xref: /freebsd/sys/cam/nvme/nvme_da.c (revision 7937bfbc0ca53fe7cdd0d54414f9296e273a518e)
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/proc.h>
47 #include <sys/reboot.h>
48 #include <sys/sbuf.h>
49 #include <geom/geom.h>
50 #include <geom/geom_disk.h>
51 #endif /* _KERNEL */
52 
53 #ifndef _KERNEL
54 #include <stdio.h>
55 #include <string.h>
56 #endif /* _KERNEL */
57 
58 #include <cam/cam.h>
59 #include <cam/cam_ccb.h>
60 #include <cam/cam_periph.h>
61 #include <cam/cam_xpt_periph.h>
62 #include <cam/cam_sim.h>
63 #include <cam/cam_iosched.h>
64 
65 #include <cam/nvme/nvme_all.h>
66 
67 typedef enum {
68 	NDA_STATE_NORMAL
69 } nda_state;
70 
71 typedef enum {
72 	NDA_FLAG_OPEN		= 0x0001,
73 	NDA_FLAG_DIRTY		= 0x0002,
74 	NDA_FLAG_SCTX_INIT	= 0x0004,
75 } nda_flags;
76 #define NDA_FLAG_STRING		\
77 	"\020"			\
78 	"\001OPEN"		\
79 	"\002DIRTY"		\
80 	"\003SCTX_INIT"
81 
82 typedef enum {
83 	NDA_Q_4K   = 0x01,
84 	NDA_Q_NONE = 0x00,
85 } nda_quirks;
86 
87 #define NDA_Q_BIT_STRING	\
88 	"\020"			\
89 	"\001Bit 0"
90 
91 typedef enum {
92 	NDA_CCB_BUFFER_IO	= 0x01,
93 	NDA_CCB_DUMP            = 0x02,
94 	NDA_CCB_TRIM            = 0x03,
95 	NDA_CCB_PASS            = 0x04,
96 	NDA_CCB_TYPE_MASK	= 0x0F,
97 } nda_ccb_state;
98 
99 /* Offsets into our private area for storing information */
100 #define ccb_state	ccb_h.ppriv_field0
101 #define ccb_bp		ccb_h.ppriv_ptr1	/* For NDA_CCB_BUFFER_IO */
102 #define ccb_trim	ccb_h.ppriv_ptr1	/* For NDA_CCB_TRIM */
103 
104 struct nda_softc {
105 	struct   cam_iosched_softc *cam_iosched;
106 	int			outstanding_cmds;	/* Number of active commands */
107 	int			refcount;		/* Active xpt_action() calls */
108 	nda_state		state;
109 	nda_flags		flags;
110 	nda_quirks		quirks;
111 	int			unmappedio;
112 	quad_t			deletes;
113 	uint32_t		nsid;			/* Namespace ID for this nda device */
114 	struct disk		*disk;
115 	struct task		sysctl_task;
116 	struct sysctl_ctx_list	sysctl_ctx;
117 	struct sysctl_oid	*sysctl_tree;
118 	uint64_t		trim_count;
119 	uint64_t		trim_ranges;
120 	uint64_t		trim_lbas;
121 #ifdef CAM_TEST_FAILURE
122 	int			force_read_error;
123 	int			force_write_error;
124 	int			periodic_read_error;
125 	int			periodic_read_count;
126 #endif
127 #ifdef CAM_IO_STATS
128 	struct sysctl_ctx_list	sysctl_stats_ctx;
129 	struct sysctl_oid	*sysctl_stats_tree;
130 	u_int			timeouts;
131 	u_int			errors;
132 	u_int			invalidations;
133 #endif
134 };
135 
136 struct nda_trim_request {
137 	struct nvme_dsm_range	dsm[NVME_MAX_DSM_TRIM / sizeof(struct nvme_dsm_range)];
138 	TAILQ_HEAD(, bio) bps;
139 };
140 _Static_assert(NVME_MAX_DSM_TRIM % sizeof(struct nvme_dsm_range) == 0,
141     "NVME_MAX_DSM_TRIM must be an integral number of ranges");
142 
143 /* Need quirk table */
144 
145 static	disk_ioctl_t	ndaioctl;
146 static	disk_strategy_t	ndastrategy;
147 static	dumper_t	ndadump;
148 static	periph_init_t	ndainit;
149 static	void		ndaasync(void *callback_arg, uint32_t code,
150 				struct cam_path *path, void *arg);
151 static	void		ndasysctlinit(void *context, int pending);
152 static	int		ndaflagssysctl(SYSCTL_HANDLER_ARGS);
153 static	periph_ctor_t	ndaregister;
154 static	periph_dtor_t	ndacleanup;
155 static	periph_start_t	ndastart;
156 static	periph_oninv_t	ndaoninvalidate;
157 static	void		ndadone(struct cam_periph *periph,
158 			       union ccb *done_ccb);
159 static  int		ndaerror(union ccb *ccb, uint32_t cam_flags,
160 				uint32_t sense_flags);
161 static void		ndashutdown(void *arg, int howto);
162 static void		ndasuspend(void *arg);
163 
164 #ifndef	NDA_DEFAULT_SEND_ORDERED
165 #define	NDA_DEFAULT_SEND_ORDERED	1
166 #endif
167 #ifndef NDA_DEFAULT_TIMEOUT
168 #define NDA_DEFAULT_TIMEOUT 30	/* Timeout in seconds */
169 #endif
170 #ifndef	NDA_DEFAULT_RETRY
171 #define	NDA_DEFAULT_RETRY	4
172 #endif
173 #ifndef NDA_MAX_TRIM_ENTRIES
174 #define NDA_MAX_TRIM_ENTRIES  (NVME_MAX_DSM_TRIM / sizeof(struct nvme_dsm_range))/* Number of DSM trims to use, max 256 */
175 #endif
176 
177 static SYSCTL_NODE(_kern_cam, OID_AUTO, nda, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
178     "CAM Direct Access Disk driver");
179 
180 //static int nda_retry_count = NDA_DEFAULT_RETRY;
181 static int nda_send_ordered = NDA_DEFAULT_SEND_ORDERED;
182 static int nda_default_timeout = NDA_DEFAULT_TIMEOUT;
183 static int nda_max_trim_entries = NDA_MAX_TRIM_ENTRIES;
184 static int nda_enable_biospeedup = 1;
185 static int nda_nvd_compat = 1;
186 SYSCTL_INT(_kern_cam_nda, OID_AUTO, max_trim, CTLFLAG_RDTUN,
187     &nda_max_trim_entries, NDA_MAX_TRIM_ENTRIES,
188     "Maximum number of BIO_DELETE to send down as a DSM TRIM.");
189 SYSCTL_INT(_kern_cam_nda, OID_AUTO, enable_biospeedup, CTLFLAG_RDTUN,
190     &nda_enable_biospeedup, 0, "Enable BIO_SPEEDUP processing.");
191 SYSCTL_INT(_kern_cam_nda, OID_AUTO, nvd_compat, CTLFLAG_RDTUN,
192     &nda_nvd_compat, 1, "Enable creation of nvd aliases.");
193 
194 /*
195  * All NVMe media is non-rotational, so all nvme device instances
196  * share this to implement the sysctl.
