xref: /freebsd/sys/cam/scsi/scsi_pass.c (revision 77c1fcec914bae1232762ca0b0462abd6957ff26)
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 	union ccb		  saved_ccb;
133 	int			  open_count;
134 	u_int		 	  maxio;
135 	struct devstat		 *device_stats;
136 	struct cdev		 *dev;
137 	struct cdev		 *alias_dev;
138 	struct task		  add_physpath_task;
139 	struct task		  shutdown_kqueue_task;
140 	struct selinfo		  read_select;
141 	TAILQ_HEAD(, pass_io_req) incoming_queue;
142 	TAILQ_HEAD(, pass_io_req) active_queue;
143 	TAILQ_HEAD(, pass_io_req) abandoned_queue;
144 	TAILQ_HEAD(, pass_io_req) done_queue;
145 	struct cam_periph	 *periph;
146 	char			  zone_name[12];
147 	char			  io_zone_name[12];
148 	uma_zone_t		  pass_zone;
149 	uma_zone_t		  pass_io_zone;
150 	size_t			  io_zone_size;
151 };
152 
153 static	d_open_t	passopen;
154 static	d_close_t	passclose;
155 static	d_ioctl_t	passioctl;
156 static	d_ioctl_t	passdoioctl;
157 static	d_poll_t	passpoll;
158 static	d_kqfilter_t	passkqfilter;
159 static	void		passreadfiltdetach(struct knote *kn);
160 static	int		passreadfilt(struct knote *kn, long hint);
161 
162 static	periph_init_t	passinit;
163 static	periph_ctor_t	passregister;
164 static	periph_oninv_t	passoninvalidate;
165 static	periph_dtor_t	passcleanup;
166 static	periph_start_t	passstart;
167 static	void		pass_shutdown_kqueue(void *context, int pending);
168 static	void		pass_add_physpath(void *context, int pending);
169 static	void		passasync(void *callback_arg, u_int32_t code,
170 				  struct cam_path *path, void *arg);
171 static	void		passdone(struct cam_periph *periph,
172 				 union ccb *done_ccb);
173 static	int		passcreatezone(struct cam_periph *periph);
174 static	void		passiocleanup(struct pass_softc *softc,
175 				      struct pass_io_req *io_req);
176 static	int		passcopysglist(struct cam_periph *periph,
177 				       struct pass_io_req *io_req,
178 				       ccb_flags direction);
179 static	int		passmemsetup(struct cam_periph *periph,
180 				     struct pass_io_req *io_req);
181 static	int		passmemdone(struct cam_periph *periph,
182 				    struct pass_io_req *io_req);
183 static	int		passerror(union ccb *ccb, u_int32_t cam_flags,
184 				  u_int32_t sense_flags);
185 static 	int		passsendccb(struct cam_periph *periph, union ccb *ccb,
186 				    union ccb *inccb);
187 
188 static struct periph_driver passdriver =
189 {
190 	passinit, "pass",
191 	TAILQ_HEAD_INITIALIZER(passdriver.units), /* generation */ 0
192 };
193 
194 PERIPHDRIVER_DECLARE(pass, passdriver);
195 
196 static struct cdevsw pass_cdevsw = {
197 	.d_version =	D_VERSION,
198 	.d_flags =	D_TRACKCLOSE,
199 	.d_open =	passopen,
200 	.d_close =	passclose,
201 	.d_ioctl =	passioctl,
202 	.d_poll = 	passpoll,
203 	.d_kqfilter = 	passkqfilter,
204 	.d_name =	"pass",
205 };
206 
207 static struct filterops passread_filtops = {
208 	.f_isfd	=	1,
209 	.f_detach =	passreadfiltdetach,
210 	.f_event =	passreadfilt
211 };
212 
213 static MALLOC_DEFINE(M_SCSIPASS, "scsi_pass", "scsi passthrough buffers");
214 
215 static void
216 passinit(void)
217 {
218 	cam_status status;
219 
220 	/*
221 	 * Install a global async callback.  This callback will
222 	 * receive async callbacks like "new device found".
223 	 */
224 	status = xpt_register_async(AC_FOUND_DEVICE, passasync, NULL, NULL);
225 
226 	if (status != CAM_REQ_CMP) {
227 		printf("pass: Failed to attach master async callback "
228 		       "due to status 0x%x!\n", status);
229 	}
230 
231 }
232 
233 static void
234 passrejectios(struct cam_periph *periph)
235 {
236 	struct pass_io_req *io_req, *io_req2;
237 	struct pass_softc *softc;
238 
239 	softc = (struct pass_softc *)periph->softc;
240 
241 	/*
242 	 * The user can no longer get status for I/O on the done queue, so
243 	 * clean up all outstanding I/O on the done queue.
244 	 */
245 	TAILQ_FOREACH_SAFE(io_req, &softc->done_queue, links, io_req2) {
246 		TAILQ_REMOVE(&softc->done_queue, io_req, links);
247 		passiocleanup(softc, io_req);
248 		uma_zfree(softc->pass_zone, io_req);
249 	}
250 
251 	/*
252 	 * The underlying device is gone, so we can't issue these I/Os.
253 	 * The devfs node has been shut down, so we can't return status to
254 	 * the user.  Free any I/O left on the incoming queue.
255 	 */
256 	TAILQ_FOREACH_SAFE(io_req, &softc->incoming_queue, links, io_req2) {
257 		TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
258 		passiocleanup(softc, io_req);
259 		uma_zfree(softc->pass_zone, io_req);
260 	}
261 
262 	/*
263 	 * Normally we would put I/Os on the abandoned queue and acquire a
264 	 * reference when we saw the final close.  But, the device went
265 	 * away and devfs may have moved everything off to deadfs by the
266 	 * time the I/O done callback is called; as a result, we won't see
267 	 * any more closes.  So, if we have any active I/Os, we need to put
268 	 * them on the abandoned queue.  When the abandoned queue is empty,
269 	 * we'll release the remaining reference (see below) to the peripheral.
270 	 */
271 	TAILQ_FOREACH_SAFE(io_req, &softc->active_queue, links, io_req2) {
272 		TAILQ_REMOVE(&softc->active_queue, io_req, links);
273 		io_req->flags |= PASS_IO_ABANDONED;
274 		TAILQ_INSERT_TAIL(&softc->abandoned_queue, io_req, links);
275 	}
276 
277 	/*
278 	 * If we put any I/O on the abandoned queue, acquire a reference.
279 	 */
280 	if ((!TAILQ_EMPTY(&softc->abandoned_queue))
281 	 && ((softc->flags & PASS_FLAG_ABANDONED_REF_SET) == 0)) {
282 		cam_periph_doacquire(periph);
283 		softc->flags |= PASS_FLAG_ABANDONED_REF_SET;
284 	}
285 }
286 
287 static void
288 passdevgonecb(void *arg)
289 {
290 	struct cam_periph *periph;
291 	struct mtx *mtx;
292 	struct pass_softc *softc;
293 	int i;
294 
295 	periph = (struct cam_periph *)arg;
296 	mtx = cam_periph_mtx(periph);
297 	mtx_lock(mtx);
298 
299 	softc = (struct pass_softc *)periph->softc;
300 	KASSERT(softc->open_count >= 0, ("Negative open count %d",
301 		softc->open_count));
302 
303 	/*
304 	 * When we get this callback, we will get no more close calls from
305 	 * devfs.  So if we have any dangling opens, we need to release the
306 	 * reference held for that particular context.
307 	 */
308 	for (i = 0; i < softc->open_count; i++)
309 		cam_periph_release_locked(periph);
310 
311 	softc->open_count = 0;
312 
313 	/*
314 	 * Release the reference held for the device node, it is gone now.
315 	 * Accordingly, inform all queued I/Os of their fate.
316 	 */
317 	cam_periph_release_locked(periph);
318 	passrejectios(periph);
319 
320 	/*
321 	 * We reference the SIM lock directly here, instead of using
322 	 * cam_periph_unlock().  The reason is that the final call to
323 	 * cam_periph_release_locked() above could result in the periph
324 	 * getting freed.  If that is the case, dereferencing the periph
325 	 * with a cam_periph_unlock() call would cause a page fault.
326 	 */
327 	mtx_unlock(mtx);
328 
329 	/*
330 	 * We have to remove our kqueue context from a thread because it
331 	 * may sleep.  It would be nice if we could get a callback from
332 	 * kqueue when it is done cleaning up resources.
333 	 */
334 	taskqueue_enqueue(taskqueue_thread, &softc->shutdown_kqueue_task);
335 }
336 
337 static void
338 passoninvalidate(struct cam_periph *periph)
339 {
340 	struct pass_softc *softc;
341 
342 	softc = (struct pass_softc *)periph->softc;
343 
344 	/*
345 	 * De-register any async callbacks.
346 	 */
347 	xpt_register_async(0, passasync, periph, periph->path);
348 
349 	softc->flags |= PASS_FLAG_INVALID;
350 
351 	/*
352 	 * Tell devfs this device has gone away, and ask for a callback
353 	 * when it has cleaned up its state.
354 	 */
355 	destroy_dev_sched_cb(softc->dev, passdevgonecb, periph);
356 }
357 
358 static void
359 passcleanup(struct cam_periph *periph)
360 {
361 	struct pass_softc *softc;
362 
363 	softc = (struct pass_softc *)periph->softc;
364 
365 	cam_periph_assert(periph, MA_OWNED);
366 	KASSERT(TAILQ_EMPTY(&softc->active_queue),
367 		("%s called when there are commands on the active queue!\n",
368 		__func__));
369 	KASSERT(TAILQ_EMPTY(&softc->abandoned_queue),
370 		("%s called when there are commands on the abandoned queue!\n",
371 		__func__));
372 	KASSERT(TAILQ_EMPTY(&softc->incoming_queue),
373 		("%s called when there are commands on the incoming queue!\n",
374 		__func__));
375 	KASSERT(TAILQ_EMPTY(&softc->done_queue),
376 		("%s called when there are commands on the done queue!\n",
377 		__func__));
378 
379 	devstat_remove_entry(softc->device_stats);
380 
381 	cam_periph_unlock(periph);
382 
383 	/*
384 	 * We call taskqueue_drain() for the physpath task to make sure it
385 	 * is complete.  We drop the lock because this can potentially
386 	 * sleep.  XXX KDM that is bad.  Need a way to get a callback when
387 	 * a taskqueue is drained.
