xref: /freebsd/sys/dev/nvme/nvme_ctrlr.c (revision 2e620256bd76c449c835c604e404483437743011)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
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 #include "opt_cam.h"
31 #include "opt_nvme.h"
32 
33 #include <sys/param.h>
34 #include <sys/systm.h>
35 #include <sys/buf.h>
36 #include <sys/bus.h>
37 #include <sys/conf.h>
38 #include <sys/ioccom.h>
39 #include <sys/proc.h>
40 #include <sys/smp.h>
41 #include <sys/uio.h>
42 #include <sys/sbuf.h>
43 #include <sys/endian.h>
44 #include <machine/stdarg.h>
45 #include <vm/vm.h>
46 
47 #include "nvme_private.h"
48 
49 #define B4_CHK_RDY_DELAY_MS	2300		/* work around controller bug */
50 
51 static void nvme_ctrlr_construct_and_submit_aer(struct nvme_controller *ctrlr,
52 						struct nvme_async_event_request *aer);
53 
54 static void
55 nvme_ctrlr_barrier(struct nvme_controller *ctrlr, int flags)
56 {
57 	bus_barrier(ctrlr->resource, 0, rman_get_size(ctrlr->resource), flags);
58 }
59 
60 static void
61 nvme_ctrlr_devctl_log(struct nvme_controller *ctrlr, const char *type, const char *msg, ...)
62 {
63 	struct sbuf sb;
64 	va_list ap;
65 	int error;
66 
67 	if (sbuf_new(&sb, NULL, 0, SBUF_AUTOEXTEND | SBUF_NOWAIT) == NULL)
68 		return;
69 	sbuf_printf(&sb, "%s: ", device_get_nameunit(ctrlr->dev));
70 	va_start(ap, msg);
71 	sbuf_vprintf(&sb, msg, ap);
72 	va_end(ap);
73 	error = sbuf_finish(&sb);
74 	if (error == 0)
75 		printf("%s\n", sbuf_data(&sb));
76 
77 	sbuf_clear(&sb);
78 	sbuf_printf(&sb, "name=\"%s\" reason=\"", device_get_nameunit(ctrlr->dev));
79 	va_start(ap, msg);
80 	sbuf_vprintf(&sb, msg, ap);
81 	va_end(ap);
82 	sbuf_printf(&sb, "\"");
83 	error = sbuf_finish(&sb);
84 	if (error == 0)
85 		devctl_notify("nvme", "controller", type, sbuf_data(&sb));
86 	sbuf_delete(&sb);
87 }
88 
89 static int
90 nvme_ctrlr_construct_admin_qpair(struct nvme_controller *ctrlr)
91 {
92 	struct nvme_qpair	*qpair;
93 	uint32_t		num_entries;
94 	int			error;
95 
96 	qpair = &ctrlr->adminq;
97 	qpair->id = 0;
98 	qpair->cpu = CPU_FFS(&cpuset_domain[ctrlr->domain]) - 1;
99 	qpair->domain = ctrlr->domain;
100 
101 	num_entries = NVME_ADMIN_ENTRIES;
102 	TUNABLE_INT_FETCH("hw.nvme.admin_entries", &num_entries);
103 	/*
104 	 * If admin_entries was overridden to an invalid value, revert it
105 	 *  back to our default value.
106 	 */
107 	if (num_entries < NVME_MIN_ADMIN_ENTRIES ||
108 	    num_entries > NVME_MAX_ADMIN_ENTRIES) {
109 		nvme_printf(ctrlr, "invalid hw.nvme.admin_entries=%d "
110 		    "specified\n", num_entries);
111 		num_entries = NVME_ADMIN_ENTRIES;
112 	}
113 
114 	/*
115 	 * The admin queue's max xfer size is treated differently than the
116 	 *  max I/O xfer size.  16KB is sufficient here - maybe even less?
117 	 */
118 	error = nvme_qpair_construct(qpair, num_entries, NVME_ADMIN_TRACKERS,
119 	     ctrlr);
120 	return (error);
121 }
122 
123 #define QP(ctrlr, c)	((c) * (ctrlr)->num_io_queues / mp_ncpus)
124 
125 static int
126 nvme_ctrlr_construct_io_qpairs(struct nvme_controller *ctrlr)
127 {
128 	struct nvme_qpair	*qpair;
129 	uint32_t		cap_lo;
130 	uint16_t		mqes;
131 	int			c, error, i, n;
132 	int			num_entries, num_trackers, max_entries;
133 
134 	/*
135 	 * NVMe spec sets a hard limit of 64K max entries, but devices may
136 	 * specify a smaller limit, so we need to check the MQES field in the
137 	 * capabilities register. We have to cap the number of entries to the
138 	 * current stride allows for in BAR 0/1, otherwise the remainder entries
139 	 * are inaccessible. MQES should reflect this, and this is just a
140 	 * fail-safe.
141 	 */
142 	max_entries =
143 	    (rman_get_size(ctrlr->resource) - nvme_mmio_offsetof(doorbell[0])) /
144 	    (1 << (ctrlr->dstrd + 1));
145 	num_entries = NVME_IO_ENTRIES;
146 	TUNABLE_INT_FETCH("hw.nvme.io_entries", &num_entries);
147 	cap_lo = nvme_mmio_read_4(ctrlr, cap_lo);
148 	mqes = NVME_CAP_LO_MQES(cap_lo);
149 	num_entries = min(num_entries, mqes + 1);
150 	num_entries = min(num_entries, max_entries);
151 
152 	num_trackers = NVME_IO_TRACKERS;
153 	TUNABLE_INT_FETCH("hw.nvme.io_trackers", &num_trackers);
154 
155 	num_trackers = max(num_trackers, NVME_MIN_IO_TRACKERS);
156 	num_trackers = min(num_trackers, NVME_MAX_IO_TRACKERS);
157 	/*
158 	 * No need to have more trackers than entries in the submit queue.  Note
159 	 * also that for a queue size of N, we can only have (N-1) commands
160 	 * outstanding, hence the "-1" here.
161 	 */
162 	num_trackers = min(num_trackers, (num_entries-1));
163 
164 	/*
165 	 * Our best estimate for the maximum number of I/Os that we should
166 	 * normally have in flight at one time. This should be viewed as a hint,
167 	 * not a hard limit and will need to be revisited when the upper layers
168 	 * of the storage system grows multi-queue support.
169 	 */
170 	ctrlr->max_hw_pend_io = num_trackers * ctrlr->num_io_queues * 3 / 4;
171 
172 	ctrlr->ioq = malloc(ctrlr->num_io_queues * sizeof(struct nvme_qpair),
173 	    M_NVME, M_ZERO | M_WAITOK);
174 
175 	for (i = c = n = 0; i < ctrlr->num_io_queues; i++, c += n) {
176 		qpair = &ctrlr->ioq[i];
177 
178 		/*
179 		 * Admin queue has ID=0. IO queues start at ID=1 -
180 		 *  hence the 'i+1' here.
181 		 */
182 		qpair->id = i + 1;
183 		if (ctrlr->num_io_queues > 1) {
184 			/* Find number of CPUs served by this queue. */
185 			for (n = 1; QP(ctrlr, c + n) == i; n++)
186 				;
187 			/* Shuffle multiple NVMe devices between CPUs. */
188 			qpair->cpu = c + (device_get_unit(ctrlr->dev)+n/2) % n;
189 			qpair->domain = pcpu_find(qpair->cpu)->pc_domain;
190 		} else {
191 			qpair->cpu = CPU_FFS(&cpuset_domain[ctrlr->domain]) - 1;
192 			qpair->domain = ctrlr->domain;
193 		}
194 
195 		/*
196 		 * For I/O queues, use the controller-wide max_xfer_size
197 		 *  calculated in nvme_attach().
198 		 */
199 		error = nvme_qpair_construct(qpair, num_entries, num_trackers,
200 		    ctrlr);
201 		if (error)
202 			return (error);
203 
204 		/*
205 		 * Do not bother binding interrupts if we only have one I/O
206 		 *  interrupt thread for this controller.
207 		 */
208 		if (ctrlr->num_io_queues > 1)
209 			bus_bind_intr(ctrlr->dev, qpair->res, qpair->cpu);
210 	}
211 
212 	return (0);
213 }
214 
215 static void
216 nvme_ctrlr_fail(struct nvme_controller *ctrlr)
217 {
218 	int i;
219 
220 	/*
221 	 * No need to disable queues before failing them. Failing is a superet
222 	 * of disabling (though pedantically we'd abort the AERs silently with
223 	 * a different error, though when we fail, that hardly matters).
224 	 */
225 	ctrlr->is_failed = true;
226 	nvme_qpair_fail(&ctrlr->adminq);
227 	if (ctrlr->ioq != NULL) {
228 		for (i = 0; i < ctrlr->num_io_queues; i++) {
229 			nvme_qpair_fail(&ctrlr->ioq[i]);
230 		}
231 	}
232 	nvme_notify_fail_consumers(ctrlr);
233 }
234 
235 /*
236  * Wait for RDY to change.
237  *
238  * Starts sleeping for 1us and geometrically increases it the longer we wait,
239  * capped at 1ms.
