xref: /freebsd/sys/dev/nvme/nvme_ctrlr.c (revision 8a802df1de2d77fd0a62996bd785ca3f1326887f)
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 "opt_nvme.h"
30 
31 #include <sys/param.h>
32 #include <sys/systm.h>
33 #include <sys/buf.h>
34 #include <sys/bus.h>
35 #include <sys/conf.h>
36 #include <sys/ioccom.h>
37 #include <sys/proc.h>
38 #include <sys/smp.h>
39 #include <sys/uio.h>
40 #include <sys/sbuf.h>
41 #include <sys/endian.h>
42 #include <machine/stdarg.h>
43 #include <vm/vm.h>
44 
45 #include "nvme_private.h"
46 
47 #define B4_CHK_RDY_DELAY_MS	2300		/* work around controller bug */
48 
49 static void nvme_ctrlr_construct_and_submit_aer(struct nvme_controller *ctrlr,
50 						struct nvme_async_event_request *aer);
51 
52 static void
53 nvme_ctrlr_barrier(struct nvme_controller *ctrlr, int flags)
54 {
55 	bus_barrier(ctrlr->resource, 0, rman_get_size(ctrlr->resource), flags);
56 }
57 
58 static void
59 nvme_ctrlr_devctl_va(struct nvme_controller *ctrlr, const char *type,
60     const char *msg, va_list ap)
61 {
62 	struct sbuf sb;
63 	int error;
64 
65 	if (sbuf_new(&sb, NULL, 0, SBUF_AUTOEXTEND | SBUF_NOWAIT) == NULL)
66 		return;
67 	sbuf_printf(&sb, "name=\"%s\" ", device_get_nameunit(ctrlr->dev));
68 	sbuf_vprintf(&sb, msg, ap);
69 	error = sbuf_finish(&sb);
70 	if (error == 0)
71 		devctl_notify("nvme", "controller", type, sbuf_data(&sb));
72 	sbuf_delete(&sb);
73 }
74 
75 static void
76 nvme_ctrlr_devctl(struct nvme_controller *ctrlr, const char *type, const char *msg, ...)
77 {
78 	va_list ap;
79 
80 	va_start(ap, msg);
81 	nvme_ctrlr_devctl_va(ctrlr, type, msg, ap);
82 	va_end(ap);
83 }
84 
85 static void
86 nvme_ctrlr_devctl_log(struct nvme_controller *ctrlr, const char *type, const char *msg, ...)
87 {
88 	struct sbuf sb;
89 	va_list ap;
90 	int error;
91 
92 	if (sbuf_new(&sb, NULL, 0, SBUF_AUTOEXTEND | SBUF_NOWAIT) == NULL)
93 		return;
94 	sbuf_printf(&sb, "%s: ", device_get_nameunit(ctrlr->dev));
95 	va_start(ap, msg);
96 	sbuf_vprintf(&sb, msg, ap);
97 	va_end(ap);
98 	error = sbuf_finish(&sb);
99 	if (error == 0)
100 		printf("%s\n", sbuf_data(&sb));
101 	sbuf_delete(&sb);
102 	va_start(ap, msg);
103 	nvme_ctrlr_devctl_va(ctrlr, type, msg, ap);
104 	va_end(ap);
105 }
106 
107 static int
108 nvme_ctrlr_construct_admin_qpair(struct nvme_controller *ctrlr)
109 {
110 	struct nvme_qpair	*qpair;
111 	uint32_t		num_entries;
112 	int			error;
113 
114 	qpair = &ctrlr->adminq;
115 	qpair->id = 0;
116 	qpair->cpu = CPU_FFS(&cpuset_domain[ctrlr->domain]) - 1;
117 	qpair->domain = ctrlr->domain;
118 
119 	num_entries = NVME_ADMIN_ENTRIES;
120 	TUNABLE_INT_FETCH("hw.nvme.admin_entries", &num_entries);
121 	/*
122 	 * If admin_entries was overridden to an invalid value, revert it
123 	 *  back to our default value.
124 	 */
125 	if (num_entries < NVME_MIN_ADMIN_ENTRIES ||
126 	    num_entries > NVME_MAX_ADMIN_ENTRIES) {
127 		nvme_printf(ctrlr, "invalid hw.nvme.admin_entries=%d "
128 		    "specified\n", num_entries);
129 		num_entries = NVME_ADMIN_ENTRIES;
130 	}
131 
132 	/*
133 	 * The admin queue's max xfer size is treated differently than the
134 	 *  max I/O xfer size.  16KB is sufficient here - maybe even less?
