xref: /freebsd/sys/cam/nvme/nvme_da.c (revision df21a004be237a1dccd03c7b47254625eea62fa9)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
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  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  *
27  * Derived from ata_da.c:
28  * Copyright (c) 2009 Alexander Motin <mav@FreeBSD.org>
29  */
30 
31 #include <sys/param.h>
32 
33 #ifdef _KERNEL
34 #include <sys/systm.h>
35 #include <sys/kernel.h>
36 #include <sys/bio.h>
37 #include <sys/sysctl.h>
38 #include <sys/taskqueue.h>
39 #include <sys/lock.h>
40 #include <sys/mutex.h>
41 #include <sys/conf.h>
42 #include <sys/devicestat.h>
43 #include <sys/eventhandler.h>
44 #include <sys/malloc.h>
45 #include <sys/cons.h>
46 #include <sys/power.h>
47 #include <sys/proc.h>
48 #include <sys/reboot.h>
49 #include <sys/sbuf.h>
50 #include <geom/geom.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 
68 typedef enum {
69 	NDA_STATE_NORMAL
70 } nda_state;
71 
72 typedef enum {
73 	NDA_FLAG_OPEN		= 0x0001,
74 	NDA_FLAG_DIRTY		= 0x0002,
75 	NDA_FLAG_SCTX_INIT	= 0x0004,
76 } nda_flags;
77 #define NDA_FLAG_STRING		\
78 	"\020"			\
79 	"\001OPEN"		\
80 	"\002DIRTY"		\
81 	"\003SCTX_INIT"
82 
83 typedef enum {
84 	NDA_Q_4K   = 0x01,
85 	NDA_Q_NONE = 0x00,
86 } nda_quirks;
87 
88 #define NDA_Q_BIT_STRING	\
89 	"\020"			\
90 	"\001Bit 0"
91 
92 typedef enum {
93 	NDA_CCB_BUFFER_IO	= 0x01,
94 	NDA_CCB_DUMP            = 0x02,
95 	NDA_CCB_TRIM            = 0x03,
96 	NDA_CCB_PASS            = 0x04,
97 	NDA_CCB_TYPE_MASK	= 0x0F,
98 } nda_ccb_state;
99 
100 /* Offsets into our private area for storing information */
101 #define ccb_state	ccb_h.ppriv_field0
102 #define ccb_bp		ccb_h.ppriv_ptr1	/* For NDA_CCB_BUFFER_IO */
103 #define ccb_trim	ccb_h.ppriv_ptr1	/* For NDA_CCB_TRIM */
104 
105 struct nda_softc {
106 	struct   cam_iosched_softc *cam_iosched;
107 	int			outstanding_cmds;	/* Number of active commands */
108 	int			refcount;		/* Active xpt_action() calls */
109 	nda_state		state;
110 	nda_flags		flags;
111 	nda_quirks		quirks;
112 	int			unmappedio;
113 	quad_t			deletes;
114 	uint32_t		nsid;			/* Namespace ID for this nda device */
115 	struct disk		*disk;
116 	struct task		sysctl_task;
117 	struct sysctl_ctx_list	sysctl_ctx;
118 	struct sysctl_oid	*sysctl_tree;
119 	uint64_t		trim_count;
120 	uint64_t		trim_ranges;
121 	uint64_t		trim_lbas;
122 #ifdef CAM_TEST_FAILURE
123 	int			force_read_error;
124 	int			force_write_error;
125 	int			periodic_read_error;
126 	int			periodic_read_count;
127 #endif
128 #ifdef CAM_IO_STATS
129 	struct sysctl_ctx_list	sysctl_stats_ctx;
130 	struct sysctl_oid	*sysctl_stats_tree;
131 	u_int			timeouts;
132 	u_int			errors;
133 	u_int			invalidations;
134 #endif
135 };
136 
137 struct nda_trim_request {
138 	struct nvme_dsm_range	dsm[NVME_MAX_DSM_TRIM / sizeof(struct nvme_dsm_range)];
139 	TAILQ_HEAD(, bio) bps;
140 };
141 _Static_assert(NVME_MAX_DSM_TRIM % sizeof(struct nvme_dsm_range) == 0,
142     "NVME_MAX_DSM_TRIM must be an integral number of ranges");
143 
144 /* Need quirk table */
145 
146 static	disk_ioctl_t	ndaioctl;
147 static	disk_strategy_t	ndastrategy;
148 static	dumper_t	ndadump;
149 static	periph_init_t	ndainit;
150 static	void		ndaasync(void *callback_arg, uint32_t code,
151 				struct cam_path *path, void *arg);
152 static	void		ndasysctlinit(void *context, int pending);
153 static	int		ndaflagssysctl(SYSCTL_HANDLER_ARGS);
154 static	periph_ctor_t	ndaregister;
155 static	periph_dtor_t	ndacleanup;
156 static	periph_start_t	ndastart;
157 static	periph_oninv_t	ndaoninvalidate;
158 static	void		ndadone(struct cam_periph *periph,
159 			       union ccb *done_ccb);
160 static  int		ndaerror(union ccb *ccb, uint32_t cam_flags,
161 				uint32_t sense_flags);
162 static void		ndashutdown(void *arg, int howto);
163 static void		ndasuspend(void *arg, enum power_stype stype);
164 
165 #ifndef	NDA_DEFAULT_SEND_ORDERED
166 #define	NDA_DEFAULT_SEND_ORDERED	1
167 #endif
168 #ifndef NDA_DEFAULT_TIMEOUT
169 #define NDA_DEFAULT_TIMEOUT 30	/* Timeout in seconds */
170 #endif
171 #ifndef	NDA_DEFAULT_RETRY
172 #define	NDA_DEFAULT_RETRY	4
173 #endif
174 #ifndef NDA_MAX_TRIM_ENTRIES
175 #define NDA_MAX_TRIM_ENTRIES  (NVME_MAX_DSM_TRIM / sizeof(struct nvme_dsm_range))/* Number of DSM trims to use, max 256 */
176 #endif
177 
178 static SYSCTL_NODE(_kern_cam, OID_AUTO, nda, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
179     "CAM Direct Access Disk driver");
180 
181 //static int nda_retry_count = NDA_DEFAULT_RETRY;
182 static int nda_send_ordered = NDA_DEFAULT_SEND_ORDERED;
183 static int nda_default_timeout = NDA_DEFAULT_TIMEOUT;
184 static int nda_max_trim_entries = NDA_MAX_TRIM_ENTRIES;
185 static int nda_enable_biospeedup = 1;
186 static int nda_nvd_compat = 1;
187 SYSCTL_INT(_kern_cam_nda, OID_AUTO, max_trim, CTLFLAG_RDTUN,
188     &nda_max_trim_entries, NDA_MAX_TRIM_ENTRIES,
189     "Maximum number of BIO_DELETE to send down as a DSM TRIM.");
190 SYSCTL_INT(_kern_cam_nda, OID_AUTO, enable_biospeedup, CTLFLAG_RDTUN,
191     &nda_enable_biospeedup, 0, "Enable BIO_SPEEDUP processing.");
192 SYSCTL_INT(_kern_cam_nda, OID_AUTO, nvd_compat, CTLFLAG_RDTUN,
193     &nda_nvd_compat, 1, "Enable creation of nvd aliases.");
194 
195 /*
196  * All NVMe media is non-rotational, so all nvme device instances
197  * share this to implement the sysctl.
