xref: /freebsd/sys/cam/scsi/scsi_pass.c (revision c03c5b1c)
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/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, &softc->alias_dev,
461 					softc->dev, softc->alias_dev, physpath);
462 		mtx_lock(mtx);
463 	}
464 
465 out:
466 	/*
467 	 * Now that we've made our alias, we no longer have to have a
468 	 * reference to the device.
469 	 */
470 	if ((softc->flags & PASS_FLAG_INITIAL_PHYSPATH) == 0)
471 		softc->flags |= PASS_FLAG_INITIAL_PHYSPATH;
472 
473 	/*
474 	 * We always acquire a reference to the periph before queueing this
475 	 * task queue function, so it won't go away before we run.
476 	 */
477 	while (pending-- > 0)
478 		cam_periph_release_locked(periph);
479 	mtx_unlock(mtx);
480 
481 	free(physpath, M_DEVBUF);
482 }
483 
484 static void
485 passasync(void *callback_arg, u_int32_t code,
486 	  struct cam_path *path, void *arg)
487 {
488 	struct cam_periph *periph;
489 
490 	periph = (struct cam_periph *)callback_arg;
491 
492 	switch (code) {
493 	case AC_FOUND_DEVICE:
494 	{
495 		struct ccb_getdev *cgd;
496 		cam_status status;
497 
498 		cgd = (struct ccb_getdev *)arg;
499 		if (cgd == NULL)
500 			break;
501 
502 		/*
503 		 * Allocate a peripheral instance for
504 		 * this device and start the probe
505 		 * process.
506 		 */
507 		status = cam_periph_alloc(passregister, passoninvalidate,
508 					  passcleanup, passstart, "pass",
509 					  CAM_PERIPH_BIO, path,
510 					  passasync, AC_FOUND_DEVICE, cgd);
511 
512 		if (status != CAM_REQ_CMP
513 		 && status != CAM_REQ_INPROG) {
514 			const struct cam_status_entry *entry;
515 
516 			entry = cam_fetch_status_entry(status);
517 
518 			printf("passasync: Unable to attach new device "
519 			       "due to status %#x: %s\n", status, entry ?
520 			       entry->status_text : "Unknown");
521 		}
522 
523 		break;
524 	}
525 	case AC_ADVINFO_CHANGED:
526 	{
527 		uintptr_t buftype;
528 
529 		buftype = (uintptr_t)arg;
530 		if (buftype == CDAI_TYPE_PHYS_PATH) {
531 			struct pass_softc *softc;
532 
533 			softc = (struct pass_softc *)periph->softc;
534 			/*
535 			 * Acquire a reference to the periph before we
536 			 * start the taskqueue, so that we don't run into
537 			 * a situation where the periph goes away before
538 			 * the task queue has a chance to run.
539 			 */
540 			if (cam_periph_acquire(periph) != 0)
541 				break;
542 
543 			taskqueue_enqueue(taskqueue_thread,
544 					  &softc->add_physpath_task);
545 		}
546 		break;
547 	}
548 	default:
549 		cam_periph_async(periph, code, path, arg);
550 		break;
551 	}
552 }
553 
554 static cam_status
555 passregister(struct cam_periph *periph, void *arg)
556 {
557 	struct pass_softc *softc;
558 	struct ccb_getdev *cgd;
559 	struct ccb_pathinq cpi;
560 	struct make_dev_args args;
561 	int error, no_tags;
562 
563 	cgd = (struct ccb_getdev *)arg;
564 	if (cgd == NULL) {
565 		printf("%s: no getdev CCB, can't register device\n", __func__);
566 		return(CAM_REQ_CMP_ERR);
567 	}
568 
569 	softc = (struct pass_softc *)malloc(sizeof(*softc),
570 					    M_DEVBUF, M_NOWAIT);
571 
572 	if (softc == NULL) {
573 		printf("%s: Unable to probe new device. "
574 		       "Unable to allocate softc\n", __func__);
575 		return(CAM_REQ_CMP_ERR);
576 	}
577 
578 	bzero(softc, sizeof(*softc));
579 	softc->state = PASS_STATE_NORMAL;
580 	if (cgd->protocol == PROTO_SCSI || cgd->protocol == PROTO_ATAPI)
581 		softc->pd_type = SID_TYPE(&cgd->inq_data);
582 	else if (cgd->protocol == PROTO_SATAPM)
583 		softc->pd_type = T_ENCLOSURE;
584 	else
585 		softc->pd_type = T_DIRECT;
586 
587 	periph->softc = softc;
588 	softc->periph = periph;
589 	TAILQ_INIT(&softc->incoming_queue);
590 	TAILQ_INIT(&softc->active_queue);
591 	TAILQ_INIT(&softc->abandoned_queue);
592 	TAILQ_INIT(&softc->done_queue);
593 	snprintf(softc->zone_name, sizeof(softc->zone_name), "%s%d",
594 		 periph->periph_name, periph->unit_number);
595 	snprintf(softc->io_zone_name, sizeof(softc->io_zone_name), "%s%dIO",
596 		 periph->periph_name, periph->unit_number);
597 	softc->io_zone_size = maxphys;
598 	knlist_init_mtx(&softc->read_select.si_note, cam_periph_mtx(periph));
599 
600 	xpt_path_inq(&cpi, periph->path);
601 
602 	if (cpi.maxio == 0)
603 		softc->maxio = DFLTPHYS;	/* traditional default */
604 	else if (cpi.maxio > maxphys)
605 		softc->maxio = maxphys;		/* for safety */
606 	else
607 		softc->maxio = cpi.maxio;	/* real value */
608 
609 	if (cpi.hba_misc & PIM_UNMAPPED)
610 		softc->flags |= PASS_FLAG_UNMAPPED_CAPABLE;
611 
612 	/*
613 	 * We pass in 0 for a blocksize, since we don't
614 	 * know what the blocksize of this device is, if
615 	 * it even has a blocksize.
616 	 */
617 	cam_periph_unlock(periph);
618 	no_tags = (cgd->inq_data.flags & SID_CmdQue) == 0;
619 	softc->device_stats = devstat_new_entry("pass",
620 			  periph->unit_number, 0,
621 			  DEVSTAT_NO_BLOCKSIZE
622 			  | (no_tags ? DEVSTAT_NO_ORDERED_TAGS : 0),
623 			  softc->pd_type |
624 			  XPORT_DEVSTAT_TYPE(cpi.transport) |
625 			  DEVSTAT_TYPE_PASS,
626 			  DEVSTAT_PRIORITY_PASS);
627 
628 	/*
629 	 * Initialize the taskqueue handler for shutting down kqueue.
630 	 */
631 	TASK_INIT(&softc->shutdown_kqueue_task, /*priority*/ 0,
632 		  pass_shutdown_kqueue, periph);
633 
634 	/*
635 	 * Acquire a reference to the periph that we can release once we've
636 	 * cleaned up the kqueue.
637 	 */
638 	if (cam_periph_acquire(periph) != 0) {
639 		xpt_print(periph->path, "%s: lost periph during "
640 			  "registration!\n", __func__);
641 		cam_periph_lock(periph);
642 		return (CAM_REQ_CMP_ERR);
643 	}
644 
645 	/*
646 	 * Acquire a reference to the periph before we create the devfs
647 	 * instance for it.  We'll release this reference once the devfs
648 	 * instance has been freed.
649 	 */
650 	if (cam_periph_acquire(periph) != 0) {
651 		xpt_print(periph->path, "%s: lost periph during "
652 			  "registration!\n", __func__);
653 		cam_periph_lock(periph);
654 		return (CAM_REQ_CMP_ERR);
655 	}
656 
657 	/* Register the device */
658 	make_dev_args_init(&args);
659 	args.mda_devsw = &pass_cdevsw;
660 	args.mda_unit = periph->unit_number;
661 	args.mda_uid = UID_ROOT;
662 	args.mda_gid = GID_OPERATOR;
663 	args.mda_mode = 0600;
664 	args.mda_si_drv1 = periph;
665 	args.mda_flags = MAKEDEV_NOWAIT;
666 	error = make_dev_s(&args, &softc->dev, "%s%d", periph->periph_name,
667 	    periph->unit_number);
668 	if (error != 0) {
669 		cam_periph_lock(periph);
670 		cam_periph_release_locked(periph);
671 		return (CAM_REQ_CMP_ERR);
672 	}
673 
674 	/*
675 	 * Hold a reference to the periph before we create the physical
676 	 * path alias so it can't go away.
677 	 */
678 	if (cam_periph_acquire(periph) != 0) {
679 		xpt_print(periph->path, "%s: lost periph during "
680 			  "registration!\n", __func__);
681 		cam_periph_lock(periph);
682 		return (CAM_REQ_CMP_ERR);
683 	}
684 
685 	cam_periph_lock(periph);
686 
687 	TASK_INIT(&softc->add_physpath_task, /*priority*/0,
688 		  pass_add_physpath, periph);
689 
690 	/*
691 	 * See if physical path information is already available.
