xref: /freebsd/sys/cam/nvme/nvme_da.c (revision 5bf5ca772c6de2d53344a78cf461447cc322ccea)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2015 Netflix, Inc
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer,
12  *    without modification, immediately at the beginning of the file.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27  *
28  * Derived from ata_da.c:
29  * Copyright (c) 2009 Alexander Motin <mav@FreeBSD.org>
30  */
31 
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34 
35 #include <sys/param.h>
36 
37 #ifdef _KERNEL
38 #include <sys/systm.h>
39 #include <sys/kernel.h>
40 #include <sys/bio.h>
41 #include <sys/sysctl.h>
42 #include <sys/taskqueue.h>
43 #include <sys/lock.h>
44 #include <sys/mutex.h>
45 #include <sys/conf.h>
46 #include <sys/devicestat.h>
47 #include <sys/eventhandler.h>
48 #include <sys/malloc.h>
49 #include <sys/cons.h>
50 #include <sys/proc.h>
51 #include <sys/reboot.h>
52 #include <geom/geom_disk.h>
53 #endif /* _KERNEL */
54 
55 #ifndef _KERNEL
56 #include <stdio.h>
57 #include <string.h>
58 #endif /* _KERNEL */
59 
60 #include <cam/cam.h>
61 #include <cam/cam_ccb.h>
62 #include <cam/cam_periph.h>
63 #include <cam/cam_xpt_periph.h>
64 #include <cam/cam_sim.h>
65 #include <cam/cam_iosched.h>
66 
67 #include <cam/nvme/nvme_all.h>
68 
69 typedef enum {
70 	NDA_STATE_NORMAL
71 } nda_state;
72 
73 typedef enum {
74 	NDA_FLAG_OPEN		= 0x0001,
75 	NDA_FLAG_DIRTY		= 0x0002,
76 	NDA_FLAG_SCTX_INIT	= 0x0004,
77 } nda_flags;
78 
79 typedef enum {
80 	NDA_Q_4K   = 0x01,
81 	NDA_Q_NONE = 0x00,
82 } nda_quirks;
83 
84 #define NDA_Q_BIT_STRING	\
85 	"\020"			\
86 	"\001Bit 0"
87 
88 typedef enum {
89 	NDA_CCB_BUFFER_IO	= 0x01,
90 	NDA_CCB_DUMP            = 0x02,
91 	NDA_CCB_TRIM            = 0x03,
92 	NDA_CCB_TYPE_MASK	= 0x0F,
93 } nda_ccb_state;
94 
95 /* Offsets into our private area for storing information */
96 #define ccb_state	ppriv_field0
97 #define ccb_bp		ppriv_ptr1
98 
99 struct trim_request {
100 	TAILQ_HEAD(, bio) bps;
101 };
102 struct nda_softc {
103 	struct   cam_iosched_softc *cam_iosched;
104 	int			outstanding_cmds;	/* Number of active commands */
105 	int			refcount;		/* Active xpt_action() calls */
106 	nda_state		state;
107 	nda_flags		flags;
108 	nda_quirks		quirks;
109 	int			unmappedio;
110 	uint32_t		nsid;			/* Namespace ID for this nda device */
111 	struct disk		*disk;
112 	struct task		sysctl_task;
113 	struct sysctl_ctx_list	sysctl_ctx;
114 	struct sysctl_oid	*sysctl_tree;
115 	struct trim_request	trim_req;
116 #ifdef CAM_IO_STATS
117 	struct sysctl_ctx_list	sysctl_stats_ctx;
118 	struct sysctl_oid	*sysctl_stats_tree;
119 	u_int			timeouts;
120 	u_int			errors;
121 	u_int			invalidations;
122 #endif
123 };
124 
125 /* Need quirk table */
126 
127 static	disk_strategy_t	ndastrategy;
128 static	dumper_t	ndadump;
129 static	periph_init_t	ndainit;
130 static	void		ndaasync(void *callback_arg, u_int32_t code,
131 				struct cam_path *path, void *arg);
132 static	void		ndasysctlinit(void *context, int pending);
133 static	periph_ctor_t	ndaregister;
134 static	periph_dtor_t	ndacleanup;
135 static	periph_start_t	ndastart;
136 static	periph_oninv_t	ndaoninvalidate;
137 static	void		ndadone(struct cam_periph *periph,
138 			       union ccb *done_ccb);
139 static  int		ndaerror(union ccb *ccb, u_int32_t cam_flags,
140 				u_int32_t sense_flags);
141 static void		ndashutdown(void *arg, int howto);
142 static void		ndasuspend(void *arg);
143 
144 #ifndef	NDA_DEFAULT_SEND_ORDERED
145 #define	NDA_DEFAULT_SEND_ORDERED	1
146 #endif
147 #ifndef NDA_DEFAULT_TIMEOUT
148 #define NDA_DEFAULT_TIMEOUT 30	/* Timeout in seconds */
149 #endif
150 #ifndef	NDA_DEFAULT_RETRY
151 #define	NDA_DEFAULT_RETRY	4
152 #endif
153 
154 
155 //static int nda_retry_count = NDA_DEFAULT_RETRY;
156 static int nda_send_ordered = NDA_DEFAULT_SEND_ORDERED;
157 static int nda_default_timeout = NDA_DEFAULT_TIMEOUT;
158 
159 /*
160  * All NVMe media is non-rotational, so all nvme device instances
161  * share this to implement the sysctl.