240  */
241 static int
242 nvme_ctrlr_wait_for_ready(struct nvme_controller *ctrlr, int desired_val)
243 {
244 	int timeout = ticks + MSEC_2_TICKS(ctrlr->ready_timeout_in_ms);
245 	sbintime_t delta_t = SBT_1US;
246 	uint32_t csts;
247 
248 	while (1) {
249 		csts = nvme_mmio_read_4(ctrlr, csts);
250 		if (csts == NVME_GONE)		/* Hot unplug. */
251 			return (ENXIO);
252 		if (((csts >> NVME_CSTS_REG_RDY_SHIFT) & NVME_CSTS_REG_RDY_MASK)
253 		    == desired_val)
254 			break;
255 		if (timeout - ticks < 0) {
256 			nvme_printf(ctrlr, "controller ready did not become %d "
257 			    "within %d ms\n", desired_val, ctrlr->ready_timeout_in_ms);
258 			return (ENXIO);
259 		}
260 
261 		pause_sbt("nvmerdy", delta_t, 0, C_PREL(1));
262 		delta_t = min(SBT_1MS, delta_t * 3 / 2);
263 	}
264 
265 	return (0);
266 }
267 
268 static int
269 nvme_ctrlr_disable(struct nvme_controller *ctrlr)
270 {
271 	uint32_t cc;
272 	uint32_t csts;
273 	uint8_t  en, rdy;
274 	int err;
275 
276 	cc = nvme_mmio_read_4(ctrlr, cc);
277 	csts = nvme_mmio_read_4(ctrlr, csts);
278 
279 	en = (cc >> NVME_CC_REG_EN_SHIFT) & NVME_CC_REG_EN_MASK;
280 	rdy = (csts >> NVME_CSTS_REG_RDY_SHIFT) & NVME_CSTS_REG_RDY_MASK;
281 
282 	/*
283 	 * Per 3.1.5 in NVME 1.3 spec, transitioning CC.EN from 0 to 1
284 	 * when CSTS.RDY is 1 or transitioning CC.EN from 1 to 0 when
285 	 * CSTS.RDY is 0 "has undefined results" So make sure that CSTS.RDY
286 	 * isn't the desired value. Short circuit if we're already disabled.
287 	 */
288 	if (en == 0) {
289 		/* Wait for RDY == 0 or timeout & fail */
290 		if (rdy == 0)
291 			return (0);
292 		return (nvme_ctrlr_wait_for_ready(ctrlr, 0));
293 	}
294 	if (rdy == 0) {
295 		/* EN == 1, wait for  RDY == 1 or timeout & fail */
296 		err = nvme_ctrlr_wait_for_ready(ctrlr, 1);
297 		if (err != 0)
298 			return (err);
299 	}
300 
301 	cc &= ~NVME_CC_REG_EN_MASK;
302 	nvme_mmio_write_4(ctrlr, cc, cc);
303 
304 	/*
305 	 * A few drives have firmware bugs that freeze the drive if we access
306 	 * the mmio too soon after we disable.
307 	 */
308 	if (ctrlr->quirks & QUIRK_DELAY_B4_CHK_RDY)
309 		pause("nvmeR", MSEC_2_TICKS(B4_CHK_RDY_DELAY_MS));
310 	return (nvme_ctrlr_wait_for_ready(ctrlr, 0));
311 }
312 
313 static int
314 nvme_ctrlr_enable(struct nvme_controller *ctrlr)
315 {
316 	uint32_t	cc;
317 	uint32_t	csts;
318 	uint32_t	aqa;
319 	uint32_t	qsize;
320 	uint8_t		en, rdy;
321 	int		err;
322 
323 	cc = nvme_mmio_read_4(ctrlr, cc);
324 	csts = nvme_mmio_read_4(ctrlr, csts);
325 
326 	en = (cc >> NVME_CC_REG_EN_SHIFT) & NVME_CC_REG_EN_MASK;
327 	rdy = (csts >> NVME_CSTS_REG_RDY_SHIFT) & NVME_CSTS_REG_RDY_MASK;
328 
329 	/*
330 	 * See note in nvme_ctrlr_disable. Short circuit if we're already enabled.
331 	 */
332 	if (en == 1) {
333 		if (rdy == 1)
334 			return (0);
335 		return (nvme_ctrlr_wait_for_ready(ctrlr, 1));
336 	}
337 
338 	/* EN == 0 already wait for RDY == 0 or timeout & fail */
339 	err = nvme_ctrlr_wait_for_ready(ctrlr, 0);
340 	if (err != 0)
341 		return (err);
342 
343 	nvme_mmio_write_8(ctrlr, asq, ctrlr->adminq.cmd_bus_addr);
344 	nvme_mmio_write_8(ctrlr, acq, ctrlr->adminq.cpl_bus_addr);
345 
346 	/* acqs and asqs are 0-based. */
347 	qsize = ctrlr->adminq.num_entries - 1;
348 
349 	aqa = 0;
350 	aqa = (qsize & NVME_AQA_REG_ACQS_MASK) << NVME_AQA_REG_ACQS_SHIFT;
351 	aqa |= (qsize & NVME_AQA_REG_ASQS_MASK) << NVME_AQA_REG_ASQS_SHIFT;
352 	nvme_mmio_write_4(ctrlr, aqa, aqa);
353 
354 	/* Initialization values for CC */
355 	cc = 0;
356 	cc |= 1 << NVME_CC_REG_EN_SHIFT;
357 	cc |= 0 << NVME_CC_REG_CSS_SHIFT;
358 	cc |= 0 << NVME_CC_REG_AMS_SHIFT;
359 	cc |= 0 << NVME_CC_REG_SHN_SHIFT;
360 	cc |= 6 << NVME_CC_REG_IOSQES_SHIFT; /* SQ entry size == 64 == 2^6 */
361 	cc |= 4 << NVME_CC_REG_IOCQES_SHIFT; /* CQ entry size == 16 == 2^4 */
362 
363 	/*
364 	 * Use the Memory Page Size selected during device initialization.  Note
365 	 * that value stored in mps is suitable to use here without adjusting by
366 	 * NVME_MPS_SHIFT.
367 	 */
368 	cc |= ctrlr->mps << NVME_CC_REG_MPS_SHIFT;
369 
370 	nvme_ctrlr_barrier(ctrlr, BUS_SPACE_BARRIER_WRITE);
371 	nvme_mmio_write_4(ctrlr, cc, cc);
372 
373 	return (nvme_ctrlr_wait_for_ready(ctrlr, 1));
374 }
375 
376 static void
377 nvme_ctrlr_disable_qpairs(struct nvme_controller *ctrlr)
378 {
379 	int i;
380 
381 	nvme_admin_qpair_disable(&ctrlr->adminq);
382 	/*
383 	 * I/O queues are not allocated before the initial HW
384 	 *  reset, so do not try to disable them.  Use is_initialized
385 	 *  to determine if this is the initial HW reset.
386 	 */
387 	if (ctrlr->is_initialized) {
388 		for (i = 0; i < ctrlr->num_io_queues; i++)
389 			nvme_io_qpair_disable(&ctrlr->ioq[i]);
390 	}
391 }
392 
393 static int
394 nvme_ctrlr_hw_reset(struct nvme_controller *ctrlr)
395 {
396 	int err;
397 
398 	TSENTER();
399 
400 	nvme_ctrlr_disable_qpairs(ctrlr);
401 
402 	err = nvme_ctrlr_disable(ctrlr);
403 	if (err != 0)
404 		goto out;
405 
406 	err = nvme_ctrlr_enable(ctrlr);
407 out:
408 
409 	TSEXIT();
410 	return (err);
411 }
412 
413 void
414 nvme_ctrlr_reset(struct nvme_controller *ctrlr)
415 {
416 	int cmpset;
417 
418 	cmpset = atomic_cmpset_32(&ctrlr->is_resetting, 0, 1);
419 
420 	if (cmpset == 0 || ctrlr->is_failed)
421 		/*
422 		 * Controller is already resetting or has failed.  Return
423 		 *  immediately since there is no need to kick off another
424 		 *  reset in these cases.
425 		 */
426 		return;
427 
428 	if (!ctrlr->is_dying)
429 		taskqueue_enqueue(ctrlr->taskqueue, &ctrlr->reset_task);
430 }
431 
432 static int
433 nvme_ctrlr_identify(struct nvme_controller *ctrlr)
434 {
435 	struct nvme_completion_poll_status	status;
436 
437 	status.done = 0;
438 	nvme_ctrlr_cmd_identify_controller(ctrlr, &ctrlr->cdata,
439 	    nvme_completion_poll_cb, &status);
440 	nvme_completion_poll(&status);
441 	if (nvme_completion_is_error(&status.cpl)) {
442 		nvme_printf(ctrlr, "nvme_identify_controller failed!\n");
443 		return (ENXIO);
444 	}
445 
446 	/* Convert data to host endian */
447 	nvme_controller_data_swapbytes(&ctrlr->cdata);
448 
449 	/*
450 	 * Use MDTS to ensure our default max_xfer_size doesn't exceed what the
451 	 *  controller supports.
