]> CyberLeo.Net >> Repos - FreeBSD/FreeBSD.git/blob - sys/cam/cam_periph.c
cam: Initialize wired to false
[FreeBSD/FreeBSD.git] / sys / cam / cam_periph.c
1 /*-
2  * Common functions for CAM "type" (peripheral) drivers.
3  *
4  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
5  *
6  * Copyright (c) 1997, 1998 Justin T. Gibbs.
7  * Copyright (c) 1997, 1998, 1999, 2000 Kenneth D. Merry.
8  * All rights reserved.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions, and the following disclaimer,
15  *    without modification, immediately at the beginning of the file.
16  * 2. The name of the author may not be used to endorse or promote products
17  *    derived from this software without specific prior written permission.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR
23  * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  */
31
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/types.h>
38 #include <sys/malloc.h>
39 #include <sys/kernel.h>
40 #include <sys/bio.h>
41 #include <sys/conf.h>
42 #include <sys/devctl.h>
43 #include <sys/lock.h>
44 #include <sys/mutex.h>
45 #include <sys/buf.h>
46 #include <sys/proc.h>
47 #include <sys/devicestat.h>
48 #include <sys/sbuf.h>
49 #include <sys/sysctl.h>
50 #include <vm/vm.h>
51 #include <vm/vm_extern.h>
52
53 #include <cam/cam.h>
54 #include <cam/cam_ccb.h>
55 #include <cam/cam_queue.h>
56 #include <cam/cam_xpt_periph.h>
57 #include <cam/cam_xpt_internal.h>
58 #include <cam/cam_periph.h>
59 #include <cam/cam_debug.h>
60 #include <cam/cam_sim.h>
61
62 #include <cam/scsi/scsi_all.h>
63 #include <cam/scsi/scsi_message.h>
64 #include <cam/scsi/scsi_pass.h>
65
66 static  u_int           camperiphnextunit(struct periph_driver *p_drv,
67                                           u_int newunit, bool wired,
68                                           path_id_t pathid, target_id_t target,
69                                           lun_id_t lun);
70 static  u_int           camperiphunit(struct periph_driver *p_drv,
71                                       path_id_t pathid, target_id_t target,
72                                       lun_id_t lun,
73                                       const char *sn);
74 static  void            camperiphdone(struct cam_periph *periph, 
75                                         union ccb *done_ccb);
76 static  void            camperiphfree(struct cam_periph *periph);
77 static int              camperiphscsistatuserror(union ccb *ccb,
78                                                 union ccb **orig_ccb,
79                                                  cam_flags camflags,
80                                                  u_int32_t sense_flags,
81                                                  int *openings,
82                                                  u_int32_t *relsim_flags,
83                                                  u_int32_t *timeout,
84                                                  u_int32_t  *action,
85                                                  const char **action_string);
86 static  int             camperiphscsisenseerror(union ccb *ccb,
87                                                 union ccb **orig_ccb,
88                                                 cam_flags camflags,
89                                                 u_int32_t sense_flags,
90                                                 int *openings,
91                                                 u_int32_t *relsim_flags,
92                                                 u_int32_t *timeout,
93                                                 u_int32_t *action,
94                                                 const char **action_string);
95 static void             cam_periph_devctl_notify(union ccb *ccb);
96
97 static int nperiph_drivers;
98 static int initialized = 0;
99 struct periph_driver **periph_drivers;
100
101 static MALLOC_DEFINE(M_CAMPERIPH, "CAM periph", "CAM peripheral buffers");
102
103 static int periph_selto_delay = 1000;
104 TUNABLE_INT("kern.cam.periph_selto_delay", &periph_selto_delay);
105 static int periph_noresrc_delay = 500;
106 TUNABLE_INT("kern.cam.periph_noresrc_delay", &periph_noresrc_delay);
107 static int periph_busy_delay = 500;
108 TUNABLE_INT("kern.cam.periph_busy_delay", &periph_busy_delay);
109
110 static u_int periph_mapmem_thresh = 65536;
111 SYSCTL_UINT(_kern_cam, OID_AUTO, mapmem_thresh, CTLFLAG_RWTUN,
112     &periph_mapmem_thresh, 0, "Threshold for user-space buffer mapping");
113
114 void
115 periphdriver_register(void *data)
116 {
117         struct periph_driver *drv = (struct periph_driver *)data;
118         struct periph_driver **newdrivers, **old;
119         int ndrivers;
120
121 again:
122         ndrivers = nperiph_drivers + 2;
123         newdrivers = malloc(sizeof(*newdrivers) * ndrivers, M_CAMPERIPH,
124                             M_WAITOK);
125         xpt_lock_buses();
126         if (ndrivers != nperiph_drivers + 2) {
127                 /*
128                  * Lost race against itself; go around.
129                  */
130                 xpt_unlock_buses();
131                 free(newdrivers, M_CAMPERIPH);
132                 goto again;
133         }
134         if (periph_drivers)
135                 bcopy(periph_drivers, newdrivers,
136                       sizeof(*newdrivers) * nperiph_drivers);
137         newdrivers[nperiph_drivers] = drv;
138         newdrivers[nperiph_drivers + 1] = NULL;
139         old = periph_drivers;
140         periph_drivers = newdrivers;
141         nperiph_drivers++;
142         xpt_unlock_buses();
143         if (old)
144                 free(old, M_CAMPERIPH);
145         /* If driver marked as early or it is late now, initialize it. */
146         if (((drv->flags & CAM_PERIPH_DRV_EARLY) != 0 && initialized > 0) ||
147             initialized > 1)
148                 (*drv->init)();
149 }
150
151 int
152 periphdriver_unregister(void *data)
153 {
154         struct periph_driver *drv = (struct periph_driver *)data;
155         int error, n;
156
157         /* If driver marked as early or it is late now, deinitialize it. */
158         if (((drv->flags & CAM_PERIPH_DRV_EARLY) != 0 && initialized > 0) ||
159             initialized > 1) {
160                 if (drv->deinit == NULL) {
161                         printf("CAM periph driver '%s' doesn't have deinit.\n",
162                             drv->driver_name);
163                         return (EOPNOTSUPP);
164                 }
165                 error = drv->deinit();
166                 if (error != 0)
167                         return (error);
168         }
169
170         xpt_lock_buses();
171         for (n = 0; n < nperiph_drivers && periph_drivers[n] != drv; n++)
172                 ;
173         KASSERT(n < nperiph_drivers,
174             ("Periph driver '%s' was not registered", drv->driver_name));
175         for (; n + 1 < nperiph_drivers; n++)
176                 periph_drivers[n] = periph_drivers[n + 1];
177         periph_drivers[n + 1] = NULL;
178         nperiph_drivers--;
179         xpt_unlock_buses();
180         return (0);
181 }
182
183 void
184 periphdriver_init(int level)
185 {
186         int     i, early;
187
188         initialized = max(initialized, level);
189         for (i = 0; periph_drivers[i] != NULL; i++) {
190                 early = (periph_drivers[i]->flags & CAM_PERIPH_DRV_EARLY) ? 1 : 2;
191                 if (early == initialized)
192                         (*periph_drivers[i]->init)();
193         }
194 }
195
196 cam_status
197 cam_periph_alloc(periph_ctor_t *periph_ctor,
198                  periph_oninv_t *periph_oninvalidate,
199                  periph_dtor_t *periph_dtor, periph_start_t *periph_start,
200                  char *name, cam_periph_type type, struct cam_path *path,
201                  ac_callback_t *ac_callback, ac_code code, void *arg)
202 {
203         struct          periph_driver **p_drv;
204         struct          cam_sim *sim;
205         struct          cam_periph *periph;
206         struct          cam_periph *cur_periph;
207         path_id_t       path_id;
208         target_id_t     target_id;
209         lun_id_t        lun_id;
210         cam_status      status;
211         u_int           init_level;
212
213         init_level = 0;
214         /*
215          * Handle Hot-Plug scenarios.  If there is already a peripheral
216          * of our type assigned to this path, we are likely waiting for
217          * final close on an old, invalidated, peripheral.  If this is
218          * the case, queue up a deferred call to the peripheral's async
219          * handler.  If it looks like a mistaken re-allocation, complain.
220          */
221         if ((periph = cam_periph_find(path, name)) != NULL) {
222                 if ((periph->flags & CAM_PERIPH_INVALID) != 0
223                  && (periph->flags & CAM_PERIPH_NEW_DEV_FOUND) == 0) {
224                         periph->flags |= CAM_PERIPH_NEW_DEV_FOUND;
225                         periph->deferred_callback = ac_callback;
226                         periph->deferred_ac = code;
227                         return (CAM_REQ_INPROG);
228                 } else {
229                         printf("cam_periph_alloc: attempt to re-allocate "
230                                "valid device %s%d rejected flags %#x "
231                                "refcount %d\n", periph->periph_name,
232                                periph->unit_number, periph->flags,
233                                periph->refcount);
234                 }
235                 return (CAM_REQ_INVALID);
236         }
237
238         periph = (struct cam_periph *)malloc(sizeof(*periph), M_CAMPERIPH,
239                                              M_NOWAIT|M_ZERO);
240
241         if (periph == NULL)
242                 return (CAM_RESRC_UNAVAIL);
243
244         init_level++;
245
246         sim = xpt_path_sim(path);
247         path_id = xpt_path_path_id(path);
248         target_id = xpt_path_target_id(path);
249         lun_id = xpt_path_lun_id(path);
250         periph->periph_start = periph_start;
251         periph->periph_dtor = periph_dtor;
252         periph->periph_oninval = periph_oninvalidate;
253         periph->type = type;
254         periph->periph_name = name;
255         periph->scheduled_priority = CAM_PRIORITY_NONE;
256         periph->immediate_priority = CAM_PRIORITY_NONE;
257         periph->refcount = 1;           /* Dropped by invalidation. */
258         periph->sim = sim;
259         SLIST_INIT(&periph->ccb_list);
260         status = xpt_create_path(&path, periph, path_id, target_id, lun_id);
261         if (status != CAM_REQ_CMP)
262                 goto failure;
263         periph->path = path;
264
265         xpt_lock_buses();
266         for (p_drv = periph_drivers; *p_drv != NULL; p_drv++) {
267                 if (strcmp((*p_drv)->driver_name, name) == 0)
268                         break;
269         }
270         if (*p_drv == NULL) {
271                 printf("cam_periph_alloc: invalid periph name '%s'\n", name);
272                 xpt_unlock_buses();
273                 xpt_free_path(periph->path);
274                 free(periph, M_CAMPERIPH);
275                 return (CAM_REQ_INVALID);
276         }
277         periph->unit_number = camperiphunit(*p_drv, path_id, target_id, lun_id,
278             path->device->serial_num);
279         cur_periph = TAILQ_FIRST(&(*p_drv)->units);
280         while (cur_periph != NULL
281             && cur_periph->unit_number < periph->unit_number)
282                 cur_periph = TAILQ_NEXT(cur_periph, unit_links);
283         if (cur_periph != NULL) {
284                 KASSERT(cur_periph->unit_number != periph->unit_number,
285                     ("duplicate units on periph list"));
286                 TAILQ_INSERT_BEFORE(cur_periph, periph, unit_links);
287         } else {
288                 TAILQ_INSERT_TAIL(&(*p_drv)->units, periph, unit_links);
289                 (*p_drv)->generation++;
290         }
291         xpt_unlock_buses();
292
293         init_level++;
294
295         status = xpt_add_periph(periph);
296         if (status != CAM_REQ_CMP)
297                 goto failure;
298
299         init_level++;
300         CAM_DEBUG(periph->path, CAM_DEBUG_INFO, ("Periph created\n"));
301
302         status = periph_ctor(periph, arg);
303
304         if (status == CAM_REQ_CMP)
305                 init_level++;
306
307 failure:
308         switch (init_level) {
309         case 4:
310                 /* Initialized successfully */
311                 break;
312         case 3:
313                 CAM_DEBUG(periph->path, CAM_DEBUG_INFO, ("Periph destroyed\n"));
314                 xpt_remove_periph(periph);
315                 /* FALLTHROUGH */
316         case 2:
317                 xpt_lock_buses();
318                 TAILQ_REMOVE(&(*p_drv)->units, periph, unit_links);
319                 xpt_unlock_buses();
320                 xpt_free_path(periph->path);
321                 /* FALLTHROUGH */
322         case 1:
323                 free(periph, M_CAMPERIPH);
324                 /* FALLTHROUGH */
325         case 0:
326                 /* No cleanup to perform. */
327                 break;
328         default:
329                 panic("%s: Unknown init level", __func__);
330         }
331         return(status);
332 }
333
334 /*
335  * Find a peripheral structure with the specified path, target, lun, 
336  * and (optionally) type.  If the name is NULL, this function will return
337  * the first peripheral driver that matches the specified path.
