xref: /freebsd/sys/cam/scsi/scsi_pass.c (revision 4d846d26)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
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/sysent.h>
49 #include <sys/taskqueue.h>
50 #include <vm/uma.h>
51 #include <vm/vm.h>
52 #include <vm/vm_extern.h>
53 
54 #include <machine/bus.h>
55 
56 #include <cam/cam.h>
57 #include <cam/cam_ccb.h>
58 #include <cam/cam_periph.h>
59 #include <cam/cam_queue.h>
60 #include <cam/cam_xpt.h>
61 #include <cam/cam_xpt_periph.h>
62 #include <cam/cam_debug.h>
63 #include <cam/cam_compat.h>
64 #include <cam/cam_xpt_periph.h>
65 
66 #include <cam/scsi/scsi_all.h>
67 #include <cam/scsi/scsi_pass.h>
68 
69 typedef enum {
70 	PASS_FLAG_OPEN			= 0x01,
71 	PASS_FLAG_LOCKED		= 0x02,
72 	PASS_FLAG_INVALID		= 0x04,
73 	PASS_FLAG_INITIAL_PHYSPATH	= 0x08,
74 	PASS_FLAG_ZONE_INPROG		= 0x10,
75 	PASS_FLAG_ZONE_VALID		= 0x20,
76 	PASS_FLAG_UNMAPPED_CAPABLE	= 0x40,
77 	PASS_FLAG_ABANDONED_REF_SET	= 0x80
78 } pass_flags;
79 
80 typedef enum {
81 	PASS_STATE_NORMAL
82 } pass_state;
83 
84 typedef enum {
85 	PASS_CCB_BUFFER_IO,
86 	PASS_CCB_QUEUED_IO
87 } pass_ccb_types;
88 
89 #define ccb_type	ppriv_field0
90 #define ccb_ioreq	ppriv_ptr1
91 
92 /*
93  * The maximum number of memory segments we preallocate.
94  */
95 #define	PASS_MAX_SEGS	16
96 
97 typedef enum {
98 	PASS_IO_NONE		= 0x00,
99 	PASS_IO_USER_SEG_MALLOC	= 0x01,
100 	PASS_IO_KERN_SEG_MALLOC	= 0x02,
101 	PASS_IO_ABANDONED	= 0x04
102 } pass_io_flags;
103 
104 struct pass_io_req {
105 	union ccb			 ccb;
106 	union ccb			*alloced_ccb;
107 	union ccb			*user_ccb_ptr;
108 	camq_entry			 user_periph_links;
109 	ccb_ppriv_area			 user_periph_priv;
110 	struct cam_periph_map_info	 mapinfo;
111 	pass_io_flags			 flags;
112 	ccb_flags			 data_flags;
113 	int				 num_user_segs;
114 	bus_dma_segment_t		 user_segs[PASS_MAX_SEGS];
115 	int				 num_kern_segs;
116 	bus_dma_segment_t		 kern_segs[PASS_MAX_SEGS];
117 	bus_dma_segment_t		*user_segptr;
118 	bus_dma_segment_t		*kern_segptr;
119 	int				 num_bufs;
120 	uint32_t			 dirs[CAM_PERIPH_MAXMAPS];
121 	uint32_t			 lengths[CAM_PERIPH_MAXMAPS];
122 	uint8_t				*user_bufs[CAM_PERIPH_MAXMAPS];
123 	uint8_t				*kern_bufs[CAM_PERIPH_MAXMAPS];
124 	struct bintime			 start_time;
125 	TAILQ_ENTRY(pass_io_req)	 links;
126 };
127 
128 struct pass_softc {
129 	pass_state		  state;
130 	pass_flags		  flags;
131 	u_int8_t		  pd_type;
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 	/*
401 	 * It should be safe to destroy the zones from here, because all
402 	 * of the references to this peripheral have been freed, and all
403 	 * I/O has been terminated and freed.  We check the zones for NULL
404 	 * because they may not have been allocated yet if the device went
405 	 * away before any asynchronous I/O has been issued.
406 	 */
407 	if (softc->pass_zone != NULL)
408 		uma_zdestroy(softc->pass_zone);
409 	if (softc->pass_io_zone != NULL)
410 		uma_zdestroy(softc->pass_io_zone);
411 
412 	cam_periph_lock(periph);
413 
414 	free(softc, M_DEVBUF);
415 }
416 
417 static void
418 pass_shutdown_kqueue(void *context, int pending)
419 {
420 	struct cam_periph *periph;
421 	struct pass_softc *softc;
422 
423 	periph = context;
424 	softc = periph->softc;
425 
426 	knlist_clear(&softc->read_select.si_note, /*is_locked*/ 0);
427 	knlist_destroy(&softc->read_select.si_note);
428 
429 	/*
430 	 * Release the reference we held for kqueue.
431 	 */
432 	cam_periph_release(periph);
433 }
434 
435 static void
436 pass_add_physpath(void *context, int pending)
437 {
438 	struct cam_periph *periph;
439 	struct pass_softc *softc;
440 	struct mtx *mtx;
441 	char *physpath;
442 
443 	/*
444 	 * If we have one, create a devfs alias for our
445 	 * physical path.
446 	 */
447 	periph = context;
448 	softc = periph->softc;
449 	physpath = malloc(MAXPATHLEN, M_DEVBUF, M_WAITOK);
450 	mtx = cam_periph_mtx(periph);
451 	mtx_lock(mtx);
452 
453 	if (periph->flags & CAM_PERIPH_INVALID)
454 		goto out;
455 
456 	if (xpt_getattr(physpath, MAXPATHLEN,
457 			"GEOM::physpath", periph->path) == 0
458 	 && strlen(physpath) != 0) {
459 		mtx_unlock(mtx);
460 		make_dev_physpath_alias(MAKEDEV_WAITOK | MAKEDEV_CHECKNAME,
461 				&softc->alias_dev, softc->dev,
462 				softc->alias_dev, physpath);
463 		mtx_lock(mtx);
464 	}
465 
466 out:
467 	/*
468 	 * Now that we've made our alias, we no longer have to have a
469 	 * reference to the device.
470 	 */
471 	if ((softc->flags & PASS_FLAG_INITIAL_PHYSPATH) == 0)
472 		softc->flags |= PASS_FLAG_INITIAL_PHYSPATH;
473 
474 	/*
475 	 * We always acquire a reference to the periph before queueing this
476 	 * task queue function, so it won't go away before we run.
477 	 */
478 	while (pending-- > 0)
479 		cam_periph_release_locked(periph);
480 	mtx_unlock(mtx);
481 
482 	free(physpath, M_DEVBUF);
483 }
484 
485 static void
486 passasync(void *callback_arg, u_int32_t code,
487 	  struct cam_path *path, void *arg)
488 {
489 	struct cam_periph *periph;
490 
491 	periph = (struct cam_periph *)callback_arg;
492 
493 	switch (code) {
494 	case AC_FOUND_DEVICE:
495 	{
496 		struct ccb_getdev *cgd;
497 		cam_status status;
498 
499 		cgd = (struct ccb_getdev *)arg;
500 		if (cgd == NULL)
501 			break;
502 
503 		/*
504 		 * Allocate a peripheral instance for
505 		 * this device and start the probe
506 		 * process.
507 		 */
508 		status = cam_periph_alloc(passregister, passoninvalidate,
509 					  passcleanup, passstart, "pass",
510 					  CAM_PERIPH_BIO, path,
511 					  passasync, AC_FOUND_DEVICE, cgd);
512 
513 		if (status != CAM_REQ_CMP
514 		 && status != CAM_REQ_INPROG) {
515 			const struct cam_status_entry *entry;
516 
517 			entry = cam_fetch_status_entry(status);
518 
519 			printf("passasync: Unable to attach new device "
520 			       "due to status %#x: %s\n", status, entry ?
521 			       entry->status_text : "Unknown");
522 		}
523 
524 		break;
525 	}
526 	case AC_ADVINFO_CHANGED:
527 	{
528 		uintptr_t buftype;
529 
530 		buftype = (uintptr_t)arg;
531 		if (buftype == CDAI_TYPE_PHYS_PATH) {
532 			struct pass_softc *softc;
533 
534 			softc = (struct pass_softc *)periph->softc;
535 			/*
536 			 * Acquire a reference to the periph before we
537 			 * start the taskqueue, so that we don't run into
538 			 * a situation where the periph goes away before
539 			 * the task queue has a chance to run.