135 	 */
136 	error = nvme_qpair_construct(qpair, num_entries, NVME_ADMIN_TRACKERS,
137 	     ctrlr);
138 	return (error);
139 }
140 
141 #define QP(ctrlr, c)	((c) * (ctrlr)->num_io_queues / mp_ncpus)
142 
143 static int
144 nvme_ctrlr_construct_io_qpairs(struct nvme_controller *ctrlr)
145 {
146 	struct nvme_qpair	*qpair;
147 	uint32_t		cap_lo;
148 	uint16_t		mqes;
149 	int			c, error, i, n;
150 	int			num_entries, num_trackers, max_entries;
151 
152 	/*
153 	 * NVMe spec sets a hard limit of 64K max entries, but devices may
154 	 * specify a smaller limit, so we need to check the MQES field in the
155 	 * capabilities register. We have to cap the number of entries to the
156 	 * current stride allows for in BAR 0/1, otherwise the remainder entries
157 	 * are inaccessible. MQES should reflect this, and this is just a
158 	 * fail-safe.
159 	 */
160 	max_entries =
161 	    (rman_get_size(ctrlr->resource) - nvme_mmio_offsetof(doorbell[0])) /
162 	    (1 << (ctrlr->dstrd + 1));
163 	num_entries = NVME_IO_ENTRIES;
164 	TUNABLE_INT_FETCH("hw.nvme.io_entries", &num_entries);
165 	cap_lo = nvme_mmio_read_4(ctrlr, cap_lo);
166 	mqes = NVME_CAP_LO_MQES(cap_lo);
167 	num_entries = min(num_entries, mqes + 1);
168 	num_entries = min(num_entries, max_entries);
169 
170 	num_trackers = NVME_IO_TRACKERS;
171 	TUNABLE_INT_FETCH("hw.nvme.io_trackers", &num_trackers);
172 
173 	num_trackers = max(num_trackers, NVME_MIN_IO_TRACKERS);
174 	num_trackers = min(num_trackers, NVME_MAX_IO_TRACKERS);
175 	/*
176 	 * No need to have more trackers than entries in the submit queue.  Note
177 	 * also that for a queue size of N, we can only have (N-1) commands
178 	 * outstanding, hence the "-1" here.
179 	 */
180 	num_trackers = min(num_trackers, (num_entries-1));
181 
182 	/*
183 	 * Our best estimate for the maximum number of I/Os that we should
184 	 * normally have in flight at one time. This should be viewed as a hint,
185 	 * not a hard limit and will need to be revisited when the upper layers
186 	 * of the storage system grows multi-queue support.
187 	 */
188 	ctrlr->max_hw_pend_io = num_trackers * ctrlr->num_io_queues * 3 / 4;
189 
190 	ctrlr->ioq = malloc(ctrlr->num_io_queues * sizeof(struct nvme_qpair),
191 	    M_NVME, M_ZERO | M_WAITOK);
192 
193 	for (i = c = n = 0; i < ctrlr->num_io_queues; i++, c += n) {
194 		qpair = &ctrlr->ioq[i];
195 
196 		/*
197 		 * Admin queue has ID=0. IO queues start at ID=1 -
198 		 *  hence the 'i+1' here.
199 		 */
200 		qpair->id = i + 1;
201 		if (ctrlr->num_io_queues > 1) {
202 			/* Find number of CPUs served by this queue. */
203 			for (n = 1; QP(ctrlr, c + n) == i; n++)
204 				;
205 			/* Shuffle multiple NVMe devices between CPUs. */
206 			qpair->cpu = c + (device_get_unit(ctrlr->dev)+n/2) % n;
207 			qpair->domain = pcpu_find(qpair->cpu)->pc_domain;
208 		} else {
209 			qpair->cpu = CPU_FFS(&cpuset_domain[ctrlr->domain]) - 1;
210 			qpair->domain = ctrlr->domain;
211 		}
212 
213 		/*
214 		 * For I/O queues, use the controller-wide max_xfer_size
215 		 *  calculated in nvme_attach().
216 		 */
217 		error = nvme_qpair_construct(qpair, num_entries, num_trackers,
218 		    ctrlr);
219 		if (error)
220 			return (error);
221 
222 		/*
223 		 * Do not bother binding interrupts if we only have one I/O
224 		 *  interrupt thread for this controller.
