xref: /freebsd/sys/cam/cam_periph.c (revision 64db83a8ab2d1f72a9b2174b39d2ef42b5b0580c)
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/bio.h>
38 #include <sys/buf.h>
39 #include <sys/proc.h>
40 #include <sys/devicestat.h>
41 #include <sys/bus.h>
42 #include <vm/vm.h>
43 #include <vm/vm_extern.h>
44 
45 #include <cam/cam.h>
46 #include <cam/cam_ccb.h>
47 #include <cam/cam_xpt_periph.h>
48 #include <cam/cam_periph.h>
49 #include <cam/cam_debug.h>
50 
51 #include <cam/scsi/scsi_all.h>
52 #include <cam/scsi/scsi_message.h>
53 #include <cam/scsi/scsi_da.h>
54 #include <cam/scsi/scsi_pass.h>
55 
56 static	u_int		camperiphnextunit(struct periph_driver *p_drv,
57 					  u_int newunit, int wired);
58 static	u_int		camperiphunit(struct periph_driver *p_drv,
59 				      path_id_t path_id_t,
60 				      target_id_t target, lun_id_t lun);
61 static	void		camperiphdone(struct cam_periph *periph,
62 					union ccb *done_ccb);
63 static  void		camperiphfree(struct cam_periph *periph);
64 
65 cam_status
66 cam_periph_alloc(periph_ctor_t *periph_ctor,
67 		 periph_oninv_t *periph_oninvalidate,
68 		 periph_dtor_t *periph_dtor, periph_start_t *periph_start,
69 		 char *name, cam_periph_type type, struct cam_path *path,
70 		 ac_callback_t *ac_callback, ac_code code, void *arg)
71 {
72 	struct		periph_driver **p_drv;
73 	struct		cam_periph *periph;
74 	struct		cam_periph *cur_periph;
75 	path_id_t	path_id;
76 	target_id_t	target_id;
77 	lun_id_t	lun_id;
78 	cam_status	status;
79 	u_int		init_level;
80 	int s;
81 
82 	init_level = 0;
83 	/*
84 	 * Handle Hot-Plug scenarios.  If there is already a peripheral
85 	 * of our type assigned to this path, we are likely waiting for
86 	 * final close on an old, invalidated, peripheral.  If this is
87 	 * the case, queue up a deferred call to the peripheral's async
88 	 * handler.  If it looks like a mistaken re-alloation, complain.
89 	 */
90 	if ((periph = cam_periph_find(path, name)) != NULL) {
91 
92 		if ((periph->flags & CAM_PERIPH_INVALID) != 0
93 		 && (periph->flags & CAM_PERIPH_NEW_DEV_FOUND) == 0) {
94 			periph->flags |= CAM_PERIPH_NEW_DEV_FOUND;
95 			periph->deferred_callback = ac_callback;
96 			periph->deferred_ac = code;
97 			return (CAM_REQ_INPROG);
98 		} else {
99 			printf("cam_periph_alloc: attempt to re-allocate "
100 			       "valid device %s%d rejected\n",
101 			       periph->periph_name, periph->unit_number);
102 		}
103 		return (CAM_REQ_INVALID);
104 	}
105 
106 	periph = (struct cam_periph *)malloc(sizeof(*periph), M_DEVBUF,
107 					     M_NOWAIT);
108 
109 	if (periph == NULL)
110 		return (CAM_RESRC_UNAVAIL);
111 
112 	init_level++;
113 
114 	for (p_drv = (struct periph_driver **)periphdriver_set.ls_items;
115 	     *p_drv != NULL; p_drv++) {
116 		if (strcmp((*p_drv)->driver_name, name) == 0)
117 			break;
118 	}
119 
120 	path_id = xpt_path_path_id(path);
121 	target_id = xpt_path_target_id(path);
122 	lun_id = xpt_path_lun_id(path);
123 	bzero(periph, sizeof(*periph));
124 	cam_init_pinfo(&periph->pinfo);
125 	periph->periph_start = periph_start;
126 	periph->periph_dtor = periph_dtor;
127 	periph->periph_oninval = periph_oninvalidate;
128 	periph->type = type;
129 	periph->periph_name = name;
130 	periph->unit_number = camperiphunit(*p_drv, path_id, target_id, lun_id);
131 	periph->immediate_priority = CAM_PRIORITY_NONE;
132 	periph->refcount = 0;
133 	SLIST_INIT(&periph->ccb_list);
134 	status = xpt_create_path(&path, periph, path_id, target_id, lun_id);
135 	if (status != CAM_REQ_CMP)
136 		goto failure;
137 
138 	periph->path = path;
139 	init_level++;
140 
141 	status = xpt_add_periph(periph);
142 
143 	if (status != CAM_REQ_CMP)
144 		goto failure;
145 
146 	s = splsoftcam();
147 	cur_periph = TAILQ_FIRST(&(*p_drv)->units);
148 	while (cur_periph != NULL
149 	    && cur_periph->unit_number < periph->unit_number)
150 		cur_periph = TAILQ_NEXT(cur_periph, unit_links);
151 
152 	if (cur_periph != NULL)
153 		TAILQ_INSERT_BEFORE(cur_periph, periph, unit_links);
154 	else {
155 		TAILQ_INSERT_TAIL(&(*p_drv)->units, periph, unit_links);
156 		(*p_drv)->generation++;
157 	}
158 
159 	splx(s);
160 
161 	init_level++;
162 
163 	status = periph_ctor(periph, arg);
164 
165 	if (status == CAM_REQ_CMP)
166 		init_level++;
167 
168 failure:
169 	switch (init_level) {
170 	case 4:
171 		/* Initialized successfully */
172 		break;
173 	case 3:
174 		s = splsoftcam();
175 		TAILQ_REMOVE(&(*p_drv)->units, periph, unit_links);
176 		splx(s);
177 		xpt_remove_periph(periph);
178 	case 2:
179 		xpt_free_path(periph->path);
180 	case 1:
181 		free(periph, M_DEVBUF);
182 	case 0:
183 		/* No cleanup to perform. */
184 		break;
185 	default:
186 		panic("cam_periph_alloc: Unkown init level");
187 	}
188 	return(status);
189 }
190 
191 /*
192  * Find a peripheral structure with the specified path, target, lun,
193  * and (optionally) type.  If the name is NULL, this function will return
194  * the first peripheral driver that matches the specified path.
195  */
196 struct cam_periph *
197 cam_periph_find(struct cam_path *path, char *name)
198 {
199 	struct periph_driver **p_drv;
200 	struct cam_periph *periph;
201 	int s;
202 
203 	for (p_drv = (struct periph_driver **)periphdriver_set.ls_items;
204 	     *p_drv != NULL; p_drv++) {
205 
206 		if (name != NULL && (strcmp((*p_drv)->driver_name, name) != 0))
207 			continue;
208 
209 		s = splsoftcam();
210 		for (periph = TAILQ_FIRST(&(*p_drv)->units); periph != NULL;
211 		     periph = TAILQ_NEXT(periph, unit_links)) {
212 			if (xpt_path_comp(periph->path, path) == 0) {
213 				splx(s);
214 				return(periph);
215 			}
216 		}
217 		splx(s);
218 		if (name != NULL)
219 			return(NULL);
220 	}
221 	return(NULL);
222 }
223 
224 cam_status
225 cam_periph_acquire(struct cam_periph *periph)
226 {
227 	int s;
228 
229 	if (periph == NULL)
230 		return(CAM_REQ_CMP_ERR);
231 
232 	s = splsoftcam();
233 	periph->refcount++;
234 	splx(s);
235 
236 	return(CAM_REQ_CMP);
237 }
238 
239 void
240 cam_periph_release(struct cam_periph *periph)
241 {
242 	int s;
243 
244 	if (periph == NULL)
245 		return;
246 
247 	s = splsoftcam();
248 	if ((--periph->refcount == 0)
249 	 && (periph->flags & CAM_PERIPH_INVALID)) {
250 		camperiphfree(periph);
251 	}
252 	splx(s);
253 
254 }
255 
256 /*
257  * Look for the next unit number that is not currently in use for this
258  * peripheral type starting at "newunit".  Also exclude unit numbers that
259  * are reserved by for future "hardwiring" unless we already know that this
260  * is a potential wired device.  Only assume that the device is "wired" the
261  * first time through the loop since after that we'll be looking at unit
262  * numbers that did not match a wiring entry.
