xref: /freebsd/sys/cam/scsi/scsi_pass.c (revision 36138969847b231cd98f48272e2bdf88a8dc08dd)
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 	.f_copy =	knote_triv_copy,
211 };
212 
213 static MALLOC_DEFINE(M_SCSIPASS, "scsi_pass", "scsi passthrough buffers");
214 
215 static void
passinit(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
passrejectios(struct cam_periph * periph)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
passdevgonecb(void * arg)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
passoninvalidate(struct cam_periph * periph)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
passcleanup(struct cam_periph * periph)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 	/*
402 	 * It should be safe to destroy the zones from here, because all
403 	 * of the references to this peripheral have been freed, and all
404 	 * I/O has been terminated and freed.  We check the zones for NULL
405 	 * because they may not have been allocated yet if the device went
406 	 * away before any asynchronous I/O has been issued.
407 	 */
408 	if (softc->pass_zone != NULL)
409 		uma_zdestroy(softc->pass_zone);
410 	if (softc->pass_io_zone != NULL)
411 		uma_zdestroy(softc->pass_io_zone);
412 
413 	cam_periph_lock(periph);
414 
415 	free(softc, M_DEVBUF);
416 }
417 
418 static void
pass_shutdown_kqueue(void * context,int pending)419 pass_shutdown_kqueue(void *context, int pending)
420 {
421 	struct cam_periph *periph;
422 	struct pass_softc *softc;
423 
424 	periph = context;
425 	softc = periph->softc;
426 
427 	knlist_clear(&softc->read_select.si_note, /*is_locked*/ 0);
428 	knlist_destroy(&softc->read_select.si_note);
429 
430 	/*
431 	 * Release the reference we held for kqueue.
432 	 */
433 	cam_periph_release(periph);
434 }
435 
436 static void
pass_add_physpath(void * context,int pending)437 pass_add_physpath(void *context, int pending)
438 {
439 	struct cam_periph *periph;
440 	struct pass_softc *softc;
441 	struct mtx *mtx;
442 	char *physpath;
443 
444 	/*
445 	 * If we have one, create a devfs alias for our
446 	 * physical path.
447 	 */
448 	periph = context;
449 	softc = periph->softc;
450 	physpath = malloc(MAXPATHLEN, M_DEVBUF, M_WAITOK);
451 	mtx = cam_periph_mtx(periph);
452 	mtx_lock(mtx);
453 
454 	if (periph->flags & CAM_PERIPH_INVALID)
455 		goto out;
456 
457 	if (xpt_getattr(physpath, MAXPATHLEN,
458 			"GEOM::physpath", periph->path) == 0
459 	 && strlen(physpath) != 0) {
460 		mtx_unlock(mtx);
461 		make_dev_physpath_alias(MAKEDEV_WAITOK | MAKEDEV_CHECKNAME,
462 				&softc->alias_dev, softc->dev,
463 				softc->alias_dev, physpath);
464 		mtx_lock(mtx);
465 	}
466 
467 out:
468 	/*
469 	 * Now that we've made our alias, we no longer have to have a
470 	 * reference to the device.
471 	 */
472 	if ((softc->flags & PASS_FLAG_INITIAL_PHYSPATH) == 0)
473 		softc->flags |= PASS_FLAG_INITIAL_PHYSPATH;
474 
475 	/*
476 	 * We always acquire a reference to the periph before queueing this
477 	 * task queue function, so it won't go away before we run.
478 	 */
479 	while (pending-- > 0)
480 		cam_periph_release_locked(periph);
481 	mtx_unlock(mtx);
482 
483 	free(physpath, M_DEVBUF);
484 }
485 
486 static void
passasync(void * callback_arg,uint32_t code,struct cam_path * path,void * arg)487 passasync(void *callback_arg, uint32_t code,
488 	  struct cam_path *path, void *arg)
489 {
490 	struct cam_periph *periph;
491 
492 	periph = (struct cam_periph *)callback_arg;
493 
494 	switch (code) {
495 	case AC_FOUND_DEVICE:
496 	{
497 		struct ccb_getdev *cgd;
498 		cam_status status;
499 
500 		cgd = (struct ccb_getdev *)arg;
501 		if (cgd == NULL)
502 			break;
503 
504 		/*
505 		 * Allocate a peripheral instance for
506 		 * this device and start the probe
507 		 * process.
508 		 */
509 		status = cam_periph_alloc(passregister, passoninvalidate,
510 					  passcleanup, passstart, PERIPH_NAME,
511 					  CAM_PERIPH_BIO, path,
512 					  passasync, AC_FOUND_DEVICE, cgd);
513 
514 		if (status != CAM_REQ_CMP
515 		 && status != CAM_REQ_INPROG) {
516 			const struct cam_status_entry *entry;
517 
518 			entry = cam_fetch_status_entry(status);
519 
520 			printf("passasync: Unable to attach new device "
521 			       "due to status %#x: %s\n", status, entry ?
522 			       entry->status_text : "Unknown");
523 		}
524 
525 		break;
526 	}
527 	case AC_ADVINFO_CHANGED:
528 	{
529 		uintptr_t buftype;
530 
531 		buftype = (uintptr_t)arg;
532 		if (buftype == CDAI_TYPE_PHYS_PATH) {
533 			struct pass_softc *softc;
534 
535 			softc = (struct pass_softc *)periph->softc;
536 			/*
537 			 * Acquire a reference to the periph before we
538 			 * start the taskqueue, so that we don't run into
539 			 * a situation where the periph goes away before
540 			 * the task queue has a chance to run.
541 			 */
542 			if (cam_periph_acquire(periph) != 0)
543 				break;
544 
545 			taskqueue_enqueue(taskqueue_thread,
546 					  &softc->add_physpath_task);
547 		}
548 		break;
549 	}
550 	default:
551 		cam_periph_async(periph, code, path, arg);
552 		break;
553 	}
554 }
555 
556 static cam_status
passregister(struct cam_periph * periph,void * arg)557 passregister(struct cam_periph *periph, void *arg)
558 {
559 	struct pass_softc *softc;
560 	struct ccb_getdev *cgd;
561 	struct ccb_pathinq cpi;
562 	struct make_dev_args args;
563 	int error, no_tags;
564 
565 	cgd = (struct ccb_getdev *)arg;
566 	if (cgd == NULL) {
567 		printf("%s: no getdev CCB, can't register device\n", __func__);
568 		return(CAM_REQ_CMP_ERR);
569 	}
570 
571 	softc = (struct pass_softc *)malloc(sizeof(*softc),
572 					    M_DEVBUF, M_NOWAIT);
573 
574 	if (softc == NULL) {
575 		printf("%s: Unable to probe new device. "
576 		       "Unable to allocate softc\n", __func__);
577 		return(CAM_REQ_CMP_ERR);
578 	}
579 
580 	bzero(softc, sizeof(*softc));
581 	softc->state = PASS_STATE_NORMAL;
582 	if (cgd->protocol == PROTO_SCSI || cgd->protocol == PROTO_ATAPI)
583 		softc->pd_type = SID_TYPE(&cgd->inq_data);
584 	else if (cgd->protocol == PROTO_SATAPM)
585 		softc->pd_type = T_ENCLOSURE;
586 	else
587 		softc->pd_type = T_DIRECT;
588 
589 	periph->softc = softc;
590 	softc->periph = periph;
591 	TAILQ_INIT(&softc->incoming_queue);
592 	TAILQ_INIT(&softc->active_queue);
593 	TAILQ_INIT(&softc->abandoned_queue);
594 	TAILQ_INIT(&softc->done_queue);
595 	snprintf(softc->zone_name, sizeof(softc->zone_name), "%s%d",
596 		 periph->periph_name, periph->unit_number);
597 	snprintf(softc->io_zone_name, sizeof(softc->io_zone_name), "%s%dIO",
598 		 periph->periph_name, periph->unit_number);
599 	softc->io_zone_size = maxphys;
600 	knlist_init_mtx(&softc->read_select.si_note, cam_periph_mtx(periph));
601 
602 	xpt_path_inq(&cpi, periph->path);
603 
604 	if (cpi.maxio == 0)
605 		softc->maxio = DFLTPHYS;	/* traditional default */
606 	else if (cpi.maxio > maxphys)
607 		softc->maxio = maxphys;		/* for safety */
608 	else
609 		softc->maxio = cpi.maxio;	/* real value */
610 
611 	if (cpi.hba_misc & PIM_UNMAPPED)
612 		softc->flags |= PASS_FLAG_UNMAPPED_CAPABLE;
613 
614 	/*
615 	 * We pass in 0 for a blocksize, since we don't know what the blocksize
616 	 * of this device is, if it even has a blocksize.
617 	 *
618 	 * Note: no_tags is valid only for SCSI peripherals, but we don't do any
619 	 * devstat accounting for tags on any other transport. SCSI is the only
620 	 * transport that uses the tag_action (ata has only vestigial references
621 	 * to it, others ignore it entirely).
622 	 */
623 	cam_periph_unlock(periph);
624 	no_tags = (cgd->inq_data.flags & SID_CmdQue) == 0;
625 	softc->device_stats = devstat_new_entry(PERIPH_NAME,
626 			  periph->unit_number, 0,
627 			  DEVSTAT_NO_BLOCKSIZE
628 			  | (no_tags ? DEVSTAT_NO_ORDERED_TAGS : 0),
629 			  softc->pd_type |
630 			  XPORT_DEVSTAT_TYPE(cpi.transport) |
631 			  DEVSTAT_TYPE_PASS,
632 			  DEVSTAT_PRIORITY_PASS);
633 
634 	/*
635 	 * Initialize the taskqueue handler for shutting down kqueue.