162  */
163 static int nda_rotating_media = 0;
164 
165 static SYSCTL_NODE(_kern_cam, OID_AUTO, nda, CTLFLAG_RD, 0,
166             "CAM Direct Access Disk driver");
167 
168 static struct periph_driver ndadriver =
169 {
170 	ndainit, "nda",
171 	TAILQ_HEAD_INITIALIZER(ndadriver.units), /* generation */ 0
172 };
173 
174 PERIPHDRIVER_DECLARE(nda, ndadriver);
175 
176 static MALLOC_DEFINE(M_NVMEDA, "nvme_da", "nvme_da buffers");
177 
178 /*
179  * nice wrappers. Maybe these belong in nvme_all.c instead of
180  * here, but this is the only place that uses these. Should
181  * we ever grow another NVME periph, we should move them
182  * all there wholesale.
183  */
184 
185 static void
186 nda_nvme_flush(struct nda_softc *softc, struct ccb_nvmeio *nvmeio)
187 {
188 	cam_fill_nvmeio(nvmeio,
189 	    0,			/* retries */
190 	    ndadone,		/* cbfcnp */
191 	    CAM_DIR_NONE,	/* flags */
192 	    NULL,		/* data_ptr */
193 	    0,			/* dxfer_len */
194 	    nda_default_timeout * 1000); /* timeout 30s */
195 	nvme_ns_flush_cmd(&nvmeio->cmd, softc->nsid);
196 }
197 
198 static void
199 nda_nvme_trim(struct nda_softc *softc, struct ccb_nvmeio *nvmeio,
200     void *payload, uint32_t num_ranges)
201 {
202 	cam_fill_nvmeio(nvmeio,
203 	    0,			/* retries */
204 	    ndadone,		/* cbfcnp */
205 	    CAM_DIR_OUT,	/* flags */
206 	    payload,		/* data_ptr */
207 	    num_ranges * sizeof(struct nvme_dsm_range), /* dxfer_len */
208 	    nda_default_timeout * 1000); /* timeout 30s */
209 	nvme_ns_trim_cmd(&nvmeio->cmd, softc->nsid, num_ranges);
210 }
211 
212 static void
213 nda_nvme_write(struct nda_softc *softc, struct ccb_nvmeio *nvmeio,
214     void *payload, uint64_t lba, uint32_t len, uint32_t count)
215 {
216 	cam_fill_nvmeio(nvmeio,
217 	    0,			/* retries */
218 	    ndadone,		/* cbfcnp */
219 	    CAM_DIR_OUT,	/* flags */
220 	    payload,		/* data_ptr */
221 	    len,		/* dxfer_len */
222 	    nda_default_timeout * 1000); /* timeout 30s */
223 	nvme_ns_write_cmd(&nvmeio->cmd, softc->nsid, lba, count);
224 }
225 
226 static void
227 nda_nvme_rw_bio(struct nda_softc *softc, struct ccb_nvmeio *nvmeio,
228     struct bio *bp, uint32_t rwcmd)
229 {
230 	int flags = rwcmd == NVME_OPC_READ ? CAM_DIR_IN : CAM_DIR_OUT;
231 	void *payload;
232 	uint64_t lba;
233 	uint32_t count;
234 
235 	if (bp->bio_flags & BIO_UNMAPPED) {
236 		flags |= CAM_DATA_BIO;
237 		payload = bp;
238 	} else {
239 		payload = bp->bio_data;
240 	}
241 
242 	lba = bp->bio_pblkno;
243 	count = bp->bio_bcount / softc->disk->d_sectorsize;
244 
245 	cam_fill_nvmeio(nvmeio,
246 	    0,			/* retries */
247 	    ndadone,		/* cbfcnp */
248 	    flags,		/* flags */
249 	    payload,		/* data_ptr */
250 	    bp->bio_bcount,	/* dxfer_len */
251 	    nda_default_timeout * 1000); /* timeout 30s */
252 	nvme_ns_rw_cmd(&nvmeio->cmd, rwcmd, softc->nsid, lba, count);
253 }
254 
255 static int
256 ndaopen(struct disk *dp)
257 {
258 	struct cam_periph *periph;
259 	struct nda_softc *softc;
260 	int error;
261 
262 	periph = (struct cam_periph *)dp->d_drv1;
263 	if (cam_periph_acquire(periph) != 0) {
264 		return(ENXIO);
265 	}
266 
267 	cam_periph_lock(periph);
268 	if ((error = cam_periph_hold(periph, PRIBIO|PCATCH)) != 0) {
269 		cam_periph_unlock(periph);
270 		cam_periph_release(periph);
271 		return (error);
272 	}
273 
274 	CAM_DEBUG(periph->path, CAM_DEBUG_TRACE | CAM_DEBUG_PERIPH,
275 	    ("ndaopen\n"));
276 
277 	softc = (struct nda_softc *)periph->softc;
278 	softc->flags |= NDA_FLAG_OPEN;
279 
280 	cam_periph_unhold(periph);
281 	cam_periph_unlock(periph);
282 	return (0);
283 }
284 
285 static int
286 ndaclose(struct disk *dp)
287 {
288 	struct	cam_periph *periph;
289 	struct	nda_softc *softc;
290 	union ccb *ccb;
291 	int error;
292 
293 	periph = (struct cam_periph *)dp->d_drv1;
294 	softc = (struct nda_softc *)periph->softc;
295 	cam_periph_lock(periph);
296 
297 	CAM_DEBUG(periph->path, CAM_DEBUG_TRACE | CAM_DEBUG_PERIPH,
298 	    ("ndaclose\n"));
299 
300 	if ((softc->flags & NDA_FLAG_DIRTY) != 0 &&
301 	    (periph->flags & CAM_PERIPH_INVALID) == 0 &&
302 	    cam_periph_hold(periph, PRIBIO) == 0) {
303 
304 		ccb = cam_periph_getccb(periph, CAM_PRIORITY_NORMAL);
305 		nda_nvme_flush(softc, &ccb->nvmeio);
306 		error = cam_periph_runccb(ccb, ndaerror, /*cam_flags*/0,
307 		    /*sense_flags*/0, softc->disk->d_devstat);
308 
309 		if (error != 0)
310 			xpt_print(periph->path, "Synchronize cache failed\n");
311 		else
312 			softc->flags &= ~NDA_FLAG_DIRTY;
313 		xpt_release_ccb(ccb);
314 		cam_periph_unhold(periph);
315 	}
316 
317 	softc->flags &= ~NDA_FLAG_OPEN;
318 
319 	while (softc->refcount != 0)
320 		cam_periph_sleep(periph, &softc->refcount, PRIBIO, "ndaclose", 1);
321 	cam_periph_unlock(periph);
322 	cam_periph_release(periph);
323 	return (0);
324 }
325 
326 static void
327 ndaschedule(struct cam_periph *periph)
328 {
329 	struct nda_softc *softc = (struct nda_softc *)periph->softc;
330 
331 	if (softc->state != NDA_STATE_NORMAL)
332 		return;
333 
334 	cam_iosched_schedule(softc->cam_iosched, periph);
335 }
336 
337 /*
338  * Actually translate the requested transfer into one the physical driver
339  * can understand.  The transfer is described by a buf and will include
340  * only one physical transfer.
