xref: /freebsd/sys/cam/scsi/scsi_pass.c (revision 06e20d1babecec1f45ffda513f55a8db5f1c0f56)
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 	args.mda_flags = MAKEDEV_NOWAIT;
656 	error = make_dev_s(&args, &softc->dev, "%s%d", periph->periph_name,
657 	    periph->unit_number);
658 	if (error != 0) {
659 		cam_periph_lock(periph);
660 		cam_periph_release_locked(periph);
661 		return (CAM_REQ_CMP_ERR);
662 	}
663 
664 	/*
665 	 * Hold a reference to the periph before we create the physical
666 	 * path alias so it can't go away.
667 	 */
668 	if (cam_periph_acquire(periph) != 0) {
669 		xpt_print(periph->path, "%s: lost periph during "
670 			  "registration!\n", __func__);
671 		cam_periph_lock(periph);
672 		return (CAM_REQ_CMP_ERR);
673 	}
674 
675 	cam_periph_lock(periph);
676 
677 	TASK_INIT(&softc->add_physpath_task, /*priority*/0,
678 		  pass_add_physpath, periph);
679 
680 	/*
681 	 * See if physical path information is already available.
682 	 */
683 	taskqueue_enqueue(taskqueue_thread, &softc->add_physpath_task);
684 
685 	/*
686 	 * Add an async callback so that we get notified if
687 	 * this device goes away or its physical path
688 	 * (stored in the advanced info data of the EDT) has
689 	 * changed.
690 	 */
691 	xpt_register_async(AC_LOST_DEVICE | AC_ADVINFO_CHANGED,
692 			   passasync, periph, periph->path);
693 
694 	if (bootverbose)
695 		xpt_announce_periph(periph, NULL);
696 
697 	return(CAM_REQ_CMP);
698 }
699 
700 static int
701 passopen(struct cdev *dev, int flags, int fmt, struct thread *td)
702 {
703 	struct cam_periph *periph;
704 	struct pass_softc *softc;
705 	int error;
706 
707 	periph = (struct cam_periph *)dev->si_drv1;
708 	if (cam_periph_acquire(periph) != 0)
709 		return (ENXIO);
710 
711 	cam_periph_lock(periph);
712 
713 	softc = (struct pass_softc *)periph->softc;
714 
715 	if (softc->flags & PASS_FLAG_INVALID) {
716 		cam_periph_release_locked(periph);
717 		cam_periph_unlock(periph);
718 		return(ENXIO);
719 	}
720 
721 	/*
722 	 * Don't allow access when we're running at a high securelevel.
723 	 */
724 	error = securelevel_gt(td->td_ucred, 1);
725 	if (error) {
726 		cam_periph_release_locked(periph);
727 		cam_periph_unlock(periph);
728 		return(error);
729 	}
730 
731 	/*
732 	 * Only allow read-write access.
733 	 */
734 	if (((flags & FWRITE) == 0) || ((flags & FREAD) == 0)) {
735 		cam_periph_release_locked(periph);
736 		cam_periph_unlock(periph);
737 		return(EPERM);
738 	}
739 
740 	/*
741 	 * We don't allow nonblocking access.
742 	 */
743 	if ((flags & O_NONBLOCK) != 0) {
744 		xpt_print(periph->path, "can't do nonblocking access\n");
745 		cam_periph_release_locked(periph);
746 		cam_periph_unlock(periph);
747 		return(EINVAL);
748 	}
749 
750 	softc->open_count++;
751 
752 	cam_periph_unlock(periph);
753 
754 	return (error);
755 }
756 
757 static int
758 passclose(struct cdev *dev, int flag, int fmt, struct thread *td)
759 {
760 	struct 	cam_periph *periph;
761 	struct  pass_softc *softc;
762 	struct mtx *mtx;
763 
764 	periph = (struct cam_periph *)dev->si_drv1;
765 	mtx = cam_periph_mtx(periph);
766 	mtx_lock(mtx);
767 
768 	softc = periph->softc;
769 	softc->open_count--;
770 
771 	if (softc->open_count == 0) {
772 		struct pass_io_req *io_req, *io_req2;
773 
774 		TAILQ_FOREACH_SAFE(io_req, &softc->done_queue, links, io_req2) {
775 			TAILQ_REMOVE(&softc->done_queue, io_req, links);
776 			passiocleanup(softc, io_req);
777 			uma_zfree(softc->pass_zone, io_req);
778 		}
779 
780 		TAILQ_FOREACH_SAFE(io_req, &softc->incoming_queue, links,
781 				   io_req2) {
782 			TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
783 			passiocleanup(softc, io_req);
784 			uma_zfree(softc->pass_zone, io_req);
785 		}
786 
787 		/*
788 		 * If there are any active I/Os, we need to forcibly acquire a
789 		 * reference to the peripheral so that we don't go away
790 		 * before they complete.  We'll release the reference when
791 		 * the abandoned queue is empty.
792 		 */
793 		io_req = TAILQ_FIRST(&softc->active_queue);
794 		if ((io_req != NULL)
795 		 && (softc->flags & PASS_FLAG_ABANDONED_REF_SET) == 0) {
796 			cam_periph_doacquire(periph);
797 			softc->flags |= PASS_FLAG_ABANDONED_REF_SET;
798 		}
799 
800 		/*
801 		 * Since the I/O in the active queue is not under our
802 		 * control, just set a flag so that we can clean it up when
803 		 * it completes and put it on the abandoned queue.  This
804 		 * will prevent our sending spurious completions in the
805 		 * event that the device is opened again before these I/Os
806 		 * complete.
807 		 */
808 		TAILQ_FOREACH_SAFE(io_req, &softc->active_queue, links,
809 				   io_req2) {
810 			TAILQ_REMOVE(&softc->active_queue, io_req, links);
811 			io_req->flags |= PASS_IO_ABANDONED;
812 			TAILQ_INSERT_TAIL(&softc->abandoned_queue, io_req,
813 					  links);
814 		}
815 	}
816 
817 	cam_periph_release_locked(periph);
818 
819 	/*
820 	 * We reference the lock directly here, instead of using
821 	 * cam_periph_unlock().  The reason is that the call to
822 	 * cam_periph_release_locked() above could result in the periph
823 	 * getting freed.  If that is the case, dereferencing the periph
824 	 * with a cam_periph_unlock() call would cause a page fault.
825 	 *
826 	 * cam_periph_release() avoids this problem using the same method,
827 	 * but we're manually acquiring and dropping the lock here to
828 	 * protect the open count and avoid another lock acquisition and
829 	 * release.
830 	 */
831 	mtx_unlock(mtx);
832 
833 	return (0);
834 }
835 
836 
837 static void
838 passstart(struct cam_periph *periph, union ccb *start_ccb)
839 {
840 	struct pass_softc *softc;
841 
842 	softc = (struct pass_softc *)periph->softc;
843 
844 	switch (softc->state) {
845 	case PASS_STATE_NORMAL: {
846 		struct pass_io_req *io_req;
847 
848 		/*
849 		 * Check for any queued I/O requests that require an
850 		 * allocated slot.
851 		 */
852 		io_req = TAILQ_FIRST(&softc->incoming_queue);
853 		if (io_req == NULL) {
854 			xpt_release_ccb(start_ccb);
855 			break;
856 		}
857 		TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
858 		TAILQ_INSERT_TAIL(&softc->active_queue, io_req, links);
859 		/*
860 		 * Merge the user's CCB into the allocated CCB.
861 		 */
862 		xpt_merge_ccb(start_ccb, &io_req->ccb);
863 		start_ccb->ccb_h.ccb_type = PASS_CCB_QUEUED_IO;
864 		start_ccb->ccb_h.ccb_ioreq = io_req;
865 		start_ccb->ccb_h.cbfcnp = passdone;
866 		io_req->alloced_ccb = start_ccb;
867 		binuptime(&io_req->start_time);
868 		devstat_start_transaction(softc->device_stats,
869 					  &io_req->start_time);
870 
871 		xpt_action(start_ccb);
872 
873 		/*
874 		 * If we have any more I/O waiting, schedule ourselves again.
