softc->flags & PASS_FLAG_ZONE_INPROG)
1098                     && ((softc->flags & PASS_FLAG_ZONE_VALID) == 0)) {
1099                         error = msleep(&softc->pass_zone,
1100                                        cam_periph_mtx(periph), PRIBIO,
1101                                        "paszon", 0);
1102                         if (error != 0)
1103                                 goto bailout;
1104                 }
1105                 /*
1106                  * If the zone creation failed, no luck for the user.
1107                  */
1108                 if ((softc->flags & PASS_FLAG_ZONE_VALID) == 0){
1109                         error = ENOMEM;
1110                         goto bailout;
1111                 }
1112         }
1113 bailout:
1114         return (error);
1115 }
1116
1117 static void
1118 passiocleanup(struct pass_softc *softc, struct pass_io_req *io_req)
1119 {
1120         union ccb *ccb;
1121         uint8_t **data_ptrs[CAM_PERIPH_MAXMAPS];
1122         int i, numbufs;
1123
1124         ccb = &io_req->ccb;
1125
1126         switch (ccb->ccb_h.func_code) {
1127         case XPT_DEV_MATCH:
1128                 numbufs = min(io_req->num_bufs, 2);
1129
1130                 if (numbufs == 1) {
1131                         data_ptrs[0] = (uint8_t **)&ccb->cdm.matches;
1132                 } else {
1133                         data_ptrs[0] = (uint8_t **)&ccb->cdm.patterns;
1134                         data_ptrs[1] = (uint8_t **)&ccb->cdm.matches;
1135                 }
1136                 break;
1137         case XPT_SCSI_IO:
1138         case XPT_CONT_TARGET_IO:
1139                 data_ptrs[0] = &ccb->csio.data_ptr;
1140                 numbufs = min(io_req->num_bufs, 1);
1141                 break;
1142         case XPT_ATA_IO:
1143                 data_ptrs[0] = &ccb->ataio.data_ptr;
1144                 numbufs = min(io_req->num_bufs, 1);
1145                 break;
1146         case XPT_SMP_IO:
1147                 numbufs = min(io_req->num_bufs, 2);
1148                 data_ptrs[0] = &ccb->smpio.smp_request;
1149                 data_ptrs[1] = &ccb->smpio.smp_response;
1150                 break;
1151         case XPT_DEV_ADVINFO:
1152                 numbufs = min(io_req->num_bufs, 1);
1153                 data_ptrs[0] = (uint8_t **)&ccb->cdai.buf;
1154                 break;
1155         case XPT_NVME_IO:
1156         case XPT_NVME_ADMIN:
1157                 data_ptrs[0] = &ccb->nvmeio.data_ptr;
1158                 numbufs = min(io_req->num_bufs, 1);
1159                 break;
1160         default:
1161                 /* allow ourselves to be swapped once again */
1162                 return;
1163                 break; /* NOTREACHED */
1164         }
1165
1166         if (io_req->flags & PASS_IO_USER_SEG_MALLOC) {
1167                 free(io_req->user_segptr, M_SCSIPASS);
1168                 io_req->user_segptr = NULL;
1169         }
1170
1171         /*
1172          * We only want to free memory we malloced.
1173          */
1174         if (io_req->data_flags == CAM_DATA_VADDR) {
1175                 for (i = 0; i < io_req->num_bufs; i++) {
1176                         if (io_req->kern_bufs[i] == NULL)
1177                                 continue;
1178
1179                         free(io_req->kern_bufs[i], M_SCSIPASS);
1180                         io_req->kern_bufs[i] = NULL;
1181                 }
1182         } else if (io_req->data_flags == CAM_DATA_SG) {
1183                 for (i = 0; i < io_req->num_kern_segs; i++) {
1184                         if ((uint8_t *)(uintptr_t)
1185                             io_req->kern_segptr[i].ds_addr == NULL)
1186                                 continue;
1187
1188                         uma_zfree(softc->pass_io_zone, (uint8_t *)(uintptr_t)
1189                             io_req->kern_segptr[i].ds_addr);
1190                         io_req->kern_segptr[i].ds_addr = 0;
1191                 }
1192         }
1193
1194         if (io_req->flags & PASS_IO_KERN_SEG_MALLOC) {
1195                 free(io_req->kern_segptr, M_SCSIPASS);
1196                 io_req->kern_segptr = NULL;
1197         }
1198
1199         if (io_req->data_flags != CAM_DATA_PADDR) {
1200                 for (i = 0; i < numbufs; i++) {
1201                         /*
1202                          * Restore the user's buffer pointers to their
1203                          * previous values.