692 	 */
693 	taskqueue_enqueue(taskqueue_thread, &softc->add_physpath_task);
694 
695 	/*
696 	 * Add an async callback so that we get notified if
697 	 * this device goes away or its physical path
698 	 * (stored in the advanced info data of the EDT) has
699 	 * changed.
700 	 */
701 	xpt_register_async(AC_LOST_DEVICE | AC_ADVINFO_CHANGED,
702 			   passasync, periph, periph->path);
703 
704 	if (bootverbose)
705 		xpt_announce_periph(periph, NULL);
706 
707 	return(CAM_REQ_CMP);
708 }
709 
710 static int
711 passopen(struct cdev *dev, int flags, int fmt, struct thread *td)
712 {
713 	struct cam_periph *periph;
714 	struct pass_softc *softc;
715 	int error;
716 
717 	periph = (struct cam_periph *)dev->si_drv1;
718 	if (cam_periph_acquire(periph) != 0)
719 		return (ENXIO);
720 
721 	cam_periph_lock(periph);
722 
723 	softc = (struct pass_softc *)periph->softc;
724 
725 	if (softc->flags & PASS_FLAG_INVALID) {
726 		cam_periph_release_locked(periph);
727 		cam_periph_unlock(periph);
728 		return(ENXIO);
729 	}
730 
731 	/*
732 	 * Don't allow access when we're running at a high securelevel.
733 	 */
734 	error = securelevel_gt(td->td_ucred, 1);
735 	if (error) {
736 		cam_periph_release_locked(periph);
737 		cam_periph_unlock(periph);
738 		return(error);
739 	}
740 
741 	/*
742 	 * Only allow read-write access.
743 	 */
744 	if (((flags & FWRITE) == 0) || ((flags & FREAD) == 0)) {
745 		cam_periph_release_locked(periph);
746 		cam_periph_unlock(periph);
747 		return(EPERM);
748 	}
749 
750 	/*
751 	 * We don't allow nonblocking access.
752 	 */
753 	if ((flags & O_NONBLOCK) != 0) {
754 		xpt_print(periph->path, "can't do nonblocking access\n");
755 		cam_periph_release_locked(periph);
756 		cam_periph_unlock(periph);
757 		return(EINVAL);
758 	}
759 
760 	softc->open_count++;
761 
762 	cam_periph_unlock(periph);
763 
764 	return (error);
765 }
766 
767 static int
768 passclose(struct cdev *dev, int flag, int fmt, struct thread *td)
769 {
770 	struct 	cam_periph *periph;
771 	struct  pass_softc *softc;
772 	struct mtx *mtx;
773 
774 	periph = (struct cam_periph *)dev->si_drv1;
775 	mtx = cam_periph_mtx(periph);
776 	mtx_lock(mtx);
777 
778 	softc = periph->softc;
779 	softc->open_count--;
780 
781 	if (softc->open_count == 0) {
782 		struct pass_io_req *io_req, *io_req2;
783 
784 		TAILQ_FOREACH_SAFE(io_req, &softc->done_queue, links, io_req2) {
785 			TAILQ_REMOVE(&softc->done_queue, io_req, links);
786 			passiocleanup(softc, io_req);
787 			uma_zfree(softc->pass_zone, io_req);
788 		}
789 
790 		TAILQ_FOREACH_SAFE(io_req, &softc->incoming_queue, links,
791 				   io_req2) {
792 			TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
793 			passiocleanup(softc, io_req);
794 			uma_zfree(softc->pass_zone, io_req);
795 		}
796 
797 		/*
798 		 * If there are any active I/Os, we need to forcibly acquire a
799 		 * reference to the peripheral so that we don't go away
800 		 * before they complete.  We'll release the reference when
801 		 * the abandoned queue is empty.
802 		 */
803 		io_req = TAILQ_FIRST(&softc->active_queue);
804 		if ((io_req != NULL)
805 		 && (softc->flags & PASS_FLAG_ABANDONED_REF_SET) == 0) {
806 			cam_periph_doacquire(periph);
807 			softc->flags |= PASS_FLAG_ABANDONED_REF_SET;
808 		}
809 
810 		/*
811 		 * Since the I/O in the active queue is not under our
812 		 * control, just set a flag so that we can clean it up when
813 		 * it completes and put it on the abandoned queue.  This
814 		 * will prevent our sending spurious completions in the
815 		 * event that the device is opened again before these I/Os
816 		 * complete.
817 		 */
818 		TAILQ_FOREACH_SAFE(io_req, &softc->active_queue, links,
819 				   io_req2) {
820 			TAILQ_REMOVE(&softc->active_queue, io_req, links);
821 			io_req->flags |= PASS_IO_ABANDONED;
822 			TAILQ_INSERT_TAIL(&softc->abandoned_queue, io_req,
823 					  links);
824 		}
825 	}
826 
827 	cam_periph_release_locked(periph);
828 
829 	/*
830 	 * We reference the lock directly here, instead of using
831 	 * cam_periph_unlock().  The reason is that the call to
832 	 * cam_periph_release_locked() above could result in the periph
833 	 * getting freed.  If that is the case, dereferencing the periph
834 	 * with a cam_periph_unlock() call would cause a page fault.
835 	 *
836 	 * cam_periph_release() avoids this problem using the same method,
837 	 * but we're manually acquiring and dropping the lock here to
838 	 * protect the open count and avoid another lock acquisition and
839 	 * release.
840 	 */
841 	mtx_unlock(mtx);
842 
843 	return (0);
844 }
845 
846 static void
847 passstart(struct cam_periph *periph, union ccb *start_ccb)
848 {
849 	struct pass_softc *softc;
850 
851 	softc = (struct pass_softc *)periph->softc;
852 
853 	switch (softc->state) {
854 	case PASS_STATE_NORMAL: {
855 		struct pass_io_req *io_req;
856 
857 		/*
858 		 * Check for any queued I/O requests that require an
859 		 * allocated slot.
860 		 */
861 		io_req = TAILQ_FIRST(&softc->incoming_queue);
862 		if (io_req == NULL) {
863 			xpt_release_ccb(start_ccb);
864 			break;
865 		}
866 		TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
867 		TAILQ_INSERT_TAIL(&softc->active_queue, io_req, links);
868 		/*
869 		 * Merge the user's CCB into the allocated CCB.
870 		 */
871 		xpt_merge_ccb(start_ccb, &io_req->ccb);
872 		start_ccb->ccb_h.ccb_type = PASS_CCB_QUEUED_IO;
873 		start_ccb->ccb_h.ccb_ioreq = io_req;
874 		start_ccb->ccb_h.cbfcnp = passdone;
875 		io_req->alloced_ccb = start_ccb;
876 		binuptime(&io_req->start_time);
877 		devstat_start_transaction(softc->device_stats,
878 					  &io_req->start_time);
879 
880 		xpt_action(start_ccb);
881 
882 		/*
883 		 * If we have any more I/O waiting, schedule ourselves again.
884 		 */
885 		if (!TAILQ_EMPTY(&softc->incoming_queue))
886 			xpt_schedule(periph, CAM_PRIORITY_NORMAL);
887 		break;
888 	}
889 	default:
890 		break;
891 	}
892 }
893 
894 static void
895 passdone(struct cam_periph *periph, union ccb *done_ccb)
896 {
897 	struct pass_softc *softc;
898 	struct ccb_scsiio *csio;
899 
900 	softc = (struct pass_softc *)periph->softc;
901 
902 	cam_periph_assert(periph, MA_OWNED);
903 
904 	csio = &done_ccb->csio;
905 	switch (csio->ccb_h.ccb_type) {
906 	case PASS_CCB_QUEUED_IO: {
907 		struct pass_io_req *io_req;
908 
909 		io_req = done_ccb->ccb_h.ccb_ioreq;
910 #if 0
911 		xpt_print(periph->path, "%s: called for user CCB %p\n",
912 			  __func__, io_req->user_ccb_ptr);
913 #endif
914 		if (((done_ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_CMP)
915 		 && (done_ccb->ccb_h.flags & CAM_PASS_ERR_RECOVER)
916 		 && ((io_req->flags & PASS_IO_ABANDONED) == 0)) {
917 			int error;
918 
919 			error = passerror(done_ccb, CAM_RETRY_SELTO,
920 					  SF_RETRY_UA | SF_NO_PRINT);
921 
922 			if (error == ERESTART) {
923 				/*
924 				 * A retry was scheduled, so
925  				 * just return.
926 				 */
927 				return;
928 			}
929 		}
930 
931 		/*
932 		 * Copy the allocated CCB contents back to the malloced CCB
933 		 * so we can give status back to the user when he requests it.
934 		 */
935 		bcopy(done_ccb, &io_req->ccb, sizeof(*done_ccb));
936 
937 		/*
938 		 * Log data/transaction completion with devstat(9).
