xref: /freebsd/sys/cam/scsi/scsi_pass.c (revision 77613ed4e3d635198014fd5b4723fb8eaa35d53e)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 1997, 1998, 2000 Justin T. Gibbs.
5  * Copyright (c) 1997, 1998, 1999 Kenneth D. Merry.
6  * All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions, and the following disclaimer,
13  *    without modification, immediately at the beginning of the file.
14  * 2. The name of the author may not be used to endorse or promote products
15  *    derived from this software without specific prior written permission.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR
21  * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 #include <sys/param.h>
31 #include <sys/systm.h>
32 #include <sys/kernel.h>
33 #include <sys/conf.h>
34 #include <sys/types.h>
35 #include <sys/bio.h>
36 #include <sys/bus.h>
37 #include <sys/devicestat.h>
38 #include <sys/errno.h>
39 #include <sys/fcntl.h>
40 #include <sys/malloc.h>
41 #include <sys/proc.h>
42 #include <sys/poll.h>
43 #include <sys/selinfo.h>
44 #include <sys/sdt.h>
45 #include <sys/sysent.h>
46 #include <sys/taskqueue.h>
47 #include <vm/uma.h>
48 #include <vm/vm.h>
49 #include <vm/vm_extern.h>
50 
51 #include <machine/bus.h>
52 
53 #include <cam/cam.h>
54 #include <cam/cam_ccb.h>
55 #include <cam/cam_periph.h>
56 #include <cam/cam_queue.h>
57 #include <cam/cam_xpt.h>
58 #include <cam/cam_xpt_periph.h>
59 #include <cam/cam_debug.h>
60 #include <cam/cam_compat.h>
61 #include <cam/cam_xpt_periph.h>
62 
63 #include <cam/scsi/scsi_pass.h>
64 
65 #define PERIPH_NAME "pass"
66 
67 typedef enum {
68 	PASS_FLAG_OPEN			= 0x01,
69 	PASS_FLAG_LOCKED		= 0x02,
70 	PASS_FLAG_INVALID		= 0x04,
71 	PASS_FLAG_INITIAL_PHYSPATH	= 0x08,
72 	PASS_FLAG_ZONE_INPROG		= 0x10,
73 	PASS_FLAG_ZONE_VALID		= 0x20,
74 	PASS_FLAG_UNMAPPED_CAPABLE	= 0x40,
75 	PASS_FLAG_ABANDONED_REF_SET	= 0x80
76 } pass_flags;
77 
78 typedef enum {
79 	PASS_STATE_NORMAL
80 } pass_state;
81 
82 typedef enum {
83 	PASS_CCB_BUFFER_IO,
84 	PASS_CCB_QUEUED_IO
85 } pass_ccb_types;
86 
87 #define ccb_type	ppriv_field0
88 #define ccb_ioreq	ppriv_ptr1
89 
90 /*
91  * The maximum number of memory segments we preallocate.
92  */
93 #define	PASS_MAX_SEGS	16
94 
95 typedef enum {
96 	PASS_IO_NONE		= 0x00,
97 	PASS_IO_USER_SEG_MALLOC	= 0x01,
98 	PASS_IO_KERN_SEG_MALLOC	= 0x02,
99 	PASS_IO_ABANDONED	= 0x04
100 } pass_io_flags;
101 
102 struct pass_io_req {
103 	union ccb			 ccb;
104 	union ccb			*alloced_ccb;
105 	union ccb			*user_ccb_ptr;
106 	camq_entry			 user_periph_links;
107 	ccb_ppriv_area			 user_periph_priv;
108 	struct cam_periph_map_info	 mapinfo;
109 	pass_io_flags			 flags;
110 	ccb_flags			 data_flags;
111 	int				 num_user_segs;
112 	bus_dma_segment_t		 user_segs[PASS_MAX_SEGS];
113 	int				 num_kern_segs;
114 	bus_dma_segment_t		 kern_segs[PASS_MAX_SEGS];
115 	bus_dma_segment_t		*user_segptr;
116 	bus_dma_segment_t		*kern_segptr;
117 	int				 num_bufs;
118 	uint32_t			 dirs[CAM_PERIPH_MAXMAPS];
119 	uint32_t			 lengths[CAM_PERIPH_MAXMAPS];
120 	uint8_t				*user_bufs[CAM_PERIPH_MAXMAPS];
121 	uint8_t				*kern_bufs[CAM_PERIPH_MAXMAPS];
122 	struct bintime			 start_time;
123 	TAILQ_ENTRY(pass_io_req)	 links;
124 };
125 
126 struct pass_softc {
127 	pass_state		  state;
128 	pass_flags		  flags;
129 	uint8_t		  pd_type;
130 	int			  open_count;
131 	u_int		 	  maxio;
132 	struct devstat		 *device_stats;
133 	struct cdev		 *dev;
134 	struct cdev		 *alias_dev;
135 	struct task		  add_physpath_task;
136 	struct task		  shutdown_kqueue_task;
137 	struct selinfo		  read_select;
138 	TAILQ_HEAD(, pass_io_req) incoming_queue;
139 	TAILQ_HEAD(, pass_io_req) active_queue;
140 	TAILQ_HEAD(, pass_io_req) abandoned_queue;
141 	TAILQ_HEAD(, pass_io_req) done_queue;
142 	struct cam_periph	 *periph;
143 	char			  zone_name[12];
144 	char			  io_zone_name[12];
145 	uma_zone_t		  pass_zone;
146 	uma_zone_t		  pass_io_zone;
147 	size_t			  io_zone_size;
148 };
149 
150 static	d_open_t	passopen;
151 static	d_close_t	passclose;
152 static	d_ioctl_t	passioctl;
153 static	d_ioctl_t	passdoioctl;
154 static	d_poll_t	passpoll;
155 static	d_kqfilter_t	passkqfilter;
156 static	void		passreadfiltdetach(struct knote *kn);
157 static	int		passreadfilt(struct knote *kn, long hint);
158 
159 static	periph_init_t	passinit;
160 static	periph_ctor_t	passregister;
161 static	periph_oninv_t	passoninvalidate;
162 static	periph_dtor_t	passcleanup;
163 static	periph_start_t	passstart;
164 static	void		pass_shutdown_kqueue(void *context, int pending);
165 static	void		pass_add_physpath(void *context, int pending);
166 static	void		passasync(void *callback_arg, uint32_t code,
167 				  struct cam_path *path, void *arg);
168 static	void		passdone(struct cam_periph *periph,
169 				 union ccb *done_ccb);
170 static	int		passcreatezone(struct cam_periph *periph);
171 static	void		passiocleanup(struct pass_softc *softc,
172 				      struct pass_io_req *io_req);
173 static	int		passcopysglist(struct cam_periph *periph,
174 				       struct pass_io_req *io_req,
175 				       ccb_flags direction);
176 static	int		passmemsetup(struct cam_periph *periph,
177 				     struct pass_io_req *io_req);
178 static	int		passmemdone(struct cam_periph *periph,
179 				    struct pass_io_req *io_req);
180 static	int		passerror(union ccb *ccb, uint32_t cam_flags,
181 				  uint32_t sense_flags);
182 static 	int		passsendccb(struct cam_periph *periph, union ccb *ccb,
183 				    union ccb *inccb);
184 static	void		passflags(union ccb *ccb, uint32_t *cam_flags,
185 				  uint32_t *sense_flags);
186 
187 static struct periph_driver passdriver =
188 {
189 	passinit, PERIPH_NAME,
190 	TAILQ_HEAD_INITIALIZER(passdriver.units), /* generation */ 0
191 };
192 
193 PERIPHDRIVER_DECLARE(pass, passdriver);
194 
195 static struct cdevsw pass_cdevsw = {
196 	.d_version =	D_VERSION,
197 	.d_flags =	D_TRACKCLOSE,
198 	.d_open =	passopen,
199 	.d_close =	passclose,
200 	.d_ioctl =	passioctl,
201 	.d_poll = 	passpoll,
202 	.d_kqfilter = 	passkqfilter,
203 	.d_name =	PERIPH_NAME,
204 };
205 
206 static const struct filterops passread_filtops = {
207 	.f_isfd	=	1,
208 	.f_detach =	passreadfiltdetach,
209 	.f_event =	passreadfilt
210 };
211 
212 static MALLOC_DEFINE(M_SCSIPASS, "scsi_pass", "scsi passthrough buffers");
213 
214 static void
passinit(void)215 passinit(void)
216 {
217 	cam_status status;
218 
219 	/*
220 	 * Install a global async callback.  This callback will
221 	 * receive async callbacks like "new device found".
222 	 */
223 	status = xpt_register_async(AC_FOUND_DEVICE, passasync, NULL, NULL);
224 
225 	if (status != CAM_REQ_CMP) {
226 		printf("pass: Failed to attach master async callback "
227 		       "due to status 0x%x!\n", status);
228 	}
229 
230 }
231 
232 static void
passrejectios(struct cam_periph * periph)233 passrejectios(struct cam_periph *periph)
234 {
235 	struct pass_io_req *io_req, *io_req2;
236 	struct pass_softc *softc;
237 
238 	softc = (struct pass_softc *)periph->softc;
239 
240 	/*
241 	 * The user can no longer get status for I/O on the done queue, so
242 	 * clean up all outstanding I/O on the done queue.
243 	 */
244 	TAILQ_FOREACH_SAFE(io_req, &softc->done_queue, links, io_req2) {
245 		TAILQ_REMOVE(&softc->done_queue, io_req, links);
246 		passiocleanup(softc, io_req);
247 		uma_zfree(softc->pass_zone, io_req);
248 	}
249 
250 	/*
251 	 * The underlying device is gone, so we can't issue these I/Os.
252 	 * The devfs node has been shut down, so we can't return status to
253 	 * the user.  Free any I/O left on the incoming queue.
254 	 */
255 	TAILQ_FOREACH_SAFE(io_req, &softc->incoming_queue, links, io_req2) {
256 		TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
257 		passiocleanup(softc, io_req);
258 		uma_zfree(softc->pass_zone, io_req);
259 	}
260 
261 	/*
262 	 * Normally we would put I/Os on the abandoned queue and acquire a
263 	 * reference when we saw the final close.  But, the device went
264 	 * away and devfs may have moved everything off to deadfs by the
265 	 * time the I/O done callback is called; as a result, we won't see
266 	 * any more closes.  So, if we have any active I/Os, we need to put
267 	 * them on the abandoned queue.  When the abandoned queue is empty,
268 	 * we'll release the remaining reference (see below) to the peripheral.
269 	 */
270 	TAILQ_FOREACH_SAFE(io_req, &softc->active_queue, links, io_req2) {
271 		TAILQ_REMOVE(&softc->active_queue, io_req, links);
272 		io_req->flags |= PASS_IO_ABANDONED;
273 		TAILQ_INSERT_TAIL(&softc->abandoned_queue, io_req, links);
274 	}
275 
276 	/*
277 	 * If we put any I/O on the abandoned queue, acquire a reference.