341  */
342 static void
343 ndastrategy(struct bio *bp)
344 {
345 	struct cam_periph *periph;
346 	struct nda_softc *softc;
347 
348 	periph = (struct cam_periph *)bp->bio_disk->d_drv1;
349 	softc = (struct nda_softc *)periph->softc;
350 
351 	cam_periph_lock(periph);
352 
353 	CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("ndastrategy(%p)\n", bp));
354 
355 	/*
356 	 * If the device has been made invalid, error out
357 	 */
358 	if ((periph->flags & CAM_PERIPH_INVALID) != 0) {
359 		cam_periph_unlock(periph);
360 		biofinish(bp, NULL, ENXIO);
361 		return;
362 	}
363 
364 	/*
365 	 * Place it in the queue of disk activities for this disk
366 	 */
367 	cam_iosched_queue_work(softc->cam_iosched, bp);
368 
369 	/*
370 	 * Schedule ourselves for performing the work.
371 	 */
372 	ndaschedule(periph);
373 	cam_periph_unlock(periph);
374 
375 	return;
376 }
377 
378 static int
379 ndadump(void *arg, void *virtual, vm_offset_t physical, off_t offset, size_t length)
380 {
381 	struct	    cam_periph *periph;
382 	struct	    nda_softc *softc;
383 	u_int	    secsize;
384 	struct ccb_nvmeio nvmeio;
385 	struct	    disk *dp;
386 	uint64_t    lba;
387 	uint32_t    count;
388 	int	    error = 0;
389 
390 	dp = arg;
391 	periph = dp->d_drv1;
392 	softc = (struct nda_softc *)periph->softc;
393 	secsize = softc->disk->d_sectorsize;
394 	lba = offset / secsize;
395 	count = length / secsize;
396 
397 	if ((periph->flags & CAM_PERIPH_INVALID) != 0)
398 		return (ENXIO);
399 
400 	/* xpt_get_ccb returns a zero'd allocation for the ccb, mimic that here */
401 	memset(&nvmeio, 0, sizeof(nvmeio));
402 	if (length > 0) {
403 		xpt_setup_ccb(&nvmeio.ccb_h, periph->path, CAM_PRIORITY_NORMAL);
404 		nvmeio.ccb_h.ccb_state = NDA_CCB_DUMP;
405 		nda_nvme_write(softc, &nvmeio, virtual, lba, length, count);
406 		error = cam_periph_runccb((union ccb *)&nvmeio, cam_periph_error,
407 		    0, SF_NO_RECOVERY | SF_NO_RETRY, NULL);
408 		if (error != 0)
409 			printf("Aborting dump due to I/O error %d.\n", error);
410 
411 		return (error);
412 	}
413 
414 	/* Flush */
415 	xpt_setup_ccb(&nvmeio.ccb_h, periph->path, CAM_PRIORITY_NORMAL);
416 
417 	nvmeio.ccb_h.ccb_state = NDA_CCB_DUMP;
418 	nda_nvme_flush(softc, &nvmeio);
419 	error = cam_periph_runccb((union ccb *)&nvmeio, cam_periph_error,
420 	    0, SF_NO_RECOVERY | SF_NO_RETRY, NULL);
421 	if (error != 0)
422 		xpt_print(periph->path, "flush cmd failed\n");
423 	return (error);
424 }
425 
426 static void
427 ndainit(void)
428 {
429 	cam_status status;
430 
431 	/*
432 	 * Install a global async callback.  This callback will
433 	 * receive async callbacks like "new device found".
434 	 */
435 	status = xpt_register_async(AC_FOUND_DEVICE, ndaasync, NULL, NULL);
436 
437 	if (status != CAM_REQ_CMP) {
438 		printf("nda: Failed to attach master async callback "
439 		       "due to status 0x%x!\n", status);
440 	} else if (nda_send_ordered) {
441 
442 		/* Register our event handlers */
443 		if ((EVENTHANDLER_REGISTER(power_suspend, ndasuspend,
444 					   NULL, EVENTHANDLER_PRI_LAST)) == NULL)
445 		    printf("ndainit: power event registration failed!\n");
446 		if ((EVENTHANDLER_REGISTER(shutdown_post_sync, ndashutdown,
447 					   NULL, SHUTDOWN_PRI_DEFAULT)) == NULL)
448 		    printf("ndainit: shutdown event registration failed!\n");
449 	}
450 }
451 
452 /*
453  * Callback from GEOM, called when it has finished cleaning up its
454  * resources.