197  */
198 static int nda_rotating_media = 0;
199 
200 static struct periph_driver ndadriver =
201 {
202 	ndainit, "nda",
203 	TAILQ_HEAD_INITIALIZER(ndadriver.units), /* generation */ 0
204 };
205 
206 PERIPHDRIVER_DECLARE(nda, ndadriver);
207 
208 static MALLOC_DEFINE(M_NVMEDA, "nvme_da", "nvme_da buffers");
209 
210 /*
211  * nice wrappers. Maybe these belong in nvme_all.c instead of
212  * here, but this is the only place that uses these. Should
213  * we ever grow another NVME periph, we should move them
214  * all there wholesale.
215  */
216 
217 static void
218 nda_nvme_flush(struct nda_softc *softc, struct ccb_nvmeio *nvmeio)
219 {
220 	cam_fill_nvmeio(nvmeio,
221 	    0,			/* retries */
222 	    ndadone,		/* cbfcnp */
223 	    CAM_DIR_NONE,	/* flags */
224 	    NULL,		/* data_ptr */
225 	    0,			/* dxfer_len */
226 	    nda_default_timeout * 1000); /* timeout 30s */
227 	nvme_ns_flush_cmd(&nvmeio->cmd, softc->nsid);
228 }
229 
230 static void
231 nda_nvme_trim(struct nda_softc *softc, struct ccb_nvmeio *nvmeio,
232     void *payload, uint32_t num_ranges)
233 {
234 	cam_fill_nvmeio(nvmeio,
235 	    0,			/* retries */
236 	    ndadone,		/* cbfcnp */
237 	    CAM_DIR_OUT,	/* flags */
238 	    payload,		/* data_ptr */
239 	    num_ranges * sizeof(struct nvme_dsm_range), /* dxfer_len */
240 	    nda_default_timeout * 1000); /* timeout 30s */
241 	nvme_ns_trim_cmd(&nvmeio->cmd, softc->nsid, num_ranges);
242 }
243 
244 static void
245 nda_nvme_write(struct nda_softc *softc, struct ccb_nvmeio *nvmeio,
246     void *payload, uint64_t lba, uint32_t len, uint32_t count)
247 {
248 	cam_fill_nvmeio(nvmeio,
249 	    0,			/* retries */
250 	    ndadone,		/* cbfcnp */
251 	    CAM_DIR_OUT,	/* flags */
252 	    payload,		/* data_ptr */
253 	    len,		/* dxfer_len */
254 	    nda_default_timeout * 1000); /* timeout 30s */
255 	nvme_ns_write_cmd(&nvmeio->cmd, softc->nsid, lba, count);
256 }
257 
258 static void
259 nda_nvme_rw_bio(struct nda_softc *softc, struct ccb_nvmeio *nvmeio,
260     struct bio *bp, uint32_t rwcmd)
261 {
262 	int flags = rwcmd == NVME_OPC_READ ? CAM_DIR_IN : CAM_DIR_OUT;
263 	void *payload;
264 	uint64_t lba;
265 	uint32_t count;
266 
267 	if (bp->bio_flags & BIO_UNMAPPED) {
268 		flags |= CAM_DATA_BIO;
269 		payload = bp;
270 	} else {
271 		payload = bp->bio_data;
272 	}
273 
274 	lba = bp->bio_pblkno;
275 	count = bp->bio_bcount / softc->disk->d_sectorsize;
276 
277 	cam_fill_nvmeio(nvmeio,
278 	    0,			/* retries */
279 	    ndadone,		/* cbfcnp */
280 	    flags,		/* flags */
281 	    payload,		/* data_ptr */
282 	    bp->bio_bcount,	/* dxfer_len */
283 	    nda_default_timeout * 1000); /* timeout 30s */
284 	nvme_ns_rw_cmd(&nvmeio->cmd, rwcmd, softc->nsid, lba, count);
285 }
286 
287 static int
288 ndaopen(struct disk *dp)
289 {
290 	struct cam_periph *periph;
291 	struct nda_softc *softc;
292 	int error;
293 
294 	periph = (struct cam_periph *)dp->d_drv1;
295 	if (cam_periph_acquire(periph) != 0) {
296 		return(ENXIO);
297 	}
298 
299 	cam_periph_lock(periph);
300 	if ((error = cam_periph_hold(periph, PRIBIO|PCATCH)) != 0) {
301 		cam_periph_unlock(periph);
302 		cam_periph_release(periph);
303 		return (error);
304 	}
305 
306 	CAM_DEBUG(periph->path, CAM_DEBUG_TRACE | CAM_DEBUG_PERIPH,
307 	    ("ndaopen\n"));
308 
309 	softc = (struct nda_softc *)periph->softc;
310 	softc->flags |= NDA_FLAG_OPEN;
311 
312 	cam_periph_unhold(periph);
313 	cam_periph_unlock(periph);
314 	return (0);
315 }
316 
317 static int
318 ndaclose(struct disk *dp)
319 {
320 	struct	cam_periph *periph;
321 	struct	nda_softc *softc;
322 	union ccb *ccb;
323 	int error;
324 
325 	periph = (struct cam_periph *)dp->d_drv1;
326 	softc = (struct nda_softc *)periph->softc;
327 	cam_periph_lock(periph);
328 
329 	CAM_DEBUG(periph->path, CAM_DEBUG_TRACE | CAM_DEBUG_PERIPH,
330 	    ("ndaclose\n"));
331 
332 	if ((softc->flags & NDA_FLAG_DIRTY) != 0 &&
333 	    (periph->flags & CAM_PERIPH_INVALID) == 0 &&
334 	    cam_periph_hold(periph, PRIBIO) == 0) {
335 		ccb = cam_periph_getccb(periph, CAM_PRIORITY_NORMAL);
336 		nda_nvme_flush(softc, &ccb->nvmeio);
337 		error = cam_periph_runccb(ccb, ndaerror, /*cam_flags*/0,
338 		    /*sense_flags*/0, softc->disk->d_devstat);
339 
340 		if (error != 0)
341 			xpt_print(periph->path, "Synchronize cache failed\n");
342 		else
343 			softc->flags &= ~NDA_FLAG_DIRTY;
344 		xpt_release_ccb(ccb);
345 		cam_periph_unhold(periph);
346 	}
347 
348 	softc->flags &= ~NDA_FLAG_OPEN;
349 
350 	while (softc->refcount != 0)
351 		cam_periph_sleep(periph, &softc->refcount, PRIBIO, "ndaclose", 1);
352 	KASSERT(softc->outstanding_cmds == 0,
353 	    ("nda %d outstanding commands", softc->outstanding_cmds));
354 	cam_periph_unlock(periph);
355 	cam_periph_release(periph);
356 	return (0);
357 }
358 
359 static void
360 ndaschedule(struct cam_periph *periph)
361 {
362 	struct nda_softc *softc = (struct nda_softc *)periph->softc;
363 
364 	if (softc->state != NDA_STATE_NORMAL)
365 		return;
366 
367 	cam_iosched_schedule(softc->cam_iosched, periph);
368 }
369 
370 static int
371 ndaioctl(struct disk *dp, u_long cmd, void *data, int fflag,
372     struct thread *td)
373 {
374 	struct cam_periph *periph;
375 
376 	periph = (struct cam_periph *)dp->d_drv1;
377 
378 	switch (cmd) {
379 	case NVME_IO_TEST:
380 	case NVME_BIO_TEST:
381 		/*
382 		 * These don't map well to the underlying CCBs, so
383 		 * they are usupported via CAM.
