xref: /freebsd/sys/cam/nvme/nvme_da.c (revision 9978ade4980452e3699285c01a4b9c4b5adc68e4)
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 <sys/sbuf.h>
52 #include <geom/geom.h>
53 #include <geom/geom_disk.h>
54 #endif /* _KERNEL */
55 
56 #ifndef _KERNEL
57 #include <stdio.h>
58 #include <string.h>
59 #endif /* _KERNEL */
60 
61 #include <cam/cam.h>
62 #include <cam/cam_ccb.h>
63 #include <cam/cam_periph.h>
64 #include <cam/cam_xpt_periph.h>
65 #include <cam/cam_sim.h>
66 #include <cam/cam_iosched.h>
67 
68 #include <cam/nvme/nvme_all.h>
69 
70 typedef enum {
71 	NDA_STATE_NORMAL
72 } nda_state;
73 
74 typedef enum {
75 	NDA_FLAG_OPEN		= 0x0001,
76 	NDA_FLAG_DIRTY		= 0x0002,
77 	NDA_FLAG_SCTX_INIT	= 0x0004,
78 } nda_flags;
79 #define NDA_FLAG_STRING		\
80 	"\020"			\
81 	"\001OPEN"		\
82 	"\002DIRTY"		\
83 	"\003SCTX_INIT"
84 
85 typedef enum {
86 	NDA_Q_4K   = 0x01,
87 	NDA_Q_NONE = 0x00,
88 } nda_quirks;
89 
90 #define NDA_Q_BIT_STRING	\
91 	"\020"			\
92 	"\001Bit 0"
93 
94 typedef enum {
95 	NDA_CCB_BUFFER_IO	= 0x01,
96 	NDA_CCB_DUMP            = 0x02,
97 	NDA_CCB_TRIM            = 0x03,
98 	NDA_CCB_PASS            = 0x04,
99 	NDA_CCB_TYPE_MASK	= 0x0F,
100 } nda_ccb_state;
101 
102 /* Offsets into our private area for storing information */
103 #define ccb_state	ccb_h.ppriv_field0
104 #define ccb_bp		ccb_h.ppriv_ptr1	/* For NDA_CCB_BUFFER_IO */
105 #define ccb_trim	ccb_h.ppriv_ptr1	/* For NDA_CCB_TRIM */
106 
107 struct nda_softc {
108 	struct   cam_iosched_softc *cam_iosched;
109 	int			outstanding_cmds;	/* Number of active commands */
110 	int			refcount;		/* Active xpt_action() calls */
111 	nda_state		state;
112 	nda_flags		flags;
113 	nda_quirks		quirks;
114 	int			unmappedio;
115 	quad_t			deletes;
116 	uint32_t		nsid;			/* Namespace ID for this nda device */
117 	struct disk		*disk;
118 	struct task		sysctl_task;
119 	struct sysctl_ctx_list	sysctl_ctx;
120 	struct sysctl_oid	*sysctl_tree;
121 	uint64_t		trim_count;
122 	uint64_t		trim_ranges;
123 	uint64_t		trim_lbas;
124 #ifdef CAM_TEST_FAILURE
125 	int			force_read_error;
126 	int			force_write_error;
127 	int			periodic_read_error;
128 	int			periodic_read_count;
129 #endif
130 #ifdef CAM_IO_STATS
131 	struct sysctl_ctx_list	sysctl_stats_ctx;
132 	struct sysctl_oid	*sysctl_stats_tree;
133 	u_int			timeouts;
134 	u_int			errors;
135 	u_int			invalidations;
136 #endif
137 };
138 
139 struct nda_trim_request {
140 	struct nvme_dsm_range	dsm[NVME_MAX_DSM_TRIM / sizeof(struct nvme_dsm_range)];
141 	TAILQ_HEAD(, bio) bps;
142 };
143 _Static_assert(NVME_MAX_DSM_TRIM % sizeof(struct nvme_dsm_range) == 0,
144     "NVME_MAX_DSM_TRIM must be an integral number of ranges");
145 
146 /* Need quirk table */
147 
148 static	disk_ioctl_t	ndaioctl;
149 static	disk_strategy_t	ndastrategy;
150 static	dumper_t	ndadump;
151 static	periph_init_t	ndainit;
152 static	void		ndaasync(void *callback_arg, u_int32_t code,
153 				struct cam_path *path, void *arg);
154 static	void		ndasysctlinit(void *context, int pending);
155 static	int		ndaflagssysctl(SYSCTL_HANDLER_ARGS);
156 static	periph_ctor_t	ndaregister;
157 static	periph_dtor_t	ndacleanup;
158 static	periph_start_t	ndastart;
159 static	periph_oninv_t	ndaoninvalidate;
160 static	void		ndadone(struct cam_periph *periph,
161 			       union ccb *done_ccb);
162 static  int		ndaerror(union ccb *ccb, u_int32_t cam_flags,
163 				u_int32_t sense_flags);
164 static void		ndashutdown(void *arg, int howto);
165 static void		ndasuspend(void *arg);
166 
167 #ifndef	NDA_DEFAULT_SEND_ORDERED
168 #define	NDA_DEFAULT_SEND_ORDERED	1
169 #endif
170 #ifndef NDA_DEFAULT_TIMEOUT
171 #define NDA_DEFAULT_TIMEOUT 30	/* Timeout in seconds */
172 #endif
173 #ifndef	NDA_DEFAULT_RETRY
174 #define	NDA_DEFAULT_RETRY	4
175 #endif
176 #ifndef NDA_MAX_TRIM_ENTRIES
177 #define NDA_MAX_TRIM_ENTRIES  (NVME_MAX_DSM_TRIM / sizeof(struct nvme_dsm_range))/* Number of DSM trims to use, max 256 */
178 #endif
179 
180 static SYSCTL_NODE(_kern_cam, OID_AUTO, nda, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
181     "CAM Direct Access Disk driver");
182 
183 //static int nda_retry_count = NDA_DEFAULT_RETRY;
184 static int nda_send_ordered = NDA_DEFAULT_SEND_ORDERED;
185 static int nda_default_timeout = NDA_DEFAULT_TIMEOUT;
186 static int nda_max_trim_entries = NDA_MAX_TRIM_ENTRIES;
187 static int nda_enable_biospeedup = 1;
188 static int nda_nvd_compat = 1;
189 SYSCTL_INT(_kern_cam_nda, OID_AUTO, max_trim, CTLFLAG_RDTUN,
190     &nda_max_trim_entries, NDA_MAX_TRIM_ENTRIES,
191     "Maximum number of BIO_DELETE to send down as a DSM TRIM.");
192 SYSCTL_INT(_kern_cam_nda, OID_AUTO, enable_biospeedup, CTLFLAG_RDTUN,
193     &nda_enable_biospeedup, 0, "Enable BIO_SPEEDUP processing.");
194 SYSCTL_INT(_kern_cam_nda, OID_AUTO, nvd_compat, CTLFLAG_RDTUN,
195     &nda_nvd_compat, 1, "Enable creation of nvd aliases.");
196 
197 /*
198  * All NVMe media is non-rotational, so all nvme device instances
199  * share this to implement the sysctl.