455  */
456 static void
457 ndadiskgonecb(struct disk *dp)
458 {
459 	struct cam_periph *periph;
460 
461 	periph = (struct cam_periph *)dp->d_drv1;
462 
463 	cam_periph_release(periph);
464 }
465 
466 static void
467 ndaoninvalidate(struct cam_periph *periph)
468 {
469 	struct nda_softc *softc;
470 
471 	softc = (struct nda_softc *)periph->softc;
472 
473 	/*
474 	 * De-register any async callbacks.
475 	 */
476 	xpt_register_async(0, ndaasync, periph, periph->path);
477 #ifdef CAM_IO_STATS
478 	softc->invalidations++;
479 #endif
480 
481 	/*
482 	 * Return all queued I/O with ENXIO.
483 	 * XXX Handle any transactions queued to the card
484 	 *     with XPT_ABORT_CCB.
485 	 */
486 	cam_iosched_flush(softc->cam_iosched, NULL, ENXIO);
487 
488 	disk_gone(softc->disk);
489 }
490 
491 static void
492 ndacleanup(struct cam_periph *periph)
493 {
494 	struct nda_softc *softc;
495 
496 	softc = (struct nda_softc *)periph->softc;
497 
498 	cam_periph_unlock(periph);
499 
500 	cam_iosched_fini(softc->cam_iosched);
501 
502 	/*
503 	 * If we can't free the sysctl tree, oh well...
504 	 */
505 	if ((softc->flags & NDA_FLAG_SCTX_INIT) != 0) {
506 #ifdef CAM_IO_STATS
507 		if (sysctl_ctx_free(&softc->sysctl_stats_ctx) != 0)
508 			xpt_print(periph->path,
509 			    "can't remove sysctl stats context\n");
510 #endif
511 		if (sysctl_ctx_free(&softc->sysctl_ctx) != 0)
512 			xpt_print(periph->path,
513 			    "can't remove sysctl context\n");
514 	}
515 
516 	disk_destroy(softc->disk);
517 	free(softc, M_DEVBUF);
518 	cam_periph_lock(periph);
519 }
520 
521 static void
522 ndaasync(void *callback_arg, u_int32_t code,
523 	struct cam_path *path, void *arg)
524 {
525 	struct cam_periph *periph;
526 
527 	periph = (struct cam_periph *)callback_arg;
528 	switch (code) {
529 	case AC_FOUND_DEVICE:
530 	{
531 		struct ccb_getdev *cgd;
532 		cam_status status;
533 
534 		cgd = (struct ccb_getdev *)arg;
535 		if (cgd == NULL)
536 			break;
537 
538 		if (cgd->protocol != PROTO_NVME)
539 			break;
540 
541 		/*
542 		 * Allocate a peripheral instance for
543 		 * this device and start the probe
544 		 * process.
545 		 */
546 		status = cam_periph_alloc(ndaregister, ndaoninvalidate,
547 					  ndacleanup, ndastart,
548 					  "nda", CAM_PERIPH_BIO,
549 					  path, ndaasync,
550 					  AC_FOUND_DEVICE, cgd);
551 
552 		if (status != CAM_REQ_CMP
553 		 && status != CAM_REQ_INPROG)
554 			printf("ndaasync: Unable to attach to new device "
555 				"due to status 0x%x\n", status);
556 		break;
557 	}
558 	case AC_ADVINFO_CHANGED:
559 	{
560 		uintptr_t buftype;
561 
562 		buftype = (uintptr_t)arg;
563 		if (buftype == CDAI_TYPE_PHYS_PATH) {
564 			struct nda_softc *softc;
565 
566 			softc = periph->softc;
567 			disk_attr_changed(softc->disk, "GEOM::physpath",
568 					  M_NOWAIT);
569 		}
570 		break;
571 	}
572 	case AC_LOST_DEVICE:
573 	default:
574 		cam_periph_async(periph, code, path, arg);
575 		break;
576 	}
577 }
578 
579 static void
580 ndasysctlinit(void *context, int pending)
581 {
582 	struct cam_periph *periph;
583 	struct nda_softc *softc;
584 	char tmpstr[32], tmpstr2[16];
585 
586 	periph = (struct cam_periph *)context;
587 
588 	/* periph was held for us when this task was enqueued */
589 	if ((periph->flags & CAM_PERIPH_INVALID) != 0) {
590 		cam_periph_release(periph);
591 		return;
592 	}
593 
594 	softc = (struct nda_softc *)periph->softc;
595 	snprintf(tmpstr, sizeof(tmpstr), "CAM NDA unit %d", periph->unit_number);
596 	snprintf(tmpstr2, sizeof(tmpstr2), "%d", periph->unit_number);
597 
598 	sysctl_ctx_init(&softc->sysctl_ctx);
599 	softc->flags |= NDA_FLAG_SCTX_INIT;
600 	softc->sysctl_tree = SYSCTL_ADD_NODE_WITH_LABEL(&softc->sysctl_ctx,
601 		SYSCTL_STATIC_CHILDREN(_kern_cam_nda), OID_AUTO, tmpstr2,
602 		CTLFLAG_RD, 0, tmpstr, "device_index");
603 	if (softc->sysctl_tree == NULL) {
604 		printf("ndasysctlinit: unable to allocate sysctl tree\n");
605 		cam_periph_release(periph);
606 		return;
607 	}
608 
609 	SYSCTL_ADD_INT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
610 		OID_AUTO, "unmapped_io", CTLFLAG_RD | CTLFLAG_MPSAFE,
611 		&softc->unmappedio, 0, "Unmapped I/O leaf");
612 
613 	SYSCTL_ADD_INT(&softc->sysctl_ctx,
614 		       SYSCTL_CHILDREN(softc->sysctl_tree),
615 		       OID_AUTO,
616 		       "rotating",
617 		       CTLFLAG_RD | CTLFLAG_MPSAFE,
618 		       &nda_rotating_media,
619 		       0,
620 		       "Rotating media");
621 
622 #ifdef CAM_IO_STATS
623 	softc->sysctl_stats_tree = SYSCTL_ADD_NODE(&softc->sysctl_stats_ctx,