939 		 */
940 		switch (done_ccb->ccb_h.func_code) {
941 		case XPT_SCSI_IO:
942 			devstat_end_transaction(softc->device_stats,
943 			    done_ccb->csio.dxfer_len - done_ccb->csio.resid,
944 			    done_ccb->csio.tag_action & 0x3,
945 			    ((done_ccb->ccb_h.flags & CAM_DIR_MASK) ==
946 			    CAM_DIR_NONE) ? DEVSTAT_NO_DATA :
947 			    (done_ccb->ccb_h.flags & CAM_DIR_OUT) ?
948 			    DEVSTAT_WRITE : DEVSTAT_READ, NULL,
949 			    &io_req->start_time);
950 			break;
951 		case XPT_ATA_IO:
952 			devstat_end_transaction(softc->device_stats,
953 			    done_ccb->ataio.dxfer_len - done_ccb->ataio.resid,
954 			    0, /* Not used in ATA */
955 			    ((done_ccb->ccb_h.flags & CAM_DIR_MASK) ==
956 			    CAM_DIR_NONE) ? DEVSTAT_NO_DATA :
957 			    (done_ccb->ccb_h.flags & CAM_DIR_OUT) ?
958 			    DEVSTAT_WRITE : DEVSTAT_READ, NULL,
959 			    &io_req->start_time);
960 			break;
961 		case XPT_SMP_IO:
962 			/*
963 			 * XXX KDM this isn't quite right, but there isn't
964 			 * currently an easy way to represent a bidirectional
965 			 * transfer in devstat.  The only way to do it
966 			 * and have the byte counts come out right would
967 			 * mean that we would have to record two
968 			 * transactions, one for the request and one for the
969 			 * response.  For now, so that we report something,
970 			 * just treat the entire thing as a read.
971 			 */
972 			devstat_end_transaction(softc->device_stats,
973 			    done_ccb->smpio.smp_request_len +
974 			    done_ccb->smpio.smp_response_len,
975 			    DEVSTAT_TAG_SIMPLE, DEVSTAT_READ, NULL,
976 			    &io_req->start_time);
977 			break;
978 		default:
979 			devstat_end_transaction(softc->device_stats, 0,
980 			    DEVSTAT_TAG_NONE, DEVSTAT_NO_DATA, NULL,
981 			    &io_req->start_time);
982 			break;
983 		}
984 
985 		/*
986 		 * In the normal case, take the completed I/O off of the
987 		 * active queue and put it on the done queue.  Notitfy the
988 		 * user that we have a completed I/O.
989 		 */
990 		if ((io_req->flags & PASS_IO_ABANDONED) == 0) {
991 			TAILQ_REMOVE(&softc->active_queue, io_req, links);
992 			TAILQ_INSERT_TAIL(&softc->done_queue, io_req, links);
993 			selwakeuppri(&softc->read_select, PRIBIO);
994 			KNOTE_LOCKED(&softc->read_select.si_note, 0);
995 		} else {
996 			/*
997 			 * In the case of an abandoned I/O (final close
998 			 * without fetching the I/O), take it off of the
999 			 * abandoned queue and free it.
1000 			 */
1001 			TAILQ_REMOVE(&softc->abandoned_queue, io_req, links);
1002 			passiocleanup(softc, io_req);
1003 			uma_zfree(softc->pass_zone, io_req);
1004 
1005 			/*
1006 			 * Release the done_ccb here, since we may wind up
1007 			 * freeing the peripheral when we decrement the
1008 			 * reference count below.
1009 			 */
1010 			xpt_release_ccb(done_ccb);
1011 
1012 			/*
1013 			 * If the abandoned queue is empty, we can release
1014 			 * our reference to the periph since we won't have
1015 			 * any more completions coming.
1016 			 */
1017 			if ((TAILQ_EMPTY(&softc->abandoned_queue))
1018 			 && (softc->flags & PASS_FLAG_ABANDONED_REF_SET)) {
1019 				softc->flags &= ~PASS_FLAG_ABANDONED_REF_SET;
1020 				cam_periph_release_locked(periph);
1021 			}
1022 
1023 			/*
1024 			 * We have already released the CCB, so we can
1025 			 * return.
1026 			 */
1027 			return;
1028 		}
1029 		break;
1030 	}
1031 	}
1032 	xpt_release_ccb(done_ccb);
1033 }
1034 
1035 static int
1036 passcreatezone(struct cam_periph *periph)
1037 {
1038 	struct pass_softc *softc;
1039 	int error;
1040 
1041 	error = 0;
1042 	softc = (struct pass_softc *)periph->softc;
1043 
1044 	cam_periph_assert(periph, MA_OWNED);
1045 	KASSERT(((softc->flags & PASS_FLAG_ZONE_VALID) == 0),
1046 		("%s called when the pass(4) zone is valid!\n", __func__));
1047 	KASSERT((softc->pass_zone == NULL),
1048 		("%s called when the pass(4) zone is allocated!\n", __func__));
1049 
1050 	if ((softc->flags & PASS_FLAG_ZONE_INPROG) == 0) {
1051 		/*
1052 		 * We're the first context through, so we need to create
1053 		 * the pass(4) UMA zone for I/O requests.
1054 		 */
1055 		softc->flags |= PASS_FLAG_ZONE_INPROG;
1056 
1057 		/*
1058 		 * uma_zcreate() does a blocking (M_WAITOK) allocation,
1059 		 * so we cannot hold a mutex while we call it.
1060 		 */
1061 		cam_periph_unlock(periph);
1062 
1063 		softc->pass_zone = uma_zcreate(softc->zone_name,
1064 		    sizeof(struct pass_io_req), NULL, NULL, NULL, NULL,
1065 		    /*align*/ 0, /*flags*/ 0);
1066 
1067 		softc->pass_io_zone = uma_zcreate(softc->io_zone_name,
1068 		    softc->io_zone_size, NULL, NULL, NULL, NULL,
1069 		    /*align*/ 0, /*flags*/ 0);
1070 
1071 		cam_periph_lock(periph);
1072 
1073 		if ((softc->pass_zone == NULL)
1074 		 || (softc->pass_io_zone == NULL)) {
1075 			if (softc->pass_zone == NULL)
1076 				xpt_print(periph->path, "unable to allocate "
1077 				    "IO Req UMA zone\n");
1078 			else
1079 				xpt_print(periph->path, "unable to allocate "
1080 				    "IO UMA zone\n");
1081 			softc->flags &= ~PASS_FLAG_ZONE_INPROG;
1082 			goto bailout;
1083 		}
1084 
1085 		/*
1086 		 * Set the flags appropriately and notify any other waiters.
1087 		 */
1088 		softc->flags &= ~PASS_FLAG_ZONE_INPROG;
1089 		softc->flags |= PASS_FLAG_ZONE_VALID;
1090 		wakeup(&softc->pass_zone);
1091 	} else {
1092 		/*
1093 		 * In this case, the UMA zone has not yet been created, but
1094 		 * another context is in the process of creating it.  We
1095 		 * need to sleep until the creation is either done or has
1096 		 * failed.
1097 		 */
1098 		while ((softc->flags & PASS_FLAG_ZONE_INPROG)
1099 		    && ((softc->flags & PASS_FLAG_ZONE_VALID) == 0)) {
1100 			error = msleep(&softc->pass_zone,
1101 				       cam_periph_mtx(periph), PRIBIO,
1102 				       "paszon", 0);
1103 			if (error != 0)
1104 				goto bailout;
1105 		}
1106 		/*
1107 		 * If the zone creation failed, no luck for the user.