452 	 */
453 	if (ctrlr->cdata.mdts > 0)
454 		ctrlr->max_xfer_size = min(ctrlr->max_xfer_size,
455 		    1 << (ctrlr->cdata.mdts + NVME_MPS_SHIFT +
456 			NVME_CAP_HI_MPSMIN(ctrlr->cap_hi)));
457 
458 	return (0);
459 }
460 
461 static int
462 nvme_ctrlr_set_num_qpairs(struct nvme_controller *ctrlr)
463 {
464 	struct nvme_completion_poll_status	status;
465 	int					cq_allocated, sq_allocated;
466 
467 	status.done = 0;
468 	nvme_ctrlr_cmd_set_num_queues(ctrlr, ctrlr->num_io_queues,
469 	    nvme_completion_poll_cb, &status);
470 	nvme_completion_poll(&status);
471 	if (nvme_completion_is_error(&status.cpl)) {
472 		nvme_printf(ctrlr, "nvme_ctrlr_set_num_qpairs failed!\n");
473 		return (ENXIO);
474 	}
475 
476 	/*
477 	 * Data in cdw0 is 0-based.
478 	 * Lower 16-bits indicate number of submission queues allocated.
479 	 * Upper 16-bits indicate number of completion queues allocated.
480 	 */
481 	sq_allocated = (status.cpl.cdw0 & 0xFFFF) + 1;
482 	cq_allocated = (status.cpl.cdw0 >> 16) + 1;
483 
484 	/*
485 	 * Controller may allocate more queues than we requested,
486 	 *  so use the minimum of the number requested and what was
487 	 *  actually allocated.
488 	 */
489 	ctrlr->num_io_queues = min(ctrlr->num_io_queues, sq_allocated);
490 	ctrlr->num_io_queues = min(ctrlr->num_io_queues, cq_allocated);
491 	if (ctrlr->num_io_queues > vm_ndomains)
492 		ctrlr->num_io_queues -= ctrlr->num_io_queues % vm_ndomains;
493 
494 	return (0);
495 }
496 
497 static int
498 nvme_ctrlr_create_qpairs(struct nvme_controller *ctrlr)
499 {
500 	struct nvme_completion_poll_status	status;
501 	struct nvme_qpair			*qpair;
502 	int					i;
503 
504 	for (i = 0; i < ctrlr->num_io_queues; i++) {
505 		qpair = &ctrlr->ioq[i];
506 
507 		status.done = 0;
508 		nvme_ctrlr_cmd_create_io_cq(ctrlr, qpair,
509 		    nvme_completion_poll_cb, &status);
510 		nvme_completion_poll(&status);
511 		if (nvme_completion_is_error(&status.cpl)) {
512 			nvme_printf(ctrlr, "nvme_create_io_cq failed!\n");
513 			return (ENXIO);
514 		}
515 
516 		status.done = 0;
517 		nvme_ctrlr_cmd_create_io_sq(ctrlr, qpair,
518 		    nvme_completion_poll_cb, &status);
519 		nvme_completion_poll(&status);
520 		if (nvme_completion_is_error(&status.cpl)) {
521 			nvme_printf(ctrlr, "nvme_create_io_sq failed!\n");
522 			return (ENXIO);
523 		}
524 	}
525 
526 	return (0);
527 }
528 
529 static int
530 nvme_ctrlr_delete_qpairs(struct nvme_controller *ctrlr)
531 {
532 	struct nvme_completion_poll_status	status;
533 	struct nvme_qpair			*qpair;
534 
535 	for (int i = 0; i < ctrlr->num_io_queues; i++) {
536 		qpair = &ctrlr->ioq[i];
537 
538 		status.done = 0;
539 		nvme_ctrlr_cmd_delete_io_sq(ctrlr, qpair,
540 		    nvme_completion_poll_cb, &status);
541 		nvme_completion_poll(&status);
542 		if (nvme_completion_is_error(&status.cpl)) {
543 			nvme_printf(ctrlr, "nvme_destroy_io_sq failed!\n");
544 			return (ENXIO);
545 		}
546 
547 		status.done = 0;
548 		nvme_ctrlr_cmd_delete_io_cq(ctrlr, qpair,
549 		    nvme_completion_poll_cb, &status);
550 		nvme_completion_poll(&status);
551 		if (nvme_completion_is_error(&status.cpl)) {
552 			nvme_printf(ctrlr, "nvme_destroy_io_cq failed!\n");
553 			return (ENXIO);
554 		}
555 	}
556 
557 	return (0);
558 }
559 
560 static int
561 nvme_ctrlr_construct_namespaces(struct nvme_controller *ctrlr)
562 {
563 	struct nvme_namespace	*ns;
564 	uint32_t 		i;
565 
566 	for (i = 0; i < min(ctrlr->cdata.nn, NVME_MAX_NAMESPACES); i++) {
567 		ns = &ctrlr->ns[i];
568 		nvme_ns_construct(ns, i+1, ctrlr);
569 	}
570 
571 	return (0);
572 }
573 
574 static bool
575 is_log_page_id_valid(uint8_t page_id)
576 {
577 
578 	switch (page_id) {
579 	case NVME_LOG_ERROR:
580 	case NVME_LOG_HEALTH_INFORMATION:
581 	case NVME_LOG_FIRMWARE_SLOT:
582 	case NVME_LOG_CHANGED_NAMESPACE:
583 	case NVME_LOG_COMMAND_EFFECT:
584 	case NVME_LOG_RES_NOTIFICATION:
585 	case NVME_LOG_SANITIZE_STATUS:
586 		return (true);
587 	}
588 
589 	return (false);
590 }
591 
592 static uint32_t
593 nvme_ctrlr_get_log_page_size(struct nvme_controller *ctrlr, uint8_t page_id)
594 {
595 	uint32_t	log_page_size;
596 
597 	switch (page_id) {
598 	case NVME_LOG_ERROR:
599 		log_page_size = min(
600 		    sizeof(struct nvme_error_information_entry) *
601 		    (ctrlr->cdata.elpe + 1), NVME_MAX_AER_LOG_SIZE);
602 		break;
603 	case NVME_LOG_HEALTH_INFORMATION:
604 		log_page_size = sizeof(struct nvme_health_information_page);
605 		break;
606 	case NVME_LOG_FIRMWARE_SLOT:
607 		log_page_size = sizeof(struct nvme_firmware_page);
608 		break;
609 	case NVME_LOG_CHANGED_NAMESPACE:
610 		log_page_size = sizeof(struct nvme_ns_list);
611 		break;
612 	case NVME_LOG_COMMAND_EFFECT:
613 		log_page_size = sizeof(struct nvme_command_effects_page);
614 		break;
615 	case NVME_LOG_RES_NOTIFICATION:
616 		log_page_size = sizeof(struct nvme_res_notification_page);
617 		break;
618 	case NVME_LOG_SANITIZE_STATUS:
619 		log_page_size = sizeof(struct nvme_sanitize_status_page);
620 		break;
621 	default:
622 		log_page_size = 0;
623 		break;
624 	}
625 
626 	return (log_page_size);
627 }
628 
629 static void
630 nvme_ctrlr_log_critical_warnings(struct nvme_controller *ctrlr,
631     uint8_t state)
632 {
633 
634 	if (state & NVME_CRIT_WARN_ST_AVAILABLE_SPARE)
635 		nvme_ctrlr_devctl_log(ctrlr, "critical",
636 		    "available spare space below threshold");
637 
638 	if (state & NVME_CRIT_WARN_ST_TEMPERATURE)
639 		nvme_ctrlr_devctl_log(ctrlr, "critical",
640 		    "temperature above threshold");
641 
642 	if (state & NVME_CRIT_WARN_ST_DEVICE_RELIABILITY)
643 		nvme_ctrlr_devctl_log(ctrlr, "critical",
644 		    "device reliability degraded");
645 
646 	if (state & NVME_CRIT_WARN_ST_READ_ONLY)
647 		nvme_ctrlr_devctl_log(ctrlr, "critical",
648 		    "media placed in read only mode");
649 
650 	if (state & NVME_CRIT_WARN_ST_VOLATILE_MEMORY_BACKUP)
651 		nvme_ctrlr_devctl_log(ctrlr, "critical",
652 		    "volatile memory backup device failed");
653 
654 	if (state & NVME_CRIT_WARN_ST_RESERVED_MASK)
655 		nvme_ctrlr_devctl_log(ctrlr, "critical",
656 		    "unknown critical warning(s): state = 0x%02x", state);
657 }
658 
659 static void
660 nvme_ctrlr_async_event_log_page_cb(void *arg, const struct nvme_completion *cpl)
661 {
662 	struct nvme_async_event_request		*aer = arg;
663 	struct nvme_health_information_page	*health_info;
664 	struct nvme_ns_list			*nsl;
665 	struct nvme_error_information_entry	*err;
666 	int i;
667 
668 	/*
669 	 * If the log page fetch for some reason completed with an error,
670 	 *  don't pass log page data to the consumers.  In practice, this case
671 	 *  should never happen.
