1/*-
2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3 *
4 * Copyright (c) 2015 Netflix, Inc.
5 *
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
8 * are met:
9 * 1. Redistributions of source code must retain the above copyright
10 *    notice, this list of conditions and the following disclaimer,
11 *    without modification, immediately at the beginning of the file.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 *    notice, this list of conditions and the following disclaimer in the
14 *    documentation and/or other materials provided with the distribution.
15 *
16 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
17 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
18 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
19 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
20 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
21 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
22 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
23 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
24 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
25 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
26 *
27 * Derived from ata_da.c:
28 * Copyright (c) 2009 Alexander Motin <mav@FreeBSD.org>
29 */
30
31#include <sys/cdefs.h>
32__FBSDID("$FreeBSD$");
33
34#include <sys/param.h>
35
36#ifdef _KERNEL
37#include <sys/systm.h>
38#include <sys/kernel.h>
39#include <sys/bio.h>
40#include <sys/sysctl.h>
41#include <sys/taskqueue.h>
42#include <sys/lock.h>
43#include <sys/mutex.h>
44#include <sys/conf.h>
45#include <sys/devicestat.h>
46#include <sys/eventhandler.h>
47#include <sys/malloc.h>
48#include <sys/cons.h>
49#include <sys/proc.h>
50#include <sys/reboot.h>
51#include <geom/geom_disk.h>
52#endif /* _KERNEL */
53
54#ifndef _KERNEL
55#include <stdio.h>
56#include <string.h>
57#endif /* _KERNEL */
58
59#include <cam/cam.h>
60#include <cam/cam_ccb.h>
61#include <cam/cam_periph.h>
62#include <cam/cam_xpt_periph.h>
63#include <cam/cam_sim.h>
64#include <cam/cam_iosched.h>
65
66#include <cam/nvme/nvme_all.h>
67
68typedef enum {
69	NDA_STATE_NORMAL
70} nda_state;
71
72typedef enum {
73	NDA_FLAG_OPEN		= 0x0001,
74	NDA_FLAG_DIRTY		= 0x0002,
75	NDA_FLAG_SCTX_INIT	= 0x0004,
76} nda_flags;
77
78typedef enum {
79	NDA_Q_4K   = 0x01,
80	NDA_Q_NONE = 0x00,
81} nda_quirks;
82
83#define NDA_Q_BIT_STRING	\
84	"\020"			\
85	"\001Bit 0"
86
87typedef enum {
88	NDA_CCB_BUFFER_IO	= 0x01,
89	NDA_CCB_DUMP            = 0x02,
90	NDA_CCB_TRIM            = 0x03,
91	NDA_CCB_PASS            = 0x04,
92	NDA_CCB_TYPE_MASK	= 0x0F,
93} nda_ccb_state;
94
95/* Offsets into our private area for storing information */
96#define ccb_state	ccb_h.ppriv_field0
97#define ccb_bp		ccb_h.ppriv_ptr1	/* For NDA_CCB_BUFFER_IO */
98#define ccb_trim	ccb_h.ppriv_ptr1	/* For NDA_CCB_TRIM */
99
100struct nda_softc {
101	struct   cam_iosched_softc *cam_iosched;
102	int			outstanding_cmds;	/* Number of active commands */
103	int			refcount;		/* Active xpt_action() calls */
104	nda_state		state;
105	nda_flags		flags;
106	nda_quirks		quirks;
107	int			unmappedio;
108	quad_t			deletes;
109	quad_t			dsm_req;
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#ifdef CAM_TEST_FAILURE
116	int			force_read_error;
117	int			force_write_error;
118	int			periodic_read_error;
119	int			periodic_read_count;
120#endif
121#ifdef CAM_IO_STATS
122	struct sysctl_ctx_list	sysctl_stats_ctx;
123	struct sysctl_oid	*sysctl_stats_tree;
124	u_int			timeouts;
125	u_int			errors;
126	u_int			invalidations;
127#endif
128};
129
130struct nda_trim_request {
131	union {
132		struct nvme_dsm_range dsm;
133		uint8_t		data[NVME_MAX_DSM_TRIM];
134	};
135	TAILQ_HEAD(, bio) bps;
136};
137
138/* Need quirk table */
139
140static	disk_ioctl_t	ndaioctl;
141static	disk_strategy_t	ndastrategy;
142static	dumper_t	ndadump;
143static	periph_init_t	ndainit;
144static	void		ndaasync(void *callback_arg, u_int32_t code,
145				struct cam_path *path, void *arg);
146static	void		ndasysctlinit(void *context, int pending);
147static	periph_ctor_t	ndaregister;
148static	periph_dtor_t	ndacleanup;
149static	periph_start_t	ndastart;
150static	periph_oninv_t	ndaoninvalidate;
151static	void		ndadone(struct cam_periph *periph,
152			       union ccb *done_ccb);
153static  int		ndaerror(union ccb *ccb, u_int32_t cam_flags,
154				u_int32_t sense_flags);
155static void		ndashutdown(void *arg, int howto);
156static void		ndasuspend(void *arg);
157
158#ifndef	NDA_DEFAULT_SEND_ORDERED
159#define	NDA_DEFAULT_SEND_ORDERED	1
160#endif
161#ifndef NDA_DEFAULT_TIMEOUT
162#define NDA_DEFAULT_TIMEOUT 30	/* Timeout in seconds */
163#endif
164#ifndef	NDA_DEFAULT_RETRY
165#define	NDA_DEFAULT_RETRY	4
166#endif
167#ifndef NDA_MAX_TRIM_ENTRIES
168#define NDA_MAX_TRIM_ENTRIES  (NVME_MAX_DSM_TRIM / sizeof(struct nvme_dsm_range))/* Number of DSM trims to use, max 256 */
169#endif
170
171static SYSCTL_NODE(_kern_cam, OID_AUTO, nda, CTLFLAG_RD, 0,
172            "CAM Direct Access Disk driver");
173
174//static int nda_retry_count = NDA_DEFAULT_RETRY;
175static int nda_send_ordered = NDA_DEFAULT_SEND_ORDERED;
176static int nda_default_timeout = NDA_DEFAULT_TIMEOUT;
177static int nda_max_trim_entries = NDA_MAX_TRIM_ENTRIES;
178SYSCTL_INT(_kern_cam_nda, OID_AUTO, max_trim, CTLFLAG_RDTUN,
179    &nda_max_trim_entries, NDA_MAX_TRIM_ENTRIES,
180    "Maximum number of BIO_DELETE to send down as a DSM TRIM.");
181
182/*
183 * All NVMe media is non-rotational, so all nvme device instances
184 * share this to implement the sysctl.
