xref: /freebsd/sys/cam/scsi/scsi_pass.c (revision 60eddb209b5ad13a549ca74a41b7cb38a31da5ef)
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 			cam_status status;
522 
523 			softc = (struct pass_softc *)periph->softc;
524 			/*
525 			 * Acquire a reference to the periph before we
526 			 * start the taskqueue, so that we don't run into
527 			 * a situation where the periph goes away before
528 			 * the task queue has a chance to run.
529 			 */
530 			status = cam_periph_acquire(periph);
531 			if (status != CAM_REQ_CMP)
532 				break;
533 
534 			taskqueue_enqueue(taskqueue_thread,
535 					  &softc->add_physpath_task);
536 		}
537 		break;
538 	}
539 	default:
540 		cam_periph_async(periph, code, path, arg);
541 		break;
542 	}
543 }
544 
545 static cam_status
546 passregister(struct cam_periph *periph, void *arg)
547 {
548 	struct pass_softc *softc;
549 	struct ccb_getdev *cgd;
550 	struct ccb_pathinq cpi;
551 	struct make_dev_args args;
552 	int error, no_tags;
553 
554 	cgd = (struct ccb_getdev *)arg;
555 	if (cgd == NULL) {
556 		printf("%s: no getdev CCB, can't register device\n", __func__);
557 		return(CAM_REQ_CMP_ERR);
558 	}
559 
560 	softc = (struct pass_softc *)malloc(sizeof(*softc),
561 					    M_DEVBUF, M_NOWAIT);
562 
563 	if (softc == NULL) {
564 		printf("%s: Unable to probe new device. "
565 		       "Unable to allocate softc\n", __func__);
566 		return(CAM_REQ_CMP_ERR);
567 	}
568 
569 	bzero(softc, sizeof(*softc));
570 	softc->state = PASS_STATE_NORMAL;
571 	if (cgd->protocol == PROTO_SCSI || cgd->protocol == PROTO_ATAPI)
572 		softc->pd_type = SID_TYPE(&cgd->inq_data);
573 	else if (cgd->protocol == PROTO_SATAPM)
574 		softc->pd_type = T_ENCLOSURE;
575 	else
576 		softc->pd_type = T_DIRECT;
577 
578 	periph->softc = softc;
579 	softc->periph = periph;
580 	TAILQ_INIT(&softc->incoming_queue);
581 	TAILQ_INIT(&softc->active_queue);
582 	TAILQ_INIT(&softc->abandoned_queue);
583 	TAILQ_INIT(&softc->done_queue);
584 	snprintf(softc->zone_name, sizeof(softc->zone_name), "%s%d",
585 		 periph->periph_name, periph->unit_number);
586 	snprintf(softc->io_zone_name, sizeof(softc->io_zone_name), "%s%dIO",
587 		 periph->periph_name, periph->unit_number);
588 	softc->io_zone_size = MAXPHYS;
589 	knlist_init_mtx(&softc->read_select.si_note, cam_periph_mtx(periph));
590 
591 	xpt_path_inq(&cpi, periph->path);
592 
593 	if (cpi.maxio == 0)
594 		softc->maxio = DFLTPHYS;	/* traditional default */
595 	else if (cpi.maxio > MAXPHYS)
596 		softc->maxio = MAXPHYS;		/* for safety */
597 	else
598 		softc->maxio = cpi.maxio;	/* real value */
599 
600 	if (cpi.hba_misc & PIM_UNMAPPED)
601 		softc->flags |= PASS_FLAG_UNMAPPED_CAPABLE;
602 
603 	/*
604 	 * We pass in 0 for a blocksize, since we don't
605 	 * know what the blocksize of this device is, if
606 	 * it even has a blocksize.
607 	 */
608 	cam_periph_unlock(periph);
609 	no_tags = (cgd->inq_data.flags & SID_CmdQue) == 0;
610 	softc->device_stats = devstat_new_entry("pass",
611 			  periph->unit_number, 0,
612 			  DEVSTAT_NO_BLOCKSIZE
613 			  | (no_tags ? DEVSTAT_NO_ORDERED_TAGS : 0),
614 			  softc->pd_type |
615 			  XPORT_DEVSTAT_TYPE(cpi.transport) |
616 			  DEVSTAT_TYPE_PASS,
617 			  DEVSTAT_PRIORITY_PASS);
618 
619 	/*
620 	 * Initialize the taskqueue handler for shutting down kqueue.
621 	 */
622 	TASK_INIT(&softc->shutdown_kqueue_task, /*priority*/ 0,
623 		  pass_shutdown_kqueue, periph);
624 
625 	/*
626 	 * Acquire a reference to the periph that we can release once we've
627 	 * cleaned up the kqueue.
628 	 */
629 	if (cam_periph_acquire(periph) != CAM_REQ_CMP) {
630 		xpt_print(periph->path, "%s: lost periph during "
631 			  "registration!\n", __func__);
632 		cam_periph_lock(periph);
633 		return (CAM_REQ_CMP_ERR);
634 	}
635 
636 	/*
637 	 * Acquire a reference to the periph before we create the devfs
638 	 * instance for it.  We'll release this reference once the devfs
639 	 * instance has been freed.
640 	 */
641 	if (cam_periph_acquire(periph) != CAM_REQ_CMP) {
642 		xpt_print(periph->path, "%s: lost periph during "
643 			  "registration!\n", __func__);
644 		cam_periph_lock(periph);
645 		return (CAM_REQ_CMP_ERR);
646 	}
647 
648 	/* Register the device */
649 	make_dev_args_init(&args);
650 	args.mda_devsw = &pass_cdevsw;
651 	args.mda_unit = periph->unit_number;
652 	args.mda_uid = UID_ROOT;
653 	args.mda_gid = GID_OPERATOR;
654 	args.mda_mode = 0600;
655 	args.mda_si_drv1 = periph;
656 	error = make_dev_s(&args, &softc->dev, "%s%d", periph->periph_name,
657 	    periph->unit_number);
658 	if (error != 0) {
659 		cam_periph_lock(periph);
660 		cam_periph_release_locked(periph);
661 		return (CAM_REQ_CMP_ERR);
662 	}
663 
664 	/*
665 	 * Hold a reference to the periph before we create the physical
666 	 * path alias so it can't go away.
667 	 */
668 	if (cam_periph_acquire(periph) != CAM_REQ_CMP) {
669 		xpt_print(periph->path, "%s: lost periph during "
670 			  "registration!\n", __func__);
671 		cam_periph_lock(periph);
672 		return (CAM_REQ_CMP_ERR);
673 	}
674 
675 	cam_periph_lock(periph);
676 
677 	TASK_INIT(&softc->add_physpath_task, /*priority*/0,
678 		  pass_add_physpath, periph);
679 
680 	/*
681 	 * See if physical path information is already available.
682 	 */
683 	taskqueue_enqueue(taskqueue_thread, &softc->add_physpath_task);
684 
685 	/*
686 	 * Add an async callback so that we get notified if
687 	 * this device goes away or its physical path
688 	 * (stored in the advanced info data of the EDT) has
689 	 * changed.
690 	 */
691 	xpt_register_async(AC_LOST_DEVICE | AC_ADVINFO_CHANGED,
692 			   passasync, periph, periph->path);
693 
694 	if (bootverbose)
695 		xpt_announce_periph(periph, NULL);
696 
697 	return(CAM_REQ_CMP);
698 }
699 
700 static int
701 passopen(struct cdev *dev, int flags, int fmt, struct thread *td)
702 {
703 	struct cam_periph *periph;
704 	struct pass_softc *softc;
705 	int error;
706 
707 	periph = (struct cam_periph *)dev->si_drv1;
708 	if (cam_periph_acquire(periph) != CAM_REQ_CMP)
709 		return (ENXIO);
710 
711 	cam_periph_lock(periph);
712 
713 	softc = (struct pass_softc *)periph->softc;
714 
715 	if (softc->flags & PASS_FLAG_INVALID) {
716 		cam_periph_release_locked(periph);
717 		cam_periph_unlock(periph);
718 		return(ENXIO);
719 	}
720 
721 	/*
722 	 * Don't allow access when we're running at a high securelevel.
723 	 */
724 	error = securelevel_gt(td->td_ucred, 1);
725 	if (error) {
726 		cam_periph_release_locked(periph);
727 		cam_periph_unlock(periph);
728 		return(error);
729 	}
730 
731 	/*
732 	 * Only allow read-write access.
733 	 */
734 	if (((flags & FWRITE) == 0) || ((flags & FREAD) == 0)) {
735 		cam_periph_release_locked(periph);
736 		cam_periph_unlock(periph);
737 		return(EPERM);
738 	}
739 
740 	/*
741 	 * We don't allow nonblocking access.