198  */
199 static int nda_rotating_media = 0;
200 
201 static struct periph_driver ndadriver =
202 {
203 	ndainit, "nda",
204 	TAILQ_HEAD_INITIALIZER(ndadriver.units), /* generation */ 0
205 };
206 
207 PERIPHDRIVER_DECLARE(nda, ndadriver);
208 
209 static MALLOC_DEFINE(M_NVMEDA, "nvme_da", "nvme_da buffers");
210 
211 /*
212  * nice wrappers. Maybe these belong in nvme_all.c instead of
213  * here, but this is the only place that uses these. Should
214  * we ever grow another NVME periph, we should move them
215  * all there wholesale.
216  */
217 
218 static void
219 nda_nvme_flush(struct nda_softc *softc, struct ccb_nvmeio *nvmeio)
220 {
221 	cam_fill_nvmeio(nvmeio,
222 	    0,			/* retries */
223 	    ndadone,		/* cbfcnp */
224 	    CAM_DIR_NONE,	/* flags */
225 	    NULL,		/* data_ptr */
226 	    0,			/* dxfer_len */
227 	    nda_default_timeout * 1000); /* timeout 30s */
228 	nvme_ns_flush_cmd(&nvmeio->cmd, softc->nsid);
229 }
230 
231 static void
232 nda_nvme_trim(struct nda_softc *softc, struct ccb_nvmeio *nvmeio,
233     void *payload, uint32_t num_ranges)
234 {
235 	cam_fill_nvmeio(nvmeio,
236 	    0,			/* retries */
237 	    ndadone,		/* cbfcnp */
238 	    CAM_DIR_OUT,	/* flags */
239 	    payload,		/* data_ptr */
240 	    num_ranges * sizeof(struct nvme_dsm_range), /* dxfer_len */
241 	    nda_default_timeout * 1000); /* timeout 30s */
242 	nvme_ns_trim_cmd(&nvmeio->cmd, softc->nsid, num_ranges);
243 }
244 
245 static void
246 nda_nvme_write(struct nda_softc *softc, struct ccb_nvmeio *nvmeio,
247     void *payload, uint64_t lba, uint32_t len, uint32_t count)
248 {
249 	cam_fill_nvmeio(nvmeio,
250 	    0,			/* retries */
251 	    ndadone,		/* cbfcnp */
252 	    CAM_DIR_OUT,	/* flags */
253 	    payload,		/* data_ptr */
254 	    len,		/* dxfer_len */
255 	    nda_default_timeout * 1000); /* timeout 30s */
256 	nvme_ns_write_cmd(&nvmeio->cmd, softc->nsid, lba, count);
257 }
258 
259 static void
260 nda_nvme_rw_bio(struct nda_softc *softc, struct ccb_nvmeio *nvmeio,
261     struct bio *bp, uint32_t rwcmd)
262 {
263 	int flags = rwcmd == NVME_OPC_READ ? CAM_DIR_IN : CAM_DIR_OUT;
264 	void *payload;
265 	uint64_t lba;
266 	uint32_t count;
267 
268 	if (bp->bio_flags & BIO_UNMAPPED) {
269 		flags |= CAM_DATA_BIO;
270 		payload = bp;
271 	} else {
272 		payload = bp->bio_data;
273 	}
274 
275 	lba = bp->bio_pblkno;
276 	count = bp->bio_bcount / softc->disk->d_sectorsize;
277 
278 	cam_fill_nvmeio(nvmeio,
279 	    0,			/* retries */
280 	    ndadone,		/* cbfcnp */
281 	    flags,		/* flags */
282 	    payload,		/* data_ptr */
283 	    bp->bio_bcount,	/* dxfer_len */
284 	    nda_default_timeout * 1000); /* timeout 30s */
285 	nvme_ns_rw_cmd(&nvmeio->cmd, rwcmd, softc->nsid, lba, count);
286 }
287 
288 static int
289 ndaopen(struct disk *dp)
290 {
291 	struct cam_periph *periph;
292 	struct nda_softc *softc;
293 	int error;
294 
295 	periph = (struct cam_periph *)dp->d_drv1;
296 	if (cam_periph_acquire(periph) != 0) {
297 		return(ENXIO);
298 	}
299 
300 	cam_periph_lock(periph);
301 	if ((error = cam_periph_hold(periph, PRIBIO|PCATCH)) != 0) {
302 		cam_periph_unlock(periph);
303 		cam_periph_release(periph);
304 		return (error);
305 	}
306 
307 	CAM_DEBUG(periph->path, CAM_DEBUG_TRACE | CAM_DEBUG_PERIPH,
308 	    ("ndaopen\n"));
309 
310 	softc = (struct nda_softc *)periph->softc;
311 	softc->flags |= NDA_FLAG_OPEN;
312 
313 	cam_periph_unhold(periph);
314 	cam_periph_unlock(periph);
315 	return (0);
316 }
317 
318 static int
319 ndaclose(struct disk *dp)
320 {
321 	struct	cam_periph *periph;
322 	struct	nda_softc *softc;
323 	union ccb *ccb;
324 	int error;
325 
326 	periph = (struct cam_periph *)dp->d_drv1;
327 	softc = (struct nda_softc *)periph->softc;
328 	cam_periph_lock(periph);
329 
330 	CAM_DEBUG(periph->path, CAM_DEBUG_TRACE | CAM_DEBUG_PERIPH,
331 	    ("ndaclose\n"));
332 
333 	if ((softc->flags & NDA_FLAG_DIRTY) != 0 &&
334 	    (periph->flags & CAM_PERIPH_INVALID) == 0 &&
335 	    cam_periph_hold(periph, PRIBIO) == 0) {
336 		ccb = cam_periph_getccb(periph, CAM_PRIORITY_NORMAL);
337 		nda_nvme_flush(softc, &ccb->nvmeio);
338 		error = cam_periph_runccb(ccb, ndaerror, /*cam_flags*/0,
339 		    /*sense_flags*/0, softc->disk->d_devstat);
340 
341 		if (error != 0)
342 			xpt_print(periph->path, "Synchronize cache failed\n");
343 		else
344 			softc->flags &= ~NDA_FLAG_DIRTY;
345 		xpt_release_ccb(ccb);
346 		cam_periph_unhold(periph);
347 	}
348 
349 	softc->flags &= ~NDA_FLAG_OPEN;
350 
351 	while (softc->refcount != 0)
352 		cam_periph_sleep(periph, &softc->refcount, PRIBIO, "ndaclose", 1);
353 	KASSERT(softc->outstanding_cmds == 0,
354 	    ("nda %d outstanding commands", softc->outstanding_cmds));
355 	cam_periph_unlock(periph);
356 	cam_periph_release(periph);
357 	return (0);
358 }
359 
360 static void
361 ndaschedule(struct cam_periph *periph)
362 {
363 	struct nda_softc *softc = (struct nda_softc *)periph->softc;
364 
365 	if (softc->state != NDA_STATE_NORMAL)
366 		return;
367 
368 	cam_iosched_schedule(softc->cam_iosched, periph);
369 }
370 
371 static int
372 ndaioctl(struct disk *dp, u_long cmd, void *data, int fflag,
373     struct thread *td)
374 {
375 	struct cam_periph *periph;
376 
377 	periph = (struct cam_periph *)dp->d_drv1;
378 
379 	switch (cmd) {
380 	case NVME_IO_TEST:
381 	case NVME_BIO_TEST:
382 		/*
383 		 * These don't map well to the underlying CCBs, so
384 		 * they are usupported via CAM.