200  */
201 static int nda_rotating_media = 0;
202 
203 static struct periph_driver ndadriver =
204 {
205 	ndainit, "nda",
206 	TAILQ_HEAD_INITIALIZER(ndadriver.units), /* generation */ 0
207 };
208 
209 PERIPHDRIVER_DECLARE(nda, ndadriver);
210 
211 static MALLOC_DEFINE(M_NVMEDA, "nvme_da", "nvme_da buffers");
212 
213 /*
214  * nice wrappers. Maybe these belong in nvme_all.c instead of
215  * here, but this is the only place that uses these. Should
216  * we ever grow another NVME periph, we should move them
217  * all there wholesale.
218  */
219 
220 static void
221 nda_nvme_flush(struct nda_softc *softc, struct ccb_nvmeio *nvmeio)
222 {
223 	cam_fill_nvmeio(nvmeio,
224 	    0,			/* retries */
225 	    ndadone,		/* cbfcnp */
226 	    CAM_DIR_NONE,	/* flags */
227 	    NULL,		/* data_ptr */
228 	    0,			/* dxfer_len */
229 	    nda_default_timeout * 1000); /* timeout 30s */
230 	nvme_ns_flush_cmd(&nvmeio->cmd, softc->nsid);
231 }
232 
233 static void
234 nda_nvme_trim(struct nda_softc *softc, struct ccb_nvmeio *nvmeio,
235     void *payload, uint32_t num_ranges)
236 {
237 	cam_fill_nvmeio(nvmeio,
238 	    0,			/* retries */
239 	    ndadone,		/* cbfcnp */
240 	    CAM_DIR_OUT,	/* flags */
241 	    payload,		/* data_ptr */
242 	    num_ranges * sizeof(struct nvme_dsm_range), /* dxfer_len */
243 	    nda_default_timeout * 1000); /* timeout 30s */
244 	nvme_ns_trim_cmd(&nvmeio->cmd, softc->nsid, num_ranges);
245 }
246 
247 static void
248 nda_nvme_write(struct nda_softc *softc, struct ccb_nvmeio *nvmeio,
249     void *payload, uint64_t lba, uint32_t len, uint32_t count)
250 {
251 	cam_fill_nvmeio(nvmeio,
252 	    0,			/* retries */
253 	    ndadone,		/* cbfcnp */
254 	    CAM_DIR_OUT,	/* flags */
255 	    payload,		/* data_ptr */
256 	    len,		/* dxfer_len */
257 	    nda_default_timeout * 1000); /* timeout 30s */
258 	nvme_ns_write_cmd(&nvmeio->cmd, softc->nsid, lba, count);
259 }
260 
261 static void
262 nda_nvme_rw_bio(struct nda_softc *softc, struct ccb_nvmeio *nvmeio,
263     struct bio *bp, uint32_t rwcmd)
264 {
265 	int flags = rwcmd == NVME_OPC_READ ? CAM_DIR_IN : CAM_DIR_OUT;
266 	void *payload;
267 	uint64_t lba;
268 	uint32_t count;
269 
270 	if (bp->bio_flags & BIO_UNMAPPED) {
271 		flags |= CAM_DATA_BIO;
272 		payload = bp;
273 	} else {
274 		payload = bp->bio_data;
275 	}
276 
277 	lba = bp->bio_pblkno;
278 	count = bp->bio_bcount / softc->disk->d_sectorsize;
279 
280 	cam_fill_nvmeio(nvmeio,
281 	    0,			/* retries */
282 	    ndadone,		/* cbfcnp */
283 	    flags,		/* flags */
284 	    payload,		/* data_ptr */
285 	    bp->bio_bcount,	/* dxfer_len */
286 	    nda_default_timeout * 1000); /* timeout 30s */
287 	nvme_ns_rw_cmd(&nvmeio->cmd, rwcmd, softc->nsid, lba, count);
288 }
289 
290 static int
291 ndaopen(struct disk *dp)
292 {
293 	struct cam_periph *periph;
294 	struct nda_softc *softc;
295 	int error;
296 
297 	periph = (struct cam_periph *)dp->d_drv1;
298 	if (cam_periph_acquire(periph) != 0) {
299 		return(ENXIO);
300 	}
301 
302 	cam_periph_lock(periph);
303 	if ((error = cam_periph_hold(periph, PRIBIO|PCATCH)) != 0) {
304 		cam_periph_unlock(periph);
305 		cam_periph_release(periph);
306 		return (error);
307 	}
308 
309 	CAM_DEBUG(periph->path, CAM_DEBUG_TRACE | CAM_DEBUG_PERIPH,
310 	    ("ndaopen\n"));
311 
312 	softc = (struct nda_softc *)periph->softc;
313 	softc->flags |= NDA_FLAG_OPEN;
314 
315 	cam_periph_unhold(periph);
316 	cam_periph_unlock(periph);
317 	return (0);
318 }
319 
320 static int
321 ndaclose(struct disk *dp)
322 {
323 	struct	cam_periph *periph;
324 	struct	nda_softc *softc;
325 	union ccb *ccb;
326 	int error;
327 
328 	periph = (struct cam_periph *)dp->d_drv1;
329 	softc = (struct nda_softc *)periph->softc;
330 	cam_periph_lock(periph);
331 
332 	CAM_DEBUG(periph->path, CAM_DEBUG_TRACE | CAM_DEBUG_PERIPH,
333 	    ("ndaclose\n"));
334 
335 	if ((softc->flags & NDA_FLAG_DIRTY) != 0 &&
336 	    (periph->flags & CAM_PERIPH_INVALID) == 0 &&
337 	    cam_periph_hold(periph, PRIBIO) == 0) {
338 		ccb = cam_periph_getccb(periph, CAM_PRIORITY_NORMAL);
339 		nda_nvme_flush(softc, &ccb->nvmeio);
340 		error = cam_periph_runccb(ccb, ndaerror, /*cam_flags*/0,
341 		    /*sense_flags*/0, softc->disk->d_devstat);
342 
343 		if (error != 0)
344 			xpt_print(periph->path, "Synchronize cache failed\n");
345 		else
346 			softc->flags &= ~NDA_FLAG_DIRTY;
347 		xpt_release_ccb(ccb);
348 		cam_periph_unhold(periph);
349 	}
350 
351 	softc->flags &= ~NDA_FLAG_OPEN;
352 
353 	while (softc->refcount != 0)
354 		cam_periph_sleep(periph, &softc->refcount, PRIBIO, "ndaclose", 1);
355 	KASSERT(softc->outstanding_cmds == 0,
356 	    ("nda %d outstanding commands", softc->outstanding_cmds));
357 	cam_periph_unlock(periph);
358 	cam_periph_release(periph);
359 	return (0);
360 }
361 
362 static void
363 ndaschedule(struct cam_periph *periph)
364 {
365 	struct nda_softc *softc = (struct nda_softc *)periph->softc;
366 
367 	if (softc->state != NDA_STATE_NORMAL)
368 		return;
369 
370 	cam_iosched_schedule(softc->cam_iosched, periph);
371 }
372 
373 static int
374 ndaioctl(struct disk *dp, u_long cmd, void *data, int fflag,
375     struct thread *td)
376 {
377 	struct cam_periph *periph;
378 
379 	periph = (struct cam_periph *)dp->d_drv1;
380 
381 	switch (cmd) {
382 	case NVME_IO_TEST:
383 	case NVME_BIO_TEST:
384 		/*
385 		 * These don't map well to the underlying CCBs, so
386 		 * they are usupported via CAM.