338  */
339 struct cam_periph *
340 cam_periph_find(struct cam_path *path, char *name)
341 {
342         struct periph_driver **p_drv;
343         struct cam_periph *periph;
344
345         xpt_lock_buses();
346         for (p_drv = periph_drivers; *p_drv != NULL; p_drv++) {
347                 if (name != NULL && (strcmp((*p_drv)->driver_name, name) != 0))
348                         continue;
349
350                 TAILQ_FOREACH(periph, &(*p_drv)->units, unit_links) {
351                         if (xpt_path_comp(periph->path, path) == 0) {
352                                 xpt_unlock_buses();
353                                 cam_periph_assert(periph, MA_OWNED);
354                                 return(periph);
355                         }
356                 }
357                 if (name != NULL) {
358                         xpt_unlock_buses();
359                         return(NULL);
360                 }
361         }
362         xpt_unlock_buses();
363         return(NULL);
364 }
365
366 /*
367  * Find peripheral driver instances attached to the specified path.
368  */
369 int
370 cam_periph_list(struct cam_path *path, struct sbuf *sb)
371 {
372         struct sbuf local_sb;
373         struct periph_driver **p_drv;
374         struct cam_periph *periph;
375         int count;
376         int sbuf_alloc_len;
377
378         sbuf_alloc_len = 16;
379 retry:
380         sbuf_new(&local_sb, NULL, sbuf_alloc_len, SBUF_FIXEDLEN);
381         count = 0;
382         xpt_lock_buses();
383         for (p_drv = periph_drivers; *p_drv != NULL; p_drv++) {
384                 TAILQ_FOREACH(periph, &(*p_drv)->units, unit_links) {
385                         if (xpt_path_comp(periph->path, path) != 0)
386                                 continue;
387
388                         if (sbuf_len(&local_sb) != 0)
389                                 sbuf_cat(&local_sb, ",");
390
391                         sbuf_printf(&local_sb, "%s%d", periph->periph_name,
392                                     periph->unit_number);
393
394                         if (sbuf_error(&local_sb) == ENOMEM) {
395                                 sbuf_alloc_len *= 2;
396                                 xpt_unlock_buses();
397                                 sbuf_delete(&local_sb);
398                                 goto retry;
399                         }
400                         count++;
401                 }
402         }
403         xpt_unlock_buses();
404         sbuf_finish(&local_sb);
405         if (sbuf_len(sb) != 0)
406                 sbuf_cat(sb, ",");
407         sbuf_cat(sb, sbuf_data(&local_sb));
408         sbuf_delete(&local_sb);
409         return (count);
410 }
411
412 int
413 cam_periph_acquire(struct cam_periph *periph)
414 {
415         int status;
416
417         if (periph == NULL)
418                 return (EINVAL);
419
420         status = ENOENT;
421         xpt_lock_buses();
422         if ((periph->flags & CAM_PERIPH_INVALID) == 0) {
423                 periph->refcount++;
424                 status = 0;
425         }
426         xpt_unlock_buses();
427
428         return (status);
429 }
430
431 void
432 cam_periph_doacquire(struct cam_periph *periph)
433 {
434
435         xpt_lock_buses();
436         KASSERT(periph->refcount >= 1,
437             ("cam_periph_doacquire() with refcount == %d", periph->refcount));
438         periph->refcount++;
439         xpt_unlock_buses();
440 }
441
442 void
443 cam_periph_release_locked_buses(struct cam_periph *periph)
444 {
445
446         cam_periph_assert(periph, MA_OWNED);
447         KASSERT(periph->refcount >= 1, ("periph->refcount >= 1"));
448         if (--periph->refcount == 0)
449                 camperiphfree(periph);
450 }
451
452 void
453 cam_periph_release_locked(struct cam_periph *periph)
454 {
455
456         if (periph == NULL)
457                 return;
458
459         xpt_lock_buses();
460         cam_periph_release_locked_buses(periph);
461         xpt_unlock_buses();
462 }
463
464 void
465 cam_periph_release(struct cam_periph *periph)
466 {
467         struct mtx *mtx;
468
469         if (periph == NULL)
470                 return;
471
472         cam_periph_assert(periph, MA_NOTOWNED);
473         mtx = cam_periph_mtx(periph);
474         mtx_lock(mtx);
475         cam_periph_release_locked(periph);
476         mtx_unlock(mtx);
477 }
478
479 /*
480  * hold/unhold act as mutual exclusion for sections of the code that
481  * need to sleep and want to make sure that other sections that
482  * will interfere are held off. This only protects exclusive sections
483  * from each other.
484  */
485 int
486 cam_periph_hold(struct cam_periph *periph, int priority)
487 {
488         int error;
489
490         /*
491          * Increment the reference count on the peripheral
492          * while we wait for our lock attempt to succeed
493          * to ensure the peripheral doesn't disappear out
494          * from user us while we sleep.
495          */
496
497         if (cam_periph_acquire(periph) != 0)
498                 return (ENXIO);
499
500         cam_periph_assert(periph, MA_OWNED);
501         while ((periph->flags & CAM_PERIPH_LOCKED) != 0) {
502                 periph->flags |= CAM_PERIPH_LOCK_WANTED;
503                 if ((error = cam_periph_sleep(periph, periph, priority,
504                     "caplck", 0)) != 0) {
505                         cam_periph_release_locked(periph);
506                         return (error);
507                 }
508                 if (periph->flags & CAM_PERIPH_INVALID) {
509                         cam_periph_release_locked(periph);
510                         return (ENXIO);
511                 }
512         }
513
514         periph->flags |= CAM_PERIPH_LOCKED;
515         return (0);
516 }
517
518 void
519 cam_periph_unhold(struct cam_periph *periph)
520 {
521
522         cam_periph_assert(periph, MA_OWNED);
523
524         periph->flags &= ~CAM_PERIPH_LOCKED;
525         if ((periph->flags & CAM_PERIPH_LOCK_WANTED) != 0) {
526                 periph->flags &= ~CAM_PERIPH_LOCK_WANTED;
527                 wakeup(periph);
528         }
529
530         cam_periph_release_locked(periph);
531 }
532
533 /*
534  * Look for the next unit number that is not currently in use for this
535  * peripheral type starting at "newunit".  Also exclude unit numbers that
536  * are reserved by for future "hardwiring" unless we already know that this
537  * is a potential wired device.  Only assume that the device is "wired" the
538  * first time through the loop since after that we'll be looking at unit
539  * numbers that did not match a wiring entry.
540  */
541 static u_int
542 camperiphnextunit(struct periph_driver *p_drv, u_int newunit, bool wired,
543                   path_id_t pathid, target_id_t target, lun_id_t lun)
544 {
545         struct  cam_periph *periph;
546         char    *periph_name;
547         int     i, val, dunit, r;
548         const char *dname, *strval;
549
550         periph_name = p_drv->driver_name;
551         for (;;newunit++) {
552                 for (periph = TAILQ_FIRST(&p_drv->units);
553                      periph != NULL && periph->unit_number != newunit;
554                      periph = TAILQ_NEXT(periph, unit_links))
555                         ;
556
557                 if (periph != NULL && periph->unit_number == newunit) {
558                         if (wired) {
559                                 xpt_print(periph->path, "Duplicate Wired "
560                                     "Device entry!\n");
561                                 xpt_print(periph->path, "Second device (%s "
562                                     "device at scbus%d target %d lun %d) will "
563                                     "not be wired\n", periph_name, pathid,
564                                     target, lun);
565                                 wired = false;
566                         }
567                         continue;
568                 }
569                 if (wired)
570                         break;
571
572                 /*
573                  * Don't allow the mere presence of any attributes of a device
574                  * means that it is for a wired down entry. Instead, insist that
575                  * one of the matching criteria from camperiphunit be present
576                  * for the device.
577                  */
578                 i = 0;
579                 dname = periph_name;
580                 for (;;) {
581                         r = resource_find_dev(&i, dname, &dunit, NULL, NULL);
582                         if (r != 0)
583                                 break;
584
585                         if (newunit != dunit)
586                                 continue;
587                         if (resource_string_value(dname, dunit, "sn", &strval) == 0 ||
588                             resource_int_value(dname, dunit, "lun", &val) == 0 ||
589                             resource_int_value(dname, dunit, "target", &val) == 0 ||
590                             resource_string_value(dname, dunit, "at", &strval) == 0)
591                                 break;
592                 }
593                 if (r != 0)
594                         break;
595         }
596         return (newunit);
597 }
598
599 static u_int
600 camperiphunit(struct periph_driver *p_drv, path_id_t pathid,
601     target_id_t target, lun_id_t lun, const char *sn)
602 {
603         bool    wired = false;
604         u_int   unit;
605         int     i, val, dunit;
606         const char *dname, *strval;
607         char    pathbuf[32], *periph_name;
608
609         periph_name = p_drv->driver_name;
610         snprintf(pathbuf, sizeof(pathbuf), "scbus%d", pathid);
611         unit = 0;
612         i = 0;
613         dname = periph_name;
614         while (resource_find_dev(&i, dname, &dunit, NULL, NULL) == 0) {
615                 wired = false;
616                 if (resource_string_value(dname, dunit, "at", &strval) == 0) {
617                         if (strcmp(strval, pathbuf) != 0)
618                                 continue;
619                         wired = true;
620                 }
621                 if (resource_int_value(dname, dunit, "target", &val) == 0) {
622                         if (val != target)
623                                 continue;
624                         wired = true;
625                 }
626                 if (resource_int_value(dname, dunit, "lun", &val) == 0) {
627                         if (val != lun)
628                                 continue;
629                         wired = true;
630                 }
631                 if (resource_string_value(dname, dunit, "sn", &strval) == 0) {
632                         if (sn == NULL || strcmp(strval, sn) != 0)
633                                 continue;
634                         wired = true;
635                 }
636                 if (wired) {
637                         unit = dunit;
638                         break;
639                 }
640         }
641
642         /*
643          * Either start from 0 looking for the next unit or from
644          * the unit number given in the resource config.  This way,
645          * if we have wildcard matches, we don't return the same
646          * unit number twice.
