xref: /freebsd/sys/cam/nvme/nvme_da.c (revision f7c32ed617858bcd22f8d1b03199099d50125721)
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 	struct nda_softc *softc;
379 
380 	periph = (struct cam_periph *)dp->d_drv1;
381 	softc = (struct nda_softc *)periph->softc;
382 
383 	switch (cmd) {
384 	case NVME_IO_TEST:
385 	case NVME_BIO_TEST:
386 		/*
387 		 * These don't map well to the underlying CCBs, so
388 		 * they are usupported via CAM.
389 		 */
390 		return (ENOTTY);
391 	case NVME_GET_NSID:
392 	{
393 		struct nvme_get_nsid *gnsid = (struct nvme_get_nsid *)data;
394 		struct ccb_pathinq cpi;
395 
396 		xpt_path_inq(&cpi, periph->path);
397 		strncpy(gnsid->cdev, cpi.xport_specific.nvme.dev_name,
398 		    sizeof(gnsid->cdev));
399 		gnsid->nsid = cpi.xport_specific.nvme.nsid;
400 		return (0);
401 	}
402 	case NVME_PASSTHROUGH_CMD:
403 	{
404 		struct nvme_pt_command *pt;
405 		union ccb *ccb;
406 		struct cam_periph_map_info mapinfo;
407 		u_int maxmap = dp->d_maxsize;
408 		int error;
409 
410 		/*
411 		 * Create a NVME_IO CCB to do the passthrough command.
412 		 */
413 		pt = (struct nvme_pt_command *)data;
414 		ccb = xpt_alloc_ccb();
415 		xpt_setup_ccb(&ccb->ccb_h, periph->path, CAM_PRIORITY_NORMAL);
416 		ccb->ccb_state = NDA_CCB_PASS;
417 		cam_fill_nvmeio(&ccb->nvmeio,
418 		    0,			/* Retries */
419 		    ndadone,
420 		    (pt->is_read ? CAM_DIR_IN : CAM_DIR_OUT) | CAM_DATA_VADDR,
421 		    pt->buf,
422 		    pt->len,
423 		    nda_default_timeout * 1000);
424 		memcpy(&ccb->nvmeio.cmd, &pt->cmd, sizeof(pt->cmd));
425 
426 		/*
427 		 * Wire the user memory in this request for the I/O
428 		 */
429 		memset(&mapinfo, 0, sizeof(mapinfo));
430 		error = cam_periph_mapmem(ccb, &mapinfo, maxmap);
431 		if (error)
432 			goto out;
433 
434 		/*
435 		 * Lock the periph and run the command.
436 		 */
437 		cam_periph_lock(periph);
438 		cam_periph_runccb(ccb, NULL, CAM_RETRY_SELTO,
439 		    SF_RETRY_UA | SF_NO_PRINT, NULL);
440 
441 		/*
442 		 * Tear down mapping and return status.
443 		 */
444 		cam_periph_unlock(periph);
445 		cam_periph_unmapmem(ccb, &mapinfo);
446 		error = (ccb->ccb_h.status == CAM_REQ_CMP) ? 0 : EIO;
447 out:
448 		cam_periph_lock(periph);
449 		xpt_release_ccb(ccb);
450 		cam_periph_unlock(periph);
451 		return (error);
452 	}
453 	default:
454 		break;
455 	}
456 	return (ENOTTY);
457 }
458 
459 /*
460  * Actually translate the requested transfer into one the physical driver
461  * can understand.  The transfer is described by a buf and will include
462  * only one physical transfer.
463  */
464 static void
465 ndastrategy(struct bio *bp)
466 {
467 	struct cam_periph *periph;
468 	struct nda_softc *softc;
469 
470 	periph = (struct cam_periph *)bp->bio_disk->d_drv1;
471 	softc = (struct nda_softc *)periph->softc;
472 
473 	cam_periph_lock(periph);
474 
475 	CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("ndastrategy(%p)\n", bp));
476 
477 	/*
478 	 * If the device has been made invalid, error out
479 	 */
480 	if ((periph->flags & CAM_PERIPH_INVALID) != 0) {
481 		cam_periph_unlock(periph);
482 		biofinish(bp, NULL, ENXIO);
483 		return;
484 	}
485 
486 	if (bp->bio_cmd == BIO_DELETE)
487 		softc->deletes++;
488 
489 	/*
490 	 * Place it in the queue of disk activities for this disk
491 	 */
492 	cam_iosched_queue_work(softc->cam_iosched, bp);
493 
494 	/*
495 	 * Schedule ourselves for performing the work.
