xref: /freebsd/sys/cam/cam_periph.c (revision c807777a43ef2b59786fa8a1a35c1f154fd069e5)
1 /*
2  * Common functions for CAM "type" (peripheral) drivers.
3  *
4  * Copyright (c) 1997, 1998 Justin T. Gibbs.
5  * Copyright (c) 1997, 1998, 1999 Kenneth D. Merry.
6  * All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions, and the following disclaimer,
13  *    without modification, immediately at the beginning of the file.
14  * 2. The name of the author may not be used to endorse or promote products
15  *    derived from this software without specific prior written permission.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR
21  * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  *
29  * $FreeBSD$
30  */
31 
32 #include <sys/param.h>
33 #include <sys/systm.h>
34 #include <sys/types.h>
35 #include <sys/malloc.h>
36 #include <sys/linker_set.h>
37 #include <sys/buf.h>
38 #include <sys/proc.h>
39 #include <sys/devicestat.h>
40 #include <vm/vm.h>
41 #include <vm/vm_extern.h>
42 
43 #include <cam/cam.h>
44 #include <cam/cam_conf.h>
45 #include <cam/cam_ccb.h>
46 #include <cam/cam_xpt_periph.h>
47 #include <cam/cam_periph.h>
48 #include <cam/cam_debug.h>
49 
50 #include <cam/scsi/scsi_all.h>
51 #include <cam/scsi/scsi_message.h>
52 #include <cam/scsi/scsi_da.h>
53 #include <cam/scsi/scsi_pass.h>
54 
55 static	u_int		camperiphnextunit(struct periph_driver *p_drv,
56 					  u_int newunit, int wired);
57 static	u_int		camperiphunit(struct periph_driver *p_drv,
58 				      path_id_t path_id_t,
59 				      target_id_t target, lun_id_t lun);
60 static	void		camperiphdone(struct cam_periph *periph,
61 					union ccb *done_ccb);
62 static  void		camperiphfree(struct cam_periph *periph);
63 
64 cam_status
65 cam_periph_alloc(periph_ctor_t *periph_ctor,
66 		 periph_oninv_t *periph_oninvalidate,
67 		 periph_dtor_t *periph_dtor, periph_start_t *periph_start,
68 		 char *name, cam_periph_type type, struct cam_path *path,
69 		 ac_callback_t *ac_callback, ac_code code, void *arg)
70 {
71 	struct		periph_driver **p_drv;
72 	struct		cam_periph *periph;
73 	struct		cam_periph *cur_periph;
74 	path_id_t	path_id;
75 	target_id_t	target_id;
76 	lun_id_t	lun_id;
77 	cam_status	status;
78 	u_int		init_level;
79 	int s;
80 
81 	init_level = 0;
82 	/*
83 	 * Handle Hot-Plug scenarios.  If there is already a peripheral
84 	 * of our type assigned to this path, we are likely waiting for
85 	 * final close on an old, invalidated, peripheral.  If this is
86 	 * the case, queue up a deferred call to the peripheral's async
87 	 * handler.  If it looks like a mistaken re-alloation, complain.
88 	 */
89 	if ((periph = cam_periph_find(path, name)) != NULL) {
90 
91 		if ((periph->flags & CAM_PERIPH_INVALID) != 0
92 		 && (periph->flags & CAM_PERIPH_NEW_DEV_FOUND) == 0) {
93 			periph->flags |= CAM_PERIPH_NEW_DEV_FOUND;
94 			periph->deferred_callback = ac_callback;
95 			periph->deferred_ac = code;
96 			return (CAM_REQ_INPROG);
97 		} else {
98 			printf("cam_periph_alloc: attempt to re-allocate "
99 			       "valid device %s%d rejected\n",
100 			       periph->periph_name, periph->unit_number);
101 		}
102 		return (CAM_REQ_INVALID);
103 	}
104 
105 	periph = (struct cam_periph *)malloc(sizeof(*periph), M_DEVBUF,
106 					     M_NOWAIT);
107 
108 	if (periph == NULL)
109 		return (CAM_RESRC_UNAVAIL);
110 
111 	init_level++;
112 
113 	for (p_drv = (struct periph_driver **)periphdriver_set.ls_items;
114 	     *p_drv != NULL; p_drv++) {
115 		if (strcmp((*p_drv)->driver_name, name) == 0)
116 			break;
117 	}
118 
119 	path_id = xpt_path_path_id(path);
120 	target_id = xpt_path_target_id(path);
121 	lun_id = xpt_path_lun_id(path);
122 	bzero(periph, sizeof(*periph));
123 	cam_init_pinfo(&periph->pinfo);
124 	periph->periph_start = periph_start;
125 	periph->periph_dtor = periph_dtor;
126 	periph->periph_oninval = periph_oninvalidate;
127 	periph->type = type;
128 	periph->periph_name = name;
129 	periph->unit_number = camperiphunit(*p_drv, path_id, target_id, lun_id);
130 	periph->immediate_priority = CAM_PRIORITY_NONE;
131 	periph->refcount = 0;
132 	SLIST_INIT(&periph->ccb_list);
133 	status = xpt_create_path(&path, periph, path_id, target_id, lun_id);
134 	if (status != CAM_REQ_CMP)
135 		goto failure;
136 
137 	periph->path = path;
138 	init_level++;
139 
140 	status = xpt_add_periph(periph);
141 
142 	if (status != CAM_REQ_CMP)
143 		goto failure;
144 
145 	s = splsoftcam();
146 	cur_periph = TAILQ_FIRST(&(*p_drv)->units);
147 	while (cur_periph != NULL
148 	    && cur_periph->unit_number < periph->unit_number)
149 		cur_periph = TAILQ_NEXT(cur_periph, unit_links);
150 
151 	if (cur_periph != NULL)
152 		TAILQ_INSERT_BEFORE(cur_periph, periph, unit_links);
153 	else {
154 		TAILQ_INSERT_TAIL(&(*p_drv)->units, periph, unit_links);
155 		(*p_drv)->generation++;
156 	}
157 
158 	splx(s);
159 
160 	init_level++;
161 
162 	status = periph_ctor(periph, arg);
163 
164 	if (status == CAM_REQ_CMP)
165 		init_level++;
166 
167 failure:
168 	switch (init_level) {
169 	case 4:
170 		/* Initialized successfully */
171 		break;
172 	case 3:
173 		s = splsoftcam();
174 		TAILQ_REMOVE(&(*p_drv)->units, periph, unit_links);
175 		splx(s);
176 		xpt_remove_periph(periph);
177 	case 2:
178 		xpt_free_path(periph->path);
179 	case 1:
180 		free(periph, M_DEVBUF);
181 	case 0:
182 		/* No cleanup to perform. */
183 		break;
184 	default:
185 		panic("cam_periph_alloc: Unkown init level");
186 	}
187 	return(status);
188 }
189 
190 /*
191  * Find a peripheral structure with the specified path, target, lun,
192  * and (optionally) type.  If the name is NULL, this function will return
193  * the first peripheral driver that matches the specified path.
