sys/cam/cam_periph.c

   1 /*-
   2  * Common functions for CAM "type" (peripheral) drivers.
   3  *
   4  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
   5  *
   6  * Copyright (c) 1997, 1998 Justin T. Gibbs.
   7  * Copyright (c) 1997, 1998, 1999, 2000 Kenneth D. Merry.
   8  * All rights reserved.
   9  *
  10  * Redistribution and use in source and binary forms, with or without
  11  * modification, are permitted provided that the following conditions
  12  * are met:
  13  * 1. Redistributions of source code must retain the above copyright
  14  *    notice, this list of conditions, and the following disclaimer,
  15  *    without modification, immediately at the beginning of the file.
  16  * 2. The name of the author may not be used to endorse or promote products
  17  *    derived from this software without specific prior written permission.
  18  *
  19  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  22  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR
  23  * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  29  * SUCH DAMAGE.
  30  */
  31
  32 #include <sys/cdefs.h>
  33 __FBSDID("$FreeBSD$");
  34
  35 #include <sys/param.h>
  36 #include <sys/systm.h>
  37 #include <sys/types.h>
  38 #include <sys/malloc.h>
  39 #include <sys/kernel.h>
  40 #include <sys/bio.h>
  41 #include <sys/conf.h>
  42 #include <sys/lock.h>
  43 #include <sys/mutex.h>
  44 #include <sys/buf.h>
  45 #include <sys/proc.h>
  46 #include <sys/devicestat.h>
  47 #include <sys/bus.h>
  48 #include <sys/sbuf.h>
  49 #include <sys/sysctl.h>
  50 #include <vm/vm.h>
  51 #include <vm/vm_extern.h>
  52
  53 #include <cam/cam.h>
  54 #include <cam/cam_ccb.h>
  55 #include <cam/cam_queue.h>
  56 #include <cam/cam_xpt_periph.h>
  57 #include <cam/cam_periph.h>
  58 #include <cam/cam_debug.h>
  59 #include <cam/cam_sim.h>
  60
  61 #include <cam/scsi/scsi_all.h>
  62 #include <cam/scsi/scsi_message.h>
  63 #include <cam/scsi/scsi_pass.h>
  64
  65 static  u_int           camperiphnextunit(struct periph_driver *p_drv,
  66                                           u_int newunit, int wired,
  67                                           path_id_t pathid, target_id_t target,
  68                                           lun_id_t lun);
  69 static  u_int           camperiphunit(struct periph_driver *p_drv,
  70                                       path_id_t pathid, target_id_t target,
  71                                       lun_id_t lun);
  72 static  void            camperiphdone(struct cam_periph *periph,
  73                                         union ccb *done_ccb);
  74 static  void            camperiphfree(struct cam_periph *periph);
  75 static int              camperiphscsistatuserror(union ccb *ccb,
  76                                                 union ccb **orig_ccb,
  77                                                  cam_flags camflags,
  78                                                  u_int32_t sense_flags,
  79                                                  int *openings,
  80                                                  u_int32_t *relsim_flags,
  81                                                  u_int32_t *timeout,
  82                                                  u_int32_t  *action,
  83                                                  const char **action_string);
  84 static  int             camperiphscsisenseerror(union ccb *ccb,
  85                                                 union ccb **orig_ccb,
  86                                                 cam_flags camflags,
  87                                                 u_int32_t sense_flags,
  88                                                 int *openings,
  89                                                 u_int32_t *relsim_flags,
  90                                                 u_int32_t *timeout,
  91                                                 u_int32_t *action,
  92                                                 const char **action_string);
  93 static void             cam_periph_devctl_notify(union ccb *ccb);
  94
  95 static int nperiph_drivers;
  96 static int initialized = 0;
  97 struct periph_driver **periph_drivers;
  98
  99 static MALLOC_DEFINE(M_CAMPERIPH, "CAM periph", "CAM peripheral buffers");
 100
 101 static int periph_selto_delay = 1000;
 102 TUNABLE_INT("kern.cam.periph_selto_delay", &periph_selto_delay);
 103 static int periph_noresrc_delay = 500;
 104 TUNABLE_INT("kern.cam.periph_noresrc_delay", &periph_noresrc_delay);
 105 static int periph_busy_delay = 500;
 106 TUNABLE_INT("kern.cam.periph_busy_delay", &periph_busy_delay);
 107
 108 static u_int periph_mapmem_thresh = 65536;
 109 SYSCTL_UINT(_kern_cam, OID_AUTO, mapmem_thresh, CTLFLAG_RWTUN,
 110     &periph_mapmem_thresh, 0, "Threshold for user-space buffer mapping");
 111
 112 void
 113 periphdriver_register(void *data)
 114 {
 115         struct periph_driver *drv = (struct periph_driver *)data;
 116         struct periph_driver **newdrivers, **old;
 117         int ndrivers;
 118
 119 again:
 120         ndrivers = nperiph_drivers + 2;
 121         newdrivers = malloc(sizeof(*newdrivers) * ndrivers, M_CAMPERIPH,
 122                             M_WAITOK);
 123         xpt_lock_buses();
 124         if (ndrivers != nperiph_drivers + 2) {
 125                 /*
 126                  * Lost race against itself; go around.
 127                  */
 128                 xpt_unlock_buses();
 129                 free(newdrivers, M_CAMPERIPH);
 130                 goto again;
 131         }
 132         if (periph_drivers)
 133                 bcopy(periph_drivers, newdrivers,
 134                       sizeof(*newdrivers) * nperiph_drivers);
 135         newdrivers[nperiph_drivers] = drv;
 136         newdrivers[nperiph_drivers + 1] = NULL;
 137         old = periph_drivers;
 138         periph_drivers = newdrivers;
 139         nperiph_drivers++;
 140         xpt_unlock_buses();
 141         if (old)
 142                 free(old, M_CAMPERIPH);
 143         /* If driver marked as early or it is late now, initialize it. */
 144         if (((drv->flags & CAM_PERIPH_DRV_EARLY) != 0 && initialized > 0) ||
 145             initialized > 1)
 146                 (*drv->init)();
 147 }
 148
 149 int
 150 periphdriver_unregister(void *data)
 151 {
 152         struct periph_driver *drv = (struct periph_driver *)data;
 153         int error, n;
 154
 155         /* If driver marked as early or it is late now, deinitialize it. */
 156         if (((drv->flags & CAM_PERIPH_DRV_EARLY) != 0 && initialized > 0) ||
 157             initialized > 1) {
 158                 if (drv->deinit == NULL) {
 159                         printf("CAM periph driver '%s' doesn't have deinit.\n",
 160                             drv->driver_name);
 161                         return (EOPNOTSUPP);
 162                 }
 163                 error = drv->deinit();
 164                 if (error != 0)
 165                         return (error);
 166         }
 167
 168         xpt_lock_buses();
 169         for (n = 0; n < nperiph_drivers && periph_drivers[n] != drv; n++)
 170                 ;
 171         KASSERT(n < nperiph_drivers,
 172             ("Periph driver '%s' was not registered", drv->driver_name));
 173         for (; n + 1 < nperiph_drivers; n++)
 174                 periph_drivers[n] = periph_drivers[n + 1];
 175         periph_drivers[n + 1] = NULL;
 176         nperiph_drivers--;
 177         xpt_unlock_buses();
 178         return (0);
 179 }
 180
 181 void
 182 periphdriver_init(int level)
 183 {
 184         int     i, early;
 185
 186         initialized = max(initialized, level);
 187         for (i = 0; periph_drivers[i] != NULL; i++) {
 188                 early = (periph_drivers[i]->flags & CAM_PERIPH_DRV_EARLY) ? 1 : 2;
 189                 if (early == initialized)
 190                         (*periph_drivers[i]->init)();
 191         }
 192 }
 193
 194 cam_status
 195 cam_periph_alloc(periph_ctor_t *periph_ctor,
 196                  periph_oninv_t *periph_oninvalidate,
 197                  periph_dtor_t *periph_dtor, periph_start_t *periph_start,
 198                  char *name, cam_periph_type type, struct cam_path *path,
 199                  ac_callback_t *ac_callback, ac_code code, void *arg)
 200 {
 201         struct          periph_driver **p_drv;
 202         struct          cam_sim *sim;
 203         struct          cam_periph *periph;
 204         struct          cam_periph *cur_periph;
 205         path_id_t       path_id;
 206         target_id_t     target_id;
 207         lun_id_t        lun_id;
 208         cam_status      status;
 209         u_int           init_level;
 210
 211         init_level = 0;
 212         /*
 213          * Handle Hot-Plug scenarios.  If there is already a peripheral
 214          * of our type assigned to this path, we are likely waiting for
 215          * final close on an old, invalidated, peripheral.  If this is
 216          * the case, queue up a deferred call to the peripheral's async
 217          * handler.  If it looks like a mistaken re-allocation, complain.
