xref: /freebsd/sys/cam/scsi/scsi_pass.c (revision 834063202a16592e1ef5c3a9fbd04ca5f1df3ed0)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 1997, 1998, 2000 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 
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
32 
33 #include <sys/param.h>
34 #include <sys/systm.h>
35 #include <sys/kernel.h>
36 #include <sys/conf.h>
37 #include <sys/types.h>
38 #include <sys/bio.h>
39 #include <sys/bus.h>
40 #include <sys/devicestat.h>
41 #include <sys/errno.h>
42 #include <sys/fcntl.h>
43 #include <sys/malloc.h>
44 #include <sys/proc.h>
45 #include <sys/poll.h>
46 #include <sys/selinfo.h>
47 #include <sys/sdt.h>
48 #include <sys/taskqueue.h>
49 #include <vm/uma.h>
50 #include <vm/vm.h>
51 #include <vm/vm_extern.h>
52 
53 #include <machine/bus.h>
54 
55 #include <cam/cam.h>
56 #include <cam/cam_ccb.h>
57 #include <cam/cam_periph.h>
58 #include <cam/cam_queue.h>
59 #include <cam/cam_xpt.h>
60 #include <cam/cam_xpt_periph.h>
61 #include <cam/cam_debug.h>
62 #include <cam/cam_compat.h>
63 #include <cam/cam_xpt_periph.h>
64 
65 #include <cam/scsi/scsi_all.h>
66 #include <cam/scsi/scsi_pass.h>
67 
68 typedef enum {
69 	PASS_FLAG_OPEN			= 0x01,
70 	PASS_FLAG_LOCKED		= 0x02,
71 	PASS_FLAG_INVALID		= 0x04,
72 	PASS_FLAG_INITIAL_PHYSPATH	= 0x08,
73 	PASS_FLAG_ZONE_INPROG		= 0x10,
74 	PASS_FLAG_ZONE_VALID		= 0x20,
75 	PASS_FLAG_UNMAPPED_CAPABLE	= 0x40,
76 	PASS_FLAG_ABANDONED_REF_SET	= 0x80
77 } pass_flags;
78 
79 typedef enum {
80 	PASS_STATE_NORMAL
81 } pass_state;
82 
83 typedef enum {
84 	PASS_CCB_BUFFER_IO,
85 	PASS_CCB_QUEUED_IO
86 } pass_ccb_types;
87 
88 #define ccb_type	ppriv_field0
89 #define ccb_ioreq	ppriv_ptr1
90 
91 /*
92  * The maximum number of memory segments we preallocate.
93  */
94 #define	PASS_MAX_SEGS	16
95 
96 typedef enum {
97 	PASS_IO_NONE		= 0x00,
98 	PASS_IO_USER_SEG_MALLOC	= 0x01,
99 	PASS_IO_KERN_SEG_MALLOC	= 0x02,
100 	PASS_IO_ABANDONED	= 0x04
101 } pass_io_flags;
102 
103 struct pass_io_req {
104 	union ccb			 ccb;
105 	union ccb			*alloced_ccb;
106 	union ccb			*user_ccb_ptr;
107 	camq_entry			 user_periph_links;
108 	ccb_ppriv_area			 user_periph_priv;
109 	struct cam_periph_map_info	 mapinfo;
110 	pass_io_flags			 flags;
111 	ccb_flags			 data_flags;
112 	int				 num_user_segs;
113 	bus_dma_segment_t		 user_segs[PASS_MAX_SEGS];
114 	int				 num_kern_segs;
115 	bus_dma_segment_t		 kern_segs[PASS_MAX_SEGS];
116 	bus_dma_segment_t		*user_segptr;
117 	bus_dma_segment_t		*kern_segptr;
118 	int				 num_bufs;
119 	uint32_t			 dirs[CAM_PERIPH_MAXMAPS];
120 	uint32_t			 lengths[CAM_PERIPH_MAXMAPS];
121 	uint8_t				*user_bufs[CAM_PERIPH_MAXMAPS];
122 	uint8_t				*kern_bufs[CAM_PERIPH_MAXMAPS];
123 	struct bintime			 start_time;
124 	TAILQ_ENTRY(pass_io_req)	 links;
125 };
126 
127 struct pass_softc {
128 	pass_state		  state;
129 	pass_flags		  flags;
130 	u_int8_t		  pd_type;
131 	union ccb		  saved_ccb;
132 	int			  open_count;
133 	u_int		 	  maxio;
134 	struct devstat		 *device_stats;
135 	struct cdev		 *dev;
136 	struct cdev		 *alias_dev;
137 	struct task		  add_physpath_task;
138 	struct task		  shutdown_kqueue_task;
139 	struct selinfo		  read_select;
140 	TAILQ_HEAD(, pass_io_req) incoming_queue;
141 	TAILQ_HEAD(, pass_io_req) active_queue;
142 	TAILQ_HEAD(, pass_io_req) abandoned_queue;
143 	TAILQ_HEAD(, pass_io_req) done_queue;
144 	struct cam_periph	 *periph;
145 	char			  zone_name[12];
146 	char			  io_zone_name[12];
147 	uma_zone_t		  pass_zone;
148 	uma_zone_t		  pass_io_zone;
149 	size_t			  io_zone_size;
150 };
151 
152 static	d_open_t	passopen;
153 static	d_close_t	passclose;
154 static	d_ioctl_t	passioctl;
155 static	d_ioctl_t	passdoioctl;
156 static	d_poll_t	passpoll;
157 static	d_kqfilter_t	passkqfilter;
158 static	void		passreadfiltdetach(struct knote *kn);
159 static	int		passreadfilt(struct knote *kn, long hint);
160 
161 static	periph_init_t	passinit;
162 static	periph_ctor_t	passregister;
163 static	periph_oninv_t	passoninvalidate;
164 static	periph_dtor_t	passcleanup;
165 static	periph_start_t	passstart;
166 static	void		pass_shutdown_kqueue(void *context, int pending);
167 static	void		pass_add_physpath(void *context, int pending);
168 static	void		passasync(void *callback_arg, u_int32_t code,
169 				  struct cam_path *path, void *arg);
170 static	void		passdone(struct cam_periph *periph,
171 				 union ccb *done_ccb);
172 static	int		passcreatezone(struct cam_periph *periph);
173 static	void		passiocleanup(struct pass_softc *softc,
174 				      struct pass_io_req *io_req);
175 static	int		passcopysglist(struct cam_periph *periph,
176 				       struct pass_io_req *io_req,
177 				       ccb_flags direction);
178 static	int		passmemsetup(struct cam_periph *periph,
179 				     struct pass_io_req *io_req);
180 static	int		passmemdone(struct cam_periph *periph,
181 				    struct pass_io_req *io_req);
182 static	int		passerror(union ccb *ccb, u_int32_t cam_flags,
183 				  u_int32_t sense_flags);
184 static 	int		passsendccb(struct cam_periph *periph, union ccb *ccb,
185 				    union ccb *inccb);
186 
187 static struct periph_driver passdriver =
188 {
189 	passinit, "pass",
190 	TAILQ_HEAD_INITIALIZER(passdriver.units), /* generation */ 0
191 };
192 
193 PERIPHDRIVER_DECLARE(pass, passdriver);
194 
195 static struct cdevsw pass_cdevsw = {
196 	.d_version =	D_VERSION,
197 	.d_flags =	D_TRACKCLOSE,
198 	.d_open =	passopen,
199 	.d_close =	passclose,
200 	.d_ioctl =	passioctl,
201 	.d_poll = 	passpoll,
202 	.d_kqfilter = 	passkqfilter,
203 	.d_name =	"pass",
204 };
205 
206 static struct filterops passread_filtops = {
207 	.f_isfd	=	1,
208 	.f_detach =	passreadfiltdetach,
209 	.f_event =	passreadfilt
210 };
211 
212 static MALLOC_DEFINE(M_SCSIPASS, "scsi_pass", "scsi passthrough buffers");
213 
214 static void
215 passinit(void)
216 {
217 	cam_status status;
218 
219 	/*
220 	 * Install a global async callback.  This callback will
221 	 * receive async callbacks like "new device found".
222 	 */
223 	status = xpt_register_async(AC_FOUND_DEVICE, passasync, NULL, NULL);
224 
225 	if (status != CAM_REQ_CMP) {
226 		printf("pass: Failed to attach master async callback "
227 		       "due to status 0x%x!\n", status);
228 	}
229 
230 }
231 
232 static void
233 passrejectios(struct cam_periph *periph)
234 {
235 	struct pass_io_req *io_req, *io_req2;
236 	struct pass_softc *softc;
237 
238 	softc = (struct pass_softc *)periph->softc;
239 
240 	/*
241 	 * The user can no longer get status for I/O on the done queue, so
242 	 * clean up all outstanding I/O on the done queue.
243 	 */
244 	TAILQ_FOREACH_SAFE(io_req, &softc->done_queue, links, io_req2) {
245 		TAILQ_REMOVE(&softc->done_queue, io_req, links);
246 		passiocleanup(softc, io_req);
247 		uma_zfree(softc->pass_zone, io_req);
248 	}
249 
250 	/*
251 	 * The underlying device is gone, so we can't issue these I/Os.
252 	 * The devfs node has been shut down, so we can't return status to
253 	 * the user.  Free any I/O left on the incoming queue.
254 	 */
255 	TAILQ_FOREACH_SAFE(io_req, &softc->incoming_queue, links, io_req2) {
256 		TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
257 		passiocleanup(softc, io_req);
258 		uma_zfree(softc->pass_zone, io_req);
259 	}
260 
261 	/*
262 	 * Normally we would put I/Os on the abandoned queue and acquire a
263 	 * reference when we saw the final close.  But, the device went
264 	 * away and devfs may have moved everything off to deadfs by the
265 	 * time the I/O done callback is called; as a result, we won't see
266 	 * any more closes.  So, if we have any active I/Os, we need to put
267 	 * them on the abandoned queue.  When the abandoned queue is empty,
268 	 * we'll release the remaining reference (see below) to the peripheral.
269 	 */
270 	TAILQ_FOREACH_SAFE(io_req, &softc->active_queue, links, io_req2) {
271 		TAILQ_REMOVE(&softc->active_queue, io_req, links);
272 		io_req->flags |= PASS_IO_ABANDONED;
273 		TAILQ_INSERT_TAIL(&softc->abandoned_queue, io_req, links);
274 	}
275 
276 	/*
277 	 * If we put any I/O on the abandoned queue, acquire a reference.
278 	 */
279 	if ((!TAILQ_EMPTY(&softc->abandoned_queue))
280 	 && ((softc->flags & PASS_FLAG_ABANDONED_REF_SET) == 0)) {
281 		cam_periph_doacquire(periph);
282 		softc->flags |= PASS_FLAG_ABANDONED_REF_SET;
283 	}
284 }
285 
286 static void
287 passdevgonecb(void *arg)
288 {
289 	struct cam_periph *periph;
290 	struct mtx *mtx;
291 	struct pass_softc *softc;
292 	int i;
293 
294 	periph = (struct cam_periph *)arg;
295 	mtx = cam_periph_mtx(periph);
296 	mtx_lock(mtx);
297 
298 	softc = (struct pass_softc *)periph->softc;
299 	KASSERT(softc->open_count >= 0, ("Negative open count %d",
300 		softc->open_count));
301 
302 	/*
303 	 * When we get this callback, we will get no more close calls from
304 	 * devfs.  So if we have any dangling opens, we need to release the
305 	 * reference held for that particular context.