194  */
195 struct cam_periph *
196 cam_periph_find(struct cam_path *path, char *name)
197 {
198 	struct periph_driver **p_drv;
199 	struct cam_periph *periph;
200 	int s;
201 
202 	for (p_drv = (struct periph_driver **)periphdriver_set.ls_items;
203 	     *p_drv != NULL; p_drv++) {
204 
205 		if (name != NULL && (strcmp((*p_drv)->driver_name, name) != 0))
206 			continue;
207 
208 		s = splsoftcam();
209 		for (periph = TAILQ_FIRST(&(*p_drv)->units); periph != NULL;
210 		     periph = TAILQ_NEXT(periph, unit_links)) {
211 			if (xpt_path_comp(periph->path, path) == 0) {
212 				splx(s);
213 				return(periph);
214 			}
215 		}
216 		splx(s);
217 		if (name != NULL)
218 			return(NULL);
219 	}
220 	return(NULL);
221 }
222 
223 cam_status
224 cam_periph_acquire(struct cam_periph *periph)
225 {
226 	int s;
227 
228 	if (periph == NULL)
229 		return(CAM_REQ_CMP_ERR);
230 
231 	s = splsoftcam();
232 	periph->refcount++;
233 	splx(s);
234 
235 	return(CAM_REQ_CMP);
236 }
237 
238 void
239 cam_periph_release(struct cam_periph *periph)
240 {
241 	int s;
242 
243 	if (periph == NULL)
244 		return;
245 
246 	s = splsoftcam();
247 	if ((--periph->refcount == 0)
248 	 && (periph->flags & CAM_PERIPH_INVALID)) {
249 		camperiphfree(periph);
250 	}
251 	splx(s);
252 
253 }
254 
255 /*
256  * Look for the next unit number that is not currently in use for this
257  * peripheral type starting at "newunit".  Also exclude unit numbers that
258  * are reserved by for future "hardwiring" unless we already know that this
259  * is a potential wired device.  Only assume that the device is "wired" the
260  * first time through the loop since after that we'll be looking at unit
261  * numbers that did not match a wiring entry.
262  */
263 static u_int
264 camperiphnextunit(struct periph_driver *p_drv, u_int newunit, int wired)
265 {
266 	struct	cam_periph *periph;
267 	struct	cam_periph_config *periph_conf;
268 	char	*periph_name;
269 	int	s;
270 
271 	s = splsoftcam();
272 	periph_name = p_drv->driver_name;
273 	for (;;newunit++) {
274 
275 		for (periph = TAILQ_FIRST(&p_drv->units);
276 		     periph != NULL && periph->unit_number != newunit;
277 		     periph = TAILQ_NEXT(periph, unit_links))
278 			;
279 
280 		if (periph != NULL && periph->unit_number == newunit) {
281 			if (wired != 0) {
282 				xpt_print_path(periph->path);
283 				printf("Duplicate Wired Device entry!\n");
284 				xpt_print_path(periph->path);
285 				printf("Second device will not be wired\n");
286 				wired = 0;
287 			}
288 			continue;
289 		}
290 
291 		for (periph_conf = cam_pinit;
292 		     wired == 0 && periph_conf->periph_name != NULL;
293 		     periph_conf++) {
294 
295 			/*
296 			 * Don't match entries like "da 4" as a wired down
297 			 * device, but do match entries like "da 4 target 5"
298 			 * or even "da 4 scbus 1".
299 			 */
300 			if (IS_SPECIFIED(periph_conf->periph_unit)
301 			 && (!strcmp(periph_name, periph_conf->periph_name))
302 			 && (IS_SPECIFIED(periph_conf->target)
303 			  || IS_SPECIFIED(periph_conf->pathid))
304 			 && (newunit == periph_conf->periph_unit))
305 				break;
306 		}
307 
308 		if (wired != 0 || periph_conf->periph_name == NULL)
309 			break;
310 	}
311 	splx(s);
312 	return (newunit);
313 }
314 
315 static u_int
316 camperiphunit(struct periph_driver *p_drv, path_id_t pathid,
317 	      target_id_t target, lun_id_t lun)
318 {
319 	struct cam_periph_config *periph_conf;
320 	u_int unit;
321 	int hit;
322 
323 	unit = 0;
324 	hit = 0;
325 
326 	for (periph_conf = cam_pinit;
327 	     periph_conf->periph_name != NULL;
328 	     periph_conf++, hit = 0) {
329 
330 		if (!strcmp(p_drv->driver_name, periph_conf->periph_name)
331 		 && IS_SPECIFIED(periph_conf->periph_unit)) {
332 
333 			if (IS_SPECIFIED(periph_conf->pathid)) {
334 
335 				if (pathid != periph_conf->pathid)
336 					continue;
337 				hit++;
338 			}
339 
340 			if (IS_SPECIFIED(periph_conf->target)) {
341 
342 				if (target != periph_conf->target)
343 					continue;
344 				hit++;
345 			}
346 
347 			if (IS_SPECIFIED(periph_conf->lun)) {
348 
349 				if (lun != periph_conf->lun)
350 					continue;
351 				hit++;
352 			}
353 
354 			if (hit != 0) {
355 				unit = periph_conf->periph_unit;
356 				break;
357 			}
358 		}
359 	}
360 
361 	/*
362 	 * Either start from 0 looking for the next unit or from
363 	 * the unit number given in the periph_conf.  This way,
364 	 * if we have wildcard matches, we don't return the same
365 	 * unit number twice.
366 	 */
367 	unit = camperiphnextunit(p_drv, unit, /*wired*/hit);
368 
369 	return (unit);
370 }
371 
372 void
373 cam_periph_invalidate(struct cam_periph *periph)
374 {
375 	int s;
376 
377 	s = splsoftcam();
378 	/*
379 	 * We only call this routine the first time a peripheral is
380 	 * invalidated.  The oninvalidate() routine is always called at
381 	 * splsoftcam().
382 	 */
383 	if (((periph->flags & CAM_PERIPH_INVALID) == 0)
384 	 && (periph->periph_oninval != NULL))
385 		periph->periph_oninval(periph);
386 
387 	periph->flags |= CAM_PERIPH_INVALID;
388 	periph->flags &= ~CAM_PERIPH_NEW_DEV_FOUND;
389 
390 	if (periph->refcount == 0)
391 		camperiphfree(periph);
392 	else if (periph->refcount < 0)
393 		printf("cam_invalidate_periph: refcount < 0!!\n");
394 	splx(s);
395 }
396 
397 static void
398 camperiphfree(struct cam_periph *periph)
399 {
400 	int s;
401 	struct periph_driver **p_drv;
402 
403 	for (p_drv = (struct periph_driver **)periphdriver_set.ls_items;
404 	     *p_drv != NULL; p_drv++) {
405 		if (strcmp((*p_drv)->driver_name, periph->periph_name) == 0)
406 			break;
407 	}
408 
409 	if (periph->periph_dtor != NULL)
410 		periph->periph_dtor(periph);
411 
412 	s = splsoftcam();
413 	TAILQ_REMOVE(&(*p_drv)->units, periph, unit_links);
414 	(*p_drv)->generation++;
415 	splx(s);
416 
417 	xpt_remove_periph(periph);
418 
419 	if (periph->flags & CAM_PERIPH_NEW_DEV_FOUND) {
420 		union ccb ccb;
421 		void *arg;
422 
423 		switch (periph->deferred_ac) {
424 		case AC_FOUND_DEVICE:
425 			ccb.ccb_h.func_code = XPT_GDEV_TYPE;
426 			xpt_setup_ccb(&ccb.ccb_h, periph->path, /*priority*/ 1);
427 			xpt_action(&ccb);
428 			arg = &ccb;
429 			break;
430 		case AC_PATH_REGISTERED:
431 			ccb.ccb_h.func_code = XPT_PATH_INQ;
432 			xpt_setup_ccb(&ccb.ccb_h, periph->path, /*priority*/ 1);
433 			xpt_action(&ccb);
434 			arg = &ccb;
435 			break;
436 		default:
437 			arg = NULL;
438 			break;
439 		}
440 		periph->deferred_callback(NULL, periph->deferred_ac,
441 					  periph->path, arg);
442 	}
443 	xpt_free_path(periph->path);
444 	free(periph, M_DEVBUF);
445 }
446 
447 /*
448  * Wait interruptibly for an exclusive lock.