225 		 */
226 		if (ctrlr->num_io_queues > 1)
227 			bus_bind_intr(ctrlr->dev, qpair->res, qpair->cpu);
228 	}
229 
230 	return (0);
231 }
232 
233 static void
234 nvme_ctrlr_fail(struct nvme_controller *ctrlr)
235 {
236 	int i;
237 
238 	/*
239 	 * No need to disable queues before failing them. Failing is a superet
240 	 * of disabling (though pedantically we'd abort the AERs silently with
241 	 * a different error, though when we fail, that hardly matters).
242 	 */
243 	ctrlr->is_failed = true;
244 	nvme_qpair_fail(&ctrlr->adminq);
245 	if (ctrlr->ioq != NULL) {
246 		for (i = 0; i < ctrlr->num_io_queues; i++) {
247 			nvme_qpair_fail(&ctrlr->ioq[i]);
248 		}
249 	}
250 	nvme_notify_fail_consumers(ctrlr);
251 }
252 
253 /*
254  * Wait for RDY to change.
255  *
256  * Starts sleeping for 1us and geometrically increases it the longer we wait,
257  * capped at 1ms.
258  */
259 static int
260 nvme_ctrlr_wait_for_ready(struct nvme_controller *ctrlr, int desired_val)
261 {
262 	int timeout = ticks + MSEC_2_TICKS(ctrlr->ready_timeout_in_ms);
263 	sbintime_t delta_t = SBT_1US;
264 	uint32_t csts;
265 
266 	while (1) {
267 		csts = nvme_mmio_read_4(ctrlr, csts);
268 		if (csts == NVME_GONE)		/* Hot unplug. */
269 			return (ENXIO);
270 		if (NVMEV(NVME_CSTS_REG_RDY, csts) == desired_val)
271 			break;
272 		if (timeout - ticks < 0) {
273 			nvme_printf(ctrlr, "controller ready did not become %d "
274 			    "within %d ms\n", desired_val, ctrlr->ready_timeout_in_ms);
275 			return (ENXIO);
276 		}
277 
278 		pause_sbt("nvmerdy", delta_t, 0, C_PREL(1));
279 		delta_t = min(SBT_1MS, delta_t * 3 / 2);
280 	}
281 
282 	return (0);
283 }
284 
285 static int
286 nvme_ctrlr_disable(struct nvme_controller *ctrlr)
287 {
288 	uint32_t cc;
289 	uint32_t csts;
290 	uint8_t  en, rdy;
291 	int err;
292 
293 	cc = nvme_mmio_read_4(ctrlr, cc);
294 	csts = nvme_mmio_read_4(ctrlr, csts);
295 
296 	en = NVMEV(NVME_CC_REG_EN, cc);
297 	rdy = NVMEV(NVME_CSTS_REG_RDY, csts);
298 
299 	/*
300 	 * Per 3.1.5 in NVME 1.3 spec, transitioning CC.EN from 0 to 1
301 	 * when CSTS.RDY is 1 or transitioning CC.EN from 1 to 0 when
302 	 * CSTS.RDY is 0 "has undefined results" So make sure that CSTS.RDY
303 	 * isn't the desired value. Short circuit if we're already disabled.
304 	 */
305 	if (en == 0) {
306 		/* Wait for RDY == 0 or timeout & fail */
307 		if (rdy == 0)
308 			return (0);
309 		return (nvme_ctrlr_wait_for_ready(ctrlr, 0));
310 	}
311 	if (rdy == 0) {
312 		/* EN == 1, wait for  RDY == 1 or timeout & fail */
313 		err = nvme_ctrlr_wait_for_ready(ctrlr, 1);
314 		if (err != 0)
315 			return (err);
316 	}
317 
318 	cc &= ~NVMEM(NVME_CC_REG_EN);
319 	nvme_mmio_write_4(ctrlr, cc, cc);
320 
321 	/*
322 	 * A few drives have firmware bugs that freeze the drive if we access
323 	 * the mmio too soon after we disable.
324 	 */
325 	if (ctrlr->quirks & QUIRK_DELAY_B4_CHK_RDY)
326 		pause("nvmeR", MSEC_2_TICKS(B4_CHK_RDY_DELAY_MS));
327 	return (nvme_ctrlr_wait_for_ready(ctrlr, 0));
328 }
329 
330 static int
331 nvme_ctrlr_enable(struct nvme_controller *ctrlr)
332 {
333 	uint32_t	cc;
334 	uint32_t	csts;
335 	uint32_t	aqa;
336 	uint32_t	qsize;
337 	uint8_t		en, rdy;
338 	int		err;
339 
340 	cc = nvme_mmio_read_4(ctrlr, cc);
341 	csts = nvme_mmio_read_4(ctrlr, csts);
342 
343 	en = NVMEV(NVME_CC_REG_EN, cc);
344 	rdy = NVMEV(NVME_CSTS_REG_RDY, csts);
345 
346 	/*
347 	 * See note in nvme_ctrlr_disable. Short circuit if we're already enabled.
