sys/cam/cam_periph.c

   1 /*-
   2  * Common functions for CAM "type" (peripheral) drivers.
   3  *
   4  * SPDX-License-Identifier: BSD-2-Clause
   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 #include <sys/param.h>
  34 #include <sys/systm.h>
  35 #include <sys/types.h>
  36 #include <sys/malloc.h>
  37 #include <sys/kernel.h>
  38 #include <sys/bio.h>
  39 #include <sys/conf.h>
  40 #include <sys/devctl.h>
  41 #include <sys/lock.h>
  42 #include <sys/mutex.h>
  43 #include <sys/buf.h>
  44 #include <sys/proc.h>
  45 #include <sys/devicestat.h>
  46 #include <sys/sbuf.h>
  47 #include <sys/sysctl.h>
  48 #include <vm/vm.h>
  49 #include <vm/vm_extern.h>
  50
  51 #include <cam/cam.h>
  52 #include <cam/cam_ccb.h>
  53 #include <cam/cam_compat.h>
  54 #include <cam/cam_queue.h>
  55 #include <cam/cam_xpt_periph.h>
  56 #include <cam/cam_xpt_internal.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, bool 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                                       const char *sn);
  73 static  void            camperiphdone(struct cam_periph *periph,
  74                                         union ccb *done_ccb);
  75 static  void            camperiphfree(struct cam_periph *periph);
  76 static int              camperiphscsistatuserror(union ccb *ccb,
  77                                                 union ccb **orig_ccb,
  78                                                  cam_flags camflags,
  79                                                  uint32_t sense_flags,
  80                                                  int *openings,
  81                                                  uint32_t *relsim_flags,
  82                                                  uint32_t *timeout,
  83                                                  uint32_t  *action,
  84                                                  const char **action_string);
  85 static  int             camperiphscsisenseerror(union ccb *ccb,
  86                                                 union ccb **orig_ccb,
  87                                                 cam_flags camflags,
  88                                                 uint32_t sense_flags,
  89                                                 int *openings,
  90                                                 uint32_t *relsim_flags,
  91                                                 uint32_t *timeout,
  92                                                 uint32_t *action,
  93                                                 const char **action_string);
  94 static void             cam_periph_devctl_notify(union ccb *ccb);
  95
  96 static int nperiph_drivers;
  97 static int initialized = 0;
  98 struct periph_driver **periph_drivers;
  99
 100 static MALLOC_DEFINE(M_CAMPERIPH, "CAM periph", "CAM peripheral buffers");
 101
 102 static int periph_selto_delay = 1000;
 103 TUNABLE_INT("kern.cam.periph_selto_delay", &periph_selto_delay);
 104 static int periph_noresrc_delay = 500;
 105 TUNABLE_INT("kern.cam.periph_noresrc_delay", &periph_noresrc_delay);
 106 static int periph_busy_delay = 500;
 107 TUNABLE_INT("kern.cam.periph_busy_delay", &periph_busy_delay);
 108
 109 static u_int periph_mapmem_thresh = 65536;
 110 SYSCTL_UINT(_kern_cam, OID_AUTO, mapmem_thresh, CTLFLAG_RWTUN,
 111     &periph_mapmem_thresh, 0, "Threshold for user-space buffer mapping");
 112
 113 void
 114 periphdriver_register(void *data)
 115 {
 116         struct periph_driver *drv = (struct periph_driver *)data;
 117         struct periph_driver **newdrivers, **old;
 118         int ndrivers;
 119
 120 again:
 121         ndrivers = nperiph_drivers + 2;
 122         newdrivers = malloc(sizeof(*newdrivers) * ndrivers, M_CAMPERIPH,
 123                             M_WAITOK);
 124         xpt_lock_buses();
 125         if (ndrivers != nperiph_drivers + 2) {
 126                 /*
 127                  * Lost race against itself; go around.
 128                  */
 129                 xpt_unlock_buses();
 130                 free(newdrivers, M_CAMPERIPH);
 131                 goto again;
 132         }
 133         if (periph_drivers)
 134                 bcopy(periph_drivers, newdrivers,
 135                       sizeof(*newdrivers) * nperiph_drivers);
 136         newdrivers[nperiph_drivers] = drv;
 137         newdrivers[nperiph_drivers + 1] = NULL;
 138         old = periph_drivers;
 139         periph_drivers = newdrivers;
 140         nperiph_drivers++;
 141         xpt_unlock_buses();
 142         if (old)
 143                 free(old, M_CAMPERIPH);
 144         /* If driver marked as early or it is late now, initialize it. */
 145         if (((drv->flags & CAM_PERIPH_DRV_EARLY) != 0 && initialized > 0) ||
 146             initialized > 1)
 147                 (*drv->init)();
 148 }
 149
 150 int
 151 periphdriver_unregister(void *data)
 152 {
 153         struct periph_driver *drv = (struct periph_driver *)data;
 154         int error, n;
 155
 156         /* If driver marked as early or it is late now, deinitialize it. */
 157         if (((drv->flags & CAM_PERIPH_DRV_EARLY) != 0 && initialized > 0) ||
 158             initialized > 1) {
 159                 if (drv->deinit == NULL) {
 160                         printf("CAM periph driver '%s' doesn't have deinit.\n",
 161                             drv->driver_name);
 162                         return (EOPNOTSUPP);
 163                 }
 164                 error = drv->deinit();
 165                 if (error != 0)
 166                         return (error);
 167         }
 168
 169         xpt_lock_buses();
 170         for (n = 0; n < nperiph_drivers && periph_drivers[n] != drv; n++)
 171                 ;
 172         KASSERT(n < nperiph_drivers,
 173             ("Periph driver '%s' was not registered", drv->driver_name));
 174         for (; n + 1 < nperiph_drivers; n++)
 175                 periph_drivers[n] = periph_drivers[n + 1];
 176         periph_drivers[n + 1] = NULL;
 177         nperiph_drivers--;
 178         xpt_unlock_buses();
 179         return (0);
 180 }
 181
 182 void
 183 periphdriver_init(int level)
 184 {
 185         int     i, early;
 186
 187         initialized = max(initialized, level);
 188         for (i = 0; periph_drivers[i] != NULL; i++) {
 189                 early = (periph_drivers[i]->flags & CAM_PERIPH_DRV_EARLY) ? 1 : 2;
 190                 if (early == initialized)
 191                         (*periph_drivers[i]->init)();
 192         }
 193 }
 194
 195 cam_status
 196 cam_periph_alloc(periph_ctor_t *periph_ctor,
 197                  periph_oninv_t *periph_oninvalidate,
 198                  periph_dtor_t *periph_dtor, periph_start_t *periph_start,
 199                  char *name, cam_periph_type type, struct cam_path *path,
 200                  ac_callback_t *ac_callback, ac_code code, void *arg)
 201 {
 202         struct          periph_driver **p_drv;
 203         struct          cam_sim *sim;
 204         struct          cam_periph *periph;
 205         struct          cam_periph *cur_periph;
 206         path_id_t       path_id;
 207         target_id_t     target_id;
 208         lun_id_t        lun_id;
 209         cam_status      status;
 210         u_int           init_level;
 211
 212         init_level = 0;
 213         /*
 214          * Handle Hot-Plug scenarios.  If there is already a peripheral
 215          * of our type assigned to this path, we are likely waiting for
 216          * final close on an old, invalidated, peripheral.  If this is
 217          * the case, queue up a deferred call to the peripheral's async
 218          * handler.  If it looks like a mistaken re-allocation, complain.
 219          */
 220         if ((periph = cam_periph_find(path, name)) != NULL) {
 221                 if ((periph->flags & CAM_PERIPH_INVALID) != 0
 222                  && (periph->flags & CAM_PERIPH_NEW_DEV_FOUND) == 0) {
 223                         periph->flags |= CAM_PERIPH_NEW_DEV_FOUND;
 224                         periph->deferred_callback = ac_callback;
 225                         periph->deferred_ac = code;
 226                         return (CAM_REQ_INPROG);
 227                 } else {
 228                         printf("cam_periph_alloc: attempt to re-allocate "
 229                                "valid device %s%d rejected flags %#x "
 230                                "refcount %d\n", periph->periph_name,
 231                                periph->unit_number, periph->flags,
 232                                periph->refcount);
 233                 }
 234                 return (CAM_REQ_INVALID);
 235         }
 236
 237         periph = (struct cam_periph *)malloc(sizeof(*periph), M_CAMPERIPH,
 238                                              M_NOWAIT|M_ZERO);
 239
 240         if (periph == NULL)
 241                 return (CAM_RESRC_UNAVAIL);
 242
 243         init_level++;
 244
 245         sim = xpt_path_sim(path);
 246         path_id = xpt_path_path_id(path);
 247         target_id = xpt_path_target_id(path);
 248         lun_id = xpt_path_lun_id(path);
 249         periph->periph_start = periph_start;
 250         periph->periph_dtor = periph_dtor;
 251         periph->periph_oninval = periph_oninvalidate;
 252         periph->type = type;
 253         periph->periph_name = name;
 254         periph->scheduled_priority = CAM_PRIORITY_NONE;
 255         periph->immediate_priority = CAM_PRIORITY_NONE;
 256         periph->refcount = 1;           /* Dropped by invalidation. */
 257         periph->sim = sim;
 258         SLIST_INIT(&periph->ccb_list);
 259         status = xpt_create_path(&path, periph, path_id, target_id, lun_id);
 260         if (status != CAM_REQ_CMP)
 261                 goto failure;
 262         periph->path = path;
 263
 264         xpt_lock_buses();
 265         for (p_drv = periph_drivers; *p_drv != NULL; p_drv++) {
 266                 if (strcmp((*p_drv)->driver_name, name) == 0)
 267                         break;
 268         }
 269         if (*p_drv == NULL) {
 270                 printf("cam_periph_alloc: invalid periph name '%s'\n", name);
 271                 xpt_unlock_buses();
 272                 xpt_free_path(periph->path);
 273                 free(periph, M_CAMPERIPH);
 274                 return (CAM_REQ_INVALID);
 275         }
 276         periph->unit_number = camperiphunit(*p_drv, path_id, target_id, lun_id,
 277             path->device->serial_num);
 278         cur_periph = TAILQ_FIRST(&(*p_drv)->units);
 279         while (cur_periph != NULL
 280             && cur_periph->unit_number < periph->unit_number)
 281                 cur_periph = TAILQ_NEXT(cur_periph, unit_links);
 282         if (cur_periph != NULL) {
 283                 KASSERT(cur_periph->unit_number != periph->unit_number,
 284                     ("duplicate units on periph list"));
 285                 TAILQ_INSERT_BEFORE(cur_periph, periph, unit_links);
 286         } else {
 287                 TAILQ_INSERT_TAIL(&(*p_drv)->units, periph, unit_links);
 288                 (*p_drv)->generation++;
 289         }
 290         xpt_unlock_buses();
 291
 292         init_level++;
 293
 294         status = xpt_add_periph(periph);
 295         if (status != CAM_REQ_CMP)
 296                 goto failure;
 297
 298         init_level++;
 299         CAM_DEBUG(periph->path, CAM_DEBUG_INFO, ("Periph created\n"));
 300
 301         status = periph_ctor(periph, arg);
 302
 303         if (status == CAM_REQ_CMP)
 304                 init_level++;
 305
 306 failure:
 307         switch (init_level) {
 308         case 4:
 309                 /* Initialized successfully */
 310                 break;
 311         case 3:
 312                 CAM_DEBUG(periph->path, CAM_DEBUG_INFO, ("Periph destroyed\n"));
 313                 xpt_remove_periph(periph);
 314                 /* FALLTHROUGH */
 315         case 2:
 316                 xpt_lock_buses();
 317                 TAILQ_REMOVE(&(*p_drv)->units, periph, unit_links);
 318                 xpt_unlock_buses();
 319                 xpt_free_path(periph->path);
 320                 /* FALLTHROUGH */
 321         case 1:
 322                 free(periph, M_CAMPERIPH);
 323                 /* FALLTHROUGH */
 324         case 0:
 325                 /* No cleanup to perform. */
 326                 break;
 327         default:
 328                 panic("%s: Unknown init level", __func__);
 329         }
 330         return(status);
 331 }
 332
 333 /*
 334  * Find a peripheral structure with the specified path, target, lun,
 335  * and (optionally) type.  If the name is NULL, this function will return
 336  * the first peripheral driver that matches the specified path.