624 		SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO, "stats",
625 		CTLFLAG_RD, 0, "Statistics");
626 	if (softc->sysctl_stats_tree == NULL) {
627 		printf("ndasysctlinit: unable to allocate sysctl tree for stats\n");
628 		cam_periph_release(periph);
629 		return;
630 	}
631 	SYSCTL_ADD_INT(&softc->sysctl_stats_ctx,
632 		SYSCTL_CHILDREN(softc->sysctl_stats_tree),
633 		OID_AUTO, "timeouts", CTLFLAG_RD | CTLFLAG_MPSAFE,
634 		&softc->timeouts, 0,
635 		"Device timeouts reported by the SIM");
636 	SYSCTL_ADD_INT(&softc->sysctl_stats_ctx,
637 		SYSCTL_CHILDREN(softc->sysctl_stats_tree),
638 		OID_AUTO, "errors", CTLFLAG_RD | CTLFLAG_MPSAFE,
639 		&softc->errors, 0,
640 		"Transport errors reported by the SIM.");
641 	SYSCTL_ADD_INT(&softc->sysctl_stats_ctx,
642 		SYSCTL_CHILDREN(softc->sysctl_stats_tree),
643 		OID_AUTO, "pack_invalidations", CTLFLAG_RD | CTLFLAG_MPSAFE,
644 		&softc->invalidations, 0,
645 		"Device pack invalidations.");
646 #endif
647 
648 	cam_iosched_sysctl_init(softc->cam_iosched, &softc->sysctl_ctx,
649 	    softc->sysctl_tree);
650 
651 	cam_periph_release(periph);
652 }
653 
654 static int
655 ndagetattr(struct bio *bp)
656 {
657 	int ret;
658 	struct cam_periph *periph;
659 
660 	periph = (struct cam_periph *)bp->bio_disk->d_drv1;
661 	cam_periph_lock(periph);
662 	ret = xpt_getattr(bp->bio_data, bp->bio_length, bp->bio_attribute,
663 	    periph->path);
664 	cam_periph_unlock(periph);
665 	if (ret == 0)
666 		bp->bio_completed = bp->bio_length;
667 	return ret;
668 }
669 
670 static cam_status
671 ndaregister(struct cam_periph *periph, void *arg)
672 {
673 	struct nda_softc *softc;
674 	struct disk *disk;
675 	struct ccb_pathinq cpi;
676 	const struct nvme_namespace_data *nsd;
677 	const struct nvme_controller_data *cd;
678 	char   announce_buf[80];
679 	uint8_t flbas_fmt, lbads, vwc_present;
680 	u_int maxio;
681 	int quirks;
682 
683 	nsd = nvme_get_identify_ns(periph);
684 	cd = nvme_get_identify_cntrl(periph);
685 
686 	softc = (struct nda_softc *)malloc(sizeof(*softc), M_DEVBUF,
687 	    M_NOWAIT | M_ZERO);
688 
689 	if (softc == NULL) {
690 		printf("ndaregister: Unable to probe new device. "
691 		    "Unable to allocate softc\n");
692 		return(CAM_REQ_CMP_ERR);
693 	}
694 
695 	if (cam_iosched_init(&softc->cam_iosched, periph) != 0) {
696 		printf("ndaregister: Unable to probe new device. "
697 		       "Unable to allocate iosched memory\n");
698 		return(CAM_REQ_CMP_ERR);
699 	}
700 
701 	/* ident_data parsing */
702 
703 	periph->softc = softc;
704 
705 	softc->quirks = NDA_Q_NONE;
706 
707 	xpt_path_inq(&cpi, periph->path);
708 
709 	TASK_INIT(&softc->sysctl_task, 0, ndasysctlinit, periph);
710 
711 	/*
712 	 * The name space ID is the lun, save it for later I/O
713 	 */
714 	softc->nsid = (uint32_t)xpt_path_lun_id(periph->path);
715 
716 	/*
717 	 * Register this media as a disk
718 	 */
719 	(void)cam_periph_hold(periph, PRIBIO);
720 	cam_periph_unlock(periph);
721 	snprintf(announce_buf, sizeof(announce_buf),
722 	    "kern.cam.nda.%d.quirks", periph->unit_number);
723 	quirks = softc->quirks;
724 	TUNABLE_INT_FETCH(announce_buf, &quirks);
725 	softc->quirks = quirks;
726 	cam_iosched_set_sort_queue(softc->cam_iosched, 0);
727 	softc->disk = disk = disk_alloc();
728 	strlcpy(softc->disk->d_descr, cd->mn,
729 	    MIN(sizeof(softc->disk->d_descr), sizeof(cd->mn)));
730 	strlcpy(softc->disk->d_ident, cd->sn,
731 	    MIN(sizeof(softc->disk->d_ident), sizeof(cd->sn)));
732 	disk->d_rotation_rate = DISK_RR_NON_ROTATING;
733 	disk->d_open = ndaopen;
734 	disk->d_close = ndaclose;
735 	disk->d_strategy = ndastrategy;
736 	disk->d_getattr = ndagetattr;
737 	disk->d_dump = ndadump;
738 	disk->d_gone = ndadiskgonecb;
739 	disk->d_name = "nda";
740 	disk->d_drv1 = periph;
741 	disk->d_unit = periph->unit_number;
742 	maxio = cpi.maxio;		/* Honor max I/O size of SIM */
743 	if (maxio == 0)
744 		maxio = DFLTPHYS;	/* traditional default */
745 	else if (maxio > MAXPHYS)
746 		maxio = MAXPHYS;	/* for safety */
747 	disk->d_maxsize = maxio;
748 	flbas_fmt = (nsd->flbas >> NVME_NS_DATA_FLBAS_FORMAT_SHIFT) &
749 		NVME_NS_DATA_FLBAS_FORMAT_MASK;
750 	lbads = (nsd->lbaf[flbas_fmt] >> NVME_NS_DATA_LBAF_LBADS_SHIFT) &
751 		NVME_NS_DATA_LBAF_LBADS_MASK;
752 	disk->d_sectorsize = 1 << lbads;
753 	disk->d_mediasize = (off_t)(disk->d_sectorsize * nsd->nsze);
754 	disk->d_delmaxsize = disk->d_mediasize;
755 	disk->d_flags = DISKFLAG_DIRECT_COMPLETION;
756 //	if (cd->oncs.dsm) // XXX broken?