496 	 */
497 	ndaschedule(periph);
498 	cam_periph_unlock(periph);
499 
500 	return;
501 }
502 
503 static int
504 ndadump(void *arg, void *virtual, vm_offset_t physical, off_t offset, size_t length)
505 {
506 	struct	    cam_periph *periph;
507 	struct	    nda_softc *softc;
508 	u_int	    secsize;
509 	struct ccb_nvmeio nvmeio;
510 	struct	    disk *dp;
511 	uint64_t    lba;
512 	uint32_t    count;
513 	int	    error = 0;
514 
515 	dp = arg;
516 	periph = dp->d_drv1;
517 	softc = (struct nda_softc *)periph->softc;
518 	secsize = softc->disk->d_sectorsize;
519 	lba = offset / secsize;
520 	count = length / secsize;
521 
522 	if ((periph->flags & CAM_PERIPH_INVALID) != 0)
523 		return (ENXIO);
524 
525 	/* xpt_get_ccb returns a zero'd allocation for the ccb, mimic that here */
526 	memset(&nvmeio, 0, sizeof(nvmeio));
527 	if (length > 0) {
528 		xpt_setup_ccb(&nvmeio.ccb_h, periph->path, CAM_PRIORITY_NORMAL);
529 		nvmeio.ccb_state = NDA_CCB_DUMP;
530 		nda_nvme_write(softc, &nvmeio, virtual, lba, length, count);
531 		error = cam_periph_runccb((union ccb *)&nvmeio, cam_periph_error,
532 		    0, SF_NO_RECOVERY | SF_NO_RETRY, NULL);
533 		if (error != 0)
534 			printf("Aborting dump due to I/O error %d.\n", error);
535 
536 		return (error);
537 	}
538 
539 	/* Flush */
540 	xpt_setup_ccb(&nvmeio.ccb_h, periph->path, CAM_PRIORITY_NORMAL);
541 
542 	nvmeio.ccb_state = NDA_CCB_DUMP;
543 	nda_nvme_flush(softc, &nvmeio);
544 	error = cam_periph_runccb((union ccb *)&nvmeio, cam_periph_error,
545 	    0, SF_NO_RECOVERY | SF_NO_RETRY, NULL);
546 	if (error != 0)
547 		xpt_print(periph->path, "flush cmd failed\n");
548 	return (error);
549 }
550 
551 static void
552 ndainit(void)
553 {
554 	cam_status status;
555 
556 	/*
557 	 * Install a global async callback.  This callback will
558 	 * receive async callbacks like "new device found".
559 	 */
560 	status = xpt_register_async(AC_FOUND_DEVICE, ndaasync, NULL, NULL);
561 
562 	if (status != CAM_REQ_CMP) {
563 		printf("nda: Failed to attach master async callback "
564 		       "due to status 0x%x!\n", status);
565 	} else if (nda_send_ordered) {
566 		/* Register our event handlers */
567 		if ((EVENTHANDLER_REGISTER(power_suspend, ndasuspend,
568 					   NULL, EVENTHANDLER_PRI_LAST)) == NULL)
569 		    printf("ndainit: power event registration failed!\n");
570 		if ((EVENTHANDLER_REGISTER(shutdown_post_sync, ndashutdown,
571 					   NULL, SHUTDOWN_PRI_DEFAULT)) == NULL)
572 		    printf("ndainit: shutdown event registration failed!\n");
573 	}
574 }
575 
576 /*
577  * Callback from GEOM, called when it has finished cleaning up its
578  * resources.
579  */
580 static void
581 ndadiskgonecb(struct disk *dp)
582 {
583 	struct cam_periph *periph;
584 
585 	periph = (struct cam_periph *)dp->d_drv1;
586 
587 	cam_periph_release(periph);
588 }
589 
590 static void
591 ndaoninvalidate(struct cam_periph *periph)
592 {
593 	struct nda_softc *softc;
594 
595 	softc = (struct nda_softc *)periph->softc;
596 
597 	/*
598 	 * De-register any async callbacks.
599 	 */
600 	xpt_register_async(0, ndaasync, periph, periph->path);
601 #ifdef CAM_IO_STATS
602 	softc->invalidations++;
603 #endif
604 
605 	/*
606 	 * Return all queued I/O with ENXIO.
607 	 * XXX Handle any transactions queued to the card
608 	 *     with XPT_ABORT_CCB.
609 	 */
610 	cam_iosched_flush(softc->cam_iosched, NULL, ENXIO);
611 
612 	disk_gone(softc->disk);
613 }
614 
615 static void
616 ndacleanup(struct cam_periph *periph)
617 {
618 	struct nda_softc *softc;
619 
620 	softc = (struct nda_softc *)periph->softc;
621 
622 	cam_periph_unlock(periph);
623 
624 	cam_iosched_fini(softc->cam_iosched);
625 
626 	/*
627 	 * If we can't free the sysctl tree, oh well...
628 	 */
629 	if ((softc->flags & NDA_FLAG_SCTX_INIT) != 0) {
630 #ifdef CAM_IO_STATS
631 		if (sysctl_ctx_free(&softc->sysctl_stats_ctx) != 0)
632 			xpt_print(periph->path,
633 			    "can't remove sysctl stats context\n");
634 #endif
635 		if (sysctl_ctx_free(&softc->sysctl_ctx) != 0)
636 			xpt_print(periph->path,
637 			    "can't remove sysctl context\n");
638 	}
639 
640 	disk_destroy(softc->disk);
641 	free(softc, M_DEVBUF);
642 	cam_periph_lock(periph);
643 }
644 
645 static void
646 ndaasync(void *callback_arg, u_int32_t code,
647 	struct cam_path *path, void *arg)
648 {
649 	struct cam_periph *periph;
650 
651 	periph = (struct cam_periph *)callback_arg;
652 	switch (code) {
653 	case AC_FOUND_DEVICE:
654 	{
655 		struct ccb_getdev *cgd;
656 		cam_status status;
657 
658 		cgd = (struct ccb_getdev *)arg;
659 		if (cgd == NULL)
660 			break;
661 
662 		if (cgd->protocol != PROTO_NVME)
663 			break;
664 
665 		/*
666 		 * Allocate a peripheral instance for
667 		 * this device and start the probe
668 		 * process.
