xref: /freebsd/sys/cam/nvme/nvme_da.c (revision 2b4dbad2db5766294ee97bb96228ec6826a9e7c3)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2015 Netflix, Inc.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  *
27  * Derived from ata_da.c:
28  * Copyright (c) 2009 Alexander Motin <mav@FreeBSD.org>
29  */
30 
31 #include <sys/param.h>
32 
33 #ifdef _KERNEL
34 #include <sys/systm.h>
35 #include <sys/kernel.h>
36 #include <sys/bio.h>
37 #include <sys/sysctl.h>
38 #include <sys/taskqueue.h>
39 #include <sys/lock.h>
40 #include <sys/mutex.h>
41 #include <sys/conf.h>
42 #include <sys/devicestat.h>
43 #include <sys/eventhandler.h>
44 #include <sys/malloc.h>
45 #include <sys/cons.h>
46 #include <sys/power.h>
47 #include <sys/proc.h>
48 #include <sys/reboot.h>
49 #include <sys/sbuf.h>
50 #include <geom/geom.h>
51 #include <geom/geom_disk.h>
52 #endif /* _KERNEL */
53 
54 #ifndef _KERNEL
55 #include <stdio.h>
56 #include <string.h>
57 #endif /* _KERNEL */
58 
59 #include <cam/cam.h>
60 #include <cam/cam_ccb.h>
61 #include <cam/cam_periph.h>
62 #include <cam/cam_xpt_periph.h>
63 #include <cam/cam_sim.h>
64 #include <cam/cam_iosched.h>
65 
66 #include <cam/nvme/nvme_all.h>
67 
68 typedef enum {
69 	NDA_STATE_NORMAL
70 } nda_state;
71 
72 typedef enum {
73 	NDA_FLAG_OPEN		= 0x0001,
74 	NDA_FLAG_DIRTY		= 0x0002,
75 	NDA_FLAG_SCTX_INIT	= 0x0004,
76 } nda_flags;
77 #define NDA_FLAG_STRING		\
78 	"\020"			\
79 	"\001OPEN"		\
80 	"\002DIRTY"		\
81 	"\003SCTX_INIT"
82 
83 typedef enum {
84 	NDA_Q_4K   = 0x01,
85 	NDA_Q_NONE = 0x00,
86 } nda_quirks;
87 
88 #define NDA_Q_BIT_STRING	\
89 	"\020"			\
90 	"\001Bit 0"
91 
92 typedef enum {
93 	NDA_CCB_BUFFER_IO	= 0x01,
94 	NDA_CCB_DUMP            = 0x02,
95 	NDA_CCB_TRIM            = 0x03,
96 	NDA_CCB_PASS            = 0x04,
97 	NDA_CCB_TYPE_MASK	= 0x0F,
98 } nda_ccb_state;
99 
100 /* Offsets into our private area for storing information */
101 #define ccb_state	ccb_h.ppriv_field0
102 #define ccb_bp		ccb_h.ppriv_ptr1	/* For NDA_CCB_BUFFER_IO */
103 #define ccb_trim	ccb_h.ppriv_ptr1	/* For NDA_CCB_TRIM */
104 
105 struct nda_softc {
106 	struct   cam_iosched_softc *cam_iosched;
107 	int			outstanding_cmds;	/* Number of active commands */
108 	int			refcount;		/* Active xpt_action() calls */
109 	nda_state		state;
110 	nda_flags		flags;
111 	nda_quirks		quirks;
112 	int			unmappedio;
113 	quad_t			deletes;
114 	uint32_t		nsid;			/* Namespace ID for this nda device */
115 	struct disk		*disk;
116 	struct task		sysctl_task;
117 	struct sysctl_ctx_list	sysctl_ctx;
118 	struct sysctl_oid	*sysctl_tree;
119 	uint64_t		trim_count;
120 	uint64_t		trim_ranges;
121 	uint64_t		trim_lbas;
122 #ifdef CAM_TEST_FAILURE
123 	int			force_read_error;
124 	int			force_write_error;
125 	int			periodic_read_error;
126 	int			periodic_read_count;
127 #endif
128 #ifdef CAM_IO_STATS
129 	struct sysctl_ctx_list	sysctl_stats_ctx;
130 	struct sysctl_oid	*sysctl_stats_tree;
131 	u_int			timeouts;
132 	u_int			errors;
133 	u_int			invalidations;
134 #endif
135 };
136 
137 struct nda_trim_request {
138 	struct nvme_dsm_range	dsm[NVME_MAX_DSM_TRIM / sizeof(struct nvme_dsm_range)];
139 	TAILQ_HEAD(, bio) bps;
140 };
141 _Static_assert(NVME_MAX_DSM_TRIM % sizeof(struct nvme_dsm_range) == 0,
142     "NVME_MAX_DSM_TRIM must be an integral number of ranges");
143 
144 /* Need quirk table */
145 
146 static	disk_ioctl_t	ndaioctl;
147 static	disk_strategy_t	ndastrategy;
148 static	dumper_t	ndadump;
149 static	periph_init_t	ndainit;
150 static	void		ndaasync(void *callback_arg, uint32_t code,
151 				struct cam_path *path, void *arg);
152 static	void		ndasysctlinit(void *context, int pending);
153 static	int		ndaflagssysctl(SYSCTL_HANDLER_ARGS);
154 static	periph_ctor_t	ndaregister;
155 static	periph_dtor_t	ndacleanup;
156 static	periph_start_t	ndastart;
157 static	periph_oninv_t	ndaoninvalidate;
158 static	void		ndadone(struct cam_periph *periph,
159 			       union ccb *done_ccb);
160 static  int		ndaerror(union ccb *ccb, uint32_t cam_flags,
161 				uint32_t sense_flags);
162 static void		ndashutdown(void *arg, int howto);
163 static void		ndasuspend(void *arg, enum power_stype stype);
164 
165 #ifndef	NDA_DEFAULT_SEND_ORDERED
166 #define	NDA_DEFAULT_SEND_ORDERED	1
167 #endif
168 #ifndef NDA_DEFAULT_TIMEOUT
169 #define NDA_DEFAULT_TIMEOUT 30	/* Timeout in seconds */
170 #endif
171 #ifndef	NDA_DEFAULT_RETRY
172 #define	NDA_DEFAULT_RETRY	4
173 #endif
174 #ifndef NDA_MAX_TRIM_ENTRIES
175 #define NDA_MAX_TRIM_ENTRIES  (NVME_MAX_DSM_TRIM / sizeof(struct nvme_dsm_range))/* Number of DSM trims to use, max 256 */
176 #endif
177 
178 static SYSCTL_NODE(_kern_cam, OID_AUTO, nda, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
179     "CAM Direct Access Disk driver");
180 
181 //static int nda_retry_count = NDA_DEFAULT_RETRY;
182 static int nda_send_ordered = NDA_DEFAULT_SEND_ORDERED;
183 static int nda_default_timeout = NDA_DEFAULT_TIMEOUT;
184 static int nda_max_trim_entries = NDA_MAX_TRIM_ENTRIES;
185 static int nda_enable_biospeedup = 1;
186 static int nda_nvd_compat = 1;
187 SYSCTL_INT(_kern_cam_nda, OID_AUTO, max_trim, CTLFLAG_RDTUN,
188     &nda_max_trim_entries, NDA_MAX_TRIM_ENTRIES,
189     "Maximum number of BIO_DELETE to send down as a DSM TRIM.");
190 SYSCTL_INT(_kern_cam_nda, OID_AUTO, enable_biospeedup, CTLFLAG_RDTUN,
191     &nda_enable_biospeedup, 0, "Enable BIO_SPEEDUP processing.");
192 SYSCTL_INT(_kern_cam_nda, OID_AUTO, nvd_compat, CTLFLAG_RDTUN,
193     &nda_nvd_compat, 1, "Enable creation of nvd aliases.");
194 
195 /*
196  * All NVMe media is non-rotational, so all nvme device instances
197  * share this to implement the sysctl.