263  */
264 static u_int
265 camperiphnextunit(struct periph_driver *p_drv, u_int newunit, int wired)
266 {
267 	struct	cam_periph *periph;
268 	char	*periph_name, *strval;
269 	int	s;
270 	int	i, val, dunit;
271 	const char *dname;
272 
273 	s = splsoftcam();
274 	periph_name = p_drv->driver_name;
275 	for (;;newunit++) {
276 
277 		for (periph = TAILQ_FIRST(&p_drv->units);
278 		     periph != NULL && periph->unit_number != newunit;
279 		     periph = TAILQ_NEXT(periph, unit_links))
280 			;
281 
282 		if (periph != NULL && periph->unit_number == newunit) {
283 			if (wired != 0) {
284 				xpt_print_path(periph->path);
285 				printf("Duplicate Wired Device entry!\n");
286 				xpt_print_path(periph->path);
287 				printf("Second device will not be wired\n");
288 				wired = 0;
289 			}
290 			continue;
291 		}
292 		if (wired)
293 			break;
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 		i = -1;
301 		while ((i = resource_locate(i, periph_name)) != -1) {
302 			dname = resource_query_name(i);
303 			dunit = resource_query_unit(i);
304 			/* if no "target" and no specific scbus, skip */
305 			if (resource_int_value(dname, dunit, "target", &val) &&
306 			    (resource_string_value(dname, dunit, "at",&strval)||
307 			     strcmp(strval, "scbus") == 0))
308 				continue;
309 			if (newunit == dunit)
310 				break;
311 		}
312 		if (i == -1)
313 			break;
314 	}
315 	splx(s);
316 	return (newunit);
317 }
318 
319 static u_int
320 camperiphunit(struct periph_driver *p_drv, path_id_t pathid,
321 	      target_id_t target, lun_id_t lun)
322 {
323 	u_int	unit;
324 	int	hit, i, val, dunit;
325 	const char *dname;
326 	char	pathbuf[32], *strval, *periph_name;
327 
328 	unit = 0;
329 	hit = 0;
330 
331 	periph_name = p_drv->driver_name;
332 	snprintf(pathbuf, sizeof(pathbuf), "scbus%d", pathid);
333 	i = -1;
334 	while ((i = resource_locate(i, periph_name)) != -1) {
335 		dname = resource_query_name(i);
336 		dunit = resource_query_unit(i);
337 		if (resource_string_value(dname, dunit, "at", &strval) == 0) {
338 			if (strcmp(strval, pathbuf) != 0)
339 				continue;
340 			hit++;
341 		}
342 		if (resource_int_value(dname, dunit, "target", &val) == 0) {
343 			if (val != target)
344 				continue;
345 			hit++;
346 		}
347 		if (resource_int_value(dname, dunit, "lun", &val) == 0) {
348 			if (val != lun)
349 				continue;
350 			hit++;
351 		}
352 		if (hit != 0) {
353 			unit = dunit;
354 			break;
355 		}
356 	}
357 
358 	/*
359 	 * Either start from 0 looking for the next unit or from
360 	 * the unit number given in the resource config.  This way,
361 	 * if we have wildcard matches, we don't return the same
362 	 * unit number twice.
363 	 */
364 	unit = camperiphnextunit(p_drv, unit, /*wired*/hit);
365 
366 	return (unit);
367 }
368 
369 void
370 cam_periph_invalidate(struct cam_periph *periph)
371 {
372 	int s;
373 
374 	s = splsoftcam();
375 	/*
376 	 * We only call this routine the first time a peripheral is
377 	 * invalidated.  The oninvalidate() routine is always called at
378 	 * splsoftcam().
379 	 */
380 	if (((periph->flags & CAM_PERIPH_INVALID) == 0)
381 	 && (periph->periph_oninval != NULL))
382 		periph->periph_oninval(periph);
383 
384 	periph->flags |= CAM_PERIPH_INVALID;
385 	periph->flags &= ~CAM_PERIPH_NEW_DEV_FOUND;
386 
387 	if (periph->refcount == 0)
388 		camperiphfree(periph);
389 	else if (periph->refcount < 0)
390 		printf("cam_invalidate_periph: refcount < 0!!\n");
391 	splx(s);
392 }
393 
394 static void
395 camperiphfree(struct cam_periph *periph)
396 {
397 	int s;
398 	struct periph_driver **p_drv;
399 
400 	for (p_drv = (struct periph_driver **)periphdriver_set.ls_items;
401 	     *p_drv != NULL; p_drv++) {
402 		if (strcmp((*p_drv)->driver_name, periph->periph_name) == 0)
403 			break;
404 	}
405 
406 	if (periph->periph_dtor != NULL)
407 		periph->periph_dtor(periph);
408 
409 	s = splsoftcam();
410 	TAILQ_REMOVE(&(*p_drv)->units, periph, unit_links);
411 	(*p_drv)->generation++;
412 	splx(s);
413 
414 	xpt_remove_periph(periph);
415 
416 	if (periph->flags & CAM_PERIPH_NEW_DEV_FOUND) {
417 		union ccb ccb;
418 		void *arg;
419 
420 		switch (periph->deferred_ac) {
421 		case AC_FOUND_DEVICE:
422 			ccb.ccb_h.func_code = XPT_GDEV_TYPE;
423 			xpt_setup_ccb(&ccb.ccb_h, periph->path, /*priority*/ 1);
424 			xpt_action(&ccb);
425 			arg = &ccb;
426 			break;
427 		case AC_PATH_REGISTERED:
428 			ccb.ccb_h.func_code = XPT_PATH_INQ;
429 			xpt_setup_ccb(&ccb.ccb_h, periph->path, /*priority*/ 1);
430 			xpt_action(&ccb);
431 			arg = &ccb;
432 			break;
433 		default:
434 			arg = NULL;
435 			break;
436 		}
437 		periph->deferred_callback(NULL, periph->deferred_ac,
438 					  periph->path, arg);
439 	}
440 	xpt_free_path(periph->path);
441 	free(periph, M_DEVBUF);
442 }
443 
444 /*
445  * Wait interruptibly for an exclusive lock.
446  */
447 int
448 cam_periph_lock(struct cam_periph *periph, int priority)
449 {
450 	int error;
451 
452 	while ((periph->flags & CAM_PERIPH_LOCKED) != 0) {
453 		periph->flags |= CAM_PERIPH_LOCK_WANTED;
454 		if ((error = tsleep(periph, priority, "caplck", 0)) != 0)
455 			return error;
456 	}
457 
458 	if (cam_periph_acquire(periph) != CAM_REQ_CMP)
459 		return(ENXIO);
460 
461 	periph->flags |= CAM_PERIPH_LOCKED;
462 	return 0;
463 }
464 
465 /*
466  * Unlock and wake up any waiters.