669 		 */
670 		status = cam_periph_alloc(ndaregister, ndaoninvalidate,
671 					  ndacleanup, ndastart,
672 					  "nda", CAM_PERIPH_BIO,
673 					  path, ndaasync,
674 					  AC_FOUND_DEVICE, cgd);
675 
676 		if (status != CAM_REQ_CMP
677 		 && status != CAM_REQ_INPROG)
678 			printf("ndaasync: Unable to attach to new device "
679 				"due to status 0x%x\n", status);
680 		break;
681 	}
682 	case AC_ADVINFO_CHANGED:
683 	{
684 		uintptr_t buftype;
685 
686 		buftype = (uintptr_t)arg;
687 		if (buftype == CDAI_TYPE_PHYS_PATH) {
688 			struct nda_softc *softc;
689 
690 			softc = periph->softc;
691 			disk_attr_changed(softc->disk, "GEOM::physpath",
692 					  M_NOWAIT);
693 		}
694 		break;
695 	}
696 	case AC_LOST_DEVICE:
697 	default:
698 		cam_periph_async(periph, code, path, arg);
699 		break;
700 	}
701 }
702 
703 static void
704 ndasysctlinit(void *context, int pending)
705 {
706 	struct cam_periph *periph;
707 	struct nda_softc *softc;
708 	char tmpstr[32], tmpstr2[16];
709 
710 	periph = (struct cam_periph *)context;
711 
712 	/* periph was held for us when this task was enqueued */
713 	if ((periph->flags & CAM_PERIPH_INVALID) != 0) {
714 		cam_periph_release(periph);
715 		return;
716 	}
717 
718 	softc = (struct nda_softc *)periph->softc;
719 	snprintf(tmpstr, sizeof(tmpstr), "CAM NDA unit %d", periph->unit_number);
720 	snprintf(tmpstr2, sizeof(tmpstr2), "%d", periph->unit_number);
721 
722 	sysctl_ctx_init(&softc->sysctl_ctx);
723 	softc->flags |= NDA_FLAG_SCTX_INIT;
724 	softc->sysctl_tree = SYSCTL_ADD_NODE_WITH_LABEL(&softc->sysctl_ctx,
725 		SYSCTL_STATIC_CHILDREN(_kern_cam_nda), OID_AUTO, tmpstr2,
726 		CTLFLAG_RD | CTLFLAG_MPSAFE, 0, tmpstr, "device_index");
727 	if (softc->sysctl_tree == NULL) {
728 		printf("ndasysctlinit: unable to allocate sysctl tree\n");
729 		cam_periph_release(periph);
730 		return;
731 	}
732 
733 	SYSCTL_ADD_INT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
734 	    OID_AUTO, "unmapped_io", CTLFLAG_RD,
735 	    &softc->unmappedio, 0, "Unmapped I/O leaf");
736 
737 	SYSCTL_ADD_QUAD(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
738 	    OID_AUTO, "deletes", CTLFLAG_RD,
739 	    &softc->deletes, "Number of BIO_DELETE requests");
740 
741 	SYSCTL_ADD_UQUAD(&softc->sysctl_ctx,
742 		SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO,
743 		"trim_count", CTLFLAG_RD, &softc->trim_count,
744 		"Total number of unmap/dsm commands sent");
745 	SYSCTL_ADD_UQUAD(&softc->sysctl_ctx,
746 		SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO,
747 		"trim_ranges", CTLFLAG_RD, &softc->trim_ranges,
748 		"Total number of ranges in unmap/dsm commands");
749 	SYSCTL_ADD_UQUAD(&softc->sysctl_ctx,
750 		SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO,
751 		"trim_lbas", CTLFLAG_RD, &softc->trim_lbas,
752 		"Total lbas in the unmap/dsm commands sent");
753 
754 	SYSCTL_ADD_INT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
755 	    OID_AUTO, "rotating", CTLFLAG_RD, &nda_rotating_media, 1,
756 	    "Rotating media");
757 
758 	SYSCTL_ADD_PROC(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
759 	    OID_AUTO, "flags", CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE,
760 	    softc, 0, ndaflagssysctl, "A",
761 	    "Flags for drive");
762 
763 #ifdef CAM_IO_STATS
764 	softc->sysctl_stats_tree = SYSCTL_ADD_NODE(&softc->sysctl_stats_ctx,
765 		SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO, "stats",
766 		CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "Statistics");
767 	if (softc->sysctl_stats_tree == NULL) {
768 		printf("ndasysctlinit: unable to allocate sysctl tree for stats\n");
769 		cam_periph_release(periph);
770 		return;
771 	}
772 	SYSCTL_ADD_INT(&softc->sysctl_stats_ctx,