 337  */
 338 struct cam_periph *
 339 cam_periph_find(struct cam_path *path, char *name)
 340 {
 341         struct periph_driver **p_drv;
 342         struct cam_periph *periph;
 343
 344         xpt_lock_buses();
 345         for (p_drv = periph_drivers; *p_drv != NULL; p_drv++) {
 346                 if (name != NULL && (strcmp((*p_drv)->driver_name, name) != 0))
 347                         continue;
 348
 349                 TAILQ_FOREACH(periph, &(*p_drv)->units, unit_links) {
 350                         if (xpt_path_comp(periph->path, path) == 0) {
 351                                 xpt_unlock_buses();
 352                                 cam_periph_assert(periph, MA_OWNED);
 353                                 return(periph);
 354                         }
 355                 }
 356                 if (name != NULL) {
 357                         xpt_unlock_buses();
 358                         return(NULL);
 359                 }
 360         }
 361         xpt_unlock_buses();
 362         return(NULL);
 363 }
 364
 365 /*
 366  * Find peripheral driver instances attached to the specified path.
 367  */
 368 int
 369 cam_periph_list(struct cam_path *path, struct sbuf *sb)
 370 {
 371         struct sbuf local_sb;
 372         struct periph_driver **p_drv;
 373         struct cam_periph *periph;
 374         int count;
 375         int sbuf_alloc_len;
 376
 377         sbuf_alloc_len = 16;
 378 retry:
 379         sbuf_new(&local_sb, NULL, sbuf_alloc_len, SBUF_FIXEDLEN);
 380         count = 0;
 381         xpt_lock_buses();
 382         for (p_drv = periph_drivers; *p_drv != NULL; p_drv++) {
 383                 TAILQ_FOREACH(periph, &(*p_drv)->units, unit_links) {
 384                         if (xpt_path_comp(periph->path, path) != 0)
 385                                 continue;
 386
 387                         if (sbuf_len(&local_sb) != 0)
 388                                 sbuf_cat(&local_sb, ",");
 389
 390                         sbuf_printf(&local_sb, "%s%d", periph->periph_name,
 391                                     periph->unit_number);
 392
 393                         if (sbuf_error(&local_sb) == ENOMEM) {
 394                                 sbuf_alloc_len *= 2;
 395                                 xpt_unlock_buses();
 396                                 sbuf_delete(&local_sb);
 397                                 goto retry;
 398                         }
 399                         count++;
 400                 }
 401         }
 402         xpt_unlock_buses();
 403         sbuf_finish(&local_sb);
 404         if (sbuf_len(sb) != 0)
 405                 sbuf_cat(sb, ",");
 406         sbuf_cat(sb, sbuf_data(&local_sb));
 407         sbuf_delete(&local_sb);
 408         return (count);
 409 }
 410
 411 int
 412 cam_periph_acquire(struct cam_periph *periph)
 413 {
 414         int status;
 415
 416         if (periph == NULL)
 417                 return (EINVAL);
 418
 419         status = ENOENT;
 420         xpt_lock_buses();
 421         if ((periph->flags & CAM_PERIPH_INVALID) == 0) {
 422                 periph->refcount++;
 423                 status = 0;
 424         }
 425         xpt_unlock_buses();
 426
 427         return (status);
 428 }
 429
 430 void
 431 cam_periph_doacquire(struct cam_periph *periph)
 432 {
 433
 434         xpt_lock_buses();
 435         KASSERT(periph->refcount >= 1,
 436             ("cam_periph_doacquire() with refcount == %d", periph->refcount));
 437         periph->refcount++;
 438         xpt_unlock_buses();
 439 }
 440
 441 void
 442 cam_periph_release_locked_buses(struct cam_periph *periph)
 443 {
 444
 445         cam_periph_assert(periph, MA_OWNED);
 446         KASSERT(periph->refcount >= 1, ("periph->refcount >= 1"));
 447         if (--periph->refcount == 0)
 448                 camperiphfree(periph);
 449 }
 450
 451 void
 452 cam_periph_release_locked(struct cam_periph *periph)
 453 {
 454
 455         if (periph == NULL)
 456                 return;
 457
 458         xpt_lock_buses();
 459         cam_periph_release_locked_buses(periph);
 460         xpt_unlock_buses();
 461 }
 462
 463 void
 464 cam_periph_release(struct cam_periph *periph)
 465 {
 466         struct mtx *mtx;
 467
 468         if (periph == NULL)
 469                 return;
 470
 471         cam_periph_assert(periph, MA_NOTOWNED);
 472         mtx = cam_periph_mtx(periph);
 473         mtx_lock(mtx);
 474         cam_periph_release_locked(periph);
 475         mtx_unlock(mtx);
 476 }
 477
 478 /*
 479  * hold/unhold act as mutual exclusion for sections of the code that
 480  * need to sleep and want to make sure that other sections that
 481  * will interfere are held off. This only protects exclusive sections
 482  * from each other.
 483  */
 484 int
 485 cam_periph_hold(struct cam_periph *periph, int priority)
 486 {
 487         int error;
 488
 489         /*
 490          * Increment the reference count on the peripheral
 491          * while we wait for our lock attempt to succeed
 492          * to ensure the peripheral doesn't disappear out
 493          * from user us while we sleep.
 494          */
 495
 496         if (cam_periph_acquire(periph) != 0)
 497                 return (ENXIO);
 498
 499         cam_periph_assert(periph, MA_OWNED);
 500         while ((periph->flags & CAM_PERIPH_LOCKED) != 0) {
 501                 periph->flags |= CAM_PERIPH_LOCK_WANTED;
 502                 if ((error = cam_periph_sleep(periph, periph, priority,
 503                     "caplck", 0)) != 0) {
 504                         cam_periph_release_locked(periph);
 505                         return (error);
 506                 }
 507                 if (periph->flags & CAM_PERIPH_INVALID) {
 508                         cam_periph_release_locked(periph);
 509                         return (ENXIO);
 510                 }
 511         }
 512
 513         periph->flags |= CAM_PERIPH_LOCKED;
 514         return (0);
 515 }
 516
 517 void
 518 cam_periph_unhold(struct cam_periph *periph)
 519 {
 520
 521         cam_periph_assert(periph, MA_OWNED);
 522
 523         periph->flags &= ~CAM_PERIPH_LOCKED;
 524         if ((periph->flags & CAM_PERIPH_LOCK_WANTED) != 0) {
 525                 periph->flags &= ~CAM_PERIPH_LOCK_WANTED;
 526                 wakeup(periph);
 527         }
 528
 529         cam_periph_release_locked(periph);
 530 }
 531
 532 void
 533 cam_periph_hold_boot(struct cam_periph *periph)
 534 {
 535
 536         root_mount_hold_token(periph->periph_name, &periph->periph_rootmount);
 537 }
 538
 539 void
 540 cam_periph_release_boot(struct cam_periph *periph)
 541 {
 542
 543         root_mount_rel(&periph->periph_rootmount);
 544 }
 545
 546 /*
 547  * Look for the next unit number that is not currently in use for this
 548  * peripheral type starting at "newunit".  Also exclude unit numbers that
 549  * are reserved by for future "hardwiring" unless we already know that this
 550  * is a potential wired device.  Only assume that the device is "wired" the
 551  * first time through the loop since after that we'll be looking at unit
 552  * numbers that did not match a wiring entry.
 553  */
 554 static u_int
 555 camperiphnextunit(struct periph_driver *p_drv, u_int newunit, bool wired,
 556                   path_id_t pathid, target_id_t target, lun_id_t lun)
 557 {
 558         struct  cam_periph *periph;
 559         char    *periph_name;
 560         int     i, val, dunit, r;
 561         const char *dname, *strval;
 562
 563         periph_name = p_drv->driver_name;
 564         for (;;newunit++) {
 565                 for (periph = TAILQ_FIRST(&p_drv->units);
 566                      periph != NULL && periph->unit_number != newunit;
 567                      periph = TAILQ_NEXT(periph, unit_links))
 568                         ;
 569
 570                 if (periph != NULL && periph->unit_number == newunit) {
 571                         if (wired) {
 572                                 xpt_print(periph->path, "Duplicate Wired "
 573                                     "Device entry!\n");
 574                                 xpt_print(periph->path, "Second device (%s "
 575                                     "device at scbus%d target %d lun %d) will "
 576                                     "not be wired\n", periph_name, pathid,
 577                                     target, lun);
 578                                 wired = false;
 579                         }
 580                         continue;
 581                 }
 582                 if (wired)
 583                         break;
 584
 585                 /*
 586                  * Don't allow the mere presence of any attributes of a device
 587                  * means that it is for a wired down entry. Instead, insist that
 588                  * one of the matching criteria from camperiphunit be present
 589                  * for the device.
 590                  */
 591                 i = 0;
 592                 dname = periph_name;
 593                 for (;;) {
 594                         r = resource_find_dev(&i, dname, &dunit, NULL, NULL);
 595                         if (r != 0)
 596                                 break;
 597
 598                         if (newunit != dunit)
 599                                 continue;
 600                         if (resource_string_value(dname, dunit, "sn", &strval) == 0 ||
 601                             resource_int_value(dname, dunit, "lun", &val) == 0 ||
 602                             resource_int_value(dname, dunit, "target", &val) == 0 ||
 603                             resource_string_value(dname, dunit, "at", &strval) == 0)
 604                                 break;
 605                 }
 606                 if (r != 0)
 607                         break;
 608         }
 609         return (newunit);
 610 }
 611
 612 static u_int
 613 camperiphunit(struct periph_driver *p_drv, path_id_t pathid,
 614     target_id_t target, lun_id_t lun, const char *sn)
 615 {
 616         bool    wired = false;
 617         u_int   unit;
 618         int     i, val, dunit;
 619         const char *dname, *strval;
 620         char    pathbuf[32], *periph_name;
 621
 622         periph_name = p_drv->driver_name;
 623         snprintf(pathbuf, sizeof(pathbuf), "scbus%d", pathid);
 624         unit = 0;
 625         i = 0;
 626         dname = periph_name;
 627
 628         for (wired = false; resource_find_dev(&i, dname, &dunit, NULL, NULL) == 0;
 629              wired = false) {
 630                 if (resource_string_value(dname, dunit, "at", &strval) == 0) {
 631                         if (strcmp(strval, pathbuf) != 0)
 632                                 continue;
 633                         wired = true;
 634                 }
 635                 if (resource_int_value(dname, dunit, "target", &val) == 0) {
 636                         if (val != target)
 637                                 continue;
 638                         wired = true;
 639                 }
 640                 if (resource_int_value(dname, dunit, "lun", &val) == 0) {
 641                         if (val != lun)
 642                                 continue;
 643                         wired = true;
 644                 }
 645                 if (resource_string_value(dname, dunit, "sn", &strval) == 0) {
 646                         if (sn == NULL || strcmp(strval, sn) != 0)
 647                                 continue;
 648                         wired = true;
 649                 }
 650                 if (wired) {
 651                         unit = dunit;
 652                         break;
 653                 }
 654         }
 655
 656         /*
 657          * Either start from 0 looking for the next unit or from
 658          * the unit number given in the resource config.  This way,
 659          * if we have wildcard matches, we don't return the same
 660          * unit number twice.