467  */
468 void
469 cam_periph_unlock(struct cam_periph *periph)
470 {
471 	periph->flags &= ~CAM_PERIPH_LOCKED;
472 	if ((periph->flags & CAM_PERIPH_LOCK_WANTED) != 0) {
473 		periph->flags &= ~CAM_PERIPH_LOCK_WANTED;
474 		wakeup(periph);
475 	}
476 
477 	cam_periph_release(periph);
478 }
479 
480 /*
481  * Map user virtual pointers into kernel virtual address space, so we can
482  * access the memory.  This won't work on physical pointers, for now it's
483  * up to the caller to check for that.  (XXX KDM -- should we do that here
484  * instead?)  This also only works for up to MAXPHYS memory.  Since we use
485  * buffers to map stuff in and out, we're limited to the buffer size.
486  */
487 int
488 cam_periph_mapmem(union ccb *ccb, struct cam_periph_map_info *mapinfo)
489 {
490 	int numbufs, i;
491 	int flags[CAM_PERIPH_MAXMAPS];
492 	u_int8_t **data_ptrs[CAM_PERIPH_MAXMAPS];
493 	u_int32_t lengths[CAM_PERIPH_MAXMAPS];
494 	u_int32_t dirs[CAM_PERIPH_MAXMAPS];
495 
496 	switch(ccb->ccb_h.func_code) {
497 	case XPT_DEV_MATCH:
498 		if (ccb->cdm.match_buf_len == 0) {
499 			printf("cam_periph_mapmem: invalid match buffer "
500 			       "length 0\n");
501 			return(EINVAL);
502 		}
503 		if (ccb->cdm.pattern_buf_len > 0) {
504 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.patterns;
505 			lengths[0] = ccb->cdm.pattern_buf_len;
506 			dirs[0] = CAM_DIR_OUT;
507 			data_ptrs[1] = (u_int8_t **)&ccb->cdm.matches;
508 			lengths[1] = ccb->cdm.match_buf_len;
509 			dirs[1] = CAM_DIR_IN;
510 			numbufs = 2;
511 		} else {
512 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.matches;
513 			lengths[0] = ccb->cdm.match_buf_len;
514 			dirs[0] = CAM_DIR_IN;
515 			numbufs = 1;
516 		}
517 		break;
518 	case XPT_SCSI_IO:
519 	case XPT_CONT_TARGET_IO:
520 		if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
521 			return(0);
522 
523 		data_ptrs[0] = &ccb->csio.data_ptr;
524 		lengths[0] = ccb->csio.dxfer_len;
525 		dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
526 		numbufs = 1;
527 		break;
528 	default:
529 		return(EINVAL);
530 		break; /* NOTREACHED */
531 	}
532 
533 	/*
534 	 * Check the transfer length and permissions first, so we don't
535 	 * have to unmap any previously mapped buffers.
536 	 */
537 	for (i = 0; i < numbufs; i++) {
538 
539 		flags[i] = 0;
540 
541 		/*
542 		 * The userland data pointer passed in may not be page
543 		 * aligned.  vmapbuf() truncates the address to a page
544 		 * boundary, so if the address isn't page aligned, we'll
545 		 * need enough space for the given transfer length, plus
546 		 * whatever extra space is necessary to make it to the page
547 		 * boundary.
548 		 */
549 		if ((lengths[i] +
550 		    (((vm_offset_t)(*data_ptrs[i])) & PAGE_MASK)) > DFLTPHYS){
551 			printf("cam_periph_mapmem: attempt to map %lu bytes, "
552 			       "which is greater than DFLTPHYS(%d)\n",
553 			       (long)(lengths[i] +
554 			       (((vm_offset_t)(*data_ptrs[i])) & PAGE_MASK)),
555 			       DFLTPHYS);
556 			return(E2BIG);
557 		}
558 
559 		if (dirs[i] & CAM_DIR_OUT) {
560 			flags[i] = BIO_WRITE;
561 			if (!useracc(*data_ptrs[i], lengths[i],
562 				     VM_PROT_READ)) {
563 				printf("cam_periph_mapmem: error, "
564 					"address %p, length %lu isn't "
565 					"user accessible for READ\n",
566 					(void *)*data_ptrs[i],
567 					(u_long)lengths[i]);
568 				return(EACCES);
569 			}
570 		}
571 
572 		if (dirs[i] & CAM_DIR_IN) {
573 			flags[i] = BIO_READ;
574 			if (!useracc(*data_ptrs[i], lengths[i],
575 				     VM_PROT_WRITE)) {
576 				printf("cam_periph_mapmem: error, "
577 					"address %p, length %lu isn't "
578 					"user accessible for WRITE\n",
579 					(void *)*data_ptrs[i],
580 					(u_long)lengths[i]);
581 
582 				return(EACCES);
583 			}
584 		}
585 
586 	}
587 
588 	/* this keeps the current process from getting swapped */
589 	/*
590 	 * XXX KDM should I use P_NOSWAP instead?
591 	 */
592 	PHOLD(curproc);
593 
594 	for (i = 0; i < numbufs; i++) {
595 		/*
596 		 * Get the buffer.
597 		 */
598 		mapinfo->bp[i] = getpbuf(NULL);
599 
600 		/* save the buffer's data address */
601 		mapinfo->bp[i]->b_saveaddr = mapinfo->bp[i]->b_data;
602 
603 		/* put our pointer in the data slot */
604 		mapinfo->bp[i]->b_data = *data_ptrs[i];
605 
606 		/* set the transfer length, we know it's < DFLTPHYS */
607 		mapinfo->bp[i]->b_bufsize = lengths[i];
608 
609 		/* set the flags */
610 		mapinfo->bp[i]->b_flags = B_PHYS;
611 
612 		/* set the direction */
613 		mapinfo->bp[i]->b_iocmd = flags[i];
614 
615 		/* map the buffer into kernel memory */
616 		vmapbuf(mapinfo->bp[i]);
617 
618 		/* set our pointer to the new mapped area */
619 		*data_ptrs[i] = mapinfo->bp[i]->b_data;
620 
621 		mapinfo->num_bufs_used++;
622 	}
623 
624 	return(0);
625 }
626 
627 /*
628  * Unmap memory segments mapped into kernel virtual address space by
629  * cam_periph_mapmem().