773 		SYSCTL_CHILDREN(softc->sysctl_stats_tree),
774 		OID_AUTO, "timeouts", CTLFLAG_RD,
775 		&softc->timeouts, 0,
776 		"Device timeouts reported by the SIM");
777 	SYSCTL_ADD_INT(&softc->sysctl_stats_ctx,
778 		SYSCTL_CHILDREN(softc->sysctl_stats_tree),
779 		OID_AUTO, "errors", CTLFLAG_RD,
780 		&softc->errors, 0,
781 		"Transport errors reported by the SIM.");
782 	SYSCTL_ADD_INT(&softc->sysctl_stats_ctx,
783 		SYSCTL_CHILDREN(softc->sysctl_stats_tree),
784 		OID_AUTO, "pack_invalidations", CTLFLAG_RD,
785 		&softc->invalidations, 0,
786 		"Device pack invalidations.");
787 #endif
788 
789 #ifdef CAM_TEST_FAILURE
790 	SYSCTL_ADD_PROC(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
791 		OID_AUTO, "invalidate", CTLTYPE_U64 | CTLFLAG_RW | CTLFLAG_MPSAFE,
792 		periph, 0, cam_periph_invalidate_sysctl, "I",
793 		"Write 1 to invalidate the drive immediately");
794 #endif
795 
796 	cam_iosched_sysctl_init(softc->cam_iosched, &softc->sysctl_ctx,
797 	    softc->sysctl_tree);
798 
799 	cam_periph_release(periph);
800 }
801 
802 static int
803 ndaflagssysctl(SYSCTL_HANDLER_ARGS)
804 {
805 	struct sbuf sbuf;
806 	struct nda_softc *softc = arg1;
807 	int error;
808 
809 	sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
810 	if (softc->flags != 0)
811 		sbuf_printf(&sbuf, "0x%b", (unsigned)softc->flags, NDA_FLAG_STRING);
812 	else
813 		sbuf_printf(&sbuf, "0");
814 	error = sbuf_finish(&sbuf);
815 	sbuf_delete(&sbuf);
816 
817 	return (error);
818 }
819 
820 static int
821 ndagetattr(struct bio *bp)
822 {
823 	int ret;
824 	struct cam_periph *periph;
825 
826 	if (g_handleattr_int(bp, "GEOM::canspeedup", nda_enable_biospeedup))
827 		return (EJUSTRETURN);
828 
829 	periph = (struct cam_periph *)bp->bio_disk->d_drv1;
830 	cam_periph_lock(periph);
831 	ret = xpt_getattr(bp->bio_data, bp->bio_length, bp->bio_attribute,
832 	    periph->path);
833 	cam_periph_unlock(periph);
834 	if (ret == 0)
835 		bp->bio_completed = bp->bio_length;
836 	return ret;
837 }
838 
839 static cam_status
840 ndaregister(struct cam_periph *periph, void *arg)
841 {
842 	struct nda_softc *softc;
843 	struct disk *disk;
844 	struct ccb_pathinq cpi;
845 	const struct nvme_namespace_data *nsd;
846 	const struct nvme_controller_data *cd;
847 	char   announce_buf[80];
848 	uint8_t flbas_fmt, lbads, vwc_present;
849 	u_int maxio;
850 	int quirks;
851 
852 	nsd = nvme_get_identify_ns(periph);
853 	cd = nvme_get_identify_cntrl(periph);
854 
855 	softc = (struct nda_softc *)malloc(sizeof(*softc), M_DEVBUF,
856 	    M_NOWAIT | M_ZERO);
857 
858 	if (softc == NULL) {
859 		printf("ndaregister: Unable to probe new device. "
860 		    "Unable to allocate softc\n");
861 		return(CAM_REQ_CMP_ERR);
862 	}
863 
864 	if (cam_iosched_init(&softc->cam_iosched, periph) != 0) {
865 		printf("ndaregister: Unable to probe new device. "
866 		       "Unable to allocate iosched memory\n");
867 		free(softc, M_DEVBUF);
868 		return(CAM_REQ_CMP_ERR);
869 	}
870 
871 	/* ident_data parsing */
872 
873 	periph->softc = softc;
874 	softc->quirks = NDA_Q_NONE;
875 	xpt_path_inq(&cpi, periph->path);
876 	TASK_INIT(&softc->sysctl_task, 0, ndasysctlinit, periph);
877 
878 	/*
879 	 * The name space ID is the lun, save it for later I/O
880 	 */
881 	softc->nsid = (uint32_t)xpt_path_lun_id(periph->path);
882 
883 	/*
884 	 * Register this media as a disk
885 	 */
886 	(void)cam_periph_hold(periph, PRIBIO);
887 	cam_periph_unlock(periph);
888 	snprintf(announce_buf, sizeof(announce_buf),
889 	    "kern.cam.nda.%d.quirks", periph->unit_number);
890 	quirks = softc->quirks;
891 	TUNABLE_INT_FETCH(announce_buf, &quirks);
892 	softc->quirks = quirks;
893 	cam_iosched_set_sort_queue(softc->cam_iosched, 0);
894 	softc->disk = disk = disk_alloc();