384 		 */
385 		return (ENOTTY);
386 	case NVME_GET_NSID:
387 	{
388 		struct nvme_get_nsid *gnsid = (struct nvme_get_nsid *)data;
389 		struct ccb_pathinq cpi;
390 
391 		xpt_path_inq(&cpi, periph->path);
392 		strncpy(gnsid->cdev, cpi.xport_specific.nvme.dev_name,
393 		    sizeof(gnsid->cdev));
394 		gnsid->nsid = cpi.xport_specific.nvme.nsid;
395 		return (0);
396 	}
397 	case NVME_PASSTHROUGH_CMD:
398 	{
399 		struct nvme_pt_command *pt;
400 		union ccb *ccb;
401 		struct cam_periph_map_info mapinfo;
402 		u_int maxmap = dp->d_maxsize;
403 		int error;
404 
405 		/*
406 		 * Create a NVME_IO CCB to do the passthrough command.
407 		 */
408 		pt = (struct nvme_pt_command *)data;
409 		ccb = xpt_alloc_ccb();
410 		xpt_setup_ccb(&ccb->ccb_h, periph->path, CAM_PRIORITY_NORMAL);
411 		ccb->ccb_state = NDA_CCB_PASS;
412 		cam_fill_nvmeio(&ccb->nvmeio,
413 		    0,			/* Retries */
414 		    ndadone,
415 		    (pt->is_read ? CAM_DIR_IN : CAM_DIR_OUT) | CAM_DATA_VADDR,
416 		    pt->buf,
417 		    pt->len,
418 		    nda_default_timeout * 1000);
419 		memcpy(&ccb->nvmeio.cmd, &pt->cmd, sizeof(pt->cmd));
420 
421 		/*
422 		 * Wire the user memory in this request for the I/O
423 		 */
424 		memset(&mapinfo, 0, sizeof(mapinfo));
425 		error = cam_periph_mapmem(ccb, &mapinfo, maxmap);
426 		if (error)
427 			goto out;
428 
429 		/*
430 		 * Lock the periph and run the command.
431 		 */
432 		cam_periph_lock(periph);
433 		cam_periph_runccb(ccb, NULL, CAM_RETRY_SELTO,
434 		    SF_RETRY_UA | SF_NO_PRINT, NULL);
435 
436 		/*
437 		 * Tear down mapping and return status.
438 		 */
439 		cam_periph_unlock(periph);
440 		error = cam_periph_unmapmem(ccb, &mapinfo);
441 		if (!cam_ccb_success(ccb))
442 			error = EIO;
443 out:
444 		cam_periph_lock(periph);
445 		xpt_release_ccb(ccb);
446 		cam_periph_unlock(periph);
447 		return (error);
448 	}
449 	default:
450 		break;
451 	}
452 	return (ENOTTY);
453 }
454 
455 /*
456  * Actually translate the requested transfer into one the physical driver
457  * can understand.  The transfer is described by a buf and will include
458  * only one physical transfer.
459  */
460 static void
461 ndastrategy(struct bio *bp)
462 {
463 	struct cam_periph *periph;
464 	struct nda_softc *softc;
465 
466 	periph = (struct cam_periph *)bp->bio_disk->d_drv1;
467 	softc = (struct nda_softc *)periph->softc;
468 
469 	cam_periph_lock(periph);
470 
471 	CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("ndastrategy(%p)\n", bp));
472 
473 	/*
474 	 * If the device has been made invalid, error out
475 	 */
476 	if ((periph->flags & CAM_PERIPH_INVALID) != 0) {
477 		cam_periph_unlock(periph);
478 		biofinish(bp, NULL, ENXIO);
479 		return;
480 	}
481 
482 	if (bp->bio_cmd == BIO_DELETE)
483 		softc->deletes++;
484 
485 	/*
486 	 * Place it in the queue of disk activities for this disk
487 	 */
488 	cam_iosched_queue_work(softc->cam_iosched, bp);
489 
490 	/*
491 	 * Schedule ourselves for performing the work.
492 	 */
493 	ndaschedule(periph);
494 	cam_periph_unlock(periph);
495 
496 	return;
497 }
498 
499 static int
500 ndadump(void *arg, void *virtual, off_t offset, size_t length)
501 {
502 	struct	    cam_periph *periph;
503 	struct	    nda_softc *softc;
504 	u_int	    secsize;
505 	struct ccb_nvmeio nvmeio;
506 	struct	    disk *dp;
507 	uint64_t    lba;
508 	uint32_t    count;
509 	int	    error = 0;
510 
511 	dp = arg;
512 	periph = dp->d_drv1;
513 	softc = (struct nda_softc *)periph->softc;
514 	secsize = softc->disk->d_sectorsize;
515 	lba = offset / secsize;
516 	count = length / secsize;
517 
518 	if ((periph->flags & CAM_PERIPH_INVALID) != 0)
519 		return (ENXIO);
520 
521 	/* xpt_get_ccb returns a zero'd allocation for the ccb, mimic that here */
522 	memset(&nvmeio, 0, sizeof(nvmeio));
523 	if (length > 0) {
524 		xpt_setup_ccb(&nvmeio.ccb_h, periph->path, CAM_PRIORITY_NORMAL);
525 		nvmeio.ccb_state = NDA_CCB_DUMP;
526 		nda_nvme_write(softc, &nvmeio, virtual, lba, length, count);
527 		error = cam_periph_runccb((union ccb *)&nvmeio, cam_periph_error,
528 		    0, SF_NO_RECOVERY | SF_NO_RETRY, NULL);
529 		if (error != 0)
530 			printf("Aborting dump due to I/O error %d.\n", error);
531 
532 		return (error);
533 	}
534 
535 	/* Flush */
536 	xpt_setup_ccb(&nvmeio.ccb_h, periph->path, CAM_PRIORITY_NORMAL);
537 
538 	nvmeio.ccb_state = NDA_CCB_DUMP;
539 	nda_nvme_flush(softc, &nvmeio);
540 	error = cam_periph_runccb((union ccb *)&nvmeio, cam_periph_error,
541 	    0, SF_NO_RECOVERY | SF_NO_RETRY, NULL);
542 	if (error != 0)
543 		xpt_print(periph->path, "flush cmd failed\n");
544 	return (error);
545 }
546 
547 static void
548 ndainit(void)
549 {
550 	cam_status status;
551 
552 	/*
553 	 * Install a global async callback.  This callback will
554 	 * receive async callbacks like "new device found".