875 		 */
876 		if (!TAILQ_EMPTY(&softc->incoming_queue))
877 			xpt_schedule(periph, CAM_PRIORITY_NORMAL);
878 		break;
879 	}
880 	default:
881 		break;
882 	}
883 }
884 
885 static void
886 passdone(struct cam_periph *periph, union ccb *done_ccb)
887 {
888 	struct pass_softc *softc;
889 	struct ccb_scsiio *csio;
890 
891 	softc = (struct pass_softc *)periph->softc;
892 
893 	cam_periph_assert(periph, MA_OWNED);
894 
895 	csio = &done_ccb->csio;
896 	switch (csio->ccb_h.ccb_type) {
897 	case PASS_CCB_QUEUED_IO: {
898 		struct pass_io_req *io_req;
899 
900 		io_req = done_ccb->ccb_h.ccb_ioreq;
901 #if 0
902 		xpt_print(periph->path, "%s: called for user CCB %p\n",
903 			  __func__, io_req->user_ccb_ptr);
904 #endif
905 		if (((done_ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_CMP)
906 		 && (done_ccb->ccb_h.flags & CAM_PASS_ERR_RECOVER)
907 		 && ((io_req->flags & PASS_IO_ABANDONED) == 0)) {
908 			int error;
909 
910 			error = passerror(done_ccb, CAM_RETRY_SELTO,
911 					  SF_RETRY_UA | SF_NO_PRINT);
912 
913 			if (error == ERESTART) {
914 				/*
915 				 * A retry was scheduled, so
916  				 * just return.
917 				 */
918 				return;
919 			}
920 		}
921 
922 		/*
923 		 * Copy the allocated CCB contents back to the malloced CCB
924 		 * so we can give status back to the user when he requests it.
925 		 */
926 		bcopy(done_ccb, &io_req->ccb, sizeof(*done_ccb));
927 
928 		/*
929 		 * Log data/transaction completion with devstat(9).
930 		 */
931 		switch (done_ccb->ccb_h.func_code) {
932 		case XPT_SCSI_IO:
933 			devstat_end_transaction(softc->device_stats,
934 			    done_ccb->csio.dxfer_len - done_ccb->csio.resid,
935 			    done_ccb->csio.tag_action & 0x3,
936 			    ((done_ccb->ccb_h.flags & CAM_DIR_MASK) ==
937 			    CAM_DIR_NONE) ? DEVSTAT_NO_DATA :
938 			    (done_ccb->ccb_h.flags & CAM_DIR_OUT) ?
939 			    DEVSTAT_WRITE : DEVSTAT_READ, NULL,
940 			    &io_req->start_time);
941 			break;
942 		case XPT_ATA_IO:
943 			devstat_end_transaction(softc->device_stats,
944 			    done_ccb->ataio.dxfer_len - done_ccb->ataio.resid,
945 			    0, /* Not used in ATA */
946 			    ((done_ccb->ccb_h.flags & CAM_DIR_MASK) ==
947 			    CAM_DIR_NONE) ? DEVSTAT_NO_DATA :
948 			    (done_ccb->ccb_h.flags & CAM_DIR_OUT) ?
949 			    DEVSTAT_WRITE : DEVSTAT_READ, NULL,
950 			    &io_req->start_time);
951 			break;
952 		case XPT_SMP_IO:
953 			/*
954 			 * XXX KDM this isn't quite right, but there isn't
955 			 * currently an easy way to represent a bidirectional
956 			 * transfer in devstat.  The only way to do it
957 			 * and have the byte counts come out right would
958 			 * mean that we would have to record two
959 			 * transactions, one for the request and one for the
960 			 * response.  For now, so that we report something,
961 			 * just treat the entire thing as a read.
962 			 */
963 			devstat_end_transaction(softc->device_stats,
964 			    done_ccb->smpio.smp_request_len +
965 			    done_ccb->smpio.smp_response_len,
966 			    DEVSTAT_TAG_SIMPLE, DEVSTAT_READ, NULL,
967 			    &io_req->start_time);
968 			break;
969 		default:
970 			devstat_end_transaction(softc->device_stats, 0,
971 			    DEVSTAT_TAG_NONE, DEVSTAT_NO_DATA, NULL,
972 			    &io_req->start_time);
973 			break;
974 		}
975 
976 		/*
977 		 * In the normal case, take the completed I/O off of the
978 		 * active queue and put it on the done queue.  Notitfy the
979 		 * user that we have a completed I/O.
980 		 */
981 		if ((io_req->flags & PASS_IO_ABANDONED) == 0) {
982 			TAILQ_REMOVE(&softc->active_queue, io_req, links);
983 			TAILQ_INSERT_TAIL(&softc->done_queue, io_req, links);
984 			selwakeuppri(&softc->read_select, PRIBIO);
985 			KNOTE_LOCKED(&softc->read_select.si_note, 0);
986 		} else {
987 			/*
988 			 * In the case of an abandoned I/O (final close
989 			 * without fetching the I/O), take it off of the
990 			 * abandoned queue and free it.
991 			 */
992 			TAILQ_REMOVE(&softc->abandoned_queue, io_req, links);
993 			passiocleanup(softc, io_req);
994 			uma_zfree(softc->pass_zone, io_req);
995 
996 			/*
997 			 * Release the done_ccb here, since we may wind up
998 			 * freeing the peripheral when we decrement the
999 			 * reference count below.
1000 			 */
1001 			xpt_release_ccb(done_ccb);
1002 
1003 			/*
1004 			 * If the abandoned queue is empty, we can release
1005 			 * our reference to the periph since we won't have
1006 			 * any more completions coming.
1007 			 */
1008 			if ((TAILQ_EMPTY(&softc->abandoned_queue))
1009 			 && (softc->flags & PASS_FLAG_ABANDONED_REF_SET)) {
1010 				softc->flags &= ~PASS_FLAG_ABANDONED_REF_SET;
1011 				cam_periph_release_locked(periph);
1012 			}
1013 
1014 			/*
1015 			 * We have already released the CCB, so we can
1016 			 * return.
1017 			 */
1018 			return;
1019 		}
1020 		break;
1021 	}
1022 	}
1023 	xpt_release_ccb(done_ccb);
1024 }
1025 
1026 static int
1027 passcreatezone(struct cam_periph *periph)
1028 {
1029 	struct pass_softc *softc;
1030 	int error;
1031 
1032 	error = 0;
1033 	softc = (struct pass_softc *)periph->softc;
1034 
1035 	cam_periph_assert(periph, MA_OWNED);
1036 	KASSERT(((softc->flags & PASS_FLAG_ZONE_VALID) == 0),
1037 		("%s called when the pass(4) zone is valid!\n", __func__));
1038 	KASSERT((softc->pass_zone == NULL),
1039 		("%s called when the pass(4) zone is allocated!\n", __func__));
1040 
1041 	if ((softc->flags & PASS_FLAG_ZONE_INPROG) == 0) {
1042 
1043 		/*
1044 		 * We're the first context through, so we need to create
1045 		 * the pass(4) UMA zone for I/O requests.
1046 		 */
1047 		softc->flags |= PASS_FLAG_ZONE_INPROG;
1048 
1049 		/*
1050 		 * uma_zcreate() does a blocking (M_WAITOK) allocation,
1051 		 * so we cannot hold a mutex while we call it.