672 	 */
673 	if (nvme_completion_is_error(cpl))
674 		nvme_notify_async_consumers(aer->ctrlr, &aer->cpl,
675 		    aer->log_page_id, NULL, 0);
676 	else {
677 		/* Convert data to host endian */
678 		switch (aer->log_page_id) {
679 		case NVME_LOG_ERROR:
680 			err = (struct nvme_error_information_entry *)aer->log_page_buffer;
681 			for (i = 0; i < (aer->ctrlr->cdata.elpe + 1); i++)
682 				nvme_error_information_entry_swapbytes(err++);
683 			break;
684 		case NVME_LOG_HEALTH_INFORMATION:
685 			nvme_health_information_page_swapbytes(
686 			    (struct nvme_health_information_page *)aer->log_page_buffer);
687 			break;
688 		case NVME_LOG_FIRMWARE_SLOT:
689 			nvme_firmware_page_swapbytes(
690 			    (struct nvme_firmware_page *)aer->log_page_buffer);
691 			break;
692 		case NVME_LOG_CHANGED_NAMESPACE:
693 			nvme_ns_list_swapbytes(
694 			    (struct nvme_ns_list *)aer->log_page_buffer);
695 			break;
696 		case NVME_LOG_COMMAND_EFFECT:
697 			nvme_command_effects_page_swapbytes(
698 			    (struct nvme_command_effects_page *)aer->log_page_buffer);
699 			break;
700 		case NVME_LOG_RES_NOTIFICATION:
701 			nvme_res_notification_page_swapbytes(
702 			    (struct nvme_res_notification_page *)aer->log_page_buffer);
703 			break;
704 		case NVME_LOG_SANITIZE_STATUS:
705 			nvme_sanitize_status_page_swapbytes(
706 			    (struct nvme_sanitize_status_page *)aer->log_page_buffer);
707 			break;
708 		case INTEL_LOG_TEMP_STATS:
709 			intel_log_temp_stats_swapbytes(
710 			    (struct intel_log_temp_stats *)aer->log_page_buffer);
711 			break;
712 		default:
713 			break;
714 		}
715 
716 		if (aer->log_page_id == NVME_LOG_HEALTH_INFORMATION) {
717 			health_info = (struct nvme_health_information_page *)
718 			    aer->log_page_buffer;
719 			nvme_ctrlr_log_critical_warnings(aer->ctrlr,
720 			    health_info->critical_warning);
721 			/*
722 			 * Critical warnings reported through the
723 			 *  SMART/health log page are persistent, so
724 			 *  clear the associated bits in the async event
725 			 *  config so that we do not receive repeated
726 			 *  notifications for the same event.
727 			 */
728 			aer->ctrlr->async_event_config &=
729 			    ~health_info->critical_warning;
730 			nvme_ctrlr_cmd_set_async_event_config(aer->ctrlr,
731 			    aer->ctrlr->async_event_config, NULL, NULL);
732 		} else if (aer->log_page_id == NVME_LOG_CHANGED_NAMESPACE &&
733 		    !nvme_use_nvd) {
734 			nsl = (struct nvme_ns_list *)aer->log_page_buffer;
735 			for (i = 0; i < nitems(nsl->ns) && nsl->ns[i] != 0; i++) {
736 				if (nsl->ns[i] > NVME_MAX_NAMESPACES)
737 					break;
738 				nvme_notify_ns(aer->ctrlr, nsl->ns[i]);
739 			}
740 		}
741 
742 		/*
743 		 * Pass the cpl data from the original async event completion,
744 		 *  not the log page fetch.
745 		 */
746 		nvme_notify_async_consumers(aer->ctrlr, &aer->cpl,
747 		    aer->log_page_id, aer->log_page_buffer, aer->log_page_size);
748 	}
749 
750 	/*
751 	 * Repost another asynchronous event request to replace the one
752 	 *  that just completed.
753 	 */
754 	nvme_ctrlr_construct_and_submit_aer(aer->ctrlr, aer);
755 }
756 
757 static void
758 nvme_ctrlr_async_event_cb(void *arg, const struct nvme_completion *cpl)
759 {
760 	struct nvme_async_event_request	*aer = arg;
761 
762 	if (nvme_completion_is_error(cpl)) {
763 		/*
764 		 *  Do not retry failed async event requests.  This avoids
765 		 *  infinite loops where a new async event request is submitted
766 		 *  to replace the one just failed, only to fail again and
767 		 *  perpetuate the loop.
768 		 */
769 		return;
770 	}
771 
772 	/* Associated log page is in bits 23:16 of completion entry dw0. */
773 	aer->log_page_id = (cpl->cdw0 & 0xFF0000) >> 16;
774 
775 	nvme_printf(aer->ctrlr, "async event occurred (type 0x%x, info 0x%02x,"
776 	    " page 0x%02x)\n", (cpl->cdw0 & 0x07), (cpl->cdw0 & 0xFF00) >> 8,
777 	    aer->log_page_id);
778 
779 	if (is_log_page_id_valid(aer->log_page_id)) {
780 		aer->log_page_size = nvme_ctrlr_get_log_page_size(aer->ctrlr,
781 		    aer->log_page_id);
782 		memcpy(&aer->cpl, cpl, sizeof(*cpl));
783 		nvme_ctrlr_cmd_get_log_page(aer->ctrlr, aer->log_page_id,
784 		    NVME_GLOBAL_NAMESPACE_TAG, aer->log_page_buffer,
785 		    aer->log_page_size, nvme_ctrlr_async_event_log_page_cb,
786 		    aer);
787 		/* Wait to notify consumers until after log page is fetched. */
788 	} else {
789 		nvme_notify_async_consumers(aer->ctrlr, cpl, aer->log_page_id,
790 		    NULL, 0);
791 
792 		/*
793 		 * Repost another asynchronous event request to replace the one
794 		 *  that just completed.
795 		 */
796 		nvme_ctrlr_construct_and_submit_aer(aer->ctrlr, aer);
797 	}
798 }
799 
800 static void
801 nvme_ctrlr_construct_and_submit_aer(struct nvme_controller *ctrlr,
802     struct nvme_async_event_request *aer)
803 {
804 	struct nvme_request *req;
805 
806 	aer->ctrlr = ctrlr;
807 	req = nvme_allocate_request_null(nvme_ctrlr_async_event_cb, aer);
808 	aer->req = req;
809 
810 	/*
811 	 * Disable timeout here, since asynchronous event requests should by
812 	 *  nature never be timed out.
813 	 */
814 	req->timeout = false;
815 	req->cmd.opc = NVME_OPC_ASYNC_EVENT_REQUEST;
816 	nvme_ctrlr_submit_admin_request(ctrlr, req);
817 }
818 
819 static void
820 nvme_ctrlr_configure_aer(struct nvme_controller *ctrlr)
821 {
822 	struct nvme_completion_poll_status	status;
823 	struct nvme_async_event_request		*aer;
824 	uint32_t				i;
825 
826 	ctrlr->async_event_config = NVME_CRIT_WARN_ST_AVAILABLE_SPARE |
827 	    NVME_CRIT_WARN_ST_DEVICE_RELIABILITY |
828 	    NVME_CRIT_WARN_ST_READ_ONLY |
829 	    NVME_CRIT_WARN_ST_VOLATILE_MEMORY_BACKUP;
830 	if (ctrlr->cdata.ver >= NVME_REV(1, 2))
831 		ctrlr->async_event_config |= NVME_ASYNC_EVENT_NS_ATTRIBUTE |
832 		    NVME_ASYNC_EVENT_FW_ACTIVATE;
833 
834 	status.done = 0;
835 	nvme_ctrlr_cmd_get_feature(ctrlr, NVME_FEAT_TEMPERATURE_THRESHOLD,
836 	    0, NULL, 0, nvme_completion_poll_cb, &status);
837 	nvme_completion_poll(&status);
838 	if (nvme_completion_is_error(&status.cpl) ||
839 	    (status.cpl.cdw0 & 0xFFFF) == 0xFFFF ||
840 	    (status.cpl.cdw0 & 0xFFFF) == 0x0000) {
841 		nvme_printf(ctrlr, "temperature threshold not supported\n");
842 	} else
843 		ctrlr->async_event_config |= NVME_CRIT_WARN_ST_TEMPERATURE;
844 
845 	nvme_ctrlr_cmd_set_async_event_config(ctrlr,
846 	    ctrlr->async_event_config, NULL, NULL);
847 
848 	/* aerl is a zero-based value, so we need to add 1 here. */
849 	ctrlr->num_aers = min(NVME_MAX_ASYNC_EVENTS, (ctrlr->cdata.aerl+1));
850 
851 	for (i = 0; i < ctrlr->num_aers; i++) {
852 		aer = &ctrlr->aer[i];
853 		nvme_ctrlr_construct_and_submit_aer(ctrlr, aer);
854 	}
855 }
856 
857 static void
858 nvme_ctrlr_configure_int_coalescing(struct nvme_controller *ctrlr)
859 {
860 
861 	ctrlr->int_coal_time = 0;
862 	TUNABLE_INT_FETCH("hw.nvme.int_coal_time",
863 	    &ctrlr->int_coal_time);
864 
865 	ctrlr->int_coal_threshold = 0;
866 	TUNABLE_INT_FETCH("hw.nvme.int_coal_threshold",
867 	    &ctrlr->int_coal_threshold);
868 
869 	nvme_ctrlr_cmd_set_interrupt_coalescing(ctrlr, ctrlr->int_coal_time,
870 	    ctrlr->int_coal_threshold, NULL, NULL);
871 }
872 
873 static void
874 nvme_ctrlr_hmb_free(struct nvme_controller *ctrlr)
875 {
876 	struct nvme_hmb_chunk *hmbc;
877 	int i;
878 
879 	if (ctrlr->hmb_desc_paddr) {
880 		bus_dmamap_unload(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_map);
881 		bus_dmamem_free(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_vaddr,
882 		    ctrlr->hmb_desc_map);
883 		ctrlr->hmb_desc_paddr = 0;
884 	}
885 	if (ctrlr->hmb_desc_tag) {
886 		bus_dma_tag_destroy(ctrlr->hmb_desc_tag);
887 		ctrlr->hmb_desc_tag = NULL;
888 	}
889 	for (i = 0; i < ctrlr->hmb_nchunks; i++) {
890 		hmbc = &ctrlr->hmb_chunks[i];
891 		bus_dmamap_unload(ctrlr->hmb_tag, hmbc->hmbc_map);
892 		bus_dmamem_free(ctrlr->hmb_tag, hmbc->hmbc_vaddr,
893 		    hmbc->hmbc_map);
894 	}
895 	ctrlr->hmb_nchunks = 0;
896 	if (ctrlr->hmb_tag) {
897 		bus_dma_tag_destroy(ctrlr->hmb_tag);
898 		ctrlr->hmb_tag = NULL;
899 	}
900 	if (ctrlr->hmb_chunks) {
901 		free(ctrlr->hmb_chunks, M_NVME);
902 		ctrlr->hmb_chunks = NULL;
903 	}
904 }
905 
906 static void
907 nvme_ctrlr_hmb_alloc(struct nvme_controller *ctrlr)
908 {
909 	struct nvme_hmb_chunk *hmbc;
910 	size_t pref, min, minc, size;
911 	int err, i;
912 	uint64_t max;
913 
914 	/* Limit HMB to 5% of RAM size per device by default. */
915 	max = (uint64_t)physmem * PAGE_SIZE / 20;
916 	TUNABLE_UINT64_FETCH("hw.nvme.hmb_max", &max);
917 
918 	/*
919 	 * Units of Host Memory Buffer in the Identify info are always in terms
920 	 * of 4k units.