636 	 */
637 	TASK_INIT(&softc->shutdown_kqueue_task, /*priority*/ 0,
638 		  pass_shutdown_kqueue, periph);
639 
640 	/*
641 	 * Acquire a reference to the periph that we can release once we've
642 	 * cleaned up the kqueue.
643 	 */
644 	if (cam_periph_acquire(periph) != 0) {
645 		xpt_print(periph->path, "%s: lost periph during "
646 			  "registration!\n", __func__);
647 		cam_periph_lock(periph);
648 		return (CAM_REQ_CMP_ERR);
649 	}
650 
651 	/*
652 	 * Acquire a reference to the periph before we create the devfs
653 	 * instance for it.  We'll release this reference once the devfs
654 	 * instance has been freed.
655 	 */
656 	if (cam_periph_acquire(periph) != 0) {
657 		xpt_print(periph->path, "%s: lost periph during "
658 			  "registration!\n", __func__);
659 		cam_periph_lock(periph);
660 		return (CAM_REQ_CMP_ERR);
661 	}
662 
663 	/* Register the device */
664 	make_dev_args_init(&args);
665 	args.mda_devsw = &pass_cdevsw;
666 	args.mda_unit = periph->unit_number;
667 	args.mda_uid = UID_ROOT;
668 	args.mda_gid = GID_OPERATOR;
669 	args.mda_mode = 0600;
670 	args.mda_si_drv1 = periph;
671 	args.mda_flags = MAKEDEV_NOWAIT;
672 	error = make_dev_s(&args, &softc->dev, "%s%d", periph->periph_name,
673 	    periph->unit_number);
674 	if (error != 0) {
675 		cam_periph_lock(periph);
676 		cam_periph_release_locked(periph);
677 		return (CAM_REQ_CMP_ERR);
678 	}
679 
680 	/*
681 	 * Hold a reference to the periph before we create the physical
682 	 * path alias so it can't go away.
683 	 */
684 	if (cam_periph_acquire(periph) != 0) {
685 		xpt_print(periph->path, "%s: lost periph during "
686 			  "registration!\n", __func__);
687 		cam_periph_lock(periph);
688 		return (CAM_REQ_CMP_ERR);
689 	}
690 
691 	cam_periph_lock(periph);
692 
693 	TASK_INIT(&softc->add_physpath_task, /*priority*/0,
694 		  pass_add_physpath, periph);
695 
696 	/*
697 	 * See if physical path information is already available.
698 	 */
699 	taskqueue_enqueue(taskqueue_thread, &softc->add_physpath_task);
700 
701 	/*
702 	 * Add an async callback so that we get notified if
703 	 * this device goes away or its physical path
704 	 * (stored in the advanced info data of the EDT) has
705 	 * changed.
706 	 */
707 	xpt_register_async(AC_LOST_DEVICE | AC_ADVINFO_CHANGED,
708 			   passasync, periph, periph->path);
709 
710 	if (bootverbose)
711 		xpt_announce_periph(periph, NULL);
712 
713 	return(CAM_REQ_CMP);
714 }
715 
716 static int
passopen(struct cdev * dev,int flags,int fmt,struct thread * td)717 passopen(struct cdev *dev, int flags, int fmt, struct thread *td)
718 {
719 	struct cam_periph *periph;
720 	struct pass_softc *softc;
721 	int error;
722 
723 	periph = (struct cam_periph *)dev->si_drv1;
724 	if (cam_periph_acquire(periph) != 0)
725 		return (ENXIO);
726 
727 	cam_periph_lock(periph);
728 
729 	softc = (struct pass_softc *)periph->softc;
730 
731 	if (softc->flags & PASS_FLAG_INVALID) {
732 		cam_periph_release_locked(periph);
733 		cam_periph_unlock(periph);
734 		return(ENXIO);
735 	}
736 
737 	/*
738 	 * Don't allow access when we're running at a high securelevel.
739 	 */
740 	error = securelevel_gt(td->td_ucred, 1);
741 	if (error) {
742 		cam_periph_release_locked(periph);
743 		cam_periph_unlock(periph);
744 		return(error);
745 	}
746 
747 	/*
748 	 * Only allow read-write access.
749 	 */
750 	if (((flags & FWRITE) == 0) || ((flags & FREAD) == 0)) {
751 		cam_periph_release_locked(periph);
752 		cam_periph_unlock(periph);
753 		return(EPERM);
754 	}
755 
756 	/*
757 	 * We don't allow nonblocking access.
758 	 */
759 	if ((flags & O_NONBLOCK) != 0) {
760 		xpt_print(periph->path, "can't do nonblocking access\n");
761 		cam_periph_release_locked(periph);
762 		cam_periph_unlock(periph);
763 		return(EINVAL);
764 	}
765 
766 	softc->open_count++;
767 
768 	cam_periph_unlock(periph);
769 
770 	return (error);
771 }
772 
773 static int
passclose(struct cdev * dev,int flag,int fmt,struct thread * td)774 passclose(struct cdev *dev, int flag, int fmt, struct thread *td)
775 {
776 	struct 	cam_periph *periph;
777 	struct  pass_softc *softc;
778 	struct mtx *mtx;
779 
780 	periph = (struct cam_periph *)dev->si_drv1;
781 	mtx = cam_periph_mtx(periph);
782 	mtx_lock(mtx);
783 
784 	softc = periph->softc;
785 	softc->open_count--;
786 
787 	if (softc->open_count == 0) {
788 		struct pass_io_req *io_req, *io_req2;
789 
790 		TAILQ_FOREACH_SAFE(io_req, &softc->done_queue, links, io_req2) {
791 			TAILQ_REMOVE(&softc->done_queue, io_req, links);
792 			passiocleanup(softc, io_req);
793 			uma_zfree(softc->pass_zone, io_req);
794 		}
795 
796 		TAILQ_FOREACH_SAFE(io_req, &softc->incoming_queue, links,
797 				   io_req2) {
798 			TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
799 			passiocleanup(softc, io_req);
800 			uma_zfree(softc->pass_zone, io_req);
801 		}
802 
803 		/*
804 		 * If there are any active I/Os, we need to forcibly acquire a
805 		 * reference to the peripheral so that we don't go away
806 		 * before they complete.  We'll release the reference when
807 		 * the abandoned queue is empty.
808 		 */
809 		io_req = TAILQ_FIRST(&softc->active_queue);
810 		if ((io_req != NULL)
811 		 && (softc->flags & PASS_FLAG_ABANDONED_REF_SET) == 0) {
812 			cam_periph_doacquire(periph);
813 			softc->flags |= PASS_FLAG_ABANDONED_REF_SET;
814 		}
815 
816 		/*
817 		 * Since the I/O in the active queue is not under our
818 		 * control, just set a flag so that we can clean it up when
819 		 * it completes and put it on the abandoned queue.  This
820 		 * will prevent our sending spurious completions in the
821 		 * event that the device is opened again before these I/Os
822 		 * complete.
823 		 */
824 		TAILQ_FOREACH_SAFE(io_req, &softc->active_queue, links,
825 				   io_req2) {
826 			TAILQ_REMOVE(&softc->active_queue, io_req, links);
827 			io_req->flags |= PASS_IO_ABANDONED;
828 			TAILQ_INSERT_TAIL(&softc->abandoned_queue, io_req,
829 					  links);
830 		}
831 	}
832 
833 	cam_periph_release_locked(periph);
834 
835 	/*
836 	 * We reference the lock directly here, instead of using
837 	 * cam_periph_unlock().  The reason is that the call to
838 	 * cam_periph_release_locked() above could result in the periph
839 	 * getting freed.  If that is the case, dereferencing the periph
840 	 * with a cam_periph_unlock() call would cause a page fault.
841 	 *
842 	 * cam_periph_release() avoids this problem using the same method,
843 	 * but we're manually acquiring and dropping the lock here to
844 	 * protect the open count and avoid another lock acquisition and
845 	 * release.
846 	 */
847 	mtx_unlock(mtx);
848 
849 	return (0);
850 }
851 
852 static void
passstart(struct cam_periph * periph,union ccb * start_ccb)853 passstart(struct cam_periph *periph, union ccb *start_ccb)
854 {
855 	struct pass_softc *softc;
856 
857 	softc = (struct pass_softc *)periph->softc;
858 
859 	switch (softc->state) {
860 	case PASS_STATE_NORMAL: {
861 		struct pass_io_req *io_req;
862 
863 		/*
864 		 * Check for any queued I/O requests that require an
865 		 * allocated slot.
866 		 */
867 		io_req = TAILQ_FIRST(&softc->incoming_queue);
868 		if (io_req == NULL) {
869 			xpt_release_ccb(start_ccb);
870 			break;
871 		}
872 		TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
873 		TAILQ_INSERT_TAIL(&softc->active_queue, io_req, links);
874 		/*
875 		 * Merge the user's CCB into the allocated CCB.
876 		 */
877 		xpt_merge_ccb(start_ccb, &io_req->ccb);
878 		start_ccb->ccb_h.ccb_type = PASS_CCB_QUEUED_IO;
879 		start_ccb->ccb_h.ccb_ioreq = io_req;
880 		start_ccb->ccb_h.cbfcnp = passdone;
881 		io_req->alloced_ccb = start_ccb;
882 		binuptime(&io_req->start_time);
883 		devstat_start_transaction(softc->device_stats,
884 					  &io_req->start_time);
885 
886 		xpt_action(start_ccb);
887 
888 		/*
889 		 * If we have any more I/O waiting, schedule ourselves again.