1052 		 */
1053 		cam_periph_unlock(periph);
1054 
1055 		softc->pass_zone = uma_zcreate(softc->zone_name,
1056 		    sizeof(struct pass_io_req), NULL, NULL, NULL, NULL,
1057 		    /*align*/ 0, /*flags*/ 0);
1058 
1059 		softc->pass_io_zone = uma_zcreate(softc->io_zone_name,
1060 		    softc->io_zone_size, NULL, NULL, NULL, NULL,
1061 		    /*align*/ 0, /*flags*/ 0);
1062 
1063 		cam_periph_lock(periph);
1064 
1065 		if ((softc->pass_zone == NULL)
1066 		 || (softc->pass_io_zone == NULL)) {
1067 			if (softc->pass_zone == NULL)
1068 				xpt_print(periph->path, "unable to allocate "
1069 				    "IO Req UMA zone\n");
1070 			else
1071 				xpt_print(periph->path, "unable to allocate "
1072 				    "IO UMA zone\n");
1073 			softc->flags &= ~PASS_FLAG_ZONE_INPROG;
1074 			goto bailout;
1075 		}
1076 
1077 		/*
1078 		 * Set the flags appropriately and notify any other waiters.
1079 		 */
1080 		softc->flags &= PASS_FLAG_ZONE_INPROG;
1081 		softc->flags |= PASS_FLAG_ZONE_VALID;
1082 		wakeup(&softc->pass_zone);
1083 	} else {
1084 		/*
1085 		 * In this case, the UMA zone has not yet been created, but
1086 		 * another context is in the process of creating it.  We
1087 		 * need to sleep until the creation is either done or has
1088 		 * failed.
1089 		 */
1090 		while ((softc->flags & PASS_FLAG_ZONE_INPROG)
1091 		    && ((softc->flags & PASS_FLAG_ZONE_VALID) == 0)) {
1092 			error = msleep(&softc->pass_zone,
1093 				       cam_periph_mtx(periph), PRIBIO,
1094 				       "paszon", 0);
1095 			if (error != 0)
1096 				goto bailout;
1097 		}
1098 		/*
1099 		 * If the zone creation failed, no luck for the user.
1100 		 */
1101 		if ((softc->flags & PASS_FLAG_ZONE_VALID) == 0){
1102 			error = ENOMEM;
1103 			goto bailout;
1104 		}
1105 	}
1106 bailout:
1107 	return (error);
1108 }
1109 
1110 static void
1111 passiocleanup(struct pass_softc *softc, struct pass_io_req *io_req)
1112 {
1113 	union ccb *ccb;
1114 	u_int8_t **data_ptrs[CAM_PERIPH_MAXMAPS];
1115 	int i, numbufs;
1116 
1117 	ccb = &io_req->ccb;
1118 
1119 	switch (ccb->ccb_h.func_code) {
1120 	case XPT_DEV_MATCH:
1121 		numbufs = min(io_req->num_bufs, 2);
1122 
1123 		if (numbufs == 1) {
1124 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.matches;
1125 		} else {
1126 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.patterns;
1127 			data_ptrs[1] = (u_int8_t **)&ccb->cdm.matches;
1128 		}
1129 		break;
1130 	case XPT_SCSI_IO:
1131 	case XPT_CONT_TARGET_IO:
1132 		data_ptrs[0] = &ccb->csio.data_ptr;
1133 		numbufs = min(io_req->num_bufs, 1);
1134 		break;
1135 	case XPT_ATA_IO:
1136 		data_ptrs[0] = &ccb->ataio.data_ptr;
1137 		numbufs = min(io_req->num_bufs, 1);
1138 		break;
1139 	case XPT_SMP_IO:
1140 		numbufs = min(io_req->num_bufs, 2);
1141 		data_ptrs[0] = &ccb->smpio.smp_request;
1142 		data_ptrs[1] = &ccb->smpio.smp_response;
1143 		break;
1144 	case XPT_DEV_ADVINFO:
1145 		numbufs = min(io_req->num_bufs, 1);
1146 		data_ptrs[0] = (uint8_t **)&ccb->cdai.buf;
1147 		break;
1148 	case XPT_NVME_IO:
1149 	case XPT_NVME_ADMIN:
1150 		data_ptrs[0] = &ccb->nvmeio.data_ptr;
1151 		numbufs = min(io_req->num_bufs, 1);
1152 		break;
1153 	default:
1154 		/* allow ourselves to be swapped once again */
1155 		return;
1156 		break; /* NOTREACHED */
1157 	}
1158 
1159 	if (io_req->flags & PASS_IO_USER_SEG_MALLOC) {
1160 		free(io_req->user_segptr, M_SCSIPASS);
1161 		io_req->user_segptr = NULL;
1162 	}
1163 
1164 	/*
1165 	 * We only want to free memory we malloced.
1166 	 */
1167 	if (io_req->data_flags == CAM_DATA_VADDR) {
1168 		for (i = 0; i < io_req->num_bufs; i++) {
1169 			if (io_req->kern_bufs[i] == NULL)
1170 				continue;
1171 
1172 			free(io_req->kern_bufs[i], M_SCSIPASS);
1173 			io_req->kern_bufs[i] = NULL;
1174 		}
1175 	} else if (io_req->data_flags == CAM_DATA_SG) {
1176 		for (i = 0; i < io_req->num_kern_segs; i++) {
1177 			if ((uint8_t *)(uintptr_t)
1178 			    io_req->kern_segptr[i].ds_addr == NULL)
1179 				continue;
1180 
1181 			uma_zfree(softc->pass_io_zone, (uint8_t *)(uintptr_t)
1182 			    io_req->kern_segptr[i].ds_addr);
1183 			io_req->kern_segptr[i].ds_addr = 0;
1184 		}
1185 	}
1186 
1187 	if (io_req->flags & PASS_IO_KERN_SEG_MALLOC) {
1188 		free(io_req->kern_segptr, M_SCSIPASS);
1189 		io_req->kern_segptr = NULL;
1190 	}
1191 
1192 	if (io_req->data_flags != CAM_DATA_PADDR) {
1193 		for (i = 0; i < numbufs; i++) {
1194 			/*
1195 			 * Restore the user's buffer pointers to their
1196 			 * previous values.