388 	 *
389  	 * Note that we don't drain the kqueue shutdown task queue.  This
390 	 * is because we hold a reference on the periph for kqueue, and
391 	 * release that reference from the kqueue shutdown task queue.  So
392 	 * we cannot come into this routine unless we've released that
393 	 * reference.  Also, because that could be the last reference, we
394 	 * could be called from the cam_periph_release() call in
395 	 * pass_shutdown_kqueue().  In that case, the taskqueue_drain()
396 	 * would deadlock.  It would be preferable if we had a way to
397 	 * get a callback when a taskqueue is done.
398 	 */
399 	taskqueue_drain(taskqueue_thread, &softc->add_physpath_task);
400 
401 	cam_periph_lock(periph);
402 
403 	free(softc, M_DEVBUF);
404 }
405 
406 static void
407 pass_shutdown_kqueue(void *context, int pending)
408 {
409 	struct cam_periph *periph;
410 	struct pass_softc *softc;
411 
412 	periph = context;
413 	softc = periph->softc;
414 
415 	knlist_clear(&softc->read_select.si_note, /*is_locked*/ 0);
416 	knlist_destroy(&softc->read_select.si_note);
417 
418 	/*
419 	 * Release the reference we held for kqueue.
420 	 */
421 	cam_periph_release(periph);
422 }
423 
424 static void
425 pass_add_physpath(void *context, int pending)
426 {
427 	struct cam_periph *periph;
428 	struct pass_softc *softc;
429 	struct mtx *mtx;
430 	char *physpath;
431 
432 	/*
433 	 * If we have one, create a devfs alias for our
434 	 * physical path.
435 	 */
436 	periph = context;
437 	softc = periph->softc;
438 	physpath = malloc(MAXPATHLEN, M_DEVBUF, M_WAITOK);
439 	mtx = cam_periph_mtx(periph);
440 	mtx_lock(mtx);
441 
442 	if (periph->flags & CAM_PERIPH_INVALID)
443 		goto out;
444 
445 	if (xpt_getattr(physpath, MAXPATHLEN,
446 			"GEOM::physpath", periph->path) == 0
447 	 && strlen(physpath) != 0) {
448 
449 		mtx_unlock(mtx);
450 		make_dev_physpath_alias(MAKEDEV_WAITOK, &softc->alias_dev,
451 					softc->dev, softc->alias_dev, physpath);
452 		mtx_lock(mtx);
453 	}
454 
455 out:
456 	/*
457 	 * Now that we've made our alias, we no longer have to have a
458 	 * reference to the device.
459 	 */
460 	if ((softc->flags & PASS_FLAG_INITIAL_PHYSPATH) == 0)
461 		softc->flags |= PASS_FLAG_INITIAL_PHYSPATH;
462 
463 	/*
464 	 * We always acquire a reference to the periph before queueing this
465 	 * task queue function, so it won't go away before we run.
466 	 */
467 	while (pending-- > 0)
468 		cam_periph_release_locked(periph);
469 	mtx_unlock(mtx);
470 
471 	free(physpath, M_DEVBUF);
472 }
473 
474 static void
475 passasync(void *callback_arg, u_int32_t code,
476 	  struct cam_path *path, void *arg)
477 {
478 	struct cam_periph *periph;
479 
480 	periph = (struct cam_periph *)callback_arg;
481 
482 	switch (code) {
483 	case AC_FOUND_DEVICE:
484 	{
485 		struct ccb_getdev *cgd;
486 		cam_status status;
487 
488 		cgd = (struct ccb_getdev *)arg;
489 		if (cgd == NULL)
490 			break;
491 
492 		/*
493 		 * Allocate a peripheral instance for
494 		 * this device and start the probe
495 		 * process.
496 		 */
497 		status = cam_periph_alloc(passregister, passoninvalidate,
498 					  passcleanup, passstart, "pass",
499 					  CAM_PERIPH_BIO, path,
500 					  passasync, AC_FOUND_DEVICE, cgd);
501 
502 		if (status != CAM_REQ_CMP
503 		 && status != CAM_REQ_INPROG) {
504 			const struct cam_status_entry *entry;
505 
506 			entry = cam_fetch_status_entry(status);
507 
508 			printf("passasync: Unable to attach new device "
509 			       "due to status %#x: %s\n", status, entry ?
510 			       entry->status_text : "Unknown");
511 		}
512 
513 		break;
514 	}
515 	case AC_ADVINFO_CHANGED:
516 	{
517 		uintptr_t buftype;
518 
519 		buftype = (uintptr_t)arg;
520 		if (buftype == CDAI_TYPE_PHYS_PATH) {
521 			struct pass_softc *softc;
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 			if (cam_periph_acquire(periph) != 0)
531 				break;
532 
533 			taskqueue_enqueue(taskqueue_thread,
534 					  &softc->add_physpath_task);
535 		}
536 		break;
537 	}
538 	default:
539 		cam_periph_async(periph, code, path, arg);
540 		break;
541 	}
542 }
543 
544 static cam_status
545 passregister(struct cam_periph *periph, void *arg)
546 {
547 	struct pass_softc *softc;
548 	struct ccb_getdev *cgd;
549 	struct ccb_pathinq cpi;
550 	struct make_dev_args args;
551 	int error, no_tags;
552 
553 	cgd = (struct ccb_getdev *)arg;
554 	if (cgd == NULL) {
555 		printf("%s: no getdev CCB, can't register device\n", __func__);
556 		return(CAM_REQ_CMP_ERR);
557 	}
558 
559 	softc = (struct pass_softc *)malloc(sizeof(*softc),
560 					    M_DEVBUF, M_NOWAIT);
561 
562 	if (softc == NULL) {
563 		printf("%s: Unable to probe new device. "
564 		       "Unable to allocate softc\n", __func__);
565 		return(CAM_REQ_CMP_ERR);
566 	}
567 
568 	bzero(softc, sizeof(*softc));
569 	softc->state = PASS_STATE_NORMAL;
570 	if (cgd->protocol == PROTO_SCSI || cgd->protocol == PROTO_ATAPI)
571 		softc->pd_type = SID_TYPE(&cgd->inq_data);
572 	else if (cgd->protocol == PROTO_SATAPM)
573 		softc->pd_type = T_ENCLOSURE;
574 	else
575 		softc->pd_type = T_DIRECT;
576 
577 	periph->softc = softc;
578 	softc->periph = periph;
579 	TAILQ_INIT(&softc->incoming_queue);
580 	TAILQ_INIT(&softc->active_queue);
581 	TAILQ_INIT(&softc->abandoned_queue);
582 	TAILQ_INIT(&softc->done_queue);
583 	snprintf(softc->zone_name, sizeof(softc->zone_name), "%s%d",
584 		 periph->periph_name, periph->unit_number);
585 	snprintf(softc->io_zone_name, sizeof(softc->io_zone_name), "%s%dIO",
586 		 periph->periph_name, periph->unit_number);
587 	softc->io_zone_size = MAXPHYS;
588 	knlist_init_mtx(&softc->read_select.si_note, cam_periph_mtx(periph));
589 
590 	xpt_path_inq(&cpi, periph->path);
591 
592 	if (cpi.maxio == 0)
593 		softc->maxio = DFLTPHYS;	/* traditional default */
594 	else if (cpi.maxio > MAXPHYS)
595 		softc->maxio = MAXPHYS;		/* for safety */
596 	else
597 		softc->maxio = cpi.maxio;	/* real value */
598 
599 	if (cpi.hba_misc & PIM_UNMAPPED)
600 		softc->flags |= PASS_FLAG_UNMAPPED_CAPABLE;
601 
602 	/*
603 	 * We pass in 0 for a blocksize, since we don't
604 	 * know what the blocksize of this device is, if
605 	 * it even has a blocksize.
606 	 */
607 	cam_periph_unlock(periph);
608 	no_tags = (cgd->inq_data.flags & SID_CmdQue) == 0;
609 	softc->device_stats = devstat_new_entry("pass",
610 			  periph->unit_number, 0,
611 			  DEVSTAT_NO_BLOCKSIZE
612 			  | (no_tags ? DEVSTAT_NO_ORDERED_TAGS : 0),
613 			  softc->pd_type |
614 			  XPORT_DEVSTAT_TYPE(cpi.transport) |
615 			  DEVSTAT_TYPE_PASS,
616 			  DEVSTAT_PRIORITY_PASS);
617 
618 	/*
619 	 * Initialize the taskqueue handler for shutting down kqueue.
620 	 */
621 	TASK_INIT(&softc->shutdown_kqueue_task, /*priority*/ 0,
622 		  pass_shutdown_kqueue, periph);
623 
624 	/*
625 	 * Acquire a reference to the periph that we can release once we've
626 	 * cleaned up the kqueue.
627 	 */
628 	if (cam_periph_acquire(periph) != 0) {
629 		xpt_print(periph->path, "%s: lost periph during "
630 			  "registration!\n", __func__);
631 		cam_periph_lock(periph);
632 		return (CAM_REQ_CMP_ERR);
633 	}
634 
635 	/*
636 	 * Acquire a reference to the periph before we create the devfs
637 	 * instance for it.  We'll release this reference once the devfs
638 	 * instance has been freed.
639 	 */
640 	if (cam_periph_acquire(periph) != 0) {
641 		xpt_print(periph->path, "%s: lost periph during "
642 			  "registration!\n", __func__);
643 		cam_periph_lock(periph);
644 		return (CAM_REQ_CMP_ERR);
645 	}
646 
647 	/* Register the device */
648 	make_dev_args_init(&args);
649 	args.mda_devsw = &pass_cdevsw;
650 	args.mda_unit = periph->unit_number;
651 	args.mda_uid = UID_ROOT;
652 	args.mda_gid = GID_OPERATOR;
653 	args.mda_mode = 0600;
654 	args.mda_si_drv1 = periph;
655 	error = make_dev_s(&args, &softc->dev, "%s%d", periph->periph_name,
656 	    periph->unit_number);
657 	if (error != 0) {
658 		cam_periph_lock(periph);
659 		cam_periph_release_locked(periph);
660 		return (CAM_REQ_CMP_ERR);
661 	}
662 
663 	/*
664 	 * Hold a reference to the periph before we create the physical
665 	 * path alias so it can't go away.
666 	 */
667 	if (cam_periph_acquire(periph) != 0) {
668 		xpt_print(periph->path, "%s: lost periph during "
669 			  "registration!\n", __func__);
670 		cam_periph_lock(periph);
671 		return (CAM_REQ_CMP_ERR);
672 	}
673 
674 	cam_periph_lock(periph);
675 
676 	TASK_INIT(&softc->add_physpath_task, /*priority*/0,
677 		  pass_add_physpath, periph);
678 
679 	/*
680 	 * See if physical path information is already available.