1108 		 */
1109 		if ((softc->flags & PASS_FLAG_ZONE_VALID) == 0){
1110 			error = ENOMEM;
1111 			goto bailout;
1112 		}
1113 	}
1114 bailout:
1115 	return (error);
1116 }
1117 
1118 static void
1119 passiocleanup(struct pass_softc *softc, struct pass_io_req *io_req)
1120 {
1121 	union ccb *ccb;
1122 	u_int8_t **data_ptrs[CAM_PERIPH_MAXMAPS];
1123 	int i, numbufs;
1124 
1125 	ccb = &io_req->ccb;
1126 
1127 	switch (ccb->ccb_h.func_code) {
1128 	case XPT_DEV_MATCH:
1129 		numbufs = min(io_req->num_bufs, 2);
1130 
1131 		if (numbufs == 1) {
1132 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.matches;
1133 		} else {
1134 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.patterns;
1135 			data_ptrs[1] = (u_int8_t **)&ccb->cdm.matches;
1136 		}
1137 		break;
1138 	case XPT_SCSI_IO:
1139 	case XPT_CONT_TARGET_IO:
1140 		data_ptrs[0] = &ccb->csio.data_ptr;
1141 		numbufs = min(io_req->num_bufs, 1);
1142 		break;
1143 	case XPT_ATA_IO:
1144 		data_ptrs[0] = &ccb->ataio.data_ptr;
1145 		numbufs = min(io_req->num_bufs, 1);
1146 		break;
1147 	case XPT_SMP_IO:
1148 		numbufs = min(io_req->num_bufs, 2);
1149 		data_ptrs[0] = &ccb->smpio.smp_request;
1150 		data_ptrs[1] = &ccb->smpio.smp_response;
1151 		break;
1152 	case XPT_DEV_ADVINFO:
1153 		numbufs = min(io_req->num_bufs, 1);
1154 		data_ptrs[0] = (uint8_t **)&ccb->cdai.buf;
1155 		break;
1156 	case XPT_NVME_IO:
1157 	case XPT_NVME_ADMIN:
1158 		data_ptrs[0] = &ccb->nvmeio.data_ptr;
1159 		numbufs = min(io_req->num_bufs, 1);
1160 		break;
1161 	default:
1162 		/* allow ourselves to be swapped once again */
1163 		return;
1164 		break; /* NOTREACHED */
1165 	}
1166 
1167 	if (io_req->flags & PASS_IO_USER_SEG_MALLOC) {
1168 		free(io_req->user_segptr, M_SCSIPASS);
1169 		io_req->user_segptr = NULL;
1170 	}
1171 
1172 	/*
1173 	 * We only want to free memory we malloced.
1174 	 */
1175 	if (io_req->data_flags == CAM_DATA_VADDR) {
1176 		for (i = 0; i < io_req->num_bufs; i++) {
1177 			if (io_req->kern_bufs[i] == NULL)
1178 				continue;
1179 
1180 			free(io_req->kern_bufs[i], M_SCSIPASS);
1181 			io_req->kern_bufs[i] = NULL;
1182 		}
1183 	} else if (io_req->data_flags == CAM_DATA_SG) {
1184 		for (i = 0; i < io_req->num_kern_segs; i++) {
1185 			if ((uint8_t *)(uintptr_t)
1186 			    io_req->kern_segptr[i].ds_addr == NULL)
1187 				continue;
1188 
1189 			uma_zfree(softc->pass_io_zone, (uint8_t *)(uintptr_t)
1190 			    io_req->kern_segptr[i].ds_addr);
1191 			io_req->kern_segptr[i].ds_addr = 0;
1192 		}
1193 	}
1194 
1195 	if (io_req->flags & PASS_IO_KERN_SEG_MALLOC) {
1196 		free(io_req->kern_segptr, M_SCSIPASS);
1197 		io_req->kern_segptr = NULL;
1198 	}
1199 
1200 	if (io_req->data_flags != CAM_DATA_PADDR) {
1201 		for (i = 0; i < numbufs; i++) {
1202 			/*
1203 			 * Restore the user's buffer pointers to their
1204 			 * previous values.
1205 			 */
1206 			if (io_req->user_bufs[i] != NULL)
1207 				*data_ptrs[i] = io_req->user_bufs[i];
1208 		}
1209 	}
1210 
1211 }
1212 
1213 static int
1214 passcopysglist(struct cam_periph *periph, struct pass_io_req *io_req,
1215 	       ccb_flags direction)
1216 {
1217 	bus_size_t kern_watermark, user_watermark, len_to_copy;
1218 	bus_dma_segment_t *user_sglist, *kern_sglist;
1219 	int i, j, error;
1220 
1221 	error = 0;
1222 	kern_watermark = 0;
1223 	user_watermark = 0;
1224 	len_to_copy = 0;
1225 	user_sglist = io_req->user_segptr;
1226 	kern_sglist = io_req->kern_segptr;
1227 
1228 	for (i = 0, j = 0; i < io_req->num_user_segs &&
1229 	     j < io_req->num_kern_segs;) {
1230 		uint8_t *user_ptr, *kern_ptr;
1231 
1232 		len_to_copy = min(user_sglist[i].ds_len -user_watermark,
1233 		    kern_sglist[j].ds_len - kern_watermark);
1234 
1235 		user_ptr = (uint8_t *)(uintptr_t)user_sglist[i].ds_addr;
1236 		user_ptr = user_ptr + user_watermark;
1237 		kern_ptr = (uint8_t *)(uintptr_t)kern_sglist[j].ds_addr;
1238 		kern_ptr = kern_ptr + kern_watermark;
1239 
1240 		user_watermark += len_to_copy;
1241 		kern_watermark += len_to_copy;
1242 
1243 		if (direction == CAM_DIR_IN) {
1244 			error = copyout(kern_ptr, user_ptr, len_to_copy);
1245 			if (error != 0) {
1246 				xpt_print(periph->path, "%s: copyout of %u "
1247 					  "bytes from %p to %p failed with "
1248 					  "error %d\n", __func__, len_to_copy,
1249 					  kern_ptr, user_ptr, error);
1250 				goto bailout;
1251 			}
1252 		} else {
1253 			error = copyin(user_ptr, kern_ptr, len_to_copy);
1254 			if (error != 0) {
1255 				xpt_print(periph->path, "%s: copyin of %u "
1256 					  "bytes from %p to %p failed with "
1257 					  "error %d\n", __func__, len_to_copy,
1258 					  user_ptr, kern_ptr, error);
1259 				goto bailout;
1260 			}
1261 		}
1262 
1263 		if (user_sglist[i].ds_len == user_watermark) {
1264 			i++;
1265 			user_watermark = 0;
1266 		}
1267 
1268 		if (kern_sglist[j].ds_len == kern_watermark) {
1269 			j++;
1270 			kern_watermark = 0;
1271 		}
1272 	}
1273 
1274 bailout:
1275 
1276 	return (error);
1277 }
1278 
1279 static int
1280 passmemsetup(struct cam_periph *periph, struct pass_io_req *io_req)
1281 {
1282 	union ccb *ccb;
1283 	struct pass_softc *softc;
1284 	int numbufs, i;
1285 	uint8_t **data_ptrs[CAM_PERIPH_MAXMAPS];
1286 	uint32_t lengths[CAM_PERIPH_MAXMAPS];
1287 	uint32_t dirs[CAM_PERIPH_MAXMAPS];
1288 	uint32_t num_segs;
1289 	uint16_t *seg_cnt_ptr;
1290 	size_t maxmap;
1291 	int error;
1292 
1293 	cam_periph_assert(periph, MA_NOTOWNED);
1294 
1295 	softc = periph->softc;
1296 
1297 	error = 0;
1298 	ccb = &io_req->ccb;
1299 	maxmap = 0;
1300 	num_segs = 0;
1301 	seg_cnt_ptr = NULL;
1302 
1303 	switch(ccb->ccb_h.func_code) {
1304 	case XPT_DEV_MATCH:
1305 		if (ccb->cdm.match_buf_len == 0) {
1306 			printf("%s: invalid match buffer length 0\n", __func__);
1307 			return(EINVAL);
1308 		}
1309 		if (ccb->cdm.pattern_buf_len > 0) {
1310 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.patterns;
1311 			lengths[0] = ccb->cdm.pattern_buf_len;
1312 			dirs[0] = CAM_DIR_OUT;
1313 			data_ptrs[1] = (u_int8_t **)&ccb->cdm.matches;
1314 			lengths[1] = ccb->cdm.match_buf_len;
1315 			dirs[1] = CAM_DIR_IN;
1316 			numbufs = 2;
1317 		} else {
1318 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.matches;
1319 			lengths[0] = ccb->cdm.match_buf_len;
1320 			dirs[0] = CAM_DIR_IN;
1321 			numbufs = 1;
1322 		}
1323 		io_req->data_flags = CAM_DATA_VADDR;
1324 		break;
1325 	case XPT_SCSI_IO:
1326 	case XPT_CONT_TARGET_IO:
1327 		if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
1328 			return(0);
1329 
1330 		/*
1331 		 * The user shouldn't be able to supply a bio.
1332 		 */
1333 		if ((ccb->ccb_h.flags & CAM_DATA_MASK) == CAM_DATA_BIO)
1334 			return (EINVAL);
1335 
1336 		io_req->data_flags = ccb->ccb_h.flags & CAM_DATA_MASK;
1337 
1338 		data_ptrs[0] = &ccb->csio.data_ptr;
1339 		lengths[0] = ccb->csio.dxfer_len;
1340 		dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1341 		num_segs = ccb->csio.sglist_cnt;
1342 		seg_cnt_ptr = &ccb->csio.sglist_cnt;
1343 		numbufs = 1;
1344 		maxmap = softc->maxio;
1345 		break;
1346 	case XPT_ATA_IO:
1347 		if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
1348 			return(0);
1349 
1350 		/*
1351 		 * We only support a single virtual address for ATA I/O.