 218          */
 219         if ((periph = cam_periph_find(path, name)) != NULL) {
 220
 221                 if ((periph->flags & CAM_PERIPH_INVALID) != 0
 222                  && (periph->flags & CAM_PERIPH_NEW_DEV_FOUND) == 0) {
 223                         periph->flags |= CAM_PERIPH_NEW_DEV_FOUND;
 224                         periph->deferred_callback = ac_callback;
 225                         periph->deferred_ac = code;
 226                         return (CAM_REQ_INPROG);
 227                 } else {
 228                         printf("cam_periph_alloc: attempt to re-allocate "
 229                                "valid device %s%d rejected flags %#x "
 230                                "refcount %d\n", periph->periph_name,
 231                                periph->unit_number, periph->flags,
 232                                periph->refcount);
 233                 }
 234                 return (CAM_REQ_INVALID);
 235         }
 236
 237         periph = (struct cam_periph *)malloc(sizeof(*periph), M_CAMPERIPH,
 238                                              M_NOWAIT|M_ZERO);
 239
 240         if (periph == NULL)
 241                 return (CAM_RESRC_UNAVAIL);
 242
 243         init_level++;
 244
 245
 246         sim = xpt_path_sim(path);
 247         path_id = xpt_path_path_id(path);
 248         target_id = xpt_path_target_id(path);
 249         lun_id = xpt_path_lun_id(path);
 250         periph->periph_start = periph_start;
 251         periph->periph_dtor = periph_dtor;
 252         periph->periph_oninval = periph_oninvalidate;
 253         periph->type = type;
 254         periph->periph_name = name;
 255         periph->scheduled_priority = CAM_PRIORITY_NONE;
 256         periph->immediate_priority = CAM_PRIORITY_NONE;
 257         periph->refcount = 1;           /* Dropped by invalidation. */
 258         periph->sim = sim;
 259         SLIST_INIT(&periph->ccb_list);
 260         status = xpt_create_path(&path, periph, path_id, target_id, lun_id);
 261         if (status != CAM_REQ_CMP)
 262                 goto failure;
 263         periph->path = path;
 264
 265         xpt_lock_buses();
 266         for (p_drv = periph_drivers; *p_drv != NULL; p_drv++) {
 267                 if (strcmp((*p_drv)->driver_name, name) == 0)
 268                         break;
 269         }
 270         if (*p_drv == NULL) {
 271                 printf("cam_periph_alloc: invalid periph name '%s'\n", name);
 272                 xpt_unlock_buses();
 273                 xpt_free_path(periph->path);
 274                 free(periph, M_CAMPERIPH);
 275                 return (CAM_REQ_INVALID);
 276         }
 277         periph->unit_number = camperiphunit(*p_drv, path_id, target_id, lun_id);
 278         cur_periph = TAILQ_FIRST(&(*p_drv)->units);
 279         while (cur_periph != NULL
 280             && cur_periph->unit_number < periph->unit_number)
 281                 cur_periph = TAILQ_NEXT(cur_periph, unit_links);
 282         if (cur_periph != NULL) {
 283                 KASSERT(cur_periph->unit_number != periph->unit_number, ("duplicate units on periph list"));
 284                 TAILQ_INSERT_BEFORE(cur_periph, periph, unit_links);
 285         } else {
 286                 TAILQ_INSERT_TAIL(&(*p_drv)->units, periph, unit_links);
 287                 (*p_drv)->generation++;
 288         }
 289         xpt_unlock_buses();
 290
 291         init_level++;
 292
 293         status = xpt_add_periph(periph);
 294         if (status != CAM_REQ_CMP)
 295                 goto failure;
 296
 297         init_level++;
 298         CAM_DEBUG(periph->path, CAM_DEBUG_INFO, ("Periph created\n"));
 299
 300         status = periph_ctor(periph, arg);
 301
 302         if (status == CAM_REQ_CMP)
 303                 init_level++;
 304
 305 failure:
 306         switch (init_level) {
 307         case 4:
 308                 /* Initialized successfully */
 309                 break;
 310         case 3:
 311                 CAM_DEBUG(periph->path, CAM_DEBUG_INFO, ("Periph destroyed\n"));
 312                 xpt_remove_periph(periph);
 313                 /* FALLTHROUGH */
 314         case 2:
 315                 xpt_lock_buses();
 316                 TAILQ_REMOVE(&(*p_drv)->units, periph, unit_links);
 317                 xpt_unlock_buses();
 318                 xpt_free_path(periph->path);
 319                 /* FALLTHROUGH */
 320         case 1:
 321                 free(periph, M_CAMPERIPH);
 322                 /* FALLTHROUGH */
 323         case 0:
 324                 /* No cleanup to perform. */
 325                 break;
 326         default:
 327                 panic("%s: Unknown init level", __func__);
 328         }
 329         return(status);
 330 }
 331
 332 /*
 333  * Find a peripheral structure with the specified path, target, lun,
 334  * and (optionally) type.  If the name is NULL, this function will return
 335  * the first peripheral driver that matches the specified path.
 336  */
 337 struct cam_periph *
 338 cam_periph_find(struct cam_path *path, char *name)
 339 {
 340         struct periph_driver **p_drv;
 341         struct cam_periph *periph;
 342
 343         xpt_lock_buses();
 344         for (p_drv = periph_drivers; *p_drv != NULL; p_drv++) {
 345
 346                 if (name != NULL && (strcmp((*p_drv)->driver_name, name) != 0))
 347                         continue;
 348
 349                 TAILQ_FOREACH(periph, &(*p_drv)->units, unit_links) {
 350                         if (xpt_path_comp(periph->path, path) == 0) {
 351                                 xpt_unlock_buses();
 352                                 cam_periph_assert(periph, MA_OWNED);
 353                                 return(periph);
 354                         }
 355                 }
 356                 if (name != NULL) {
 357                         xpt_unlock_buses();
 358                         return(NULL);
 359                 }
 360         }
 361         xpt_unlock_buses();
 362         return(NULL);
 363 }
 364
 365 /*
 366  * Find peripheral driver instances attached to the specified path.
 367  */
 368 int
 369 cam_periph_list(struct cam_path *path, struct sbuf *sb)
 370 {
 371         struct sbuf local_sb;
 372         struct periph_driver **p_drv;
 373         struct cam_periph *periph;
 374         int count;
 375         int sbuf_alloc_len;
 376
 377         sbuf_alloc_len = 16;
 378 retry:
 379         sbuf_new(&local_sb, NULL, sbuf_alloc_len, SBUF_FIXEDLEN);
 380         count = 0;
 381         xpt_lock_buses();
 382         for (p_drv = periph_drivers; *p_drv != NULL; p_drv++) {
 383
 384                 TAILQ_FOREACH(periph, &(*p_drv)->units, unit_links) {
 385                         if (xpt_path_comp(periph->path, path) != 0)
 386                                 continue;
 387
 388                         if (sbuf_len(&local_sb) != 0)
 389                                 sbuf_cat(&local_sb, ",");
 390
 391                         sbuf_printf(&local_sb, "%s%d", periph->periph_name,
 392                                     periph->unit_number);
 393
 394                         if (sbuf_error(&local_sb) == ENOMEM) {
 395                                 sbuf_alloc_len *= 2;
 396                                 xpt_unlock_buses();
 397                                 sbuf_delete(&local_sb);
 398                                 goto retry;
 399                         }
 400                         count++;
 401                 }
 402         }
 403         xpt_unlock_buses();
 404         sbuf_finish(&local_sb);
 405         if (sbuf_len(sb) != 0)
 406                 sbuf_cat(sb, ",");
 407         sbuf_cat(sb, sbuf_data(&local_sb));
 408         sbuf_delete(&local_sb);
 409         return (count);
 410 }
 411
 412 int
 413 cam_periph_acquire(struct cam_periph *periph)
 414 {
 415         int status;
 416
 417         if (periph == NULL)
 418                 return (EINVAL);
 419
 420         status = ENOENT;
 421         xpt_lock_buses();
 422         if ((periph->flags & CAM_PERIPH_INVALID) == 0) {
 423                 periph->refcount++;
 424                 status = 0;
 425         }
 426         xpt_unlock_buses();
 427
 428         return (status);
 429 }
 430
 431 void
 432 cam_periph_doacquire(struct cam_periph *periph)
 433 {
 434
 435         xpt_lock_buses();
 436         KASSERT(periph->refcount >= 1,
 437             ("cam_periph_doacquire() with refcount == %d", periph->refcount));
 438         periph->refcount++;
 439         xpt_unlock_buses();
 440 }
 441
 442 void
 443 cam_periph_release_locked_buses(struct cam_periph *periph)
 444 {
 445
 446         cam_periph_assert(periph, MA_OWNED);
 447         KASSERT(periph->refcount >= 1, ("periph->refcount >= 1"));
 448         if (--periph->refcount == 0)
 449                 camperiphfree(periph);
 450 }
 451
 452 void
 453 cam_periph_release_locked(struct cam_periph *periph)
 454 {
 455
 456         if (periph == NULL)
 457                 return;
 458
 459         xpt_lock_buses();
 460         cam_periph_release_locked_buses(periph);
 461         xpt_unlock_buses();
 462 }
 463
 464 void
 465 cam_periph_release(struct cam_periph *periph)
 466 {
 467         struct mtx *mtx;
 468
 469         if (periph == NULL)
 470                 return;
 471
 472         cam_periph_assert(periph, MA_NOTOWNED);
 473         mtx = cam_periph_mtx(periph);
 474         mtx_lock(mtx);
 475         cam_periph_release_locked(periph);
 476         mtx_unlock(mtx);
 477 }
 478
 479 /*
 480  * hold/unhold act as mutual exclusion for sections of the code that
 481  * need to sleep and want to make sure that other sections that
 482  * will interfere are held off. This only protects exclusive sections
 483  * from each other.
 484  */
 485 int
 486 cam_periph_hold(struct cam_periph *periph, int priority)
 487 {
 488         int error;
 489
 490         /*
 491          * Increment the reference count on the peripheral
 492          * while we wait for our lock attempt to succeed
 493          * to ensure the peripheral doesn't disappear out
 494          * from user us while we sleep.
 495          */
 496
 497         if (cam_periph_acquire(periph) != 0)
 498                 return (ENXIO);
 499
 500         cam_periph_assert(periph, MA_OWNED);
 501         while ((periph->flags & CAM_PERIPH_LOCKED) != 0) {
 502                 periph->flags |= CAM_PERIPH_LOCK_WANTED;
 503                 if ((error = cam_periph_sleep(periph, periph, priority,
 504                     "caplck", 0)) != 0) {
 505                         cam_periph_release_locked(periph);
 506                         return (error);
 507                 }
 508                 if (periph->flags & CAM_PERIPH_INVALID) {
 509                         cam_periph_release_locked(periph);
 510                         return (ENXIO);
 511                 }
 512         }
 513
 514         periph->flags |= CAM_PERIPH_LOCKED;
 515         return (0);
 516 }
 517
 518 void
 519 cam_periph_unhold(struct cam_periph *periph)
 520 {
 521
 522         cam_periph_assert(periph, MA_OWNED);
 523
 524         periph->flags &= ~CAM_PERIPH_LOCKED;
 525         if ((periph->flags & CAM_PERIPH_LOCK_WANTED) != 0) {
 526                 periph->flags &= ~CAM_PERIPH_LOCK_WANTED;
 527                 wakeup(periph);
 528         }
 529
 530         cam_periph_release_locked(periph);
 531 }
 532
 533 /*
 534  * Look for the next unit number that is not currently in use for this
 535  * peripheral type starting at "newunit".  Also exclude unit numbers that
 536  * are reserved by for future "hardwiring" unless we already know that this
 537  * is a potential wired device.  Only assume that the device is "wired" the
 538  * first time through the loop since after that we'll be looking at unit
 539  * numbers that did not match a wiring entry.
 540  */
 541 static u_int
 542 camperiphnextunit(struct periph_driver *p_drv, u_int newunit, int wired,
 543                   path_id_t pathid, target_id_t target, lun_id_t lun)
 544 {
 545         struct  cam_periph *periph;
 546         char    *periph_name;
 547         int     i, val, dunit, r;
 548         const char *dname, *strval;
 549
 550         periph_name = p_drv->driver_name;
 551         for (;;newunit++) {
 552
 553                 for (periph = TAILQ_FIRST(&p_drv->units);
 554                      periph != NULL && periph->unit_number != newunit;
 555                      periph = TAILQ_NEXT(periph, unit_links))
 556                         ;
 557
 558                 if (periph != NULL && periph->unit_number == newunit) {
 559                         if (wired != 0) {
 560                                 xpt_print(periph->path, "Duplicate Wired "
 561                                     "Device entry!\n");
 562                                 xpt_print(periph->path, "Second device (%s "
 563                                     "device at scbus%d target %d lun %d) will "
 564                                     "not be wired\n", periph_name, pathid,
 565                                     target, lun);
 566                                 wired = 0;
 567                         }
 568                         continue;
 569                 }
 570                 if (wired)
 571                         break;
 572
 573                 /*
 574                  * Don't match entries like "da 4" as a wired down
 575                  * device, but do match entries like "da 4 target 5"
 576                  * or even "da 4 scbus 1".