757 		disk->d_flags |= DISKFLAG_CANDELETE;
758 	vwc_present = (cd->vwc >> NVME_CTRLR_DATA_VWC_PRESENT_SHIFT) &
759 		NVME_CTRLR_DATA_VWC_PRESENT_MASK;
760 	if (vwc_present)
761 		disk->d_flags |= DISKFLAG_CANFLUSHCACHE;
762 	if ((cpi.hba_misc & PIM_UNMAPPED) != 0) {
763 		disk->d_flags |= DISKFLAG_UNMAPPED_BIO;
764 		softc->unmappedio = 1;
765 	}
766 	/*
767 	 * d_ident and d_descr are both far bigger than the length of either
768 	 *  the serial or model number strings.
769 	 */
770 	nvme_strvis(disk->d_descr, cd->mn,
771 	    sizeof(disk->d_descr), NVME_MODEL_NUMBER_LENGTH);
772 	nvme_strvis(disk->d_ident, cd->sn,
773 	    sizeof(disk->d_ident), NVME_SERIAL_NUMBER_LENGTH);
774 	disk->d_hba_vendor = cpi.hba_vendor;
775 	disk->d_hba_device = cpi.hba_device;
776 	disk->d_hba_subvendor = cpi.hba_subvendor;
777 	disk->d_hba_subdevice = cpi.hba_subdevice;
778 	disk->d_stripesize = disk->d_sectorsize;
779 	disk->d_stripeoffset = 0;
780 	disk->d_devstat = devstat_new_entry(periph->periph_name,
781 	    periph->unit_number, disk->d_sectorsize,
782 	    DEVSTAT_ALL_SUPPORTED,
783 	    DEVSTAT_TYPE_DIRECT | XPORT_DEVSTAT_TYPE(cpi.transport),
784 	    DEVSTAT_PRIORITY_DISK);
785 	/*
786 	 * Add alias for older nvd drives to ease transition.
787 	 */
788 	/* disk_add_alias(disk, "nvd"); Have reports of this causing problems */
789 
790 	/*
791 	 * Acquire a reference to the periph before we register with GEOM.
792 	 * We'll release this reference once GEOM calls us back (via
793 	 * ndadiskgonecb()) telling us that our provider has been freed.
794 	 */
795 	if (cam_periph_acquire(periph) != 0) {
796 		xpt_print(periph->path, "%s: lost periph during "
797 			  "registration!\n", __func__);
798 		cam_periph_lock(periph);
799 		return (CAM_REQ_CMP_ERR);
800 	}
801 	disk_create(softc->disk, DISK_VERSION);
802 	cam_periph_lock(periph);
803 	cam_periph_unhold(periph);
804 
805 	snprintf(announce_buf, sizeof(announce_buf),
806 		"%juMB (%ju %u byte sectors)",
807 	    (uintmax_t)((uintmax_t)disk->d_mediasize / (1024*1024)),
808 		(uintmax_t)disk->d_mediasize / disk->d_sectorsize,
809 		disk->d_sectorsize);
810 	xpt_announce_periph(periph, announce_buf);
811 	xpt_announce_quirks(periph, softc->quirks, NDA_Q_BIT_STRING);
812 
813 	/*
814 	 * Create our sysctl variables, now that we know
815 	 * we have successfully attached.
816 	 */
817 	if (cam_periph_acquire(periph) == 0)
818 		taskqueue_enqueue(taskqueue_thread, &softc->sysctl_task);
819 
820 	/*
821 	 * Register for device going away and info about the drive
822 	 * changing (though with NVMe, it can't)
823 	 */
824 	xpt_register_async(AC_LOST_DEVICE | AC_ADVINFO_CHANGED,
825 	    ndaasync, periph, periph->path);
826 
827 	softc->state = NDA_STATE_NORMAL;
828 	return(CAM_REQ_CMP);
829 }
830 
831 static void
832 ndastart(struct cam_periph *periph, union ccb *start_ccb)
833 {
834 	struct nda_softc *softc = (struct nda_softc *)periph->softc;
835 	struct ccb_nvmeio *nvmeio = &start_ccb->nvmeio;
836 
837 	CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("ndastart\n"));
838 
839 	switch (softc->state) {
840 	case NDA_STATE_NORMAL:
841 	{
842 		struct bio *bp;
843 
844 		bp = cam_iosched_next_bio(softc->cam_iosched);
845 		CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("ndastart: bio %p\n", bp));
846 		if (bp == NULL) {
847 			xpt_release_ccb(start_ccb);
848 			break;
849 		}
850 
851 		switch (bp->bio_cmd) {
852 		case BIO_WRITE:
853 			softc->flags |= NDA_FLAG_DIRTY;
854 			/* FALLTHROUGH */
855 		case BIO_READ:
856 		{
857 #ifdef NDA_TEST_FAILURE
858 			int fail = 0;
859 
860 			/*
861 			 * Support the failure ioctls.  If the command is a
862 			 * read, and there are pending forced read errors, or
863 			 * if a write and pending write errors, then fail this
864 			 * operation with EIO.  This is useful for testing
865 			 * purposes.  Also, support having every Nth read fail.
