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