xref: /freebsd/sys/dev/nvme/nvme_ctrlr.c (revision dd41de95a84d979615a2ef11df6850622bf6184e)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (C) 2012-2016 Intel Corporation
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
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 AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  */
28 
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
31 
32 #include "opt_cam.h"
33 
34 #include <sys/param.h>
35 #include <sys/systm.h>
36 #include <sys/buf.h>
37 #include <sys/bus.h>
38 #include <sys/conf.h>
39 #include <sys/ioccom.h>
40 #include <sys/proc.h>
41 #include <sys/smp.h>
42 #include <sys/uio.h>
43 #include <sys/sbuf.h>
44 #include <sys/endian.h>
45 #include <machine/stdarg.h>
46 #include <vm/vm.h>
47 
48 #include "nvme_private.h"
49 
50 #define B4_CHK_RDY_DELAY_MS	2300		/* work around controller bug */
51 
52 static void nvme_ctrlr_construct_and_submit_aer(struct nvme_controller *ctrlr,
53 						struct nvme_async_event_request *aer);
54 
55 static void
56 nvme_ctrlr_devctl_log(struct nvme_controller *ctrlr, const char *type, const char *msg, ...)
57 {
58 	struct sbuf sb;
59 	va_list ap;
60 	int error;
61 
62 	if (sbuf_new(&sb, NULL, 0, SBUF_AUTOEXTEND | SBUF_NOWAIT) == NULL)
63 		return;
64 	sbuf_printf(&sb, "%s: ", device_get_nameunit(ctrlr->dev));
65 	va_start(ap, msg);
66 	sbuf_vprintf(&sb, msg, ap);
67 	va_end(ap);
68 	error = sbuf_finish(&sb);
69 	if (error == 0)
70 		printf("%s\n", sbuf_data(&sb));
71 
72 	sbuf_clear(&sb);
73 	sbuf_printf(&sb, "name=\"%s\" reason=\"", device_get_nameunit(ctrlr->dev));
74 	va_start(ap, msg);
75 	sbuf_vprintf(&sb, msg, ap);
76 	va_end(ap);
77 	sbuf_printf(&sb, "\"");
78 	error = sbuf_finish(&sb);
79 	if (error == 0)
80 		devctl_notify("nvme", "controller", type, sbuf_data(&sb));
81 	sbuf_delete(&sb);
82 }
83 
84 static int
85 nvme_ctrlr_construct_admin_qpair(struct nvme_controller *ctrlr)
86 {
87 	struct nvme_qpair	*qpair;
88 	uint32_t		num_entries;
89 	int			error;
90 
91 	qpair = &ctrlr->adminq;
92 	qpair->id = 0;
93 	qpair->cpu = CPU_FFS(&cpuset_domain[ctrlr->domain]) - 1;
94 	qpair->domain = ctrlr->domain;
95 
96 	num_entries = NVME_ADMIN_ENTRIES;
97 	TUNABLE_INT_FETCH("hw.nvme.admin_entries", &num_entries);
98 	/*
99 	 * If admin_entries was overridden to an invalid value, revert it
100 	 *  back to our default value.
101 	 */
102 	if (num_entries < NVME_MIN_ADMIN_ENTRIES ||
103 	    num_entries > NVME_MAX_ADMIN_ENTRIES) {
104 		nvme_printf(ctrlr, "invalid hw.nvme.admin_entries=%d "
105 		    "specified\n", num_entries);
106 		num_entries = NVME_ADMIN_ENTRIES;
107 	}
108 
109 	/*
110 	 * The admin queue's max xfer size is treated differently than the
111 	 *  max I/O xfer size.  16KB is sufficient here - maybe even less?
112 	 */
113 	error = nvme_qpair_construct(qpair, num_entries, NVME_ADMIN_TRACKERS,
114 	     ctrlr);
115 	return (error);
116 }
117 
118 #define QP(ctrlr, c)	((c) * (ctrlr)->num_io_queues / mp_ncpus)
119 
120 static int
121 nvme_ctrlr_construct_io_qpairs(struct nvme_controller *ctrlr)
122 {
123 	struct nvme_qpair	*qpair;
124 	uint32_t		cap_lo;
125 	uint16_t		mqes;
126 	int			c, error, i, n;
127 	int			num_entries, num_trackers, max_entries;
128 
129 	/*
130 	 * NVMe spec sets a hard limit of 64K max entries, but devices may
131 	 * specify a smaller limit, so we need to check the MQES field in the
132 	 * capabilities register. We have to cap the number of entries to the
133 	 * current stride allows for in BAR 0/1, otherwise the remainder entries
134 	 * are inaccessable. MQES should reflect this, and this is just a
135 	 * fail-safe.
136 	 */
137 	max_entries =
138 	    (rman_get_size(ctrlr->resource) - nvme_mmio_offsetof(doorbell[0])) /
139 	    (1 << (ctrlr->dstrd + 1));
140 	num_entries = NVME_IO_ENTRIES;
141 	TUNABLE_INT_FETCH("hw.nvme.io_entries", &num_entries);
142 	cap_lo = nvme_mmio_read_4(ctrlr, cap_lo);
143 	mqes = NVME_CAP_LO_MQES(cap_lo);
144 	num_entries = min(num_entries, mqes + 1);
145 	num_entries = min(num_entries, max_entries);
146 
147 	num_trackers = NVME_IO_TRACKERS;
148 	TUNABLE_INT_FETCH("hw.nvme.io_trackers", &num_trackers);
149 
150 	num_trackers = max(num_trackers, NVME_MIN_IO_TRACKERS);
151 	num_trackers = min(num_trackers, NVME_MAX_IO_TRACKERS);
152 	/*
153 	 * No need to have more trackers than entries in the submit queue.  Note
154 	 * also that for a queue size of N, we can only have (N-1) commands
155 	 * outstanding, hence the "-1" here.
156 	 */
157 	num_trackers = min(num_trackers, (num_entries-1));
158 
159 	/*
160 	 * Our best estimate for the maximum number of I/Os that we should
161 	 * normally have in flight at one time. This should be viewed as a hint,
162 	 * not a hard limit and will need to be revisited when the upper layers
163 	 * of the storage system grows multi-queue support.
164 	 */
165 	ctrlr->max_hw_pend_io = num_trackers * ctrlr->num_io_queues * 3 / 4;
166 
167 	ctrlr->ioq = malloc(ctrlr->num_io_queues * sizeof(struct nvme_qpair),
168 	    M_NVME, M_ZERO | M_WAITOK);
169 
170 	for (i = c = n = 0; i < ctrlr->num_io_queues; i++, c += n) {
171 		qpair = &ctrlr->ioq[i];
172 
173 		/*
174 		 * Admin queue has ID=0. IO queues start at ID=1 -
175 		 *  hence the 'i+1' here.
176 		 */
177 		qpair->id = i + 1;
178 		if (ctrlr->num_io_queues > 1) {
179 			/* Find number of CPUs served by this queue. */
180 			for (n = 1; QP(ctrlr, c + n) == i; n++)
181 				;
182 			/* Shuffle multiple NVMe devices between CPUs. */
183 			qpair->cpu = c + (device_get_unit(ctrlr->dev)+n/2) % n;
184 			qpair->domain = pcpu_find(qpair->cpu)->pc_domain;
185 		} else {
186 			qpair->cpu = CPU_FFS(&cpuset_domain[ctrlr->domain]) - 1;
187 			qpair->domain = ctrlr->domain;
188 		}
189 
190 		/*
191 		 * For I/O queues, use the controller-wide max_xfer_size
192 		 *  calculated in nvme_attach().
193 		 */
194 		error = nvme_qpair_construct(qpair, num_entries, num_trackers,
195 		    ctrlr);
196 		if (error)
197 			return (error);
198 
199 		/*
200 		 * Do not bother binding interrupts if we only have one I/O
201 		 *  interrupt thread for this controller.