 661          */
 662         unit = camperiphnextunit(p_drv, unit, wired, pathid, target, lun);
 663
 664         return (unit);
 665 }
 666
 667 void
 668 cam_periph_invalidate(struct cam_periph *periph)
 669 {
 670
 671         cam_periph_assert(periph, MA_OWNED);
 672         /*
 673          * We only tear down the device the first time a peripheral is
 674          * invalidated.
 675          */
 676         if ((periph->flags & CAM_PERIPH_INVALID) != 0)
 677                 return;
 678
 679         CAM_DEBUG(periph->path, CAM_DEBUG_INFO, ("Periph invalidated\n"));
 680         if ((periph->flags & CAM_PERIPH_ANNOUNCED) && !rebooting) {
 681                 struct sbuf sb;
 682                 char buffer[160];
 683
 684                 sbuf_new(&sb, buffer, 160, SBUF_FIXEDLEN);
 685                 xpt_denounce_periph_sbuf(periph, &sb);
 686                 sbuf_finish(&sb);
 687                 sbuf_putbuf(&sb);
 688         }
 689         periph->flags |= CAM_PERIPH_INVALID;
 690         periph->flags &= ~CAM_PERIPH_NEW_DEV_FOUND;
 691         if (periph->periph_oninval != NULL)
 692                 periph->periph_oninval(periph);
 693         cam_periph_release_locked(periph);
 694 }
 695
 696 static void
 697 camperiphfree(struct cam_periph *periph)
 698 {
 699         struct periph_driver **p_drv;
 700         struct periph_driver *drv;
 701
 702         cam_periph_assert(periph, MA_OWNED);
 703         KASSERT(periph->periph_allocating == 0, ("%s%d: freed while allocating",
 704             periph->periph_name, periph->unit_number));
 705         for (p_drv = periph_drivers; *p_drv != NULL; p_drv++) {
 706                 if (strcmp((*p_drv)->driver_name, periph->periph_name) == 0)
 707                         break;
 708         }
 709         if (*p_drv == NULL) {
 710                 printf("camperiphfree: attempt to free non-existant periph\n");
 711                 return;
 712         }
 713         /*
 714          * Cache a pointer to the periph_driver structure.  If a
 715          * periph_driver is added or removed from the array (see
 716          * periphdriver_register()) while we drop the toplogy lock
 717          * below, p_drv may change.  This doesn't protect against this
 718          * particular periph_driver going away.  That will require full
 719          * reference counting in the periph_driver infrastructure.
 720          */
 721         drv = *p_drv;
 722
 723         /*
 724          * We need to set this flag before dropping the topology lock, to
 725          * let anyone who is traversing the list that this peripheral is
 726          * about to be freed, and there will be no more reference count
 727          * checks.
 728          */
 729         periph->flags |= CAM_PERIPH_FREE;
 730
 731         /*
 732          * The peripheral destructor semantics dictate calling with only the
 733          * SIM mutex held.  Since it might sleep, it should not be called
 734          * with the topology lock held.
 735          */
 736         xpt_unlock_buses();
 737
 738         /*
 739          * We need to call the peripheral destructor prior to removing the
 740          * peripheral from the list.  Otherwise, we risk running into a
 741          * scenario where the peripheral unit number may get reused
 742          * (because it has been removed from the list), but some resources
 743          * used by the peripheral are still hanging around.  In particular,
 744          * the devfs nodes used by some peripherals like the pass(4) driver
 745          * aren't fully cleaned up until the destructor is run.  If the
 746          * unit number is reused before the devfs instance is fully gone,
 747          * devfs will panic.
 748          */
 749         if (periph->periph_dtor != NULL)
 750                 periph->periph_dtor(periph);
 751
 752         /*
 753          * The peripheral list is protected by the topology lock. We have to
 754          * remove the periph from the drv list before we call deferred_ac. The
 755          * AC_FOUND_DEVICE callback won't create a new periph if it's still there.
 756          */
 757         xpt_lock_buses();
 758
 759         TAILQ_REMOVE(&drv->units, periph, unit_links);
 760         drv->generation++;
 761
 762         xpt_remove_periph(periph);
 763
 764         xpt_unlock_buses();
 765         if ((periph->flags & CAM_PERIPH_ANNOUNCED) && !rebooting)
 766                 xpt_print(periph->path, "Periph destroyed\n");
 767         else
 768                 CAM_DEBUG(periph->path, CAM_DEBUG_INFO, ("Periph destroyed\n"));
 769
 770         if (periph->flags & CAM_PERIPH_NEW_DEV_FOUND) {
 771                 union ccb ccb;
 772                 void *arg;
 773
 774                 memset(&ccb, 0, sizeof(ccb));
 775                 switch (periph->deferred_ac) {
 776                 case AC_FOUND_DEVICE:
 777                         ccb.ccb_h.func_code = XPT_GDEV_TYPE;
 778                         xpt_setup_ccb(&ccb.ccb_h, periph->path, CAM_PRIORITY_NORMAL);
 779                         xpt_action(&ccb);
 780                         arg = &ccb;
 781                         break;
 782                 case AC_PATH_REGISTERED:
 783                         xpt_path_inq(&ccb.cpi, periph->path);
 784                         arg = &ccb;
 785                         break;
 786                 default:
 787                         arg = NULL;
 788                         break;
 789                 }
 790                 periph->deferred_callback(NULL, periph->deferred_ac,
 791                                           periph->path, arg);
 792         }
 793         xpt_free_path(periph->path);
 794         free(periph, M_CAMPERIPH);
 795         xpt_lock_buses();
 796 }
 797
 798 /*
 799  * Map user virtual pointers into kernel virtual address space, so we can
 800  * access the memory.  This is now a generic function that centralizes most
 801  * of the sanity checks on the data flags, if any.
 802  * This also only works for up to maxphys memory.  Since we use
 803  * buffers to map stuff in and out, we're limited to the buffer size.
 804  */
 805 int
 806 cam_periph_mapmem(union ccb *ccb, struct cam_periph_map_info *mapinfo,
 807     u_int maxmap)
 808 {
 809         int numbufs, i;
 810         uint8_t **data_ptrs[CAM_PERIPH_MAXMAPS];
 811         uint32_t lengths[CAM_PERIPH_MAXMAPS];
 812         uint32_t dirs[CAM_PERIPH_MAXMAPS];
 813
 814         bzero(mapinfo, sizeof(*mapinfo));
 815         if (maxmap == 0)
 816                 maxmap = DFLTPHYS;      /* traditional default */
 817         else if (maxmap > maxphys)
 818                 maxmap = maxphys;       /* for safety */
 819         switch(ccb->ccb_h.func_code) {
 820         case XPT_DEV_MATCH:
 821                 if (ccb->cdm.match_buf_len == 0) {
 822                         printf("cam_periph_mapmem: invalid match buffer "
 823                                "length 0\n");
 824                         return(EINVAL);
 825                 }
 826                 if (ccb->cdm.pattern_buf_len > 0) {
 827                         data_ptrs[0] = (uint8_t **)&ccb->cdm.patterns;
 828                         lengths[0] = ccb->cdm.pattern_buf_len;
 829                         dirs[0] = CAM_DIR_OUT;
 830                         data_ptrs[1] = (uint8_t **)&ccb->cdm.matches;
 831                         lengths[1] = ccb->cdm.match_buf_len;
 832                         dirs[1] = CAM_DIR_IN;
 833                         numbufs = 2;
 834                 } else {
 835                         data_ptrs[0] = (uint8_t **)&ccb->cdm.matches;
 836                         lengths[0] = ccb->cdm.match_buf_len;
 837                         dirs[0] = CAM_DIR_IN;
 838                         numbufs = 1;
 839                 }
 840                 /*
 841                  * This request will not go to the hardware, no reason
 842                  * to be so strict. vmapbuf() is able to map up to maxphys.
 843                  */
 844                 maxmap = maxphys;
 845                 break;
 846         case XPT_SCSI_IO:
 847         case XPT_CONT_TARGET_IO:
 848                 if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
 849                         return(0);
 850                 if ((ccb->ccb_h.flags & CAM_DATA_MASK) != CAM_DATA_VADDR)
 851                         return (EINVAL);
 852                 data_ptrs[0] = &ccb->csio.data_ptr;
 853                 lengths[0] = ccb->csio.dxfer_len;
 854                 dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
 855                 numbufs = 1;
 856                 break;
 857         case XPT_ATA_IO:
 858                 if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
 859                         return(0);
 860                 if ((ccb->ccb_h.flags & CAM_DATA_MASK) != CAM_DATA_VADDR)
 861                         return (EINVAL);
 862                 data_ptrs[0] = &ccb->ataio.data_ptr;
 863                 lengths[0] = ccb->ataio.dxfer_len;
 864                 dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
 865                 numbufs = 1;
 866                 break;
 867         case XPT_MMC_IO:
 868                 if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
 869                         return(0);
 870                 /* Two mappings: one for cmd->data and one for cmd->data->data */
 871                 data_ptrs[0] = (unsigned char **)&ccb->mmcio.cmd.data;
 872                 lengths[0] = sizeof(struct mmc_data *);
 873                 dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
 874                 data_ptrs[1] = (unsigned char **)&ccb->mmcio.cmd.data->data;
 875                 lengths[1] = ccb->mmcio.cmd.data->len;
 876                 dirs[1] = ccb->ccb_h.flags & CAM_DIR_MASK;
 877                 numbufs = 2;
 878                 break;
 879         case XPT_SMP_IO:
 880                 data_ptrs[0] = &ccb->smpio.smp_request;
 881                 lengths[0] = ccb->smpio.smp_request_len;
 882                 dirs[0] = CAM_DIR_OUT;
 883                 data_ptrs[1] = &ccb->smpio.smp_response;
 884                 lengths[1] = ccb->smpio.smp_response_len;
 885                 dirs[1] = CAM_DIR_IN;
 886                 numbufs = 2;
 887                 break;
 888         case XPT_NVME_IO:
 889         case XPT_NVME_ADMIN:
 890                 if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
 891                         return (0);
 892                 if ((ccb->ccb_h.flags & CAM_DATA_MASK) != CAM_DATA_VADDR)
 893                         return (EINVAL);
 894                 data_ptrs[0] = &ccb->nvmeio.data_ptr;
 895                 lengths[0] = ccb->nvmeio.dxfer_len;
 896                 dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
 897                 numbufs = 1;
 898                 break;
 899         case XPT_DEV_ADVINFO:
 900                 if (ccb->cdai.bufsiz == 0)
 901                         return (0);
 902
 903                 data_ptrs[0] = (uint8_t **)&ccb->cdai.buf;
 904                 lengths[0] = ccb->cdai.bufsiz;
 905                 dirs[0] = CAM_DIR_IN;
 906                 numbufs = 1;
 907
 908                 /*
 909                  * This request will not go to the hardware, no reason
 910                  * to be so strict. vmapbuf() is able to map up to maxphys.