1352 		 */
1353 		if ((ccb->ccb_h.flags & CAM_DATA_MASK) != CAM_DATA_VADDR)
1354 			return (EINVAL);
1355 
1356 		io_req->data_flags = CAM_DATA_VADDR;
1357 
1358 		data_ptrs[0] = &ccb->ataio.data_ptr;
1359 		lengths[0] = ccb->ataio.dxfer_len;
1360 		dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1361 		numbufs = 1;
1362 		maxmap = softc->maxio;
1363 		break;
1364 	case XPT_SMP_IO:
1365 		io_req->data_flags = CAM_DATA_VADDR;
1366 
1367 		data_ptrs[0] = &ccb->smpio.smp_request;
1368 		lengths[0] = ccb->smpio.smp_request_len;
1369 		dirs[0] = CAM_DIR_OUT;
1370 		data_ptrs[1] = &ccb->smpio.smp_response;
1371 		lengths[1] = ccb->smpio.smp_response_len;
1372 		dirs[1] = CAM_DIR_IN;
1373 		numbufs = 2;
1374 		maxmap = softc->maxio;
1375 		break;
1376 	case XPT_DEV_ADVINFO:
1377 		if (ccb->cdai.bufsiz == 0)
1378 			return (0);
1379 
1380 		io_req->data_flags = CAM_DATA_VADDR;
1381 
1382 		data_ptrs[0] = (uint8_t **)&ccb->cdai.buf;
1383 		lengths[0] = ccb->cdai.bufsiz;
1384 		dirs[0] = CAM_DIR_IN;
1385 		numbufs = 1;
1386 		break;
1387 	case XPT_NVME_ADMIN:
1388 	case XPT_NVME_IO:
1389 		if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
1390 			return (0);
1391 
1392 		io_req->data_flags = ccb->ccb_h.flags & CAM_DATA_MASK;
1393 
1394 		data_ptrs[0] = &ccb->nvmeio.data_ptr;
1395 		lengths[0] = ccb->nvmeio.dxfer_len;
1396 		dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1397 		num_segs = ccb->nvmeio.sglist_cnt;
1398 		seg_cnt_ptr = &ccb->nvmeio.sglist_cnt;
1399 		numbufs = 1;
1400 		maxmap = softc->maxio;
1401 		break;
1402 	default:
1403 		return(EINVAL);
1404 		break; /* NOTREACHED */
1405 	}
1406 
1407 	io_req->num_bufs = numbufs;
1408 
1409 	/*
1410 	 * If there is a maximum, check to make sure that the user's
1411 	 * request fits within the limit.  In general, we should only have
1412 	 * a maximum length for requests that go to hardware.  Otherwise it
1413 	 * is whatever we're able to malloc.
1414 	 */
1415 	for (i = 0; i < numbufs; i++) {
1416 		io_req->user_bufs[i] = *data_ptrs[i];
1417 		io_req->dirs[i] = dirs[i];
1418 		io_req->lengths[i] = lengths[i];
1419 
1420 		if (maxmap == 0)
1421 			continue;
1422 
1423 		if (lengths[i] <= maxmap)
1424 			continue;
1425 
1426 		xpt_print(periph->path, "%s: data length %u > max allowed %u "
1427 			  "bytes\n", __func__, lengths[i], maxmap);
1428 		error = EINVAL;
1429 		goto bailout;
1430 	}
1431 
1432 	switch (io_req->data_flags) {
1433 	case CAM_DATA_VADDR:
1434 		/* Map or copy the buffer into kernel address space */
1435 		for (i = 0; i < numbufs; i++) {
1436 			uint8_t *tmp_buf;
1437 
1438 			/*
1439 			 * If for some reason no length is specified, we
1440 			 * don't need to allocate anything.
1441 			 */
1442 			if (io_req->lengths[i] == 0)
1443 				continue;
1444 
1445 			tmp_buf = malloc(lengths[i], M_SCSIPASS,
1446 					 M_WAITOK | M_ZERO);
1447 			io_req->kern_bufs[i] = tmp_buf;
1448 			*data_ptrs[i] = tmp_buf;
1449 
1450 #if 0
1451 			xpt_print(periph->path, "%s: malloced %p len %u, user "
1452 				  "buffer %p, operation: %s\n", __func__,
1453 				  tmp_buf, lengths[i], io_req->user_bufs[i],
1454 				  (dirs[i] == CAM_DIR_IN) ? "read" : "write");
1455 #endif
1456 			/*
1457 			 * We only need to copy in if the user is writing.
1458 			 */
1459 			if (dirs[i] != CAM_DIR_OUT)
1460 				continue;
1461 
1462 			error = copyin(io_req->user_bufs[i],
1463 				       io_req->kern_bufs[i], lengths[i]);
1464 			if (error != 0) {
1465 				xpt_print(periph->path, "%s: copy of user "
1466 					  "buffer from %p to %p failed with "
1467 					  "error %d\n", __func__,
1468 					  io_req->user_bufs[i],
1469 					  io_req->kern_bufs[i], error);
1470 				goto bailout;
1471 			}
1472 		}
1473 		break;
1474 	case CAM_DATA_PADDR:
1475 		/* Pass down the pointer as-is */
1476 		break;
1477 	case CAM_DATA_SG: {
1478 		size_t sg_length, size_to_go, alloc_size;
1479 		uint32_t num_segs_needed;
1480 
1481 		/*
1482 		 * Copy the user S/G list in, and then copy in the
1483 		 * individual segments.
1484 		 */
1485 		/*
1486 		 * We shouldn't see this, but check just in case.
1487 		 */
1488 		if (numbufs != 1) {
1489 			xpt_print(periph->path, "%s: cannot currently handle "
1490 				  "more than one S/G list per CCB\n", __func__);
1491 			error = EINVAL;
1492 			goto bailout;
1493 		}
1494 
1495 		/*
1496 		 * We have to have at least one segment.
1497 		 */
1498 		if (num_segs == 0) {
1499 			xpt_print(periph->path, "%s: CAM_DATA_SG flag set, "
1500 				  "but sglist_cnt=0!\n", __func__);
1501 			error = EINVAL;
1502 			goto bailout;
1503 		}
1504 
1505 		/*
1506 		 * Make sure the user specified the total length and didn't
1507 		 * just leave it to us to decode the S/G list.
1508 		 */
1509 		if (lengths[0] == 0) {
1510 			xpt_print(periph->path, "%s: no dxfer_len specified, "
1511 				  "but CAM_DATA_SG flag is set!\n", __func__);
1512 			error = EINVAL;
1513 			goto bailout;
1514 		}
1515 
1516 		/*
1517 		 * We allocate buffers in io_zone_size increments for an
1518 		 * S/G list.  This will generally be maxphys.
1519 		 */
1520 		if (lengths[0] <= softc->io_zone_size)
1521 			num_segs_needed = 1;
1522 		else {
1523 			num_segs_needed = lengths[0] / softc->io_zone_size;
1524 			if ((lengths[0] % softc->io_zone_size) != 0)
1525 				num_segs_needed++;
1526 		}
1527 
1528 		/* Figure out the size of the S/G list */
1529 		sg_length = num_segs * sizeof(bus_dma_segment_t);
1530 		io_req->num_user_segs = num_segs;
1531 		io_req->num_kern_segs = num_segs_needed;
1532 
1533 		/* Save the user's S/G list pointer for later restoration */
1534 		io_req->user_bufs[0] = *data_ptrs[0];
1535 
1536 		/*
1537 		 * If we have enough segments allocated by default to handle
1538 		 * the length of the user's S/G list,
1539 		 */
1540 		if (num_segs > PASS_MAX_SEGS) {
1541 			io_req->user_segptr = malloc(sizeof(bus_dma_segment_t) *
1542 			    num_segs, M_SCSIPASS, M_WAITOK | M_ZERO);
1543 			io_req->flags |= PASS_IO_USER_SEG_MALLOC;
1544 		} else
1545 			io_req->user_segptr = io_req->user_segs;
1546 
1547 		error = copyin(*data_ptrs[0], io_req->user_segptr, sg_length);
1548 		if (error != 0) {
1549 			xpt_print(periph->path, "%s: copy of user S/G list "
1550 				  "from %p to %p failed with error %d\n",
1551 				  __func__, *data_ptrs[0], io_req->user_segptr,
1552 				  error);
1553 			goto bailout;
1554 		}
1555 
1556 		if (num_segs_needed > PASS_MAX_SEGS) {
1557 			io_req->kern_segptr = malloc(sizeof(bus_dma_segment_t) *
1558 			    num_segs_needed, M_SCSIPASS, M_WAITOK | M_ZERO);
1559 			io_req->flags |= PASS_IO_KERN_SEG_MALLOC;
1560 		} else {
1561 			io_req->kern_segptr = io_req->kern_segs;
1562 		}
1563 
1564 		/*
1565 		 * Allocate the kernel S/G list.