185 */
186static int nda_rotating_media = 0;
187
188static struct periph_driver ndadriver =
189{
190	ndainit, "nda",
191	TAILQ_HEAD_INITIALIZER(ndadriver.units), /* generation */ 0
192};
193
194PERIPHDRIVER_DECLARE(nda, ndadriver);
195
196static MALLOC_DEFINE(M_NVMEDA, "nvme_da", "nvme_da buffers");
197
198/*
199 * nice wrappers. Maybe these belong in nvme_all.c instead of
200 * here, but this is the only place that uses these. Should
201 * we ever grow another NVME periph, we should move them
202 * all there wholesale.
203 */
204
205static void
206nda_nvme_flush(struct nda_softc *softc, struct ccb_nvmeio *nvmeio)
207{
208	cam_fill_nvmeio(nvmeio,
209	    0,			/* retries */
210	    ndadone,		/* cbfcnp */
211	    CAM_DIR_NONE,	/* flags */
212	    NULL,		/* data_ptr */
213	    0,			/* dxfer_len */
214	    nda_default_timeout * 1000); /* timeout 30s */
215	nvme_ns_flush_cmd(&nvmeio->cmd, softc->nsid);
216}
217
218static void
219nda_nvme_trim(struct nda_softc *softc, struct ccb_nvmeio *nvmeio,
220    void *payload, uint32_t num_ranges)
221{
222	cam_fill_nvmeio(nvmeio,
223	    0,			/* retries */
224	    ndadone,		/* cbfcnp */
225	    CAM_DIR_OUT,	/* flags */
226	    payload,		/* data_ptr */
227	    num_ranges * sizeof(struct nvme_dsm_range), /* dxfer_len */
228	    nda_default_timeout * 1000); /* timeout 30s */
229	nvme_ns_trim_cmd(&nvmeio->cmd, softc->nsid, num_ranges);
230}
231
232static void
233nda_nvme_write(struct nda_softc *softc, struct ccb_nvmeio *nvmeio,
234    void *payload, uint64_t lba, uint32_t len, uint32_t count)
235{
236	cam_fill_nvmeio(nvmeio,
237	    0,			/* retries */
238	    ndadone,		/* cbfcnp */
239	    CAM_DIR_OUT,	/* flags */
240	    payload,		/* data_ptr */
241	    len,		/* dxfer_len */
242	    nda_default_timeout * 1000); /* timeout 30s */
243	nvme_ns_write_cmd(&nvmeio->cmd, softc->nsid, lba, count);
244}
245
246static void
247nda_nvme_rw_bio(struct nda_softc *softc, struct ccb_nvmeio *nvmeio,
248    struct bio *bp, uint32_t rwcmd)
249{
250	int flags = rwcmd == NVME_OPC_READ ? CAM_DIR_IN : CAM_DIR_OUT;
251	void *payload;
252	uint64_t lba;
253	uint32_t count;
254
255	if (bp->bio_flags & BIO_UNMAPPED) {
256		flags |= CAM_DATA_BIO;
257		payload = bp;
258	} else {
259		payload = bp->bio_data;
260	}
261
262	lba = bp->bio_pblkno;
263	count = bp->bio_bcount / softc->disk->d_sectorsize;
264
265	cam_fill_nvmeio(nvmeio,
266	    0,			/* retries */
267	    ndadone,		/* cbfcnp */
268	    flags,		/* flags */
269	    payload,		/* data_ptr */
270	    bp->bio_bcount,	/* dxfer_len */
271	    nda_default_timeout * 1000); /* timeout 30s */
272	nvme_ns_rw_cmd(&nvmeio->cmd, rwcmd, softc->nsid, lba, count);
273}
274
275static int
276ndaopen(struct disk *dp)
277{
278	struct cam_periph *periph;
279	struct nda_softc *softc;
280	int error;
281
282	periph = (struct cam_periph *)dp->d_drv1;
283	if (cam_periph_acquire(periph) != 0) {
284		return(ENXIO);
285	}
286
287	cam_periph_lock(periph);
288	if ((error = cam_periph_hold(periph, PRIBIO|PCATCH)) != 0) {
289		cam_periph_unlock(periph);
290		cam_periph_release(periph);
291		return (error);
292	}
293
294	CAM_DEBUG(periph->path, CAM_DEBUG_TRACE | CAM_DEBUG_PERIPH,
295	    ("ndaopen\n"));
296
297	softc = (struct nda_softc *)periph->softc;
298	softc->flags |= NDA_FLAG_OPEN;
299
300	cam_periph_unhold(periph);
301	cam_periph_unlock(periph);
302	return (0);
303}
304
305static int
306ndaclose(struct disk *dp)
307{
308	struct	cam_periph *periph;
309	struct	nda_softc *softc;
310	union ccb *ccb;
311	int error;
312
313	periph = (struct cam_periph *)dp->d_drv1;
314	softc = (struct nda_softc *)periph->softc;
315	cam_periph_lock(periph);
316
317	CAM_DEBUG(periph->path, CAM_DEBUG_TRACE | CAM_DEBUG_PERIPH,
318	    ("ndaclose\n"));
319
320	if ((softc->flags & NDA_FLAG_DIRTY) != 0 &&
321	    (periph->flags & CAM_PERIPH_INVALID) == 0 &&
322	    cam_periph_hold(periph, PRIBIO) == 0) {
323
324		ccb = cam_periph_getccb(periph, CAM_PRIORITY_NORMAL);
325		nda_nvme_flush(softc, &ccb->nvmeio);
326		error = cam_periph_runccb(ccb, ndaerror, /*cam_flags*/0,
327		    /*sense_flags*/0, softc->disk->d_devstat);
328
329		if (error != 0)
330			xpt_print(periph->path, "Synchronize cache failed\n");
331		else
332			softc->flags &= ~NDA_FLAG_DIRTY;
333		xpt_release_ccb(ccb);
334		cam_periph_unhold(periph);
335	}
336
337	softc->flags &= ~NDA_FLAG_OPEN;
338
339	while (softc->refcount != 0)
340		cam_periph_sleep(periph, &softc->refcount, PRIBIO, "ndaclose", 1);
341	KASSERT(softc->outstanding_cmds == 0,
342	    ("nda %d outstanding commands", softc->outstanding_cmds));
343	cam_periph_unlock(periph);
344	cam_periph_release(periph);
345	return (0);
346}
347
348static void
349ndaschedule(struct cam_periph *periph)
350{
351	struct nda_softc *softc = (struct nda_softc *)periph->softc;
352
353	if (softc->state != NDA_STATE_NORMAL)
354		return;
355
356	cam_iosched_schedule(softc->cam_iosched, periph);
357}
358
359static int
360ndaioctl(struct disk *dp, u_long cmd, void *data, int fflag,
361    struct thread *td)
362{
363	struct cam_periph *periph;
364	struct nda_softc *softc;
365
366	periph = (struct cam_periph *)dp->d_drv1;
367	softc = (struct nda_softc *)periph->softc;
368
369	switch (cmd) {
370	case NVME_IO_TEST:
371	case NVME_BIO_TEST:
372		/*
373		 * These don't map well to the underlying CCBs, so
374		 * they are usupported via CAM.