198  */
199 static int nda_rotating_media = 0;
200 
201 static struct periph_driver ndadriver =
202 {
203 	ndainit, "nda",
204 	TAILQ_HEAD_INITIALIZER(ndadriver.units), /* generation */ 0
205 };
206 
207 PERIPHDRIVER_DECLARE(nda, ndadriver);
208 
209 static MALLOC_DEFINE(M_NVMEDA, "nvme_da", "nvme_da buffers");
210 
211 /*
212  * nice wrappers. Maybe these belong in nvme_all.c instead of
213  * here, but this is the only place that uses these. Should
214  * we ever grow another NVME periph, we should move them
215  * all there wholesale.
216  */
217 
218 static void
nda_nvme_flush(struct nda_softc * softc,struct ccb_nvmeio * nvmeio)219 nda_nvme_flush(struct nda_softc *softc, struct ccb_nvmeio *nvmeio)
220 {
221 	cam_fill_nvmeio(nvmeio,
222 	    0,			/* retries */
223 	    ndadone,		/* cbfcnp */
224 	    CAM_DIR_NONE,	/* flags */
225 	    NULL,		/* data_ptr */
226 	    0,			/* dxfer_len */
227 	    nda_default_timeout * 1000); /* timeout 30s */
228 	nvme_ns_flush_cmd(&nvmeio->cmd, softc->nsid);
229 }
230 
231 static void
nda_nvme_trim(struct nda_softc * softc,struct ccb_nvmeio * nvmeio,void * payload,uint32_t num_ranges)232 nda_nvme_trim(struct nda_softc *softc, struct ccb_nvmeio *nvmeio,
233     void *payload, uint32_t num_ranges)
234 {
235 	cam_fill_nvmeio(nvmeio,
236 	    0,			/* retries */
237 	    ndadone,		/* cbfcnp */
238 	    CAM_DIR_OUT,	/* flags */
239 	    payload,		/* data_ptr */
240 	    num_ranges * sizeof(struct nvme_dsm_range), /* dxfer_len */
241 	    nda_default_timeout * 1000); /* timeout 30s */
242 	nvme_ns_trim_cmd(&nvmeio->cmd, softc->nsid, num_ranges);
243 }
244 
245 static void
nda_nvme_write(struct nda_softc * softc,struct ccb_nvmeio * nvmeio,void * payload,uint64_t lba,uint32_t len,uint32_t count)246 nda_nvme_write(struct nda_softc *softc, struct ccb_nvmeio *nvmeio,
247     void *payload, uint64_t lba, uint32_t len, uint32_t count)
248 {
249 	cam_fill_nvmeio(nvmeio,
250 	    0,			/* retries */
251 	    ndadone,		/* cbfcnp */
252 	    CAM_DIR_OUT,	/* flags */
253 	    payload,		/* data_ptr */
254 	    len,		/* dxfer_len */
255 	    nda_default_timeout * 1000); /* timeout 30s */
256 	nvme_ns_write_cmd(&nvmeio->cmd, softc->nsid, lba, count);
257 }
258 
259 static void
nda_nvme_rw_bio(struct nda_softc * softc,struct ccb_nvmeio * nvmeio,struct bio * bp,uint32_t rwcmd)260 nda_nvme_rw_bio(struct nda_softc *softc, struct ccb_nvmeio *nvmeio,
261     struct bio *bp, uint32_t rwcmd)
262 {
263 	int flags = rwcmd == NVME_OPC_READ ? CAM_DIR_IN : CAM_DIR_OUT;
264 	void *payload;
265 	uint64_t lba;
266 	uint32_t count;
267 
268 	if (bp->bio_flags & BIO_UNMAPPED) {
269 		flags |= CAM_DATA_BIO;
270 		payload = bp;
271 	} else {
272 		payload = bp->bio_data;
273 	}
274 
275 	lba = bp->bio_pblkno;
276 	count = bp->bio_bcount / softc->disk->d_sectorsize;
277 
278 	cam_fill_nvmeio(nvmeio,
279 	    0,			/* retries */
280 	    ndadone,		/* cbfcnp */
281 	    flags,		/* flags */
282 	    payload,		/* data_ptr */
283 	    bp->bio_bcount,	/* dxfer_len */
284 	    nda_default_timeout * 1000); /* timeout 30s */
285 	nvme_ns_rw_cmd(&nvmeio->cmd, rwcmd, softc->nsid, lba, count);
286 }
287 
288 static int
ndaopen(struct disk * dp)289 ndaopen(struct disk *dp)
290 {
291 	struct cam_periph *periph;
292 	struct nda_softc *softc;
293 	int error;
294 
295 	periph = (struct cam_periph *)dp->d_drv1;
296 	if (cam_periph_acquire(periph) != 0) {
297 		return(ENXIO);
298 	}
299 
300 	cam_periph_lock(periph);
301 	if ((error = cam_periph_hold(periph, PRIBIO|PCATCH)) != 0) {
302 		cam_periph_unlock(periph);
303 		cam_periph_release(periph);
304 		return (error);
305 	}
306 
307 	CAM_DEBUG(periph->path, CAM_DEBUG_TRACE | CAM_DEBUG_PERIPH,
308 	    ("ndaopen\n"));
309 
310 	softc = (struct nda_softc *)periph->softc;
311 	softc->flags |= NDA_FLAG_OPEN;
312 
313 	cam_periph_unhold(periph);
314 	cam_periph_unlock(periph);
315 	return (0);
316 }
317 
318 static int
ndaclose(struct disk * dp)319 ndaclose(struct disk *dp)
320 {
321 	struct	cam_periph *periph;
322 	struct	nda_softc *softc;
323 	union ccb *ccb;
324 	int error;
325 
326 	periph = (struct cam_periph *)dp->d_drv1;
327 	softc = (struct nda_softc *)periph->softc;
328 	cam_periph_lock(periph);
329 
330 	CAM_DEBUG(periph->path, CAM_DEBUG_TRACE | CAM_DEBUG_PERIPH,
331 	    ("ndaclose\n"));
332 
333 	if ((softc->flags & NDA_FLAG_DIRTY) != 0 &&
334 	    (periph->flags & CAM_PERIPH_INVALID) == 0 &&
335 	    cam_periph_hold(periph, PRIBIO) == 0) {
336 		ccb = cam_periph_getccb(periph, CAM_PRIORITY_NORMAL);
337 		nda_nvme_flush(softc, &ccb->nvmeio);
338 		error = cam_periph_runccb(ccb, ndaerror, /*cam_flags*/0,
339 		    /*sense_flags*/0, softc->disk->d_devstat);
340 
341 		if (error != 0)
342 			xpt_print(periph->path, "Synchronize cache failed\n");
343 		else
344 			softc->flags &= ~NDA_FLAG_DIRTY;
345 		xpt_release_ccb(ccb);
346 		cam_periph_unhold(periph);
347 	}
348 
349 	softc->flags &= ~NDA_FLAG_OPEN;
350 
351 	while (softc->refcount != 0)
352 		cam_periph_sleep(periph, &softc->refcount, PRIBIO, "ndaclose", 1);
353 	KASSERT(softc->outstanding_cmds == 0,
354 	    ("nda %d outstanding commands", softc->outstanding_cmds));
355 	cam_periph_unlock(periph);
356 	cam_periph_release(periph);
357 	return (0);
358 }
359 
360 static void
ndaschedule(struct cam_periph * periph)361 ndaschedule(struct cam_periph *periph)
362 {
363 	struct nda_softc *softc = (struct nda_softc *)periph->softc;
364 
365 	if (softc->state != NDA_STATE_NORMAL)
366 		return;
367 
368 	cam_iosched_schedule(softc->cam_iosched, periph);
369 }
370 
371 static int
ndaioctl(struct disk * dp,u_long cmd,void * data,int fflag,struct thread * td)372 ndaioctl(struct disk *dp, u_long cmd, void *data, int fflag,
373     struct thread *td)
374 {
375 	struct cam_periph *periph;
376 
377 	periph = (struct cam_periph *)dp->d_drv1;
378 
379 	switch (cmd) {
380 	case NVME_IO_TEST:
381 	case NVME_BIO_TEST:
382 		/*
383 		 * These don't map well to the underlying CCBs, so
384 		 * they are usupported via CAM.