1204                          */
1205                         if (io_req->user_bufs[i] != NULL)
1206                                 *data_ptrs[i] = io_req->user_bufs[i];
1207                 }
1208         }
1209
1210 }
1211
1212 static int
1213 passcopysglist(struct cam_periph *periph, struct pass_io_req *io_req,
1214                ccb_flags direction)
1215 {
1216         bus_size_t kern_watermark, user_watermark, len_to_copy;
1217         bus_dma_segment_t *user_sglist, *kern_sglist;
1218         int i, j, error;
1219
1220         error = 0;
1221         kern_watermark = 0;
1222         user_watermark = 0;
1223         len_to_copy = 0;
1224         user_sglist = io_req->user_segptr;
1225         kern_sglist = io_req->kern_segptr;
1226
1227         for (i = 0, j = 0; i < io_req->num_user_segs &&
1228              j < io_req->num_kern_segs;) {
1229                 uint8_t *user_ptr, *kern_ptr;
1230
1231                 len_to_copy = min(user_sglist[i].ds_len -user_watermark,
1232                     kern_sglist[j].ds_len - kern_watermark);
1233
1234                 user_ptr = (uint8_t *)(uintptr_t)user_sglist[i].ds_addr;
1235                 user_ptr = user_ptr + user_watermark;
1236                 kern_ptr = (uint8_t *)(uintptr_t)kern_sglist[j].ds_addr;
1237                 kern_ptr = kern_ptr + kern_watermark;
1238
1239                 user_watermark += len_to_copy;
1240                 kern_watermark += len_to_copy;
1241
1242                 if (direction == CAM_DIR_IN) {
1243                         error = copyout(kern_ptr, user_ptr, len_to_copy);
1244                         if (error != 0) {
1245                                 xpt_print(periph->path, "%s: copyout of %u "
1246                                           "bytes from %p to %p failed with "
1247                                           "error %d\n", __func__, len_to_copy,
1248                                           kern_ptr, user_ptr, error);
1249                                 goto bailout;
1250                         }
1251                 } else {
1252                         error = copyin(user_ptr, kern_ptr, len_to_copy);
1253                         if (error != 0) {
1254                                 xpt_print(periph->path, "%s: copyin of %u "
1255                                           "bytes from %p to %p failed with "
1256                                           "error %d\n", __func__, len_to_copy,
1257                                           user_ptr, kern_ptr, error);
1258                                 goto bailout;
1259                         }
1260                 }
1261
1262                 if (user_sglist[i].ds_len == user_watermark) {
1263                         i++;
1264                         user_watermark = 0;
1265                 }
1266
1267                 if (kern_sglist[j].ds_len == kern_watermark) {
1268                         j++;
1269                         kern_watermark = 0;
1270                 }
1271         }
1272
1273 bailout:
1274
1275         return (error);
1276 }
1277
1278 static int
1279 passmemsetup(struct cam_periph *periph, struct pass_io_req *io_req)
1280 {
1281         union ccb *ccb;
1282         struct pass_softc *softc;
1283         int numbufs, i;
1284         uint8_t **data_ptrs[CAM_PERIPH_MAXMAPS];
1285         uint32_t lengths[CAM_PERIPH_MAXMAPS];
1286         uint32_t dirs[CAM_PERIPH_MAXMAPS];
1287         uint32_t num_segs;
1288         uint16_t *seg_cnt_ptr;
1289         size_t maxmap;
1290         int error;
1291
1292         cam_periph_assert(periph, MA_NOTOWNED);
1293
1294         softc = periph->softc;
1295
1296         error = 0;
1297         ccb = &io_req->ccb;
1298         maxmap = 0;
1299         num_segs = 0;
1300         seg_cnt_ptr = NULL;
1301
1302         switch(ccb->ccb_h.func_code) {
1303         case XPT_DEV_MATCH:
1304                 if (ccb->cdm.match_buf_len == 0) {
1305                         printf("%s: invalid match buffer length 0\n", __func__);
1306                         return(EINVAL);
1307                 }
1308                 if (ccb->cdm.pattern_buf_len > 0) {
1309                         data_ptrs[0] = (uint8_t **)&ccb->cdm.patterns;
1310                         lengths[0] = ccb->cdm.pattern_buf_len;
1311                         dirs[0] = CAM_DIR_OUT;
1312                         data_ptrs[1] = (uint8_t **)&ccb->cdm.matches;
1313                         lengths[1] = ccb->cdm.match_buf_len;
1314                         dirs[1] = CAM_DIR_IN;
1315                         numbufs = 2;
1316                 } else {
1317                         data_ptrs[0] = (uint8_t **)&ccb->cdm.matches;
1318                         lengths[0] = ccb->cdm.match_buf_len;
1319                         dirs[0] = CAM_DIR_IN;
1320                         numbufs = 1;
1321                 }
1322                 io_req->data_flags = CAM_DATA_VADDR;
1323                 break;
1324         case XPT_SCSI_IO:
1325         case XPT_CONT_TARGET_IO:
1326                 if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
1327                         return(0);
1328
1329                 /*
1330                  * The user shouldn't be able to supply a bio.
1331                  */
1332                 if ((ccb->ccb_h.flags & CAM_DATA_MASK) == CAM_DATA_BIO)
1333                         return (EINVAL);
1334
1335                 io_req->data_flags = ccb->ccb_h.flags & CAM_DATA_MASK;
1336
1337                 data_ptrs[0] = &ccb->csio.data_ptr;
1338                 lengths[0] = ccb->csio.dxfer_len;
1339                 dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1340                 num_segs = ccb->csio.sglist_cnt;
1341                 seg_cnt_ptr = &ccb->csio.sglist_cnt;
1342                 numbufs = 1;
1343                 maxmap = softc->maxio;
1344                 break;
1345         case XPT_ATA_IO:
1346                 if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
1347                         return(0);
1348
1349                 /*
1350                  * We only support a single virtual address for ATA I/O.