306 	 */
307 	for (i = 0; i < softc->open_count; i++)
308 		cam_periph_release_locked(periph);
309 
310 	softc->open_count = 0;
311 
312 	/*
313 	 * Release the reference held for the device node, it is gone now.
314 	 * Accordingly, inform all queued I/Os of their fate.
315 	 */
316 	cam_periph_release_locked(periph);
317 	passrejectios(periph);
318 
319 	/*
320 	 * We reference the SIM lock directly here, instead of using
321 	 * cam_periph_unlock().  The reason is that the final call to
322 	 * cam_periph_release_locked() above could result in the periph
323 	 * getting freed.  If that is the case, dereferencing the periph
324 	 * with a cam_periph_unlock() call would cause a page fault.
325 	 */
326 	mtx_unlock(mtx);
327 
328 	/*
329 	 * We have to remove our kqueue context from a thread because it
330 	 * may sleep.  It would be nice if we could get a callback from
331 	 * kqueue when it is done cleaning up resources.
332 	 */
333 	taskqueue_enqueue(taskqueue_thread, &softc->shutdown_kqueue_task);
334 }
335 
336 static void
337 passoninvalidate(struct cam_periph *periph)
338 {
339 	struct pass_softc *softc;
340 
341 	softc = (struct pass_softc *)periph->softc;
342 
343 	/*
344 	 * De-register any async callbacks.
345 	 */
346 	xpt_register_async(0, passasync, periph, periph->path);
347 
348 	softc->flags |= PASS_FLAG_INVALID;
349 
350 	/*
351 	 * Tell devfs this device has gone away, and ask for a callback
352 	 * when it has cleaned up its state.
353 	 */
354 	destroy_dev_sched_cb(softc->dev, passdevgonecb, periph);
355 }
356 
357 static void
358 passcleanup(struct cam_periph *periph)
359 {
360 	struct pass_softc *softc;
361 
362 	softc = (struct pass_softc *)periph->softc;
363 
364 	cam_periph_assert(periph, MA_OWNED);
365 	KASSERT(TAILQ_EMPTY(&softc->active_queue),
366 		("%s called when there are commands on the active queue!\n",
367 		__func__));
368 	KASSERT(TAILQ_EMPTY(&softc->abandoned_queue),
369 		("%s called when there are commands on the abandoned queue!\n",
370 		__func__));
371 	KASSERT(TAILQ_EMPTY(&softc->incoming_queue),
372 		("%s called when there are commands on the incoming queue!\n",
373 		__func__));
374 	KASSERT(TAILQ_EMPTY(&softc->done_queue),
375 		("%s called when there are commands on the done queue!\n",
376 		__func__));
377 
378 	devstat_remove_entry(softc->device_stats);
379 
380 	cam_periph_unlock(periph);
381 
382 	/*
383 	 * We call taskqueue_drain() for the physpath task to make sure it
384 	 * is complete.  We drop the lock because this can potentially
385 	 * sleep.  XXX KDM that is bad.  Need a way to get a callback when
386 	 * a taskqueue is drained.
387 	 *
388  	 * Note that we don't drain the kqueue shutdown task queue.  This
389 	 * is because we hold a reference on the periph for kqueue, and
390 	 * release that reference from the kqueue shutdown task queue.  So
391 	 * we cannot come into this routine unless we've released that
392 	 * reference.  Also, because that could be the last reference, we
393 	 * could be called from the cam_periph_release() call in
394 	 * pass_shutdown_kqueue().  In that case, the taskqueue_drain()
395 	 * would deadlock.  It would be preferable if we had a way to
396 	 * get a callback when a taskqueue is done.
397 	 */
398 	taskqueue_drain(taskqueue_thread, &softc->add_physpath_task);
399 
400 	cam_periph_lock(periph);
401 
402 	free(softc, M_DEVBUF);
403 }
404 
405 static void
406 pass_shutdown_kqueue(void *context, int pending)
407 {
408 	struct cam_periph *periph;
409 	struct pass_softc *softc;
410 
411 	periph = context;
412 	softc = periph->softc;
413 
414 	knlist_clear(&softc->read_select.si_note, /*is_locked*/ 0);
415 	knlist_destroy(&softc->read_select.si_note);
416 
417 	/*
418 	 * Release the reference we held for kqueue.
419 	 */
420 	cam_periph_release(periph);
421 }
422 
423 static void
424 pass_add_physpath(void *context, int pending)
425 {
426 	struct cam_periph *periph;
427 	struct pass_softc *softc;
428 	struct mtx *mtx;
429 	char *physpath;
430 
431 	/*
432 	 * If we have one, create a devfs alias for our
433 	 * physical path.
434 	 */
435 	periph = context;
436 	softc = periph->softc;
437 	physpath = malloc(MAXPATHLEN, M_DEVBUF, M_WAITOK);
438 	mtx = cam_periph_mtx(periph);
439 	mtx_lock(mtx);
440 
441 	if (periph->flags & CAM_PERIPH_INVALID)
442 		goto out;
443 
444 	if (xpt_getattr(physpath, MAXPATHLEN,
445 			"GEOM::physpath", periph->path) == 0
446 	 && strlen(physpath) != 0) {
447 
448 		mtx_unlock(mtx);
449 		make_dev_physpath_alias(MAKEDEV_WAITOK, &softc->alias_dev,
450 					softc->dev, softc->alias_dev, physpath);
451 		mtx_lock(mtx);
452 	}
453 
454 out:
455 	/*
456 	 * Now that we've made our alias, we no longer have to have a
457 	 * reference to the device.
458 	 */
459 	if ((softc->flags & PASS_FLAG_INITIAL_PHYSPATH) == 0)
460 		softc->flags |= PASS_FLAG_INITIAL_PHYSPATH;
461 
462 	/*
463 	 * We always acquire a reference to the periph before queueing this
464 	 * task queue function, so it won't go away before we run.
465 	 */
466 	while (pending-- > 0)
467 		cam_periph_release_locked(periph);
468 	mtx_unlock(mtx);
469 
470 	free(physpath, M_DEVBUF);
471 }
472 
473 static void
474 passasync(void *callback_arg, u_int32_t code,
475 	  struct cam_path *path, void *arg)
476 {
477 	struct cam_periph *periph;
478 
479 	periph = (struct cam_periph *)callback_arg;
480 
481 	switch (code) {
482 	case AC_FOUND_DEVICE:
483 	{
484 		struct ccb_getdev *cgd;
485 		cam_status status;
486 
487 		cgd = (struct ccb_getdev *)arg;
488 		if (cgd == NULL)
489 			break;
490 
491 		/*
492 		 * Allocate a peripheral instance for
493 		 * this device and start the probe
494 		 * process.
495 		 */
496 		status = cam_periph_alloc(passregister, passoninvalidate,
497 					  passcleanup, passstart, "pass",
498 					  CAM_PERIPH_BIO, path,
499 					  passasync, AC_FOUND_DEVICE, cgd);
500 
501 		if (status != CAM_REQ_CMP
502 		 && status != CAM_REQ_INPROG) {
503 			const struct cam_status_entry *entry;
504 
505 			entry = cam_fetch_status_entry(status);
506 
507 			printf("passasync: Unable to attach new device "
508 			       "due to status %#x: %s\n", status, entry ?
509 			       entry->status_text : "Unknown");
510 		}
511 
512 		break;
513 	}
514 	case AC_ADVINFO_CHANGED:
515 	{
516 		uintptr_t buftype;
517 
518 		buftype = (uintptr_t)arg;
519 		if (buftype == CDAI_TYPE_PHYS_PATH) {
520 			struct pass_softc *softc;
521 
522 			softc = (struct pass_softc *)periph->softc;
523 			/*
524 			 * Acquire a reference to the periph before we
525 			 * start the taskqueue, so that we don't run into
526 			 * a situation where the periph goes away before
527 			 * the task queue has a chance to run.
528 			 */
529 			if (cam_periph_acquire(periph) != 0)
530 				break;
531 
532 			taskqueue_enqueue(taskqueue_thread,
533 					  &softc->add_physpath_task);
534 		}
535 		break;
536 	}
537 	default:
538 		cam_periph_async(periph, code, path, arg);
539 		break;
540 	}
541 }
542 
543 static cam_status
544 passregister(struct cam_periph *periph, void *arg)
545 {
546 	struct pass_softc *softc;
547 	struct ccb_getdev *cgd;
548 	struct ccb_pathinq cpi;
549 	struct make_dev_args args;
550 	int error, no_tags;
551 
552 	cgd = (struct ccb_getdev *)arg;
553 	if (cgd == NULL) {
554 		printf("%s: no getdev CCB, can't register device\n", __func__);
555 		return(CAM_REQ_CMP_ERR);
556 	}
557 
558 	softc = (struct pass_softc *)malloc(sizeof(*softc),
559 					    M_DEVBUF, M_NOWAIT);
560 
561 	if (softc == NULL) {
562 		printf("%s: Unable to probe new device. "
563 		       "Unable to allocate softc\n", __func__);
564 		return(CAM_REQ_CMP_ERR);
565 	}
566 
567 	bzero(softc, sizeof(*softc));
568 	softc->state = PASS_STATE_NORMAL;
569 	if (cgd->protocol == PROTO_SCSI || cgd->protocol == PROTO_ATAPI)
570 		softc->pd_type = SID_TYPE(&cgd->inq_data);
571 	else if (cgd->protocol == PROTO_SATAPM)
572 		softc->pd_type = T_ENCLOSURE;
573 	else
574 		softc->pd_type = T_DIRECT;
575 
576 	periph->softc = softc;
577 	softc->periph = periph;
578 	TAILQ_INIT(&softc->incoming_queue);
579 	TAILQ_INIT(&softc->active_queue);
580 	TAILQ_INIT(&softc->abandoned_queue);
581 	TAILQ_INIT(&softc->done_queue);
582 	snprintf(softc->zone_name, sizeof(softc->zone_name), "%s%d",
583 		 periph->periph_name, periph->unit_number);
584 	snprintf(softc->io_zone_name, sizeof(softc->io_zone_name), "%s%dIO",
585 		 periph->periph_name, periph->unit_number);
586 	softc->io_zone_size = MAXPHYS;
587 	knlist_init_mtx(&softc->read_select.si_note, cam_periph_mtx(periph));
588 
589 	xpt_path_inq(&cpi, periph->path);
590 
591 	if (cpi.maxio == 0)
592 		softc->maxio = DFLTPHYS;	/* traditional default */
593 	else if (cpi.maxio > MAXPHYS)
594 		softc->maxio = MAXPHYS;		/* for safety */
595 	else
596 		softc->maxio = cpi.maxio;	/* real value */
597 
598 	if (cpi.hba_misc & PIM_UNMAPPED)
599 		softc->flags |= PASS_FLAG_UNMAPPED_CAPABLE;
600 
601 	/*
602 	 * We pass in 0 for a blocksize, since we don't
603 	 * know what the blocksize of this device is, if
604 	 * it even has a blocksize.