890 		 */
891 		if (!TAILQ_EMPTY(&softc->incoming_queue))
892 			xpt_schedule(periph, CAM_PRIORITY_NORMAL);
893 		break;
894 	}
895 	default:
896 		break;
897 	}
898 }
899 
900 static void
passdone(struct cam_periph * periph,union ccb * done_ccb)901 passdone(struct cam_periph *periph, union ccb *done_ccb)
902 {
903 	struct pass_softc *softc;
904 	struct ccb_hdr *hdr;
905 
906 	softc = (struct pass_softc *)periph->softc;
907 
908 	cam_periph_assert(periph, MA_OWNED);
909 
910 	hdr = &done_ccb->ccb_h;
911 	switch (hdr->ccb_type) {
912 	case PASS_CCB_QUEUED_IO: {
913 		struct pass_io_req *io_req;
914 
915 		io_req = hdr->ccb_ioreq;
916 #if 0
917 		xpt_print(periph->path, "%s: called for user CCB %p\n",
918 			  __func__, io_req->user_ccb_ptr);
919 #endif
920 		if (((hdr->status & CAM_STATUS_MASK) != CAM_REQ_CMP) &&
921 		    ((io_req->flags & PASS_IO_ABANDONED) == 0)) {
922 			int error;
923 			uint32_t cam_flags, sense_flags;
924 
925 			passflags(done_ccb, &cam_flags, &sense_flags);
926 			error = passerror(done_ccb, cam_flags, sense_flags);
927 
928 			if (error == ERESTART) {
929 				KASSERT(((sense_flags & SF_NO_RETRY) == 0),
930 				    ("passerror returned ERESTART with no retry requested\n"));
931 				return;
932 			}
933 		}
934 
935 		/*
936 		 * Copy the allocated CCB contents back to the malloced CCB
937 		 * so we can give status back to the user when he requests it.
938 		 */
939 		bcopy(done_ccb, &io_req->ccb, sizeof(*done_ccb));
940 
941 		/*
942 		 * Log data/transaction completion with devstat(9).
943 		 */
944 		switch (hdr->func_code) {
945 		case XPT_SCSI_IO:
946 			devstat_end_transaction(softc->device_stats,
947 			    done_ccb->csio.dxfer_len - done_ccb->csio.resid,
948 			    done_ccb->csio.tag_action & 0x3,
949 			    ((hdr->flags & CAM_DIR_MASK) ==
950 			    CAM_DIR_NONE) ? DEVSTAT_NO_DATA :
951 			    (hdr->flags & CAM_DIR_OUT) ?
952 			    DEVSTAT_WRITE : DEVSTAT_READ, NULL,
953 			    &io_req->start_time);
954 			break;
955 		case XPT_ATA_IO:
956 			devstat_end_transaction(softc->device_stats,
957 			    done_ccb->ataio.dxfer_len - done_ccb->ataio.resid,
958 			    0, /* Not used in ATA */
959 			    ((hdr->flags & CAM_DIR_MASK) ==
960 			    CAM_DIR_NONE) ? DEVSTAT_NO_DATA :
961 			    (hdr->flags & CAM_DIR_OUT) ?
962 			    DEVSTAT_WRITE : DEVSTAT_READ, NULL,
963 			    &io_req->start_time);
964 			break;
965 		case XPT_SMP_IO:
966 			/*
967 			 * XXX KDM this isn't quite right, but there isn't
968 			 * currently an easy way to represent a bidirectional
969 			 * transfer in devstat.  The only way to do it
970 			 * and have the byte counts come out right would
971 			 * mean that we would have to record two
972 			 * transactions, one for the request and one for the
973 			 * response.  For now, so that we report something,
974 			 * just treat the entire thing as a read.
975 			 */
976 			devstat_end_transaction(softc->device_stats,
977 			    done_ccb->smpio.smp_request_len +
978 			    done_ccb->smpio.smp_response_len,
979 			    DEVSTAT_TAG_SIMPLE, DEVSTAT_READ, NULL,
980 			    &io_req->start_time);
981 			break;
982 		/* XXX XPT_NVME_IO and XPT_NVME_ADMIN need cases here for resid */
983 		default:
984 			devstat_end_transaction(softc->device_stats, 0,
985 			    DEVSTAT_TAG_NONE, DEVSTAT_NO_DATA, NULL,
986 			    &io_req->start_time);
987 			break;
988 		}
989 
990 		/*
991 		 * In the normal case, take the completed I/O off of the
992 		 * active queue and put it on the done queue.  Notitfy the
993 		 * user that we have a completed I/O.
994 		 */
995 		if ((io_req->flags & PASS_IO_ABANDONED) == 0) {
996 			TAILQ_REMOVE(&softc->active_queue, io_req, links);
997 			TAILQ_INSERT_TAIL(&softc->done_queue, io_req, links);
998 			selwakeuppri(&softc->read_select, PRIBIO);
999 			KNOTE_LOCKED(&softc->read_select.si_note, 0);
1000 		} else {
1001 			/*
1002 			 * In the case of an abandoned I/O (final close
1003 			 * without fetching the I/O), take it off of the
1004 			 * abandoned queue and free it.
1005 			 */
1006 			TAILQ_REMOVE(&softc->abandoned_queue, io_req, links);
1007 			passiocleanup(softc, io_req);
1008 			uma_zfree(softc->pass_zone, io_req);
1009 
1010 			/*
1011 			 * Release the done_ccb here, since we may wind up
1012 			 * freeing the peripheral when we decrement the
1013 			 * reference count below.
1014 			 */
1015 			xpt_release_ccb(done_ccb);
1016 
1017 			/*
1018 			 * If the abandoned queue is empty, we can release
1019 			 * our reference to the periph since we won't have
1020 			 * any more completions coming.
1021 			 */
1022 			if ((TAILQ_EMPTY(&softc->abandoned_queue))
1023 			 && (softc->flags & PASS_FLAG_ABANDONED_REF_SET)) {
1024 				softc->flags &= ~PASS_FLAG_ABANDONED_REF_SET;
1025 				cam_periph_release_locked(periph);
1026 			}
1027 
1028 			/*
1029 			 * We have already released the CCB, so we can
1030 			 * return.
1031 			 */
1032 			return;
1033 		}
1034 		break;
1035 	}
1036 	}
1037 	xpt_release_ccb(done_ccb);
1038 }
1039 
1040 static int
passcreatezone(struct cam_periph * periph)1041 passcreatezone(struct cam_periph *periph)
1042 {
1043 	struct pass_softc *softc;
1044 	int error;
1045 
1046 	error = 0;
1047 	softc = (struct pass_softc *)periph->softc;
1048 
1049 	cam_periph_assert(periph, MA_OWNED);
1050 	KASSERT(((softc->flags & PASS_FLAG_ZONE_VALID) == 0),
1051 		("%s called when the pass(4) zone is valid!\n", __func__));
1052 	KASSERT((softc->pass_zone == NULL),
1053 		("%s called when the pass(4) zone is allocated!\n", __func__));
1054 
1055 	if ((softc->flags & PASS_FLAG_ZONE_INPROG) == 0) {
1056 		/*
1057 		 * We're the first context through, so we need to create
1058 		 * the pass(4) UMA zone for I/O requests.
1059 		 */
1060 		softc->flags |= PASS_FLAG_ZONE_INPROG;
1061 
1062 		/*
1063 		 * uma_zcreate() does a blocking (M_WAITOK) allocation,
1064 		 * so we cannot hold a mutex while we call it.
1065 		 */
1066 		cam_periph_unlock(periph);
1067 
1068 		softc->pass_zone = uma_zcreate(softc->zone_name,
1069 		    sizeof(struct pass_io_req), NULL, NULL, NULL, NULL,
1070 		    /*align*/ 0, /*flags*/ 0);
1071 
1072 		softc->pass_io_zone = uma_zcreate(softc->io_zone_name,
1073 		    softc->io_zone_size, NULL, NULL, NULL, NULL,
1074 		    /*align*/ 0, /*flags*/ 0);
1075 
1076 		cam_periph_lock(periph);
1077 
1078 		if ((softc->pass_zone == NULL)
1079 		 || (softc->pass_io_zone == NULL)) {
1080 			if (softc->pass_zone == NULL)
1081 				xpt_print(periph->path, "unable to allocate "
1082 				    "IO Req UMA zone\n");
1083 			else
1084 				xpt_print(periph->path, "unable to allocate "
1085 				    "IO UMA zone\n");
1086 			softc->flags &= ~PASS_FLAG_ZONE_INPROG;
1087 			goto bailout;
1088 		}
1089 
1090 		/*
1091 		 * Set the flags appropriately and notify any other waiters.
1092 		 */
1093 		softc->flags &= ~PASS_FLAG_ZONE_INPROG;
1094 		softc->flags |= PASS_FLAG_ZONE_VALID;
1095 		wakeup(&softc->pass_zone);
1096 	} else {
1097 		/*
1098 		 * In this case, the UMA zone has not yet been created, but
1099 		 * another context is in the process of creating it.  We
1100 		 * need to sleep until the creation is either done or has
1101 		 * failed.
1102 		 */
1103 		while ((softc->flags & PASS_FLAG_ZONE_INPROG)
1104 		    && ((softc->flags & PASS_FLAG_ZONE_VALID) == 0)) {
1105 			error = msleep(&softc->pass_zone,
1106 				       cam_periph_mtx(periph), PRIBIO,
1107 				       "paszon", 0);
1108 			if (error != 0)
1109 				goto bailout;
1110 		}
1111 		/*
1112 		 * If the zone creation failed, no luck for the user.