1351                  */
1352                 if ((ccb->ccb_h.flags & CAM_DATA_MASK) != CAM_DATA_VADDR)
1353                         return (EINVAL);
1354
1355                 io_req->data_flags = CAM_DATA_VADDR;
1356
1357                 data_ptrs[0] = &ccb->ataio.data_ptr;
1358                 lengths[0] = ccb->ataio.dxfer_len;
1359                 dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1360                 numbufs = 1;
1361                 maxmap = softc->maxio;
1362                 break;
1363         case XPT_SMP_IO:
1364                 io_req->data_flags = CAM_DATA_VADDR;
1365
1366                 data_ptrs[0] = &ccb->smpio.smp_request;
1367                 lengths[0] = ccb->smpio.smp_request_len;
1368                 dirs[0] = CAM_DIR_OUT;
1369                 data_ptrs[1] = &ccb->smpio.smp_response;
1370                 lengths[1] = ccb->smpio.smp_response_len;
1371                 dirs[1] = CAM_DIR_IN;
1372                 numbufs = 2;
1373                 maxmap = softc->maxio;
1374                 break;
1375         case XPT_DEV_ADVINFO:
1376                 if (ccb->cdai.bufsiz == 0)
1377                         return (0);
1378
1379                 io_req->data_flags = CAM_DATA_VADDR;
1380
1381                 data_ptrs[0] = (uint8_t **)&ccb->cdai.buf;
1382                 lengths[0] = ccb->cdai.bufsiz;
1383                 dirs[0] = CAM_DIR_IN;
1384                 numbufs = 1;
1385                 break;
1386         case XPT_NVME_ADMIN:
1387         case XPT_NVME_IO:
1388                 if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
1389                         return (0);
1390
1391                 io_req->data_flags = ccb->ccb_h.flags & CAM_DATA_MASK;
1392
1393                 data_ptrs[0] = &ccb->nvmeio.data_ptr;
1394                 lengths[0] = ccb->nvmeio.dxfer_len;
1395                 dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1396                 num_segs = ccb->nvmeio.sglist_cnt;
1397                 seg_cnt_ptr = &ccb->nvmeio.sglist_cnt;
1398                 numbufs = 1;
1399                 maxmap = softc->maxio;
1400                 break;
1401         default:
1402                 return(EINVAL);
1403                 break; /* NOTREACHED */
1404         }
1405
1406         io_req->num_bufs = numbufs;
1407
1408         /*
1409          * If there is a maximum, check to make sure that the user's
1410          * request fits within the limit.  In general, we should only have
1411          * a maximum length for requests that go to hardware.  Otherwise it
1412          * is whatever we're able to malloc.
1413          */
1414         for (i = 0; i < numbufs; i++) {
1415                 io_req->user_bufs[i] = *data_ptrs[i];
1416                 io_req->dirs[i] = dirs[i];
1417                 io_req->lengths[i] = lengths[i];
1418
1419                 if (maxmap == 0)
1420                         continue;
1421
1422                 if (lengths[i] <= maxmap)
1423                         continue;
1424
1425                 xpt_print(periph->path, "%s: data length %u > max allowed %u "
1426                           "bytes\n", __func__, lengths[i], maxmap);
1427                 error = EINVAL;
1428                 goto bailout;
1429         }
1430
1431         switch (io_req->data_flags) {
1432         case CAM_DATA_VADDR:
1433                 /* Map or copy the buffer into kernel address space */
1434                 for (i = 0; i < numbufs; i++) {
1435                         uint8_t *tmp_buf;
1436
1437                         /*
1438                          * If for some reason no length is specified, we
1439                          * don't need to allocate anything.
1440                          */
1441                         if (io_req->lengths[i] == 0)
1442                                 continue;
1443
1444                         tmp_buf = malloc(lengths[i], M_SCSIPASS,
1445                                          M_WAITOK | M_ZERO);
1446                         io_req->kern_bufs[i] = tmp_buf;
1447                         *data_ptrs[i] = tmp_buf;
1448
1449 #if 0
1450                         xpt_print(periph->path, "%s: malloced %p len %u, user "
1451                                   "buffer %p, operation: %s\n", __func__,
1452                                   tmp_buf, lengths[i], io_req->user_bufs[i],
1453                                   (dirs[i] == CAM_DIR_IN) ? "read" : "write");
1454 #endif
1455                         /*
1456                          * We only need to copy in if the user is writing.
1457                          */
1458                         if (dirs[i] != CAM_DIR_OUT)
1459                                 continue;
1460
1461                         error = copyin(io_req->user_bufs[i],
1462                                        io_req->kern_bufs[i], lengths[i]);
1463                         if (error != 0) {
1464                                 xpt_print(periph->path, "%s: copy of user "
1465                                           "buffer from %p to %p failed with "
1466                                           "error %d\n", __func__,
1467                                           io_req->user_bufs[i],
1468                                           io_req->kern_bufs[i], error);
1469                                 goto bailout;
1470                         }
1471                 }
1472                 break;
1473         case CAM_DATA_PADDR:
1474                 /* Pass down the pointer as-is */
1475                 break;
1476         case CAM_DATA_SG: {
1477                 size_t sg_length, size_to_go, alloc_size;
1478                 uint32_t num_segs_needed;
1479
1480                 /*
1481                  * Copy the user S/G list in, and then copy in the
1482                  * individual segments.
1483                  */
1484                 /*
1485                  * We shouldn't see this, but check just in case.
1486                  */
1487                 if (numbufs != 1) {
1488                         xpt_print(periph->path, "%s: cannot currently handle "
1489                                   "more than one S/G list per CCB\n", __func__);
1490                         error = EINVAL;
1491                         goto bailout;
1492                 }
1493
1494                 /*
1495                  * We have to have at least one segment.
1496                  */
1497                 if (num_segs == 0) {
1498                         xpt_print(periph->path, "%s: CAM_DATA_SG flag set, "
1499                                   "but sglist_cnt=0!\n", __func__);
1500                         error = EINVAL;
1501                         goto bailout;
1502                 }
1503
1504                 /*
1505                  * Make sure the user specified the total length and didn't
1506                  * just leave it to us to decode the S/G list.
1507                  */
1508                 if (lengths[0] == 0) {
1509                         xpt_print(periph->path, "%s: no dxfer_len specified, "
1510                                   "but CAM_DATA_SG flag is set!\n", __func__);
1511                         error = EINVAL;
1512                         goto bailout;
1513                 }
1514
1515                 /*
1516                  * We allocate buffers in io_zone_size increments for an
1517                  * S/G list.  This will generally be maxphys.