681 	 */
682 	taskqueue_enqueue(taskqueue_thread, &softc->add_physpath_task);
683 
684 	/*
685 	 * Add an async callback so that we get notified if
686 	 * this device goes away or its physical path
687 	 * (stored in the advanced info data of the EDT) has
688 	 * changed.
689 	 */
690 	xpt_register_async(AC_LOST_DEVICE | AC_ADVINFO_CHANGED,
691 			   passasync, periph, periph->path);
692 
693 	if (bootverbose)
694 		xpt_announce_periph(periph, NULL);
695 
696 	return(CAM_REQ_CMP);
697 }
698 
699 static int
700 passopen(struct cdev *dev, int flags, int fmt, struct thread *td)
701 {
702 	struct cam_periph *periph;
703 	struct pass_softc *softc;
704 	int error;
705 
706 	periph = (struct cam_periph *)dev->si_drv1;
707 	if (cam_periph_acquire(periph) != 0)
708 		return (ENXIO);
709 
710 	cam_periph_lock(periph);
711 
712 	softc = (struct pass_softc *)periph->softc;
713 
714 	if (softc->flags & PASS_FLAG_INVALID) {
715 		cam_periph_release_locked(periph);
716 		cam_periph_unlock(periph);
717 		return(ENXIO);
718 	}
719 
720 	/*
721 	 * Don't allow access when we're running at a high securelevel.
722 	 */
723 	error = securelevel_gt(td->td_ucred, 1);
724 	if (error) {
725 		cam_periph_release_locked(periph);
726 		cam_periph_unlock(periph);
727 		return(error);
728 	}
729 
730 	/*
731 	 * Only allow read-write access.
732 	 */
733 	if (((flags & FWRITE) == 0) || ((flags & FREAD) == 0)) {
734 		cam_periph_release_locked(periph);
735 		cam_periph_unlock(periph);
736 		return(EPERM);
737 	}
738 
739 	/*
740 	 * We don't allow nonblocking access.
741 	 */
742 	if ((flags & O_NONBLOCK) != 0) {
743 		xpt_print(periph->path, "can't do nonblocking access\n");
744 		cam_periph_release_locked(periph);
745 		cam_periph_unlock(periph);
746 		return(EINVAL);
747 	}
748 
749 	softc->open_count++;
750 
751 	cam_periph_unlock(periph);
752 
753 	return (error);
754 }
755 
756 static int
757 passclose(struct cdev *dev, int flag, int fmt, struct thread *td)
758 {
759 	struct 	cam_periph *periph;
760 	struct  pass_softc *softc;
761 	struct mtx *mtx;
762 
763 	periph = (struct cam_periph *)dev->si_drv1;
764 	mtx = cam_periph_mtx(periph);
765 	mtx_lock(mtx);
766 
767 	softc = periph->softc;
768 	softc->open_count--;
769 
770 	if (softc->open_count == 0) {
771 		struct pass_io_req *io_req, *io_req2;
772 
773 		TAILQ_FOREACH_SAFE(io_req, &softc->done_queue, links, io_req2) {
774 			TAILQ_REMOVE(&softc->done_queue, io_req, links);
775 			passiocleanup(softc, io_req);
776 			uma_zfree(softc->pass_zone, io_req);
777 		}
778 
779 		TAILQ_FOREACH_SAFE(io_req, &softc->incoming_queue, links,
780 				   io_req2) {
781 			TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
782 			passiocleanup(softc, io_req);
783 			uma_zfree(softc->pass_zone, io_req);
784 		}
785 
786 		/*
787 		 * If there are any active I/Os, we need to forcibly acquire a
788 		 * reference to the peripheral so that we don't go away
789 		 * before they complete.  We'll release the reference when
790 		 * the abandoned queue is empty.
791 		 */
792 		io_req = TAILQ_FIRST(&softc->active_queue);
793 		if ((io_req != NULL)
794 		 && (softc->flags & PASS_FLAG_ABANDONED_REF_SET) == 0) {
795 			cam_periph_doacquire(periph);
796 			softc->flags |= PASS_FLAG_ABANDONED_REF_SET;
797 		}
798 
799 		/*
800 		 * Since the I/O in the active queue is not under our
801 		 * control, just set a flag so that we can clean it up when
802 		 * it completes and put it on the abandoned queue.  This
803 		 * will prevent our sending spurious completions in the
804 		 * event that the device is opened again before these I/Os
805 		 * complete.
806 		 */
807 		TAILQ_FOREACH_SAFE(io_req, &softc->active_queue, links,
808 				   io_req2) {
809 			TAILQ_REMOVE(&softc->active_queue, io_req, links);
810 			io_req->flags |= PASS_IO_ABANDONED;
811 			TAILQ_INSERT_TAIL(&softc->abandoned_queue, io_req,
812 					  links);
813 		}
814 	}
815 
816 	cam_periph_release_locked(periph);
817 
818 	/*
819 	 * We reference the lock directly here, instead of using
820 	 * cam_periph_unlock().  The reason is that the call to
821 	 * cam_periph_release_locked() above could result in the periph
822 	 * getting freed.  If that is the case, dereferencing the periph
823 	 * with a cam_periph_unlock() call would cause a page fault.
824 	 *
825 	 * cam_periph_release() avoids this problem using the same method,
826 	 * but we're manually acquiring and dropping the lock here to
827 	 * protect the open count and avoid another lock acquisition and
828 	 * release.
829 	 */
830 	mtx_unlock(mtx);
831 
832 	return (0);
833 }
834 
835 
836 static void
837 passstart(struct cam_periph *periph, union ccb *start_ccb)
838 {
839 	struct pass_softc *softc;
840 
841 	softc = (struct pass_softc *)periph->softc;
842 
843 	switch (softc->state) {
844 	case PASS_STATE_NORMAL: {
845 		struct pass_io_req *io_req;
846 
847 		/*
848 		 * Check for any queued I/O requests that require an
849 		 * allocated slot.
850 		 */
851 		io_req = TAILQ_FIRST(&softc->incoming_queue);
852 		if (io_req == NULL) {
853 			xpt_release_ccb(start_ccb);
854 			break;
855 		}
856 		TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
857 		TAILQ_INSERT_TAIL(&softc->active_queue, io_req, links);
858 		/*
859 		 * Merge the user's CCB into the allocated CCB.
860 		 */
861 		xpt_merge_ccb(start_ccb, &io_req->ccb);
862 		start_ccb->ccb_h.ccb_type = PASS_CCB_QUEUED_IO;
863 		start_ccb->ccb_h.ccb_ioreq = io_req;
864 		start_ccb->ccb_h.cbfcnp = passdone;
865 		io_req->alloced_ccb = start_ccb;
866 		binuptime(&io_req->start_time);
867 		devstat_start_transaction(softc->device_stats,
868 					  &io_req->start_time);
869 
870 		xpt_action(start_ccb);
871 
872 		/*
873 		 * If we have any more I/O waiting, schedule ourselves again.
874 		 */
875 		if (!TAILQ_EMPTY(&softc->incoming_queue))
876 			xpt_schedule(periph, CAM_PRIORITY_NORMAL);
877 		break;
878 	}
879 	default:
880 		break;
881 	}
882 }
883 
884 static void
885 passdone(struct cam_periph *periph, union ccb *done_ccb)
886 {
887 	struct pass_softc *softc;
888 	struct ccb_scsiio *csio;
889 
890 	softc = (struct pass_softc *)periph->softc;
891 
892 	cam_periph_assert(periph, MA_OWNED);
893 
894 	csio = &done_ccb->csio;
895 	switch (csio->ccb_h.ccb_type) {
896 	case PASS_CCB_QUEUED_IO: {
897 		struct pass_io_req *io_req;
898 
899 		io_req = done_ccb->ccb_h.ccb_ioreq;
900 #if 0
901 		xpt_print(periph->path, "%s: called for user CCB %p\n",
902 			  __func__, io_req->user_ccb_ptr);
903 #endif
904 		if (((done_ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_CMP)
905 		 && (done_ccb->ccb_h.flags & CAM_PASS_ERR_RECOVER)
906 		 && ((io_req->flags & PASS_IO_ABANDONED) == 0)) {
907 			int error;
908 
909 			error = passerror(done_ccb, CAM_RETRY_SELTO,
910 					  SF_RETRY_UA | SF_NO_PRINT);
911 
912 			if (error == ERESTART) {
913 				/*
914 				 * A retry was scheduled, so
915  				 * just return.
916 				 */
917 				return;
918 			}
919 		}
920 
921 		/*
922 		 * Copy the allocated CCB contents back to the malloced CCB
923 		 * so we can give status back to the user when he requests it.
924 		 */
925 		bcopy(done_ccb, &io_req->ccb, sizeof(*done_ccb));
926 
927 		/*
928 		 * Log data/transaction completion with devstat(9).
929 		 */
930 		switch (done_ccb->ccb_h.func_code) {
931 		case XPT_SCSI_IO:
932 			devstat_end_transaction(softc->device_stats,
933 			    done_ccb->csio.dxfer_len - done_ccb->csio.resid,
934 			    done_ccb->csio.tag_action & 0x3,
935 			    ((done_ccb->ccb_h.flags & CAM_DIR_MASK) ==
936 			    CAM_DIR_NONE) ? DEVSTAT_NO_DATA :
937 			    (done_ccb->ccb_h.flags & CAM_DIR_OUT) ?
938 			    DEVSTAT_WRITE : DEVSTAT_READ, NULL,
939 			    &io_req->start_time);
940 			break;
941 		case XPT_ATA_IO:
942 			devstat_end_transaction(softc->device_stats,
943 			    done_ccb->ataio.dxfer_len - done_ccb->ataio.resid,
944 			    0, /* Not used in ATA */
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_SMP_IO:
952 			/*
953 			 * XXX KDM this isn't quite right, but there isn't
954 			 * currently an easy way to represent a bidirectional
955 			 * transfer in devstat.  The only way to do it
956 			 * and have the byte counts come out right would
957 			 * mean that we would have to record two
958 			 * transactions, one for the request and one for the
959 			 * response.  For now, so that we report something,
960 			 * just treat the entire thing as a read.