202 		 */
203 		if (ctrlr->num_io_queues > 1)
204 			bus_bind_intr(ctrlr->dev, qpair->res, qpair->cpu);
205 	}
206 
207 	return (0);
208 }
209 
210 static void
211 nvme_ctrlr_fail(struct nvme_controller *ctrlr)
212 {
213 	int i;
214 
215 	ctrlr->is_failed = true;
216 	nvme_admin_qpair_disable(&ctrlr->adminq);
217 	nvme_qpair_fail(&ctrlr->adminq);
218 	if (ctrlr->ioq != NULL) {
219 		for (i = 0; i < ctrlr->num_io_queues; i++) {
220 			nvme_io_qpair_disable(&ctrlr->ioq[i]);
221 			nvme_qpair_fail(&ctrlr->ioq[i]);
222 		}
223 	}
224 	nvme_notify_fail_consumers(ctrlr);
225 }
226 
227 void
228 nvme_ctrlr_post_failed_request(struct nvme_controller *ctrlr,
229     struct nvme_request *req)
230 {
231 
232 	mtx_lock(&ctrlr->lock);
233 	STAILQ_INSERT_TAIL(&ctrlr->fail_req, req, stailq);
234 	mtx_unlock(&ctrlr->lock);
235 	taskqueue_enqueue(ctrlr->taskqueue, &ctrlr->fail_req_task);
236 }
237 
238 static void
239 nvme_ctrlr_fail_req_task(void *arg, int pending)
240 {
241 	struct nvme_controller	*ctrlr = arg;
242 	struct nvme_request	*req;
243 
244 	mtx_lock(&ctrlr->lock);
245 	while ((req = STAILQ_FIRST(&ctrlr->fail_req)) != NULL) {
246 		STAILQ_REMOVE_HEAD(&ctrlr->fail_req, stailq);
247 		mtx_unlock(&ctrlr->lock);
248 		nvme_qpair_manual_complete_request(req->qpair, req,
249 		    NVME_SCT_GENERIC, NVME_SC_ABORTED_BY_REQUEST);
250 		mtx_lock(&ctrlr->lock);
251 	}
252 	mtx_unlock(&ctrlr->lock);
253 }
254 
255 static int
256 nvme_ctrlr_wait_for_ready(struct nvme_controller *ctrlr, int desired_val)
257 {
258 	int timeout = ticks + (uint64_t)ctrlr->ready_timeout_in_ms * hz / 1000;
259 	uint32_t csts;
260 
261 	while (1) {
262 		csts = nvme_mmio_read_4(ctrlr, csts);
263 		if (csts == NVME_GONE)		/* Hot unplug. */
264 			return (ENXIO);
265 		if (((csts >> NVME_CSTS_REG_RDY_SHIFT) & NVME_CSTS_REG_RDY_MASK)
266 		    == desired_val)
267 			break;
268 		if (timeout - ticks < 0) {
269 			nvme_printf(ctrlr, "controller ready did not become %d "
270 			    "within %d ms\n", desired_val, ctrlr->ready_timeout_in_ms);
271 			return (ENXIO);
272 		}
273 		pause("nvmerdy", 1);
274 	}
275 
276 	return (0);
277 }
278 
279 static int
280 nvme_ctrlr_disable(struct nvme_controller *ctrlr)
281 {
282 	uint32_t cc;
283 	uint32_t csts;
284 	uint8_t  en, rdy;
285 	int err;
286 
287 	cc = nvme_mmio_read_4(ctrlr, cc);
288 	csts = nvme_mmio_read_4(ctrlr, csts);
289 
290 	en = (cc >> NVME_CC_REG_EN_SHIFT) & NVME_CC_REG_EN_MASK;
291 	rdy = (csts >> NVME_CSTS_REG_RDY_SHIFT) & NVME_CSTS_REG_RDY_MASK;
292 
293 	/*
294 	 * Per 3.1.5 in NVME 1.3 spec, transitioning CC.EN from 0 to 1
295 	 * when CSTS.RDY is 1 or transitioning CC.EN from 1 to 0 when
296 	 * CSTS.RDY is 0 "has undefined results" So make sure that CSTS.RDY
297 	 * isn't the desired value. Short circuit if we're already disabled.
298 	 */
299 	if (en == 1) {
300 		if (rdy == 0) {
301 			/* EN == 1, wait for  RDY == 1 or fail */
302 			err = nvme_ctrlr_wait_for_ready(ctrlr, 1);
303 			if (err != 0)
304 				return (err);
305 		}
306 	} else {
307 		/* EN == 0 already wait for RDY == 0 */
308 		if (rdy == 0)
309 			return (0);
310 		else
311 			return (nvme_ctrlr_wait_for_ready(ctrlr, 0));
312 	}
313 
314 	cc &= ~NVME_CC_REG_EN_MASK;
315 	nvme_mmio_write_4(ctrlr, cc, cc);
316 	/*
317 	 * Some drives have issues with accessing the mmio after we
318 	 * disable, so delay for a bit after we write the bit to
319 	 * cope with these issues.
320 	 */
321 	if (ctrlr->quirks & QUIRK_DELAY_B4_CHK_RDY)
322 		pause("nvmeR", B4_CHK_RDY_DELAY_MS * hz / 1000);
323 	return (nvme_ctrlr_wait_for_ready(ctrlr, 0));
324 }
325 
326 static int
327 nvme_ctrlr_enable(struct nvme_controller *ctrlr)
328 {
329 	uint32_t	cc;
330 	uint32_t	csts;
331 	uint32_t	aqa;
332 	uint32_t	qsize;
333 	uint8_t		en, rdy;
334 	int		err;
335 
336 	cc = nvme_mmio_read_4(ctrlr, cc);
337 	csts = nvme_mmio_read_4(ctrlr, csts);
338 
339 	en = (cc >> NVME_CC_REG_EN_SHIFT) & NVME_CC_REG_EN_MASK;
340 	rdy = (csts >> NVME_CSTS_REG_RDY_SHIFT) & NVME_CSTS_REG_RDY_MASK;
341 
342 	/*
343 	 * See note in nvme_ctrlr_disable. Short circuit if we're already enabled.
344 	 */
345 	if (en == 1) {
346 		if (rdy == 1)
347 			return (0);
348 		else
349 			return (nvme_ctrlr_wait_for_ready(ctrlr, 1));
350 	} else {
351 		/* EN == 0 already wait for RDY == 0 or fail */
352 		err = nvme_ctrlr_wait_for_ready(ctrlr, 0);
353 		if (err != 0)
354 			return (err);
355 	}
356 
357 	nvme_mmio_write_8(ctrlr, asq, ctrlr->adminq.cmd_bus_addr);
358 	DELAY(5000);
359 	nvme_mmio_write_8(ctrlr, acq, ctrlr->adminq.cpl_bus_addr);
360 	DELAY(5000);
361 
362 	/* acqs and asqs are 0-based. */
363 	qsize = ctrlr->adminq.num_entries - 1;
364 
365 	aqa = 0;
366 	aqa = (qsize & NVME_AQA_REG_ACQS_MASK) << NVME_AQA_REG_ACQS_SHIFT;
367 	aqa |= (qsize & NVME_AQA_REG_ASQS_MASK) << NVME_AQA_REG_ASQS_SHIFT;
368 	nvme_mmio_write_4(ctrlr, aqa, aqa);
369 	DELAY(5000);
370 
371 	/* Initialization values for CC */
372 	cc = 0;
373 	cc |= 1 << NVME_CC_REG_EN_SHIFT;
374 	cc |= 0 << NVME_CC_REG_CSS_SHIFT;
375 	cc |= 0 << NVME_CC_REG_AMS_SHIFT;
376 	cc |= 0 << NVME_CC_REG_SHN_SHIFT;
377 	cc |= 6 << NVME_CC_REG_IOSQES_SHIFT; /* SQ entry size == 64 == 2^6 */
378 	cc |= 4 << NVME_CC_REG_IOCQES_SHIFT; /* CQ entry size == 16 == 2^4 */
379 
380 	/* This evaluates to 0, which is according to spec. */
381 	cc |= (PAGE_SIZE >> 13) << NVME_CC_REG_MPS_SHIFT;
382 
383 	nvme_mmio_write_4(ctrlr, cc, cc);
384 
385 	return (nvme_ctrlr_wait_for_ready(ctrlr, 1));
386 }
387 
388 static void
389 nvme_ctrlr_disable_qpairs(struct nvme_controller *ctrlr)
390 {
391 	int i;
392 
393 	nvme_admin_qpair_disable(&ctrlr->adminq);
394 	/*
395 	 * I/O queues are not allocated before the initial HW
396 	 *  reset, so do not try to disable them.  Use is_initialized
397 	 *  to determine if this is the initial HW reset.
398 	 */
399 	if (ctrlr->is_initialized) {
400 		for (i = 0; i < ctrlr->num_io_queues; i++)
401 			nvme_io_qpair_disable(&ctrlr->ioq[i]);
402 	}
403 }
404 
405 static int
406 nvme_ctrlr_hw_reset(struct nvme_controller *ctrlr)
407 {
408 	int err;
409 
410 	nvme_ctrlr_disable_qpairs(ctrlr);
411 
412 	pause("nvmehwreset", hz / 10);
413 
414 	err = nvme_ctrlr_disable(ctrlr);
415 	if (err != 0)
416 		return err;
417 	return (nvme_ctrlr_enable(ctrlr));
418 }
419 
420 void
421 nvme_ctrlr_reset(struct nvme_controller *ctrlr)
422 {
423 	int cmpset;
424 
425 	cmpset = atomic_cmpset_32(&ctrlr->is_resetting, 0, 1);
426 
427 	if (cmpset == 0 || ctrlr->is_failed)
428 		/*
429 		 * Controller is already resetting or has failed.  Return
430 		 *  immediately since there is no need to kick off another
431 		 *  reset in these cases.
432 		 */
433 		return;
434 
435 	taskqueue_enqueue(ctrlr->taskqueue, &ctrlr->reset_task);
436 }
437 
438 static int
439 nvme_ctrlr_identify(struct nvme_controller *ctrlr)
440 {
441 	struct nvme_completion_poll_status	status;
442 
443 	status.done = 0;
444 	nvme_ctrlr_cmd_identify_controller(ctrlr, &ctrlr->cdata,
445 	    nvme_completion_poll_cb, &status);
446 	nvme_completion_poll(&status);
447 	if (nvme_completion_is_error(&status.cpl)) {
448 		nvme_printf(ctrlr, "nvme_identify_controller failed!\n");
449 		return (ENXIO);
450 	}
451 
452 	/* Convert data to host endian */
453 	nvme_controller_data_swapbytes(&ctrlr->cdata);
454 
455 	/*
456 	 * Use MDTS to ensure our default max_xfer_size doesn't exceed what the
457 	 *  controller supports.