1197 			 */
1198 			if (io_req->user_bufs[i] != NULL)
1199 				*data_ptrs[i] = io_req->user_bufs[i];
1200 		}
1201 	}
1202 
1203 }
1204 
1205 static int
1206 passcopysglist(struct cam_periph *periph, struct pass_io_req *io_req,
1207 	       ccb_flags direction)
1208 {
1209 	bus_size_t kern_watermark, user_watermark, len_copied, len_to_copy;
1210 	bus_dma_segment_t *user_sglist, *kern_sglist;
1211 	int i, j, error;
1212 
1213 	error = 0;
1214 	kern_watermark = 0;
1215 	user_watermark = 0;
1216 	len_to_copy = 0;
1217 	len_copied = 0;
1218 	user_sglist = io_req->user_segptr;
1219 	kern_sglist = io_req->kern_segptr;
1220 
1221 	for (i = 0, j = 0; i < io_req->num_user_segs &&
1222 	     j < io_req->num_kern_segs;) {
1223 		uint8_t *user_ptr, *kern_ptr;
1224 
1225 		len_to_copy = min(user_sglist[i].ds_len -user_watermark,
1226 		    kern_sglist[j].ds_len - kern_watermark);
1227 
1228 		user_ptr = (uint8_t *)(uintptr_t)user_sglist[i].ds_addr;
1229 		user_ptr = user_ptr + user_watermark;
1230 		kern_ptr = (uint8_t *)(uintptr_t)kern_sglist[j].ds_addr;
1231 		kern_ptr = kern_ptr + kern_watermark;
1232 
1233 		user_watermark += len_to_copy;
1234 		kern_watermark += len_to_copy;
1235 
1236 		if (direction == CAM_DIR_IN) {
1237 			error = copyout(kern_ptr, user_ptr, len_to_copy);
1238 			if (error != 0) {
1239 				xpt_print(periph->path, "%s: copyout of %u "
1240 					  "bytes from %p to %p failed with "
1241 					  "error %d\n", __func__, len_to_copy,
1242 					  kern_ptr, user_ptr, error);
1243 				goto bailout;
1244 			}
1245 		} else {
1246 			error = copyin(user_ptr, kern_ptr, len_to_copy);
1247 			if (error != 0) {
1248 				xpt_print(periph->path, "%s: copyin of %u "
1249 					  "bytes from %p to %p failed with "
1250 					  "error %d\n", __func__, len_to_copy,
1251 					  user_ptr, kern_ptr, error);
1252 				goto bailout;
1253 			}
1254 		}
1255 
1256 		len_copied += len_to_copy;
1257 
1258 		if (user_sglist[i].ds_len == user_watermark) {
1259 			i++;
1260 			user_watermark = 0;
1261 		}
1262 
1263 		if (kern_sglist[j].ds_len == kern_watermark) {
1264 			j++;
1265 			kern_watermark = 0;
1266 		}
1267 	}
1268 
1269 bailout:
1270 
1271 	return (error);
1272 }
1273 
1274 static int
1275 passmemsetup(struct cam_periph *periph, struct pass_io_req *io_req)
1276 {
1277 	union ccb *ccb;
1278 	struct pass_softc *softc;
1279 	int numbufs, i;
1280 	uint8_t **data_ptrs[CAM_PERIPH_MAXMAPS];
1281 	uint32_t lengths[CAM_PERIPH_MAXMAPS];
1282 	uint32_t dirs[CAM_PERIPH_MAXMAPS];
1283 	uint32_t num_segs;
1284 	uint16_t *seg_cnt_ptr;
1285 	size_t maxmap;
1286 	int error;
1287 
1288 	cam_periph_assert(periph, MA_NOTOWNED);
1289 
1290 	softc = periph->softc;
1291 
1292 	error = 0;
1293 	ccb = &io_req->ccb;
1294 	maxmap = 0;
1295 	num_segs = 0;
1296 	seg_cnt_ptr = NULL;
1297 
1298 	switch(ccb->ccb_h.func_code) {
1299 	case XPT_DEV_MATCH:
1300 		if (ccb->cdm.match_buf_len == 0) {
1301 			printf("%s: invalid match buffer length 0\n", __func__);
1302 			return(EINVAL);
1303 		}
1304 		if (ccb->cdm.pattern_buf_len > 0) {
1305 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.patterns;
1306 			lengths[0] = ccb->cdm.pattern_buf_len;
1307 			dirs[0] = CAM_DIR_OUT;
1308 			data_ptrs[1] = (u_int8_t **)&ccb->cdm.matches;
1309 			lengths[1] = ccb->cdm.match_buf_len;
1310 			dirs[1] = CAM_DIR_IN;
1311 			numbufs = 2;
1312 		} else {
1313 			data_ptrs[0] = (u_int8_t **)&ccb->cdm.matches;
1314 			lengths[0] = ccb->cdm.match_buf_len;
1315 			dirs[0] = CAM_DIR_IN;
1316 			numbufs = 1;
1317 		}
1318 		io_req->data_flags = CAM_DATA_VADDR;
1319 		break;
1320 	case XPT_SCSI_IO:
1321 	case XPT_CONT_TARGET_IO:
1322 		if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
1323 			return(0);
1324 
1325 		/*
1326 		 * The user shouldn't be able to supply a bio.
1327 		 */
1328 		if ((ccb->ccb_h.flags & CAM_DATA_MASK) == CAM_DATA_BIO)
1329 			return (EINVAL);
1330 
1331 		io_req->data_flags = ccb->ccb_h.flags & CAM_DATA_MASK;
1332 
1333 		data_ptrs[0] = &ccb->csio.data_ptr;
1334 		lengths[0] = ccb->csio.dxfer_len;
1335 		dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1336 		num_segs = ccb->csio.sglist_cnt;
1337 		seg_cnt_ptr = &ccb->csio.sglist_cnt;
1338 		numbufs = 1;
1339 		maxmap = softc->maxio;
1340 		break;
1341 	case XPT_ATA_IO:
1342 		if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
1343 			return(0);
1344 
1345 		/*
1346 		 * We only support a single virtual address for ATA I/O.
1347 		 */
1348 		if ((ccb->ccb_h.flags & CAM_DATA_MASK) != CAM_DATA_VADDR)
1349 			return (EINVAL);
1350 
1351 		io_req->data_flags = CAM_DATA_VADDR;
1352 
1353 		data_ptrs[0] = &ccb->ataio.data_ptr;
1354 		lengths[0] = ccb->ataio.dxfer_len;
1355 		dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1356 		numbufs = 1;
1357 		maxmap = softc->maxio;
1358 		break;
1359 	case XPT_SMP_IO:
1360 		io_req->data_flags = CAM_DATA_VADDR;
1361 
1362 		data_ptrs[0] = &ccb->smpio.smp_request;
1363 		lengths[0] = ccb->smpio.smp_request_len;
1364 		dirs[0] = CAM_DIR_OUT;
1365 		data_ptrs[1] = &ccb->smpio.smp_response;
1366 		lengths[1] = ccb->smpio.smp_response_len;
1367 		dirs[1] = CAM_DIR_IN;
1368 		numbufs = 2;
1369 		maxmap = softc->maxio;
1370 		break;
1371 	case XPT_DEV_ADVINFO:
1372 		if (ccb->cdai.bufsiz == 0)
1373 			return (0);
1374 
1375 		io_req->data_flags = CAM_DATA_VADDR;
1376 
1377 		data_ptrs[0] = (uint8_t **)&ccb->cdai.buf;
1378 		lengths[0] = ccb->cdai.bufsiz;
1379 		dirs[0] = CAM_DIR_IN;
1380 		numbufs = 1;
1381 		break;
1382 	case XPT_NVME_ADMIN:
1383 	case XPT_NVME_IO:
1384 		if ((ccb->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_NONE)
1385 			return (0);
1386 
1387 		io_req->data_flags = ccb->ccb_h.flags & CAM_DATA_MASK;
1388 
1389 		data_ptrs[0] = &ccb->nvmeio.data_ptr;
1390 		lengths[0] = ccb->nvmeio.dxfer_len;
1391 		dirs[0] = ccb->ccb_h.flags & CAM_DIR_MASK;
1392 		num_segs = ccb->nvmeio.sglist_cnt;
1393 		seg_cnt_ptr = &ccb->nvmeio.sglist_cnt;
1394 		numbufs = 1;
1395 		maxmap = softc->maxio;
1396 		break;
1397 	default:
1398 		return(EINVAL);
1399 		break; /* NOTREACHED */
1400 	}
1401 
1402 	io_req->num_bufs = numbufs;
1403 
1404 	/*
1405 	 * If there is a maximum, check to make sure that the user's
1406 	 * request fits within the limit.  In general, we should only have
1407 	 * a maximum length for requests that go to hardware.  Otherwise it
1408 	 * is whatever we're able to malloc.
1409 	 */
1410 	for (i = 0; i < numbufs; i++) {
1411 		io_req->user_bufs[i] = *data_ptrs[i];
1412 		io_req->dirs[i] = dirs[i];
1413 		io_req->lengths[i] = lengths[i];
1414 
1415 		if (maxmap == 0)
1416 			continue;
1417 
1418 		if (lengths[i] <= maxmap)
1419 			continue;
1420 
1421 		xpt_print(periph->path, "%s: data length %u > max allowed %u "
1422 			  "bytes\n", __func__, lengths[i], maxmap);
1423 		error = EINVAL;
1424 		goto bailout;
1425 	}
1426 
1427 	switch (io_req->data_flags) {
1428 	case CAM_DATA_VADDR:
1429 		/* Map or copy the buffer into kernel address space */
1430 		for (i = 0; i < numbufs; i++) {
1431 			uint8_t *tmp_buf;
1432 
1433 			/*
1434 			 * If for some reason no length is specified, we
1435 			 * don't need to allocate anything.
1436 			 */
1437 			if (io_req->lengths[i] == 0)
1438 				continue;
1439 
1440 			tmp_buf = malloc(lengths[i], M_SCSIPASS,
1441 					 M_WAITOK | M_ZERO);
1442 			io_req->kern_bufs[i] = tmp_buf;
1443 			*data_ptrs[i] = tmp_buf;
1444 
1445 #if 0
1446 			xpt_print(periph->path, "%s: malloced %p len %u, user "
1447 				  "buffer %p, operation: %s\n", __func__,
1448 				  tmp_buf, lengths[i], io_req->user_bufs[i],
1449 				  (dirs[i] == CAM_DIR_IN) ? "read" : "write");
1450 #endif
1451 			/*
1452 			 * We only need to copy in if the user is writing.