555 	 */
556 	status = xpt_register_async(AC_FOUND_DEVICE, ndaasync, NULL, NULL);
557 
558 	if (status != CAM_REQ_CMP) {
559 		printf("nda: Failed to attach master async callback "
560 		       "due to status 0x%x!\n", status);
561 	} else if (nda_send_ordered) {
562 		/* Register our event handlers */
563 		if ((EVENTHANDLER_REGISTER(power_suspend, ndasuspend,
564 					   NULL, EVENTHANDLER_PRI_LAST)) == NULL)
565 		    printf("ndainit: power event registration failed!\n");
566 		if ((EVENTHANDLER_REGISTER(shutdown_post_sync, ndashutdown,
567 					   NULL, SHUTDOWN_PRI_DEFAULT)) == NULL)
568 		    printf("ndainit: shutdown event registration failed!\n");
569 	}
570 }
571 
572 /*
573  * Callback from GEOM, called when it has finished cleaning up its
574  * resources.
575  */
576 static void
577 ndadiskgonecb(struct disk *dp)
578 {
579 	struct cam_periph *periph;
580 
581 	periph = (struct cam_periph *)dp->d_drv1;
582 
583 	cam_periph_release(periph);
584 }
585 
586 static void
587 ndaoninvalidate(struct cam_periph *periph)
588 {
589 	struct nda_softc *softc;
590 
591 	softc = (struct nda_softc *)periph->softc;
592 
593 	/*
594 	 * De-register any async callbacks.
595 	 */
596 	xpt_register_async(0, ndaasync, periph, periph->path);
597 #ifdef CAM_IO_STATS
598 	softc->invalidations++;
599 #endif
600 
601 	/*
602 	 * Return all queued I/O with ENXIO. Transactions may be queued up here
603 	 * for retry (since we are called while there's other transactions
604 	 * pending). Any requests in the hardware will drain before ndacleanup
605 	 * is called.
606 	 */
607 	cam_iosched_flush(softc->cam_iosched, NULL, ENXIO);
608 
609 	/*
610 	 * Tell GEOM that we've gone away, we'll get a callback when it is
611 	 * done cleaning up its resources.
612 	 */
613 	disk_gone(softc->disk);
614 }
615 
616 static void
617 ndacleanup(struct cam_periph *periph)
618 {
619 	struct nda_softc *softc;
620 
621 	softc = (struct nda_softc *)periph->softc;
622 
623 	cam_periph_unlock(periph);
624 
625 	cam_iosched_fini(softc->cam_iosched);
626 
627 	/*
628 	 * If we can't free the sysctl tree, oh well...
629 	 */
630 	if ((softc->flags & NDA_FLAG_SCTX_INIT) != 0) {
631 #ifdef CAM_IO_STATS
632 		if (sysctl_ctx_free(&softc->sysctl_stats_ctx) != 0)
633 			xpt_print(periph->path,
634 			    "can't remove sysctl stats context\n");
635 #endif
636 		if (sysctl_ctx_free(&softc->sysctl_ctx) != 0)
637 			xpt_print(periph->path,
638 			    "can't remove sysctl context\n");
639 	}
640 
641 	disk_destroy(softc->disk);
642 	free(softc, M_DEVBUF);
643 	cam_periph_lock(periph);
644 }
645 
646 static void
647 ndaasync(void *callback_arg, uint32_t code,
648 	struct cam_path *path, void *arg)
649 {
650 	struct cam_periph *periph;
651 
652 	periph = (struct cam_periph *)callback_arg;
653 	switch (code) {
654 	case AC_FOUND_DEVICE:
655 	{
656 		struct ccb_getdev *cgd;
657 		cam_status status;
658 
659 		cgd = (struct ccb_getdev *)arg;
660 		if (cgd == NULL)
661 			break;
662 
663 		if (cgd->protocol != PROTO_NVME)
664 			break;
665 
666 		/*
667 		 * Allocate a peripheral instance for
668 		 * this device and start the probe
669 		 * process.
670 		 */
671 		status = cam_periph_alloc(ndaregister, ndaoninvalidate,
672 					  ndacleanup, ndastart,
673 					  "nda", CAM_PERIPH_BIO,
674 					  path, ndaasync,
675 					  AC_FOUND_DEVICE, cgd);
676 
677 		if (status != CAM_REQ_CMP
678 		 && status != CAM_REQ_INPROG)
679 			printf("ndaasync: Unable to attach to new device "
680 				"due to status 0x%x\n", status);
681 		break;
682 	}
683 	case AC_ADVINFO_CHANGED:
684 	{
685 		uintptr_t buftype;
686 
687 		buftype = (uintptr_t)arg;
688 		if (buftype == CDAI_TYPE_PHYS_PATH) {
689 			struct nda_softc *softc;
690 
691 			softc = periph->softc;
692 			disk_attr_changed(softc->disk, "GEOM::physpath",
693 					  M_NOWAIT);
694 		}
695 		break;
696 	}
697 	case AC_LOST_DEVICE:
698 	default:
699 		break;
700 	}
701 	cam_periph_async(periph, code, path, arg);
702 }
703 
704 static void
705 ndasysctlinit(void *context, int pending)
706 {
707 	struct cam_periph *periph;
708 	struct nda_softc *softc;
709 	char tmpstr[32], tmpstr2[16];
710 
711 	periph = (struct cam_periph *)context;
712 
713 	/* periph was held for us when this task was enqueued */
714 	if ((periph->flags & CAM_PERIPH_INVALID) != 0) {
715 		cam_periph_release(periph);
716 		return;
717 	}
718 
719 	softc = (struct nda_softc *)periph->softc;
720 	snprintf(tmpstr, sizeof(tmpstr), "CAM NDA unit %d", periph->unit_number);
721 	snprintf(tmpstr2, sizeof(tmpstr2), "%d", periph->unit_number);
722 
723 	sysctl_ctx_init(&softc->sysctl_ctx);
724 	softc->flags |= NDA_FLAG_SCTX_INIT;
725 	softc->sysctl_tree = SYSCTL_ADD_NODE_WITH_LABEL(&softc->sysctl_ctx,
726 		SYSCTL_STATIC_CHILDREN(_kern_cam_nda), OID_AUTO, tmpstr2,
727 		CTLFLAG_RD | CTLFLAG_MPSAFE, 0, tmpstr, "device_index");
728 	if (softc->sysctl_tree == NULL) {
729 		printf("ndasysctlinit: unable to allocate sysctl tree\n");
730 		cam_periph_release(periph);
731 		return;
732 	}
733 
734 	SYSCTL_ADD_INT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
735 	    OID_AUTO, "unmapped_io", CTLFLAG_RD,
736 	    &softc->unmappedio, 0, "Unmapped I/O leaf");
737 
738 	SYSCTL_ADD_QUAD(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
739 	    OID_AUTO, "deletes", CTLFLAG_RD,
740 	    &softc->deletes, "Number of BIO_DELETE requests");
741 
742 	SYSCTL_ADD_UQUAD(&softc->sysctl_ctx,
743 		SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO,
744 		"trim_count", CTLFLAG_RD, &softc->trim_count,
745 		"Total number of unmap/dsm commands sent");
746 	SYSCTL_ADD_UQUAD(&softc->sysctl_ctx,
747 		SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO,
748 		"trim_ranges", CTLFLAG_RD, &softc->trim_ranges,
749 		"Total number of ranges in unmap/dsm commands");
750 	SYSCTL_ADD_UQUAD(&softc->sysctl_ctx,
751 		SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO,