605 	 */
606 	cam_periph_unlock(periph);
607 	no_tags = (cgd->inq_data.flags & SID_CmdQue) == 0;
608 	softc->device_stats = devstat_new_entry("pass",
609 			  periph->unit_number, 0,
610 			  DEVSTAT_NO_BLOCKSIZE
611 			  | (no_tags ? DEVSTAT_NO_ORDERED_TAGS : 0),
612 			  softc->pd_type |
613 			  XPORT_DEVSTAT_TYPE(cpi.transport) |
614 			  DEVSTAT_TYPE_PASS,
615 			  DEVSTAT_PRIORITY_PASS);
616 
617 	/*
618 	 * Initialize the taskqueue handler for shutting down kqueue.
619 	 */
620 	TASK_INIT(&softc->shutdown_kqueue_task, /*priority*/ 0,
621 		  pass_shutdown_kqueue, periph);
622 
623 	/*
624 	 * Acquire a reference to the periph that we can release once we've
625 	 * cleaned up the kqueue.
626 	 */
627 	if (cam_periph_acquire(periph) != 0) {
628 		xpt_print(periph->path, "%s: lost periph during "
629 			  "registration!\n", __func__);
630 		cam_periph_lock(periph);
631 		return (CAM_REQ_CMP_ERR);
632 	}
633 
634 	/*
635 	 * Acquire a reference to the periph before we create the devfs
636 	 * instance for it.  We'll release this reference once the devfs
637 	 * instance has been freed.
638 	 */
639 	if (cam_periph_acquire(periph) != 0) {
640 		xpt_print(periph->path, "%s: lost periph during "
641 			  "registration!\n", __func__);
642 		cam_periph_lock(periph);
643 		return (CAM_REQ_CMP_ERR);
644 	}
645 
646 	/* Register the device */
647 	make_dev_args_init(&args);
648 	args.mda_devsw = &pass_cdevsw;
649 	args.mda_unit = periph->unit_number;
650 	args.mda_uid = UID_ROOT;
651 	args.mda_gid = GID_OPERATOR;
652 	args.mda_mode = 0600;
653 	args.mda_si_drv1 = periph;
654 	error = make_dev_s(&args, &softc->dev, "%s%d", periph->periph_name,
655 	    periph->unit_number);
656 	if (error != 0) {
657 		cam_periph_lock(periph);
658 		cam_periph_release_locked(periph);
659 		return (CAM_REQ_CMP_ERR);
660 	}
661 
662 	/*
663 	 * Hold a reference to the periph before we create the physical
664 	 * path alias so it can't go away.
665 	 */
666 	if (cam_periph_acquire(periph) != 0) {
667 		xpt_print(periph->path, "%s: lost periph during "
668 			  "registration!\n", __func__);
669 		cam_periph_lock(periph);
670 		return (CAM_REQ_CMP_ERR);
671 	}
672 
673 	cam_periph_lock(periph);
674 
675 	TASK_INIT(&softc->add_physpath_task, /*priority*/0,
676 		  pass_add_physpath, periph);
677 
678 	/*
679 	 * See if physical path information is already available.
680 	 */
681 	taskqueue_enqueue(taskqueue_thread, &softc->add_physpath_task);
682 
683 	/*
684 	 * Add an async callback so that we get notified if
685 	 * this device goes away or its physical path
686 	 * (stored in the advanced info data of the EDT) has
687 	 * changed.
688 	 */
689 	xpt_register_async(AC_LOST_DEVICE | AC_ADVINFO_CHANGED,
690 			   passasync, periph, periph->path);
691 
692 	if (bootverbose)
693 		xpt_announce_periph(periph, NULL);
694 
695 	return(CAM_REQ_CMP);
696 }
697 
698 static int
699 passopen(struct cdev *dev, int flags, int fmt, struct thread *td)
700 {
701 	struct cam_periph *periph;
702 	struct pass_softc *softc;
703 	int error;
704 
705 	periph = (struct cam_periph *)dev->si_drv1;
706 	if (cam_periph_acquire(periph) != 0)
707 		return (ENXIO);
708 
709 	cam_periph_lock(periph);
710 
711 	softc = (struct pass_softc *)periph->softc;
712 
713 	if (softc->flags & PASS_FLAG_INVALID) {
714 		cam_periph_release_locked(periph);
715 		cam_periph_unlock(periph);
716 		return(ENXIO);
717 	}
718 
719 	/*
720 	 * Don't allow access when we're running at a high securelevel.
721 	 */
722 	error = securelevel_gt(td->td_ucred, 1);
723 	if (error) {
724 		cam_periph_release_locked(periph);
725 		cam_periph_unlock(periph);
726 		return(error);
727 	}
728 
729 	/*
730 	 * Only allow read-write access.
731 	 */
732 	if (((flags & FWRITE) == 0) || ((flags & FREAD) == 0)) {
733 		cam_periph_release_locked(periph);
734 		cam_periph_unlock(periph);
735 		return(EPERM);
736 	}
737 
738 	/*
739 	 * We don't allow nonblocking access.
740 	 */
741 	if ((flags & O_NONBLOCK) != 0) {
742 		xpt_print(periph->path, "can't do nonblocking access\n");
743 		cam_periph_release_locked(periph);
744 		cam_periph_unlock(periph);
745 		return(EINVAL);
746 	}
747 
748 	softc->open_count++;
749 
750 	cam_periph_unlock(periph);
751 
752 	return (error);
753 }
754 
755 static int
756 passclose(struct cdev *dev, int flag, int fmt, struct thread *td)
757 {
758 	struct 	cam_periph *periph;
759 	struct  pass_softc *softc;
760 	struct mtx *mtx;
761 
762 	periph = (struct cam_periph *)dev->si_drv1;
763 	mtx = cam_periph_mtx(periph);
764 	mtx_lock(mtx);
765 
766 	softc = periph->softc;
767 	softc->open_count--;
768 
769 	if (softc->open_count == 0) {
770 		struct pass_io_req *io_req, *io_req2;
771 
772 		TAILQ_FOREACH_SAFE(io_req, &softc->done_queue, links, io_req2) {
773 			TAILQ_REMOVE(&softc->done_queue, io_req, links);
774 			passiocleanup(softc, io_req);
775 			uma_zfree(softc->pass_zone, io_req);
776 		}
777 
778 		TAILQ_FOREACH_SAFE(io_req, &softc->incoming_queue, links,
779 				   io_req2) {
780 			TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
781 			passiocleanup(softc, io_req);
782 			uma_zfree(softc->pass_zone, io_req);
783 		}
784 
785 		/*
786 		 * If there are any active I/Os, we need to forcibly acquire a
787 		 * reference to the peripheral so that we don't go away
788 		 * before they complete.  We'll release the reference when
789 		 * the abandoned queue is empty.
790 		 */
791 		io_req = TAILQ_FIRST(&softc->active_queue);
792 		if ((io_req != NULL)
793 		 && (softc->flags & PASS_FLAG_ABANDONED_REF_SET) == 0) {
794 			cam_periph_doacquire(periph);
795 			softc->flags |= PASS_FLAG_ABANDONED_REF_SET;
796 		}
797 
798 		/*
799 		 * Since the I/O in the active queue is not under our
800 		 * control, just set a flag so that we can clean it up when
801 		 * it completes and put it on the abandoned queue.  This
802 		 * will prevent our sending spurious completions in the
803 		 * event that the device is opened again before these I/Os
804 		 * complete.
805 		 */
806 		TAILQ_FOREACH_SAFE(io_req, &softc->active_queue, links,
807 				   io_req2) {
808 			TAILQ_REMOVE(&softc->active_queue, io_req, links);
809 			io_req->flags |= PASS_IO_ABANDONED;
810 			TAILQ_INSERT_TAIL(&softc->abandoned_queue, io_req,
811 					  links);
812 		}
813 	}
814 
815 	cam_periph_release_locked(periph);
816 
817 	/*
818 	 * We reference the lock directly here, instead of using
819 	 * cam_periph_unlock().  The reason is that the call to
820 	 * cam_periph_release_locked() above could result in the periph
821 	 * getting freed.  If that is the case, dereferencing the periph
822 	 * with a cam_periph_unlock() call would cause a page fault.
823 	 *
824 	 * cam_periph_release() avoids this problem using the same method,
825 	 * but we're manually acquiring and dropping the lock here to
826 	 * protect the open count and avoid another lock acquisition and
827 	 * release.
828 	 */
829 	mtx_unlock(mtx);
830 
831 	return (0);
832 }
833 
834 
835 static void
836 passstart(struct cam_periph *periph, union ccb *start_ccb)
837 {
838 	struct pass_softc *softc;
839 
840 	softc = (struct pass_softc *)periph->softc;
841 
842 	switch (softc->state) {
843 	case PASS_STATE_NORMAL: {
844 		struct pass_io_req *io_req;
845 
846 		/*
847 		 * Check for any queued I/O requests that require an
848 		 * allocated slot.
849 		 */
850 		io_req = TAILQ_FIRST(&softc->incoming_queue);
851 		if (io_req == NULL) {
852 			xpt_release_ccb(start_ccb);
853 			break;
854 		}
855 		TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
856 		TAILQ_INSERT_TAIL(&softc->active_queue, io_req, links);
857 		/*
858 		 * Merge the user's CCB into the allocated CCB.
859 		 */
860 		xpt_merge_ccb(start_ccb, &io_req->ccb);
861 		start_ccb->ccb_h.ccb_type = PASS_CCB_QUEUED_IO;
862 		start_ccb->ccb_h.ccb_ioreq = io_req;
863 		start_ccb->ccb_h.cbfcnp = passdone;
864 		io_req->alloced_ccb = start_ccb;
865 		binuptime(&io_req->start_time);
866 		devstat_start_transaction(softc->device_stats,
867 					  &io_req->start_time);
868 
869 		xpt_action(start_ccb);
870 
871 		/*
872 		 * If we have any more I/O waiting, schedule ourselves again.
873 		 */
874 		if (!TAILQ_EMPTY(&softc->incoming_queue))
875 			xpt_schedule(periph, CAM_PRIORITY_NORMAL);
876 		break;
877 	}
878 	default:
879 		break;
880 	}
881 }
882 
883 static void
884 passdone(struct cam_periph *periph, union ccb *done_ccb)
885 {
886 	struct pass_softc *softc;
887 	struct ccb_scsiio *csio;
888 
889 	softc = (struct pass_softc *)periph->softc;
890 
891 	cam_periph_assert(periph, MA_OWNED);
892 
893 	csio = &done_ccb->csio;
894 	switch (csio->ccb_h.ccb_type) {
895 	case PASS_CCB_QUEUED_IO: {
896 		struct pass_io_req *io_req;
897 
898 		io_req = done_ccb->ccb_h.ccb_ioreq;
899 #if 0
900 		xpt_print(periph->path, "%s: called for user CCB %p\n",
901 			  __func__, io_req->user_ccb_ptr);
902 #endif
903 		if (((done_ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_CMP)
904 		 && (done_ccb->ccb_h.flags & CAM_PASS_ERR_RECOVER)
905 		 && ((io_req->flags & PASS_IO_ABANDONED) == 0)) {
906 			int error;
907 
908 			error = passerror(done_ccb, CAM_RETRY_SELTO,
909 					  SF_RETRY_UA | SF_NO_PRINT);
910 
911 			if (error == ERESTART) {
912 				/*
913 				 * A retry was scheduled, so
914  				 * just return.