 911                  */
 912                 maxmap = maxphys;
 913                 break;
 914         default:
 915                 return(EINVAL);
 916                 break; /* NOTREACHED */
 917         }
 918
 919         /*
 920          * Check the transfer length and permissions first, so we don't
 921          * have to unmap any previously mapped buffers.
 922          */
 923         for (i = 0; i < numbufs; i++) {
 924                 if (lengths[i] > maxmap) {
 925                         printf("cam_periph_mapmem: attempt to map %lu bytes, "
 926                                "which is greater than %lu\n",
 927                                (long)(lengths[i]), (u_long)maxmap);
 928                         return (E2BIG);
 929                 }
 930         }
 931
 932         /*
 933          * This keeps the kernel stack of current thread from getting
 934          * swapped.  In low-memory situations where the kernel stack might
 935          * otherwise get swapped out, this holds it and allows the thread
 936          * to make progress and release the kernel mapped pages sooner.
 937          *
 938          * XXX KDM should I use P_NOSWAP instead?
 939          */
 940         PHOLD(curproc);
 941
 942         for (i = 0; i < numbufs; i++) {
 943                 /* Save the user's data address. */
 944                 mapinfo->orig[i] = *data_ptrs[i];
 945
 946                 /*
 947                  * For small buffers use malloc+copyin/copyout instead of
 948                  * mapping to KVA to avoid expensive TLB shootdowns.  For
 949                  * small allocations malloc is backed by UMA, and so much
 950                  * cheaper on SMP systems.
 951                  */
 952                 if (lengths[i] <= periph_mapmem_thresh &&
 953                     ccb->ccb_h.func_code != XPT_MMC_IO) {
 954                         *data_ptrs[i] = malloc(lengths[i], M_CAMPERIPH,
 955                             M_WAITOK);
 956                         if (dirs[i] != CAM_DIR_IN) {
 957                                 if (copyin(mapinfo->orig[i], *data_ptrs[i],
 958                                     lengths[i]) != 0) {
 959                                         free(*data_ptrs[i], M_CAMPERIPH);
 960                                         *data_ptrs[i] = mapinfo->orig[i];
 961                                         goto fail;
 962                                 }
 963                         } else
 964                                 bzero(*data_ptrs[i], lengths[i]);
 965                         continue;
 966                 }
 967
 968                 /*
 969                  * Get the buffer.
 970                  */
 971                 mapinfo->bp[i] = uma_zalloc(pbuf_zone, M_WAITOK);
 972
 973                 /* set the direction */
 974                 mapinfo->bp[i]->b_iocmd = (dirs[i] == CAM_DIR_OUT) ?
 975                     BIO_WRITE : BIO_READ;
 976
 977                 /* Map the buffer into kernel memory. */
 978                 if (vmapbuf(mapinfo->bp[i], *data_ptrs[i], lengths[i], 1) < 0) {
 979                         uma_zfree(pbuf_zone, mapinfo->bp[i]);
 980                         goto fail;
 981                 }
 982
 983                 /* set our pointer to the new mapped area */
 984                 *data_ptrs[i] = mapinfo->bp[i]->b_data;
 985         }
 986
 987         /*
 988          * Now that we've gotten this far, change ownership to the kernel
 989          * of the buffers so that we don't run afoul of returning to user
 990          * space with locks (on the buffer) held.
 991          */
 992         for (i = 0; i < numbufs; i++) {
 993                 if (mapinfo->bp[i])
 994                         BUF_KERNPROC(mapinfo->bp[i]);
 995         }
 996
 997         mapinfo->num_bufs_used = numbufs;
 998         return(0);
 999
1000 fail:
1001         for (i--; i >= 0; i--) {
1002                 if (mapinfo->bp[i]) {
1003                         vunmapbuf(mapinfo->bp[i]);
1004                         uma_zfree(pbuf_zone, mapinfo->bp[i]);
1005                 } else
1006                         free(*data_ptrs[i], M_CAMPERIPH);
1007                 *data_ptrs[i] = mapinfo->orig[i];
1008         }
1009         PRELE(curproc);
1010         return(EACCES);
1011 }
1012
1013 /*
1014  * Unmap memory segments mapped into kernel virtual address space by
1015  * cam_periph_mapmem().
1016  */
1017 void
1018 cam_periph_unmapmem(union ccb *ccb, struct cam_periph_map_info *mapinfo)
1019 {
1020         int numbufs, i;
1021         uint8_t **data_ptrs[CAM_PERIPH_MAXMAPS];
1022         uint32_t lengths[CAM_PERIPH_MAXMAPS];
1023         uint32_t dirs[CAM_PERIPH_MAXMAPS];
1024
1025         if (mapinfo->num_bufs_used <= 0) {
1026                 /* nothing to free and the process wasn't held. */
1027                 return;
1028         }
1029
1030         switch (ccb->ccb_h.func_code) {
1031         case XPT_DEV_MATCH:
1032                 if (ccb->cdm.pattern_buf_len > 0) {
1033                         data_ptrs[0] = (uint8_t **)&ccb->cdm.patterns;
1034                         lengths[0] = ccb->cdm.pattern_buf_len;
1035                         dirs[0] = CAM_DIR_OUT;
1036                         data_ptrs[1] = (uint8_t **)&ccb->cdm.matches;
1037                         lengths[1] = ccb->cdm.match_buf_len;
1038                         dirs[1] = CAM_DIR_IN;
1039                         numbufs = 2;
1040                 } else {
1041                         data_ptrs[0] = (uint8_t **)&ccb->cdm.matches;
1042                         lengths[0] = ccb->cdm.match_buf_len;
1043                         dirs[0] = CAM_DIR_IN;
1044                         numbufs = 1;
1045                 }
1046                 break;
1047         case XPT_SCSI_IO:
1048         case XPT_CONT_TARGET_IO:
1049                 data_ptrs[0] = &ccb->csio.data_ptr;
1050                 lengths[0] = ccb->csio.dxfer_len;
1051                 dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1052                 numbufs = 1;
1053                 break;
1054         case XPT_ATA_IO:
1055                 data_ptrs[0] = &ccb->ataio.data_ptr;
1056                 lengths[0] = ccb->ataio.dxfer_len;
1057                 dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1058                 numbufs = 1;
1059                 break;
1060         case XPT_MMC_IO:
1061                 data_ptrs[0] = (uint8_t **)&ccb->mmcio.cmd.data;
1062                 lengths[0] = sizeof(struct mmc_data *);
1063                 dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1064                 data_ptrs[1] = (uint8_t **)&ccb->mmcio.cmd.data->data;
1065                 lengths[1] = ccb->mmcio.cmd.data->len;
1066                 dirs[1] = ccb->ccb_h.flags & CAM_DIR_MASK;
1067                 numbufs = 2;
1068                 break;
1069         case XPT_SMP_IO:
1070                 data_ptrs[0] = &ccb->smpio.smp_request;
1071                 lengths[0] = ccb->smpio.smp_request_len;
1072                 dirs[0] = CAM_DIR_OUT;
1073                 data_ptrs[1] = &ccb->smpio.smp_response;
1074                 lengths[1] = ccb->smpio.smp_response_len;
1075                 dirs[1] = CAM_DIR_IN;
1076                 numbufs = 2;
1077                 break;
1078         case XPT_NVME_IO:
1079         case XPT_NVME_ADMIN:
1080                 data_ptrs[0] = &ccb->nvmeio.data_ptr;
1081                 lengths[0] = ccb->nvmeio.dxfer_len;
1082                 dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1083                 numbufs = 1;
1084                 break;
1085         case XPT_DEV_ADVINFO:
1086                 data_ptrs[0] = (uint8_t **)&ccb->cdai.buf;
1087                 lengths[0] = ccb->cdai.bufsiz;
1088                 dirs[0] = CAM_DIR_IN;
1089                 numbufs = 1;
1090                 break;
1091         default:
1092                 /* allow ourselves to be swapped once again */
1093                 PRELE(curproc);
1094                 return;
1095                 break; /* NOTREACHED */
1096         }
1097
1098         for (i = 0; i < numbufs; i++) {
1099                 if (mapinfo->bp[i]) {
1100                         /* unmap the buffer */
1101                         vunmapbuf(mapinfo->bp[i]);
1102
1103                         /* release the buffer */
1104                         uma_zfree(pbuf_zone, mapinfo->bp[i]);
1105                 } else {
1106                         if (dirs[i] != CAM_DIR_OUT) {
1107                                 copyout(*data_ptrs[i], mapinfo->orig[i],
1108                                     lengths[i]);
1109                         }
1110                         free(*data_ptrs[i], M_CAMPERIPH);
1111                 }
1112
1113                 /* Set the user's pointer back to the original value */
1114                 *data_ptrs[i] = mapinfo->orig[i];
1115         }
1116
1117         /* allow ourselves to be swapped once again */
1118         PRELE(curproc);
1119 }
1120
1121 int
1122 cam_periph_ioctl(struct cam_periph *periph, u_long cmd, caddr_t addr,
1123                  int (*error_routine)(union ccb *ccb,
1124                                       cam_flags camflags,
1125                                       uint32_t sense_flags))
1126 {
1127         union ccb            *ccb;
1128         int                  error;
1129         int                  found;
1130
1131         error = found = 0;
1132
1133         switch(cmd){
1134         case CAMGETPASSTHRU_0x19:
1135         case CAMGETPASSTHRU:
1136                 ccb = cam_periph_getccb(periph, CAM_PRIORITY_NORMAL);
1137                 xpt_setup_ccb(&ccb->ccb_h,
1138                               ccb->ccb_h.path,
1139                               CAM_PRIORITY_NORMAL);
1140                 ccb->ccb_h.func_code = XPT_GDEVLIST;
1141
1142                 /*
1143                  * Basically, the point of this is that we go through
1144                  * getting the list of devices, until we find a passthrough
1145                  * device.  In the current version of the CAM code, the
1146                  * only way to determine what type of device we're dealing
1147                  * with is by its name.