540 			 */
541 			if (cam_periph_acquire(periph) != 0)
542 				break;
543 
544 			taskqueue_enqueue(taskqueue_thread,
545 					  &softc->add_physpath_task);
546 		}
547 		break;
548 	}
549 	default:
550 		cam_periph_async(periph, code, path, arg);
551 		break;
552 	}
553 }
554 
555 static cam_status
556 passregister(struct cam_periph *periph, void *arg)
557 {
558 	struct pass_softc *softc;
559 	struct ccb_getdev *cgd;
560 	struct ccb_pathinq cpi;
561 	struct make_dev_args args;
562 	int error, no_tags;
563 
564 	cgd = (struct ccb_getdev *)arg;
565 	if (cgd == NULL) {
566 		printf("%s: no getdev CCB, can't register device\n", __func__);
567 		return(CAM_REQ_CMP_ERR);
568 	}
569 
570 	softc = (struct pass_softc *)malloc(sizeof(*softc),
571 					    M_DEVBUF, M_NOWAIT);
572 
573 	if (softc == NULL) {
574 		printf("%s: Unable to probe new device. "
575 		       "Unable to allocate softc\n", __func__);
576 		return(CAM_REQ_CMP_ERR);
577 	}
578 
579 	bzero(softc, sizeof(*softc));
580 	softc->state = PASS_STATE_NORMAL;
581 	if (cgd->protocol == PROTO_SCSI || cgd->protocol == PROTO_ATAPI)
582 		softc->pd_type = SID_TYPE(&cgd->inq_data);
583 	else if (cgd->protocol == PROTO_SATAPM)
584 		softc->pd_type = T_ENCLOSURE;
585 	else
586 		softc->pd_type = T_DIRECT;
587 
588 	periph->softc = softc;
589 	softc->periph = periph;
590 	TAILQ_INIT(&softc->incoming_queue);
591 	TAILQ_INIT(&softc->active_queue);
592 	TAILQ_INIT(&softc->abandoned_queue);
593 	TAILQ_INIT(&softc->done_queue);
594 	snprintf(softc->zone_name, sizeof(softc->zone_name), "%s%d",
595 		 periph->periph_name, periph->unit_number);
596 	snprintf(softc->io_zone_name, sizeof(softc->io_zone_name), "%s%dIO",
597 		 periph->periph_name, periph->unit_number);
598 	softc->io_zone_size = maxphys;
599 	knlist_init_mtx(&softc->read_select.si_note, cam_periph_mtx(periph));
600 
601 	xpt_path_inq(&cpi, periph->path);
602 
603 	if (cpi.maxio == 0)
604 		softc->maxio = DFLTPHYS;	/* traditional default */
605 	else if (cpi.maxio > maxphys)
606 		softc->maxio = maxphys;		/* for safety */
607 	else
608 		softc->maxio = cpi.maxio;	/* real value */
609 
610 	if (cpi.hba_misc & PIM_UNMAPPED)
611 		softc->flags |= PASS_FLAG_UNMAPPED_CAPABLE;
612 
613 	/*
614 	 * We pass in 0 for a blocksize, since we don't
615 	 * know what the blocksize of this device is, if
616 	 * it even has a blocksize.
617 	 */
618 	cam_periph_unlock(periph);
619 	no_tags = (cgd->inq_data.flags & SID_CmdQue) == 0;
620 	softc->device_stats = devstat_new_entry("pass",
621 			  periph->unit_number, 0,
622 			  DEVSTAT_NO_BLOCKSIZE
623 			  | (no_tags ? DEVSTAT_NO_ORDERED_TAGS : 0),
624 			  softc->pd_type |
625 			  XPORT_DEVSTAT_TYPE(cpi.transport) |
626 			  DEVSTAT_TYPE_PASS,
627 			  DEVSTAT_PRIORITY_PASS);
628 
629 	/*
630 	 * Initialize the taskqueue handler for shutting down kqueue.
631 	 */
632 	TASK_INIT(&softc->shutdown_kqueue_task, /*priority*/ 0,
633 		  pass_shutdown_kqueue, periph);
634 
635 	/*
636 	 * Acquire a reference to the periph that we can release once we've
637 	 * cleaned up the kqueue.
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 	/*
647 	 * Acquire a reference to the periph before we create the devfs
648 	 * instance for it.  We'll release this reference once the devfs
649 	 * instance has been freed.
650 	 */
651 	if (cam_periph_acquire(periph) != 0) {
652 		xpt_print(periph->path, "%s: lost periph during "
653 			  "registration!\n", __func__);
654 		cam_periph_lock(periph);
655 		return (CAM_REQ_CMP_ERR);
656 	}
657 
658 	/* Register the device */
659 	make_dev_args_init(&args);
660 	args.mda_devsw = &pass_cdevsw;
661 	args.mda_unit = periph->unit_number;
662 	args.mda_uid = UID_ROOT;
663 	args.mda_gid = GID_OPERATOR;
664 	args.mda_mode = 0600;
665 	args.mda_si_drv1 = periph;
666 	args.mda_flags = MAKEDEV_NOWAIT;
667 	error = make_dev_s(&args, &softc->dev, "%s%d", periph->periph_name,
668 	    periph->unit_number);
669 	if (error != 0) {
670 		cam_periph_lock(periph);
671 		cam_periph_release_locked(periph);
672 		return (CAM_REQ_CMP_ERR);
673 	}
674 
675 	/*
676 	 * Hold a reference to the periph before we create the physical
677 	 * path alias so it can't go away.
678 	 */
679 	if (cam_periph_acquire(periph) != 0) {
680 		xpt_print(periph->path, "%s: lost periph during "
681 			  "registration!\n", __func__);
682 		cam_periph_lock(periph);
683 		return (CAM_REQ_CMP_ERR);
684 	}
685 
686 	cam_periph_lock(periph);
687 
688 	TASK_INIT(&softc->add_physpath_task, /*priority*/0,
689 		  pass_add_physpath, periph);
690 
691 	/*
692 	 * See if physical path information is already available.
693 	 */
694 	taskqueue_enqueue(taskqueue_thread, &softc->add_physpath_task);
695 
696 	/*
697 	 * Add an async callback so that we get notified if
698 	 * this device goes away or its physical path
699 	 * (stored in the advanced info data of the EDT) has
700 	 * changed.
701 	 */
702 	xpt_register_async(AC_LOST_DEVICE | AC_ADVINFO_CHANGED,
703 			   passasync, periph, periph->path);
704 
705 	if (bootverbose)
706 		xpt_announce_periph(periph, NULL);
707 
708 	return(CAM_REQ_CMP);
709 }
710 
711 static int
712 passopen(struct cdev *dev, int flags, int fmt, struct thread *td)
713 {
714 	struct cam_periph *periph;
715 	struct pass_softc *softc;
716 	int error;
717 
718 	periph = (struct cam_periph *)dev->si_drv1;
719 	if (cam_periph_acquire(periph) != 0)
720 		return (ENXIO);
721 
722 	cam_periph_lock(periph);
723 
724 	softc = (struct pass_softc *)periph->softc;
725 
726 	if (softc->flags & PASS_FLAG_INVALID) {
727 		cam_periph_release_locked(periph);
728 		cam_periph_unlock(periph);
729 		return(ENXIO);
730 	}
731 
732 	/*
733 	 * Don't allow access when we're running at a high securelevel.
734 	 */
735 	error = securelevel_gt(td->td_ucred, 1);
736 	if (error) {
737 		cam_periph_release_locked(periph);
738 		cam_periph_unlock(periph);
739 		return(error);
740 	}
741 
742 	/*
743 	 * Only allow read-write access.
744 	 */
745 	if (((flags & FWRITE) == 0) || ((flags & FREAD) == 0)) {
746 		cam_periph_release_locked(periph);
747 		cam_periph_unlock(periph);
748 		return(EPERM);
749 	}
750 
751 	/*
752 	 * We don't allow nonblocking access.
753 	 */
754 	if ((flags & O_NONBLOCK) != 0) {
755 		xpt_print(periph->path, "can't do nonblocking access\n");
756 		cam_periph_release_locked(periph);
757 		cam_periph_unlock(periph);
758 		return(EINVAL);
759 	}
760 
761 	softc->open_count++;
762 
763 	cam_periph_unlock(periph);
764 
765 	return (error);
766 }
767 
768 static int
769 passclose(struct cdev *dev, int flag, int fmt, struct thread *td)
770 {
771 	struct 	cam_periph *periph;
772 	struct  pass_softc *softc;
773 	struct mtx *mtx;
774 
775 	periph = (struct cam_periph *)dev->si_drv1;
776 	mtx = cam_periph_mtx(periph);
777 	mtx_lock(mtx);
778 
779 	softc = periph->softc;
780 	softc->open_count--;
781 
782 	if (softc->open_count == 0) {
783 		struct pass_io_req *io_req, *io_req2;
784 
785 		TAILQ_FOREACH_SAFE(io_req, &softc->done_queue, links, io_req2) {
786 			TAILQ_REMOVE(&softc->done_queue, io_req, links);
787 			passiocleanup(softc, io_req);
788 			uma_zfree(softc->pass_zone, io_req);
789 		}
790 
791 		TAILQ_FOREACH_SAFE(io_req, &softc->incoming_queue, links,
792 				   io_req2) {
793 			TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
794 			passiocleanup(softc, io_req);
795 			uma_zfree(softc->pass_zone, io_req);
796 		}
797 
798 		/*
799 		 * If there are any active I/Os, we need to forcibly acquire a
800 		 * reference to the peripheral so that we don't go away
801 		 * before they complete.  We'll release the reference when
802 		 * the abandoned queue is empty.
803 		 */
804 		io_req = TAILQ_FIRST(&softc->active_queue);
805 		if ((io_req != NULL)
806 		 && (softc->flags & PASS_FLAG_ABANDONED_REF_SET) == 0) {
807 			cam_periph_doacquire(periph);
808 			softc->flags |= PASS_FLAG_ABANDONED_REF_SET;
809 		}
810 
811 		/*
812 		 * Since the I/O in the active queue is not under our
813 		 * control, just set a flag so that we can clean it up when
814 		 * it completes and put it on the abandoned queue.  This
815 		 * will prevent our sending spurious completions in the
816 		 * event that the device is opened again before these I/Os
817 		 * complete.