385 		 */
386 		return (ENOTTY);
387 	case NVME_GET_NSID:
388 	{
389 		struct nvme_get_nsid *gnsid = (struct nvme_get_nsid *)data;
390 		struct ccb_pathinq cpi;
391 
392 		xpt_path_inq(&cpi, periph->path);
393 		strncpy(gnsid->cdev, cpi.xport_specific.nvme.dev_name,
394 		    sizeof(gnsid->cdev));
395 		gnsid->nsid = cpi.xport_specific.nvme.nsid;
396 		return (0);
397 	}
398 	case NVME_PASSTHROUGH_CMD:
399 	{
400 		struct nvme_pt_command *pt;
401 		union ccb *ccb;
402 		struct cam_periph_map_info mapinfo;
403 		u_int maxmap = dp->d_maxsize;
404 		int error;
405 
406 		/*
407 		 * Create a NVME_IO CCB to do the passthrough command.
408 		 */
409 		pt = (struct nvme_pt_command *)data;
410 		ccb = xpt_alloc_ccb();
411 		xpt_setup_ccb(&ccb->ccb_h, periph->path, CAM_PRIORITY_NORMAL);
412 		ccb->ccb_state = NDA_CCB_PASS;
413 		cam_fill_nvmeio(&ccb->nvmeio,
414 		    0,			/* Retries */
415 		    ndadone,
416 		    (pt->is_read ? CAM_DIR_IN : CAM_DIR_OUT) | CAM_DATA_VADDR,
417 		    pt->buf,
418 		    pt->len,
419 		    nda_default_timeout * 1000);
420 		memcpy(&ccb->nvmeio.cmd, &pt->cmd, sizeof(pt->cmd));
421 
422 		/*
423 		 * Wire the user memory in this request for the I/O
424 		 */
425 		memset(&mapinfo, 0, sizeof(mapinfo));
426 		error = cam_periph_mapmem(ccb, &mapinfo, maxmap);
427 		if (error)
428 			goto out;
429 
430 		/*
431 		 * Lock the periph and run the command.
432 		 */
433 		cam_periph_lock(periph);
434 		cam_periph_runccb(ccb, NULL, CAM_RETRY_SELTO,
435 		    SF_RETRY_UA | SF_NO_PRINT, NULL);
436 
437 		/*
438 		 * Tear down mapping and return status.
439 		 */
440 		cam_periph_unlock(periph);
441 		error = cam_periph_unmapmem(ccb, &mapinfo);
442 		if (!cam_ccb_success(ccb))
443 			error = EIO;
444 out:
445 		cam_periph_lock(periph);
446 		xpt_release_ccb(ccb);
447 		cam_periph_unlock(periph);
448 		return (error);
449 	}
450 	default:
451 		break;
452 	}
453 	return (ENOTTY);
454 }
455 
456 /*
457  * Actually translate the requested transfer into one the physical driver
458  * can understand.  The transfer is described by a buf and will include
459  * only one physical transfer.
460  */
461 static void
462 ndastrategy(struct bio *bp)
463 {
464 	struct cam_periph *periph;
465 	struct nda_softc *softc;
466 
467 	periph = (struct cam_periph *)bp->bio_disk->d_drv1;
468 	softc = (struct nda_softc *)periph->softc;
469 
470 	cam_periph_lock(periph);
471 
472 	CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("ndastrategy(%p)\n", bp));
473 
474 	/*
475 	 * If the device has been made invalid, error out
476 	 */
477 	if ((periph->flags & CAM_PERIPH_INVALID) != 0) {
478 		cam_periph_unlock(periph);
479 		biofinish(bp, NULL, ENXIO);
480 		return;
481 	}
482 
483 	if (bp->bio_cmd == BIO_DELETE)
484 		softc->deletes++;
485 
486 	/*
487 	 * Place it in the queue of disk activities for this disk
488 	 */
489 	cam_iosched_queue_work(softc->cam_iosched, bp);
490 
491 	/*
492 	 * Schedule ourselves for performing the work.
493 	 */
494 	ndaschedule(periph);
495 	cam_periph_unlock(periph);
496 
497 	return;
498 }
499 
500 static int
501 ndadump(void *arg, void *virtual, off_t offset, size_t length)
502 {
503 	struct	    cam_periph *periph;
504 	struct	    nda_softc *softc;
505 	u_int	    secsize;
506 	struct ccb_nvmeio nvmeio;
507 	struct	    disk *dp;
508 	uint64_t    lba;
509 	uint32_t    count;
510 	int	    error = 0;
511 
512 	dp = arg;
513 	periph = dp->d_drv1;
514 	softc = (struct nda_softc *)periph->softc;
515 	secsize = softc->disk->d_sectorsize;
516 	lba = offset / secsize;
517 	count = length / secsize;
518 
519 	if ((periph->flags & CAM_PERIPH_INVALID) != 0)
520 		return (ENXIO);
521 
522 	/* xpt_get_ccb returns a zero'd allocation for the ccb, mimic that here */
523 	memset(&nvmeio, 0, sizeof(nvmeio));
524 	if (length > 0) {
525 		xpt_setup_ccb(&nvmeio.ccb_h, periph->path, CAM_PRIORITY_NORMAL);
526 		nvmeio.ccb_state = NDA_CCB_DUMP;
527 		nda_nvme_write(softc, &nvmeio, virtual, lba, length, count);
528 		error = cam_periph_runccb((union ccb *)&nvmeio, cam_periph_error,
529 		    0, SF_NO_RECOVERY | SF_NO_RETRY, NULL);
530 		if (error != 0)
531 			printf("Aborting dump due to I/O error %d.\n", error);
532 
533 		return (error);
534 	}
535 
536 	/* Flush */
537 	xpt_setup_ccb(&nvmeio.ccb_h, periph->path, CAM_PRIORITY_NORMAL);
538 
539 	nvmeio.ccb_state = NDA_CCB_DUMP;
540 	nda_nvme_flush(softc, &nvmeio);
541 	error = cam_periph_runccb((union ccb *)&nvmeio, cam_periph_error,
542 	    0, SF_NO_RECOVERY | SF_NO_RETRY, NULL);
543 	if (error != 0)
544 		xpt_print(periph->path, "flush cmd failed\n");
545 	return (error);
546 }
547 
548 static void
549 ndainit(void)
550 {
551 	cam_status status;
552 
553 	/*
554 	 * Install a global async callback.  This callback will
555 	 * receive async callbacks like "new device found".