647          */
648         unit = camperiphnextunit(p_drv, unit, wired, pathid, target, lun);
649
650         return (unit);
651 }
652
653 void
654 cam_periph_invalidate(struct cam_periph *periph)
655 {
656
657         cam_periph_assert(periph, MA_OWNED);
658         /*
659          * We only tear down the device the first time a peripheral is
660          * invalidated.
661          */
662         if ((periph->flags & CAM_PERIPH_INVALID) != 0)
663                 return;
664
665         CAM_DEBUG(periph->path, CAM_DEBUG_INFO, ("Periph invalidated\n"));
666         if ((periph->flags & CAM_PERIPH_ANNOUNCED) && !rebooting) {
667                 struct sbuf sb;
668                 char buffer[160];
669
670                 sbuf_new(&sb, buffer, 160, SBUF_FIXEDLEN);
671                 xpt_denounce_periph_sbuf(periph, &sb);
672                 sbuf_finish(&sb);
673                 sbuf_putbuf(&sb);
674         }
675         periph->flags |= CAM_PERIPH_INVALID;
676         periph->flags &= ~CAM_PERIPH_NEW_DEV_FOUND;
677         if (periph->periph_oninval != NULL)
678                 periph->periph_oninval(periph);
679         cam_periph_release_locked(periph);
680 }
681
682 static void
683 camperiphfree(struct cam_periph *periph)
684 {
685         struct periph_driver **p_drv;
686         struct periph_driver *drv;
687
688         cam_periph_assert(periph, MA_OWNED);
689         KASSERT(periph->periph_allocating == 0, ("%s%d: freed while allocating",
690             periph->periph_name, periph->unit_number));
691         for (p_drv = periph_drivers; *p_drv != NULL; p_drv++) {
692                 if (strcmp((*p_drv)->driver_name, periph->periph_name) == 0)
693                         break;
694         }
695         if (*p_drv == NULL) {
696                 printf("camperiphfree: attempt to free non-existant periph\n");
697                 return;
698         }
699         /*
700          * Cache a pointer to the periph_driver structure.  If a
701          * periph_driver is added or removed from the array (see
702          * periphdriver_register()) while we drop the toplogy lock
703          * below, p_drv may change.  This doesn't protect against this
704          * particular periph_driver going away.  That will require full
705          * reference counting in the periph_driver infrastructure.
706          */
707         drv = *p_drv;
708
709         /*
710          * We need to set this flag before dropping the topology lock, to
711          * let anyone who is traversing the list that this peripheral is
712          * about to be freed, and there will be no more reference count
713          * checks.
714          */
715         periph->flags |= CAM_PERIPH_FREE;
716
717         /*
718          * The peripheral destructor semantics dictate calling with only the
719          * SIM mutex held.  Since it might sleep, it should not be called
720          * with the topology lock held.
721          */
722         xpt_unlock_buses();
723
724         /*
725          * We need to call the peripheral destructor prior to removing the
726          * peripheral from the list.  Otherwise, we risk running into a
727          * scenario where the peripheral unit number may get reused
728          * (because it has been removed from the list), but some resources
729          * used by the peripheral are still hanging around.  In particular,
730          * the devfs nodes used by some peripherals like the pass(4) driver
731          * aren't fully cleaned up until the destructor is run.  If the
732          * unit number is reused before the devfs instance is fully gone,
733          * devfs will panic.
734          */
735         if (periph->periph_dtor != NULL)
736                 periph->periph_dtor(periph);
737
738         /*
739          * The peripheral list is protected by the topology lock. We have to
740          * remove the periph from the drv list before we call deferred_ac. The
741          * AC_FOUND_DEVICE callback won't create a new periph if it's still there.
742          */
743         xpt_lock_buses();
744
745         TAILQ_REMOVE(&drv->units, periph, unit_links);
746         drv->generation++;
747
748         xpt_remove_periph(periph);
749
750         xpt_unlock_buses();
751         if ((periph->flags & CAM_PERIPH_ANNOUNCED) && !rebooting)
752                 xpt_print(periph->path, "Periph destroyed\n");
753         else
754                 CAM_DEBUG(periph->path, CAM_DEBUG_INFO, ("Periph destroyed\n"));
755
756         if (periph->flags & CAM_PERIPH_NEW_DEV_FOUND) {
757                 union ccb ccb;
758                 void *arg;
759
760                 memset(&ccb, 0, sizeof(ccb));
761                 switch (periph->deferred_ac) {
762                 case AC_FOUND_DEVICE:
763                         ccb.ccb_h.func_code = XPT_GDEV_TYPE;
764                         xpt_setup_ccb(&ccb.ccb_h, periph->path, CAM_PRIORITY_NORMAL);
765                         xpt_action(&ccb);
766                         arg = &ccb;
767                         break;
768                 case AC_PATH_REGISTERED:
769                         xpt_path_inq(&ccb.cpi, periph->path);
770                         arg = &ccb;
771                         break;
772                 default:
773                         arg = NULL;
774                         break;
775                 }
776                 periph->deferred_callback(NULL, periph->deferred_ac,
777                                           periph->path, arg);
778         }
779         xpt_free_path(periph->path);
780         free(periph, M_CAMPERIPH);
781         xpt_lock_buses();
782 }
783
784 /*
785  * Map user virtual pointers into kernel virtual address space, so we can
786  * access the memory.  This is now a generic function that centralizes most
787  * of the sanity checks on the data flags, if any.
788  * This also only works for up to maxphys memory.  Since we use
789  * buffers to map stuff in and out, we're limited to the buffer size.
790  */
791 int
792 cam_periph_mapmem(union ccb *ccb, struct cam_periph_map_info *mapinfo,
793     u_int maxmap)
794 {
795         int numbufs, i;
796         u_int8_t **data_ptrs[CAM_PERIPH_MAXMAPS];
797         u_int32_t lengths[CAM_PERIPH_MAXMAPS];
798         u_int32_t dirs[CAM_PERIPH_MAXMAPS];
799
800         bzero(mapinfo, sizeof(*mapinfo));
801         if (maxmap == 0)
802                 maxmap = DFLTPHYS;      /* traditional default */
803         else if (maxmap > maxphys)
804                 maxmap = maxphys;       /* for safety */
805         switch(ccb->ccb_h.func_code) {
806         case XPT_DEV_MATCH:
807                 if (ccb->cdm.match_buf_len == 0) {
808                         printf("cam_periph_mapmem: invalid match buffer "
809                                "length 0\n");
810                         return(EINVAL);
811                 }
812                 if (ccb->cdm.pattern_buf_len > 0) {
813                         data_ptrs[0] = (u_int8_t **)&ccb->cdm.patterns;
814                         lengths[0] = ccb->cdm.pattern_buf_len;
815                         dirs[0] = CAM_DIR_OUT;
816                         data_ptrs[1] = (u_int8_t **)&ccb->cdm.matches;
817                         lengths[1] = ccb->cdm.match_buf_len;
818                         dirs[1] = CAM_DIR_IN;
819                         numbufs = 2;
820                 } else {
821                         data_ptrs[0] = (u_int8_t **)&ccb->cdm.matches;
822                         lengths[0] = ccb->cdm.match_buf_len;
823                         dirs[0] = CAM_DIR_IN;
824                         numbufs = 1;
825                 }
826                 /*
827                  * This request will not go to the hardware, no reason
828                  * to be so strict. vmapbuf() is able to map up to maxphys.
829                  */
830                 maxmap = maxphys;
831                 break;
832         case XPT_SCSI_IO:
833         case XPT_CONT_TARGET_IO:
834                 if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
835                         return(0);
836                 if ((ccb->ccb_h.flags & CAM_DATA_MASK) != CAM_DATA_VADDR)
837                         return (EINVAL);
838                 data_ptrs[0] = &ccb->csio.data_ptr;
839                 lengths[0] = ccb->csio.dxfer_len;
840                 dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
841                 numbufs = 1;
842                 break;
843         case XPT_ATA_IO:
844                 if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
845                         return(0);
846                 if ((ccb->ccb_h.flags & CAM_DATA_MASK) != CAM_DATA_VADDR)
847                         return (EINVAL);
848                 data_ptrs[0] = &ccb->ataio.data_ptr;
849                 lengths[0] = ccb->ataio.dxfer_len;
850                 dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
851                 numbufs = 1;
852                 break;
853         case XPT_MMC_IO:
854                 if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
855                         return(0);
856                 /* Two mappings: one for cmd->data and one for cmd->data->data */
857                 data_ptrs[0] = (unsigned char **)&ccb->mmcio.cmd.data;
858                 lengths[0] = sizeof(struct mmc_data *);
859                 dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
860                 data_ptrs[1] = (unsigned char **)&ccb->mmcio.cmd.data->data;
861                 lengths[1] = ccb->mmcio.cmd.data->len;
862                 dirs[1] = ccb->ccb_h.flags & CAM_DIR_MASK;
863                 numbufs = 2;
864                 break;
865         case XPT_SMP_IO:
866                 data_ptrs[0] = &ccb->smpio.smp_request;
867                 lengths[0] = ccb->smpio.smp_request_len;
868                 dirs[0] = CAM_DIR_OUT;
869                 data_ptrs[1] = &ccb->smpio.smp_response;
870                 lengths[1] = ccb->smpio.smp_response_len;
871                 dirs[1] = CAM_DIR_IN;
872                 numbufs = 2;
873                 break;
874         case XPT_NVME_IO:
875         case XPT_NVME_ADMIN:
876                 if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
877                         return (0);
878                 if ((ccb->ccb_h.flags & CAM_DATA_MASK) != CAM_DATA_VADDR)
879                         return (EINVAL);
880                 data_ptrs[0] = &ccb->nvmeio.data_ptr;
881                 lengths[0] = ccb->nvmeio.dxfer_len;
882                 dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
883                 numbufs = 1;
884                 break;
885         case XPT_DEV_ADVINFO:
886                 if (ccb->cdai.bufsiz == 0)
887                         return (0);
888
889                 data_ptrs[0] = (uint8_t **)&ccb->cdai.buf;
890                 lengths[0] = ccb->cdai.bufsiz;
891                 dirs[0] = CAM_DIR_IN;
892                 numbufs = 1;
893
894                 /*
895                  * This request will not go to the hardware, no reason
896                  * to be so strict. vmapbuf() is able to map up to maxphys.