1453 			 */
1454 			if (dirs[i] != CAM_DIR_OUT)
1455 				continue;
1456 
1457 			error = copyin(io_req->user_bufs[i],
1458 				       io_req->kern_bufs[i], lengths[i]);
1459 			if (error != 0) {
1460 				xpt_print(periph->path, "%s: copy of user "
1461 					  "buffer from %p to %p failed with "
1462 					  "error %d\n", __func__,
1463 					  io_req->user_bufs[i],
1464 					  io_req->kern_bufs[i], error);
1465 				goto bailout;
1466 			}
1467 		}
1468 		break;
1469 	case CAM_DATA_PADDR:
1470 		/* Pass down the pointer as-is */
1471 		break;
1472 	case CAM_DATA_SG: {
1473 		size_t sg_length, size_to_go, alloc_size;
1474 		uint32_t num_segs_needed;
1475 
1476 		/*
1477 		 * Copy the user S/G list in, and then copy in the
1478 		 * individual segments.
1479 		 */
1480 		/*
1481 		 * We shouldn't see this, but check just in case.
1482 		 */
1483 		if (numbufs != 1) {
1484 			xpt_print(periph->path, "%s: cannot currently handle "
1485 				  "more than one S/G list per CCB\n", __func__);
1486 			error = EINVAL;
1487 			goto bailout;
1488 		}
1489 
1490 		/*
1491 		 * We have to have at least one segment.
1492 		 */
1493 		if (num_segs == 0) {
1494 			xpt_print(periph->path, "%s: CAM_DATA_SG flag set, "
1495 				  "but sglist_cnt=0!\n", __func__);
1496 			error = EINVAL;
1497 			goto bailout;
1498 		}
1499 
1500 		/*
1501 		 * Make sure the user specified the total length and didn't
1502 		 * just leave it to us to decode the S/G list.
1503 		 */
1504 		if (lengths[0] == 0) {
1505 			xpt_print(periph->path, "%s: no dxfer_len specified, "
1506 				  "but CAM_DATA_SG flag is set!\n", __func__);
1507 			error = EINVAL;
1508 			goto bailout;
1509 		}
1510 
1511 		/*
1512 		 * We allocate buffers in io_zone_size increments for an
1513 		 * S/G list.  This will generally be MAXPHYS.
1514 		 */
1515 		if (lengths[0] <= softc->io_zone_size)
1516 			num_segs_needed = 1;
1517 		else {
1518 			num_segs_needed = lengths[0] / softc->io_zone_size;
1519 			if ((lengths[0] % softc->io_zone_size) != 0)
1520 				num_segs_needed++;
1521 		}
1522 
1523 		/* Figure out the size of the S/G list */
1524 		sg_length = num_segs * sizeof(bus_dma_segment_t);
1525 		io_req->num_user_segs = num_segs;
1526 		io_req->num_kern_segs = num_segs_needed;
1527 
1528 		/* Save the user's S/G list pointer for later restoration */
1529 		io_req->user_bufs[0] = *data_ptrs[0];
1530 
1531 		/*
1532 		 * If we have enough segments allocated by default to handle
1533 		 * the length of the user's S/G list,
1534 		 */
1535 		if (num_segs > PASS_MAX_SEGS) {
1536 			io_req->user_segptr = malloc(sizeof(bus_dma_segment_t) *
1537 			    num_segs, M_SCSIPASS, M_WAITOK | M_ZERO);
1538 			io_req->flags |= PASS_IO_USER_SEG_MALLOC;
1539 		} else
1540 			io_req->user_segptr = io_req->user_segs;
1541 
1542 		error = copyin(*data_ptrs[0], io_req->user_segptr, sg_length);
1543 		if (error != 0) {
1544 			xpt_print(periph->path, "%s: copy of user S/G list "
1545 				  "from %p to %p failed with error %d\n",
1546 				  __func__, *data_ptrs[0], io_req->user_segptr,
1547 				  error);
1548 			goto bailout;
1549 		}
1550 
1551 		if (num_segs_needed > PASS_MAX_SEGS) {
1552 			io_req->kern_segptr = malloc(sizeof(bus_dma_segment_t) *
1553 			    num_segs_needed, M_SCSIPASS, M_WAITOK | M_ZERO);
1554 			io_req->flags |= PASS_IO_KERN_SEG_MALLOC;
1555 		} else {
1556 			io_req->kern_segptr = io_req->kern_segs;
1557 		}
1558 
1559 		/*
1560 		 * Allocate the kernel S/G list.
1561 		 */
1562 		for (size_to_go = lengths[0], i = 0;
1563 		     size_to_go > 0 && i < num_segs_needed;
1564 		     i++, size_to_go -= alloc_size) {
1565 			uint8_t *kern_ptr;
1566 
1567 			alloc_size = min(size_to_go, softc->io_zone_size);
1568 			kern_ptr = uma_zalloc(softc->pass_io_zone, M_WAITOK);
1569 			io_req->kern_segptr[i].ds_addr =
1570 			    (bus_addr_t)(uintptr_t)kern_ptr;
1571 			io_req->kern_segptr[i].ds_len = alloc_size;
1572 		}
1573 		if (size_to_go > 0) {
1574 			printf("%s: size_to_go = %zu, software error!\n",
1575 			       __func__, size_to_go);
1576 			error = EINVAL;
1577 			goto bailout;
1578 		}
1579 
1580 		*data_ptrs[0] = (uint8_t *)io_req->kern_segptr;
1581 		*seg_cnt_ptr = io_req->num_kern_segs;
1582 
1583 		/*
1584 		 * We only need to copy data here if the user is writing.
1585 		 */
1586 		if (dirs[0] == CAM_DIR_OUT)
1587 			error = passcopysglist(periph, io_req, dirs[0]);
1588 		break;
1589 	}
1590 	case CAM_DATA_SG_PADDR: {
1591 		size_t sg_length;
1592 
1593 		/*
1594 		 * We shouldn't see this, but check just in case.
1595 		 */
1596 		if (numbufs != 1) {
1597 			printf("%s: cannot currently handle more than one "
1598 			       "S/G list per CCB\n", __func__);
1599 			error = EINVAL;
1600 			goto bailout;
1601 		}
1602 
1603 		/*
1604 		 * We have to have at least one segment.
1605 		 */
1606 		if (num_segs == 0) {
1607 			xpt_print(periph->path, "%s: CAM_DATA_SG_PADDR flag "
1608 				  "set, but sglist_cnt=0!\n", __func__);
1609 			error = EINVAL;
1610 			goto bailout;
1611 		}
1612 
1613 		/*
1614 		 * Make sure the user specified the total length and didn't
1615 		 * just leave it to us to decode the S/G list.
1616 		 */
1617 		if (lengths[0] == 0) {
1618 			xpt_print(periph->path, "%s: no dxfer_len specified, "
1619 				  "but CAM_DATA_SG flag is set!\n", __func__);
1620 			error = EINVAL;
1621 			goto bailout;
1622 		}
1623 
1624 		/* Figure out the size of the S/G list */
1625 		sg_length = num_segs * sizeof(bus_dma_segment_t);
1626 		io_req->num_user_segs = num_segs;
1627 		io_req->num_kern_segs = io_req->num_user_segs;
1628 
1629 		/* Save the user's S/G list pointer for later restoration */
1630 		io_req->user_bufs[0] = *data_ptrs[0];
1631 
1632 		if (num_segs > PASS_MAX_SEGS) {
1633 			io_req->user_segptr = malloc(sizeof(bus_dma_segment_t) *
1634 			    num_segs, M_SCSIPASS, M_WAITOK | M_ZERO);
1635 			io_req->flags |= PASS_IO_USER_SEG_MALLOC;
1636 		} else
1637 			io_req->user_segptr = io_req->user_segs;
1638 
1639 		io_req->kern_segptr = io_req->user_segptr;
1640 
1641 		error = copyin(*data_ptrs[0], io_req->user_segptr, sg_length);
1642 		if (error != 0) {
1643 			xpt_print(periph->path, "%s: copy of user S/G list "
1644 				  "from %p to %p failed with error %d\n",
1645 				  __func__, *data_ptrs[0], io_req->user_segptr,
1646 				  error);
1647 			goto bailout;
1648 		}
1649 		break;
1650 	}
1651 	default:
1652 	case CAM_DATA_BIO:
1653 		/*
1654 		 * A user shouldn't be attaching a bio to the CCB.  It
1655 		 * isn't a user-accessible structure.