556 	 */
557 	status = xpt_register_async(AC_FOUND_DEVICE, ndaasync, NULL, NULL);
558 
559 	if (status != CAM_REQ_CMP) {
560 		printf("nda: Failed to attach master async callback "
561 		       "due to status 0x%x!\n", status);
562 	} else if (nda_send_ordered) {
563 		/* Register our event handlers */
564 		if ((EVENTHANDLER_REGISTER(power_suspend, ndasuspend,
565 					   NULL, EVENTHANDLER_PRI_LAST)) == NULL)
566 		    printf("ndainit: power event registration failed!\n");
567 		if ((EVENTHANDLER_REGISTER(shutdown_post_sync, ndashutdown,
568 					   NULL, SHUTDOWN_PRI_DEFAULT)) == NULL)
569 		    printf("ndainit: shutdown event registration failed!\n");
570 	}
571 }
572 
573 /*
574  * Callback from GEOM, called when it has finished cleaning up its
575  * resources.
576  */
577 static void
578 ndadiskgonecb(struct disk *dp)
579 {
580 	struct cam_periph *periph;
581 
582 	periph = (struct cam_periph *)dp->d_drv1;
583 
584 	cam_periph_release(periph);
585 }
586 
587 static void
588 ndaoninvalidate(struct cam_periph *periph)
589 {
590 	struct nda_softc *softc;
591 
592 	softc = (struct nda_softc *)periph->softc;
593 
594 	/*
595 	 * De-register any async callbacks.
596 	 */
597 	xpt_register_async(0, ndaasync, periph, periph->path);
598 #ifdef CAM_IO_STATS
599 	softc->invalidations++;
600 #endif
601 
602 	/*
603 	 * Return all queued I/O with ENXIO. Transactions may be queued up here
604 	 * for retry (since we are called while there's other transactions
605 	 * pending). Any requests in the hardware will drain before ndacleanup
606 	 * is called.
607 	 */
608 	cam_iosched_flush(softc->cam_iosched, NULL, ENXIO);
609 
610 	/*
611 	 * Tell GEOM that we've gone away, we'll get a callback when it is
612 	 * done cleaning up its resources.
613 	 */
614 	disk_gone(softc->disk);
615 }
616 
617 static void
618 ndacleanup(struct cam_periph *periph)
619 {
620 	struct nda_softc *softc;
621 
622 	softc = (struct nda_softc *)periph->softc;
623 
624 	cam_periph_unlock(periph);
625 
626 	cam_iosched_fini(softc->cam_iosched);
627 
628 	/*
629 	 * If we can't free the sysctl tree, oh well...
630 	 */
631 	if ((softc->flags & NDA_FLAG_SCTX_INIT) != 0) {
632 #ifdef CAM_IO_STATS
633 		if (sysctl_ctx_free(&softc->sysctl_stats_ctx) != 0)
634 			xpt_print(periph->path,
635 			    "can't remove sysctl stats context\n");
636 #endif
637 		if (sysctl_ctx_free(&softc->sysctl_ctx) != 0)
638 			xpt_print(periph->path,
639 			    "can't remove sysctl context\n");
640 	}
641 
642 	disk_destroy(softc->disk);
643 	free(softc, M_DEVBUF);
644 	cam_periph_lock(periph);
645 }
646 
647 static void
648 ndaasync(void *callback_arg, uint32_t code,
649 	struct cam_path *path, void *arg)
650 {
651 	struct cam_periph *periph;
652 
653 	periph = (struct cam_periph *)callback_arg;
654 	switch (code) {
655 	case AC_FOUND_DEVICE:
656 	{
657 		struct ccb_getdev *cgd;
658 		cam_status status;
659 
660 		cgd = (struct ccb_getdev *)arg;
661 		if (cgd == NULL)
662 			break;
663 
664 		if (cgd->protocol != PROTO_NVME)
665 			break;
666 
667 		/*
668 		 * Allocate a peripheral instance for
669 		 * this device and start the probe
670 		 * process.
671 		 */
672 		status = cam_periph_alloc(ndaregister, ndaoninvalidate,
673 					  ndacleanup, ndastart,
674 					  "nda", CAM_PERIPH_BIO,
675 					  path, ndaasync,
676 					  AC_FOUND_DEVICE, cgd);
677 
678 		if (status != CAM_REQ_CMP
679 		 && status != CAM_REQ_INPROG)
680 			printf("ndaasync: Unable to attach to new device "
681 				"due to status 0x%x\n", status);
682 		break;
683 	}
684 	case AC_ADVINFO_CHANGED:
685 	{
686 		uintptr_t buftype;
687 
688 		buftype = (uintptr_t)arg;
689 		if (buftype == CDAI_TYPE_PHYS_PATH) {
690 			struct nda_softc *softc;
691 
692 			softc = periph->softc;
693 			disk_attr_changed(softc->disk, "GEOM::physpath",
694 					  M_NOWAIT);
695 		}
696 		break;
697 	}
698 	case AC_LOST_DEVICE:
699 	default:
700 		break;
701 	}
702 	cam_periph_async(periph, code, path, arg);
703 }
704 
705 static void
706 ndasysctlinit(void *context, int pending)
707 {
708 	struct cam_periph *periph;
709 	struct nda_softc *softc;
710 	char tmpstr[32], tmpstr2[16];
711 
712 	periph = (struct cam_periph *)context;
713 
714 	/* periph was held for us when this task was enqueued */
715 	if ((periph->flags & CAM_PERIPH_INVALID) != 0) {
716 		cam_periph_release(periph);
717 		return;
718 	}
719 
720 	softc = (struct nda_softc *)periph->softc;
721 	snprintf(tmpstr, sizeof(tmpstr), "CAM NDA unit %d", periph->unit_number);
722 	snprintf(tmpstr2, sizeof(tmpstr2), "%d", periph->unit_number);
723 
724 	sysctl_ctx_init(&softc->sysctl_ctx);
725 	softc->flags |= NDA_FLAG_SCTX_INIT;
726 	softc->sysctl_tree = SYSCTL_ADD_NODE_WITH_LABEL(&softc->sysctl_ctx,
727 		SYSCTL_STATIC_CHILDREN(_kern_cam_nda), OID_AUTO, tmpstr2,
728 		CTLFLAG_RD | CTLFLAG_MPSAFE, 0, tmpstr, "device_index");
729 	if (softc->sysctl_tree == NULL) {
730 		printf("ndasysctlinit: unable to allocate sysctl tree\n");
731 		cam_periph_release(periph);
732 		return;
733 	}
734 
735 	SYSCTL_ADD_INT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
736 	    OID_AUTO, "unmapped_io", CTLFLAG_RD,
737 	    &softc->unmappedio, 0, "Unmapped I/O leaf");
738 
739 	SYSCTL_ADD_QUAD(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
740 	    OID_AUTO, "deletes", CTLFLAG_RD,
741 	    &softc->deletes, "Number of BIO_DELETE requests");
742 
743 	SYSCTL_ADD_UQUAD(&softc->sysctl_ctx,
744 		SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO,
745 		"trim_count", CTLFLAG_RD, &softc->trim_count,
746 		"Total number of unmap/dsm commands sent");
747 	SYSCTL_ADD_UQUAD(&softc->sysctl_ctx,
748 		SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO,
749 		"trim_ranges", CTLFLAG_RD, &softc->trim_ranges,
750 		"Total number of ranges in unmap/dsm commands");
751 	SYSCTL_ADD_UQUAD(&softc->sysctl_ctx,
752 		SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO,