387 		 */
388 		return (ENOTTY);
389 	case NVME_GET_NSID:
390 	{
391 		struct nvme_get_nsid *gnsid = (struct nvme_get_nsid *)data;
392 		struct ccb_pathinq cpi;
393 
394 		xpt_path_inq(&cpi, periph->path);
395 		strncpy(gnsid->cdev, cpi.xport_specific.nvme.dev_name,
396 		    sizeof(gnsid->cdev));
397 		gnsid->nsid = cpi.xport_specific.nvme.nsid;
398 		return (0);
399 	}
400 	case NVME_PASSTHROUGH_CMD:
401 	{
402 		struct nvme_pt_command *pt;
403 		union ccb *ccb;
404 		struct cam_periph_map_info mapinfo;
405 		u_int maxmap = dp->d_maxsize;
406 		int error;
407 
408 		/*
409 		 * Create a NVME_IO CCB to do the passthrough command.
410 		 */
411 		pt = (struct nvme_pt_command *)data;
412 		ccb = xpt_alloc_ccb();
413 		xpt_setup_ccb(&ccb->ccb_h, periph->path, CAM_PRIORITY_NORMAL);
414 		ccb->ccb_state = NDA_CCB_PASS;
415 		cam_fill_nvmeio(&ccb->nvmeio,
416 		    0,			/* Retries */
417 		    ndadone,
418 		    (pt->is_read ? CAM_DIR_IN : CAM_DIR_OUT) | CAM_DATA_VADDR,
419 		    pt->buf,
420 		    pt->len,
421 		    nda_default_timeout * 1000);
422 		memcpy(&ccb->nvmeio.cmd, &pt->cmd, sizeof(pt->cmd));
423 
424 		/*
425 		 * Wire the user memory in this request for the I/O
426 		 */
427 		memset(&mapinfo, 0, sizeof(mapinfo));
428 		error = cam_periph_mapmem(ccb, &mapinfo, maxmap);
429 		if (error)
430 			goto out;
431 
432 		/*
433 		 * Lock the periph and run the command.
434 		 */
435 		cam_periph_lock(periph);
436 		cam_periph_runccb(ccb, NULL, CAM_RETRY_SELTO,
437 		    SF_RETRY_UA | SF_NO_PRINT, NULL);
438 
439 		/*
440 		 * Tear down mapping and return status.
441 		 */
442 		cam_periph_unlock(periph);
443 		cam_periph_unmapmem(ccb, &mapinfo);
444 		error = (ccb->ccb_h.status == CAM_REQ_CMP) ? 0 : EIO;
445 out:
446 		cam_periph_lock(periph);
447 		xpt_release_ccb(ccb);
448 		cam_periph_unlock(periph);
449 		return (error);
450 	}
451 	default:
452 		break;
453 	}
454 	return (ENOTTY);
455 }
456 
457 /*
458  * Actually translate the requested transfer into one the physical driver
459  * can understand.  The transfer is described by a buf and will include
460  * only one physical transfer.
461  */
462 static void
463 ndastrategy(struct bio *bp)
464 {
465 	struct cam_periph *periph;
466 	struct nda_softc *softc;
467 
468 	periph = (struct cam_periph *)bp->bio_disk->d_drv1;
469 	softc = (struct nda_softc *)periph->softc;
470 
471 	cam_periph_lock(periph);
472 
473 	CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("ndastrategy(%p)\n", bp));
474 
475 	/*
476 	 * If the device has been made invalid, error out
477 	 */
478 	if ((periph->flags & CAM_PERIPH_INVALID) != 0) {
479 		cam_periph_unlock(periph);
480 		biofinish(bp, NULL, ENXIO);
481 		return;
482 	}
483 
484 	if (bp->bio_cmd == BIO_DELETE)
485 		softc->deletes++;
486 
487 	/*
488 	 * Place it in the queue of disk activities for this disk
489 	 */
490 	cam_iosched_queue_work(softc->cam_iosched, bp);
491 
492 	/*
493 	 * Schedule ourselves for performing the work.
494 	 */
495 	ndaschedule(periph);
496 	cam_periph_unlock(periph);
497 
498 	return;
499 }
500 
501 static int
502 ndadump(void *arg, void *virtual, vm_offset_t physical, off_t offset, size_t length)
503 {
504 	struct	    cam_periph *periph;
505 	struct	    nda_softc *softc;
506 	u_int	    secsize;
507 	struct ccb_nvmeio nvmeio;
508 	struct	    disk *dp;
509 	uint64_t    lba;
510 	uint32_t    count;
511 	int	    error = 0;
512 
513 	dp = arg;
514 	periph = dp->d_drv1;
515 	softc = (struct nda_softc *)periph->softc;
516 	secsize = softc->disk->d_sectorsize;
517 	lba = offset / secsize;
518 	count = length / secsize;
519 
520 	if ((periph->flags & CAM_PERIPH_INVALID) != 0)
521 		return (ENXIO);
522 
523 	/* xpt_get_ccb returns a zero'd allocation for the ccb, mimic that here */
524 	memset(&nvmeio, 0, sizeof(nvmeio));
525 	if (length > 0) {
526 		xpt_setup_ccb(&nvmeio.ccb_h, periph->path, CAM_PRIORITY_NORMAL);
527 		nvmeio.ccb_state = NDA_CCB_DUMP;
528 		nda_nvme_write(softc, &nvmeio, virtual, lba, length, count);
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 			printf("Aborting dump due to I/O error %d.\n", error);
533 
534 		return (error);
535 	}
536 
537 	/* Flush */
538 	xpt_setup_ccb(&nvmeio.ccb_h, periph->path, CAM_PRIORITY_NORMAL);
539 
540 	nvmeio.ccb_state = NDA_CCB_DUMP;
541 	nda_nvme_flush(softc, &nvmeio);
542 	error = cam_periph_runccb((union ccb *)&nvmeio, cam_periph_error,
543 	    0, SF_NO_RECOVERY | SF_NO_RETRY, NULL);
544 	if (error != 0)
545 		xpt_print(periph->path, "flush cmd failed\n");
546 	return (error);
547 }
548 
549 static void
550 ndainit(void)
551 {
552 	cam_status status;
553 
554 	/*
555 	 * Install a global async callback.  This callback will
556 	 * receive async callbacks like "new device found".