818 		 */
819 		TAILQ_FOREACH_SAFE(io_req, &softc->active_queue, links,
820 				   io_req2) {
821 			TAILQ_REMOVE(&softc->active_queue, io_req, links);
822 			io_req->flags |= PASS_IO_ABANDONED;
823 			TAILQ_INSERT_TAIL(&softc->abandoned_queue, io_req,
824 					  links);
825 		}
826 	}
827 
828 	cam_periph_release_locked(periph);
829 
830 	/*
831 	 * We reference the lock directly here, instead of using
832 	 * cam_periph_unlock().  The reason is that the call to
833 	 * cam_periph_release_locked() above could result in the periph
834 	 * getting freed.  If that is the case, dereferencing the periph
835 	 * with a cam_periph_unlock() call would cause a page fault.
836 	 *
837 	 * cam_periph_release() avoids this problem using the same method,
838 	 * but we're manually acquiring and dropping the lock here to
839 	 * protect the open count and avoid another lock acquisition and
840 	 * release.
841 	 */
842 	mtx_unlock(mtx);
843 
844 	return (0);
845 }
846 
847 static void
848 passstart(struct cam_periph *periph, union ccb *start_ccb)
849 {
850 	struct pass_softc *softc;
851 
852 	softc = (struct pass_softc *)periph->softc;
853 
854 	switch (softc->state) {
855 	case PASS_STATE_NORMAL: {
856 		struct pass_io_req *io_req;
857 
858 		/*
859 		 * Check for any queued I/O requests that require an
860 		 * allocated slot.
861 		 */
862 		io_req = TAILQ_FIRST(&softc->incoming_queue);
863 		if (io_req == NULL) {
864 			xpt_release_ccb(start_ccb);
865 			break;
866 		}
867 		TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
868 		TAILQ_INSERT_TAIL(&softc->active_queue, io_req, links);
869 		/*
870 		 * Merge the user's CCB into the allocated CCB.
871 		 */
872 		xpt_merge_ccb(start_ccb, &io_req->ccb);
873 		start_ccb->ccb_h.ccb_type = PASS_CCB_QUEUED_IO;
874 		start_ccb->ccb_h.ccb_ioreq = io_req;
875 		start_ccb->ccb_h.cbfcnp = passdone;
876 		io_req->alloced_ccb = start_ccb;
877 		binuptime(&io_req->start_time);
878 		devstat_start_transaction(softc->device_stats,
879 					  &io_req->start_time);
880 
881 		xpt_action(start_ccb);
882 
883 		/*
884 		 * If we have any more I/O waiting, schedule ourselves again.
885 		 */
886 		if (!TAILQ_EMPTY(&softc->incoming_queue))
887 			xpt_schedule(periph, CAM_PRIORITY_NORMAL);
888 		break;
889 	}
890 	default:
891 		break;
892 	}
893 }
894 
895 static void
896 passdone(struct cam_periph *periph, union ccb *done_ccb)
897 {
898 	struct pass_softc *softc;
899 	struct ccb_scsiio *csio;
900 
901 	softc = (struct pass_softc *)periph->softc;
902 
903 	cam_periph_assert(periph, MA_OWNED);
904 
905 	csio = &done_ccb->csio;
906 	switch (csio->ccb_h.ccb_type) {
907 	case PASS_CCB_QUEUED_IO: {
908 		struct pass_io_req *io_req;
909 
910 		io_req = done_ccb->ccb_h.ccb_ioreq;
911 #if 0
912 		xpt_print(periph->path, "%s: called for user CCB %p\n",
913 			  __func__, io_req->user_ccb_ptr);
914 #endif
915 		if (((done_ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_CMP)
916 		 && (done_ccb->ccb_h.flags & CAM_PASS_ERR_RECOVER)
917 		 && ((io_req->flags & PASS_IO_ABANDONED) == 0)) {
918 			int error;
919 
920 			error = passerror(done_ccb, CAM_RETRY_SELTO,
921 					  SF_RETRY_UA | SF_NO_PRINT);
922 
923 			if (error == ERESTART) {
924 				/*
925 				 * A retry was scheduled, so
926  				 * just return.
927 				 */
928 				return;
929 			}
930 		}
931 
932 		/*
933 		 * Copy the allocated CCB contents back to the malloced CCB
934 		 * so we can give status back to the user when he requests it.
935 		 */
936 		bcopy(done_ccb, &io_req->ccb, sizeof(*done_ccb));
937 
938 		/*
939 		 * Log data/transaction completion with devstat(9).
940 		 */
941 		switch (done_ccb->ccb_h.func_code) {
942 		case XPT_SCSI_IO:
943 			devstat_end_transaction(softc->device_stats,
944 			    done_ccb->csio.dxfer_len - done_ccb->csio.resid,
945 			    done_ccb->csio.tag_action & 0x3,
946 			    ((done_ccb->ccb_h.flags & CAM_DIR_MASK) ==
947 			    CAM_DIR_NONE) ? DEVSTAT_NO_DATA :
948 			    (done_ccb->ccb_h.flags & CAM_DIR_OUT) ?
949 			    DEVSTAT_WRITE : DEVSTAT_READ, NULL,
950 			    &io_req->start_time);
951 			break;
952 		case XPT_ATA_IO:
953 			devstat_end_transaction(softc->device_stats,
954 			    done_ccb->ataio.dxfer_len - done_ccb->ataio.resid,
955 			    0, /* Not used in ATA */
956 			    ((done_ccb->ccb_h.flags & CAM_DIR_MASK) ==
957 			    CAM_DIR_NONE) ? DEVSTAT_NO_DATA :
958 			    (done_ccb->ccb_h.flags & CAM_DIR_OUT) ?
959 			    DEVSTAT_WRITE : DEVSTAT_READ, NULL,
960 			    &io_req->start_time);
961 			break;
962 		case XPT_SMP_IO:
963 			/*
964 			 * XXX KDM this isn't quite right, but there isn't
965 			 * currently an easy way to represent a bidirectional
966 			 * transfer in devstat.  The only way to do it
967 			 * and have the byte counts come out right would
968 			 * mean that we would have to record two
969 			 * transactions, one for the request and one for the
970 			 * response.  For now, so that we report something,
971 			 * just treat the entire thing as a read.
972 			 */
973 			devstat_end_transaction(softc->device_stats,
974 			    done_ccb->smpio.smp_request_len +
975 			    done_ccb->smpio.smp_response_len,
976 			    DEVSTAT_TAG_SIMPLE, DEVSTAT_READ, NULL,
977 			    &io_req->start_time);
978 			break;
979 		default:
980 			devstat_end_transaction(softc->device_stats, 0,
981 			    DEVSTAT_TAG_NONE, DEVSTAT_NO_DATA, NULL,
982 			    &io_req->start_time);
983 			break;
984 		}
985 
986 		/*
987 		 * In the normal case, take the completed I/O off of the
988 		 * active queue and put it on the done queue.  Notitfy the
989 		 * user that we have a completed I/O.
990 		 */
991 		if ((io_req->flags & PASS_IO_ABANDONED) == 0) {
992 			TAILQ_REMOVE(&softc->active_queue, io_req, links);
993 			TAILQ_INSERT_TAIL(&softc->done_queue, io_req, links);
994 			selwakeuppri(&softc->read_select, PRIBIO);
995 			KNOTE_LOCKED(&softc->read_select.si_note, 0);
996 		} else {
997 			/*
998 			 * In the case of an abandoned I/O (final close
999 			 * without fetching the I/O), take it off of the
1000 			 * abandoned queue and free it.
1001 			 */
1002 			TAILQ_REMOVE(&softc->abandoned_queue, io_req, links);
1003 			passiocleanup(softc, io_req);
1004 			uma_zfree(softc->pass_zone, io_req);
1005 
1006 			/*
1007 			 * Release the done_ccb here, since we may wind up
1008 			 * freeing the peripheral when we decrement the
1009 			 * reference count below.
1010 			 */
1011 			xpt_release_ccb(done_ccb);
1012 
1013 			/*
1014 			 * If the abandoned queue is empty, we can release
1015 			 * our reference to the periph since we won't have
1016 			 * any more completions coming.
1017 			 */
1018 			if ((TAILQ_EMPTY(&softc->abandoned_queue))
1019 			 && (softc->flags & PASS_FLAG_ABANDONED_REF_SET)) {
1020 				softc->flags &= ~PASS_FLAG_ABANDONED_REF_SET;
1021 				cam_periph_release_locked(periph);
1022 			}
1023 
1024 			/*
1025 			 * We have already released the CCB, so we can
1026 			 * return.
1027 			 */
1028 			return;
1029 		}
1030 		break;
1031 	}
1032 	}
1033 	xpt_release_ccb(done_ccb);
1034 }
1035 
1036 static int
1037 passcreatezone(struct cam_periph *periph)
1038 {
1039 	struct pass_softc *softc;
1040 	int error;
1041 
1042 	error = 0;
1043 	softc = (struct pass_softc *)periph->softc;
1044 
1045 	cam_periph_assert(periph, MA_OWNED);
1046 	KASSERT(((softc->flags & PASS_FLAG_ZONE_VALID) == 0),
1047 		("%s called when the pass(4) zone is valid!\n", __func__));
1048 	KASSERT((softc->pass_zone == NULL),
1049 		("%s called when the pass(4) zone is allocated!\n", __func__));
1050 
1051 	if ((softc->flags & PASS_FLAG_ZONE_INPROG) == 0) {
1052 		/*
1053 		 * We're the first context through, so we need to create
1054 		 * the pass(4) UMA zone for I/O requests.
1055 		 */
1056 		softc->flags |= PASS_FLAG_ZONE_INPROG;
1057 
1058 		/*
1059 		 * uma_zcreate() does a blocking (M_WAITOK) allocation,
1060 		 * so we cannot hold a mutex while we call it.