961 			 */
962 			devstat_end_transaction(softc->device_stats,
963 			    done_ccb->smpio.smp_request_len +
964 			    done_ccb->smpio.smp_response_len,
965 			    DEVSTAT_TAG_SIMPLE, DEVSTAT_READ, NULL,
966 			    &io_req->start_time);
967 			break;
968 		default:
969 			devstat_end_transaction(softc->device_stats, 0,
970 			    DEVSTAT_TAG_NONE, DEVSTAT_NO_DATA, NULL,
971 			    &io_req->start_time);
972 			break;
973 		}
974 
975 		/*
976 		 * In the normal case, take the completed I/O off of the
977 		 * active queue and put it on the done queue.  Notitfy the
978 		 * user that we have a completed I/O.
979 		 */
980 		if ((io_req->flags & PASS_IO_ABANDONED) == 0) {
981 			TAILQ_REMOVE(&softc->active_queue, io_req, links);
982 			TAILQ_INSERT_TAIL(&softc->done_queue, io_req, links);
983 			selwakeuppri(&softc->read_select, PRIBIO);
984 			KNOTE_LOCKED(&softc->read_select.si_note, 0);
985 		} else {
986 			/*
987 			 * In the case of an abandoned I/O (final close
988 			 * without fetching the I/O), take it off of the
989 			 * abandoned queue and free it.
990 			 */
991 			TAILQ_REMOVE(&softc->abandoned_queue, io_req, links);
992 			passiocleanup(softc, io_req);
993 			uma_zfree(softc->pass_zone, io_req);
994 
995 			/*
996 			 * Release the done_ccb here, since we may wind up
997 			 * freeing the peripheral when we decrement the
998 			 * reference count below.
999 			 */
1000 			xpt_release_ccb(done_ccb);
1001 
1002 			/*
1003 			 * If the abandoned queue is empty, we can release
1004 			 * our reference to the periph since we won't have
1005 			 * any more completions coming.
1006 			 */
1007 			if ((TAILQ_EMPTY(&softc->abandoned_queue))
1008 			 && (softc->flags & PASS_FLAG_ABANDONED_REF_SET)) {
1009 				softc->flags &= ~PASS_FLAG_ABANDONED_REF_SET;
1010 				cam_periph_release_locked(periph);
1011 			}
1012 
1013 			/*
1014 			 * We have already released the CCB, so we can
1015 			 * return.
1016 			 */
1017 			return;
1018 		}
1019 		break;
1020 	}
1021 	}
1022 	xpt_release_ccb(done_ccb);
1023 }
1024 
1025 static int
1026 passcreatezone(struct cam_periph *periph)
1027 {
1028 	struct pass_softc *softc;
1029 	int error;
1030 
1031 	error = 0;
1032 	softc = (struct pass_softc *)periph->softc;
1033 
1034 	cam_periph_assert(periph, MA_OWNED);
1035 	KASSERT(((softc->flags & PASS_FLAG_ZONE_VALID) == 0),
1036 		("%s called when the pass(4) zone is valid!\n", __func__));
1037 	KASSERT((softc->pass_zone == NULL),
1038 		("%s called when the pass(4) zone is allocated!\n", __func__));
1039 
1040 	if ((softc->flags & PASS_FLAG_ZONE_INPROG) == 0) {
1041 
1042 		/*
1043 		 * We're the first context through, so we need to create
1044 		 * the pass(4) UMA zone for I/O requests.
1045 		 */
1046 		softc->flags |= PASS_FLAG_ZONE_INPROG;
1047 
1048 		/*
1049 		 * uma_zcreate() does a blocking (M_WAITOK) allocation,
1050 		 * so we cannot hold a mutex while we call it.
1051 		 */
1052 		cam_periph_unlock(periph);
1053 
1054 		softc->pass_zone = uma_zcreate(softc->zone_name,
1055 		    sizeof(struct pass_io_req), NULL, NULL, NULL, NULL,
1056 		    /*align*/ 0, /*flags*/ 0);
1057 
1058 		softc->pass_io_zone = uma_zcreate(softc->io_zone_name,
1059 		    softc->io_zone_size, NULL, NULL, NULL, NULL,
1060 		    /*align*/ 0, /*flags*/ 0);
1061 
1062 		cam_periph_lock(periph);
1063 
1064 		if ((softc->pass_zone == NULL)
1065 		 || (softc->pass_io_zone == NULL)) {
1066 			if (softc->pass_zone == NULL)
1067 				xpt_print(periph->path, "unable to allocate "
1068 				    "IO Req UMA zone\n");
1069 			else
1070 				xpt_print(periph->path, "unable to allocate "
1071 				    "IO UMA zone\n");
1072 			softc->flags &= ~PASS_FLAG_ZONE_INPROG;
1073 			goto bailout;
1074 		}
1075 
1076 		/*
1077 		 * Set the flags appropriately and notify any other waiters.
1078 		 */
1079 		softc->flags &= PASS_FLAG_ZONE_INPROG;
1080 		softc->flags |= PASS_FLAG_ZONE_VALID;
1081 		wakeup(&softc->pass_zone);
1082 	} else {
1083 		/*
1084 		 * In this case, the UMA zone has not yet been created, but
1085 		 * another context is in the process of creating it.  We
1086 		 * need to sleep until the creation is either done or has
1087 		 * failed.
1088 		 */
1089 		while ((softc->flags & PASS_FLAG_ZONE_INPROG)
1090 		    && ((softc->flags & PASS_FLAG_ZONE_VALID) == 0)) {
1091 			error = msleep(&softc->pass_zone,
1092 				       cam_periph_mtx(periph), PRIBIO,
1093 				       "paszon", 0);
1094 			if (error != 0)
1095 				goto bailout;
1096 		}
1097 		/*
1098 		 * If the zone creation failed, no luck for the user.
1099 		 */
1100 		if ((softc->flags & PASS_FLAG_ZONE_VALID) == 0){
1101 			error = ENOMEM;
1102 			goto bailout;
1103 		}
1104 	}
1105 bailout:
1106 	return (error);
1107 }
1108 
1109 static void
1110 passiocleanup(struct pass_softc *softc, struct pass_io_req *io_req)
1111 {
1112 	union ccb *ccb;
1113 	u_int8_t **data_ptrs[CAM_PERIPH_MAXMAPS];
1114 	int i, numbufs;
1115 
1116 	ccb = &io_req->ccb;
1117 
1118 	switch (ccb->ccb_h.func_code) {
1119 	case XPT_DEV_MATCH:
1120 		numbufs = min(io_req->num_bufs, 2);
1121 
1122 		if (numbufs == 1) {
1123 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.matches;
1124 		} else {
1125 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.patterns;
1126 			data_ptrs[1] = (u_int8_t **)&ccb->cdm.matches;
1127 		}
1128 		break;
1129 	case XPT_SCSI_IO:
1130 	case XPT_CONT_TARGET_IO:
1131 		data_ptrs[0] = &ccb->csio.data_ptr;
1132 		numbufs = min(io_req->num_bufs, 1);
1133 		break;
1134 	case XPT_ATA_IO:
1135 		data_ptrs[0] = &ccb->ataio.data_ptr;
1136 		numbufs = min(io_req->num_bufs, 1);
1137 		break;
1138 	case XPT_SMP_IO:
1139 		numbufs = min(io_req->num_bufs, 2);
1140 		data_ptrs[0] = &ccb->smpio.smp_request;
1141 		data_ptrs[1] = &ccb->smpio.smp_response;
1142 		break;
1143 	case XPT_DEV_ADVINFO:
1144 		numbufs = min(io_req->num_bufs, 1);
1145 		data_ptrs[0] = (uint8_t **)&ccb->cdai.buf;
1146 		break;
1147 	case XPT_NVME_IO:
1148 	case XPT_NVME_ADMIN:
1149 		data_ptrs[0] = &ccb->nvmeio.data_ptr;
1150 		numbufs = min(io_req->num_bufs, 1);
1151 		break;
1152 	default:
1153 		/* allow ourselves to be swapped once again */
1154 		return;
1155 		break; /* NOTREACHED */
1156 	}
1157 
1158 	if (io_req->flags & PASS_IO_USER_SEG_MALLOC) {
1159 		free(io_req->user_segptr, M_SCSIPASS);
1160 		io_req->user_segptr = NULL;
1161 	}
1162 
1163 	/*
1164 	 * We only want to free memory we malloced.
1165 	 */
1166 	if (io_req->data_flags == CAM_DATA_VADDR) {
1167 		for (i = 0; i < io_req->num_bufs; i++) {
1168 			if (io_req->kern_bufs[i] == NULL)
1169 				continue;
1170 
1171 			free(io_req->kern_bufs[i], M_SCSIPASS);
1172 			io_req->kern_bufs[i] = NULL;
1173 		}
1174 	} else if (io_req->data_flags == CAM_DATA_SG) {
1175 		for (i = 0; i < io_req->num_kern_segs; i++) {
1176 			if ((uint8_t *)(uintptr_t)
1177 			    io_req->kern_segptr[i].ds_addr == NULL)
1178 				continue;
1179 
1180 			uma_zfree(softc->pass_io_zone, (uint8_t *)(uintptr_t)
1181 			    io_req->kern_segptr[i].ds_addr);
1182 			io_req->kern_segptr[i].ds_addr = 0;
1183 		}
1184 	}
1185 
1186 	if (io_req->flags & PASS_IO_KERN_SEG_MALLOC) {
1187 		free(io_req->kern_segptr, M_SCSIPASS);
1188 		io_req->kern_segptr = NULL;
1189 	}
1190 
1191 	if (io_req->data_flags != CAM_DATA_PADDR) {
1192 		for (i = 0; i < numbufs; i++) {
1193 			/*
1194 			 * Restore the user's buffer pointers to their
1195 			 * previous values.