742 	 */
743 	if ((flags & O_NONBLOCK) != 0) {
744 		xpt_print(periph->path, "can't do nonblocking access\n");
745 		cam_periph_release_locked(periph);
746 		cam_periph_unlock(periph);
747 		return(EINVAL);
748 	}
749 
750 	softc->open_count++;
751 
752 	cam_periph_unlock(periph);
753 
754 	return (error);
755 }
756 
757 static int
758 passclose(struct cdev *dev, int flag, int fmt, struct thread *td)
759 {
760 	struct 	cam_periph *periph;
761 	struct  pass_softc *softc;
762 	struct mtx *mtx;
763 
764 	periph = (struct cam_periph *)dev->si_drv1;
765 	mtx = cam_periph_mtx(periph);
766 	mtx_lock(mtx);
767 
768 	softc = periph->softc;
769 	softc->open_count--;
770 
771 	if (softc->open_count == 0) {
772 		struct pass_io_req *io_req, *io_req2;
773 
774 		TAILQ_FOREACH_SAFE(io_req, &softc->done_queue, links, io_req2) {
775 			TAILQ_REMOVE(&softc->done_queue, io_req, links);
776 			passiocleanup(softc, io_req);
777 			uma_zfree(softc->pass_zone, io_req);
778 		}
779 
780 		TAILQ_FOREACH_SAFE(io_req, &softc->incoming_queue, links,
781 				   io_req2) {
782 			TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
783 			passiocleanup(softc, io_req);
784 			uma_zfree(softc->pass_zone, io_req);
785 		}
786 
787 		/*
788 		 * If there are any active I/Os, we need to forcibly acquire a
789 		 * reference to the peripheral so that we don't go away
790 		 * before they complete.  We'll release the reference when
791 		 * the abandoned queue is empty.
792 		 */
793 		io_req = TAILQ_FIRST(&softc->active_queue);
794 		if ((io_req != NULL)
795 		 && (softc->flags & PASS_FLAG_ABANDONED_REF_SET) == 0) {
796 			cam_periph_doacquire(periph);
797 			softc->flags |= PASS_FLAG_ABANDONED_REF_SET;
798 		}
799 
800 		/*
801 		 * Since the I/O in the active queue is not under our
802 		 * control, just set a flag so that we can clean it up when
803 		 * it completes and put it on the abandoned queue.  This
804 		 * will prevent our sending spurious completions in the
805 		 * event that the device is opened again before these I/Os
806 		 * complete.
807 		 */
808 		TAILQ_FOREACH_SAFE(io_req, &softc->active_queue, links,
809 				   io_req2) {
810 			TAILQ_REMOVE(&softc->active_queue, io_req, links);
811 			io_req->flags |= PASS_IO_ABANDONED;
812 			TAILQ_INSERT_TAIL(&softc->abandoned_queue, io_req,
813 					  links);
814 		}
815 	}
816 
817 	cam_periph_release_locked(periph);
818 
819 	/*
820 	 * We reference the lock directly here, instead of using
821 	 * cam_periph_unlock().  The reason is that the call to
822 	 * cam_periph_release_locked() above could result in the periph
823 	 * getting freed.  If that is the case, dereferencing the periph
824 	 * with a cam_periph_unlock() call would cause a page fault.
825 	 *
826 	 * cam_periph_release() avoids this problem using the same method,
827 	 * but we're manually acquiring and dropping the lock here to
828 	 * protect the open count and avoid another lock acquisition and
829 	 * release.
830 	 */
831 	mtx_unlock(mtx);
832 
833 	return (0);
834 }
835 
836 
837 static void
838 passstart(struct cam_periph *periph, union ccb *start_ccb)
839 {
840 	struct pass_softc *softc;
841 
842 	softc = (struct pass_softc *)periph->softc;
843 
844 	switch (softc->state) {
845 	case PASS_STATE_NORMAL: {
846 		struct pass_io_req *io_req;
847 
848 		/*
849 		 * Check for any queued I/O requests that require an
850 		 * allocated slot.
851 		 */
852 		io_req = TAILQ_FIRST(&softc->incoming_queue);
853 		if (io_req == NULL) {
854 			xpt_release_ccb(start_ccb);
855 			break;
856 		}
857 		TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
858 		TAILQ_INSERT_TAIL(&softc->active_queue, io_req, links);
859 		/*
860 		 * Merge the user's CCB into the allocated CCB.
861 		 */
862 		xpt_merge_ccb(start_ccb, &io_req->ccb);
863 		start_ccb->ccb_h.ccb_type = PASS_CCB_QUEUED_IO;
864 		start_ccb->ccb_h.ccb_ioreq = io_req;
865 		start_ccb->ccb_h.cbfcnp = passdone;
866 		io_req->alloced_ccb = start_ccb;
867 		binuptime(&io_req->start_time);
868 		devstat_start_transaction(softc->device_stats,
869 					  &io_req->start_time);
870 
871 		xpt_action(start_ccb);
872 
873 		/*
874 		 * If we have any more I/O waiting, schedule ourselves again.
875 		 */
876 		if (!TAILQ_EMPTY(&softc->incoming_queue))
877 			xpt_schedule(periph, CAM_PRIORITY_NORMAL);
878 		break;
879 	}
880 	default:
881 		break;
882 	}
883 }
884 
885 static void
886 passdone(struct cam_periph *periph, union ccb *done_ccb)
887 {
888 	struct pass_softc *softc;
889 	struct ccb_scsiio *csio;
890 
891 	softc = (struct pass_softc *)periph->softc;
892 
893 	cam_periph_assert(periph, MA_OWNED);
894 
895 	csio = &done_ccb->csio;
896 	switch (csio->ccb_h.ccb_type) {
897 	case PASS_CCB_QUEUED_IO: {
898 		struct pass_io_req *io_req;
899 
900 		io_req = done_ccb->ccb_h.ccb_ioreq;
901 #if 0
902 		xpt_print(periph->path, "%s: called for user CCB %p\n",
903 			  __func__, io_req->user_ccb_ptr);
904 #endif
905 		if (((done_ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_CMP)
906 		 && (done_ccb->ccb_h.flags & CAM_PASS_ERR_RECOVER)
907 		 && ((io_req->flags & PASS_IO_ABANDONED) == 0)) {
908 			int error;
909 
910 			error = passerror(done_ccb, CAM_RETRY_SELTO,
911 					  SF_RETRY_UA | SF_NO_PRINT);
912 
913 			if (error == ERESTART) {
914 				/*
915 				 * A retry was scheduled, so
916  				 * just return.
917 				 */
918 				return;
919 			}
920 		}
921 
922 		/*
923 		 * Copy the allocated CCB contents back to the malloced CCB
924 		 * so we can give status back to the user when he requests it.
925 		 */
926 		bcopy(done_ccb, &io_req->ccb, sizeof(*done_ccb));
927 
928 		/*
929 		 * Log data/transaction completion with devstat(9).
930 		 */
931 		switch (done_ccb->ccb_h.func_code) {
932 		case XPT_SCSI_IO:
933 			devstat_end_transaction(softc->device_stats,
934 			    done_ccb->csio.dxfer_len - done_ccb->csio.resid,
935 			    done_ccb->csio.tag_action & 0x3,
936 			    ((done_ccb->ccb_h.flags & CAM_DIR_MASK) ==
937 			    CAM_DIR_NONE) ? DEVSTAT_NO_DATA :
938 			    (done_ccb->ccb_h.flags & CAM_DIR_OUT) ?
939 			    DEVSTAT_WRITE : DEVSTAT_READ, NULL,
940 			    &io_req->start_time);
941 			break;
942 		case XPT_ATA_IO:
943 			devstat_end_transaction(softc->device_stats,
944 			    done_ccb->ataio.dxfer_len - done_ccb->ataio.resid,
945 			    0, /* Not used in ATA */
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_SMP_IO:
953 			/*
954 			 * XXX KDM this isn't quite right, but there isn't
955 			 * currently an easy way to represent a bidirectional
956 			 * transfer in devstat.  The only way to do it
957 			 * and have the byte counts come out right would
958 			 * mean that we would have to record two
959 			 * transactions, one for the request and one for the
960 			 * response.  For now, so that we report something,
961 			 * just treat the entire thing as a read.
962 			 */
963 			devstat_end_transaction(softc->device_stats,
964 			    done_ccb->smpio.smp_request_len +
965 			    done_ccb->smpio.smp_response_len,
966 			    DEVSTAT_TAG_SIMPLE, DEVSTAT_READ, NULL,
967 			    &io_req->start_time);
968 			break;
969 		default:
970 			devstat_end_transaction(softc->device_stats, 0,
971 			    DEVSTAT_TAG_NONE, DEVSTAT_NO_DATA, NULL,
972 			    &io_req->start_time);
973 			break;
974 		}
975 
976 		/*
977 		 * In the normal case, take the completed I/O off of the
978 		 * active queue and put it on the done queue.  Notitfy the
979 		 * user that we have a completed I/O.