458 	 */
459 	if (ctrlr->cdata.mdts > 0)
460 		ctrlr->max_xfer_size = min(ctrlr->max_xfer_size,
461 		    ctrlr->min_page_size * (1 << (ctrlr->cdata.mdts)));
462 
463 	return (0);
464 }
465 
466 static int
467 nvme_ctrlr_set_num_qpairs(struct nvme_controller *ctrlr)
468 {
469 	struct nvme_completion_poll_status	status;
470 	int					cq_allocated, sq_allocated;
471 
472 	status.done = 0;
473 	nvme_ctrlr_cmd_set_num_queues(ctrlr, ctrlr->num_io_queues,
474 	    nvme_completion_poll_cb, &status);
475 	nvme_completion_poll(&status);
476 	if (nvme_completion_is_error(&status.cpl)) {
477 		nvme_printf(ctrlr, "nvme_ctrlr_set_num_qpairs failed!\n");
478 		return (ENXIO);
479 	}
480 
481 	/*
482 	 * Data in cdw0 is 0-based.
483 	 * Lower 16-bits indicate number of submission queues allocated.
484 	 * Upper 16-bits indicate number of completion queues allocated.
485 	 */
486 	sq_allocated = (status.cpl.cdw0 & 0xFFFF) + 1;
487 	cq_allocated = (status.cpl.cdw0 >> 16) + 1;
488 
489 	/*
490 	 * Controller may allocate more queues than we requested,
491 	 *  so use the minimum of the number requested and what was
492 	 *  actually allocated.
493 	 */
494 	ctrlr->num_io_queues = min(ctrlr->num_io_queues, sq_allocated);
495 	ctrlr->num_io_queues = min(ctrlr->num_io_queues, cq_allocated);
496 	if (ctrlr->num_io_queues > vm_ndomains)
497 		ctrlr->num_io_queues -= ctrlr->num_io_queues % vm_ndomains;
498 
499 	return (0);
500 }
501 
502 static int
503 nvme_ctrlr_create_qpairs(struct nvme_controller *ctrlr)
504 {
505 	struct nvme_completion_poll_status	status;
506 	struct nvme_qpair			*qpair;
507 	int					i;
508 
509 	for (i = 0; i < ctrlr->num_io_queues; i++) {
510 		qpair = &ctrlr->ioq[i];
511 
512 		status.done = 0;
513 		nvme_ctrlr_cmd_create_io_cq(ctrlr, qpair,
514 		    nvme_completion_poll_cb, &status);
515 		nvme_completion_poll(&status);
516 		if (nvme_completion_is_error(&status.cpl)) {
517 			nvme_printf(ctrlr, "nvme_create_io_cq failed!\n");
518 			return (ENXIO);
519 		}
520 
521 		status.done = 0;
522 		nvme_ctrlr_cmd_create_io_sq(ctrlr, qpair,
523 		    nvme_completion_poll_cb, &status);
524 		nvme_completion_poll(&status);
525 		if (nvme_completion_is_error(&status.cpl)) {
526 			nvme_printf(ctrlr, "nvme_create_io_sq failed!\n");
527 			return (ENXIO);
528 		}
529 	}
530 
531 	return (0);
532 }
533 
534 static int
535 nvme_ctrlr_delete_qpairs(struct nvme_controller *ctrlr)
536 {
537 	struct nvme_completion_poll_status	status;
538 	struct nvme_qpair			*qpair;
539 
540 	for (int i = 0; i < ctrlr->num_io_queues; i++) {
541 		qpair = &ctrlr->ioq[i];
542 
543 		status.done = 0;
544 		nvme_ctrlr_cmd_delete_io_sq(ctrlr, qpair,
545 		    nvme_completion_poll_cb, &status);
546 		nvme_completion_poll(&status);
547 		if (nvme_completion_is_error(&status.cpl)) {
548 			nvme_printf(ctrlr, "nvme_destroy_io_sq failed!\n");
549 			return (ENXIO);
550 		}
551 
552 		status.done = 0;
553 		nvme_ctrlr_cmd_delete_io_cq(ctrlr, qpair,
554 		    nvme_completion_poll_cb, &status);
555 		nvme_completion_poll(&status);
556 		if (nvme_completion_is_error(&status.cpl)) {
557 			nvme_printf(ctrlr, "nvme_destroy_io_cq failed!\n");
558 			return (ENXIO);
559 		}
560 	}
561 
562 	return (0);
563 }
564 
565 static int
566 nvme_ctrlr_construct_namespaces(struct nvme_controller *ctrlr)
567 {
568 	struct nvme_namespace	*ns;
569 	uint32_t 		i;
570 
571 	for (i = 0; i < min(ctrlr->cdata.nn, NVME_MAX_NAMESPACES); i++) {
572 		ns = &ctrlr->ns[i];
573 		nvme_ns_construct(ns, i+1, ctrlr);
574 	}
575 
576 	return (0);
577 }
578 
579 static bool
580 is_log_page_id_valid(uint8_t page_id)
581 {
582 
583 	switch (page_id) {
584 	case NVME_LOG_ERROR:
585 	case NVME_LOG_HEALTH_INFORMATION:
586 	case NVME_LOG_FIRMWARE_SLOT:
587 	case NVME_LOG_CHANGED_NAMESPACE:
588 	case NVME_LOG_COMMAND_EFFECT:
589 	case NVME_LOG_RES_NOTIFICATION:
590 	case NVME_LOG_SANITIZE_STATUS:
591 		return (true);
592 	}
593 
594 	return (false);
595 }
596 
597 static uint32_t
598 nvme_ctrlr_get_log_page_size(struct nvme_controller *ctrlr, uint8_t page_id)
599 {
600 	uint32_t	log_page_size;
601 
602 	switch (page_id) {
603 	case NVME_LOG_ERROR:
604 		log_page_size = min(
605 		    sizeof(struct nvme_error_information_entry) *
606 		    (ctrlr->cdata.elpe + 1), NVME_MAX_AER_LOG_SIZE);
607 		break;
608 	case NVME_LOG_HEALTH_INFORMATION:
609 		log_page_size = sizeof(struct nvme_health_information_page);
610 		break;
611 	case NVME_LOG_FIRMWARE_SLOT:
612 		log_page_size = sizeof(struct nvme_firmware_page);
613 		break;
614 	case NVME_LOG_CHANGED_NAMESPACE:
615 		log_page_size = sizeof(struct nvme_ns_list);
616 		break;
617 	case NVME_LOG_COMMAND_EFFECT:
618 		log_page_size = sizeof(struct nvme_command_effects_page);
619 		break;
620 	case NVME_LOG_RES_NOTIFICATION:
621 		log_page_size = sizeof(struct nvme_res_notification_page);
622 		break;
623 	case NVME_LOG_SANITIZE_STATUS:
624 		log_page_size = sizeof(struct nvme_sanitize_status_page);
625 		break;
626 	default:
627 		log_page_size = 0;
628 		break;
629 	}
630 
631 	return (log_page_size);
632 }
633 
634 static void
635 nvme_ctrlr_log_critical_warnings(struct nvme_controller *ctrlr,
636     uint8_t state)
637 {
638 
639 	if (state & NVME_CRIT_WARN_ST_AVAILABLE_SPARE)
640 		nvme_ctrlr_devctl_log(ctrlr, "critical",
641 		    "available spare space below threshold");
642 
643 	if (state & NVME_CRIT_WARN_ST_TEMPERATURE)
644 		nvme_ctrlr_devctl_log(ctrlr, "critical",
645 		    "temperature above threshold");
646 
647 	if (state & NVME_CRIT_WARN_ST_DEVICE_RELIABILITY)
648 		nvme_ctrlr_devctl_log(ctrlr, "critical",
649 		    "device reliability degraded");
650 
651 	if (state & NVME_CRIT_WARN_ST_READ_ONLY)
652 		nvme_ctrlr_devctl_log(ctrlr, "critical",
653 		    "media placed in read only mode");
654 
655 	if (state & NVME_CRIT_WARN_ST_VOLATILE_MEMORY_BACKUP)
656 		nvme_ctrlr_devctl_log(ctrlr, "critical",
657 		    "volatile memory backup device failed");
658 
659 	if (state & NVME_CRIT_WARN_ST_RESERVED_MASK)
660 		nvme_ctrlr_devctl_log(ctrlr, "critical",
661 		    "unknown critical warning(s): state = 0x%02x", state);
662 }
663 
664 static void
665 nvme_ctrlr_async_event_log_page_cb(void *arg, const struct nvme_completion *cpl)
666 {
667 	struct nvme_async_event_request		*aer = arg;
668 	struct nvme_health_information_page	*health_info;
669 	struct nvme_ns_list			*nsl;
670 	struct nvme_error_information_entry	*err;
671 	int i;
672 
673 	/*
674 	 * If the log page fetch for some reason completed with an error,
675 	 *  don't pass log page data to the consumers.  In practice, this case
676 	 *  should never happen.