1061 		 */
1062 		cam_periph_unlock(periph);
1063 
1064 		softc->pass_zone = uma_zcreate(softc->zone_name,
1065 		    sizeof(struct pass_io_req), NULL, NULL, NULL, NULL,
1066 		    /*align*/ 0, /*flags*/ 0);
1067 
1068 		softc->pass_io_zone = uma_zcreate(softc->io_zone_name,
1069 		    softc->io_zone_size, NULL, NULL, NULL, NULL,
1070 		    /*align*/ 0, /*flags*/ 0);
1071 
1072 		cam_periph_lock(periph);
1073 
1074 		if ((softc->pass_zone == NULL)
1075 		 || (softc->pass_io_zone == NULL)) {
1076 			if (softc->pass_zone == NULL)
1077 				xpt_print(periph->path, "unable to allocate "
1078 				    "IO Req UMA zone\n");
1079 			else
1080 				xpt_print(periph->path, "unable to allocate "
1081 				    "IO UMA zone\n");
1082 			softc->flags &= ~PASS_FLAG_ZONE_INPROG;
1083 			goto bailout;
1084 		}
1085 
1086 		/*
1087 		 * Set the flags appropriately and notify any other waiters.
1088 		 */
1089 		softc->flags &= ~PASS_FLAG_ZONE_INPROG;
1090 		softc->flags |= PASS_FLAG_ZONE_VALID;
1091 		wakeup(&softc->pass_zone);
1092 	} else {
1093 		/*
1094 		 * In this case, the UMA zone has not yet been created, but
1095 		 * another context is in the process of creating it.  We
1096 		 * need to sleep until the creation is either done or has
1097 		 * failed.
1098 		 */
1099 		while ((softc->flags & PASS_FLAG_ZONE_INPROG)
1100 		    && ((softc->flags & PASS_FLAG_ZONE_VALID) == 0)) {
1101 			error = msleep(&softc->pass_zone,
1102 				       cam_periph_mtx(periph), PRIBIO,
1103 				       "paszon", 0);
1104 			if (error != 0)
1105 				goto bailout;
1106 		}
1107 		/*
1108 		 * If the zone creation failed, no luck for the user.
1109 		 */
1110 		if ((softc->flags & PASS_FLAG_ZONE_VALID) == 0){
1111 			error = ENOMEM;
1112 			goto bailout;
1113 		}
1114 	}
1115 bailout:
1116 	return (error);
1117 }
1118 
1119 static void
1120 passiocleanup(struct pass_softc *softc, struct pass_io_req *io_req)
1121 {
1122 	union ccb *ccb;
1123 	u_int8_t **data_ptrs[CAM_PERIPH_MAXMAPS];
1124 	int i, numbufs;
1125 
1126 	ccb = &io_req->ccb;
1127 
1128 	switch (ccb->ccb_h.func_code) {
1129 	case XPT_DEV_MATCH:
1130 		numbufs = min(io_req->num_bufs, 2);
1131 
1132 		if (numbufs == 1) {
1133 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.matches;
1134 		} else {
1135 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.patterns;
1136 			data_ptrs[1] = (u_int8_t **)&ccb->cdm.matches;
1137 		}
1138 		break;
1139 	case XPT_SCSI_IO:
1140 	case XPT_CONT_TARGET_IO:
1141 		data_ptrs[0] = &ccb->csio.data_ptr;
1142 		numbufs = min(io_req->num_bufs, 1);
1143 		break;
1144 	case XPT_ATA_IO:
1145 		data_ptrs[0] = &ccb->ataio.data_ptr;
1146 		numbufs = min(io_req->num_bufs, 1);
1147 		break;
1148 	case XPT_SMP_IO:
1149 		numbufs = min(io_req->num_bufs, 2);
1150 		data_ptrs[0] = &ccb->smpio.smp_request;
1151 		data_ptrs[1] = &ccb->smpio.smp_response;
1152 		break;
1153 	case XPT_DEV_ADVINFO:
1154 		numbufs = min(io_req->num_bufs, 1);
1155 		data_ptrs[0] = (uint8_t **)&ccb->cdai.buf;
1156 		break;
1157 	case XPT_NVME_IO:
1158 	case XPT_NVME_ADMIN:
1159 		data_ptrs[0] = &ccb->nvmeio.data_ptr;
1160 		numbufs = min(io_req->num_bufs, 1);
1161 		break;
1162 	default:
1163 		/* allow ourselves to be swapped once again */
1164 		return;
1165 		break; /* NOTREACHED */
1166 	}
1167 
1168 	if (io_req->flags & PASS_IO_USER_SEG_MALLOC) {
1169 		free(io_req->user_segptr, M_SCSIPASS);
1170 		io_req->user_segptr = NULL;
1171 	}
1172 
1173 	/*
1174 	 * We only want to free memory we malloced.
1175 	 */
1176 	if (io_req->data_flags == CAM_DATA_VADDR) {
1177 		for (i = 0; i < io_req->num_bufs; i++) {
1178 			if (io_req->kern_bufs[i] == NULL)
1179 				continue;
1180 
1181 			free(io_req->kern_bufs[i], M_SCSIPASS);
1182 			io_req->kern_bufs[i] = NULL;
1183 		}
1184 	} else if (io_req->data_flags == CAM_DATA_SG) {
1185 		for (i = 0; i < io_req->num_kern_segs; i++) {
1186 			if ((uint8_t *)(uintptr_t)
1187 			    io_req->kern_segptr[i].ds_addr == NULL)
1188 				continue;
1189 
1190 			uma_zfree(softc->pass_io_zone, (uint8_t *)(uintptr_t)
1191 			    io_req->kern_segptr[i].ds_addr);
1192 			io_req->kern_segptr[i].ds_addr = 0;
1193 		}
1194 	}
1195 
1196 	if (io_req->flags & PASS_IO_KERN_SEG_MALLOC) {
1197 		free(io_req->kern_segptr, M_SCSIPASS);
1198 		io_req->kern_segptr = NULL;
1199 	}
1200 
1201 	if (io_req->data_flags != CAM_DATA_PADDR) {
1202 		for (i = 0; i < numbufs; i++) {
1203 			/*
1204 			 * Restore the user's buffer pointers to their
1205 			 * previous values.
1206 			 */
1207 			if (io_req->user_bufs[i] != NULL)
1208 				*data_ptrs[i] = io_req->user_bufs[i];
1209 		}
1210 	}
1211 
1212 }
1213 
1214 static int
1215 passcopysglist(struct cam_periph *periph, struct pass_io_req *io_req,
1216 	       ccb_flags direction)
1217 {
1218 	bus_size_t kern_watermark, user_watermark, len_to_copy;
1219 	bus_dma_segment_t *user_sglist, *kern_sglist;
1220 	int i, j, error;
1221 
1222 	error = 0;
1223 	kern_watermark = 0;
1224 	user_watermark = 0;
1225 	len_to_copy = 0;
1226 	user_sglist = io_req->user_segptr;
1227 	kern_sglist = io_req->kern_segptr;
1228 
1229 	for (i = 0, j = 0; i < io_req->num_user_segs &&
1230 	     j < io_req->num_kern_segs;) {
1231 		uint8_t *user_ptr, *kern_ptr;
1232 
1233 		len_to_copy = min(user_sglist[i].ds_len -user_watermark,
1234 		    kern_sglist[j].ds_len - kern_watermark);
1235 
1236 		user_ptr = (uint8_t *)(uintptr_t)user_sglist[i].ds_addr;
1237 		user_ptr = user_ptr + user_watermark;
1238 		kern_ptr = (uint8_t *)(uintptr_t)kern_sglist[j].ds_addr;
1239 		kern_ptr = kern_ptr + kern_watermark;
1240 
1241 		user_watermark += len_to_copy;
1242 		kern_watermark += len_to_copy;
1243 
1244 		if (direction == CAM_DIR_IN) {
1245 			error = copyout(kern_ptr, user_ptr, len_to_copy);
1246 			if (error != 0) {
1247 				xpt_print(periph->path, "%s: copyout of %u "
1248 					  "bytes from %p to %p failed with "
1249 					  "error %d\n", __func__, len_to_copy,
1250 					  kern_ptr, user_ptr, error);
1251 				goto bailout;
1252 			}
1253 		} else {
1254 			error = copyin(user_ptr, kern_ptr, len_to_copy);
1255 			if (error != 0) {
1256 				xpt_print(periph->path, "%s: copyin of %u "
1257 					  "bytes from %p to %p failed with "
1258 					  "error %d\n", __func__, len_to_copy,
1259 					  user_ptr, kern_ptr, error);
1260 				goto bailout;
1261 			}
1262 		}
1263 
1264 		if (user_sglist[i].ds_len == user_watermark) {
1265 			i++;
1266 			user_watermark = 0;
1267 		}
1268 
1269 		if (kern_sglist[j].ds_len == kern_watermark) {
1270 			j++;
1271 			kern_watermark = 0;
1272 		}
1273 	}
1274 
1275 bailout:
1276 
1277 	return (error);
1278 }
1279 
1280 static int
1281 passmemsetup(struct cam_periph *periph, struct pass_io_req *io_req)
1282 {
1283 	union ccb *ccb;
1284 	struct pass_softc *softc;
1285 	int numbufs, i;
1286 	uint8_t **data_ptrs[CAM_PERIPH_MAXMAPS];
1287 	uint32_t lengths[CAM_PERIPH_MAXMAPS];
1288 	uint32_t dirs[CAM_PERIPH_MAXMAPS];
1289 	uint32_t num_segs;
1290 	uint16_t *seg_cnt_ptr;
1291 	size_t maxmap;
1292 	int error;
1293 
1294 	cam_periph_assert(periph, MA_NOTOWNED);
1295 
1296 	softc = periph->softc;
1297 
1298 	error = 0;
1299 	ccb = &io_req->ccb;
1300 	maxmap = 0;
1301 	num_segs = 0;
1302 	seg_cnt_ptr = NULL;
1303 
1304 	switch(ccb->ccb_h.func_code) {
1305 	case XPT_DEV_MATCH:
1306 		if (ccb->cdm.match_buf_len == 0) {
1307 			printf("%s: invalid match buffer length 0\n", __func__);
1308 			return(EINVAL);
1309 		}
1310 		if (ccb->cdm.pattern_buf_len > 0) {
1311 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.patterns;
1312 			lengths[0] = ccb->cdm.pattern_buf_len;
1313 			dirs[0] = CAM_DIR_OUT;
1314 			data_ptrs[1] = (u_int8_t **)&ccb->cdm.matches;
1315 			lengths[1] = ccb->cdm.match_buf_len;
1316 			dirs[1] = CAM_DIR_IN;
1317 			numbufs = 2;
1318 		} else {
1319 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.matches;
1320 			lengths[0] = ccb->cdm.match_buf_len;
1321 			dirs[0] = CAM_DIR_IN;
1322 			numbufs = 1;
1323 		}
1324 		io_req->data_flags = CAM_DATA_VADDR;
1325 		break;
1326 	case XPT_SCSI_IO:
1327 	case XPT_CONT_TARGET_IO:
1328 		if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
1329 			return(0);
1330 
1331 		/*
1332 		 * The user shouldn't be able to supply a bio.