895 	disk->d_rotation_rate = DISK_RR_NON_ROTATING;
896 	disk->d_open = ndaopen;
897 	disk->d_close = ndaclose;
898 	disk->d_strategy = ndastrategy;
899 	disk->d_ioctl = ndaioctl;
900 	disk->d_getattr = ndagetattr;
901 	if (cam_sim_pollable(periph->sim))
902 		disk->d_dump = ndadump;
903 	disk->d_gone = ndadiskgonecb;
904 	disk->d_name = "nda";
905 	disk->d_drv1 = periph;
906 	disk->d_unit = periph->unit_number;
907 	maxio = cpi.maxio;		/* Honor max I/O size of SIM */
908 	if (maxio == 0)
909 		maxio = DFLTPHYS;	/* traditional default */
910 	else if (maxio > maxphys)
911 		maxio = maxphys;	/* for safety */
912 	disk->d_maxsize = maxio;
913 	flbas_fmt = (nsd->flbas >> NVME_NS_DATA_FLBAS_FORMAT_SHIFT) &
914 		NVME_NS_DATA_FLBAS_FORMAT_MASK;
915 	lbads = (nsd->lbaf[flbas_fmt] >> NVME_NS_DATA_LBAF_LBADS_SHIFT) &
916 		NVME_NS_DATA_LBAF_LBADS_MASK;
917 	disk->d_sectorsize = 1 << lbads;
918 	disk->d_mediasize = (off_t)(disk->d_sectorsize * nsd->nsze);
919 	disk->d_delmaxsize = disk->d_mediasize;
920 	disk->d_flags = DISKFLAG_DIRECT_COMPLETION;
921 	if (nvme_ctrlr_has_dataset_mgmt(cd))
922 		disk->d_flags |= DISKFLAG_CANDELETE;
923 	vwc_present = (cd->vwc >> NVME_CTRLR_DATA_VWC_PRESENT_SHIFT) &
924 		NVME_CTRLR_DATA_VWC_PRESENT_MASK;
925 	if (vwc_present)
926 		disk->d_flags |= DISKFLAG_CANFLUSHCACHE;
927 	if ((cpi.hba_misc & PIM_UNMAPPED) != 0) {
928 		disk->d_flags |= DISKFLAG_UNMAPPED_BIO;
929 		softc->unmappedio = 1;
930 	}
931 	/*
932 	 * d_ident and d_descr are both far bigger than the length of either
933 	 *  the serial or model number strings.
934 	 */
935 	cam_strvis(disk->d_descr, cd->mn,
936 	    NVME_MODEL_NUMBER_LENGTH, sizeof(disk->d_descr));
937 	cam_strvis(disk->d_ident, cd->sn,
938 	    NVME_SERIAL_NUMBER_LENGTH, sizeof(disk->d_ident));
939 	disk->d_hba_vendor = cpi.hba_vendor;
940 	disk->d_hba_device = cpi.hba_device;
941 	disk->d_hba_subvendor = cpi.hba_subvendor;
942 	disk->d_hba_subdevice = cpi.hba_subdevice;
943 	snprintf(disk->d_attachment, sizeof(disk->d_attachment),
944 	    "%s%d", cpi.dev_name, cpi.unit_number);
945 	if (((nsd->nsfeat >> NVME_NS_DATA_NSFEAT_NPVALID_SHIFT) &
946 	    NVME_NS_DATA_NSFEAT_NPVALID_MASK) != 0 && nsd->npwg != 0)
947 		disk->d_stripesize = ((nsd->npwg + 1) * disk->d_sectorsize);
948 	else
949 		disk->d_stripesize = nsd->noiob * disk->d_sectorsize;
950 	disk->d_stripeoffset = 0;
951 	disk->d_devstat = devstat_new_entry(periph->periph_name,
952 	    periph->unit_number, disk->d_sectorsize,
953 	    DEVSTAT_ALL_SUPPORTED,
954 	    DEVSTAT_TYPE_DIRECT | XPORT_DEVSTAT_TYPE(cpi.transport),
955 	    DEVSTAT_PRIORITY_DISK);
956 	/*
957 	 * Add alias for older nvd drives to ease transition.
958 	 */
959 	if (nda_nvd_compat)
960 		disk_add_alias(disk, "nvd");
961 
962 	/*
963 	 * Acquire a reference to the periph before we register with GEOM.
964 	 * We'll release this reference once GEOM calls us back (via
965 	 * ndadiskgonecb()) telling us that our provider has been freed.
966 	 */
967 	if (cam_periph_acquire(periph) != 0) {
968 		xpt_print(periph->path, "%s: lost periph during "
969 			  "registration!\n", __func__);
970 		cam_periph_lock(periph);
971 		return (CAM_REQ_CMP_ERR);
972 	}
973 	disk_create(softc->disk, DISK_VERSION);
974 	cam_periph_lock(periph);
975 	cam_periph_unhold(periph);
976 
977 	snprintf(announce_buf, sizeof(announce_buf),
978 		"%juMB (%ju %u byte sectors)",
979 	    (uintmax_t)((uintmax_t)disk->d_mediasize / (1024*1024)),
980 		(uintmax_t)disk->d_mediasize / disk->d_sectorsize,
981 		disk->d_sectorsize);
982 	xpt_announce_periph(periph, announce_buf);
983 	xpt_announce_quirks(periph, softc->quirks, NDA_Q_BIT_STRING);
984 
985 	/*
986 	 * Create our sysctl variables, now that we know
987 	 * we have successfully attached.