1518                  */
1519                 if (lengths[0] <= softc->io_zone_size)
1520                         num_segs_needed = 1;
1521                 else {
1522                         num_segs_needed = lengths[0] / softc->io_zone_size;
1523                         if ((lengths[0] % softc->io_zone_size) != 0)
1524                                 num_segs_needed++;
1525                 }
1526
1527                 /* Figure out the size of the S/G list */
1528                 sg_length = num_segs * sizeof(bus_dma_segment_t);
1529                 io_req->num_user_segs = num_segs;
1530                 io_req->num_kern_segs = num_segs_needed;
1531
1532                 /* Save the user's S/G list pointer for later restoration */
1533                 io_req->user_bufs[0] = *data_ptrs[0];
1534
1535                 /*
1536                  * If we have enough segments allocated by default to handle
1537                  * the length of the user's S/G list,
1538                  */
1539                 if (num_segs > PASS_MAX_SEGS) {
1540                         io_req->user_segptr = malloc(sizeof(bus_dma_segment_t) *
1541                             num_segs, M_SCSIPASS, M_WAITOK | M_ZERO);
1542                         io_req->flags |= PASS_IO_USER_SEG_MALLOC;
1543                 } else
1544                         io_req->user_segptr = io_req->user_segs;
1545
1546                 error = copyin(*data_ptrs[0], io_req->user_segptr, sg_length);
1547                 if (error != 0) {
1548                         xpt_print(periph->path, "%s: copy of user S/G list "
1549                                   "from %p to %p failed with error %d\n",
1550                                   __func__, *data_ptrs[0], io_req->user_segptr,
1551                                   error);
1552                         goto bailout;
1553                 }
1554
1555                 if (num_segs_needed > PASS_MAX_SEGS) {
1556                         io_req->kern_segptr = malloc(sizeof(bus_dma_segment_t) *
1557                             num_segs_needed, M_SCSIPASS, M_WAITOK | M_ZERO);
1558                         io_req->flags |= PASS_IO_KERN_SEG_MALLOC;
1559                 } else {
1560                         io_req->kern_segptr = io_req->kern_segs;
1561                 }
1562
1563                 /*
1564                  * Allocate the kernel S/G list.
1565                  */
1566                 for (size_to_go = lengths[0], i = 0;
1567                      size_to_go > 0 && i < num_segs_needed;
1568                      i++, size_to_go -= alloc_size) {
1569                         uint8_t *kern_ptr;
1570
1571                         alloc_size = min(size_to_go, softc->io_zone_size);
1572                         kern_ptr = uma_zalloc(softc->pass_io_zone, M_WAITOK);
1573                         io_req->kern_segptr[i].ds_addr =
1574                             (bus_addr_t)(uintptr_t)kern_ptr;
1575                         io_req->kern_segptr[i].ds_len = alloc_size;
1576                 }
1577                 if (size_to_go > 0) {
1578                         printf("%s: size_to_go = %zu, software error!\n",
1579                                __func__, size_to_go);
1580                         error = EINVAL;
1581                         goto bailout;
1582                 }
1583
1584                 *data_ptrs[0] = (uint8_t *)io_req->kern_segptr;
1585                 *seg_cnt_ptr = io_req->num_kern_segs;
1586
1587                 /*
1588                  * We only need to copy data here if the user is writing.
1589                  */
1590                 if (dirs[0] == CAM_DIR_OUT)
1591                         error = passcopysglist(periph, io_req, dirs[0]);
1592                 break;
1593         }
1594         case CAM_DATA_SG_PADDR: {
1595                 size_t sg_length;
1596
1597                 /*
1598                  * We shouldn't see this, but check just in case.
1599                  */
1600                 if (numbufs != 1) {
1601                         printf("%s: cannot currently handle more than one "
1602                                "S/G list per CCB\n", __func__);
1603                         error = EINVAL;
1604                         goto bailout;
1605                 }
1606
1607                 /*
1608                  * We have to have at least one segment.
1609                  */
1610                 if (num_segs == 0) {
1611                         xpt_print(periph->path, "%s: CAM_DATA_SG_PADDR flag "
1612                                   "set, but sglist_cnt=0!\n", __func__);
1613                         error = EINVAL;
1614                         goto bailout;
1615                 }
1616
1617                 /*
1618                  * Make sure the user specified the total length and didn't
1619                  * just leave it to us to decode the S/G list.
1620                  */
1621                 if (lengths[0] == 0) {
1622                         xpt_print(periph->path, "%s: no dxfer_len specified, "
1623                                   "but CAM_DATA_SG flag is set!\n", __func__);
1624                         error = EINVAL;
1625                         goto bailout;
1626                 }
1627
1628                 /* Figure out the size of the S/G list */
1629                 sg_length = num_segs * sizeof(bus_dma_segment_t);
1630                 io_req->num_user_segs = num_segs;
1631                 io_req->num_kern_segs = io_req->num_user_segs;
1632
1633                 /* Save the user's S/G list pointer for later restoration */
1634                 io_req->user_bufs[0] = *data_ptrs[0];
1635
1636                 if (num_segs > PASS_MAX_SEGS) {
1637                         io_req->user_segptr = malloc(sizeof(bus_dma_segment_t) *
1638                             num_segs, M_SCSIPASS, M_WAITOK | M_ZERO);
1639                         io_req->flags |= PASS_IO_USER_SEG_MALLOC;
1640                 } else
1641                         io_req->user_segptr = io_req->user_segs;
1642
1643                 io_req->kern_segptr = io_req->user_segptr;
1644
1645                 error = copyin(*data_ptrs[0], io_req->user_segptr, sg_length);
1646                 if (error != 0) {
1647                         xpt_print(periph->path, "%s: copy of user S/G list "
1648                                   "from %p to %p failed with error %d\n",
1649                                   __func__, *data_ptrs[0], io_req->user_segptr,
1650                                   error);
1651                         goto bailout;
1652                 }
1653                 break;
1654         }
1655         default:
1656         case CAM_DATA_BIO:
1657                 /*
1658                  * A user shouldn't be attaching a bio to the CCB.  It
1659                  * isn't a user-accessible structure.