1148                  */
1149                 while (found == 0) {
1150                         ccb->cgdl.index = 0;
1151                         ccb->cgdl.status = CAM_GDEVLIST_MORE_DEVS;
1152                         while (ccb->cgdl.status == CAM_GDEVLIST_MORE_DEVS) {
1153                                 /* we want the next device in the list */
1154                                 xpt_action(ccb);
1155                                 if (strncmp(ccb->cgdl.periph_name,
1156                                     "pass", 4) == 0){
1157                                         found = 1;
1158                                         break;
1159                                 }
1160                         }
1161                         if ((ccb->cgdl.status == CAM_GDEVLIST_LAST_DEVICE) &&
1162                             (found == 0)) {
1163                                 ccb->cgdl.periph_name[0] = '\0';
1164                                 ccb->cgdl.unit_number = 0;
1165                                 break;
1166                         }
1167                 }
1168
1169                 /* copy the result back out */
1170                 bcopy(ccb, addr, sizeof(union ccb));
1171
1172                 /* and release the ccb */
1173                 xpt_release_ccb(ccb);
1174
1175                 break;
1176         default:
1177                 error = ENOTTY;
1178                 break;
1179         }
1180         return(error);
1181 }
1182
1183 static void
1184 cam_periph_done_panic(struct cam_periph *periph, union ccb *done_ccb)
1185 {
1186
1187         panic("%s: already done with ccb %p", __func__, done_ccb);
1188 }
1189
1190 static void
1191 cam_periph_done(struct cam_periph *periph, union ccb *done_ccb)
1192 {
1193
1194         /* Caller will release the CCB */
1195         xpt_path_assert(done_ccb->ccb_h.path, MA_OWNED);
1196         done_ccb->ccb_h.cbfcnp = cam_periph_done_panic;
1197         wakeup(&done_ccb->ccb_h.cbfcnp);
1198 }
1199
1200 static void
1201 cam_periph_ccbwait(union ccb *ccb)
1202 {
1203
1204         if ((ccb->ccb_h.func_code & XPT_FC_QUEUED) != 0) {
1205                 while (ccb->ccb_h.cbfcnp != cam_periph_done_panic)
1206                         xpt_path_sleep(ccb->ccb_h.path, &ccb->ccb_h.cbfcnp,
1207                             PRIBIO, "cbwait", 0);
1208         }
1209         KASSERT(ccb->ccb_h.pinfo.index == CAM_UNQUEUED_INDEX &&
1210             (ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_INPROG,
1211             ("%s: proceeding with incomplete ccb: ccb=%p, func_code=%#x, "
1212              "status=%#x, index=%d", __func__, ccb, ccb->ccb_h.func_code,
1213              ccb->ccb_h.status, ccb->ccb_h.pinfo.index));
1214 }
1215
1216 /*
1217  * Dispatch a CCB and wait for it to complete.  If the CCB has set a
1218  * callback function (ccb->ccb_h.cbfcnp), it will be overwritten and lost.
1219  */
1220 int
1221 cam_periph_runccb(union ccb *ccb,
1222                   int (*error_routine)(union ccb *ccb,
1223                                        cam_flags camflags,
1224                                        uint32_t sense_flags),
1225                   cam_flags camflags, uint32_t sense_flags,
1226                   struct devstat *ds)
1227 {
1228         struct bintime *starttime;
1229         struct bintime ltime;
1230         int error;
1231         bool must_poll;
1232         uint32_t timeout = 1;
1233
1234         starttime = NULL;
1235         xpt_path_assert(ccb->ccb_h.path, MA_OWNED);
1236         KASSERT((ccb->ccb_h.flags & CAM_UNLOCKED) == 0,
1237             ("%s: ccb=%p, func_code=%#x, flags=%#x", __func__, ccb,
1238              ccb->ccb_h.func_code, ccb->ccb_h.flags));
1239
1240         /*
1241          * If the user has supplied a stats structure, and if we understand
1242          * this particular type of ccb, record the transaction start.
1243          */
1244         if (ds != NULL &&
1245             (ccb->ccb_h.func_code == XPT_SCSI_IO ||
1246             ccb->ccb_h.func_code == XPT_ATA_IO ||
1247             ccb->ccb_h.func_code == XPT_NVME_IO)) {
1248                 starttime = &ltime;
1249                 binuptime(starttime);
1250                 devstat_start_transaction(ds, starttime);
1251         }
1252
1253         /*
1254          * We must poll the I/O while we're dumping. The scheduler is normally
1255          * stopped for dumping, except when we call doadump from ddb. While the
1256          * scheduler is running in this case, we still need to poll the I/O to
1257          * avoid sleeping waiting for the ccb to complete.
1258          *
1259          * A panic triggered dump stops the scheduler, any callback from the
1260          * shutdown_post_sync event will run with the scheduler stopped, but
1261          * before we're officially dumping. To avoid hanging in adashutdown
1262          * initiated commands (or other similar situations), we have to test for
1263          * either dumping or SCHEDULER_STOPPED() here.
1264          *
1265          * To avoid locking problems, dumping/polling callers must call
1266          * without a periph lock held.
1267          */
1268         must_poll = dumping || SCHEDULER_STOPPED();
1269         ccb->ccb_h.cbfcnp = cam_periph_done;
1270
1271         /*
1272          * If we're polling, then we need to ensure that we have ample resources
1273          * in the periph.  cam_periph_error can reschedule the ccb by calling
1274          * xpt_action and returning ERESTART, so we have to effect the polling
1275          * in the do loop below.
1276          */
1277         if (must_poll) {
1278                 if (cam_sim_pollable(ccb->ccb_h.path->bus->sim))
1279                         timeout = xpt_poll_setup(ccb);
1280                 else
1281                         timeout = 0;
1282         }
1283
1284         if (timeout == 0) {
1285                 ccb->ccb_h.status = CAM_RESRC_UNAVAIL;
1286                 error = EBUSY;
1287         } else {
1288                 xpt_action(ccb);
1289                 do {
1290                         if (must_poll) {
1291                                 xpt_pollwait(ccb, timeout);
1292                                 timeout = ccb->ccb_h.timeout * 10;
1293                         } else {
1294                                 cam_periph_ccbwait(ccb);
1295                         }
1296                         if ((ccb->ccb_h.status & CAM_STATUS_MASK) == CAM_REQ_CMP)
1297                                 error = 0;
1298                         else if (error_routine != NULL) {
1299                                 /*
1300                                  * cbfcnp is modified by cam_periph_ccbwait so
1301                                  * reset it before we call the error routine
1302                                  * which may call xpt_done.
1303                                  */
1304                                 ccb->ccb_h.cbfcnp = cam_periph_done;
1305                                 error = (*error_routine)(ccb, camflags, sense_flags);
1306                         } else
1307                                 error = 0;
1308                 } while (error == ERESTART);
1309         }
1310
1311         if ((ccb->ccb_h.status & CAM_DEV_QFRZN) != 0) {
1312                 cam_release_devq(ccb->ccb_h.path,
1313                                  /* relsim_flags */0,
1314                                  /* openings */0,
1315                                  /* timeout */0,
1316                                  /* getcount_only */ FALSE);
1317                 ccb->ccb_h.status &= ~CAM_DEV_QFRZN;
1318         }
1319
1320         if (ds != NULL) {
1321                 uint32_t bytes;
1322                 devstat_tag_type tag;
1323                 bool valid = true;
1324
1325                 if (ccb->ccb_h.func_code == XPT_SCSI_IO) {
1326                         bytes = ccb->csio.dxfer_len - ccb->csio.resid;
1327                         tag = (devstat_tag_type)(ccb->csio.tag_action & 0x3);
1328                 } else if (ccb->ccb_h.func_code == XPT_ATA_IO) {
1329                         bytes = ccb->ataio.dxfer_len - ccb->ataio.resid;
1330                         tag = (devstat_tag_type)0;
1331                 } else if (ccb->ccb_h.func_code == XPT_NVME_IO) {
1332                         bytes = ccb->nvmeio.dxfer_len; /* NB: resid no possible */
1333                         tag = (devstat_tag_type)0;
1334                 } else {
1335                         valid = false;
1336                 }
1337                 if (valid)
1338                         devstat_end_transaction(ds, bytes, tag,
1339                             ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE) ?
1340                             DEVSTAT_NO_DATA : (ccb->ccb_h.flags & CAM_DIR_OUT) ?
1341                             DEVSTAT_WRITE : DEVSTAT_READ, NULL, starttime);
1342         }
1343
1344         return(error);
1345 }
1346
1347 void
1348 cam_freeze_devq(struct cam_path *path)
1349 {
1350         struct ccb_hdr ccb_h;
1351
1352         CAM_DEBUG(path, CAM_DEBUG_TRACE, ("cam_freeze_devq\n"));
1353         memset(&ccb_h, 0, sizeof(ccb_h));
1354         xpt_setup_ccb(&ccb_h, path, /*priority*/1);
1355         ccb_h.func_code = XPT_NOOP;
1356         ccb_h.flags = CAM_DEV_QFREEZE;
1357         xpt_action((union ccb *)&ccb_h);
1358 }
1359
1360 uint32_t
1361 cam_release_devq(struct cam_path *path, uint32_t relsim_flags,
1362                  uint32_t openings, uint32_t arg,
1363                  int getcount_only)
1364 {
1365         struct ccb_relsim crs;
1366
1367         CAM_DEBUG(path, CAM_DEBUG_TRACE, ("cam_release_devq(%u, %u, %u, %d)\n",
1368             relsim_flags, openings, arg, getcount_only));
1369         memset(&crs, 0, sizeof(crs));
1370         xpt_setup_ccb(&crs.ccb_h, path, CAM_PRIORITY_NORMAL);
1371         crs.ccb_h.func_code = XPT_REL_SIMQ;
1372         crs.ccb_h.flags = getcount_only ? CAM_DEV_QFREEZE : 0;
1373         crs.release_flags = relsim_flags;
1374         crs.openings = openings;
1375         crs.release_timeout = arg;
1376         xpt_action((union ccb *)&crs);
1377         return (crs.qfrozen_cnt);
1378 }
1379
1380 #define saved_ccb_ptr ppriv_ptr0
1381 static void
1382 camperiphdone(struct cam_periph *periph, union ccb *done_ccb)
1383 {
1384         union ccb      *saved_ccb;
1385         cam_status      status;
1386         struct scsi_start_stop_unit *scsi_cmd;
1387         int             error = 0, error_code, sense_key, asc, ascq;
1388         uint16_t        done_flags;
1389
1390         scsi_cmd = (struct scsi_start_stop_unit *)
1391             &done_ccb->csio.cdb_io.cdb_bytes;
1392         status = done_ccb->ccb_h.status;
1393
1394         if ((status & CAM_STATUS_MASK) != CAM_REQ_CMP) {
1395                 if (scsi_extract_sense_ccb(done_ccb,
1396                     &error_code, &sense_key, &asc, &ascq)) {
1397                         /*
1398                          * If the error is "invalid field in CDB",
1399                          * and the load/eject flag is set, turn the
1400                          * flag off and try again.  This is just in
1401                          * case the drive in question barfs on the
1402                          * load eject flag.  The CAM code should set
1403                          * the load/eject flag by default for
1404                          * removable media.
1405                          */
1406                         if ((scsi_cmd->opcode == START_STOP_UNIT) &&
1407                             ((scsi_cmd->how & SSS_LOEJ) != 0) &&
1408                              (asc == 0x24) && (ascq == 0x00)) {
1409                                 scsi_cmd->how &= ~SSS_LOEJ;
1410                                 if (status & CAM_DEV_QFRZN) {
1411                                         cam_release_devq(done_ccb->ccb_h.path,
1412                                             0, 0, 0, 0);
1413                                         done_ccb->ccb_h.status &=
1414                                             ~CAM_DEV_QFRZN;
1415                                 }
1416                                 xpt_action(done_ccb);
1417                                 goto out;
1418                         }
1419                 }
1420                 error = cam_periph_error(done_ccb, 0,
1421                     SF_RETRY_UA | SF_NO_PRINT);
1422                 if (error == ERESTART)
1423                         goto out;
1424                 if (done_ccb->ccb_h.status & CAM_DEV_QFRZN) {
1425                         cam_release_devq(done_ccb->ccb_h.path, 0, 0, 0, 0);
1426                         done_ccb->ccb_h.status &= ~CAM_DEV_QFRZN;
1427                 }
1428         } else {
1429                 /*
1430                  * If we have successfully taken a device from the not
1431                  * ready to ready state, re-scan the device and re-get
1432                  * the inquiry information.  Many devices (mostly disks)
1433                  * don't properly report their inquiry information unless
1434                  * they are spun up.