1656 		 */
1657 		error = EINVAL;
1658 		break;
1659 	}
1660 
1661 bailout:
1662 	if (error != 0)
1663 		passiocleanup(softc, io_req);
1664 
1665 	return (error);
1666 }
1667 
1668 static int
1669 passmemdone(struct cam_periph *periph, struct pass_io_req *io_req)
1670 {
1671 	struct pass_softc *softc;
1672 	int error;
1673 	int i;
1674 
1675 	error = 0;
1676 	softc = (struct pass_softc *)periph->softc;
1677 
1678 	switch (io_req->data_flags) {
1679 	case CAM_DATA_VADDR:
1680 		/*
1681 		 * Copy back to the user buffer if this was a read.
1682 		 */
1683 		for (i = 0; i < io_req->num_bufs; i++) {
1684 			if (io_req->dirs[i] != CAM_DIR_IN)
1685 				continue;
1686 
1687 			error = copyout(io_req->kern_bufs[i],
1688 			    io_req->user_bufs[i], io_req->lengths[i]);
1689 			if (error != 0) {
1690 				xpt_print(periph->path, "Unable to copy %u "
1691 					  "bytes from %p to user address %p\n",
1692 					  io_req->lengths[i],
1693 					  io_req->kern_bufs[i],
1694 					  io_req->user_bufs[i]);
1695 				goto bailout;
1696 			}
1697 
1698 		}
1699 		break;
1700 	case CAM_DATA_PADDR:
1701 		/* Do nothing.  The pointer is a physical address already */
1702 		break;
1703 	case CAM_DATA_SG:
1704 		/*
1705 		 * Copy back to the user buffer if this was a read.
1706 		 * Restore the user's S/G list buffer pointer.
1707 		 */
1708 		if (io_req->dirs[0] == CAM_DIR_IN)
1709 			error = passcopysglist(periph, io_req, io_req->dirs[0]);
1710 		break;
1711 	case CAM_DATA_SG_PADDR:
1712 		/*
1713 		 * Restore the user's S/G list buffer pointer.  No need to
1714 		 * copy.
1715 		 */
1716 		break;
1717 	default:
1718 	case CAM_DATA_BIO:
1719 		error = EINVAL;
1720 		break;
1721 	}
1722 
1723 bailout:
1724 	/*
1725 	 * Reset the user's pointers to their original values and free
1726 	 * allocated memory.
1727 	 */
1728 	passiocleanup(softc, io_req);
1729 
1730 	return (error);
1731 }
1732 
1733 static int
1734 passioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flag, struct thread *td)
1735 {
1736 	int error;
1737 
1738 	if ((error = passdoioctl(dev, cmd, addr, flag, td)) == ENOTTY) {
1739 		error = cam_compat_ioctl(dev, cmd, addr, flag, td, passdoioctl);
1740 	}
1741 	return (error);
1742 }
1743 
1744 static int
1745 passdoioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flag, struct thread *td)
1746 {
1747 	struct	cam_periph *periph;
1748 	struct	pass_softc *softc;
1749 	int	error;
1750 	uint32_t priority;
1751 
1752 	periph = (struct cam_periph *)dev->si_drv1;
1753 	cam_periph_lock(periph);
1754 	softc = (struct pass_softc *)periph->softc;
1755 
1756 	error = 0;
1757 
1758 	switch (cmd) {
1759 
1760 	case CAMIOCOMMAND:
1761 	{
1762 		union ccb *inccb;
1763 		union ccb *ccb;
1764 		int ccb_malloced;
1765 
1766 		inccb = (union ccb *)addr;
1767 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
1768 		if (inccb->ccb_h.func_code == XPT_SCSI_IO)
1769 			inccb->csio.bio = NULL;
1770 #endif
1771 
1772 		if (inccb->ccb_h.flags & CAM_UNLOCKED) {
1773 			error = EINVAL;
1774 			break;
1775 		}
1776 
1777 		/*
1778 		 * Some CCB types, like scan bus and scan lun can only go
1779 		 * through the transport layer device.
1780 		 */
1781 		if (inccb->ccb_h.func_code & XPT_FC_XPT_ONLY) {
1782 			xpt_print(periph->path, "CCB function code %#x is "
1783 			    "restricted to the XPT device\n",
1784 			    inccb->ccb_h.func_code);
1785 			error = ENODEV;
1786 			break;
1787 		}
1788 
1789 		/* Compatibility for RL/priority-unaware code. */
1790 		priority = inccb->ccb_h.pinfo.priority;
1791 		if (priority <= CAM_PRIORITY_OOB)
1792 		    priority += CAM_PRIORITY_OOB + 1;
1793 
1794 		/*
1795 		 * Non-immediate CCBs need a CCB from the per-device pool
1796 		 * of CCBs, which is scheduled by the transport layer.
1797 		 * Immediate CCBs and user-supplied CCBs should just be
1798 		 * malloced.
1799 		 */
1800 		if ((inccb->ccb_h.func_code & XPT_FC_QUEUED)
1801 		 && ((inccb->ccb_h.func_code & XPT_FC_USER_CCB) == 0)) {
1802 			ccb = cam_periph_getccb(periph, priority);
1803 			ccb_malloced = 0;
1804 		} else {
1805 			ccb = xpt_alloc_ccb_nowait();
1806 
1807 			if (ccb != NULL)
1808 				xpt_setup_ccb(&ccb->ccb_h, periph->path,
1809 					      priority);
1810 			ccb_malloced = 1;
1811 		}
1812 
1813 		if (ccb == NULL) {
1814 			xpt_print(periph->path, "unable to allocate CCB\n");
1815 			error = ENOMEM;
1816 			break;
1817 		}
1818 
1819 		error = passsendccb(periph, ccb, inccb);
1820 
1821 		if (ccb_malloced)
1822 			xpt_free_ccb(ccb);
1823 		else
1824 			xpt_release_ccb(ccb);
1825 
1826 		break;
1827 	}
1828 	case CAMIOQUEUE:
1829 	{
1830 		struct pass_io_req *io_req;
1831 		union ccb **user_ccb, *ccb;
1832 		xpt_opcode fc;
1833 
1834 #ifdef COMPAT_FREEBSD32
1835 		if (SV_PROC_FLAG(td->td_proc, SV_ILP32)) {
1836 			error = ENOTTY;
1837 			goto bailout;
1838 		}
1839 #endif
1840 		if ((softc->flags & PASS_FLAG_ZONE_VALID) == 0) {
1841 			error = passcreatezone(periph);
1842 			if (error != 0)
1843 				goto bailout;
1844 		}
1845 
1846 		/*
1847 		 * We're going to do a blocking allocation for this I/O
1848 		 * request, so we have to drop the lock.
1849 		 */
1850 		cam_periph_unlock(periph);
1851 
1852 		io_req = uma_zalloc(softc->pass_zone, M_WAITOK | M_ZERO);
1853 		ccb = &io_req->ccb;
1854 		user_ccb = (union ccb **)addr;
1855 
1856 		/*
1857 		 * Unlike the CAMIOCOMMAND ioctl above, we only have a
1858 		 * pointer to the user's CCB, so we have to copy the whole
1859 		 * thing in to a buffer we have allocated (above) instead
1860 		 * of allowing the ioctl code to malloc a buffer and copy
1861 		 * it in.