1660                  */
1661                 error = EINVAL;
1662                 break;
1663         }
1664
1665 bailout:
1666         if (error != 0)
1667                 passiocleanup(softc, io_req);
1668
1669         return (error);
1670 }
1671
1672 static int
1673 passmemdone(struct cam_periph *periph, struct pass_io_req *io_req)
1674 {
1675         struct pass_softc *softc;
1676         int error;
1677         int i;
1678
1679         error = 0;
1680         softc = (struct pass_softc *)periph->softc;
1681
1682         switch (io_req->data_flags) {
1683         case CAM_DATA_VADDR:
1684                 /*
1685                  * Copy back to the user buffer if this was a read.
1686                  */
1687                 for (i = 0; i < io_req->num_bufs; i++) {
1688                         if (io_req->dirs[i] != CAM_DIR_IN)
1689                                 continue;
1690
1691                         error = copyout(io_req->kern_bufs[i],
1692                             io_req->user_bufs[i], io_req->lengths[i]);
1693                         if (error != 0) {
1694                                 xpt_print(periph->path, "Unable to copy %u "
1695                                           "bytes from %p to user address %p\n",
1696                                           io_req->lengths[i],
1697                                           io_req->kern_bufs[i],
1698                                           io_req->user_bufs[i]);
1699                                 goto bailout;
1700                         }
1701                 }
1702                 break;
1703         case CAM_DATA_PADDR:
1704                 /* Do nothing.  The pointer is a physical address already */
1705                 break;
1706         case CAM_DATA_SG:
1707                 /*
1708                  * Copy back to the user buffer if this was a read.
1709                  * Restore the user's S/G list buffer pointer.
1710                  */
1711                 if (io_req->dirs[0] == CAM_DIR_IN)
1712                         error = passcopysglist(periph, io_req, io_req->dirs[0]);
1713                 break;
1714         case CAM_DATA_SG_PADDR:
1715                 /*
1716                  * Restore the user's S/G list buffer pointer.  No need to
1717                  * copy.
1718                  */
1719                 break;
1720         default:
1721         case CAM_DATA_BIO:
1722                 error = EINVAL;
1723                 break;
1724         }
1725
1726 bailout:
1727         /*
1728          * Reset the user's pointers to their original values and free
1729          * allocated memory.
1730          */
1731         passiocleanup(softc, io_req);
1732
1733         return (error);
1734 }
1735
1736 static int
1737 passioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flag, struct thread *td)
1738 {
1739         int error;
1740
1741         if ((error = passdoioctl(dev, cmd, addr, flag, td)) == ENOTTY) {
1742                 error = cam_compat_ioctl(dev, cmd, addr, flag, td, passdoioctl);
1743         }
1744         return (error);
1745 }
1746
1747 static int
1748 passdoioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flag, struct thread *td)
1749 {
1750         struct  cam_periph *periph;
1751         struct  pass_softc *softc;
1752         int     error;
1753         uint32_t priority;
1754
1755         periph = (struct cam_periph *)dev->si_drv1;
1756         cam_periph_lock(periph);
1757         softc = (struct pass_softc *)periph->softc;
1758
1759         error = 0;
1760
1761         switch (cmd) {
1762         case CAMIOCOMMAND:
1763         {
1764                 union ccb *inccb;
1765                 union ccb *ccb;
1766                 int ccb_malloced;
1767
1768                 inccb = (union ccb *)addr;
1769 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
1770                 if (inccb->ccb_h.func_code == XPT_SCSI_IO)
1771                         inccb->csio.bio = NULL;
1772 #endif
1773
1774                 if (inccb->ccb_h.flags & CAM_UNLOCKED) {
1775                         error = EINVAL;
1776                         break;
1777                 }
1778
1779                 /*
1780                  * Some CCB types, like scan bus and scan lun can only go
1781                  * through the transport layer device.
1782                  */
1783                 if (inccb->ccb_h.func_code & XPT_FC_XPT_ONLY) {
1784                         xpt_print(periph->path, "CCB function code %#x is "
1785                             "restricted to the XPT device\n",
1786                             inccb->ccb_h.func_code);
1787                         error = ENODEV;
1788                         break;
1789                 }
1790
1791                 /* Compatibility for RL/priority-unaware code. */
1792                 priority = inccb->ccb_h.pinfo.priority;
1793                 if (priority <= CAM_PRIORITY_OOB)
1794                     priority += CAM_PRIORITY_OOB + 1;
1795
1796                 /*
1797                  * Non-immediate CCBs need a CCB from the per-device pool
1798                  * of CCBs, which is scheduled by the transport layer.
1799                  * Immediate CCBs and user-supplied CCBs should just be
1800                  * malloced.
1801                  */
1802                 if ((inccb->ccb_h.func_code & XPT_FC_QUEUED)
1803                  && ((inccb->ccb_h.func_code & XPT_FC_USER_CCB) == 0)) {
1804                         ccb = cam_periph_getccb(periph, priority);
1805                         ccb_malloced = 0;
1806                 } else {
1807                         ccb = xpt_alloc_ccb_nowait();
1808
1809                         if (ccb != NULL)
1810                                 xpt_setup_ccb(&ccb->ccb_h, periph->path,
1811                                               priority);
1812                         ccb_malloced = 1;
1813                 }
1814
1815                 if (ccb == NULL) {
1816                         xpt_print(periph->path, "unable to allocate CCB\n");
1817                         error = ENOMEM;
1818                         break;
1819                 }
1820
1821                 error = passsendccb(periph, ccb, inccb);
1822
1823                 if (ccb_malloced)
1824                         xpt_free_ccb(ccb);
1825                 else
1826                         xpt_release_ccb(ccb);
1827
1828                 break;
1829         }
1830         case CAMIOQUEUE:
1831         {
1832                 struct pass_io_req *io_req;
1833                 union ccb **user_ccb, *ccb;
1834                 xpt_opcode fc;
1835
1836 #ifdef COMPAT_FREEBSD32
1837                 if (SV_PROC_FLAG(td->td_proc, SV_ILP32)) {
1838                         error = ENOTTY;
1839                         goto bailout;
1840                 }
1841 #endif
1842                 if ((softc->flags & PASS_FLAG_ZONE_VALID) == 0) {
1843                         error = passcreatezone(periph);
1844                         if (error != 0)
1845                                 goto bailout;
1846                 }
1847
1848                 /*
1849                  * We're going to do a blocking allocation for this I/O
1850                  * request, so we have to drop the lock.