1113 		 */
1114 		if ((softc->flags & PASS_FLAG_ZONE_VALID) == 0){
1115 			error = ENOMEM;
1116 			goto bailout;
1117 		}
1118 	}
1119 bailout:
1120 	return (error);
1121 }
1122 
1123 static void
passiocleanup(struct pass_softc * softc,struct pass_io_req * io_req)1124 passiocleanup(struct pass_softc *softc, struct pass_io_req *io_req)
1125 {
1126 	union ccb *ccb;
1127 	struct ccb_hdr *hdr;
1128 	uint8_t **data_ptrs[CAM_PERIPH_MAXMAPS];
1129 	int i, numbufs;
1130 
1131 	ccb = &io_req->ccb;
1132 	hdr = &ccb->ccb_h;
1133 
1134 	switch (hdr->func_code) {
1135 	case XPT_DEV_MATCH:
1136 		numbufs = min(io_req->num_bufs, 2);
1137 
1138 		if (numbufs == 1) {
1139 			data_ptrs[0] = (uint8_t **)&ccb->cdm.matches;
1140 		} else {
1141 			data_ptrs[0] = (uint8_t **)&ccb->cdm.patterns;
1142 			data_ptrs[1] = (uint8_t **)&ccb->cdm.matches;
1143 		}
1144 		break;
1145 	case XPT_SCSI_IO:
1146 	case XPT_CONT_TARGET_IO:
1147 		data_ptrs[0] = &ccb->csio.data_ptr;
1148 		numbufs = min(io_req->num_bufs, 1);
1149 		break;
1150 	case XPT_ATA_IO:
1151 		data_ptrs[0] = &ccb->ataio.data_ptr;
1152 		numbufs = min(io_req->num_bufs, 1);
1153 		break;
1154 	case XPT_SMP_IO:
1155 		numbufs = min(io_req->num_bufs, 2);
1156 		data_ptrs[0] = &ccb->smpio.smp_request;
1157 		data_ptrs[1] = &ccb->smpio.smp_response;
1158 		break;
1159 	case XPT_DEV_ADVINFO:
1160 		numbufs = min(io_req->num_bufs, 1);
1161 		data_ptrs[0] = (uint8_t **)&ccb->cdai.buf;
1162 		break;
1163 	case XPT_NVME_IO:
1164 	case XPT_NVME_ADMIN:
1165 		data_ptrs[0] = &ccb->nvmeio.data_ptr;
1166 		numbufs = min(io_req->num_bufs, 1);
1167 		break;
1168 	default:
1169 		/* allow ourselves to be swapped once again */
1170 		return;
1171 		break; /* NOTREACHED */
1172 	}
1173 
1174 	if (io_req->flags & PASS_IO_USER_SEG_MALLOC) {
1175 		free(io_req->user_segptr, M_SCSIPASS);
1176 		io_req->user_segptr = NULL;
1177 	}
1178 
1179 	/*
1180 	 * We only want to free memory we malloced.
1181 	 */
1182 	if (io_req->data_flags == CAM_DATA_VADDR) {
1183 		for (i = 0; i < io_req->num_bufs; i++) {
1184 			if (io_req->kern_bufs[i] == NULL)
1185 				continue;
1186 
1187 			free(io_req->kern_bufs[i], M_SCSIPASS);
1188 			io_req->kern_bufs[i] = NULL;
1189 		}
1190 	} else if (io_req->data_flags == CAM_DATA_SG) {
1191 		for (i = 0; i < io_req->num_kern_segs; i++) {
1192 			if ((uint8_t *)(uintptr_t)
1193 			    io_req->kern_segptr[i].ds_addr == NULL)
1194 				continue;
1195 
1196 			uma_zfree(softc->pass_io_zone, (uint8_t *)(uintptr_t)
1197 			    io_req->kern_segptr[i].ds_addr);
1198 			io_req->kern_segptr[i].ds_addr = 0;
1199 		}
1200 	}
1201 
1202 	if (io_req->flags & PASS_IO_KERN_SEG_MALLOC) {
1203 		free(io_req->kern_segptr, M_SCSIPASS);
1204 		io_req->kern_segptr = NULL;
1205 	}
1206 
1207 	if (io_req->data_flags != CAM_DATA_PADDR) {
1208 		for (i = 0; i < numbufs; i++) {
1209 			/*
1210 			 * Restore the user's buffer pointers to their
1211 			 * previous values.
1212 			 */
1213 			if (io_req->user_bufs[i] != NULL)
1214 				*data_ptrs[i] = io_req->user_bufs[i];
1215 		}
1216 	}
1217 
1218 }
1219 
1220 static int
passcopysglist(struct cam_periph * periph,struct pass_io_req * io_req,ccb_flags direction)1221 passcopysglist(struct cam_periph *periph, struct pass_io_req *io_req,
1222 	       ccb_flags direction)
1223 {
1224 	bus_size_t kern_watermark, user_watermark, len_to_copy;
1225 	bus_dma_segment_t *user_sglist, *kern_sglist;
1226 	int i, j, error;
1227 
1228 	error = 0;
1229 	kern_watermark = 0;
1230 	user_watermark = 0;
1231 	len_to_copy = 0;
1232 	user_sglist = io_req->user_segptr;
1233 	kern_sglist = io_req->kern_segptr;
1234 
1235 	for (i = 0, j = 0; i < io_req->num_user_segs &&
1236 	     j < io_req->num_kern_segs;) {
1237 		uint8_t *user_ptr, *kern_ptr;
1238 
1239 		len_to_copy = min(user_sglist[i].ds_len -user_watermark,
1240 		    kern_sglist[j].ds_len - kern_watermark);
1241 
1242 		user_ptr = (uint8_t *)(uintptr_t)user_sglist[i].ds_addr;
1243 		user_ptr = user_ptr + user_watermark;
1244 		kern_ptr = (uint8_t *)(uintptr_t)kern_sglist[j].ds_addr;
1245 		kern_ptr = kern_ptr + kern_watermark;
1246 
1247 		user_watermark += len_to_copy;
1248 		kern_watermark += len_to_copy;
1249 
1250 		if (direction == CAM_DIR_IN) {
1251 			error = copyout(kern_ptr, user_ptr, len_to_copy);
1252 			if (error != 0) {
1253 				xpt_print(periph->path, "%s: copyout of %u "
1254 					  "bytes from %p to %p failed with "
1255 					  "error %d\n", __func__, len_to_copy,
1256 					  kern_ptr, user_ptr, error);
1257 				goto bailout;
1258 			}
1259 		} else {
1260 			error = copyin(user_ptr, kern_ptr, len_to_copy);
1261 			if (error != 0) {
1262 				xpt_print(periph->path, "%s: copyin of %u "
1263 					  "bytes from %p to %p failed with "
1264 					  "error %d\n", __func__, len_to_copy,
1265 					  user_ptr, kern_ptr, error);
1266 				goto bailout;
1267 			}
1268 		}
1269 
1270 		if (user_sglist[i].ds_len == user_watermark) {
1271 			i++;
1272 			user_watermark = 0;
1273 		}
1274 
1275 		if (kern_sglist[j].ds_len == kern_watermark) {
1276 			j++;
1277 			kern_watermark = 0;
1278 		}
1279 	}
1280 
1281 bailout:
1282 
1283 	return (error);
1284 }
1285 
1286 static int
passmemsetup(struct cam_periph * periph,struct pass_io_req * io_req)1287 passmemsetup(struct cam_periph *periph, struct pass_io_req *io_req)
1288 {
1289 	union ccb *ccb;
1290 	struct ccb_hdr *hdr;
1291 	struct pass_softc *softc;
1292 	int numbufs, i;
1293 	uint8_t **data_ptrs[CAM_PERIPH_MAXMAPS];
1294 	uint32_t lengths[CAM_PERIPH_MAXMAPS];
1295 	uint32_t dirs[CAM_PERIPH_MAXMAPS];
1296 	uint32_t num_segs;
1297 	uint16_t *seg_cnt_ptr;
1298 	size_t maxmap;
1299 	int error;
1300 
1301 	cam_periph_assert(periph, MA_NOTOWNED);
1302 
1303 	softc = periph->softc;
1304 
1305 	error = 0;
1306 	ccb = &io_req->ccb;
1307 	hdr = &ccb->ccb_h;
1308 	maxmap = 0;
1309 	num_segs = 0;
1310 	seg_cnt_ptr = NULL;
1311 
1312 	switch(hdr->func_code) {
1313 	case XPT_DEV_MATCH:
1314 		if (ccb->cdm.match_buf_len == 0) {
1315 			printf("%s: invalid match buffer length 0\n", __func__);
1316 			return(EINVAL);
1317 		}
1318 		if (ccb->cdm.pattern_buf_len > 0) {
1319 			data_ptrs[0] = (uint8_t **)&ccb->cdm.patterns;
1320 			lengths[0] = ccb->cdm.pattern_buf_len;
1321 			dirs[0] = CAM_DIR_OUT;
1322 			data_ptrs[1] = (uint8_t **)&ccb->cdm.matches;
1323 			lengths[1] = ccb->cdm.match_buf_len;
1324 			dirs[1] = CAM_DIR_IN;
1325 			numbufs = 2;
1326 		} else {
1327 			data_ptrs[0] = (uint8_t **)&ccb->cdm.matches;
1328 			lengths[0] = ccb->cdm.match_buf_len;
1329 			dirs[0] = CAM_DIR_IN;
1330 			numbufs = 1;
1331 		}
1332 		io_req->data_flags = CAM_DATA_VADDR;
1333 		break;
1334 	case XPT_SCSI_IO:
1335 	case XPT_CONT_TARGET_IO:
1336 		if ((hdr->flags & CAM_DIR_MASK) == CAM_DIR_NONE)
1337 			return(0);
1338 
1339 		/*
1340 		 * The user shouldn't be able to supply a bio.