1196 			 */
1197 			if (io_req->user_bufs[i] != NULL)
1198 				*data_ptrs[i] = io_req->user_bufs[i];
1199 		}
1200 	}
1201 
1202 }
1203 
1204 static int
1205 passcopysglist(struct cam_periph *periph, struct pass_io_req *io_req,
1206 	       ccb_flags direction)
1207 {
1208 	bus_size_t kern_watermark, user_watermark, len_copied, len_to_copy;
1209 	bus_dma_segment_t *user_sglist, *kern_sglist;
1210 	int i, j, error;
1211 
1212 	error = 0;
1213 	kern_watermark = 0;
1214 	user_watermark = 0;
1215 	len_to_copy = 0;
1216 	len_copied = 0;
1217 	user_sglist = io_req->user_segptr;
1218 	kern_sglist = io_req->kern_segptr;
1219 
1220 	for (i = 0, j = 0; i < io_req->num_user_segs &&
1221 	     j < io_req->num_kern_segs;) {
1222 		uint8_t *user_ptr, *kern_ptr;
1223 
1224 		len_to_copy = min(user_sglist[i].ds_len -user_watermark,
1225 		    kern_sglist[j].ds_len - kern_watermark);
1226 
1227 		user_ptr = (uint8_t *)(uintptr_t)user_sglist[i].ds_addr;
1228 		user_ptr = user_ptr + user_watermark;
1229 		kern_ptr = (uint8_t *)(uintptr_t)kern_sglist[j].ds_addr;
1230 		kern_ptr = kern_ptr + kern_watermark;
1231 
1232 		user_watermark += len_to_copy;
1233 		kern_watermark += len_to_copy;
1234 
1235 		if (!useracc(user_ptr, len_to_copy,
1236 		    (direction == CAM_DIR_IN) ? VM_PROT_WRITE : VM_PROT_READ)) {
1237 			xpt_print(periph->path, "%s: unable to access user "
1238 				  "S/G list element %p len %zu\n", __func__,
1239 				  user_ptr, len_to_copy);
1240 			error = EFAULT;
1241 			goto bailout;
1242 		}
1243 
1244 		if (direction == CAM_DIR_IN) {
1245 			error = copyout(kern_ptr, user_ptr, len_to_copy);
1246 			if (error != 0) {
1247 				xpt_print(periph->path, "%s: copyout of %u "
1248 					  "bytes from %p to %p failed with "
1249 					  "error %d\n", __func__, len_to_copy,
1250 					  kern_ptr, user_ptr, error);
1251 				goto bailout;
1252 			}
1253 		} else {
1254 			error = copyin(user_ptr, kern_ptr, len_to_copy);
1255 			if (error != 0) {
1256 				xpt_print(periph->path, "%s: copyin of %u "
1257 					  "bytes from %p to %p failed with "
1258 					  "error %d\n", __func__, len_to_copy,
1259 					  user_ptr, kern_ptr, error);
1260 				goto bailout;
1261 			}
1262 		}
1263 
1264 		len_copied += len_to_copy;
1265 
1266 		if (user_sglist[i].ds_len == user_watermark) {
1267 			i++;
1268 			user_watermark = 0;
1269 		}
1270 
1271 		if (kern_sglist[j].ds_len == kern_watermark) {
1272 			j++;
1273 			kern_watermark = 0;
1274 		}
1275 	}
1276 
1277 bailout:
1278 
1279 	return (error);
1280 }
1281 
1282 static int
1283 passmemsetup(struct cam_periph *periph, struct pass_io_req *io_req)
1284 {
1285 	union ccb *ccb;
1286 	struct pass_softc *softc;
1287 	int numbufs, i;
1288 	uint8_t **data_ptrs[CAM_PERIPH_MAXMAPS];
1289 	uint32_t lengths[CAM_PERIPH_MAXMAPS];
1290 	uint32_t dirs[CAM_PERIPH_MAXMAPS];
1291 	uint32_t num_segs;
1292 	uint16_t *seg_cnt_ptr;
1293 	size_t maxmap;
1294 	int error;
1295 
1296 	cam_periph_assert(periph, MA_NOTOWNED);
1297 
1298 	softc = periph->softc;
1299 
1300 	error = 0;
1301 	ccb = &io_req->ccb;
1302 	maxmap = 0;
1303 	num_segs = 0;
1304 	seg_cnt_ptr = NULL;
1305 
1306 	switch(ccb->ccb_h.func_code) {
1307 	case XPT_DEV_MATCH:
1308 		if (ccb->cdm.match_buf_len == 0) {
1309 			printf("%s: invalid match buffer length 0\n", __func__);
1310 			return(EINVAL);
1311 		}
1312 		if (ccb->cdm.pattern_buf_len > 0) {
1313 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.patterns;
1314 			lengths[0] = ccb->cdm.pattern_buf_len;
1315 			dirs[0] = CAM_DIR_OUT;
1316 			data_ptrs[1] = (u_int8_t **)&ccb->cdm.matches;
1317 			lengths[1] = ccb->cdm.match_buf_len;
1318 			dirs[1] = CAM_DIR_IN;
1319 			numbufs = 2;
1320 		} else {
1321 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.matches;
1322 			lengths[0] = ccb->cdm.match_buf_len;
1323 			dirs[0] = CAM_DIR_IN;
1324 			numbufs = 1;
1325 		}
1326 		io_req->data_flags = CAM_DATA_VADDR;
1327 		break;
1328 	case XPT_SCSI_IO:
1329 	case XPT_CONT_TARGET_IO:
1330 		if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
1331 			return(0);
1332 
1333 		/*
1334 		 * The user shouldn't be able to supply a bio.
1335 		 */
1336 		if ((ccb->ccb_h.flags & CAM_DATA_MASK) == CAM_DATA_BIO)
1337 			return (EINVAL);
1338 
1339 		io_req->data_flags = ccb->ccb_h.flags & CAM_DATA_MASK;
1340 
1341 		data_ptrs[0] = &ccb->csio.data_ptr;
1342 		lengths[0] = ccb->csio.dxfer_len;
1343 		dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1344 		num_segs = ccb->csio.sglist_cnt;
1345 		seg_cnt_ptr = &ccb->csio.sglist_cnt;
1346 		numbufs = 1;
1347 		maxmap = softc->maxio;
1348 		break;
1349 	case XPT_ATA_IO:
1350 		if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
1351 			return(0);
1352 
1353 		/*
1354 		 * We only support a single virtual address for ATA I/O.
1355 		 */
1356 		if ((ccb->ccb_h.flags & CAM_DATA_MASK) != CAM_DATA_VADDR)
1357 			return (EINVAL);
1358 
1359 		io_req->data_flags = CAM_DATA_VADDR;
1360 
1361 		data_ptrs[0] = &ccb->ataio.data_ptr;
1362 		lengths[0] = ccb->ataio.dxfer_len;
1363 		dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1364 		numbufs = 1;
1365 		maxmap = softc->maxio;
1366 		break;
1367 	case XPT_SMP_IO:
1368 		io_req->data_flags = CAM_DATA_VADDR;
1369 
1370 		data_ptrs[0] = &ccb->smpio.smp_request;
1371 		lengths[0] = ccb->smpio.smp_request_len;
1372 		dirs[0] = CAM_DIR_OUT;
1373 		data_ptrs[1] = &ccb->smpio.smp_response;
1374 		lengths[1] = ccb->smpio.smp_response_len;
1375 		dirs[1] = CAM_DIR_IN;
1376 		numbufs = 2;
1377 		maxmap = softc->maxio;
1378 		break;
1379 	case XPT_DEV_ADVINFO:
1380 		if (ccb->cdai.bufsiz == 0)
1381 			return (0);
1382 
1383 		io_req->data_flags = CAM_DATA_VADDR;
1384 
1385 		data_ptrs[0] = (uint8_t **)&ccb->cdai.buf;
1386 		lengths[0] = ccb->cdai.bufsiz;
1387 		dirs[0] = CAM_DIR_IN;
1388 		numbufs = 1;
1389 		break;
1390 	case XPT_NVME_ADMIN:
1391 	case XPT_NVME_IO:
1392 		if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
1393 			return (0);
1394 
1395 		io_req->data_flags = ccb->ccb_h.flags & CAM_DATA_MASK;
1396 
1397 		data_ptrs[0] = &ccb->nvmeio.data_ptr;
1398 		lengths[0] = ccb->nvmeio.dxfer_len;
1399 		dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1400 		num_segs = ccb->nvmeio.sglist_cnt;
1401 		seg_cnt_ptr = &ccb->nvmeio.sglist_cnt;
1402 		numbufs = 1;
1403 		maxmap = softc->maxio;
1404 		break;
1405 	default:
1406 		return(EINVAL);
1407 		break; /* NOTREACHED */
1408 	}
1409 
1410 	io_req->num_bufs = numbufs;
1411 
1412 	/*
1413 	 * If there is a maximum, check to make sure that the user's
1414 	 * request fits within the limit.  In general, we should only have
1415 	 * a maximum length for requests that go to hardware.  Otherwise it
1416 	 * is whatever we're able to malloc.
1417 	 */
1418 	for (i = 0; i < numbufs; i++) {
1419 		io_req->user_bufs[i] = *data_ptrs[i];
1420 		io_req->dirs[i] = dirs[i];
1421 		io_req->lengths[i] = lengths[i];
1422 
1423 		if (maxmap == 0)
1424 			continue;
1425 
1426 		if (lengths[i] <= maxmap)
1427 			continue;
1428 
1429 		xpt_print(periph->path, "%s: data length %u > max allowed %u "
1430 			  "bytes\n", __func__, lengths[i], maxmap);
1431 		error = EINVAL;
1432 		goto bailout;
1433 	}
1434 
1435 	switch (io_req->data_flags) {
1436 	case CAM_DATA_VADDR:
1437 		/* Map or copy the buffer into kernel address space */
1438 		for (i = 0; i < numbufs; i++) {
1439 			uint8_t *tmp_buf;
1440 
1441 			/*
1442 			 * If for some reason no length is specified, we
1443 			 * don't need to allocate anything.
1444 			 */
1445 			if (io_req->lengths[i] == 0)
1446 				continue;
1447 
1448 			/*
1449 			 * Make sure that the user's buffer is accessible
1450 			 * to that process.
1451 			 */
1452 			if (!useracc(io_req->user_bufs[i], io_req->lengths[i],
1453 			    (io_req->dirs[i] == CAM_DIR_IN) ? VM_PROT_WRITE :
1454 			     VM_PROT_READ)) {
1455 				xpt_print(periph->path, "%s: user address %p "
1456 				    "length %u is not accessible\n", __func__,
1457 				    io_req->user_bufs[i], io_req->lengths[i]);
1458 				error = EFAULT;
1459 				goto bailout;
1460 			}
1461 
1462 			tmp_buf = malloc(lengths[i], M_SCSIPASS,
1463 					 M_WAITOK | M_ZERO);
1464 			io_req->kern_bufs[i] = tmp_buf;
1465 			*data_ptrs[i] = tmp_buf;
1466 
1467 #if 0
1468 			xpt_print(periph->path, "%s: malloced %p len %u, user "
1469 				  "buffer %p, operation: %s\n", __func__,
1470 				  tmp_buf, lengths[i], io_req->user_bufs[i],
1471 				  (dirs[i] == CAM_DIR_IN) ? "read" : "write");
1472 #endif
1473 			/*
1474 			 * We only need to copy in if the user is writing.