630  */
631 void
632 cam_periph_unmapmem(union ccb *ccb, struct cam_periph_map_info *mapinfo)
633 {
634 	int numbufs, i;
635 	u_int8_t **data_ptrs[CAM_PERIPH_MAXMAPS];
636 
637 	if (mapinfo->num_bufs_used <= 0) {
638 		/* allow ourselves to be swapped once again */
639 		PRELE(curproc);
640 		return;
641 	}
642 
643 	switch (ccb->ccb_h.func_code) {
644 	case XPT_DEV_MATCH:
645 		numbufs = min(mapinfo->num_bufs_used, 2);
646 
647 		if (numbufs == 1) {
648 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.matches;
649 		} else {
650 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.patterns;
651 			data_ptrs[1] = (u_int8_t **)&ccb->cdm.matches;
652 		}
653 		break;
654 	case XPT_SCSI_IO:
655 	case XPT_CONT_TARGET_IO:
656 		data_ptrs[0] = &ccb->csio.data_ptr;
657 		numbufs = min(mapinfo->num_bufs_used, 1);
658 		break;
659 	default:
660 		/* allow ourselves to be swapped once again */
661 		PRELE(curproc);
662 		return;
663 		break; /* NOTREACHED */
664 	}
665 
666 	for (i = 0; i < numbufs; i++) {
667 		/* Set the user's pointer back to the original value */
668 		*data_ptrs[i] = mapinfo->bp[i]->b_saveaddr;
669 
670 		/* unmap the buffer */
671 		vunmapbuf(mapinfo->bp[i]);
672 
673 		/* clear the flags we set above */
674 		mapinfo->bp[i]->b_flags &= ~B_PHYS;
675 
676 		/* release the buffer */
677 		relpbuf(mapinfo->bp[i], NULL);
678 	}
679 
680 	/* allow ourselves to be swapped once again */
681 	PRELE(curproc);
682 }
683 
684 union ccb *
685 cam_periph_getccb(struct cam_periph *periph, u_int32_t priority)
686 {
687 	struct ccb_hdr *ccb_h;
688 	int s;
689 
690 	CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("entering cdgetccb\n"));
691 
692 	s = splsoftcam();
693 
694 	while (periph->ccb_list.slh_first == NULL) {
695 		if (periph->immediate_priority > priority)
696 			periph->immediate_priority = priority;
697 		xpt_schedule(periph, priority);
698 		if ((periph->ccb_list.slh_first != NULL)
699 		 && (periph->ccb_list.slh_first->pinfo.priority == priority))
700 			break;
701 		tsleep(&periph->ccb_list, PRIBIO, "cgticb", 0);
702 	}
703 
704 	ccb_h = periph->ccb_list.slh_first;
705 	SLIST_REMOVE_HEAD(&periph->ccb_list, periph_links.sle);
706 	splx(s);
707 	return ((union ccb *)ccb_h);
708 }
709 
710 void
711 cam_periph_ccbwait(union ccb *ccb)
712 {
713 	int s;
714 
715 	s = splsoftcam();
716 	if ((ccb->ccb_h.pinfo.index != CAM_UNQUEUED_INDEX)
717 	 || ((ccb->ccb_h.status & CAM_STATUS_MASK) == CAM_REQ_INPROG))
718 		tsleep(&ccb->ccb_h.cbfcnp, PRIBIO, "cbwait", 0);
719 
720 	splx(s);
721 }
722 
723 int
724 cam_periph_ioctl(struct cam_periph *periph, int cmd, caddr_t addr,
725 		 int (*error_routine)(union ccb *ccb,
726 				      cam_flags camflags,
727 				      u_int32_t sense_flags))
728 {
729 	union ccb 	     *ccb;
730 	int 		     error;
731 	int		     found;
732 
733 	error = found = 0;
734 
735 	switch(cmd){
736 	case CAMGETPASSTHRU:
737 		ccb = cam_periph_getccb(periph, /* priority */ 1);
738 		xpt_setup_ccb(&ccb->ccb_h,
739 			      ccb->ccb_h.path,
740 			      /*priority*/1);
741 		ccb->ccb_h.func_code = XPT_GDEVLIST;
742 
743 		/*
744 		 * Basically, the point of this is that we go through
745 		 * getting the list of devices, until we find a passthrough
746 		 * device.  In the current version of the CAM code, the
747 		 * only way to determine what type of device we're dealing
748 		 * with is by its name.
749 		 */
750 		while (found == 0) {
751 			ccb->cgdl.index = 0;
752 			ccb->cgdl.status = CAM_GDEVLIST_MORE_DEVS;
753 			while (ccb->cgdl.status == CAM_GDEVLIST_MORE_DEVS) {
754 
755 				/* we want the next device in the list */
756 				xpt_action(ccb);
757 				if (strncmp(ccb->cgdl.periph_name,
758 				    "pass", 4) == 0){
759 					found = 1;
760 					break;
761 				}
762 			}
763 			if ((ccb->cgdl.status == CAM_GDEVLIST_LAST_DEVICE) &&
764 			    (found == 0)) {
765 				ccb->cgdl.periph_name[0] = '\0';
766 				ccb->cgdl.unit_number = 0;
767 				break;
768 			}
769 		}
770 
771 		/* copy the result back out */
772 		bcopy(ccb, addr, sizeof(union ccb));
773 
774 		/* and release the ccb */
775 		xpt_release_ccb(ccb);
776 
777 		break;
778 	default:
779 		error = ENOTTY;
780 		break;
781 	}
782 	return(error);
783 }
784 
785 int
786 cam_periph_runccb(union ccb *ccb,
787 		  int (*error_routine)(union ccb *ccb,
788 				       cam_flags camflags,
789 				       u_int32_t sense_flags),
790 		  cam_flags camflags, u_int32_t sense_flags,
791 		  struct devstat *ds)
792 {
793 	int error;
794 
795 	error = 0;
796 
797 	/*
798 	 * If the user has supplied a stats structure, and if we understand
799 	 * this particular type of ccb, record the transaction start.
800 	 */
801 	if ((ds != NULL) && (ccb->ccb_h.func_code == XPT_SCSI_IO))
802 		devstat_start_transaction(ds);
803 
804 	xpt_action(ccb);
805 
806 	do {
807 		cam_periph_ccbwait(ccb);
808 		if ((ccb->ccb_h.status & CAM_STATUS_MASK) == CAM_REQ_CMP)
809 			error = 0;
810 		else if (error_routine != NULL)
811 			error = (*error_routine)(ccb, camflags, sense_flags);
812 		else
813 			error = 0;
814 
815 	} while (error == ERESTART);
816 
817 	if ((ccb->ccb_h.status & CAM_DEV_QFRZN) != 0)
818 		cam_release_devq(ccb->ccb_h.path,
819 				 /* relsim_flags */0,
820 				 /* openings */0,
821 				 /* timeout */0,
822 				 /* getcount_only */ FALSE);
823 
824 	if ((ds != NULL) && (ccb->ccb_h.func_code == XPT_SCSI_IO))
825 		devstat_end_transaction(ds,
826 					ccb->csio.dxfer_len,
827 					ccb->csio.tag_action & 0xf,
828 					((ccb->ccb_h.flags & CAM_DIR_MASK) ==
829 					CAM_DIR_NONE) ?  DEVSTAT_NO_DATA :
830 					(ccb->ccb_h.flags & CAM_DIR_OUT) ?