 577                  */
 578                 i = 0;
 579                 dname = periph_name;
 580                 for (;;) {
 581                         r = resource_find_dev(&i, dname, &dunit, NULL, NULL);
 582                         if (r != 0)
 583                                 break;
 584                         /* if no "target" and no specific scbus, skip */
 585                         if (resource_int_value(dname, dunit, "target", &val) &&
 586                             (resource_string_value(dname, dunit, "at",&strval)||
 587                              strcmp(strval, "scbus") == 0))
 588                                 continue;
 589                         if (newunit == dunit)
 590                                 break;
 591                 }
 592                 if (r != 0)
 593                         break;
 594         }
 595         return (newunit);
 596 }
 597
 598 static u_int
 599 camperiphunit(struct periph_driver *p_drv, path_id_t pathid,
 600               target_id_t target, lun_id_t lun)
 601 {
 602         u_int   unit;
 603         int     wired, i, val, dunit;
 604         const char *dname, *strval;
 605         char    pathbuf[32], *periph_name;
 606
 607         periph_name = p_drv->driver_name;
 608         snprintf(pathbuf, sizeof(pathbuf), "scbus%d", pathid);
 609         unit = 0;
 610         i = 0;
 611         dname = periph_name;
 612         for (wired = 0; resource_find_dev(&i, dname, &dunit, NULL, NULL) == 0;
 613              wired = 0) {
 614                 if (resource_string_value(dname, dunit, "at", &strval) == 0) {
 615                         if (strcmp(strval, pathbuf) != 0)
 616                                 continue;
 617                         wired++;
 618                 }
 619                 if (resource_int_value(dname, dunit, "target", &val) == 0) {
 620                         if (val != target)
 621                                 continue;
 622                         wired++;
 623                 }
 624                 if (resource_int_value(dname, dunit, "lun", &val) == 0) {
 625                         if (val != lun)
 626                                 continue;
 627                         wired++;
 628                 }
 629                 if (wired != 0) {
 630                         unit = dunit;
 631                         break;
 632                 }
 633         }
 634
 635         /*
 636          * Either start from 0 looking for the next unit or from
 637          * the unit number given in the resource config.  This way,
 638          * if we have wildcard matches, we don't return the same
 639          * unit number twice.
 640          */
 641         unit = camperiphnextunit(p_drv, unit, wired, pathid, target, lun);
 642
 643         return (unit);
 644 }
 645
 646 void
 647 cam_periph_invalidate(struct cam_periph *periph)
 648 {
 649
 650         cam_periph_assert(periph, MA_OWNED);
 651         /*
 652          * We only call this routine the first time a peripheral is
 653          * invalidated.
 654          */
 655         if ((periph->flags & CAM_PERIPH_INVALID) != 0)
 656                 return;
 657
 658         CAM_DEBUG(periph->path, CAM_DEBUG_INFO, ("Periph invalidated\n"));
 659         if ((periph->flags & CAM_PERIPH_ANNOUNCED) && !rebooting) {
 660                 struct sbuf sb;
 661                 char buffer[160];
 662
 663                 sbuf_new(&sb, buffer, 160, SBUF_FIXEDLEN);
 664                 xpt_denounce_periph_sbuf(periph, &sb);
 665                 sbuf_finish(&sb);
 666                 sbuf_putbuf(&sb);
 667         }
 668         periph->flags |= CAM_PERIPH_INVALID;
 669         periph->flags &= ~CAM_PERIPH_NEW_DEV_FOUND;
 670         if (periph->periph_oninval != NULL)
 671                 periph->periph_oninval(periph);
 672         cam_periph_release_locked(periph);
 673 }
 674
 675 static void
 676 camperiphfree(struct cam_periph *periph)
 677 {
 678         struct periph_driver **p_drv;
 679         struct periph_driver *drv;
 680
 681         cam_periph_assert(periph, MA_OWNED);
 682         KASSERT(periph->periph_allocating == 0, ("%s%d: freed while allocating",
 683             periph->periph_name, periph->unit_number));
 684         for (p_drv = periph_drivers; *p_drv != NULL; p_drv++) {
 685                 if (strcmp((*p_drv)->driver_name, periph->periph_name) == 0)
 686                         break;
 687         }
 688         if (*p_drv == NULL) {
 689                 printf("camperiphfree: attempt to free non-existant periph\n");
 690                 return;
 691         }
 692         /*
 693          * Cache a pointer to the periph_driver structure.  If a
 694          * periph_driver is added or removed from the array (see
 695          * periphdriver_register()) while we drop the toplogy lock
 696          * below, p_drv may change.  This doesn't protect against this
 697          * particular periph_driver going away.  That will require full
 698          * reference counting in the periph_driver infrastructure.
 699          */
 700         drv = *p_drv;
 701
 702         /*
 703          * We need to set this flag before dropping the topology lock, to
 704          * let anyone who is traversing the list that this peripheral is
 705          * about to be freed, and there will be no more reference count
 706          * checks.
 707          */
 708         periph->flags |= CAM_PERIPH_FREE;
 709
 710         /*
 711          * The peripheral destructor semantics dictate calling with only the
 712          * SIM mutex held.  Since it might sleep, it should not be called
 713          * with the topology lock held.
 714          */
 715         xpt_unlock_buses();
 716
 717         /*
 718          * We need to call the peripheral destructor prior to removing the
 719          * peripheral from the list.  Otherwise, we risk running into a
 720          * scenario where the peripheral unit number may get reused
 721          * (because it has been removed from the list), but some resources
 722          * used by the peripheral are still hanging around.  In particular,
 723          * the devfs nodes used by some peripherals like the pass(4) driver
 724          * aren't fully cleaned up until the destructor is run.  If the
 725          * unit number is reused before the devfs instance is fully gone,
 726          * devfs will panic.
 727          */
 728         if (periph->periph_dtor != NULL)
 729                 periph->periph_dtor(periph);
 730
 731         /*
 732          * The peripheral list is protected by the topology lock.
 733          */
 734         xpt_lock_buses();
 735
 736         TAILQ_REMOVE(&drv->units, periph, unit_links);
 737         drv->generation++;
 738
 739         xpt_remove_periph(periph);
 740
 741         xpt_unlock_buses();
 742         if ((periph->flags & CAM_PERIPH_ANNOUNCED) && !rebooting)
 743                 xpt_print(periph->path, "Periph destroyed\n");
 744         else
 745                 CAM_DEBUG(periph->path, CAM_DEBUG_INFO, ("Periph destroyed\n"));
 746
 747         if (periph->flags & CAM_PERIPH_NEW_DEV_FOUND) {
 748                 union ccb ccb;
 749                 void *arg;
 750
 751                 switch (periph->deferred_ac) {
 752                 case AC_FOUND_DEVICE:
 753                         ccb.ccb_h.func_code = XPT_GDEV_TYPE;
 754                         xpt_setup_ccb(&ccb.ccb_h, periph->path, CAM_PRIORITY_NORMAL);
 755                         xpt_action(&ccb);
 756                         arg = &ccb;
 757                         break;
 758                 case AC_PATH_REGISTERED:
 759                         xpt_path_inq(&ccb.cpi, periph->path);
 760                         arg = &ccb;
 761                         break;
 762                 default:
 763                         arg = NULL;
 764                         break;
 765                 }
 766                 periph->deferred_callback(NULL, periph->deferred_ac,
 767                                           periph->path, arg);
 768         }
 769         xpt_free_path(periph->path);
 770         free(periph, M_CAMPERIPH);
 771         xpt_lock_buses();
 772 }
 773
 774 /*
 775  * Map user virtual pointers into kernel virtual address space, so we can
 776  * access the memory.  This is now a generic function that centralizes most
 777  * of the sanity checks on the data flags, if any.
 778  * This also only works for up to MAXPHYS memory.  Since we use
 779  * buffers to map stuff in and out, we're limited to the buffer size.
 780  */
 781 int
 782 cam_periph_mapmem(union ccb *ccb, struct cam_periph_map_info *mapinfo,
 783     u_int maxmap)
 784 {
 785         int numbufs, i;
 786         u_int8_t **data_ptrs[CAM_PERIPH_MAXMAPS];
 787         u_int32_t lengths[CAM_PERIPH_MAXMAPS];
 788         u_int32_t dirs[CAM_PERIPH_MAXMAPS];
 789
 790         bzero(mapinfo, sizeof(*mapinfo));
 791         if (maxmap == 0)
 792                 maxmap = DFLTPHYS;      /* traditional default */
 793         else if (maxmap > MAXPHYS)
 794                 maxmap = MAXPHYS;       /* for safety */
 795         switch(ccb->ccb_h.func_code) {
 796         case XPT_DEV_MATCH:
 797                 if (ccb->cdm.match_buf_len == 0) {
 798                         printf("cam_periph_mapmem: invalid match buffer "
 799                                "length 0\n");
 800                         return(EINVAL);
 801                 }
 802                 if (ccb->cdm.pattern_buf_len > 0) {
 803                         data_ptrs[0] = (u_int8_t **)&ccb->cdm.patterns;
 804                         lengths[0] = ccb->cdm.pattern_buf_len;
 805                         dirs[0] = CAM_DIR_OUT;
 806                         data_ptrs[1] = (u_int8_t **)&ccb->cdm.matches;
 807                         lengths[1] = ccb->cdm.match_buf_len;
 808                         dirs[1] = CAM_DIR_IN;
 809                         numbufs = 2;
 810                 } else {
 811                         data_ptrs[0] = (u_int8_t **)&ccb->cdm.matches;
 812                         lengths[0] = ccb->cdm.match_buf_len;
 813                         dirs[0] = CAM_DIR_IN;
 814                         numbufs = 1;
 815                 }
 816                 /*
 817                  * This request will not go to the hardware, no reason
 818                  * to be so strict. vmapbuf() is able to map up to MAXPHYS.
 819                  */
 820                 maxmap = MAXPHYS;
 821                 break;
 822         case XPT_SCSI_IO:
 823         case XPT_CONT_TARGET_IO:
 824                 if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
 825                         return(0);
 826                 if ((ccb->ccb_h.flags & CAM_DATA_MASK) != CAM_DATA_VADDR)
 827                         return (EINVAL);
 828                 data_ptrs[0] = &ccb->csio.data_ptr;
 829                 lengths[0] = ccb->csio.dxfer_len;
 830                 dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
 831                 numbufs = 1;
 832                 break;
 833         case XPT_ATA_IO:
 834                 if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
 835                         return(0);
 836                 if ((ccb->ccb_h.flags & CAM_DATA_MASK) != CAM_DATA_VADDR)
 837                         return (EINVAL);
 838                 data_ptrs[0] = &ccb->ataio.data_ptr;
 839                 lengths[0] = ccb->ataio.dxfer_len;
 840                 dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
 841                 numbufs = 1;
 842                 break;
 843         case XPT_MMC_IO:
 844                 if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
 845                         return(0);
 846                 /* Two mappings: one for cmd->data and one for cmd->data->data */
 847                 data_ptrs[0] = (unsigned char **)&ccb->mmcio.cmd.data;
 848                 lengths[0] = sizeof(struct mmc_data *);
 849                 dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
 850                 data_ptrs[1] = (unsigned char **)&ccb->mmcio.cmd.data->data;
 851                 lengths[1] = ccb->mmcio.cmd.data->len;
 852                 dirs[1] = ccb->ccb_h.flags & CAM_DIR_MASK;
 853                 numbufs = 2;
 854                 break;
 855         case XPT_SMP_IO:
 856                 data_ptrs[0] = &ccb->smpio.smp_request;
 857                 lengths[0] = ccb->smpio.smp_request_len;
 858                 dirs[0] = CAM_DIR_OUT;
 859                 data_ptrs[1] = &ccb->smpio.smp_response;
 860                 lengths[1] = ccb->smpio.smp_response_len;
 861                 dirs[1] = CAM_DIR_IN;
 862                 numbufs = 2;
 863                 break;
 864         case XPT_NVME_IO:
 865         case XPT_NVME_ADMIN:
 866                 if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
 867                         return (0);
 868                 if ((ccb->ccb_h.flags & CAM_DATA_MASK) != CAM_DATA_VADDR)
 869                         return (EINVAL);
 870                 data_ptrs[0] = &ccb->nvmeio.data_ptr;
 871                 lengths[0] = ccb->nvmeio.dxfer_len;
 872                 dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
 873                 numbufs = 1;
 874                 break;
 875         case XPT_DEV_ADVINFO:
 876                 if (ccb->cdai.bufsiz == 0)
 877                         return (0);
 878
 879                 data_ptrs[0] = (uint8_t **)&ccb->cdai.buf;
 880                 lengths[0] = ccb->cdai.bufsiz;
 881                 dirs[0] = CAM_DIR_IN;
 882                 numbufs = 1;
 883
 884                 /*
 885                  * This request will not go to the hardware, no reason
 886                  * to be so strict. vmapbuf() is able to map up to MAXPHYS.