1475 			 */
1476 			if (dirs[i] != CAM_DIR_OUT)
1477 				continue;
1478 
1479 			error = copyin(io_req->user_bufs[i],
1480 				       io_req->kern_bufs[i], lengths[i]);
1481 			if (error != 0) {
1482 				xpt_print(periph->path, "%s: copy of user "
1483 					  "buffer from %p to %p failed with "
1484 					  "error %d\n", __func__,
1485 					  io_req->user_bufs[i],
1486 					  io_req->kern_bufs[i], error);
1487 				goto bailout;
1488 			}
1489 		}
1490 		break;
1491 	case CAM_DATA_PADDR:
1492 		/* Pass down the pointer as-is */
1493 		break;
1494 	case CAM_DATA_SG: {
1495 		size_t sg_length, size_to_go, alloc_size;
1496 		uint32_t num_segs_needed;
1497 
1498 		/*
1499 		 * Copy the user S/G list in, and then copy in the
1500 		 * individual segments.
1501 		 */
1502 		/*
1503 		 * We shouldn't see this, but check just in case.
1504 		 */
1505 		if (numbufs != 1) {
1506 			xpt_print(periph->path, "%s: cannot currently handle "
1507 				  "more than one S/G list per CCB\n", __func__);
1508 			error = EINVAL;
1509 			goto bailout;
1510 		}
1511 
1512 		/*
1513 		 * We have to have at least one segment.
1514 		 */
1515 		if (num_segs == 0) {
1516 			xpt_print(periph->path, "%s: CAM_DATA_SG flag set, "
1517 				  "but sglist_cnt=0!\n", __func__);
1518 			error = EINVAL;
1519 			goto bailout;
1520 		}
1521 
1522 		/*
1523 		 * Make sure the user specified the total length and didn't
1524 		 * just leave it to us to decode the S/G list.
1525 		 */
1526 		if (lengths[0] == 0) {
1527 			xpt_print(periph->path, "%s: no dxfer_len specified, "
1528 				  "but CAM_DATA_SG flag is set!\n", __func__);
1529 			error = EINVAL;
1530 			goto bailout;
1531 		}
1532 
1533 		/*
1534 		 * We allocate buffers in io_zone_size increments for an
1535 		 * S/G list.  This will generally be MAXPHYS.
1536 		 */
1537 		if (lengths[0] <= softc->io_zone_size)
1538 			num_segs_needed = 1;
1539 		else {
1540 			num_segs_needed = lengths[0] / softc->io_zone_size;
1541 			if ((lengths[0] % softc->io_zone_size) != 0)
1542 				num_segs_needed++;
1543 		}
1544 
1545 		/* Figure out the size of the S/G list */
1546 		sg_length = num_segs * sizeof(bus_dma_segment_t);
1547 		io_req->num_user_segs = num_segs;
1548 		io_req->num_kern_segs = num_segs_needed;
1549 
1550 		/* Save the user's S/G list pointer for later restoration */
1551 		io_req->user_bufs[0] = *data_ptrs[0];
1552 
1553 		/*
1554 		 * If we have enough segments allocated by default to handle
1555 		 * the length of the user's S/G list,
1556 		 */
1557 		if (num_segs > PASS_MAX_SEGS) {
1558 			io_req->user_segptr = malloc(sizeof(bus_dma_segment_t) *
1559 			    num_segs, M_SCSIPASS, M_WAITOK | M_ZERO);
1560 			io_req->flags |= PASS_IO_USER_SEG_MALLOC;
1561 		} else
1562 			io_req->user_segptr = io_req->user_segs;
1563 
1564 		if (!useracc(*data_ptrs[0], sg_length, VM_PROT_READ)) {
1565 			xpt_print(periph->path, "%s: unable to access user "
1566 				  "S/G list at %p\n", __func__, *data_ptrs[0]);
1567 			error = EFAULT;
1568 			goto bailout;
1569 		}
1570 
1571 		error = copyin(*data_ptrs[0], io_req->user_segptr, sg_length);
1572 		if (error != 0) {
1573 			xpt_print(periph->path, "%s: copy of user S/G list "
1574 				  "from %p to %p failed with error %d\n",
1575 				  __func__, *data_ptrs[0], io_req->user_segptr,
1576 				  error);
1577 			goto bailout;
1578 		}
1579 
1580 		if (num_segs_needed > PASS_MAX_SEGS) {
1581 			io_req->kern_segptr = malloc(sizeof(bus_dma_segment_t) *
1582 			    num_segs_needed, M_SCSIPASS, M_WAITOK | M_ZERO);
1583 			io_req->flags |= PASS_IO_KERN_SEG_MALLOC;
1584 		} else {
1585 			io_req->kern_segptr = io_req->kern_segs;
1586 		}
1587 
1588 		/*
1589 		 * Allocate the kernel S/G list.
1590 		 */
1591 		for (size_to_go = lengths[0], i = 0;
1592 		     size_to_go > 0 && i < num_segs_needed;
1593 		     i++, size_to_go -= alloc_size) {
1594 			uint8_t *kern_ptr;
1595 
1596 			alloc_size = min(size_to_go, softc->io_zone_size);
1597 			kern_ptr = uma_zalloc(softc->pass_io_zone, M_WAITOK);
1598 			io_req->kern_segptr[i].ds_addr =
1599 			    (bus_addr_t)(uintptr_t)kern_ptr;
1600 			io_req->kern_segptr[i].ds_len = alloc_size;
1601 		}
1602 		if (size_to_go > 0) {
1603 			printf("%s: size_to_go = %zu, software error!\n",
1604 			       __func__, size_to_go);
1605 			error = EINVAL;
1606 			goto bailout;
1607 		}
1608 
1609 		*data_ptrs[0] = (uint8_t *)io_req->kern_segptr;
1610 		*seg_cnt_ptr = io_req->num_kern_segs;
1611 
1612 		/*
1613 		 * We only need to copy data here if the user is writing.
1614 		 */
1615 		if (dirs[0] == CAM_DIR_OUT)
1616 			error = passcopysglist(periph, io_req, dirs[0]);
1617 		break;
1618 	}
1619 	case CAM_DATA_SG_PADDR: {
1620 		size_t sg_length;
1621 
1622 		/*
1623 		 * We shouldn't see this, but check just in case.
1624 		 */
1625 		if (numbufs != 1) {
1626 			printf("%s: cannot currently handle more than one "
1627 			       "S/G list per CCB\n", __func__);
1628 			error = EINVAL;
1629 			goto bailout;
1630 		}
1631 
1632 		/*
1633 		 * We have to have at least one segment.
1634 		 */
1635 		if (num_segs == 0) {
1636 			xpt_print(periph->path, "%s: CAM_DATA_SG_PADDR flag "
1637 				  "set, but sglist_cnt=0!\n", __func__);
1638 			error = EINVAL;
1639 			goto bailout;
1640 		}
1641 
1642 		/*
1643 		 * Make sure the user specified the total length and didn't
1644 		 * just leave it to us to decode the S/G list.
1645 		 */
1646 		if (lengths[0] == 0) {
1647 			xpt_print(periph->path, "%s: no dxfer_len specified, "
1648 				  "but CAM_DATA_SG flag is set!\n", __func__);
1649 			error = EINVAL;
1650 			goto bailout;
1651 		}
1652 
1653 		/* Figure out the size of the S/G list */
1654 		sg_length = num_segs * sizeof(bus_dma_segment_t);
1655 		io_req->num_user_segs = num_segs;
1656 		io_req->num_kern_segs = io_req->num_user_segs;
1657 
1658 		/* Save the user's S/G list pointer for later restoration */
1659 		io_req->user_bufs[0] = *data_ptrs[0];
1660 
1661 		if (num_segs > PASS_MAX_SEGS) {
1662 			io_req->user_segptr = malloc(sizeof(bus_dma_segment_t) *
1663 			    num_segs, M_SCSIPASS, M_WAITOK | M_ZERO);
1664 			io_req->flags |= PASS_IO_USER_SEG_MALLOC;
1665 		} else
1666 			io_req->user_segptr = io_req->user_segs;
1667 
1668 		io_req->kern_segptr = io_req->user_segptr;
1669 
1670 		error = copyin(*data_ptrs[0], io_req->user_segptr, sg_length);
1671 		if (error != 0) {
1672 			xpt_print(periph->path, "%s: copy of user S/G list "
1673 				  "from %p to %p failed with error %d\n",
1674 				  __func__, *data_ptrs[0], io_req->user_segptr,
1675 				  error);
1676 			goto bailout;
1677 		}
1678 		break;
1679 	}
1680 	default:
1681 	case CAM_DATA_BIO:
1682 		/*
1683 		 * A user shouldn't be attaching a bio to the CCB.  It
1684 		 * isn't a user-accessible structure.
1685 		 */
1686 		error = EINVAL;
1687 		break;
1688 	}
1689 
1690 bailout:
1691 	if (error != 0)
1692 		passiocleanup(softc, io_req);
1693 
1694 	return (error);
1695 }
1696 
1697 static int
1698 passmemdone(struct cam_periph *periph, struct pass_io_req *io_req)
1699 {
1700 	struct pass_softc *softc;
1701 	int error;
1702 	int i;
1703 
1704 	error = 0;
1705 	softc = (struct pass_softc *)periph->softc;
1706 
1707 	switch (io_req->data_flags) {
1708 	case CAM_DATA_VADDR:
1709 		/*
1710 		 * Copy back to the user buffer if this was a read.
1711 		 */
1712 		for (i = 0; i < io_req->num_bufs; i++) {
1713 			if (io_req->dirs[i] != CAM_DIR_IN)
1714 				continue;
1715 
1716 			error = copyout(io_req->kern_bufs[i],
1717 			    io_req->user_bufs[i], io_req->lengths[i]);
1718 			if (error != 0) {
1719 				xpt_print(periph->path, "Unable to copy %u "
1720 					  "bytes from %p to user address %p\n",
1721 					  io_req->lengths[i],
1722 					  io_req->kern_bufs[i],
1723 					  io_req->user_bufs[i]);
1724 				goto bailout;
1725 			}
1726 
1727 		}
1728 		break;
1729 	case CAM_DATA_PADDR:
1730 		/* Do nothing.  The pointer is a physical address already */
1731 		break;
1732 	case CAM_DATA_SG:
1733 		/*
1734 		 * Copy back to the user buffer if this was a read.
1735 		 * Restore the user's S/G list buffer pointer.