831 					DEVSTAT_WRITE :
832 					DEVSTAT_READ);
833 
834 	return(error);
835 }
836 
837 void
838 cam_freeze_devq(struct cam_path *path)
839 {
840 	struct ccb_hdr ccb_h;
841 
842 	xpt_setup_ccb(&ccb_h, path, /*priority*/1);
843 	ccb_h.func_code = XPT_NOOP;
844 	ccb_h.flags = CAM_DEV_QFREEZE;
845 	xpt_action((union ccb *)&ccb_h);
846 }
847 
848 u_int32_t
849 cam_release_devq(struct cam_path *path, u_int32_t relsim_flags,
850 		 u_int32_t openings, u_int32_t timeout,
851 		 int getcount_only)
852 {
853 	struct ccb_relsim crs;
854 
855 	xpt_setup_ccb(&crs.ccb_h, path,
856 		      /*priority*/1);
857 	crs.ccb_h.func_code = XPT_REL_SIMQ;
858 	crs.ccb_h.flags = getcount_only ? CAM_DEV_QFREEZE : 0;
859 	crs.release_flags = relsim_flags;
860 	crs.openings = openings;
861 	crs.release_timeout = timeout;
862 	xpt_action((union ccb *)&crs);
863 	return (crs.qfrozen_cnt);
864 }
865 
866 #define saved_ccb_ptr ppriv_ptr0
867 static void
868 camperiphdone(struct cam_periph *periph, union ccb *done_ccb)
869 {
870 	cam_status	status;
871 	int		frozen;
872 	int		sense;
873 	struct scsi_start_stop_unit *scsi_cmd;
874 	u_int32_t	relsim_flags, timeout;
875 	u_int32_t	qfrozen_cnt;
876 
877 	status = done_ccb->ccb_h.status;
878 	frozen = (status & CAM_DEV_QFRZN) != 0;
879 	sense  = (status & CAM_AUTOSNS_VALID) != 0;
880 	status &= CAM_STATUS_MASK;
881 
882 	timeout = 0;
883 	relsim_flags = 0;
884 
885 	/*
886 	 * Unfreeze the queue once if it is already frozen..
887 	 */
888 	if (frozen != 0) {
889 		qfrozen_cnt = cam_release_devq(done_ccb->ccb_h.path,
890 					      /*relsim_flags*/0,
891 					      /*openings*/0,
892 					      /*timeout*/0,
893 					      /*getcount_only*/0);
894 	}
895 
896 	switch (status) {
897 
898 	case CAM_REQ_CMP:
899 
900 		/*
901 		 * If we have successfully taken a device from the not
902 		 * ready to ready state, re-scan the device and re-get the
903 		 * inquiry information.  Many devices (mostly disks) don't
904 		 * properly report their inquiry information unless they
905 		 * are spun up.
906 		 */
907 		if (done_ccb->ccb_h.func_code == XPT_SCSI_IO) {
908 			scsi_cmd = (struct scsi_start_stop_unit *)
909 					&done_ccb->csio.cdb_io.cdb_bytes;
910 
911 		 	if (scsi_cmd->opcode == START_STOP_UNIT)
912 				xpt_async(AC_INQ_CHANGED,
913 					  done_ccb->ccb_h.path, NULL);
914 		}
915 		bcopy(done_ccb->ccb_h.saved_ccb_ptr, done_ccb,
916 		      sizeof(union ccb));
917 
918 		periph->flags &= ~CAM_PERIPH_RECOVERY_INPROG;
919 
920 		xpt_action(done_ccb);
921 
922 		break;
923 	case CAM_SCSI_STATUS_ERROR:
924 		scsi_cmd = (struct scsi_start_stop_unit *)
925 				&done_ccb->csio.cdb_io.cdb_bytes;
926 		if (sense != 0) {
927 			struct scsi_sense_data *sense;
928 			int    error_code, sense_key, asc, ascq;
929 
930 			sense = &done_ccb->csio.sense_data;
931 			scsi_extract_sense(sense, &error_code,
932 					   &sense_key, &asc, &ascq);
933 
934 			/*
935 	 		 * If the error is "invalid field in CDB",
936 			 * and the load/eject flag is set, turn the
937 			 * flag off and try again.  This is just in
938 			 * case the drive in question barfs on the
939 			 * load eject flag.  The CAM code should set
940 			 * the load/eject flag by default for
941 			 * removable media.
942 			 */
943 
944 			/* XXX KDM
945 			 * Should we check to see what the specific
946 			 * scsi status is??  Or does it not matter
947 			 * since we already know that there was an
948 			 * error, and we know what the specific
949 			 * error code was, and we know what the
950 			 * opcode is..
951 			 */
952 			if ((scsi_cmd->opcode == START_STOP_UNIT) &&
953 			    ((scsi_cmd->how & SSS_LOEJ) != 0) &&
954 			     (asc == 0x24) && (ascq == 0x00) &&
955 			     (done_ccb->ccb_h.retry_count > 0)) {
956 
957 				scsi_cmd->how &= ~SSS_LOEJ;
958 
959 				xpt_action(done_ccb);
960 
961 			} else if (done_ccb->ccb_h.retry_count > 0) {
962 				/*
963 				 * In this case, the error recovery
964 				 * command failed, but we've got
965 				 * some retries left on it.  Give
966 				 * it another try.
967 				 */
968 
969 				/* set the timeout to .5 sec */
970 				relsim_flags =
971 					RELSIM_RELEASE_AFTER_TIMEOUT;
972 				timeout = 500;
973 
974 				xpt_action(done_ccb);
975 
976 				break;
977 
978 			} else {
979 				/*
980 				 * Copy the original CCB back and
981 				 * send it back to the caller.
982 				 */
983 				bcopy(done_ccb->ccb_h.saved_ccb_ptr,
984 				      done_ccb, sizeof(union ccb));
985 
986 				periph->flags &= ~CAM_PERIPH_RECOVERY_INPROG;
987 
988 				xpt_action(done_ccb);
989 			}
990 		} else {
991 			/*
992 			 * Eh??  The command failed, but we don't
993 			 * have any sense.  What's up with that?
994 			 * Fire the CCB again to return it to the
995 			 * caller.
996 			 */
997 			bcopy(done_ccb->ccb_h.saved_ccb_ptr,
998 			      done_ccb, sizeof(union ccb));
999 
1000 			periph->flags &= ~CAM_PERIPH_RECOVERY_INPROG;
1001 
1002 			xpt_action(done_ccb);
1003 
1004 		}
1005 		break;
1006 	default:
1007 		bcopy(done_ccb->ccb_h.saved_ccb_ptr, done_ccb,
1008 		      sizeof(union ccb));
1009 
1010 		periph->flags &= ~CAM_PERIPH_RECOVERY_INPROG;
1011 
1012 		xpt_action(done_ccb);
1013 
1014 		break;
1015 	}
1016 
1017 	/* decrement the retry count */
1018 	if (done_ccb->ccb_h.retry_count > 0)
1019 		done_ccb->ccb_h.retry_count--;
1020 
1021 	qfrozen_cnt = cam_release_devq(done_ccb->ccb_h.path,
1022 				      /*relsim_flags*/relsim_flags,
1023 				      /*openings*/0,
1024 				      /*timeout*/timeout,
1025 				      /*getcount_only*/0);
1026 }
1027 
1028 /*
1029  * Generic Async Event handler.  Peripheral drivers usually
1030  * filter out the events that require personal attention,
1031  * and leave the rest to this function.