915 				 */
916 				return;
917 			}
918 		}
919 
920 		/*
921 		 * Copy the allocated CCB contents back to the malloced CCB
922 		 * so we can give status back to the user when he requests it.
923 		 */
924 		bcopy(done_ccb, &io_req->ccb, sizeof(*done_ccb));
925 
926 		/*
927 		 * Log data/transaction completion with devstat(9).
928 		 */
929 		switch (done_ccb->ccb_h.func_code) {
930 		case XPT_SCSI_IO:
931 			devstat_end_transaction(softc->device_stats,
932 			    done_ccb->csio.dxfer_len - done_ccb->csio.resid,
933 			    done_ccb->csio.tag_action & 0x3,
934 			    ((done_ccb->ccb_h.flags & CAM_DIR_MASK) ==
935 			    CAM_DIR_NONE) ? DEVSTAT_NO_DATA :
936 			    (done_ccb->ccb_h.flags & CAM_DIR_OUT) ?
937 			    DEVSTAT_WRITE : DEVSTAT_READ, NULL,
938 			    &io_req->start_time);
939 			break;
940 		case XPT_ATA_IO:
941 			devstat_end_transaction(softc->device_stats,
942 			    done_ccb->ataio.dxfer_len - done_ccb->ataio.resid,
943 			    0, /* Not used in ATA */
944 			    ((done_ccb->ccb_h.flags & CAM_DIR_MASK) ==
945 			    CAM_DIR_NONE) ? DEVSTAT_NO_DATA :
946 			    (done_ccb->ccb_h.flags & CAM_DIR_OUT) ?
947 			    DEVSTAT_WRITE : DEVSTAT_READ, NULL,
948 			    &io_req->start_time);
949 			break;
950 		case XPT_SMP_IO:
951 			/*
952 			 * XXX KDM this isn't quite right, but there isn't
953 			 * currently an easy way to represent a bidirectional
954 			 * transfer in devstat.  The only way to do it
955 			 * and have the byte counts come out right would
956 			 * mean that we would have to record two
957 			 * transactions, one for the request and one for the
958 			 * response.  For now, so that we report something,
959 			 * just treat the entire thing as a read.
960 			 */
961 			devstat_end_transaction(softc->device_stats,
962 			    done_ccb->smpio.smp_request_len +
963 			    done_ccb->smpio.smp_response_len,
964 			    DEVSTAT_TAG_SIMPLE, DEVSTAT_READ, NULL,
965 			    &io_req->start_time);
966 			break;
967 		default:
968 			devstat_end_transaction(softc->device_stats, 0,
969 			    DEVSTAT_TAG_NONE, DEVSTAT_NO_DATA, NULL,
970 			    &io_req->start_time);
971 			break;
972 		}
973 
974 		/*
975 		 * In the normal case, take the completed I/O off of the
976 		 * active queue and put it on the done queue.  Notitfy the
977 		 * user that we have a completed I/O.
978 		 */
979 		if ((io_req->flags & PASS_IO_ABANDONED) == 0) {
980 			TAILQ_REMOVE(&softc->active_queue, io_req, links);
981 			TAILQ_INSERT_TAIL(&softc->done_queue, io_req, links);
982 			selwakeuppri(&softc->read_select, PRIBIO);
983 			KNOTE_LOCKED(&softc->read_select.si_note, 0);
984 		} else {
985 			/*
986 			 * In the case of an abandoned I/O (final close
987 			 * without fetching the I/O), take it off of the
988 			 * abandoned queue and free it.
989 			 */
990 			TAILQ_REMOVE(&softc->abandoned_queue, io_req, links);
991 			passiocleanup(softc, io_req);
992 			uma_zfree(softc->pass_zone, io_req);
993 
994 			/*
995 			 * Release the done_ccb here, since we may wind up
996 			 * freeing the peripheral when we decrement the
997 			 * reference count below.
998 			 */
999 			xpt_release_ccb(done_ccb);
1000 
1001 			/*
1002 			 * If the abandoned queue is empty, we can release
1003 			 * our reference to the periph since we won't have
1004 			 * any more completions coming.
1005 			 */
1006 			if ((TAILQ_EMPTY(&softc->abandoned_queue))
1007 			 && (softc->flags & PASS_FLAG_ABANDONED_REF_SET)) {
1008 				softc->flags &= ~PASS_FLAG_ABANDONED_REF_SET;
1009 				cam_periph_release_locked(periph);
1010 			}
1011 
1012 			/*
1013 			 * We have already released the CCB, so we can
1014 			 * return.
1015 			 */
1016 			return;
1017 		}
1018 		break;
1019 	}
1020 	}
1021 	xpt_release_ccb(done_ccb);
1022 }
1023 
1024 static int
1025 passcreatezone(struct cam_periph *periph)
1026 {
1027 	struct pass_softc *softc;
1028 	int error;
1029 
1030 	error = 0;
1031 	softc = (struct pass_softc *)periph->softc;
1032 
1033 	cam_periph_assert(periph, MA_OWNED);
1034 	KASSERT(((softc->flags & PASS_FLAG_ZONE_VALID) == 0),
1035 		("%s called when the pass(4) zone is valid!\n", __func__));
1036 	KASSERT((softc->pass_zone == NULL),
1037 		("%s called when the pass(4) zone is allocated!\n", __func__));
1038 
1039 	if ((softc->flags & PASS_FLAG_ZONE_INPROG) == 0) {
1040 
1041 		/*
1042 		 * We're the first context through, so we need to create
1043 		 * the pass(4) UMA zone for I/O requests.
1044 		 */
1045 		softc->flags |= PASS_FLAG_ZONE_INPROG;
1046 
1047 		/*
1048 		 * uma_zcreate() does a blocking (M_WAITOK) allocation,
1049 		 * so we cannot hold a mutex while we call it.
1050 		 */
1051 		cam_periph_unlock(periph);
1052 
1053 		softc->pass_zone = uma_zcreate(softc->zone_name,
1054 		    sizeof(struct pass_io_req), NULL, NULL, NULL, NULL,
1055 		    /*align*/ 0, /*flags*/ 0);
1056 
1057 		softc->pass_io_zone = uma_zcreate(softc->io_zone_name,
1058 		    softc->io_zone_size, NULL, NULL, NULL, NULL,
1059 		    /*align*/ 0, /*flags*/ 0);
1060 
1061 		cam_periph_lock(periph);
1062 
1063 		if ((softc->pass_zone == NULL)
1064 		 || (softc->pass_io_zone == NULL)) {
1065 			if (softc->pass_zone == NULL)
1066 				xpt_print(periph->path, "unable to allocate "
1067 				    "IO Req UMA zone\n");
1068 			else
1069 				xpt_print(periph->path, "unable to allocate "
1070 				    "IO UMA zone\n");
1071 			softc->flags &= ~PASS_FLAG_ZONE_INPROG;
1072 			goto bailout;
1073 		}
1074 
1075 		/*
1076 		 * Set the flags appropriately and notify any other waiters.
1077 		 */
1078 		softc->flags &= PASS_FLAG_ZONE_INPROG;
1079 		softc->flags |= PASS_FLAG_ZONE_VALID;
1080 		wakeup(&softc->pass_zone);
1081 	} else {
1082 		/*
1083 		 * In this case, the UMA zone has not yet been created, but
1084 		 * another context is in the process of creating it.  We
1085 		 * need to sleep until the creation is either done or has
1086 		 * failed.
1087 		 */
1088 		while ((softc->flags & PASS_FLAG_ZONE_INPROG)
1089 		    && ((softc->flags & PASS_FLAG_ZONE_VALID) == 0)) {
1090 			error = msleep(&softc->pass_zone,
1091 				       cam_periph_mtx(periph), PRIBIO,
1092 				       "paszon", 0);
1093 			if (error != 0)
1094 				goto bailout;
1095 		}
1096 		/*
1097 		 * If the zone creation failed, no luck for the user.
1098 		 */
1099 		if ((softc->flags & PASS_FLAG_ZONE_VALID) == 0){
1100 			error = ENOMEM;
1101 			goto bailout;
1102 		}
1103 	}
1104 bailout:
1105 	return (error);
1106 }
1107 
1108 static void
1109 passiocleanup(struct pass_softc *softc, struct pass_io_req *io_req)
1110 {
1111 	union ccb *ccb;
1112 	u_int8_t **data_ptrs[CAM_PERIPH_MAXMAPS];
1113 	int i, numbufs;
1114 
1115 	ccb = &io_req->ccb;
1116 
1117 	switch (ccb->ccb_h.func_code) {
1118 	case XPT_DEV_MATCH:
1119 		numbufs = min(io_req->num_bufs, 2);
1120 
1121 		if (numbufs == 1) {
1122 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.matches;
1123 		} else {
1124 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.patterns;
1125 			data_ptrs[1] = (u_int8_t **)&ccb->cdm.matches;
1126 		}
1127 		break;
1128 	case XPT_SCSI_IO:
1129 	case XPT_CONT_TARGET_IO:
1130 		data_ptrs[0] = &ccb->csio.data_ptr;
1131 		numbufs = min(io_req->num_bufs, 1);
1132 		break;
1133 	case XPT_ATA_IO:
1134 		data_ptrs[0] = &ccb->ataio.data_ptr;
1135 		numbufs = min(io_req->num_bufs, 1);
1136 		break;
1137 	case XPT_SMP_IO:
1138 		numbufs = min(io_req->num_bufs, 2);
1139 		data_ptrs[0] = &ccb->smpio.smp_request;
1140 		data_ptrs[1] = &ccb->smpio.smp_response;
1141 		break;
1142 	case XPT_DEV_ADVINFO:
1143 		numbufs = min(io_req->num_bufs, 1);
1144 		data_ptrs[0] = (uint8_t **)&ccb->cdai.buf;
1145 		break;
1146 	case XPT_NVME_IO:
1147 	case XPT_NVME_ADMIN:
1148 		data_ptrs[0] = &ccb->nvmeio.data_ptr;
1149 		numbufs = min(io_req->num_bufs, 1);
1150 		break;
1151 	default:
1152 		/* allow ourselves to be swapped once again */
1153 		return;
1154 		break; /* NOTREACHED */
1155 	}
1156 
1157 	if (io_req->flags & PASS_IO_USER_SEG_MALLOC) {
1158 		free(io_req->user_segptr, M_SCSIPASS);
1159 		io_req->user_segptr = NULL;
1160 	}
1161 
1162 	/*
1163 	 * We only want to free memory we malloced.