375		 */
376		return (ENOTTY);
377	case NVME_GET_NSID:
378	{
379		struct nvme_get_nsid *gnsid = (struct nvme_get_nsid *)data;
380		struct ccb_pathinq cpi;
381
382		xpt_path_inq(&cpi, periph->path);
383		strncpy(gnsid->cdev, cpi.xport_specific.nvme.dev_name,
384		    sizeof(gnsid->cdev));
385		gnsid->nsid = cpi.xport_specific.nvme.nsid;
386		return (0);
387	}
388	case NVME_PASSTHROUGH_CMD:
389	{
390		struct nvme_pt_command *pt;
391		union ccb *ccb;
392		struct cam_periph_map_info mapinfo;
393		u_int maxmap = MAXPHYS;	/* XXX is this right */
394		int error;
395
396		/*
397		 * Create a NVME_IO CCB to do the passthrough command.
398		 */
399		pt = (struct nvme_pt_command *)data;
400		ccb = xpt_alloc_ccb();
401		xpt_setup_ccb(&ccb->ccb_h, periph->path, CAM_PRIORITY_NORMAL);
402		ccb->ccb_state = NDA_CCB_PASS;
403		cam_fill_nvmeio(&ccb->nvmeio,
404		    0,			/* Retries */
405		    ndadone,
406		    (pt->is_read ? CAM_DIR_IN : CAM_DIR_OUT) | CAM_DATA_VADDR,
407		    pt->buf,
408		    pt->len,
409		    nda_default_timeout * 1000);
410		memcpy(&ccb->nvmeio.cmd, &pt->cmd, sizeof(pt->cmd));
411
412		/*
413		 * Wire the user memory in this request for the I/O
414		 */
415		memset(&mapinfo, 0, sizeof(mapinfo));
416		error = cam_periph_mapmem(ccb, &mapinfo, maxmap);
417		if (error)
418			return (error);
419
420		/*
421		 * Lock the periph and run the command. XXX do we need
422		 * to lock the periph?
423		 */
424		cam_periph_lock(periph);
425		cam_periph_runccb(ccb, NULL, CAM_RETRY_SELTO, SF_RETRY_UA | SF_NO_PRINT,
426		    NULL);
427		cam_periph_unlock(periph);
428
429		/*
430		 * Tear down mapping and return status.
431		 */
432		cam_periph_unmapmem(ccb, &mapinfo);
433		cam_periph_lock(periph);
434		error = (ccb->ccb_h.status == CAM_REQ_CMP) ? 0 : EIO;
435		xpt_release_ccb(ccb);
436		return (error);
437	}
438	default:
439		break;
440	}
441	return (ENOTTY);
442}
443
444/*
445 * Actually translate the requested transfer into one the physical driver
446 * can understand.  The transfer is described by a buf and will include
447 * only one physical transfer.
448 */
449static void
450ndastrategy(struct bio *bp)
451{
452	struct cam_periph *periph;
453	struct nda_softc *softc;
454
455	periph = (struct cam_periph *)bp->bio_disk->d_drv1;
456	softc = (struct nda_softc *)periph->softc;
457
458	cam_periph_lock(periph);
459
460	CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("ndastrategy(%p)\n", bp));
461
462	/*
463	 * If the device has been made invalid, error out
464	 */
465	if ((periph->flags & CAM_PERIPH_INVALID) != 0) {
466		cam_periph_unlock(periph);
467		biofinish(bp, NULL, ENXIO);
468		return;
469	}
470
471	if (bp->bio_cmd == BIO_DELETE)
472		softc->deletes++;
473
474	/*
475	 * Place it in the queue of disk activities for this disk
476	 */
477	cam_iosched_queue_work(softc->cam_iosched, bp);
478
479	/*
480	 * Schedule ourselves for performing the work.
481	 */
482	ndaschedule(periph);
483	cam_periph_unlock(periph);
484
485	return;
486}
487
488static int
489ndadump(void *arg, void *virtual, vm_offset_t physical, off_t offset, size_t length)
490{
491	struct	    cam_periph *periph;
492	struct	    nda_softc *softc;
493	u_int	    secsize;
494	struct ccb_nvmeio nvmeio;
495	struct	    disk *dp;
496	uint64_t    lba;
497	uint32_t    count;
498	int	    error = 0;
499
500	dp = arg;
501	periph = dp->d_drv1;
502	softc = (struct nda_softc *)periph->softc;
503	secsize = softc->disk->d_sectorsize;
504	lba = offset / secsize;
505	count = length / secsize;
506
507	if ((periph->flags & CAM_PERIPH_INVALID) != 0)
508		return (ENXIO);
509
510	/* xpt_get_ccb returns a zero'd allocation for the ccb, mimic that here */
511	memset(&nvmeio, 0, sizeof(nvmeio));
512	if (length > 0) {
513		xpt_setup_ccb(&nvmeio.ccb_h, periph->path, CAM_PRIORITY_NORMAL);
514		nvmeio.ccb_state = NDA_CCB_DUMP;
515		nda_nvme_write(softc, &nvmeio, virtual, lba, length, count);
516		error = cam_periph_runccb((union ccb *)&nvmeio, cam_periph_error,
517		    0, SF_NO_RECOVERY | SF_NO_RETRY, NULL);
518		if (error != 0)
519			printf("Aborting dump due to I/O error %d.\n", error);
520
521		return (error);
522	}
523
524	/* Flush */
525	xpt_setup_ccb(&nvmeio.ccb_h, periph->path, CAM_PRIORITY_NORMAL);
526
527	nvmeio.ccb_state = NDA_CCB_DUMP;
528	nda_nvme_flush(softc, &nvmeio);
529	error = cam_periph_runccb((union ccb *)&nvmeio, cam_periph_error,
530	    0, SF_NO_RECOVERY | SF_NO_RETRY, NULL);
531	if (error != 0)
532		xpt_print(periph->path, "flush cmd failed\n");
533	return (error);
534}
535
536static void
537ndainit(void)
538{
539	cam_status status;
540
541	/*
542	 * Install a global async callback.  This callback will
543	 * receive async callbacks like "new device found".