449  */
450 int
451 cam_periph_lock(struct cam_periph *periph, int priority)
452 {
453 	int error;
454 
455 	while ((periph->flags & CAM_PERIPH_LOCKED) != 0) {
456 		periph->flags |= CAM_PERIPH_LOCK_WANTED;
457 		if ((error = tsleep(periph, priority, "caplck", 0)) != 0)
458 			return error;
459 	}
460 
461 	if (cam_periph_acquire(periph) != CAM_REQ_CMP)
462 		return(ENXIO);
463 
464 	periph->flags |= CAM_PERIPH_LOCKED;
465 	return 0;
466 }
467 
468 /*
469  * Unlock and wake up any waiters.
470  */
471 void
472 cam_periph_unlock(struct cam_periph *periph)
473 {
474 	periph->flags &= ~CAM_PERIPH_LOCKED;
475 	if ((periph->flags & CAM_PERIPH_LOCK_WANTED) != 0) {
476 		periph->flags &= ~CAM_PERIPH_LOCK_WANTED;
477 		wakeup(periph);
478 	}
479 
480 	cam_periph_release(periph);
481 }
482 
483 /*
484  * Map user virtual pointers into kernel virtual address space, so we can
485  * access the memory.  This won't work on physical pointers, for now it's
486  * up to the caller to check for that.  (XXX KDM -- should we do that here
487  * instead?)  This also only works for up to MAXPHYS memory.  Since we use
488  * buffers to map stuff in and out, we're limited to the buffer size.
489  */
490 int
491 cam_periph_mapmem(union ccb *ccb, struct cam_periph_map_info *mapinfo)
492 {
493 	int numbufs, i;
494 	int flags[CAM_PERIPH_MAXMAPS];
495 	u_int8_t **data_ptrs[CAM_PERIPH_MAXMAPS];
496 	u_int32_t lengths[CAM_PERIPH_MAXMAPS];
497 	u_int32_t dirs[CAM_PERIPH_MAXMAPS];
498 
499 	switch(ccb->ccb_h.func_code) {
500 	case XPT_DEV_MATCH:
501 		if (ccb->cdm.match_buf_len == 0) {
502 			printf("cam_periph_mapmem: invalid match buffer "
503 			       "length 0\n");
504 			return(EINVAL);
505 		}
506 		if (ccb->cdm.pattern_buf_len > 0) {
507 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.patterns;
508 			lengths[0] = ccb->cdm.pattern_buf_len;
509 			dirs[0] = CAM_DIR_OUT;
510 			data_ptrs[1] = (u_int8_t **)&ccb->cdm.matches;
511 			lengths[1] = ccb->cdm.match_buf_len;
512 			dirs[1] = CAM_DIR_IN;
513 			numbufs = 2;
514 		} else {
515 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.matches;
516 			lengths[0] = ccb->cdm.match_buf_len;
517 			dirs[0] = CAM_DIR_IN;
518 			numbufs = 1;
519 		}
520 		break;
521 	case XPT_SCSI_IO:
522 	case XPT_CONT_TARGET_IO:
523 		if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
524 			return(0);
525 
526 		data_ptrs[0] = &ccb->csio.data_ptr;
527 		lengths[0] = ccb->csio.dxfer_len;
528 		dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
529 		numbufs = 1;
530 		break;
531 	default:
532 		return(EINVAL);
533 		break; /* NOTREACHED */
534 	}
535 
536 	/*
537 	 * Check the transfer length and permissions first, so we don't
538 	 * have to unmap any previously mapped buffers.
539 	 */
540 	for (i = 0; i < numbufs; i++) {
541 
542 		flags[i] = 0;
543 
544 		/*
545 		 * The userland data pointer passed in may not be page
546 		 * aligned.  vmapbuf() truncates the address to a page
547 		 * boundary, so if the address isn't page aligned, we'll
548 		 * need enough space for the given transfer length, plus
549 		 * whatever extra space is necessary to make it to the page
550 		 * boundary.
551 		 */
552 		if ((lengths[i] +
553 		    (((vm_offset_t)(*data_ptrs[i])) & PAGE_MASK)) > DFLTPHYS){
554 			printf("cam_periph_mapmem: attempt to map %lu bytes, "
555 			       "which is greater than DFLTPHYS(%d)\n",
556 			       (long)(lengths[i] +
557 			       (((vm_offset_t)(*data_ptrs[i])) & PAGE_MASK)),
558 			       DFLTPHYS);
559 			return(E2BIG);
560 		}
561 
562 		if (dirs[i] & CAM_DIR_OUT) {
563 			flags[i] = B_WRITE;
564 			if (!useracc(*data_ptrs[i], lengths[i],
565 				     VM_PROT_READ)) {
566 				printf("cam_periph_mapmem: error, "
567 					"address %p, length %lu isn't "
568 					"user accessible for READ\n",
569 					(void *)*data_ptrs[i],
570 					(u_long)lengths[i]);
571 				return(EACCES);
572 			}
573 		}
574 
575 		/*
576 		 * XXX this check is really bogus, since B_WRITE currently
577 		 * is all 0's, and so it is "set" all the time.
578 		 */
579 		if (dirs[i] & CAM_DIR_IN) {
580 			flags[i] |= B_READ;
581 			if (!useracc(*data_ptrs[i], lengths[i],
582 				     VM_PROT_WRITE)) {
583 				printf("cam_periph_mapmem: error, "
584 					"address %p, length %lu isn't "
585 					"user accessible for WRITE\n",
586 					(void *)*data_ptrs[i],
587 					(u_long)lengths[i]);
588 
589 				return(EACCES);
590 			}
591 		}
592 
593 	}
594 
595 	/* this keeps the current process from getting swapped */
596 	/*
597 	 * XXX KDM should I use P_NOSWAP instead?
598 	 */
599 	PHOLD(curproc);
600 
601 	for (i = 0; i < numbufs; i++) {
602 		/*
603 		 * Get the buffer.
604 		 */
605 		mapinfo->bp[i] = getpbuf(NULL);
606 
607 		/* save the buffer's data address */
608 		mapinfo->bp[i]->b_saveaddr = mapinfo->bp[i]->b_data;
609 
610 		/* put our pointer in the data slot */
611 		mapinfo->bp[i]->b_data = *data_ptrs[i];
612 
613 		/* set the transfer length, we know it's < DFLTPHYS */
614 		mapinfo->bp[i]->b_bufsize = lengths[i];
615 
616 		/* set the flags */
617 		mapinfo->bp[i]->b_flags = flags[i] | B_PHYS;
618 
619 		/* map the buffer into kernel memory */
620 		vmapbuf(mapinfo->bp[i]);
621 
622 		/* set our pointer to the new mapped area */
623 		*data_ptrs[i] = mapinfo->bp[i]->b_data;
624 
625 		mapinfo->num_bufs_used++;
626 	}
627 
628 	return(0);
629 }
630 
631 /*
632  * Unmap memory segments mapped into kernel virtual address space by
633  * cam_periph_mapmem().