1164 	 */
1165 	if (io_req->data_flags == CAM_DATA_VADDR) {
1166 		for (i = 0; i < io_req->num_bufs; i++) {
1167 			if (io_req->kern_bufs[i] == NULL)
1168 				continue;
1169 
1170 			free(io_req->kern_bufs[i], M_SCSIPASS);
1171 			io_req->kern_bufs[i] = NULL;
1172 		}
1173 	} else if (io_req->data_flags == CAM_DATA_SG) {
1174 		for (i = 0; i < io_req->num_kern_segs; i++) {
1175 			if ((uint8_t *)(uintptr_t)
1176 			    io_req->kern_segptr[i].ds_addr == NULL)
1177 				continue;
1178 
1179 			uma_zfree(softc->pass_io_zone, (uint8_t *)(uintptr_t)
1180 			    io_req->kern_segptr[i].ds_addr);
1181 			io_req->kern_segptr[i].ds_addr = 0;
1182 		}
1183 	}
1184 
1185 	if (io_req->flags & PASS_IO_KERN_SEG_MALLOC) {
1186 		free(io_req->kern_segptr, M_SCSIPASS);
1187 		io_req->kern_segptr = NULL;
1188 	}
1189 
1190 	if (io_req->data_flags != CAM_DATA_PADDR) {
1191 		for (i = 0; i < numbufs; i++) {
1192 			/*
1193 			 * Restore the user's buffer pointers to their
1194 			 * previous values.
1195 			 */
1196 			if (io_req->user_bufs[i] != NULL)
1197 				*data_ptrs[i] = io_req->user_bufs[i];
1198 		}
1199 	}
1200 
1201 }
1202 
1203 static int
1204 passcopysglist(struct cam_periph *periph, struct pass_io_req *io_req,
1205 	       ccb_flags direction)
1206 {
1207 	bus_size_t kern_watermark, user_watermark, len_copied, len_to_copy;
1208 	bus_dma_segment_t *user_sglist, *kern_sglist;
1209 	int i, j, error;
1210 
1211 	error = 0;
1212 	kern_watermark = 0;
1213 	user_watermark = 0;
1214 	len_to_copy = 0;
1215 	len_copied = 0;
1216 	user_sglist = io_req->user_segptr;
1217 	kern_sglist = io_req->kern_segptr;
1218 
1219 	for (i = 0, j = 0; i < io_req->num_user_segs &&
1220 	     j < io_req->num_kern_segs;) {
1221 		uint8_t *user_ptr, *kern_ptr;
1222 
1223 		len_to_copy = min(user_sglist[i].ds_len -user_watermark,
1224 		    kern_sglist[j].ds_len - kern_watermark);
1225 
1226 		user_ptr = (uint8_t *)(uintptr_t)user_sglist[i].ds_addr;
1227 		user_ptr = user_ptr + user_watermark;
1228 		kern_ptr = (uint8_t *)(uintptr_t)kern_sglist[j].ds_addr;
1229 		kern_ptr = kern_ptr + kern_watermark;
1230 
1231 		user_watermark += len_to_copy;
1232 		kern_watermark += len_to_copy;
1233 
1234 		if (!useracc(user_ptr, len_to_copy,
1235 		    (direction == CAM_DIR_IN) ? VM_PROT_WRITE : VM_PROT_READ)) {
1236 			xpt_print(periph->path, "%s: unable to access user "
1237 				  "S/G list element %p len %zu\n", __func__,
1238 				  user_ptr, len_to_copy);
1239 			error = EFAULT;
1240 			goto bailout;
1241 		}
1242 
1243 		if (direction == CAM_DIR_IN) {
1244 			error = copyout(kern_ptr, user_ptr, len_to_copy);
1245 			if (error != 0) {
1246 				xpt_print(periph->path, "%s: copyout of %u "
1247 					  "bytes from %p to %p failed with "
1248 					  "error %d\n", __func__, len_to_copy,
1249 					  kern_ptr, user_ptr, error);
1250 				goto bailout;
1251 			}
1252 		} else {
1253 			error = copyin(user_ptr, kern_ptr, len_to_copy);
1254 			if (error != 0) {
1255 				xpt_print(periph->path, "%s: copyin of %u "
1256 					  "bytes from %p to %p failed with "
1257 					  "error %d\n", __func__, len_to_copy,
1258 					  user_ptr, kern_ptr, error);
1259 				goto bailout;
1260 			}
1261 		}
1262 
1263 		len_copied += len_to_copy;
1264 
1265 		if (user_sglist[i].ds_len == user_watermark) {
1266 			i++;
1267 			user_watermark = 0;
1268 		}
1269 
1270 		if (kern_sglist[j].ds_len == kern_watermark) {
1271 			j++;
1272 			kern_watermark = 0;
1273 		}
1274 	}
1275 
1276 bailout:
1277 
1278 	return (error);
1279 }
1280 
1281 static int
1282 passmemsetup(struct cam_periph *periph, struct pass_io_req *io_req)
1283 {
1284 	union ccb *ccb;
1285 	struct pass_softc *softc;
1286 	int numbufs, i;
1287 	uint8_t **data_ptrs[CAM_PERIPH_MAXMAPS];
1288 	uint32_t lengths[CAM_PERIPH_MAXMAPS];
1289 	uint32_t dirs[CAM_PERIPH_MAXMAPS];
1290 	uint32_t num_segs;
1291 	uint16_t *seg_cnt_ptr;
1292 	size_t maxmap;
1293 	int error;
1294 
1295 	cam_periph_assert(periph, MA_NOTOWNED);
1296 
1297 	softc = periph->softc;
1298 
1299 	error = 0;
1300 	ccb = &io_req->ccb;
1301 	maxmap = 0;
1302 	num_segs = 0;
1303 	seg_cnt_ptr = NULL;
1304 
1305 	switch(ccb->ccb_h.func_code) {
1306 	case XPT_DEV_MATCH:
1307 		if (ccb->cdm.match_buf_len == 0) {
1308 			printf("%s: invalid match buffer length 0\n", __func__);
1309 			return(EINVAL);
1310 		}
1311 		if (ccb->cdm.pattern_buf_len > 0) {
1312 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.patterns;
1313 			lengths[0] = ccb->cdm.pattern_buf_len;
1314 			dirs[0] = CAM_DIR_OUT;
1315 			data_ptrs[1] = (u_int8_t **)&ccb->cdm.matches;
1316 			lengths[1] = ccb->cdm.match_buf_len;
1317 			dirs[1] = CAM_DIR_IN;
1318 			numbufs = 2;
1319 		} else {
1320 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.matches;
1321 			lengths[0] = ccb->cdm.match_buf_len;
1322 			dirs[0] = CAM_DIR_IN;
1323 			numbufs = 1;
1324 		}
1325 		io_req->data_flags = CAM_DATA_VADDR;
1326 		break;
1327 	case XPT_SCSI_IO:
1328 	case XPT_CONT_TARGET_IO:
1329 		if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
1330 			return(0);
1331 
1332 		/*
1333 		 * The user shouldn't be able to supply a bio.
1334 		 */
1335 		if ((ccb->ccb_h.flags & CAM_DATA_MASK) == CAM_DATA_BIO)
1336 			return (EINVAL);
1337 
1338 		io_req->data_flags = ccb->ccb_h.flags & CAM_DATA_MASK;
1339 
1340 		data_ptrs[0] = &ccb->csio.data_ptr;
1341 		lengths[0] = ccb->csio.dxfer_len;
1342 		dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1343 		num_segs = ccb->csio.sglist_cnt;
1344 		seg_cnt_ptr = &ccb->csio.sglist_cnt;
1345 		numbufs = 1;
1346 		maxmap = softc->maxio;
1347 		break;
1348 	case XPT_ATA_IO:
1349 		if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
1350 			return(0);
1351 
1352 		/*
1353 		 * We only support a single virtual address for ATA I/O.
1354 		 */
1355 		if ((ccb->ccb_h.flags & CAM_DATA_MASK) != CAM_DATA_VADDR)
1356 			return (EINVAL);
1357 
1358 		io_req->data_flags = CAM_DATA_VADDR;
1359 
1360 		data_ptrs[0] = &ccb->ataio.data_ptr;
1361 		lengths[0] = ccb->ataio.dxfer_len;
1362 		dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1363 		numbufs = 1;
1364 		maxmap = softc->maxio;
1365 		break;
1366 	case XPT_SMP_IO:
1367 		io_req->data_flags = CAM_DATA_VADDR;
1368 
1369 		data_ptrs[0] = &ccb->smpio.smp_request;
1370 		lengths[0] = ccb->smpio.smp_request_len;
1371 		dirs[0] = CAM_DIR_OUT;
1372 		data_ptrs[1] = &ccb->smpio.smp_response;
1373 		lengths[1] = ccb->smpio.smp_response_len;
1374 		dirs[1] = CAM_DIR_IN;
1375 		numbufs = 2;
1376 		maxmap = softc->maxio;
1377 		break;
1378 	case XPT_DEV_ADVINFO:
1379 		if (ccb->cdai.bufsiz == 0)
1380 			return (0);
1381 
1382 		io_req->data_flags = CAM_DATA_VADDR;
1383 
1384 		data_ptrs[0] = (uint8_t **)&ccb->cdai.buf;
1385 		lengths[0] = ccb->cdai.bufsiz;
1386 		dirs[0] = CAM_DIR_IN;
1387 		numbufs = 1;
1388 		break;
1389 	case XPT_NVME_ADMIN:
1390 	case XPT_NVME_IO:
1391 		if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
1392 			return (0);
1393 
1394 		io_req->data_flags = ccb->ccb_h.flags & CAM_DATA_MASK;
1395 
1396 		data_ptrs[0] = &ccb->nvmeio.data_ptr;
1397 		lengths[0] = ccb->nvmeio.dxfer_len;
1398 		dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1399 		num_segs = ccb->nvmeio.sglist_cnt;
1400 		seg_cnt_ptr = &ccb->nvmeio.sglist_cnt;
1401 		numbufs = 1;
1402 		maxmap = softc->maxio;
1403 		break;
1404 	default:
1405 		return(EINVAL);
1406 		break; /* NOTREACHED */
1407 	}
1408 
1409 	io_req->num_bufs = numbufs;
1410 
1411 	/*
1412 	 * If there is a maximum, check to make sure that the user's
1413 	 * request fits within the limit.  In general, we should only have
1414 	 * a maximum length for requests that go to hardware.  Otherwise it
1415 	 * is whatever we're able to malloc.