866 			 *
867 			 * This is a rather blunt tool.
868 			 */
869 			if (bp->bio_cmd == BIO_READ) {
870 				if (softc->force_read_error) {
871 					softc->force_read_error--;
872 					fail = 1;
873 				}
874 				if (softc->periodic_read_error > 0) {
875 					if (++softc->periodic_read_count >=
876 					    softc->periodic_read_error) {
877 						softc->periodic_read_count = 0;
878 						fail = 1;
879 					}
880 				}
881 			} else {
882 				if (softc->force_write_error) {
883 					softc->force_write_error--;
884 					fail = 1;
885 				}
886 			}
887 			if (fail) {
888 				biofinish(bp, NULL, EIO);
889 				xpt_release_ccb(start_ccb);
890 				ndaschedule(periph);
891 				return;
892 			}
893 #endif
894 			KASSERT((bp->bio_flags & BIO_UNMAPPED) == 0 ||
895 			    round_page(bp->bio_bcount + bp->bio_ma_offset) /
896 			    PAGE_SIZE == bp->bio_ma_n,
897 			    ("Short bio %p", bp));
898 			nda_nvme_rw_bio(softc, &start_ccb->nvmeio, bp, bp->bio_cmd == BIO_READ ?
899 			    NVME_OPC_READ : NVME_OPC_WRITE);
900 			break;
901 		}
902 		case BIO_DELETE:
903 		{
904 			struct nvme_dsm_range *dsm_range;
905 
906 			dsm_range =
907 			    malloc(sizeof(*dsm_range), M_NVMEDA, M_ZERO | M_NOWAIT);
908 			if (dsm_range == NULL) {
909 				biofinish(bp, NULL, ENOMEM);
910 				xpt_release_ccb(start_ccb);
911 				ndaschedule(periph);
912 				return;
913 			}
914 			dsm_range->length =
915 			    htole32(bp->bio_bcount / softc->disk->d_sectorsize);
916 			dsm_range->starting_lba =
917 			    htole64(bp->bio_offset / softc->disk->d_sectorsize);
918 			bp->bio_driver2 = dsm_range;
919 			nda_nvme_trim(softc, &start_ccb->nvmeio, dsm_range, 1);
920 			start_ccb->ccb_h.ccb_state = NDA_CCB_TRIM;
921 			start_ccb->ccb_h.flags |= CAM_UNLOCKED;
922 			/*
923 			 * Note: We can have multiple TRIMs in flight, so we don't call
924 			 * cam_iosched_submit_trim(softc->cam_iosched);
925 			 * since that forces the I/O scheduler to only schedule one at a time.
926 			 * On NVMe drives, this is a performance disaster.
927 			 */
928 			goto out;
929 		}
930 		case BIO_FLUSH:
931 			nda_nvme_flush(softc, nvmeio);
932 			break;
933 		}
934 		start_ccb->ccb_h.ccb_state = NDA_CCB_BUFFER_IO;
935 		start_ccb->ccb_h.flags |= CAM_UNLOCKED;
936 out:
937 		start_ccb->ccb_h.ccb_bp = bp;
938 		softc->outstanding_cmds++;
939 		softc->refcount++;
940 		cam_periph_unlock(periph);
941 		xpt_action(start_ccb);
942 		cam_periph_lock(periph);
943 		softc->refcount--;
944 
945 		/* May have more work to do, so ensure we stay scheduled */
946 		ndaschedule(periph);
947 		break;
948 		}
949 	}
950 }
951 
952 static void
953 ndadone(struct cam_periph *periph, union ccb *done_ccb)
954 {
955 	struct nda_softc *softc;
956 	struct ccb_nvmeio *nvmeio = &done_ccb->nvmeio;
957 	struct cam_path *path;
958 	int state;
959 
960 	softc = (struct nda_softc *)periph->softc;
961 	path = done_ccb->ccb_h.path;
962 
963 	CAM_DEBUG(path, CAM_DEBUG_TRACE, ("ndadone\n"));
964 
965 	state = nvmeio->ccb_h.ccb_state & NDA_CCB_TYPE_MASK;
966 	switch (state) {
967 	case NDA_CCB_BUFFER_IO:
968 	case NDA_CCB_TRIM:
969 	{
970 		struct bio *bp;
971 		int error;
972 
973 		cam_periph_lock(periph);
974 		bp = (struct bio *)done_ccb->ccb_h.ccb_bp;
975 		if ((done_ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_CMP) {
976 			error = ndaerror(done_ccb, 0, 0);
977 			if (error == ERESTART) {
978 				/* A retry was scheduled, so just return. */
979 				cam_periph_unlock(periph);
980 				return;
981 			}
982 			if ((done_ccb->ccb_h.status & CAM_DEV_QFRZN) != 0)
983 				cam_release_devq(path,
984 						 /*relsim_flags*/0,
985 						 /*reduction*/0,
986 						 /*timeout*/0,
987 						 /*getcount_only*/0);
988 		} else {
989 			if ((done_ccb->ccb_h.status & CAM_DEV_QFRZN) != 0)
990 				panic("REQ_CMP with QFRZN");
991 			error = 0;
992 		}
993 		bp->bio_error = error;
994 		if (error != 0) {
995 			bp->bio_resid = bp->bio_bcount;
996 			bp->bio_flags |= BIO_ERROR;
997 		} else {
998 			bp->bio_resid = 0;
999 		}
1000 		if (state == NDA_CCB_TRIM)
1001 			free(bp->bio_driver2, M_NVMEDA);
1002 		softc->outstanding_cmds--;
1003 
1004 		/*
1005 		 * We need to call cam_iosched before we call biodone so that we
1006 		 * don't measure any activity that happens in the completion
1007 		 * routine, which in the case of sendfile can be quite
1008 		 * extensive.