385 		 */
386 		return (ENOTTY);
387 	case NVME_GET_NSID:
388 	{
389 		struct nvme_get_nsid *gnsid = (struct nvme_get_nsid *)data;
390 		struct ccb_pathinq cpi;
391 
392 		xpt_path_inq(&cpi, periph->path);
393 		strncpy(gnsid->cdev, cpi.xport_specific.nvme.dev_name,
394 		    sizeof(gnsid->cdev));
395 		gnsid->nsid = cpi.xport_specific.nvme.nsid;
396 		return (0);
397 	}
398 	case NVME_PASSTHROUGH_CMD:
399 	{
400 		struct nvme_pt_command *pt;
401 		union ccb *ccb;
402 		struct cam_periph_map_info mapinfo;
403 		u_int maxmap = dp->d_maxsize;
404 		int error;
405 
406 		/*
407 		 * Create a NVME_IO CCB to do the passthrough command.
408 		 */
409 		pt = (struct nvme_pt_command *)data;
410 		ccb = xpt_alloc_ccb();
411 		xpt_setup_ccb(&ccb->ccb_h, periph->path, CAM_PRIORITY_NORMAL);
412 		ccb->ccb_state = NDA_CCB_PASS;
413 		cam_fill_nvmeio(&ccb->nvmeio,
414 		    0,			/* Retries */
415 		    ndadone,
416 		    (pt->is_read ? CAM_DIR_IN : CAM_DIR_OUT) | CAM_DATA_VADDR,
417 		    pt->buf,
418 		    pt->len,
419 		    nda_default_timeout * 1000);
420 		memcpy(&ccb->nvmeio.cmd, &pt->cmd, sizeof(pt->cmd));
421 
422 		/*
423 		 * Wire the user memory in this request for the I/O
424 		 */
425 		memset(&mapinfo, 0, sizeof(mapinfo));
426 		error = cam_periph_mapmem(ccb, &mapinfo, maxmap);
427 		if (error)
428 			goto out;
429 
430 		/*
431 		 * Lock the periph and run the command.
432 		 */
433 		cam_periph_lock(periph);
434 		cam_periph_runccb(ccb, NULL, CAM_RETRY_SELTO,
435 		    SF_RETRY_UA | SF_NO_PRINT, NULL);
436 
437 		/*
438 		 * Tear down mapping and return status.
439 		 */
440 		cam_periph_unlock(periph);
441 		error = cam_periph_unmapmem(ccb, &mapinfo);
442 		if (!cam_ccb_success(ccb))
443 			error = EIO;
444 out:
445 		cam_periph_lock(periph);
446 		xpt_release_ccb(ccb);
447 		cam_periph_unlock(periph);
448 		return (error);
449 	}
450 	default:
451 		break;
452 	}
453 	return (ENOTTY);
454 }
455 
456 /*
457  * Actually translate the requested transfer into one the physical driver
458  * can understand.  The transfer is described by a buf and will include
459  * only one physical transfer.
460  */
461 static void
ndastrategy(struct bio * bp)462 ndastrategy(struct bio *bp)
463 {
464 	struct cam_periph *periph;
465 	struct nda_softc *softc;
466 
467 	periph = (struct cam_periph *)bp->bio_disk->d_drv1;
468 	softc = (struct nda_softc *)periph->softc;
469 
470 	cam_periph_lock(periph);
471 
472 	CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("ndastrategy(%p)\n", bp));
473 
474 	/*
475 	 * If the device has been made invalid, error out
476 	 */
477 	if ((periph->flags & CAM_PERIPH_INVALID) != 0) {
478 		cam_periph_unlock(periph);
479 		biofinish(bp, NULL, ENXIO);
480 		return;
481 	}
482 
483 	if (bp->bio_cmd == BIO_DELETE)
484 		softc->deletes++;
485 
486 	/*
487 	 * Place it in the queue of disk activities for this disk
488 	 */
489 	cam_iosched_queue_work(softc->cam_iosched, bp);
490 
491 	/*
492 	 * Schedule ourselves for performing the work.
493 	 */
494 	ndaschedule(periph);
495 	cam_periph_unlock(periph);
496 
497 	return;
498 }
499 
500 static int
ndadump(void * arg,void * virtual,off_t offset,size_t length)501 ndadump(void *arg, void *virtual, off_t offset, size_t length)
502 {
503 	struct	    cam_periph *periph;
504 	struct	    nda_softc *softc;
505 	u_int	    secsize;
506 	struct ccb_nvmeio nvmeio;
507 	struct	    disk *dp;
508 	uint64_t    lba;
509 	uint32_t    count;
510 	int	    error = 0;
511 
512 	dp = arg;
513 	periph = dp->d_drv1;
514 	softc = (struct nda_softc *)periph->softc;
515 	secsize = softc->disk->d_sectorsize;
516 	lba = offset / secsize;
517 	count = length / secsize;
518 
519 	if ((periph->flags & CAM_PERIPH_INVALID) != 0)
520 		return (ENXIO);
521 
522 	/* xpt_get_ccb returns a zero'd allocation for the ccb, mimic that here */
523 	memset(&nvmeio, 0, sizeof(nvmeio));
524 	if (length > 0) {
525 		xpt_setup_ccb(&nvmeio.ccb_h, periph->path, CAM_PRIORITY_NORMAL);
526 		nvmeio.ccb_state = NDA_CCB_DUMP;
527 		nda_nvme_write(softc, &nvmeio, virtual, lba, length, count);
528 		error = cam_periph_runccb((union ccb *)&nvmeio, cam_periph_error,
529 		    0, SF_NO_RECOVERY | SF_NO_RETRY, NULL);
530 		if (error != 0)
531 			printf("Aborting dump due to I/O error %d.\n", error);
532 
533 		return (error);
534 	}
535 
536 	/* Flush */
537 	xpt_setup_ccb(&nvmeio.ccb_h, periph->path, CAM_PRIORITY_NORMAL);
538 
539 	nvmeio.ccb_state = NDA_CCB_DUMP;
540 	nda_nvme_flush(softc, &nvmeio);
541 	error = cam_periph_runccb((union ccb *)&nvmeio, cam_periph_error,
542 	    0, SF_NO_RECOVERY | SF_NO_RETRY, NULL);
543 	if (error != 0)
544 		xpt_print(periph->path, "flush cmd failed\n");
545 	return (error);
546 }
547 
548 static void
ndainit(void)549 ndainit(void)
550 {
551 	cam_status status;
552 
553 	/*
554 	 * Install a global async callback.  This callback will
555 	 * receive async callbacks like "new device found".