752 		"trim_lbas", CTLFLAG_RD, &softc->trim_lbas,
753 		"Total lbas in the unmap/dsm commands sent");
754 
755 	SYSCTL_ADD_INT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
756 	    OID_AUTO, "rotating", CTLFLAG_RD, &nda_rotating_media, 1,
757 	    "Rotating media");
758 
759 	SYSCTL_ADD_PROC(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
760 	    OID_AUTO, "flags", CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE,
761 	    softc, 0, ndaflagssysctl, "A",
762 	    "Flags for drive");
763 
764 #ifdef CAM_IO_STATS
765 	softc->sysctl_stats_tree = SYSCTL_ADD_NODE(&softc->sysctl_stats_ctx,
766 		SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO, "stats",
767 		CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "Statistics");
768 	if (softc->sysctl_stats_tree == NULL) {
769 		printf("ndasysctlinit: unable to allocate sysctl tree for stats\n");
770 		cam_periph_release(periph);
771 		return;
772 	}
773 	SYSCTL_ADD_INT(&softc->sysctl_stats_ctx,
774 		SYSCTL_CHILDREN(softc->sysctl_stats_tree),
775 		OID_AUTO, "timeouts", CTLFLAG_RD,
776 		&softc->timeouts, 0,
777 		"Device timeouts reported by the SIM");
778 	SYSCTL_ADD_INT(&softc->sysctl_stats_ctx,
779 		SYSCTL_CHILDREN(softc->sysctl_stats_tree),
780 		OID_AUTO, "errors", CTLFLAG_RD,
781 		&softc->errors, 0,
782 		"Transport errors reported by the SIM.");
783 	SYSCTL_ADD_INT(&softc->sysctl_stats_ctx,
784 		SYSCTL_CHILDREN(softc->sysctl_stats_tree),
785 		OID_AUTO, "pack_invalidations", CTLFLAG_RD,
786 		&softc->invalidations, 0,
787 		"Device pack invalidations.");
788 #endif
789 
790 #ifdef CAM_TEST_FAILURE
791 	SYSCTL_ADD_PROC(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
792 		OID_AUTO, "invalidate", CTLTYPE_U64 | CTLFLAG_RW | CTLFLAG_MPSAFE,
793 		periph, 0, cam_periph_invalidate_sysctl, "I",
794 		"Write 1 to invalidate the drive immediately");
795 #endif
796 
797 	cam_iosched_sysctl_init(softc->cam_iosched, &softc->sysctl_ctx,
798 	    softc->sysctl_tree);
799 
800 	cam_periph_release(periph);
801 }
802 
803 static int
804 ndaflagssysctl(SYSCTL_HANDLER_ARGS)
805 {
806 	struct sbuf sbuf;
807 	struct nda_softc *softc = arg1;
808 	int error;
809 
810 	sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
811 	if (softc->flags != 0)
812 		sbuf_printf(&sbuf, "0x%b", (unsigned)softc->flags, NDA_FLAG_STRING);
813 	else
814 		sbuf_putc(&sbuf, '0');
815 	error = sbuf_finish(&sbuf);
816 	sbuf_delete(&sbuf);
817 
818 	return (error);
819 }
820 
821 static int
822 ndagetattr(struct bio *bp)
823 {
824 	int ret;
825 	struct cam_periph *periph;
826 
827 	if (g_handleattr_int(bp, "GEOM::canspeedup", nda_enable_biospeedup))
828 		return (EJUSTRETURN);
829 
830 	periph = (struct cam_periph *)bp->bio_disk->d_drv1;
831 	cam_periph_lock(periph);
832 	ret = xpt_getattr(bp->bio_data, bp->bio_length, bp->bio_attribute,
833 	    periph->path);
834 	cam_periph_unlock(periph);
835 	if (ret == 0)
836 		bp->bio_completed = bp->bio_length;
837 	return ret;
838 }
839 
840 static cam_status
841 ndaregister(struct cam_periph *periph, void *arg)
842 {
843 	struct nda_softc *softc;
844 	struct disk *disk;
845 	struct ccb_pathinq cpi;
846 	const struct nvme_namespace_data *nsd;
847 	const struct nvme_controller_data *cd;
848 	char   announce_buf[80];
849 	uint8_t flbas_fmt, lbads, vwc_present;
850 	u_int maxio;
851 	int quirks;
852 
853 	nsd = nvme_get_identify_ns(periph);
854 	cd = nvme_get_identify_cntrl(periph);
855 
856 	softc = (struct nda_softc *)malloc(sizeof(*softc), M_DEVBUF,
857 	    M_NOWAIT | M_ZERO);
858 
859 	if (softc == NULL) {
860 		printf("ndaregister: Unable to probe new device. "
861 		    "Unable to allocate softc\n");
862 		return(CAM_REQ_CMP_ERR);
863 	}
864 
865 	/* ident_data parsing */
866 
867 	periph->softc = softc;
868 	softc->quirks = NDA_Q_NONE;
869 	xpt_path_inq(&cpi, periph->path);
870 	TASK_INIT(&softc->sysctl_task, 0, ndasysctlinit, periph);
871 
872 	/*
873 	 * The name space ID is the lun, save it for later I/O
874 	 */
875 	softc->nsid = (uint32_t)xpt_path_lun_id(periph->path);
876 
877 	/*
878 	 * Register this media as a disk
879 	 */
880 	(void)cam_periph_acquire(periph);
881 	cam_periph_unlock(periph);
882 	snprintf(announce_buf, sizeof(announce_buf),
883 	    "kern.cam.nda.%d.quirks", periph->unit_number);
884 	quirks = softc->quirks;
885 	TUNABLE_INT_FETCH(announce_buf, &quirks);
886 	softc->quirks = quirks;
887 	softc->disk = disk = disk_alloc();
888 	disk->d_rotation_rate = DISK_RR_NON_ROTATING;
889 	disk->d_open = ndaopen;
890 	disk->d_close = ndaclose;
891 	disk->d_strategy = ndastrategy;
892 	disk->d_ioctl = ndaioctl;
893 	disk->d_getattr = ndagetattr;
894 	if (cam_sim_pollable(periph->sim))
895 		disk->d_dump = ndadump;
896 	disk->d_gone = ndadiskgonecb;
897 	disk->d_name = "nda";
898 	disk->d_drv1 = periph;
899 	disk->d_unit = periph->unit_number;
900 	maxio = cpi.maxio;		/* Honor max I/O size of SIM */
901 	if (maxio == 0)
902 		maxio = DFLTPHYS;	/* traditional default */
903 	else if (maxio > maxphys)
904 		maxio = maxphys;	/* for safety */
905 	disk->d_maxsize = maxio;
906 	flbas_fmt = NVMEV(NVME_NS_DATA_FLBAS_FORMAT, nsd->flbas);
907 	lbads = NVMEV(NVME_NS_DATA_LBAF_LBADS, nsd->lbaf[flbas_fmt]);
908 	disk->d_sectorsize = 1 << lbads;
909 	disk->d_mediasize = (off_t)(disk->d_sectorsize * nsd->nsze);
910 	disk->d_delmaxsize = disk->d_mediasize;
911 	disk->d_flags = DISKFLAG_DIRECT_COMPLETION;
912 	if (nvme_ctrlr_has_dataset_mgmt(cd))
913 		disk->d_flags |= DISKFLAG_CANDELETE;
914 	vwc_present = NVMEV(NVME_CTRLR_DATA_VWC_PRESENT, cd->vwc);
915 	if (vwc_present)
916 		disk->d_flags |= DISKFLAG_CANFLUSHCACHE;
917 	if ((cpi.hba_misc & PIM_UNMAPPED) != 0) {
918 		disk->d_flags |= DISKFLAG_UNMAPPED_BIO;
919 		softc->unmappedio = 1;
920 	}
921 	/*
922 	 * d_ident and d_descr are both far bigger than the length of either
923 	 *  the serial or model number strings.