557 	 */
558 	status = xpt_register_async(AC_FOUND_DEVICE, ndaasync, NULL, NULL);
559 
560 	if (status != CAM_REQ_CMP) {
561 		printf("nda: Failed to attach master async callback "
562 		       "due to status 0x%x!\n", status);
563 	} else if (nda_send_ordered) {
564 		/* Register our event handlers */
565 		if ((EVENTHANDLER_REGISTER(power_suspend, ndasuspend,
566 					   NULL, EVENTHANDLER_PRI_LAST)) == NULL)
567 		    printf("ndainit: power event registration failed!\n");
568 		if ((EVENTHANDLER_REGISTER(shutdown_post_sync, ndashutdown,
569 					   NULL, SHUTDOWN_PRI_DEFAULT)) == NULL)
570 		    printf("ndainit: shutdown event registration failed!\n");
571 	}
572 }
573 
574 /*
575  * Callback from GEOM, called when it has finished cleaning up its
576  * resources.
577  */
578 static void
579 ndadiskgonecb(struct disk *dp)
580 {
581 	struct cam_periph *periph;
582 
583 	periph = (struct cam_periph *)dp->d_drv1;
584 
585 	cam_periph_release(periph);
586 }
587 
588 static void
589 ndaoninvalidate(struct cam_periph *periph)
590 {
591 	struct nda_softc *softc;
592 
593 	softc = (struct nda_softc *)periph->softc;
594 
595 	/*
596 	 * De-register any async callbacks.
597 	 */
598 	xpt_register_async(0, ndaasync, periph, periph->path);
599 #ifdef CAM_IO_STATS
600 	softc->invalidations++;
601 #endif
602 
603 	/*
604 	 * Return all queued I/O with ENXIO. Transactions may be queued up here
605 	 * for retry (since we are called while there's other transactions
606 	 * pending). Any requests in the hardware will drain before ndacleanup
607 	 * is called.
608 	 */
609 	cam_iosched_flush(softc->cam_iosched, NULL, ENXIO);
610 
611 	/*
612 	 * Tell GEOM that we've gone away, we'll get a callback when it is
613 	 * done cleaning up its resources.
614 	 */
615 	disk_gone(softc->disk);
616 }
617 
618 static void
619 ndacleanup(struct cam_periph *periph)
620 {
621 	struct nda_softc *softc;
622 
623 	softc = (struct nda_softc *)periph->softc;
624 
625 	cam_periph_unlock(periph);
626 
627 	cam_iosched_fini(softc->cam_iosched);
628 
629 	/*
630 	 * If we can't free the sysctl tree, oh well...
631 	 */
632 	if ((softc->flags & NDA_FLAG_SCTX_INIT) != 0) {
633 #ifdef CAM_IO_STATS
634 		if (sysctl_ctx_free(&softc->sysctl_stats_ctx) != 0)
635 			xpt_print(periph->path,
636 			    "can't remove sysctl stats context\n");
637 #endif
638 		if (sysctl_ctx_free(&softc->sysctl_ctx) != 0)
639 			xpt_print(periph->path,
640 			    "can't remove sysctl context\n");
641 	}
642 
643 	disk_destroy(softc->disk);
644 	free(softc, M_DEVBUF);
645 	cam_periph_lock(periph);
646 }
647 
648 static void
649 ndaasync(void *callback_arg, u_int32_t code,
650 	struct cam_path *path, void *arg)
651 {
652 	struct cam_periph *periph;
653 
654 	periph = (struct cam_periph *)callback_arg;
655 	switch (code) {
656 	case AC_FOUND_DEVICE:
657 	{
658 		struct ccb_getdev *cgd;
659 		cam_status status;
660 
661 		cgd = (struct ccb_getdev *)arg;
662 		if (cgd == NULL)
663 			break;
664 
665 		if (cgd->protocol != PROTO_NVME)
666 			break;
667 
668 		/*
669 		 * Allocate a peripheral instance for
670 		 * this device and start the probe
671 		 * process.