980 		 */
981 		if ((io_req->flags & PASS_IO_ABANDONED) == 0) {
982 			TAILQ_REMOVE(&softc->active_queue, io_req, links);
983 			TAILQ_INSERT_TAIL(&softc->done_queue, io_req, links);
984 			selwakeuppri(&softc->read_select, PRIBIO);
985 			KNOTE_LOCKED(&softc->read_select.si_note, 0);
986 		} else {
987 			/*
988 			 * In the case of an abandoned I/O (final close
989 			 * without fetching the I/O), take it off of the
990 			 * abandoned queue and free it.
991 			 */
992 			TAILQ_REMOVE(&softc->abandoned_queue, io_req, links);
993 			passiocleanup(softc, io_req);
994 			uma_zfree(softc->pass_zone, io_req);
995 
996 			/*
997 			 * Release the done_ccb here, since we may wind up
998 			 * freeing the peripheral when we decrement the
999 			 * reference count below.
1000 			 */
1001 			xpt_release_ccb(done_ccb);
1002 
1003 			/*
1004 			 * If the abandoned queue is empty, we can release
1005 			 * our reference to the periph since we won't have
1006 			 * any more completions coming.
1007 			 */
1008 			if ((TAILQ_EMPTY(&softc->abandoned_queue))
1009 			 && (softc->flags & PASS_FLAG_ABANDONED_REF_SET)) {
1010 				softc->flags &= ~PASS_FLAG_ABANDONED_REF_SET;
1011 				cam_periph_release_locked(periph);
1012 			}
1013 
1014 			/*
1015 			 * We have already released the CCB, so we can
1016 			 * return.
1017 			 */
1018 			return;
1019 		}
1020 		break;
1021 	}
1022 	}
1023 	xpt_release_ccb(done_ccb);
1024 }
1025 
1026 static int
1027 passcreatezone(struct cam_periph *periph)
1028 {
1029 	struct pass_softc *softc;
1030 	int error;
1031 
1032 	error = 0;
1033 	softc = (struct pass_softc *)periph->softc;
1034 
1035 	cam_periph_assert(periph, MA_OWNED);
1036 	KASSERT(((softc->flags & PASS_FLAG_ZONE_VALID) == 0),
1037 		("%s called when the pass(4) zone is valid!\n", __func__));
1038 	KASSERT((softc->pass_zone == NULL),
1039 		("%s called when the pass(4) zone is allocated!\n", __func__));
1040 
1041 	if ((softc->flags & PASS_FLAG_ZONE_INPROG) == 0) {
1042 
1043 		/*
1044 		 * We're the first context through, so we need to create
1045 		 * the pass(4) UMA zone for I/O requests.
1046 		 */
1047 		softc->flags |= PASS_FLAG_ZONE_INPROG;
1048 
1049 		/*
1050 		 * uma_zcreate() does a blocking (M_WAITOK) allocation,
1051 		 * so we cannot hold a mutex while we call it.
1052 		 */
1053 		cam_periph_unlock(periph);
1054 
1055 		softc->pass_zone = uma_zcreate(softc->zone_name,
1056 		    sizeof(struct pass_io_req), NULL, NULL, NULL, NULL,
1057 		    /*align*/ 0, /*flags*/ 0);
1058 
1059 		softc->pass_io_zone = uma_zcreate(softc->io_zone_name,
1060 		    softc->io_zone_size, NULL, NULL, NULL, NULL,
1061 		    /*align*/ 0, /*flags*/ 0);
1062 
1063 		cam_periph_lock(periph);
1064 
1065 		if ((softc->pass_zone == NULL)
1066 		 || (softc->pass_io_zone == NULL)) {
1067 			if (softc->pass_zone == NULL)
1068 				xpt_print(periph->path, "unable to allocate "
1069 				    "IO Req UMA zone\n");
1070 			else
1071 				xpt_print(periph->path, "unable to allocate "
1072 				    "IO UMA zone\n");
1073 			softc->flags &= ~PASS_FLAG_ZONE_INPROG;
1074 			goto bailout;
1075 		}
1076 
1077 		/*
1078 		 * Set the flags appropriately and notify any other waiters.
1079 		 */
1080 		softc->flags &= PASS_FLAG_ZONE_INPROG;
1081 		softc->flags |= PASS_FLAG_ZONE_VALID;
1082 		wakeup(&softc->pass_zone);
1083 	} else {
1084 		/*
1085 		 * In this case, the UMA zone has not yet been created, but
1086 		 * another context is in the process of creating it.  We
1087 		 * need to sleep until the creation is either done or has
1088 		 * failed.
1089 		 */
1090 		while ((softc->flags & PASS_FLAG_ZONE_INPROG)
1091 		    && ((softc->flags & PASS_FLAG_ZONE_VALID) == 0)) {
1092 			error = msleep(&softc->pass_zone,
1093 				       cam_periph_mtx(periph), PRIBIO,
1094 				       "paszon", 0);
1095 			if (error != 0)
1096 				goto bailout;
1097 		}
1098 		/*
1099 		 * If the zone creation failed, no luck for the user.
1100 		 */
1101 		if ((softc->flags & PASS_FLAG_ZONE_VALID) == 0){
1102 			error = ENOMEM;
1103 			goto bailout;
1104 		}
1105 	}
1106 bailout:
1107 	return (error);
1108 }
1109 
1110 static void
1111 passiocleanup(struct pass_softc *softc, struct pass_io_req *io_req)
1112 {
1113 	union ccb *ccb;
1114 	u_int8_t **data_ptrs[CAM_PERIPH_MAXMAPS];
1115 	int i, numbufs;
1116 
1117 	ccb = &io_req->ccb;
1118 
1119 	switch (ccb->ccb_h.func_code) {
1120 	case XPT_DEV_MATCH:
1121 		numbufs = min(io_req->num_bufs, 2);
1122 
1123 		if (numbufs == 1) {
1124 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.matches;
1125 		} else {
1126 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.patterns;
1127 			data_ptrs[1] = (u_int8_t **)&ccb->cdm.matches;
1128 		}
1129 		break;
1130 	case XPT_SCSI_IO:
1131 	case XPT_CONT_TARGET_IO:
1132 		data_ptrs[0] = &ccb->csio.data_ptr;
1133 		numbufs = min(io_req->num_bufs, 1);
1134 		break;
1135 	case XPT_ATA_IO:
1136 		data_ptrs[0] = &ccb->ataio.data_ptr;
1137 		numbufs = min(io_req->num_bufs, 1);
1138 		break;
1139 	case XPT_SMP_IO:
1140 		numbufs = min(io_req->num_bufs, 2);
1141 		data_ptrs[0] = &ccb->smpio.smp_request;
1142 		data_ptrs[1] = &ccb->smpio.smp_response;
1143 		break;
1144 	case XPT_DEV_ADVINFO:
1145 		numbufs = min(io_req->num_bufs, 1);
1146 		data_ptrs[0] = (uint8_t **)&ccb->cdai.buf;
1147 		break;
1148 	case XPT_NVME_IO:
1149 	case XPT_NVME_ADMIN:
1150 		data_ptrs[0] = &ccb->nvmeio.data_ptr;
1151 		numbufs = min(io_req->num_bufs, 1);
1152 		break;
1153 	default:
1154 		/* allow ourselves to be swapped once again */
1155 		return;
1156 		break; /* NOTREACHED */
1157 	}
1158 
1159 	if (io_req->flags & PASS_IO_USER_SEG_MALLOC) {
1160 		free(io_req->user_segptr, M_SCSIPASS);
1161 		io_req->user_segptr = NULL;
1162 	}
1163 
1164 	/*
1165 	 * We only want to free memory we malloced.
1166 	 */
1167 	if (io_req->data_flags == CAM_DATA_VADDR) {
1168 		for (i = 0; i < io_req->num_bufs; i++) {
1169 			if (io_req->kern_bufs[i] == NULL)
1170 				continue;
1171 
1172 			free(io_req->kern_bufs[i], M_SCSIPASS);
1173 			io_req->kern_bufs[i] = NULL;
1174 		}
1175 	} else if (io_req->data_flags == CAM_DATA_SG) {
1176 		for (i = 0; i < io_req->num_kern_segs; i++) {
1177 			if ((uint8_t *)(uintptr_t)
1178 			    io_req->kern_segptr[i].ds_addr == NULL)
1179 				continue;
1180 
1181 			uma_zfree(softc->pass_io_zone, (uint8_t *)(uintptr_t)
1182 			    io_req->kern_segptr[i].ds_addr);
1183 			io_req->kern_segptr[i].ds_addr = 0;
1184 		}
1185 	}
1186 
1187 	if (io_req->flags & PASS_IO_KERN_SEG_MALLOC) {
1188 		free(io_req->kern_segptr, M_SCSIPASS);
1189 		io_req->kern_segptr = NULL;
1190 	}
1191 
1192 	if (io_req->data_flags != CAM_DATA_PADDR) {
1193 		for (i = 0; i < numbufs; i++) {
1194 			/*
1195 			 * Restore the user's buffer pointers to their
1196 			 * previous values.