1851                  */
1852                 cam_periph_unlock(periph);
1853
1854                 io_req = uma_zalloc(softc->pass_zone, M_WAITOK | M_ZERO);
1855                 ccb = &io_req->ccb;
1856                 user_ccb = (union ccb **)addr;
1857
1858                 /*
1859                  * Unlike the CAMIOCOMMAND ioctl above, we only have a
1860                  * pointer to the user's CCB, so we have to copy the whole
1861                  * thing in to a buffer we have allocated (above) instead
1862                  * of allowing the ioctl code to malloc a buffer and copy
1863                  * it in.
1864                  *
1865                  * This is an advantage for this asynchronous interface,
1866                  * since we don't want the memory to get freed while the
1867                  * CCB is outstanding.
1868                  */
1869 #if 0
1870                 xpt_print(periph->path, "Copying user CCB %p to "
1871                           "kernel address %p\n", *user_ccb, ccb);
1872 #endif
1873                 error = copyin(*user_ccb, ccb, sizeof(*ccb));
1874                 if (error != 0) {
1875                         xpt_print(periph->path, "Copy of user CCB %p to "
1876                                   "kernel address %p failed with error %d\n",
1877                                   *user_ccb, ccb, error);
1878                         goto camioqueue_error;
1879                 }
1880 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
1881                 if (ccb->ccb_h.func_code == XPT_SCSI_IO)
1882                         ccb->csio.bio = NULL;
1883 #endif
1884
1885                 if (ccb->ccb_h.flags & CAM_UNLOCKED) {
1886                         error = EINVAL;
1887                         goto camioqueue_error;
1888                 }
1889
1890                 if (ccb->ccb_h.flags & CAM_CDB_POINTER) {
1891                         if (ccb->csio.cdb_len > IOCDBLEN) {
1892                                 error = EINVAL;
1893                                 goto camioqueue_error;
1894                         }
1895                         error = copyin(ccb->csio.cdb_io.cdb_ptr,
1896                             ccb->csio.cdb_io.cdb_bytes, ccb->csio.cdb_len);
1897                         if (error != 0)
1898                                 goto camioqueue_error;
1899                         ccb->ccb_h.flags &= ~CAM_CDB_POINTER;
1900                 }
1901
1902                 /*
1903                  * Some CCB types, like scan bus and scan lun can only go
1904                  * through the transport layer device.
1905                  */
1906                 if (ccb->ccb_h.func_code & XPT_FC_XPT_ONLY) {
1907                         xpt_print(periph->path, "CCB function code %#x is "
1908                             "restricted to the XPT device\n",
1909                             ccb->ccb_h.func_code);
1910                         error = ENODEV;
1911                         goto camioqueue_error;
1912                 }
1913
1914                 /*
1915                  * Save the user's CCB pointer as well as his linked list
1916                  * pointers and peripheral private area so that we can
1917                  * restore these later.
1918                  */
1919                 io_req->user_ccb_ptr = *user_ccb;
1920                 io_req->user_periph_links = ccb->ccb_h.periph_links;
1921                 io_req->user_periph_priv = ccb->ccb_h.periph_priv;
1922
1923                 /*
1924                  * Now that we've saved the user's values, we can set our
1925                  * own peripheral private entry.
1926                  */
1927                 ccb->ccb_h.ccb_ioreq = io_req;
1928
1929                 /* Compatibility for RL/priority-unaware code. */
1930                 priority = ccb->ccb_h.pinfo.priority;
1931                 if (priority <= CAM_PRIORITY_OOB)
1932                     priority += CAM_PRIORITY_OOB + 1;
1933
1934                 /*
1935                  * Setup fields in the CCB like the path and the priority.
1936                  * The path in particular cannot be done in userland, since
1937                  * it is a pointer to a kernel data structure.
1938                  */
1939                 xpt_setup_ccb_flags(&ccb->ccb_h, periph->path, priority,
1940                                     ccb->ccb_h.flags);
1941
1942                 /*
1943                  * Setup our done routine.  There is no way for the user to
1944                  * have a valid pointer here.
1945                  */
1946                 ccb->ccb_h.cbfcnp = passdone;
1947
1948                 fc = ccb->ccb_h.func_code;
1949                 /*
1950                  * If this function code has memory that can be mapped in
1951                  * or out, we need to call passmemsetup().
1952                  */
1953                 if ((fc == XPT_SCSI_IO) || (fc == XPT_ATA_IO)
1954                  || (fc == XPT_SMP_IO) || (fc == XPT_DEV_MATCH)
1955                  || (fc == XPT_DEV_ADVINFO)
1956                  || (fc == XPT_NVME_ADMIN) || (fc == XPT_NVME_IO)) {
1957                         error = passmemsetup(periph, io_req);
1958                         if (error != 0)
1959                                 goto camioqueue_error;
1960                 } else
1961                         io_req->mapinfo.num_bufs_used = 0;
1962
1963                 cam_periph_lock(periph);
1964
1965                 /*
1966                  * Everything goes on the incoming queue initially.