1009 		 */
1010 		cam_iosched_bio_complete(softc->cam_iosched, bp, done_ccb);
1011 		xpt_release_ccb(done_ccb);
1012 		if (state == NDA_CCB_TRIM) {
1013 #ifdef notyet
1014 			TAILQ_HEAD(, bio) queue;
1015 			struct bio *bp1;
1016 
1017 			TAILQ_INIT(&queue);
1018 			TAILQ_CONCAT(&queue, &softc->trim_req.bps, bio_queue);
1019 #endif
1020 			/*
1021 			 * Since we can have multiple trims in flight, we don't
1022 			 * need to call this here.
1023 			 * cam_iosched_trim_done(softc->cam_iosched);
1024 			 */
1025 			ndaschedule(periph);
1026 			cam_periph_unlock(periph);
1027 #ifdef notyet
1028 /* Not yet collapsing several BIO_DELETE requests into one TRIM */
1029 			while ((bp1 = TAILQ_FIRST(&queue)) != NULL) {
1030 				TAILQ_REMOVE(&queue, bp1, bio_queue);
1031 				bp1->bio_error = error;
1032 				if (error != 0) {
1033 					bp1->bio_flags |= BIO_ERROR;
1034 					bp1->bio_resid = bp1->bio_bcount;
1035 				} else
1036 					bp1->bio_resid = 0;
1037 				biodone(bp1);
1038 			}
1039 #else
1040 			biodone(bp);
1041 #endif
1042 		} else {
1043 			ndaschedule(periph);
1044 			cam_periph_unlock(periph);
1045 			biodone(bp);
1046 		}
1047 		return;
1048 	}
1049 	case NDA_CCB_DUMP:
1050 		/* No-op.  We're polling */
1051 		return;
1052 	default:
1053 		break;
1054 	}
1055 	xpt_release_ccb(done_ccb);
1056 }
1057 
1058 static int
1059 ndaerror(union ccb *ccb, u_int32_t cam_flags, u_int32_t sense_flags)
1060 {
1061 	struct nda_softc *softc;
1062 	struct cam_periph *periph;
1063 
1064 	periph = xpt_path_periph(ccb->ccb_h.path);
1065 	softc = (struct nda_softc *)periph->softc;
1066 
1067 	switch (ccb->ccb_h.status & CAM_STATUS_MASK) {
1068 	case CAM_CMD_TIMEOUT:
1069 #ifdef CAM_IO_STATS
1070 		softc->timeouts++;
1071 #endif
1072 		break;
1073 	case CAM_REQ_ABORTED:
1074 	case CAM_REQ_CMP_ERR:
1075 	case CAM_REQ_TERMIO:
1076 	case CAM_UNREC_HBA_ERROR:
1077 	case CAM_DATA_RUN_ERR:
1078 	case CAM_ATA_STATUS_ERROR:
1079 #ifdef CAM_IO_STATS
1080 		softc->errors++;
1081 #endif
1082 		break;
1083 	default:
1084 		break;
1085 	}
1086 
1087 	return(cam_periph_error(ccb, cam_flags, sense_flags));
1088 }
1089 
1090 /*
1091  * Step through all NDA peripheral drivers, and if the device is still open,
1092  * sync the disk cache to physical media.
1093  */
1094 static void
1095 ndaflush(void)
1096 {
1097 	struct cam_periph *periph;
1098 	struct nda_softc *softc;
1099 	union ccb *ccb;
1100 	int error;
1101 
1102 	CAM_PERIPH_FOREACH(periph, &ndadriver) {
1103 		softc = (struct nda_softc *)periph->softc;
1104 
1105 		if (SCHEDULER_STOPPED()) {
1106 			/*
1107 			 * If we paniced with the lock held or the periph is not
1108 			 * open, do not recurse.  Otherwise, call ndadump since
1109 			 * that avoids the sleeping cam_periph_getccb does if no
1110 			 * CCBs are available.
1111 			 */
1112 			if (!cam_periph_owned(periph) &&
1113 			    (softc->flags & NDA_FLAG_OPEN)) {
1114 				ndadump(softc->disk, NULL, 0, 0, 0);
1115 			}
1116 			continue;
1117 		}
1118 
1119 		/*
1120 		 * We only sync the cache if the drive is still open
1121 		 */
1122 		cam_periph_lock(periph);
1123 		if ((softc->flags & NDA_FLAG_OPEN) == 0) {
1124 			cam_periph_unlock(periph);
1125 			continue;
1126 		}
1127 
1128 		ccb = cam_periph_getccb(periph, CAM_PRIORITY_NORMAL);
1129 		nda_nvme_flush(softc, &ccb->nvmeio);
1130 		error = cam_periph_runccb(ccb, ndaerror, /*cam_flags*/0,
1131 		    /*sense_flags*/ SF_NO_RECOVERY | SF_NO_RETRY,
1132 		    softc->disk->d_devstat);
1133 		if (error != 0)
1134 			xpt_print(periph->path, "Synchronize cache failed\n");
1135 		xpt_release_ccb(ccb);
1136 		cam_periph_unlock(periph);
1137 	}
1138 }
1139 
1140 static void
1141 ndashutdown(void *arg, int howto)
1142 {
1143 
1144 	ndaflush();
1145 }
1146 
1147 static void
1148 ndasuspend(void *arg)
1149 {
1150 
1151 	ndaflush();
1152 }
1153