1197 			 */
1198 			if (io_req->user_bufs[i] != NULL)
1199 				*data_ptrs[i] = io_req->user_bufs[i];
1200 		}
1201 	}
1202 
1203 }
1204 
1205 static int
1206 passcopysglist(struct cam_periph *periph, struct pass_io_req *io_req,
1207 	       ccb_flags direction)
1208 {
1209 	bus_size_t kern_watermark, user_watermark, len_copied, len_to_copy;
1210 	bus_dma_segment_t *user_sglist, *kern_sglist;
1211 	int i, j, error;
1212 
1213 	error = 0;
1214 	kern_watermark = 0;
1215 	user_watermark = 0;
1216 	len_to_copy = 0;
1217 	len_copied = 0;
1218 	user_sglist = io_req->user_segptr;
1219 	kern_sglist = io_req->kern_segptr;
1220 
1221 	for (i = 0, j = 0; i < io_req->num_user_segs &&
1222 	     j < io_req->num_kern_segs;) {
1223 		uint8_t *user_ptr, *kern_ptr;
1224 
1225 		len_to_copy = min(user_sglist[i].ds_len -user_watermark,
1226 		    kern_sglist[j].ds_len - kern_watermark);
1227 
1228 		user_ptr = (uint8_t *)(uintptr_t)user_sglist[i].ds_addr;
1229 		user_ptr = user_ptr + user_watermark;
1230 		kern_ptr = (uint8_t *)(uintptr_t)kern_sglist[j].ds_addr;
1231 		kern_ptr = kern_ptr + kern_watermark;
1232 
1233 		user_watermark += len_to_copy;
1234 		kern_watermark += len_to_copy;
1235 
1236 		if (!useracc(user_ptr, len_to_copy,
1237 		    (direction == CAM_DIR_IN) ? VM_PROT_WRITE : VM_PROT_READ)) {
1238 			xpt_print(periph->path, "%s: unable to access user "
1239 				  "S/G list element %p len %zu\n", __func__,
1240 				  user_ptr, len_to_copy);
1241 			error = EFAULT;
1242 			goto bailout;
1243 		}
1244 
1245 		if (direction == CAM_DIR_IN) {
1246 			error = copyout(kern_ptr, user_ptr, len_to_copy);
1247 			if (error != 0) {
1248 				xpt_print(periph->path, "%s: copyout of %u "
1249 					  "bytes from %p to %p failed with "
1250 					  "error %d\n", __func__, len_to_copy,
1251 					  kern_ptr, user_ptr, error);
1252 				goto bailout;
1253 			}
1254 		} else {
1255 			error = copyin(user_ptr, kern_ptr, len_to_copy);
1256 			if (error != 0) {
1257 				xpt_print(periph->path, "%s: copyin of %u "
1258 					  "bytes from %p to %p failed with "
1259 					  "error %d\n", __func__, len_to_copy,
1260 					  user_ptr, kern_ptr, error);
1261 				goto bailout;
1262 			}
1263 		}
1264 
1265 		len_copied += len_to_copy;
1266 
1267 		if (user_sglist[i].ds_len == user_watermark) {
1268 			i++;
1269 			user_watermark = 0;
1270 		}
1271 
1272 		if (kern_sglist[j].ds_len == kern_watermark) {
1273 			j++;
1274 			kern_watermark = 0;
1275 		}
1276 	}
1277 
1278 bailout:
1279 
1280 	return (error);
1281 }
1282 
1283 static int
1284 passmemsetup(struct cam_periph *periph, struct pass_io_req *io_req)
1285 {
1286 	union ccb *ccb;
1287 	struct pass_softc *softc;
1288 	int numbufs, i;
1289 	uint8_t **data_ptrs[CAM_PERIPH_MAXMAPS];
1290 	uint32_t lengths[CAM_PERIPH_MAXMAPS];
1291 	uint32_t dirs[CAM_PERIPH_MAXMAPS];
1292 	uint32_t num_segs;
1293 	uint16_t *seg_cnt_ptr;
1294 	size_t maxmap;
1295 	int error;
1296 
1297 	cam_periph_assert(periph, MA_NOTOWNED);
1298 
1299 	softc = periph->softc;
1300 
1301 	error = 0;
1302 	ccb = &io_req->ccb;
1303 	maxmap = 0;
1304 	num_segs = 0;
1305 	seg_cnt_ptr = NULL;
1306 
1307 	switch(ccb->ccb_h.func_code) {
1308 	case XPT_DEV_MATCH:
1309 		if (ccb->cdm.match_buf_len == 0) {
1310 			printf("%s: invalid match buffer length 0\n", __func__);
1311 			return(EINVAL);
1312 		}
1313 		if (ccb->cdm.pattern_buf_len > 0) {
1314 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.patterns;
1315 			lengths[0] = ccb->cdm.pattern_buf_len;
1316 			dirs[0] = CAM_DIR_OUT;
1317 			data_ptrs[1] = (u_int8_t **)&ccb->cdm.matches;
1318 			lengths[1] = ccb->cdm.match_buf_len;
1319 			dirs[1] = CAM_DIR_IN;
1320 			numbufs = 2;
1321 		} else {
1322 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.matches;
1323 			lengths[0] = ccb->cdm.match_buf_len;
1324 			dirs[0] = CAM_DIR_IN;
1325 			numbufs = 1;
1326 		}
1327 		io_req->data_flags = CAM_DATA_VADDR;
1328 		break;
1329 	case XPT_SCSI_IO:
1330 	case XPT_CONT_TARGET_IO:
1331 		if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
1332 			return(0);
1333 
1334 		/*
1335 		 * The user shouldn't be able to supply a bio.
1336 		 */
1337 		if ((ccb->ccb_h.flags & CAM_DATA_MASK) == CAM_DATA_BIO)
1338 			return (EINVAL);
1339 
1340 		io_req->data_flags = ccb->ccb_h.flags & CAM_DATA_MASK;
1341 
1342 		data_ptrs[0] = &ccb->csio.data_ptr;
1343 		lengths[0] = ccb->csio.dxfer_len;
1344 		dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1345 		num_segs = ccb->csio.sglist_cnt;
1346 		seg_cnt_ptr = &ccb->csio.sglist_cnt;
1347 		numbufs = 1;
1348 		maxmap = softc->maxio;
1349 		break;
1350 	case XPT_ATA_IO:
1351 		if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
1352 			return(0);
1353 
1354 		/*
1355 		 * We only support a single virtual address for ATA I/O.
1356 		 */
1357 		if ((ccb->ccb_h.flags & CAM_DATA_MASK) != CAM_DATA_VADDR)
1358 			return (EINVAL);
1359 
1360 		io_req->data_flags = CAM_DATA_VADDR;
1361 
1362 		data_ptrs[0] = &ccb->ataio.data_ptr;
1363 		lengths[0] = ccb->ataio.dxfer_len;
1364 		dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1365 		numbufs = 1;
1366 		maxmap = softc->maxio;
1367 		break;
1368 	case XPT_SMP_IO:
1369 		io_req->data_flags = CAM_DATA_VADDR;
1370 
1371 		data_ptrs[0] = &ccb->smpio.smp_request;
1372 		lengths[0] = ccb->smpio.smp_request_len;
1373 		dirs[0] = CAM_DIR_OUT;
1374 		data_ptrs[1] = &ccb->smpio.smp_response;
1375 		lengths[1] = ccb->smpio.smp_response_len;
1376 		dirs[1] = CAM_DIR_IN;
1377 		numbufs = 2;
1378 		maxmap = softc->maxio;
1379 		break;
1380 	case XPT_DEV_ADVINFO:
1381 		if (ccb->cdai.bufsiz == 0)
1382 			return (0);
1383 
1384 		io_req->data_flags = CAM_DATA_VADDR;
1385 
1386 		data_ptrs[0] = (uint8_t **)&ccb->cdai.buf;
1387 		lengths[0] = ccb->cdai.bufsiz;
1388 		dirs[0] = CAM_DIR_IN;
1389 		numbufs = 1;
1390 		break;
1391 	case XPT_NVME_ADMIN:
1392 	case XPT_NVME_IO:
1393 		if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
1394 			return (0);
1395 
1396 		io_req->data_flags = ccb->ccb_h.flags & CAM_DATA_MASK;
1397 
1398 		data_ptrs[0] = &ccb->nvmeio.data_ptr;
1399 		lengths[0] = ccb->nvmeio.dxfer_len;
1400 		dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1401 		num_segs = ccb->nvmeio.sglist_cnt;
1402 		seg_cnt_ptr = &ccb->nvmeio.sglist_cnt;
1403 		numbufs = 1;
1404 		maxmap = softc->maxio;
1405 		break;
1406 	default:
1407 		return(EINVAL);
1408 		break; /* NOTREACHED */
1409 	}
1410 
1411 	io_req->num_bufs = numbufs;
1412 
1413 	/*
1414 	 * If there is a maximum, check to make sure that the user's
1415 	 * request fits within the limit.  In general, we should only have
1416 	 * a maximum length for requests that go to hardware.  Otherwise it
1417 	 * is whatever we're able to malloc.