672 		 */
673 		status = cam_periph_alloc(ndaregister, ndaoninvalidate,
674 					  ndacleanup, ndastart,
675 					  "nda", CAM_PERIPH_BIO,
676 					  path, ndaasync,
677 					  AC_FOUND_DEVICE, cgd);
678 
679 		if (status != CAM_REQ_CMP
680 		 && status != CAM_REQ_INPROG)
681 			printf("ndaasync: Unable to attach to new device "
682 				"due to status 0x%x\n", status);
683 		break;
684 	}
685 	case AC_ADVINFO_CHANGED:
686 	{
687 		uintptr_t buftype;
688 
689 		buftype = (uintptr_t)arg;
690 		if (buftype == CDAI_TYPE_PHYS_PATH) {
691 			struct nda_softc *softc;
692 
693 			softc = periph->softc;
694 			disk_attr_changed(softc->disk, "GEOM::physpath",
695 					  M_NOWAIT);
696 		}
697 		break;
698 	}
699 	case AC_LOST_DEVICE:
700 	default:
701 		break;
702 	}
703 	cam_periph_async(periph, code, path, arg);
704 }
705 
706 static void
707 ndasysctlinit(void *context, int pending)
708 {
709 	struct cam_periph *periph;
710 	struct nda_softc *softc;
711 	char tmpstr[32], tmpstr2[16];
712 
713 	periph = (struct cam_periph *)context;
714 
715 	/* periph was held for us when this task was enqueued */
716 	if ((periph->flags & CAM_PERIPH_INVALID) != 0) {
717 		cam_periph_release(periph);
718 		return;
719 	}
720 
721 	softc = (struct nda_softc *)periph->softc;
722 	snprintf(tmpstr, sizeof(tmpstr), "CAM NDA unit %d", periph->unit_number);
723 	snprintf(tmpstr2, sizeof(tmpstr2), "%d", periph->unit_number);
724 
725 	sysctl_ctx_init(&softc->sysctl_ctx);
726 	softc->flags |= NDA_FLAG_SCTX_INIT;
727 	softc->sysctl_tree = SYSCTL_ADD_NODE_WITH_LABEL(&softc->sysctl_ctx,
728 		SYSCTL_STATIC_CHILDREN(_kern_cam_nda), OID_AUTO, tmpstr2,
729 		CTLFLAG_RD | CTLFLAG_MPSAFE, 0, tmpstr, "device_index");
730 	if (softc->sysctl_tree == NULL) {
731 		printf("ndasysctlinit: unable to allocate sysctl tree\n");
732 		cam_periph_release(periph);
733 		return;
734 	}
735 
736 	SYSCTL_ADD_INT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
737 	    OID_AUTO, "unmapped_io", CTLFLAG_RD,
738 	    &softc->unmappedio, 0, "Unmapped I/O leaf");
739 
740 	SYSCTL_ADD_QUAD(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
741 	    OID_AUTO, "deletes", CTLFLAG_RD,
742 	    &softc->deletes, "Number of BIO_DELETE requests");
743 
744 	SYSCTL_ADD_UQUAD(&softc->sysctl_ctx,
745 		SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO,
746 		"trim_count", CTLFLAG_RD, &softc->trim_count,
747 		"Total number of unmap/dsm commands sent");
748 	SYSCTL_ADD_UQUAD(&softc->sysctl_ctx,
749 		SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO,
750 		"trim_ranges", CTLFLAG_RD, &softc->trim_ranges,
751 		"Total number of ranges in unmap/dsm commands");
752 	SYSCTL_ADD_UQUAD(&softc->sysctl_ctx,
753 		SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO,
754 		"trim_lbas", CTLFLAG_RD, &softc->trim_lbas,
755 		"Total lbas in the unmap/dsm commands sent");
756 
757 	SYSCTL_ADD_INT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
758 	    OID_AUTO, "rotating", CTLFLAG_RD, &nda_rotating_media, 1,
759 	    "Rotating media");
760 
761 	SYSCTL_ADD_PROC(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
762 	    OID_AUTO, "flags", CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE,
763 	    softc, 0, ndaflagssysctl, "A",
764 	    "Flags for drive");
765 
766 #ifdef CAM_IO_STATS
767 	softc->sysctl_stats_tree = SYSCTL_ADD_NODE(&softc->sysctl_stats_ctx,
768 		SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO, "stats",
769 		CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "Statistics");
770 	if (softc->sysctl_stats_tree == NULL) {
771 		printf("ndasysctlinit: unable to allocate sysctl tree for stats\n");
772 		cam_periph_release(periph);
773 		return;
774 	}
775 	SYSCTL_ADD_INT(&softc->sysctl_stats_ctx,
776 		SYSCTL_CHILDREN(softc->sysctl_stats_tree),
777 		OID_AUTO, "timeouts", CTLFLAG_RD,
778 		&softc->timeouts, 0,
779 		"Device timeouts reported by the SIM");
780 	SYSCTL_ADD_INT(&softc->sysctl_stats_ctx,
781 		SYSCTL_CHILDREN(softc->sysctl_stats_tree),
782 		OID_AUTO, "errors", CTLFLAG_RD,
783 		&softc->errors, 0,
784 		"Transport errors reported by the SIM.");
785 	SYSCTL_ADD_INT(&softc->sysctl_stats_ctx,
786 		SYSCTL_CHILDREN(softc->sysctl_stats_tree),
787 		OID_AUTO, "pack_invalidations", CTLFLAG_RD,
788 		&softc->invalidations, 0,
789 		"Device pack invalidations.");
790 #endif
791 
792 #ifdef CAM_TEST_FAILURE
793 	SYSCTL_ADD_PROC(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
794 		OID_AUTO, "invalidate", CTLTYPE_U64 | CTLFLAG_RW | CTLFLAG_MPSAFE,
795 		periph, 0, cam_periph_invalidate_sysctl, "I",
796 		"Write 1 to invalidate the drive immediately");
797 #endif
798 
799 	cam_iosched_sysctl_init(softc->cam_iosched, &softc->sysctl_ctx,
800 	    softc->sysctl_tree);
801 
802 	cam_periph_release(periph);
803 }
804 
805 static int
806 ndaflagssysctl(SYSCTL_HANDLER_ARGS)
807 {
808 	struct sbuf sbuf;
809 	struct nda_softc *softc = arg1;
810 	int error;
811 
812 	sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
813 	if (softc->flags != 0)
814 		sbuf_printf(&sbuf, "0x%b", (unsigned)softc->flags, NDA_FLAG_STRING);
815 	else
816 		sbuf_printf(&sbuf, "0");
817 	error = sbuf_finish(&sbuf);
818 	sbuf_delete(&sbuf);
819 
820 	return (error);
821 }
822 
823 static int
824 ndagetattr(struct bio *bp)
825 {
826 	int ret;
827 	struct cam_periph *periph;
828 
829 	if (g_handleattr_int(bp, "GEOM::canspeedup", nda_enable_biospeedup))
830 		return (EJUSTRETURN);
831 
832 	periph = (struct cam_periph *)bp->bio_disk->d_drv1;
833 	cam_periph_lock(periph);
834 	ret = xpt_getattr(bp->bio_data, bp->bio_length, bp->bio_attribute,
835 	    periph->path);
836 	cam_periph_unlock(periph);
837 	if (ret == 0)