348 	 */
349 	if (en == 1) {
350 		if (rdy == 1)
351 			return (0);
352 		return (nvme_ctrlr_wait_for_ready(ctrlr, 1));
353 	}
354 
355 	/* EN == 0 already wait for RDY == 0 or timeout & fail */
356 	err = nvme_ctrlr_wait_for_ready(ctrlr, 0);
357 	if (err != 0)
358 		return (err);
359 
360 	nvme_mmio_write_8(ctrlr, asq, ctrlr->adminq.cmd_bus_addr);
361 	nvme_mmio_write_8(ctrlr, acq, ctrlr->adminq.cpl_bus_addr);
362 
363 	/* acqs and asqs are 0-based. */
364 	qsize = ctrlr->adminq.num_entries - 1;
365 
366 	aqa = 0;
367 	aqa |= NVMEF(NVME_AQA_REG_ACQS, qsize);
368 	aqa |= NVMEF(NVME_AQA_REG_ASQS, qsize);
369 	nvme_mmio_write_4(ctrlr, aqa, aqa);
370 
371 	/* Initialization values for CC */
372 	cc = 0;
373 	cc |= NVMEF(NVME_CC_REG_EN, 1);
374 	cc |= NVMEF(NVME_CC_REG_CSS, 0);
375 	cc |= NVMEF(NVME_CC_REG_AMS, 0);
376 	cc |= NVMEF(NVME_CC_REG_SHN, 0);
377 	cc |= NVMEF(NVME_CC_REG_IOSQES, 6); /* SQ entry size == 64 == 2^6 */
378 	cc |= NVMEF(NVME_CC_REG_IOCQES, 4); /* CQ entry size == 16 == 2^4 */
379 
380 	/*
381 	 * Use the Memory Page Size selected during device initialization.  Note
382 	 * that value stored in mps is suitable to use here without adjusting by
383 	 * NVME_MPS_SHIFT.
384 	 */
385 	cc |= NVMEF(NVME_CC_REG_MPS, ctrlr->mps);
386 
387 	nvme_ctrlr_barrier(ctrlr, BUS_SPACE_BARRIER_WRITE);
388 	nvme_mmio_write_4(ctrlr, cc, cc);
389 
390 	return (nvme_ctrlr_wait_for_ready(ctrlr, 1));
391 }
392 
393 static void
394 nvme_ctrlr_disable_qpairs(struct nvme_controller *ctrlr)
395 {
396 	int i;
397 
398 	nvme_admin_qpair_disable(&ctrlr->adminq);
399 	/*
400 	 * I/O queues are not allocated before the initial HW
401 	 *  reset, so do not try to disable them.  Use is_initialized
402 	 *  to determine if this is the initial HW reset.
403 	 */
404 	if (ctrlr->is_initialized) {
405 		for (i = 0; i < ctrlr->num_io_queues; i++)
406 			nvme_io_qpair_disable(&ctrlr->ioq[i]);
407 	}
408 }
409 
410 static int
411 nvme_ctrlr_hw_reset(struct nvme_controller *ctrlr)
412 {
413 	int err;
414 
415 	TSENTER();
416 
417 	nvme_ctrlr_disable_qpairs(ctrlr);
418 
419 	err = nvme_ctrlr_disable(ctrlr);
420 	if (err != 0)
421 		goto out;
422 
423 	err = nvme_ctrlr_enable(ctrlr);
424 out:
425 
426 	TSEXIT();
427 	return (err);
428 }
429 
430 void
431 nvme_ctrlr_reset(struct nvme_controller *ctrlr)
432 {
433 	int cmpset;
434 
435 	cmpset = atomic_cmpset_32(&ctrlr->is_resetting, 0, 1);
436 
437 	if (cmpset == 0 || ctrlr->is_failed)
438 		/*
439 		 * Controller is already resetting or has failed.  Return
440 		 *  immediately since there is no need to kick off another
441 		 *  reset in these cases.