1032  */
1033 void
1034 cam_periph_async(struct cam_periph *periph, u_int32_t code,
1035 		 struct cam_path *path, void *arg)
1036 {
1037 	switch (code) {
1038 	case AC_LOST_DEVICE:
1039 		cam_periph_invalidate(periph);
1040 		break;
1041 	case AC_SENT_BDR:
1042 	case AC_BUS_RESET:
1043 	{
1044 		cam_periph_bus_settle(periph, SCSI_DELAY);
1045 		break;
1046 	}
1047 	default:
1048 		break;
1049 	}
1050 }
1051 
1052 void
1053 cam_periph_bus_settle(struct cam_periph *periph, u_int bus_settle)
1054 {
1055 	struct ccb_getdevstats cgds;
1056 
1057 	xpt_setup_ccb(&cgds.ccb_h, periph->path, /*priority*/1);
1058 	cgds.ccb_h.func_code = XPT_GDEV_STATS;
1059 	xpt_action((union ccb *)&cgds);
1060 	cam_periph_freeze_after_event(periph, &cgds.last_reset, bus_settle);
1061 }
1062 
1063 void
1064 cam_periph_freeze_after_event(struct cam_periph *periph,
1065 			      struct timeval* event_time, u_int duration_ms)
1066 {
1067 	struct timeval delta;
1068 	struct timeval duration_tv;
1069 	int s;
1070 
1071 	s = splclock();
1072 	microtime(&delta);
1073 	splx(s);
1074 	timevalsub(&delta, event_time);
1075 	duration_tv.tv_sec = duration_ms / 1000;
1076 	duration_tv.tv_usec = (duration_ms % 1000) * 1000;
1077 	if (timevalcmp(&delta, &duration_tv, <)) {
1078 		timevalsub(&duration_tv, &delta);
1079 
1080 		duration_ms = duration_tv.tv_sec * 1000;
1081 		duration_ms += duration_tv.tv_usec / 1000;
1082 		cam_freeze_devq(periph->path);
1083 		cam_release_devq(periph->path,
1084 				RELSIM_RELEASE_AFTER_TIMEOUT,
1085 				/*reduction*/0,
1086 				/*timeout*/duration_ms,
1087 				/*getcount_only*/0);
1088 	}
1089 
1090 }
1091 
1092 /*
1093  * Generic error handler.  Peripheral drivers usually filter
1094  * out the errors that they handle in a unique mannor, then
1095  * call this function.
1096  */
1097 int
1098 cam_periph_error(union ccb *ccb, cam_flags camflags,
1099 		 u_int32_t sense_flags, union ccb *save_ccb)
1100 {
1101 	cam_status status;
1102 	int	   frozen;
1103 	int	   sense;
1104 	int	   error;
1105 	int        openings;
1106 	int	   retry;
1107 	u_int32_t  relsim_flags;
1108 	u_int32_t  timeout;
1109 
1110 	status = ccb->ccb_h.status;
1111 	frozen = (status & CAM_DEV_QFRZN) != 0;
1112 	sense  = (status & CAM_AUTOSNS_VALID) != 0;
1113 	status &= CAM_STATUS_MASK;
1114 	relsim_flags = 0;
1115 
1116 	switch (status) {
1117 	case CAM_REQ_CMP:
1118 		/* decrement the number of retries */
1119 		retry = ccb->ccb_h.retry_count > 0;
1120 		if (retry)
1121 			ccb->ccb_h.retry_count--;
1122 		error = 0;
1123 		break;
1124 	case CAM_AUTOSENSE_FAIL:
1125 	case CAM_SCSI_STATUS_ERROR:
1126 
1127 		switch (ccb->csio.scsi_status) {
1128 		case SCSI_STATUS_OK:
1129 		case SCSI_STATUS_COND_MET:
1130 		case SCSI_STATUS_INTERMED:
1131 		case SCSI_STATUS_INTERMED_COND_MET:
1132 			error = 0;
1133 			break;
1134 		case SCSI_STATUS_CMD_TERMINATED:
1135 		case SCSI_STATUS_CHECK_COND:
1136 			if (sense != 0) {
1137 				struct scsi_sense_data *sense;
1138 				int    error_code, sense_key, asc, ascq;
1139 				struct cam_periph *periph;
1140 				scsi_sense_action err_action;
1141 				struct ccb_getdev cgd;
1142 
1143 				sense = &ccb->csio.sense_data;
1144 				scsi_extract_sense(sense, &error_code,
1145 						   &sense_key, &asc, &ascq);
1146 				periph = xpt_path_periph(ccb->ccb_h.path);
1147 
1148 				/*
1149 				 * Grab the inquiry data for this device.
1150 				 */
1151 				xpt_setup_ccb(&cgd.ccb_h, ccb->ccb_h.path,
1152 					      /*priority*/ 1);
1153 				cgd.ccb_h.func_code = XPT_GDEV_TYPE;
1154 				xpt_action((union ccb *)&cgd);
1155 
1156 				err_action = scsi_error_action(asc, ascq,
1157 							       &cgd.inq_data);
1158 
1159 				/*
1160 				 * Send a Test Unit Ready to the device.
1161 				 * If the 'many' flag is set, we send 120
1162 				 * test unit ready commands, one every half
1163 				 * second.  Otherwise, we just send one TUR.
1164 				 * We only want to do this if the retry
1165 				 * count has not been exhausted.
1166 				 */
1167 				if (((err_action & SS_MASK) == SS_TUR)
1168 				 && save_ccb != NULL
1169 				 && ccb->ccb_h.retry_count > 0) {
1170 
1171 					/*
1172 					 * Since error recovery is already
1173 					 * in progress, don't attempt to
1174 					 * process this error.  It is probably
1175 					 * related to the error that caused
1176 					 * the currently active error recovery
1177 					 * action.  Also, we only have
1178 					 * space for one saved CCB, so if we
1179 					 * had two concurrent error recovery
1180 					 * actions, we would end up
1181 					 * over-writing one error recovery
1182 					 * CCB with another one.
1183 					 */
1184 					if (periph->flags &
1185 					    CAM_PERIPH_RECOVERY_INPROG) {
1186 						error = ERESTART;
1187 						break;
1188 					}
1189 
1190 					periph->flags |=
1191 						CAM_PERIPH_RECOVERY_INPROG;
1192 
1193 					/* decrement the number of retries */
1194 					if ((err_action &
1195 					     SSQ_DECREMENT_COUNT) != 0) {
1196 						retry = 1;
1197 						ccb->ccb_h.retry_count--;
1198 					}
1199 
1200 					bcopy(ccb, save_ccb, sizeof(*save_ccb));
1201 
1202 					/*
1203 					 * We retry this one every half
1204 					 * second for a minute.  If the
1205 					 * device hasn't become ready in a
1206 					 * minute's time, it's unlikely to
1207 					 * ever become ready.  If the table
1208 					 * doesn't specify SSQ_MANY, we can
1209 					 * only try this once.  Oh well.
1210 					 */
1211 					if ((err_action & SSQ_MANY) != 0)
1212 						scsi_test_unit_ready(&ccb->csio,
1213 							       /*retries*/120,
1214 							       camperiphdone,
1215 						 	       MSG_SIMPLE_Q_TAG,
1216 							       SSD_FULL_SIZE,
1217 							       /*timeout*/5000);
1218 					else
1219 						scsi_test_unit_ready(&ccb->csio,
1220 							       /*retries*/1,
1221 							       camperiphdone,
1222 						 	       MSG_SIMPLE_Q_TAG,
1223 							       SSD_FULL_SIZE,
1224 							       /*timeout*/5000);
1225 
1226 					/* release the queue after .5 sec.  */
1227 					relsim_flags =
1228 						RELSIM_RELEASE_AFTER_TIMEOUT;
1229 					timeout = 500;
1230 					/*
1231 					 * Drop the priority to 0 so that
1232 					 * we are the first to execute.  Also
1233 					 * freeze the queue after this command
1234 					 * is sent so that we can restore the
1235 					 * old csio and have it queued in the
1236 					 * proper order before we let normal
1237 					 * transactions go to the drive.
1238 					 */
1239 					ccb->ccb_h.pinfo.priority = 0;
1240 					ccb->ccb_h.flags |= CAM_DEV_QFREEZE;
1241 
1242 					/*
1243 					 * Save a pointer to the original
1244 					 * CCB in the new CCB.