1416 	 */
1417 	for (i = 0; i < numbufs; i++) {
1418 		io_req->user_bufs[i] = *data_ptrs[i];
1419 		io_req->dirs[i] = dirs[i];
1420 		io_req->lengths[i] = lengths[i];
1421 
1422 		if (maxmap == 0)
1423 			continue;
1424 
1425 		if (lengths[i] <= maxmap)
1426 			continue;
1427 
1428 		xpt_print(periph->path, "%s: data length %u > max allowed %u "
1429 			  "bytes\n", __func__, lengths[i], maxmap);
1430 		error = EINVAL;
1431 		goto bailout;
1432 	}
1433 
1434 	switch (io_req->data_flags) {
1435 	case CAM_DATA_VADDR:
1436 		/* Map or copy the buffer into kernel address space */
1437 		for (i = 0; i < numbufs; i++) {
1438 			uint8_t *tmp_buf;
1439 
1440 			/*
1441 			 * If for some reason no length is specified, we
1442 			 * don't need to allocate anything.
1443 			 */
1444 			if (io_req->lengths[i] == 0)
1445 				continue;
1446 
1447 			/*
1448 			 * Make sure that the user's buffer is accessible
1449 			 * to that process.
1450 			 */
1451 			if (!useracc(io_req->user_bufs[i], io_req->lengths[i],
1452 			    (io_req->dirs[i] == CAM_DIR_IN) ? VM_PROT_WRITE :
1453 			     VM_PROT_READ)) {
1454 				xpt_print(periph->path, "%s: user address %p "
1455 				    "length %u is not accessible\n", __func__,
1456 				    io_req->user_bufs[i], io_req->lengths[i]);
1457 				error = EFAULT;
1458 				goto bailout;
1459 			}
1460 
1461 			tmp_buf = malloc(lengths[i], M_SCSIPASS,
1462 					 M_WAITOK | M_ZERO);
1463 			io_req->kern_bufs[i] = tmp_buf;
1464 			*data_ptrs[i] = tmp_buf;
1465 
1466 #if 0
1467 			xpt_print(periph->path, "%s: malloced %p len %u, user "
1468 				  "buffer %p, operation: %s\n", __func__,
1469 				  tmp_buf, lengths[i], io_req->user_bufs[i],
1470 				  (dirs[i] == CAM_DIR_IN) ? "read" : "write");
1471 #endif
1472 			/*
1473 			 * We only need to copy in if the user is writing.
1474 			 */
1475 			if (dirs[i] != CAM_DIR_OUT)
1476 				continue;
1477 
1478 			error = copyin(io_req->user_bufs[i],
1479 				       io_req->kern_bufs[i], lengths[i]);
1480 			if (error != 0) {
1481 				xpt_print(periph->path, "%s: copy of user "
1482 					  "buffer from %p to %p failed with "
1483 					  "error %d\n", __func__,
1484 					  io_req->user_bufs[i],
1485 					  io_req->kern_bufs[i], error);
1486 				goto bailout;
1487 			}
1488 		}
1489 		break;
1490 	case CAM_DATA_PADDR:
1491 		/* Pass down the pointer as-is */
1492 		break;
1493 	case CAM_DATA_SG: {
1494 		size_t sg_length, size_to_go, alloc_size;
1495 		uint32_t num_segs_needed;
1496 
1497 		/*
1498 		 * Copy the user S/G list in, and then copy in the
1499 		 * individual segments.
1500 		 */
1501 		/*
1502 		 * We shouldn't see this, but check just in case.
1503 		 */
1504 		if (numbufs != 1) {
1505 			xpt_print(periph->path, "%s: cannot currently handle "
1506 				  "more than one S/G list per CCB\n", __func__);
1507 			error = EINVAL;
1508 			goto bailout;
1509 		}
1510 
1511 		/*
1512 		 * We have to have at least one segment.
1513 		 */
1514 		if (num_segs == 0) {
1515 			xpt_print(periph->path, "%s: CAM_DATA_SG flag set, "
1516 				  "but sglist_cnt=0!\n", __func__);
1517 			error = EINVAL;
1518 			goto bailout;
1519 		}
1520 
1521 		/*
1522 		 * Make sure the user specified the total length and didn't
1523 		 * just leave it to us to decode the S/G list.
1524 		 */
1525 		if (lengths[0] == 0) {
1526 			xpt_print(periph->path, "%s: no dxfer_len specified, "
1527 				  "but CAM_DATA_SG flag is set!\n", __func__);
1528 			error = EINVAL;
1529 			goto bailout;
1530 		}
1531 
1532 		/*
1533 		 * We allocate buffers in io_zone_size increments for an
1534 		 * S/G list.  This will generally be MAXPHYS.
1535 		 */
1536 		if (lengths[0] <= softc->io_zone_size)
1537 			num_segs_needed = 1;
1538 		else {
1539 			num_segs_needed = lengths[0] / softc->io_zone_size;
1540 			if ((lengths[0] % softc->io_zone_size) != 0)
1541 				num_segs_needed++;
1542 		}
1543 
1544 		/* Figure out the size of the S/G list */
1545 		sg_length = num_segs * sizeof(bus_dma_segment_t);
1546 		io_req->num_user_segs = num_segs;
1547 		io_req->num_kern_segs = num_segs_needed;
1548 
1549 		/* Save the user's S/G list pointer for later restoration */
1550 		io_req->user_bufs[0] = *data_ptrs[0];
1551 
1552 		/*
1553 		 * If we have enough segments allocated by default to handle
1554 		 * the length of the user's S/G list,
1555 		 */
1556 		if (num_segs > PASS_MAX_SEGS) {
1557 			io_req->user_segptr = malloc(sizeof(bus_dma_segment_t) *
1558 			    num_segs, M_SCSIPASS, M_WAITOK | M_ZERO);
1559 			io_req->flags |= PASS_IO_USER_SEG_MALLOC;
1560 		} else
1561 			io_req->user_segptr = io_req->user_segs;
1562 
1563 		if (!useracc(*data_ptrs[0], sg_length, VM_PROT_READ)) {
1564 			xpt_print(periph->path, "%s: unable to access user "
1565 				  "S/G list at %p\n", __func__, *data_ptrs[0]);
1566 			error = EFAULT;
1567 			goto bailout;
1568 		}
1569 
1570 		error = copyin(*data_ptrs[0], io_req->user_segptr, sg_length);
1571 		if (error != 0) {
1572 			xpt_print(periph->path, "%s: copy of user S/G list "
1573 				  "from %p to %p failed with error %d\n",
1574 				  __func__, *data_ptrs[0], io_req->user_segptr,
1575 				  error);
1576 			goto bailout;
1577 		}
1578 
1579 		if (num_segs_needed > PASS_MAX_SEGS) {
1580 			io_req->kern_segptr = malloc(sizeof(bus_dma_segment_t) *
1581 			    num_segs_needed, M_SCSIPASS, M_WAITOK | M_ZERO);
1582 			io_req->flags |= PASS_IO_KERN_SEG_MALLOC;
1583 		} else {
1584 			io_req->kern_segptr = io_req->kern_segs;
1585 		}
1586 
1587 		/*
1588 		 * Allocate the kernel S/G list.
1589 		 */
1590 		for (size_to_go = lengths[0], i = 0;
1591 		     size_to_go > 0 && i < num_segs_needed;
1592 		     i++, size_to_go -= alloc_size) {
1593 			uint8_t *kern_ptr;
1594 
1595 			alloc_size = min(size_to_go, softc->io_zone_size);
1596 			kern_ptr = uma_zalloc(softc->pass_io_zone, M_WAITOK);
1597 			io_req->kern_segptr[i].ds_addr =
1598 			    (bus_addr_t)(uintptr_t)kern_ptr;
1599 			io_req->kern_segptr[i].ds_len = alloc_size;
1600 		}
1601 		if (size_to_go > 0) {
1602 			printf("%s: size_to_go = %zu, software error!\n",
1603 			       __func__, size_to_go);
1604 			error = EINVAL;
1605 			goto bailout;
1606 		}
1607 
1608 		*data_ptrs[0] = (uint8_t *)io_req->kern_segptr;
1609 		*seg_cnt_ptr = io_req->num_kern_segs;
1610 
1611 		/*
1612 		 * We only need to copy data here if the user is writing.
1613 		 */
1614 		if (dirs[0] == CAM_DIR_OUT)
1615 			error = passcopysglist(periph, io_req, dirs[0]);
1616 		break;
1617 	}
1618 	case CAM_DATA_SG_PADDR: {
1619 		size_t sg_length;
1620 
1621 		/*
1622 		 * We shouldn't see this, but check just in case.
1623 		 */
1624 		if (numbufs != 1) {
1625 			printf("%s: cannot currently handle more than one "
1626 			       "S/G list per CCB\n", __func__);
1627 			error = EINVAL;
1628 			goto bailout;
1629 		}
1630 
1631 		/*
1632 		 * We have to have at least one segment.
1633 		 */
1634 		if (num_segs == 0) {
1635 			xpt_print(periph->path, "%s: CAM_DATA_SG_PADDR flag "
1636 				  "set, but sglist_cnt=0!\n", __func__);
1637 			error = EINVAL;
1638 			goto bailout;
1639 		}
1640 
1641 		/*
1642 		 * Make sure the user specified the total length and didn't
1643 		 * just leave it to us to decode the S/G list.
1644 		 */
1645 		if (lengths[0] == 0) {
1646 			xpt_print(periph->path, "%s: no dxfer_len specified, "
1647 				  "but CAM_DATA_SG flag is set!\n", __func__);
1648 			error = EINVAL;
1649 			goto bailout;
1650 		}
1651 
1652 		/* Figure out the size of the S/G list */
1653 		sg_length = num_segs * sizeof(bus_dma_segment_t);
1654 		io_req->num_user_segs = num_segs;
1655 		io_req->num_kern_segs = io_req->num_user_segs;
1656 
1657 		/* Save the user's S/G list pointer for later restoration */
1658 		io_req->user_bufs[0] = *data_ptrs[0];
1659 
1660 		if (num_segs > PASS_MAX_SEGS) {
1661 			io_req->user_segptr = malloc(sizeof(bus_dma_segment_t) *
1662 			    num_segs, M_SCSIPASS, M_WAITOK | M_ZERO);
1663 			io_req->flags |= PASS_IO_USER_SEG_MALLOC;
1664 		} else
1665 			io_req->user_segptr = io_req->user_segs;
1666 
1667 		io_req->kern_segptr = io_req->user_segptr;
1668 
1669 		error = copyin(*data_ptrs[0], io_req->user_segptr, sg_length);
1670 		if (error != 0) {
1671 			xpt_print(periph->path, "%s: copy of user S/G list "
1672 				  "from %p to %p failed with error %d\n",
1673 				  __func__, *data_ptrs[0], io_req->user_segptr,
1674 				  error);
1675 			goto bailout;
1676 		}
1677 		break;
1678 	}
1679 	default:
1680 	case CAM_DATA_BIO:
1681 		/*
1682 		 * A user shouldn't be attaching a bio to the CCB.  It
1683 		 * isn't a user-accessible structure.
1684 		 */
1685 		error = EINVAL;
1686 		break;
1687 	}
1688 
1689 bailout:
1690 	if (error != 0)
1691 		passiocleanup(softc, io_req);
1692 
1693 	return (error);
1694 }
1695 
1696 static int
1697 passmemdone(struct cam_periph *periph, struct pass_io_req *io_req)
1698 {
1699 	struct pass_softc *softc;
1700 	int error;
1701 	int i;
1702 
1703 	error = 0;
1704 	softc = (struct pass_softc *)periph->softc;
1705 
1706 	switch (io_req->data_flags) {
1707 	case CAM_DATA_VADDR:
1708 		/*
1709 		 * Copy back to the user buffer if this was a read.