1418 	 */
1419 	for (i = 0; i < numbufs; i++) {
1420 		io_req->user_bufs[i] = *data_ptrs[i];
1421 		io_req->dirs[i] = dirs[i];
1422 		io_req->lengths[i] = lengths[i];
1423 
1424 		if (maxmap == 0)
1425 			continue;
1426 
1427 		if (lengths[i] <= maxmap)
1428 			continue;
1429 
1430 		xpt_print(periph->path, "%s: data length %u > max allowed %u "
1431 			  "bytes\n", __func__, lengths[i], maxmap);
1432 		error = EINVAL;
1433 		goto bailout;
1434 	}
1435 
1436 	switch (io_req->data_flags) {
1437 	case CAM_DATA_VADDR:
1438 		/* Map or copy the buffer into kernel address space */
1439 		for (i = 0; i < numbufs; i++) {
1440 			uint8_t *tmp_buf;
1441 
1442 			/*
1443 			 * If for some reason no length is specified, we
1444 			 * don't need to allocate anything.
1445 			 */
1446 			if (io_req->lengths[i] == 0)
1447 				continue;
1448 
1449 			/*
1450 			 * Make sure that the user's buffer is accessible
1451 			 * to that process.
1452 			 */
1453 			if (!useracc(io_req->user_bufs[i], io_req->lengths[i],
1454 			    (io_req->dirs[i] == CAM_DIR_IN) ? VM_PROT_WRITE :
1455 			     VM_PROT_READ)) {
1456 				xpt_print(periph->path, "%s: user address %p "
1457 				    "length %u is not accessible\n", __func__,
1458 				    io_req->user_bufs[i], io_req->lengths[i]);
1459 				error = EFAULT;
1460 				goto bailout;
1461 			}
1462 
1463 			tmp_buf = malloc(lengths[i], M_SCSIPASS,
1464 					 M_WAITOK | M_ZERO);
1465 			io_req->kern_bufs[i] = tmp_buf;
1466 			*data_ptrs[i] = tmp_buf;
1467 
1468 #if 0
1469 			xpt_print(periph->path, "%s: malloced %p len %u, user "
1470 				  "buffer %p, operation: %s\n", __func__,
1471 				  tmp_buf, lengths[i], io_req->user_bufs[i],
1472 				  (dirs[i] == CAM_DIR_IN) ? "read" : "write");
1473 #endif
1474 			/*
1475 			 * We only need to copy in if the user is writing.
1476 			 */
1477 			if (dirs[i] != CAM_DIR_OUT)
1478 				continue;
1479 
1480 			error = copyin(io_req->user_bufs[i],
1481 				       io_req->kern_bufs[i], lengths[i]);
1482 			if (error != 0) {
1483 				xpt_print(periph->path, "%s: copy of user "
1484 					  "buffer from %p to %p failed with "
1485 					  "error %d\n", __func__,
1486 					  io_req->user_bufs[i],
1487 					  io_req->kern_bufs[i], error);
1488 				goto bailout;
1489 			}
1490 		}
1491 		break;
1492 	case CAM_DATA_PADDR:
1493 		/* Pass down the pointer as-is */
1494 		break;
1495 	case CAM_DATA_SG: {
1496 		size_t sg_length, size_to_go, alloc_size;
1497 		uint32_t num_segs_needed;
1498 
1499 		/*
1500 		 * Copy the user S/G list in, and then copy in the
1501 		 * individual segments.
1502 		 */
1503 		/*
1504 		 * We shouldn't see this, but check just in case.
1505 		 */
1506 		if (numbufs != 1) {
1507 			xpt_print(periph->path, "%s: cannot currently handle "
1508 				  "more than one S/G list per CCB\n", __func__);
1509 			error = EINVAL;
1510 			goto bailout;
1511 		}
1512 
1513 		/*
1514 		 * We have to have at least one segment.
1515 		 */
1516 		if (num_segs == 0) {
1517 			xpt_print(periph->path, "%s: CAM_DATA_SG flag set, "
1518 				  "but sglist_cnt=0!\n", __func__);
1519 			error = EINVAL;
1520 			goto bailout;
1521 		}
1522 
1523 		/*
1524 		 * Make sure the user specified the total length and didn't
1525 		 * just leave it to us to decode the S/G list.
1526 		 */
1527 		if (lengths[0] == 0) {
1528 			xpt_print(periph->path, "%s: no dxfer_len specified, "
1529 				  "but CAM_DATA_SG flag is set!\n", __func__);
1530 			error = EINVAL;
1531 			goto bailout;
1532 		}
1533 
1534 		/*
1535 		 * We allocate buffers in io_zone_size increments for an
1536 		 * S/G list.  This will generally be MAXPHYS.
1537 		 */
1538 		if (lengths[0] <= softc->io_zone_size)
1539 			num_segs_needed = 1;
1540 		else {
1541 			num_segs_needed = lengths[0] / softc->io_zone_size;
1542 			if ((lengths[0] % softc->io_zone_size) != 0)
1543 				num_segs_needed++;
1544 		}
1545 
1546 		/* Figure out the size of the S/G list */
1547 		sg_length = num_segs * sizeof(bus_dma_segment_t);
1548 		io_req->num_user_segs = num_segs;
1549 		io_req->num_kern_segs = num_segs_needed;
1550 
1551 		/* Save the user's S/G list pointer for later restoration */
1552 		io_req->user_bufs[0] = *data_ptrs[0];
1553 
1554 		/*
1555 		 * If we have enough segments allocated by default to handle
1556 		 * the length of the user's S/G list,
1557 		 */
1558 		if (num_segs > PASS_MAX_SEGS) {
1559 			io_req->user_segptr = malloc(sizeof(bus_dma_segment_t) *
1560 			    num_segs, M_SCSIPASS, M_WAITOK | M_ZERO);
1561 			io_req->flags |= PASS_IO_USER_SEG_MALLOC;
1562 		} else
1563 			io_req->user_segptr = io_req->user_segs;
1564 
1565 		if (!useracc(*data_ptrs[0], sg_length, VM_PROT_READ)) {
1566 			xpt_print(periph->path, "%s: unable to access user "
1567 				  "S/G list at %p\n", __func__, *data_ptrs[0]);
1568 			error = EFAULT;
1569 			goto bailout;
1570 		}
1571 
1572 		error = copyin(*data_ptrs[0], io_req->user_segptr, sg_length);
1573 		if (error != 0) {
1574 			xpt_print(periph->path, "%s: copy of user S/G list "
1575 				  "from %p to %p failed with error %d\n",
1576 				  __func__, *data_ptrs[0], io_req->user_segptr,
1577 				  error);
1578 			goto bailout;
1579 		}
1580 
1581 		if (num_segs_needed > PASS_MAX_SEGS) {
1582 			io_req->kern_segptr = malloc(sizeof(bus_dma_segment_t) *
1583 			    num_segs_needed, M_SCSIPASS, M_WAITOK | M_ZERO);
1584 			io_req->flags |= PASS_IO_KERN_SEG_MALLOC;
1585 		} else {
1586 			io_req->kern_segptr = io_req->kern_segs;
1587 		}
1588 
1589 		/*
1590 		 * Allocate the kernel S/G list.