 887                  */
 888                 maxmap = MAXPHYS;
 889                 break;
 890         default:
 891                 return(EINVAL);
 892                 break; /* NOTREACHED */
 893         }
 894
 895         /*
 896          * Check the transfer length and permissions first, so we don't
 897          * have to unmap any previously mapped buffers.
 898          */
 899         for (i = 0; i < numbufs; i++) {
 900
 901                 /*
 902                  * The userland data pointer passed in may not be page
 903                  * aligned.  vmapbuf() truncates the address to a page
 904                  * boundary, so if the address isn't page aligned, we'll
 905                  * need enough space for the given transfer length, plus
 906                  * whatever extra space is necessary to make it to the page
 907                  * boundary.
 908                  */
 909                 if ((lengths[i] +
 910                     (((vm_offset_t)(*data_ptrs[i])) & PAGE_MASK)) > maxmap){
 911                         printf("cam_periph_mapmem: attempt to map %lu bytes, "
 912                                "which is greater than %lu\n",
 913                                (long)(lengths[i] +
 914                                (((vm_offset_t)(*data_ptrs[i])) & PAGE_MASK)),
 915                                (u_long)maxmap);
 916                         return(E2BIG);
 917                 }
 918         }
 919
 920         /*
 921          * This keeps the kernel stack of current thread from getting
 922          * swapped.  In low-memory situations where the kernel stack might
 923          * otherwise get swapped out, this holds it and allows the thread
 924          * to make progress and release the kernel mapped pages sooner.
 925          *
 926          * XXX KDM should I use P_NOSWAP instead?
 927          */
 928         PHOLD(curproc);
 929
 930         for (i = 0; i < numbufs; i++) {
 931
 932                 /* Save the user's data address. */
 933                 mapinfo->orig[i] = *data_ptrs[i];
 934
 935                 /*
 936                  * For small buffers use malloc+copyin/copyout instead of
 937                  * mapping to KVA to avoid expensive TLB shootdowns.  For
 938                  * small allocations malloc is backed by UMA, and so much
 939                  * cheaper on SMP systems.
 940                  */
 941                 if (lengths[i] <= periph_mapmem_thresh &&
 942                     ccb->ccb_h.func_code != XPT_MMC_IO) {
 943                         *data_ptrs[i] = malloc(lengths[i], M_CAMPERIPH,
 944                             M_WAITOK);
 945                         if (dirs[i] != CAM_DIR_IN) {
 946                                 if (copyin(mapinfo->orig[i], *data_ptrs[i],
 947                                     lengths[i]) != 0) {
 948                                         free(*data_ptrs[i], M_CAMPERIPH);
 949                                         *data_ptrs[i] = mapinfo->orig[i];
 950                                         goto fail;
 951                                 }
 952                         } else
 953                                 bzero(*data_ptrs[i], lengths[i]);
 954                         continue;
 955                 }
 956
 957                 /*
 958                  * Get the buffer.
 959                  */
 960                 mapinfo->bp[i] = getpbuf(NULL);
 961
 962                 /* put our pointer in the data slot */
 963                 mapinfo->bp[i]->b_data = *data_ptrs[i];
 964
 965                 /* set the transfer length, we know it's < MAXPHYS */
 966                 mapinfo->bp[i]->b_bufsize = lengths[i];
 967
 968                 /* set the direction */
 969                 mapinfo->bp[i]->b_iocmd = (dirs[i] == CAM_DIR_OUT) ?
 970                     BIO_WRITE : BIO_READ;
 971
 972                 /*
 973                  * Map the buffer into kernel memory.
 974                  *
 975                  * Note that useracc() alone is not a  sufficient test.
 976                  * vmapbuf() can still fail due to a smaller file mapped
 977                  * into a larger area of VM, or if userland races against
 978                  * vmapbuf() after the useracc() check.
 979                  */
 980                 if (vmapbuf(mapinfo->bp[i], 1) < 0) {
 981                         relpbuf(mapinfo->bp[i], NULL);
 982                         goto fail;
 983                 }
 984
 985                 /* set our pointer to the new mapped area */
 986                 *data_ptrs[i] = mapinfo->bp[i]->b_data;
 987         }
 988
 989         /*
 990          * Now that we've gotten this far, change ownership to the kernel
 991          * of the buffers so that we don't run afoul of returning to user
 992          * space with locks (on the buffer) held.
 993          */
 994         for (i = 0; i < numbufs; i++) {
 995                 if (mapinfo->bp[i])
 996                         BUF_KERNPROC(mapinfo->bp[i]);
 997         }
 998
 999         mapinfo->num_bufs_used = numbufs;
1000         return(0);
1001
1002 fail:
1003         for (i--; i >= 0; i--) {
1004                 if (mapinfo->bp[i]) {
1005                         vunmapbuf(mapinfo->bp[i]);
1006                         relpbuf(mapinfo->bp[i], NULL);
1007                 } else
1008                         free(*data_ptrs[i], M_CAMPERIPH);
1009                 *data_ptrs[i] = mapinfo->orig[i];
1010         }
1011         PRELE(curproc);
1012         return(EACCES);
1013 }
1014
1015 /*
1016  * Unmap memory segments mapped into kernel virtual address space by
1017  * cam_periph_mapmem().
1018  */
1019 void
1020 cam_periph_unmapmem(union ccb *ccb, struct cam_periph_map_info *mapinfo)
1021 {
1022         int numbufs, i;
1023         u_int8_t **data_ptrs[CAM_PERIPH_MAXMAPS];
1024         u_int32_t lengths[CAM_PERIPH_MAXMAPS];
1025         u_int32_t dirs[CAM_PERIPH_MAXMAPS];
1026
1027         if (mapinfo->num_bufs_used <= 0) {
1028                 /* nothing to free and the process wasn't held. */
1029                 return;
1030         }
1031
1032         switch (ccb->ccb_h.func_code) {
1033         case XPT_DEV_MATCH:
1034                 if (ccb->cdm.pattern_buf_len > 0) {
1035                         data_ptrs[0] = (u_int8_t **)&ccb->cdm.patterns;
1036                         lengths[0] = ccb->cdm.pattern_buf_len;
1037                         dirs[0] = CAM_DIR_OUT;
1038                         data_ptrs[1] = (u_int8_t **)&ccb->cdm.matches;
1039                         lengths[1] = ccb->cdm.match_buf_len;
1040                         dirs[1] = CAM_DIR_IN;
1041                         numbufs = 2;
1042                 } else {
1043                         data_ptrs[0] = (u_int8_t **)&ccb->cdm.matches;
1044                         lengths[0] = ccb->cdm.match_buf_len;
1045                         dirs[0] = CAM_DIR_IN;
1046                         numbufs = 1;
1047                 }
1048                 break;
1049         case XPT_SCSI_IO:
1050         case XPT_CONT_TARGET_IO:
1051                 data_ptrs[0] = &ccb->csio.data_ptr;
1052                 lengths[0] = ccb->csio.dxfer_len;
1053                 dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1054                 numbufs = 1;
1055                 break;
1056         case XPT_ATA_IO:
1057                 data_ptrs[0] = &ccb->ataio.data_ptr;
1058                 lengths[0] = ccb->ataio.dxfer_len;
1059                 dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1060                 numbufs = 1;
1061                 break;
1062         case XPT_MMC_IO:
1063                 data_ptrs[0] = (u_int8_t **)&ccb->mmcio.cmd.data;
1064                 lengths[0] = sizeof(struct mmc_data *);
1065                 dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1066                 data_ptrs[1] = (u_int8_t **)&ccb->mmcio.cmd.data->data;
1067                 lengths[1] = ccb->mmcio.cmd.data->len;
1068                 dirs[1] = ccb->ccb_h.flags & CAM_DIR_MASK;
1069                 numbufs = 2;
1070                 break;
1071         case XPT_SMP_IO:
1072                 data_ptrs[0] = &ccb->smpio.smp_request;
1073                 lengths[0] = ccb->smpio.smp_request_len;
1074                 dirs[0] = CAM_DIR_OUT;
1075                 data_ptrs[1] = &ccb->smpio.smp_response;
1076                 lengths[1] = ccb->smpio.smp_response_len;
1077                 dirs[1] = CAM_DIR_IN;
1078                 numbufs = 2;
1079                 break;
1080         case XPT_NVME_IO:
1081         case XPT_NVME_ADMIN:
1082                 data_ptrs[0] = &ccb->nvmeio.data_ptr;
1083                 lengths[0] = ccb->nvmeio.dxfer_len;
1084                 dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1085                 numbufs = 1;
1086                 break;
1087         case XPT_DEV_ADVINFO:
1088                 data_ptrs[0] = (uint8_t **)&ccb->cdai.buf;
1089                 lengths[0] = ccb->cdai.bufsiz;
1090                 dirs[0] = CAM_DIR_IN;
1091                 numbufs = 1;
1092                 break;
1093         default:
1094                 /* allow ourselves to be swapped once again */
1095                 PRELE(curproc);
1096                 return;
1097                 break; /* NOTREACHED */
1098         }
1099
1100         for (i = 0; i < numbufs; i++) {
1101                 if (mapinfo->bp[i]) {
1102                         /* unmap the buffer */
1103                         vunmapbuf(mapinfo->bp[i]);
1104
1105                         /* release the buffer */
1106                         relpbuf(mapinfo->bp[i], NULL);
1107                 } else {
1108                         if (dirs[i] != CAM_DIR_OUT) {
1109                                 copyout(*data_ptrs[i], mapinfo->orig[i],
1110                                     lengths[i]);
1111                         }
1112                         free(*data_ptrs[i], M_CAMPERIPH);
1113                 }
1114
1115                 /* Set the user's pointer back to the original value */
1116                 *data_ptrs[i] = mapinfo->orig[i];
1117         }
1118
1119         /* allow ourselves to be swapped once again */
1120         PRELE(curproc);
1121 }
1122
1123 int
1124 cam_periph_ioctl(struct cam_periph *periph, u_long cmd, caddr_t addr,
1125                  int (*error_routine)(union ccb *ccb,
1126                                       cam_flags camflags,
1127                                       u_int32_t sense_flags))
1128 {
1129         union ccb            *ccb;
1130         int                  error;
1131         int                  found;
1132
1133         error = found = 0;
1134
1135         switch(cmd){
1136         case CAMGETPASSTHRU:
1137                 ccb = cam_periph_getccb(periph, CAM_PRIORITY_NORMAL);
1138                 xpt_setup_ccb(&ccb->ccb_h,
1139                               ccb->ccb_h.path,
1140                               CAM_PRIORITY_NORMAL);
1141                 ccb->ccb_h.func_code = XPT_GDEVLIST;
1142
1143                 /*
1144                  * Basically, the point of this is that we go through
1145                  * getting the list of devices, until we find a passthrough
1146                  * device.  In the current version of the CAM code, the
1147                  * only way to determine what type of device we're dealing
1148                  * with is by its name.