924 	 */
925 	cam_strvis_flag(disk->d_descr, cd->mn, NVME_MODEL_NUMBER_LENGTH,
926 	    sizeof(disk->d_descr), CAM_STRVIS_FLAG_NONASCII_SPC);
927 
928 	cam_strvis_flag(disk->d_ident, cd->sn, NVME_SERIAL_NUMBER_LENGTH,
929 	    sizeof(disk->d_ident), CAM_STRVIS_FLAG_NONASCII_SPC);
930 
931 	disk->d_hba_vendor = cpi.hba_vendor;
932 	disk->d_hba_device = cpi.hba_device;
933 	disk->d_hba_subvendor = cpi.hba_subvendor;
934 	disk->d_hba_subdevice = cpi.hba_subdevice;
935 	snprintf(disk->d_attachment, sizeof(disk->d_attachment),
936 	    "%s%d", cpi.dev_name, cpi.unit_number);
937 	if (NVMEV(NVME_NS_DATA_NSFEAT_NPVALID, nsd->nsfeat) != 0 &&
938 	    nsd->npwg != 0)
939 		disk->d_stripesize = ((nsd->npwg + 1) * disk->d_sectorsize);
940 	else
941 		disk->d_stripesize = nsd->noiob * disk->d_sectorsize;
942 	disk->d_stripeoffset = 0;
943 	disk->d_devstat = devstat_new_entry(periph->periph_name,
944 	    periph->unit_number, disk->d_sectorsize,
945 	    DEVSTAT_ALL_SUPPORTED,
946 	    DEVSTAT_TYPE_DIRECT | XPORT_DEVSTAT_TYPE(cpi.transport),
947 	    DEVSTAT_PRIORITY_DISK);
948 
949 	if (cam_iosched_init(&softc->cam_iosched, periph, disk,
950 	    ndaschedule) != 0) {
951 		printf("ndaregister: Unable to probe new device. "
952 		       "Unable to allocate iosched memory\n");
953 		free(softc, M_DEVBUF);
954 		return(CAM_REQ_CMP_ERR);
955 	}
956 	cam_iosched_set_sort_queue(softc->cam_iosched, 0);
957 
958 	/*
959 	 * Add alias for older nvd drives to ease transition.
960 	 */
961 	if (nda_nvd_compat)
962 		disk_add_alias(disk, "nvd");
963 
964 	cam_periph_lock(periph);
965 
966 	snprintf(announce_buf, sizeof(announce_buf),
967 		"%juMB (%ju %u byte sectors)",
968 	    (uintmax_t)((uintmax_t)disk->d_mediasize / (1024*1024)),
969 		(uintmax_t)disk->d_mediasize / disk->d_sectorsize,
970 		disk->d_sectorsize);
971 	xpt_announce_periph(periph, announce_buf);
972 	xpt_announce_quirks(periph, softc->quirks, NDA_Q_BIT_STRING);
973 
974 	/*
975 	 * Create our sysctl variables, now that we know
976 	 * we have successfully attached.
977 	 */
978 	if (cam_periph_acquire(periph) == 0)
979 		taskqueue_enqueue(taskqueue_thread, &softc->sysctl_task);
980 
981 	/*
982 	 * Register for device going away and info about the drive
983 	 * changing (though with NVMe, it can't)
984 	 */
985 	xpt_register_async(AC_LOST_DEVICE | AC_ADVINFO_CHANGED,
986 	    ndaasync, periph, periph->path);
987 
988 	softc->state = NDA_STATE_NORMAL;
989 
990 	/*
991 	 * We'll release this reference once GEOM calls us back via
992 	 * ndadiskgonecb(), telling us that our provider has been freed.
993 	 */
994 	if (cam_periph_acquire(periph) == 0)
995 		disk_create(softc->disk, DISK_VERSION);
996 
997 	cam_periph_release_locked(periph);
998 	return(CAM_REQ_CMP);
999 }
1000 
1001 static void
1002 ndastart(struct cam_periph *periph, union ccb *start_ccb)
1003 {
1004 	struct nda_softc *softc = (struct nda_softc *)periph->softc;
1005 	struct ccb_nvmeio *nvmeio = &start_ccb->nvmeio;
1006 
1007 	CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("ndastart\n"));
1008 
1009 	switch (softc->state) {
1010 	case NDA_STATE_NORMAL:
1011 	{
1012 		struct bio *bp;
1013 
1014 		bp = cam_iosched_next_bio(softc->cam_iosched);
1015 		CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("ndastart: bio %p\n", bp));
1016 		if (bp == NULL) {
1017 			xpt_release_ccb(start_ccb);
1018 			break;
1019 		}
1020 
1021 		switch (bp->bio_cmd) {
1022 		case BIO_WRITE:
1023 			softc->flags |= NDA_FLAG_DIRTY;
1024 			/* FALLTHROUGH */
1025 		case BIO_READ:
1026 		{
1027 #ifdef CAM_TEST_FAILURE
1028 			int fail = 0;
1029 
1030 			/*
1031 			 * Support the failure ioctls.  If the command is a
1032 			 * read, and there are pending forced read errors, or
1033 			 * if a write and pending write errors, then fail this
1034 			 * operation with EIO.  This is useful for testing
1035 			 * purposes.  Also, support having every Nth read fail.