921 	 */
922 	min = (long long unsigned)ctrlr->cdata.hmmin * NVME_HMB_UNITS;
923 	if (max == 0 || max < min)
924 		return;
925 	pref = MIN((long long unsigned)ctrlr->cdata.hmpre * NVME_HMB_UNITS, max);
926 	minc = MAX(ctrlr->cdata.hmminds * NVME_HMB_UNITS, ctrlr->page_size);
927 	if (min > 0 && ctrlr->cdata.hmmaxd > 0)
928 		minc = MAX(minc, min / ctrlr->cdata.hmmaxd);
929 	ctrlr->hmb_chunk = pref;
930 
931 again:
932 	/*
933 	 * However, the chunk sizes, number of chunks, and alignment of chunks
934 	 * are all based on the current MPS (ctrlr->page_size).
935 	 */
936 	ctrlr->hmb_chunk = roundup2(ctrlr->hmb_chunk, ctrlr->page_size);
937 	ctrlr->hmb_nchunks = howmany(pref, ctrlr->hmb_chunk);
938 	if (ctrlr->cdata.hmmaxd > 0 && ctrlr->hmb_nchunks > ctrlr->cdata.hmmaxd)
939 		ctrlr->hmb_nchunks = ctrlr->cdata.hmmaxd;
940 	ctrlr->hmb_chunks = malloc(sizeof(struct nvme_hmb_chunk) *
941 	    ctrlr->hmb_nchunks, M_NVME, M_WAITOK);
942 	err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev),
943 	    ctrlr->page_size, 0, BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
944 	    ctrlr->hmb_chunk, 1, ctrlr->hmb_chunk, 0, NULL, NULL, &ctrlr->hmb_tag);
945 	if (err != 0) {
946 		nvme_printf(ctrlr, "HMB tag create failed %d\n", err);
947 		nvme_ctrlr_hmb_free(ctrlr);
948 		return;
949 	}
950 
951 	for (i = 0; i < ctrlr->hmb_nchunks; i++) {
952 		hmbc = &ctrlr->hmb_chunks[i];
953 		if (bus_dmamem_alloc(ctrlr->hmb_tag,
954 		    (void **)&hmbc->hmbc_vaddr, BUS_DMA_NOWAIT,
955 		    &hmbc->hmbc_map)) {
956 			nvme_printf(ctrlr, "failed to alloc HMB\n");
957 			break;
958 		}
959 		if (bus_dmamap_load(ctrlr->hmb_tag, hmbc->hmbc_map,
960 		    hmbc->hmbc_vaddr, ctrlr->hmb_chunk, nvme_single_map,
961 		    &hmbc->hmbc_paddr, BUS_DMA_NOWAIT) != 0) {
962 			bus_dmamem_free(ctrlr->hmb_tag, hmbc->hmbc_vaddr,
963 			    hmbc->hmbc_map);
964 			nvme_printf(ctrlr, "failed to load HMB\n");
965 			break;
966 		}
967 		bus_dmamap_sync(ctrlr->hmb_tag, hmbc->hmbc_map,
968 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
969 	}
970 
971 	if (i < ctrlr->hmb_nchunks && i * ctrlr->hmb_chunk < min &&
972 	    ctrlr->hmb_chunk / 2 >= minc) {
973 		ctrlr->hmb_nchunks = i;
974 		nvme_ctrlr_hmb_free(ctrlr);
975 		ctrlr->hmb_chunk /= 2;
976 		goto again;
977 	}
978 	ctrlr->hmb_nchunks = i;
979 	if (ctrlr->hmb_nchunks * ctrlr->hmb_chunk < min) {
980 		nvme_ctrlr_hmb_free(ctrlr);
981 		return;
982 	}
983 
984 	size = sizeof(struct nvme_hmb_desc) * ctrlr->hmb_nchunks;
985 	err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev),
986 	    16, 0, BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
987 	    size, 1, size, 0, NULL, NULL, &ctrlr->hmb_desc_tag);
988 	if (err != 0) {
989 		nvme_printf(ctrlr, "HMB desc tag create failed %d\n", err);
990 		nvme_ctrlr_hmb_free(ctrlr);
991 		return;
992 	}
993 	if (bus_dmamem_alloc(ctrlr->hmb_desc_tag,
994 	    (void **)&ctrlr->hmb_desc_vaddr, BUS_DMA_WAITOK,
995 	    &ctrlr->hmb_desc_map)) {
996 		nvme_printf(ctrlr, "failed to alloc HMB desc\n");
997 		nvme_ctrlr_hmb_free(ctrlr);
998 		return;
999 	}
1000 	if (bus_dmamap_load(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_map,
1001 	    ctrlr->hmb_desc_vaddr, size, nvme_single_map,
1002 	    &ctrlr->hmb_desc_paddr, BUS_DMA_NOWAIT) != 0) {
1003 		bus_dmamem_free(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_vaddr,
1004 		    ctrlr->hmb_desc_map);
1005 		nvme_printf(ctrlr, "failed to load HMB desc\n");
1006 		nvme_ctrlr_hmb_free(ctrlr);
1007 		return;
1008 	}
1009 
1010 	for (i = 0; i < ctrlr->hmb_nchunks; i++) {
1011 		ctrlr->hmb_desc_vaddr[i].addr =
1012 		    htole64(ctrlr->hmb_chunks[i].hmbc_paddr);
1013 		ctrlr->hmb_desc_vaddr[i].size = htole32(ctrlr->hmb_chunk / ctrlr->page_size);
1014 	}
1015 	bus_dmamap_sync(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_map,
1016 	    BUS_DMASYNC_PREWRITE);
1017 
1018 	nvme_printf(ctrlr, "Allocated %lluMB host memory buffer\n",
1019 	    (long long unsigned)ctrlr->hmb_nchunks * ctrlr->hmb_chunk
1020 	    / 1024 / 1024);
1021 }
1022 
1023 static void
1024 nvme_ctrlr_hmb_enable(struct nvme_controller *ctrlr, bool enable, bool memret)
1025 {
1026 	struct nvme_completion_poll_status	status;
1027 	uint32_t cdw11;
1028 
1029 	cdw11 = 0;
1030 	if (enable)
1031 		cdw11 |= 1;
1032 	if (memret)
1033 		cdw11 |= 2;
1034 	status.done = 0;
1035 	nvme_ctrlr_cmd_set_feature(ctrlr, NVME_FEAT_HOST_MEMORY_BUFFER, cdw11,
1036 	    ctrlr->hmb_nchunks * ctrlr->hmb_chunk / ctrlr->page_size,
1037 	    ctrlr->hmb_desc_paddr, ctrlr->hmb_desc_paddr >> 32,
1038 	    ctrlr->hmb_nchunks, NULL, 0,
1039 	    nvme_completion_poll_cb, &status);
1040 	nvme_completion_poll(&status);
1041 	if (nvme_completion_is_error(&status.cpl))
1042 		nvme_printf(ctrlr, "nvme_ctrlr_hmb_enable failed!\n");
1043 }
1044 
1045 static void
1046 nvme_ctrlr_start(void *ctrlr_arg, bool resetting)
1047 {
1048 	struct nvme_controller *ctrlr = ctrlr_arg;
1049 	uint32_t old_num_io_queues;
1050 	int i;
1051 
1052 	TSENTER();
1053 
1054 	/*
1055 	 * Only reset adminq here when we are restarting the
1056 	 *  controller after a reset.  During initialization,
1057 	 *  we have already submitted admin commands to get
1058 	 *  the number of I/O queues supported, so cannot reset
1059 	 *  the adminq again here.