556 	 */
557 	status = xpt_register_async(AC_FOUND_DEVICE, ndaasync, NULL, NULL);
558 
559 	if (status != CAM_REQ_CMP) {
560 		printf("nda: Failed to attach master async callback "
561 		       "due to status 0x%x!\n", status);
562 	} else if (nda_send_ordered) {
563 		/* Register our event handlers */
564 		if ((EVENTHANDLER_REGISTER(power_suspend, ndasuspend,
565 					   NULL, EVENTHANDLER_PRI_LAST)) == NULL)
566 		    printf("ndainit: power event registration failed!\n");
567 		if ((EVENTHANDLER_REGISTER(shutdown_post_sync, ndashutdown,
568 					   NULL, SHUTDOWN_PRI_DEFAULT)) == NULL)
569 		    printf("ndainit: shutdown event registration failed!\n");
570 	}
571 }
572 
573 /*
574  * Callback from GEOM, called when it has finished cleaning up its
575  * resources.
576  */
577 static void
ndadiskgonecb(struct disk * dp)578 ndadiskgonecb(struct disk *dp)
579 {
580 	struct cam_periph *periph;
581 
582 	periph = (struct cam_periph *)dp->d_drv1;
583 
584 	cam_periph_release(periph);
585 }
586 
587 static void
ndaoninvalidate(struct cam_periph * periph)588 ndaoninvalidate(struct cam_periph *periph)
589 {
590 	struct nda_softc *softc;
591 
592 	softc = (struct nda_softc *)periph->softc;
593 
594 	/*
595 	 * De-register any async callbacks.
596 	 */
597 	xpt_register_async(0, ndaasync, periph, periph->path);
598 #ifdef CAM_IO_STATS
599 	softc->invalidations++;
600 #endif
601 
602 	/*
603 	 * Return all queued I/O with ENXIO. Transactions may be queued up here
604 	 * for retry (since we are called while there's other transactions
605 	 * pending). Any requests in the hardware will drain before ndacleanup
606 	 * is called.
607 	 */
608 	cam_iosched_flush(softc->cam_iosched, NULL, ENXIO);
609 
610 	/*
611 	 * Tell GEOM that we've gone away, we'll get a callback when it is
612 	 * done cleaning up its resources.
613 	 */
614 	disk_gone(softc->disk);
615 }
616 
617 static void
ndacleanup(struct cam_periph * periph)618 ndacleanup(struct cam_periph *periph)
619 {
620 	struct nda_softc *softc;
621 
622 	softc = (struct nda_softc *)periph->softc;
623 
624 	cam_periph_unlock(periph);
625 
626 	cam_iosched_fini(softc->cam_iosched);
627 
628 	/*
629 	 * If we can't free the sysctl tree, oh well...
630 	 */
631 	if ((softc->flags & NDA_FLAG_SCTX_INIT) != 0) {
632 #ifdef CAM_IO_STATS
633 		if (sysctl_ctx_free(&softc->sysctl_stats_ctx) != 0)
634 			xpt_print(periph->path,
635 			    "can't remove sysctl stats context\n");
636 #endif
637 		if (sysctl_ctx_free(&softc->sysctl_ctx) != 0)
638 			xpt_print(periph->path,
639 			    "can't remove sysctl context\n");
640 	}
641 
642 	disk_destroy(softc->disk);
643 	free(softc, M_DEVBUF);
644 	cam_periph_lock(periph);
645 }
646 
647 static void
ndasetgeom(struct nda_softc * softc,struct cam_periph * periph)648 ndasetgeom(struct nda_softc *softc, struct cam_periph *periph)
649 {
650 	struct disk *disk = softc->disk;
651 	const struct nvme_namespace_data *nsd;
652 	const struct nvme_controller_data *cd;
653 	uint8_t flbas_fmt, lbads, vwc_present;
654 
655 	nsd = nvme_get_identify_ns(periph);
656         cd = nvme_get_identify_cntrl(periph);
657 
658 	flbas_fmt = NVMEV(NVME_NS_DATA_FLBAS_FORMAT, nsd->flbas);
659 	lbads = NVMEV(NVME_NS_DATA_LBAF_LBADS, nsd->lbaf[flbas_fmt]);
660 	disk->d_sectorsize = 1 << lbads;
661 	disk->d_mediasize = (off_t)(disk->d_sectorsize * nsd->nsze);
662 	disk->d_delmaxsize = disk->d_mediasize;
663 	disk->d_flags = DISKFLAG_DIRECT_COMPLETION;
664 	if (nvme_ctrlr_has_dataset_mgmt(cd))
665 		disk->d_flags |= DISKFLAG_CANDELETE;
666 	vwc_present = NVMEV(NVME_CTRLR_DATA_VWC_PRESENT, cd->vwc);
667 	if (vwc_present)
668 		disk->d_flags |= DISKFLAG_CANFLUSHCACHE;
669 }
670 
671 static void
ndaasync(void * callback_arg,uint32_t code,struct cam_path * path,void * arg)672 ndaasync(void *callback_arg, uint32_t code, struct cam_path *path, void *arg)
673 {
674 	struct cam_periph *periph = callback_arg;
675 	struct nda_softc *softc;
676 
677 	switch (code) {
678 	case AC_FOUND_DEVICE:
679 	{
680 		struct ccb_getdev *cgd;
681 		cam_status status;
682 
683 		cgd = (struct ccb_getdev *)arg;
684 		if (cgd == NULL)
685 			break;
686 
687 		if (cgd->protocol != PROTO_NVME)
688 			break;
689 
690 		/*
691 		 * Allocate a peripheral instance for
692 		 * this device and start the probe
693 		 * process.