1710 		 */
1711 		for (i = 0; i < io_req->num_bufs; i++) {
1712 			if (io_req->dirs[i] != CAM_DIR_IN)
1713 				continue;
1714 
1715 			error = copyout(io_req->kern_bufs[i],
1716 			    io_req->user_bufs[i], io_req->lengths[i]);
1717 			if (error != 0) {
1718 				xpt_print(periph->path, "Unable to copy %u "
1719 					  "bytes from %p to user address %p\n",
1720 					  io_req->lengths[i],
1721 					  io_req->kern_bufs[i],
1722 					  io_req->user_bufs[i]);
1723 				goto bailout;
1724 			}
1725 
1726 		}
1727 		break;
1728 	case CAM_DATA_PADDR:
1729 		/* Do nothing.  The pointer is a physical address already */
1730 		break;
1731 	case CAM_DATA_SG:
1732 		/*
1733 		 * Copy back to the user buffer if this was a read.
1734 		 * Restore the user's S/G list buffer pointer.
1735 		 */
1736 		if (io_req->dirs[0] == CAM_DIR_IN)
1737 			error = passcopysglist(periph, io_req, io_req->dirs[0]);
1738 		break;
1739 	case CAM_DATA_SG_PADDR:
1740 		/*
1741 		 * Restore the user's S/G list buffer pointer.  No need to
1742 		 * copy.
1743 		 */
1744 		break;
1745 	default:
1746 	case CAM_DATA_BIO:
1747 		error = EINVAL;
1748 		break;
1749 	}
1750 
1751 bailout:
1752 	/*
1753 	 * Reset the user's pointers to their original values and free
1754 	 * allocated memory.
1755 	 */
1756 	passiocleanup(softc, io_req);
1757 
1758 	return (error);
1759 }
1760 
1761 static int
1762 passioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flag, struct thread *td)
1763 {
1764 	int error;
1765 
1766 	if ((error = passdoioctl(dev, cmd, addr, flag, td)) == ENOTTY) {
1767 		error = cam_compat_ioctl(dev, cmd, addr, flag, td, passdoioctl);
1768 	}
1769 	return (error);
1770 }
1771 
1772 static int
1773 passdoioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flag, struct thread *td)
1774 {
1775 	struct	cam_periph *periph;
1776 	struct	pass_softc *softc;
1777 	int	error;
1778 	uint32_t priority;
1779 
1780 	periph = (struct cam_periph *)dev->si_drv1;
1781 	cam_periph_lock(periph);
1782 	softc = (struct pass_softc *)periph->softc;
1783 
1784 	error = 0;
1785 
1786 	switch (cmd) {
1787 
1788 	case CAMIOCOMMAND:
1789 	{
1790 		union ccb *inccb;
1791 		union ccb *ccb;
1792 		int ccb_malloced;
1793 
1794 		inccb = (union ccb *)addr;
1795 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
1796 		if (inccb->ccb_h.func_code == XPT_SCSI_IO)
1797 			inccb->csio.bio = NULL;
1798 #endif
1799 
1800 		if (inccb->ccb_h.flags & CAM_UNLOCKED) {
1801 			error = EINVAL;
1802 			break;
1803 		}
1804 
1805 		/*
1806 		 * Some CCB types, like scan bus and scan lun can only go
1807 		 * through the transport layer device.
1808 		 */
1809 		if (inccb->ccb_h.func_code & XPT_FC_XPT_ONLY) {
1810 			xpt_print(periph->path, "CCB function code %#x is "
1811 			    "restricted to the XPT device\n",
1812 			    inccb->ccb_h.func_code);
1813 			error = ENODEV;
1814 			break;
1815 		}
1816 
1817 		/* Compatibility for RL/priority-unaware code. */
1818 		priority = inccb->ccb_h.pinfo.priority;
1819 		if (priority <= CAM_PRIORITY_OOB)
1820 		    priority += CAM_PRIORITY_OOB + 1;
1821 
1822 		/*
1823 		 * Non-immediate CCBs need a CCB from the per-device pool
1824 		 * of CCBs, which is scheduled by the transport layer.
1825 		 * Immediate CCBs and user-supplied CCBs should just be
1826 		 * malloced.
1827 		 */
1828 		if ((inccb->ccb_h.func_code & XPT_FC_QUEUED)
1829 		 && ((inccb->ccb_h.func_code & XPT_FC_USER_CCB) == 0)) {
1830 			ccb = cam_periph_getccb(periph, priority);
1831 			ccb_malloced = 0;
1832 		} else {
1833 			ccb = xpt_alloc_ccb_nowait();
1834 
1835 			if (ccb != NULL)
1836 				xpt_setup_ccb(&ccb->ccb_h, periph->path,
1837 					      priority);
1838 			ccb_malloced = 1;
1839 		}
1840 
1841 		if (ccb == NULL) {
1842 			xpt_print(periph->path, "unable to allocate CCB\n");
1843 			error = ENOMEM;
1844 			break;
1845 		}
1846 
1847 		error = passsendccb(periph, ccb, inccb);
1848 
1849 		if (ccb_malloced)
1850 			xpt_free_ccb(ccb);
1851 		else
1852 			xpt_release_ccb(ccb);
1853 
1854 		break;
1855 	}
1856 	case CAMIOQUEUE:
1857 	{
1858 		struct pass_io_req *io_req;
1859 		union ccb **user_ccb, *ccb;
1860 		xpt_opcode fc;
1861 
1862 		if ((softc->flags & PASS_FLAG_ZONE_VALID) == 0) {
1863 			error = passcreatezone(periph);
1864 			if (error != 0)
1865 				goto bailout;
1866 		}
1867 
1868 		/*
1869 		 * We're going to do a blocking allocation for this I/O
1870 		 * request, so we have to drop the lock.
1871 		 */
1872 		cam_periph_unlock(periph);
1873 
1874 		io_req = uma_zalloc(softc->pass_zone, M_WAITOK | M_ZERO);
1875 		ccb = &io_req->ccb;
1876 		user_ccb = (union ccb **)addr;
1877 
1878 		/*
1879 		 * Unlike the CAMIOCOMMAND ioctl above, we only have a
1880 		 * pointer to the user's CCB, so we have to copy the whole
1881 		 * thing in to a buffer we have allocated (above) instead
1882 		 * of allowing the ioctl code to malloc a buffer and copy
1883 		 * it in.
1884 		 *
1885 		 * This is an advantage for this asynchronous interface,
1886 		 * since we don't want the memory to get freed while the
1887 		 * CCB is outstanding.
1888 		 */
1889 #if 0
1890 		xpt_print(periph->path, "Copying user CCB %p to "
1891 			  "kernel address %p\n", *user_ccb, ccb);
1892 #endif
1893 		error = copyin(*user_ccb, ccb, sizeof(*ccb));
1894 		if (error != 0) {
1895 			xpt_print(periph->path, "Copy of user CCB %p to "
1896 				  "kernel address %p failed with error %d\n",
1897 				  *user_ccb, ccb, error);
1898 			goto camioqueue_error;
1899 		}
1900 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
1901 		if (ccb->ccb_h.func_code == XPT_SCSI_IO)
1902 			ccb->csio.bio = NULL;
1903 #endif
1904 
1905 		if (ccb->ccb_h.flags & CAM_UNLOCKED) {
1906 			error = EINVAL;
1907 			goto camioqueue_error;
1908 		}
1909 
1910 		if (ccb->ccb_h.flags & CAM_CDB_POINTER) {
1911 			if (ccb->csio.cdb_len > IOCDBLEN) {
1912 				error = EINVAL;
1913 				goto camioqueue_error;
1914 			}
1915 			error = copyin(ccb->csio.cdb_io.cdb_ptr,
1916 			    ccb->csio.cdb_io.cdb_bytes, ccb->csio.cdb_len);
1917 			if (error != 0)
1918 				goto camioqueue_error;
1919 			ccb->ccb_h.flags &= ~CAM_CDB_POINTER;
1920 		}
1921 
1922 		/*
1923 		 * Some CCB types, like scan bus and scan lun can only go
1924 		 * through the transport layer device.
1925 		 */
1926 		if (ccb->ccb_h.func_code & XPT_FC_XPT_ONLY) {
1927 			xpt_print(periph->path, "CCB function code %#x is "
1928 			    "restricted to the XPT device\n",
1929 			    ccb->ccb_h.func_code);
1930 			error = ENODEV;
1931 			goto camioqueue_error;
1932 		}
1933 
1934 		/*
1935 		 * Save the user's CCB pointer as well as his linked list
1936 		 * pointers and peripheral private area so that we can
1937 		 * restore these later.
1938 		 */
1939 		io_req->user_ccb_ptr = *user_ccb;
1940 		io_req->user_periph_links = ccb->ccb_h.periph_links;
1941 		io_req->user_periph_priv = ccb->ccb_h.periph_priv;
1942 
1943 		/*
1944 		 * Now that we've saved the user's values, we can set our
1945 		 * own peripheral private entry.
1946 		 */
1947 		ccb->ccb_h.ccb_ioreq = io_req;
1948 
1949 		/* Compatibility for RL/priority-unaware code. */
1950 		priority = ccb->ccb_h.pinfo.priority;
1951 		if (priority <= CAM_PRIORITY_OOB)
1952 		    priority += CAM_PRIORITY_OOB + 1;
1953 
1954 		/*
1955 		 * Setup fields in the CCB like the path and the priority.
1956 		 * The path in particular cannot be done in userland, since
1957 		 * it is a pointer to a kernel data structure.
1958 		 */
1959 		xpt_setup_ccb_flags(&ccb->ccb_h, periph->path, priority,
1960 				    ccb->ccb_h.flags);
1961 
1962 		/*
1963 		 * Setup our done routine.  There is no way for the user to
1964 		 * have a valid pointer here.
1965 		 */
1966 		ccb->ccb_h.cbfcnp = passdone;
1967 
1968 		fc = ccb->ccb_h.func_code;
1969 		/*
1970 		 * If this function code has memory that can be mapped in
1971 		 * or out, we need to call passmemsetup().
1972 		 */
1973 		if ((fc == XPT_SCSI_IO) || (fc == XPT_ATA_IO)
1974 		 || (fc == XPT_SMP_IO) || (fc == XPT_DEV_MATCH)
1975 		 || (fc == XPT_DEV_ADVINFO)
1976 		 || (fc == XPT_NVME_ADMIN) || (fc == XPT_NVME_IO)) {
1977 			error = passmemsetup(periph, io_req);
1978 			if (error != 0)
1979 				goto camioqueue_error;
1980 		} else
1981 			io_req->mapinfo.num_bufs_used = 0;
1982 
1983 		cam_periph_lock(periph);
1984 
1985 		/*
1986 		 * Everything goes on the incoming queue initially.