677 	 */
678 	if (nvme_completion_is_error(cpl))
679 		nvme_notify_async_consumers(aer->ctrlr, &aer->cpl,
680 		    aer->log_page_id, NULL, 0);
681 	else {
682 		/* Convert data to host endian */
683 		switch (aer->log_page_id) {
684 		case NVME_LOG_ERROR:
685 			err = (struct nvme_error_information_entry *)aer->log_page_buffer;
686 			for (i = 0; i < (aer->ctrlr->cdata.elpe + 1); i++)
687 				nvme_error_information_entry_swapbytes(err++);
688 			break;
689 		case NVME_LOG_HEALTH_INFORMATION:
690 			nvme_health_information_page_swapbytes(
691 			    (struct nvme_health_information_page *)aer->log_page_buffer);
692 			break;
693 		case NVME_LOG_FIRMWARE_SLOT:
694 			nvme_firmware_page_swapbytes(
695 			    (struct nvme_firmware_page *)aer->log_page_buffer);
696 			break;
697 		case NVME_LOG_CHANGED_NAMESPACE:
698 			nvme_ns_list_swapbytes(
699 			    (struct nvme_ns_list *)aer->log_page_buffer);
700 			break;
701 		case NVME_LOG_COMMAND_EFFECT:
702 			nvme_command_effects_page_swapbytes(
703 			    (struct nvme_command_effects_page *)aer->log_page_buffer);
704 			break;
705 		case NVME_LOG_RES_NOTIFICATION:
706 			nvme_res_notification_page_swapbytes(
707 			    (struct nvme_res_notification_page *)aer->log_page_buffer);
708 			break;
709 		case NVME_LOG_SANITIZE_STATUS:
710 			nvme_sanitize_status_page_swapbytes(
711 			    (struct nvme_sanitize_status_page *)aer->log_page_buffer);
712 			break;
713 		case INTEL_LOG_TEMP_STATS:
714 			intel_log_temp_stats_swapbytes(
715 			    (struct intel_log_temp_stats *)aer->log_page_buffer);
716 			break;
717 		default:
718 			break;
719 		}
720 
721 		if (aer->log_page_id == NVME_LOG_HEALTH_INFORMATION) {
722 			health_info = (struct nvme_health_information_page *)
723 			    aer->log_page_buffer;
724 			nvme_ctrlr_log_critical_warnings(aer->ctrlr,
725 			    health_info->critical_warning);
726 			/*
727 			 * Critical warnings reported through the
728 			 *  SMART/health log page are persistent, so
729 			 *  clear the associated bits in the async event
730 			 *  config so that we do not receive repeated
731 			 *  notifications for the same event.
732 			 */
733 			aer->ctrlr->async_event_config &=
734 			    ~health_info->critical_warning;
735 			nvme_ctrlr_cmd_set_async_event_config(aer->ctrlr,
736 			    aer->ctrlr->async_event_config, NULL, NULL);
737 		} else if (aer->log_page_id == NVME_LOG_CHANGED_NAMESPACE &&
738 		    !nvme_use_nvd) {
739 			nsl = (struct nvme_ns_list *)aer->log_page_buffer;
740 			for (i = 0; i < nitems(nsl->ns) && nsl->ns[i] != 0; i++) {
741 				if (nsl->ns[i] > NVME_MAX_NAMESPACES)
742 					break;
743 				nvme_notify_ns(aer->ctrlr, nsl->ns[i]);
744 			}
745 		}
746 
747 		/*
748 		 * Pass the cpl data from the original async event completion,
749 		 *  not the log page fetch.
750 		 */
751 		nvme_notify_async_consumers(aer->ctrlr, &aer->cpl,
752 		    aer->log_page_id, aer->log_page_buffer, aer->log_page_size);
753 	}
754 
755 	/*
756 	 * Repost another asynchronous event request to replace the one
757 	 *  that just completed.
758 	 */
759 	nvme_ctrlr_construct_and_submit_aer(aer->ctrlr, aer);
760 }
761 
762 static void
763 nvme_ctrlr_async_event_cb(void *arg, const struct nvme_completion *cpl)
764 {
765 	struct nvme_async_event_request	*aer = arg;
766 
767 	if (nvme_completion_is_error(cpl)) {
768 		/*
769 		 *  Do not retry failed async event requests.  This avoids
770 		 *  infinite loops where a new async event request is submitted
771 		 *  to replace the one just failed, only to fail again and
772 		 *  perpetuate the loop.
773 		 */
774 		return;
775 	}
776 
777 	/* Associated log page is in bits 23:16 of completion entry dw0. */
778 	aer->log_page_id = (cpl->cdw0 & 0xFF0000) >> 16;
779 
780 	nvme_printf(aer->ctrlr, "async event occurred (type 0x%x, info 0x%02x,"
781 	    " page 0x%02x)\n", (cpl->cdw0 & 0x07), (cpl->cdw0 & 0xFF00) >> 8,
782 	    aer->log_page_id);
783 
784 	if (is_log_page_id_valid(aer->log_page_id)) {
785 		aer->log_page_size = nvme_ctrlr_get_log_page_size(aer->ctrlr,
786 		    aer->log_page_id);
787 		memcpy(&aer->cpl, cpl, sizeof(*cpl));
788 		nvme_ctrlr_cmd_get_log_page(aer->ctrlr, aer->log_page_id,
789 		    NVME_GLOBAL_NAMESPACE_TAG, aer->log_page_buffer,
790 		    aer->log_page_size, nvme_ctrlr_async_event_log_page_cb,
791 		    aer);
792 		/* Wait to notify consumers until after log page is fetched. */
793 	} else {
794 		nvme_notify_async_consumers(aer->ctrlr, cpl, aer->log_page_id,
795 		    NULL, 0);
796 
797 		/*
798 		 * Repost another asynchronous event request to replace the one
799 		 *  that just completed.
800 		 */
801 		nvme_ctrlr_construct_and_submit_aer(aer->ctrlr, aer);
802 	}
803 }
804 
805 static void
806 nvme_ctrlr_construct_and_submit_aer(struct nvme_controller *ctrlr,
807     struct nvme_async_event_request *aer)
808 {
809 	struct nvme_request *req;
810 
811 	aer->ctrlr = ctrlr;
812 	req = nvme_allocate_request_null(nvme_ctrlr_async_event_cb, aer);
813 	aer->req = req;
814 
815 	/*
816 	 * Disable timeout here, since asynchronous event requests should by
817 	 *  nature never be timed out.
818 	 */
819 	req->timeout = false;
820 	req->cmd.opc = NVME_OPC_ASYNC_EVENT_REQUEST;
821 	nvme_ctrlr_submit_admin_request(ctrlr, req);
822 }
823 
824 static void
825 nvme_ctrlr_configure_aer(struct nvme_controller *ctrlr)
826 {
827 	struct nvme_completion_poll_status	status;
828 	struct nvme_async_event_request		*aer;
829 	uint32_t				i;
830 
831 	ctrlr->async_event_config = NVME_CRIT_WARN_ST_AVAILABLE_SPARE |
832 	    NVME_CRIT_WARN_ST_DEVICE_RELIABILITY |
833 	    NVME_CRIT_WARN_ST_READ_ONLY |
834 	    NVME_CRIT_WARN_ST_VOLATILE_MEMORY_BACKUP;
835 	if (ctrlr->cdata.ver >= NVME_REV(1, 2))
836 		ctrlr->async_event_config |= NVME_ASYNC_EVENT_NS_ATTRIBUTE |
837 		    NVME_ASYNC_EVENT_FW_ACTIVATE;
838 
839 	status.done = 0;
840 	nvme_ctrlr_cmd_get_feature(ctrlr, NVME_FEAT_TEMPERATURE_THRESHOLD,
841 	    0, NULL, 0, nvme_completion_poll_cb, &status);
842 	nvme_completion_poll(&status);
843 	if (nvme_completion_is_error(&status.cpl) ||
844 	    (status.cpl.cdw0 & 0xFFFF) == 0xFFFF ||
845 	    (status.cpl.cdw0 & 0xFFFF) == 0x0000) {
846 		nvme_printf(ctrlr, "temperature threshold not supported\n");
847 	} else
848 		ctrlr->async_event_config |= NVME_CRIT_WARN_ST_TEMPERATURE;
849 
850 	nvme_ctrlr_cmd_set_async_event_config(ctrlr,
851 	    ctrlr->async_event_config, NULL, NULL);
852 
853 	/* aerl is a zero-based value, so we need to add 1 here. */
854 	ctrlr->num_aers = min(NVME_MAX_ASYNC_EVENTS, (ctrlr->cdata.aerl+1));
855 
856 	for (i = 0; i < ctrlr->num_aers; i++) {
857 		aer = &ctrlr->aer[i];
858 		nvme_ctrlr_construct_and_submit_aer(ctrlr, aer);
859 	}
860 }
861 
862 static void
863 nvme_ctrlr_configure_int_coalescing(struct nvme_controller *ctrlr)
864 {
865 
866 	ctrlr->int_coal_time = 0;
867 	TUNABLE_INT_FETCH("hw.nvme.int_coal_time",
868 	    &ctrlr->int_coal_time);
869 
870 	ctrlr->int_coal_threshold = 0;
871 	TUNABLE_INT_FETCH("hw.nvme.int_coal_threshold",