897                  */
898                 maxmap = maxphys;
899                 break;
900         default:
901                 return(EINVAL);
902                 break; /* NOTREACHED */
903         }
904
905         /*
906          * Check the transfer length and permissions first, so we don't
907          * have to unmap any previously mapped buffers.
908          */
909         for (i = 0; i < numbufs; i++) {
910                 if (lengths[i] > maxmap) {
911                         printf("cam_periph_mapmem: attempt to map %lu bytes, "
912                                "which is greater than %lu\n",
913                                (long)(lengths[i]), (u_long)maxmap);
914                         return (E2BIG);
915                 }
916         }
917
918         /*
919          * This keeps the kernel stack of current thread from getting
920          * swapped.  In low-memory situations where the kernel stack might
921          * otherwise get swapped out, this holds it and allows the thread
922          * to make progress and release the kernel mapped pages sooner.
923          *
924          * XXX KDM should I use P_NOSWAP instead?
925          */
926         PHOLD(curproc);
927
928         for (i = 0; i < numbufs; i++) {
929                 /* Save the user's data address. */
930                 mapinfo->orig[i] = *data_ptrs[i];
931
932                 /*
933                  * For small buffers use malloc+copyin/copyout instead of
934                  * mapping to KVA to avoid expensive TLB shootdowns.  For
935                  * small allocations malloc is backed by UMA, and so much
936                  * cheaper on SMP systems.
937                  */
938                 if (lengths[i] <= periph_mapmem_thresh &&
939                     ccb->ccb_h.func_code != XPT_MMC_IO) {
940                         *data_ptrs[i] = malloc(lengths[i], M_CAMPERIPH,
941                             M_WAITOK);
942                         if (dirs[i] != CAM_DIR_IN) {
943                                 if (copyin(mapinfo->orig[i], *data_ptrs[i],
944                                     lengths[i]) != 0) {
945                                         free(*data_ptrs[i], M_CAMPERIPH);
946                                         *data_ptrs[i] = mapinfo->orig[i];
947                                         goto fail;
948                                 }
949                         } else
950                                 bzero(*data_ptrs[i], lengths[i]);
951                         continue;
952                 }
953
954                 /*
955                  * Get the buffer.
956                  */
957                 mapinfo->bp[i] = uma_zalloc(pbuf_zone, M_WAITOK);
958
959                 /* set the direction */
960                 mapinfo->bp[i]->b_iocmd = (dirs[i] == CAM_DIR_OUT) ?
961                     BIO_WRITE : BIO_READ;
962
963                 /* Map the buffer into kernel memory. */
964                 if (vmapbuf(mapinfo->bp[i], *data_ptrs[i], lengths[i], 1) < 0) {
965                         uma_zfree(pbuf_zone, mapinfo->bp[i]);
966                         goto fail;
967                 }
968
969                 /* set our pointer to the new mapped area */
970                 *data_ptrs[i] = mapinfo->bp[i]->b_data;
971         }
972
973         /*
974          * Now that we've gotten this far, change ownership to the kernel
975          * of the buffers so that we don't run afoul of returning to user
976          * space with locks (on the buffer) held.
977          */
978         for (i = 0; i < numbufs; i++) {
979                 if (mapinfo->bp[i])
980                         BUF_KERNPROC(mapinfo->bp[i]);
981         }
982
983         mapinfo->num_bufs_used = numbufs;
984         return(0);
985
986 fail:
987         for (i--; i >= 0; i--) {
988                 if (mapinfo->bp[i]) {
989                         vunmapbuf(mapinfo->bp[i]);
990                         uma_zfree(pbuf_zone, mapinfo->bp[i]);
991                 } else
992                         free(*data_ptrs[i], M_CAMPERIPH);
993                 *data_ptrs[i] = mapinfo->orig[i];
994         }
995         PRELE(curproc);
996         return(EACCES);
997 }
998
999 /*
1000  * Unmap memory segments mapped into kernel virtual address space by
1001  * cam_periph_mapmem().
1002  */
1003 void
1004 cam_periph_unmapmem(union ccb *ccb, struct cam_periph_map_info *mapinfo)
1005 {
1006         int numbufs, i;
1007         u_int8_t **data_ptrs[CAM_PERIPH_MAXMAPS];
1008         u_int32_t lengths[CAM_PERIPH_MAXMAPS];
1009         u_int32_t dirs[CAM_PERIPH_MAXMAPS];
1010
1011         if (mapinfo->num_bufs_used <= 0) {
1012                 /* nothing to free and the process wasn't held. */
1013                 return;
1014         }
1015
1016         switch (ccb->ccb_h.func_code) {
1017         case XPT_DEV_MATCH:
1018                 if (ccb->cdm.pattern_buf_len > 0) {
1019                         data_ptrs[0] = (u_int8_t **)&ccb->cdm.patterns;
1020                         lengths[0] = ccb->cdm.pattern_buf_len;
1021                         dirs[0] = CAM_DIR_OUT;
1022                         data_ptrs[1] = (u_int8_t **)&ccb->cdm.matches;
1023                         lengths[1] = ccb->cdm.match_buf_len;
1024                         dirs[1] = CAM_DIR_IN;
1025                         numbufs = 2;
1026                 } else {
1027                         data_ptrs[0] = (u_int8_t **)&ccb->cdm.matches;
1028                         lengths[0] = ccb->cdm.match_buf_len;
1029                         dirs[0] = CAM_DIR_IN;
1030                         numbufs = 1;
1031                 }
1032                 break;
1033         case XPT_SCSI_IO:
1034         case XPT_CONT_TARGET_IO:
1035                 data_ptrs[0] = &ccb->csio.data_ptr;
1036                 lengths[0] = ccb->csio.dxfer_len;
1037                 dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1038                 numbufs = 1;
1039                 break;
1040         case XPT_ATA_IO:
1041                 data_ptrs[0] = &ccb->ataio.data_ptr;
1042                 lengths[0] = ccb->ataio.dxfer_len;
1043                 dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1044                 numbufs = 1;
1045                 break;
1046         case XPT_MMC_IO:
1047                 data_ptrs[0] = (u_int8_t **)&ccb->mmcio.cmd.data;
1048                 lengths[0] = sizeof(struct mmc_data *);
1049                 dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1050                 data_ptrs[1] = (u_int8_t **)&ccb->mmcio.cmd.data->data;
1051                 lengths[1] = ccb->mmcio.cmd.data->len;
1052                 dirs[1] = ccb->ccb_h.flags & CAM_DIR_MASK;
1053                 numbufs = 2;
1054                 break;
1055         case XPT_SMP_IO:
1056                 data_ptrs[0] = &ccb->smpio.smp_request;
1057                 lengths[0] = ccb->smpio.smp_request_len;
1058                 dirs[0] = CAM_DIR_OUT;
1059                 data_ptrs[1] = &ccb->smpio.smp_response;
1060                 lengths[1] = ccb->smpio.smp_response_len;
1061                 dirs[1] = CAM_DIR_IN;
1062                 numbufs = 2;
1063                 break;
1064         case XPT_NVME_IO:
1065         case XPT_NVME_ADMIN:
1066                 data_ptrs[0] = &ccb->nvmeio.data_ptr;
1067                 lengths[0] = ccb->nvmeio.dxfer_len;
1068                 dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1069                 numbufs = 1;
1070                 break;
1071         case XPT_DEV_ADVINFO:
1072                 data_ptrs[0] = (uint8_t **)&ccb->cdai.buf;
1073                 lengths[0] = ccb->cdai.bufsiz;
1074                 dirs[0] = CAM_DIR_IN;
1075                 numbufs = 1;
1076                 break;
1077         default:
1078                 /* allow ourselves to be swapped once again */
1079                 PRELE(curproc);
1080                 return;
1081                 break; /* NOTREACHED */ 
1082         }
1083
1084         for (i = 0; i < numbufs; i++) {
1085                 if (mapinfo->bp[i]) {
1086                         /* unmap the buffer */
1087                         vunmapbuf(mapinfo->bp[i]);
1088
1089                         /* release the buffer */
1090                         uma_zfree(pbuf_zone, mapinfo->bp[i]);
1091                 } else {
1092                         if (dirs[i] != CAM_DIR_OUT) {
1093                                 copyout(*data_ptrs[i], mapinfo->orig[i],
1094                                     lengths[i]);
1095                         }
1096                         free(*data_ptrs[i], M_CAMPERIPH);
1097                 }
1098
1099                 /* Set the user's pointer back to the original value */
1100                 *data_ptrs[i] = mapinfo->orig[i];
1101         }
1102
1103         /* allow ourselves to be swapped once again */
1104         PRELE(curproc);
1105 }
1106
1107 int
1108 cam_periph_ioctl(struct cam_periph *periph, u_long cmd, caddr_t addr,
1109                  int (*error_routine)(union ccb *ccb, 
1110                                       cam_flags camflags,
1111                                       u_int32_t sense_flags))
1112 {
1113         union ccb            *ccb;
1114         int                  error;
1115         int                  found;
1116
1117         error = found = 0;
1118
1119         switch(cmd){
1120         case CAMGETPASSTHRU:
1121                 ccb = cam_periph_getccb(periph, CAM_PRIORITY_NORMAL);
1122                 xpt_setup_ccb(&ccb->ccb_h,
1123                               ccb->ccb_h.path,
1124                               CAM_PRIORITY_NORMAL);
1125                 ccb->ccb_h.func_code = XPT_GDEVLIST;
1126
1127                 /*
1128                  * Basically, the point of this is that we go through
1129                  * getting the list of devices, until we find a passthrough
1130                  * device.  In the current version of the CAM code, the
1131                  * only way to determine what type of device we're dealing
1132                  * with is by its name.
1133                  */
1134                 while (found == 0) {
1135                         ccb->cgdl.index = 0;
1136                         ccb->cgdl.status = CAM_GDEVLIST_MORE_DEVS;
1137                         while (ccb->cgdl.status == CAM_GDEVLIST_MORE_DEVS) {
1138                                 /* we want the next device in the list */
1139                                 xpt_action(ccb);
1140                                 if (strncmp(ccb->cgdl.periph_name, 
1141                                     "pass", 4) == 0){
1142                                         found = 1;
1143                                         break;
1144                                 }
1145                         }
1146                         if ((ccb->cgdl.status == CAM_GDEVLIST_LAST_DEVICE) &&
1147                             (found == 0)) {
1148                                 ccb->cgdl.periph_name[0] = '\0';
1149                                 ccb->cgdl.unit_number = 0;
1150                                 break;
1151                         }
1152                 }
1153
1154                 /* copy the result back out */  
1155                 bcopy(ccb, addr, sizeof(union ccb));
1156
1157                 /* and release the ccb */
1158                 xpt_release_ccb(ccb);
1159
1160                 break;
1161         default:
1162                 error = ENOTTY;
1163                 break;
1164         }
1165         return(error);
1166 }
1167
1168 static void
1169 cam_periph_done_panic(struct cam_periph *periph, union ccb *done_ccb)
1170 {
1171
1172         panic("%s: already done with ccb %p", __func__, done_ccb);
1173 }
1174
1175 static void
1176 cam_periph_done(struct cam_periph *periph, union ccb *done_ccb)
1177 {
1178
1179         /* Caller will release the CCB */
1180         xpt_path_assert(done_ccb->ccb_h.path, MA_OWNED);
1181         done_ccb->ccb_h.cbfcnp = cam_periph_done_panic;
1182         wakeup(&done_ccb->ccb_h.cbfcnp);
1183 }
1184
1185 static void
1186 cam_periph_ccbwait(union ccb *ccb)
1187 {
1188
1189         if ((ccb->ccb_h.func_code & XPT_FC_QUEUED) != 0) {
1190                 while (ccb->ccb_h.cbfcnp != cam_periph_done_panic)
1191                         xpt_path_sleep(ccb->ccb_h.path, &ccb->ccb_h.cbfcnp,
1192                             PRIBIO, "cbwait", 0);
1193         }
1194         KASSERT(ccb->ccb_h.pinfo.index == CAM_UNQUEUED_INDEX &&
1195             (ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_INPROG,
1196             ("%s: proceeding with incomplete ccb: ccb=%p, func_code=%#x, "
1197              "status=%#x, index=%d", __func__, ccb, ccb->ccb_h.func_code,
1198              ccb->ccb_h.status, ccb->ccb_h.pinfo.index));
1199 }
1200
1201 /*
1202  * Dispatch a CCB and wait for it to complete.  If the CCB has set a
1203  * callback function (ccb->ccb_h.cbfcnp), it will be overwritten and lost.