1036 			 *
1037 			 * This is a rather blunt tool.
1038 			 */
1039 			if (bp->bio_cmd == BIO_READ) {
1040 				if (softc->force_read_error) {
1041 					softc->force_read_error--;
1042 					fail = 1;
1043 				}
1044 				if (softc->periodic_read_error > 0) {
1045 					if (++softc->periodic_read_count >=
1046 					    softc->periodic_read_error) {
1047 						softc->periodic_read_count = 0;
1048 						fail = 1;
1049 					}
1050 				}
1051 			} else {
1052 				if (softc->force_write_error) {
1053 					softc->force_write_error--;
1054 					fail = 1;
1055 				}
1056 			}
1057 			if (fail) {
1058 				biofinish(bp, NULL, EIO);
1059 				xpt_release_ccb(start_ccb);
1060 				ndaschedule(periph);
1061 				return;
1062 			}
1063 #endif
1064 			KASSERT((bp->bio_flags & BIO_UNMAPPED) == 0 ||
1065 			    round_page(bp->bio_bcount + bp->bio_ma_offset) /
1066 			    PAGE_SIZE == bp->bio_ma_n,
1067 			    ("Short bio %p", bp));
1068 			nda_nvme_rw_bio(softc, &start_ccb->nvmeio, bp, bp->bio_cmd == BIO_READ ?
1069 			    NVME_OPC_READ : NVME_OPC_WRITE);
1070 			break;
1071 		}
1072 		case BIO_DELETE:
1073 		{
1074 			struct nvme_dsm_range *dsm_range, *dsm_end;
1075 			struct nda_trim_request *trim;
1076 			struct bio *bp1;
1077 			int ents;
1078 			uint32_t totalcount = 0, ranges = 0;
1079 
1080 			trim = malloc(sizeof(*trim), M_NVMEDA, M_ZERO | M_NOWAIT);
1081 			if (trim == NULL) {
1082 				/*
1083 				 * We have to drop the periph lock when
1084 				 * returning ENOMEM. g_io_deliver treats these
1085 				 * request differently and will recursively call
1086 				 * the start routine which causes us to get into
1087 				 * ndastrategy with the periph lock held,
1088 				 * leading to a panic when its acquired again.
1089 				 */
1090 				cam_periph_unlock(periph);
1091 				biofinish(bp, NULL, ENOMEM);
1092 				cam_periph_lock(periph);
1093 				xpt_release_ccb(start_ccb);
1094 				ndaschedule(periph);
1095 				return;
1096 			}
1097 			TAILQ_INIT(&trim->bps);
1098 			bp1 = bp;
1099 			ents = min(nitems(trim->dsm), nda_max_trim_entries);
1100 			ents = max(ents, 1);
1101 			dsm_range = trim->dsm;
1102 			dsm_end = dsm_range + ents;
1103 			do {
1104 				TAILQ_INSERT_TAIL(&trim->bps, bp1, bio_queue);
1105 				dsm_range->length =
1106 				    htole32(bp1->bio_bcount / softc->disk->d_sectorsize);
1107 				dsm_range->starting_lba =
1108 				    htole64(bp1->bio_offset / softc->disk->d_sectorsize);
1109 				ranges++;
1110 				totalcount += dsm_range->length;
1111 				dsm_range++;
1112 				if (dsm_range >= dsm_end)
1113 					break;
1114 				bp1 = cam_iosched_next_trim(softc->cam_iosched);
1115 				/* XXX -- Could collapse adjacent ranges, but we don't for now */
1116 				/* XXX -- Could limit based on total payload size */
1117 			} while (bp1 != NULL);
1118 			start_ccb->ccb_trim = trim;
1119 			nda_nvme_trim(softc, &start_ccb->nvmeio, trim->dsm,
1120 			    dsm_range - trim->dsm);
1121 			start_ccb->ccb_state = NDA_CCB_TRIM;
1122 			softc->trim_count++;
1123 			softc->trim_ranges += ranges;
1124 			softc->trim_lbas += totalcount;
1125 			/*
1126 			 * Note: We can have multiple TRIMs in flight, so we don't call
1127 			 * cam_iosched_submit_trim(softc->cam_iosched);
1128 			 * since that forces the I/O scheduler to only schedule one at a time.
1129 			 * On NVMe drives, this is a performance disaster.
1130 			 */
1131 			goto out;
1132 		}
1133 		case BIO_FLUSH:
1134 			nda_nvme_flush(softc, nvmeio);
1135 			break;
1136 		default:
1137 			biofinish(bp, NULL, EOPNOTSUPP);
1138 			xpt_release_ccb(start_ccb);
1139 			ndaschedule(periph);
1140 			return;
1141 		}
1142 		start_ccb->ccb_state = NDA_CCB_BUFFER_IO;
1143 		start_ccb->ccb_bp = bp;
1144 out:
1145 		start_ccb->ccb_h.flags |= CAM_UNLOCKED;
1146 		softc->outstanding_cmds++;
1147 		softc->refcount++;			/* For submission only */
1148 		cam_periph_unlock(periph);
1149 		xpt_action(start_ccb);
1150 		cam_periph_lock(periph);
1151 		softc->refcount--;			/* Submission done */
1152 
1153 		/* May have more work to do, so ensure we stay scheduled */
1154 		ndaschedule(periph);
1155 		break;
1156 		}
1157 	}
1158 }
1159 
1160 static void
1161 ndadone(struct cam_periph *periph, union ccb *done_ccb)
1162 {
1163 	struct nda_softc *softc;
1164 	struct ccb_nvmeio *nvmeio = &done_ccb->nvmeio;
1165 	struct cam_path *path;
1166 	int state;
1167 
1168 	softc = (struct nda_softc *)periph->softc;
1169 	path = done_ccb->ccb_h.path;
1170 
1171 	CAM_DEBUG(path, CAM_DEBUG_TRACE, ("ndadone\n"));
1172 
1173 	state = nvmeio->ccb_state & NDA_CCB_TYPE_MASK;
1174 	switch (state) {
1175 	case NDA_CCB_BUFFER_IO:
1176 	case NDA_CCB_TRIM:
1177 	{
1178 		int error;
1179 
1180 		cam_periph_lock(periph);
1181 		if ((done_ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_CMP) {
1182 			error = ndaerror(done_ccb, 0, 0);
1183 			if (error == ERESTART) {
1184 				/* A retry was scheduled, so just return. */
1185 				cam_periph_unlock(periph);
1186 				return;
1187 			}
1188 			if ((done_ccb->ccb_h.status & CAM_DEV_QFRZN) != 0)
1189 				cam_release_devq(path,
1190 						 /*relsim_flags*/0,
1191 						 /*reduction*/0,
1192 						 /*timeout*/0,
1193 						 /*getcount_only*/0);
1194 		} else {
1195 			if ((done_ccb->ccb_h.status & CAM_DEV_QFRZN) != 0)
1196 				panic("REQ_CMP with QFRZN");
1197 			error = 0;
1198 		}
1199 		if (state == NDA_CCB_BUFFER_IO) {
1200 			struct bio *bp;
1201 
1202 			bp = (struct bio *)done_ccb->ccb_bp;
1203 			bp->bio_error = error;
1204 			if (error != 0) {
1205 				bp->bio_resid = bp->bio_bcount;
1206 				bp->bio_flags |= BIO_ERROR;
1207 			} else {
1208 				bp->bio_resid = 0;
1209 			}
1210 			softc->outstanding_cmds--;
1211 
1212 			/*
1213 			 * We need to call cam_iosched before we call biodone so that we
1214 			 * don't measure any activity that happens in the completion
1215 			 * routine, which in the case of sendfile can be quite
1216 			 * extensive.