634  */
635 void
636 cam_periph_unmapmem(union ccb *ccb, struct cam_periph_map_info *mapinfo)
637 {
638 	int numbufs, i;
639 	u_int8_t **data_ptrs[CAM_PERIPH_MAXMAPS];
640 
641 	if (mapinfo->num_bufs_used <= 0) {
642 		/* allow ourselves to be swapped once again */
643 		PRELE(curproc);
644 		return;
645 	}
646 
647 	switch (ccb->ccb_h.func_code) {
648 	case XPT_DEV_MATCH:
649 		numbufs = min(mapinfo->num_bufs_used, 2);
650 
651 		if (numbufs == 1) {
652 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.matches;
653 		} else {
654 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.patterns;
655 			data_ptrs[1] = (u_int8_t **)&ccb->cdm.matches;
656 		}
657 		break;
658 	case XPT_SCSI_IO:
659 	case XPT_CONT_TARGET_IO:
660 		data_ptrs[0] = &ccb->csio.data_ptr;
661 		numbufs = min(mapinfo->num_bufs_used, 1);
662 		break;
663 	default:
664 		/* allow ourselves to be swapped once again */
665 		PRELE(curproc);
666 		return;
667 		break; /* NOTREACHED */
668 	}
669 
670 	for (i = 0; i < numbufs; i++) {
671 		/* Set the user's pointer back to the original value */
672 		*data_ptrs[i] = mapinfo->bp[i]->b_saveaddr;
673 
674 		/* unmap the buffer */
675 		vunmapbuf(mapinfo->bp[i]);
676 
677 		/* clear the flags we set above */
678 		mapinfo->bp[i]->b_flags &= ~B_PHYS;
679 
680 		/* release the buffer */
681 		relpbuf(mapinfo->bp[i], NULL);
682 	}
683 
684 	/* allow ourselves to be swapped once again */
685 	PRELE(curproc);
686 }
687 
688 union ccb *
689 cam_periph_getccb(struct cam_periph *periph, u_int32_t priority)
690 {
691 	struct ccb_hdr *ccb_h;
692 	int s;
693 
694 	CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("entering cdgetccb\n"));
695 
696 	s = splsoftcam();
697 
698 	while (periph->ccb_list.slh_first == NULL) {
699 		if (periph->immediate_priority > priority)
700 			periph->immediate_priority = priority;
701 		xpt_schedule(periph, priority);
702 		if ((periph->ccb_list.slh_first != NULL)
703 		 && (periph->ccb_list.slh_first->pinfo.priority == priority))
704 			break;
705 		tsleep(&periph->ccb_list, PRIBIO, "cgticb", 0);
706 	}
707 
708 	ccb_h = periph->ccb_list.slh_first;
709 	SLIST_REMOVE_HEAD(&periph->ccb_list, periph_links.sle);
710 	splx(s);
711 	return ((union ccb *)ccb_h);
712 }
713 
714 void
715 cam_periph_ccbwait(union ccb *ccb)
716 {
717 	int s;
718 
719 	s = splsoftcam();
720 	if ((ccb->ccb_h.pinfo.index != CAM_UNQUEUED_INDEX)
721 	 || ((ccb->ccb_h.status & CAM_STATUS_MASK) == CAM_REQ_INPROG))
722 		tsleep(&ccb->ccb_h.cbfcnp, PRIBIO, "cbwait", 0);
723 
724 	splx(s);
725 }
726 
727 int
728 cam_periph_ioctl(struct cam_periph *periph, int cmd, caddr_t addr,
729 		 int (*error_routine)(union ccb *ccb,
730 				      cam_flags camflags,
731 				      u_int32_t sense_flags))
732 {
733 	union ccb 	     *ccb;
734 	int 		     error;
735 	int		     found;
736 
737 	error = found = 0;
738 
739 	switch(cmd){
740 	case CAMGETPASSTHRU:
741 		ccb = cam_periph_getccb(periph, /* priority */ 1);
742 		xpt_setup_ccb(&ccb->ccb_h,
743 			      ccb->ccb_h.path,
744 			      /*priority*/1);
745 		ccb->ccb_h.func_code = XPT_GDEVLIST;
746 
747 		/*
748 		 * Basically, the point of this is that we go through
749 		 * getting the list of devices, until we find a passthrough
750 		 * device.  In the current version of the CAM code, the
751 		 * only way to determine what type of device we're dealing
752 		 * with is by its name.
753 		 */
754 		while (found == 0) {
755 			ccb->cgdl.index = 0;
756 			ccb->cgdl.status = CAM_GDEVLIST_MORE_DEVS;
757 			while (ccb->cgdl.status == CAM_GDEVLIST_MORE_DEVS) {
758 
759 				/* we want the next device in the list */
760 				xpt_action(ccb);
761 				if (strncmp(ccb->cgdl.periph_name,
762 				    "pass", 4) == 0){
763 					found = 1;
764 					break;
765 				}
766 			}
767 			if ((ccb->cgdl.status == CAM_GDEVLIST_LAST_DEVICE) &&
768 			    (found == 0)) {
769 				ccb->cgdl.periph_name[0] = '\0';
770 				ccb->cgdl.unit_number = 0;
771 				break;
772 			}
773 		}
774 
775 		/* copy the result back out */
776 		bcopy(ccb, addr, sizeof(union ccb));
777 
778 		/* and release the ccb */
779 		xpt_release_ccb(ccb);
780 
781 		break;
782 	default:
783 		error = ENOTTY;
784 		break;
785 	}
786 	return(error);
787 }
788 
789 int
790 cam_periph_runccb(union ccb *ccb,
791 		  int (*error_routine)(union ccb *ccb,
792 				       cam_flags camflags,
793 				       u_int32_t sense_flags),
794 		  cam_flags camflags, u_int32_t sense_flags,
795 		  struct devstat *ds)
796 {
797 	int error;
798 
799 	error = 0;
800 
801 	/*
802 	 * If the user has supplied a stats structure, and if we understand
803 	 * this particular type of ccb, record the transaction start.
804 	 */
805 	if ((ds != NULL) && (ccb->ccb_h.func_code == XPT_SCSI_IO))
806 		devstat_start_transaction(ds);
807 
808 	xpt_action(ccb);
809 
810 	do {
811 		cam_periph_ccbwait(ccb);
812 		if ((ccb->ccb_h.status & CAM_STATUS_MASK) == CAM_REQ_CMP)
813 			error = 0;
814 		else if (error_routine != NULL)
815 			error = (*error_routine)(ccb, camflags, sense_flags);
816 		else
817 			error = 0;
818 
819 	} while (error == ERESTART);
820 
821 	if ((ccb->ccb_h.status & CAM_DEV_QFRZN) != 0)
822 		cam_release_devq(ccb->ccb_h.path,
823 				 /* relsim_flags */0,
824 				 /* openings */0,
825 				 /* timeout */0,
826 				 /* getcount_only */ FALSE);
827 
828 	if ((ds != NULL) && (ccb->ccb_h.func_code == XPT_SCSI_IO))
829 		devstat_end_transaction(ds,
830 					ccb->csio.dxfer_len,
831 					ccb->csio.tag_action & 0xf,
832 					((ccb->ccb_h.flags & CAM_DIR_MASK) ==
833 					CAM_DIR_NONE) ?  DEVSTAT_NO_DATA :
834 					(ccb->ccb_h.flags & CAM_DIR_OUT) ?