1204  */
1205 int
1206 cam_periph_runccb(union ccb *ccb,
1207                   int (*error_routine)(union ccb *ccb,
1208                                        cam_flags camflags,
1209                                        u_int32_t sense_flags),
1210                   cam_flags camflags, u_int32_t sense_flags,
1211                   struct devstat *ds)
1212 {
1213         struct bintime *starttime;
1214         struct bintime ltime;
1215         int error;
1216         bool must_poll;
1217         uint32_t timeout = 1;
1218
1219         starttime = NULL;
1220         xpt_path_assert(ccb->ccb_h.path, MA_OWNED);
1221         KASSERT((ccb->ccb_h.flags & CAM_UNLOCKED) == 0,
1222             ("%s: ccb=%p, func_code=%#x, flags=%#x", __func__, ccb,
1223              ccb->ccb_h.func_code, ccb->ccb_h.flags));
1224
1225         /*
1226          * If the user has supplied a stats structure, and if we understand
1227          * this particular type of ccb, record the transaction start.
1228          */
1229         if (ds != NULL &&
1230             (ccb->ccb_h.func_code == XPT_SCSI_IO ||
1231             ccb->ccb_h.func_code == XPT_ATA_IO ||
1232             ccb->ccb_h.func_code == XPT_NVME_IO)) {
1233                 starttime = &ltime;
1234                 binuptime(starttime);
1235                 devstat_start_transaction(ds, starttime);
1236         }
1237
1238         /*
1239          * We must poll the I/O while we're dumping. The scheduler is normally
1240          * stopped for dumping, except when we call doadump from ddb. While the
1241          * scheduler is running in this case, we still need to poll the I/O to
1242          * avoid sleeping waiting for the ccb to complete.
1243          *
1244          * A panic triggered dump stops the scheduler, any callback from the
1245          * shutdown_post_sync event will run with the scheduler stopped, but
1246          * before we're officially dumping. To avoid hanging in adashutdown
1247          * initiated commands (or other similar situations), we have to test for
1248          * either SCHEDULER_STOPPED() here as well.
1249          *
1250          * To avoid locking problems, dumping/polling callers must call
1251          * without a periph lock held.
1252          */
1253         must_poll = dumping || SCHEDULER_STOPPED();
1254         ccb->ccb_h.cbfcnp = cam_periph_done;
1255
1256         /*
1257          * If we're polling, then we need to ensure that we have ample resources
1258          * in the periph.  cam_periph_error can reschedule the ccb by calling
1259          * xpt_action and returning ERESTART, so we have to effect the polling
1260          * in the do loop below.
1261          */
1262         if (must_poll) {
1263                 if (cam_sim_pollable(ccb->ccb_h.path->bus->sim))
1264                         timeout = xpt_poll_setup(ccb);
1265                 else
1266                         timeout = 0;
1267         }
1268
1269         if (timeout == 0) {
1270                 ccb->ccb_h.status = CAM_RESRC_UNAVAIL;
1271                 error = EBUSY;
1272         } else {
1273                 xpt_action(ccb);
1274                 do {
1275                         if (must_poll) {
1276                                 xpt_pollwait(ccb, timeout);
1277                                 timeout = ccb->ccb_h.timeout * 10;
1278                         } else {
1279                                 cam_periph_ccbwait(ccb);
1280                         }
1281                         if ((ccb->ccb_h.status & CAM_STATUS_MASK) == CAM_REQ_CMP)
1282                                 error = 0;
1283                         else if (error_routine != NULL) {
1284                                 ccb->ccb_h.cbfcnp = cam_periph_done;
1285                                 error = (*error_routine)(ccb, camflags, sense_flags);
1286                         } else
1287                                 error = 0;
1288                 } while (error == ERESTART);
1289         }
1290
1291         if ((ccb->ccb_h.status & CAM_DEV_QFRZN) != 0) {
1292                 cam_release_devq(ccb->ccb_h.path,
1293                                  /* relsim_flags */0,
1294                                  /* openings */0,
1295                                  /* timeout */0,
1296                                  /* getcount_only */ FALSE);
1297                 ccb->ccb_h.status &= ~CAM_DEV_QFRZN;
1298         }
1299
1300         if (ds != NULL) {
1301                 uint32_t bytes;
1302                 devstat_tag_type tag;
1303                 bool valid = true;
1304
1305                 if (ccb->ccb_h.func_code == XPT_SCSI_IO) {
1306                         bytes = ccb->csio.dxfer_len - ccb->csio.resid;
1307                         tag = (devstat_tag_type)(ccb->csio.tag_action & 0x3);
1308                 } else if (ccb->ccb_h.func_code == XPT_ATA_IO) {
1309                         bytes = ccb->ataio.dxfer_len - ccb->ataio.resid;
1310                         tag = (devstat_tag_type)0;
1311                 } else if (ccb->ccb_h.func_code == XPT_NVME_IO) {
1312                         bytes = ccb->nvmeio.dxfer_len; /* NB: resid no possible */
1313                         tag = (devstat_tag_type)0;
1314                 } else {
1315                         valid = false;
1316                 }
1317                 if (valid)
1318                         devstat_end_transaction(ds, bytes, tag,
1319                             ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE) ?
1320                             DEVSTAT_NO_DATA : (ccb->ccb_h.flags & CAM_DIR_OUT) ?
1321                             DEVSTAT_WRITE : DEVSTAT_READ, NULL, starttime);
1322         }
1323
1324         return(error);
1325 }
1326
1327 void
1328 cam_freeze_devq(struct cam_path *path)
1329 {
1330         struct ccb_hdr ccb_h;
1331
1332         CAM_DEBUG(path, CAM_DEBUG_TRACE, ("cam_freeze_devq\n"));
1333         memset(&ccb_h, 0, sizeof(ccb_h));
1334         xpt_setup_ccb(&ccb_h, path, /*priority*/1);
1335         ccb_h.func_code = XPT_NOOP;
1336         ccb_h.flags = CAM_DEV_QFREEZE;
1337         xpt_action((union ccb *)&ccb_h);
1338 }
1339
1340 u_int32_t
1341 cam_release_devq(struct cam_path *path, u_int32_t relsim_flags,
1342                  u_int32_t openings, u_int32_t arg,
1343                  int getcount_only)
1344 {
1345         struct ccb_relsim crs;
1346
1347         CAM_DEBUG(path, CAM_DEBUG_TRACE, ("cam_release_devq(%u, %u, %u, %d)\n",
1348             relsim_flags, openings, arg, getcount_only));
1349         memset(&crs, 0, sizeof(crs));
1350         xpt_setup_ccb(&crs.ccb_h, path, CAM_PRIORITY_NORMAL);
1351         crs.ccb_h.func_code = XPT_REL_SIMQ;
1352         crs.ccb_h.flags = getcount_only ? CAM_DEV_QFREEZE : 0;
1353         crs.release_flags = relsim_flags;
1354         crs.openings = openings;
1355         crs.release_timeout = arg;
1356         xpt_action((union ccb *)&crs);
1357         return (crs.qfrozen_cnt);
1358 }
1359
1360 #define saved_ccb_ptr ppriv_ptr0
1361 static void
1362 camperiphdone(struct cam_periph *periph, union ccb *done_ccb)
1363 {
1364         union ccb      *saved_ccb;
1365         cam_status      status;
1366         struct scsi_start_stop_unit *scsi_cmd;
1367         int             error = 0, error_code, sense_key, asc, ascq;
1368         u_int16_t       done_flags;
1369
1370         scsi_cmd = (struct scsi_start_stop_unit *)
1371             &done_ccb->csio.cdb_io.cdb_bytes;
1372         status = done_ccb->ccb_h.status;
1373
1374         if ((status & CAM_STATUS_MASK) != CAM_REQ_CMP) {
1375                 if (scsi_extract_sense_ccb(done_ccb,
1376                     &error_code, &sense_key, &asc, &ascq)) {
1377                         /*
1378                          * If the error is "invalid field in CDB",
1379                          * and the load/eject flag is set, turn the
1380                          * flag off and try again.  This is just in
1381                          * case the drive in question barfs on the
1382                          * load eject flag.  The CAM code should set
1383                          * the load/eject flag by default for
1384                          * removable media.
1385                          */
1386                         if ((scsi_cmd->opcode == START_STOP_UNIT) &&
1387                             ((scsi_cmd->how & SSS_LOEJ) != 0) &&
1388                              (asc == 0x24) && (ascq == 0x00)) {
1389                                 scsi_cmd->how &= ~SSS_LOEJ;
1390                                 if (status & CAM_DEV_QFRZN) {
1391                                         cam_release_devq(done_ccb->ccb_h.path,
1392                                             0, 0, 0, 0);
1393                                         done_ccb->ccb_h.status &=
1394                                             ~CAM_DEV_QFRZN;
1395                                 }
1396                                 xpt_action(done_ccb);
1397                                 goto out;
1398                         }
1399                 }
1400                 error = cam_periph_error(done_ccb, 0,
1401                     SF_RETRY_UA | SF_NO_PRINT);
1402                 if (error == ERESTART)
1403                         goto out;
1404                 if (done_ccb->ccb_h.status & CAM_DEV_QFRZN) {
1405                         cam_release_devq(done_ccb->ccb_h.path, 0, 0, 0, 0);
1406                         done_ccb->ccb_h.status &= ~CAM_DEV_QFRZN;
1407                 }
1408         } else {
1409                 /*
1410                  * If we have successfully taken a device from the not
1411                  * ready to ready state, re-scan the device and re-get
1412                  * the inquiry information.  Many devices (mostly disks)
1413                  * don't properly report their inquiry information unless
1414                  * they are spun up.