1736 		 */
1737 		if (io_req->dirs[0] == CAM_DIR_IN)
1738 			error = passcopysglist(periph, io_req, io_req->dirs[0]);
1739 		break;
1740 	case CAM_DATA_SG_PADDR:
1741 		/*
1742 		 * Restore the user's S/G list buffer pointer.  No need to
1743 		 * copy.
1744 		 */
1745 		break;
1746 	default:
1747 	case CAM_DATA_BIO:
1748 		error = EINVAL;
1749 		break;
1750 	}
1751 
1752 bailout:
1753 	/*
1754 	 * Reset the user's pointers to their original values and free
1755 	 * allocated memory.
1756 	 */
1757 	passiocleanup(softc, io_req);
1758 
1759 	return (error);
1760 }
1761 
1762 static int
1763 passioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flag, struct thread *td)
1764 {
1765 	int error;
1766 
1767 	if ((error = passdoioctl(dev, cmd, addr, flag, td)) == ENOTTY) {
1768 		error = cam_compat_ioctl(dev, cmd, addr, flag, td, passdoioctl);
1769 	}
1770 	return (error);
1771 }
1772 
1773 static int
1774 passdoioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flag, struct thread *td)
1775 {
1776 	struct	cam_periph *periph;
1777 	struct	pass_softc *softc;
1778 	int	error;
1779 	uint32_t priority;
1780 
1781 	periph = (struct cam_periph *)dev->si_drv1;
1782 	cam_periph_lock(periph);
1783 	softc = (struct pass_softc *)periph->softc;
1784 
1785 	error = 0;
1786 
1787 	switch (cmd) {
1788 
1789 	case CAMIOCOMMAND:
1790 	{
1791 		union ccb *inccb;
1792 		union ccb *ccb;
1793 		int ccb_malloced;
1794 
1795 		inccb = (union ccb *)addr;
1796 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
1797 		if (inccb->ccb_h.func_code == XPT_SCSI_IO)
1798 			inccb->csio.bio = NULL;
1799 #endif
1800 
1801 		if (inccb->ccb_h.flags & CAM_UNLOCKED) {
1802 			error = EINVAL;
1803 			break;
1804 		}
1805 
1806 		/*
1807 		 * Some CCB types, like scan bus and scan lun can only go
1808 		 * through the transport layer device.
1809 		 */
1810 		if (inccb->ccb_h.func_code & XPT_FC_XPT_ONLY) {
1811 			xpt_print(periph->path, "CCB function code %#x is "
1812 			    "restricted to the XPT device\n",
1813 			    inccb->ccb_h.func_code);
1814 			error = ENODEV;
1815 			break;
1816 		}
1817 
1818 		/* Compatibility for RL/priority-unaware code. */
1819 		priority = inccb->ccb_h.pinfo.priority;
1820 		if (priority <= CAM_PRIORITY_OOB)
1821 		    priority += CAM_PRIORITY_OOB + 1;
1822 
1823 		/*
1824 		 * Non-immediate CCBs need a CCB from the per-device pool
1825 		 * of CCBs, which is scheduled by the transport layer.
1826 		 * Immediate CCBs and user-supplied CCBs should just be
1827 		 * malloced.
1828 		 */
1829 		if ((inccb->ccb_h.func_code & XPT_FC_QUEUED)
1830 		 && ((inccb->ccb_h.func_code & XPT_FC_USER_CCB) == 0)) {
1831 			ccb = cam_periph_getccb(periph, priority);
1832 			ccb_malloced = 0;
1833 		} else {
1834 			ccb = xpt_alloc_ccb_nowait();
1835 
1836 			if (ccb != NULL)
1837 				xpt_setup_ccb(&ccb->ccb_h, periph->path,
1838 					      priority);
1839 			ccb_malloced = 1;
1840 		}
1841 
1842 		if (ccb == NULL) {
1843 			xpt_print(periph->path, "unable to allocate CCB\n");
1844 			error = ENOMEM;
1845 			break;
1846 		}
1847 
1848 		error = passsendccb(periph, ccb, inccb);
1849 
1850 		if (ccb_malloced)
1851 			xpt_free_ccb(ccb);
1852 		else
1853 			xpt_release_ccb(ccb);
1854 
1855 		break;
1856 	}
1857 	case CAMIOQUEUE:
1858 	{
1859 		struct pass_io_req *io_req;
1860 		union ccb **user_ccb, *ccb;
1861 		xpt_opcode fc;
1862 
1863 #ifdef COMPAT_FREEBSD32
1864 		if (SV_PROC_FLAG(td->td_proc, SV_ILP32)) {
1865 			error = ENOTTY;
1866 			goto bailout;
1867 		}
1868 #endif
1869 		if ((softc->flags & PASS_FLAG_ZONE_VALID) == 0) {
1870 			error = passcreatezone(periph);
1871 			if (error != 0)
1872 				goto bailout;
1873 		}
1874 
1875 		/*
1876 		 * We're going to do a blocking allocation for this I/O
1877 		 * request, so we have to drop the lock.
1878 		 */
1879 		cam_periph_unlock(periph);
1880 
1881 		io_req = uma_zalloc(softc->pass_zone, M_WAITOK | M_ZERO);
1882 		ccb = &io_req->ccb;
1883 		user_ccb = (union ccb **)addr;
1884 
1885 		/*
1886 		 * Unlike the CAMIOCOMMAND ioctl above, we only have a
1887 		 * pointer to the user's CCB, so we have to copy the whole
1888 		 * thing in to a buffer we have allocated (above) instead
1889 		 * of allowing the ioctl code to malloc a buffer and copy
1890 		 * it in.
1891 		 *
1892 		 * This is an advantage for this asynchronous interface,
1893 		 * since we don't want the memory to get freed while the
1894 		 * CCB is outstanding.
1895 		 */
1896 #if 0
1897 		xpt_print(periph->path, "Copying user CCB %p to "
1898 			  "kernel address %p\n", *user_ccb, ccb);
1899 #endif
1900 		error = copyin(*user_ccb, ccb, sizeof(*ccb));
1901 		if (error != 0) {
1902 			xpt_print(periph->path, "Copy of user CCB %p to "
1903 				  "kernel address %p failed with error %d\n",
1904 				  *user_ccb, ccb, error);
1905 			goto camioqueue_error;
1906 		}
1907 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
1908 		if (ccb->ccb_h.func_code == XPT_SCSI_IO)
1909 			ccb->csio.bio = NULL;
1910 #endif
1911 
1912 		if (ccb->ccb_h.flags & CAM_UNLOCKED) {
1913 			error = EINVAL;
1914 			goto camioqueue_error;
1915 		}
1916 
1917 		if (ccb->ccb_h.flags & CAM_CDB_POINTER) {
1918 			if (ccb->csio.cdb_len > IOCDBLEN) {
1919 				error = EINVAL;
1920 				goto camioqueue_error;
1921 			}
1922 			error = copyin(ccb->csio.cdb_io.cdb_ptr,
1923 			    ccb->csio.cdb_io.cdb_bytes, ccb->csio.cdb_len);
1924 			if (error != 0)
1925 				goto camioqueue_error;
1926 			ccb->ccb_h.flags &= ~CAM_CDB_POINTER;
1927 		}
1928 
1929 		/*
1930 		 * Some CCB types, like scan bus and scan lun can only go
1931 		 * through the transport layer device.
1932 		 */
1933 		if (ccb->ccb_h.func_code & XPT_FC_XPT_ONLY) {
1934 			xpt_print(periph->path, "CCB function code %#x is "
1935 			    "restricted to the XPT device\n",
1936 			    ccb->ccb_h.func_code);
1937 			error = ENODEV;
1938 			goto camioqueue_error;
1939 		}
1940 
1941 		/*
1942 		 * Save the user's CCB pointer as well as his linked list
1943 		 * pointers and peripheral private area so that we can
1944 		 * restore these later.
1945 		 */
1946 		io_req->user_ccb_ptr = *user_ccb;
1947 		io_req->user_periph_links = ccb->ccb_h.periph_links;
1948 		io_req->user_periph_priv = ccb->ccb_h.periph_priv;
1949 
1950 		/*
1951 		 * Now that we've saved the user's values, we can set our
1952 		 * own peripheral private entry.
1953 		 */
1954 		ccb->ccb_h.ccb_ioreq = io_req;
1955 
1956 		/* Compatibility for RL/priority-unaware code. */
1957 		priority = ccb->ccb_h.pinfo.priority;
1958 		if (priority <= CAM_PRIORITY_OOB)
1959 		    priority += CAM_PRIORITY_OOB + 1;
1960 
1961 		/*
1962 		 * Setup fields in the CCB like the path and the priority.
1963 		 * The path in particular cannot be done in userland, since
1964 		 * it is a pointer to a kernel data structure.
1965 		 */
1966 		xpt_setup_ccb_flags(&ccb->ccb_h, periph->path, priority,
1967 				    ccb->ccb_h.flags);
1968 
1969 		/*
1970 		 * Setup our done routine.  There is no way for the user to
1971 		 * have a valid pointer here.
1972 		 */
1973 		ccb->ccb_h.cbfcnp = passdone;
1974 
1975 		fc = ccb->ccb_h.func_code;
1976 		/*
1977 		 * If this function code has memory that can be mapped in
1978 		 * or out, we need to call passmemsetup().
1979 		 */
1980 		if ((fc == XPT_SCSI_IO) || (fc == XPT_ATA_IO)
1981 		 || (fc == XPT_SMP_IO) || (fc == XPT_DEV_MATCH)
1982 		 || (fc == XPT_DEV_ADVINFO)
1983 		 || (fc == XPT_NVME_ADMIN) || (fc == XPT_NVME_IO)) {
1984 			error = passmemsetup(periph, io_req);
1985 			if (error != 0)
1986 				goto camioqueue_error;
1987 		} else
1988 			io_req->mapinfo.num_bufs_used = 0;
1989 
1990 		cam_periph_lock(periph);
1991 
1992 		/*
1993 		 * Everything goes on the incoming queue initially.
1994 		 */
1995 		TAILQ_INSERT_TAIL(&softc->incoming_queue, io_req, links);
1996 
1997 		/*
1998 		 * If the CCB is queued, and is not a user CCB, then
1999 		 * we need to allocate a slot for it.  Call xpt_schedule()
2000 		 * so that our start routine will get called when a CCB is
2001 		 * available.
2002 		 */
2003 		if ((fc & XPT_FC_QUEUED)
2004 		 && ((fc & XPT_FC_USER_CCB) == 0)) {
2005 			xpt_schedule(periph, priority);
2006 			break;
2007 		}
2008 
2009 		/*
2010 		 * At this point, the CCB in question is either an
2011 		 * immediate CCB (like XPT_DEV_ADVINFO) or it is a user CCB
2012 		 * and therefore should be malloced, not allocated via a slot.