544	 */
545	status = xpt_register_async(AC_FOUND_DEVICE, ndaasync, NULL, NULL);
546
547	if (status != CAM_REQ_CMP) {
548		printf("nda: Failed to attach master async callback "
549		       "due to status 0x%x!\n", status);
550	} else if (nda_send_ordered) {
551
552		/* Register our event handlers */
553		if ((EVENTHANDLER_REGISTER(power_suspend, ndasuspend,
554					   NULL, EVENTHANDLER_PRI_LAST)) == NULL)
555		    printf("ndainit: power event registration failed!\n");
556		if ((EVENTHANDLER_REGISTER(shutdown_post_sync, ndashutdown,
557					   NULL, SHUTDOWN_PRI_DEFAULT)) == NULL)
558		    printf("ndainit: shutdown event registration failed!\n");
559	}
560}
561
562/*
563 * Callback from GEOM, called when it has finished cleaning up its
564 * resources.
565 */
566static void
567ndadiskgonecb(struct disk *dp)
568{
569	struct cam_periph *periph;
570
571	periph = (struct cam_periph *)dp->d_drv1;
572
573	cam_periph_release(periph);
574}
575
576static void
577ndaoninvalidate(struct cam_periph *periph)
578{
579	struct nda_softc *softc;
580
581	softc = (struct nda_softc *)periph->softc;
582
583	/*
584	 * De-register any async callbacks.
585	 */
586	xpt_register_async(0, ndaasync, periph, periph->path);
587#ifdef CAM_IO_STATS
588	softc->invalidations++;
589#endif
590
591	/*
592	 * Return all queued I/O with ENXIO.
593	 * XXX Handle any transactions queued to the card
594	 *     with XPT_ABORT_CCB.
595	 */
596	cam_iosched_flush(softc->cam_iosched, NULL, ENXIO);
597
598	disk_gone(softc->disk);
599}
600
601static void
602ndacleanup(struct cam_periph *periph)
603{
604	struct nda_softc *softc;
605
606	softc = (struct nda_softc *)periph->softc;
607
608	cam_periph_unlock(periph);
609
610	cam_iosched_fini(softc->cam_iosched);
611
612	/*
613	 * If we can't free the sysctl tree, oh well...
614	 */
615	if ((softc->flags & NDA_FLAG_SCTX_INIT) != 0) {
616#ifdef CAM_IO_STATS
617		if (sysctl_ctx_free(&softc->sysctl_stats_ctx) != 0)
618			xpt_print(periph->path,
619			    "can't remove sysctl stats context\n");
620#endif
621		if (sysctl_ctx_free(&softc->sysctl_ctx) != 0)
622			xpt_print(periph->path,
623			    "can't remove sysctl context\n");
624	}
625
626	disk_destroy(softc->disk);
627	free(softc, M_DEVBUF);
628	cam_periph_lock(periph);
629}
630
631static void
632ndaasync(void *callback_arg, u_int32_t code,
633	struct cam_path *path, void *arg)
634{
635	struct cam_periph *periph;
636
637	periph = (struct cam_periph *)callback_arg;
638	switch (code) {
639	case AC_FOUND_DEVICE:
640	{
641		struct ccb_getdev *cgd;
642		cam_status status;
643
644		cgd = (struct ccb_getdev *)arg;
645		if (cgd == NULL)
646			break;
647
648		if (cgd->protocol != PROTO_NVME)
649			break;
650
651		/*
652		 * Allocate a peripheral instance for
653		 * this device and start the probe
654		 * process.