1435                  */
1436                 if (scsi_cmd->opcode == START_STOP_UNIT)
1437                         xpt_async(AC_INQ_CHANGED, done_ccb->ccb_h.path, NULL);
1438         }
1439
1440         /* If we tried long wait and still failed, remember that. */
1441         if ((periph->flags & CAM_PERIPH_RECOVERY_WAIT) &&
1442             (done_ccb->csio.cdb_io.cdb_bytes[0] == TEST_UNIT_READY)) {
1443                 periph->flags &= ~CAM_PERIPH_RECOVERY_WAIT;
1444                 if (error != 0 && done_ccb->ccb_h.retry_count == 0)
1445                         periph->flags |= CAM_PERIPH_RECOVERY_WAIT_FAILED;
1446         }
1447
1448         /*
1449          * After recovery action(s) completed, return to the original CCB.
1450          * If the recovery CCB has failed, considering its own possible
1451          * retries and recovery, assume we are back in state where we have
1452          * been originally, but without recovery hopes left.  In such case,
1453          * after the final attempt below, we cancel any further retries,
1454          * blocking by that also any new recovery attempts for this CCB,
1455          * and the result will be the final one returned to the CCB owher.
1456          */
1457         saved_ccb = (union ccb *)done_ccb->ccb_h.saved_ccb_ptr;
1458         KASSERT(saved_ccb->ccb_h.func_code == XPT_SCSI_IO,
1459             ("%s: saved_ccb func_code %#x != XPT_SCSI_IO",
1460              __func__, saved_ccb->ccb_h.func_code));
1461         KASSERT(done_ccb->ccb_h.func_code == XPT_SCSI_IO,
1462             ("%s: done_ccb func_code %#x != XPT_SCSI_IO",
1463              __func__, done_ccb->ccb_h.func_code));
1464         saved_ccb->ccb_h.periph_links = done_ccb->ccb_h.periph_links;
1465         done_flags = done_ccb->ccb_h.alloc_flags;
1466         bcopy(saved_ccb, done_ccb, sizeof(struct ccb_scsiio));
1467         done_ccb->ccb_h.alloc_flags = done_flags;
1468         xpt_free_ccb(saved_ccb);
1469         if (done_ccb->ccb_h.cbfcnp != camperiphdone)
1470                 periph->flags &= ~CAM_PERIPH_RECOVERY_INPROG;
1471         if (error != 0)
1472                 done_ccb->ccb_h.retry_count = 0;
1473         xpt_action(done_ccb);
1474
1475 out:
1476         /* Drop freeze taken due to CAM_DEV_QFREEZE flag set. */
1477         cam_release_devq(done_ccb->ccb_h.path, 0, 0, 0, 0);
1478 }
1479
1480 /*
1481  * Generic Async Event handler.  Peripheral drivers usually
1482  * filter out the events that require personal attention,
1483  * and leave the rest to this function.
1484  */
1485 void
1486 cam_periph_async(struct cam_periph *periph, uint32_t code,
1487                  struct cam_path *path, void *arg)
1488 {
1489         switch (code) {
1490         case AC_LOST_DEVICE:
1491                 cam_periph_invalidate(periph);
1492                 break;
1493         default:
1494                 break;
1495         }
1496 }
1497
1498 void
1499 cam_periph_bus_settle(struct cam_periph *periph, u_int bus_settle)
1500 {
1501         struct ccb_getdevstats cgds;
1502
1503         memset(&cgds, 0, sizeof(cgds));
1504         xpt_setup_ccb(&cgds.ccb_h, periph->path, CAM_PRIORITY_NORMAL);
1505         cgds.ccb_h.func_code = XPT_GDEV_STATS;
1506         xpt_action((union ccb *)&cgds);
1507         cam_periph_freeze_after_event(periph, &cgds.last_reset, bus_settle);
1508 }
1509
1510 void
1511 cam_periph_freeze_after_event(struct cam_periph *periph,
1512                               struct timeval* event_time, u_int duration_ms)
1513 {
1514         struct timeval delta;
1515         struct timeval duration_tv;
1516
1517         if (!timevalisset(event_time))
1518                 return;
1519
1520         microtime(&delta);
1521         timevalsub(&delta, event_time);
1522         duration_tv.tv_sec = duration_ms / 1000;
1523         duration_tv.tv_usec = (duration_ms % 1000) * 1000;
1524         if (timevalcmp(&delta, &duration_tv, <)) {
1525                 timevalsub(&duration_tv, &delta);
1526
1527                 duration_ms = duration_tv.tv_sec * 1000;
1528                 duration_ms += duration_tv.tv_usec / 1000;
1529                 cam_freeze_devq(periph->path);
1530                 cam_release_devq(periph->path,
1531                                 RELSIM_RELEASE_AFTER_TIMEOUT,
1532                                 /*reduction*/0,
1533                                 /*timeout*/duration_ms,
1534                                 /*getcount_only*/0);
1535         }
1536
1537 }
1538
1539 static int
1540 camperiphscsistatuserror(union ccb *ccb, union ccb **orig_ccb,
1541     cam_flags camflags, uint32_t sense_flags,
1542     int *openings, uint32_t *relsim_flags,
1543     uint32_t *timeout, uint32_t *action, const char **action_string)
1544 {
1545         struct cam_periph *periph;
1546         int error;
1547
1548         switch (ccb->csio.scsi_status) {
1549         case SCSI_STATUS_OK:
1550         case SCSI_STATUS_COND_MET:
1551         case SCSI_STATUS_INTERMED:
1552         case SCSI_STATUS_INTERMED_COND_MET:
1553                 error = 0;
1554                 break;
1555         case SCSI_STATUS_CMD_TERMINATED:
1556         case SCSI_STATUS_CHECK_COND:
1557                 error = camperiphscsisenseerror(ccb, orig_ccb,
1558                                                 camflags,
1559                                                 sense_flags,
1560                                                 openings,
1561                                                 relsim_flags,
1562                                                 timeout,
1563                                                 action,
1564                                                 action_string);
1565                 break;
1566         case SCSI_STATUS_QUEUE_FULL:
1567         {
1568                 /* no decrement */
1569                 struct ccb_getdevstats cgds;
1570
1571                 /*
1572                  * First off, find out what the current
1573                  * transaction counts are.
1574                  */
1575                 memset(&cgds, 0, sizeof(cgds));
1576                 xpt_setup_ccb(&cgds.ccb_h,
1577                               ccb->ccb_h.path,
1578                               CAM_PRIORITY_NORMAL);
1579                 cgds.ccb_h.func_code = XPT_GDEV_STATS;
1580                 xpt_action((union ccb *)&cgds);
1581
1582                 /*
1583                  * If we were the only transaction active, treat
1584                  * the QUEUE FULL as if it were a BUSY condition.
1585                  */
1586                 if (cgds.dev_active != 0) {
1587                         int total_openings;
1588
1589                         /*
1590                          * Reduce the number of openings to
1591                          * be 1 less than the amount it took
1592                          * to get a queue full bounded by the
1593                          * minimum allowed tag count for this
1594                          * device.
1595                          */
1596                         total_openings = cgds.dev_active + cgds.dev_openings;
1597                         *openings = cgds.dev_active;
1598                         if (*openings < cgds.mintags)
1599                                 *openings = cgds.mintags;
1600                         if (*openings < total_openings)
1601                                 *relsim_flags = RELSIM_ADJUST_OPENINGS;
1602                         else {
1603                                 /*
1604                                  * Some devices report queue full for
1605                                  * temporary resource shortages.  For
1606                                  * this reason, we allow a minimum
1607                                  * tag count to be entered via a
1608                                  * quirk entry to prevent the queue
1609                                  * count on these devices from falling
1610                                  * to a pessimisticly low value.  We
1611                                  * still wait for the next successful
1612                                  * completion, however, before queueing
1613                                  * more transactions to the device.
1614                                  */
1615                                 *relsim_flags = RELSIM_RELEASE_AFTER_CMDCMPLT;
1616                         }
1617                         *timeout = 0;
1618                         error = ERESTART;
1619                         *action &= ~SSQ_PRINT_SENSE;
1620                         break;
1621                 }
1622                 /* FALLTHROUGH */
1623         }
1624         case SCSI_STATUS_BUSY:
1625                 /*
1626                  * Restart the queue after either another
1627                  * command completes or a 1 second timeout.
1628                  */
1629                 periph = xpt_path_periph(ccb->ccb_h.path);
1630                 if (periph->flags & CAM_PERIPH_INVALID) {
1631                         error = ENXIO;
1632                         *action_string = "Periph was invalidated";
1633                 } else if ((sense_flags & SF_RETRY_BUSY) != 0 ||
1634                     ccb->ccb_h.retry_count > 0) {
1635                         if ((sense_flags & SF_RETRY_BUSY) == 0)
1636                                 ccb->ccb_h.retry_count--;
1637                         error = ERESTART;
1638                         *relsim_flags = RELSIM_RELEASE_AFTER_TIMEOUT
1639                                       | RELSIM_RELEASE_AFTER_CMDCMPLT;
1640                         *timeout = 1000;
1641                 } else {
1642                         error = EIO;
1643                         *action_string = "Retries exhausted";
1644                 }
1645                 break;
1646         case SCSI_STATUS_RESERV_CONFLICT:
1647         default:
1648                 error = EIO;
1649                 break;
1650         }
1651         return (error);
1652 }
1653
1654 static int
1655 camperiphscsisenseerror(union ccb *ccb, union ccb **orig,
1656     cam_flags camflags, uint32_t sense_flags,
1657     int *openings, uint32_t *relsim_flags,
1658     uint32_t *timeout, uint32_t *action, const char **action_string)
1659 {
1660         struct cam_periph *periph;
1661         union ccb *orig_ccb = ccb;
1662         int error, recoveryccb;
1663         uint16_t flags;
1664
1665 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
1666         if (ccb->ccb_h.func_code == XPT_SCSI_IO && ccb->csio.bio != NULL)
1667                 biotrack(ccb->csio.bio, __func__);
1668 #endif
1669
1670         periph = xpt_path_periph(ccb->ccb_h.path);
1671         recoveryccb = (ccb->ccb_h.cbfcnp == camperiphdone);
1672         if ((periph->flags & CAM_PERIPH_RECOVERY_INPROG) && !recoveryccb) {
1673                 /*
1674                  * If error recovery is already in progress, don't attempt
1675                  * to process this error, but requeue it unconditionally
1676                  * and attempt to process it once error recovery has
1677                  * completed.  This failed command is probably related to
1678                  * the error that caused the currently active error recovery
1679                  * action so our  current recovery efforts should also
1680                  * address this command.  Be aware that the error recovery
1681                  * code assumes that only one recovery action is in progress
1682                  * on a particular peripheral instance at any given time
1683                  * (e.g. only one saved CCB for error recovery) so it is
1684                  * imperitive that we don't violate this assumption.