1149                  */
1150                 while (found == 0) {
1151                         ccb->cgdl.index = 0;
1152                         ccb->cgdl.status = CAM_GDEVLIST_MORE_DEVS;
1153                         while (ccb->cgdl.status == CAM_GDEVLIST_MORE_DEVS) {
1154
1155                                 /* we want the next device in the list */
1156                                 xpt_action(ccb);
1157                                 if (strncmp(ccb->cgdl.periph_name,
1158                                     "pass", 4) == 0){
1159                                         found = 1;
1160                                         break;
1161                                 }
1162                         }
1163                         if ((ccb->cgdl.status == CAM_GDEVLIST_LAST_DEVICE) &&
1164                             (found == 0)) {
1165                                 ccb->cgdl.periph_name[0] = '\0';
1166                                 ccb->cgdl.unit_number = 0;
1167                                 break;
1168                         }
1169                 }
1170
1171                 /* copy the result back out */
1172                 bcopy(ccb, addr, sizeof(union ccb));
1173
1174                 /* and release the ccb */
1175                 xpt_release_ccb(ccb);
1176
1177                 break;
1178         default:
1179                 error = ENOTTY;
1180                 break;
1181         }
1182         return(error);
1183 }
1184
1185 static void
1186 cam_periph_done_panic(struct cam_periph *periph, union ccb *done_ccb)
1187 {
1188
1189         panic("%s: already done with ccb %p", __func__, done_ccb);
1190 }
1191
1192 static void
1193 cam_periph_done(struct cam_periph *periph, union ccb *done_ccb)
1194 {
1195
1196         /* Caller will release the CCB */
1197         xpt_path_assert(done_ccb->ccb_h.path, MA_OWNED);
1198         done_ccb->ccb_h.cbfcnp = cam_periph_done_panic;
1199         wakeup(&done_ccb->ccb_h.cbfcnp);
1200 }
1201
1202 static void
1203 cam_periph_ccbwait(union ccb *ccb)
1204 {
1205
1206         if ((ccb->ccb_h.func_code & XPT_FC_QUEUED) != 0) {
1207                 while (ccb->ccb_h.cbfcnp != cam_periph_done_panic)
1208                         xpt_path_sleep(ccb->ccb_h.path, &ccb->ccb_h.cbfcnp,
1209                             PRIBIO, "cbwait", 0);
1210         }
1211         KASSERT(ccb->ccb_h.pinfo.index == CAM_UNQUEUED_INDEX &&
1212             (ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_INPROG,
1213             ("%s: proceeding with incomplete ccb: ccb=%p, func_code=%#x, "
1214              "status=%#x, index=%d", __func__, ccb, ccb->ccb_h.func_code,
1215              ccb->ccb_h.status, ccb->ccb_h.pinfo.index));
1216 }
1217
1218 /*
1219  * Dispatch a CCB and wait for it to complete.  If the CCB has set a
1220  * callback function (ccb->ccb_h.cbfcnp), it will be overwritten and lost.
1221  */
1222 int
1223 cam_periph_runccb(union ccb *ccb,
1224                   int (*error_routine)(union ccb *ccb,
1225                                        cam_flags camflags,
1226                                        u_int32_t sense_flags),
1227                   cam_flags camflags, u_int32_t sense_flags,
1228                   struct devstat *ds)
1229 {
1230         struct bintime *starttime;
1231         struct bintime ltime;
1232         int error;
1233         bool must_poll;
1234         uint32_t timeout = 1;
1235
1236         starttime = NULL;
1237         xpt_path_assert(ccb->ccb_h.path, MA_OWNED);
1238         KASSERT((ccb->ccb_h.flags & CAM_UNLOCKED) == 0,
1239             ("%s: ccb=%p, func_code=%#x, flags=%#x", __func__, ccb,
1240              ccb->ccb_h.func_code, ccb->ccb_h.flags));
1241
1242         /*
1243          * If the user has supplied a stats structure, and if we understand
1244          * this particular type of ccb, record the transaction start.
1245          */
1246         if (ds != NULL &&
1247             (ccb->ccb_h.func_code == XPT_SCSI_IO ||
1248             ccb->ccb_h.func_code == XPT_ATA_IO ||
1249             ccb->ccb_h.func_code == XPT_NVME_IO)) {
1250                 starttime = &ltime;
1251                 binuptime(starttime);
1252                 devstat_start_transaction(ds, starttime);
1253         }
1254
1255         /*
1256          * We must poll the I/O while we're dumping. The scheduler is normally
1257          * stopped for dumping, except when we call doadump from ddb. While the
1258          * scheduler is running in this case, we still need to poll the I/O to
1259          * avoid sleeping waiting for the ccb to complete.
1260          *
1261          * A panic triggered dump stops the scheduler, any callback from the
1262          * shutdown_post_sync event will run with the scheduler stopped, but
1263          * before we're officially dumping. To avoid hanging in adashutdown
1264          * initiated commands (or other similar situations), we have to test for
1265          * either SCHEDULER_STOPPED() here as well.
1266          *
1267          * To avoid locking problems, dumping/polling callers must call
1268          * without a periph lock held.
1269          */
1270         must_poll = dumping || SCHEDULER_STOPPED();
1271         ccb->ccb_h.cbfcnp = cam_periph_done;
1272
1273         /*
1274          * If we're polling, then we need to ensure that we have ample resources
1275          * in the periph.  cam_periph_error can reschedule the ccb by calling
1276          * xpt_action and returning ERESTART, so we have to effect the polling
1277          * in the do loop below.
1278          */
1279         if (must_poll) {
1280                 timeout = xpt_poll_setup(ccb);
1281         }
1282
1283         if (timeout == 0) {
1284                 ccb->ccb_h.status = CAM_RESRC_UNAVAIL;
1285                 error = EBUSY;
1286         } else {
1287                 xpt_action(ccb);
1288                 do {
1289                         if (must_poll) {
1290                                 xpt_pollwait(ccb, timeout);
1291                                 timeout = ccb->ccb_h.timeout * 10;
1292                         } else {
1293                                 cam_periph_ccbwait(ccb);
1294                         }
1295                         if ((ccb->ccb_h.status & CAM_STATUS_MASK) == CAM_REQ_CMP)
1296                                 error = 0;
1297                         else if (error_routine != NULL) {
1298                                 ccb->ccb_h.cbfcnp = cam_periph_done;
1299                                 error = (*error_routine)(ccb, camflags, sense_flags);
1300                         } else
1301                                 error = 0;
1302                 } while (error == ERESTART);
1303         }
1304
1305         if ((ccb->ccb_h.status & CAM_DEV_QFRZN) != 0) {
1306                 cam_release_devq(ccb->ccb_h.path,
1307                                  /* relsim_flags */0,
1308                                  /* openings */0,
1309                                  /* timeout */0,
1310                                  /* getcount_only */ FALSE);
1311                 ccb->ccb_h.status &= ~CAM_DEV_QFRZN;
1312         }
1313
1314         if (ds != NULL) {
1315                 uint32_t bytes;
1316                 devstat_tag_type tag;
1317                 bool valid = true;
1318
1319                 if (ccb->ccb_h.func_code == XPT_SCSI_IO) {
1320                         bytes = ccb->csio.dxfer_len - ccb->csio.resid;
1321                         tag = (devstat_tag_type)(ccb->csio.tag_action & 0x3);
1322                 } else if (ccb->ccb_h.func_code == XPT_ATA_IO) {
1323                         bytes = ccb->ataio.dxfer_len - ccb->ataio.resid;
1324                         tag = (devstat_tag_type)0;
1325                 } else if (ccb->ccb_h.func_code == XPT_NVME_IO) {
1326                         bytes = ccb->nvmeio.dxfer_len; /* NB: resid no possible */
1327                         tag = (devstat_tag_type)0;
1328                 } else {
1329                         valid = false;
1330                 }
1331                 if (valid)
1332                         devstat_end_transaction(ds, bytes, tag,
1333                             ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE) ?
1334                             DEVSTAT_NO_DATA : (ccb->ccb_h.flags & CAM_DIR_OUT) ?
1335                             DEVSTAT_WRITE : DEVSTAT_READ, NULL, starttime);
1336         }
1337
1338         return(error);
1339 }
1340
1341 void
1342 cam_freeze_devq(struct cam_path *path)
1343 {
1344         struct ccb_hdr ccb_h;
1345
1346         CAM_DEBUG(path, CAM_DEBUG_TRACE, ("cam_freeze_devq\n"));
1347         xpt_setup_ccb(&ccb_h, path, /*priority*/1);
1348         ccb_h.func_code = XPT_NOOP;
1349         ccb_h.flags = CAM_DEV_QFREEZE;
1350         xpt_action((union ccb *)&ccb_h);
1351 }
1352
1353 u_int32_t
1354 cam_release_devq(struct cam_path *path, u_int32_t relsim_flags,
1355                  u_int32_t openings, u_int32_t arg,
1356                  int getcount_only)
1357 {
1358         struct ccb_relsim crs;
1359
1360         CAM_DEBUG(path, CAM_DEBUG_TRACE, ("cam_release_devq(%u, %u, %u, %d)\n",
1361             relsim_flags, openings, arg, getcount_only));
1362         xpt_setup_ccb(&crs.ccb_h, path, CAM_PRIORITY_NORMAL);
1363         crs.ccb_h.func_code = XPT_REL_SIMQ;
1364         crs.ccb_h.flags = getcount_only ? CAM_DEV_QFREEZE : 0;
1365         crs.release_flags = relsim_flags;
1366         crs.openings = openings;
1367         crs.release_timeout = arg;
1368         xpt_action((union ccb *)&crs);
1369         return (crs.qfrozen_cnt);
1370 }
1371
1372 #define saved_ccb_ptr ppriv_ptr0
1373 static void
1374 camperiphdone(struct cam_periph *periph, union ccb *done_ccb)
1375 {
1376         union ccb      *saved_ccb;
1377         cam_status      status;
1378         struct scsi_start_stop_unit *scsi_cmd;
1379         int             error = 0, error_code, sense_key, asc, ascq;
1380
1381         scsi_cmd = (struct scsi_start_stop_unit *)
1382             &done_ccb->csio.cdb_io.cdb_bytes;
1383         status = done_ccb->ccb_h.status;
1384
1385         if ((status & CAM_STATUS_MASK) != CAM_REQ_CMP) {
1386                 if (scsi_extract_sense_ccb(done_ccb,
1387                     &error_code, &sense_key, &asc, &ascq)) {
1388                         /*
1389                          * If the error is "invalid field in CDB",
1390                          * and the load/eject flag is set, turn the
1391                          * flag off and try again.  This is just in
1392                          * case the drive in question barfs on the
1393                          * load eject flag.  The CAM code should set
1394                          * the load/eject flag by default for
1395                          * removable media.