655		 */
656		status = cam_periph_alloc(ndaregister, ndaoninvalidate,
657					  ndacleanup, ndastart,
658					  "nda", CAM_PERIPH_BIO,
659					  path, ndaasync,
660					  AC_FOUND_DEVICE, cgd);
661
662		if (status != CAM_REQ_CMP
663		 && status != CAM_REQ_INPROG)
664			printf("ndaasync: Unable to attach to new device "
665				"due to status 0x%x\n", status);
666		break;
667	}
668	case AC_ADVINFO_CHANGED:
669	{
670		uintptr_t buftype;
671
672		buftype = (uintptr_t)arg;
673		if (buftype == CDAI_TYPE_PHYS_PATH) {
674			struct nda_softc *softc;
675
676			softc = periph->softc;
677			disk_attr_changed(softc->disk, "GEOM::physpath",
678					  M_NOWAIT);
679		}
680		break;
681	}
682	case AC_LOST_DEVICE:
683	default:
684		cam_periph_async(periph, code, path, arg);
685		break;
686	}
687}
688
689static void
690ndasysctlinit(void *context, int pending)
691{
692	struct cam_periph *periph;
693	struct nda_softc *softc;
694	char tmpstr[32], tmpstr2[16];
695
696	periph = (struct cam_periph *)context;
697
698	/* periph was held for us when this task was enqueued */
699	if ((periph->flags & CAM_PERIPH_INVALID) != 0) {
700		cam_periph_release(periph);
701		return;
702	}
703
704	softc = (struct nda_softc *)periph->softc;
705	snprintf(tmpstr, sizeof(tmpstr), "CAM NDA unit %d", periph->unit_number);
706	snprintf(tmpstr2, sizeof(tmpstr2), "%d", periph->unit_number);
707
708	sysctl_ctx_init(&softc->sysctl_ctx);
709	softc->flags |= NDA_FLAG_SCTX_INIT;
710	softc->sysctl_tree = SYSCTL_ADD_NODE_WITH_LABEL(&softc->sysctl_ctx,
711		SYSCTL_STATIC_CHILDREN(_kern_cam_nda), OID_AUTO, tmpstr2,
712		CTLFLAG_RD, 0, tmpstr, "device_index");
713	if (softc->sysctl_tree == NULL) {
714		printf("ndasysctlinit: unable to allocate sysctl tree\n");
715		cam_periph_release(periph);
716		return;
717	}
718
719	SYSCTL_ADD_INT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
720	    OID_AUTO, "unmapped_io", CTLFLAG_RD,
721	    &softc->unmappedio, 0, "Unmapped I/O leaf");
722
723	SYSCTL_ADD_QUAD(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
724	    OID_AUTO, "deletes", CTLFLAG_RD,
725	    &softc->deletes, "Number of BIO_DELETE requests");
726
727	SYSCTL_ADD_QUAD(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
728	    OID_AUTO, "dsm_req", CTLFLAG_RD,
729	    &softc->dsm_req, "Number of DSM requests sent to SIM");
730
731	SYSCTL_ADD_INT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
732	    OID_AUTO, "rotating", CTLFLAG_RD, &nda_rotating_media, 1,
733	    "Rotating media");
734
735#ifdef CAM_IO_STATS
736	softc->sysctl_stats_tree = SYSCTL_ADD_NODE(&softc->sysctl_stats_ctx,
737		SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO, "stats",
738		CTLFLAG_RD, 0, "Statistics");
739	if (softc->sysctl_stats_tree == NULL) {
740		printf("ndasysctlinit: unable to allocate sysctl tree for stats\n");
741		cam_periph_release(periph);
742		return;
743	}
744	SYSCTL_ADD_INT(&softc->sysctl_stats_ctx,
745		SYSCTL_CHILDREN(softc->sysctl_stats_tree),
746		OID_AUTO, "timeouts", CTLFLAG_RD,
747		&softc->timeouts, 0,
748		"Device timeouts reported by the SIM");
749	SYSCTL_ADD_INT(&softc->sysctl_stats_ctx,
750		SYSCTL_CHILDREN(softc->sysctl_stats_tree),
751		OID_AUTO, "errors", CTLFLAG_RD,
752		&softc->errors, 0,
753		"Transport errors reported by the SIM.");
754	SYSCTL_ADD_INT(&softc->sysctl_stats_ctx,
755		SYSCTL_CHILDREN(softc->sysctl_stats_tree),
756		OID_AUTO, "pack_invalidations", CTLFLAG_RD,
757		&softc->invalidations, 0,
758		"Device pack invalidations.");
759#endif
760
761#ifdef CAM_TEST_FAILURE
762	SYSCTL_ADD_PROC(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
763		OID_AUTO, "invalidate", CTLTYPE_U64 | CTLFLAG_RW | CTLFLAG_MPSAFE,
764		periph, 0, cam_periph_invalidate_sysctl, "I",
765		"Write 1 to invalidate the drive immediately");
766#endif
767
768	cam_iosched_sysctl_init(softc->cam_iosched, &softc->sysctl_ctx,
769	    softc->sysctl_tree);
770
771	cam_periph_release(periph);
772}
773
774static int
775ndagetattr(struct bio *bp)
776{
777	int ret;
778	struct cam_periph *periph;
779
780	periph = (struct cam_periph *)bp->bio_disk->d_drv1;
781	cam_periph_lock(periph);
782	ret = xpt_getattr(bp->bio_data, bp->bio_length, bp->bio_attribute,
783	    periph->path);
784	cam_periph_unlock(periph);
785	if (ret == 0)
786		bp->bio_completed = bp->bio_length;
787	return ret;
788}
789
790static cam_status
791ndaregister(struct cam_periph *periph, void *arg)
792{
793	struct nda_softc *softc;
794	struct disk *disk;
795	struct ccb_pathinq cpi;
796	const struct nvme_namespace_data *nsd;
797	const struct nvme_controller_data *cd;
798	char   announce_buf[80];
799	uint8_t flbas_fmt, lbads, vwc_present;
800	u_int maxio;
801	int quirks;
802
803	nsd = nvme_get_identify_ns(periph);
804	cd = nvme_get_identify_cntrl(periph);
805
806	softc = (struct nda_softc *)malloc(sizeof(*softc), M_DEVBUF,
807	    M_NOWAIT | M_ZERO);
808
809	if (softc == NULL) {
810		printf("ndaregister: Unable to probe new device. "
811		    "Unable to allocate softc\n");
812		return(CAM_REQ_CMP_ERR);
813	}
814
815	if (cam_iosched_init(&softc->cam_iosched, periph) != 0) {
816		printf("ndaregister: Unable to probe new device. "
817		       "Unable to allocate iosched memory\n");
818		free(softc, M_DEVBUF);
819		return(CAM_REQ_CMP_ERR);
820	}
821
822	/* ident_data parsing */
823
824	periph->softc = softc;
825	softc->quirks = NDA_Q_NONE;
826	xpt_path_inq(&cpi, periph->path);
827	TASK_INIT(&softc->sysctl_task, 0, ndasysctlinit, periph);
828
829	/*
830	 * The name space ID is the lun, save it for later I/O
831	 */