1685                  */
1686                 error = ERESTART;
1687                 *action &= ~SSQ_PRINT_SENSE;
1688         } else {
1689                 scsi_sense_action err_action;
1690                 struct ccb_getdev cgd;
1691
1692                 /*
1693                  * Grab the inquiry data for this device.
1694                  */
1695                 memset(&cgd, 0, sizeof(cgd));
1696                 xpt_setup_ccb(&cgd.ccb_h, ccb->ccb_h.path, CAM_PRIORITY_NORMAL);
1697                 cgd.ccb_h.func_code = XPT_GDEV_TYPE;
1698                 xpt_action((union ccb *)&cgd);
1699
1700                 err_action = scsi_error_action(&ccb->csio, &cgd.inq_data,
1701                     sense_flags);
1702                 error = err_action & SS_ERRMASK;
1703
1704                 /*
1705                  * Do not autostart sequential access devices
1706                  * to avoid unexpected tape loading.
1707                  */
1708                 if ((err_action & SS_MASK) == SS_START &&
1709                     SID_TYPE(&cgd.inq_data) == T_SEQUENTIAL) {
1710                         *action_string = "Will not autostart a "
1711                             "sequential access device";
1712                         goto sense_error_done;
1713                 }
1714
1715                 /*
1716                  * Avoid recovery recursion if recovery action is the same.
1717                  */
1718                 if ((err_action & SS_MASK) >= SS_START && recoveryccb) {
1719                         if (((err_action & SS_MASK) == SS_START &&
1720                              ccb->csio.cdb_io.cdb_bytes[0] == START_STOP_UNIT) ||
1721                             ((err_action & SS_MASK) == SS_TUR &&
1722                              (ccb->csio.cdb_io.cdb_bytes[0] == TEST_UNIT_READY))) {
1723                                 err_action = SS_RETRY|SSQ_DECREMENT_COUNT|EIO;
1724                                 *relsim_flags = RELSIM_RELEASE_AFTER_TIMEOUT;
1725                                 *timeout = 500;
1726                         }
1727                 }
1728
1729                 /*
1730                  * If the recovery action will consume a retry,
1731                  * make sure we actually have retries available.
1732                  */
1733                 if ((err_action & SSQ_DECREMENT_COUNT) != 0) {
1734                         if (ccb->ccb_h.retry_count > 0 &&
1735                             (periph->flags & CAM_PERIPH_INVALID) == 0)
1736                                 ccb->ccb_h.retry_count--;
1737                         else {
1738                                 *action_string = "Retries exhausted";
1739                                 goto sense_error_done;
1740                         }
1741                 }
1742
1743                 if ((err_action & SS_MASK) >= SS_START) {
1744                         /*
1745                          * Do common portions of commands that
1746                          * use recovery CCBs.
1747                          */
1748                         orig_ccb = xpt_alloc_ccb_nowait();
1749                         if (orig_ccb == NULL) {
1750                                 *action_string = "Can't allocate recovery CCB";
1751                                 goto sense_error_done;
1752                         }
1753                         /*
1754                          * Clear freeze flag for original request here, as
1755                          * this freeze will be dropped as part of ERESTART.
1756                          */
1757                         ccb->ccb_h.status &= ~CAM_DEV_QFRZN;
1758
1759                         KASSERT(ccb->ccb_h.func_code == XPT_SCSI_IO,
1760                             ("%s: ccb func_code %#x != XPT_SCSI_IO",
1761                              __func__, ccb->ccb_h.func_code));
1762                         flags = orig_ccb->ccb_h.alloc_flags;
1763                         bcopy(ccb, orig_ccb, sizeof(struct ccb_scsiio));
1764                         orig_ccb->ccb_h.alloc_flags = flags;
1765                 }
1766
1767                 switch (err_action & SS_MASK) {
1768                 case SS_NOP:
1769                         *action_string = "No recovery action needed";
1770                         error = 0;
1771                         break;
1772                 case SS_RETRY:
1773                         *action_string = "Retrying command (per sense data)";
1774                         error = ERESTART;
1775                         break;
1776                 case SS_FAIL:
1777                         *action_string = "Unretryable error";
1778                         break;
1779                 case SS_START:
1780                 {
1781                         int le;
1782
1783                         /*
1784                          * Send a start unit command to the device, and
1785                          * then retry the command.
1786                          */
1787                         *action_string = "Attempting to start unit";
1788                         periph->flags |= CAM_PERIPH_RECOVERY_INPROG;
1789
1790                         /*
1791                          * Check for removable media and set
1792                          * load/eject flag appropriately.
1793                          */
1794                         if (SID_IS_REMOVABLE(&cgd.inq_data))
1795                                 le = TRUE;
1796                         else
1797                                 le = FALSE;
1798
1799                         scsi_start_stop(&ccb->csio,
1800                                         /*retries*/1,
1801                                         camperiphdone,
1802                                         MSG_SIMPLE_Q_TAG,
1803                                         /*start*/TRUE,
1804                                         /*load/eject*/le,
1805                                         /*immediate*/FALSE,
1806                                         SSD_FULL_SIZE,
1807                                         /*timeout*/50000);
1808                         break;
1809                 }
1810                 case SS_TUR:
1811                 {
1812                         /*
1813                          * Send a Test Unit Ready to the device.
1814                          * If the 'many' flag is set, we send 120
1815                          * test unit ready commands, one every half
1816                          * second.  Otherwise, we just send one TUR.
1817                          * We only want to do this if the retry
1818                          * count has not been exhausted.
1819                          */
1820                         int retries;
1821
1822                         if ((err_action & SSQ_MANY) != 0 && (periph->flags &
1823                              CAM_PERIPH_RECOVERY_WAIT_FAILED) == 0) {
1824                                 periph->flags |= CAM_PERIPH_RECOVERY_WAIT;
1825                                 *action_string = "Polling device for readiness";
1826                                 retries = 120;
1827                         } else {
1828                                 *action_string = "Testing device for readiness";
1829                                 retries = 1;
1830                         }
1831                         periph->flags |= CAM_PERIPH_RECOVERY_INPROG;
1832                         scsi_test_unit_ready(&ccb->csio,
1833                                              retries,
1834                                              camperiphdone,
1835                                              MSG_SIMPLE_Q_TAG,
1836                                              SSD_FULL_SIZE,
1837                                              /*timeout*/5000);
1838
1839                         /*
1840                          * Accomplish our 500ms delay by deferring
1841                          * the release of our device queue appropriately.
1842                          */
1843                         *relsim_flags = RELSIM_RELEASE_AFTER_TIMEOUT;
1844                         *timeout = 500;
1845                         break;
1846                 }
1847                 default:
1848                         panic("Unhandled error action %x", err_action);
1849                 }
1850
1851                 if ((err_action & SS_MASK) >= SS_START) {
1852                         /*
1853                          * Drop the priority, so that the recovery
1854                          * CCB is the first to execute.  Freeze the queue
1855                          * after this command is sent so that we can
1856                          * restore the old csio and have it queued in
1857                          * the proper order before we release normal
1858                          * transactions to the device.
1859                          */
1860                         ccb->ccb_h.pinfo.priority--;
1861                         ccb->ccb_h.flags |= CAM_DEV_QFREEZE;
1862                         ccb->ccb_h.saved_ccb_ptr = orig_ccb;
1863                         error = ERESTART;
1864                         *orig = orig_ccb;
1865                 }
1866
1867 sense_error_done:
1868                 *action = err_action;
1869         }
1870         return (error);
1871 }
1872
1873 /*
1874  * Generic error handler.  Peripheral drivers usually filter
1875  * out the errors that they handle in a unique manner, then
1876  * call this function.
1877  */
1878 int
1879 cam_periph_error(union ccb *ccb, cam_flags camflags,
1880                  uint32_t sense_flags)
1881 {
1882         struct cam_path *newpath;
1883         union ccb  *orig_ccb, *scan_ccb;
1884         struct cam_periph *periph;
1885         const char *action_string;
1886         cam_status  status;
1887         int         frozen, error, openings, devctl_err;
1888         uint32_t   action, relsim_flags, timeout;
1889
1890         action = SSQ_PRINT_SENSE;
1891         periph = xpt_path_periph(ccb->ccb_h.path);
1892         action_string = NULL;
1893         status = ccb->ccb_h.status;
1894         frozen = (status & CAM_DEV_QFRZN) != 0;
1895         status &= CAM_STATUS_MASK;
1896         devctl_err = openings = relsim_flags = timeout = 0;
1897         orig_ccb = ccb;
1898
1899         /* Filter the errors that should be reported via devctl */
1900         switch (ccb->ccb_h.status & CAM_STATUS_MASK) {
1901         case CAM_CMD_TIMEOUT:
1902         case CAM_REQ_ABORTED:
1903         case CAM_REQ_CMP_ERR:
1904         case CAM_REQ_TERMIO:
1905         case CAM_UNREC_HBA_ERROR:
1906         case CAM_DATA_RUN_ERR:
1907         case CAM_SCSI_STATUS_ERROR:
1908         case CAM_ATA_STATUS_ERROR:
1909         case CAM_SMP_STATUS_ERROR:
1910         case CAM_DEV_NOT_THERE:
1911         case CAM_NVME_STATUS_ERROR:
1912                 devctl_err++;
1913                 break;
1914         default:
1915                 break;
1916         }
1917
1918         switch (status) {
1919         case CAM_REQ_CMP:
1920                 error = 0;
1921                 action &= ~SSQ_PRINT_SENSE;
1922                 break;
1923         case CAM_SCSI_STATUS_ERROR:
1924                 error = camperiphscsistatuserror(ccb, &orig_ccb,
1925                     camflags, sense_flags, &openings, &relsim_flags,
1926                     &timeout, &action, &action_string);
1927                 break;
1928         case CAM_AUTOSENSE_FAIL:
1929                 error = EIO;    /* we have to kill the command */
1930                 break;
1931         case CAM_UA_ABORT:
1932         case CAM_UA_TERMIO:
1933         case CAM_MSG_REJECT_REC:
1934                 /* XXX Don't know that these are correct */
1935                 error = EIO;
1936                 break;
1937         case CAM_SEL_TIMEOUT:
1938                 if ((camflags & CAM_RETRY_SELTO) != 0) {
1939                         if (ccb->ccb_h.retry_count > 0 &&
1940                             (periph->flags & CAM_PERIPH_INVALID) == 0) {
1941                                 ccb->ccb_h.retry_count--;
1942                                 error = ERESTART;
1943
1944                                 /*
1945                                  * Wait a bit to give the device
1946                                  * time to recover before we try again.
1947                                  */
1948                                 relsim_flags = RELSIM_RELEASE_AFTER_TIMEOUT;
1949                                 timeout = periph_selto_delay;
1950                                 break;
1951                         }
1952                         action_string = "Retries exhausted";
1953                 }
1954                 /* FALLTHROUGH */
1955         case CAM_DEV_NOT_THERE:
1956                 error = ENXIO;
1957                 action = SSQ_LOST;
1958                 break;
1959         case CAM_REQ_INVALID:
1960         case CAM_PATH_INVALID:
1961         case CAM_NO_HBA:
1962         case CAM_PROVIDE_FAIL:
1963         case CAM_REQ_TOO_BIG:
1964         case CAM_LUN_INVALID:
1965         case CAM_TID_INVALID:
1966         case CAM_FUNC_NOTAVAIL:
1967                 error = EINVAL;
1968                 break;
1969         case CAM_SCSI_BUS_RESET:
1970         case CAM_BDR_SENT:
1971                 /*
1972                  * Commands that repeatedly timeout and cause these
1973                  * kinds of error recovery actions, should return
1974                  * CAM_CMD_TIMEOUT, which allows us to safely assume
1975                  * that this command was an innocent bystander to
1976                  * these events and should be unconditionally
1977                  * retried.