278 	 */
279 	if ((!TAILQ_EMPTY(&softc->abandoned_queue))
280 	 && ((softc->flags & PASS_FLAG_ABANDONED_REF_SET) == 0)) {
281 		cam_periph_doacquire(periph);
282 		softc->flags |= PASS_FLAG_ABANDONED_REF_SET;
283 	}
284 }
285 
286 static void
passdevgonecb(void * arg)287 passdevgonecb(void *arg)
288 {
289 	struct cam_periph *periph;
290 	struct mtx *mtx;
291 	struct pass_softc *softc;
292 	int i;
293 
294 	periph = (struct cam_periph *)arg;
295 	mtx = cam_periph_mtx(periph);
296 	mtx_lock(mtx);
297 
298 	softc = (struct pass_softc *)periph->softc;
299 	KASSERT(softc->open_count >= 0, ("Negative open count %d",
300 		softc->open_count));
301 
302 	/*
303 	 * When we get this callback, we will get no more close calls from
304 	 * devfs.  So if we have any dangling opens, we need to release the
305 	 * reference held for that particular context.
306 	 */
307 	for (i = 0; i < softc->open_count; i++)
308 		cam_periph_release_locked(periph);
309 
310 	softc->open_count = 0;
311 
312 	/*
313 	 * Release the reference held for the device node, it is gone now.
314 	 * Accordingly, inform all queued I/Os of their fate.
315 	 */
316 	cam_periph_release_locked(periph);
317 	passrejectios(periph);
318 
319 	/*
320 	 * We reference the SIM lock directly here, instead of using
321 	 * cam_periph_unlock().  The reason is that the final call to
322 	 * cam_periph_release_locked() above could result in the periph
323 	 * getting freed.  If that is the case, dereferencing the periph
324 	 * with a cam_periph_unlock() call would cause a page fault.
325 	 */
326 	mtx_unlock(mtx);
327 
328 	/*
329 	 * We have to remove our kqueue context from a thread because it
330 	 * may sleep.  It would be nice if we could get a callback from
331 	 * kqueue when it is done cleaning up resources.
332 	 */
333 	taskqueue_enqueue(taskqueue_thread, &softc->shutdown_kqueue_task);
334 }
335 
336 static void
passoninvalidate(struct cam_periph * periph)337 passoninvalidate(struct cam_periph *periph)
338 {
339 	struct pass_softc *softc;
340 
341 	softc = (struct pass_softc *)periph->softc;
342 
343 	/*
344 	 * De-register any async callbacks.
345 	 */
346 	xpt_register_async(0, passasync, periph, periph->path);
347 
348 	softc->flags |= PASS_FLAG_INVALID;
349 
350 	/*
351 	 * Tell devfs this device has gone away, and ask for a callback
352 	 * when it has cleaned up its state.
353 	 */
354 	destroy_dev_sched_cb(softc->dev, passdevgonecb, periph);
355 }
356 
357 static void
passcleanup(struct cam_periph * periph)358 passcleanup(struct cam_periph *periph)
359 {
360 	struct pass_softc *softc;
361 
362 	softc = (struct pass_softc *)periph->softc;
363 
364 	cam_periph_assert(periph, MA_OWNED);
365 	KASSERT(TAILQ_EMPTY(&softc->active_queue),
366 		("%s called when there are commands on the active queue!\n",
367 		__func__));
368 	KASSERT(TAILQ_EMPTY(&softc->abandoned_queue),
369 		("%s called when there are commands on the abandoned queue!\n",
370 		__func__));
371 	KASSERT(TAILQ_EMPTY(&softc->incoming_queue),
372 		("%s called when there are commands on the incoming queue!\n",
373 		__func__));
374 	KASSERT(TAILQ_EMPTY(&softc->done_queue),
375 		("%s called when there are commands on the done queue!\n",
376 		__func__));
377 
378 	devstat_remove_entry(softc->device_stats);
379 
380 	cam_periph_unlock(periph);
381 
382 	/*
383 	 * We call taskqueue_drain() for the physpath task to make sure it
384 	 * is complete.  We drop the lock because this can potentially
385 	 * sleep.  XXX KDM that is bad.  Need a way to get a callback when
386 	 * a taskqueue is drained.
387 	 *
388  	 * Note that we don't drain the kqueue shutdown task queue.  This
389 	 * is because we hold a reference on the periph for kqueue, and
390 	 * release that reference from the kqueue shutdown task queue.  So
391 	 * we cannot come into this routine unless we've released that
392 	 * reference.  Also, because that could be the last reference, we
393 	 * could be called from the cam_periph_release() call in
394 	 * pass_shutdown_kqueue().  In that case, the taskqueue_drain()
395 	 * would deadlock.  It would be preferable if we had a way to
396 	 * get a callback when a taskqueue is done.
397 	 */
398 	taskqueue_drain(taskqueue_thread, &softc->add_physpath_task);
399 
400 	/*
401 	 * It should be safe to destroy the zones from here, because all
402 	 * of the references to this peripheral have been freed, and all
403 	 * I/O has been terminated and freed.  We check the zones for NULL
404 	 * because they may not have been allocated yet if the device went
405 	 * away before any asynchronous I/O has been issued.
406 	 */
407 	if (softc->pass_zone != NULL)
408 		uma_zdestroy(softc->pass_zone);
409 	if (softc->pass_io_zone != NULL)
410 		uma_zdestroy(softc->pass_io_zone);
411 
412 	cam_periph_lock(periph);
413 
414 	free(softc, M_DEVBUF);
415 }
416 
417 static void
pass_shutdown_kqueue(void * context,int pending)418 pass_shutdown_kqueue(void *context, int pending)
419 {
420 	struct cam_periph *periph;
421 	struct pass_softc *softc;
422 
423 	periph = context;
424 	softc = periph->softc;
425 
426 	knlist_clear(&softc->read_select.si_note, /*is_locked*/ 0);
427 	knlist_destroy(&softc->read_select.si_note);
428 
429 	/*
430 	 * Release the reference we held for kqueue.
431 	 */
432 	cam_periph_release(periph);
433 }
434 
435 static void
pass_add_physpath(void * context,int pending)436 pass_add_physpath(void *context, int pending)
437 {
438 	struct cam_periph *periph;
439 	struct pass_softc *softc;
440 	struct mtx *mtx;
441 	char *physpath;
442 
443 	/*
444 	 * If we have one, create a devfs alias for our
445 	 * physical path.
446 	 */
447 	periph = context;
448 	softc = periph->softc;
449 	physpath = malloc(MAXPATHLEN, M_DEVBUF, M_WAITOK);
450 	mtx = cam_periph_mtx(periph);
451 	mtx_lock(mtx);
452 
453 	if (periph->flags & CAM_PERIPH_INVALID)
454 		goto out;
455 
456 	if (xpt_getattr(physpath, MAXPATHLEN,
457 			"GEOM::physpath", periph->path) == 0
458 	 && strlen(physpath) != 0) {
459 		mtx_unlock(mtx);
460 		make_dev_physpath_alias(MAKEDEV_WAITOK | MAKEDEV_CHECKNAME,
461 				&softc->alias_dev, softc->dev,
462 				softc->alias_dev, physpath);
463 		mtx_lock(mtx);
464 	}
465 
466 out:
467 	/*
468 	 * Now that we've made our alias, we no longer have to have a
469 	 * reference to the device.
470 	 */
471 	if ((softc->flags & PASS_FLAG_INITIAL_PHYSPATH) == 0)
472 		softc->flags |= PASS_FLAG_INITIAL_PHYSPATH;
473 
474 	/*
475 	 * We always acquire a reference to the periph before queueing this
476 	 * task queue function, so it won't go away before we run.
477 	 */
478 	while (pending-- > 0)
479 		cam_periph_release_locked(periph);
480 	mtx_unlock(mtx);
481 
482 	free(physpath, M_DEVBUF);
483 }
484 
485 static void
passasync(void * callback_arg,uint32_t code,struct cam_path * path,void * arg)486 passasync(void *callback_arg, uint32_t code,
487 	  struct cam_path *path, void *arg)
488 {
489 	struct cam_periph *periph;
490 
491 	periph = (struct cam_periph *)callback_arg;
492 
493 	switch (code) {
494 	case AC_FOUND_DEVICE:
495 	{
496 		struct ccb_getdev *cgd;
497 		cam_status status;
498 
499 		cgd = (struct ccb_getdev *)arg;
500 		if (cgd == NULL)
501 			break;
502 
503 		/*
504 		 * Allocate a peripheral instance for
505 		 * this device and start the probe
506 		 * process.
507 		 */
508 		status = cam_periph_alloc(passregister, passoninvalidate,
509 					  passcleanup, passstart, PERIPH_NAME,
510 					  CAM_PERIPH_BIO, path,
511 					  passasync, AC_FOUND_DEVICE, cgd);
512 
513 		if (status != CAM_REQ_CMP
514 		 && status != CAM_REQ_INPROG) {
515 			const struct cam_status_entry *entry;
516 
517 			entry = cam_fetch_status_entry(status);
518 
519 			printf("passasync: Unable to attach new device "
520 			       "due to status %#x: %s\n", status, entry ?
521 			       entry->status_text : "Unknown");
522 		}
523 
524 		break;
525 	}
526 	case AC_ADVINFO_CHANGED:
527 	{
528 		uintptr_t buftype;
529 
530 		buftype = (uintptr_t)arg;
531 		if (buftype == CDAI_TYPE_PHYS_PATH) {
532 			struct pass_softc *softc;
533 
534 			softc = (struct pass_softc *)periph->softc;
535 			/*
536 			 * Acquire a reference to the periph before we
537 			 * start the taskqueue, so that we don't run into
538 			 * a situation where the periph goes away before
539 			 * the task queue has a chance to run.
540 			 */
541 			if (cam_periph_acquire(periph) != 0)
542 				break;
543 
544 			taskqueue_enqueue(taskqueue_thread,
545 					  &softc->add_physpath_task);
546 		}
547 		break;
548 	}
549 	default:
550 		cam_periph_async(periph, code, path, arg);
551 		break;
552 	}
553 }
554 
555 static cam_status
passregister(struct cam_periph * periph,void * arg)556 passregister(struct cam_periph *periph, void *arg)
557 {
558 	struct pass_softc *softc;
559 	struct ccb_getdev *cgd;
560 	struct ccb_pathinq cpi;
561 	struct make_dev_args args;
562 	int error, no_tags;
563 
564 	cgd = (struct ccb_getdev *)arg;
565 	if (cgd == NULL) {
566 		printf("%s: no getdev CCB, can't register device\n", __func__);
567 		return(CAM_REQ_CMP_ERR);
568 	}
569 
570 	softc = (struct pass_softc *)malloc(sizeof(*softc),
571 					    M_DEVBUF, M_NOWAIT);
572 
573 	if (softc == NULL) {
574 		printf("%s: Unable to probe new device. "
575 		       "Unable to allocate softc\n", __func__);
576 		return(CAM_REQ_CMP_ERR);
577 	}
578 
579 	bzero(softc, sizeof(*softc));
580 	softc->state = PASS_STATE_NORMAL;
581 	if (cgd->protocol == PROTO_SCSI || cgd->protocol == PROTO_ATAPI)
582 		softc->pd_type = SID_TYPE(&cgd->inq_data);
583 	else if (cgd->protocol == PROTO_SATAPM)
584 		softc->pd_type = T_ENCLOSURE;
585 	else
586 		softc->pd_type = T_DIRECT;
587 
588 	periph->softc = softc;
589 	softc->periph = periph;
590 	TAILQ_INIT(&softc->incoming_queue);
591 	TAILQ_INIT(&softc->active_queue);
592 	TAILQ_INIT(&softc->abandoned_queue);
593 	TAILQ_INIT(&softc->done_queue);
594 	snprintf(softc->zone_name, sizeof(softc->zone_name), "%s%d",
595 		 periph->periph_name, periph->unit_number);
596 	snprintf(softc->io_zone_name, sizeof(softc->io_zone_name), "%s%dIO",
597 		 periph->periph_name, periph->unit_number);
598 	softc->io_zone_size = maxphys;
599 	knlist_init_mtx(&softc->read_select.si_note, cam_periph_mtx(periph));
600 
601 	xpt_path_inq(&cpi, periph->path);
602 
603 	if (cpi.maxio == 0)
604 		softc->maxio = DFLTPHYS;	/* traditional default */
605 	else if (cpi.maxio > maxphys)
606 		softc->maxio = maxphys;		/* for safety */
607 	else
608 		softc->maxio = cpi.maxio;	/* real value */
609 
610 	if (cpi.hba_misc & PIM_UNMAPPED)
611 		softc->flags |= PASS_FLAG_UNMAPPED_CAPABLE;
612 
613 	/*
614 	 * We pass in 0 for a blocksize, since we don't know what the blocksize
615 	 * of this device is, if it even has a blocksize.