1341 		 */
1342 		if ((hdr->flags & CAM_DATA_MASK) == CAM_DATA_BIO)
1343 			return (EINVAL);
1344 
1345 		io_req->data_flags = hdr->flags & CAM_DATA_MASK;
1346 
1347 		data_ptrs[0] = &ccb->csio.data_ptr;
1348 		lengths[0] = ccb->csio.dxfer_len;
1349 		dirs[0] = hdr->flags & CAM_DIR_MASK;
1350 		num_segs = ccb->csio.sglist_cnt;
1351 		seg_cnt_ptr = &ccb->csio.sglist_cnt;
1352 		numbufs = 1;
1353 		maxmap = softc->maxio;
1354 		break;
1355 	case XPT_ATA_IO:
1356 		if ((hdr->flags & CAM_DIR_MASK) == CAM_DIR_NONE)
1357 			return(0);
1358 
1359 		/*
1360 		 * We only support a single virtual address for ATA I/O.
1361 		 */
1362 		if ((hdr->flags & CAM_DATA_MASK) != CAM_DATA_VADDR)
1363 			return (EINVAL);
1364 
1365 		io_req->data_flags = CAM_DATA_VADDR;
1366 
1367 		data_ptrs[0] = &ccb->ataio.data_ptr;
1368 		lengths[0] = ccb->ataio.dxfer_len;
1369 		dirs[0] = hdr->flags & CAM_DIR_MASK;
1370 		numbufs = 1;
1371 		maxmap = softc->maxio;
1372 		break;
1373 	case XPT_SMP_IO:
1374 		io_req->data_flags = CAM_DATA_VADDR;
1375 
1376 		data_ptrs[0] = &ccb->smpio.smp_request;
1377 		lengths[0] = ccb->smpio.smp_request_len;
1378 		dirs[0] = CAM_DIR_OUT;
1379 		data_ptrs[1] = &ccb->smpio.smp_response;
1380 		lengths[1] = ccb->smpio.smp_response_len;
1381 		dirs[1] = CAM_DIR_IN;
1382 		numbufs = 2;
1383 		maxmap = softc->maxio;
1384 		break;
1385 	case XPT_DEV_ADVINFO:
1386 		if (ccb->cdai.bufsiz == 0)
1387 			return (0);
1388 
1389 		io_req->data_flags = CAM_DATA_VADDR;
1390 
1391 		data_ptrs[0] = (uint8_t **)&ccb->cdai.buf;
1392 		lengths[0] = ccb->cdai.bufsiz;
1393 		dirs[0] = CAM_DIR_IN;
1394 		numbufs = 1;
1395 		break;
1396 	case XPT_NVME_ADMIN:
1397 	case XPT_NVME_IO:
1398 		if ((hdr->flags & CAM_DIR_MASK) == CAM_DIR_NONE)
1399 			return (0);
1400 
1401 		io_req->data_flags = hdr->flags & CAM_DATA_MASK;
1402 
1403 		data_ptrs[0] = &ccb->nvmeio.data_ptr;
1404 		lengths[0] = ccb->nvmeio.dxfer_len;
1405 		dirs[0] = hdr->flags & CAM_DIR_MASK;
1406 		num_segs = ccb->nvmeio.sglist_cnt;
1407 		seg_cnt_ptr = &ccb->nvmeio.sglist_cnt;
1408 		numbufs = 1;
1409 		maxmap = softc->maxio;
1410 		break;
1411 	default:
1412 		return(EINVAL);
1413 		break; /* NOTREACHED */
1414 	}
1415 
1416 	io_req->num_bufs = numbufs;
1417 
1418 	/*
1419 	 * If there is a maximum, check to make sure that the user's
1420 	 * request fits within the limit.  In general, we should only have
1421 	 * a maximum length for requests that go to hardware.  Otherwise it
1422 	 * is whatever we're able to malloc.
1423 	 */
1424 	for (i = 0; i < numbufs; i++) {
1425 		io_req->user_bufs[i] = *data_ptrs[i];
1426 		io_req->dirs[i] = dirs[i];
1427 		io_req->lengths[i] = lengths[i];
1428 
1429 		if (maxmap == 0)
1430 			continue;
1431 
1432 		if (lengths[i] <= maxmap)
1433 			continue;
1434 
1435 		xpt_print(periph->path, "%s: data length %u > max allowed %u "
1436 			  "bytes\n", __func__, lengths[i], maxmap);
1437 		error = EINVAL;
1438 		goto bailout;
1439 	}
1440 
1441 	switch (io_req->data_flags) {
1442 	case CAM_DATA_VADDR:
1443 		/* Map or copy the buffer into kernel address space */
1444 		for (i = 0; i < numbufs; i++) {
1445 			uint8_t *tmp_buf;
1446 
1447 			/*
1448 			 * If for some reason no length is specified, we
1449 			 * don't need to allocate anything.
1450 			 */
1451 			if (io_req->lengths[i] == 0)
1452 				continue;
1453 
1454 			tmp_buf = malloc(lengths[i], M_SCSIPASS,
1455 					 M_WAITOK | M_ZERO);
1456 			io_req->kern_bufs[i] = tmp_buf;
1457 			*data_ptrs[i] = tmp_buf;
1458 
1459 #if 0
1460 			xpt_print(periph->path, "%s: malloced %p len %u, user "
1461 				  "buffer %p, operation: %s\n", __func__,
1462 				  tmp_buf, lengths[i], io_req->user_bufs[i],
1463 				  (dirs[i] == CAM_DIR_IN) ? "read" : "write");
1464 #endif
1465 			/*
1466 			 * We only need to copy in if the user is writing.
1467 			 */
1468 			if (dirs[i] != CAM_DIR_OUT)
1469 				continue;
1470 
1471 			error = copyin(io_req->user_bufs[i],
1472 				       io_req->kern_bufs[i], lengths[i]);
1473 			if (error != 0) {
1474 				xpt_print(periph->path, "%s: copy of user "
1475 					  "buffer from %p to %p failed with "
1476 					  "error %d\n", __func__,
1477 					  io_req->user_bufs[i],
1478 					  io_req->kern_bufs[i], error);
1479 				goto bailout;
1480 			}
1481 		}
1482 		break;
1483 	case CAM_DATA_PADDR:
1484 		/* Pass down the pointer as-is */
1485 		break;
1486 	case CAM_DATA_SG: {
1487 		size_t sg_length, size_to_go, alloc_size;
1488 		uint32_t num_segs_needed;
1489 
1490 		/*
1491 		 * Copy the user S/G list in, and then copy in the
1492 		 * individual segments.
1493 		 */
1494 		/*
1495 		 * We shouldn't see this, but check just in case.
1496 		 */
1497 		if (numbufs != 1) {
1498 			xpt_print(periph->path, "%s: cannot currently handle "
1499 				  "more than one S/G list per CCB\n", __func__);
1500 			error = EINVAL;
1501 			goto bailout;
1502 		}
1503 
1504 		/*
1505 		 * We have to have at least one segment.
1506 		 */
1507 		if (num_segs == 0) {
1508 			xpt_print(periph->path, "%s: CAM_DATA_SG flag set, "
1509 				  "but sglist_cnt=0!\n", __func__);
1510 			error = EINVAL;
1511 			goto bailout;
1512 		}
1513 
1514 		/*
1515 		 * Make sure the user specified the total length and didn't
1516 		 * just leave it to us to decode the S/G list.
1517 		 */
1518 		if (lengths[0] == 0) {
1519 			xpt_print(periph->path, "%s: no dxfer_len specified, "
1520 				  "but CAM_DATA_SG flag is set!\n", __func__);
1521 			error = EINVAL;
1522 			goto bailout;
1523 		}
1524 
1525 		/*
1526 		 * We allocate buffers in io_zone_size increments for an
1527 		 * S/G list.  This will generally be maxphys.
1528 		 */
1529 		if (lengths[0] <= softc->io_zone_size)
1530 			num_segs_needed = 1;
1531 		else {
1532 			num_segs_needed = lengths[0] / softc->io_zone_size;
1533 			if ((lengths[0] % softc->io_zone_size) != 0)
1534 				num_segs_needed++;
1535 		}
1536 
1537 		/* Figure out the size of the S/G list */
1538 		sg_length = num_segs * sizeof(bus_dma_segment_t);
1539 		io_req->num_user_segs = num_segs;
1540 		io_req->num_kern_segs = num_segs_needed;
1541 
1542 		/* Save the user's S/G list pointer for later restoration */
1543 		io_req->user_bufs[0] = *data_ptrs[0];
1544 
1545 		/*
1546 		 * If we have enough segments allocated by default to handle
1547 		 * the length of the user's S/G list,
1548 		 */
1549 		if (num_segs > PASS_MAX_SEGS) {
1550 			io_req->user_segptr = malloc(sizeof(bus_dma_segment_t) *
1551 			    num_segs, M_SCSIPASS, M_WAITOK | M_ZERO);
1552 			io_req->flags |= PASS_IO_USER_SEG_MALLOC;
1553 		} else
1554 			io_req->user_segptr = io_req->user_segs;
1555 
1556 		error = copyin(*data_ptrs[0], io_req->user_segptr, sg_length);
1557 		if (error != 0) {
1558 			xpt_print(periph->path, "%s: copy of user S/G list "
1559 				  "from %p to %p failed with error %d\n",
1560 				  __func__, *data_ptrs[0], io_req->user_segptr,
1561 				  error);
1562 			goto bailout;
1563 		}
1564 
1565 		if (num_segs_needed > PASS_MAX_SEGS) {
1566 			io_req->kern_segptr = malloc(sizeof(bus_dma_segment_t) *
1567 			    num_segs_needed, M_SCSIPASS, M_WAITOK | M_ZERO);
1568 			io_req->flags |= PASS_IO_KERN_SEG_MALLOC;
1569 		} else {
1570 			io_req->kern_segptr = io_req->kern_segs;
1571 		}
1572 
1573 		/*
1574 		 * Allocate the kernel S/G list.