442 		 */
443 		return;
444 
445 	if (!ctrlr->is_dying)
446 		taskqueue_enqueue(ctrlr->taskqueue, &ctrlr->reset_task);
447 }
448 
449 static int
450 nvme_ctrlr_identify(struct nvme_controller *ctrlr)
451 {
452 	struct nvme_completion_poll_status	status;
453 
454 	status.done = 0;
455 	nvme_ctrlr_cmd_identify_controller(ctrlr, &ctrlr->cdata,
456 	    nvme_completion_poll_cb, &status);
457 	nvme_completion_poll(&status);
458 	if (nvme_completion_is_error(&status.cpl)) {
459 		nvme_printf(ctrlr, "nvme_identify_controller failed!\n");
460 		return (ENXIO);
461 	}
462 
463 	/* Convert data to host endian */
464 	nvme_controller_data_swapbytes(&ctrlr->cdata);
465 
466 	/*
467 	 * Use MDTS to ensure our default max_xfer_size doesn't exceed what the
468 	 *  controller supports.
469 	 */
470 	if (ctrlr->cdata.mdts > 0)
471 		ctrlr->max_xfer_size = min(ctrlr->max_xfer_size,
472 		    1 << (ctrlr->cdata.mdts + NVME_MPS_SHIFT +
473 			NVME_CAP_HI_MPSMIN(ctrlr->cap_hi)));
474 
475 	return (0);
476 }
477 
478 static int
479 nvme_ctrlr_set_num_qpairs(struct nvme_controller *ctrlr)
480 {
481 	struct nvme_completion_poll_status	status;
482 	int					cq_allocated, sq_allocated;
483 
484 	status.done = 0;
485 	nvme_ctrlr_cmd_set_num_queues(ctrlr, ctrlr->num_io_queues,
486 	    nvme_completion_poll_cb, &status);
487 	nvme_completion_poll(&status);
488 	if (nvme_completion_is_error(&status.cpl)) {
489 		nvme_printf(ctrlr, "nvme_ctrlr_set_num_qpairs failed!\n");
490 		return (ENXIO);
491 	}
492 
493 	/*
494 	 * Data in cdw0 is 0-based.
495 	 * Lower 16-bits indicate number of submission queues allocated.
496 	 * Upper 16-bits indicate number of completion queues allocated.
497 	 */
498 	sq_allocated = (status.cpl.cdw0 & 0xFFFF) + 1;
499 	cq_allocated = (status.cpl.cdw0 >> 16) + 1;
500 
501 	/*
502 	 * Controller may allocate more queues than we requested,
503 	 *  so use the minimum of the number requested and what was
504 	 *  actually allocated.
505 	 */
506 	ctrlr->num_io_queues = min(ctrlr->num_io_queues, sq_allocated);
507 	ctrlr->num_io_queues = min(ctrlr->num_io_queues, cq_allocated);
508 	if (ctrlr->num_io_queues > vm_ndomains)
509 		ctrlr->num_io_queues -= ctrlr->num_io_queues % vm_ndomains;
510 
511 	return (0);
512 }
513 
514 static int
515 nvme_ctrlr_create_qpairs(struct nvme_controller *ctrlr)
516 {
517 	struct nvme_completion_poll_status	status;
518 	struct nvme_qpair			*qpair;
519 	int					i;
520 
521 	for (i = 0; i < ctrlr->num_io_queues; i++) {
522 		qpair = &ctrlr->ioq[i];
523 
524 		status.done = 0;
525 		nvme_ctrlr_cmd_create_io_cq(ctrlr, qpair,
526 		    nvme_completion_poll_cb, &status);
527 		nvme_completion_poll(&status);
528 		if (nvme_completion_is_error(&status.cpl)) {
529 			nvme_printf(ctrlr, "nvme_create_io_cq failed!\n");
530 			return (ENXIO);
531 		}
532 
533 		status.done = 0;
534 		nvme_ctrlr_cmd_create_io_sq(ctrlr, qpair,
535 		    nvme_completion_poll_cb, &status);
536 		nvme_completion_poll(&status);
537 		if (nvme_completion_is_error(&status.cpl)) {
538 			nvme_printf(ctrlr, "nvme_create_io_sq failed!\n");
539 			return (ENXIO);
540 		}
541 	}
542 
543 	return (0);
544 }
545 
546 static int
547 nvme_ctrlr_delete_qpairs(struct nvme_controller *ctrlr)
548 {
549 	struct nvme_completion_poll_status	status;
550 	struct nvme_qpair			*qpair;
551 
552 	for (int i = 0; i < ctrlr->num_io_queues; i++) {
553 		qpair = &ctrlr->ioq[i];
554 
555 		status.done = 0;
556 		nvme_ctrlr_cmd_delete_io_sq(ctrlr, qpair,
557 		    nvme_completion_poll_cb, &status);
558 		nvme_completion_poll(&status);
559 		if (nvme_completion_is_error(&status.cpl)) {
560 			nvme_printf(ctrlr, "nvme_destroy_io_sq failed!\n");
561 			return (ENXIO);
562 		}
563 
564 		status.done = 0;
565 		nvme_ctrlr_cmd_delete_io_cq(ctrlr, qpair,
566 		    nvme_completion_poll_cb, &status);