832	softc->nsid = (uint32_t)xpt_path_lun_id(periph->path);
833
834	/*
835	 * Register this media as a disk
836	 */
837	(void)cam_periph_hold(periph, PRIBIO);
838	cam_periph_unlock(periph);
839	snprintf(announce_buf, sizeof(announce_buf),
840	    "kern.cam.nda.%d.quirks", periph->unit_number);
841	quirks = softc->quirks;
842	TUNABLE_INT_FETCH(announce_buf, &quirks);
843	softc->quirks = quirks;
844	cam_iosched_set_sort_queue(softc->cam_iosched, 0);
845	softc->disk = disk = disk_alloc();
846	disk->d_rotation_rate = DISK_RR_NON_ROTATING;
847	disk->d_open = ndaopen;
848	disk->d_close = ndaclose;
849	disk->d_strategy = ndastrategy;
850	disk->d_ioctl = ndaioctl;
851	disk->d_getattr = ndagetattr;
852	disk->d_dump = ndadump;
853	disk->d_gone = ndadiskgonecb;
854	disk->d_name = "nda";
855	disk->d_drv1 = periph;
856	disk->d_unit = periph->unit_number;
857	maxio = cpi.maxio;		/* Honor max I/O size of SIM */
858	if (maxio == 0)
859		maxio = DFLTPHYS;	/* traditional default */
860	else if (maxio > MAXPHYS)
861		maxio = MAXPHYS;	/* for safety */
862	disk->d_maxsize = maxio;
863	flbas_fmt = (nsd->flbas >> NVME_NS_DATA_FLBAS_FORMAT_SHIFT) &
864		NVME_NS_DATA_FLBAS_FORMAT_MASK;
865	lbads = (nsd->lbaf[flbas_fmt] >> NVME_NS_DATA_LBAF_LBADS_SHIFT) &
866		NVME_NS_DATA_LBAF_LBADS_MASK;
867	disk->d_sectorsize = 1 << lbads;
868	disk->d_mediasize = (off_t)(disk->d_sectorsize * nsd->nsze);
869	disk->d_delmaxsize = disk->d_mediasize;
870	disk->d_flags = DISKFLAG_DIRECT_COMPLETION;
871	if (nvme_ctrlr_has_dataset_mgmt(cd))
872		disk->d_flags |= DISKFLAG_CANDELETE;
873	vwc_present = (cd->vwc >> NVME_CTRLR_DATA_VWC_PRESENT_SHIFT) &
874		NVME_CTRLR_DATA_VWC_PRESENT_MASK;
875	if (vwc_present)
876		disk->d_flags |= DISKFLAG_CANFLUSHCACHE;
877	if ((cpi.hba_misc & PIM_UNMAPPED) != 0) {
878		disk->d_flags |= DISKFLAG_UNMAPPED_BIO;
879		softc->unmappedio = 1;
880	}
881	/*
882	 * d_ident and d_descr are both far bigger than the length of either
883	 *  the serial or model number strings.
884	 */
885	cam_strvis(disk->d_descr, cd->mn,
886	    NVME_MODEL_NUMBER_LENGTH, sizeof(disk->d_descr));
887	cam_strvis(disk->d_ident, cd->sn,
888	    NVME_SERIAL_NUMBER_LENGTH, sizeof(disk->d_ident));
889	disk->d_hba_vendor = cpi.hba_vendor;
890	disk->d_hba_device = cpi.hba_device;
891	disk->d_hba_subvendor = cpi.hba_subvendor;
892	disk->d_hba_subdevice = cpi.hba_subdevice;
893	if (((nsd->nsfeat >> NVME_NS_DATA_NSFEAT_NPVALID_SHIFT) &
894	    NVME_NS_DATA_NSFEAT_NPVALID_MASK) != 0 && nsd->npwg != 0)
895		disk->d_stripesize = ((nsd->npwg + 1) * disk->d_sectorsize);
896	else
897		disk->d_stripesize = nsd->noiob * disk->d_sectorsize;
898	disk->d_stripeoffset = 0;
899	disk->d_devstat = devstat_new_entry(periph->periph_name,
900	    periph->unit_number, disk->d_sectorsize,
901	    DEVSTAT_ALL_SUPPORTED,
902	    DEVSTAT_TYPE_DIRECT | XPORT_DEVSTAT_TYPE(cpi.transport),
903	    DEVSTAT_PRIORITY_DISK);
904	/*
905	 * Add alias for older nvd drives to ease transition.
906	 */
907	/* disk_add_alias(disk, "nvd"); Have reports of this causing problems */
908
909	/*
910	 * Acquire a reference to the periph before we register with GEOM.
911	 * We'll release this reference once GEOM calls us back (via
912	 * ndadiskgonecb()) telling us that our provider has been freed.
913	 */
914	if (cam_periph_acquire(periph) != 0) {
915		xpt_print(periph->path, "%s: lost periph during "
916			  "registration!\n", __func__);
917		cam_periph_lock(periph);
918		return (CAM_REQ_CMP_ERR);
919	}
920	disk_create(softc->disk, DISK_VERSION);
921	cam_periph_lock(periph);
922	cam_periph_unhold(periph);
923
924	snprintf(announce_buf, sizeof(announce_buf),
925		"%juMB (%ju %u byte sectors)",
926	    (uintmax_t)((uintmax_t)disk->d_mediasize / (1024*1024)),
927		(uintmax_t)disk->d_mediasize / disk->d_sectorsize,
928		disk->d_sectorsize);
929	xpt_announce_periph(periph, announce_buf);
930	xpt_announce_quirks(periph, softc->quirks, NDA_Q_BIT_STRING);
931
932	/*
933	 * Create our sysctl variables, now that we know
934	 * we have successfully attached.
935	 */
936	if (cam_periph_acquire(periph) == 0)
937		taskqueue_enqueue(taskqueue_thread, &softc->sysctl_task);
938
939	/*
940	 * Register for device going away and info about the drive
941	 * changing (though with NVMe, it can't)
942	 */
943	xpt_register_async(AC_LOST_DEVICE | AC_ADVINFO_CHANGED,
944	    ndaasync, periph, periph->path);
945
946	softc->state = NDA_STATE_NORMAL;
947	return(CAM_REQ_CMP);
948}
949
950static void
951ndastart(struct cam_periph *periph, union ccb *start_ccb)
952{
953	struct nda_softc *softc = (struct nda_softc *)periph->softc;
954	struct ccb_nvmeio *nvmeio = &start_ccb->nvmeio;
955
956	CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("ndastart\n"));
957
958	switch (softc->state) {
959	case NDA_STATE_NORMAL:
960	{
961		struct bio *bp;
962
963		bp = cam_iosched_next_bio(softc->cam_iosched);
964		CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("ndastart: bio %p\n", bp));
965		if (bp == NULL) {
966			xpt_release_ccb(start_ccb);
967			break;
968		}
969
970		switch (bp->bio_cmd) {
971		case BIO_WRITE:
972			softc->flags |= NDA_FLAG_DIRTY;
973			/* FALLTHROUGH */
974		case BIO_READ:
975		{
976#ifdef CAM_TEST_FAILURE
977			int fail = 0;
978
979			/*
980			 * Support the failure ioctls.  If the command is a
981			 * read, and there are pending forced read errors, or
982			 * if a write and pending write errors, then fail this
983			 * operation with EIO.  This is useful for testing
984			 * purposes.  Also, support having every Nth read fail.