1060 	 */
1061 	if (resetting) {
1062 		nvme_qpair_reset(&ctrlr->adminq);
1063 		nvme_admin_qpair_enable(&ctrlr->adminq);
1064 	}
1065 
1066 	if (ctrlr->ioq != NULL) {
1067 		for (i = 0; i < ctrlr->num_io_queues; i++)
1068 			nvme_qpair_reset(&ctrlr->ioq[i]);
1069 	}
1070 
1071 	/*
1072 	 * If it was a reset on initialization command timeout, just
1073 	 * return here, letting initialization code fail gracefully.
1074 	 */
1075 	if (resetting && !ctrlr->is_initialized)
1076 		return;
1077 
1078 	if (resetting && nvme_ctrlr_identify(ctrlr) != 0) {
1079 		nvme_ctrlr_fail(ctrlr);
1080 		return;
1081 	}
1082 
1083 	/*
1084 	 * The number of qpairs are determined during controller initialization,
1085 	 *  including using NVMe SET_FEATURES/NUMBER_OF_QUEUES to determine the
1086 	 *  HW limit.  We call SET_FEATURES again here so that it gets called
1087 	 *  after any reset for controllers that depend on the driver to
1088 	 *  explicit specify how many queues it will use.  This value should
1089 	 *  never change between resets, so panic if somehow that does happen.
1090 	 */
1091 	if (resetting) {
1092 		old_num_io_queues = ctrlr->num_io_queues;
1093 		if (nvme_ctrlr_set_num_qpairs(ctrlr) != 0) {
1094 			nvme_ctrlr_fail(ctrlr);
1095 			return;
1096 		}
1097 
1098 		if (old_num_io_queues != ctrlr->num_io_queues) {
1099 			panic("num_io_queues changed from %u to %u",
1100 			      old_num_io_queues, ctrlr->num_io_queues);
1101 		}
1102 	}
1103 
1104 	if (ctrlr->cdata.hmpre > 0 && ctrlr->hmb_nchunks == 0) {
1105 		nvme_ctrlr_hmb_alloc(ctrlr);
1106 		if (ctrlr->hmb_nchunks > 0)
1107 			nvme_ctrlr_hmb_enable(ctrlr, true, false);
1108 	} else if (ctrlr->hmb_nchunks > 0)
1109 		nvme_ctrlr_hmb_enable(ctrlr, true, true);
1110 
1111 	if (nvme_ctrlr_create_qpairs(ctrlr) != 0) {
1112 		nvme_ctrlr_fail(ctrlr);
1113 		return;
1114 	}
1115 
1116 	if (nvme_ctrlr_construct_namespaces(ctrlr) != 0) {
1117 		nvme_ctrlr_fail(ctrlr);
1118 		return;
1119 	}
1120 
1121 	nvme_ctrlr_configure_aer(ctrlr);
1122 	nvme_ctrlr_configure_int_coalescing(ctrlr);
1123 
1124 	for (i = 0; i < ctrlr->num_io_queues; i++)
1125 		nvme_io_qpair_enable(&ctrlr->ioq[i]);
1126 	TSEXIT();
1127 }
1128 
1129 void
1130 nvme_ctrlr_start_config_hook(void *arg)
1131 {
1132 	struct nvme_controller *ctrlr = arg;
1133 
1134 	TSENTER();
1135 
1136 	if (nvme_ctrlr_hw_reset(ctrlr) != 0) {
1137 fail:
1138 		nvme_ctrlr_fail(ctrlr);
1139 		config_intrhook_disestablish(&ctrlr->config_hook);
1140 		return;
1141 	}
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 	TSEXIT();
1159 }
1160 
1161 static void
1162 nvme_ctrlr_reset_task(void *arg, int pending)
1163 {
1164 	struct nvme_controller	*ctrlr = arg;
1165 	int			status;
1166 
1167 	nvme_ctrlr_devctl_log(ctrlr, "RESET", "resetting controller");
1168 	status = nvme_ctrlr_hw_reset(ctrlr);
1169 	if (status == 0)
1170 		nvme_ctrlr_start(ctrlr, true);
1171 	else
1172 		nvme_ctrlr_fail(ctrlr);
1173 
1174 	atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
1175 }
1176 
1177 /*
1178  * Poll all the queues enabled on the device for completion.
1179  */
1180 void
1181 nvme_ctrlr_poll(struct nvme_controller *ctrlr)
1182 {
1183 	int i;
1184 
1185 	nvme_qpair_process_completions(&ctrlr->adminq);
1186 
1187 	for (i = 0; i < ctrlr->num_io_queues; i++)
1188 		if (ctrlr->ioq && ctrlr->ioq[i].cpl)
1189 			nvme_qpair_process_completions(&ctrlr->ioq[i]);
1190 }
1191 
1192 /*
1193  * Poll the single-vector interrupt case: num_io_queues will be 1 and
1194  * there's only a single vector. While we're polling, we mask further
1195  * interrupts in the controller.
1196  */
1197 void
1198 nvme_ctrlr_shared_handler(void *arg)
1199 {
1200 	struct nvme_controller *ctrlr = arg;
1201 
1202 	nvme_mmio_write_4(ctrlr, intms, 1);
1203 	nvme_ctrlr_poll(ctrlr);
1204 	nvme_mmio_write_4(ctrlr, intmc, 1);
1205 }
1206 
1207 static void
1208 nvme_pt_done(void *arg, const struct nvme_completion *cpl)
1209 {
1210 	struct nvme_pt_command *pt = arg;
1211 	struct mtx *mtx = pt->driver_lock;
1212 	uint16_t status;
1213 
1214 	bzero(&pt->cpl, sizeof(pt->cpl));
1215 	pt->cpl.cdw0 = cpl->cdw0;
1216 
1217 	status = cpl->status;
1218 	status &= ~NVME_STATUS_P_MASK;
1219 	pt->cpl.status = status;
1220 
1221 	mtx_lock(mtx);
1222 	pt->driver_lock = NULL;
1223 	wakeup(pt);
1224 	mtx_unlock(mtx);
1225 }
1226 
1227 int
1228 nvme_ctrlr_passthrough_cmd(struct nvme_controller *ctrlr,
1229     struct nvme_pt_command *pt, uint32_t nsid, int is_user_buffer,
1230     int is_admin_cmd)
1231 {
1232 	struct nvme_request	*req;
1233 	struct mtx		*mtx;
1234 	struct buf		*buf = NULL;
1235 	int			ret = 0;
1236 
1237 	if (pt->len > 0) {
1238 		if (pt->len > ctrlr->max_xfer_size) {
1239 			nvme_printf(ctrlr, "pt->len (%d) "
1240 			    "exceeds max_xfer_size (%d)\n", pt->len,
1241 			    ctrlr->max_xfer_size);
1242 			return EIO;
1243 		}
1244 		if (is_user_buffer) {
1245 			/*
1246 			 * Ensure the user buffer is wired for the duration of
1247 			 *  this pass-through command.