872 	    &ctrlr->int_coal_threshold);
873 
874 	nvme_ctrlr_cmd_set_interrupt_coalescing(ctrlr, ctrlr->int_coal_time,
875 	    ctrlr->int_coal_threshold, NULL, NULL);
876 }
877 
878 static void
879 nvme_ctrlr_hmb_free(struct nvme_controller *ctrlr)
880 {
881 	struct nvme_hmb_chunk *hmbc;
882 	int i;
883 
884 	if (ctrlr->hmb_desc_paddr) {
885 		bus_dmamap_unload(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_map);
886 		bus_dmamem_free(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_vaddr,
887 		    ctrlr->hmb_desc_map);
888 		ctrlr->hmb_desc_paddr = 0;
889 	}
890 	if (ctrlr->hmb_desc_tag) {
891 		bus_dma_tag_destroy(ctrlr->hmb_desc_tag);
892 		ctrlr->hmb_desc_tag = NULL;
893 	}
894 	for (i = 0; i < ctrlr->hmb_nchunks; i++) {
895 		hmbc = &ctrlr->hmb_chunks[i];
896 		bus_dmamap_unload(ctrlr->hmb_tag, hmbc->hmbc_map);
897 		bus_dmamem_free(ctrlr->hmb_tag, hmbc->hmbc_vaddr,
898 		    hmbc->hmbc_map);
899 	}
900 	ctrlr->hmb_nchunks = 0;
901 	if (ctrlr->hmb_tag) {
902 		bus_dma_tag_destroy(ctrlr->hmb_tag);
903 		ctrlr->hmb_tag = NULL;
904 	}
905 	if (ctrlr->hmb_chunks) {
906 		free(ctrlr->hmb_chunks, M_NVME);
907 		ctrlr->hmb_chunks = NULL;
908 	}
909 }
910 
911 static void
912 nvme_ctrlr_hmb_alloc(struct nvme_controller *ctrlr)
913 {
914 	struct nvme_hmb_chunk *hmbc;
915 	size_t pref, min, minc, size;
916 	int err, i;
917 	uint64_t max;
918 
919 	/* Limit HMB to 5% of RAM size per device by default. */
920 	max = (uint64_t)physmem * PAGE_SIZE / 20;
921 	TUNABLE_UINT64_FETCH("hw.nvme.hmb_max", &max);
922 
923 	min = (long long unsigned)ctrlr->cdata.hmmin * 4096;
924 	if (max == 0 || max < min)
925 		return;
926 	pref = MIN((long long unsigned)ctrlr->cdata.hmpre * 4096, max);
927 	minc = MAX(ctrlr->cdata.hmminds * 4096, PAGE_SIZE);
928 	if (min > 0 && ctrlr->cdata.hmmaxd > 0)
929 		minc = MAX(minc, min / ctrlr->cdata.hmmaxd);
930 	ctrlr->hmb_chunk = pref;
931 
932 again:
933 	ctrlr->hmb_chunk = roundup2(ctrlr->hmb_chunk, PAGE_SIZE);
934 	ctrlr->hmb_nchunks = howmany(pref, ctrlr->hmb_chunk);
935 	if (ctrlr->cdata.hmmaxd > 0 && ctrlr->hmb_nchunks > ctrlr->cdata.hmmaxd)
936 		ctrlr->hmb_nchunks = ctrlr->cdata.hmmaxd;
937 	ctrlr->hmb_chunks = malloc(sizeof(struct nvme_hmb_chunk) *
938 	    ctrlr->hmb_nchunks, M_NVME, M_WAITOK);
939 	err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev),
940 	    PAGE_SIZE, 0, BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
941 	    ctrlr->hmb_chunk, 1, ctrlr->hmb_chunk, 0, NULL, NULL, &ctrlr->hmb_tag);
942 	if (err != 0) {
943 		nvme_printf(ctrlr, "HMB tag create failed %d\n", err);
944 		nvme_ctrlr_hmb_free(ctrlr);
945 		return;
946 	}
947 
948 	for (i = 0; i < ctrlr->hmb_nchunks; i++) {
949 		hmbc = &ctrlr->hmb_chunks[i];
950 		if (bus_dmamem_alloc(ctrlr->hmb_tag,
951 		    (void **)&hmbc->hmbc_vaddr, BUS_DMA_NOWAIT,
952 		    &hmbc->hmbc_map)) {
953 			nvme_printf(ctrlr, "failed to alloc HMB\n");
954 			break;
955 		}
956 		if (bus_dmamap_load(ctrlr->hmb_tag, hmbc->hmbc_map,
957 		    hmbc->hmbc_vaddr, ctrlr->hmb_chunk, nvme_single_map,
958 		    &hmbc->hmbc_paddr, BUS_DMA_NOWAIT) != 0) {
959 			bus_dmamem_free(ctrlr->hmb_tag, hmbc->hmbc_vaddr,
960 			    hmbc->hmbc_map);
961 			nvme_printf(ctrlr, "failed to load HMB\n");
962 			break;
963 		}
964 		bus_dmamap_sync(ctrlr->hmb_tag, hmbc->hmbc_map,
965 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
966 	}
967 
968 	if (i < ctrlr->hmb_nchunks && i * ctrlr->hmb_chunk < min &&
969 	    ctrlr->hmb_chunk / 2 >= minc) {
970 		ctrlr->hmb_nchunks = i;
971 		nvme_ctrlr_hmb_free(ctrlr);
972 		ctrlr->hmb_chunk /= 2;
973 		goto again;
974 	}
975 	ctrlr->hmb_nchunks = i;
976 	if (ctrlr->hmb_nchunks * ctrlr->hmb_chunk < min) {
977 		nvme_ctrlr_hmb_free(ctrlr);
978 		return;
979 	}
980 
981 	size = sizeof(struct nvme_hmb_desc) * ctrlr->hmb_nchunks;
982 	err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev),
983 	    16, 0, BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
984 	    size, 1, size, 0, NULL, NULL, &ctrlr->hmb_desc_tag);
985 	if (err != 0) {
986 		nvme_printf(ctrlr, "HMB desc tag create failed %d\n", err);
987 		nvme_ctrlr_hmb_free(ctrlr);
988 		return;
989 	}
990 	if (bus_dmamem_alloc(ctrlr->hmb_desc_tag,
991 	    (void **)&ctrlr->hmb_desc_vaddr, BUS_DMA_WAITOK,
992 	    &ctrlr->hmb_desc_map)) {
993 		nvme_printf(ctrlr, "failed to alloc HMB desc\n");
994 		nvme_ctrlr_hmb_free(ctrlr);
995 		return;
996 	}
997 	if (bus_dmamap_load(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_map,
998 	    ctrlr->hmb_desc_vaddr, size, nvme_single_map,
999 	    &ctrlr->hmb_desc_paddr, BUS_DMA_NOWAIT) != 0) {
1000 		bus_dmamem_free(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_vaddr,
1001 		    ctrlr->hmb_desc_map);
1002 		nvme_printf(ctrlr, "failed to load HMB desc\n");
1003 		nvme_ctrlr_hmb_free(ctrlr);
1004 		return;
1005 	}
1006 
1007 	for (i = 0; i < ctrlr->hmb_nchunks; i++) {
1008 		ctrlr->hmb_desc_vaddr[i].addr =
1009 		    htole64(ctrlr->hmb_chunks[i].hmbc_paddr);
1010 		ctrlr->hmb_desc_vaddr[i].size = htole32(ctrlr->hmb_chunk / 4096);
1011 	}
1012 	bus_dmamap_sync(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_map,
1013 	    BUS_DMASYNC_PREWRITE);
1014 
1015 	nvme_printf(ctrlr, "Allocated %lluMB host memory buffer\n",
1016 	    (long long unsigned)ctrlr->hmb_nchunks * ctrlr->hmb_chunk
1017 	    / 1024 / 1024);
1018 }
1019 
1020 static void
1021 nvme_ctrlr_hmb_enable(struct nvme_controller *ctrlr, bool enable, bool memret)
1022 {
1023 	struct nvme_completion_poll_status	status;
1024 	uint32_t cdw11;
1025 
1026 	cdw11 = 0;
1027 	if (enable)
1028 		cdw11 |= 1;
1029 	if (memret)
1030 		cdw11 |= 2;
1031 	status.done = 0;
1032 	nvme_ctrlr_cmd_set_feature(ctrlr, NVME_FEAT_HOST_MEMORY_BUFFER, cdw11,
1033 	    ctrlr->hmb_nchunks * ctrlr->hmb_chunk / 4096, ctrlr->hmb_desc_paddr,
1034 	    ctrlr->hmb_desc_paddr >> 32, ctrlr->hmb_nchunks, NULL, 0,
1035 	    nvme_completion_poll_cb, &status);
1036 	nvme_completion_poll(&status);
1037 	if (nvme_completion_is_error(&status.cpl))
1038 		nvme_printf(ctrlr, "nvme_ctrlr_hmb_enable failed!\n");
1039 }
1040 
1041 static void
1042 nvme_ctrlr_start(void *ctrlr_arg, bool resetting)
1043 {
1044 	struct nvme_controller *ctrlr = ctrlr_arg;
1045 	uint32_t old_num_io_queues;
1046 	int i;
1047 
1048 	/*
1049 	 * Only reset adminq here when we are restarting the
1050 	 *  controller after a reset.  During initialization,
1051 	 *  we have already submitted admin commands to get
1052 	 *  the number of I/O queues supported, so cannot reset
1053 	 *  the adminq again here.
1054 	 */
1055 	if (resetting) {
1056 		nvme_qpair_reset(&ctrlr->adminq);
1057 		nvme_admin_qpair_enable(&ctrlr->adminq);
1058 	}
1059 
1060 	if (ctrlr->ioq != NULL) {
1061 		for (i = 0; i < ctrlr->num_io_queues; i++)
1062 			nvme_qpair_reset(&ctrlr->ioq[i]);
1063 	}
1064 
1065 	/*
1066 	 * If it was a reset on initialization command timeout, just
1067 	 * return here, letting initialization code fail gracefully.