1987 		 */
1988 		TAILQ_INSERT_TAIL(&softc->incoming_queue, io_req, links);
1989 
1990 		/*
1991 		 * If the CCB is queued, and is not a user CCB, then
1992 		 * we need to allocate a slot for it.  Call xpt_schedule()
1993 		 * so that our start routine will get called when a CCB is
1994 		 * available.
1995 		 */
1996 		if ((fc & XPT_FC_QUEUED)
1997 		 && ((fc & XPT_FC_USER_CCB) == 0)) {
1998 			xpt_schedule(periph, priority);
1999 			break;
2000 		}
2001 
2002 		/*
2003 		 * At this point, the CCB in question is either an
2004 		 * immediate CCB (like XPT_DEV_ADVINFO) or it is a user CCB
2005 		 * and therefore should be malloced, not allocated via a slot.
2006 		 * Remove the CCB from the incoming queue and add it to the
2007 		 * active queue.
2008 		 */
2009 		TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
2010 		TAILQ_INSERT_TAIL(&softc->active_queue, io_req, links);
2011 
2012 		xpt_action(ccb);
2013 
2014 		/*
2015 		 * If this is not a queued CCB (i.e. it is an immediate CCB),
2016 		 * then it is already done.  We need to put it on the done
2017 		 * queue for the user to fetch.
2018 		 */
2019 		if ((fc & XPT_FC_QUEUED) == 0) {
2020 			TAILQ_REMOVE(&softc->active_queue, io_req, links);
2021 			TAILQ_INSERT_TAIL(&softc->done_queue, io_req, links);
2022 		}
2023 		break;
2024 
2025 camioqueue_error:
2026 		uma_zfree(softc->pass_zone, io_req);
2027 		cam_periph_lock(periph);
2028 		break;
2029 	}
2030 	case CAMIOGET:
2031 	{
2032 		union ccb **user_ccb;
2033 		struct pass_io_req *io_req;
2034 		int old_error;
2035 
2036 		user_ccb = (union ccb **)addr;
2037 		old_error = 0;
2038 
2039 		io_req = TAILQ_FIRST(&softc->done_queue);
2040 		if (io_req == NULL) {
2041 			error = ENOENT;
2042 			break;
2043 		}
2044 
2045 		/*
2046 		 * Remove the I/O from the done queue.
2047 		 */
2048 		TAILQ_REMOVE(&softc->done_queue, io_req, links);
2049 
2050 		/*
2051 		 * We have to drop the lock during the copyout because the
2052 		 * copyout can result in VM faults that require sleeping.
2053 		 */
2054 		cam_periph_unlock(periph);
2055 
2056 		/*
2057 		 * Do any needed copies (e.g. for reads) and revert the
2058 		 * pointers in the CCB back to the user's pointers.
2059 		 */
2060 		error = passmemdone(periph, io_req);
2061 
2062 		old_error = error;
2063 
2064 		io_req->ccb.ccb_h.periph_links = io_req->user_periph_links;
2065 		io_req->ccb.ccb_h.periph_priv = io_req->user_periph_priv;
2066 
2067 #if 0
2068 		xpt_print(periph->path, "Copying to user CCB %p from "
2069 			  "kernel address %p\n", *user_ccb, &io_req->ccb);
2070 #endif
2071 
2072 		error = copyout(&io_req->ccb, *user_ccb, sizeof(union ccb));
2073 		if (error != 0) {
2074 			xpt_print(periph->path, "Copy to user CCB %p from "
2075 				  "kernel address %p failed with error %d\n",
2076 				  *user_ccb, &io_req->ccb, error);
2077 		}
2078 
2079 		/*
2080 		 * Prefer the first error we got back, and make sure we
2081 		 * don't overwrite bad status with good.
2082 		 */
2083 		if (old_error != 0)
2084 			error = old_error;
2085 
2086 		cam_periph_lock(periph);
2087 
2088 		/*
2089 		 * At this point, if there was an error, we could potentially
2090 		 * re-queue the I/O and try again.  But why?  The error
2091 		 * would almost certainly happen again.  We might as well
2092 		 * not leak memory.
2093 		 */
2094 		uma_zfree(softc->pass_zone, io_req);
2095 		break;
2096 	}
2097 	default:
2098 		error = cam_periph_ioctl(periph, cmd, addr, passerror);
2099 		break;
2100 	}
2101 
2102 bailout:
2103 	cam_periph_unlock(periph);
2104 
2105 	return(error);
2106 }
2107 
2108 static int
2109 passpoll(struct cdev *dev, int poll_events, struct thread *td)
2110 {
2111 	struct cam_periph *periph;
2112 	struct pass_softc *softc;
2113 	int revents;
2114 
2115 	periph = (struct cam_periph *)dev->si_drv1;
2116 	softc = (struct pass_softc *)periph->softc;
2117 
2118 	revents = poll_events & (POLLOUT | POLLWRNORM);
2119 	if ((poll_events & (POLLIN | POLLRDNORM)) != 0) {
2120 		cam_periph_lock(periph);
2121 
2122 		if (!TAILQ_EMPTY(&softc->done_queue)) {
2123 			revents |= poll_events & (POLLIN | POLLRDNORM);
2124 		}
2125 		cam_periph_unlock(periph);
2126 		if (revents == 0)
2127 			selrecord(td, &softc->read_select);
2128 	}
2129 
2130 	return (revents);
2131 }
2132 
2133 static int
2134 passkqfilter(struct cdev *dev, struct knote *kn)
2135 {
2136 	struct cam_periph *periph;
2137 	struct pass_softc *softc;
2138 
2139 	periph = (struct cam_periph *)dev->si_drv1;
2140 	softc = (struct pass_softc *)periph->softc;
2141 
2142 	kn->kn_hook = (caddr_t)periph;
2143 	kn->kn_fop = &passread_filtops;
2144 	knlist_add(&softc->read_select.si_note, kn, 0);
2145 
2146 	return (0);
2147 }
2148 
2149 static void
2150 passreadfiltdetach(struct knote *kn)
2151 {
2152 	struct cam_periph *periph;
2153 	struct pass_softc *softc;
2154 
2155 	periph = (struct cam_periph *)kn->kn_hook;
2156 	softc = (struct pass_softc *)periph->softc;
2157 
2158 	knlist_remove(&softc->read_select.si_note, kn, 0);
2159 }
2160 
2161 static int
2162 passreadfilt(struct knote *kn, long hint)
2163 {
2164 	struct cam_periph *periph;
2165 	struct pass_softc *softc;
2166 	int retval;
2167 
2168 	periph = (struct cam_periph *)kn->kn_hook;
2169 	softc = (struct pass_softc *)periph->softc;
2170 
2171 	cam_periph_assert(periph, MA_OWNED);
2172 
2173 	if (TAILQ_EMPTY(&softc->done_queue))
2174 		retval = 0;
2175 	else
2176 		retval = 1;
2177 
2178 	return (retval);
2179 }
2180 
2181 /*
2182  * Generally, "ccb" should be the CCB supplied by the kernel.  "inccb"
2183  * should be the CCB that is copied in from the user.
2184  */
2185 static int
2186 passsendccb(struct cam_periph *periph, union ccb *ccb, union ccb *inccb)
2187 {
2188 	struct pass_softc *softc;
2189 	struct cam_periph_map_info mapinfo;
2190 	uint8_t *cmd;
2191 	xpt_opcode fc;
2192 	int error;
2193 
2194 	softc = (struct pass_softc *)periph->softc;
2195 
2196 	/*
2197 	 * There are some fields in the CCB header that need to be
2198 	 * preserved, the rest we get from the user.
2199 	 */
2200 	xpt_merge_ccb(ccb, inccb);
2201 
2202 	if (ccb->ccb_h.flags & CAM_CDB_POINTER) {
2203 		cmd = __builtin_alloca(ccb->csio.cdb_len);
2204 		error = copyin(ccb->csio.cdb_io.cdb_ptr, cmd, ccb->csio.cdb_len);
2205 		if (error)
2206 			return (error);
2207 		ccb->csio.cdb_io.cdb_ptr = cmd;
2208 	}
2209 
2210 	/*
2211 	 */
2212 	ccb->ccb_h.cbfcnp = passdone;
2213 
2214 	/*
2215 	 * Let cam_periph_mapmem do a sanity check on the data pointer format.
2216 	 * Even if no data transfer is needed, it's a cheap check and it
2217 	 * simplifies the code.
2218 	 */
2219 	fc = ccb->ccb_h.func_code;
2220 	if ((fc == XPT_SCSI_IO) || (fc == XPT_ATA_IO) || (fc == XPT_SMP_IO)
2221             || (fc == XPT_DEV_MATCH) || (fc == XPT_DEV_ADVINFO) || (fc == XPT_MMC_IO)
2222             || (fc == XPT_NVME_ADMIN) || (fc == XPT_NVME_IO)) {
2223 
2224 		bzero(&mapinfo, sizeof(mapinfo));
2225 
2226 		/*
2227 		 * cam_periph_mapmem calls into proc and vm functions that can
2228 		 * sleep as well as trigger I/O, so we can't hold the lock.
2229 		 * Dropping it here is reasonably safe.
2230 		 */
2231 		cam_periph_unlock(periph);
2232 		error = cam_periph_mapmem(ccb, &mapinfo, softc->maxio);
2233 		cam_periph_lock(periph);
2234 
2235 		/*
2236 		 * cam_periph_mapmem returned an error, we can't continue.
2237 		 * Return the error to the user.
2238 		 */
2239 		if (error)
2240 			return(error);
2241 	} else
2242 		/* Ensure that the unmap call later on is a no-op. */
2243 		mapinfo.num_bufs_used = 0;
2244 
2245 	/*
2246 	 * If the user wants us to perform any error recovery, then honor
2247 	 * that request.  Otherwise, it's up to the user to perform any
2248 	 * error recovery.
2249 	 */
2250 	cam_periph_runccb(ccb, (ccb->ccb_h.flags & CAM_PASS_ERR_RECOVER) ?
2251 	    passerror : NULL, /* cam_flags */ CAM_RETRY_SELTO,
2252 	    /* sense_flags */ SF_RETRY_UA | SF_NO_PRINT,
2253 	    softc->device_stats);
2254 
2255 	cam_periph_unmapmem(ccb, &mapinfo);
2256 
2257 	ccb->ccb_h.cbfcnp = NULL;
2258 	ccb->ccb_h.periph_priv = inccb->ccb_h.periph_priv;
2259 	bcopy(ccb, inccb, sizeof(union ccb));
2260 
2261 	return(0);
2262 }
2263 
2264 static int
2265 passerror(union ccb *ccb, u_int32_t cam_flags, u_int32_t sense_flags)
2266 {
2267 	struct cam_periph *periph;
2268 	struct pass_softc *softc;
2269 
2270 	periph = xpt_path_periph(ccb->ccb_h.path);
2271 	softc = (struct pass_softc *)periph->softc;
2272 
2273 	return(cam_periph_error(ccb, cam_flags, sense_flags));
2274 }
2275