1248 			 */
1249 			PHOLD(curproc);
1250 			buf = uma_zalloc(pbuf_zone, M_WAITOK);
1251 			buf->b_iocmd = pt->is_read ? BIO_READ : BIO_WRITE;
1252 			if (vmapbuf(buf, pt->buf, pt->len, 1) < 0) {
1253 				ret = EFAULT;
1254 				goto err;
1255 			}
1256 			req = nvme_allocate_request_vaddr(buf->b_data, pt->len,
1257 			    nvme_pt_done, pt);
1258 		} else
1259 			req = nvme_allocate_request_vaddr(pt->buf, pt->len,
1260 			    nvme_pt_done, pt);
1261 	} else
1262 		req = nvme_allocate_request_null(nvme_pt_done, pt);
1263 
1264 	/* Assume user space already converted to little-endian */
1265 	req->cmd.opc = pt->cmd.opc;
1266 	req->cmd.fuse = pt->cmd.fuse;
1267 	req->cmd.rsvd2 = pt->cmd.rsvd2;
1268 	req->cmd.rsvd3 = pt->cmd.rsvd3;
1269 	req->cmd.cdw10 = pt->cmd.cdw10;
1270 	req->cmd.cdw11 = pt->cmd.cdw11;
1271 	req->cmd.cdw12 = pt->cmd.cdw12;
1272 	req->cmd.cdw13 = pt->cmd.cdw13;
1273 	req->cmd.cdw14 = pt->cmd.cdw14;
1274 	req->cmd.cdw15 = pt->cmd.cdw15;
1275 
1276 	req->cmd.nsid = htole32(nsid);
1277 
1278 	mtx = mtx_pool_find(mtxpool_sleep, pt);
1279 	pt->driver_lock = mtx;
1280 
1281 	if (is_admin_cmd)
1282 		nvme_ctrlr_submit_admin_request(ctrlr, req);
1283 	else
1284 		nvme_ctrlr_submit_io_request(ctrlr, req);
1285 
1286 	mtx_lock(mtx);
1287 	while (pt->driver_lock != NULL)
1288 		mtx_sleep(pt, mtx, PRIBIO, "nvme_pt", 0);
1289 	mtx_unlock(mtx);
1290 
1291 	if (buf != NULL) {
1292 		vunmapbuf(buf);
1293 err:
1294 		uma_zfree(pbuf_zone, buf);
1295 		PRELE(curproc);
1296 	}
1297 
1298 	return (ret);
1299 }
1300 
1301 static int
1302 nvme_ctrlr_ioctl(struct cdev *cdev, u_long cmd, caddr_t arg, int flag,
1303     struct thread *td)
1304 {
1305 	struct nvme_controller			*ctrlr;
1306 	struct nvme_pt_command			*pt;
1307 
1308 	ctrlr = cdev->si_drv1;
1309 
1310 	switch (cmd) {
1311 	case NVME_RESET_CONTROLLER:
1312 		nvme_ctrlr_reset(ctrlr);
1313 		break;
1314 	case NVME_PASSTHROUGH_CMD:
1315 		pt = (struct nvme_pt_command *)arg;
1316 		return (nvme_ctrlr_passthrough_cmd(ctrlr, pt, le32toh(pt->cmd.nsid),
1317 		    1 /* is_user_buffer */, 1 /* is_admin_cmd */));
1318 	case NVME_GET_NSID:
1319 	{
1320 		struct nvme_get_nsid *gnsid = (struct nvme_get_nsid *)arg;
1321 		strncpy(gnsid->cdev, device_get_nameunit(ctrlr->dev),
1322 		    sizeof(gnsid->cdev));
1323 		gnsid->cdev[sizeof(gnsid->cdev) - 1] = '\0';
1324 		gnsid->nsid = 0;
1325 		break;
1326 	}
1327 	case NVME_GET_MAX_XFER_SIZE:
1328 		*(uint64_t *)arg = ctrlr->max_xfer_size;
1329 		break;
1330 	default:
1331 		return (ENOTTY);
1332 	}
1333 
1334 	return (0);
1335 }
1336 
1337 static struct cdevsw nvme_ctrlr_cdevsw = {
1338 	.d_version =	D_VERSION,
1339 	.d_flags =	0,
1340 	.d_ioctl =	nvme_ctrlr_ioctl
1341 };
1342 
1343 int
1344 nvme_ctrlr_construct(struct nvme_controller *ctrlr, device_t dev)
1345 {
1346 	struct make_dev_args	md_args;
1347 	uint32_t	cap_lo;
1348 	uint32_t	cap_hi;
1349 	uint32_t	to, vs, pmrcap;
1350 	int		status, timeout_period;
1351 
1352 	ctrlr->dev = dev;
1353 
1354 	mtx_init(&ctrlr->lock, "nvme ctrlr lock", NULL, MTX_DEF);
1355 	if (bus_get_domain(dev, &ctrlr->domain) != 0)
1356 		ctrlr->domain = 0;
1357 
1358 	ctrlr->cap_lo = cap_lo = nvme_mmio_read_4(ctrlr, cap_lo);
1359 	if (bootverbose) {
1360 		device_printf(dev, "CapLo: 0x%08x: MQES %u%s%s%s%s, TO %u\n",
1361 		    cap_lo, NVME_CAP_LO_MQES(cap_lo),
1362 		    NVME_CAP_LO_CQR(cap_lo) ? ", CQR" : "",
1363 		    NVME_CAP_LO_AMS(cap_lo) ? ", AMS" : "",
1364 		    (NVME_CAP_LO_AMS(cap_lo) & 0x1) ? " WRRwUPC" : "",
1365 		    (NVME_CAP_LO_AMS(cap_lo) & 0x2) ? " VS" : "",
1366 		    NVME_CAP_LO_TO(cap_lo));
1367 	}
1368 	ctrlr->cap_hi = cap_hi = nvme_mmio_read_4(ctrlr, cap_hi);
1369 	if (bootverbose) {
1370 		device_printf(dev, "CapHi: 0x%08x: DSTRD %u%s, CSS %x%s, "
1371 		    "MPSMIN %u, MPSMAX %u%s%s\n", cap_hi,
1372 		    NVME_CAP_HI_DSTRD(cap_hi),
1373 		    NVME_CAP_HI_NSSRS(cap_hi) ? ", NSSRS" : "",
1374 		    NVME_CAP_HI_CSS(cap_hi),
1375 		    NVME_CAP_HI_BPS(cap_hi) ? ", BPS" : "",
1376 		    NVME_CAP_HI_MPSMIN(cap_hi),
1377 		    NVME_CAP_HI_MPSMAX(cap_hi),
1378 		    NVME_CAP_HI_PMRS(cap_hi) ? ", PMRS" : "",
1379 		    NVME_CAP_HI_CMBS(cap_hi) ? ", CMBS" : "");
1380 	}
1381 	if (bootverbose) {
1382 		vs = nvme_mmio_read_4(ctrlr, vs);
1383 		device_printf(dev, "Version: 0x%08x: %d.%d\n", vs,
1384 		    NVME_MAJOR(vs), NVME_MINOR(vs));
1385 	}
1386 	if (bootverbose && NVME_CAP_HI_PMRS(cap_hi)) {
1387 		pmrcap = nvme_mmio_read_4(ctrlr, pmrcap);
1388 		device_printf(dev, "PMRCap: 0x%08x: BIR %u%s%s, PMRTU %u, "
1389 		    "PMRWBM %x, PMRTO %u%s\n", pmrcap,
1390 		    NVME_PMRCAP_BIR(pmrcap),
1391 		    NVME_PMRCAP_RDS(pmrcap) ? ", RDS" : "",
1392 		    NVME_PMRCAP_WDS(pmrcap) ? ", WDS" : "",
1393 		    NVME_PMRCAP_PMRTU(pmrcap),
1394 		    NVME_PMRCAP_PMRWBM(pmrcap),
1395 		    NVME_PMRCAP_PMRTO(pmrcap),
1396 		    NVME_PMRCAP_CMSS(pmrcap) ? ", CMSS" : "");
1397 	}
1398 
1399 	ctrlr->dstrd = NVME_CAP_HI_DSTRD(cap_hi) + 2;
1400 
1401 	ctrlr->mps = NVME_CAP_HI_MPSMIN(cap_hi);
1402 	ctrlr->page_size = 1 << (NVME_MPS_SHIFT + ctrlr->mps);
1403 
1404 	/* Get ready timeout value from controller, in units of 500ms. */
1405 	to = NVME_CAP_LO_TO(cap_lo) + 1;
1406 	ctrlr->ready_timeout_in_ms = to * 500;
1407 
1408 	timeout_period = NVME_DEFAULT_TIMEOUT_PERIOD;
1409 	TUNABLE_INT_FETCH("hw.nvme.timeout_period", &timeout_period);
1410 	timeout_period = min(timeout_period, NVME_MAX_TIMEOUT_PERIOD);
1411 	timeout_period = max(timeout_period, NVME_MIN_TIMEOUT_PERIOD);
1412 	ctrlr->timeout_period = timeout_period;
1413 
1414 	nvme_retry_count = NVME_DEFAULT_RETRY_COUNT;
1415 	TUNABLE_INT_FETCH("hw.nvme.retry_count", &nvme_retry_count);
1416 
1417 	ctrlr->enable_aborts = 0;
1418 	TUNABLE_INT_FETCH("hw.nvme.enable_aborts", &ctrlr->enable_aborts);
1419 
1420 	/* Cap transfers by the maximum addressable by page-sized PRP (4KB pages -> 2MB). */
1421 	ctrlr->max_xfer_size = MIN(maxphys, (ctrlr->page_size / 8 * ctrlr->page_size));
1422 	if (nvme_ctrlr_construct_admin_qpair(ctrlr) != 0)
1423 		return (ENXIO);
1424 
1425 	/*
1426 	 * Create 2 threads for the taskqueue. The reset thread will block when
1427 	 * it detects that the controller has failed until all I/O has been
1428 	 * failed up the stack. The fail_req task needs to be able to run in
1429 	 * this case to finish the request failure for some cases.
1430 	 *
1431 	 * We could partially solve this race by draining the failed requeust
1432 	 * queue before proceding to free the sim, though nothing would stop
1433 	 * new I/O from coming in after we do that drain, but before we reach
1434 	 * cam_sim_free, so this big hammer is used instead.
1435 	 */
1436 	ctrlr->taskqueue = taskqueue_create("nvme_taskq", M_WAITOK,
1437 	    taskqueue_thread_enqueue, &ctrlr->taskqueue);
1438 	taskqueue_start_threads(&ctrlr->taskqueue, 2, PI_DISK, "nvme taskq");
1439 
1440 	ctrlr->is_resetting = 0;
1441 	ctrlr->is_initialized = 0;
1442 	ctrlr->notification_sent = 0;
1443 	TASK_INIT(&ctrlr->reset_task, 0, nvme_ctrlr_reset_task, ctrlr);
1444 	STAILQ_INIT(&ctrlr->fail_req);
1445 	ctrlr->is_failed = false;
1446 
1447 	make_dev_args_init(&md_args);
1448 	md_args.mda_devsw = &nvme_ctrlr_cdevsw;
1449 	md_args.mda_uid = UID_ROOT;
1450 	md_args.mda_gid = GID_WHEEL;
1451 	md_args.mda_mode = 0600;
1452 	md_args.mda_unit = device_get_unit(dev);
1453 	md_args.mda_si_drv1 = (void *)ctrlr;
1454 	status = make_dev_s(&md_args, &ctrlr->cdev, "nvme%d",
1455 	    device_get_unit(dev));
1456 	if (status != 0)
1457 		return (ENXIO);
1458 
1459 	return (0);
1460 }
1461 
1462 void
1463 nvme_ctrlr_destruct(struct nvme_controller *ctrlr, device_t dev)
1464 {
1465 	int	gone, i;
1466 
1467 	ctrlr->is_dying = true;
1468 
1469 	if (ctrlr->resource == NULL)
1470 		goto nores;
1471 	if (!mtx_initialized(&ctrlr->adminq.lock))
1472 		goto noadminq;
1473 
1474 	/*
1475 	 * Check whether it is a hot unplug or a clean driver detach.