985			 *
986			 * This is a rather blunt tool.
987			 */
988			if (bp->bio_cmd == BIO_READ) {
989				if (softc->force_read_error) {
990					softc->force_read_error--;
991					fail = 1;
992				}
993				if (softc->periodic_read_error > 0) {
994					if (++softc->periodic_read_count >=
995					    softc->periodic_read_error) {
996						softc->periodic_read_count = 0;
997						fail = 1;
998					}
999				}
1000			} else {
1001				if (softc->force_write_error) {
1002					softc->force_write_error--;
1003					fail = 1;
1004				}
1005			}
1006			if (fail) {
1007				biofinish(bp, NULL, EIO);
1008				xpt_release_ccb(start_ccb);
1009				ndaschedule(periph);
1010				return;
1011			}
1012#endif
1013			KASSERT((bp->bio_flags & BIO_UNMAPPED) == 0 ||
1014			    round_page(bp->bio_bcount + bp->bio_ma_offset) /
1015			    PAGE_SIZE == bp->bio_ma_n,
1016			    ("Short bio %p", bp));
1017			nda_nvme_rw_bio(softc, &start_ccb->nvmeio, bp, bp->bio_cmd == BIO_READ ?
1018			    NVME_OPC_READ : NVME_OPC_WRITE);
1019			break;
1020		}
1021		case BIO_DELETE:
1022		{
1023			struct nvme_dsm_range *dsm_range, *dsm_end;
1024			struct nda_trim_request *trim;
1025			struct bio *bp1;
1026			int ents;
1027
1028			trim = malloc(sizeof(*trim), M_NVMEDA, M_ZERO | M_NOWAIT);
1029			if (trim == NULL) {
1030				biofinish(bp, NULL, ENOMEM);
1031				xpt_release_ccb(start_ccb);
1032				ndaschedule(periph);
1033				return;
1034			}
1035			TAILQ_INIT(&trim->bps);
1036			bp1 = bp;
1037			ents = sizeof(trim->data) / sizeof(struct nvme_dsm_range);
1038			ents = min(ents, nda_max_trim_entries);
1039			dsm_range = &trim->dsm;
1040			dsm_end = dsm_range + ents;
1041			do {
1042				TAILQ_INSERT_TAIL(&trim->bps, bp1, bio_queue);
1043				dsm_range->length =
1044				    htole32(bp1->bio_bcount / softc->disk->d_sectorsize);
1045				dsm_range->starting_lba =
1046				    htole64(bp1->bio_offset / softc->disk->d_sectorsize);
1047				dsm_range++;
1048				if (dsm_range >= dsm_end)
1049					break;
1050				bp1 = cam_iosched_next_trim(softc->cam_iosched);
1051				/* XXX -- Could collapse adjacent ranges, but we don't for now */
1052				/* XXX -- Could limit based on total payload size */
1053			} while (bp1 != NULL);
1054			start_ccb->ccb_trim = trim;
1055			softc->dsm_req++;
1056			nda_nvme_trim(softc, &start_ccb->nvmeio, &trim->dsm,
1057			    dsm_range - &trim->dsm);
1058			start_ccb->ccb_state = NDA_CCB_TRIM;
1059			/*
1060			 * Note: We can have multiple TRIMs in flight, so we don't call
1061			 * cam_iosched_submit_trim(softc->cam_iosched);
1062			 * since that forces the I/O scheduler to only schedule one at a time.
1063			 * On NVMe drives, this is a performance disaster.
1064			 */
1065			goto out;
1066		}
1067		case BIO_FLUSH:
1068			nda_nvme_flush(softc, nvmeio);
1069			break;
1070		}
1071		start_ccb->ccb_state = NDA_CCB_BUFFER_IO;
1072		start_ccb->ccb_bp = bp;
1073out:
1074		start_ccb->ccb_h.flags |= CAM_UNLOCKED;
1075		softc->outstanding_cmds++;
1076		softc->refcount++;			/* For submission only */
1077		cam_periph_unlock(periph);
1078		xpt_action(start_ccb);
1079		cam_periph_lock(periph);
1080		softc->refcount--;			/* Submission done */
1081
1082		/* May have more work to do, so ensure we stay scheduled */
1083		ndaschedule(periph);
1084		break;
1085		}
1086	}
1087}
1088
1089static void
1090ndadone(struct cam_periph *periph, union ccb *done_ccb)
1091{
1092	struct nda_softc *softc;
1093	struct ccb_nvmeio *nvmeio = &done_ccb->nvmeio;
1094	struct cam_path *path;
1095	int state;
1096
1097	softc = (struct nda_softc *)periph->softc;
1098	path = done_ccb->ccb_h.path;
1099
1100	CAM_DEBUG(path, CAM_DEBUG_TRACE, ("ndadone\n"));
1101
1102	state = nvmeio->ccb_state & NDA_CCB_TYPE_MASK;
1103	switch (state) {
1104	case NDA_CCB_BUFFER_IO:
1105	case NDA_CCB_TRIM:
1106	{
1107		int error;
1108
1109		cam_periph_lock(periph);
1110		if ((done_ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_CMP) {
1111			error = ndaerror(done_ccb, 0, 0);
1112			if (error == ERESTART) {
1113				/* A retry was scheduled, so just return. */
1114				cam_periph_unlock(periph);
1115				return;
1116			}
1117			if ((done_ccb->ccb_h.status & CAM_DEV_QFRZN) != 0)
1118				cam_release_devq(path,
1119						 /*relsim_flags*/0,
1120						 /*reduction*/0,
1121						 /*timeout*/0,
1122						 /*getcount_only*/0);
1123		} else {
1124			if ((done_ccb->ccb_h.status & CAM_DEV_QFRZN) != 0)
1125				panic("REQ_CMP with QFRZN");
1126			error = 0;
1127		}
1128		if (state == NDA_CCB_BUFFER_IO) {
1129			struct bio *bp;
1130
1131			bp = (struct bio *)done_ccb->ccb_bp;
1132			bp->bio_error = error;
1133			if (error != 0) {
1134				bp->bio_resid = bp->bio_bcount;
1135				bp->bio_flags |= BIO_ERROR;
1136			} else {
1137				bp->bio_resid = 0;
1138			}
1139			softc->outstanding_cmds--;
1140
1141			/*
1142			 * We need to call cam_iosched before we call biodone so that we
1143			 * don't measure any activity that happens in the completion
1144			 * routine, which in the case of sendfile can be quite
1145			 * extensive.