1862 		 *
1863 		 * This is an advantage for this asynchronous interface,
1864 		 * since we don't want the memory to get freed while the
1865 		 * CCB is outstanding.
1866 		 */
1867 #if 0
1868 		xpt_print(periph->path, "Copying user CCB %p to "
1869 			  "kernel address %p\n", *user_ccb, ccb);
1870 #endif
1871 		error = copyin(*user_ccb, ccb, sizeof(*ccb));
1872 		if (error != 0) {
1873 			xpt_print(periph->path, "Copy of user CCB %p to "
1874 				  "kernel address %p failed with error %d\n",
1875 				  *user_ccb, ccb, error);
1876 			goto camioqueue_error;
1877 		}
1878 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
1879 		if (ccb->ccb_h.func_code == XPT_SCSI_IO)
1880 			ccb->csio.bio = NULL;
1881 #endif
1882 
1883 		if (ccb->ccb_h.flags & CAM_UNLOCKED) {
1884 			error = EINVAL;
1885 			goto camioqueue_error;
1886 		}
1887 
1888 		if (ccb->ccb_h.flags & CAM_CDB_POINTER) {
1889 			if (ccb->csio.cdb_len > IOCDBLEN) {
1890 				error = EINVAL;
1891 				goto camioqueue_error;
1892 			}
1893 			error = copyin(ccb->csio.cdb_io.cdb_ptr,
1894 			    ccb->csio.cdb_io.cdb_bytes, ccb->csio.cdb_len);
1895 			if (error != 0)
1896 				goto camioqueue_error;
1897 			ccb->ccb_h.flags &= ~CAM_CDB_POINTER;
1898 		}
1899 
1900 		/*
1901 		 * Some CCB types, like scan bus and scan lun can only go
1902 		 * through the transport layer device.
1903 		 */
1904 		if (ccb->ccb_h.func_code & XPT_FC_XPT_ONLY) {
1905 			xpt_print(periph->path, "CCB function code %#x is "
1906 			    "restricted to the XPT device\n",
1907 			    ccb->ccb_h.func_code);
1908 			error = ENODEV;
1909 			goto camioqueue_error;
1910 		}
1911 
1912 		/*
1913 		 * Save the user's CCB pointer as well as his linked list
1914 		 * pointers and peripheral private area so that we can
1915 		 * restore these later.
1916 		 */
1917 		io_req->user_ccb_ptr = *user_ccb;
1918 		io_req->user_periph_links = ccb->ccb_h.periph_links;
1919 		io_req->user_periph_priv = ccb->ccb_h.periph_priv;
1920 
1921 		/*
1922 		 * Now that we've saved the user's values, we can set our
1923 		 * own peripheral private entry.
1924 		 */
1925 		ccb->ccb_h.ccb_ioreq = io_req;
1926 
1927 		/* Compatibility for RL/priority-unaware code. */
1928 		priority = ccb->ccb_h.pinfo.priority;
1929 		if (priority <= CAM_PRIORITY_OOB)
1930 		    priority += CAM_PRIORITY_OOB + 1;
1931 
1932 		/*
1933 		 * Setup fields in the CCB like the path and the priority.
1934 		 * The path in particular cannot be done in userland, since
1935 		 * it is a pointer to a kernel data structure.
1936 		 */
1937 		xpt_setup_ccb_flags(&ccb->ccb_h, periph->path, priority,
1938 				    ccb->ccb_h.flags);
1939 
1940 		/*
1941 		 * Setup our done routine.  There is no way for the user to
1942 		 * have a valid pointer here.
1943 		 */
1944 		ccb->ccb_h.cbfcnp = passdone;
1945 
1946 		fc = ccb->ccb_h.func_code;
1947 		/*
1948 		 * If this function code has memory that can be mapped in
1949 		 * or out, we need to call passmemsetup().
1950 		 */
1951 		if ((fc == XPT_SCSI_IO) || (fc == XPT_ATA_IO)
1952 		 || (fc == XPT_SMP_IO) || (fc == XPT_DEV_MATCH)
1953 		 || (fc == XPT_DEV_ADVINFO)
1954 		 || (fc == XPT_NVME_ADMIN) || (fc == XPT_NVME_IO)) {
1955 			error = passmemsetup(periph, io_req);
1956 			if (error != 0)
1957 				goto camioqueue_error;
1958 		} else
1959 			io_req->mapinfo.num_bufs_used = 0;
1960 
1961 		cam_periph_lock(periph);
1962 
1963 		/*
1964 		 * Everything goes on the incoming queue initially.
1965 		 */
1966 		TAILQ_INSERT_TAIL(&softc->incoming_queue, io_req, links);
1967 
1968 		/*
1969 		 * If the CCB is queued, and is not a user CCB, then
1970 		 * we need to allocate a slot for it.  Call xpt_schedule()
1971 		 * so that our start routine will get called when a CCB is
1972 		 * available.
1973 		 */
1974 		if ((fc & XPT_FC_QUEUED)
1975 		 && ((fc & XPT_FC_USER_CCB) == 0)) {
1976 			xpt_schedule(periph, priority);
1977 			break;
1978 		}
1979 
1980 		/*
1981 		 * At this point, the CCB in question is either an
1982 		 * immediate CCB (like XPT_DEV_ADVINFO) or it is a user CCB
1983 		 * and therefore should be malloced, not allocated via a slot.
1984 		 * Remove the CCB from the incoming queue and add it to the
1985 		 * active queue.
1986 		 */
1987 		TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
1988 		TAILQ_INSERT_TAIL(&softc->active_queue, io_req, links);
1989 
1990 		xpt_action(ccb);
1991 
1992 		/*
1993 		 * If this is not a queued CCB (i.e. it is an immediate CCB),
1994 		 * then it is already done.  We need to put it on the done
1995 		 * queue for the user to fetch.
1996 		 */
1997 		if ((fc & XPT_FC_QUEUED) == 0) {
1998 			TAILQ_REMOVE(&softc->active_queue, io_req, links);
1999 			TAILQ_INSERT_TAIL(&softc->done_queue, io_req, links);
2000 		}
2001 		break;
2002 
2003 camioqueue_error:
2004 		uma_zfree(softc->pass_zone, io_req);
2005 		cam_periph_lock(periph);
2006 		break;
2007 	}
2008 	case CAMIOGET:
2009 	{
2010 		union ccb **user_ccb;
2011 		struct pass_io_req *io_req;
2012 		int old_error;
2013 
2014 #ifdef COMPAT_FREEBSD32
2015 		if (SV_PROC_FLAG(td->td_proc, SV_ILP32)) {
2016 			error = ENOTTY;
2017 			goto bailout;
2018 		}
2019 #endif
2020 		user_ccb = (union ccb **)addr;
2021 		old_error = 0;
2022 
2023 		io_req = TAILQ_FIRST(&softc->done_queue);
2024 		if (io_req == NULL) {
2025 			error = ENOENT;
2026 			break;
2027 		}
2028 
2029 		/*
2030 		 * Remove the I/O from the done queue.
2031 		 */
2032 		TAILQ_REMOVE(&softc->done_queue, io_req, links);
2033 
2034 		/*
2035 		 * We have to drop the lock during the copyout because the
2036 		 * copyout can result in VM faults that require sleeping.
2037 		 */
2038 		cam_periph_unlock(periph);
2039 
2040 		/*
2041 		 * Do any needed copies (e.g. for reads) and revert the
2042 		 * pointers in the CCB back to the user's pointers.