694 		 */
695 		status = cam_periph_alloc(ndaregister, ndaoninvalidate,
696 					  ndacleanup, ndastart,
697 					  "nda", CAM_PERIPH_BIO,
698 					  path, ndaasync,
699 					  AC_FOUND_DEVICE, cgd);
700 
701 		if (status != CAM_REQ_CMP
702 		 && status != CAM_REQ_INPROG)
703 			printf("ndaasync: Unable to attach to new device "
704 				"due to status 0x%x\n", status);
705 		break;
706 	}
707 	case AC_GETDEV_CHANGED:
708 	{
709 		int error;
710 
711 		softc = periph->softc;
712 		ndasetgeom(softc, periph);
713 		error = disk_resize(softc->disk, M_NOWAIT);
714 		if (error != 0) {
715 			xpt_print(periph->path, "disk_resize(9) failed, error = %d\n", error);
716 			break;
717 		}
718 		break;
719 
720 	}
721 	case AC_ADVINFO_CHANGED:
722 	{
723 		uintptr_t buftype;
724 
725 		softc = periph->softc;
726 		buftype = (uintptr_t)arg;
727 		if (buftype == CDAI_TYPE_PHYS_PATH) {
728 			disk_attr_changed(softc->disk, "GEOM::physpath",
729 			    M_NOWAIT);
730 		}
731 		break;
732 	}
733 	case AC_LOST_DEVICE:
734 	default:
735 		break;
736 	}
737 	cam_periph_async(periph, code, path, arg);
738 }
739 
740 static void
ndasysctlinit(void * context,int pending)741 ndasysctlinit(void *context, int pending)
742 {
743 	struct cam_periph *periph;
744 	struct nda_softc *softc;
745 	char tmpstr[32], tmpstr2[16];
746 
747 	periph = (struct cam_periph *)context;
748 
749 	/* periph was held for us when this task was enqueued */
750 	if ((periph->flags & CAM_PERIPH_INVALID) != 0) {
751 		cam_periph_release(periph);
752 		return;
753 	}
754 
755 	softc = (struct nda_softc *)periph->softc;
756 	snprintf(tmpstr, sizeof(tmpstr), "CAM NDA unit %d", periph->unit_number);
757 	snprintf(tmpstr2, sizeof(tmpstr2), "%d", periph->unit_number);
758 
759 	sysctl_ctx_init(&softc->sysctl_ctx);
760 	softc->flags |= NDA_FLAG_SCTX_INIT;
761 	softc->sysctl_tree = SYSCTL_ADD_NODE_WITH_LABEL(&softc->sysctl_ctx,
762 		SYSCTL_STATIC_CHILDREN(_kern_cam_nda), OID_AUTO, tmpstr2,
763 		CTLFLAG_RD | CTLFLAG_MPSAFE, 0, tmpstr, "device_index");
764 	if (softc->sysctl_tree == NULL) {
765 		printf("ndasysctlinit: unable to allocate sysctl tree\n");
766 		cam_periph_release(periph);
767 		return;
768 	}
769 
770 	SYSCTL_ADD_INT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
771 	    OID_AUTO, "unmapped_io", CTLFLAG_RD,
772 	    &softc->unmappedio, 0, "Unmapped I/O leaf");
773 
774 	SYSCTL_ADD_QUAD(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
775 	    OID_AUTO, "deletes", CTLFLAG_RD,
776 	    &softc->deletes, "Number of BIO_DELETE requests");
777 
778 	SYSCTL_ADD_UQUAD(&softc->sysctl_ctx,
779 		SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO,
780 		"trim_count", CTLFLAG_RD, &softc->trim_count,
781 		"Total number of unmap/dsm commands sent");
782 	SYSCTL_ADD_UQUAD(&softc->sysctl_ctx,
783 		SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO,
784 		"trim_ranges", CTLFLAG_RD, &softc->trim_ranges,
785 		"Total number of ranges in unmap/dsm commands");
786 	SYSCTL_ADD_UQUAD(&softc->sysctl_ctx,
787 		SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO,
788 		"trim_lbas", CTLFLAG_RD, &softc->trim_lbas,
789 		"Total lbas in the unmap/dsm commands sent");
790 
791 	SYSCTL_ADD_INT(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
792 	    OID_AUTO, "rotating", CTLFLAG_RD, &nda_rotating_media, 1,
793 	    "Rotating media");
794 
795 	SYSCTL_ADD_PROC(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
796 	    OID_AUTO, "flags", CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE,
797 	    softc, 0, ndaflagssysctl, "A",
798 	    "Flags for drive");
799 
800 #ifdef CAM_IO_STATS
801 	softc->sysctl_stats_tree = SYSCTL_ADD_NODE(&softc->sysctl_stats_ctx,
802 		SYSCTL_CHILDREN(softc->sysctl_tree), OID_AUTO, "stats",
803 		CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "Statistics");
804 	if (softc->sysctl_stats_tree == NULL) {
805 		printf("ndasysctlinit: unable to allocate sysctl tree for stats\n");
806 		cam_periph_release(periph);
807 		return;
808 	}
809 	SYSCTL_ADD_INT(&softc->sysctl_stats_ctx,
810 		SYSCTL_CHILDREN(softc->sysctl_stats_tree),
811 		OID_AUTO, "timeouts", CTLFLAG_RD,
812 		&softc->timeouts, 0,
813 		"Device timeouts reported by the SIM");
814 	SYSCTL_ADD_INT(&softc->sysctl_stats_ctx,
815 		SYSCTL_CHILDREN(softc->sysctl_stats_tree),
816 		OID_AUTO, "errors", CTLFLAG_RD,
817 		&softc->errors, 0,
818 		"Transport errors reported by the SIM.");
819 	SYSCTL_ADD_INT(&softc->sysctl_stats_ctx,
820 		SYSCTL_CHILDREN(softc->sysctl_stats_tree),
821 		OID_AUTO, "pack_invalidations", CTLFLAG_RD,
822 		&softc->invalidations, 0,
823 		"Device pack invalidations.");
824 #endif
825 
826 #ifdef CAM_TEST_FAILURE
827 	SYSCTL_ADD_PROC(&softc->sysctl_ctx, SYSCTL_CHILDREN(softc->sysctl_tree),
828 		OID_AUTO, "invalidate", CTLTYPE_U64 | CTLFLAG_RW | CTLFLAG_MPSAFE,
829 		periph, 0, cam_periph_invalidate_sysctl, "I",
830 		"Write 1 to invalidate the drive immediately");
831 #endif
832 
833 	cam_iosched_sysctl_init(softc->cam_iosched, &softc->sysctl_ctx,
834 	    softc->sysctl_tree);
835 
836 	cam_periph_release(periph);
837 }
838 
839 static int
ndaflagssysctl(SYSCTL_HANDLER_ARGS)840 ndaflagssysctl(SYSCTL_HANDLER_ARGS)
841 {
842 	struct sbuf sbuf;
843 	struct nda_softc *softc = arg1;
844 	int error;
845 
846 	sbuf_new_for_sysctl(&sbuf, NULL, 0, req);
847 	if (softc->flags != 0)
848 		sbuf_printf(&sbuf, "0x%b", (unsigned)softc->flags, NDA_FLAG_STRING);
849 	else
850 		sbuf_putc(&sbuf, '0');
851 	error = sbuf_finish(&sbuf);
852 	sbuf_delete(&sbuf);
853 
854 	return (error);
855 }
856 
857 static int
ndagetattr(struct bio * bp)858 ndagetattr(struct bio *bp)
859 {
860 	int ret;
861 	struct cam_periph *periph;
862 
863 	if (g_handleattr_int(bp, "GEOM::canspeedup", nda_enable_biospeedup))
864 		return (EJUSTRETURN);
865 
866 	periph = (struct cam_periph *)bp->bio_disk->d_drv1;
867 	cam_periph_lock(periph);
868 	ret = xpt_getattr(bp->bio_data, bp->bio_length, bp->bio_attribute,
869 	    periph->path);
870 	cam_periph_unlock(periph);
871 	if (ret == 0)
872 		bp->bio_completed = bp->bio_length;
873 	return ret;
874 }
875 
876 static cam_status
ndaregister(struct cam_periph * periph,void * arg)877 ndaregister(struct cam_periph *periph, void *arg)
878 {
879 	struct nda_softc *softc;