1333 		 */
1334 		if ((ccb->ccb_h.flags & CAM_DATA_MASK) == CAM_DATA_BIO)
1335 			return (EINVAL);
1336 
1337 		io_req->data_flags = ccb->ccb_h.flags & CAM_DATA_MASK;
1338 
1339 		data_ptrs[0] = &ccb->csio.data_ptr;
1340 		lengths[0] = ccb->csio.dxfer_len;
1341 		dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1342 		num_segs = ccb->csio.sglist_cnt;
1343 		seg_cnt_ptr = &ccb->csio.sglist_cnt;
1344 		numbufs = 1;
1345 		maxmap = softc->maxio;
1346 		break;
1347 	case XPT_ATA_IO:
1348 		if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
1349 			return(0);
1350 
1351 		/*
1352 		 * We only support a single virtual address for ATA I/O.
1353 		 */
1354 		if ((ccb->ccb_h.flags & CAM_DATA_MASK) != CAM_DATA_VADDR)
1355 			return (EINVAL);
1356 
1357 		io_req->data_flags = CAM_DATA_VADDR;
1358 
1359 		data_ptrs[0] = &ccb->ataio.data_ptr;
1360 		lengths[0] = ccb->ataio.dxfer_len;
1361 		dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1362 		numbufs = 1;
1363 		maxmap = softc->maxio;
1364 		break;
1365 	case XPT_SMP_IO:
1366 		io_req->data_flags = CAM_DATA_VADDR;
1367 
1368 		data_ptrs[0] = &ccb->smpio.smp_request;
1369 		lengths[0] = ccb->smpio.smp_request_len;
1370 		dirs[0] = CAM_DIR_OUT;
1371 		data_ptrs[1] = &ccb->smpio.smp_response;
1372 		lengths[1] = ccb->smpio.smp_response_len;
1373 		dirs[1] = CAM_DIR_IN;
1374 		numbufs = 2;
1375 		maxmap = softc->maxio;
1376 		break;
1377 	case XPT_DEV_ADVINFO:
1378 		if (ccb->cdai.bufsiz == 0)
1379 			return (0);
1380 
1381 		io_req->data_flags = CAM_DATA_VADDR;
1382 
1383 		data_ptrs[0] = (uint8_t **)&ccb->cdai.buf;
1384 		lengths[0] = ccb->cdai.bufsiz;
1385 		dirs[0] = CAM_DIR_IN;
1386 		numbufs = 1;
1387 		break;
1388 	case XPT_NVME_ADMIN:
1389 	case XPT_NVME_IO:
1390 		if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
1391 			return (0);
1392 
1393 		io_req->data_flags = ccb->ccb_h.flags & CAM_DATA_MASK;
1394 
1395 		data_ptrs[0] = &ccb->nvmeio.data_ptr;
1396 		lengths[0] = ccb->nvmeio.dxfer_len;
1397 		dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1398 		num_segs = ccb->nvmeio.sglist_cnt;
1399 		seg_cnt_ptr = &ccb->nvmeio.sglist_cnt;
1400 		numbufs = 1;
1401 		maxmap = softc->maxio;
1402 		break;
1403 	default:
1404 		return(EINVAL);
1405 		break; /* NOTREACHED */
1406 	}
1407 
1408 	io_req->num_bufs = numbufs;
1409 
1410 	/*
1411 	 * If there is a maximum, check to make sure that the user's
1412 	 * request fits within the limit.  In general, we should only have
1413 	 * a maximum length for requests that go to hardware.  Otherwise it
1414 	 * is whatever we're able to malloc.
1415 	 */
1416 	for (i = 0; i < numbufs; i++) {
1417 		io_req->user_bufs[i] = *data_ptrs[i];
1418 		io_req->dirs[i] = dirs[i];
1419 		io_req->lengths[i] = lengths[i];
1420 
1421 		if (maxmap == 0)
1422 			continue;
1423 
1424 		if (lengths[i] <= maxmap)
1425 			continue;
1426 
1427 		xpt_print(periph->path, "%s: data length %u > max allowed %u "
1428 			  "bytes\n", __func__, lengths[i], maxmap);
1429 		error = EINVAL;
1430 		goto bailout;
1431 	}
1432 
1433 	switch (io_req->data_flags) {
1434 	case CAM_DATA_VADDR:
1435 		/* Map or copy the buffer into kernel address space */
1436 		for (i = 0; i < numbufs; i++) {
1437 			uint8_t *tmp_buf;
1438 
1439 			/*
1440 			 * If for some reason no length is specified, we
1441 			 * don't need to allocate anything.
1442 			 */
1443 			if (io_req->lengths[i] == 0)
1444 				continue;
1445 
1446 			tmp_buf = malloc(lengths[i], M_SCSIPASS,
1447 					 M_WAITOK | M_ZERO);
1448 			io_req->kern_bufs[i] = tmp_buf;
1449 			*data_ptrs[i] = tmp_buf;
1450 
1451 #if 0
1452 			xpt_print(periph->path, "%s: malloced %p len %u, user "
1453 				  "buffer %p, operation: %s\n", __func__,
1454 				  tmp_buf, lengths[i], io_req->user_bufs[i],
1455 				  (dirs[i] == CAM_DIR_IN) ? "read" : "write");
1456 #endif
1457 			/*
1458 			 * We only need to copy in if the user is writing.
1459 			 */
1460 			if (dirs[i] != CAM_DIR_OUT)
1461 				continue;
1462 
1463 			error = copyin(io_req->user_bufs[i],
1464 				       io_req->kern_bufs[i], lengths[i]);
1465 			if (error != 0) {
1466 				xpt_print(periph->path, "%s: copy of user "
1467 					  "buffer from %p to %p failed with "
1468 					  "error %d\n", __func__,
1469 					  io_req->user_bufs[i],
1470 					  io_req->kern_bufs[i], error);
1471 				goto bailout;
1472 			}
1473 		}
1474 		break;
1475 	case CAM_DATA_PADDR:
1476 		/* Pass down the pointer as-is */
1477 		break;
1478 	case CAM_DATA_SG: {
1479 		size_t sg_length, size_to_go, alloc_size;
1480 		uint32_t num_segs_needed;
1481 
1482 		/*
1483 		 * Copy the user S/G list in, and then copy in the
1484 		 * individual segments.
1485 		 */
1486 		/*
1487 		 * We shouldn't see this, but check just in case.
1488 		 */
1489 		if (numbufs != 1) {
1490 			xpt_print(periph->path, "%s: cannot currently handle "
1491 				  "more than one S/G list per CCB\n", __func__);
1492 			error = EINVAL;
1493 			goto bailout;
1494 		}
1495 
1496 		/*
1497 		 * We have to have at least one segment.
1498 		 */
1499 		if (num_segs == 0) {
1500 			xpt_print(periph->path, "%s: CAM_DATA_SG flag set, "
1501 				  "but sglist_cnt=0!\n", __func__);
1502 			error = EINVAL;
1503 			goto bailout;
1504 		}
1505 
1506 		/*
1507 		 * Make sure the user specified the total length and didn't
1508 		 * just leave it to us to decode the S/G list.
1509 		 */
1510 		if (lengths[0] == 0) {
1511 			xpt_print(periph->path, "%s: no dxfer_len specified, "
1512 				  "but CAM_DATA_SG flag is set!\n", __func__);
1513 			error = EINVAL;
1514 			goto bailout;
1515 		}
1516 
1517 		/*
1518 		 * We allocate buffers in io_zone_size increments for an
1519 		 * S/G list.  This will generally be maxphys.