1245 					 */
1246 					ccb->ccb_h.saved_ccb_ptr = save_ccb;
1247 
1248 					error = ERESTART;
1249 				}
1250 				/*
1251 				 * Send a start unit command to the device,
1252 				 * and then retry the command.  We only
1253 				 * want to do this if the retry count has
1254 				 * not been exhausted.  If the user
1255 				 * specified 0 retries, then we follow
1256 				 * their request and do not retry.
1257 				 */
1258 				else if (((err_action & SS_MASK) == SS_START)
1259 				      && save_ccb != NULL
1260 				      && ccb->ccb_h.retry_count > 0) {
1261 					int le;
1262 
1263 					/*
1264 					 * Only one error recovery action
1265 					 * at a time.  See above.
1266 					 */
1267 					if (periph->flags &
1268 					    CAM_PERIPH_RECOVERY_INPROG) {
1269 						error = ERESTART;
1270 						break;
1271 					}
1272 
1273 					periph->flags |=
1274 						CAM_PERIPH_RECOVERY_INPROG;
1275 
1276 					/* decrement the number of retries */
1277 					retry = 1;
1278 					ccb->ccb_h.retry_count--;
1279 
1280 					/*
1281 					 * Check for removable media and
1282 					 * set load/eject flag
1283 					 * appropriately.
1284 					 */
1285 					if (SID_IS_REMOVABLE(&cgd.inq_data))
1286 						le = TRUE;
1287 					else
1288 						le = FALSE;
1289 
1290 					/*
1291 					 * Attempt to start the drive up.
1292 					 *
1293 					 * Save the current ccb so it can
1294 					 * be restored and retried once the
1295 					 * drive is started up.
1296 					 */
1297 					bcopy(ccb, save_ccb, sizeof(*save_ccb));
1298 
1299 					scsi_start_stop(&ccb->csio,
1300 							/*retries*/1,
1301 							camperiphdone,
1302 							MSG_SIMPLE_Q_TAG,
1303 							/*start*/TRUE,
1304 							/*load/eject*/le,
1305 							/*immediate*/FALSE,
1306 							SSD_FULL_SIZE,
1307 							/*timeout*/50000);
1308 					/*
1309 					 * Drop the priority to 0 so that
1310 					 * we are the first to execute.  Also
1311 					 * freeze the queue after this command
1312 					 * is sent so that we can restore the
1313 					 * old csio and have it queued in the
1314 					 * proper order before we let normal
1315 					 * transactions go to the drive.
1316 					 */
1317 					ccb->ccb_h.pinfo.priority = 0;
1318 					ccb->ccb_h.flags |= CAM_DEV_QFREEZE;
1319 
1320 					/*
1321 					 * Save a pointer to the original
1322 					 * CCB in the new CCB.
1323 					 */
1324 					ccb->ccb_h.saved_ccb_ptr = save_ccb;
1325 
1326 					error = ERESTART;
1327 				} else if ((sense_flags & SF_RETRY_UA) != 0) {
1328 					/*
1329 					 * XXX KDM this is a *horrible*
1330 					 * hack.
1331 					 */
1332 					error = scsi_interpret_sense(ccb,
1333 								  sense_flags,
1334 								  &relsim_flags,
1335 								  &openings,
1336 								  &timeout,
1337 								  err_action);
1338 				}
1339 
1340 				/*
1341 				 * Theoretically, this code should send a
1342 				 * test unit ready to the given device, and
1343 				 * if it returns and error, send a start
1344 				 * unit command.  Since we don't yet have
1345 				 * the capability to do two-command error
1346 				 * recovery, just send a start unit.
1347 				 * XXX KDM fix this!
1348 				 */
1349 				else if (((err_action & SS_MASK) == SS_TURSTART)
1350 				      && save_ccb != NULL
1351 				      && ccb->ccb_h.retry_count > 0) {
1352 					int le;
1353 
1354 					/*
1355 					 * Only one error recovery action
1356 					 * at a time.  See above.
1357 					 */
1358 					if (periph->flags &
1359 					    CAM_PERIPH_RECOVERY_INPROG) {
1360 						error = ERESTART;
1361 						break;
1362 					}
1363 
1364 					periph->flags |=
1365 						CAM_PERIPH_RECOVERY_INPROG;
1366 
1367 					/* decrement the number of retries */
1368 					retry = 1;
1369 					ccb->ccb_h.retry_count--;
1370 
1371 					/*
1372 					 * Check for removable media and
1373 					 * set load/eject flag
1374 					 * appropriately.
1375 					 */
1376 					if (SID_IS_REMOVABLE(&cgd.inq_data))
1377 						le = TRUE;
1378 					else
1379 						le = FALSE;
1380 
1381 					/*
1382 					 * Attempt to start the drive up.
1383 					 *
1384 					 * Save the current ccb so it can
1385 					 * be restored and retried once the
1386 					 * drive is started up.
1387 					 */
1388 					bcopy(ccb, save_ccb, sizeof(*save_ccb));
1389 
1390 					scsi_start_stop(&ccb->csio,
1391 							/*retries*/1,
1392 							camperiphdone,
1393 							MSG_SIMPLE_Q_TAG,
1394 							/*start*/TRUE,
1395 							/*load/eject*/le,
1396 							/*immediate*/FALSE,
1397 							SSD_FULL_SIZE,
1398 							/*timeout*/50000);
1399 
1400 					/* release the queue after .5 sec.  */
1401 					relsim_flags =
1402 						RELSIM_RELEASE_AFTER_TIMEOUT;
1403 					timeout = 500;
1404 					/*
1405 					 * Drop the priority to 0 so that
1406 					 * we are the first to execute.  Also
1407 					 * freeze the queue after this command
1408 					 * is sent so that we can restore the
1409 					 * old csio and have it queued in the
1410 					 * proper order before we let normal
1411 					 * transactions go to the drive.
1412 					 */
1413 					ccb->ccb_h.pinfo.priority = 0;
1414 					ccb->ccb_h.flags |= CAM_DEV_QFREEZE;
1415 
1416 					/*
1417 					 * Save a pointer to the original
1418 					 * CCB in the new CCB.
1419 					 */
1420 					ccb->ccb_h.saved_ccb_ptr = save_ccb;
1421 
1422 					error = ERESTART;
1423 				} else {
1424 					error = scsi_interpret_sense(ccb,
1425 								  sense_flags,
1426 								  &relsim_flags,
1427 								  &openings,
1428 								  &timeout,
1429 								  err_action);
1430 				}
1431 			} else if (ccb->csio.scsi_status ==
1432 				   SCSI_STATUS_CHECK_COND
1433 				&& status != CAM_AUTOSENSE_FAIL) {
1434 				/* no point in decrementing the retry count */
1435 				panic("cam_periph_error: scsi status of "
1436 				      "CHECK COND returned but no sense "
1437 				      "information is availible.  "
1438 				      "Controller should have returned "
1439 				      "CAM_AUTOSENSE_FAILED");
1440 				/* NOTREACHED */
1441 				error = EIO;
1442 			} else if (ccb->ccb_h.retry_count == 0) {
1443 				/*
1444 				 * XXX KDM shouldn't there be a better
1445 				 * argument to return??
1446 				 */
1447 				error = EIO;
1448 			} else {
1449 				/* decrement the number of retries */
1450 				retry = ccb->ccb_h.retry_count > 0;
1451 				if (retry)
1452 					ccb->ccb_h.retry_count--;
1453 				/*
1454 				 * If it was aborted with no
1455 				 * clue as to the reason, just
1456 				 * retry it again.