753 		"trim_lbas", CTLFLAG_RD, &softc->trim_lbas,
754 		"Total lbas in the unmap/dsm commands sent");
755 
756 	SYSCTL_ADD_INT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
757 	    OID_AUTO, "rotating", CTLFLAG_RD, &nda_rotating_media, 1,
758 	    "Rotating media");
759 
760 	SYSCTL_ADD_PROC(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
761 	    OID_AUTO, "flags", CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE,
762 	    softc, 0, ndaflagssysctl, "A",
763 	    "Flags for drive");
764 
765 #ifdef CAM_IO_STATS
766 	softc->sysctl_stats_tree = SYSCTL_ADD_NODE(&softc->sysctl_stats_ctx,
767 		SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO, "stats",
768 		CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "Statistics");
769 	if (softc->sysctl_stats_tree == NULL) {
770 		printf("ndasysctlinit: unable to allocate sysctl tree for stats\n");
771 		cam_periph_release(periph);
772 		return;
773 	}
774 	SYSCTL_ADD_INT(&softc->sysctl_stats_ctx,
775 		SYSCTL_CHILDREN(softc->sysctl_stats_tree),
776 		OID_AUTO, "timeouts", CTLFLAG_RD,
777 		&softc->timeouts, 0,
778 		"Device timeouts reported by the SIM");
779 	SYSCTL_ADD_INT(&softc->sysctl_stats_ctx,
780 		SYSCTL_CHILDREN(softc->sysctl_stats_tree),
781 		OID_AUTO, "errors", CTLFLAG_RD,
782 		&softc->errors, 0,
783 		"Transport errors reported by the SIM.");
784 	SYSCTL_ADD_INT(&softc->sysctl_stats_ctx,
785 		SYSCTL_CHILDREN(softc->sysctl_stats_tree),
786 		OID_AUTO, "pack_invalidations", CTLFLAG_RD,
787 		&softc->invalidations, 0,
788 		"Device pack invalidations.");
789 #endif
790 
791 #ifdef CAM_TEST_FAILURE
792 	SYSCTL_ADD_PROC(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
793 		OID_AUTO, "invalidate", CTLTYPE_U64 | CTLFLAG_RW | CTLFLAG_MPSAFE,
794 		periph, 0, cam_periph_invalidate_sysctl, "I",
795 		"Write 1 to invalidate the drive immediately");
796 #endif
797 
798 	cam_iosched_sysctl_init(softc->cam_iosched, &softc->sysctl_ctx,
799 	    softc->sysctl_tree);
800 
801 	cam_periph_release(periph);
802 }
803 
804 static int
805 ndaflagssysctl(SYSCTL_HANDLER_ARGS)
806 {
807 	struct sbuf sbuf;
808 	struct nda_softc *softc = arg1;
809 	int error;
810 
811 	sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
812 	if (softc->flags != 0)
813 		sbuf_printf(&sbuf, "0x%b", (unsigned)softc->flags, NDA_FLAG_STRING);
814 	else
815 		sbuf_putc(&sbuf, '0');
816 	error = sbuf_finish(&sbuf);
817 	sbuf_delete(&sbuf);
818 
819 	return (error);
820 }
821 
822 static int
823 ndagetattr(struct bio *bp)
824 {
825 	int ret;
826 	struct cam_periph *periph;
827 
828 	if (g_handleattr_int(bp, "GEOM::canspeedup", nda_enable_biospeedup))
829 		return (EJUSTRETURN);
830 
831 	periph = (struct cam_periph *)bp->bio_disk->d_drv1;
832 	cam_periph_lock(periph);
833 	ret = xpt_getattr(bp->bio_data, bp->bio_length, bp->bio_attribute,
834 	    periph->path);
835 	cam_periph_unlock(periph);
836 	if (ret == 0)
837 		bp->bio_completed = bp->bio_length;
838 	return ret;
839 }
840 
841 static cam_status
842 ndaregister(struct cam_periph *periph, void *arg)
843 {
844 	struct nda_softc *softc;
845 	struct disk *disk;
846 	struct ccb_pathinq cpi;
847 	const struct nvme_namespace_data *nsd;
848 	const struct nvme_controller_data *cd;
849 	char   announce_buf[80];
850 	uint8_t flbas_fmt, lbads, vwc_present;
851 	u_int maxio;
852 	int quirks;
853 
854 	nsd = nvme_get_identify_ns(periph);
855 	cd = nvme_get_identify_cntrl(periph);
856 
857 	softc = (struct nda_softc *)malloc(sizeof(*softc), M_DEVBUF,
858 	    M_NOWAIT | M_ZERO);
859 
860 	if (softc == NULL) {
861 		printf("ndaregister: Unable to probe new device. "
862 		    "Unable to allocate softc\n");
863 		return(CAM_REQ_CMP_ERR);
864 	}
865 
866 	/* ident_data parsing */
867 
868 	periph->softc = softc;
869 	softc->quirks = NDA_Q_NONE;
870 	xpt_path_inq(&cpi, periph->path);
871 	TASK_INIT(&softc->sysctl_task, 0, ndasysctlinit, periph);
872 
873 	/*
874 	 * The name space ID is the lun, save it for later I/O
875 	 */
876 	softc->nsid = (uint32_t)xpt_path_lun_id(periph->path);
877 
878 	/*
879 	 * Register this media as a disk
880 	 */
881 	(void)cam_periph_acquire(periph);
882 	cam_periph_unlock(periph);
883 	snprintf(announce_buf, sizeof(announce_buf),
884 	    "kern.cam.nda.%d.quirks", periph->unit_number);
885 	quirks = softc->quirks;
886 	TUNABLE_INT_FETCH(announce_buf, &quirks);
887 	softc->quirks = quirks;
888 	softc->disk = disk = disk_alloc();
889 	disk->d_rotation_rate = DISK_RR_NON_ROTATING;
890 	disk->d_open = ndaopen;
891 	disk->d_close = ndaclose;
892 	disk->d_strategy = ndastrategy;
893 	disk->d_ioctl = ndaioctl;
894 	disk->d_getattr = ndagetattr;
895 	if (cam_sim_pollable(periph->sim))
896 		disk->d_dump = ndadump;
897 	disk->d_gone = ndadiskgonecb;
898 	disk->d_name = "nda";
899 	disk->d_drv1 = periph;
900 	disk->d_unit = periph->unit_number;
901 	maxio = cpi.maxio;		/* Honor max I/O size of SIM */
902 	if (maxio == 0)
903 		maxio = DFLTPHYS;	/* traditional default */
904 	else if (maxio > maxphys)
905 		maxio = maxphys;	/* for safety */
906 	disk->d_maxsize = maxio;
907 	flbas_fmt = NVMEV(NVME_NS_DATA_FLBAS_FORMAT, nsd->flbas);
908 	lbads = NVMEV(NVME_NS_DATA_LBAF_LBADS, nsd->lbaf[flbas_fmt]);
909 	disk->d_sectorsize = 1 << lbads;
910 	disk->d_mediasize = (off_t)(disk->d_sectorsize * nsd->nsze);
911 	disk->d_delmaxsize = disk->d_mediasize;
912 	disk->d_flags = DISKFLAG_DIRECT_COMPLETION;
913 	if (nvme_ctrlr_has_dataset_mgmt(cd))
914 		disk->d_flags |= DISKFLAG_CANDELETE;
915 	vwc_present = NVMEV(NVME_CTRLR_DATA_VWC_PRESENT, cd->vwc);
916 	if (vwc_present)
917 		disk->d_flags |= DISKFLAG_CANFLUSHCACHE;
918 	if ((cpi.hba_misc & PIM_UNMAPPED) != 0) {
919 		disk->d_flags |= DISKFLAG_UNMAPPED_BIO;
920 		softc->unmappedio = 1;
921 	}
922 	/*
923 	 * d_ident and d_descr are both far bigger than the length of either
924 	 *  the serial or model number strings.