1068 	 */
1069 	if (resetting && !ctrlr->is_initialized)
1070 		return;
1071 
1072 	if (resetting && nvme_ctrlr_identify(ctrlr) != 0) {
1073 		nvme_ctrlr_fail(ctrlr);
1074 		return;
1075 	}
1076 
1077 	/*
1078 	 * The number of qpairs are determined during controller initialization,
1079 	 *  including using NVMe SET_FEATURES/NUMBER_OF_QUEUES to determine the
1080 	 *  HW limit.  We call SET_FEATURES again here so that it gets called
1081 	 *  after any reset for controllers that depend on the driver to
1082 	 *  explicit specify how many queues it will use.  This value should
1083 	 *  never change between resets, so panic if somehow that does happen.
1084 	 */
1085 	if (resetting) {
1086 		old_num_io_queues = ctrlr->num_io_queues;
1087 		if (nvme_ctrlr_set_num_qpairs(ctrlr) != 0) {
1088 			nvme_ctrlr_fail(ctrlr);
1089 			return;
1090 		}
1091 
1092 		if (old_num_io_queues != ctrlr->num_io_queues) {
1093 			panic("num_io_queues changed from %u to %u",
1094 			      old_num_io_queues, ctrlr->num_io_queues);
1095 		}
1096 	}
1097 
1098 	if (ctrlr->cdata.hmpre > 0 && ctrlr->hmb_nchunks == 0) {
1099 		nvme_ctrlr_hmb_alloc(ctrlr);
1100 		if (ctrlr->hmb_nchunks > 0)
1101 			nvme_ctrlr_hmb_enable(ctrlr, true, false);
1102 	} else if (ctrlr->hmb_nchunks > 0)
1103 		nvme_ctrlr_hmb_enable(ctrlr, true, true);
1104 
1105 	if (nvme_ctrlr_create_qpairs(ctrlr) != 0) {
1106 		nvme_ctrlr_fail(ctrlr);
1107 		return;
1108 	}
1109 
1110 	if (nvme_ctrlr_construct_namespaces(ctrlr) != 0) {
1111 		nvme_ctrlr_fail(ctrlr);
1112 		return;
1113 	}
1114 
1115 	nvme_ctrlr_configure_aer(ctrlr);
1116 	nvme_ctrlr_configure_int_coalescing(ctrlr);
1117 
1118 	for (i = 0; i < ctrlr->num_io_queues; i++)
1119 		nvme_io_qpair_enable(&ctrlr->ioq[i]);
1120 }
1121 
1122 void
1123 nvme_ctrlr_start_config_hook(void *arg)
1124 {
1125 	struct nvme_controller *ctrlr = arg;
1126 
1127 	/*
1128 	 * Reset controller twice to ensure we do a transition from cc.en==1 to
1129 	 * cc.en==0.  This is because we don't really know what status the
1130 	 * controller was left in when boot handed off to OS.  Linux doesn't do
1131 	 * this, however. If we adopt that policy, see also nvme_ctrlr_resume().
1132 	 */
1133 	if (nvme_ctrlr_hw_reset(ctrlr) != 0) {
1134 fail:
1135 		nvme_ctrlr_fail(ctrlr);
1136 		config_intrhook_disestablish(&ctrlr->config_hook);
1137 		return;
1138 	}
1139 
1140 	if (nvme_ctrlr_hw_reset(ctrlr) != 0)
1141 		goto fail;
1142 
1143 	nvme_qpair_reset(&ctrlr->adminq);
1144 	nvme_admin_qpair_enable(&ctrlr->adminq);
1145 
1146 	if (nvme_ctrlr_identify(ctrlr) == 0 &&
1147 	    nvme_ctrlr_set_num_qpairs(ctrlr) == 0 &&
1148 	    nvme_ctrlr_construct_io_qpairs(ctrlr) == 0)
1149 		nvme_ctrlr_start(ctrlr, false);
1150 	else
1151 		goto fail;
1152 
1153 	nvme_sysctl_initialize_ctrlr(ctrlr);
1154 	config_intrhook_disestablish(&ctrlr->config_hook);
1155 
1156 	ctrlr->is_initialized = 1;
1157 	nvme_notify_new_controller(ctrlr);
1158 }
1159 
1160 static void
1161 nvme_ctrlr_reset_task(void *arg, int pending)
1162 {
1163 	struct nvme_controller	*ctrlr = arg;
1164 	int			status;
1165 
1166 	nvme_ctrlr_devctl_log(ctrlr, "RESET", "resetting controller");
1167 	status = nvme_ctrlr_hw_reset(ctrlr);
1168 	/*
1169 	 * Use pause instead of DELAY, so that we yield to any nvme interrupt
1170 	 *  handlers on this CPU that were blocked on a qpair lock. We want
1171 	 *  all nvme interrupts completed before proceeding with restarting the
1172 	 *  controller.
1173 	 *
1174 	 * XXX - any way to guarantee the interrupt handlers have quiesced?
1175 	 */
1176 	pause("nvmereset", hz / 10);
1177 	if (status == 0)
1178 		nvme_ctrlr_start(ctrlr, true);
1179 	else
1180 		nvme_ctrlr_fail(ctrlr);
1181 
1182 	atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
1183 }
1184 
1185 /*
1186  * Poll all the queues enabled on the device for completion.
1187  */
1188 void
1189 nvme_ctrlr_poll(struct nvme_controller *ctrlr)
1190 {
1191 	int i;
1192 
1193 	nvme_qpair_process_completions(&ctrlr->adminq);
1194 
1195 	for (i = 0; i < ctrlr->num_io_queues; i++)
1196 		if (ctrlr->ioq && ctrlr->ioq[i].cpl)
1197 			nvme_qpair_process_completions(&ctrlr->ioq[i]);
1198 }
1199 
1200 /*
1201  * Poll the single-vector interrupt case: num_io_queues will be 1 and
1202  * there's only a single vector. While we're polling, we mask further
1203  * interrupts in the controller.
1204  */
1205 void
1206 nvme_ctrlr_intx_handler(void *arg)
1207 {
1208 	struct nvme_controller *ctrlr = arg;
1209 
1210 	nvme_mmio_write_4(ctrlr, intms, 1);
1211 	nvme_ctrlr_poll(ctrlr);
1212 	nvme_mmio_write_4(ctrlr, intmc, 1);
1213 }
1214 
1215 static void
1216 nvme_pt_done(void *arg, const struct nvme_completion *cpl)
1217 {
1218 	struct nvme_pt_command *pt = arg;
1219 	struct mtx *mtx = pt->driver_lock;
1220 	uint16_t status;
1221 
1222 	bzero(&pt->cpl, sizeof(pt->cpl));
1223 	pt->cpl.cdw0 = cpl->cdw0;
1224 
1225 	status = cpl->status;
1226 	status &= ~NVME_STATUS_P_MASK;
1227 	pt->cpl.status = status;
1228 
1229 	mtx_lock(mtx);
1230 	pt->driver_lock = NULL;
1231 	wakeup(pt);
1232 	mtx_unlock(mtx);
1233 }
1234 
1235 int
1236 nvme_ctrlr_passthrough_cmd(struct nvme_controller *ctrlr,
1237     struct nvme_pt_command *pt, uint32_t nsid, int is_user_buffer,
1238     int is_admin_cmd)
1239 {
1240 	struct nvme_request	*req;
1241 	struct mtx		*mtx;
1242 	struct buf		*buf = NULL;
1243 	int			ret = 0;
1244 
1245 	if (pt->len > 0) {
1246 		if (pt->len > ctrlr->max_xfer_size) {
1247 			nvme_printf(ctrlr, "pt->len (%d) "
1248 			    "exceeds max_xfer_size (%d)\n", pt->len,
1249 			    ctrlr->max_xfer_size);
1250 			return EIO;
1251 		}
1252 		if (is_user_buffer) {
1253 			/*
1254 			 * Ensure the user buffer is wired for the duration of
1255 			 *  this pass-through command.