835 					DEVSTAT_WRITE :
836 					DEVSTAT_READ);
837 
838 	return(error);
839 }
840 
841 void
842 cam_freeze_devq(struct cam_path *path)
843 {
844 	struct ccb_hdr ccb_h;
845 
846 	xpt_setup_ccb(&ccb_h, path, /*priority*/1);
847 	ccb_h.func_code = XPT_NOOP;
848 	ccb_h.flags = CAM_DEV_QFREEZE;
849 	xpt_action((union ccb *)&ccb_h);
850 }
851 
852 u_int32_t
853 cam_release_devq(struct cam_path *path, u_int32_t relsim_flags,
854 		 u_int32_t openings, u_int32_t timeout,
855 		 int getcount_only)
856 {
857 	struct ccb_relsim crs;
858 
859 	xpt_setup_ccb(&crs.ccb_h, path,
860 		      /*priority*/1);
861 	crs.ccb_h.func_code = XPT_REL_SIMQ;
862 	crs.ccb_h.flags = getcount_only ? CAM_DEV_QFREEZE : 0;
863 	crs.release_flags = relsim_flags;
864 	crs.openings = openings;
865 	crs.release_timeout = timeout;
866 	xpt_action((union ccb *)&crs);
867 	return (crs.qfrozen_cnt);
868 }
869 
870 #define saved_ccb_ptr ppriv_ptr0
871 static void
872 camperiphdone(struct cam_periph *periph, union ccb *done_ccb)
873 {
874 	cam_status	status;
875 	int		frozen;
876 	int		sense;
877 	struct scsi_start_stop_unit *scsi_cmd;
878 	u_int32_t	relsim_flags, timeout;
879 	u_int32_t	qfrozen_cnt;
880 
881 	status = done_ccb->ccb_h.status;
882 	frozen = (status & CAM_DEV_QFRZN) != 0;
883 	sense  = (status & CAM_AUTOSNS_VALID) != 0;
884 	status &= CAM_STATUS_MASK;
885 
886 	timeout = 0;
887 	relsim_flags = 0;
888 
889 	/*
890 	 * Unfreeze the queue once if it is already frozen..
891 	 */
892 	if (frozen != 0) {
893 		qfrozen_cnt = cam_release_devq(done_ccb->ccb_h.path,
894 					      /*relsim_flags*/0,
895 					      /*openings*/0,
896 					      /*timeout*/0,
897 					      /*getcount_only*/0);
898 	}
899 
900 	switch (status) {
901 
902 	case CAM_REQ_CMP:
903 
904 		/*
905 		 * If we have successfully taken a device from the not
906 		 * ready to ready state, re-scan the device and re-get the
907 		 * inquiry information.  Many devices (mostly disks) don't
908 		 * properly report their inquiry information unless they
909 		 * are spun up.
910 		 */
911 		if (done_ccb->ccb_h.func_code == XPT_SCSI_IO) {
912 			scsi_cmd = (struct scsi_start_stop_unit *)
913 					&done_ccb->csio.cdb_io.cdb_bytes;
914 
915 		 	if (scsi_cmd->opcode == START_STOP_UNIT)
916 				xpt_async(AC_INQ_CHANGED,
917 					  done_ccb->ccb_h.path, NULL);
918 		}
919 		bcopy(done_ccb->ccb_h.saved_ccb_ptr, done_ccb,
920 		      sizeof(union ccb));
921 
922 		periph->flags &= ~CAM_PERIPH_RECOVERY_INPROG;
923 
924 		xpt_action(done_ccb);
925 
926 		break;
927 	case CAM_SCSI_STATUS_ERROR:
928 		scsi_cmd = (struct scsi_start_stop_unit *)
929 				&done_ccb->csio.cdb_io.cdb_bytes;
930 		if (sense != 0) {
931 			struct scsi_sense_data *sense;
932 			int    error_code, sense_key, asc, ascq;
933 
934 			sense = &done_ccb->csio.sense_data;
935 			scsi_extract_sense(sense, &error_code,
936 					   &sense_key, &asc, &ascq);
937 
938 			/*
939 	 		 * If the error is "invalid field in CDB",
940 			 * and the load/eject flag is set, turn the
941 			 * flag off and try again.  This is just in
942 			 * case the drive in question barfs on the
943 			 * load eject flag.  The CAM code should set
944 			 * the load/eject flag by default for
945 			 * removable media.
946 			 */
947 
948 			/* XXX KDM
949 			 * Should we check to see what the specific
950 			 * scsi status is??  Or does it not matter
951 			 * since we already know that there was an
952 			 * error, and we know what the specific
953 			 * error code was, and we know what the
954 			 * opcode is..
955 			 */
956 			if ((scsi_cmd->opcode == START_STOP_UNIT) &&
957 			    ((scsi_cmd->how & SSS_LOEJ) != 0) &&
958 			     (asc == 0x24) && (ascq == 0x00) &&
959 			     (done_ccb->ccb_h.retry_count > 0)) {
960 
961 				scsi_cmd->how &= ~SSS_LOEJ;
962 
963 				xpt_action(done_ccb);
964 
965 			} else if (done_ccb->ccb_h.retry_count > 0) {
966 				/*
967 				 * In this case, the error recovery
968 				 * command failed, but we've got
969 				 * some retries left on it.  Give
970 				 * it another try.
971 				 */
972 
973 				/* set the timeout to .5 sec */
974 				relsim_flags =
975 					RELSIM_RELEASE_AFTER_TIMEOUT;
976 				timeout = 500;
977 
978 				xpt_action(done_ccb);
979 
980 				break;
981 
982 			} else {
983 				/*
984 				 * Copy the original CCB back and
985 				 * send it back to the caller.
986 				 */
987 				bcopy(done_ccb->ccb_h.saved_ccb_ptr,
988 				      done_ccb, sizeof(union ccb));
989 
990 				periph->flags &= ~CAM_PERIPH_RECOVERY_INPROG;
991 
992 				xpt_action(done_ccb);
993 			}
994 		} else {
995 			/*
996 			 * Eh??  The command failed, but we don't
997 			 * have any sense.  What's up with that?
998 			 * Fire the CCB again to return it to the
999 			 * caller.
1000 			 */
1001 			bcopy(done_ccb->ccb_h.saved_ccb_ptr,
1002 			      done_ccb, sizeof(union ccb));
1003 
1004 			periph->flags &= ~CAM_PERIPH_RECOVERY_INPROG;
1005 
1006 			xpt_action(done_ccb);
1007 
1008 		}
1009 		break;
1010 	default:
1011 		bcopy(done_ccb->ccb_h.saved_ccb_ptr, done_ccb,
1012 		      sizeof(union ccb));
1013 
1014 		periph->flags &= ~CAM_PERIPH_RECOVERY_INPROG;
1015 
1016 		xpt_action(done_ccb);
1017 
1018 		break;
1019 	}
1020 
1021 	/* decrement the retry count */
1022 	if (done_ccb->ccb_h.retry_count > 0)
1023 		done_ccb->ccb_h.retry_count--;
1024 
1025 	qfrozen_cnt = cam_release_devq(done_ccb->ccb_h.path,
1026 				      /*relsim_flags*/relsim_flags,
1027 				      /*openings*/0,
1028 				      /*timeout*/timeout,
1029 				      /*getcount_only*/0);
1030 }
1031 
1032 /*
1033  * Generic Async Event handler.  Peripheral drivers usually
1034  * filter out the events that require personal attention,
1035  * and leave the rest to this function.