1575 		 */
1576 		for (size_to_go = lengths[0], i = 0;
1577 		     size_to_go > 0 && i < num_segs_needed;
1578 		     i++, size_to_go -= alloc_size) {
1579 			uint8_t *kern_ptr;
1580 
1581 			alloc_size = min(size_to_go, softc->io_zone_size);
1582 			kern_ptr = uma_zalloc(softc->pass_io_zone, M_WAITOK);
1583 			io_req->kern_segptr[i].ds_addr =
1584 			    (bus_addr_t)(uintptr_t)kern_ptr;
1585 			io_req->kern_segptr[i].ds_len = alloc_size;
1586 		}
1587 		if (size_to_go > 0) {
1588 			printf("%s: size_to_go = %zu, software error!\n",
1589 			       __func__, size_to_go);
1590 			error = EINVAL;
1591 			goto bailout;
1592 		}
1593 
1594 		*data_ptrs[0] = (uint8_t *)io_req->kern_segptr;
1595 		*seg_cnt_ptr = io_req->num_kern_segs;
1596 
1597 		/*
1598 		 * We only need to copy data here if the user is writing.
1599 		 */
1600 		if (dirs[0] == CAM_DIR_OUT)
1601 			error = passcopysglist(periph, io_req, dirs[0]);
1602 		break;
1603 	}
1604 	case CAM_DATA_SG_PADDR: {
1605 		size_t sg_length;
1606 
1607 		/*
1608 		 * We shouldn't see this, but check just in case.
1609 		 */
1610 		if (numbufs != 1) {
1611 			printf("%s: cannot currently handle more than one "
1612 			       "S/G list per CCB\n", __func__);
1613 			error = EINVAL;
1614 			goto bailout;
1615 		}
1616 
1617 		/*
1618 		 * We have to have at least one segment.
1619 		 */
1620 		if (num_segs == 0) {
1621 			xpt_print(periph->path, "%s: CAM_DATA_SG_PADDR flag "
1622 				  "set, but sglist_cnt=0!\n", __func__);
1623 			error = EINVAL;
1624 			goto bailout;
1625 		}
1626 
1627 		/*
1628 		 * Make sure the user specified the total length and didn't
1629 		 * just leave it to us to decode the S/G list.
1630 		 */
1631 		if (lengths[0] == 0) {
1632 			xpt_print(periph->path, "%s: no dxfer_len specified, "
1633 				  "but CAM_DATA_SG flag is set!\n", __func__);
1634 			error = EINVAL;
1635 			goto bailout;
1636 		}
1637 
1638 		/* Figure out the size of the S/G list */
1639 		sg_length = num_segs * sizeof(bus_dma_segment_t);
1640 		io_req->num_user_segs = num_segs;
1641 		io_req->num_kern_segs = io_req->num_user_segs;
1642 
1643 		/* Save the user's S/G list pointer for later restoration */
1644 		io_req->user_bufs[0] = *data_ptrs[0];
1645 
1646 		if (num_segs > PASS_MAX_SEGS) {
1647 			io_req->user_segptr = malloc(sizeof(bus_dma_segment_t) *
1648 			    num_segs, M_SCSIPASS, M_WAITOK | M_ZERO);
1649 			io_req->flags |= PASS_IO_USER_SEG_MALLOC;
1650 		} else
1651 			io_req->user_segptr = io_req->user_segs;
1652 
1653 		io_req->kern_segptr = io_req->user_segptr;
1654 
1655 		error = copyin(*data_ptrs[0], io_req->user_segptr, sg_length);
1656 		if (error != 0) {
1657 			xpt_print(periph->path, "%s: copy of user S/G list "
1658 				  "from %p to %p failed with error %d\n",
1659 				  __func__, *data_ptrs[0], io_req->user_segptr,
1660 				  error);
1661 			goto bailout;
1662 		}
1663 		break;
1664 	}
1665 	default:
1666 	case CAM_DATA_BIO:
1667 		/*
1668 		 * A user shouldn't be attaching a bio to the CCB.  It
1669 		 * isn't a user-accessible structure.
1670 		 */
1671 		error = EINVAL;
1672 		break;
1673 	}
1674 
1675 bailout:
1676 	if (error != 0)
1677 		passiocleanup(softc, io_req);
1678 
1679 	return (error);
1680 }
1681 
1682 static int
passmemdone(struct cam_periph * periph,struct pass_io_req * io_req)1683 passmemdone(struct cam_periph *periph, struct pass_io_req *io_req)
1684 {
1685 	struct pass_softc *softc;
1686 	int error;
1687 	int i;
1688 
1689 	error = 0;
1690 	softc = (struct pass_softc *)periph->softc;
1691 
1692 	switch (io_req->data_flags) {
1693 	case CAM_DATA_VADDR:
1694 		/*
1695 		 * Copy back to the user buffer if this was a read.
1696 		 */
1697 		for (i = 0; i < io_req->num_bufs; i++) {
1698 			if (io_req->dirs[i] != CAM_DIR_IN)
1699 				continue;
1700 
1701 			error = copyout(io_req->kern_bufs[i],
1702 			    io_req->user_bufs[i], io_req->lengths[i]);
1703 			if (error != 0) {
1704 				xpt_print(periph->path, "Unable to copy %u "
1705 					  "bytes from %p to user address %p\n",
1706 					  io_req->lengths[i],
1707 					  io_req->kern_bufs[i],
1708 					  io_req->user_bufs[i]);
1709 				goto bailout;
1710 			}
1711 		}
1712 		break;
1713 	case CAM_DATA_PADDR:
1714 		/* Do nothing.  The pointer is a physical address already */
1715 		break;
1716 	case CAM_DATA_SG:
1717 		/*
1718 		 * Copy back to the user buffer if this was a read.
1719 		 * Restore the user's S/G list buffer pointer.
1720 		 */
1721 		if (io_req->dirs[0] == CAM_DIR_IN)
1722 			error = passcopysglist(periph, io_req, io_req->dirs[0]);
1723 		break;
1724 	case CAM_DATA_SG_PADDR:
1725 		/*
1726 		 * Restore the user's S/G list buffer pointer.  No need to
1727 		 * copy.
1728 		 */
1729 		break;
1730 	default:
1731 	case CAM_DATA_BIO:
1732 		error = EINVAL;
1733 		break;
1734 	}
1735 
1736 bailout:
1737 	/*
1738 	 * Reset the user's pointers to their original values and free
1739 	 * allocated memory.
1740 	 */
1741 	passiocleanup(softc, io_req);
1742 
1743 	return (error);
1744 }
1745 
1746 static int
passioctl(struct cdev * dev,u_long cmd,caddr_t addr,int flag,struct thread * td)1747 passioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flag, struct thread *td)
1748 {
1749 	int error;
1750 
1751 	if ((error = passdoioctl(dev, cmd, addr, flag, td)) == ENOTTY) {
1752 		error = cam_compat_ioctl(dev, cmd, addr, flag, td, passdoioctl);
1753 	}
1754 	return (error);
1755 }
1756 
1757 static int
passdoioctl(struct cdev * dev,u_long cmd,caddr_t addr,int flag,struct thread * td)1758 passdoioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flag, struct thread *td)
1759 {
1760 	struct	cam_periph *periph;
1761 	struct	pass_softc *softc;
1762 	int	error;
1763 	uint32_t priority;
1764 
1765 	periph = (struct cam_periph *)dev->si_drv1;
1766 	cam_periph_lock(periph);
1767 	softc = (struct pass_softc *)periph->softc;
1768 
1769 	error = 0;
1770 
1771 	switch (cmd) {
1772 	case CAMIOCOMMAND:
1773 	{
1774 		union ccb *inccb;
1775 		union ccb *ccb;
1776 		int ccb_malloced;
1777 
1778 		inccb = (union ccb *)addr;
1779 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
1780 		if (inccb->ccb_h.func_code == XPT_SCSI_IO)
1781 			inccb->csio.bio = NULL;
1782 #endif
1783 
1784 		if (inccb->ccb_h.flags & CAM_UNLOCKED) {
1785 			error = EINVAL;
1786 			break;
1787 		}
1788 
1789 		/*
1790 		 * Some CCB types, like scan bus and scan lun can only go
1791 		 * through the transport layer device.
1792 		 */
1793 		if (inccb->ccb_h.func_code & XPT_FC_XPT_ONLY) {
1794 			xpt_print(periph->path, "CCB function code %#x is "
1795 			    "restricted to the XPT device\n",
1796 			    inccb->ccb_h.func_code);
1797 			error = ENODEV;
1798 			break;
1799 		}
1800 
1801 		/* Compatibility for RL/priority-unaware code. */
1802 		priority = inccb->ccb_h.pinfo.priority;
1803 		if (priority <= CAM_PRIORITY_OOB)
1804 		    priority += CAM_PRIORITY_OOB + 1;
1805 
1806 		/*
1807 		 * Non-immediate CCBs need a CCB from the per-device pool
1808 		 * of CCBs, which is scheduled by the transport layer.
1809 		 * Immediate CCBs and user-supplied CCBs should just be
1810 		 * malloced.