1967                  */
1968                 TAILQ_INSERT_TAIL(&softc->incoming_queue, io_req, links);
1969
1970                 /*
1971                  * If the CCB is queued, and is not a user CCB, then
1972                  * we need to allocate a slot for it.  Call xpt_schedule()
1973                  * so that our start routine will get called when a CCB is
1974                  * available.
1975                  */
1976                 if ((fc & XPT_FC_QUEUED)
1977                  && ((fc & XPT_FC_USER_CCB) == 0)) {
1978                         xpt_schedule(periph, priority);
1979                         break;
1980                 }
1981
1982                 /*
1983                  * At this point, the CCB in question is either an
1984                  * immediate CCB (like XPT_DEV_ADVINFO) or it is a user CCB
1985                  * and therefore should be malloced, not allocated via a slot.
1986                  * Remove the CCB from the incoming queue and add it to the
1987                  * active queue.
1988                  */
1989                 TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
1990                 TAILQ_INSERT_TAIL(&softc->active_queue, io_req, links);
1991
1992                 xpt_action(ccb);
1993
1994                 /*
1995                  * If this is not a queued CCB (i.e. it is an immediate CCB),
1996                  * then it is already done.  We need to put it on the done
1997                  * queue for the user to fetch.
1998                  */
1999                 if ((fc & XPT_FC_QUEUED) == 0) {
2000                         TAILQ_REMOVE(&softc->active_queue, io_req, links);
2001                         TAILQ_INSERT_TAIL(&softc->done_queue, io_req, links);
2002                 }
2003                 break;
2004
2005 camioqueue_error:
2006                 uma_zfree(softc->pass_zone, io_req);
2007                 cam_periph_lock(periph);
2008                 break;
2009         }
2010         case CAMIOGET:
2011         {
2012                 union ccb **user_ccb;
2013                 struct pass_io_req *io_req;
2014                 int old_error;
2015
2016 #ifdef COMPAT_FREEBSD32
2017                 if (SV_PROC_FLAG(td->td_proc, SV_ILP32)) {
2018                         error = ENOTTY;
2019                         goto bailout;
2020                 }
2021 #endif
2022                 user_ccb = (union ccb **)addr;
2023                 old_error = 0;
2024
2025                 io_req = TAILQ_FIRST(&softc->done_queue);
2026                 if (io_req == NULL) {
2027                         error = ENOENT;
2028                         break;
2029                 }
2030
2031                 /*
2032                  * Remove the I/O from the done queue.
2033                  */
2034                 TAILQ_REMOVE(&softc->done_queue, io_req, links);
2035
2036                 /*
2037                  * We have to drop the lock during the copyout because the
2038                  * copyout can result in VM faults that require sleeping.
2039                  */
2040                 cam_periph_unlock(periph);
2041
2042                 /*
2043                  * Do any needed copies (e.g. for reads) and revert the
2044                  * pointers in the CCB back to the user's pointers.
2045                  */
2046                 error = passmemdone(periph, io_req);
2047
2048                 old_error = error;
2049
2050                 io_req->ccb.ccb_h.periph_links = io_req->user_periph_links;
2051                 io_req->ccb.ccb_h.periph_priv = io_req->user_periph_priv;
2052
2053 #if 0
2054                 xpt_print(periph->path, "Copying to user CCB %p from "
2055                           "kernel address %p\n", *user_ccb, &io_req->ccb);
2056 #endif
2057
2058                 error = copyout(&io_req->ccb, *user_ccb, sizeof(union ccb));
2059                 if (error != 0) {
2060                         xpt_print(periph->path, "Copy to user CCB %p from "
2061                                   "kernel address %p failed with error %d\n",
2062                                   *user_ccb, &io_req->ccb, error);
2063                 }
2064
2065                 /*
2066                  * Prefer the first error we got back, and make sure we
2067                  * don't overwrite bad status with good.
2068                  */
2069                 if (old_error != 0)
2070                         error = old_error;
2071
2072                 cam_periph_lock(periph);
2073
2074                 /*
2075                  * At this point, if there was an error, we could potentially
2076                  * re-queue the I/O and try again.  But why?  The error
2077                  * would almost certainly happen again.  We might as well
2078                  * not leak memory.
2079                  */
2080                 uma_zfree(softc->pass_zone, io_req);
2081                 break;
2082         }
2083         default:
2084                 error = cam_periph_ioctl(periph, cmd, addr, passerror);
2085                 break;
2086         }
2087
2088 bailout:
2089         cam_periph_unlock(periph);
2090
2091         return(error);
2092 }
2093
2094 static int
2095 passpoll(struct cdev *dev, int poll_events, struct thread *td)
2096 {
2097         struct cam_periph *periph;
2098         struct pass_softc *softc;
2099         int revents;
2100
2101         periph = (struct cam_periph *)dev->si_drv1;
2102         softc = (struct pass_softc *)periph->softc;
2103
2104         revents = poll_events & (POLLOUT | POLLWRNORM);
2105         if ((poll_events & (POLLIN | POLLRDNORM)) != 0) {
2106                 cam_periph_lock(periph);
2107
2108                 if (!TAILQ_EMPTY(&softc->done_queue)) {
2109                         revents |= poll_events & (POLLIN | POLLRDNORM);
2110                 }
2111                 cam_periph_unlock(periph);
2112                 if (revents == 0)
2113                         selrecord(td, &softc->read_select);
2114         }
2115
2116         return (revents);
2117 }
2118
2119 static int
2120 passkqfilter(struct cdev *dev, struct knote *kn)
2121 {
2122         struct cam_periph *periph;
2123         struct pass_softc *softc;
2124
2125         periph = (struct cam_periph *)dev->si_drv1;
2126         softc = (struct pass_softc *)periph->softc;
2127
2128         kn->kn_hook = (caddr_t)periph;
2129         kn->kn_fop = &passread_filtops;
2130         knlist_add(&softc->read_select.si_note, kn, 0);
2131
2132         return (0);
2133 }
2134
2135 static void
2136 passreadfiltdetach(struct knote *kn)
2137 {
2138         struct cam_periph *periph;
2139         struct pass_softc *softc;
2140
2141         periph = (struct cam_periph *)kn->kn_hook;
2142         softc = (struct pass_softc *)periph->softc;
2143
2144         knlist_remove(&softc->read_select.si_note, kn, 0);
2145 }
2146
2147 static int
2148 passreadfilt(struct knote *kn, long hint)
2149 {
2150         struct cam_periph *periph;
2151         struct pass_softc *softc;
2152         int retval;
2153
2154         periph = (struct cam_periph *)kn->kn_hook;
2155         softc = (struct pass_softc *)periph->softc;
2156
2157         cam_periph_assert(periph, MA_OWNED);
2158
2159         if (TAILQ_EMPTY(&softc->done_queue))
2160                 retval = 0;
2161         else
2162                 retval = 1;
2163
2164         return (retval);
2165 }
2166
2167 /*
2168  * Generally, "ccb" should be the CCB supplied by the kernel.  "inccb"
2169  * should be the CCB that is copied in from the user.