1591 		 */
1592 		for (size_to_go = lengths[0], i = 0;
1593 		     size_to_go > 0 && i < num_segs_needed;
1594 		     i++, size_to_go -= alloc_size) {
1595 			uint8_t *kern_ptr;
1596 
1597 			alloc_size = min(size_to_go, softc->io_zone_size);
1598 			kern_ptr = uma_zalloc(softc->pass_io_zone, M_WAITOK);
1599 			io_req->kern_segptr[i].ds_addr =
1600 			    (bus_addr_t)(uintptr_t)kern_ptr;
1601 			io_req->kern_segptr[i].ds_len = alloc_size;
1602 		}
1603 		if (size_to_go > 0) {
1604 			printf("%s: size_to_go = %zu, software error!\n",
1605 			       __func__, size_to_go);
1606 			error = EINVAL;
1607 			goto bailout;
1608 		}
1609 
1610 		*data_ptrs[0] = (uint8_t *)io_req->kern_segptr;
1611 		*seg_cnt_ptr = io_req->num_kern_segs;
1612 
1613 		/*
1614 		 * We only need to copy data here if the user is writing.
1615 		 */
1616 		if (dirs[0] == CAM_DIR_OUT)
1617 			error = passcopysglist(periph, io_req, dirs[0]);
1618 		break;
1619 	}
1620 	case CAM_DATA_SG_PADDR: {
1621 		size_t sg_length;
1622 
1623 		/*
1624 		 * We shouldn't see this, but check just in case.
1625 		 */
1626 		if (numbufs != 1) {
1627 			printf("%s: cannot currently handle more than one "
1628 			       "S/G list per CCB\n", __func__);
1629 			error = EINVAL;
1630 			goto bailout;
1631 		}
1632 
1633 		/*
1634 		 * We have to have at least one segment.
1635 		 */
1636 		if (num_segs == 0) {
1637 			xpt_print(periph->path, "%s: CAM_DATA_SG_PADDR flag "
1638 				  "set, but sglist_cnt=0!\n", __func__);
1639 			error = EINVAL;
1640 			goto bailout;
1641 		}
1642 
1643 		/*
1644 		 * Make sure the user specified the total length and didn't
1645 		 * just leave it to us to decode the S/G list.
1646 		 */
1647 		if (lengths[0] == 0) {
1648 			xpt_print(periph->path, "%s: no dxfer_len specified, "
1649 				  "but CAM_DATA_SG flag is set!\n", __func__);
1650 			error = EINVAL;
1651 			goto bailout;
1652 		}
1653 
1654 		/* Figure out the size of the S/G list */
1655 		sg_length = num_segs * sizeof(bus_dma_segment_t);
1656 		io_req->num_user_segs = num_segs;
1657 		io_req->num_kern_segs = io_req->num_user_segs;
1658 
1659 		/* Save the user's S/G list pointer for later restoration */
1660 		io_req->user_bufs[0] = *data_ptrs[0];
1661 
1662 		if (num_segs > PASS_MAX_SEGS) {
1663 			io_req->user_segptr = malloc(sizeof(bus_dma_segment_t) *
1664 			    num_segs, M_SCSIPASS, M_WAITOK | M_ZERO);
1665 			io_req->flags |= PASS_IO_USER_SEG_MALLOC;
1666 		} else
1667 			io_req->user_segptr = io_req->user_segs;
1668 
1669 		io_req->kern_segptr = io_req->user_segptr;
1670 
1671 		error = copyin(*data_ptrs[0], io_req->user_segptr, sg_length);
1672 		if (error != 0) {
1673 			xpt_print(periph->path, "%s: copy of user S/G list "
1674 				  "from %p to %p failed with error %d\n",
1675 				  __func__, *data_ptrs[0], io_req->user_segptr,
1676 				  error);
1677 			goto bailout;
1678 		}
1679 		break;
1680 	}
1681 	default:
1682 	case CAM_DATA_BIO:
1683 		/*
1684 		 * A user shouldn't be attaching a bio to the CCB.  It
1685 		 * isn't a user-accessible structure.
1686 		 */
1687 		error = EINVAL;
1688 		break;
1689 	}
1690 
1691 bailout:
1692 	if (error != 0)
1693 		passiocleanup(softc, io_req);
1694 
1695 	return (error);
1696 }
1697 
1698 static int
1699 passmemdone(struct cam_periph *periph, struct pass_io_req *io_req)
1700 {
1701 	struct pass_softc *softc;
1702 	int error;
1703 	int i;
1704 
1705 	error = 0;
1706 	softc = (struct pass_softc *)periph->softc;
1707 
1708 	switch (io_req->data_flags) {
1709 	case CAM_DATA_VADDR:
1710 		/*
1711 		 * Copy back to the user buffer if this was a read.
1712 		 */
1713 		for (i = 0; i < io_req->num_bufs; i++) {
1714 			if (io_req->dirs[i] != CAM_DIR_IN)
1715 				continue;
1716 
1717 			error = copyout(io_req->kern_bufs[i],
1718 			    io_req->user_bufs[i], io_req->lengths[i]);
1719 			if (error != 0) {
1720 				xpt_print(periph->path, "Unable to copy %u "
1721 					  "bytes from %p to user address %p\n",
1722 					  io_req->lengths[i],
1723 					  io_req->kern_bufs[i],
1724 					  io_req->user_bufs[i]);
1725 				goto bailout;
1726 			}
1727 
1728 		}
1729 		break;
1730 	case CAM_DATA_PADDR:
1731 		/* Do nothing.  The pointer is a physical address already */
1732 		break;
1733 	case CAM_DATA_SG:
1734 		/*
1735 		 * Copy back to the user buffer if this was a read.
1736 		 * Restore the user's S/G list buffer pointer.
1737 		 */
1738 		if (io_req->dirs[0] == CAM_DIR_IN)
1739 			error = passcopysglist(periph, io_req, io_req->dirs[0]);
1740 		break;
1741 	case CAM_DATA_SG_PADDR:
1742 		/*
1743 		 * Restore the user's S/G list buffer pointer.  No need to
1744 		 * copy.
1745 		 */
1746 		break;
1747 	default:
1748 	case CAM_DATA_BIO:
1749 		error = EINVAL;
1750 		break;
1751 	}
1752 
1753 bailout:
1754 	/*
1755 	 * Reset the user's pointers to their original values and free
1756 	 * allocated memory.
1757 	 */
1758 	passiocleanup(softc, io_req);
1759 
1760 	return (error);
1761 }
1762 
1763 static int
1764 passioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flag, struct thread *td)
1765 {
1766 	int error;
1767 
1768 	if ((error = passdoioctl(dev, cmd, addr, flag, td)) == ENOTTY) {
1769 		error = cam_compat_ioctl(dev, cmd, addr, flag, td, passdoioctl);
1770 	}
1771 	return (error);
1772 }
1773 
1774 static int
1775 passdoioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flag, struct thread *td)
1776 {
1777 	struct	cam_periph *periph;
1778 	struct	pass_softc *softc;
1779 	int	error;
1780 	uint32_t priority;
1781 
1782 	periph = (struct cam_periph *)dev->si_drv1;
1783 	cam_periph_lock(periph);
1784 	softc = (struct pass_softc *)periph->softc;
1785 
1786 	error = 0;
1787 
1788 	switch (cmd) {
1789 
1790 	case CAMIOCOMMAND:
1791 	{
1792 		union ccb *inccb;
1793 		union ccb *ccb;
1794 		int ccb_malloced;
1795 
1796 		inccb = (union ccb *)addr;
1797 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
1798 		if (inccb->ccb_h.func_code == XPT_SCSI_IO)
1799 			inccb->csio.bio = NULL;
1800 #endif
1801 
1802 		if (inccb->ccb_h.flags & CAM_UNLOCKED) {
1803 			error = EINVAL;
1804 			break;
1805 		}
1806 
1807 		/*
1808 		 * Some CCB types, like scan bus and scan lun can only go
1809 		 * through the transport layer device.
1810 		 */
1811 		if (inccb->ccb_h.func_code & XPT_FC_XPT_ONLY) {
1812 			xpt_print(periph->path, "CCB function code %#x is "
1813 			    "restricted to the XPT device\n",
1814 			    inccb->ccb_h.func_code);
1815 			error = ENODEV;
1816 			break;
1817 		}
1818 
1819 		/* Compatibility for RL/priority-unaware code. */
1820 		priority = inccb->ccb_h.pinfo.priority;
1821 		if (priority <= CAM_PRIORITY_OOB)
1822 		    priority += CAM_PRIORITY_OOB + 1;
1823 
1824 		/*
1825 		 * Non-immediate CCBs need a CCB from the per-device pool
1826 		 * of CCBs, which is scheduled by the transport layer.
1827 		 * Immediate CCBs and user-supplied CCBs should just be
1828 		 * malloced.