988 	 */
989 	if (cam_periph_acquire(periph) == 0)
990 		taskqueue_enqueue(taskqueue_thread, &softc->sysctl_task);
991 
992 	/*
993 	 * Register for device going away and info about the drive
994 	 * changing (though with NVMe, it can't)
995 	 */
996 	xpt_register_async(AC_LOST_DEVICE | AC_ADVINFO_CHANGED,
997 	    ndaasync, periph, periph->path);
998 
999 	softc->state = NDA_STATE_NORMAL;
1000 	return(CAM_REQ_CMP);
1001 }
1002 
1003 static void
1004 ndastart(struct cam_periph *periph, union ccb *start_ccb)
1005 {
1006 	struct nda_softc *softc = (struct nda_softc *)periph->softc;
1007 	struct ccb_nvmeio *nvmeio = &start_ccb->nvmeio;
1008 
1009 	CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("ndastart\n"));
1010 
1011 	switch (softc->state) {
1012 	case NDA_STATE_NORMAL:
1013 	{
1014 		struct bio *bp;
1015 
1016 		bp = cam_iosched_next_bio(softc->cam_iosched);
1017 		CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("ndastart: bio %p\n", bp));
1018 		if (bp == NULL) {
1019 			xpt_release_ccb(start_ccb);
1020 			break;
1021 		}
1022 
1023 		switch (bp->bio_cmd) {
1024 		case BIO_WRITE:
1025 			softc->flags |= NDA_FLAG_DIRTY;
1026 			/* FALLTHROUGH */
1027 		case BIO_READ:
1028 		{
1029 #ifdef CAM_TEST_FAILURE
1030 			int fail = 0;
1031 
1032 			/*
1033 			 * Support the failure ioctls.  If the command is a
1034 			 * read, and there are pending forced read errors, or
1035 			 * if a write and pending write errors, then fail this
1036 			 * operation with EIO.  This is useful for testing
1037 			 * purposes.  Also, support having every Nth read fail.
1038 			 *
1039 			 * This is a rather blunt tool.
1040 			 */
1041 			if (bp->bio_cmd == BIO_READ) {
1042 				if (softc->force_read_error) {
1043 					softc->force_read_error--;
1044 					fail = 1;
1045 				}
1046 				if (softc->periodic_read_error > 0) {
1047 					if (++softc->periodic_read_count >=
1048 					    softc->periodic_read_error) {
1049 						softc->periodic_read_count = 0;
1050 						fail = 1;
1051 					}
1052 				}
1053 			} else {
1054 				if (softc->force_write_error) {
1055 					softc->force_write_error--;
1056 					fail = 1;
1057 				}
1058 			}
1059 			if (fail) {
1060 				biofinish(bp, NULL, EIO);
1061 				xpt_release_ccb(start_ccb);
1062 				ndaschedule(periph);
1063 				return;
1064 			}
1065 #endif
1066 			KASSERT((bp->bio_flags & BIO_UNMAPPED) == 0 ||
1067 			    round_page(bp->bio_bcount + bp->bio_ma_offset) /
1068 			    PAGE_SIZE == bp->bio_ma_n,
1069 			    ("Short bio %p", bp));
1070 			nda_nvme_rw_bio(softc, &start_ccb->nvmeio, bp, bp->bio_cmd == BIO_READ ?
1071 			    NVME_OPC_READ : NVME_OPC_WRITE);
1072 			break;
1073 		}
1074 		case BIO_DELETE:
1075 		{
1076 			struct nvme_dsm_range *dsm_range, *dsm_end;
1077 			struct nda_trim_request *trim;
1078 			struct bio *bp1;
1079 			int ents;
1080 			uint32_t totalcount = 0, ranges = 0;
1081 
1082 			trim = malloc(sizeof(*trim), M_NVMEDA, M_ZERO | M_NOWAIT);
1083 			if (trim == NULL) {
1084 				biofinish(bp, NULL, ENOMEM);
1085 				xpt_release_ccb(start_ccb);
1086 				ndaschedule(periph);
1087 				return;
1088 			}
1089 			TAILQ_INIT(&trim->bps);
1090 			bp1 = bp;
1091 			ents = min(nitems(trim->dsm), nda_max_trim_entries);
1092 			ents = max(ents, 1);
1093 			dsm_range = trim->dsm;
1094 			dsm_end = dsm_range + ents;
1095 			do {
1096 				TAILQ_INSERT_TAIL(&trim->bps, bp1, bio_queue);
1097 				dsm_range->length =
1098 				    htole32(bp1->bio_bcount / softc->disk->d_sectorsize);
1099 				dsm_range->starting_lba =
1100 				    htole64(bp1->bio_offset / softc->disk->d_sectorsize);
1101 				ranges++;
1102 				totalcount += dsm_range->length;
1103 				dsm_range++;
1104 				if (dsm_range >= dsm_end)
1105 					break;
1106 				bp1 = cam_iosched_next_trim(softc->cam_iosched);
1107 				/* XXX -- Could collapse adjacent ranges, but we don't for now */
1108 				/* XXX -- Could limit based on total payload size */
1109 			} while (bp1 != NULL);
1110 			start_ccb->ccb_trim = trim;
1111 			nda_nvme_trim(softc, &start_ccb->nvmeio, trim->dsm,
1112 			    dsm_range - trim->dsm);
1113 			start_ccb->ccb_state = NDA_CCB_TRIM;
1114 			softc->trim_count++;
1115 			softc->trim_ranges += ranges;
1116 			softc->trim_lbas += totalcount;
1117 			/*
1118 			 * Note: We can have multiple TRIMs in flight, so we don't call
1119 			 * cam_iosched_submit_trim(softc->cam_iosched);
1120 			 * since that forces the I/O scheduler to only schedule one at a time.