1520 		 */
1521 		if (lengths[0] <= softc->io_zone_size)
1522 			num_segs_needed = 1;
1523 		else {
1524 			num_segs_needed = lengths[0] / softc->io_zone_size;
1525 			if ((lengths[0] % softc->io_zone_size) != 0)
1526 				num_segs_needed++;
1527 		}
1528 
1529 		/* Figure out the size of the S/G list */
1530 		sg_length = num_segs * sizeof(bus_dma_segment_t);
1531 		io_req->num_user_segs = num_segs;
1532 		io_req->num_kern_segs = num_segs_needed;
1533 
1534 		/* Save the user's S/G list pointer for later restoration */
1535 		io_req->user_bufs[0] = *data_ptrs[0];
1536 
1537 		/*
1538 		 * If we have enough segments allocated by default to handle
1539 		 * the length of the user's S/G list,
1540 		 */
1541 		if (num_segs > PASS_MAX_SEGS) {
1542 			io_req->user_segptr = malloc(sizeof(bus_dma_segment_t) *
1543 			    num_segs, M_SCSIPASS, M_WAITOK | M_ZERO);
1544 			io_req->flags |= PASS_IO_USER_SEG_MALLOC;
1545 		} else
1546 			io_req->user_segptr = io_req->user_segs;
1547 
1548 		error = copyin(*data_ptrs[0], io_req->user_segptr, sg_length);
1549 		if (error != 0) {
1550 			xpt_print(periph->path, "%s: copy of user S/G list "
1551 				  "from %p to %p failed with error %d\n",
1552 				  __func__, *data_ptrs[0], io_req->user_segptr,
1553 				  error);
1554 			goto bailout;
1555 		}
1556 
1557 		if (num_segs_needed > PASS_MAX_SEGS) {
1558 			io_req->kern_segptr = malloc(sizeof(bus_dma_segment_t) *
1559 			    num_segs_needed, M_SCSIPASS, M_WAITOK | M_ZERO);
1560 			io_req->flags |= PASS_IO_KERN_SEG_MALLOC;
1561 		} else {
1562 			io_req->kern_segptr = io_req->kern_segs;
1563 		}
1564 
1565 		/*
1566 		 * Allocate the kernel S/G list.
1567 		 */
1568 		for (size_to_go = lengths[0], i = 0;
1569 		     size_to_go > 0 && i < num_segs_needed;
1570 		     i++, size_to_go -= alloc_size) {
1571 			uint8_t *kern_ptr;
1572 
1573 			alloc_size = min(size_to_go, softc->io_zone_size);
1574 			kern_ptr = uma_zalloc(softc->pass_io_zone, M_WAITOK);
1575 			io_req->kern_segptr[i].ds_addr =
1576 			    (bus_addr_t)(uintptr_t)kern_ptr;
1577 			io_req->kern_segptr[i].ds_len = alloc_size;
1578 		}
1579 		if (size_to_go > 0) {
1580 			printf("%s: size_to_go = %zu, software error!\n",
1581 			       __func__, size_to_go);
1582 			error = EINVAL;
1583 			goto bailout;
1584 		}
1585 
1586 		*data_ptrs[0] = (uint8_t *)io_req->kern_segptr;
1587 		*seg_cnt_ptr = io_req->num_kern_segs;
1588 
1589 		/*
1590 		 * We only need to copy data here if the user is writing.
1591 		 */
1592 		if (dirs[0] == CAM_DIR_OUT)
1593 			error = passcopysglist(periph, io_req, dirs[0]);
1594 		break;
1595 	}
1596 	case CAM_DATA_SG_PADDR: {
1597 		size_t sg_length;
1598 
1599 		/*
1600 		 * We shouldn't see this, but check just in case.
1601 		 */
1602 		if (numbufs != 1) {
1603 			printf("%s: cannot currently handle more than one "
1604 			       "S/G list per CCB\n", __func__);
1605 			error = EINVAL;
1606 			goto bailout;
1607 		}
1608 
1609 		/*
1610 		 * We have to have at least one segment.
1611 		 */
1612 		if (num_segs == 0) {
1613 			xpt_print(periph->path, "%s: CAM_DATA_SG_PADDR flag "
1614 				  "set, but sglist_cnt=0!\n", __func__);
1615 			error = EINVAL;
1616 			goto bailout;
1617 		}
1618 
1619 		/*
1620 		 * Make sure the user specified the total length and didn't
1621 		 * just leave it to us to decode the S/G list.
1622 		 */
1623 		if (lengths[0] == 0) {
1624 			xpt_print(periph->path, "%s: no dxfer_len specified, "
1625 				  "but CAM_DATA_SG flag is set!\n", __func__);
1626 			error = EINVAL;
1627 			goto bailout;
1628 		}
1629 
1630 		/* Figure out the size of the S/G list */
1631 		sg_length = num_segs * sizeof(bus_dma_segment_t);
1632 		io_req->num_user_segs = num_segs;
1633 		io_req->num_kern_segs = io_req->num_user_segs;
1634 
1635 		/* Save the user's S/G list pointer for later restoration */
1636 		io_req->user_bufs[0] = *data_ptrs[0];
1637 
1638 		if (num_segs > PASS_MAX_SEGS) {
1639 			io_req->user_segptr = malloc(sizeof(bus_dma_segment_t) *
1640 			    num_segs, M_SCSIPASS, M_WAITOK | M_ZERO);
1641 			io_req->flags |= PASS_IO_USER_SEG_MALLOC;
1642 		} else
1643 			io_req->user_segptr = io_req->user_segs;
1644 
1645 		io_req->kern_segptr = io_req->user_segptr;
1646 
1647 		error = copyin(*data_ptrs[0], io_req->user_segptr, sg_length);
1648 		if (error != 0) {
1649 			xpt_print(periph->path, "%s: copy of user S/G list "
1650 				  "from %p to %p failed with error %d\n",
1651 				  __func__, *data_ptrs[0], io_req->user_segptr,
1652 				  error);
1653 			goto bailout;
1654 		}
1655 		break;
1656 	}
1657 	default:
1658 	case CAM_DATA_BIO:
1659 		/*
1660 		 * A user shouldn't be attaching a bio to the CCB.  It
1661 		 * isn't a user-accessible structure.
1662 		 */
1663 		error = EINVAL;
1664 		break;
1665 	}
1666 
1667 bailout:
1668 	if (error != 0)
1669 		passiocleanup(softc, io_req);
1670 
1671 	return (error);
1672 }
1673 
1674 static int
1675 passmemdone(struct cam_periph *periph, struct pass_io_req *io_req)
1676 {
1677 	struct pass_softc *softc;
1678 	int error;
1679 	int i;
1680 
1681 	error = 0;
1682 	softc = (struct pass_softc *)periph->softc;
1683 
1684 	switch (io_req->data_flags) {
1685 	case CAM_DATA_VADDR:
1686 		/*
1687 		 * Copy back to the user buffer if this was a read.
1688 		 */
1689 		for (i = 0; i < io_req->num_bufs; i++) {
1690 			if (io_req->dirs[i] != CAM_DIR_IN)
1691 				continue;
1692 
1693 			error = copyout(io_req->kern_bufs[i],
1694 			    io_req->user_bufs[i], io_req->lengths[i]);
1695 			if (error != 0) {
1696 				xpt_print(periph->path, "Unable to copy %u "
1697 					  "bytes from %p to user address %p\n",
1698 					  io_req->lengths[i],
1699 					  io_req->kern_bufs[i],
1700 					  io_req->user_bufs[i]);
1701 				goto bailout;
1702 			}
1703 		}
1704 		break;
1705 	case CAM_DATA_PADDR:
1706 		/* Do nothing.  The pointer is a physical address already */
1707 		break;
1708 	case CAM_DATA_SG:
1709 		/*
1710 		 * Copy back to the user buffer if this was a read.
1711 		 * Restore the user's S/G list buffer pointer.
1712 		 */
1713 		if (io_req->dirs[0] == CAM_DIR_IN)
1714 			error = passcopysglist(periph, io_req, io_req->dirs[0]);
1715 		break;
1716 	case CAM_DATA_SG_PADDR:
1717 		/*
1718 		 * Restore the user's S/G list buffer pointer.  No need to
1719 		 * copy.
1720 		 */
1721 		break;
1722 	default:
1723 	case CAM_DATA_BIO:
1724 		error = EINVAL;
1725 		break;
1726 	}
1727 
1728 bailout:
1729 	/*
1730 	 * Reset the user's pointers to their original values and free
1731 	 * allocated memory.
1732 	 */
1733 	passiocleanup(softc, io_req);
1734 
1735 	return (error);
1736 }
1737 
1738 static int
1739 passioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flag, struct thread *td)
1740 {
1741 	int error;
1742 
1743 	if ((error = passdoioctl(dev, cmd, addr, flag, td)) == ENOTTY) {
1744 		error = cam_compat_ioctl(dev, cmd, addr, flag, td, passdoioctl);
1745 	}
1746 	return (error);
1747 }
1748 
1749 static int
1750 passdoioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flag, struct thread *td)
1751 {
1752 	struct	cam_periph *periph;
1753 	struct	pass_softc *softc;
1754 	int	error;
1755 	uint32_t priority;
1756 
1757 	periph = (struct cam_periph *)dev->si_drv1;
1758 	cam_periph_lock(periph);
1759 	softc = (struct pass_softc *)periph->softc;
1760 
1761 	error = 0;
1762 
1763 	switch (cmd) {
1764 	case CAMIOCOMMAND:
1765 	{
1766 		union ccb *inccb;
1767 		union ccb *ccb;
1768 		int ccb_malloced;
1769 
1770 		inccb = (union ccb *)addr;
1771 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
1772 		if (inccb->ccb_h.func_code == XPT_SCSI_IO)
1773 			inccb->csio.bio = NULL;
1774 #endif
1775 
1776 		if (inccb->ccb_h.flags & CAM_UNLOCKED) {
1777 			error = EINVAL;
1778 			break;
1779 		}
1780 
1781 		/*
1782 		 * Some CCB types, like scan bus and scan lun can only go
1783 		 * through the transport layer device.