1566 		 */
1567 		for (size_to_go = lengths[0], i = 0;
1568 		     size_to_go > 0 && i < num_segs_needed;
1569 		     i++, size_to_go -= alloc_size) {
1570 			uint8_t *kern_ptr;
1571 
1572 			alloc_size = min(size_to_go, softc->io_zone_size);
1573 			kern_ptr = uma_zalloc(softc->pass_io_zone, M_WAITOK);
1574 			io_req->kern_segptr[i].ds_addr =
1575 			    (bus_addr_t)(uintptr_t)kern_ptr;
1576 			io_req->kern_segptr[i].ds_len = alloc_size;
1577 		}
1578 		if (size_to_go > 0) {
1579 			printf("%s: size_to_go = %zu, software error!\n",
1580 			       __func__, size_to_go);
1581 			error = EINVAL;
1582 			goto bailout;
1583 		}
1584 
1585 		*data_ptrs[0] = (uint8_t *)io_req->kern_segptr;
1586 		*seg_cnt_ptr = io_req->num_kern_segs;
1587 
1588 		/*
1589 		 * We only need to copy data here if the user is writing.
1590 		 */
1591 		if (dirs[0] == CAM_DIR_OUT)
1592 			error = passcopysglist(periph, io_req, dirs[0]);
1593 		break;
1594 	}
1595 	case CAM_DATA_SG_PADDR: {
1596 		size_t sg_length;
1597 
1598 		/*
1599 		 * We shouldn't see this, but check just in case.
1600 		 */
1601 		if (numbufs != 1) {
1602 			printf("%s: cannot currently handle more than one "
1603 			       "S/G list per CCB\n", __func__);
1604 			error = EINVAL;
1605 			goto bailout;
1606 		}
1607 
1608 		/*
1609 		 * We have to have at least one segment.
1610 		 */
1611 		if (num_segs == 0) {
1612 			xpt_print(periph->path, "%s: CAM_DATA_SG_PADDR flag "
1613 				  "set, but sglist_cnt=0!\n", __func__);
1614 			error = EINVAL;
1615 			goto bailout;
1616 		}
1617 
1618 		/*
1619 		 * Make sure the user specified the total length and didn't
1620 		 * just leave it to us to decode the S/G list.
1621 		 */
1622 		if (lengths[0] == 0) {
1623 			xpt_print(periph->path, "%s: no dxfer_len specified, "
1624 				  "but CAM_DATA_SG flag is set!\n", __func__);
1625 			error = EINVAL;
1626 			goto bailout;
1627 		}
1628 
1629 		/* Figure out the size of the S/G list */
1630 		sg_length = num_segs * sizeof(bus_dma_segment_t);
1631 		io_req->num_user_segs = num_segs;
1632 		io_req->num_kern_segs = io_req->num_user_segs;
1633 
1634 		/* Save the user's S/G list pointer for later restoration */
1635 		io_req->user_bufs[0] = *data_ptrs[0];
1636 
1637 		if (num_segs > PASS_MAX_SEGS) {
1638 			io_req->user_segptr = malloc(sizeof(bus_dma_segment_t) *
1639 			    num_segs, M_SCSIPASS, M_WAITOK | M_ZERO);
1640 			io_req->flags |= PASS_IO_USER_SEG_MALLOC;
1641 		} else
1642 			io_req->user_segptr = io_req->user_segs;
1643 
1644 		io_req->kern_segptr = io_req->user_segptr;
1645 
1646 		error = copyin(*data_ptrs[0], io_req->user_segptr, sg_length);
1647 		if (error != 0) {
1648 			xpt_print(periph->path, "%s: copy of user S/G list "
1649 				  "from %p to %p failed with error %d\n",
1650 				  __func__, *data_ptrs[0], io_req->user_segptr,
1651 				  error);
1652 			goto bailout;
1653 		}
1654 		break;
1655 	}
1656 	default:
1657 	case CAM_DATA_BIO:
1658 		/*
1659 		 * A user shouldn't be attaching a bio to the CCB.  It
1660 		 * isn't a user-accessible structure.
1661 		 */
1662 		error = EINVAL;
1663 		break;
1664 	}
1665 
1666 bailout:
1667 	if (error != 0)
1668 		passiocleanup(softc, io_req);
1669 
1670 	return (error);
1671 }
1672 
1673 static int
1674 passmemdone(struct cam_periph *periph, struct pass_io_req *io_req)
1675 {
1676 	struct pass_softc *softc;
1677 	int error;
1678 	int i;
1679 
1680 	error = 0;
1681 	softc = (struct pass_softc *)periph->softc;
1682 
1683 	switch (io_req->data_flags) {
1684 	case CAM_DATA_VADDR:
1685 		/*
1686 		 * Copy back to the user buffer if this was a read.
1687 		 */
1688 		for (i = 0; i < io_req->num_bufs; i++) {
1689 			if (io_req->dirs[i] != CAM_DIR_IN)
1690 				continue;
1691 
1692 			error = copyout(io_req->kern_bufs[i],
1693 			    io_req->user_bufs[i], io_req->lengths[i]);
1694 			if (error != 0) {
1695 				xpt_print(periph->path, "Unable to copy %u "
1696 					  "bytes from %p to user address %p\n",
1697 					  io_req->lengths[i],
1698 					  io_req->kern_bufs[i],
1699 					  io_req->user_bufs[i]);
1700 				goto bailout;
1701 			}
1702 		}
1703 		break;
1704 	case CAM_DATA_PADDR:
1705 		/* Do nothing.  The pointer is a physical address already */
1706 		break;
1707 	case CAM_DATA_SG:
1708 		/*
1709 		 * Copy back to the user buffer if this was a read.
1710 		 * Restore the user's S/G list buffer pointer.
1711 		 */
1712 		if (io_req->dirs[0] == CAM_DIR_IN)
1713 			error = passcopysglist(periph, io_req, io_req->dirs[0]);
1714 		break;
1715 	case CAM_DATA_SG_PADDR:
1716 		/*
1717 		 * Restore the user's S/G list buffer pointer.  No need to
1718 		 * copy.
1719 		 */
1720 		break;
1721 	default:
1722 	case CAM_DATA_BIO:
1723 		error = EINVAL;
1724 		break;
1725 	}
1726 
1727 bailout:
1728 	/*
1729 	 * Reset the user's pointers to their original values and free
1730 	 * allocated memory.
1731 	 */
1732 	passiocleanup(softc, io_req);
1733 
1734 	return (error);
1735 }
1736 
1737 static int
1738 passioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flag, struct thread *td)
1739 {
1740 	int error;
1741 
1742 	if ((error = passdoioctl(dev, cmd, addr, flag, td)) == ENOTTY) {
1743 		error = cam_compat_ioctl(dev, cmd, addr, flag, td, passdoioctl);
1744 	}
1745 	return (error);
1746 }
1747 
1748 static int
1749 passdoioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flag, struct thread *td)
1750 {
1751 	struct	cam_periph *periph;
1752 	struct	pass_softc *softc;
1753 	int	error;
1754 	uint32_t priority;
1755 
1756 	periph = (struct cam_periph *)dev->si_drv1;
1757 	cam_periph_lock(periph);
1758 	softc = (struct pass_softc *)periph->softc;
1759 
1760 	error = 0;
1761 
1762 	switch (cmd) {
1763 	case CAMIOCOMMAND:
1764 	{
1765 		union ccb *inccb;
1766 		union ccb *ccb;
1767 		int ccb_malloced;
1768 
1769 		inccb = (union ccb *)addr;
1770 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
1771 		if (inccb->ccb_h.func_code == XPT_SCSI_IO)
1772 			inccb->csio.bio = NULL;
1773 #endif
1774 
1775 		if (inccb->ccb_h.flags & CAM_UNLOCKED) {
1776 			error = EINVAL;
1777 			break;
1778 		}
1779 
1780 		/*
1781 		 * Some CCB types, like scan bus and scan lun can only go
1782 		 * through the transport layer device.
1783 		 */
1784 		if (inccb->ccb_h.func_code & XPT_FC_XPT_ONLY) {
1785 			xpt_print(periph->path, "CCB function code %#x is "
1786 			    "restricted to the XPT device\n",
1787 			    inccb->ccb_h.func_code);
1788 			error = ENODEV;
1789 			break;
1790 		}
1791 
1792 		/* Compatibility for RL/priority-unaware code. */
1793 		priority = inccb->ccb_h.pinfo.priority;
1794 		if (priority <= CAM_PRIORITY_OOB)
1795 		    priority += CAM_PRIORITY_OOB + 1;
1796 
1797 		/*
1798 		 * Non-immediate CCBs need a CCB from the per-device pool
1799 		 * of CCBs, which is scheduled by the transport layer.
1800 		 * Immediate CCBs and user-supplied CCBs should just be
1801 		 * malloced.