1036  */
1037 void
1038 cam_periph_async(struct cam_periph *periph, u_int32_t code,
1039 		 struct cam_path *path, void *arg)
1040 {
1041 	switch (code) {
1042 	case AC_LOST_DEVICE:
1043 		cam_periph_invalidate(periph);
1044 		break;
1045 	case AC_SENT_BDR:
1046 	case AC_BUS_RESET:
1047 	{
1048 		cam_periph_bus_settle(periph, SCSI_DELAY);
1049 		break;
1050 	}
1051 	default:
1052 		break;
1053 	}
1054 }
1055 
1056 void
1057 cam_periph_bus_settle(struct cam_periph *periph, u_int bus_settle)
1058 {
1059 	struct ccb_getdevstats cgds;
1060 
1061 	xpt_setup_ccb(&cgds.ccb_h, periph->path, /*priority*/1);
1062 	cgds.ccb_h.func_code = XPT_GDEV_STATS;
1063 	xpt_action((union ccb *)&cgds);
1064 	cam_periph_freeze_after_event(periph, &cgds.last_reset, bus_settle);
1065 }
1066 
1067 void
1068 cam_periph_freeze_after_event(struct cam_periph *periph,
1069 			      struct timeval* event_time, u_int duration_ms)
1070 {
1071 	struct timeval delta;
1072 	struct timeval duration_tv;
1073 	int s;
1074 
1075 	s = splclock();
1076 	microtime(&delta);
1077 	splx(s);
1078 	timevalsub(&delta, event_time);
1079 	duration_tv.tv_sec = duration_ms / 1000;
1080 	duration_tv.tv_usec = (duration_ms % 1000) * 1000;
1081 	if (timevalcmp(&delta, &duration_tv, <)) {
1082 		timevalsub(&duration_tv, &delta);
1083 
1084 		duration_ms = duration_tv.tv_sec * 1000;
1085 		duration_ms += duration_tv.tv_usec / 1000;
1086 		cam_freeze_devq(periph->path);
1087 		cam_release_devq(periph->path,
1088 				RELSIM_RELEASE_AFTER_TIMEOUT,
1089 				/*reduction*/0,
1090 				/*timeout*/duration_ms,
1091 				/*getcount_only*/0);
1092 	}
1093 
1094 }
1095 
1096 /*
1097  * Generic error handler.  Peripheral drivers usually filter
1098  * out the errors that they handle in a unique mannor, then
1099  * call this function.
1100  */
1101 int
1102 cam_periph_error(union ccb *ccb, cam_flags camflags,
1103 		 u_int32_t sense_flags, union ccb *save_ccb)
1104 {
1105 	cam_status status;
1106 	int	   frozen;
1107 	int	   sense;
1108 	int	   error;
1109 	int        openings;
1110 	int	   retry;
1111 	u_int32_t  relsim_flags;
1112 	u_int32_t  timeout;
1113 
1114 	status = ccb->ccb_h.status;
1115 	frozen = (status & CAM_DEV_QFRZN) != 0;
1116 	sense  = (status & CAM_AUTOSNS_VALID) != 0;
1117 	status &= CAM_STATUS_MASK;
1118 	relsim_flags = 0;
1119 
1120 	switch (status) {
1121 	case CAM_REQ_CMP:
1122 		/* decrement the number of retries */
1123 		retry = ccb->ccb_h.retry_count > 0;
1124 		if (retry)
1125 			ccb->ccb_h.retry_count--;
1126 		error = 0;
1127 		break;
1128 	case CAM_AUTOSENSE_FAIL:
1129 	case CAM_SCSI_STATUS_ERROR:
1130 
1131 		switch (ccb->csio.scsi_status) {
1132 		case SCSI_STATUS_OK:
1133 		case SCSI_STATUS_COND_MET:
1134 		case SCSI_STATUS_INTERMED:
1135 		case SCSI_STATUS_INTERMED_COND_MET:
1136 			error = 0;
1137 			break;
1138 		case SCSI_STATUS_CMD_TERMINATED:
1139 		case SCSI_STATUS_CHECK_COND:
1140 			if (sense != 0) {
1141 				struct scsi_sense_data *sense;
1142 				int    error_code, sense_key, asc, ascq;
1143 				struct cam_periph *periph;
1144 				scsi_sense_action err_action;
1145 				struct ccb_getdev cgd;
1146 
1147 				sense = &ccb->csio.sense_data;
1148 				scsi_extract_sense(sense, &error_code,
1149 						   &sense_key, &asc, &ascq);
1150 				periph = xpt_path_periph(ccb->ccb_h.path);
1151 
1152 				/*
1153 				 * Grab the inquiry data for this device.
1154 				 */
1155 				xpt_setup_ccb(&cgd.ccb_h, ccb->ccb_h.path,
1156 					      /*priority*/ 1);
1157 				cgd.ccb_h.func_code = XPT_GDEV_TYPE;
1158 				xpt_action((union ccb *)&cgd);
1159 
1160 				err_action = scsi_error_action(asc, ascq,
1161 							       &cgd.inq_data);
1162 
1163 				/*
1164 				 * Send a Test Unit Ready to the device.
1165 				 * If the 'many' flag is set, we send 120
1166 				 * test unit ready commands, one every half
1167 				 * second.  Otherwise, we just send one TUR.
1168 				 * We only want to do this if the retry
1169 				 * count has not been exhausted.
1170 				 */
1171 				if (((err_action & SS_MASK) == SS_TUR)
1172 				 && save_ccb != NULL
1173 				 && ccb->ccb_h.retry_count > 0) {
1174 
1175 					/*
1176 					 * Since error recovery is already
1177 					 * in progress, don't attempt to
1178 					 * process this error.  It is probably
1179 					 * related to the error that caused
1180 					 * the currently active error recovery
1181 					 * action.  Also, we only have
1182 					 * space for one saved CCB, so if we
1183 					 * had two concurrent error recovery
1184 					 * actions, we would end up
1185 					 * over-writing one error recovery
1186 					 * CCB with another one.
1187 					 */
1188 					if (periph->flags &
1189 					    CAM_PERIPH_RECOVERY_INPROG) {
1190 						error = ERESTART;
1191 						break;
1192 					}
1193 
1194 					periph->flags |=
1195 						CAM_PERIPH_RECOVERY_INPROG;
1196 
1197 					/* decrement the number of retries */
1198 					if ((err_action &
1199 					     SSQ_DECREMENT_COUNT) != 0) {
1200 						retry = 1;
1201 						ccb->ccb_h.retry_count--;
1202 					}
1203 
1204 					bcopy(ccb, save_ccb, sizeof(*save_ccb));
1205 
1206 					/*
1207 					 * We retry this one every half
1208 					 * second for a minute.  If the
1209 					 * device hasn't become ready in a
1210 					 * minute's time, it's unlikely to
1211 					 * ever become ready.  If the table
1212 					 * doesn't specify SSQ_MANY, we can
1213 					 * only try this once.  Oh well.
1214 					 */
1215 					if ((err_action & SSQ_MANY) != 0)
1216 						scsi_test_unit_ready(&ccb->csio,
1217 							       /*retries*/120,
1218 							       camperiphdone,
1219 						 	       MSG_SIMPLE_Q_TAG,
1220 							       SSD_FULL_SIZE,
1221 							       /*timeout*/5000);
1222 					else
1223 						scsi_test_unit_ready(&ccb->csio,
1224 							       /*retries*/1,
1225 							       camperiphdone,
1226 						 	       MSG_SIMPLE_Q_TAG,
1227 							       SSD_FULL_SIZE,
1228 							       /*timeout*/5000);
1229 
1230 					/* release the queue after .5 sec.  */
1231 					relsim_flags =
1232 						RELSIM_RELEASE_AFTER_TIMEOUT;
1233 					timeout = 500;
1234 					/*
1235 					 * Drop the priority to 0 so that
1236 					 * we are the first to execute.  Also
1237 					 * freeze the queue after this command
1238 					 * is sent so that we can restore the
1239 					 * old csio and have it queued in the
1240 					 * proper order before we let normal
1241 					 * transactions go to the drive.