838 		bp->bio_completed = bp->bio_length;
839 	return ret;
840 }
841 
842 static cam_status
843 ndaregister(struct cam_periph *periph, void *arg)
844 {
845 	struct nda_softc *softc;
846 	struct disk *disk;
847 	struct ccb_pathinq cpi;
848 	const struct nvme_namespace_data *nsd;
849 	const struct nvme_controller_data *cd;
850 	char   announce_buf[80];
851 	uint8_t flbas_fmt, lbads, vwc_present;
852 	u_int maxio;
853 	int quirks;
854 
855 	nsd = nvme_get_identify_ns(periph);
856 	cd = nvme_get_identify_cntrl(periph);
857 
858 	softc = (struct nda_softc *)malloc(sizeof(*softc), M_DEVBUF,
859 	    M_NOWAIT | M_ZERO);
860 
861 	if (softc == NULL) {
862 		printf("ndaregister: Unable to probe new device. "
863 		    "Unable to allocate softc\n");
864 		return(CAM_REQ_CMP_ERR);
865 	}
866 
867 	if (cam_iosched_init(&softc->cam_iosched, periph) != 0) {
868 		printf("ndaregister: Unable to probe new device. "
869 		       "Unable to allocate iosched memory\n");
870 		free(softc, M_DEVBUF);
871 		return(CAM_REQ_CMP_ERR);
872 	}
873 
874 	/* ident_data parsing */
875 
876 	periph->softc = softc;
877 	softc->quirks = NDA_Q_NONE;
878 	xpt_path_inq(&cpi, periph->path);
879 	TASK_INIT(&softc->sysctl_task, 0, ndasysctlinit, periph);
880 
881 	/*
882 	 * The name space ID is the lun, save it for later I/O
883 	 */
884 	softc->nsid = (uint32_t)xpt_path_lun_id(periph->path);
885 
886 	/*
887 	 * Register this media as a disk
888 	 */
889 	(void)cam_periph_hold(periph, PRIBIO);
890 	cam_periph_unlock(periph);
891 	snprintf(announce_buf, sizeof(announce_buf),
892 	    "kern.cam.nda.%d.quirks", periph->unit_number);
893 	quirks = softc->quirks;
894 	TUNABLE_INT_FETCH(announce_buf, &quirks);
895 	softc->quirks = quirks;
896 	cam_iosched_set_sort_queue(softc->cam_iosched, 0);
897 	softc->disk = disk = disk_alloc();
898 	disk->d_rotation_rate = DISK_RR_NON_ROTATING;
899 	disk->d_open = ndaopen;
900 	disk->d_close = ndaclose;
901 	disk->d_strategy = ndastrategy;
902 	disk->d_ioctl = ndaioctl;
903 	disk->d_getattr = ndagetattr;
904 	if (cam_sim_pollable(periph->sim))
905 		disk->d_dump = ndadump;
906 	disk->d_gone = ndadiskgonecb;
907 	disk->d_name = "nda";
908 	disk->d_drv1 = periph;
909 	disk->d_unit = periph->unit_number;
910 	maxio = cpi.maxio;		/* Honor max I/O size of SIM */
911 	if (maxio == 0)
912 		maxio = DFLTPHYS;	/* traditional default */
913 	else if (maxio > maxphys)
914 		maxio = maxphys;	/* for safety */
915 	disk->d_maxsize = maxio;
916 	flbas_fmt = (nsd->flbas >> NVME_NS_DATA_FLBAS_FORMAT_SHIFT) &
917 		NVME_NS_DATA_FLBAS_FORMAT_MASK;
918 	lbads = (nsd->lbaf[flbas_fmt] >> NVME_NS_DATA_LBAF_LBADS_SHIFT) &
919 		NVME_NS_DATA_LBAF_LBADS_MASK;
920 	disk->d_sectorsize = 1 << lbads;
921 	disk->d_mediasize = (off_t)(disk->d_sectorsize * nsd->nsze);
922 	disk->d_delmaxsize = disk->d_mediasize;
923 	disk->d_flags = DISKFLAG_DIRECT_COMPLETION;
924 	if (nvme_ctrlr_has_dataset_mgmt(cd))
925 		disk->d_flags |= DISKFLAG_CANDELETE;
926 	vwc_present = (cd->vwc >> NVME_CTRLR_DATA_VWC_PRESENT_SHIFT) &
927 		NVME_CTRLR_DATA_VWC_PRESENT_MASK;
928 	if (vwc_present)
929 		disk->d_flags |= DISKFLAG_CANFLUSHCACHE;
930 	if ((cpi.hba_misc & PIM_UNMAPPED) != 0) {
931 		disk->d_flags |= DISKFLAG_UNMAPPED_BIO;
932 		softc->unmappedio = 1;
933 	}
934 	/*
935 	 * d_ident and d_descr are both far bigger than the length of either
936 	 *  the serial or model number strings.
937 	 */
938 	cam_strvis_flag(disk->d_descr, cd->mn, NVME_MODEL_NUMBER_LENGTH,
939 	    sizeof(disk->d_descr), CAM_STRVIS_FLAG_NONASCII_SPC);
940 
941 	cam_strvis_flag(disk->d_ident, cd->sn, NVME_SERIAL_NUMBER_LENGTH,
942 	    sizeof(disk->d_ident), CAM_STRVIS_FLAG_NONASCII_SPC);
943 
944 	disk->d_hba_vendor = cpi.hba_vendor;
945 	disk->d_hba_device = cpi.hba_device;
946 	disk->d_hba_subvendor = cpi.hba_subvendor;
947 	disk->d_hba_subdevice = cpi.hba_subdevice;
948 	snprintf(disk->d_attachment, sizeof(disk->d_attachment),
949 	    "%s%d", cpi.dev_name, cpi.unit_number);
950 	if (((nsd->nsfeat >> NVME_NS_DATA_NSFEAT_NPVALID_SHIFT) &
951 	    NVME_NS_DATA_NSFEAT_NPVALID_MASK) != 0 && nsd->npwg != 0)
952 		disk->d_stripesize = ((nsd->npwg + 1) * disk->d_sectorsize);
953 	else
954 		disk->d_stripesize = nsd->noiob * disk->d_sectorsize;
955 	disk->d_stripeoffset = 0;
956 	disk->d_devstat = devstat_new_entry(periph->periph_name,
957 	    periph->unit_number, disk->d_sectorsize,
958 	    DEVSTAT_ALL_SUPPORTED,
959 	    DEVSTAT_TYPE_DIRECT | XPORT_DEVSTAT_TYPE(cpi.transport),
960 	    DEVSTAT_PRIORITY_DISK);
961 	/*
962 	 * Add alias for older nvd drives to ease transition.
963 	 */
964 	if (nda_nvd_compat)
965 		disk_add_alias(disk, "nvd");
966 
967 	/*
968 	 * Acquire a reference to the periph before we register with GEOM.
969 	 * We'll release this reference once GEOM calls us back (via
970 	 * ndadiskgonecb()) telling us that our provider has been freed.
971 	 */
972 	if (cam_periph_acquire(periph) != 0) {
973 		xpt_print(periph->path, "%s: lost periph during "
974 			  "registration!\n", __func__);
975 		cam_periph_lock(periph);
976 		return (CAM_REQ_CMP_ERR);
977 	}
978 	disk_create(softc->disk, DISK_VERSION);
979 	cam_periph_lock(periph);
980 	cam_periph_unhold(periph);
981 
982 	snprintf(announce_buf, sizeof(announce_buf),
983 		"%juMB (%ju %u byte sectors)",
984 	    (uintmax_t)((uintmax_t)disk->d_mediasize / (1024*1024)),
985 		(uintmax_t)disk->d_mediasize / disk->d_sectorsize,
986 		disk->d_sectorsize);
987 	xpt_announce_periph(periph, announce_buf);
988 	xpt_announce_quirks(periph, softc->quirks, NDA_Q_BIT_STRING);
989 
990 	/*
991 	 * Create our sysctl variables, now that we know
992 	 * we have successfully attached.