925 	 */
926 	cam_strvis_flag(disk->d_descr, cd->mn, NVME_MODEL_NUMBER_LENGTH,
927 	    sizeof(disk->d_descr), CAM_STRVIS_FLAG_NONASCII_SPC);
928 
929 	cam_strvis_flag(disk->d_ident, cd->sn, NVME_SERIAL_NUMBER_LENGTH,
930 	    sizeof(disk->d_ident), CAM_STRVIS_FLAG_NONASCII_SPC);
931 
932 	disk->d_hba_vendor = cpi.hba_vendor;
933 	disk->d_hba_device = cpi.hba_device;
934 	disk->d_hba_subvendor = cpi.hba_subvendor;
935 	disk->d_hba_subdevice = cpi.hba_subdevice;
936 	snprintf(disk->d_attachment, sizeof(disk->d_attachment),
937 	    "%s%d", cpi.dev_name, cpi.unit_number);
938 	if (NVMEV(NVME_NS_DATA_NSFEAT_NPVALID, nsd->nsfeat) != 0 &&
939 	    nsd->npwg != 0)
940 		disk->d_stripesize = ((nsd->npwg + 1) * disk->d_sectorsize);
941 	else
942 		disk->d_stripesize = nsd->noiob * disk->d_sectorsize;
943 	disk->d_stripeoffset = 0;
944 	disk->d_devstat = devstat_new_entry(periph->periph_name,
945 	    periph->unit_number, disk->d_sectorsize,
946 	    DEVSTAT_ALL_SUPPORTED,
947 	    DEVSTAT_TYPE_DIRECT | XPORT_DEVSTAT_TYPE(cpi.transport),
948 	    DEVSTAT_PRIORITY_DISK);
949 
950 	if (cam_iosched_init(&softc->cam_iosched, periph, disk,
951 	    ndaschedule) != 0) {
952 		printf("ndaregister: Unable to probe new device. "
953 		       "Unable to allocate iosched memory\n");
954 		free(softc, M_DEVBUF);
955 		return(CAM_REQ_CMP_ERR);
956 	}
957 	cam_iosched_set_sort_queue(softc->cam_iosched, 0);
958 
959 	/*
960 	 * Add alias for older nvd drives to ease transition.
961 	 */
962 	if (nda_nvd_compat)
963 		disk_add_alias(disk, "nvd");
964 
965 	cam_periph_lock(periph);
966 
967 	snprintf(announce_buf, sizeof(announce_buf),
968 		"%juMB (%ju %u byte sectors)",
969 	    (uintmax_t)((uintmax_t)disk->d_mediasize / (1024*1024)),
970 		(uintmax_t)disk->d_mediasize / disk->d_sectorsize,
971 		disk->d_sectorsize);
972 	xpt_announce_periph(periph, announce_buf);
973 	xpt_announce_quirks(periph, softc->quirks, NDA_Q_BIT_STRING);
974 
975 	/*
976 	 * Create our sysctl variables, now that we know
977 	 * we have successfully attached.
978 	 */
979 	if (cam_periph_acquire(periph) == 0)
980 		taskqueue_enqueue(taskqueue_thread, &softc->sysctl_task);
981 
982 	/*
983 	 * Register for device going away and info about the drive
984 	 * changing (though with NVMe, it can't)
985 	 */
986 	xpt_register_async(AC_LOST_DEVICE | AC_ADVINFO_CHANGED,
987 	    ndaasync, periph, periph->path);
988 
989 	softc->state = NDA_STATE_NORMAL;
990 
991 	/*
992 	 * We'll release this reference once GEOM calls us back via
993 	 * ndadiskgonecb(), telling us that our provider has been freed.
994 	 */
995 	if (cam_periph_acquire(periph) == 0)
996 		disk_create(softc->disk, DISK_VERSION);
997 
998 	cam_periph_release_locked(periph);
999 	return(CAM_REQ_CMP);
1000 }
1001 
1002 static void
1003 ndastart(struct cam_periph *periph, union ccb *start_ccb)
1004 {
1005 	struct nda_softc *softc = (struct nda_softc *)periph->softc;
1006 	struct ccb_nvmeio *nvmeio = &start_ccb->nvmeio;
1007 
1008 	CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("ndastart\n"));
1009 
1010 	switch (softc->state) {
1011 	case NDA_STATE_NORMAL:
1012 	{
1013 		struct bio *bp;
1014 
1015 		bp = cam_iosched_next_bio(softc->cam_iosched);
1016 		CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("ndastart: bio %p\n", bp));
1017 		if (bp == NULL) {
1018 			xpt_release_ccb(start_ccb);
1019 			break;
1020 		}
1021 
1022 		switch (bp->bio_cmd) {
1023 		case BIO_WRITE:
1024 			softc->flags |= NDA_FLAG_DIRTY;
1025 			/* FALLTHROUGH */
1026 		case BIO_READ:
1027 		{
1028 #ifdef CAM_TEST_FAILURE
1029 			int fail = 0;
1030 
1031 			/*
1032 			 * Support the failure ioctls.  If the command is a
1033 			 * read, and there are pending forced read errors, or
1034 			 * if a write and pending write errors, then fail this
1035 			 * operation with EIO.  This is useful for testing
1036 			 * purposes.  Also, support having every Nth read fail.