567 		nvme_completion_poll(&status);
568 		if (nvme_completion_is_error(&status.cpl)) {
569 			nvme_printf(ctrlr, "nvme_destroy_io_cq failed!\n");
570 			return (ENXIO);
571 		}
572 	}
573 
574 	return (0);
575 }
576 
577 static int
578 nvme_ctrlr_construct_namespaces(struct nvme_controller *ctrlr)
579 {
580 	struct nvme_namespace	*ns;
581 	uint32_t 		i;
582 
583 	for (i = 0; i < min(ctrlr->cdata.nn, NVME_MAX_NAMESPACES); i++) {
584 		ns = &ctrlr->ns[i];
585 		nvme_ns_construct(ns, i+1, ctrlr);
586 	}
587 
588 	return (0);
589 }
590 
591 static bool
592 is_log_page_id_valid(uint8_t page_id)
593 {
594 
595 	switch (page_id) {
596 	case NVME_LOG_ERROR:
597 	case NVME_LOG_HEALTH_INFORMATION:
598 	case NVME_LOG_FIRMWARE_SLOT:
599 	case NVME_LOG_CHANGED_NAMESPACE:
600 	case NVME_LOG_COMMAND_EFFECT:
601 	case NVME_LOG_RES_NOTIFICATION:
602 	case NVME_LOG_SANITIZE_STATUS:
603 		return (true);
604 	}
605 
606 	return (false);
607 }
608 
609 static uint32_t
610 nvme_ctrlr_get_log_page_size(struct nvme_controller *ctrlr, uint8_t page_id)
611 {
612 	uint32_t	log_page_size;
613 
614 	switch (page_id) {
615 	case NVME_LOG_ERROR:
616 		log_page_size = min(
617 		    sizeof(struct nvme_error_information_entry) *
618 		    (ctrlr->cdata.elpe + 1), NVME_MAX_AER_LOG_SIZE);
619 		break;
620 	case NVME_LOG_HEALTH_INFORMATION:
621 		log_page_size = sizeof(struct nvme_health_information_page);
622 		break;
623 	case NVME_LOG_FIRMWARE_SLOT:
624 		log_page_size = sizeof(struct nvme_firmware_page);
625 		break;
626 	case NVME_LOG_CHANGED_NAMESPACE:
627 		log_page_size = sizeof(struct nvme_ns_list);
628 		break;
629 	case NVME_LOG_COMMAND_EFFECT:
630 		log_page_size = sizeof(struct nvme_command_effects_page);
631 		break;
632 	case NVME_LOG_RES_NOTIFICATION:
633 		log_page_size = sizeof(struct nvme_res_notification_page);
634 		break;
635 	case NVME_LOG_SANITIZE_STATUS:
636 		log_page_size = sizeof(struct nvme_sanitize_status_page);
637 		break;
638 	default:
639 		log_page_size = 0;
640 		break;
641 	}
642 
643 	return (log_page_size);
644 }
645 
646 static void
647 nvme_ctrlr_log_critical_warnings(struct nvme_controller *ctrlr,
648     uint8_t state)
649 {
650 
651 	if (state & NVME_CRIT_WARN_ST_AVAILABLE_SPARE)
652 		nvme_printf(ctrlr, "SMART WARNING: available spare space below threshold\n");
653 
654 	if (state & NVME_CRIT_WARN_ST_TEMPERATURE)
655 		nvme_printf(ctrlr, "SMART WARNING: temperature above threshold\n");
656 
657 	if (state & NVME_CRIT_WARN_ST_DEVICE_RELIABILITY)
658 		nvme_printf(ctrlr, "SMART WARNING: device reliability degraded\n");
659 
660 	if (state & NVME_CRIT_WARN_ST_READ_ONLY)
661 		nvme_printf(ctrlr, "SMART WARNING: media placed in read only mode\n");
662 
663 	if (state & NVME_CRIT_WARN_ST_VOLATILE_MEMORY_BACKUP)
664 		nvme_printf(ctrlr, "SMART WARNING: volatile memory backup device failed\n");
665 
666 	if (state & NVME_CRIT_WARN_ST_PERSISTENT_MEMORY_REGION)
667 		nvme_printf(ctrlr, "SMART WARNING: persistent memory read only or unreliable\n");
668 
669 	if (state & NVME_CRIT_WARN_ST_RESERVED_MASK)
670 		nvme_printf(ctrlr, "SMART WARNING: unknown critical warning(s): state = 0x%02x\n",
671 		    state & NVME_CRIT_WARN_ST_RESERVED_MASK);
672 
673 	nvme_ctrlr_devctl(ctrlr, "critical", "SMART_ERROR", "state=0x%02x", state);
674 }
675 
676 static void
677 nvme_ctrlr_async_event_log_page_cb(void *arg, const struct nvme_completion *cpl)
678 {
679 	struct nvme_async_event_request		*aer = arg;
680 	struct nvme_health_information_page	*health_info;
681 	struct nvme_ns_list			*nsl;
682 	struct nvme_error_information_entry	*err;
683 	int i;
684 
685 	/*
686 	 * If the log page fetch for some reason completed with an error,
687 	 *  don't pass log page data to the consumers.  In practice, this case
688 	 *  should never happen.