616 	 *
617 	 * Note: no_tags is valid only for SCSI peripherals, but we don't do any
618 	 * devstat accounting for tags on any other transport. SCSI is the only
619 	 * transport that uses the tag_action (ata has only vestigial references
620 	 * to it, others ignore it entirely).
621 	 */
622 	cam_periph_unlock(periph);
623 	no_tags = (cgd->inq_data.flags & SID_CmdQue) == 0;
624 	softc->device_stats = devstat_new_entry(PERIPH_NAME,
625 			  periph->unit_number, 0,
626 			  DEVSTAT_NO_BLOCKSIZE
627 			  | (no_tags ? DEVSTAT_NO_ORDERED_TAGS : 0),
628 			  softc->pd_type |
629 			  XPORT_DEVSTAT_TYPE(cpi.transport) |
630 			  DEVSTAT_TYPE_PASS,
631 			  DEVSTAT_PRIORITY_PASS);
632 
633 	/*
634 	 * Initialize the taskqueue handler for shutting down kqueue.
635 	 */
636 	TASK_INIT(&softc->shutdown_kqueue_task, /*priority*/ 0,
637 		  pass_shutdown_kqueue, periph);
638 
639 	/*
640 	 * Acquire a reference to the periph that we can release once we've
641 	 * cleaned up the kqueue.
642 	 */
643 	if (cam_periph_acquire(periph) != 0) {
644 		xpt_print(periph->path, "%s: lost periph during "
645 			  "registration!\n", __func__);
646 		cam_periph_lock(periph);
647 		return (CAM_REQ_CMP_ERR);
648 	}
649 
650 	/*
651 	 * Acquire a reference to the periph before we create the devfs
652 	 * instance for it.  We'll release this reference once the devfs
653 	 * instance has been freed.
654 	 */
655 	if (cam_periph_acquire(periph) != 0) {
656 		xpt_print(periph->path, "%s: lost periph during "
657 			  "registration!\n", __func__);
658 		cam_periph_lock(periph);
659 		return (CAM_REQ_CMP_ERR);
660 	}
661 
662 	/* Register the device */
663 	make_dev_args_init(&args);
664 	args.mda_devsw = &pass_cdevsw;
665 	args.mda_unit = periph->unit_number;
666 	args.mda_uid = UID_ROOT;
667 	args.mda_gid = GID_OPERATOR;
668 	args.mda_mode = 0600;
669 	args.mda_si_drv1 = periph;
670 	args.mda_flags = MAKEDEV_NOWAIT;
671 	error = make_dev_s(&args, &softc->dev, "%s%d", periph->periph_name,
672 	    periph->unit_number);
673 	if (error != 0) {
674 		cam_periph_lock(periph);
675 		cam_periph_release_locked(periph);
676 		return (CAM_REQ_CMP_ERR);
677 	}
678 
679 	/*
680 	 * Hold a reference to the periph before we create the physical
681 	 * path alias so it can't go away.
682 	 */
683 	if (cam_periph_acquire(periph) != 0) {
684 		xpt_print(periph->path, "%s: lost periph during "
685 			  "registration!\n", __func__);
686 		cam_periph_lock(periph);
687 		return (CAM_REQ_CMP_ERR);
688 	}
689 
690 	cam_periph_lock(periph);
691 
692 	TASK_INIT(&softc->add_physpath_task, /*priority*/0,
693 		  pass_add_physpath, periph);
694 
695 	/*
696 	 * See if physical path information is already available.
697 	 */
698 	taskqueue_enqueue(taskqueue_thread, &softc->add_physpath_task);
699 
700 	/*
701 	 * Add an async callback so that we get notified if
702 	 * this device goes away or its physical path
703 	 * (stored in the advanced info data of the EDT) has
704 	 * changed.
705 	 */
706 	xpt_register_async(AC_LOST_DEVICE | AC_ADVINFO_CHANGED,
707 			   passasync, periph, periph->path);
708 
709 	if (bootverbose)
710 		xpt_announce_periph(periph, NULL);
711 
712 	return(CAM_REQ_CMP);
713 }
714 
715 static int
passopen(struct cdev * dev,int flags,int fmt,struct thread * td)716 passopen(struct cdev *dev, int flags, int fmt, struct thread *td)
717 {
718 	struct cam_periph *periph;
719 	struct pass_softc *softc;
720 	int error;
721 
722 	periph = (struct cam_periph *)dev->si_drv1;
723 	if (cam_periph_acquire(periph) != 0)
724 		return (ENXIO);
725 
726 	cam_periph_lock(periph);
727 
728 	softc = (struct pass_softc *)periph->softc;
729 
730 	if (softc->flags & PASS_FLAG_INVALID) {
731 		cam_periph_release_locked(periph);
732 		cam_periph_unlock(periph);
733 		return(ENXIO);
734 	}
735 
736 	/*
737 	 * Don't allow access when we're running at a high securelevel.
738 	 */
739 	error = securelevel_gt(td->td_ucred, 1);
740 	if (error) {
741 		cam_periph_release_locked(periph);
742 		cam_periph_unlock(periph);
743 		return(error);
744 	}
745 
746 	/*
747 	 * Only allow read-write access.
748 	 */
749 	if (((flags & FWRITE) == 0) || ((flags & FREAD) == 0)) {
750 		cam_periph_release_locked(periph);
751 		cam_periph_unlock(periph);
752 		return(EPERM);
753 	}
754 
755 	/*
756 	 * We don't allow nonblocking access.
757 	 */
758 	if ((flags & O_NONBLOCK) != 0) {
759 		xpt_print(periph->path, "can't do nonblocking access\n");
760 		cam_periph_release_locked(periph);
761 		cam_periph_unlock(periph);
762 		return(EINVAL);
763 	}
764 
765 	softc->open_count++;
766 
767 	cam_periph_unlock(periph);
768 
769 	return (error);
770 }
771 
772 static int
passclose(struct cdev * dev,int flag,int fmt,struct thread * td)773 passclose(struct cdev *dev, int flag, int fmt, struct thread *td)
774 {
775 	struct 	cam_periph *periph;
776 	struct  pass_softc *softc;
777 	struct mtx *mtx;
778 
779 	periph = (struct cam_periph *)dev->si_drv1;
780 	mtx = cam_periph_mtx(periph);
781 	mtx_lock(mtx);
782 
783 	softc = periph->softc;
784 	softc->open_count--;
785 
786 	if (softc->open_count == 0) {
787 		struct pass_io_req *io_req, *io_req2;
788 
789 		TAILQ_FOREACH_SAFE(io_req, &softc->done_queue, links, io_req2) {
790 			TAILQ_REMOVE(&softc->done_queue, io_req, links);
791 			passiocleanup(softc, io_req);
792 			uma_zfree(softc->pass_zone, io_req);
793 		}
794 
795 		TAILQ_FOREACH_SAFE(io_req, &softc->incoming_queue, links,
796 				   io_req2) {
797 			TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
798 			passiocleanup(softc, io_req);
799 			uma_zfree(softc->pass_zone, io_req);
800 		}
801 
802 		/*
803 		 * If there are any active I/Os, we need to forcibly acquire a
804 		 * reference to the peripheral so that we don't go away
805 		 * before they complete.  We'll release the reference when
806 		 * the abandoned queue is empty.
807 		 */
808 		io_req = TAILQ_FIRST(&softc->active_queue);
809 		if ((io_req != NULL)
810 		 && (softc->flags & PASS_FLAG_ABANDONED_REF_SET) == 0) {
811 			cam_periph_doacquire(periph);
812 			softc->flags |= PASS_FLAG_ABANDONED_REF_SET;
813 		}
814 
815 		/*
816 		 * Since the I/O in the active queue is not under our
817 		 * control, just set a flag so that we can clean it up when
818 		 * it completes and put it on the abandoned queue.  This
819 		 * will prevent our sending spurious completions in the
820 		 * event that the device is opened again before these I/Os
821 		 * complete.
822 		 */
823 		TAILQ_FOREACH_SAFE(io_req, &softc->active_queue, links,
824 				   io_req2) {
825 			TAILQ_REMOVE(&softc->active_queue, io_req, links);
826 			io_req->flags |= PASS_IO_ABANDONED;
827 			TAILQ_INSERT_TAIL(&softc->abandoned_queue, io_req,
828 					  links);
829 		}
830 	}
831 
832 	cam_periph_release_locked(periph);
833 
834 	/*
835 	 * We reference the lock directly here, instead of using
836 	 * cam_periph_unlock().  The reason is that the call to
837 	 * cam_periph_release_locked() above could result in the periph
838 	 * getting freed.  If that is the case, dereferencing the periph
839 	 * with a cam_periph_unlock() call would cause a page fault.
840 	 *
841 	 * cam_periph_release() avoids this problem using the same method,
842 	 * but we're manually acquiring and dropping the lock here to
843 	 * protect the open count and avoid another lock acquisition and
844 	 * release.
845 	 */
846 	mtx_unlock(mtx);
847 
848 	return (0);
849 }
850 
851 static void
passstart(struct cam_periph * periph,union ccb * start_ccb)852 passstart(struct cam_periph *periph, union ccb *start_ccb)
853 {
854 	struct pass_softc *softc;
855 
856 	softc = (struct pass_softc *)periph->softc;
857 
858 	switch (softc->state) {
859 	case PASS_STATE_NORMAL: {
860 		struct pass_io_req *io_req;
861 
862 		/*
863 		 * Check for any queued I/O requests that require an
864 		 * allocated slot.
865 		 */
866 		io_req = TAILQ_FIRST(&softc->incoming_queue);
867 		if (io_req == NULL) {
868 			xpt_release_ccb(start_ccb);
869 			break;
870 		}
871 		TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
872 		TAILQ_INSERT_TAIL(&softc->active_queue, io_req, links);
873 		/*
874 		 * Merge the user's CCB into the allocated CCB.
875 		 */
876 		xpt_merge_ccb(start_ccb, &io_req->ccb);
877 		start_ccb->ccb_h.ccb_type = PASS_CCB_QUEUED_IO;
878 		start_ccb->ccb_h.ccb_ioreq = io_req;
879 		start_ccb->ccb_h.cbfcnp = passdone;
880 		io_req->alloced_ccb = start_ccb;
881 		binuptime(&io_req->start_time);
882 		devstat_start_transaction(softc->device_stats,
883 					  &io_req->start_time);
884 
885 		xpt_action(start_ccb);
886 
887 		/*
888 		 * If we have any more I/O waiting, schedule ourselves again.