1802 		 */
1803 		if ((inccb->ccb_h.func_code & XPT_FC_QUEUED)
1804 		 && ((inccb->ccb_h.func_code & XPT_FC_USER_CCB) == 0)) {
1805 			ccb = cam_periph_getccb(periph, priority);
1806 			ccb_malloced = 0;
1807 		} else {
1808 			ccb = xpt_alloc_ccb_nowait();
1809 
1810 			if (ccb != NULL)
1811 				xpt_setup_ccb(&ccb->ccb_h, periph->path,
1812 					      priority);
1813 			ccb_malloced = 1;
1814 		}
1815 
1816 		if (ccb == NULL) {
1817 			xpt_print(periph->path, "unable to allocate CCB\n");
1818 			error = ENOMEM;
1819 			break;
1820 		}
1821 
1822 		error = passsendccb(periph, ccb, inccb);
1823 
1824 		if (ccb_malloced)
1825 			xpt_free_ccb(ccb);
1826 		else
1827 			xpt_release_ccb(ccb);
1828 
1829 		break;
1830 	}
1831 	case CAMIOQUEUE:
1832 	{
1833 		struct pass_io_req *io_req;
1834 		union ccb **user_ccb, *ccb;
1835 		xpt_opcode fc;
1836 
1837 #ifdef COMPAT_FREEBSD32
1838 		if (SV_PROC_FLAG(td->td_proc, SV_ILP32)) {
1839 			error = ENOTTY;
1840 			goto bailout;
1841 		}
1842 #endif
1843 		if ((softc->flags & PASS_FLAG_ZONE_VALID) == 0) {
1844 			error = passcreatezone(periph);
1845 			if (error != 0)
1846 				goto bailout;
1847 		}
1848 
1849 		/*
1850 		 * We're going to do a blocking allocation for this I/O
1851 		 * request, so we have to drop the lock.
1852 		 */
1853 		cam_periph_unlock(periph);
1854 
1855 		io_req = uma_zalloc(softc->pass_zone, M_WAITOK | M_ZERO);
1856 		ccb = &io_req->ccb;
1857 		user_ccb = (union ccb **)addr;
1858 
1859 		/*
1860 		 * Unlike the CAMIOCOMMAND ioctl above, we only have a
1861 		 * pointer to the user's CCB, so we have to copy the whole
1862 		 * thing in to a buffer we have allocated (above) instead
1863 		 * of allowing the ioctl code to malloc a buffer and copy
1864 		 * it in.
1865 		 *
1866 		 * This is an advantage for this asynchronous interface,
1867 		 * since we don't want the memory to get freed while the
1868 		 * CCB is outstanding.
1869 		 */
1870 #if 0
1871 		xpt_print(periph->path, "Copying user CCB %p to "
1872 			  "kernel address %p\n", *user_ccb, ccb);
1873 #endif
1874 		error = copyin(*user_ccb, ccb, sizeof(*ccb));
1875 		if (error != 0) {
1876 			xpt_print(periph->path, "Copy of user CCB %p to "
1877 				  "kernel address %p failed with error %d\n",
1878 				  *user_ccb, ccb, error);
1879 			goto camioqueue_error;
1880 		}
1881 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
1882 		if (ccb->ccb_h.func_code == XPT_SCSI_IO)
1883 			ccb->csio.bio = NULL;
1884 #endif
1885 
1886 		if (ccb->ccb_h.flags & CAM_UNLOCKED) {
1887 			error = EINVAL;
1888 			goto camioqueue_error;
1889 		}
1890 
1891 		if (ccb->ccb_h.flags & CAM_CDB_POINTER) {
1892 			if (ccb->csio.cdb_len > IOCDBLEN) {
1893 				error = EINVAL;
1894 				goto camioqueue_error;
1895 			}
1896 			error = copyin(ccb->csio.cdb_io.cdb_ptr,
1897 			    ccb->csio.cdb_io.cdb_bytes, ccb->csio.cdb_len);
1898 			if (error != 0)
1899 				goto camioqueue_error;
1900 			ccb->ccb_h.flags &= ~CAM_CDB_POINTER;
1901 		}
1902 
1903 		/*
1904 		 * Some CCB types, like scan bus and scan lun can only go
1905 		 * through the transport layer device.
1906 		 */
1907 		if (ccb->ccb_h.func_code & XPT_FC_XPT_ONLY) {
1908 			xpt_print(periph->path, "CCB function code %#x is "
1909 			    "restricted to the XPT device\n",
1910 			    ccb->ccb_h.func_code);
1911 			error = ENODEV;
1912 			goto camioqueue_error;
1913 		}
1914 
1915 		/*
1916 		 * Save the user's CCB pointer as well as his linked list
1917 		 * pointers and peripheral private area so that we can
1918 		 * restore these later.
1919 		 */
1920 		io_req->user_ccb_ptr = *user_ccb;
1921 		io_req->user_periph_links = ccb->ccb_h.periph_links;
1922 		io_req->user_periph_priv = ccb->ccb_h.periph_priv;
1923 
1924 		/*
1925 		 * Now that we've saved the user's values, we can set our
1926 		 * own peripheral private entry.
1927 		 */
1928 		ccb->ccb_h.ccb_ioreq = io_req;
1929 
1930 		/* Compatibility for RL/priority-unaware code. */
1931 		priority = ccb->ccb_h.pinfo.priority;
1932 		if (priority <= CAM_PRIORITY_OOB)
1933 		    priority += CAM_PRIORITY_OOB + 1;
1934 
1935 		/*
1936 		 * Setup fields in the CCB like the path and the priority.
1937 		 * The path in particular cannot be done in userland, since
1938 		 * it is a pointer to a kernel data structure.
1939 		 */
1940 		xpt_setup_ccb_flags(&ccb->ccb_h, periph->path, priority,
1941 				    ccb->ccb_h.flags);
1942 
1943 		/*
1944 		 * Setup our done routine.  There is no way for the user to
1945 		 * have a valid pointer here.
1946 		 */
1947 		ccb->ccb_h.cbfcnp = passdone;
1948 
1949 		fc = ccb->ccb_h.func_code;
1950 		/*
1951 		 * If this function code has memory that can be mapped in
1952 		 * or out, we need to call passmemsetup().
1953 		 */
1954 		if ((fc == XPT_SCSI_IO) || (fc == XPT_ATA_IO)
1955 		 || (fc == XPT_SMP_IO) || (fc == XPT_DEV_MATCH)
1956 		 || (fc == XPT_DEV_ADVINFO)
1957 		 || (fc == XPT_NVME_ADMIN) || (fc == XPT_NVME_IO)) {
1958 			error = passmemsetup(periph, io_req);
1959 			if (error != 0)
1960 				goto camioqueue_error;
1961 		} else
1962 			io_req->mapinfo.num_bufs_used = 0;
1963 
1964 		cam_periph_lock(periph);
1965 
1966 		/*
1967 		 * Everything goes on the incoming queue initially.
1968 		 */
1969 		TAILQ_INSERT_TAIL(&softc->incoming_queue, io_req, links);
1970 
1971 		/*
1972 		 * If the CCB is queued, and is not a user CCB, then
1973 		 * we need to allocate a slot for it.  Call xpt_schedule()
1974 		 * so that our start routine will get called when a CCB is
1975 		 * available.
1976 		 */
1977 		if ((fc & XPT_FC_QUEUED)
1978 		 && ((fc & XPT_FC_USER_CCB) == 0)) {
1979 			xpt_schedule(periph, priority);
1980 			break;
1981 		}
1982 
1983 		/*
1984 		 * At this point, the CCB in question is either an
1985 		 * immediate CCB (like XPT_DEV_ADVINFO) or it is a user CCB
1986 		 * and therefore should be malloced, not allocated via a slot.
1987 		 * Remove the CCB from the incoming queue and add it to the
1988 		 * active queue.
1989 		 */
1990 		TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
1991 		TAILQ_INSERT_TAIL(&softc->active_queue, io_req, links);
1992 
1993 		xpt_action(ccb);
1994 
1995 		/*
1996 		 * If this is not a queued CCB (i.e. it is an immediate CCB),
1997 		 * then it is already done.  We need to put it on the done
1998 		 * queue for the user to fetch.
1999 		 */
2000 		if ((fc & XPT_FC_QUEUED) == 0) {
2001 			TAILQ_REMOVE(&softc->active_queue, io_req, links);
2002 			TAILQ_INSERT_TAIL(&softc->done_queue, io_req, links);
2003 		}
2004 		break;
2005 
2006 camioqueue_error:
2007 		uma_zfree(softc->pass_zone, io_req);
2008 		cam_periph_lock(periph);
2009 		break;
2010 	}
2011 	case CAMIOGET:
2012 	{
2013 		union ccb **user_ccb;
2014 		struct pass_io_req *io_req;
2015 		int old_error;
2016 
2017 #ifdef COMPAT_FREEBSD32
2018 		if (SV_PROC_FLAG(td->td_proc, SV_ILP32)) {
2019 			error = ENOTTY;
2020 			goto bailout;
2021 		}
2022 #endif
2023 		user_ccb = (union ccb **)addr;
2024 		old_error = 0;
2025 
2026 		io_req = TAILQ_FIRST(&softc->done_queue);
2027 		if (io_req == NULL) {
2028 			error = ENOENT;
2029 			break;
2030 		}
2031 
2032 		/*
2033 		 * Remove the I/O from the done queue.