1396                          */
1397                         if ((scsi_cmd->opcode == START_STOP_UNIT) &&
1398                             ((scsi_cmd->how & SSS_LOEJ) != 0) &&
1399                              (asc == 0x24) && (ascq == 0x00)) {
1400                                 scsi_cmd->how &= ~SSS_LOEJ;
1401                                 if (status & CAM_DEV_QFRZN) {
1402                                         cam_release_devq(done_ccb->ccb_h.path,
1403                                             0, 0, 0, 0);
1404                                         done_ccb->ccb_h.status &=
1405                                             ~CAM_DEV_QFRZN;
1406                                 }
1407                                 xpt_action(done_ccb);
1408                                 goto out;
1409                         }
1410                 }
1411                 error = cam_periph_error(done_ccb, 0,
1412                     SF_RETRY_UA | SF_NO_PRINT);
1413                 if (error == ERESTART)
1414                         goto out;
1415                 if (done_ccb->ccb_h.status & CAM_DEV_QFRZN) {
1416                         cam_release_devq(done_ccb->ccb_h.path, 0, 0, 0, 0);
1417                         done_ccb->ccb_h.status &= ~CAM_DEV_QFRZN;
1418                 }
1419         } else {
1420                 /*
1421                  * If we have successfully taken a device from the not
1422                  * ready to ready state, re-scan the device and re-get
1423                  * the inquiry information.  Many devices (mostly disks)
1424                  * don't properly report their inquiry information unless
1425                  * they are spun up.
1426                  */
1427                 if (scsi_cmd->opcode == START_STOP_UNIT)
1428                         xpt_async(AC_INQ_CHANGED, done_ccb->ccb_h.path, NULL);
1429         }
1430
1431         /*
1432          * After recovery action(s) completed, return to the original CCB.
1433          * If the recovery CCB has failed, considering its own possible
1434          * retries and recovery, assume we are back in state where we have
1435          * been originally, but without recovery hopes left.  In such case,
1436          * after the final attempt below, we cancel any further retries,
1437          * blocking by that also any new recovery attempts for this CCB,
1438          * and the result will be the final one returned to the CCB owher.
1439          */
1440         saved_ccb = (union ccb *)done_ccb->ccb_h.saved_ccb_ptr;
1441         bcopy(saved_ccb, done_ccb, sizeof(*done_ccb));
1442         xpt_free_ccb(saved_ccb);
1443         if (done_ccb->ccb_h.cbfcnp != camperiphdone)
1444                 periph->flags &= ~CAM_PERIPH_RECOVERY_INPROG;
1445         if (error != 0)
1446                 done_ccb->ccb_h.retry_count = 0;
1447         xpt_action(done_ccb);
1448
1449 out:
1450         /* Drop freeze taken due to CAM_DEV_QFREEZE flag set. */
1451         cam_release_devq(done_ccb->ccb_h.path, 0, 0, 0, 0);
1452 }
1453
1454 /*
1455  * Generic Async Event handler.  Peripheral drivers usually
1456  * filter out the events that require personal attention,
1457  * and leave the rest to this function.
1458  */
1459 void
1460 cam_periph_async(struct cam_periph *periph, u_int32_t code,
1461                  struct cam_path *path, void *arg)
1462 {
1463         switch (code) {
1464         case AC_LOST_DEVICE:
1465                 cam_periph_invalidate(periph);
1466                 break;
1467         default:
1468                 break;
1469         }
1470 }
1471
1472 void
1473 cam_periph_bus_settle(struct cam_periph *periph, u_int bus_settle)
1474 {
1475         struct ccb_getdevstats cgds;
1476
1477         xpt_setup_ccb(&cgds.ccb_h, periph->path, CAM_PRIORITY_NORMAL);
1478         cgds.ccb_h.func_code = XPT_GDEV_STATS;
1479         xpt_action((union ccb *)&cgds);
1480         cam_periph_freeze_after_event(periph, &cgds.last_reset, bus_settle);
1481 }
1482
1483 void
1484 cam_periph_freeze_after_event(struct cam_periph *periph,
1485                               struct timeval* event_time, u_int duration_ms)
1486 {
1487         struct timeval delta;
1488         struct timeval duration_tv;
1489
1490         if (!timevalisset(event_time))
1491                 return;
1492
1493         microtime(&delta);
1494         timevalsub(&delta, event_time);
1495         duration_tv.tv_sec = duration_ms / 1000;
1496         duration_tv.tv_usec = (duration_ms % 1000) * 1000;
1497         if (timevalcmp(&delta, &duration_tv, <)) {
1498                 timevalsub(&duration_tv, &delta);
1499
1500                 duration_ms = duration_tv.tv_sec * 1000;
1501                 duration_ms += duration_tv.tv_usec / 1000;
1502                 cam_freeze_devq(periph->path);
1503                 cam_release_devq(periph->path,
1504                                 RELSIM_RELEASE_AFTER_TIMEOUT,
1505                                 /*reduction*/0,
1506                                 /*timeout*/duration_ms,
1507                                 /*getcount_only*/0);
1508         }
1509
1510 }
1511
1512 static int
1513 camperiphscsistatuserror(union ccb *ccb, union ccb **orig_ccb,
1514     cam_flags camflags, u_int32_t sense_flags,
1515     int *openings, u_int32_t *relsim_flags,
1516     u_int32_t *timeout, u_int32_t *action, const char **action_string)
1517 {
1518         struct cam_periph *periph;
1519         int error;
1520
1521         switch (ccb->csio.scsi_status) {
1522         case SCSI_STATUS_OK:
1523         case SCSI_STATUS_COND_MET:
1524         case SCSI_STATUS_INTERMED:
1525         case SCSI_STATUS_INTERMED_COND_MET:
1526                 error = 0;
1527                 break;
1528         case SCSI_STATUS_CMD_TERMINATED:
1529         case SCSI_STATUS_CHECK_COND:
1530                 error = camperiphscsisenseerror(ccb, orig_ccb,
1531                                                 camflags,
1532                                                 sense_flags,
1533                                                 openings,
1534                                                 relsim_flags,
1535                                                 timeout,
1536                                                 action,
1537                                                 action_string);
1538                 break;
1539         case SCSI_STATUS_QUEUE_FULL:
1540         {
1541                 /* no decrement */
1542                 struct ccb_getdevstats cgds;
1543
1544                 /*
1545                  * First off, find out what the current
1546                  * transaction counts are.
1547                  */
1548                 xpt_setup_ccb(&cgds.ccb_h,
1549                               ccb->ccb_h.path,
1550                               CAM_PRIORITY_NORMAL);
1551                 cgds.ccb_h.func_code = XPT_GDEV_STATS;
1552                 xpt_action((union ccb *)&cgds);
1553
1554                 /*
1555                  * If we were the only transaction active, treat
1556                  * the QUEUE FULL as if it were a BUSY condition.
1557                  */
1558                 if (cgds.dev_active != 0) {
1559                         int total_openings;
1560
1561                         /*
1562                          * Reduce the number of openings to
1563                          * be 1 less than the amount it took
1564                          * to get a queue full bounded by the
1565                          * minimum allowed tag count for this
1566                          * device.
1567                          */
1568                         total_openings = cgds.dev_active + cgds.dev_openings;
1569                         *openings = cgds.dev_active;
1570                         if (*openings < cgds.mintags)
1571                                 *openings = cgds.mintags;
1572                         if (*openings < total_openings)
1573                                 *relsim_flags = RELSIM_ADJUST_OPENINGS;
1574                         else {
1575                                 /*
1576                                  * Some devices report queue full for
1577                                  * temporary resource shortages.  For
1578                                  * this reason, we allow a minimum
1579                                  * tag count to be entered via a
1580                                  * quirk entry to prevent the queue
1581                                  * count on these devices from falling
1582                                  * to a pessimisticly low value.  We
1583                                  * still wait for the next successful
1584                                  * completion, however, before queueing
1585                                  * more transactions to the device.
1586                                  */
1587                                 *relsim_flags = RELSIM_RELEASE_AFTER_CMDCMPLT;
1588                         }
1589                         *timeout = 0;
1590                         error = ERESTART;
1591                         *action &= ~SSQ_PRINT_SENSE;
1592                         break;
1593                 }
1594                 /* FALLTHROUGH */
1595         }
1596         case SCSI_STATUS_BUSY:
1597                 /*
1598                  * Restart the queue after either another
1599                  * command completes or a 1 second timeout.
1600                  */
1601                 periph = xpt_path_periph(ccb->ccb_h.path);
1602                 if (periph->flags & CAM_PERIPH_INVALID) {
1603                         error = EIO;
1604                         *action_string = "Periph was invalidated";
1605                 } else if ((sense_flags & SF_RETRY_BUSY) != 0 ||
1606                     ccb->ccb_h.retry_count > 0) {
1607                         if ((sense_flags & SF_RETRY_BUSY) == 0)
1608                                 ccb->ccb_h.retry_count--;
1609                         error = ERESTART;
1610                         *relsim_flags = RELSIM_RELEASE_AFTER_TIMEOUT
1611                                       | RELSIM_RELEASE_AFTER_CMDCMPLT;
1612                         *timeout = 1000;
1613                 } else {
1614                         error = EIO;
1615                         *action_string = "Retries exhausted";
1616                 }
1617                 break;
1618         case SCSI_STATUS_RESERV_CONFLICT:
1619         default:
1620                 error = EIO;
1621                 break;
1622         }
1623         return (error);
1624 }
1625
1626 static int
1627 camperiphscsisenseerror(union ccb *ccb, union ccb **orig,
1628     cam_flags camflags, u_int32_t sense_flags,
1629     int *openings, u_int32_t *relsim_flags,
1630     u_int32_t *timeout, u_int32_t *action, const char **action_string)
1631 {
1632         struct cam_periph *periph;
1633         union ccb *orig_ccb = ccb;
1634         int error, recoveryccb;
1635
1636 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
1637         if (ccb->ccb_h.func_code == XPT_SCSI_IO && ccb->csio.bio != NULL)
1638                 biotrack(ccb->csio.bio, __func__);
1639 #endif
1640
1641         periph = xpt_path_periph(ccb->ccb_h.path);
1642         recoveryccb = (ccb->ccb_h.cbfcnp == camperiphdone);
1643         if ((periph->flags & CAM_PERIPH_RECOVERY_INPROG) && !recoveryccb) {
1644                 /*
1645                  * If error recovery is already in progress, don't attempt
1646                  * to process this error, but requeue it unconditionally
1647                  * and attempt to process it once error recovery has
1648                  * completed.  This failed command is probably related to
1649                  * the error that caused the currently active error recovery
1650                  * action so our  current recovery efforts should also
1651                  * address this command.  Be aware that the error recovery
1652                  * code assumes that only one recovery action is in progress
1653                  * on a particular peripheral instance at any given time
1654                  * (e.g. only one saved CCB for error recovery) so it is
1655                  * imperitive that we don't violate this assumption.