1784 		 */
1785 		if (inccb->ccb_h.func_code & XPT_FC_XPT_ONLY) {
1786 			xpt_print(periph->path, "CCB function code %#x is "
1787 			    "restricted to the XPT device\n",
1788 			    inccb->ccb_h.func_code);
1789 			error = ENODEV;
1790 			break;
1791 		}
1792 
1793 		/* Compatibility for RL/priority-unaware code. */
1794 		priority = inccb->ccb_h.pinfo.priority;
1795 		if (priority <= CAM_PRIORITY_OOB)
1796 		    priority += CAM_PRIORITY_OOB + 1;
1797 
1798 		/*
1799 		 * Non-immediate CCBs need a CCB from the per-device pool
1800 		 * of CCBs, which is scheduled by the transport layer.
1801 		 * Immediate CCBs and user-supplied CCBs should just be
1802 		 * malloced.
1803 		 */
1804 		if ((inccb->ccb_h.func_code & XPT_FC_QUEUED)
1805 		 && ((inccb->ccb_h.func_code & XPT_FC_USER_CCB) == 0)) {
1806 			ccb = cam_periph_getccb(periph, priority);
1807 			ccb_malloced = 0;
1808 		} else {
1809 			ccb = xpt_alloc_ccb_nowait();
1810 
1811 			if (ccb != NULL)
1812 				xpt_setup_ccb(&ccb->ccb_h, periph->path,
1813 					      priority);
1814 			ccb_malloced = 1;
1815 		}
1816 
1817 		if (ccb == NULL) {
1818 			xpt_print(periph->path, "unable to allocate CCB\n");
1819 			error = ENOMEM;
1820 			break;
1821 		}
1822 
1823 		error = passsendccb(periph, ccb, inccb);
1824 
1825 		if (ccb_malloced)
1826 			xpt_free_ccb(ccb);
1827 		else
1828 			xpt_release_ccb(ccb);
1829 
1830 		break;
1831 	}
1832 	case CAMIOQUEUE:
1833 	{
1834 		struct pass_io_req *io_req;
1835 		union ccb **user_ccb, *ccb;
1836 		xpt_opcode fc;
1837 
1838 #ifdef COMPAT_FREEBSD32
1839 		if (SV_PROC_FLAG(td->td_proc, SV_ILP32)) {
1840 			error = ENOTTY;
1841 			goto bailout;
1842 		}
1843 #endif
1844 		if ((softc->flags & PASS_FLAG_ZONE_VALID) == 0) {
1845 			error = passcreatezone(periph);
1846 			if (error != 0)
1847 				goto bailout;
1848 		}
1849 
1850 		/*
1851 		 * We're going to do a blocking allocation for this I/O
1852 		 * request, so we have to drop the lock.
1853 		 */
1854 		cam_periph_unlock(periph);
1855 
1856 		io_req = uma_zalloc(softc->pass_zone, M_WAITOK | M_ZERO);
1857 		ccb = &io_req->ccb;
1858 		user_ccb = (union ccb **)addr;
1859 
1860 		/*
1861 		 * Unlike the CAMIOCOMMAND ioctl above, we only have a
1862 		 * pointer to the user's CCB, so we have to copy the whole
1863 		 * thing in to a buffer we have allocated (above) instead
1864 		 * of allowing the ioctl code to malloc a buffer and copy
1865 		 * it in.
1866 		 *
1867 		 * This is an advantage for this asynchronous interface,
1868 		 * since we don't want the memory to get freed while the
1869 		 * CCB is outstanding.
1870 		 */
1871 #if 0
1872 		xpt_print(periph->path, "Copying user CCB %p to "
1873 			  "kernel address %p\n", *user_ccb, ccb);
1874 #endif
1875 		error = copyin(*user_ccb, ccb, sizeof(*ccb));
1876 		if (error != 0) {
1877 			xpt_print(periph->path, "Copy of user CCB %p to "
1878 				  "kernel address %p failed with error %d\n",
1879 				  *user_ccb, ccb, error);
1880 			goto camioqueue_error;
1881 		}
1882 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
1883 		if (ccb->ccb_h.func_code == XPT_SCSI_IO)
1884 			ccb->csio.bio = NULL;
1885 #endif
1886 
1887 		if (ccb->ccb_h.flags & CAM_UNLOCKED) {
1888 			error = EINVAL;
1889 			goto camioqueue_error;
1890 		}
1891 
1892 		if (ccb->ccb_h.flags & CAM_CDB_POINTER) {
1893 			if (ccb->csio.cdb_len > IOCDBLEN) {
1894 				error = EINVAL;
1895 				goto camioqueue_error;
1896 			}
1897 			error = copyin(ccb->csio.cdb_io.cdb_ptr,
1898 			    ccb->csio.cdb_io.cdb_bytes, ccb->csio.cdb_len);
1899 			if (error != 0)
1900 				goto camioqueue_error;
1901 			ccb->ccb_h.flags &= ~CAM_CDB_POINTER;
1902 		}
1903 
1904 		/*
1905 		 * Some CCB types, like scan bus and scan lun can only go
1906 		 * through the transport layer device.
1907 		 */
1908 		if (ccb->ccb_h.func_code & XPT_FC_XPT_ONLY) {
1909 			xpt_print(periph->path, "CCB function code %#x is "
1910 			    "restricted to the XPT device\n",
1911 			    ccb->ccb_h.func_code);
1912 			error = ENODEV;
1913 			goto camioqueue_error;
1914 		}
1915 
1916 		/*
1917 		 * Save the user's CCB pointer as well as his linked list
1918 		 * pointers and peripheral private area so that we can
1919 		 * restore these later.
1920 		 */
1921 		io_req->user_ccb_ptr = *user_ccb;
1922 		io_req->user_periph_links = ccb->ccb_h.periph_links;
1923 		io_req->user_periph_priv = ccb->ccb_h.periph_priv;
1924 
1925 		/*
1926 		 * Now that we've saved the user's values, we can set our
1927 		 * own peripheral private entry.
1928 		 */
1929 		ccb->ccb_h.ccb_ioreq = io_req;
1930 
1931 		/* Compatibility for RL/priority-unaware code. */
1932 		priority = ccb->ccb_h.pinfo.priority;
1933 		if (priority <= CAM_PRIORITY_OOB)
1934 		    priority += CAM_PRIORITY_OOB + 1;
1935 
1936 		/*
1937 		 * Setup fields in the CCB like the path and the priority.
1938 		 * The path in particular cannot be done in userland, since
1939 		 * it is a pointer to a kernel data structure.
1940 		 */
1941 		xpt_setup_ccb_flags(&ccb->ccb_h, periph->path, priority,
1942 				    ccb->ccb_h.flags);
1943 
1944 		/*
1945 		 * Setup our done routine.  There is no way for the user to
1946 		 * have a valid pointer here.
1947 		 */
1948 		ccb->ccb_h.cbfcnp = passdone;
1949 
1950 		fc = ccb->ccb_h.func_code;
1951 		/*
1952 		 * If this function code has memory that can be mapped in
1953 		 * or out, we need to call passmemsetup().
1954 		 */
1955 		if ((fc == XPT_SCSI_IO) || (fc == XPT_ATA_IO)
1956 		 || (fc == XPT_SMP_IO) || (fc == XPT_DEV_MATCH)
1957 		 || (fc == XPT_DEV_ADVINFO)
1958 		 || (fc == XPT_NVME_ADMIN) || (fc == XPT_NVME_IO)) {
1959 			error = passmemsetup(periph, io_req);
1960 			if (error != 0)
1961 				goto camioqueue_error;
1962 		} else
1963 			io_req->mapinfo.num_bufs_used = 0;
1964 
1965 		cam_periph_lock(periph);
1966 
1967 		/*
1968 		 * Everything goes on the incoming queue initially.
1969 		 */
1970 		TAILQ_INSERT_TAIL(&softc->incoming_queue, io_req, links);
1971 
1972 		/*
1973 		 * If the CCB is queued, and is not a user CCB, then
1974 		 * we need to allocate a slot for it.  Call xpt_schedule()
1975 		 * so that our start routine will get called when a CCB is
1976 		 * available.
1977 		 */
1978 		if ((fc & XPT_FC_QUEUED)
1979 		 && ((fc & XPT_FC_USER_CCB) == 0)) {
1980 			xpt_schedule(periph, priority);
1981 			break;
1982 		}
1983 
1984 		/*
1985 		 * At this point, the CCB in question is either an
1986 		 * immediate CCB (like XPT_DEV_ADVINFO) or it is a user CCB
1987 		 * and therefore should be malloced, not allocated via a slot.
1988 		 * Remove the CCB from the incoming queue and add it to the
1989 		 * active queue.
1990 		 */
1991 		TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
1992 		TAILQ_INSERT_TAIL(&softc->active_queue, io_req, links);
1993 
1994 		xpt_action(ccb);
1995 
1996 		/*
1997 		 * If this is not a queued CCB (i.e. it is an immediate CCB),
1998 		 * then it is already done.  We need to put it on the done
1999 		 * queue for the user to fetch.