1037 			 *
1038 			 * This is a rather blunt tool.
1039 			 */
1040 			if (bp->bio_cmd == BIO_READ) {
1041 				if (softc->force_read_error) {
1042 					softc->force_read_error--;
1043 					fail = 1;
1044 				}
1045 				if (softc->periodic_read_error > 0) {
1046 					if (++softc->periodic_read_count >=
1047 					    softc->periodic_read_error) {
1048 						softc->periodic_read_count = 0;
1049 						fail = 1;
1050 					}
1051 				}
1052 			} else {
1053 				if (softc->force_write_error) {
1054 					softc->force_write_error--;
1055 					fail = 1;
1056 				}
1057 			}
1058 			if (fail) {
1059 				biofinish(bp, NULL, EIO);
1060 				xpt_release_ccb(start_ccb);
1061 				ndaschedule(periph);
1062 				return;
1063 			}
1064 #endif
1065 			KASSERT((bp->bio_flags & BIO_UNMAPPED) == 0 ||
1066 			    round_page(bp->bio_bcount + bp->bio_ma_offset) /
1067 			    PAGE_SIZE == bp->bio_ma_n,
1068 			    ("Short bio %p", bp));
1069 			nda_nvme_rw_bio(softc, &start_ccb->nvmeio, bp, bp->bio_cmd == BIO_READ ?
1070 			    NVME_OPC_READ : NVME_OPC_WRITE);
1071 			break;
1072 		}
1073 		case BIO_DELETE:
1074 		{
1075 			struct nvme_dsm_range *dsm_range, *dsm_end;
1076 			struct nda_trim_request *trim;
1077 			struct bio *bp1;
1078 			int ents;
1079 			uint32_t totalcount = 0, ranges = 0;
1080 
1081 			trim = malloc(sizeof(*trim), M_NVMEDA, M_ZERO | M_NOWAIT);
1082 			if (trim == NULL) {
1083 				/*
1084 				 * We have to drop the periph lock when
1085 				 * returning ENOMEM. g_io_deliver treats these
1086 				 * request differently and will recursively call
1087 				 * the start routine which causes us to get into
1088 				 * ndastrategy with the periph lock held,
1089 				 * leading to a panic when its acquired again.
1090 				 */
1091 				cam_periph_unlock(periph);
1092 				biofinish(bp, NULL, ENOMEM);
1093 				cam_periph_lock(periph);
1094 				xpt_release_ccb(start_ccb);
1095 				ndaschedule(periph);
1096 				return;
1097 			}
1098 			TAILQ_INIT(&trim->bps);
1099 			bp1 = bp;
1100 			ents = min(nitems(trim->dsm), nda_max_trim_entries);
1101 			ents = max(ents, 1);
1102 			dsm_range = trim->dsm;
1103 			dsm_end = dsm_range + ents;
1104 			do {
1105 				TAILQ_INSERT_TAIL(&trim->bps, bp1, bio_queue);
1106 				dsm_range->length =
1107 				    htole32(bp1->bio_bcount / softc->disk->d_sectorsize);
1108 				dsm_range->starting_lba =
1109 				    htole64(bp1->bio_offset / softc->disk->d_sectorsize);
1110 				ranges++;
1111 				totalcount += dsm_range->length;
1112 				dsm_range++;
1113 				if (dsm_range >= dsm_end)
1114 					break;
1115 				bp1 = cam_iosched_next_trim(softc->cam_iosched);
1116 				/* XXX -- Could collapse adjacent ranges, but we don't for now */
1117 				/* XXX -- Could limit based on total payload size */
1118 			} while (bp1 != NULL);
1119 			start_ccb->ccb_trim = trim;
1120 			nda_nvme_trim(softc, &start_ccb->nvmeio, trim->dsm,
1121 			    dsm_range - trim->dsm);
1122 			start_ccb->ccb_state = NDA_CCB_TRIM;
1123 			softc->trim_count++;
1124 			softc->trim_ranges += ranges;
1125 			softc->trim_lbas += totalcount;
1126 			/*
1127 			 * Note: We can have multiple TRIMs in flight, so we don't call
1128 			 * cam_iosched_submit_trim(softc->cam_iosched);
1129 			 * since that forces the I/O scheduler to only schedule one at a time.
1130 			 * On NVMe drives, this is a performance disaster.
1131 			 */
1132 			goto out;
1133 		}
1134 		case BIO_FLUSH:
1135 			nda_nvme_flush(softc, nvmeio);
1136 			break;
1137 		default:
1138 			biofinish(bp, NULL, EOPNOTSUPP);
1139 			xpt_release_ccb(start_ccb);
1140 			ndaschedule(periph);
1141 			return;
1142 		}
1143 		start_ccb->ccb_state = NDA_CCB_BUFFER_IO;
1144 		start_ccb->ccb_bp = bp;
1145 out:
1146 		start_ccb->ccb_h.flags |= CAM_UNLOCKED;
1147 		softc->outstanding_cmds++;
1148 		softc->refcount++;			/* For submission only */
1149 		cam_periph_unlock(periph);
1150 		xpt_action(start_ccb);
1151 		cam_periph_lock(periph);
1152 		softc->refcount--;			/* Submission done */
1153 
1154 		/* May have more work to do, so ensure we stay scheduled */
1155 		ndaschedule(periph);
1156 		break;
1157 		}
1158 	}
1159 }
1160 
1161 static void
1162 ndadone(struct cam_periph *periph, union ccb *done_ccb)
1163 {
1164 	struct nda_softc *softc;
1165 	struct ccb_nvmeio *nvmeio = &done_ccb->nvmeio;
1166 	struct cam_path *path;
1167 	int state;
1168 
1169 	softc = (struct nda_softc *)periph->softc;
1170 	path = done_ccb->ccb_h.path;
1171 
1172 	CAM_DEBUG(path, CAM_DEBUG_TRACE, ("ndadone\n"));
1173 
1174 	state = nvmeio->ccb_state & NDA_CCB_TYPE_MASK;
1175 	switch (state) {
1176 	case NDA_CCB_BUFFER_IO:
1177 	case NDA_CCB_TRIM:
1178 	{
1179 		int error;
1180 
1181 		cam_periph_lock(periph);
1182 		if ((done_ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_CMP) {
1183 			error = ndaerror(done_ccb, 0, 0);
1184 			if (error == ERESTART) {
1185 				/* A retry was scheduled, so just return. */
1186 				cam_periph_unlock(periph);
1187 				return;
1188 			}
1189 			if ((done_ccb->ccb_h.status & CAM_DEV_QFRZN) != 0)
1190 				cam_release_devq(path,
1191 						 /*relsim_flags*/0,
1192 						 /*reduction*/0,
1193 						 /*timeout*/0,
1194 						 /*getcount_only*/0);
1195 		} else {
1196 			if ((done_ccb->ccb_h.status & CAM_DEV_QFRZN) != 0)
1197 				panic("REQ_CMP with QFRZN");
1198 			error = 0;
1199 		}
1200 		if (state == NDA_CCB_BUFFER_IO) {
1201 			struct bio *bp;
1202 
1203 			bp = (struct bio *)done_ccb->ccb_bp;
1204 			bp->bio_error = error;
1205 			if (error != 0) {
1206 				bp->bio_resid = bp->bio_bcount;
1207 				bp->bio_flags |= BIO_ERROR;
1208 			} else {
1209 				bp->bio_resid = 0;
1210 			}
1211 			softc->outstanding_cmds--;
1212 
1213 			/*
1214 			 * We need to call cam_iosched before we call biodone so that we
1215 			 * don't measure any activity that happens in the completion
1216 			 * routine, which in the case of sendfile can be quite
1217 			 * extensive.