1415                  */
1416                 if (scsi_cmd->opcode == START_STOP_UNIT)
1417                         xpt_async(AC_INQ_CHANGED, done_ccb->ccb_h.path, NULL);
1418         }
1419
1420         /* If we tried long wait and still failed, remember that. */
1421         if ((periph->flags & CAM_PERIPH_RECOVERY_WAIT) &&
1422             (done_ccb->csio.cdb_io.cdb_bytes[0] == TEST_UNIT_READY)) {
1423                 periph->flags &= ~CAM_PERIPH_RECOVERY_WAIT;
1424                 if (error != 0 && done_ccb->ccb_h.retry_count == 0)
1425                         periph->flags |= CAM_PERIPH_RECOVERY_WAIT_FAILED;
1426         }
1427
1428         /*
1429          * After recovery action(s) completed, return to the original CCB.
1430          * If the recovery CCB has failed, considering its own possible
1431          * retries and recovery, assume we are back in state where we have
1432          * been originally, but without recovery hopes left.  In such case,
1433          * after the final attempt below, we cancel any further retries,
1434          * blocking by that also any new recovery attempts for this CCB,
1435          * and the result will be the final one returned to the CCB owher.
1436          */
1437
1438         /*
1439          * Copy the CCB back, preserving the alloc_flags field.  Things
1440          * will crash horribly if the CCBs are not of the same size.
1441          */
1442         saved_ccb = (union ccb *)done_ccb->ccb_h.saved_ccb_ptr;
1443         KASSERT(saved_ccb->ccb_h.func_code == XPT_SCSI_IO,
1444             ("%s: saved_ccb func_code %#x != XPT_SCSI_IO",
1445              __func__, saved_ccb->ccb_h.func_code));
1446         KASSERT(done_ccb->ccb_h.func_code == XPT_SCSI_IO,
1447             ("%s: done_ccb func_code %#x != XPT_SCSI_IO",
1448              __func__, done_ccb->ccb_h.func_code));
1449         done_flags = done_ccb->ccb_h.alloc_flags;
1450         bcopy(saved_ccb, done_ccb, sizeof(struct ccb_scsiio));
1451         done_ccb->ccb_h.alloc_flags = done_flags;
1452         xpt_free_ccb(saved_ccb);
1453         if (done_ccb->ccb_h.cbfcnp != camperiphdone)
1454                 periph->flags &= ~CAM_PERIPH_RECOVERY_INPROG;
1455         if (error != 0)
1456                 done_ccb->ccb_h.retry_count = 0;
1457         xpt_action(done_ccb);
1458
1459 out:
1460         /* Drop freeze taken due to CAM_DEV_QFREEZE flag set. */
1461         cam_release_devq(done_ccb->ccb_h.path, 0, 0, 0, 0);
1462 }
1463
1464 /*
1465  * Generic Async Event handler.  Peripheral drivers usually
1466  * filter out the events that require personal attention,
1467  * and leave the rest to this function.
1468  */
1469 void
1470 cam_periph_async(struct cam_periph *periph, u_int32_t code,
1471                  struct cam_path *path, void *arg)
1472 {
1473         switch (code) {
1474         case AC_LOST_DEVICE:
1475                 cam_periph_invalidate(periph);
1476                 break; 
1477         default:
1478                 break;
1479         }
1480 }
1481
1482 void
1483 cam_periph_bus_settle(struct cam_periph *periph, u_int bus_settle)
1484 {
1485         struct ccb_getdevstats cgds;
1486
1487         memset(&cgds, 0, sizeof(cgds));
1488         xpt_setup_ccb(&cgds.ccb_h, periph->path, CAM_PRIORITY_NORMAL);
1489         cgds.ccb_h.func_code = XPT_GDEV_STATS;
1490         xpt_action((union ccb *)&cgds);
1491         cam_periph_freeze_after_event(periph, &cgds.last_reset, bus_settle);
1492 }
1493
1494 void
1495 cam_periph_freeze_after_event(struct cam_periph *periph,
1496                               struct timeval* event_time, u_int duration_ms)
1497 {
1498         struct timeval delta;
1499         struct timeval duration_tv;
1500
1501         if (!timevalisset(event_time))
1502                 return;
1503
1504         microtime(&delta);
1505         timevalsub(&delta, event_time);
1506         duration_tv.tv_sec = duration_ms / 1000;
1507         duration_tv.tv_usec = (duration_ms % 1000) * 1000;
1508         if (timevalcmp(&delta, &duration_tv, <)) {
1509                 timevalsub(&duration_tv, &delta);
1510
1511                 duration_ms = duration_tv.tv_sec * 1000;
1512                 duration_ms += duration_tv.tv_usec / 1000;
1513                 cam_freeze_devq(periph->path); 
1514                 cam_release_devq(periph->path,
1515                                 RELSIM_RELEASE_AFTER_TIMEOUT,
1516                                 /*reduction*/0,
1517                                 /*timeout*/duration_ms,
1518                                 /*getcount_only*/0);
1519         }
1520
1521 }
1522
1523 static int
1524 camperiphscsistatuserror(union ccb *ccb, union ccb **orig_ccb,
1525     cam_flags camflags, u_int32_t sense_flags,
1526     int *openings, u_int32_t *relsim_flags,
1527     u_int32_t *timeout, u_int32_t *action, const char **action_string)
1528 {
1529         struct cam_periph *periph;
1530         int error;
1531
1532         switch (ccb->csio.scsi_status) {
1533         case SCSI_STATUS_OK:
1534         case SCSI_STATUS_COND_MET:
1535         case SCSI_STATUS_INTERMED:
1536         case SCSI_STATUS_INTERMED_COND_MET:
1537                 error = 0;
1538                 break;
1539         case SCSI_STATUS_CMD_TERMINATED:
1540         case SCSI_STATUS_CHECK_COND:
1541                 error = camperiphscsisenseerror(ccb, orig_ccb,
1542                                                 camflags,
1543                                                 sense_flags,
1544                                                 openings,
1545                                                 relsim_flags,
1546                                                 timeout,
1547                                                 action,
1548                                                 action_string);
1549                 break;
1550         case SCSI_STATUS_QUEUE_FULL:
1551         {
1552                 /* no decrement */
1553                 struct ccb_getdevstats cgds;
1554
1555                 /*
1556                  * First off, find out what the current
1557                  * transaction counts are.
1558                  */
1559                 memset(&cgds, 0, sizeof(cgds));
1560                 xpt_setup_ccb(&cgds.ccb_h,
1561                               ccb->ccb_h.path,
1562                               CAM_PRIORITY_NORMAL);
1563                 cgds.ccb_h.func_code = XPT_GDEV_STATS;
1564                 xpt_action((union ccb *)&cgds);
1565
1566                 /*
1567                  * If we were the only transaction active, treat
1568                  * the QUEUE FULL as if it were a BUSY condition.
1569                  */
1570                 if (cgds.dev_active != 0) {
1571                         int total_openings;
1572
1573                         /*
1574                          * Reduce the number of openings to
1575                          * be 1 less than the amount it took
1576                          * to get a queue full bounded by the
1577                          * minimum allowed tag count for this
1578                          * device.
1579                          */
1580                         total_openings = cgds.dev_active + cgds.dev_openings;
1581                         *openings = cgds.dev_active;
1582                         if (*openings < cgds.mintags)
1583                                 *openings = cgds.mintags;
1584                         if (*openings < total_openings)
1585                                 *relsim_flags = RELSIM_ADJUST_OPENINGS;
1586                         else {
1587                                 /*
1588                                  * Some devices report queue full for
1589                                  * temporary resource shortages.  For
1590                                  * this reason, we allow a minimum
1591                                  * tag count to be entered via a
1592                                  * quirk entry to prevent the queue
1593                                  * count on these devices from falling
1594                                  * to a pessimisticly low value.  We
1595                                  * still wait for the next successful
1596                                  * completion, however, before queueing
1597                                  * more transactions to the device.
1598                                  */
1599                                 *relsim_flags = RELSIM_RELEASE_AFTER_CMDCMPLT;
1600                         }
1601                         *timeout = 0;
1602                         error = ERESTART;
1603                         *action &= ~SSQ_PRINT_SENSE;
1604                         break;
1605                 }
1606                 /* FALLTHROUGH */
1607         }
1608         case SCSI_STATUS_BUSY:
1609                 /*
1610                  * Restart the queue after either another
1611                  * command completes or a 1 second timeout.
1612                  */
1613                 periph = xpt_path_periph(ccb->ccb_h.path);
1614                 if (periph->flags & CAM_PERIPH_INVALID) {
1615                         error = EIO;
1616                         *action_string = "Periph was invalidated";
1617                 } else if ((sense_flags & SF_RETRY_BUSY) != 0 ||
1618                     ccb->ccb_h.retry_count > 0) {
1619                         if ((sense_flags & SF_RETRY_BUSY) == 0)
1620                                 ccb->ccb_h.retry_count--;
1621                         error = ERESTART;
1622                         *relsim_flags = RELSIM_RELEASE_AFTER_TIMEOUT
1623                                       | RELSIM_RELEASE_AFTER_CMDCMPLT;
1624                         *timeout = 1000;
1625                 } else {
1626                         error = EIO;
1627                         *action_string = "Retries exhausted";
1628                 }
1629                 break;
1630         case SCSI_STATUS_RESERV_CONFLICT:
1631         default:
1632                 error = EIO;
1633                 break;
1634         }
1635         return (error);
1636 }
1637
1638 static int
1639 camperiphscsisenseerror(union ccb *ccb, union ccb **orig,
1640     cam_flags camflags, u_int32_t sense_flags,
1641     int *openings, u_int32_t *relsim_flags,
1642     u_int32_t *timeout, u_int32_t *action, const char **action_string)
1643 {
1644         struct cam_periph *periph;
1645         union ccb *orig_ccb = ccb;
1646         int error, recoveryccb;
1647         u_int16_t flags;
1648
1649 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
1650         if (ccb->ccb_h.func_code == XPT_SCSI_IO && ccb->csio.bio != NULL)
1651                 biotrack(ccb->csio.bio, __func__);
1652 #endif
1653
1654         periph = xpt_path_periph(ccb->ccb_h.path);
1655         recoveryccb = (ccb->ccb_h.cbfcnp == camperiphdone);
1656         if ((periph->flags & CAM_PERIPH_RECOVERY_INPROG) && !recoveryccb) {
1657                 /*
1658                  * If error recovery is already in progress, don't attempt
1659                  * to process this error, but requeue it unconditionally
1660                  * and attempt to process it once error recovery has
1661                  * completed.  This failed command is probably related to
1662                  * the error that caused the currently active error recovery
1663                  * action so our  current recovery efforts should also
1664                  * address this command.  Be aware that the error recovery
1665                  * code assumes that only one recovery action is in progress
1666                  * on a particular peripheral instance at any given time
1667                  * (e.g. only one saved CCB for error recovery) so it is
1668                  * imperitive that we don't violate this assumption.
1669                  */
1670                 error = ERESTART;
1671                 *action &= ~SSQ_PRINT_SENSE;
1672         } else {
1673                 scsi_sense_action err_action;
1674                 struct ccb_getdev cgd;
1675
1676                 /*
1677                  * Grab the inquiry data for this device.