1811 		 */
1812 		if ((inccb->ccb_h.func_code & XPT_FC_QUEUED)
1813 		 && ((inccb->ccb_h.func_code & XPT_FC_USER_CCB) == 0)) {
1814 			ccb = cam_periph_getccb(periph, priority);
1815 			ccb_malloced = 0;
1816 		} else {
1817 			ccb = xpt_alloc_ccb_nowait();
1818 
1819 			if (ccb != NULL)
1820 				xpt_setup_ccb(&ccb->ccb_h, periph->path,
1821 					      priority);
1822 			ccb_malloced = 1;
1823 		}
1824 
1825 		if (ccb == NULL) {
1826 			xpt_print(periph->path, "unable to allocate CCB\n");
1827 			error = ENOMEM;
1828 			break;
1829 		}
1830 
1831 		error = passsendccb(periph, ccb, inccb);
1832 
1833 		if (ccb_malloced)
1834 			xpt_free_ccb(ccb);
1835 		else
1836 			xpt_release_ccb(ccb);
1837 
1838 		break;
1839 	}
1840 	case CAMIOQUEUE:
1841 	{
1842 		struct pass_io_req *io_req;
1843 		union ccb **user_ccb, *ccb;
1844 		xpt_opcode fc;
1845 
1846 #ifdef COMPAT_FREEBSD32
1847 		if (SV_PROC_FLAG(td->td_proc, SV_ILP32)) {
1848 			error = ENOTTY;
1849 			goto bailout;
1850 		}
1851 #endif
1852 		if ((softc->flags & PASS_FLAG_ZONE_VALID) == 0) {
1853 			error = passcreatezone(periph);
1854 			if (error != 0)
1855 				goto bailout;
1856 		}
1857 
1858 		/*
1859 		 * We're going to do a blocking allocation for this I/O
1860 		 * request, so we have to drop the lock.
1861 		 */
1862 		cam_periph_unlock(periph);
1863 
1864 		io_req = uma_zalloc(softc->pass_zone, M_WAITOK | M_ZERO);
1865 		ccb = &io_req->ccb;
1866 		user_ccb = (union ccb **)addr;
1867 
1868 		/*
1869 		 * Unlike the CAMIOCOMMAND ioctl above, we only have a
1870 		 * pointer to the user's CCB, so we have to copy the whole
1871 		 * thing in to a buffer we have allocated (above) instead
1872 		 * of allowing the ioctl code to malloc a buffer and copy
1873 		 * it in.
1874 		 *
1875 		 * This is an advantage for this asynchronous interface,
1876 		 * since we don't want the memory to get freed while the
1877 		 * CCB is outstanding.
1878 		 */
1879 #if 0
1880 		xpt_print(periph->path, "Copying user CCB %p to "
1881 			  "kernel address %p\n", *user_ccb, ccb);
1882 #endif
1883 		error = copyin(*user_ccb, ccb, sizeof(*ccb));
1884 		if (error != 0) {
1885 			xpt_print(periph->path, "Copy of user CCB %p to "
1886 				  "kernel address %p failed with error %d\n",
1887 				  *user_ccb, ccb, error);
1888 			goto camioqueue_error;
1889 		}
1890 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
1891 		if (ccb->ccb_h.func_code == XPT_SCSI_IO)
1892 			ccb->csio.bio = NULL;
1893 #endif
1894 
1895 		if (ccb->ccb_h.flags & CAM_UNLOCKED) {
1896 			error = EINVAL;
1897 			goto camioqueue_error;
1898 		}
1899 
1900 		if (ccb->ccb_h.flags & CAM_CDB_POINTER) {
1901 			if (ccb->csio.cdb_len > IOCDBLEN) {
1902 				error = EINVAL;
1903 				goto camioqueue_error;
1904 			}
1905 			error = copyin(ccb->csio.cdb_io.cdb_ptr,
1906 			    ccb->csio.cdb_io.cdb_bytes, ccb->csio.cdb_len);
1907 			if (error != 0)
1908 				goto camioqueue_error;
1909 			ccb->ccb_h.flags &= ~CAM_CDB_POINTER;
1910 		}
1911 
1912 		/*
1913 		 * Some CCB types, like scan bus and scan lun can only go
1914 		 * through the transport layer device.
1915 		 */
1916 		if (ccb->ccb_h.func_code & XPT_FC_XPT_ONLY) {
1917 			xpt_print(periph->path, "CCB function code %#x is "
1918 			    "restricted to the XPT device\n",
1919 			    ccb->ccb_h.func_code);
1920 			error = ENODEV;
1921 			goto camioqueue_error;
1922 		}
1923 
1924 		/*
1925 		 * Save the user's CCB pointer as well as his linked list
1926 		 * pointers and peripheral private area so that we can
1927 		 * restore these later.
1928 		 */
1929 		io_req->user_ccb_ptr = *user_ccb;
1930 		io_req->user_periph_links = ccb->ccb_h.periph_links;
1931 		io_req->user_periph_priv = ccb->ccb_h.periph_priv;
1932 
1933 		/*
1934 		 * Now that we've saved the user's values, we can set our
1935 		 * own peripheral private entry.
1936 		 */
1937 		ccb->ccb_h.ccb_ioreq = io_req;
1938 
1939 		/* Compatibility for RL/priority-unaware code. */
1940 		priority = ccb->ccb_h.pinfo.priority;
1941 		if (priority <= CAM_PRIORITY_OOB)
1942 		    priority += CAM_PRIORITY_OOB + 1;
1943 
1944 		/*
1945 		 * Setup fields in the CCB like the path and the priority.
1946 		 * The path in particular cannot be done in userland, since
1947 		 * it is a pointer to a kernel data structure.
1948 		 */
1949 		xpt_setup_ccb_flags(&ccb->ccb_h, periph->path, priority,
1950 				    ccb->ccb_h.flags);
1951 
1952 		/*
1953 		 * Setup our done routine.  There is no way for the user to
1954 		 * have a valid pointer here.
1955 		 */
1956 		ccb->ccb_h.cbfcnp = passdone;
1957 
1958 		fc = ccb->ccb_h.func_code;
1959 		/*
1960 		 * If this function code has memory that can be mapped in
1961 		 * or out, we need to call passmemsetup().
1962 		 */
1963 		if ((fc == XPT_SCSI_IO) || (fc == XPT_ATA_IO)
1964 		 || (fc == XPT_SMP_IO) || (fc == XPT_DEV_MATCH)
1965 		 || (fc == XPT_DEV_ADVINFO)
1966 		 || (fc == XPT_NVME_ADMIN) || (fc == XPT_NVME_IO)) {
1967 			error = passmemsetup(periph, io_req);
1968 			if (error != 0)
1969 				goto camioqueue_error;
1970 		} else
1971 			io_req->mapinfo.num_bufs_used = 0;
1972 
1973 		cam_periph_lock(periph);
1974 
1975 		/*
1976 		 * Everything goes on the incoming queue initially.
1977 		 */
1978 		TAILQ_INSERT_TAIL(&softc->incoming_queue, io_req, links);
1979 
1980 		/*
1981 		 * If the CCB is queued, and is not a user CCB, then
1982 		 * we need to allocate a slot for it.  Call xpt_schedule()
1983 		 * so that our start routine will get called when a CCB is
1984 		 * available.
1985 		 */
1986 		if ((fc & XPT_FC_QUEUED)
1987 		 && ((fc & XPT_FC_USER_CCB) == 0)) {
1988 			xpt_schedule(periph, priority);
1989 			break;
1990 		}
1991 
1992 		/*
1993 		 * At this point, the CCB in question is either an
1994 		 * immediate CCB (like XPT_DEV_ADVINFO) or it is a user CCB
1995 		 * and therefore should be malloced, not allocated via a slot.
1996 		 * Remove the CCB from the incoming queue and add it to the
1997 		 * active queue.
1998 		 */
1999 		TAILQ_REMOVE(&softc->incoming_queue, io_req, links);
2000 		TAILQ_INSERT_TAIL(&softc->active_queue, io_req, links);
2001 
2002 		xpt_action(ccb);
2003 
2004 		/*
2005 		 * If this is not a queued CCB (i.e. it is an immediate CCB),
2006 		 * then it is already done.  We need to put it on the done
2007 		 * queue for the user to fetch.
2008 		 */
2009 		if ((fc & XPT_FC_QUEUED) == 0) {
2010 			TAILQ_REMOVE(&softc->active_queue, io_req, links);
2011 			TAILQ_INSERT_TAIL(&softc->done_queue, io_req, links);
2012 		}
2013 		break;
2014 
2015 camioqueue_error:
2016 		uma_zfree(softc->pass_zone, io_req);
2017 		cam_periph_lock(periph);
2018 		break;
2019 	}
2020 	case CAMIOGET:
2021 	{
2022 		union ccb **user_ccb;
2023 		struct pass_io_req *io_req;
2024 		int old_error;
2025 
2026 #ifdef COMPAT_FREEBSD32
2027 		if (SV_PROC_FLAG(td->td_proc, SV_ILP32)) {
2028 			error = ENOTTY;
2029 			goto bailout;
2030 		}
2031 #endif
2032 		user_ccb = (union ccb **)addr;
2033 		old_error = 0;
2034 
2035 		io_req = TAILQ_FIRST(&softc->done_queue);
2036 		if (io_req == NULL) {
2037 			error = ENOENT;
2038 			break;
2039 		}
2040 
2041 		/*
2042 		 * Remove the I/O from the done queue.
2043 		 */
2044 		TAILQ_REMOVE(&softc->done_queue, io_req, links);
2045 
2046 		/*
2047 		 * We have to drop the lock during the copyout because the
2048 		 * copyout can result in VM faults that require sleeping.