1242 					 */
1243 					ccb->ccb_h.pinfo.priority = 0;
1244 					ccb->ccb_h.flags |= CAM_DEV_QFREEZE;
1245 
1246 					/*
1247 					 * Save a pointer to the original
1248 					 * CCB in the new CCB.
1249 					 */
1250 					ccb->ccb_h.saved_ccb_ptr = save_ccb;
1251 
1252 					error = ERESTART;
1253 				}
1254 				/*
1255 				 * Send a start unit command to the device,
1256 				 * and then retry the command.  We only
1257 				 * want to do this if the retry count has
1258 				 * not been exhausted.  If the user
1259 				 * specified 0 retries, then we follow
1260 				 * their request and do not retry.
1261 				 */
1262 				else if (((err_action & SS_MASK) == SS_START)
1263 				      && save_ccb != NULL
1264 				      && ccb->ccb_h.retry_count > 0) {
1265 					int le;
1266 
1267 					/*
1268 					 * Only one error recovery action
1269 					 * at a time.  See above.
1270 					 */
1271 					if (periph->flags &
1272 					    CAM_PERIPH_RECOVERY_INPROG) {
1273 						error = ERESTART;
1274 						break;
1275 					}
1276 
1277 					periph->flags |=
1278 						CAM_PERIPH_RECOVERY_INPROG;
1279 
1280 					/* decrement the number of retries */
1281 					retry = 1;
1282 					ccb->ccb_h.retry_count--;
1283 
1284 					/*
1285 					 * Check for removable media and
1286 					 * set load/eject flag
1287 					 * appropriately.
1288 					 */
1289 					if (SID_IS_REMOVABLE(&cgd.inq_data))
1290 						le = TRUE;
1291 					else
1292 						le = FALSE;
1293 
1294 					/*
1295 					 * Attempt to start the drive up.
1296 					 *
1297 					 * Save the current ccb so it can
1298 					 * be restored and retried once the
1299 					 * drive is started up.
1300 					 */
1301 					bcopy(ccb, save_ccb, sizeof(*save_ccb));
1302 
1303 					scsi_start_stop(&ccb->csio,
1304 							/*retries*/1,
1305 							camperiphdone,
1306 							MSG_SIMPLE_Q_TAG,
1307 							/*start*/TRUE,
1308 							/*load/eject*/le,
1309 							/*immediate*/FALSE,
1310 							SSD_FULL_SIZE,
1311 							/*timeout*/50000);
1312 					/*
1313 					 * Drop the priority to 0 so that
1314 					 * we are the first to execute.  Also
1315 					 * freeze the queue after this command
1316 					 * is sent so that we can restore the
1317 					 * old csio and have it queued in the
1318 					 * proper order before we let normal
1319 					 * transactions go to the drive.
1320 					 */
1321 					ccb->ccb_h.pinfo.priority = 0;
1322 					ccb->ccb_h.flags |= CAM_DEV_QFREEZE;
1323 
1324 					/*
1325 					 * Save a pointer to the original
1326 					 * CCB in the new CCB.
1327 					 */
1328 					ccb->ccb_h.saved_ccb_ptr = save_ccb;
1329 
1330 					error = ERESTART;
1331 				} else if ((sense_flags & SF_RETRY_UA) != 0) {
1332 					/*
1333 					 * XXX KDM this is a *horrible*
1334 					 * hack.
1335 					 */
1336 					error = scsi_interpret_sense(ccb,
1337 								  sense_flags,
1338 								  &relsim_flags,
1339 								  &openings,
1340 								  &timeout,
1341 								  err_action);
1342 				}
1343 
1344 				/*
1345 				 * Theoretically, this code should send a
1346 				 * test unit ready to the given device, and
1347 				 * if it returns and error, send a start
1348 				 * unit command.  Since we don't yet have
1349 				 * the capability to do two-command error
1350 				 * recovery, just send a start unit.
1351 				 * XXX KDM fix this!
1352 				 */
1353 				else if (((err_action & SS_MASK) == SS_TURSTART)
1354 				      && save_ccb != NULL
1355 				      && ccb->ccb_h.retry_count > 0) {
1356 					int le;
1357 
1358 					/*
1359 					 * Only one error recovery action
1360 					 * at a time.  See above.
1361 					 */
1362 					if (periph->flags &
1363 					    CAM_PERIPH_RECOVERY_INPROG) {
1364 						error = ERESTART;
1365 						break;
1366 					}
1367 
1368 					periph->flags |=
1369 						CAM_PERIPH_RECOVERY_INPROG;
1370 
1371 					/* decrement the number of retries */
1372 					retry = 1;
1373 					ccb->ccb_h.retry_count--;
1374 
1375 					/*
1376 					 * Check for removable media and
1377 					 * set load/eject flag
1378 					 * appropriately.
1379 					 */
1380 					if (SID_IS_REMOVABLE(&cgd.inq_data))
1381 						le = TRUE;
1382 					else
1383 						le = FALSE;
1384 
1385 					/*
1386 					 * Attempt to start the drive up.
1387 					 *
1388 					 * Save the current ccb so it can
1389 					 * be restored and retried once the
1390 					 * drive is started up.
1391 					 */
1392 					bcopy(ccb, save_ccb, sizeof(*save_ccb));
1393 
1394 					scsi_start_stop(&ccb->csio,
1395 							/*retries*/1,
1396 							camperiphdone,
1397 							MSG_SIMPLE_Q_TAG,
1398 							/*start*/TRUE,
1399 							/*load/eject*/le,
1400 							/*immediate*/FALSE,
1401 							SSD_FULL_SIZE,
1402 							/*timeout*/50000);
1403 
1404 					/* release the queue after .5 sec.  */
1405 					relsim_flags =
1406 						RELSIM_RELEASE_AFTER_TIMEOUT;
1407 					timeout = 500;
1408 					/*
1409 					 * Drop the priority to 0 so that
1410 					 * we are the first to execute.  Also
1411 					 * freeze the queue after this command
1412 					 * is sent so that we can restore the
1413 					 * old csio and have it queued in the
1414 					 * proper order before we let normal
1415 					 * transactions go to the drive.
1416 					 */
1417 					ccb->ccb_h.pinfo.priority = 0;
1418 					ccb->ccb_h.flags |= CAM_DEV_QFREEZE;
1419 
1420 					/*
1421 					 * Save a pointer to the original
1422 					 * CCB in the new CCB.
1423 					 */
1424 					ccb->ccb_h.saved_ccb_ptr = save_ccb;
1425 
1426 					error = ERESTART;
1427 				} else {
1428 					error = scsi_interpret_sense(ccb,
1429 								  sense_flags,
1430 								  &relsim_flags,
1431 								  &openings,
1432 								  &timeout,
1433 								  err_action);
1434 				}
1435 			} else if (ccb->csio.scsi_status ==
1436 				   SCSI_STATUS_CHECK_COND
1437 				&& status != CAM_AUTOSENSE_FAIL) {
1438 				/* no point in decrementing the retry count */
1439 				panic("cam_periph_error: scsi status of "
1440 				      "CHECK COND returned but no sense "
1441 				      "information is availible.  "
1442 				      "Controller should have returned "
1443 				      "CAM_AUTOSENSE_FAILED");
1444 				/* NOTREACHED */
1445 				error = EIO;
1446 			} else if (ccb->ccb_h.retry_count > 0) {
1447 				/*
1448 				 * XXX KDM shouldn't there be a better
1449 				 * argument to return??