889 		 */
890 		if (!TAILQ_EMPTY(&softc->incoming_queue))
891 			xpt_schedule(periph, CAM_PRIORITY_NORMAL);
892 		break;
893 	}
894 	default:
895 		break;
896 	}
897 }
898 
899 static void
passdone(struct cam_periph * periph,union ccb * done_ccb)900 passdone(struct cam_periph *periph, union ccb *done_ccb)
901 {
902 	struct pass_softc *softc;
903 	struct ccb_hdr *hdr;
904 
905 	softc = (struct pass_softc *)periph->softc;
906 
907 	cam_periph_assert(periph, MA_OWNED);
908 
909 	hdr = &done_ccb->ccb_h;
910 	switch (hdr->ccb_type) {
911 	case PASS_CCB_QUEUED_IO: {
912 		struct pass_io_req *io_req;
913 
914 		io_req = hdr->ccb_ioreq;
915 #if 0
916 		xpt_print(periph->path, "%s: called for user CCB %p\n",
917 			  __func__, io_req->user_ccb_ptr);
918 #endif
919 		if (((hdr->status & CAM_STATUS_MASK) != CAM_REQ_CMP) &&
920 		    ((io_req->flags & PASS_IO_ABANDONED) == 0)) {
921 			int error;
922 			uint32_t cam_flags, sense_flags;
923 
924 			passflags(done_ccb, &cam_flags, &sense_flags);
925 			error = passerror(done_ccb, cam_flags, sense_flags);
926 
927 			if (error == ERESTART) {
928 				KASSERT(((sense_flags & SF_NO_RETRY) == 0),
929 				    ("passerror returned ERESTART with no retry requested\n"));
930 				return;
931 			}
932 		}
933 
934 		/*
935 		 * Copy the allocated CCB contents back to the malloced CCB
936 		 * so we can give status back to the user when he requests it.
937 		 */
938 		bcopy(done_ccb, &io_req->ccb, sizeof(*done_ccb));
939 
940 		/*
941 		 * Log data/transaction completion with devstat(9).
942 		 */
943 		switch (hdr->func_code) {
944 		case XPT_SCSI_IO:
945 			devstat_end_transaction(softc->device_stats,
946 			    done_ccb->csio.dxfer_len - done_ccb->csio.resid,
947 			    done_ccb->csio.tag_action & 0x3,
948 			    ((hdr->flags & CAM_DIR_MASK) ==
949 			    CAM_DIR_NONE) ? DEVSTAT_NO_DATA :
950 			    (hdr->flags & CAM_DIR_OUT) ?
951 			    DEVSTAT_WRITE : DEVSTAT_READ, NULL,
952 			    &io_req->start_time);
953 			break;
954 		case XPT_ATA_IO:
955 			devstat_end_transaction(softc->device_stats,
956 			    done_ccb->ataio.dxfer_len - done_ccb->ataio.resid,
957 			    0, /* Not used in ATA */
958 			    ((hdr->flags & CAM_DIR_MASK) ==
959 			    CAM_DIR_NONE) ? DEVSTAT_NO_DATA :
960 			    (hdr->flags & CAM_DIR_OUT) ?
961 			    DEVSTAT_WRITE : DEVSTAT_READ, NULL,
962 			    &io_req->start_time);
963 			break;
964 		case XPT_SMP_IO:
965 			/*
966 			 * XXX KDM this isn't quite right, but there isn't
967 			 * currently an easy way to represent a bidirectional
968 			 * transfer in devstat.  The only way to do it
969 			 * and have the byte counts come out right would
970 			 * mean that we would have to record two
971 			 * transactions, one for the request and one for the
972 			 * response.  For now, so that we report something,
973 			 * just treat the entire thing as a read.
974 			 */
975 			devstat_end_transaction(softc->device_stats,
976 			    done_ccb->smpio.smp_request_len +
977 			    done_ccb->smpio.smp_response_len,
978 			    DEVSTAT_TAG_SIMPLE, DEVSTAT_READ, NULL,
979 			    &io_req->start_time);
980 			break;
981 		/* XXX XPT_NVME_IO and XPT_NVME_ADMIN need cases here for resid */
982 		default:
983 			devstat_end_transaction(softc->device_stats, 0,
984 			    DEVSTAT_TAG_NONE, DEVSTAT_NO_DATA, NULL,
985 			    &io_req->start_time);
986 			break;
987 		}
988 
989 		/*
990 		 * In the normal case, take the completed I/O off of the
991 		 * active queue and put it on the done queue.  Notitfy the
992 		 * user that we have a completed I/O.
993 		 */
994 		if ((io_req->flags & PASS_IO_ABANDONED) == 0) {
995 			TAILQ_REMOVE(&softc->active_queue, io_req, links);
996 			TAILQ_INSERT_TAIL(&softc->done_queue, io_req, links);
997 			selwakeuppri(&softc->read_select, PRIBIO);
998 			KNOTE_LOCKED(&softc->read_select.si_note, 0);
999 		} else {
1000 			/*
1001 			 * In the case of an abandoned I/O (final close
1002 			 * without fetching the I/O), take it off of the
1003 			 * abandoned queue and free it.
1004 			 */
1005 			TAILQ_REMOVE(&softc->abandoned_queue, io_req, links);
1006 			passiocleanup(softc, io_req);
1007 			uma_zfree(softc->pass_zone, io_req);
1008 
1009 			/*
1010 			 * Release the done_ccb here, since we may wind up
1011 			 * freeing the peripheral when we decrement the
1012 			 * reference count below.
1013 			 */
1014 			xpt_release_ccb(done_ccb);
1015 
1016 			/*
1017 			 * If the abandoned queue is empty, we can release
1018 			 * our reference to the periph since we won't have
1019 			 * any more completions coming.
1020 			 */
1021 			if ((TAILQ_EMPTY(&softc->abandoned_queue))
1022 			 && (softc->flags & PASS_FLAG_ABANDONED_REF_SET)) {
1023 				softc->flags &= ~PASS_FLAG_ABANDONED_REF_SET;
1024 				cam_periph_release_locked(periph);
1025 			}
1026 
1027 			/*
1028 			 * We have already released the CCB, so we can
1029 			 * return.
1030 			 */
1031 			return;
1032 		}
1033 		break;
1034 	}
1035 	}
1036 	xpt_release_ccb(done_ccb);
1037 }
1038 
1039 static int
passcreatezone(struct cam_periph * periph)1040 passcreatezone(struct cam_periph *periph)
1041 {
1042 	struct pass_softc *softc;
1043 	int error;
1044 
1045 	error = 0;
1046 	softc = (struct pass_softc *)periph->softc;
1047 
1048 	cam_periph_assert(periph, MA_OWNED);
1049 	KASSERT(((softc->flags & PASS_FLAG_ZONE_VALID) == 0),
1050 		("%s called when the pass(4) zone is valid!\n", __func__));
1051 	KASSERT((softc->pass_zone == NULL),
1052 		("%s called when the pass(4) zone is allocated!\n", __func__));
1053 
1054 	if ((softc->flags & PASS_FLAG_ZONE_INPROG) == 0) {
1055 		/*
1056 		 * We're the first context through, so we need to create
1057 		 * the pass(4) UMA zone for I/O requests.
1058 		 */
1059 		softc->flags |= PASS_FLAG_ZONE_INPROG;
1060 
1061 		/*
1062 		 * uma_zcreate() does a blocking (M_WAITOK) allocation,
1063 		 * so we cannot hold a mutex while we call it.
1064 		 */
1065 		cam_periph_unlock(periph);
1066 
1067 		softc->pass_zone = uma_zcreate(softc->zone_name,
1068 		    sizeof(struct pass_io_req), NULL, NULL, NULL, NULL,
1069 		    /*align*/ 0, /*flags*/ 0);
1070 
1071 		softc->pass_io_zone = uma_zcreate(softc->io_zone_name,
1072 		    softc->io_zone_size, NULL, NULL, NULL, NULL,
1073 		    /*align*/ 0, /*flags*/ 0);
1074 
1075 		cam_periph_lock(periph);
1076 
1077 		if ((softc->pass_zone == NULL)
1078 		 || (softc->pass_io_zone == NULL)) {
1079 			if (softc->pass_zone == NULL)
1080 				xpt_print(periph->path, "unable to allocate "
1081 				    "IO Req UMA zone\n");
1082 			else
1083 				xpt_print(periph->path, "unable to allocate "
1084 				    "IO UMA zone\n");
1085 			softc->flags &= ~PASS_FLAG_ZONE_INPROG;
1086 			goto bailout;
1087 		}
1088 
1089 		/*
1090 		 * Set the flags appropriately and notify any other waiters.
1091 		 */
1092 		softc->flags &= ~PASS_FLAG_ZONE_INPROG;
1093 		softc->flags |= PASS_FLAG_ZONE_VALID;
1094 		wakeup(&softc->pass_zone);
1095 	} else {
1096 		/*
1097 		 * In this case, the UMA zone has not yet been created, but
1098 		 * another context is in the process of creating it.  We
1099 		 * need to sleep until the creation is either done or has
1100 		 * failed.
1101 		 */
1102 		while ((softc->flags & PASS_FLAG_ZONE_INPROG)
1103 		    && ((softc->flags & PASS_FLAG_ZONE_VALID) == 0)) {
1104 			error = msleep(&softc->pass_zone,
1105 				       cam_periph_mtx(periph), PRIBIO,
1106 				       "paszon", 0);
1107 			if (error != 0)
1108 				goto bailout;
1109 		}
1110 		/*
1111 		 * If the zone creation failed, no luck for the user.
1112 		 */
1113 		if ((softc->flags & PASS_FLAG_ZONE_VALID) == 0){
1114 			error = ENOMEM;
1115 			goto bailout;
1116 		}
1117 	}
1118 bailout:
1119 	return (error);
1120 }
1121 
1122 static void
passiocleanup(struct pass_softc * softc,struct pass_io_req * io_req)1123 passiocleanup(struct pass_softc *softc, struct pass_io_req *io_req)
1124 {
1125 	union ccb *ccb;
1126 	struct ccb_hdr *hdr;
1127 	uint8_t **data_ptrs[CAM_PERIPH_MAXMAPS];
1128 	int i, numbufs;
1129 
1130 	ccb = &io_req->ccb;
1131 	hdr = &ccb->ccb_h;
1132 
1133 	switch (hdr->func_code) {
1134 	case XPT_DEV_MATCH:
1135 		numbufs = min(io_req->num_bufs, 2);
1136 
1137 		if (numbufs == 1) {
1138 			data_ptrs[0] = (uint8_t **)&ccb->cdm.matches;
1139 		} else {
1140 			data_ptrs[0] = (uint8_t **)&ccb->cdm.patterns;
1141 			data_ptrs[1] = (uint8_t **)&ccb->cdm.matches;
1142 		}
1143 		break;
1144 	case XPT_SCSI_IO:
1145 	case XPT_CONT_TARGET_IO:
1146 		data_ptrs[0] = &ccb->csio.data_ptr;
1147 		numbufs = min(io_req->num_bufs, 1);
1148 		break;
1149 	case XPT_ATA_IO:
1150 		data_ptrs[0] = &ccb->ataio.data_ptr;
1151 		numbufs = min(io_req->num_bufs, 1);
1152 		break;
1153 	case XPT_SMP_IO:
1154 		numbufs = min(io_req->num_bufs, 2);
1155 		data_ptrs[0] = &ccb->smpio.smp_request;
1156 		data_ptrs[1] = &ccb->smpio.smp_response;
1157 		break;
1158 	case XPT_DEV_ADVINFO:
1159 		numbufs = min(io_req->num_bufs, 1);
1160 		data_ptrs[0] = (uint8_t **)&ccb->cdai.buf;
1161 		break;
1162 	case XPT_NVME_IO:
1163 	case XPT_NVME_ADMIN:
1164 		data_ptrs[0] = &ccb->nvmeio.data_ptr;
1165 		numbufs = min(io_req->num_bufs, 1);
1166 		break;
1167 	default:
1168 		/* allow ourselves to be swapped once again */
1169 		return;
1170 		break; /* NOTREACHED */
1171 	}
1172 
1173 	if (io_req->flags & PASS_IO_USER_SEG_MALLOC) {
1174 		free(io_req->user_segptr, M_SCSIPASS);
1175 		io_req->user_segptr = NULL;
1176 	}
1177 
1178 	/*
1179 	 * We only want to free memory we malloced.