880 	struct disk *disk;
881 	struct ccb_pathinq cpi;
882 	const struct nvme_namespace_data *nsd;
883 	const struct nvme_controller_data *cd;
884 	char   announce_buf[80];
885 	u_int maxio;
886 	int quirks;
887 
888 	nsd = nvme_get_identify_ns(periph);
889 	cd = nvme_get_identify_cntrl(periph);
890 
891 	softc = (struct nda_softc *)malloc(sizeof(*softc), M_DEVBUF,
892 	    M_NOWAIT | M_ZERO);
893 
894 	if (softc == NULL) {
895 		printf("ndaregister: Unable to probe new device. "
896 		    "Unable to allocate softc\n");
897 		return(CAM_REQ_CMP_ERR);
898 	}
899 
900 	/* ident_data parsing */
901 
902 	periph->softc = softc;
903 	softc->quirks = NDA_Q_NONE;
904 	xpt_path_inq(&cpi, periph->path);
905 	TASK_INIT(&softc->sysctl_task, 0, ndasysctlinit, periph);
906 
907 	/*
908 	 * The name space ID is the lun, save it for later I/O
909 	 */
910 	softc->nsid = (uint32_t)xpt_path_lun_id(periph->path);
911 
912 	/*
913 	 * Register this media as a disk
914 	 */
915 	(void)cam_periph_acquire(periph);
916 	cam_periph_unlock(periph);
917 	snprintf(announce_buf, sizeof(announce_buf),
918 	    "kern.cam.nda.%d.quirks", periph->unit_number);
919 	quirks = softc->quirks;
920 	TUNABLE_INT_FETCH(announce_buf, &quirks);
921 	softc->quirks = quirks;
922 	softc->disk = disk = disk_alloc();
923 	disk->d_rotation_rate = DISK_RR_NON_ROTATING;
924 	disk->d_open = ndaopen;
925 	disk->d_close = ndaclose;
926 	disk->d_strategy = ndastrategy;
927 	disk->d_ioctl = ndaioctl;
928 	disk->d_getattr = ndagetattr;
929 	if (cam_sim_pollable(periph->sim))
930 		disk->d_dump = ndadump;
931 	disk->d_gone = ndadiskgonecb;
932 	disk->d_name = "nda";
933 	disk->d_drv1 = periph;
934 	disk->d_unit = periph->unit_number;
935 	maxio = cpi.maxio;		/* Honor max I/O size of SIM */
936 	if (maxio == 0)
937 		maxio = DFLTPHYS;	/* traditional default */
938 	else if (maxio > maxphys)
939 		maxio = maxphys;	/* for safety */
940 	disk->d_maxsize = maxio;
941 	ndasetgeom(softc, periph);
942 	if ((cpi.hba_misc & PIM_UNMAPPED) != 0) {
943 		disk->d_flags |= DISKFLAG_UNMAPPED_BIO;
944 		softc->unmappedio = 1;
945 	}
946 
947 	/*
948 	 * d_ident and d_descr are both far bigger than the length of either
949 	 *  the serial or model number strings.
950 	 */
951 	cam_strvis_flag(disk->d_descr, cd->mn, NVME_MODEL_NUMBER_LENGTH,
952 	    sizeof(disk->d_descr), CAM_STRVIS_FLAG_NONASCII_SPC);
953 
954 	cam_strvis_flag(disk->d_ident, cd->sn, NVME_SERIAL_NUMBER_LENGTH,
955 	    sizeof(disk->d_ident), CAM_STRVIS_FLAG_NONASCII_SPC);
956 
957 	disk->d_hba_vendor = cpi.hba_vendor;
958 	disk->d_hba_device = cpi.hba_device;
959 	disk->d_hba_subvendor = cpi.hba_subvendor;
960 	disk->d_hba_subdevice = cpi.hba_subdevice;
961 	snprintf(disk->d_attachment, sizeof(disk->d_attachment),
962 	    "%s%d", cpi.dev_name, cpi.unit_number);
963 	if (NVMEV(NVME_NS_DATA_NSFEAT_NPVALID, nsd->nsfeat) != 0 &&
964 	    nsd->npwg != 0)
965 		disk->d_stripesize = ((nsd->npwg + 1) * disk->d_sectorsize);
966 	else
967 		disk->d_stripesize = nsd->noiob * disk->d_sectorsize;
968 	disk->d_stripeoffset = 0;
969 	disk->d_devstat = devstat_new_entry(periph->periph_name,
970 	    periph->unit_number, disk->d_sectorsize,
971 	    DEVSTAT_ALL_SUPPORTED,
972 	    DEVSTAT_TYPE_DIRECT | XPORT_DEVSTAT_TYPE(cpi.transport),
973 	    DEVSTAT_PRIORITY_DISK);
974 
975 	if (cam_iosched_init(&softc->cam_iosched, periph, disk,
976 	    ndaschedule) != 0) {
977 		printf("ndaregister: Unable to probe new device. "
978 		       "Unable to allocate iosched memory\n");
979 		free(softc, M_DEVBUF);
980 		return(CAM_REQ_CMP_ERR);
981 	}
982 	cam_iosched_set_sort_queue(softc->cam_iosched, 0);
983 
984 	/*
985 	 * Add alias for older nvd drives to ease transition.
986 	 */
987 	if (nda_nvd_compat)
988 		disk_add_alias(disk, "nvd");
989 
990 	cam_periph_lock(periph);
991 
992 	snprintf(announce_buf, sizeof(announce_buf),
993 		"%juMB (%ju %u byte sectors)",
994 	    (uintmax_t)((uintmax_t)disk->d_mediasize / (1024*1024)),
995 		(uintmax_t)disk->d_mediasize / disk->d_sectorsize,
996 		disk->d_sectorsize);
997 	xpt_announce_periph(periph, announce_buf);
998 	xpt_announce_quirks(periph, softc->quirks, NDA_Q_BIT_STRING);
999 
1000 	/*
1001 	 * Create our sysctl variables, now that we know
1002 	 * we have successfully attached.
1003 	 */
1004 	if (cam_periph_acquire(periph) == 0)
1005 		taskqueue_enqueue(taskqueue_thread, &softc->sysctl_task);
1006 
1007 	/*
1008 	 * Register for device going away and info about the drive
1009 	 * changing (though with NVMe, it can't)
1010 	 */
1011 	xpt_register_async(AC_LOST_DEVICE | AC_ADVINFO_CHANGED | AC_GETDEV_CHANGED,
1012 	    ndaasync, periph, periph->path);
1013 
1014 	softc->state = NDA_STATE_NORMAL;
1015 
1016 	/*
1017 	 * We'll release this reference once GEOM calls us back via
1018 	 * ndadiskgonecb(), telling us that our provider has been freed.
1019 	 */
1020 	if (cam_periph_acquire(periph) == 0)
1021 		disk_create(softc->disk, DISK_VERSION);
1022 
1023 	cam_periph_release_locked(periph);
1024 	return(CAM_REQ_CMP);
1025 }
1026 
1027 static void
ndastart(struct cam_periph * periph,union ccb * start_ccb)1028 ndastart(struct cam_periph *periph, union ccb *start_ccb)
1029 {
1030 	struct nda_softc *softc = (struct nda_softc *)periph->softc;
1031 	struct ccb_nvmeio *nvmeio = &start_ccb->nvmeio;
1032 
1033 	CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("ndastart\n"));
1034 
1035 	switch (softc->state) {
1036 	case NDA_STATE_NORMAL:
1037 	{
1038 		struct bio *bp;
1039 
1040 		bp = cam_iosched_next_bio(softc->cam_iosched);
1041 		CAM_DEBUG(periph->path, CAM_DEBUG_TRACE, ("ndastart: bio %p\n", bp));
1042 		if (bp == NULL) {
1043 			xpt_release_ccb(start_ccb);
1044 			break;
1045 		}
1046 
1047 		switch (bp->bio_cmd) {
1048 		case BIO_WRITE:
1049 			softc->flags |= NDA_FLAG_DIRTY;
1050 			/* FALLTHROUGH */
1051 		case BIO_READ:
1052 		{
1053 #ifdef CAM_TEST_FAILURE
1054 			int fail = 0;
1055 
1056 			/*
1057 			 * Support the failure ioctls.  If the command is a
1058 			 * read, and there are pending forced read errors, or
1059 			 * if a write and pending write errors, then fail this
1060 			 * operation with EIO.  This is useful for testing
1061 			 * purposes.  Also, support having every Nth read fail.