2043 		 */
2044 		error = passmemdone(periph, io_req);
2045 
2046 		old_error = error;
2047 
2048 		io_req->ccb.ccb_h.periph_links = io_req->user_periph_links;
2049 		io_req->ccb.ccb_h.periph_priv = io_req->user_periph_priv;
2050 
2051 #if 0
2052 		xpt_print(periph->path, "Copying to user CCB %p from "
2053 			  "kernel address %p\n", *user_ccb, &io_req->ccb);
2054 #endif
2055 
2056 		error = copyout(&io_req->ccb, *user_ccb, sizeof(union ccb));
2057 		if (error != 0) {
2058 			xpt_print(periph->path, "Copy to user CCB %p from "
2059 				  "kernel address %p failed with error %d\n",
2060 				  *user_ccb, &io_req->ccb, error);
2061 		}
2062 
2063 		/*
2064 		 * Prefer the first error we got back, and make sure we
2065 		 * don't overwrite bad status with good.
2066 		 */
2067 		if (old_error != 0)
2068 			error = old_error;
2069 
2070 		cam_periph_lock(periph);
2071 
2072 		/*
2073 		 * At this point, if there was an error, we could potentially
2074 		 * re-queue the I/O and try again.  But why?  The error
2075 		 * would almost certainly happen again.  We might as well
2076 		 * not leak memory.
2077 		 */
2078 		uma_zfree(softc->pass_zone, io_req);
2079 		break;
2080 	}
2081 	default:
2082 		error = cam_periph_ioctl(periph, cmd, addr, passerror);
2083 		break;
2084 	}
2085 
2086 bailout:
2087 	cam_periph_unlock(periph);
2088 
2089 	return(error);
2090 }
2091 
2092 static int
2093 passpoll(struct cdev *dev, int poll_events, struct thread *td)
2094 {
2095 	struct cam_periph *periph;
2096 	struct pass_softc *softc;
2097 	int revents;
2098 
2099 	periph = (struct cam_periph *)dev->si_drv1;
2100 	softc = (struct pass_softc *)periph->softc;
2101 
2102 	revents = poll_events & (POLLOUT | POLLWRNORM);
2103 	if ((poll_events & (POLLIN | POLLRDNORM)) != 0) {
2104 		cam_periph_lock(periph);
2105 
2106 		if (!TAILQ_EMPTY(&softc->done_queue)) {
2107 			revents |= poll_events & (POLLIN | POLLRDNORM);
2108 		}
2109 		cam_periph_unlock(periph);
2110 		if (revents == 0)
2111 			selrecord(td, &softc->read_select);
2112 	}
2113 
2114 	return (revents);
2115 }
2116 
2117 static int
2118 passkqfilter(struct cdev *dev, struct knote *kn)
2119 {
2120 	struct cam_periph *periph;
2121 	struct pass_softc *softc;
2122 
2123 	periph = (struct cam_periph *)dev->si_drv1;
2124 	softc = (struct pass_softc *)periph->softc;
2125 
2126 	kn->kn_hook = (caddr_t)periph;
2127 	kn->kn_fop = &passread_filtops;
2128 	knlist_add(&softc->read_select.si_note, kn, 0);
2129 
2130 	return (0);
2131 }
2132 
2133 static void
2134 passreadfiltdetach(struct knote *kn)
2135 {
2136 	struct cam_periph *periph;
2137 	struct pass_softc *softc;
2138 
2139 	periph = (struct cam_periph *)kn->kn_hook;
2140 	softc = (struct pass_softc *)periph->softc;
2141 
2142 	knlist_remove(&softc->read_select.si_note, kn, 0);
2143 }
2144 
2145 static int
2146 passreadfilt(struct knote *kn, long hint)
2147 {
2148 	struct cam_periph *periph;
2149 	struct pass_softc *softc;
2150 	int retval;
2151 
2152 	periph = (struct cam_periph *)kn->kn_hook;
2153 	softc = (struct pass_softc *)periph->softc;
2154 
2155 	cam_periph_assert(periph, MA_OWNED);
2156 
2157 	if (TAILQ_EMPTY(&softc->done_queue))
2158 		retval = 0;
2159 	else
2160 		retval = 1;
2161 
2162 	return (retval);
2163 }
2164 
2165 /*
2166  * Generally, "ccb" should be the CCB supplied by the kernel.  "inccb"
2167  * should be the CCB that is copied in from the user.
2168  */
2169 static int
2170 passsendccb(struct cam_periph *periph, union ccb *ccb, union ccb *inccb)
2171 {
2172 	struct pass_softc *softc;
2173 	struct cam_periph_map_info mapinfo;
2174 	uint8_t *cmd;
2175 	xpt_opcode fc;
2176 	int error;
2177 
2178 	softc = (struct pass_softc *)periph->softc;
2179 
2180 	/*
2181 	 * There are some fields in the CCB header that need to be
2182 	 * preserved, the rest we get from the user.
2183 	 */
2184 	xpt_merge_ccb(ccb, inccb);
2185 
2186 	if (ccb->ccb_h.flags & CAM_CDB_POINTER) {
2187 		cmd = __builtin_alloca(ccb->csio.cdb_len);
2188 		error = copyin(ccb->csio.cdb_io.cdb_ptr, cmd, ccb->csio.cdb_len);
2189 		if (error)
2190 			return (error);
2191 		ccb->csio.cdb_io.cdb_ptr = cmd;
2192 	}
2193 
2194 	/*
2195 	 * Let cam_periph_mapmem do a sanity check on the data pointer format.
2196 	 * Even if no data transfer is needed, it's a cheap check and it
2197 	 * simplifies the code.
2198 	 */
2199 	fc = ccb->ccb_h.func_code;
2200 	if ((fc == XPT_SCSI_IO) || (fc == XPT_ATA_IO) || (fc == XPT_SMP_IO)
2201             || (fc == XPT_DEV_MATCH) || (fc == XPT_DEV_ADVINFO) || (fc == XPT_MMC_IO)
2202             || (fc == XPT_NVME_ADMIN) || (fc == XPT_NVME_IO)) {
2203 
2204 		bzero(&mapinfo, sizeof(mapinfo));
2205 
2206 		/*
2207 		 * cam_periph_mapmem calls into proc and vm functions that can
2208 		 * sleep as well as trigger I/O, so we can't hold the lock.
2209 		 * Dropping it here is reasonably safe.
2210 		 */
2211 		cam_periph_unlock(periph);
2212 		error = cam_periph_mapmem(ccb, &mapinfo, softc->maxio);
2213 		cam_periph_lock(periph);
2214 
2215 		/*
2216 		 * cam_periph_mapmem returned an error, we can't continue.
2217 		 * Return the error to the user.
2218 		 */
2219 		if (error)
2220 			return(error);
2221 	} else
2222 		/* Ensure that the unmap call later on is a no-op. */
2223 		mapinfo.num_bufs_used = 0;
2224 
2225 	/*
2226 	 * If the user wants us to perform any error recovery, then honor
2227 	 * that request.  Otherwise, it's up to the user to perform any
2228 	 * error recovery.
2229 	 */
2230 	cam_periph_runccb(ccb, (ccb->ccb_h.flags & CAM_PASS_ERR_RECOVER) ?
2231 	    passerror : NULL, /* cam_flags */ CAM_RETRY_SELTO,
2232 	    /* sense_flags */ SF_RETRY_UA | SF_NO_PRINT,
2233 	    softc->device_stats);
2234 
2235 	cam_periph_unlock(periph);
2236 	cam_periph_unmapmem(ccb, &mapinfo);
2237 	cam_periph_lock(periph);
2238 
2239 	ccb->ccb_h.cbfcnp = NULL;
2240 	ccb->ccb_h.periph_priv = inccb->ccb_h.periph_priv;
2241 	bcopy(ccb, inccb, sizeof(union ccb));
2242 
2243 	return(0);
2244 }
2245 
2246 static int
2247 passerror(union ccb *ccb, u_int32_t cam_flags, u_int32_t sense_flags)
2248 {
2249 	struct cam_periph *periph;
2250 	struct pass_softc *softc;
2251 
2252 	periph = xpt_path_periph(ccb->ccb_h.path);
2253 	softc = (struct pass_softc *)periph->softc;
2254 
2255 	return(cam_periph_error(ccb, cam_flags, sense_flags));
2256 }
2257