2000 		 */
2001 		if ((fc & XPT_FC_QUEUED) == 0) {
2002 			TAILQ_REMOVE(&softc->active_queue, io_req, links);
2003 			TAILQ_INSERT_TAIL(&softc->done_queue, io_req, links);
2004 		}
2005 		break;
2006 
2007 camioqueue_error:
2008 		uma_zfree(softc->pass_zone, io_req);
2009 		cam_periph_lock(periph);
2010 		break;
2011 	}
2012 	case CAMIOGET:
2013 	{
2014 		union ccb **user_ccb;
2015 		struct pass_io_req *io_req;
2016 		int old_error;
2017 
2018 #ifdef COMPAT_FREEBSD32
2019 		if (SV_PROC_FLAG(td->td_proc, SV_ILP32)) {
2020 			error = ENOTTY;
2021 			goto bailout;
2022 		}
2023 #endif
2024 		user_ccb = (union ccb **)addr;
2025 		old_error = 0;
2026 
2027 		io_req = TAILQ_FIRST(&softc->done_queue);
2028 		if (io_req == NULL) {
2029 			error = ENOENT;
2030 			break;
2031 		}
2032 
2033 		/*
2034 		 * Remove the I/O from the done queue.
2035 		 */
2036 		TAILQ_REMOVE(&softc->done_queue, io_req, links);
2037 
2038 		/*
2039 		 * We have to drop the lock during the copyout because the
2040 		 * copyout can result in VM faults that require sleeping.
2041 		 */
2042 		cam_periph_unlock(periph);
2043 
2044 		/*
2045 		 * Do any needed copies (e.g. for reads) and revert the
2046 		 * pointers in the CCB back to the user's pointers.
2047 		 */
2048 		error = passmemdone(periph, io_req);
2049 
2050 		old_error = error;
2051 
2052 		io_req->ccb.ccb_h.periph_links = io_req->user_periph_links;
2053 		io_req->ccb.ccb_h.periph_priv = io_req->user_periph_priv;
2054 
2055 #if 0
2056 		xpt_print(periph->path, "Copying to user CCB %p from "
2057 			  "kernel address %p\n", *user_ccb, &io_req->ccb);
2058 #endif
2059 
2060 		error = copyout(&io_req->ccb, *user_ccb, sizeof(union ccb));
2061 		if (error != 0) {
2062 			xpt_print(periph->path, "Copy to user CCB %p from "
2063 				  "kernel address %p failed with error %d\n",
2064 				  *user_ccb, &io_req->ccb, error);
2065 		}
2066 
2067 		/*
2068 		 * Prefer the first error we got back, and make sure we
2069 		 * don't overwrite bad status with good.
2070 		 */
2071 		if (old_error != 0)
2072 			error = old_error;
2073 
2074 		cam_periph_lock(periph);
2075 
2076 		/*
2077 		 * At this point, if there was an error, we could potentially
2078 		 * re-queue the I/O and try again.  But why?  The error
2079 		 * would almost certainly happen again.  We might as well
2080 		 * not leak memory.
2081 		 */
2082 		uma_zfree(softc->pass_zone, io_req);
2083 		break;
2084 	}
2085 	default:
2086 		error = cam_periph_ioctl(periph, cmd, addr, passerror);
2087 		break;
2088 	}
2089 
2090 bailout:
2091 	cam_periph_unlock(periph);
2092 
2093 	return(error);
2094 }
2095 
2096 static int
2097 passpoll(struct cdev *dev, int poll_events, struct thread *td)
2098 {
2099 	struct cam_periph *periph;
2100 	struct pass_softc *softc;
2101 	int revents;
2102 
2103 	periph = (struct cam_periph *)dev->si_drv1;
2104 	softc = (struct pass_softc *)periph->softc;
2105 
2106 	revents = poll_events & (POLLOUT | POLLWRNORM);
2107 	if ((poll_events & (POLLIN | POLLRDNORM)) != 0) {
2108 		cam_periph_lock(periph);
2109 
2110 		if (!TAILQ_EMPTY(&softc->done_queue)) {
2111 			revents |= poll_events & (POLLIN | POLLRDNORM);
2112 		}
2113 		cam_periph_unlock(periph);
2114 		if (revents == 0)
2115 			selrecord(td, &softc->read_select);
2116 	}
2117 
2118 	return (revents);
2119 }
2120 
2121 static int
2122 passkqfilter(struct cdev *dev, struct knote *kn)
2123 {
2124 	struct cam_periph *periph;
2125 	struct pass_softc *softc;
2126 
2127 	periph = (struct cam_periph *)dev->si_drv1;
2128 	softc = (struct pass_softc *)periph->softc;
2129 
2130 	kn->kn_hook = (caddr_t)periph;
2131 	kn->kn_fop = &passread_filtops;
2132 	knlist_add(&softc->read_select.si_note, kn, 0);
2133 
2134 	return (0);
2135 }
2136 
2137 static void
2138 passreadfiltdetach(struct knote *kn)
2139 {
2140 	struct cam_periph *periph;
2141 	struct pass_softc *softc;
2142 
2143 	periph = (struct cam_periph *)kn->kn_hook;
2144 	softc = (struct pass_softc *)periph->softc;
2145 
2146 	knlist_remove(&softc->read_select.si_note, kn, 0);
2147 }
2148 
2149 static int
2150 passreadfilt(struct knote *kn, long hint)
2151 {
2152 	struct cam_periph *periph;
2153 	struct pass_softc *softc;
2154 	int retval;
2155 
2156 	periph = (struct cam_periph *)kn->kn_hook;
2157 	softc = (struct pass_softc *)periph->softc;
2158 
2159 	cam_periph_assert(periph, MA_OWNED);
2160 
2161 	if (TAILQ_EMPTY(&softc->done_queue))
2162 		retval = 0;
2163 	else
2164 		retval = 1;
2165 
2166 	return (retval);
2167 }
2168 
2169 /*
2170  * Generally, "ccb" should be the CCB supplied by the kernel.  "inccb"
2171  * should be the CCB that is copied in from the user.
2172  */
2173 static int
2174 passsendccb(struct cam_periph *periph, union ccb *ccb, union ccb *inccb)
2175 {
2176 	struct pass_softc *softc;
2177 	struct cam_periph_map_info mapinfo;
2178 	uint8_t *cmd;
2179 	xpt_opcode fc;
2180 	int error;
2181 
2182 	softc = (struct pass_softc *)periph->softc;
2183 
2184 	/*
2185 	 * There are some fields in the CCB header that need to be
2186 	 * preserved, the rest we get from the user.
2187 	 */
2188 	xpt_merge_ccb(ccb, inccb);
2189 
2190 	if (ccb->ccb_h.flags & CAM_CDB_POINTER) {
2191 		cmd = __builtin_alloca(ccb->csio.cdb_len);
2192 		error = copyin(ccb->csio.cdb_io.cdb_ptr, cmd, ccb->csio.cdb_len);
2193 		if (error)
2194 			return (error);
2195 		ccb->csio.cdb_io.cdb_ptr = cmd;
2196 	}
2197 
2198 	/*
2199 	 * Let cam_periph_mapmem do a sanity check on the data pointer format.
2200 	 * Even if no data transfer is needed, it's a cheap check and it
2201 	 * simplifies the code.
2202 	 */
2203 	fc = ccb->ccb_h.func_code;
2204 	if ((fc == XPT_SCSI_IO) || (fc == XPT_ATA_IO) || (fc == XPT_SMP_IO)
2205             || (fc == XPT_DEV_MATCH) || (fc == XPT_DEV_ADVINFO) || (fc == XPT_MMC_IO)
2206             || (fc == XPT_NVME_ADMIN) || (fc == XPT_NVME_IO)) {
2207 		bzero(&mapinfo, sizeof(mapinfo));
2208 
2209 		/*
2210 		 * cam_periph_mapmem calls into proc and vm functions that can
2211 		 * sleep as well as trigger I/O, so we can't hold the lock.
2212 		 * Dropping it here is reasonably safe.
2213 		 */
2214 		cam_periph_unlock(periph);
2215 		error = cam_periph_mapmem(ccb, &mapinfo, softc->maxio);
2216 		cam_periph_lock(periph);
2217 
2218 		/*
2219 		 * cam_periph_mapmem returned an error, we can't continue.
2220 		 * Return the error to the user.
2221 		 */
2222 		if (error)
2223 			return(error);
2224 	} else
2225 		/* Ensure that the unmap call later on is a no-op. */
2226 		mapinfo.num_bufs_used = 0;
2227 
2228 	/*
2229 	 * If the user wants us to perform any error recovery, then honor
2230 	 * that request.  Otherwise, it's up to the user to perform any
2231 	 * error recovery.
2232 	 */
2233 	cam_periph_runccb(ccb, (ccb->ccb_h.flags & CAM_PASS_ERR_RECOVER) ?
2234 	    passerror : NULL, /* cam_flags */ CAM_RETRY_SELTO,
2235 	    /* sense_flags */ SF_RETRY_UA | SF_NO_PRINT,
2236 	    softc->device_stats);
2237 
2238 	cam_periph_unlock(periph);
2239 	cam_periph_unmapmem(ccb, &mapinfo);
2240 	cam_periph_lock(periph);
2241 
2242 	ccb->ccb_h.cbfcnp = NULL;
2243 	ccb->ccb_h.periph_priv = inccb->ccb_h.periph_priv;
2244 	bcopy(ccb, inccb, sizeof(union ccb));
2245 
2246 	return(0);
2247 }
2248 
2249 static int
2250 passerror(union ccb *ccb, u_int32_t cam_flags, u_int32_t sense_flags)
2251 {
2252 
2253 	return(cam_periph_error(ccb, cam_flags, sense_flags));
2254 }
2255