1062 			 *
1063 			 * This is a rather blunt tool.
1064 			 */
1065 			if (bp->bio_cmd == BIO_READ) {
1066 				if (softc->force_read_error) {
1067 					softc->force_read_error--;
1068 					fail = 1;
1069 				}
1070 				if (softc->periodic_read_error > 0) {
1071 					if (++softc->periodic_read_count >=
1072 					    softc->periodic_read_error) {
1073 						softc->periodic_read_count = 0;
1074 						fail = 1;
1075 					}
1076 				}
1077 			} else {
1078 				if (softc->force_write_error) {
1079 					softc->force_write_error--;
1080 					fail = 1;
1081 				}
1082 			}
1083 			if (fail) {
1084 				biofinish(bp, NULL, EIO);
1085 				xpt_release_ccb(start_ccb);
1086 				ndaschedule(periph);
1087 				return;
1088 			}
1089 #endif
1090 			KASSERT((bp->bio_flags & BIO_UNMAPPED) == 0 ||
1091 			    round_page(bp->bio_bcount + bp->bio_ma_offset) /
1092 			    PAGE_SIZE == bp->bio_ma_n,
1093 			    ("Short bio %p", bp));
1094 			nda_nvme_rw_bio(softc, &start_ccb->nvmeio, bp, bp->bio_cmd == BIO_READ ?
1095 			    NVME_OPC_READ : NVME_OPC_WRITE);
1096 			break;
1097 		}
1098 		case BIO_DELETE:
1099 		{
1100 			struct nvme_dsm_range *dsm_range, *dsm_end;
1101 			struct nda_trim_request *trim;
1102 			struct bio *bp1;
1103 			int ents;
1104 			uint32_t totalcount = 0, ranges = 0;
1105 
1106 			trim = malloc(sizeof(*trim), M_NVMEDA, M_ZERO | M_NOWAIT);
1107 			if (trim == NULL) {
1108 				/*
1109 				 * We have to drop the periph lock when
1110 				 * returning ENOMEM. g_io_deliver treats these
1111 				 * request differently and will recursively call
1112 				 * the start routine which causes us to get into
1113 				 * ndastrategy with the periph lock held,
1114 				 * leading to a panic when its acquired again.
1115 				 */
1116 				cam_periph_unlock(periph);
1117 				biofinish(bp, NULL, ENOMEM);
1118 				cam_periph_lock(periph);
1119 				xpt_release_ccb(start_ccb);
1120 				ndaschedule(periph);
1121 				return;
1122 			}
1123 			TAILQ_INIT(&trim->bps);
1124 			bp1 = bp;
1125 			ents = min(nitems(trim->dsm), nda_max_trim_entries);
1126 			ents = max(ents, 1);
1127 			dsm_range = trim->dsm;
1128 			dsm_end = dsm_range + ents;
1129 			do {
1130 				TAILQ_INSERT_TAIL(&trim->bps, bp1, bio_queue);
1131 				dsm_range->length =
1132 				    htole32(bp1->bio_bcount / softc->disk->d_sectorsize);
1133 				dsm_range->starting_lba =
1134 				    htole64(bp1->bio_offset / softc->disk->d_sectorsize);
1135 				ranges++;
1136 				totalcount += dsm_range->length;
1137 				dsm_range++;
1138 				if (dsm_range >= dsm_end)
1139 					break;
1140 				bp1 = cam_iosched_next_trim(softc->cam_iosched);
1141 				/* XXX -- Could collapse adjacent ranges, but we don't for now */
1142 				/* XXX -- Could limit based on total payload size */
1143 			} while (bp1 != NULL);
1144 			start_ccb->ccb_trim = trim;
1145 			nda_nvme_trim(softc, &start_ccb->nvmeio, trim->dsm,
1146 			    dsm_range - trim->dsm);
1147 			start_ccb->ccb_state = NDA_CCB_TRIM;
1148 			softc->trim_count++;
1149 			softc->trim_ranges += ranges;
1150 			softc->trim_lbas += totalcount;
1151 			/*
1152 			 * Note: We can have multiple TRIMs in flight, so we don't call
1153 			 * cam_iosched_submit_trim(softc->cam_iosched);
1154 			 * since that forces the I/O scheduler to only schedule one at a time.
1155 			 * On NVMe drives, this is a performance disaster.
1156 			 */
1157 			goto out;
1158 		}
1159 		case BIO_FLUSH:
1160 			nda_nvme_flush(softc, nvmeio);
1161 			break;
1162 		default:
1163 			biofinish(bp, NULL, EOPNOTSUPP);
1164 			xpt_release_ccb(start_ccb);
1165 			ndaschedule(periph);
1166 			return;
1167 		}
1168 		start_ccb->ccb_state = NDA_CCB_BUFFER_IO;
1169 		start_ccb->ccb_bp = bp;
1170 out:
1171 		start_ccb->ccb_h.flags |= CAM_UNLOCKED;
1172 		softc->outstanding_cmds++;
1173 		softc->refcount++;			/* For submission only */
1174 		cam_periph_unlock(periph);
1175 		xpt_action(start_ccb);
1176 		cam_periph_lock(periph);
1177 		softc->refcount--;			/* Submission done */
1178 
1179 		/* May have more work to do, so ensure we stay scheduled */
1180 		ndaschedule(periph);
1181 		break;
1182 		}
1183 	}
1184 }
1185 
1186 static void
ndadone(struct cam_periph * periph,union ccb * done_ccb)1187 ndadone(struct cam_periph *periph, union ccb *done_ccb)
1188 {
1189 	struct nda_softc *softc;
1190 	struct ccb_nvmeio *nvmeio = &done_ccb->nvmeio;
1191 	struct cam_path *path;
1192 	int state;
1193 
1194 	softc = (struct nda_softc *)periph->softc;
1195 	path = done_ccb->ccb_h.path;
1196 
1197 	CAM_DEBUG(path, CAM_DEBUG_TRACE, ("ndadone\n"));
1198 
1199 	state = nvmeio->ccb_state & NDA_CCB_TYPE_MASK;
1200 	switch (state) {
1201 	case NDA_CCB_BUFFER_IO:
1202 	case NDA_CCB_TRIM:
1203 	{
1204 		int error;
1205 
1206 		cam_periph_lock(periph);
1207 		if ((done_ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_CMP) {
1208 			error = ndaerror(done_ccb, 0, 0);
1209 			if (error == ERESTART) {
1210 				/* A retry was scheduled, so just return. */
1211 				cam_periph_unlock(periph);
1212 				return;
1213 			}
1214 			if ((done_ccb->ccb_h.status & CAM_DEV_QFRZN) != 0)
1215 				cam_release_devq(path,
1216 						 /*relsim_flags*/0,
1217 						 /*reduction*/0,
1218 						 /*timeout*/0,
1219 						 /*getcount_only*/0);
1220 		} else {
1221 			if ((done_ccb->ccb_h.status & CAM_DEV_QFRZN) != 0)
1222 				panic("REQ_CMP with QFRZN");
1223 			error = 0;
1224 		}
1225 		if (state == NDA_CCB_BUFFER_IO) {
1226 			struct bio *bp;
1227 
1228 			bp = (struct bio *)done_ccb->ccb_bp;
1229 			bp->bio_error = error;
1230 			if (error != 0) {
1231 				bp->bio_resid = bp->bio_bcount;
1232 				bp->bio_flags |= BIO_ERROR;
1233 			} else {
1234 				bp->bio_resid = 0;
1235 			}
1236 			softc->outstanding_cmds--;
1237 
1238 			/*
1239 			 * We need to call cam_iosched before we call biodone so that we
1240 			 * don't measure any activity that happens in the completion
1241 			 * routine, which in the case of sendfile can be quite
1242 			 * extensive.