2049 		 */
2050 		cam_periph_unlock(periph);
2051 
2052 		/*
2053 		 * Do any needed copies (e.g. for reads) and revert the
2054 		 * pointers in the CCB back to the user's pointers.
2055 		 */
2056 		error = passmemdone(periph, io_req);
2057 
2058 		old_error = error;
2059 
2060 		io_req->ccb.ccb_h.periph_links = io_req->user_periph_links;
2061 		io_req->ccb.ccb_h.periph_priv = io_req->user_periph_priv;
2062 
2063 #if 0
2064 		xpt_print(periph->path, "Copying to user CCB %p from "
2065 			  "kernel address %p\n", *user_ccb, &io_req->ccb);
2066 #endif
2067 
2068 		error = copyout(&io_req->ccb, *user_ccb, sizeof(union ccb));
2069 		if (error != 0) {
2070 			xpt_print(periph->path, "Copy to user CCB %p from "
2071 				  "kernel address %p failed with error %d\n",
2072 				  *user_ccb, &io_req->ccb, error);
2073 		}
2074 
2075 		/*
2076 		 * Prefer the first error we got back, and make sure we
2077 		 * don't overwrite bad status with good.
2078 		 */
2079 		if (old_error != 0)
2080 			error = old_error;
2081 
2082 		cam_periph_lock(periph);
2083 
2084 		/*
2085 		 * At this point, if there was an error, we could potentially
2086 		 * re-queue the I/O and try again.  But why?  The error
2087 		 * would almost certainly happen again.  We might as well
2088 		 * not leak memory.
2089 		 */
2090 		uma_zfree(softc->pass_zone, io_req);
2091 		break;
2092 	}
2093 	default:
2094 		error = cam_periph_ioctl(periph, cmd, addr, passerror);
2095 		break;
2096 	}
2097 
2098 bailout:
2099 	cam_periph_unlock(periph);
2100 
2101 	return(error);
2102 }
2103 
2104 static int
passpoll(struct cdev * dev,int poll_events,struct thread * td)2105 passpoll(struct cdev *dev, int poll_events, struct thread *td)
2106 {
2107 	struct cam_periph *periph;
2108 	struct pass_softc *softc;
2109 	int revents;
2110 
2111 	periph = (struct cam_periph *)dev->si_drv1;
2112 	softc = (struct pass_softc *)periph->softc;
2113 
2114 	revents = poll_events & (POLLOUT | POLLWRNORM);
2115 	if ((poll_events & (POLLIN | POLLRDNORM)) != 0) {
2116 		cam_periph_lock(periph);
2117 
2118 		if (!TAILQ_EMPTY(&softc->done_queue)) {
2119 			revents |= poll_events & (POLLIN | POLLRDNORM);
2120 		}
2121 		cam_periph_unlock(periph);
2122 		if (revents == 0)
2123 			selrecord(td, &softc->read_select);
2124 	}
2125 
2126 	return (revents);
2127 }
2128 
2129 static int
passkqfilter(struct cdev * dev,struct knote * kn)2130 passkqfilter(struct cdev *dev, struct knote *kn)
2131 {
2132 	struct cam_periph *periph;
2133 	struct pass_softc *softc;
2134 
2135 	periph = (struct cam_periph *)dev->si_drv1;
2136 	softc = (struct pass_softc *)periph->softc;
2137 
2138 	kn->kn_hook = (caddr_t)periph;
2139 	kn->kn_fop = &passread_filtops;
2140 	knlist_add(&softc->read_select.si_note, kn, 0);
2141 
2142 	return (0);
2143 }
2144 
2145 static void
passreadfiltdetach(struct knote * kn)2146 passreadfiltdetach(struct knote *kn)
2147 {
2148 	struct cam_periph *periph;
2149 	struct pass_softc *softc;
2150 
2151 	periph = (struct cam_periph *)kn->kn_hook;
2152 	softc = (struct pass_softc *)periph->softc;
2153 
2154 	knlist_remove(&softc->read_select.si_note, kn, 0);
2155 }
2156 
2157 static int
passreadfilt(struct knote * kn,long hint)2158 passreadfilt(struct knote *kn, long hint)
2159 {
2160 	struct cam_periph *periph;
2161 	struct pass_softc *softc;
2162 	int retval;
2163 
2164 	periph = (struct cam_periph *)kn->kn_hook;
2165 	softc = (struct pass_softc *)periph->softc;
2166 
2167 	cam_periph_assert(periph, MA_OWNED);
2168 
2169 	if (TAILQ_EMPTY(&softc->done_queue))
2170 		retval = 0;
2171 	else
2172 		retval = 1;
2173 
2174 	return (retval);
2175 }
2176 
2177 /*
2178  * Generally, "ccb" should be the CCB supplied by the kernel.  "inccb"
2179  * should be the CCB that is copied in from the user.
2180  */
2181 static int
passsendccb(struct cam_periph * periph,union ccb * ccb,union ccb * inccb)2182 passsendccb(struct cam_periph *periph, union ccb *ccb, union ccb *inccb)
2183 {
2184 	struct pass_softc *softc;
2185 	struct cam_periph_map_info mapinfo;
2186 	uint8_t *cmd;
2187 	xpt_opcode fc;
2188 	int error;
2189 
2190 	softc = (struct pass_softc *)periph->softc;
2191 
2192 	/*
2193 	 * There are some fields in the CCB header that need to be
2194 	 * preserved, the rest we get from the user.
2195 	 */
2196 	xpt_merge_ccb(ccb, inccb);
2197 
2198 	if (ccb->ccb_h.flags & CAM_CDB_POINTER) {
2199 		cmd = __builtin_alloca(ccb->csio.cdb_len);
2200 		error = copyin(ccb->csio.cdb_io.cdb_ptr, cmd, ccb->csio.cdb_len);
2201 		if (error)
2202 			return (error);
2203 		ccb->csio.cdb_io.cdb_ptr = cmd;
2204 	}
2205 
2206 	/*
2207 	 * Let cam_periph_mapmem do a sanity check on the data pointer format.
2208 	 * Even if no data transfer is needed, it's a cheap check and it
2209 	 * simplifies the code.
2210 	 */
2211 	fc = ccb->ccb_h.func_code;
2212 	if ((fc == XPT_SCSI_IO) || (fc == XPT_ATA_IO) || (fc == XPT_SMP_IO)
2213             || (fc == XPT_DEV_MATCH) || (fc == XPT_DEV_ADVINFO) || (fc == XPT_MMC_IO)
2214             || (fc == XPT_NVME_ADMIN) || (fc == XPT_NVME_IO)) {
2215 		bzero(&mapinfo, sizeof(mapinfo));
2216 
2217 		/*
2218 		 * cam_periph_mapmem calls into proc and vm functions that can
2219 		 * sleep as well as trigger I/O, so we can't hold the lock.
2220 		 * Dropping it here is reasonably safe.
2221 		 */
2222 		cam_periph_unlock(periph);
2223 		error = cam_periph_mapmem(ccb, &mapinfo, softc->maxio);
2224 		cam_periph_lock(periph);
2225 
2226 		/*
2227 		 * cam_periph_mapmem returned an error, we can't continue.
2228 		 * Return the error to the user.
2229 		 */
2230 		if (error)
2231 			return(error);
2232 	} else
2233 		/* Ensure that the unmap call later on is a no-op. */
2234 		mapinfo.num_bufs_used = 0;
2235 
2236 	/*
2237 	 * If the user wants us to perform any error recovery, then honor
2238 	 * that request.  Otherwise, it's up to the user to perform any
2239 	 * error recovery.
2240 	 */
2241 	{
2242 		uint32_t cam_flags, sense_flags;
2243 
2244 		passflags(ccb, &cam_flags, &sense_flags);
2245 		cam_periph_runccb(ccb,  passerror, cam_flags,
2246 		    sense_flags, softc->device_stats);
2247 	}
2248 
2249 	cam_periph_unlock(periph);
2250 	error = cam_periph_unmapmem(ccb, &mapinfo);
2251 	cam_periph_lock(periph);
2252 
2253 	ccb->ccb_h.cbfcnp = NULL;
2254 	ccb->ccb_h.periph_priv = inccb->ccb_h.periph_priv;
2255 	bcopy(ccb, inccb, sizeof(union ccb));
2256 
2257 	return (error);
2258 }
2259 
2260 /*
2261  * Set the cam_flags and sense_flags based on whether or not the request wants
2262  * error recovery. In order to log errors via devctl, we need to do at least
2263  * minimal recovery. We do this by not retrying unit attention (we let the
2264  * requester do it, or not, if appropriate) and specifically asking for no
2265  * recovery, like we do during device probing.
2266  */
2267 static void
passflags(union ccb * ccb,uint32_t * cam_flags,uint32_t * sense_flags)2268 passflags(union ccb *ccb, uint32_t *cam_flags, uint32_t *sense_flags)
2269 {
2270 	if ((ccb->ccb_h.flags & CAM_PASS_ERR_RECOVER) != 0) {
2271 		*cam_flags = CAM_RETRY_SELTO;
2272 		*sense_flags = SF_RETRY_UA | SF_NO_PRINT;
2273 	} else {
2274 		*cam_flags = 0;
2275 		*sense_flags = SF_NO_RETRY | SF_NO_RECOVERY | SF_NO_PRINT;
2276 	}
2277 }
2278 
2279 static int
passerror(union ccb * ccb,uint32_t cam_flags,uint32_t sense_flags)2280 passerror(union ccb *ccb, uint32_t cam_flags, uint32_t sense_flags)
2281 {
2282 
2283 	return(cam_periph_error(ccb, cam_flags, sense_flags));
2284 }
2285