993 	 */
994 	if (cam_periph_acquire(periph) == 0)
995 		taskqueue_enqueue(taskqueue_thread, &softc->sysctl_task);
996 
997 	/*
998 	 * Register for device going away and info about the drive
999 	 * changing (though with NVMe, it can't)
1000 	 */
1001 	xpt_register_async(AC_LOST_DEVICE | AC_ADVINFO_CHANGED,
1002 	    ndaasync, periph, periph->path);
1003 
1004 	softc->state = NDA_STATE_NORMAL;
1005 	return(CAM_REQ_CMP);
1006 }
1007 
1008 static void
1009 ndastart(struct cam_periph *periph, union ccb *start_ccb)
1010 {
1011 	struct nda_softc *softc = (struct nda_softc *)periph->softc;
1012 	struct ccb_nvmeio *nvmeio = &start_ccb->nvmeio;
1013 
1014 	CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("ndastart\n"));
1015 
1016 	switch (softc->state) {
1017 	case NDA_STATE_NORMAL:
1018 	{
1019 		struct bio *bp;
1020 
1021 		bp = cam_iosched_next_bio(softc->cam_iosched);
1022 		CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("ndastart: bio %p\n", bp));
1023 		if (bp == NULL) {
1024 			xpt_release_ccb(start_ccb);
1025 			break;
1026 		}
1027 
1028 		switch (bp->bio_cmd) {
1029 		case BIO_WRITE:
1030 			softc->flags |= NDA_FLAG_DIRTY;
1031 			/* FALLTHROUGH */
1032 		case BIO_READ:
1033 		{
1034 #ifdef CAM_TEST_FAILURE
1035 			int fail = 0;
1036 
1037 			/*
1038 			 * Support the failure ioctls.  If the command is a
1039 			 * read, and there are pending forced read errors, or
1040 			 * if a write and pending write errors, then fail this
1041 			 * operation with EIO.  This is useful for testing
1042 			 * purposes.  Also, support having every Nth read fail.
1043 			 *
1044 			 * This is a rather blunt tool.
1045 			 */
1046 			if (bp->bio_cmd == BIO_READ) {
1047 				if (softc->force_read_error) {
1048 					softc->force_read_error--;
1049 					fail = 1;
1050 				}
1051 				if (softc->periodic_read_error > 0) {
1052 					if (++softc->periodic_read_count >=
1053 					    softc->periodic_read_error) {
1054 						softc->periodic_read_count = 0;
1055 						fail = 1;
1056 					}
1057 				}
1058 			} else {
1059 				if (softc->force_write_error) {
1060 					softc->force_write_error--;
1061 					fail = 1;
1062 				}
1063 			}
1064 			if (fail) {
1065 				biofinish(bp, NULL, EIO);
1066 				xpt_release_ccb(start_ccb);
1067 				ndaschedule(periph);
1068 				return;
1069 			}
1070 #endif
1071 			KASSERT((bp->bio_flags & BIO_UNMAPPED) == 0 ||
1072 			    round_page(bp->bio_bcount + bp->bio_ma_offset) /
1073 			    PAGE_SIZE == bp->bio_ma_n,
1074 			    ("Short bio %p", bp));
1075 			nda_nvme_rw_bio(softc, &start_ccb->nvmeio, bp, bp->bio_cmd == BIO_READ ?
1076 			    NVME_OPC_READ : NVME_OPC_WRITE);
1077 			break;
1078 		}
1079 		case BIO_DELETE:
1080 		{
1081 			struct nvme_dsm_range *dsm_range, *dsm_end;
1082 			struct nda_trim_request *trim;
1083 			struct bio *bp1;
1084 			int ents;
1085 			uint32_t totalcount = 0, ranges = 0;
1086 
1087 			trim = malloc(sizeof(*trim), M_NVMEDA, M_ZERO | M_NOWAIT);
1088 			if (trim == NULL) {
1089 				biofinish(bp, NULL, ENOMEM);
1090 				xpt_release_ccb(start_ccb);
1091 				ndaschedule(periph);
1092 				return;
1093 			}
1094 			TAILQ_INIT(&trim->bps);
1095 			bp1 = bp;
1096 			ents = min(nitems(trim->dsm), nda_max_trim_entries);
1097 			ents = max(ents, 1);
1098 			dsm_range = trim->dsm;
1099 			dsm_end = dsm_range + ents;
1100 			do {
1101 				TAILQ_INSERT_TAIL(&trim->bps, bp1, bio_queue);
1102 				dsm_range->length =
1103 				    htole32(bp1->bio_bcount / softc->disk->d_sectorsize);
1104 				dsm_range->starting_lba =
1105 				    htole64(bp1->bio_offset / softc->disk->d_sectorsize);
1106 				ranges++;
1107 				totalcount += dsm_range->length;
1108 				dsm_range++;
1109 				if (dsm_range >= dsm_end)
1110 					break;
1111 				bp1 = cam_iosched_next_trim(softc->cam_iosched);
1112 				/* XXX -- Could collapse adjacent ranges, but we don't for now */
1113 				/* XXX -- Could limit based on total payload size */
1114 			} while (bp1 != NULL);
1115 			start_ccb->ccb_trim = trim;
1116 			nda_nvme_trim(softc, &start_ccb->nvmeio, trim->dsm,
1117 			    dsm_range - trim->dsm);
1118 			start_ccb->ccb_state = NDA_CCB_TRIM;
1119 			softc->trim_count++;
1120 			softc->trim_ranges += ranges;
1121 			softc->trim_lbas += totalcount;
1122 			/*
1123 			 * Note: We can have multiple TRIMs in flight, so we don't call
1124 			 * cam_iosched_submit_trim(softc->cam_iosched);
1125 			 * since that forces the I/O scheduler to only schedule one at a time.
1126 			 * On NVMe drives, this is a performance disaster.