1180 	 */
1181 	if (io_req->data_flags == CAM_DATA_VADDR) {
1182 		for (i = 0; i < io_req->num_bufs; i++) {
1183 			if (io_req->kern_bufs[i] == NULL)
1184 				continue;
1185 
1186 			free(io_req->kern_bufs[i], M_SCSIPASS);
1187 			io_req->kern_bufs[i] = NULL;
1188 		}
1189 	} else if (io_req->data_flags == CAM_DATA_SG) {
1190 		for (i = 0; i < io_req->num_kern_segs; i++) {
1191 			if ((uint8_t *)(uintptr_t)
1192 			    io_req->kern_segptr[i].ds_addr == NULL)
1193 				continue;
1194 
1195 			uma_zfree(softc->pass_io_zone, (uint8_t *)(uintptr_t)
1196 			    io_req->kern_segptr[i].ds_addr);
1197 			io_req->kern_segptr[i].ds_addr = 0;
1198 		}
1199 	}
1200 
1201 	if (io_req->flags & PASS_IO_KERN_SEG_MALLOC) {
1202 		free(io_req->kern_segptr, M_SCSIPASS);
1203 		io_req->kern_segptr = NULL;
1204 	}
1205 
1206 	if (io_req->data_flags != CAM_DATA_PADDR) {
1207 		for (i = 0; i < numbufs; i++) {
1208 			/*
1209 			 * Restore the user's buffer pointers to their
1210 			 * previous values.
1211 			 */
1212 			if (io_req->user_bufs[i] != NULL)
1213 				*data_ptrs[i] = io_req->user_bufs[i];
1214 		}
1215 	}
1216 
1217 }
1218 
1219 static int
passcopysglist(struct cam_periph * periph,struct pass_io_req * io_req,ccb_flags direction)1220 passcopysglist(struct cam_periph *periph, struct pass_io_req *io_req,
1221 	       ccb_flags direction)
1222 {
1223 	bus_size_t kern_watermark, user_watermark, len_to_copy;
1224 	bus_dma_segment_t *user_sglist, *kern_sglist;
1225 	int i, j, error;
1226 
1227 	error = 0;
1228 	kern_watermark = 0;
1229 	user_watermark = 0;
1230 	len_to_copy = 0;
1231 	user_sglist = io_req->user_segptr;
1232 	kern_sglist = io_req->kern_segptr;
1233 
1234 	for (i = 0, j = 0; i < io_req->num_user_segs &&
1235 	     j < io_req->num_kern_segs;) {
1236 		uint8_t *user_ptr, *kern_ptr;
1237 
1238 		len_to_copy = min(user_sglist[i].ds_len -user_watermark,
1239 		    kern_sglist[j].ds_len - kern_watermark);
1240 
1241 		user_ptr = (uint8_t *)(uintptr_t)user_sglist[i].ds_addr;
1242 		user_ptr = user_ptr + user_watermark;
1243 		kern_ptr = (uint8_t *)(uintptr_t)kern_sglist[j].ds_addr;
1244 		kern_ptr = kern_ptr + kern_watermark;
1245 
1246 		user_watermark += len_to_copy;
1247 		kern_watermark += len_to_copy;
1248 
1249 		if (direction == CAM_DIR_IN) {
1250 			error = copyout(kern_ptr, user_ptr, len_to_copy);
1251 			if (error != 0) {
1252 				xpt_print(periph->path, "%s: copyout of %u "
1253 					  "bytes from %p to %p failed with "
1254 					  "error %d\n", __func__, len_to_copy,
1255 					  kern_ptr, user_ptr, error);
1256 				goto bailout;
1257 			}
1258 		} else {
1259 			error = copyin(user_ptr, kern_ptr, len_to_copy);
1260 			if (error != 0) {
1261 				xpt_print(periph->path, "%s: copyin of %u "
1262 					  "bytes from %p to %p failed with "
1263 					  "error %d\n", __func__, len_to_copy,
1264 					  user_ptr, kern_ptr, error);
1265 				goto bailout;
1266 			}
1267 		}
1268 
1269 		if (user_sglist[i].ds_len == user_watermark) {
1270 			i++;
1271 			user_watermark = 0;
1272 		}
1273 
1274 		if (kern_sglist[j].ds_len == kern_watermark) {
1275 			j++;
1276 			kern_watermark = 0;
1277 		}
1278 	}
1279 
1280 bailout:
1281 
1282 	return (error);
1283 }
1284 
1285 static int
passmemsetup(struct cam_periph * periph,struct pass_io_req * io_req)1286 passmemsetup(struct cam_periph *periph, struct pass_io_req *io_req)
1287 {
1288 	union ccb *ccb;
1289 	struct ccb_hdr *hdr;
1290 	struct pass_softc *softc;
1291 	int numbufs, i;
1292 	uint8_t **data_ptrs[CAM_PERIPH_MAXMAPS];
1293 	uint32_t lengths[CAM_PERIPH_MAXMAPS];
1294 	uint32_t dirs[CAM_PERIPH_MAXMAPS];
1295 	uint32_t num_segs;
1296 	uint16_t *seg_cnt_ptr;
1297 	size_t maxmap;
1298 	int error;
1299 
1300 	cam_periph_assert(periph, MA_NOTOWNED);
1301 
1302 	softc = periph->softc;
1303 
1304 	error = 0;
1305 	ccb = &io_req->ccb;
1306 	hdr = &ccb->ccb_h;
1307 	maxmap = 0;
1308 	num_segs = 0;
1309 	seg_cnt_ptr = NULL;
1310 
1311 	switch(hdr->func_code) {
1312 	case XPT_DEV_MATCH:
1313 		if (ccb->cdm.match_buf_len == 0) {
1314 			printf("%s: invalid match buffer length 0\n", __func__);
1315 			return(EINVAL);
1316 		}
1317 		if (ccb->cdm.pattern_buf_len > 0) {
1318 			data_ptrs[0] = (uint8_t **)&ccb->cdm.patterns;
1319 			lengths[0] = ccb->cdm.pattern_buf_len;
1320 			dirs[0] = CAM_DIR_OUT;
1321 			data_ptrs[1] = (uint8_t **)&ccb->cdm.matches;
1322 			lengths[1] = ccb->cdm.match_buf_len;
1323 			dirs[1] = CAM_DIR_IN;
1324 			numbufs = 2;
1325 		} else {
1326 			data_ptrs[0] = (uint8_t **)&ccb->cdm.matches;
1327 			lengths[0] = ccb->cdm.match_buf_len;
1328 			dirs[0] = CAM_DIR_IN;
1329 			numbufs = 1;
1330 		}
1331 		io_req->data_flags = CAM_DATA_VADDR;
1332 		break;
1333 	case XPT_SCSI_IO:
1334 	case XPT_CONT_TARGET_IO:
1335 		if ((hdr->flags & CAM_DIR_MASK) == CAM_DIR_NONE)
1336 			return(0);
1337 
1338 		/*
1339 		 * The user shouldn't be able to supply a bio.
1340 		 */
1341 		if ((hdr->flags & CAM_DATA_MASK) == CAM_DATA_BIO)
1342 			return (EINVAL);
1343 
1344 		io_req->data_flags = hdr->flags & CAM_DATA_MASK;
1345 
1346 		data_ptrs[0] = &ccb->csio.data_ptr;
1347 		lengths[0] = ccb->csio.dxfer_len;
1348 		dirs[0] = hdr->flags & CAM_DIR_MASK;
1349 		num_segs = ccb->csio.sglist_cnt;
1350 		seg_cnt_ptr = &ccb->csio.sglist_cnt;
1351 		numbufs = 1;
1352 		maxmap = softc->maxio;
1353 		break;
1354 	case XPT_ATA_IO:
1355 		if ((hdr->flags & CAM_DIR_MASK) == CAM_DIR_NONE)
1356 			return(0);
1357 
1358 		/*
1359 		 * We only support a single virtual address for ATA I/O.
1360 		 */
1361 		if ((hdr->flags & CAM_DATA_MASK) != CAM_DATA_VADDR)
1362 			return (EINVAL);
1363 
1364 		io_req->data_flags = CAM_DATA_VADDR;
1365 
1366 		data_ptrs[0] = &ccb->ataio.data_ptr;
1367 		lengths[0] = ccb->ataio.dxfer_len;
1368 		dirs[0] = hdr->flags & CAM_DIR_MASK;
1369 		numbufs = 1;
1370 		maxmap = softc->maxio;
1371 		break;
1372 	case XPT_SMP_IO:
1373 		io_req->data_flags = CAM_DATA_VADDR;
1374 
1375 		data_ptrs[0] = &ccb->smpio.smp_request;
1376 		lengths[0] = ccb->smpio.smp_request_len;
1377 		dirs[0] = CAM_DIR_OUT;
1378 		data_ptrs[1] = &ccb->smpio.smp_response;
1379 		lengths[1] = ccb->smpio.smp_response_len;
1380 		dirs[1] = CAM_DIR_IN;
1381 		numbufs = 2;
1382 		maxmap = softc->maxio;
1383 		break;
1384 	case XPT_DEV_ADVINFO:
1385 		if (ccb->cdai.bufsiz == 0)
1386 			return (0);
1387 
1388 		io_req->data_flags = CAM_DATA_VADDR;
1389 
1390 		data_ptrs[0] = (uint8_t **)&ccb->cdai.buf;
1391 		lengths[0] = ccb->cdai.bufsiz;
1392 		dirs[0] = CAM_DIR_IN;
1393 		numbufs = 1;
1394 		break;
1395 	case XPT_NVME_ADMIN:
1396 	case XPT_NVME_IO:
1397 		if ((hdr->flags & CAM_DIR_MASK) == CAM_DIR_NONE)
1398 			return (0);
1399 
1400 		io_req->data_flags = hdr->flags & CAM_DATA_MASK;
1401 
1402 		data_ptrs[0] = &ccb->nvmeio.data_ptr;
1403 		lengths[0] = ccb->nvmeio.dxfer_len;
1404 		dirs[0] = hdr->flags & CAM_DIR_MASK;
1405 		num_segs = ccb->nvmeio.sglist_cnt;
1406 		seg_cnt_ptr = &ccb->nvmeio.sglist_cnt;
1407 		numbufs = 1;
1408 		maxmap = softc->maxio;
1409 		break;
1410 	default:
1411 		return(EINVAL);
1412 		break; /* NOTREACHED */
1413 	}
1414 
1415 	io_req->num_bufs = numbufs;
1416 
1417 	/*
1418 	 * If there is a maximum, check to make sure that the user's
1419 	 * request fits within the limit.  In general, we should only have
1420 	 * a maximum length for requests that go to hardware.  Otherwise it
1421 	 * is whatever we're able to malloc.