689 	 */
690 	if (nvme_completion_is_error(cpl))
691 		nvme_notify_async_consumers(aer->ctrlr, &aer->cpl,
692 		    aer->log_page_id, NULL, 0);
693 	else {
694 		/* Convert data to host endian */
695 		switch (aer->log_page_id) {
696 		case NVME_LOG_ERROR:
697 			err = (struct nvme_error_information_entry *)aer->log_page_buffer;
698 			for (i = 0; i < (aer->ctrlr->cdata.elpe + 1); i++)
699 				nvme_error_information_entry_swapbytes(err++);
700 			break;
701 		case NVME_LOG_HEALTH_INFORMATION:
702 			nvme_health_information_page_swapbytes(
703 			    (struct nvme_health_information_page *)aer->log_page_buffer);
704 			break;
705 		case NVME_LOG_CHANGED_NAMESPACE:
706 			nvme_ns_list_swapbytes(
707 			    (struct nvme_ns_list *)aer->log_page_buffer);
708 			break;
709 		case NVME_LOG_COMMAND_EFFECT:
710 			nvme_command_effects_page_swapbytes(
711 			    (struct nvme_command_effects_page *)aer->log_page_buffer);
712 			break;
713 		case NVME_LOG_RES_NOTIFICATION:
714 			nvme_res_notification_page_swapbytes(
715 			    (struct nvme_res_notification_page *)aer->log_page_buffer);
716 			break;
717 		case NVME_LOG_SANITIZE_STATUS:
718 			nvme_sanitize_status_page_swapbytes(
719 			    (struct nvme_sanitize_status_page *)aer->log_page_buffer);
720 			break;
721 		case INTEL_LOG_TEMP_STATS:
722 			intel_log_temp_stats_swapbytes(
723 			    (struct intel_log_temp_stats *)aer->log_page_buffer);
724 			break;
725 		default:
726 			break;
727 		}
728 
729 		if (aer->log_page_id == NVME_LOG_HEALTH_INFORMATION) {
730 			health_info = (struct nvme_health_information_page *)
731 			    aer->log_page_buffer;
732 			nvme_ctrlr_log_critical_warnings(aer->ctrlr,
733 			    health_info->critical_warning);
734 			/*
735 			 * Critical warnings reported through the
736 			 *  SMART/health log page are persistent, so
737 			 *  clear the associated bits in the async event
738 			 *  config so that we do not receive repeated
739 			 *  notifications for the same event.
740 			 */
741 			aer->ctrlr->async_event_config &=
742 			    ~health_info->critical_warning;
743 			nvme_ctrlr_cmd_set_async_event_config(aer->ctrlr,
744 			    aer->ctrlr->async_event_config, NULL, NULL);
745 		} else if (aer->log_page_id == NVME_LOG_CHANGED_NAMESPACE &&
746 		    !nvme_use_nvd) {
747 			nsl = (struct nvme_ns_list *)aer->log_page_buffer;
748 			for (i = 0; i < nitems(nsl->ns) && nsl->ns[i] != 0; i++) {
749 				if (nsl->ns[i] > NVME_MAX_NAMESPACES)
750 					break;
751 				nvme_notify_ns(aer->ctrlr, nsl->ns[i]);
752 			}
753 		}
754 
755 		/*
756 		 * Pass the cpl data from the original async event completion,
757 		 *  not the log page fetch.
758 		 */
759 		nvme_notify_async_consumers(aer->ctrlr, &aer->cpl,
760 		    aer->log_page_id, aer->log_page_buffer, aer->log_page_size);
761 	}
762 
763 	/*
764 	 * Repost another asynchronous event request to replace the one
765 	 *  that just completed.