1450 				 */
1451 				error = EIO;
1452 			} else {
1453 				/* decrement the number of retries */
1454 				retry = ccb->ccb_h.retry_count > 0;
1455 				if (retry)
1456 					ccb->ccb_h.retry_count--;
1457 				/*
1458 				 * If it was aborted with no
1459 				 * clue as to the reason, just
1460 				 * retry it again.
1461 				 */
1462 				error = ERESTART;
1463 			}
1464 			break;
1465 		case SCSI_STATUS_QUEUE_FULL:
1466 		{
1467 			/* no decrement */
1468 			struct ccb_getdevstats cgds;
1469 
1470 			/*
1471 			 * First off, find out what the current
1472 			 * transaction counts are.
1473 			 */
1474 			xpt_setup_ccb(&cgds.ccb_h,
1475 				      ccb->ccb_h.path,
1476 				      /*priority*/1);
1477 			cgds.ccb_h.func_code = XPT_GDEV_STATS;
1478 			xpt_action((union ccb *)&cgds);
1479 
1480 			/*
1481 			 * If we were the only transaction active, treat
1482 			 * the QUEUE FULL as if it were a BUSY condition.
1483 			 */
1484 			if (cgds.dev_active != 0) {
1485 				int total_openings;
1486 
1487 				/*
1488 			 	 * Reduce the number of openings to
1489 				 * be 1 less than the amount it took
1490 				 * to get a queue full bounded by the
1491 				 * minimum allowed tag count for this
1492 				 * device.
1493 			 	 */
1494 				total_openings =
1495 				    cgds.dev_active+cgds.dev_openings;
1496 				openings = cgds.dev_active;
1497 				if (openings < cgds.mintags)
1498 					openings = cgds.mintags;
1499 				if (openings < total_openings)
1500 					relsim_flags = RELSIM_ADJUST_OPENINGS;
1501 				else {
1502 					/*
1503 					 * Some devices report queue full for
1504 					 * temporary resource shortages.  For
1505 					 * this reason, we allow a minimum
1506 					 * tag count to be entered via a
1507 					 * quirk entry to prevent the queue
1508 					 * count on these devices from falling
1509 					 * to a pessimisticly low value.  We
1510 					 * still wait for the next successful
1511 					 * completion, however, before queueing
1512 					 * more transactions to the device.
1513 					 */
1514 					relsim_flags =
1515 					    RELSIM_RELEASE_AFTER_CMDCMPLT;
1516 				}
1517 				timeout = 0;
1518 				error = ERESTART;
1519 				break;
1520 			}
1521 			/* FALLTHROUGH */
1522 		}
1523 		case SCSI_STATUS_BUSY:
1524 			/*
1525 			 * Restart the queue after either another
1526 			 * command completes or a 1 second timeout.
1527 			 * If we have any retries left, that is.
1528 			 */
1529 			retry = ccb->ccb_h.retry_count > 0;
1530 			if (retry) {
1531 				ccb->ccb_h.retry_count--;
1532 				error = ERESTART;
1533 				relsim_flags = RELSIM_RELEASE_AFTER_TIMEOUT
1534 					     | RELSIM_RELEASE_AFTER_CMDCMPLT;
1535 				timeout = 1000;
1536 			} else {
1537 				error = EIO;
1538 			}
1539 			break;
1540 		case SCSI_STATUS_RESERV_CONFLICT:
1541 			error = EIO;
1542 			break;
1543 		default:
1544 			error = EIO;
1545 			break;
1546 		}
1547 		break;
1548 	case CAM_REQ_CMP_ERR:
1549 	case CAM_CMD_TIMEOUT:
1550 	case CAM_UNEXP_BUSFREE:
1551 	case CAM_UNCOR_PARITY:
1552 	case CAM_DATA_RUN_ERR:
1553 		/* decrement the number of retries */
1554 		retry = ccb->ccb_h.retry_count > 0;
1555 		if (retry) {
1556 			ccb->ccb_h.retry_count--;
1557 			error = ERESTART;
1558 		} else {
1559 			error = EIO;
1560 		}
1561 		break;
1562 	case CAM_UA_ABORT:
1563 	case CAM_UA_TERMIO:
1564 	case CAM_MSG_REJECT_REC:
1565 		/* XXX Don't know that these are correct */
1566 		error = EIO;
1567 		break;
1568 	case CAM_SEL_TIMEOUT:
1569 	{
1570 		/*
1571 		 * XXX
1572 		 * A single selection timeout should not be enough
1573 		 * to invalidate a device.  We should retry for multiple
1574 		 * seconds assuming this isn't a probe.  We'll probably
1575 		 * need a special flag for that.
1576 		 */
1577 #if 0
1578 		struct cam_path *newpath;
1579 
1580 		/* Should we do more if we can't create the path?? */
1581 		if (xpt_create_path(&newpath, xpt_path_periph(ccb->ccb_h.path),
1582 				    xpt_path_path_id(ccb->ccb_h.path),
1583 				    xpt_path_target_id(ccb->ccb_h.path),
1584 				    CAM_LUN_WILDCARD) != CAM_REQ_CMP)
1585 			break;
1586 		/*
1587 		 * Let peripheral drivers know that this device has gone
1588 		 * away.
1589 		 */
1590 		xpt_async(AC_LOST_DEVICE, newpath, NULL);
1591 		xpt_free_path(newpath);
1592 #endif
1593 		if ((sense_flags & SF_RETRY_SELTO) != 0) {
1594 			retry = ccb->ccb_h.retry_count > 0;
1595 			if (retry) {
1596 				ccb->ccb_h.retry_count--;
1597 				error = ERESTART;
1598 				/*
1599 				 * Wait half a second to give the device
1600 				 * time to recover before we try again.
1601 				 */
1602 				relsim_flags = RELSIM_RELEASE_AFTER_TIMEOUT;
1603 				timeout = 500;
1604 			} else {
1605 				error = ENXIO;
1606 			}
1607 		} else {
1608 			error = ENXIO;
1609 		}
1610 		break;
1611 	}
1612 	case CAM_REQ_INVALID:
1613 	case CAM_PATH_INVALID:
1614 	case CAM_DEV_NOT_THERE:
1615 	case CAM_NO_HBA:
1616 	case CAM_PROVIDE_FAIL:
1617 	case CAM_REQ_TOO_BIG:
1618 		error = EINVAL;
1619 		break;
1620 	case CAM_SCSI_BUS_RESET:
1621 	case CAM_BDR_SENT:
1622 	case CAM_REQUEUE_REQ:
1623 		/* Unconditional requeue, dammit */
1624 		error = ERESTART;
1625 		break;
1626 	case CAM_RESRC_UNAVAIL:
1627 	case CAM_BUSY:
1628 		/* timeout??? */
1629 	default:
1630 		/* decrement the number of retries */
1631 		retry = ccb->ccb_h.retry_count > 0;
1632 		if (retry) {
1633 			ccb->ccb_h.retry_count--;
1634 			error = ERESTART;
1635 		} else {
1636 			/* Check the sense codes */
1637 			error = EIO;
1638 		}
1639 		break;
1640 	}
1641 
1642 	/* Attempt a retry */
1643 	if (error == ERESTART || error == 0) {
1644 		if (frozen != 0)
1645 			ccb->ccb_h.status &= ~CAM_DEV_QFRZN;
1646 
1647 		if (error == ERESTART)
1648 			xpt_action(ccb);
1649 
1650 		if (frozen != 0) {
1651 			cam_release_devq(ccb->ccb_h.path,
1652 					 relsim_flags,
1653 					 openings,
1654 					 timeout,
1655 					 /*getcount_only*/0);
1656 		}
1657 	}
1658 
1659 
1660 	return (error);
1661 }
1662