2170  */
2171 static int
2172 passsendccb(struct cam_periph *periph, union ccb *ccb, union ccb *inccb)
2173 {
2174         struct pass_softc *softc;
2175         struct cam_periph_map_info mapinfo;
2176         uint8_t *cmd;
2177         xpt_opcode fc;
2178         int error;
2179
2180         softc = (struct pass_softc *)periph->softc;
2181
2182         /*
2183          * There are some fields in the CCB header that need to be
2184          * preserved, the rest we get from the user.
2185          */
2186         xpt_merge_ccb(ccb, inccb);
2187
2188         if (ccb->ccb_h.flags & CAM_CDB_POINTER) {
2189                 cmd = __builtin_alloca(ccb->csio.cdb_len);
2190                 error = copyin(ccb->csio.cdb_io.cdb_ptr, cmd, ccb->csio.cdb_len);
2191                 if (error)
2192                         return (error);
2193                 ccb->csio.cdb_io.cdb_ptr = cmd;
2194         }
2195
2196         /*
2197          * Let cam_periph_mapmem do a sanity check on the data pointer format.
2198          * Even if no data transfer is needed, it's a cheap check and it
2199          * simplifies the code.
2200          */
2201         fc = ccb->ccb_h.func_code;
2202         if ((fc == XPT_SCSI_IO) || (fc == XPT_ATA_IO) || (fc == XPT_SMP_IO)
2203             || (fc == XPT_DEV_MATCH) || (fc == XPT_DEV_ADVINFO) || (fc == XPT_MMC_IO)
2204             || (fc == XPT_NVME_ADMIN) || (fc == XPT_NVME_IO)) {
2205                 bzero(&mapinfo, sizeof(mapinfo));
2206
2207                 /*
2208                  * cam_periph_mapmem calls into proc and vm functions that can
2209                  * sleep as well as trigger I/O, so we can't hold the lock.
2210                  * Dropping it here is reasonably safe.
2211                  */
2212                 cam_periph_unlock(periph);
2213                 error = cam_periph_mapmem(ccb, &mapinfo, softc->maxio);
2214                 cam_periph_lock(periph);
2215
2216                 /*
2217                  * cam_periph_mapmem returned an error, we can't continue.
2218                  * Return the error to the user.
2219                  */
2220                 if (error)
2221                         return(error);
2222         } else
2223                 /* Ensure that the unmap call later on is a no-op. */
2224                 mapinfo.num_bufs_used = 0;
2225
2226         /*
2227          * If the user wants us to perform any error recovery, then honor
2228          * that request.  Otherwise, it's up to the user to perform any
2229          * error recovery.
2230          */
2231         {
2232                 uint32_t cam_flags, sense_flags;
2233
2234                 passflags(ccb, &cam_flags, &sense_flags);
2235                 cam_periph_runccb(ccb,  passerror, cam_flags,
2236                     sense_flags, softc->device_stats);
2237         }
2238
2239         cam_periph_unlock(periph);
2240         cam_periph_unmapmem(ccb, &mapinfo);
2241         cam_periph_lock(periph);
2242
2243         ccb->ccb_h.cbfcnp = NULL;
2244         ccb->ccb_h.periph_priv = inccb->ccb_h.periph_priv;
2245         bcopy(ccb, inccb, sizeof(union ccb));
2246
2247         return(0);
2248 }
2249
2250 /*
2251  * Set the cam_flags and sense_flags based on whether or not the request wants
2252  * error recovery. In order to log errors via devctl, we need to do at least
2253  * minimal recovery. We do this by not retrying unit attention (we let the
2254  * requester do it, or not, if appropriate) and specifically asking for no
2255  * recovery, like we do during device probing.
2256  */
2257 static void
2258 passflags(union ccb *ccb, uint32_t *cam_flags, uint32_t *sense_flags)
2259 {
2260         if ((ccb->ccb_h.flags & CAM_PASS_ERR_RECOVER) != 0) {
2261                 *cam_flags = CAM_RETRY_SELTO;
2262                 *sense_flags = SF_RETRY_UA | SF_NO_PRINT;
2263         } else {
2264                 *cam_flags = 0;
2265                 *sense_flags = SF_NO_RETRY | SF_NO_RECOVERY | SF_NO_PRINT;
2266         }
2267 }
2268
2269 static int
2270 passerror(union ccb *ccb, uint32_t cam_flags, uint32_t sense_flags)
2271 {
2272
2273         return(cam_periph_error(ccb, cam_flags, sense_flags));
2274 }