1422 	 */
1423 	for (i = 0; i < numbufs; i++) {
1424 		io_req->user_bufs[i] = *data_ptrs[i];
1425 		io_req->dirs[i] = dirs[i];
1426 		io_req->lengths[i] = lengths[i];
1427 
1428 		if (maxmap == 0)
1429 			continue;
1430 
1431 		if (lengths[i] <= maxmap)
1432 			continue;
1433 
1434 		xpt_print(periph->path, "%s: data length %u > max allowed %u "
1435 			  "bytes\n", __func__, lengths[i], maxmap);
1436 		error = EINVAL;
1437 		goto bailout;
1438 	}
1439 
1440 	switch (io_req->data_flags) {
1441 	case CAM_DATA_VADDR:
1442 		/* Map or copy the buffer into kernel address space */
1443 		for (i = 0; i < numbufs; i++) {
1444 			uint8_t *tmp_buf;
1445 
1446 			/*
1447 			 * If for some reason no length is specified, we
1448 			 * don't need to allocate anything.
1449 			 */
1450 			if (io_req->lengths[i] == 0)
1451 				continue;
1452 
1453 			tmp_buf = malloc(lengths[i], M_SCSIPASS,
1454 					 M_WAITOK | M_ZERO);
1455 			io_req->kern_bufs[i] = tmp_buf;
1456 			*data_ptrs[i] = tmp_buf;
1457 
1458 #if 0
1459 			xpt_print(periph->path, "%s: malloced %p len %u, user "
1460 				  "buffer %p, operation: %s\n", __func__,
1461 				  tmp_buf, lengths[i], io_req->user_bufs[i],
1462 				  (dirs[i] == CAM_DIR_IN) ? "read" : "write");
1463 #endif
1464 			/*
1465 			 * We only need to copy in if the user is writing.
1466 			 */
1467 			if (dirs[i] != CAM_DIR_OUT)
1468 				continue;
1469 
1470 			error = copyin(io_req->user_bufs[i],
1471 				       io_req->kern_bufs[i], lengths[i]);
1472 			if (error != 0) {
1473 				xpt_print(periph->path, "%s: copy of user "
1474 					  "buffer from %p to %p failed with "
1475 					  "error %d\n", __func__,
1476 					  io_req->user_bufs[i],
1477 					  io_req->kern_bufs[i], error);
1478 				goto bailout;
1479 			}
1480 		}
1481 		break;
1482 	case CAM_DATA_PADDR:
1483 		/* Pass down the pointer as-is */
1484 		break;
1485 	case CAM_DATA_SG: {
1486 		size_t sg_length, size_to_go, alloc_size;
1487 		uint32_t num_segs_needed;
1488 
1489 		/*
1490 		 * Copy the user S/G list in, and then copy in the
1491 		 * individual segments.
1492 		 */
1493 		/*
1494 		 * We shouldn't see this, but check just in case.
1495 		 */
1496 		if (numbufs != 1) {
1497 			xpt_print(periph->path, "%s: cannot currently handle "
1498 				  "more than one S/G list per CCB\n", __func__);
1499 			error = EINVAL;
1500 			goto bailout;
1501 		}
1502 
1503 		/*
1504 		 * We have to have at least one segment.
1505 		 */
1506 		if (num_segs == 0) {
1507 			xpt_print(periph->path, "%s: CAM_DATA_SG flag set, "
1508 				  "but sglist_cnt=0!\n", __func__);
1509 			error = EINVAL;
1510 			goto bailout;
1511 		}
1512 
1513 		/*
1514 		 * Make sure the user specified the total length and didn't
1515 		 * just leave it to us to decode the S/G list.
1516 		 */
1517 		if (lengths[0] == 0) {
1518 			xpt_print(periph->path, "%s: no dxfer_len specified, "
1519 				  "but CAM_DATA_SG flag is set!\n", __func__);
1520 			error = EINVAL;
1521 			goto bailout;
1522 		}
1523 
1524 		/*
1525 		 * We allocate buffers in io_zone_size increments for an
1526 		 * S/G list.  This will generally be maxphys.
1527 		 */
1528 		if (lengths[0] <= softc->io_zone_size)
1529 			num_segs_needed = 1;
1530 		else {
1531 			num_segs_needed = lengths[0] / softc->io_zone_size;
1532 			if ((lengths[0] % softc->io_zone_size) != 0)
1533 				num_segs_needed++;
1534 		}
1535 
1536 		/* Figure out the size of the S/G list */
1537 		sg_length = num_segs * sizeof(bus_dma_segment_t);
1538 		io_req->num_user_segs = num_segs;
1539 		io_req->num_kern_segs = num_segs_needed;
1540 
1541 		/* Save the user's S/G list pointer for later restoration */
1542 		io_req->user_bufs[0] = *data_ptrs[0];
1543 
1544 		/*
1545 		 * If we have enough segments allocated by default to handle
1546 		 * the length of the user's S/G list,
1547 		 */
1548 		if (num_segs > PASS_MAX_SEGS) {
1549 			io_req->user_segptr = malloc(sizeof(bus_dma_segment_t) *
1550 			    num_segs, M_SCSIPASS, M_WAITOK | M_ZERO);
1551 			io_req->flags |= PASS_IO_USER_SEG_MALLOC;
1552 		} else
1553 			io_req->user_segptr = io_req->user_segs;
1554 
1555 		error = copyin(*data_ptrs[0], io_req->user_segptr, sg_length);
1556 		if (error != 0) {
1557 			xpt_print(periph->path, "%s: copy of user S/G list "
1558 				  "from %p to %p failed with error %d\n",
1559 				  __func__, *data_ptrs[0], io_req->user_segptr,
1560 				  error);
1561 			goto bailout;
1562 		}
1563 
1564 		if (num_segs_needed > PASS_MAX_SEGS) {
1565 			io_req->kern_segptr = malloc(sizeof(bus_dma_segment_t) *
1566 			    num_segs_needed, M_SCSIPASS, M_WAITOK | M_ZERO);
1567 			io_req->flags |= PASS_IO_KERN_SEG_MALLOC;
1568 		} else {
1569 			io_req->kern_segptr = io_req->kern_segs;
1570 		}
1571 
1572 		/*
1573 		 * Allocate the kernel S/G list.
1574 		 */
1575 		for (size_to_go = lengths[0], i = 0;
1576 		     size_to_go > 0 && i < num_segs_needed;
1577 		     i++, size_to_go -= alloc_size) {
1578 			uint8_t *kern_ptr;
1579 
1580 			alloc_size = min(size_to_go, softc->io_zone_size);
1581 			kern_ptr = uma_zalloc(softc->pass_io_zone, M_WAITOK);
1582 			io_req->kern_segptr[i].ds_addr =
1583 			    (bus_addr_t)(uintptr_t)kern_ptr;
1584 			io_req->kern_segptr[i].ds_len = alloc_size;
1585 		}
1586 		if (size_to_go > 0) {
1587 			printf("%s: size_to_go = %zu, software error!\n",
1588 			       __func__, size_to_go);
1589 			error = EINVAL;
1590 			goto bailout;
1591 		}
1592 
1593 		*data_ptrs[0] = (uint8_t *)io_req->kern_segptr;
1594 		*seg_cnt_ptr = io_req->num_kern_segs;
1595 
1596 		/*
1597 		 * We only need to copy data here if the user is writing.
1598 		 */
1599 		if (dirs[0] == CAM_DIR_OUT)
1600 			error = passcopysglist(periph, io_req, dirs[0]);
1601 		break;
1602 	}
1603 	case CAM_DATA_SG_PADDR: {
1604 		size_t sg_length;
1605 
1606 		/*
1607 		 * We shouldn't see this, but check just in case.
1608 		 */
1609 		if (numbufs != 1) {
1610 			printf("%s: cannot currently handle more than one "
1611 			       "S/G list per CCB\n", __func__);
1612 			error = EINVAL;
1613 			goto bailout;
1614 		}
1615 
1616 		/*
1617 		 * We have to have at least one segment.
1618 		 */
1619 		if (num_segs == 0) {
1620 			xpt_print(periph->path, "%s: CAM_DATA_SG_PADDR flag "
1621 				  "set, but sglist_cnt=0!\n", __func__);
1622 			error = EINVAL;
1623 			goto bailout;
1624 		}
1625 
1626 		/*
1627 		 * Make sure the user specified the total length and didn't
1628 		 * just leave it to us to decode the S/G list.
1629 		 */
1630 		if (lengths[0] == 0) {
1631 			xpt_print(periph->path, "%s: no dxfer_len specified, "
1632 				  "but CAM_DATA_SG flag is set!\n", __func__);
1633 			error = EINVAL;
1634 			goto bailout;
1635 		}
1636 
1637 		/* Figure out the size of the S/G list */
1638 		sg_length = num_segs * sizeof(bus_dma_segment_t);
1639 		io_req->num_user_segs = num_segs;
1640 		io_req->num_kern_segs = io_req->num_user_segs;
1641 
1642 		/* Save the user's S/G list pointer for later restoration */
1643 		io_req->user_bufs[0] = *data_ptrs[0];
1644 
1645 		if (num_segs > PASS_MAX_SEGS) {
1646 			io_req->user_segptr = malloc(sizeof(bus_dma_segment_t) *
1647 			    num_segs, M_SCSIPASS, M_WAITOK | M_ZERO);
1648 			io_req->flags |= PASS_IO_USER_SEG_MALLOC;
1649 		} else
1650 			io_req->user_segptr = io_req->user_segs;
1651 
1652 		io_req->kern_segptr = io_req->user_segptr;
1653 
1654 		error = copyin(*data_ptrs[0], io_req->user_segptr, sg_length);
1655 		if (error != 0) {
1656 			xpt_print(periph->path, "%s: copy of user S/G list "
1657 				  "from %p to %p failed with error %d\n",
1658 				  __func__, *data_ptrs[0], io_req->user_segptr,
1659 				  error);
1660 			goto bailout;
1661 		}
1662 		break;
1663 	}
1664 	default:
1665 	case CAM_DATA_BIO:
1666 		/*
1667 		 * A user shouldn't be attaching a bio to the CCB.  It
1668 		 * isn't a user-accessible structure.
1669 		 */
1670 		error = EINVAL;
1671 		break;
1672 	}
1673 
1674 bailout:
1675 	if (error != 0)
1676 		passiocleanup(softc, io_req);
1677 
1678 	return (error);
1679 }
1680 
1681 static int
passmemdone(struct cam_periph * periph,struct pass_io_req * io_req)1682 passmemdone(struct cam_periph *periph, struct pass_io_req *io_req)
1683 {
1684 	struct pass_softc *softc;
1685 	int error;
1686 	int i;
1687 
1688 	error = 0;
1689 	softc = (struct pass_softc *)periph->softc;
1690 
1691 	switch (io_req->data_flags) {
1692 	case CAM_DATA_VADDR:
1693 		/*
1694 		 * Copy back to the user buffer if this was a read.
1695 		 */
1696 		for (i = 0; i < io_req->num_bufs; i++) {
1697 			if (io_req->dirs[i] != CAM_DIR_IN)
1698 				continue;
1699 
1700 			error = copyout(io_req->kern_bufs[i],
1701 			    io_req->user_bufs[i], io_req->lengths[i]);
1702 			if (error != 0) {
1703 				xpt_print(periph->path, "Unable to copy %u "
1704 					  "bytes from %p to user address %p\n",
1705 					  io_req->lengths[i],
1706 					  io_req->kern_bufs[i],
1707 					  io_req->user_bufs[i]);
1708 				goto bailout;
1709 			}
1710 		}
1711 		break;
1712 	case CAM_DATA_PADDR:
1713 		/* Do nothing.  The pointer is a physical address already */
1714 		break;
1715 	case CAM_DATA_SG:
1716 		/*
1717 		 * Copy back to the user buffer if this was a read.