1829 		 */
1830 		if ((inccb->ccb_h.func_code & XPT_FC_QUEUED)
1831 		 && ((inccb->ccb_h.func_code & XPT_FC_USER_CCB) == 0)) {
1832 			ccb = cam_periph_getccb(periph, priority);
1833 			ccb_malloced = 0;
1834 		} else {
1835 			ccb = xpt_alloc_ccb_nowait();
1836 
1837 			if (ccb != NULL)
1838 				xpt_setup_ccb(&ccb->ccb_h, periph->path,
1839 					      priority);
1840 			ccb_malloced = 1;
1841 		}
1842 
1843 		if (ccb == NULL) {
1844 			xpt_print(periph->path, "unable to allocate CCB\n");
1845 			error = ENOMEM;
1846 			break;
1847 		}
1848 
1849 		error = passsendccb(periph, ccb, inccb);
1850 
1851 		if (ccb_malloced)
1852 			xpt_free_ccb(ccb);
1853 		else
1854 			xpt_release_ccb(ccb);
1855 
1856 		break;
1857 	}
1858 	case CAMIOQUEUE:
1859 	{
1860 		struct pass_io_req *io_req;
1861 		union ccb **user_ccb, *ccb;
1862 		xpt_opcode fc;
1863 
1864 		if ((softc->flags & PASS_FLAG_ZONE_VALID) == 0) {
1865 			error = passcreatezone(periph);
1866 			if (error != 0)
1867 				goto bailout;
1868 		}
1869 
1870 		/*
1871 		 * We're going to do a blocking allocation for this I/O
1872 		 * request, so we have to drop the lock.
1873 		 */
1874 		cam_periph_unlock(periph);
1875 
1876 		io_req = uma_zalloc(softc->pass_zone, M_WAITOK | M_ZERO);
1877 		ccb = &io_req->ccb;
1878 		user_ccb = (union ccb **)addr;
1879 
1880 		/*
1881 		 * Unlike the CAMIOCOMMAND ioctl above, we only have a
1882 		 * pointer to the user's CCB, so we have to copy the whole
1883 		 * thing in to a buffer we have allocated (above) instead
1884 		 * of allowing the ioctl code to malloc a buffer and copy
1885 		 * it in.
1886 		 *
1887 		 * This is an advantage for this asynchronous interface,
1888 		 * since we don't want the memory to get freed while the
1889 		 * CCB is outstanding.
1890 		 */
1891 #if 0
1892 		xpt_print(periph->path, "Copying user CCB %p to "
1893 			  "kernel address %p\n", *user_ccb, ccb);
1894 #endif
1895 		error = copyin(*user_ccb, ccb, sizeof(*ccb));
1896 		if (error != 0) {
1897 			xpt_print(periph->path, "Copy of user CCB %p to "
1898 				  "kernel address %p failed with error %d\n",
1899 				  *user_ccb, ccb, error);
1900 			goto camioqueue_error;
1901 		}
1902 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
1903 		if (ccb->ccb_h.func_code == XPT_SCSI_IO)
1904 			ccb->csio.bio = NULL;
1905 #endif
1906 
1907 		if (ccb->ccb_h.flags & CAM_UNLOCKED) {
1908 			error = EINVAL;
1909 			goto camioqueue_error;
1910 		}
1911 
1912 		if (ccb->ccb_h.flags & CAM_CDB_POINTER) {
1913 			if (ccb->csio.cdb_len > IOCDBLEN) {
1914 				error = EINVAL;
1915 				goto camioqueue_error;
1916 			}
1917 			error = copyin(ccb->csio.cdb_io.cdb_ptr,
1918 			    ccb->csio.cdb_io.cdb_bytes, ccb->csio.cdb_len);
1919 			if (error != 0)
1920 				goto camioqueue_error;
1921 			ccb->ccb_h.flags &= ~CAM_CDB_POINTER;
1922 		}
1923 
1924 		/*
1925 		 * Some CCB types, like scan bus and scan lun can only go
1926 		 * through the transport layer device.
1927 		 */
1928 		if (ccb->ccb_h.func_code & XPT_FC_XPT_ONLY) {
1929 			xpt_print(periph->path, "CCB function code %#x is "
1930 			    "restricted to the XPT device\n",
1931 			    ccb->ccb_h.func_code);
1932 			error = ENODEV;
1933 			goto camioqueue_error;
1934 		}
1935 
1936 		/*
1937 		 * Save the user's CCB pointer as well as his linked list
1938 		 * pointers and peripheral private area so that we can
1939 		 * restore these later.
1940 		 */
1941 		io_req->user_ccb_ptr = *user_ccb;
1942 		io_req->user_periph_links = ccb->ccb_h.periph_links;
1943 		io_req->user_periph_priv = ccb->ccb_h.periph_priv;
1944 
1945 		/*
1946 		 * Now that we've saved the user's values, we can set our
1947 		 * own peripheral private entry.
1948 		 */
1949 		ccb->ccb_h.ccb_ioreq = io_req;
1950 
1951 		/* Compatibility for RL/priority-unaware code. */
1952 		priority = ccb->ccb_h.pinfo.priority;
1953 		if (priority <= CAM_PRIORITY_OOB)
1954 		    priority += CAM_PRIORITY_OOB + 1;
1955 
1956 		/*
1957 		 * Setup fields in the CCB like the path and the priority.
1958 		 * The path in particular cannot be done in userland, since
1959 		 * it is a pointer to a kernel data structure.
1960 		 */
1961 		xpt_setup_ccb_flags(&ccb->ccb_h, periph->path, priority,
1962 				    ccb->ccb_h.flags);
1963 
1964 		/*
1965 		 * Setup our done routine.  There is no way for the user to
1966 		 * have a valid pointer here.
1967 		 */
1968 		ccb->ccb_h.cbfcnp = passdone;
1969 
1970 		fc = ccb->ccb_h.func_code;
1971 		/*
1972 		 * If this function code has memory that can be mapped in
1973 		 * or out, we need to call passmemsetup().
1974 		 */
1975 		if ((fc == XPT_SCSI_IO) || (fc == XPT_ATA_IO)
1976 		 || (fc == XPT_SMP_IO) || (fc == XPT_DEV_MATCH)
1977 		 || (fc == XPT_DEV_ADVINFO)
1978 		 || (fc == XPT_NVME_ADMIN) || (fc == XPT_NVME_IO)) {
1979 			error = passmemsetup(periph, io_req);
1980 			if (error != 0)
1981 				goto camioqueue_error;
1982 		} else
1983 			io_req->mapinfo.num_bufs_used = 0;
1984 
1985 		cam_periph_lock(periph);
1986 
1987 		/*
1988 		 * Everything goes on the incoming queue initially.
1989 		 */
1990 		TAILQ_INSERT_TAIL(&softc->incoming_queue, io_req, links);
1991 
1992 		/*
1993 		 * If the CCB is queued, and is not a user CCB, then
1994 		 * we need to allocate a slot for it.  Call xpt_schedule()
1995 		 * so that our start routine will get called when a CCB is
1996 		 * available.
1997 		 */
1998 		if ((fc & XPT_FC_QUEUED)
1999 		 && ((fc & XPT_FC_USER_CCB) == 0)) {
2000 			xpt_schedule(periph, priority);
2001 			break;
2002 		}
2003 
2004 		/*
2005 		 * At this point, the CCB in question is either an
2006 		 * immediate CCB (like XPT_DEV_ADVINFO) or it is a user CCB
2007 		 * and therefore should be malloced, not allocated via a slot.
2008 		 * Remove the CCB from the incoming queue and add it to the
2009 		 * active queue.
2010 		 */
2011 		TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
2012 		TAILQ_INSERT_TAIL(&softc->active_queue, io_req, links);
2013 
2014 		xpt_action(ccb);
2015 
2016 		/*
2017 		 * If this is not a queued CCB (i.e. it is an immediate CCB),
2018 		 * then it is already done.  We need to put it on the done
2019 		 * queue for the user to fetch.
2020 		 */
2021 		if ((fc & XPT_FC_QUEUED) == 0) {
2022 			TAILQ_REMOVE(&softc->active_queue, io_req, links);
2023 			TAILQ_INSERT_TAIL(&softc->done_queue, io_req, links);
2024 		}
2025 		break;
2026 
2027 camioqueue_error:
2028 		uma_zfree(softc->pass_zone, io_req);
2029 		cam_periph_lock(periph);
2030 		break;
2031 	}
2032 	case CAMIOGET:
2033 	{
2034 		union ccb **user_ccb;
2035 		struct pass_io_req *io_req;
2036 		int old_error;
2037 
2038 		user_ccb = (union ccb **)addr;
2039 		old_error = 0;
2040 
2041 		io_req = TAILQ_FIRST(&softc->done_queue);
2042 		if (io_req == NULL) {
2043 			error = ENOENT;
2044 			break;
2045 		}
2046 
2047 		/*
2048 		 * Remove the I/O from the done queue.