2013 		 * Remove the CCB from the incoming queue and add it to the
2014 		 * active queue.
2015 		 */
2016 		TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
2017 		TAILQ_INSERT_TAIL(&softc->active_queue, io_req, links);
2018 
2019 		xpt_action(ccb);
2020 
2021 		/*
2022 		 * If this is not a queued CCB (i.e. it is an immediate CCB),
2023 		 * then it is already done.  We need to put it on the done
2024 		 * queue for the user to fetch.
2025 		 */
2026 		if ((fc & XPT_FC_QUEUED) == 0) {
2027 			TAILQ_REMOVE(&softc->active_queue, io_req, links);
2028 			TAILQ_INSERT_TAIL(&softc->done_queue, io_req, links);
2029 		}
2030 		break;
2031 
2032 camioqueue_error:
2033 		uma_zfree(softc->pass_zone, io_req);
2034 		cam_periph_lock(periph);
2035 		break;
2036 	}
2037 	case CAMIOGET:
2038 	{
2039 		union ccb **user_ccb;
2040 		struct pass_io_req *io_req;
2041 		int old_error;
2042 
2043 #ifdef COMPAT_FREEBSD32
2044 		if (SV_PROC_FLAG(td->td_proc, SV_ILP32)) {
2045 			error = ENOTTY;
2046 			goto bailout;
2047 		}
2048 #endif
2049 		user_ccb = (union ccb **)addr;
2050 		old_error = 0;
2051 
2052 		io_req = TAILQ_FIRST(&softc->done_queue);
2053 		if (io_req == NULL) {
2054 			error = ENOENT;
2055 			break;
2056 		}
2057 
2058 		/*
2059 		 * Remove the I/O from the done queue.
2060 		 */
2061 		TAILQ_REMOVE(&softc->done_queue, io_req, links);
2062 
2063 		/*
2064 		 * We have to drop the lock during the copyout because the
2065 		 * copyout can result in VM faults that require sleeping.
2066 		 */
2067 		cam_periph_unlock(periph);
2068 
2069 		/*
2070 		 * Do any needed copies (e.g. for reads) and revert the
2071 		 * pointers in the CCB back to the user's pointers.
2072 		 */
2073 		error = passmemdone(periph, io_req);
2074 
2075 		old_error = error;
2076 
2077 		io_req->ccb.ccb_h.periph_links = io_req->user_periph_links;
2078 		io_req->ccb.ccb_h.periph_priv = io_req->user_periph_priv;
2079 
2080 #if 0
2081 		xpt_print(periph->path, "Copying to user CCB %p from "
2082 			  "kernel address %p\n", *user_ccb, &io_req->ccb);
2083 #endif
2084 
2085 		error = copyout(&io_req->ccb, *user_ccb, sizeof(union ccb));
2086 		if (error != 0) {
2087 			xpt_print(periph->path, "Copy to user CCB %p from "
2088 				  "kernel address %p failed with error %d\n",
2089 				  *user_ccb, &io_req->ccb, error);
2090 		}
2091 
2092 		/*
2093 		 * Prefer the first error we got back, and make sure we
2094 		 * don't overwrite bad status with good.
2095 		 */
2096 		if (old_error != 0)
2097 			error = old_error;
2098 
2099 		cam_periph_lock(periph);
2100 
2101 		/*
2102 		 * At this point, if there was an error, we could potentially
2103 		 * re-queue the I/O and try again.  But why?  The error
2104 		 * would almost certainly happen again.  We might as well
2105 		 * not leak memory.
2106 		 */
2107 		uma_zfree(softc->pass_zone, io_req);
2108 		break;
2109 	}
2110 	default:
2111 		error = cam_periph_ioctl(periph, cmd, addr, passerror);
2112 		break;
2113 	}
2114 
2115 bailout:
2116 	cam_periph_unlock(periph);
2117 
2118 	return(error);
2119 }
2120 
2121 static int
2122 passpoll(struct cdev *dev, int poll_events, struct thread *td)
2123 {
2124 	struct cam_periph *periph;
2125 	struct pass_softc *softc;
2126 	int revents;
2127 
2128 	periph = (struct cam_periph *)dev->si_drv1;
2129 	softc = (struct pass_softc *)periph->softc;
2130 
2131 	revents = poll_events & (POLLOUT | POLLWRNORM);
2132 	if ((poll_events & (POLLIN | POLLRDNORM)) != 0) {
2133 		cam_periph_lock(periph);
2134 
2135 		if (!TAILQ_EMPTY(&softc->done_queue)) {
2136 			revents |= poll_events & (POLLIN | POLLRDNORM);
2137 		}
2138 		cam_periph_unlock(periph);
2139 		if (revents == 0)
2140 			selrecord(td, &softc->read_select);
2141 	}
2142 
2143 	return (revents);
2144 }
2145 
2146 static int
2147 passkqfilter(struct cdev *dev, struct knote *kn)
2148 {
2149 	struct cam_periph *periph;
2150 	struct pass_softc *softc;
2151 
2152 	periph = (struct cam_periph *)dev->si_drv1;
2153 	softc = (struct pass_softc *)periph->softc;
2154 
2155 	kn->kn_hook = (caddr_t)periph;
2156 	kn->kn_fop = &passread_filtops;
2157 	knlist_add(&softc->read_select.si_note, kn, 0);
2158 
2159 	return (0);
2160 }
2161 
2162 static void
2163 passreadfiltdetach(struct knote *kn)
2164 {
2165 	struct cam_periph *periph;
2166 	struct pass_softc *softc;
2167 
2168 	periph = (struct cam_periph *)kn->kn_hook;
2169 	softc = (struct pass_softc *)periph->softc;
2170 
2171 	knlist_remove(&softc->read_select.si_note, kn, 0);
2172 }
2173 
2174 static int
2175 passreadfilt(struct knote *kn, long hint)
2176 {
2177 	struct cam_periph *periph;
2178 	struct pass_softc *softc;
2179 	int retval;
2180 
2181 	periph = (struct cam_periph *)kn->kn_hook;
2182 	softc = (struct pass_softc *)periph->softc;
2183 
2184 	cam_periph_assert(periph, MA_OWNED);
2185 
2186 	if (TAILQ_EMPTY(&softc->done_queue))
2187 		retval = 0;
2188 	else
2189 		retval = 1;
2190 
2191 	return (retval);
2192 }
2193 
2194 /*
2195  * Generally, "ccb" should be the CCB supplied by the kernel.  "inccb"
2196  * should be the CCB that is copied in from the user.
2197  */
2198 static int
2199 passsendccb(struct cam_periph *periph, union ccb *ccb, union ccb *inccb)
2200 {
2201 	struct pass_softc *softc;
2202 	struct cam_periph_map_info mapinfo;
2203 	uint8_t *cmd;
2204 	xpt_opcode fc;
2205 	int error;
2206 
2207 	softc = (struct pass_softc *)periph->softc;
2208 
2209 	/*
2210 	 * There are some fields in the CCB header that need to be
2211 	 * preserved, the rest we get from the user.
2212 	 */
2213 	xpt_merge_ccb(ccb, inccb);
2214 
2215 	if (ccb->ccb_h.flags & CAM_CDB_POINTER) {
2216 		cmd = __builtin_alloca(ccb->csio.cdb_len);
2217 		error = copyin(ccb->csio.cdb_io.cdb_ptr, cmd, ccb->csio.cdb_len);
2218 		if (error)
2219 			return (error);
2220 		ccb->csio.cdb_io.cdb_ptr = cmd;
2221 	}
2222 
2223 	/*
2224 	 * Let cam_periph_mapmem do a sanity check on the data pointer format.
2225 	 * Even if no data transfer is needed, it's a cheap check and it
2226 	 * simplifies the code.
2227 	 */
2228 	fc = ccb->ccb_h.func_code;
2229 	if ((fc == XPT_SCSI_IO) || (fc == XPT_ATA_IO) || (fc == XPT_SMP_IO)
2230             || (fc == XPT_DEV_MATCH) || (fc == XPT_DEV_ADVINFO) || (fc == XPT_MMC_IO)
2231             || (fc == XPT_NVME_ADMIN) || (fc == XPT_NVME_IO)) {
2232 
2233 		bzero(&mapinfo, sizeof(mapinfo));
2234 
2235 		/*
2236 		 * cam_periph_mapmem calls into proc and vm functions that can
2237 		 * sleep as well as trigger I/O, so we can't hold the lock.
2238 		 * Dropping it here is reasonably safe.
2239 		 */
2240 		cam_periph_unlock(periph);
2241 		error = cam_periph_mapmem(ccb, &mapinfo, softc->maxio);
2242 		cam_periph_lock(periph);
2243 
2244 		/*
2245 		 * cam_periph_mapmem returned an error, we can't continue.
2246 		 * Return the error to the user.
2247 		 */
2248 		if (error)
2249 			return(error);
2250 	} else
2251 		/* Ensure that the unmap call later on is a no-op. */
2252 		mapinfo.num_bufs_used = 0;
2253 
2254 	/*
2255 	 * If the user wants us to perform any error recovery, then honor
2256 	 * that request.  Otherwise, it's up to the user to perform any
2257 	 * error recovery.
2258 	 */
2259 	cam_periph_runccb(ccb, (ccb->ccb_h.flags & CAM_PASS_ERR_RECOVER) ?
2260 	    passerror : NULL, /* cam_flags */ CAM_RETRY_SELTO,
2261 	    /* sense_flags */ SF_RETRY_UA | SF_NO_PRINT,
2262 	    softc->device_stats);
2263 
2264 	cam_periph_unmapmem(ccb, &mapinfo);
2265 
2266 	ccb->ccb_h.cbfcnp = NULL;
2267 	ccb->ccb_h.periph_priv = inccb->ccb_h.periph_priv;
2268 	bcopy(ccb, inccb, sizeof(union ccb));
2269 
2270 	return(0);
2271 }
2272 
2273 static int
2274 passerror(union ccb *ccb, u_int32_t cam_flags, u_int32_t sense_flags)
2275 {
2276 	struct cam_periph *periph;
2277 	struct pass_softc *softc;
2278 
2279 	periph = xpt_path_periph(ccb->ccb_h.path);
2280 	softc = (struct pass_softc *)periph->softc;
2281 
2282 	return(cam_periph_error(ccb, cam_flags, sense_flags));
2283 }
2284