1656                  */
1657                 error = ERESTART;
1658                 *action &= ~SSQ_PRINT_SENSE;
1659         } else {
1660                 scsi_sense_action err_action;
1661                 struct ccb_getdev cgd;
1662
1663                 /*
1664                  * Grab the inquiry data for this device.
1665                  */
1666                 xpt_setup_ccb(&cgd.ccb_h, ccb->ccb_h.path, CAM_PRIORITY_NORMAL);
1667                 cgd.ccb_h.func_code = XPT_GDEV_TYPE;
1668                 xpt_action((union ccb *)&cgd);
1669
1670                 err_action = scsi_error_action(&ccb->csio, &cgd.inq_data,
1671                     sense_flags);
1672                 error = err_action & SS_ERRMASK;
1673
1674                 /*
1675                  * Do not autostart sequential access devices
1676                  * to avoid unexpected tape loading.
1677                  */
1678                 if ((err_action & SS_MASK) == SS_START &&
1679                     SID_TYPE(&cgd.inq_data) == T_SEQUENTIAL) {
1680                         *action_string = "Will not autostart a "
1681                             "sequential access device";
1682                         goto sense_error_done;
1683                 }
1684
1685                 /*
1686                  * Avoid recovery recursion if recovery action is the same.
1687                  */
1688                 if ((err_action & SS_MASK) >= SS_START && recoveryccb) {
1689                         if (((err_action & SS_MASK) == SS_START &&
1690                              ccb->csio.cdb_io.cdb_bytes[0] == START_STOP_UNIT) ||
1691                             ((err_action & SS_MASK) == SS_TUR &&
1692                              (ccb->csio.cdb_io.cdb_bytes[0] == TEST_UNIT_READY))) {
1693                                 err_action = SS_RETRY|SSQ_DECREMENT_COUNT|EIO;
1694                                 *relsim_flags = RELSIM_RELEASE_AFTER_TIMEOUT;
1695                                 *timeout = 500;
1696                         }
1697                 }
1698
1699                 /*
1700                  * If the recovery action will consume a retry,
1701                  * make sure we actually have retries available.
1702                  */
1703                 if ((err_action & SSQ_DECREMENT_COUNT) != 0) {
1704                         if (ccb->ccb_h.retry_count > 0 &&
1705                             (periph->flags & CAM_PERIPH_INVALID) == 0)
1706                                 ccb->ccb_h.retry_count--;
1707                         else {
1708                                 *action_string = "Retries exhausted";
1709                                 goto sense_error_done;
1710                         }
1711                 }
1712
1713                 if ((err_action & SS_MASK) >= SS_START) {
1714                         /*
1715                          * Do common portions of commands that
1716                          * use recovery CCBs.
1717                          */
1718                         orig_ccb = xpt_alloc_ccb_nowait();
1719                         if (orig_ccb == NULL) {
1720                                 *action_string = "Can't allocate recovery CCB";
1721                                 goto sense_error_done;
1722                         }
1723                         /*
1724                          * Clear freeze flag for original request here, as
1725                          * this freeze will be dropped as part of ERESTART.
1726                          */
1727                         ccb->ccb_h.status &= ~CAM_DEV_QFRZN;
1728                         bcopy(ccb, orig_ccb, sizeof(*orig_ccb));
1729                 }
1730
1731                 switch (err_action & SS_MASK) {
1732                 case SS_NOP:
1733                         *action_string = "No recovery action needed";
1734                         error = 0;
1735                         break;
1736                 case SS_RETRY:
1737                         *action_string = "Retrying command (per sense data)";
1738                         error = ERESTART;
1739                         break;
1740                 case SS_FAIL:
1741                         *action_string = "Unretryable error";
1742                         break;
1743                 case SS_START:
1744                 {
1745                         int le;
1746
1747                         /*
1748                          * Send a start unit command to the device, and
1749                          * then retry the command.
1750                          */
1751                         *action_string = "Attempting to start unit";
1752                         periph->flags |= CAM_PERIPH_RECOVERY_INPROG;
1753
1754                         /*
1755                          * Check for removable media and set
1756                          * load/eject flag appropriately.
1757                          */
1758                         if (SID_IS_REMOVABLE(&cgd.inq_data))
1759                                 le = TRUE;
1760                         else
1761                                 le = FALSE;
1762
1763                         scsi_start_stop(&ccb->csio,
1764                                         /*retries*/1,
1765                                         camperiphdone,
1766                                         MSG_SIMPLE_Q_TAG,
1767                                         /*start*/TRUE,
1768                                         /*load/eject*/le,
1769                                         /*immediate*/FALSE,
1770                                         SSD_FULL_SIZE,
1771                                         /*timeout*/50000);
1772                         break;
1773                 }
1774                 case SS_TUR:
1775                 {
1776                         /*
1777                          * Send a Test Unit Ready to the device.
1778                          * If the 'many' flag is set, we send 120
1779                          * test unit ready commands, one every half
1780                          * second.  Otherwise, we just send one TUR.
1781                          * We only want to do this if the retry
1782                          * count has not been exhausted.
1783                          */
1784                         int retries;
1785
1786                         if ((err_action & SSQ_MANY) != 0) {
1787                                 *action_string = "Polling device for readiness";
1788                                 retries = 120;
1789                         } else {
1790                                 *action_string = "Testing device for readiness";
1791                                 retries = 1;
1792                         }
1793                         periph->flags |= CAM_PERIPH_RECOVERY_INPROG;
1794                         scsi_test_unit_ready(&ccb->csio,
1795                                              retries,
1796                                              camperiphdone,
1797                                              MSG_SIMPLE_Q_TAG,
1798                                              SSD_FULL_SIZE,
1799                                              /*timeout*/5000);
1800
1801                         /*
1802                          * Accomplish our 500ms delay by deferring
1803                          * the release of our device queue appropriately.
1804                          */
1805                         *relsim_flags = RELSIM_RELEASE_AFTER_TIMEOUT;
1806                         *timeout = 500;
1807                         break;
1808                 }
1809                 default:
1810                         panic("Unhandled error action %x", err_action);
1811                 }
1812
1813                 if ((err_action & SS_MASK) >= SS_START) {
1814                         /*
1815                          * Drop the priority, so that the recovery
1816                          * CCB is the first to execute.  Freeze the queue
1817                          * after this command is sent so that we can
1818                          * restore the old csio and have it queued in
1819                          * the proper order before we release normal
1820                          * transactions to the device.
1821                          */
1822                         ccb->ccb_h.pinfo.priority--;
1823                         ccb->ccb_h.flags |= CAM_DEV_QFREEZE;
1824                         ccb->ccb_h.saved_ccb_ptr = orig_ccb;
1825                         error = ERESTART;
1826                         *orig = orig_ccb;
1827                 }
1828
1829 sense_error_done:
1830                 *action = err_action;
1831         }
1832         return (error);
1833 }
1834
1835 /*
1836  * Generic error handler.  Peripheral drivers usually filter
1837  * out the errors that they handle in a unique manner, then
1838  * call this function.
1839  */
1840 int
1841 cam_periph_error(union ccb *ccb, cam_flags camflags,
1842                  u_int32_t sense_flags)
1843 {
1844         struct cam_path *newpath;
1845         union ccb  *orig_ccb, *scan_ccb;
1846         struct cam_periph *periph;
1847         const char *action_string;
1848         cam_status  status;
1849         int         frozen, error, openings, devctl_err;
1850         u_int32_t   action, relsim_flags, timeout;
1851
1852         action = SSQ_PRINT_SENSE;
1853         periph = xpt_path_periph(ccb->ccb_h.path);
1854         action_string = NULL;
1855         status = ccb->ccb_h.status;
1856         frozen = (status & CAM_DEV_QFRZN) != 0;
1857         status &= CAM_STATUS_MASK;
1858         devctl_err = openings = relsim_flags = timeout = 0;
1859         orig_ccb = ccb;
1860
1861         /* Filter the errors that should be reported via devctl */
1862         switch (ccb->ccb_h.status & CAM_STATUS_MASK) {
1863         case CAM_CMD_TIMEOUT:
1864         case CAM_REQ_ABORTED:
1865         case CAM_REQ_CMP_ERR:
1866         case CAM_REQ_TERMIO:
1867         case CAM_UNREC_HBA_ERROR:
1868         case CAM_DATA_RUN_ERR:
1869         case CAM_SCSI_STATUS_ERROR:
1870         case CAM_ATA_STATUS_ERROR:
1871         case CAM_SMP_STATUS_ERROR:
1872                 devctl_err++;
1873                 break;
1874         default:
1875                 break;
1876         }
1877
1878         switch (status) {
1879         case CAM_REQ_CMP:
1880                 error = 0;
1881                 action &= ~SSQ_PRINT_SENSE;
1882                 break;
1883         case CAM_SCSI_STATUS_ERROR:
1884                 error = camperiphscsistatuserror(ccb, &orig_ccb,
1885                     camflags, sense_flags, &openings, &relsim_flags,
1886                     &timeout, &action, &action_string);
1887                 break;
1888         case CAM_AUTOSENSE_FAIL:
1889                 error = EIO;    /* we have to kill the command */
1890                 break;
1891         case CAM_UA_ABORT:
1892         case CAM_UA_TERMIO:
1893         case CAM_MSG_REJECT_REC:
1894                 /* XXX Don't know that these are correct */
1895                 error = EIO;
1896                 break;
1897         case CAM_SEL_TIMEOUT:
1898                 if ((camflags & CAM_RETRY_SELTO) != 0) {
1899                         if (ccb->ccb_h.retry_count > 0 &&
1900                             (periph->flags & CAM_PERIPH_INVALID) == 0) {
1901                                 ccb->ccb_h.retry_count--;
1902                                 error = ERESTART;
1903
1904                                 /*
1905                                  * Wait a bit to give the device
1906                                  * time to recover before we try again.
1907                                  */
1908                                 relsim_flags = RELSIM_RELEASE_AFTER_TIMEOUT;
1909                                 timeout = periph_selto_delay;
1910                                 break;
1911                         }
1912                         action_string = "Retries exhausted";
1913                 }
1914                 /* FALLTHROUGH */
1915         case CAM_DEV_NOT_THERE:
1916                 error = ENXIO;
1917                 action = SSQ_LOST;
1918                 break;
1919         case CAM_REQ_INVALID:
1920         case CAM_PATH_INVALID:
1921         case CAM_NO_HBA:
1922         case CAM_PROVIDE_FAIL:
1923         case CAM_REQ_TOO_BIG:
1924         case CAM_LUN_INVALID:
1925         case CAM_TID_INVALID:
1926         case CAM_FUNC_NOTAVAIL:
1927                 error = EINVAL;
1928                 break;
1929         case CAM_SCSI_BUS_RESET:
1930         case CAM_BDR_SENT:
1931                 /*
1932                  * Commands that repeatedly timeout and cause these
1933                  * kinds of error recovery actions, should return
1934                  * CAM_CMD_TIMEOUT, which allows us to safely assume
1935                  * that this command was an innocent bystander to
1936                  * these events and should be unconditionally
1937                  * retried.