2049 		 */
2050 		TAILQ_REMOVE(&softc->done_queue, io_req, links);
2051 
2052 		/*
2053 		 * We have to drop the lock during the copyout because the
2054 		 * copyout can result in VM faults that require sleeping.
2055 		 */
2056 		cam_periph_unlock(periph);
2057 
2058 		/*
2059 		 * Do any needed copies (e.g. for reads) and revert the
2060 		 * pointers in the CCB back to the user's pointers.
2061 		 */
2062 		error = passmemdone(periph, io_req);
2063 
2064 		old_error = error;
2065 
2066 		io_req->ccb.ccb_h.periph_links = io_req->user_periph_links;
2067 		io_req->ccb.ccb_h.periph_priv = io_req->user_periph_priv;
2068 
2069 #if 0
2070 		xpt_print(periph->path, "Copying to user CCB %p from "
2071 			  "kernel address %p\n", *user_ccb, &io_req->ccb);
2072 #endif
2073 
2074 		error = copyout(&io_req->ccb, *user_ccb, sizeof(union ccb));
2075 		if (error != 0) {
2076 			xpt_print(periph->path, "Copy to user CCB %p from "
2077 				  "kernel address %p failed with error %d\n",
2078 				  *user_ccb, &io_req->ccb, error);
2079 		}
2080 
2081 		/*
2082 		 * Prefer the first error we got back, and make sure we
2083 		 * don't overwrite bad status with good.
2084 		 */
2085 		if (old_error != 0)
2086 			error = old_error;
2087 
2088 		cam_periph_lock(periph);
2089 
2090 		/*
2091 		 * At this point, if there was an error, we could potentially
2092 		 * re-queue the I/O and try again.  But why?  The error
2093 		 * would almost certainly happen again.  We might as well
2094 		 * not leak memory.
2095 		 */
2096 		uma_zfree(softc->pass_zone, io_req);
2097 		break;
2098 	}
2099 	default:
2100 		error = cam_periph_ioctl(periph, cmd, addr, passerror);
2101 		break;
2102 	}
2103 
2104 bailout:
2105 	cam_periph_unlock(periph);
2106 
2107 	return(error);
2108 }
2109 
2110 static int
2111 passpoll(struct cdev *dev, int poll_events, struct thread *td)
2112 {
2113 	struct cam_periph *periph;
2114 	struct pass_softc *softc;
2115 	int revents;
2116 
2117 	periph = (struct cam_periph *)dev->si_drv1;
2118 	softc = (struct pass_softc *)periph->softc;
2119 
2120 	revents = poll_events & (POLLOUT | POLLWRNORM);
2121 	if ((poll_events & (POLLIN | POLLRDNORM)) != 0) {
2122 		cam_periph_lock(periph);
2123 
2124 		if (!TAILQ_EMPTY(&softc->done_queue)) {
2125 			revents |= poll_events & (POLLIN | POLLRDNORM);
2126 		}
2127 		cam_periph_unlock(periph);
2128 		if (revents == 0)
2129 			selrecord(td, &softc->read_select);
2130 	}
2131 
2132 	return (revents);
2133 }
2134 
2135 static int
2136 passkqfilter(struct cdev *dev, struct knote *kn)
2137 {
2138 	struct cam_periph *periph;
2139 	struct pass_softc *softc;
2140 
2141 	periph = (struct cam_periph *)dev->si_drv1;
2142 	softc = (struct pass_softc *)periph->softc;
2143 
2144 	kn->kn_hook = (caddr_t)periph;
2145 	kn->kn_fop = &passread_filtops;
2146 	knlist_add(&softc->read_select.si_note, kn, 0);
2147 
2148 	return (0);
2149 }
2150 
2151 static void
2152 passreadfiltdetach(struct knote *kn)
2153 {
2154 	struct cam_periph *periph;
2155 	struct pass_softc *softc;
2156 
2157 	periph = (struct cam_periph *)kn->kn_hook;
2158 	softc = (struct pass_softc *)periph->softc;
2159 
2160 	knlist_remove(&softc->read_select.si_note, kn, 0);
2161 }
2162 
2163 static int
2164 passreadfilt(struct knote *kn, long hint)
2165 {
2166 	struct cam_periph *periph;
2167 	struct pass_softc *softc;
2168 	int retval;
2169 
2170 	periph = (struct cam_periph *)kn->kn_hook;
2171 	softc = (struct pass_softc *)periph->softc;
2172 
2173 	cam_periph_assert(periph, MA_OWNED);
2174 
2175 	if (TAILQ_EMPTY(&softc->done_queue))
2176 		retval = 0;
2177 	else
2178 		retval = 1;
2179 
2180 	return (retval);
2181 }
2182 
2183 /*
2184  * Generally, "ccb" should be the CCB supplied by the kernel.  "inccb"
2185  * should be the CCB that is copied in from the user.
2186  */
2187 static int
2188 passsendccb(struct cam_periph *periph, union ccb *ccb, union ccb *inccb)
2189 {
2190 	struct pass_softc *softc;
2191 	struct cam_periph_map_info mapinfo;
2192 	uint8_t *cmd;
2193 	xpt_opcode fc;
2194 	int error;
2195 
2196 	softc = (struct pass_softc *)periph->softc;
2197 
2198 	/*
2199 	 * There are some fields in the CCB header that need to be
2200 	 * preserved, the rest we get from the user.
2201 	 */
2202 	xpt_merge_ccb(ccb, inccb);
2203 
2204 	if (ccb->ccb_h.flags & CAM_CDB_POINTER) {
2205 		cmd = __builtin_alloca(ccb->csio.cdb_len);
2206 		error = copyin(ccb->csio.cdb_io.cdb_ptr, cmd, ccb->csio.cdb_len);
2207 		if (error)
2208 			return (error);
2209 		ccb->csio.cdb_io.cdb_ptr = cmd;
2210 	}
2211 
2212 	/*
2213 	 */
2214 	ccb->ccb_h.cbfcnp = passdone;
2215 
2216 	/*
2217 	 * Let cam_periph_mapmem do a sanity check on the data pointer format.
2218 	 * Even if no data transfer is needed, it's a cheap check and it
2219 	 * simplifies the code.
2220 	 */
2221 	fc = ccb->ccb_h.func_code;
2222 	if ((fc == XPT_SCSI_IO) || (fc == XPT_ATA_IO) || (fc == XPT_SMP_IO)
2223             || (fc == XPT_DEV_MATCH) || (fc == XPT_DEV_ADVINFO) || (fc == XPT_MMC_IO)
2224             || (fc == XPT_NVME_ADMIN) || (fc == XPT_NVME_IO)) {
2225 
2226 		bzero(&mapinfo, sizeof(mapinfo));
2227 
2228 		/*
2229 		 * cam_periph_mapmem calls into proc and vm functions that can
2230 		 * sleep as well as trigger I/O, so we can't hold the lock.
2231 		 * Dropping it here is reasonably safe.
2232 		 */
2233 		cam_periph_unlock(periph);
2234 		error = cam_periph_mapmem(ccb, &mapinfo, softc->maxio);
2235 		cam_periph_lock(periph);
2236 
2237 		/*
2238 		 * cam_periph_mapmem returned an error, we can't continue.
2239 		 * Return the error to the user.
2240 		 */
2241 		if (error)
2242 			return(error);
2243 	} else
2244 		/* Ensure that the unmap call later on is a no-op. */
2245 		mapinfo.num_bufs_used = 0;
2246 
2247 	/*
2248 	 * If the user wants us to perform any error recovery, then honor
2249 	 * that request.  Otherwise, it's up to the user to perform any
2250 	 * error recovery.
2251 	 */
2252 	cam_periph_runccb(ccb, (ccb->ccb_h.flags & CAM_PASS_ERR_RECOVER) ?
2253 	    passerror : NULL, /* cam_flags */ CAM_RETRY_SELTO,
2254 	    /* sense_flags */ SF_RETRY_UA | SF_NO_PRINT,
2255 	    softc->device_stats);
2256 
2257 	cam_periph_unmapmem(ccb, &mapinfo);
2258 
2259 	ccb->ccb_h.cbfcnp = NULL;
2260 	ccb->ccb_h.periph_priv = inccb->ccb_h.periph_priv;
2261 	bcopy(ccb, inccb, sizeof(union ccb));
2262 
2263 	return(0);
2264 }
2265 
2266 static int
2267 passerror(union ccb *ccb, u_int32_t cam_flags, u_int32_t sense_flags)
2268 {
2269 	struct cam_periph *periph;
2270 	struct pass_softc *softc;
2271 
2272 	periph = xpt_path_periph(ccb->ccb_h.path);
2273 	softc = (struct pass_softc *)periph->softc;
2274 
2275 	return(cam_periph_error(ccb, cam_flags, sense_flags));
2276 }
2277