1678                  */
1679                 memset(&cgd, 0, sizeof(cgd));
1680                 xpt_setup_ccb(&cgd.ccb_h, ccb->ccb_h.path, CAM_PRIORITY_NORMAL);
1681                 cgd.ccb_h.func_code = XPT_GDEV_TYPE;
1682                 xpt_action((union ccb *)&cgd);
1683
1684                 err_action = scsi_error_action(&ccb->csio, &cgd.inq_data,
1685                     sense_flags);
1686                 error = err_action & SS_ERRMASK;
1687
1688                 /*
1689                  * Do not autostart sequential access devices
1690                  * to avoid unexpected tape loading.
1691                  */
1692                 if ((err_action & SS_MASK) == SS_START &&
1693                     SID_TYPE(&cgd.inq_data) == T_SEQUENTIAL) {
1694                         *action_string = "Will not autostart a "
1695                             "sequential access device";
1696                         goto sense_error_done;
1697                 }
1698
1699                 /*
1700                  * Avoid recovery recursion if recovery action is the same.
1701                  */
1702                 if ((err_action & SS_MASK) >= SS_START && recoveryccb) {
1703                         if (((err_action & SS_MASK) == SS_START &&
1704                              ccb->csio.cdb_io.cdb_bytes[0] == START_STOP_UNIT) ||
1705                             ((err_action & SS_MASK) == SS_TUR &&
1706                              (ccb->csio.cdb_io.cdb_bytes[0] == TEST_UNIT_READY))) {
1707                                 err_action = SS_RETRY|SSQ_DECREMENT_COUNT|EIO;
1708                                 *relsim_flags = RELSIM_RELEASE_AFTER_TIMEOUT;
1709                                 *timeout = 500;
1710                         }
1711                 }
1712
1713                 /*
1714                  * If the recovery action will consume a retry,
1715                  * make sure we actually have retries available.
1716                  */
1717                 if ((err_action & SSQ_DECREMENT_COUNT) != 0) {
1718                         if (ccb->ccb_h.retry_count > 0 &&
1719                             (periph->flags & CAM_PERIPH_INVALID) == 0)
1720                                 ccb->ccb_h.retry_count--;
1721                         else {
1722                                 *action_string = "Retries exhausted";
1723                                 goto sense_error_done;
1724                         }
1725                 }
1726
1727                 if ((err_action & SS_MASK) >= SS_START) {
1728                         /*
1729                          * Do common portions of commands that
1730                          * use recovery CCBs.
1731                          */
1732                         orig_ccb = xpt_alloc_ccb_nowait();
1733                         if (orig_ccb == NULL) {
1734                                 *action_string = "Can't allocate recovery CCB";
1735                                 goto sense_error_done;
1736                         }
1737                         /*
1738                          * Clear freeze flag for original request here, as
1739                          * this freeze will be dropped as part of ERESTART.
1740                          */
1741                         ccb->ccb_h.status &= ~CAM_DEV_QFRZN;
1742
1743                         KASSERT(ccb->ccb_h.func_code == XPT_SCSI_IO,
1744                             ("%s: ccb func_code %#x != XPT_SCSI_IO",
1745                              __func__, ccb->ccb_h.func_code));
1746                         flags = orig_ccb->ccb_h.alloc_flags;
1747                         bcopy(ccb, orig_ccb, sizeof(struct ccb_scsiio));
1748                         orig_ccb->ccb_h.alloc_flags = flags;
1749                 }
1750
1751                 switch (err_action & SS_MASK) {
1752                 case SS_NOP:
1753                         *action_string = "No recovery action needed";
1754                         error = 0;
1755                         break;
1756                 case SS_RETRY:
1757                         *action_string = "Retrying command (per sense data)";
1758                         error = ERESTART;
1759                         break;
1760                 case SS_FAIL:
1761                         *action_string = "Unretryable error";
1762                         break;
1763                 case SS_START:
1764                 {
1765                         int le;
1766
1767                         /*
1768                          * Send a start unit command to the device, and
1769                          * then retry the command.
1770                          */
1771                         *action_string = "Attempting to start unit";
1772                         periph->flags |= CAM_PERIPH_RECOVERY_INPROG;
1773
1774                         /*
1775                          * Check for removable media and set
1776                          * load/eject flag appropriately.
1777                          */
1778                         if (SID_IS_REMOVABLE(&cgd.inq_data))
1779                                 le = TRUE;
1780                         else
1781                                 le = FALSE;
1782
1783                         scsi_start_stop(&ccb->csio,
1784                                         /*retries*/1,
1785                                         camperiphdone,
1786                                         MSG_SIMPLE_Q_TAG,
1787                                         /*start*/TRUE,
1788                                         /*load/eject*/le,
1789                                         /*immediate*/FALSE,
1790                                         SSD_FULL_SIZE,
1791                                         /*timeout*/50000);
1792                         break;
1793                 }
1794                 case SS_TUR:
1795                 {
1796                         /*
1797                          * Send a Test Unit Ready to the device.
1798                          * If the 'many' flag is set, we send 120
1799                          * test unit ready commands, one every half 
1800                          * second.  Otherwise, we just send one TUR.
1801                          * We only want to do this if the retry 
1802                          * count has not been exhausted.
1803                          */
1804                         int retries;
1805
1806                         if ((err_action & SSQ_MANY) != 0 && (periph->flags &
1807                              CAM_PERIPH_RECOVERY_WAIT_FAILED) == 0) {
1808                                 periph->flags |= CAM_PERIPH_RECOVERY_WAIT;
1809                                 *action_string = "Polling device for readiness";
1810                                 retries = 120;
1811                         } else {
1812                                 *action_string = "Testing device for readiness";
1813                                 retries = 1;
1814                         }
1815                         periph->flags |= CAM_PERIPH_RECOVERY_INPROG;
1816                         scsi_test_unit_ready(&ccb->csio,
1817                                              retries,
1818                                              camperiphdone,
1819                                              MSG_SIMPLE_Q_TAG,
1820                                              SSD_FULL_SIZE,
1821                                              /*timeout*/5000);
1822
1823                         /*
1824                          * Accomplish our 500ms delay by deferring
1825                          * the release of our device queue appropriately.
1826                          */
1827                         *relsim_flags = RELSIM_RELEASE_AFTER_TIMEOUT;
1828                         *timeout = 500;
1829                         break;
1830                 }
1831                 default:
1832                         panic("Unhandled error action %x", err_action);
1833                 }
1834                 
1835                 if ((err_action & SS_MASK) >= SS_START) {
1836                         /*
1837                          * Drop the priority, so that the recovery
1838                          * CCB is the first to execute.  Freeze the queue
1839                          * after this command is sent so that we can
1840                          * restore the old csio and have it queued in
1841                          * the proper order before we release normal 
1842                          * transactions to the device.
1843                          */
1844                         ccb->ccb_h.pinfo.priority--;
1845                         ccb->ccb_h.flags |= CAM_DEV_QFREEZE;
1846                         ccb->ccb_h.saved_ccb_ptr = orig_ccb;
1847                         error = ERESTART;
1848                         *orig = orig_ccb;
1849                 }
1850
1851 sense_error_done:
1852                 *action = err_action;
1853         }
1854         return (error);
1855 }
1856
1857 /*
1858  * Generic error handler.  Peripheral drivers usually filter
1859  * out the errors that they handle in a unique manner, then
1860  * call this function.
1861  */
1862 int
1863 cam_periph_error(union ccb *ccb, cam_flags camflags,
1864                  u_int32_t sense_flags)
1865 {
1866         struct cam_path *newpath;
1867         union ccb  *orig_ccb, *scan_ccb;
1868         struct cam_periph *periph;
1869         const char *action_string;
1870         cam_status  status;
1871         int         frozen, error, openings, devctl_err;
1872         u_int32_t   action, relsim_flags, timeout;
1873
1874         action = SSQ_PRINT_SENSE;
1875         periph = xpt_path_periph(ccb->ccb_h.path);
1876         action_string = NULL;
1877         status = ccb->ccb_h.status;
1878         frozen = (status & CAM_DEV_QFRZN) != 0;
1879         status &= CAM_STATUS_MASK;
1880         devctl_err = openings = relsim_flags = timeout = 0;
1881         orig_ccb = ccb;
1882
1883         /* Filter the errors that should be reported via devctl */
1884         switch (ccb->ccb_h.status & CAM_STATUS_MASK) {
1885         case CAM_CMD_TIMEOUT:
1886         case CAM_REQ_ABORTED:
1887         case CAM_REQ_CMP_ERR:
1888         case CAM_REQ_TERMIO:
1889         case CAM_UNREC_HBA_ERROR:
1890         case CAM_DATA_RUN_ERR:
1891         case CAM_SCSI_STATUS_ERROR:
1892         case CAM_ATA_STATUS_ERROR:
1893         case CAM_SMP_STATUS_ERROR:
1894                 devctl_err++;
1895                 break;
1896         default:
1897                 break;
1898         }
1899
1900         switch (status) {
1901         case CAM_REQ_CMP:
1902                 error = 0;
1903                 action &= ~SSQ_PRINT_SENSE;
1904                 break;
1905         case CAM_SCSI_STATUS_ERROR:
1906                 error = camperiphscsistatuserror(ccb, &orig_ccb,
1907                     camflags, sense_flags, &openings, &relsim_flags,
1908                     &timeout, &action, &action_string);
1909                 break;
1910         case CAM_AUTOSENSE_FAIL:
1911                 error = EIO;    /* we have to kill the command */
1912                 break;
1913         case CAM_UA_ABORT:
1914         case CAM_UA_TERMIO:
1915         case CAM_MSG_REJECT_REC:
1916                 /* XXX Don't know that these are correct */
1917                 error = EIO;
1918                 break;
1919         case CAM_SEL_TIMEOUT:
1920                 if ((camflags & CAM_RETRY_SELTO) != 0) {
1921                         if (ccb->ccb_h.retry_count > 0 &&
1922                             (periph->flags & CAM_PERIPH_INVALID) == 0) {
1923                                 ccb->ccb_h.retry_count--;
1924                                 error = ERESTART;
1925
1926                                 /*
1927                                  * Wait a bit to give the device
1928                                  * time to recover before we try again.
1929                                  */
1930                                 relsim_flags = RELSIM_RELEASE_AFTER_TIMEOUT;
1931                                 timeout = periph_selto_delay;
1932                                 break;
1933                         }
1934                         action_string = "Retries exhausted";
1935                 }
1936                 /* FALLTHROUGH */
1937         case CAM_DEV_NOT_THERE:
1938                 error = ENXIO;
1939                 action = SSQ_LOST;
1940                 break;
1941         case CAM_REQ_INVALID:
1942         case CAM_PATH_INVALID:
1943         case CAM_NO_HBA:
1944         case CAM_PROVIDE_FAIL:
1945         case CAM_REQ_TOO_BIG:
1946         case CAM_LUN_INVALID:
1947         case CAM_TID_INVALID:
1948         case CAM_FUNC_NOTAVAIL:
1949                 error = EINVAL;
1950                 break;
1951         case CAM_SCSI_BUS_RESET:
1952         case CAM_BDR_SENT:
1953                 /*
1954                  * Commands that repeatedly timeout and cause these
1955                  * kinds of error recovery actions, should return
1956                  * CAM_CMD_TIMEOUT, which allows us to safely assume
1957                  * that this command was an innocent bystander to
1958                  * these events and should be unconditionally
1959                  * retried.