1457 				 */
1458 				error = ERESTART;
1459 			}
1460 			break;
1461 		case SCSI_STATUS_QUEUE_FULL:
1462 		{
1463 			/* no decrement */
1464 			struct ccb_getdevstats cgds;
1465 
1466 			/*
1467 			 * First off, find out what the current
1468 			 * transaction counts are.
1469 			 */
1470 			xpt_setup_ccb(&cgds.ccb_h,
1471 				      ccb->ccb_h.path,
1472 				      /*priority*/1);
1473 			cgds.ccb_h.func_code = XPT_GDEV_STATS;
1474 			xpt_action((union ccb *)&cgds);
1475 
1476 			/*
1477 			 * If we were the only transaction active, treat
1478 			 * the QUEUE FULL as if it were a BUSY condition.
1479 			 */
1480 			if (cgds.dev_active != 0) {
1481 				int total_openings;
1482 
1483 				/*
1484 			 	 * Reduce the number of openings to
1485 				 * be 1 less than the amount it took
1486 				 * to get a queue full bounded by the
1487 				 * minimum allowed tag count for this
1488 				 * device.
1489 			 	 */
1490 				total_openings =
1491 				    cgds.dev_active+cgds.dev_openings;
1492 				openings = cgds.dev_active;
1493 				if (openings < cgds.mintags)
1494 					openings = cgds.mintags;
1495 				if (openings < total_openings)
1496 					relsim_flags = RELSIM_ADJUST_OPENINGS;
1497 				else {
1498 					/*
1499 					 * Some devices report queue full for
1500 					 * temporary resource shortages.  For
1501 					 * this reason, we allow a minimum
1502 					 * tag count to be entered via a
1503 					 * quirk entry to prevent the queue
1504 					 * count on these devices from falling
1505 					 * to a pessimisticly low value.  We
1506 					 * still wait for the next successful
1507 					 * completion, however, before queueing
1508 					 * more transactions to the device.
1509 					 */
1510 					relsim_flags =
1511 					    RELSIM_RELEASE_AFTER_CMDCMPLT;
1512 				}
1513 				timeout = 0;
1514 				error = ERESTART;
1515 				break;
1516 			}
1517 			/* FALLTHROUGH */
1518 		}
1519 		case SCSI_STATUS_BUSY:
1520 			/*
1521 			 * Restart the queue after either another
1522 			 * command completes or a 1 second timeout.
1523 			 * If we have any retries left, that is.
1524 			 */
1525 			retry = ccb->ccb_h.retry_count > 0;
1526 			if (retry) {
1527 				ccb->ccb_h.retry_count--;
1528 				error = ERESTART;
1529 				relsim_flags = RELSIM_RELEASE_AFTER_TIMEOUT
1530 					     | RELSIM_RELEASE_AFTER_CMDCMPLT;
1531 				timeout = 1000;
1532 			} else {
1533 				error = EIO;
1534 			}
1535 			break;
1536 		case SCSI_STATUS_RESERV_CONFLICT:
1537 			error = EIO;
1538 			break;
1539 		default:
1540 			error = EIO;
1541 			break;
1542 		}
1543 		break;
1544 	case CAM_REQ_CMP_ERR:
1545 	case CAM_CMD_TIMEOUT:
1546 	case CAM_UNEXP_BUSFREE:
1547 	case CAM_UNCOR_PARITY:
1548 	case CAM_DATA_RUN_ERR:
1549 		/* decrement the number of retries */
1550 		retry = ccb->ccb_h.retry_count > 0;
1551 		if (retry) {
1552 			ccb->ccb_h.retry_count--;
1553 			error = ERESTART;
1554 		} else {
1555 			error = EIO;
1556 		}
1557 		break;
1558 	case CAM_UA_ABORT:
1559 	case CAM_UA_TERMIO:
1560 	case CAM_MSG_REJECT_REC:
1561 		/* XXX Don't know that these are correct */
1562 		error = EIO;
1563 		break;
1564 	case CAM_SEL_TIMEOUT:
1565 	{
1566 		/*
1567 		 * XXX
1568 		 * A single selection timeout should not be enough
1569 		 * to invalidate a device.  We should retry for multiple
1570 		 * seconds assuming this isn't a probe.  We'll probably
1571 		 * need a special flag for that.
1572 		 */
1573 #if 0
1574 		struct cam_path *newpath;
1575 
1576 		/* Should we do more if we can't create the path?? */
1577 		if (xpt_create_path(&newpath, xpt_path_periph(ccb->ccb_h.path),
1578 				    xpt_path_path_id(ccb->ccb_h.path),
1579 				    xpt_path_target_id(ccb->ccb_h.path),
1580 				    CAM_LUN_WILDCARD) != CAM_REQ_CMP)
1581 			break;
1582 		/*
1583 		 * Let peripheral drivers know that this device has gone
1584 		 * away.
1585 		 */
1586 		xpt_async(AC_LOST_DEVICE, newpath, NULL);
1587 		xpt_free_path(newpath);
1588 #endif
1589 		if ((sense_flags & SF_RETRY_SELTO) != 0) {
1590 			retry = ccb->ccb_h.retry_count > 0;
1591 			if (retry) {
1592 				ccb->ccb_h.retry_count--;
1593 				error = ERESTART;
1594 				/*
1595 				 * Wait half a second to give the device
1596 				 * time to recover before we try again.
1597 				 */
1598 				relsim_flags = RELSIM_RELEASE_AFTER_TIMEOUT;
1599 				timeout = 500;
1600 			} else {
1601 				error = ENXIO;
1602 			}
1603 		} else {
1604 			error = ENXIO;
1605 		}
1606 		break;
1607 	}
1608 	case CAM_REQ_INVALID:
1609 	case CAM_PATH_INVALID:
1610 	case CAM_DEV_NOT_THERE:
1611 	case CAM_NO_HBA:
1612 	case CAM_PROVIDE_FAIL:
1613 	case CAM_REQ_TOO_BIG:
1614 		error = EINVAL;
1615 		break;
1616 	case CAM_SCSI_BUS_RESET:
1617 	case CAM_BDR_SENT:
1618 	case CAM_REQUEUE_REQ:
1619 		/* Unconditional requeue, dammit */
1620 		error = ERESTART;
1621 		break;
1622 	case CAM_RESRC_UNAVAIL:
1623 	case CAM_BUSY:
1624 		/* timeout??? */
1625 	default:
1626 		/* decrement the number of retries */
1627 		retry = ccb->ccb_h.retry_count > 0;
1628 		if (retry) {
1629 			ccb->ccb_h.retry_count--;
1630 			error = ERESTART;
1631 		} else {
1632 			/* Check the sense codes */
1633 			error = EIO;
1634 		}
1635 		break;
1636 	}
1637 
1638 	/* Attempt a retry */
1639 	if (error == ERESTART || error == 0) {
1640 		if (frozen != 0)
1641 			ccb->ccb_h.status &= ~CAM_DEV_QFRZN;
1642 
1643 		if (error == ERESTART)
1644 			xpt_action(ccb);
1645 
1646 		if (frozen != 0) {
1647 			cam_release_devq(ccb->ccb_h.path,
1648 					 relsim_flags,
1649 					 openings,
1650 					 timeout,
1651 					 /*getcount_only*/0);
1652 		}
1653 	}
1654 
1655 
1656 	return (error);
1657 }
1658