1978                  */
1979         case CAM_REQUEUE_REQ:
1980                 /* Unconditional requeue if device is still there */
1981                 if (periph->flags & CAM_PERIPH_INVALID) {
1982                         action_string = "Periph was invalidated";
1983                         error = ENXIO;
1984                 } else if (sense_flags & SF_NO_RETRY) {
1985                         error = EIO;
1986                         action_string = "Retry was blocked";
1987                 } else {
1988                         error = ERESTART;
1989                         action &= ~SSQ_PRINT_SENSE;
1990                 }
1991                 break;
1992         case CAM_RESRC_UNAVAIL:
1993                 /* Wait a bit for the resource shortage to abate. */
1994                 timeout = periph_noresrc_delay;
1995                 /* FALLTHROUGH */
1996         case CAM_BUSY:
1997                 if (timeout == 0) {
1998                         /* Wait a bit for the busy condition to abate. */
1999                         timeout = periph_busy_delay;
2000                 }
2001                 relsim_flags = RELSIM_RELEASE_AFTER_TIMEOUT;
2002                 /* FALLTHROUGH */
2003         case CAM_ATA_STATUS_ERROR:
2004         case CAM_NVME_STATUS_ERROR:
2005         case CAM_SMP_STATUS_ERROR:
2006         case CAM_REQ_CMP_ERR:
2007         case CAM_CMD_TIMEOUT:
2008         case CAM_UNEXP_BUSFREE:
2009         case CAM_UNCOR_PARITY:
2010         case CAM_DATA_RUN_ERR:
2011         default:
2012                 if (periph->flags & CAM_PERIPH_INVALID) {
2013                         error = ENXIO;
2014                         action_string = "Periph was invalidated";
2015                 } else if (ccb->ccb_h.retry_count == 0) {
2016                         error = EIO;
2017                         action_string = "Retries exhausted";
2018                 } else if (sense_flags & SF_NO_RETRY) {
2019                         error = EIO;
2020                         action_string = "Retry was blocked";
2021                 } else {
2022                         ccb->ccb_h.retry_count--;
2023                         error = ERESTART;
2024                 }
2025                 break;
2026         }
2027
2028         if ((sense_flags & SF_PRINT_ALWAYS) ||
2029             CAM_DEBUGGED(ccb->ccb_h.path, CAM_DEBUG_INFO))
2030                 action |= SSQ_PRINT_SENSE;
2031         else if (sense_flags & SF_NO_PRINT)
2032                 action &= ~SSQ_PRINT_SENSE;
2033         if ((action & SSQ_PRINT_SENSE) != 0)
2034                 cam_error_print(orig_ccb, CAM_ESF_ALL, CAM_EPF_ALL);
2035         if (error != 0 && (action & SSQ_PRINT_SENSE) != 0) {
2036                 if (error != ERESTART) {
2037                         if (action_string == NULL)
2038                                 action_string = "Unretryable error";
2039                         xpt_print(ccb->ccb_h.path, "Error %d, %s\n",
2040                             error, action_string);
2041                 } else if (action_string != NULL)
2042                         xpt_print(ccb->ccb_h.path, "%s\n", action_string);
2043                 else {
2044                         xpt_print(ccb->ccb_h.path,
2045                             "Retrying command, %d more tries remain\n",
2046                             ccb->ccb_h.retry_count);
2047                 }
2048         }
2049
2050         if (devctl_err && (error != 0 || (action & SSQ_PRINT_SENSE) != 0))
2051                 cam_periph_devctl_notify(orig_ccb);
2052
2053         if ((action & SSQ_LOST) != 0) {
2054                 lun_id_t lun_id;
2055
2056                 /*
2057                  * For a selection timeout, we consider all of the LUNs on
2058                  * the target to be gone.  If the status is CAM_DEV_NOT_THERE,
2059                  * then we only get rid of the device(s) specified by the
2060                  * path in the original CCB.
2061                  */
2062                 if (status == CAM_SEL_TIMEOUT)
2063                         lun_id = CAM_LUN_WILDCARD;
2064                 else
2065                         lun_id = xpt_path_lun_id(ccb->ccb_h.path);
2066
2067                 /* Should we do more if we can't create the path?? */
2068                 if (xpt_create_path(&newpath, periph,
2069                                     xpt_path_path_id(ccb->ccb_h.path),
2070                                     xpt_path_target_id(ccb->ccb_h.path),
2071                                     lun_id) == CAM_REQ_CMP) {
2072                         /*
2073                          * Let peripheral drivers know that this
2074                          * device has gone away.
2075                          */
2076                         xpt_async(AC_LOST_DEVICE, newpath, NULL);
2077                         xpt_free_path(newpath);
2078                 }
2079         }
2080
2081         /* Broadcast UNIT ATTENTIONs to all periphs. */
2082         if ((action & SSQ_UA) != 0)
2083                 xpt_async(AC_UNIT_ATTENTION, orig_ccb->ccb_h.path, orig_ccb);
2084
2085         /* Rescan target on "Reported LUNs data has changed" */
2086         if ((action & SSQ_RESCAN) != 0) {
2087                 if (xpt_create_path(&newpath, NULL,
2088                                     xpt_path_path_id(ccb->ccb_h.path),
2089                                     xpt_path_target_id(ccb->ccb_h.path),
2090                                     CAM_LUN_WILDCARD) == CAM_REQ_CMP) {
2091                         scan_ccb = xpt_alloc_ccb_nowait();
2092                         if (scan_ccb != NULL) {
2093                                 scan_ccb->ccb_h.path = newpath;
2094                                 scan_ccb->ccb_h.func_code = XPT_SCAN_TGT;
2095                                 scan_ccb->crcn.flags = 0;
2096                                 xpt_rescan(scan_ccb);
2097                         } else {
2098                                 xpt_print(newpath,
2099                                     "Can't allocate CCB to rescan target\n");
2100                                 xpt_free_path(newpath);
2101                         }
2102                 }
2103         }
2104
2105         /* Attempt a retry */
2106         if (error == ERESTART || error == 0) {
2107                 if (frozen != 0)
2108                         ccb->ccb_h.status &= ~CAM_DEV_QFRZN;
2109                 if (error == ERESTART)
2110                         xpt_action(ccb);
2111                 if (frozen != 0)
2112                         cam_release_devq(ccb->ccb_h.path,
2113                                          relsim_flags,
2114                                          openings,
2115                                          timeout,
2116                                          /*getcount_only*/0);
2117         }
2118
2119         return (error);
2120 }
2121
2122 #define CAM_PERIPH_DEVD_MSG_SIZE        256
2123
2124 static void
2125 cam_periph_devctl_notify(union ccb *ccb)
2126 {
2127         struct cam_periph *periph;
2128         struct ccb_getdev *cgd;
2129         struct sbuf sb;
2130         int serr, sk, asc, ascq;
2131         char *sbmsg, *type;
2132
2133         sbmsg = malloc(CAM_PERIPH_DEVD_MSG_SIZE, M_CAMPERIPH, M_NOWAIT);
2134         if (sbmsg == NULL)
2135                 return;
2136
2137         sbuf_new(&sb, sbmsg, CAM_PERIPH_DEVD_MSG_SIZE, SBUF_FIXEDLEN);
2138
2139         periph = xpt_path_periph(ccb->ccb_h.path);
2140         sbuf_printf(&sb, "device=%s%d ", periph->periph_name,
2141             periph->unit_number);
2142
2143         sbuf_printf(&sb, "serial=\"");
2144         if ((cgd = (struct ccb_getdev *)xpt_alloc_ccb_nowait()) != NULL) {
2145                 xpt_setup_ccb(&cgd->ccb_h, ccb->ccb_h.path,
2146                     CAM_PRIORITY_NORMAL);
2147                 cgd->ccb_h.func_code = XPT_GDEV_TYPE;
2148                 xpt_action((union ccb *)cgd);
2149
2150                 if (cgd->ccb_h.status == CAM_REQ_CMP)
2151                         sbuf_bcat(&sb, cgd->serial_num, cgd->serial_num_len);
2152                 xpt_free_ccb((union ccb *)cgd);
2153         }
2154         sbuf_printf(&sb, "\" ");
2155         sbuf_printf(&sb, "cam_status=\"0x%x\" ", ccb->ccb_h.status);
2156
2157         switch (ccb->ccb_h.status & CAM_STATUS_MASK) {
2158         case CAM_CMD_TIMEOUT:
2159                 sbuf_printf(&sb, "timeout=%d ", ccb->ccb_h.timeout);
2160                 type = "timeout";
2161                 break;
2162         case CAM_SCSI_STATUS_ERROR:
2163                 sbuf_printf(&sb, "scsi_status=%d ", ccb->csio.scsi_status);
2164                 if (scsi_extract_sense_ccb(ccb, &serr, &sk, &asc, &ascq))
2165                         sbuf_printf(&sb, "scsi_sense=\"%02x %02x %02x %02x\" ",
2166                             serr, sk, asc, ascq);
2167                 type = "error";
2168                 break;
2169         case CAM_ATA_STATUS_ERROR:
2170                 sbuf_printf(&sb, "RES=\"");
2171                 ata_res_sbuf(&ccb->ataio.res, &sb);
2172                 sbuf_printf(&sb, "\" ");
2173                 type = "error";
2174                 break;
2175         default:
2176                 type = "error";
2177                 break;
2178         }
2179
2180         if (ccb->ccb_h.func_code == XPT_SCSI_IO) {
2181                 sbuf_printf(&sb, "CDB=\"");
2182                 scsi_cdb_sbuf(scsiio_cdb_ptr(&ccb->csio), &sb);
2183                 sbuf_printf(&sb, "\" ");
2184         } else if (ccb->ccb_h.func_code == XPT_ATA_IO) {
2185                 sbuf_printf(&sb, "ACB=\"");
2186                 ata_cmd_sbuf(&ccb->ataio.cmd, &sb);
2187                 sbuf_printf(&sb, "\" ");
2188         }
2189
2190         if (sbuf_finish(&sb) == 0)
2191                 devctl_notify("CAM", "periph", type, sbuf_data(&sb));
2192         sbuf_delete(&sb);
2193         free(sbmsg, M_CAMPERIPH);
2194 }
2195
2196 /*
2197  * Sysctl to force an invalidation of the drive right now. Can be
2198  * called with CTLFLAG_MPSAFE since we take periph lock.
2199  */
2200 int
2201 cam_periph_invalidate_sysctl(SYSCTL_HANDLER_ARGS)
2202 {
2203         struct cam_periph *periph;
2204         int error, value;
2205
2206         periph = arg1;
2207         value = 0;
2208         error = sysctl_handle_int(oidp, &value, 0, req);
2209         if (error != 0 || req->newptr == NULL || value != 1)
2210                 return (error);
2211
2212         cam_periph_lock(periph);
2213         cam_periph_invalidate(periph);
2214         cam_periph_unlock(periph);
2215
2216         return (0);
2217 }