1127 			 */
1128 			goto out;
1129 		}
1130 		case BIO_FLUSH:
1131 			nda_nvme_flush(softc, nvmeio);
1132 			break;
1133 		default:
1134 			biofinish(bp, NULL, EOPNOTSUPP);
1135 			xpt_release_ccb(start_ccb);
1136 			ndaschedule(periph);
1137 			return;
1138 		}
1139 		start_ccb->ccb_state = NDA_CCB_BUFFER_IO;
1140 		start_ccb->ccb_bp = bp;
1141 out:
1142 		start_ccb->ccb_h.flags |= CAM_UNLOCKED;
1143 		softc->outstanding_cmds++;
1144 		softc->refcount++;			/* For submission only */
1145 		cam_periph_unlock(periph);
1146 		xpt_action(start_ccb);
1147 		cam_periph_lock(periph);
1148 		softc->refcount--;			/* Submission done */
1149 
1150 		/* May have more work to do, so ensure we stay scheduled */
1151 		ndaschedule(periph);
1152 		break;
1153 		}
1154 	}
1155 }
1156 
1157 static void
1158 ndadone(struct cam_periph *periph, union ccb *done_ccb)
1159 {
1160 	struct nda_softc *softc;
1161 	struct ccb_nvmeio *nvmeio = &done_ccb->nvmeio;
1162 	struct cam_path *path;
1163 	int state;
1164 
1165 	softc = (struct nda_softc *)periph->softc;
1166 	path = done_ccb->ccb_h.path;
1167 
1168 	CAM_DEBUG(path, CAM_DEBUG_TRACE, ("ndadone\n"));
1169 
1170 	state = nvmeio->ccb_state & NDA_CCB_TYPE_MASK;
1171 	switch (state) {
1172 	case NDA_CCB_BUFFER_IO:
1173 	case NDA_CCB_TRIM:
1174 	{
1175 		int error;
1176 
1177 		cam_periph_lock(periph);
1178 		if ((done_ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_CMP) {
1179 			error = ndaerror(done_ccb, 0, 0);
1180 			if (error == ERESTART) {
1181 				/* A retry was scheduled, so just return. */
1182 				cam_periph_unlock(periph);
1183 				return;
1184 			}
1185 			if ((done_ccb->ccb_h.status & CAM_DEV_QFRZN) != 0)
1186 				cam_release_devq(path,
1187 						 /*relsim_flags*/0,
1188 						 /*reduction*/0,
1189 						 /*timeout*/0,
1190 						 /*getcount_only*/0);
1191 		} else {
1192 			if ((done_ccb->ccb_h.status & CAM_DEV_QFRZN) != 0)
1193 				panic("REQ_CMP with QFRZN");
1194 			error = 0;
1195 		}
1196 		if (state == NDA_CCB_BUFFER_IO) {
1197 			struct bio *bp;
1198 
1199 			bp = (struct bio *)done_ccb->ccb_bp;
1200 			bp->bio_error = error;
1201 			if (error != 0) {
1202 				bp->bio_resid = bp->bio_bcount;
1203 				bp->bio_flags |= BIO_ERROR;
1204 			} else {
1205 				bp->bio_resid = 0;
1206 			}
1207 			softc->outstanding_cmds--;
1208 
1209 			/*
1210 			 * We need to call cam_iosched before we call biodone so that we
1211 			 * don't measure any activity that happens in the completion
1212 			 * routine, which in the case of sendfile can be quite
1213 			 * extensive.
1214 			 */
1215 			cam_iosched_bio_complete(softc->cam_iosched, bp, done_ccb);
1216 			xpt_release_ccb(done_ccb);
1217 			ndaschedule(periph);
1218 			cam_periph_unlock(periph);
1219 			biodone(bp);
1220 		} else { /* state == NDA_CCB_TRIM */
1221 			struct nda_trim_request *trim;
1222 			struct bio *bp1, *bp2;
1223 			TAILQ_HEAD(, bio) queue;
1224 
1225 			trim = nvmeio->ccb_trim;
1226 			TAILQ_INIT(&queue);
1227 			TAILQ_CONCAT(&queue, &trim->bps, bio_queue);
1228 			free(trim, M_NVMEDA);
1229 
1230 			/*
1231 			 * Since we can have multiple trims in flight, we don't
1232 			 * need to call this here.
1233 			 * cam_iosched_trim_done(softc->cam_iosched);
1234 			 */
1235 			/*
1236 			 * The the I/O scheduler that we're finishing the I/O
1237 			 * so we can keep book. The first one we pass in the CCB
1238 			 * which has the timing information. The rest we pass in NULL
1239 			 * so we can keep proper counts.
1240 			 */
1241 			bp1 = TAILQ_FIRST(&queue);
1242 			cam_iosched_bio_complete(softc->cam_iosched, bp1, done_ccb);
1243 			xpt_release_ccb(done_ccb);
1244 			softc->outstanding_cmds--;
1245 			ndaschedule(periph);
1246 			cam_periph_unlock(periph);
1247 			while ((bp2 = TAILQ_FIRST(&queue)) != NULL) {
1248 				TAILQ_REMOVE(&queue, bp2, bio_queue);
1249 				bp2->bio_error = error;
1250 				if (error != 0) {
1251 					bp2->bio_flags |= BIO_ERROR;
1252 					bp2->bio_resid = bp1->bio_bcount;
1253 				} else
1254 					bp2->bio_resid = 0;
1255 				if (bp1 != bp2)
1256 					cam_iosched_bio_complete(softc->cam_iosched, bp2, NULL);
1257 				biodone(bp2);
1258 			}
1259 		}
1260 		return;
1261 	}
1262 	case NDA_CCB_DUMP:
1263 		/* No-op.  We're polling */
1264 		return;
1265 	case NDA_CCB_PASS:
1266 		/* NVME_PASSTHROUGH_CMD runs this CCB and releases it */
1267 		return;
1268 	default:
1269 		break;
1270 	}
1271 	xpt_release_ccb(done_ccb);
1272 }
1273 
1274 static int
1275 ndaerror(union ccb *ccb, u_int32_t cam_flags, u_int32_t sense_flags)
1276 {
1277 #ifdef CAM_IO_STATS
1278 	struct nda_softc *softc;
1279 	struct cam_periph *periph;
1280 
1281 	periph = xpt_path_periph(ccb->ccb_h.path);
1282 	softc = (struct nda_softc *)periph->softc;
1283 #endif
1284 
1285 	switch (ccb->ccb_h.status & CAM_STATUS_MASK) {
1286 	case CAM_CMD_TIMEOUT:
1287 #ifdef CAM_IO_STATS
1288 		softc->timeouts++;
1289 #endif
1290 		break;
1291 	case CAM_REQ_ABORTED:
1292 	case CAM_REQ_CMP_ERR:
1293 	case CAM_REQ_TERMIO:
1294 	case CAM_UNREC_HBA_ERROR:
1295 	case CAM_DATA_RUN_ERR:
1296 	case CAM_ATA_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, 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)
1370 {
1371 
1372 	ndaflush();
1373 }
1374