1217 			 */
1218 			cam_iosched_bio_complete(softc->cam_iosched, bp, done_ccb);
1219 			xpt_release_ccb(done_ccb);
1220 			ndaschedule(periph);
1221 			cam_periph_unlock(periph);
1222 			biodone(bp);
1223 		} else { /* state == NDA_CCB_TRIM */
1224 			struct nda_trim_request *trim;
1225 			struct bio *bp1, *bp2;
1226 			TAILQ_HEAD(, bio) queue;
1227 
1228 			trim = nvmeio->ccb_trim;
1229 			TAILQ_INIT(&queue);
1230 			TAILQ_CONCAT(&queue, &trim->bps, bio_queue);
1231 			free(trim, M_NVMEDA);
1232 
1233 			/*
1234 			 * Since we can have multiple trims in flight, we don't
1235 			 * need to call this here.
1236 			 * cam_iosched_trim_done(softc->cam_iosched);
1237 			 */
1238 			/*
1239 			 * The the I/O scheduler that we're finishing the I/O
1240 			 * so we can keep book. The first one we pass in the CCB
1241 			 * which has the timing information. The rest we pass in NULL
1242 			 * so we can keep proper counts.
1243 			 */
1244 			bp1 = TAILQ_FIRST(&queue);
1245 			cam_iosched_bio_complete(softc->cam_iosched, bp1, done_ccb);
1246 			xpt_release_ccb(done_ccb);
1247 			softc->outstanding_cmds--;
1248 			ndaschedule(periph);
1249 			cam_periph_unlock(periph);
1250 			while ((bp2 = TAILQ_FIRST(&queue)) != NULL) {
1251 				TAILQ_REMOVE(&queue, bp2, bio_queue);
1252 				bp2->bio_error = error;
1253 				if (error != 0) {
1254 					bp2->bio_flags |= BIO_ERROR;
1255 					bp2->bio_resid = bp1->bio_bcount;
1256 				} else
1257 					bp2->bio_resid = 0;
1258 				if (bp1 != bp2)
1259 					cam_iosched_bio_complete(softc->cam_iosched, bp2, NULL);
1260 				biodone(bp2);
1261 			}
1262 		}
1263 		return;
1264 	}
1265 	case NDA_CCB_DUMP:
1266 		/* No-op.  We're polling */
1267 		return;
1268 	case NDA_CCB_PASS:
1269 		/* NVME_PASSTHROUGH_CMD runs this CCB and releases it */
1270 		return;
1271 	default:
1272 		break;
1273 	}
1274 	xpt_release_ccb(done_ccb);
1275 }
1276 
1277 static int
1278 ndaerror(union ccb *ccb, uint32_t cam_flags, uint32_t sense_flags)
1279 {
1280 #ifdef CAM_IO_STATS
1281 	struct nda_softc *softc;
1282 	struct cam_periph *periph;
1283 
1284 	periph = xpt_path_periph(ccb->ccb_h.path);
1285 	softc = (struct nda_softc *)periph->softc;
1286 #endif
1287 
1288 	switch (ccb->ccb_h.status & CAM_STATUS_MASK) {
1289 	case CAM_CMD_TIMEOUT:
1290 #ifdef CAM_IO_STATS
1291 		softc->timeouts++;
1292 #endif
1293 		break;
1294 	case CAM_REQ_CMP_ERR:
1295 	case CAM_NVME_STATUS_ERROR:
1296 #ifdef CAM_IO_STATS
1297 		softc->errors++;
1298 #endif
1299 		break;
1300 	default:
1301 		break;
1302 	}
1303 
1304 	return(cam_periph_error(ccb, cam_flags, sense_flags));
1305 }
1306 
1307 /*
1308  * Step through all NDA peripheral drivers, and if the device is still open,
1309  * sync the disk cache to physical media.
1310  */
1311 static void
1312 ndaflush(void)
1313 {
1314 	struct cam_periph *periph;
1315 	struct nda_softc *softc;
1316 	union ccb *ccb;
1317 	int error;
1318 
1319 	CAM_PERIPH_FOREACH(periph, &ndadriver) {
1320 		softc = (struct nda_softc *)periph->softc;
1321 
1322 		if (SCHEDULER_STOPPED()) {
1323 			/*
1324 			 * If we panicked with the lock held or the periph is not
1325 			 * open, do not recurse.  Otherwise, call ndadump since
1326 			 * that avoids the sleeping cam_periph_getccb does if no
1327 			 * CCBs are available.
1328 			 */
1329 			if (!cam_periph_owned(periph) &&
1330 			    (softc->flags & NDA_FLAG_OPEN)) {
1331 				ndadump(softc->disk, NULL, 0, 0);
1332 			}
1333 			continue;
1334 		}
1335 
1336 		/*
1337 		 * We only sync the cache if the drive is still open
1338 		 */
1339 		cam_periph_lock(periph);
1340 		if ((softc->flags & NDA_FLAG_OPEN) == 0) {
1341 			cam_periph_unlock(periph);
1342 			continue;
1343 		}
1344 
1345 		ccb = cam_periph_getccb(periph, CAM_PRIORITY_NORMAL);
1346 		nda_nvme_flush(softc, &ccb->nvmeio);
1347 		error = cam_periph_runccb(ccb, ndaerror, /*cam_flags*/0,
1348 		    /*sense_flags*/ SF_NO_RECOVERY | SF_NO_RETRY,
1349 		    softc->disk->d_devstat);
1350 		if (error != 0)
1351 			xpt_print(periph->path, "Synchronize cache failed\n");
1352 		xpt_release_ccb(ccb);
1353 		cam_periph_unlock(periph);
1354 	}
1355 }
1356 
1357 static void
1358 ndashutdown(void *arg, int howto)
1359 {
1360 
1361 	if ((howto & RB_NOSYNC) != 0)
1362 		return;
1363 
1364 	ndaflush();
1365 }
1366 
1367 static void
1368 ndasuspend(void *arg)
1369 {
1370 
1371 	ndaflush();
1372 }
1373