1256 			 */
1257 			PHOLD(curproc);
1258 			buf = uma_zalloc(pbuf_zone, M_WAITOK);
1259 			buf->b_iocmd = pt->is_read ? BIO_READ : BIO_WRITE;
1260 			if (vmapbuf(buf, pt->buf, pt->len, 1) < 0) {
1261 				ret = EFAULT;
1262 				goto err;
1263 			}
1264 			req = nvme_allocate_request_vaddr(buf->b_data, pt->len,
1265 			    nvme_pt_done, pt);
1266 		} else
1267 			req = nvme_allocate_request_vaddr(pt->buf, pt->len,
1268 			    nvme_pt_done, pt);
1269 	} else
1270 		req = nvme_allocate_request_null(nvme_pt_done, pt);
1271 
1272 	/* Assume user space already converted to little-endian */
1273 	req->cmd.opc = pt->cmd.opc;
1274 	req->cmd.fuse = pt->cmd.fuse;
1275 	req->cmd.rsvd2 = pt->cmd.rsvd2;
1276 	req->cmd.rsvd3 = pt->cmd.rsvd3;
1277 	req->cmd.cdw10 = pt->cmd.cdw10;
1278 	req->cmd.cdw11 = pt->cmd.cdw11;
1279 	req->cmd.cdw12 = pt->cmd.cdw12;
1280 	req->cmd.cdw13 = pt->cmd.cdw13;
1281 	req->cmd.cdw14 = pt->cmd.cdw14;
1282 	req->cmd.cdw15 = pt->cmd.cdw15;
1283 
1284 	req->cmd.nsid = htole32(nsid);
1285 
1286 	mtx = mtx_pool_find(mtxpool_sleep, pt);
1287 	pt->driver_lock = mtx;
1288 
1289 	if (is_admin_cmd)
1290 		nvme_ctrlr_submit_admin_request(ctrlr, req);
1291 	else
1292 		nvme_ctrlr_submit_io_request(ctrlr, req);
1293 
1294 	mtx_lock(mtx);
1295 	while (pt->driver_lock != NULL)
1296 		mtx_sleep(pt, mtx, PRIBIO, "nvme_pt", 0);
1297 	mtx_unlock(mtx);
1298 
1299 err:
1300 	if (buf != NULL) {
1301 		uma_zfree(pbuf_zone, buf);
1302 		PRELE(curproc);
1303 	}
1304 
1305 	return (ret);
1306 }
1307 
1308 static int
1309 nvme_ctrlr_ioctl(struct cdev *cdev, u_long cmd, caddr_t arg, int flag,
1310     struct thread *td)
1311 {
1312 	struct nvme_controller			*ctrlr;
1313 	struct nvme_pt_command			*pt;
1314 
1315 	ctrlr = cdev->si_drv1;
1316 
1317 	switch (cmd) {
1318 	case NVME_RESET_CONTROLLER:
1319 		nvme_ctrlr_reset(ctrlr);
1320 		break;
1321 	case NVME_PASSTHROUGH_CMD:
1322 		pt = (struct nvme_pt_command *)arg;
1323 		return (nvme_ctrlr_passthrough_cmd(ctrlr, pt, le32toh(pt->cmd.nsid),
1324 		    1 /* is_user_buffer */, 1 /* is_admin_cmd */));
1325 	case NVME_GET_NSID:
1326 	{
1327 		struct nvme_get_nsid *gnsid = (struct nvme_get_nsid *)arg;
1328 		strncpy(gnsid->cdev, device_get_nameunit(ctrlr->dev),
1329 		    sizeof(gnsid->cdev));
1330 		gnsid->cdev[sizeof(gnsid->cdev) - 1] = '\0';
1331 		gnsid->nsid = 0;
1332 		break;
1333 	}
1334 	case NVME_GET_MAX_XFER_SIZE:
1335 		*(uint64_t *)arg = ctrlr->max_xfer_size;
1336 		break;
1337 	default:
1338 		return (ENOTTY);
1339 	}
1340 
1341 	return (0);
1342 }
1343 
1344 static struct cdevsw nvme_ctrlr_cdevsw = {
1345 	.d_version =	D_VERSION,
1346 	.d_flags =	0,
1347 	.d_ioctl =	nvme_ctrlr_ioctl
1348 };
1349 
1350 int
1351 nvme_ctrlr_construct(struct nvme_controller *ctrlr, device_t dev)
1352 {
1353 	struct make_dev_args	md_args;
1354 	uint32_t	cap_lo;
1355 	uint32_t	cap_hi;
1356 	uint32_t	to, vs, pmrcap;
1357 	uint8_t		mpsmin;
1358 	int		status, timeout_period;
1359 
1360 	ctrlr->dev = dev;
1361 
1362 	mtx_init(&ctrlr->lock, "nvme ctrlr lock", NULL, MTX_DEF);
1363 	if (bus_get_domain(dev, &ctrlr->domain) != 0)
1364 		ctrlr->domain = 0;
1365 
1366 	cap_lo = nvme_mmio_read_4(ctrlr, cap_lo);
1367 	if (bootverbose) {
1368 		device_printf(dev, "CapLo: 0x%08x: MQES %u%s%s%s%s, TO %u\n",
1369 		    cap_lo, NVME_CAP_LO_MQES(cap_lo),
1370 		    NVME_CAP_LO_CQR(cap_lo) ? ", CQR" : "",
1371 		    NVME_CAP_LO_AMS(cap_lo) ? ", AMS" : "",
1372 		    (NVME_CAP_LO_AMS(cap_lo) & 0x1) ? " WRRwUPC" : "",
1373 		    (NVME_CAP_LO_AMS(cap_lo) & 0x2) ? " VS" : "",
1374 		    NVME_CAP_LO_TO(cap_lo));
1375 	}
1376 	cap_hi = nvme_mmio_read_4(ctrlr, cap_hi);
1377 	if (bootverbose) {
1378 		device_printf(dev, "CapHi: 0x%08x: DSTRD %u%s, CSS %x%s, "
1379 		    "MPSMIN %u, MPSMAX %u%s%s\n", cap_hi,
1380 		    NVME_CAP_HI_DSTRD(cap_hi),
1381 		    NVME_CAP_HI_NSSRS(cap_hi) ? ", NSSRS" : "",
1382 		    NVME_CAP_HI_CSS(cap_hi),
1383 		    NVME_CAP_HI_BPS(cap_hi) ? ", BPS" : "",
1384 		    NVME_CAP_HI_MPSMIN(cap_hi),
1385 		    NVME_CAP_HI_MPSMAX(cap_hi),
1386 		    NVME_CAP_HI_PMRS(cap_hi) ? ", PMRS" : "",
1387 		    NVME_CAP_HI_CMBS(cap_hi) ? ", CMBS" : "");
1388 	}
1389 	if (bootverbose) {
1390 		vs = nvme_mmio_read_4(ctrlr, vs);
1391 		device_printf(dev, "Version: 0x%08x: %d.%d\n", vs,
1392 		    NVME_MAJOR(vs), NVME_MINOR(vs));
1393 	}
1394 	if (bootverbose && NVME_CAP_HI_PMRS(cap_hi)) {
1395 		pmrcap = nvme_mmio_read_4(ctrlr, pmrcap);
1396 		device_printf(dev, "PMRCap: 0x%08x: BIR %u%s%s, PMRTU %u, "
1397 		    "PMRWBM %x, PMRTO %u%s\n", pmrcap,
1398 		    NVME_PMRCAP_BIR(pmrcap),
1399 		    NVME_PMRCAP_RDS(pmrcap) ? ", RDS" : "",
1400 		    NVME_PMRCAP_WDS(pmrcap) ? ", WDS" : "",
1401 		    NVME_PMRCAP_PMRTU(pmrcap),
1402 		    NVME_PMRCAP_PMRWBM(pmrcap),
1403 		    NVME_PMRCAP_PMRTO(pmrcap),
1404 		    NVME_PMRCAP_CMSS(pmrcap) ? ", CMSS" : "");
1405 	}
1406 
1407 	ctrlr->dstrd = NVME_CAP_HI_DSTRD(cap_hi) + 2;
1408 
1409 	mpsmin = NVME_CAP_HI_MPSMIN(cap_hi);
1410 	ctrlr->min_page_size = 1 << (12 + mpsmin);
1411 
1412 	/* Get ready timeout value from controller, in units of 500ms. */
1413 	to = NVME_CAP_LO_TO(cap_lo) + 1;
1414 	ctrlr->ready_timeout_in_ms = to * 500;
1415 
1416 	timeout_period = NVME_DEFAULT_TIMEOUT_PERIOD;
1417 	TUNABLE_INT_FETCH("hw.nvme.timeout_period", &timeout_period);
1418 	timeout_period = min(timeout_period, NVME_MAX_TIMEOUT_PERIOD);
1419 	timeout_period = max(timeout_period, NVME_MIN_TIMEOUT_PERIOD);
1420 	ctrlr->timeout_period = timeout_period;
1421 
1422 	nvme_retry_count = NVME_DEFAULT_RETRY_COUNT;
1423 	TUNABLE_INT_FETCH("hw.nvme.retry_count", &nvme_retry_count);
1424 
1425 	ctrlr->enable_aborts = 0;
1426 	TUNABLE_INT_FETCH("hw.nvme.enable_aborts", &ctrlr->enable_aborts);
1427 
1428 	ctrlr->max_xfer_size = NVME_MAX_XFER_SIZE;
1429 	if (nvme_ctrlr_construct_admin_qpair(ctrlr) != 0)
1430 		return (ENXIO);
1431 
1432 	/*
1433 	 * Create 2 threads for the taskqueue. The reset thread will block when
1434 	 * it detects that the controller has failed until all I/O has been
1435 	 * failed up the stack. The fail_req task needs to be able to run in
1436 	 * this case to finish the request failure for some cases.
1437 	 *
1438 	 * We could partially solve this race by draining the failed requeust
1439 	 * queue before proceding to free the sim, though nothing would stop
1440 	 * new I/O from coming in after we do that drain, but before we reach
1441 	 * cam_sim_free, so this big hammer is used instead.
1442 	 */
1443 	ctrlr->taskqueue = taskqueue_create("nvme_taskq", M_WAITOK,
1444 	    taskqueue_thread_enqueue, &ctrlr->taskqueue);
1445 	taskqueue_start_threads(&ctrlr->taskqueue, 2, PI_DISK, "nvme taskq");
1446 
1447 	ctrlr->is_resetting = 0;
1448 	ctrlr->is_initialized = 0;
1449 	ctrlr->notification_sent = 0;
1450 	TASK_INIT(&ctrlr->reset_task, 0, nvme_ctrlr_reset_task, ctrlr);
1451 	TASK_INIT(&ctrlr->fail_req_task, 0, nvme_ctrlr_fail_req_task, ctrlr);
1452 	STAILQ_INIT(&ctrlr->fail_req);
1453 	ctrlr->is_failed = false;
1454 
1455 	make_dev_args_init(&md_args);
1456 	md_args.mda_devsw = &nvme_ctrlr_cdevsw;
1457 	md_args.mda_uid = UID_ROOT;
1458 	md_args.mda_gid = GID_WHEEL;
1459 	md_args.mda_mode = 0600;
1460 	md_args.mda_unit = device_get_unit(dev);
1461 	md_args.mda_si_drv1 = (void *)ctrlr;
1462 	status = make_dev_s(&md_args, &ctrlr->cdev, "nvme%d",
1463 	    device_get_unit(dev));
1464 	if (status != 0)
1465 		return (ENXIO);
1466 
1467 	return (0);
1468 }
1469 
1470 void
1471 nvme_ctrlr_destruct(struct nvme_controller *ctrlr, device_t dev)
1472 {
1473 	int	gone, i;
1474 
1475 	if (ctrlr->resource == NULL)
1476 		goto nores;
1477 
1478 	/*
1479 	 * Check whether it is a hot unplug or a clean driver detach.