1938                  */
1939         case CAM_REQUEUE_REQ:
1940                 /* Unconditional requeue if device is still there */
1941                 if (periph->flags & CAM_PERIPH_INVALID) {
1942                         action_string = "Periph was invalidated";
1943                         error = EIO;
1944                 } else if (sense_flags & SF_NO_RETRY) {
1945                         error = EIO;
1946                         action_string = "Retry was blocked";
1947                 } else {
1948                         error = ERESTART;
1949                         action &= ~SSQ_PRINT_SENSE;
1950                 }
1951                 break;
1952         case CAM_RESRC_UNAVAIL:
1953                 /* Wait a bit for the resource shortage to abate. */
1954                 timeout = periph_noresrc_delay;
1955                 /* FALLTHROUGH */
1956         case CAM_BUSY:
1957                 if (timeout == 0) {
1958                         /* Wait a bit for the busy condition to abate. */
1959                         timeout = periph_busy_delay;
1960                 }
1961                 relsim_flags = RELSIM_RELEASE_AFTER_TIMEOUT;
1962                 /* FALLTHROUGH */
1963         case CAM_ATA_STATUS_ERROR:
1964         case CAM_REQ_CMP_ERR:
1965         case CAM_CMD_TIMEOUT:
1966         case CAM_UNEXP_BUSFREE:
1967         case CAM_UNCOR_PARITY:
1968         case CAM_DATA_RUN_ERR:
1969         default:
1970                 if (periph->flags & CAM_PERIPH_INVALID) {
1971                         error = EIO;
1972                         action_string = "Periph was invalidated";
1973                 } else if (ccb->ccb_h.retry_count == 0) {
1974                         error = EIO;
1975                         action_string = "Retries exhausted";
1976                 } else if (sense_flags & SF_NO_RETRY) {
1977                         error = EIO;
1978                         action_string = "Retry was blocked";
1979                 } else {
1980                         ccb->ccb_h.retry_count--;
1981                         error = ERESTART;
1982                 }
1983                 break;
1984         }
1985
1986         if ((sense_flags & SF_PRINT_ALWAYS) ||
1987             CAM_DEBUGGED(ccb->ccb_h.path, CAM_DEBUG_INFO))
1988                 action |= SSQ_PRINT_SENSE;
1989         else if (sense_flags & SF_NO_PRINT)
1990                 action &= ~SSQ_PRINT_SENSE;
1991         if ((action & SSQ_PRINT_SENSE) != 0)
1992                 cam_error_print(orig_ccb, CAM_ESF_ALL, CAM_EPF_ALL);
1993         if (error != 0 && (action & SSQ_PRINT_SENSE) != 0) {
1994                 if (error != ERESTART) {
1995                         if (action_string == NULL)
1996                                 action_string = "Unretryable error";
1997                         xpt_print(ccb->ccb_h.path, "Error %d, %s\n",
1998                             error, action_string);
1999                 } else if (action_string != NULL)
2000                         xpt_print(ccb->ccb_h.path, "%s\n", action_string);
2001                 else {
2002                         xpt_print(ccb->ccb_h.path,
2003                             "Retrying command, %d more tries remain\n",
2004                             ccb->ccb_h.retry_count);
2005                 }
2006         }
2007
2008         if (devctl_err && (error != 0 || (action & SSQ_PRINT_SENSE) != 0))
2009                 cam_periph_devctl_notify(orig_ccb);
2010
2011         if ((action & SSQ_LOST) != 0) {
2012                 lun_id_t lun_id;
2013
2014                 /*
2015                  * For a selection timeout, we consider all of the LUNs on
2016                  * the target to be gone.  If the status is CAM_DEV_NOT_THERE,
2017                  * then we only get rid of the device(s) specified by the
2018                  * path in the original CCB.
2019                  */
2020                 if (status == CAM_SEL_TIMEOUT)
2021                         lun_id = CAM_LUN_WILDCARD;
2022                 else
2023                         lun_id = xpt_path_lun_id(ccb->ccb_h.path);
2024
2025                 /* Should we do more if we can't create the path?? */
2026                 if (xpt_create_path(&newpath, periph,
2027                                     xpt_path_path_id(ccb->ccb_h.path),
2028                                     xpt_path_target_id(ccb->ccb_h.path),
2029                                     lun_id) == CAM_REQ_CMP) {
2030
2031                         /*
2032                          * Let peripheral drivers know that this
2033                          * device has gone away.
2034                          */
2035                         xpt_async(AC_LOST_DEVICE, newpath, NULL);
2036                         xpt_free_path(newpath);
2037                 }
2038         }
2039
2040         /* Broadcast UNIT ATTENTIONs to all periphs. */
2041         if ((action & SSQ_UA) != 0)
2042                 xpt_async(AC_UNIT_ATTENTION, orig_ccb->ccb_h.path, orig_ccb);
2043
2044         /* Rescan target on "Reported LUNs data has changed" */
2045         if ((action & SSQ_RESCAN) != 0) {
2046                 if (xpt_create_path(&newpath, NULL,
2047                                     xpt_path_path_id(ccb->ccb_h.path),
2048                                     xpt_path_target_id(ccb->ccb_h.path),
2049                                     CAM_LUN_WILDCARD) == CAM_REQ_CMP) {
2050
2051                         scan_ccb = xpt_alloc_ccb_nowait();
2052                         if (scan_ccb != NULL) {
2053                                 scan_ccb->ccb_h.path = newpath;
2054                                 scan_ccb->ccb_h.func_code = XPT_SCAN_TGT;
2055                                 scan_ccb->crcn.flags = 0;
2056                                 xpt_rescan(scan_ccb);
2057                         } else {
2058                                 xpt_print(newpath,
2059                                     "Can't allocate CCB to rescan target\n");
2060                                 xpt_free_path(newpath);
2061                         }
2062                 }
2063         }
2064
2065         /* Attempt a retry */
2066         if (error == ERESTART || error == 0) {
2067                 if (frozen != 0)
2068                         ccb->ccb_h.status &= ~CAM_DEV_QFRZN;
2069                 if (error == ERESTART)
2070                         xpt_action(ccb);
2071                 if (frozen != 0)
2072                         cam_release_devq(ccb->ccb_h.path,
2073                                          relsim_flags,
2074                                          openings,
2075                                          timeout,
2076                                          /*getcount_only*/0);
2077         }
2078
2079         return (error);
2080 }
2081
2082 #define CAM_PERIPH_DEVD_MSG_SIZE        256
2083
2084 static void
2085 cam_periph_devctl_notify(union ccb *ccb)
2086 {
2087         struct cam_periph *periph;
2088         struct ccb_getdev *cgd;
2089         struct sbuf sb;
2090         int serr, sk, asc, ascq;
2091         char *sbmsg, *type;
2092
2093         sbmsg = malloc(CAM_PERIPH_DEVD_MSG_SIZE, M_CAMPERIPH, M_NOWAIT);
2094         if (sbmsg == NULL)
2095                 return;
2096
2097         sbuf_new(&sb, sbmsg, CAM_PERIPH_DEVD_MSG_SIZE, SBUF_FIXEDLEN);
2098
2099         periph = xpt_path_periph(ccb->ccb_h.path);
2100         sbuf_printf(&sb, "device=%s%d ", periph->periph_name,
2101             periph->unit_number);
2102
2103         sbuf_printf(&sb, "serial=\"");
2104         if ((cgd = (struct ccb_getdev *)xpt_alloc_ccb_nowait()) != NULL) {
2105                 xpt_setup_ccb(&cgd->ccb_h, ccb->ccb_h.path,
2106                     CAM_PRIORITY_NORMAL);
2107                 cgd->ccb_h.func_code = XPT_GDEV_TYPE;
2108                 xpt_action((union ccb *)cgd);
2109
2110                 if (cgd->ccb_h.status == CAM_REQ_CMP)
2111                         sbuf_bcat(&sb, cgd->serial_num, cgd->serial_num_len);
2112                 xpt_free_ccb((union ccb *)cgd);
2113         }
2114         sbuf_printf(&sb, "\" ");
2115         sbuf_printf(&sb, "cam_status=\"0x%x\" ", ccb->ccb_h.status);
2116
2117         switch (ccb->ccb_h.status & CAM_STATUS_MASK) {
2118         case CAM_CMD_TIMEOUT:
2119                 sbuf_printf(&sb, "timeout=%d ", ccb->ccb_h.timeout);
2120                 type = "timeout";
2121                 break;
2122         case CAM_SCSI_STATUS_ERROR:
2123                 sbuf_printf(&sb, "scsi_status=%d ", ccb->csio.scsi_status);
2124                 if (scsi_extract_sense_ccb(ccb, &serr, &sk, &asc, &ascq))
2125                         sbuf_printf(&sb, "scsi_sense=\"%02x %02x %02x %02x\" ",
2126                             serr, sk, asc, ascq);
2127                 type = "error";
2128                 break;
2129         case CAM_ATA_STATUS_ERROR:
2130                 sbuf_printf(&sb, "RES=\"");
2131                 ata_res_sbuf(&ccb->ataio.res, &sb);
2132                 sbuf_printf(&sb, "\" ");
2133                 type = "error";
2134                 break;
2135         default:
2136                 type = "error";
2137                 break;
2138         }
2139
2140         if (ccb->ccb_h.func_code == XPT_SCSI_IO) {
2141                 sbuf_printf(&sb, "CDB=\"");
2142                 scsi_cdb_sbuf(scsiio_cdb_ptr(&ccb->csio), &sb);
2143                 sbuf_printf(&sb, "\" ");
2144         } else if (ccb->ccb_h.func_code == XPT_ATA_IO) {
2145                 sbuf_printf(&sb, "ACB=\"");
2146                 ata_cmd_sbuf(&ccb->ataio.cmd, &sb);
2147                 sbuf_printf(&sb, "\" ");
2148         }
2149
2150         if (sbuf_finish(&sb) == 0)
2151                 devctl_notify("CAM", "periph", type, sbuf_data(&sb));
2152         sbuf_delete(&sb);
2153         free(sbmsg, M_CAMPERIPH);
2154 }
2155
2156 /*
2157  * Sysctl to force an invalidation of the drive right now. Can be
2158  * called with CTLFLAG_MPSAFE since we take periph lock.
2159  */
2160 int
2161 cam_periph_invalidate_sysctl(SYSCTL_HANDLER_ARGS)
2162 {
2163         struct cam_periph *periph;
2164         int error, value;
2165
2166         periph = arg1;
2167         value = 0;
2168         error = sysctl_handle_int(oidp, &value, 0, req);
2169         if (error != 0 || req->newptr == NULL || value != 1)
2170                 return (error);
2171
2172         cam_periph_lock(periph);
2173         cam_periph_invalidate(periph);
2174         cam_periph_unlock(periph);
2175
2176         return (0);
2177 }