1960                  */
1961         case CAM_REQUEUE_REQ:
1962                 /* Unconditional requeue if device is still there */
1963                 if (periph->flags & CAM_PERIPH_INVALID) {
1964                         action_string = "Periph was invalidated";
1965                         error = EIO;
1966                 } else if (sense_flags & SF_NO_RETRY) {
1967                         error = EIO;
1968                         action_string = "Retry was blocked";
1969                 } else {
1970                         error = ERESTART;
1971                         action &= ~SSQ_PRINT_SENSE;
1972                 }
1973                 break;
1974         case CAM_RESRC_UNAVAIL:
1975                 /* Wait a bit for the resource shortage to abate. */
1976                 timeout = periph_noresrc_delay;
1977                 /* FALLTHROUGH */
1978         case CAM_BUSY:
1979                 if (timeout == 0) {
1980                         /* Wait a bit for the busy condition to abate. */
1981                         timeout = periph_busy_delay;
1982                 }
1983                 relsim_flags = RELSIM_RELEASE_AFTER_TIMEOUT;
1984                 /* FALLTHROUGH */
1985         case CAM_ATA_STATUS_ERROR:
1986         case CAM_REQ_CMP_ERR:
1987         case CAM_CMD_TIMEOUT:
1988         case CAM_UNEXP_BUSFREE:
1989         case CAM_UNCOR_PARITY:
1990         case CAM_DATA_RUN_ERR:
1991         default:
1992                 if (periph->flags & CAM_PERIPH_INVALID) {
1993                         error = EIO;
1994                         action_string = "Periph was invalidated";
1995                 } else if (ccb->ccb_h.retry_count == 0) {
1996                         error = EIO;
1997                         action_string = "Retries exhausted";
1998                 } else if (sense_flags & SF_NO_RETRY) {
1999                         error = EIO;
2000                         action_string = "Retry was blocked";
2001                 } else {
2002                         ccb->ccb_h.retry_count--;
2003                         error = ERESTART;
2004                 }
2005                 break;
2006         }
2007
2008         if ((sense_flags & SF_PRINT_ALWAYS) ||
2009             CAM_DEBUGGED(ccb->ccb_h.path, CAM_DEBUG_INFO))
2010                 action |= SSQ_PRINT_SENSE;
2011         else if (sense_flags & SF_NO_PRINT)
2012                 action &= ~SSQ_PRINT_SENSE;
2013         if ((action & SSQ_PRINT_SENSE) != 0)
2014                 cam_error_print(orig_ccb, CAM_ESF_ALL, CAM_EPF_ALL);
2015         if (error != 0 && (action & SSQ_PRINT_SENSE) != 0) {
2016                 if (error != ERESTART) {
2017                         if (action_string == NULL)
2018                                 action_string = "Unretryable error";
2019                         xpt_print(ccb->ccb_h.path, "Error %d, %s\n",
2020                             error, action_string);
2021                 } else if (action_string != NULL)
2022                         xpt_print(ccb->ccb_h.path, "%s\n", action_string);
2023                 else {
2024                         xpt_print(ccb->ccb_h.path,
2025                             "Retrying command, %d more tries remain\n",
2026                             ccb->ccb_h.retry_count);
2027                 }
2028         }
2029
2030         if (devctl_err && (error != 0 || (action & SSQ_PRINT_SENSE) != 0))
2031                 cam_periph_devctl_notify(orig_ccb);
2032
2033         if ((action & SSQ_LOST) != 0) {
2034                 lun_id_t lun_id;
2035
2036                 /*
2037                  * For a selection timeout, we consider all of the LUNs on
2038                  * the target to be gone.  If the status is CAM_DEV_NOT_THERE,
2039                  * then we only get rid of the device(s) specified by the
2040                  * path in the original CCB.
2041                  */
2042                 if (status == CAM_SEL_TIMEOUT)
2043                         lun_id = CAM_LUN_WILDCARD;
2044                 else
2045                         lun_id = xpt_path_lun_id(ccb->ccb_h.path);
2046
2047                 /* Should we do more if we can't create the path?? */
2048                 if (xpt_create_path(&newpath, periph,
2049                                     xpt_path_path_id(ccb->ccb_h.path),
2050                                     xpt_path_target_id(ccb->ccb_h.path),
2051                                     lun_id) == CAM_REQ_CMP) {
2052                         /*
2053                          * Let peripheral drivers know that this
2054                          * device has gone away.
2055                          */
2056                         xpt_async(AC_LOST_DEVICE, newpath, NULL);
2057                         xpt_free_path(newpath);
2058                 }
2059         }
2060
2061         /* Broadcast UNIT ATTENTIONs to all periphs. */
2062         if ((action & SSQ_UA) != 0)
2063                 xpt_async(AC_UNIT_ATTENTION, orig_ccb->ccb_h.path, orig_ccb);
2064
2065         /* Rescan target on "Reported LUNs data has changed" */
2066         if ((action & SSQ_RESCAN) != 0) {
2067                 if (xpt_create_path(&newpath, NULL,
2068                                     xpt_path_path_id(ccb->ccb_h.path),
2069                                     xpt_path_target_id(ccb->ccb_h.path),
2070                                     CAM_LUN_WILDCARD) == CAM_REQ_CMP) {
2071                         scan_ccb = xpt_alloc_ccb_nowait();
2072                         if (scan_ccb != NULL) {
2073                                 scan_ccb->ccb_h.path = newpath;
2074                                 scan_ccb->ccb_h.func_code = XPT_SCAN_TGT;
2075                                 scan_ccb->crcn.flags = 0;
2076                                 xpt_rescan(scan_ccb);
2077                         } else {
2078                                 xpt_print(newpath,
2079                                     "Can't allocate CCB to rescan target\n");
2080                                 xpt_free_path(newpath);
2081                         }
2082                 }
2083         }
2084
2085         /* Attempt a retry */
2086         if (error == ERESTART || error == 0) {
2087                 if (frozen != 0)
2088                         ccb->ccb_h.status &= ~CAM_DEV_QFRZN;
2089                 if (error == ERESTART)
2090                         xpt_action(ccb);
2091                 if (frozen != 0)
2092                         cam_release_devq(ccb->ccb_h.path,
2093                                          relsim_flags,
2094                                          openings,
2095                                          timeout,
2096                                          /*getcount_only*/0);
2097         }
2098
2099         return (error);
2100 }
2101
2102 #define CAM_PERIPH_DEVD_MSG_SIZE        256
2103
2104 static void
2105 cam_periph_devctl_notify(union ccb *ccb)
2106 {
2107         struct cam_periph *periph;
2108         struct ccb_getdev *cgd;
2109         struct sbuf sb;
2110         int serr, sk, asc, ascq;
2111         char *sbmsg, *type;
2112
2113         sbmsg = malloc(CAM_PERIPH_DEVD_MSG_SIZE, M_CAMPERIPH, M_NOWAIT);
2114         if (sbmsg == NULL)
2115                 return;
2116
2117         sbuf_new(&sb, sbmsg, CAM_PERIPH_DEVD_MSG_SIZE, SBUF_FIXEDLEN);
2118
2119         periph = xpt_path_periph(ccb->ccb_h.path);
2120         sbuf_printf(&sb, "device=%s%d ", periph->periph_name,
2121             periph->unit_number);
2122
2123         sbuf_printf(&sb, "serial=\"");
2124         if ((cgd = (struct ccb_getdev *)xpt_alloc_ccb_nowait()) != NULL) {
2125                 xpt_setup_ccb(&cgd->ccb_h, ccb->ccb_h.path,
2126                     CAM_PRIORITY_NORMAL);
2127                 cgd->ccb_h.func_code = XPT_GDEV_TYPE;
2128                 xpt_action((union ccb *)cgd);
2129
2130                 if (cgd->ccb_h.status == CAM_REQ_CMP)
2131                         sbuf_bcat(&sb, cgd->serial_num, cgd->serial_num_len);
2132                 xpt_free_ccb((union ccb *)cgd);
2133         }
2134         sbuf_printf(&sb, "\" ");
2135         sbuf_printf(&sb, "cam_status=\"0x%x\" ", ccb->ccb_h.status);
2136
2137         switch (ccb->ccb_h.status & CAM_STATUS_MASK) {
2138         case CAM_CMD_TIMEOUT:
2139                 sbuf_printf(&sb, "timeout=%d ", ccb->ccb_h.timeout);
2140                 type = "timeout";
2141                 break;
2142         case CAM_SCSI_STATUS_ERROR:
2143                 sbuf_printf(&sb, "scsi_status=%d ", ccb->csio.scsi_status);
2144                 if (scsi_extract_sense_ccb(ccb, &serr, &sk, &asc, &ascq))
2145                         sbuf_printf(&sb, "scsi_sense=\"%02x %02x %02x %02x\" ",
2146                             serr, sk, asc, ascq);
2147                 type = "error";
2148                 break;
2149         case CAM_ATA_STATUS_ERROR:
2150                 sbuf_printf(&sb, "RES=\"");
2151                 ata_res_sbuf(&ccb->ataio.res, &sb);
2152                 sbuf_printf(&sb, "\" ");
2153                 type = "error";
2154                 break;
2155         default:
2156                 type = "error";
2157                 break;
2158         }
2159
2160         if (ccb->ccb_h.func_code == XPT_SCSI_IO) {
2161                 sbuf_printf(&sb, "CDB=\"");
2162                 scsi_cdb_sbuf(scsiio_cdb_ptr(&ccb->csio), &sb);
2163                 sbuf_printf(&sb, "\" ");
2164         } else if (ccb->ccb_h.func_code == XPT_ATA_IO) {
2165                 sbuf_printf(&sb, "ACB=\"");
2166                 ata_cmd_sbuf(&ccb->ataio.cmd, &sb);
2167                 sbuf_printf(&sb, "\" ");
2168         }
2169
2170         if (sbuf_finish(&sb) == 0)
2171                 devctl_notify("CAM", "periph", type, sbuf_data(&sb));
2172         sbuf_delete(&sb);
2173         free(sbmsg, M_CAMPERIPH);
2174 }
2175
2176 /*
2177  * Sysctl to force an invalidation of the drive right now. Can be
2178  * called with CTLFLAG_MPSAFE since we take periph lock.
2179  */
2180 int
2181 cam_periph_invalidate_sysctl(SYSCTL_HANDLER_ARGS)
2182 {
2183         struct cam_periph *periph;
2184         int error, value;
2185
2186         periph = arg1;
2187         value = 0;
2188         error = sysctl_handle_int(oidp, &value, 0, req);
2189         if (error != 0 || req->newptr == NULL || value != 1)
2190                 return (error);
2191
2192         cam_periph_lock(periph);
2193         cam_periph_invalidate(periph);
2194         cam_periph_unlock(periph);
2195
2196         return (0);
2197 }