1718 		 * Restore the user's S/G list buffer pointer.
1719 		 */
1720 		if (io_req->dirs[0] == CAM_DIR_IN)
1721 			error = passcopysglist(periph, io_req, io_req->dirs[0]);
1722 		break;
1723 	case CAM_DATA_SG_PADDR:
1724 		/*
1725 		 * Restore the user's S/G list buffer pointer.  No need to
1726 		 * copy.
1727 		 */
1728 		break;
1729 	default:
1730 	case CAM_DATA_BIO:
1731 		error = EINVAL;
1732 		break;
1733 	}
1734 
1735 bailout:
1736 	/*
1737 	 * Reset the user's pointers to their original values and free
1738 	 * allocated memory.
1739 	 */
1740 	passiocleanup(softc, io_req);
1741 
1742 	return (error);
1743 }
1744 
1745 static int
passioctl(struct cdev * dev,u_long cmd,caddr_t addr,int flag,struct thread * td)1746 passioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flag, struct thread *td)
1747 {
1748 	int error;
1749 
1750 	if ((error = passdoioctl(dev, cmd, addr, flag, td)) == ENOTTY) {
1751 		error = cam_compat_ioctl(dev, cmd, addr, flag, td, passdoioctl);
1752 	}
1753 	return (error);
1754 }
1755 
1756 static int
passdoioctl(struct cdev * dev,u_long cmd,caddr_t addr,int flag,struct thread * td)1757 passdoioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flag, struct thread *td)
1758 {
1759 	struct	cam_periph *periph;
1760 	struct	pass_softc *softc;
1761 	int	error;
1762 	uint32_t priority;
1763 
1764 	periph = (struct cam_periph *)dev->si_drv1;
1765 	cam_periph_lock(periph);
1766 	softc = (struct pass_softc *)periph->softc;
1767 
1768 	error = 0;
1769 
1770 	switch (cmd) {
1771 	case CAMIOCOMMAND:
1772 	{
1773 		union ccb *inccb;
1774 		union ccb *ccb;
1775 		int ccb_malloced;
1776 
1777 		inccb = (union ccb *)addr;
1778 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
1779 		if (inccb->ccb_h.func_code == XPT_SCSI_IO)
1780 			inccb->csio.bio = NULL;
1781 #endif
1782 
1783 		if (inccb->ccb_h.flags & CAM_UNLOCKED) {
1784 			error = EINVAL;
1785 			break;
1786 		}
1787 
1788 		/*
1789 		 * Some CCB types, like scan bus and scan lun can only go
1790 		 * through the transport layer device.
1791 		 */
1792 		if (inccb->ccb_h.func_code & XPT_FC_XPT_ONLY) {
1793 			xpt_print(periph->path, "CCB function code %#x is "
1794 			    "restricted to the XPT device\n",
1795 			    inccb->ccb_h.func_code);
1796 			error = ENODEV;
1797 			break;
1798 		}
1799 
1800 		/* Compatibility for RL/priority-unaware code. */
1801 		priority = inccb->ccb_h.pinfo.priority;
1802 		if (priority <= CAM_PRIORITY_OOB)
1803 		    priority += CAM_PRIORITY_OOB + 1;
1804 
1805 		/*
1806 		 * Non-immediate CCBs need a CCB from the per-device pool
1807 		 * of CCBs, which is scheduled by the transport layer.
1808 		 * Immediate CCBs and user-supplied CCBs should just be
1809 		 * malloced.
1810 		 */
1811 		if ((inccb->ccb_h.func_code & XPT_FC_QUEUED)
1812 		 && ((inccb->ccb_h.func_code & XPT_FC_USER_CCB) == 0)) {
1813 			ccb = cam_periph_getccb(periph, priority);
1814 			ccb_malloced = 0;
1815 		} else {
1816 			ccb = xpt_alloc_ccb_nowait();
1817 
1818 			if (ccb != NULL)
1819 				xpt_setup_ccb(&ccb->ccb_h, periph->path,
1820 					      priority);
1821 			ccb_malloced = 1;
1822 		}
1823 
1824 		if (ccb == NULL) {
1825 			xpt_print(periph->path, "unable to allocate CCB\n");
1826 			error = ENOMEM;
1827 			break;
1828 		}
1829 
1830 		error = passsendccb(periph, ccb, inccb);
1831 
1832 		if (ccb_malloced)
1833 			xpt_free_ccb(ccb);
1834 		else
1835 			xpt_release_ccb(ccb);
1836 
1837 		break;
1838 	}
1839 	case CAMIOQUEUE:
1840 	{
1841 		struct pass_io_req *io_req;
1842 		union ccb **user_ccb, *ccb;
1843 		xpt_opcode fc;
1844 
1845 #ifdef COMPAT_FREEBSD32
1846 		if (SV_PROC_FLAG(td->td_proc, SV_ILP32)) {
1847 			error = ENOTTY;
1848 			goto bailout;
1849 		}
1850 #endif
1851 		if ((softc->flags & PASS_FLAG_ZONE_VALID) == 0) {
1852 			error = passcreatezone(periph);
1853 			if (error != 0)
1854 				goto bailout;
1855 		}
1856 
1857 		/*
1858 		 * We're going to do a blocking allocation for this I/O
1859 		 * request, so we have to drop the lock.
1860 		 */
1861 		cam_periph_unlock(periph);
1862 
1863 		io_req = uma_zalloc(softc->pass_zone, M_WAITOK | M_ZERO);
1864 		ccb = &io_req->ccb;
1865 		user_ccb = (union ccb **)addr;
1866 
1867 		/*
1868 		 * Unlike the CAMIOCOMMAND ioctl above, we only have a
1869 		 * pointer to the user's CCB, so we have to copy the whole
1870 		 * thing in to a buffer we have allocated (above) instead
1871 		 * of allowing the ioctl code to malloc a buffer and copy
1872 		 * it in.
1873 		 *
1874 		 * This is an advantage for this asynchronous interface,
1875 		 * since we don't want the memory to get freed while the
1876 		 * CCB is outstanding.
1877 		 */
1878 #if 0
1879 		xpt_print(periph->path, "Copying user CCB %p to "
1880 			  "kernel address %p\n", *user_ccb, ccb);
1881 #endif
1882 		error = copyin(*user_ccb, ccb, sizeof(*ccb));
1883 		if (error != 0) {
1884 			xpt_print(periph->path, "Copy of user CCB %p to "
1885 				  "kernel address %p failed with error %d\n",
1886 				  *user_ccb, ccb, error);
1887 			goto camioqueue_error;
1888 		}
1889 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
1890 		if (ccb->ccb_h.func_code == XPT_SCSI_IO)
1891 			ccb->csio.bio = NULL;
1892 #endif
1893 
1894 		if (ccb->ccb_h.flags & CAM_UNLOCKED) {
1895 			error = EINVAL;
1896 			goto camioqueue_error;
1897 		}
1898 
1899 		if (ccb->ccb_h.flags & CAM_CDB_POINTER) {
1900 			if (ccb->csio.cdb_len > IOCDBLEN) {
1901 				error = EINVAL;
1902 				goto camioqueue_error;
1903 			}
1904 			error = copyin(ccb->csio.cdb_io.cdb_ptr,
1905 			    ccb->csio.cdb_io.cdb_bytes, ccb->csio.cdb_len);
1906 			if (error != 0)
1907 				goto camioqueue_error;
1908 			ccb->ccb_h.flags &= ~CAM_CDB_POINTER;
1909 		}
1910 
1911 		/*
1912 		 * Some CCB types, like scan bus and scan lun can only go
1913 		 * through the transport layer device.
1914 		 */
1915 		if (ccb->ccb_h.func_code & XPT_FC_XPT_ONLY) {
1916 			xpt_print(periph->path, "CCB function code %#x is "
1917 			    "restricted to the XPT device\n",
1918 			    ccb->ccb_h.func_code);
1919 			error = ENODEV;
1920 			goto camioqueue_error;
1921 		}
1922 
1923 		/*
1924 		 * Save the user's CCB pointer as well as his linked list
1925 		 * pointers and peripheral private area so that we can
1926 		 * restore these later.
1927 		 */
1928 		io_req->user_ccb_ptr = *user_ccb;
1929 		io_req->user_periph_links = ccb->ccb_h.periph_links;
1930 		io_req->user_periph_priv = ccb->ccb_h.periph_priv;
1931 
1932 		/*
1933 		 * Now that we've saved the user's values, we can set our
1934 		 * own peripheral private entry.
1935 		 */
1936 		ccb->ccb_h.ccb_ioreq = io_req;
1937 
1938 		/* Compatibility for RL/priority-unaware code. */
1939 		priority = ccb->ccb_h.pinfo.priority;
1940 		if (priority <= CAM_PRIORITY_OOB)
1941 		    priority += CAM_PRIORITY_OOB + 1;
1942 
1943 		/*
1944 		 * Setup fields in the CCB like the path and the priority.
1945 		 * The path in particular cannot be done in userland, since
1946 		 * it is a pointer to a kernel data structure.
1947 		 */
1948 		xpt_setup_ccb_flags(&ccb->ccb_h, periph->path, priority,
1949 				    ccb->ccb_h.flags);
1950 
1951 		/*
1952 		 * Setup our done routine.  There is no way for the user to
1953 		 * have a valid pointer here.
1954 		 */
1955 		ccb->ccb_h.cbfcnp = passdone;
1956 
1957 		fc = ccb->ccb_h.func_code;
1958 		/*
1959 		 * If this function code has memory that can be mapped in
1960 		 * or out, we need to call passmemsetup().
1961 		 */
1962 		if ((fc == XPT_SCSI_IO) || (fc == XPT_ATA_IO)
1963 		 || (fc == XPT_SMP_IO) || (fc == XPT_DEV_MATCH)
1964 		 || (fc == XPT_DEV_ADVINFO)
1965 		 || (fc == XPT_NVME_ADMIN) || (fc == XPT_NVME_IO)) {
1966 			error = passmemsetup(periph, io_req);
1967 			if (error != 0)
1968 				goto camioqueue_error;
1969 		} else
1970 			io_req->mapinfo.num_bufs_used = 0;
1971 
1972 		cam_periph_lock(periph);
1973 
1974 		/*
1975 		 * Everything goes on the incoming queue initially.
1976 		 */
1977 		TAILQ_INSERT_TAIL(&softc->incoming_queue, io_req, links);
1978 
1979 		/*
1980 		 * If the CCB is queued, and is not a user CCB, then
1981 		 * we need to allocate a slot for it.  Call xpt_schedule()
1982 		 * so that our start routine will get called when a CCB is
1983 		 * available.
1984 		 */
1985 		if ((fc & XPT_FC_QUEUED)
1986 		 && ((fc & XPT_FC_USER_CCB) == 0)) {
1987 			xpt_schedule(periph, priority);
1988 			break;
1989 		}
1990 
1991 		/*
1992 		 * At this point, the CCB in question is either an
1993 		 * immediate CCB (like XPT_DEV_ADVINFO) or it is a user CCB
1994 		 * and therefore should be malloced, not allocated via a slot.
