sys/cam/cam_periph.c

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