1121 			 * On NVMe drives, this is a performance disaster.
1122 			 */
1123 			goto out;
1124 		}
1125 		case BIO_FLUSH:
1126 			nda_nvme_flush(softc, nvmeio);
1127 			break;
1128 		default:
1129 			biofinish(bp, NULL, EOPNOTSUPP);
1130 			xpt_release_ccb(start_ccb);
1131 			ndaschedule(periph);
1132 			return;
1133 		}
1134 		start_ccb->ccb_state = NDA_CCB_BUFFER_IO;
1135 		start_ccb->ccb_bp = bp;
1136 out:
1137 		start_ccb->ccb_h.flags |= CAM_UNLOCKED;
1138 		softc->outstanding_cmds++;
1139 		softc->refcount++;			/* For submission only */
1140 		cam_periph_unlock(periph);
1141 		xpt_action(start_ccb);
1142 		cam_periph_lock(periph);
1143 		softc->refcount--;			/* Submission done */
1144 
1145 		/* May have more work to do, so ensure we stay scheduled */
1146 		ndaschedule(periph);
1147 		break;
1148 		}
1149 	}
1150 }
1151 
1152 static void
1153 ndadone(struct cam_periph *periph, union ccb *done_ccb)
1154 {
1155 	struct nda_softc *softc;
1156 	struct ccb_nvmeio *nvmeio = &done_ccb->nvmeio;
1157 	struct cam_path *path;
1158 	int state;
1159 
1160 	softc = (struct nda_softc *)periph->softc;
1161 	path = done_ccb->ccb_h.path;
1162 
1163 	CAM_DEBUG(path, CAM_DEBUG_TRACE, ("ndadone\n"));
1164 
1165 	state = nvmeio->ccb_state & NDA_CCB_TYPE_MASK;
1166 	switch (state) {
1167 	case NDA_CCB_BUFFER_IO:
1168 	case NDA_CCB_TRIM:
1169 	{
1170 		int error;
1171 
1172 		cam_periph_lock(periph);
1173 		if ((done_ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_CMP) {
1174 			error = ndaerror(done_ccb, 0, 0);
1175 			if (error == ERESTART) {
1176 				/* A retry was scheduled, so just return. */
1177 				cam_periph_unlock(periph);
1178 				return;
1179 			}
1180 			if ((done_ccb->ccb_h.status & CAM_DEV_QFRZN) != 0)
1181 				cam_release_devq(path,
1182 						 /*relsim_flags*/0,
1183 						 /*reduction*/0,
1184 						 /*timeout*/0,
1185 						 /*getcount_only*/0);
1186 		} else {
1187 			if ((done_ccb->ccb_h.status & CAM_DEV_QFRZN) != 0)
1188 				panic("REQ_CMP with QFRZN");
1189 			error = 0;
1190 		}
1191 		if (state == NDA_CCB_BUFFER_IO) {
1192 			struct bio *bp;
1193 
1194 			bp = (struct bio *)done_ccb->ccb_bp;
1195 			bp->bio_error = error;
1196 			if (error != 0) {
1197 				bp->bio_resid = bp->bio_bcount;
1198 				bp->bio_flags |= BIO_ERROR;
1199 			} else {
1200 				bp->bio_resid = 0;
1201 			}
1202 			softc->outstanding_cmds--;
1203 
1204 			/*
1205 			 * We need to call cam_iosched before we call biodone so that we
1206 			 * don't measure any activity that happens in the completion
1207 			 * routine, which in the case of sendfile can be quite
1208 			 * extensive.
1209 			 */
1210 			cam_iosched_bio_complete(softc->cam_iosched, bp, done_ccb);
1211 			xpt_release_ccb(done_ccb);
1212 			ndaschedule(periph);
1213 			cam_periph_unlock(periph);
1214 			biodone(bp);
1215 		} else { /* state == NDA_CCB_TRIM */
1216 			struct nda_trim_request *trim;
1217 			struct bio *bp1, *bp2;
1218 			TAILQ_HEAD(, bio) queue;
1219 
1220 			trim = nvmeio->ccb_trim;
1221 			TAILQ_INIT(&queue);
1222 			TAILQ_CONCAT(&queue, &trim->bps, bio_queue);
1223 			free(trim, M_NVMEDA);
1224 
1225 			/*
1226 			 * Since we can have multiple trims in flight, we don't
1227 			 * need to call this here.