1218 			 */
1219 			cam_iosched_bio_complete(softc->cam_iosched, bp, done_ccb);
1220 			xpt_release_ccb(done_ccb);
1221 			ndaschedule(periph);
1222 			cam_periph_unlock(periph);
1223 			biodone(bp);
1224 		} else { /* state == NDA_CCB_TRIM */
1225 			struct nda_trim_request *trim;
1226 			struct bio *bp1, *bp2;
1227 			TAILQ_HEAD(, bio) queue;
1228 
1229 			trim = nvmeio->ccb_trim;
1230 			TAILQ_INIT(&queue);
1231 			TAILQ_CONCAT(&queue, &trim->bps, bio_queue);
1232 			free(trim, M_NVMEDA);
1233 
1234 			/*
1235 			 * Since we can have multiple trims in flight, we don't
1236 			 * need to call this here.
1237 			 * cam_iosched_trim_done(softc->cam_iosched);
1238 			 */
1239 			/*
1240 			 * The the I/O scheduler that we're finishing the I/O
1241 			 * so we can keep book. The first one we pass in the CCB
1242 			 * which has the timing information. The rest we pass in NULL
1243 			 * so we can keep proper counts.
1244 			 */
1245 			bp1 = TAILQ_FIRST(&queue);
1246 			cam_iosched_bio_complete(softc->cam_iosched, bp1, done_ccb);
1247 			xpt_release_ccb(done_ccb);
1248 			softc->outstanding_cmds--;
1249 			ndaschedule(periph);
1250 			cam_periph_unlock(periph);
1251 			while ((bp2 = TAILQ_FIRST(&queue)) != NULL) {
1252 				TAILQ_REMOVE(&queue, bp2, bio_queue);
1253 				bp2->bio_error = error;
1254 				if (error != 0) {
1255 					bp2->bio_flags |= BIO_ERROR;
1256 					bp2->bio_resid = bp1->bio_bcount;
1257 				} else
1258 					bp2->bio_resid = 0;
1259 				if (bp1 != bp2)
1260 					cam_iosched_bio_complete(softc->cam_iosched, bp2, NULL);
1261 				biodone(bp2);
1262 			}
1263 		}
1264 		return;
1265 	}
1266 	case NDA_CCB_DUMP:
1267 		/* No-op.  We're polling */
1268 		return;
1269 	case NDA_CCB_PASS:
1270 		/* NVME_PASSTHROUGH_CMD runs this CCB and releases it */
1271 		return;
1272 	default:
1273 		break;
1274 	}
1275 	xpt_release_ccb(done_ccb);
1276 }
1277 
1278 static int
1279 ndaerror(union ccb *ccb, uint32_t cam_flags, uint32_t sense_flags)
1280 {
1281 #ifdef CAM_IO_STATS
1282 	struct nda_softc *softc;
1283 	struct cam_periph *periph;
1284 
1285 	periph = xpt_path_periph(ccb->ccb_h.path);
1286 	softc = (struct nda_softc *)periph->softc;
1287 #endif
1288 
1289 	switch (ccb->ccb_h.status & CAM_STATUS_MASK) {
1290 	case CAM_CMD_TIMEOUT:
1291 #ifdef CAM_IO_STATS
1292 		softc->timeouts++;
1293 #endif
1294 		break;
1295 	case CAM_REQ_CMP_ERR:
1296 	case CAM_NVME_STATUS_ERROR:
1297 #ifdef CAM_IO_STATS
1298 		softc->errors++;
1299 #endif
1300 		break;
1301 	default:
1302 		break;
1303 	}
1304 
1305 	return(cam_periph_error(ccb, cam_flags, sense_flags));
1306 }
1307 
1308 /*
1309  * Step through all NDA peripheral drivers, and if the device is still open,
1310  * sync the disk cache to physical media.
1311  */
1312 static void
1313 ndaflush(void)
1314 {
1315 	struct cam_periph *periph;
1316 	struct nda_softc *softc;
1317 	union ccb *ccb;
1318 	int error;
1319 
1320 	CAM_PERIPH_FOREACH(periph, &ndadriver) {
1321 		softc = (struct nda_softc *)periph->softc;
1322 
1323 		if (SCHEDULER_STOPPED()) {
1324 			/*
1325 			 * If we panicked with the lock held or the periph is not
1326 			 * open, do not recurse.  Otherwise, call ndadump since
1327 			 * that avoids the sleeping cam_periph_getccb does if no
1328 			 * CCBs are available.
1329 			 */
1330 			if (!cam_periph_owned(periph) &&
1331 			    (softc->flags & NDA_FLAG_OPEN)) {
1332 				ndadump(softc->disk, NULL, 0, 0);
1333 			}
1334 			continue;
1335 		}
1336 
1337 		/*
1338 		 * We only sync the cache if the drive is still open
1339 		 */
1340 		cam_periph_lock(periph);
1341 		if ((softc->flags & NDA_FLAG_OPEN) == 0) {
1342 			cam_periph_unlock(periph);
1343 			continue;
1344 		}
1345 
1346 		ccb = cam_periph_getccb(periph, CAM_PRIORITY_NORMAL);
1347 		nda_nvme_flush(softc, &ccb->nvmeio);
1348 		error = cam_periph_runccb(ccb, ndaerror, /*cam_flags*/0,
1349 		    /*sense_flags*/ SF_NO_RECOVERY | SF_NO_RETRY,
1350 		    softc->disk->d_devstat);
1351 		if (error != 0)
1352 			xpt_print(periph->path, "Synchronize cache failed\n");
1353 		xpt_release_ccb(ccb);
1354 		cam_periph_unlock(periph);
1355 	}
1356 }
1357 
1358 static void
1359 ndashutdown(void *arg, int howto)
1360 {
1361 
1362 	if ((howto & RB_NOSYNC) != 0)
1363 		return;
1364 
1365 	ndaflush();
1366 }
1367 
1368 static void
1369 ndasuspend(void *arg, enum power_stype stype)
1370 {
1371 
1372 	ndaflush();
1373 }
1374