1476 	 * If device is not there any more, skip any shutdown commands.
1477 	 */
1478 	gone = (nvme_mmio_read_4(ctrlr, csts) == NVME_GONE);
1479 	if (gone)
1480 		nvme_ctrlr_fail(ctrlr);
1481 	else
1482 		nvme_notify_fail_consumers(ctrlr);
1483 
1484 	for (i = 0; i < NVME_MAX_NAMESPACES; i++)
1485 		nvme_ns_destruct(&ctrlr->ns[i]);
1486 
1487 	if (ctrlr->cdev)
1488 		destroy_dev(ctrlr->cdev);
1489 
1490 	if (ctrlr->is_initialized) {
1491 		if (!gone) {
1492 			if (ctrlr->hmb_nchunks > 0)
1493 				nvme_ctrlr_hmb_enable(ctrlr, false, false);
1494 			nvme_ctrlr_delete_qpairs(ctrlr);
1495 		}
1496 		nvme_ctrlr_hmb_free(ctrlr);
1497 	}
1498 	if (ctrlr->ioq != NULL) {
1499 		for (i = 0; i < ctrlr->num_io_queues; i++)
1500 			nvme_io_qpair_destroy(&ctrlr->ioq[i]);
1501 		free(ctrlr->ioq, M_NVME);
1502 	}
1503 	nvme_admin_qpair_destroy(&ctrlr->adminq);
1504 
1505 	/*
1506 	 *  Notify the controller of a shutdown, even though this is due to
1507 	 *   a driver unload, not a system shutdown (this path is not invoked
1508 	 *   during shutdown).  This ensures the controller receives a
1509 	 *   shutdown notification in case the system is shutdown before
1510 	 *   reloading the driver.
1511 	 */
1512 	if (!gone)
1513 		nvme_ctrlr_shutdown(ctrlr);
1514 
1515 	if (!gone)
1516 		nvme_ctrlr_disable(ctrlr);
1517 
1518 noadminq:
1519 	if (ctrlr->taskqueue)
1520 		taskqueue_free(ctrlr->taskqueue);
1521 
1522 	if (ctrlr->tag)
1523 		bus_teardown_intr(ctrlr->dev, ctrlr->res, ctrlr->tag);
1524 
1525 	if (ctrlr->res)
1526 		bus_release_resource(ctrlr->dev, SYS_RES_IRQ,
1527 		    rman_get_rid(ctrlr->res), ctrlr->res);
1528 
1529 	if (ctrlr->bar4_resource != NULL) {
1530 		bus_release_resource(dev, SYS_RES_MEMORY,
1531 		    ctrlr->bar4_resource_id, ctrlr->bar4_resource);
1532 	}
1533 
1534 	bus_release_resource(dev, SYS_RES_MEMORY,
1535 	    ctrlr->resource_id, ctrlr->resource);
1536 
1537 nores:
1538 	mtx_destroy(&ctrlr->lock);
1539 }
1540 
1541 void
1542 nvme_ctrlr_shutdown(struct nvme_controller *ctrlr)
1543 {
1544 	uint32_t	cc;
1545 	uint32_t	csts;
1546 	int		timeout;
1547 
1548 	cc = nvme_mmio_read_4(ctrlr, cc);
1549 	cc &= ~(NVME_CC_REG_SHN_MASK << NVME_CC_REG_SHN_SHIFT);
1550 	cc |= NVME_SHN_NORMAL << NVME_CC_REG_SHN_SHIFT;
1551 	nvme_mmio_write_4(ctrlr, cc, cc);
1552 
1553 	timeout = ticks + (ctrlr->cdata.rtd3e == 0 ? 5 * hz :
1554 	    ((uint64_t)ctrlr->cdata.rtd3e * hz + 999999) / 1000000);
1555 	while (1) {
1556 		csts = nvme_mmio_read_4(ctrlr, csts);
1557 		if (csts == NVME_GONE)		/* Hot unplug. */
1558 			break;
1559 		if (NVME_CSTS_GET_SHST(csts) == NVME_SHST_COMPLETE)
1560 			break;
1561 		if (timeout - ticks < 0) {
1562 			nvme_printf(ctrlr, "shutdown timeout\n");
1563 			break;
1564 		}
1565 		pause("nvmeshut", 1);
1566 	}
1567 }
1568 
1569 void
1570 nvme_ctrlr_submit_admin_request(struct nvme_controller *ctrlr,
1571     struct nvme_request *req)
1572 {
1573 
1574 	nvme_qpair_submit_request(&ctrlr->adminq, req);
1575 }
1576 
1577 void
1578 nvme_ctrlr_submit_io_request(struct nvme_controller *ctrlr,
1579     struct nvme_request *req)
1580 {
1581 	struct nvme_qpair       *qpair;
1582 
1583 	qpair = &ctrlr->ioq[QP(ctrlr, curcpu)];
1584 	nvme_qpair_submit_request(qpair, req);
1585 }
1586 
1587 device_t
1588 nvme_ctrlr_get_device(struct nvme_controller *ctrlr)
1589 {
1590 
1591 	return (ctrlr->dev);
1592 }
1593 
1594 const struct nvme_controller_data *
1595 nvme_ctrlr_get_data(struct nvme_controller *ctrlr)
1596 {
1597 
1598 	return (&ctrlr->cdata);
1599 }
1600 
1601 int
1602 nvme_ctrlr_suspend(struct nvme_controller *ctrlr)
1603 {
1604 	int to = hz;
1605 
1606 	/*
1607 	 * Can't touch failed controllers, so it's already suspended.
1608 	 */
1609 	if (ctrlr->is_failed)
1610 		return (0);
1611 
1612 	/*
1613 	 * We don't want the reset taskqueue running, since it does similar
1614 	 * things, so prevent it from running after we start. Wait for any reset
1615 	 * that may have been started to complete. The reset process we follow
1616 	 * will ensure that any new I/O will queue and be given to the hardware
1617 	 * after we resume (though there should be none).
1618 	 */
1619 	while (atomic_cmpset_32(&ctrlr->is_resetting, 0, 1) == 0 && to-- > 0)
1620 		pause("nvmesusp", 1);
1621 	if (to <= 0) {
1622 		nvme_printf(ctrlr,
1623 		    "Competing reset task didn't finish. Try again later.\n");
1624 		return (EWOULDBLOCK);
1625 	}
1626 
1627 	if (ctrlr->hmb_nchunks > 0)
1628 		nvme_ctrlr_hmb_enable(ctrlr, false, false);
1629 
1630 	/*
1631 	 * Per Section 7.6.2 of NVMe spec 1.4, to properly suspend, we need to
1632 	 * delete the hardware I/O queues, and then shutdown. This properly
1633 	 * flushes any metadata the drive may have stored so it can survive
1634 	 * having its power removed and prevents the unsafe shutdown count from
1635 	 * incriminating. Once we delete the qpairs, we have to disable them
1636 	 * before shutting down.
1637 	 */
1638 	nvme_ctrlr_delete_qpairs(ctrlr);
1639 	nvme_ctrlr_disable_qpairs(ctrlr);
1640 	nvme_ctrlr_shutdown(ctrlr);
1641 
1642 	return (0);
1643 }
1644 
1645 int
1646 nvme_ctrlr_resume(struct nvme_controller *ctrlr)
1647 {
1648 
1649 	/*
1650 	 * Can't touch failed controllers, so nothing to do to resume.
1651 	 */
1652 	if (ctrlr->is_failed)
1653 		return (0);
1654 
1655 	if (nvme_ctrlr_hw_reset(ctrlr) != 0)
1656 		goto fail;
1657 
1658 	/*
1659 	 * Now that we've reset the hardware, we can restart the controller. Any
1660 	 * I/O that was pending is requeued. Any admin commands are aborted with
1661 	 * an error. Once we've restarted, take the controller out of reset.
1662 	 */
1663 	nvme_ctrlr_start(ctrlr, true);
1664 	(void)atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
1665 
1666 	return (0);
1667 fail:
1668 	/*
1669 	 * Since we can't bring the controller out of reset, announce and fail
1670 	 * the controller. However, we have to return success for the resume
1671 	 * itself, due to questionable APIs.
1672 	 */
1673 	nvme_printf(ctrlr, "Failed to reset on resume, failing.\n");
1674 	nvme_ctrlr_fail(ctrlr);
1675 	(void)atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
1676 	return (0);
1677 }
1678