1228 			 * cam_iosched_trim_done(softc->cam_iosched);
1229 			 */
1230 			/*
1231 			 * The the I/O scheduler that we're finishing the I/O
1232 			 * so we can keep book. The first one we pass in the CCB
1233 			 * which has the timing information. The rest we pass in NULL
1234 			 * so we can keep proper counts.
1235 			 */
1236 			bp1 = TAILQ_FIRST(&queue);
1237 			cam_iosched_bio_complete(softc->cam_iosched, bp1, done_ccb);
1238 			xpt_release_ccb(done_ccb);
1239 			softc->outstanding_cmds--;
1240 			ndaschedule(periph);
1241 			cam_periph_unlock(periph);
1242 			while ((bp2 = TAILQ_FIRST(&queue)) != NULL) {
1243 				TAILQ_REMOVE(&queue, bp2, bio_queue);
1244 				bp2->bio_error = error;
1245 				if (error != 0) {
1246 					bp2->bio_flags |= BIO_ERROR;
1247 					bp2->bio_resid = bp1->bio_bcount;
1248 				} else
1249 					bp2->bio_resid = 0;
1250 				if (bp1 != bp2)
1251 					cam_iosched_bio_complete(softc->cam_iosched, bp2, NULL);
1252 				biodone(bp2);
1253 			}
1254 		}
1255 		return;
1256 	}
1257 	case NDA_CCB_DUMP:
1258 		/* No-op.  We're polling */
1259 		return;
1260 	case NDA_CCB_PASS:
1261 		/* NVME_PASSTHROUGH_CMD runs this CCB and releases it */
1262 		return;
1263 	default:
1264 		break;
1265 	}
1266 	xpt_release_ccb(done_ccb);
1267 }
1268 
1269 static int
1270 ndaerror(union ccb *ccb, u_int32_t cam_flags, u_int32_t sense_flags)
1271 {
1272 	struct nda_softc *softc;
1273 	struct cam_periph *periph;
1274 
1275 	periph = xpt_path_periph(ccb->ccb_h.path);
1276 	softc = (struct nda_softc *)periph->softc;
1277 
1278 	switch (ccb->ccb_h.status & CAM_STATUS_MASK) {
1279 	case CAM_CMD_TIMEOUT:
1280 #ifdef CAM_IO_STATS
1281 		softc->timeouts++;
1282 #endif
1283 		break;
1284 	case CAM_REQ_ABORTED:
1285 	case CAM_REQ_CMP_ERR:
1286 	case CAM_REQ_TERMIO:
1287 	case CAM_UNREC_HBA_ERROR:
1288 	case CAM_DATA_RUN_ERR:
1289 	case CAM_ATA_STATUS_ERROR:
1290 #ifdef CAM_IO_STATS
1291 		softc->errors++;
1292 #endif
1293 		break;
1294 	default:
1295 		break;
1296 	}
1297 
1298 	return(cam_periph_error(ccb, cam_flags, sense_flags));
1299 }
1300 
1301 /*
1302  * Step through all NDA peripheral drivers, and if the device is still open,
1303  * sync the disk cache to physical media.
1304  */
1305 static void
1306 ndaflush(void)
1307 {
1308 	struct cam_periph *periph;
1309 	struct nda_softc *softc;
1310 	union ccb *ccb;
1311 	int error;
1312 
1313 	CAM_PERIPH_FOREACH(periph, &ndadriver) {
1314 		softc = (struct nda_softc *)periph->softc;
1315 
1316 		if (SCHEDULER_STOPPED()) {
1317 			/*
1318 			 * If we paniced with the lock held or the periph is not
1319 			 * open, do not recurse.  Otherwise, call ndadump since
1320 			 * that avoids the sleeping cam_periph_getccb does if no
1321 			 * CCBs are available.
1322 			 */
1323 			if (!cam_periph_owned(periph) &&
1324 			    (softc->flags & NDA_FLAG_OPEN)) {
1325 				ndadump(softc->disk, NULL, 0, 0, 0);
1326 			}
1327 			continue;
1328 		}
1329 
1330 		/*
1331 		 * We only sync the cache if the drive is still open
1332 		 */
1333 		cam_periph_lock(periph);
1334 		if ((softc->flags & NDA_FLAG_OPEN) == 0) {
1335 			cam_periph_unlock(periph);
1336 			continue;
1337 		}
1338 
1339 		ccb = cam_periph_getccb(periph, CAM_PRIORITY_NORMAL);
1340 		nda_nvme_flush(softc, &ccb->nvmeio);
1341 		error = cam_periph_runccb(ccb, ndaerror, /*cam_flags*/0,
1342 		    /*sense_flags*/ SF_NO_RECOVERY | SF_NO_RETRY,
1343 		    softc->disk->d_devstat);
1344 		if (error != 0)
1345 			xpt_print(periph->path, "Synchronize cache failed\n");
1346 		xpt_release_ccb(ccb);
1347 		cam_periph_unlock(periph);
1348 	}
1349 }
1350 
1351 static void
1352 ndashutdown(void *arg, int howto)
1353 {
1354 
1355 	ndaflush();
1356 }
1357 
1358 static void
1359 ndasuspend(void *arg)
1360 {
1361 
1362 	ndaflush();
1363 }
1364