1243 			 */
1244 			cam_iosched_bio_complete(softc->cam_iosched, bp, done_ccb);
1245 			xpt_release_ccb(done_ccb);
1246 			ndaschedule(periph);
1247 			cam_periph_unlock(periph);
1248 			biodone(bp);
1249 		} else { /* state == NDA_CCB_TRIM */
1250 			struct nda_trim_request *trim;
1251 			struct bio *bp1, *bp2;
1252 			TAILQ_HEAD(, bio) queue;
1253 
1254 			trim = nvmeio->ccb_trim;
1255 			TAILQ_INIT(&queue);
1256 			TAILQ_CONCAT(&queue, &trim->bps, bio_queue);
1257 			free(trim, M_NVMEDA);
1258 
1259 			/*
1260 			 * Since we can have multiple trims in flight, we don't
1261 			 * need to call this here.
1262 			 * cam_iosched_trim_done(softc->cam_iosched);
1263 			 */
1264 			/*
1265 			 * The the I/O scheduler that we're finishing the I/O
1266 			 * so we can keep book. The first one we pass in the CCB
1267 			 * which has the timing information. The rest we pass in NULL
1268 			 * so we can keep proper counts.
1269 			 */
1270 			bp1 = TAILQ_FIRST(&queue);
1271 			cam_iosched_bio_complete(softc->cam_iosched, bp1, done_ccb);
1272 			xpt_release_ccb(done_ccb);
1273 			softc->outstanding_cmds--;
1274 			ndaschedule(periph);
1275 			cam_periph_unlock(periph);
1276 			while ((bp2 = TAILQ_FIRST(&queue)) != NULL) {
1277 				TAILQ_REMOVE(&queue, bp2, bio_queue);
1278 				bp2->bio_error = error;
1279 				if (error != 0) {
1280 					bp2->bio_flags |= BIO_ERROR;
1281 					bp2->bio_resid = bp1->bio_bcount;
1282 				} else
1283 					bp2->bio_resid = 0;
1284 				if (bp1 != bp2)
1285 					cam_iosched_bio_complete(softc->cam_iosched, bp2, NULL);
1286 				biodone(bp2);
1287 			}
1288 		}
1289 		return;
1290 	}
1291 	case NDA_CCB_DUMP:
1292 		/* No-op.  We're polling */
1293 		return;
1294 	case NDA_CCB_PASS:
1295 		/* NVME_PASSTHROUGH_CMD runs this CCB and releases it */
1296 		return;
1297 	default:
1298 		break;
1299 	}
1300 	xpt_release_ccb(done_ccb);
1301 }
1302 
1303 static int
ndaerror(union ccb * ccb,uint32_t cam_flags,uint32_t sense_flags)1304 ndaerror(union ccb *ccb, uint32_t cam_flags, uint32_t sense_flags)
1305 {
1306 #ifdef CAM_IO_STATS
1307 	struct nda_softc *softc;
1308 	struct cam_periph *periph;
1309 
1310 	periph = xpt_path_periph(ccb->ccb_h.path);
1311 	softc = (struct nda_softc *)periph->softc;
1312 #endif
1313 
1314 	switch (ccb->ccb_h.status & CAM_STATUS_MASK) {
1315 	case CAM_CMD_TIMEOUT:
1316 #ifdef CAM_IO_STATS
1317 		softc->timeouts++;
1318 #endif
1319 		break;
1320 	case CAM_REQ_CMP_ERR:
1321 	case CAM_NVME_STATUS_ERROR:
1322 #ifdef CAM_IO_STATS
1323 		softc->errors++;
1324 #endif
1325 		break;
1326 	default:
1327 		break;
1328 	}
1329 
1330 	return(cam_periph_error(ccb, cam_flags, sense_flags));
1331 }
1332 
1333 /*
1334  * Step through all NDA peripheral drivers, and if the device is still open,
1335  * sync the disk cache to physical media.
1336  */
1337 static void
ndaflush(void)1338 ndaflush(void)
1339 {
1340 	struct cam_periph *periph;
1341 	struct nda_softc *softc;
1342 	union ccb *ccb;
1343 	int error;
1344 
1345 	CAM_PERIPH_FOREACH(periph, &ndadriver) {
1346 		softc = (struct nda_softc *)periph->softc;
1347 
1348 		if (SCHEDULER_STOPPED()) {
1349 			/*
1350 			 * If we panicked with the lock held or the periph is not
1351 			 * open, do not recurse.  Otherwise, call ndadump since
1352 			 * that avoids the sleeping cam_periph_getccb does if no
1353 			 * CCBs are available.
1354 			 */
1355 			if (!cam_periph_owned(periph) &&
1356 			    (softc->flags & NDA_FLAG_OPEN)) {
1357 				ndadump(softc->disk, NULL, 0, 0);
1358 			}
1359 			continue;
1360 		}
1361 
1362 		/*
1363 		 * We only sync the cache if the drive is still open
1364 		 */
1365 		cam_periph_lock(periph);
1366 		if ((softc->flags & NDA_FLAG_OPEN) == 0) {
1367 			cam_periph_unlock(periph);
1368 			continue;
1369 		}
1370 
1371 		ccb = cam_periph_getccb(periph, CAM_PRIORITY_NORMAL);
1372 		nda_nvme_flush(softc, &ccb->nvmeio);
1373 		error = cam_periph_runccb(ccb, ndaerror, /*cam_flags*/0,
1374 		    /*sense_flags*/ SF_NO_RECOVERY | SF_NO_RETRY,
1375 		    softc->disk->d_devstat);
1376 		if (error != 0)
1377 			xpt_print(periph->path, "Synchronize cache failed\n");
1378 		xpt_release_ccb(ccb);
1379 		cam_periph_unlock(periph);
1380 	}
1381 }
1382 
1383 static void
ndashutdown(void * arg,int howto)1384 ndashutdown(void *arg, int howto)
1385 {
1386 
1387 	if ((howto & RB_NOSYNC) != 0)
1388 		return;
1389 
1390 	ndaflush();
1391 }
1392 
1393 static void
ndasuspend(void * arg,enum power_stype stype)1394 ndasuspend(void *arg, enum power_stype stype)
1395 {
1396 
1397 	ndaflush();
1398 }
1399