2034 		 */
2035 		TAILQ_REMOVE(&softc->done_queue, io_req, links);
2036 
2037 		/*
2038 		 * We have to drop the lock during the copyout because the
2039 		 * copyout can result in VM faults that require sleeping.
2040 		 */
2041 		cam_periph_unlock(periph);
2042 
2043 		/*
2044 		 * Do any needed copies (e.g. for reads) and revert the
2045 		 * pointers in the CCB back to the user's pointers.
2046 		 */
2047 		error = passmemdone(periph, io_req);
2048 
2049 		old_error = error;
2050 
2051 		io_req->ccb.ccb_h.periph_links = io_req->user_periph_links;
2052 		io_req->ccb.ccb_h.periph_priv = io_req->user_periph_priv;
2053 
2054 #if 0
2055 		xpt_print(periph->path, "Copying to user CCB %p from "
2056 			  "kernel address %p\n", *user_ccb, &io_req->ccb);
2057 #endif
2058 
2059 		error = copyout(&io_req->ccb, *user_ccb, sizeof(union ccb));
2060 		if (error != 0) {
2061 			xpt_print(periph->path, "Copy to user CCB %p from "
2062 				  "kernel address %p failed with error %d\n",
2063 				  *user_ccb, &io_req->ccb, error);
2064 		}
2065 
2066 		/*
2067 		 * Prefer the first error we got back, and make sure we
2068 		 * don't overwrite bad status with good.
2069 		 */
2070 		if (old_error != 0)
2071 			error = old_error;
2072 
2073 		cam_periph_lock(periph);
2074 
2075 		/*
2076 		 * At this point, if there was an error, we could potentially
2077 		 * re-queue the I/O and try again.  But why?  The error
2078 		 * would almost certainly happen again.  We might as well
2079 		 * not leak memory.
2080 		 */
2081 		uma_zfree(softc->pass_zone, io_req);
2082 		break;
2083 	}
2084 	default:
2085 		error = cam_periph_ioctl(periph, cmd, addr, passerror);
2086 		break;
2087 	}
2088 
2089 bailout:
2090 	cam_periph_unlock(periph);
2091 
2092 	return(error);
2093 }
2094 
2095 static int
2096 passpoll(struct cdev *dev, int poll_events, struct thread *td)
2097 {
2098 	struct cam_periph *periph;
2099 	struct pass_softc *softc;
2100 	int revents;
2101 
2102 	periph = (struct cam_periph *)dev->si_drv1;
2103 	softc = (struct pass_softc *)periph->softc;
2104 
2105 	revents = poll_events & (POLLOUT | POLLWRNORM);
2106 	if ((poll_events & (POLLIN | POLLRDNORM)) != 0) {
2107 		cam_periph_lock(periph);
2108 
2109 		if (!TAILQ_EMPTY(&softc->done_queue)) {
2110 			revents |= poll_events & (POLLIN | POLLRDNORM);
2111 		}
2112 		cam_periph_unlock(periph);
2113 		if (revents == 0)
2114 			selrecord(td, &softc->read_select);
2115 	}
2116 
2117 	return (revents);
2118 }
2119 
2120 static int
2121 passkqfilter(struct cdev *dev, struct knote *kn)
2122 {
2123 	struct cam_periph *periph;
2124 	struct pass_softc *softc;
2125 
2126 	periph = (struct cam_periph *)dev->si_drv1;
2127 	softc = (struct pass_softc *)periph->softc;
2128 
2129 	kn->kn_hook = (caddr_t)periph;
2130 	kn->kn_fop = &passread_filtops;
2131 	knlist_add(&softc->read_select.si_note, kn, 0);
2132 
2133 	return (0);
2134 }
2135 
2136 static void
2137 passreadfiltdetach(struct knote *kn)
2138 {
2139 	struct cam_periph *periph;
2140 	struct pass_softc *softc;
2141 
2142 	periph = (struct cam_periph *)kn->kn_hook;
2143 	softc = (struct pass_softc *)periph->softc;
2144 
2145 	knlist_remove(&softc->read_select.si_note, kn, 0);
2146 }
2147 
2148 static int
2149 passreadfilt(struct knote *kn, long hint)
2150 {
2151 	struct cam_periph *periph;
2152 	struct pass_softc *softc;
2153 	int retval;
2154 
2155 	periph = (struct cam_periph *)kn->kn_hook;
2156 	softc = (struct pass_softc *)periph->softc;
2157 
2158 	cam_periph_assert(periph, MA_OWNED);
2159 
2160 	if (TAILQ_EMPTY(&softc->done_queue))
2161 		retval = 0;
2162 	else
2163 		retval = 1;
2164 
2165 	return (retval);
2166 }
2167 
2168 /*
2169  * Generally, "ccb" should be the CCB supplied by the kernel.  "inccb"
2170  * should be the CCB that is copied in from the user.
2171  */
2172 static int
2173 passsendccb(struct cam_periph *periph, union ccb *ccb, union ccb *inccb)
2174 {
2175 	struct pass_softc *softc;
2176 	struct cam_periph_map_info mapinfo;
2177 	uint8_t *cmd;
2178 	xpt_opcode fc;
2179 	int error;
2180 
2181 	softc = (struct pass_softc *)periph->softc;
2182 
2183 	/*
2184 	 * There are some fields in the CCB header that need to be
2185 	 * preserved, the rest we get from the user.
2186 	 */
2187 	xpt_merge_ccb(ccb, inccb);
2188 
2189 	if (ccb->ccb_h.flags & CAM_CDB_POINTER) {
2190 		cmd = __builtin_alloca(ccb->csio.cdb_len);
2191 		error = copyin(ccb->csio.cdb_io.cdb_ptr, cmd, ccb->csio.cdb_len);
2192 		if (error)
2193 			return (error);
2194 		ccb->csio.cdb_io.cdb_ptr = cmd;
2195 	}
2196 
2197 	/*
2198 	 * Let cam_periph_mapmem do a sanity check on the data pointer format.
2199 	 * Even if no data transfer is needed, it's a cheap check and it
2200 	 * simplifies the code.
2201 	 */
2202 	fc = ccb->ccb_h.func_code;
2203 	if ((fc == XPT_SCSI_IO) || (fc == XPT_ATA_IO) || (fc == XPT_SMP_IO)
2204             || (fc == XPT_DEV_MATCH) || (fc == XPT_DEV_ADVINFO) || (fc == XPT_MMC_IO)
2205             || (fc == XPT_NVME_ADMIN) || (fc == XPT_NVME_IO)) {
2206 		bzero(&mapinfo, sizeof(mapinfo));
2207 
2208 		/*
2209 		 * cam_periph_mapmem calls into proc and vm functions that can
2210 		 * sleep as well as trigger I/O, so we can't hold the lock.
2211 		 * Dropping it here is reasonably safe.
2212 		 */
2213 		cam_periph_unlock(periph);
2214 		error = cam_periph_mapmem(ccb, &mapinfo, softc->maxio);
2215 		cam_periph_lock(periph);
2216 
2217 		/*
2218 		 * cam_periph_mapmem returned an error, we can't continue.
2219 		 * Return the error to the user.
2220 		 */
2221 		if (error)
2222 			return(error);
2223 	} else
2224 		/* Ensure that the unmap call later on is a no-op. */
2225 		mapinfo.num_bufs_used = 0;
2226 
2227 	/*
2228 	 * If the user wants us to perform any error recovery, then honor
2229 	 * that request.  Otherwise, it's up to the user to perform any
2230 	 * error recovery.
2231 	 */
2232 	cam_periph_runccb(ccb, (ccb->ccb_h.flags & CAM_PASS_ERR_RECOVER) ?
2233 	    passerror : NULL, /* cam_flags */ CAM_RETRY_SELTO,
2234 	    /* sense_flags */ SF_RETRY_UA | SF_NO_PRINT,
2235 	    softc->device_stats);
2236 
2237 	cam_periph_unlock(periph);
2238 	cam_periph_unmapmem(ccb, &mapinfo);
2239 	cam_periph_lock(periph);
2240 
2241 	ccb->ccb_h.cbfcnp = NULL;
2242 	ccb->ccb_h.periph_priv = inccb->ccb_h.periph_priv;
2243 	bcopy(ccb, inccb, sizeof(union ccb));
2244 
2245 	return(0);
2246 }
2247 
2248 static int
2249 passerror(union ccb *ccb, u_int32_t cam_flags, u_int32_t sense_flags)
2250 {
2251 
2252 	return(cam_periph_error(ccb, cam_flags, sense_flags));
2253 }
2254