1146			 */
1147			cam_iosched_bio_complete(softc->cam_iosched, bp, done_ccb);
1148			xpt_release_ccb(done_ccb);
1149			ndaschedule(periph);
1150			cam_periph_unlock(periph);
1151			biodone(bp);
1152		} else { /* state == NDA_CCB_TRIM */
1153			struct nda_trim_request *trim;
1154			struct bio *bp1, *bp2;
1155			TAILQ_HEAD(, bio) queue;
1156
1157			trim = nvmeio->ccb_trim;
1158			TAILQ_INIT(&queue);
1159			TAILQ_CONCAT(&queue, &trim->bps, bio_queue);
1160			free(trim, M_NVMEDA);
1161
1162			/*
1163			 * Since we can have multiple trims in flight, we don't
1164			 * need to call this here.
1165			 * cam_iosched_trim_done(softc->cam_iosched);
1166			 */
1167			/*
1168			 * The the I/O scheduler that we're finishing the I/O
1169			 * so we can keep book. The first one we pass in the CCB
1170			 * which has the timing information. The rest we pass in NULL
1171			 * so we can keep proper counts.
1172			 */
1173			bp1 = TAILQ_FIRST(&queue);
1174			cam_iosched_bio_complete(softc->cam_iosched, bp1, done_ccb);
1175			xpt_release_ccb(done_ccb);
1176			softc->outstanding_cmds--;
1177			ndaschedule(periph);
1178			cam_periph_unlock(periph);
1179			while ((bp2 = TAILQ_FIRST(&queue)) != NULL) {
1180				TAILQ_REMOVE(&queue, bp2, bio_queue);
1181				bp2->bio_error = error;
1182				if (error != 0) {
1183					bp2->bio_flags |= BIO_ERROR;
1184					bp2->bio_resid = bp1->bio_bcount;
1185				} else
1186					bp2->bio_resid = 0;
1187				if (bp1 != bp2)
1188					cam_iosched_bio_complete(softc->cam_iosched, bp2, NULL);
1189				biodone(bp2);
1190			}
1191		}
1192		return;
1193	}
1194	case NDA_CCB_DUMP:
1195		/* No-op.  We're polling */
1196		return;
1197	case NDA_CCB_PASS:
1198		return;
1199	default:
1200		break;
1201	}
1202	xpt_release_ccb(done_ccb);
1203}
1204
1205static int
1206ndaerror(union ccb *ccb, u_int32_t cam_flags, u_int32_t sense_flags)
1207{
1208	struct nda_softc *softc;
1209	struct cam_periph *periph;
1210
1211	periph = xpt_path_periph(ccb->ccb_h.path);
1212	softc = (struct nda_softc *)periph->softc;
1213
1214	switch (ccb->ccb_h.status & CAM_STATUS_MASK) {
1215	case CAM_CMD_TIMEOUT:
1216#ifdef CAM_IO_STATS
1217		softc->timeouts++;
1218#endif
1219		break;
1220	case CAM_REQ_ABORTED:
1221	case CAM_REQ_CMP_ERR:
1222	case CAM_REQ_TERMIO:
1223	case CAM_UNREC_HBA_ERROR:
1224	case CAM_DATA_RUN_ERR:
1225	case CAM_ATA_STATUS_ERROR:
1226#ifdef CAM_IO_STATS
1227		softc->errors++;
1228#endif
1229		break;
1230	default:
1231		break;
1232	}
1233
1234	return(cam_periph_error(ccb, cam_flags, sense_flags));
1235}
1236
1237/*
1238 * Step through all NDA peripheral drivers, and if the device is still open,
1239 * sync the disk cache to physical media.
1240 */
1241static void
1242ndaflush(void)
1243{
1244	struct cam_periph *periph;
1245	struct nda_softc *softc;
1246	union ccb *ccb;
1247	int error;
1248
1249	CAM_PERIPH_FOREACH(periph, &ndadriver) {
1250		softc = (struct nda_softc *)periph->softc;
1251
1252		if (SCHEDULER_STOPPED()) {
1253			/*
1254			 * If we paniced with the lock held or the periph is not
1255			 * open, do not recurse.  Otherwise, call ndadump since
1256			 * that avoids the sleeping cam_periph_getccb does if no
1257			 * CCBs are available.
1258			 */
1259			if (!cam_periph_owned(periph) &&
1260			    (softc->flags & NDA_FLAG_OPEN)) {
1261				ndadump(softc->disk, NULL, 0, 0, 0);
1262			}
1263			continue;
1264		}
1265
1266		/*
1267		 * We only sync the cache if the drive is still open
1268		 */
1269		cam_periph_lock(periph);
1270		if ((softc->flags & NDA_FLAG_OPEN) == 0) {
1271			cam_periph_unlock(periph);
1272			continue;
1273		}
1274
1275		ccb = cam_periph_getccb(periph, CAM_PRIORITY_NORMAL);
1276		nda_nvme_flush(softc, &ccb->nvmeio);
1277		error = cam_periph_runccb(ccb, ndaerror, /*cam_flags*/0,
1278		    /*sense_flags*/ SF_NO_RECOVERY | SF_NO_RETRY,
1279		    softc->disk->d_devstat);
1280		if (error != 0)
1281			xpt_print(periph->path, "Synchronize cache failed\n");
1282		xpt_release_ccb(ccb);
1283		cam_periph_unlock(periph);
1284	}
1285}
1286
1287static void
1288ndashutdown(void *arg, int howto)
1289{
1290
1291	ndaflush();
1292}
1293
1294static void
1295ndasuspend(void *arg)
1296{
1297
1298	ndaflush();
1299}
1300