1480 	 * If device is not there any more, skip any shutdown commands.
1481 	 */
1482 	gone = (nvme_mmio_read_4(ctrlr, csts) == NVME_GONE);
1483 	if (gone)
1484 		nvme_ctrlr_fail(ctrlr);
1485 	else
1486 		nvme_notify_fail_consumers(ctrlr);
1487 
1488 	for (i = 0; i < NVME_MAX_NAMESPACES; i++)
1489 		nvme_ns_destruct(&ctrlr->ns[i]);
1490 
1491 	if (ctrlr->cdev)
1492 		destroy_dev(ctrlr->cdev);
1493 
1494 	if (ctrlr->is_initialized) {
1495 		if (!gone) {
1496 			if (ctrlr->hmb_nchunks > 0)
1497 				nvme_ctrlr_hmb_enable(ctrlr, false, false);
1498 			nvme_ctrlr_delete_qpairs(ctrlr);
1499 		}
1500 		nvme_ctrlr_hmb_free(ctrlr);
1501 	}
1502 	if (ctrlr->ioq != NULL) {
1503 		for (i = 0; i < ctrlr->num_io_queues; i++)
1504 			nvme_io_qpair_destroy(&ctrlr->ioq[i]);
1505 		free(ctrlr->ioq, M_NVME);
1506 	}
1507 	nvme_admin_qpair_destroy(&ctrlr->adminq);
1508 
1509 	/*
1510 	 *  Notify the controller of a shutdown, even though this is due to
1511 	 *   a driver unload, not a system shutdown (this path is not invoked
1512 	 *   during shutdown).  This ensures the controller receives a
1513 	 *   shutdown notification in case the system is shutdown before
1514 	 *   reloading the driver.
1515 	 */
1516 	if (!gone)
1517 		nvme_ctrlr_shutdown(ctrlr);
1518 
1519 	if (!gone)
1520 		nvme_ctrlr_disable(ctrlr);
1521 
1522 	if (ctrlr->taskqueue)
1523 		taskqueue_free(ctrlr->taskqueue);
1524 
1525 	if (ctrlr->tag)
1526 		bus_teardown_intr(ctrlr->dev, ctrlr->res, ctrlr->tag);
1527 
1528 	if (ctrlr->res)
1529 		bus_release_resource(ctrlr->dev, SYS_RES_IRQ,
1530 		    rman_get_rid(ctrlr->res), ctrlr->res);
1531 
1532 	if (ctrlr->bar4_resource != NULL) {
1533 		bus_release_resource(dev, SYS_RES_MEMORY,
1534 		    ctrlr->bar4_resource_id, ctrlr->bar4_resource);
1535 	}
1536 
1537 	bus_release_resource(dev, SYS_RES_MEMORY,
1538 	    ctrlr->resource_id, ctrlr->resource);
1539 
1540 nores:
1541 	mtx_destroy(&ctrlr->lock);
1542 }
1543 
1544 void
1545 nvme_ctrlr_shutdown(struct nvme_controller *ctrlr)
1546 {
1547 	uint32_t	cc;
1548 	uint32_t	csts;
1549 	int		timeout;
1550 
1551 	cc = nvme_mmio_read_4(ctrlr, cc);
1552 	cc &= ~(NVME_CC_REG_SHN_MASK << NVME_CC_REG_SHN_SHIFT);
1553 	cc |= NVME_SHN_NORMAL << NVME_CC_REG_SHN_SHIFT;
1554 	nvme_mmio_write_4(ctrlr, cc, cc);
1555 
1556 	timeout = ticks + (ctrlr->cdata.rtd3e == 0 ? 5 * hz :
1557 	    ((uint64_t)ctrlr->cdata.rtd3e * hz + 999999) / 1000000);
1558 	while (1) {
1559 		csts = nvme_mmio_read_4(ctrlr, csts);
1560 		if (csts == NVME_GONE)		/* Hot unplug. */
1561 			break;
1562 		if (NVME_CSTS_GET_SHST(csts) == NVME_SHST_COMPLETE)
1563 			break;
1564 		if (timeout - ticks < 0) {
1565 			nvme_printf(ctrlr, "shutdown timeout\n");
1566 			break;
1567 		}
1568 		pause("nvmeshut", 1);
1569 	}
1570 }
1571 
1572 void
1573 nvme_ctrlr_submit_admin_request(struct nvme_controller *ctrlr,
1574     struct nvme_request *req)
1575 {
1576 
1577 	nvme_qpair_submit_request(&ctrlr->adminq, req);
1578 }
1579 
1580 void
1581 nvme_ctrlr_submit_io_request(struct nvme_controller *ctrlr,
1582     struct nvme_request *req)
1583 {
1584 	struct nvme_qpair       *qpair;
1585 
1586 	qpair = &ctrlr->ioq[QP(ctrlr, curcpu)];
1587 	nvme_qpair_submit_request(qpair, req);
1588 }
1589 
1590 device_t
1591 nvme_ctrlr_get_device(struct nvme_controller *ctrlr)
1592 {
1593 
1594 	return (ctrlr->dev);
1595 }
1596 
1597 const struct nvme_controller_data *
1598 nvme_ctrlr_get_data(struct nvme_controller *ctrlr)
1599 {
1600 
1601 	return (&ctrlr->cdata);
1602 }
1603 
1604 int
1605 nvme_ctrlr_suspend(struct nvme_controller *ctrlr)
1606 {
1607 	int to = hz;
1608 
1609 	/*
1610 	 * Can't touch failed controllers, so it's already suspended.
1611 	 */
1612 	if (ctrlr->is_failed)
1613 		return (0);
1614 
1615 	/*
1616 	 * We don't want the reset taskqueue running, since it does similar
1617 	 * things, so prevent it from running after we start. Wait for any reset
1618 	 * that may have been started to complete. The reset process we follow
1619 	 * will ensure that any new I/O will queue and be given to the hardware
1620 	 * after we resume (though there should be none).
1621 	 */
1622 	while (atomic_cmpset_32(&ctrlr->is_resetting, 0, 1) == 0 && to-- > 0)
1623 		pause("nvmesusp", 1);
1624 	if (to <= 0) {
1625 		nvme_printf(ctrlr,
1626 		    "Competing reset task didn't finish. Try again later.\n");
1627 		return (EWOULDBLOCK);
1628 	}
1629 
1630 	if (ctrlr->hmb_nchunks > 0)
1631 		nvme_ctrlr_hmb_enable(ctrlr, false, false);
1632 
1633 	/*
1634 	 * Per Section 7.6.2 of NVMe spec 1.4, to properly suspend, we need to
1635 	 * delete the hardware I/O queues, and then shutdown. This properly
1636 	 * flushes any metadata the drive may have stored so it can survive
1637 	 * having its power removed and prevents the unsafe shutdown count from
1638 	 * incriminating. Once we delete the qpairs, we have to disable them
1639 	 * before shutting down. The delay is out of paranoia in
1640 	 * nvme_ctrlr_hw_reset, and is repeated here (though we should have no
1641 	 * pending I/O that the delay copes with).
1642 	 */
1643 	nvme_ctrlr_delete_qpairs(ctrlr);
1644 	nvme_ctrlr_disable_qpairs(ctrlr);
1645 	pause("nvmesusp", hz / 10);
1646 	nvme_ctrlr_shutdown(ctrlr);
1647 
1648 	return (0);
1649 }
1650 
1651 int
1652 nvme_ctrlr_resume(struct nvme_controller *ctrlr)
1653 {
1654 
1655 	/*
1656 	 * Can't touch failed controllers, so nothing to do to resume.
1657 	 */
1658 	if (ctrlr->is_failed)
1659 		return (0);
1660 
1661 	/*
1662 	 * Have to reset the hardware twice, just like we do on attach. See
1663 	 * nmve_attach() for why.
1664 	 */
1665 	if (nvme_ctrlr_hw_reset(ctrlr) != 0)
1666 		goto fail;
1667 	if (nvme_ctrlr_hw_reset(ctrlr) != 0)
1668 		goto fail;
1669 
1670 	/*
1671 	 * Now that we've reset the hardware, we can restart the controller. Any
1672 	 * I/O that was pending is requeued. Any admin commands are aborted with
1673 	 * an error. Once we've restarted, take the controller out of reset.
1674 	 */
1675 	nvme_ctrlr_start(ctrlr, true);
1676 	(void)atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
1677 
1678 	return (0);
1679 fail:
1680 	/*
1681 	 * Since we can't bring the controller out of reset, announce and fail
1682 	 * the controller. However, we have to return success for the resume
1683 	 * itself, due to questionable APIs.
1684 	 */
1685 	nvme_printf(ctrlr, "Failed to reset on resume, failing.\n");
1686 	nvme_ctrlr_fail(ctrlr);
1687 	(void)atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
1688 	return (0);
1689 }
1690