1995 		 * Remove the CCB from the incoming queue and add it to the
1996 		 * active queue.
1997 		 */
1998 		TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
1999 		TAILQ_INSERT_TAIL(&softc->active_queue, io_req, links);
2000 
2001 		xpt_action(ccb);
2002 
2003 		/*
2004 		 * If this is not a queued CCB (i.e. it is an immediate CCB),
2005 		 * then it is already done.  We need to put it on the done
2006 		 * queue for the user to fetch.
2007 		 */
2008 		if ((fc & XPT_FC_QUEUED) == 0) {
2009 			TAILQ_REMOVE(&softc->active_queue, io_req, links);
2010 			TAILQ_INSERT_TAIL(&softc->done_queue, io_req, links);
2011 		}
2012 		break;
2013 
2014 camioqueue_error:
2015 		uma_zfree(softc->pass_zone, io_req);
2016 		cam_periph_lock(periph);
2017 		break;
2018 	}
2019 	case CAMIOGET:
2020 	{
2021 		union ccb **user_ccb;
2022 		struct pass_io_req *io_req;
2023 		int old_error;
2024 
2025 #ifdef COMPAT_FREEBSD32
2026 		if (SV_PROC_FLAG(td->td_proc, SV_ILP32)) {
2027 			error = ENOTTY;
2028 			goto bailout;
2029 		}
2030 #endif
2031 		user_ccb = (union ccb **)addr;
2032 		old_error = 0;
2033 
2034 		io_req = TAILQ_FIRST(&softc->done_queue);
2035 		if (io_req == NULL) {
2036 			error = ENOENT;
2037 			break;
2038 		}
2039 
2040 		/*
2041 		 * Remove the I/O from the done queue.
2042 		 */
2043 		TAILQ_REMOVE(&softc->done_queue, io_req, links);
2044 
2045 		/*
2046 		 * We have to drop the lock during the copyout because the
2047 		 * copyout can result in VM faults that require sleeping.
2048 		 */
2049 		cam_periph_unlock(periph);
2050 
2051 		/*
2052 		 * Do any needed copies (e.g. for reads) and revert the
2053 		 * pointers in the CCB back to the user's pointers.
2054 		 */
2055 		error = passmemdone(periph, io_req);
2056 
2057 		old_error = error;
2058 
2059 		io_req->ccb.ccb_h.periph_links = io_req->user_periph_links;
2060 		io_req->ccb.ccb_h.periph_priv = io_req->user_periph_priv;
2061 
2062 #if 0
2063 		xpt_print(periph->path, "Copying to user CCB %p from "
2064 			  "kernel address %p\n", *user_ccb, &io_req->ccb);
2065 #endif
2066 
2067 		error = copyout(&io_req->ccb, *user_ccb, sizeof(union ccb));
2068 		if (error != 0) {
2069 			xpt_print(periph->path, "Copy to user CCB %p from "
2070 				  "kernel address %p failed with error %d\n",
2071 				  *user_ccb, &io_req->ccb, error);
2072 		}
2073 
2074 		/*
2075 		 * Prefer the first error we got back, and make sure we
2076 		 * don't overwrite bad status with good.
2077 		 */
2078 		if (old_error != 0)
2079 			error = old_error;
2080 
2081 		cam_periph_lock(periph);
2082 
2083 		/*
2084 		 * At this point, if there was an error, we could potentially
2085 		 * re-queue the I/O and try again.  But why?  The error
2086 		 * would almost certainly happen again.  We might as well
2087 		 * not leak memory.
2088 		 */
2089 		uma_zfree(softc->pass_zone, io_req);
2090 		break;
2091 	}
2092 	default:
2093 		error = cam_periph_ioctl(periph, cmd, addr, passerror);
2094 		break;
2095 	}
2096 
2097 bailout:
2098 	cam_periph_unlock(periph);
2099 
2100 	return(error);
2101 }
2102 
2103 static int
passpoll(struct cdev * dev,int poll_events,struct thread * td)2104 passpoll(struct cdev *dev, int poll_events, struct thread *td)
2105 {
2106 	struct cam_periph *periph;
2107 	struct pass_softc *softc;
2108 	int revents;
2109 
2110 	periph = (struct cam_periph *)dev->si_drv1;
2111 	softc = (struct pass_softc *)periph->softc;
2112 
2113 	revents = poll_events & (POLLOUT | POLLWRNORM);
2114 	if ((poll_events & (POLLIN | POLLRDNORM)) != 0) {
2115 		cam_periph_lock(periph);
2116 
2117 		if (!TAILQ_EMPTY(&softc->done_queue)) {
2118 			revents |= poll_events & (POLLIN | POLLRDNORM);
2119 		}
2120 		cam_periph_unlock(periph);
2121 		if (revents == 0)
2122 			selrecord(td, &softc->read_select);
2123 	}
2124 
2125 	return (revents);
2126 }
2127 
2128 static int
passkqfilter(struct cdev * dev,struct knote * kn)2129 passkqfilter(struct cdev *dev, struct knote *kn)
2130 {
2131 	struct cam_periph *periph;
2132 	struct pass_softc *softc;
2133 
2134 	periph = (struct cam_periph *)dev->si_drv1;
2135 	softc = (struct pass_softc *)periph->softc;
2136 
2137 	kn->kn_hook = (caddr_t)periph;
2138 	kn->kn_fop = &passread_filtops;
2139 	knlist_add(&softc->read_select.si_note, kn, 0);
2140 
2141 	return (0);
2142 }
2143 
2144 static void
passreadfiltdetach(struct knote * kn)2145 passreadfiltdetach(struct knote *kn)
2146 {
2147 	struct cam_periph *periph;
2148 	struct pass_softc *softc;
2149 
2150 	periph = (struct cam_periph *)kn->kn_hook;
2151 	softc = (struct pass_softc *)periph->softc;
2152 
2153 	knlist_remove(&softc->read_select.si_note, kn, 0);
2154 }
2155 
2156 static int
passreadfilt(struct knote * kn,long hint)2157 passreadfilt(struct knote *kn, long hint)
2158 {
2159 	struct cam_periph *periph;
2160 	struct pass_softc *softc;
2161 	int retval;
2162 
2163 	periph = (struct cam_periph *)kn->kn_hook;
2164 	softc = (struct pass_softc *)periph->softc;
2165 
2166 	cam_periph_assert(periph, MA_OWNED);
2167 
2168 	if (TAILQ_EMPTY(&softc->done_queue))
2169 		retval = 0;
2170 	else
2171 		retval = 1;
2172 
2173 	return (retval);
2174 }
2175 
2176 /*
2177  * Generally, "ccb" should be the CCB supplied by the kernel.  "inccb"
2178  * should be the CCB that is copied in from the user.
2179  */
2180 static int
passsendccb(struct cam_periph * periph,union ccb * ccb,union ccb * inccb)2181 passsendccb(struct cam_periph *periph, union ccb *ccb, union ccb *inccb)
2182 {
2183 	struct pass_softc *softc;
2184 	struct cam_periph_map_info mapinfo;
2185 	uint8_t *cmd;
2186 	xpt_opcode fc;
2187 	int error;
2188 
2189 	softc = (struct pass_softc *)periph->softc;
2190 
2191 	/*
2192 	 * There are some fields in the CCB header that need to be
2193 	 * preserved, the rest we get from the user.
2194 	 */
2195 	xpt_merge_ccb(ccb, inccb);
2196 
2197 	if (ccb->ccb_h.flags & CAM_CDB_POINTER) {
2198 		cmd = __builtin_alloca(ccb->csio.cdb_len);
2199 		error = copyin(ccb->csio.cdb_io.cdb_ptr, cmd, ccb->csio.cdb_len);
2200 		if (error)
2201 			return (error);
2202 		ccb->csio.cdb_io.cdb_ptr = cmd;
2203 	}
2204 
2205 	/*
2206 	 * Let cam_periph_mapmem do a sanity check on the data pointer format.
2207 	 * Even if no data transfer is needed, it's a cheap check and it
2208 	 * simplifies the code.
2209 	 */
2210 	fc = ccb->ccb_h.func_code;
2211 	if ((fc == XPT_SCSI_IO) || (fc == XPT_ATA_IO) || (fc == XPT_SMP_IO)
2212             || (fc == XPT_DEV_MATCH) || (fc == XPT_DEV_ADVINFO) || (fc == XPT_MMC_IO)
2213             || (fc == XPT_NVME_ADMIN) || (fc == XPT_NVME_IO)) {
2214 		bzero(&mapinfo, sizeof(mapinfo));
2215 
2216 		/*
2217 		 * cam_periph_mapmem calls into proc and vm functions that can
2218 		 * sleep as well as trigger I/O, so we can't hold the lock.
2219 		 * Dropping it here is reasonably safe.
2220 		 */
2221 		cam_periph_unlock(periph);
2222 		error = cam_periph_mapmem(ccb, &mapinfo, softc->maxio);
2223 		cam_periph_lock(periph);
2224 
2225 		/*
2226 		 * cam_periph_mapmem returned an error, we can't continue.
2227 		 * Return the error to the user.
2228 		 */
2229 		if (error)
2230 			return(error);
2231 	} else
2232 		/* Ensure that the unmap call later on is a no-op. */
2233 		mapinfo.num_bufs_used = 0;
2234 
2235 	/*
2236 	 * If the user wants us to perform any error recovery, then honor
2237 	 * that request.  Otherwise, it's up to the user to perform any
2238 	 * error recovery.
2239 	 */
2240 	{
2241 		uint32_t cam_flags, sense_flags;
2242 
2243 		passflags(ccb, &cam_flags, &sense_flags);
2244 		cam_periph_runccb(ccb,  passerror, cam_flags,
2245 		    sense_flags, softc->device_stats);
2246 	}
2247 
2248 	cam_periph_unlock(periph);
2249 	error = cam_periph_unmapmem(ccb, &mapinfo);
2250 	cam_periph_lock(periph);
2251 
2252 	ccb->ccb_h.cbfcnp = NULL;
2253 	ccb->ccb_h.periph_priv = inccb->ccb_h.periph_priv;
2254 	bcopy(ccb, inccb, sizeof(union ccb));
2255 
2256 	return (error);
2257 }
2258 
2259 /*
2260  * Set the cam_flags and sense_flags based on whether or not the request wants
2261  * error recovery. In order to log errors via devctl, we need to do at least
2262  * minimal recovery. We do this by not retrying unit attention (we let the
2263  * requester do it, or not, if appropriate) and specifically asking for no
2264  * recovery, like we do during device probing.
2265  */
2266 static void
passflags(union ccb * ccb,uint32_t * cam_flags,uint32_t * sense_flags)2267 passflags(union ccb *ccb, uint32_t *cam_flags, uint32_t *sense_flags)
2268 {
2269 	if ((ccb->ccb_h.flags & CAM_PASS_ERR_RECOVER) != 0) {
2270 		*cam_flags = CAM_RETRY_SELTO;
2271 		*sense_flags = SF_RETRY_UA | SF_NO_PRINT;
2272 	} else {
2273 		*cam_flags = 0;
2274 		*sense_flags = SF_NO_RETRY | SF_NO_RECOVERY | SF_NO_PRINT;
2275 	}
2276 }
2277 
2278 static int
passerror(union ccb * ccb,uint32_t cam_flags,uint32_t sense_flags)2279 passerror(union ccb *ccb, uint32_t cam_flags, uint32_t sense_flags)
2280 {
2281 
2282 	return(cam_periph_error(ccb, cam_flags, sense_flags));
2283 }
2284