766 	 */
767 	nvme_ctrlr_construct_and_submit_aer(aer->ctrlr, aer);
768 }
769 
770 static void
771 nvme_ctrlr_async_event_cb(void *arg, const struct nvme_completion *cpl)
772 {
773 	struct nvme_async_event_request	*aer = arg;
774 
775 	if (nvme_completion_is_error(cpl)) {
776 		/*
777 		 *  Do not retry failed async event requests.  This avoids
778 		 *  infinite loops where a new async event request is submitted
779 		 *  to replace the one just failed, only to fail again and
780 		 *  perpetuate the loop.
781 		 */
782 		return;
783 	}
784 
785 	/* Associated log page is in bits 23:16 of completion entry dw0. */
786 	aer->log_page_id = (cpl->cdw0 & 0xFF0000) >> 16;
787 
788 	nvme_printf(aer->ctrlr, "async event occurred (type 0x%x, info 0x%02x,"
789 	    " page 0x%02x)\n", (cpl->cdw0 & 0x07), (cpl->cdw0 & 0xFF00) >> 8,
790 	    aer->log_page_id);
791 
792 	if (is_log_page_id_valid(aer->log_page_id)) {
793 		aer->log_page_size = nvme_ctrlr_get_log_page_size(aer->ctrlr,
794 		    aer->log_page_id);
795 		memcpy(&aer->cpl, cpl, sizeof(*cpl));
796 		nvme_ctrlr_cmd_get_log_page(aer->ctrlr, aer->log_page_id,
797 		    NVME_GLOBAL_NAMESPACE_TAG, aer->log_page_buffer,
798 		    aer->log_page_size, nvme_ctrlr_async_event_log_page_cb,
799 		    aer);
800 		/* Wait to notify consumers until after log page is fetched. */
801 	} else {
802 		nvme_notify_async_consumers(aer->ctrlr, cpl, aer->log_page_id,
803 		    NULL, 0);
804 
805 		/*
806 		 * Repost another asynchronous event request to replace the one
807 		 *  that just completed.
808 		 */
809 		nvme_ctrlr_construct_and_submit_aer(aer->ctrlr, aer);
810 	}
811 }
812 
813 static void
814 nvme_ctrlr_construct_and_submit_aer(struct nvme_controller *ctrlr,
815     struct nvme_async_event_request *aer)
816 {
817 	struct nvme_request *req;
818 
819 	aer->ctrlr = ctrlr;
820 	req = nvme_allocate_request_null(nvme_ctrlr_async_event_cb, aer);
821 	aer->req = req;
822 
823 	/*
824 	 * Disable timeout here, since asynchronous event requests should by
825 	 *  nature never be timed out.
826 	 */
827 	req->timeout = false;
828 	req->cmd.opc = NVME_OPC_ASYNC_EVENT_REQUEST;
829 	nvme_ctrlr_submit_admin_request(ctrlr, req);
830 }
831 
832 static void
833 nvme_ctrlr_configure_aer(struct nvme_controller *ctrlr)
834 {
835 	struct nvme_completion_poll_status	status;
836 	struct nvme_async_event_request		*aer;
837 	uint32_t				i;
838 
839 	ctrlr->async_event_config = NVME_CRIT_WARN_ST_AVAILABLE_SPARE |
840 	    NVME_CRIT_WARN_ST_DEVICE_RELIABILITY |
841 	    NVME_CRIT_WARN_ST_READ_ONLY |
842 	    NVME_CRIT_WARN_ST_VOLATILE_MEMORY_BACKUP;
843 	if (ctrlr->cdata.ver >= NVME_REV(1, 2))
844 		ctrlr->async_event_config |=
845 		    ctrlr->cdata.oaes & (NVME_ASYNC_EVENT_NS_ATTRIBUTE |
846 			NVME_ASYNC_EVENT_FW_ACTIVATE);
847 
848 	status.done = 0;
849 	nvme_ctrlr_cmd_get_feature(ctrlr, NVME_FEAT_TEMPERATURE_THRESHOLD,
850 	    0, NULL, 0, nvme_completion_poll_cb, &status);
851 	nvme_completion_poll(&status);
852 	if (nvme_completion_is_error(&status.cpl) ||
853 	    (status.cpl.cdw0 & 0xFFFF) == 0xFFFF ||
854 	    (status.cpl.cdw0 & 0xFFFF) == 0x0000) {
855 		nvme_printf(ctrlr, "temperature threshold not supported\n");
856 	} else
857 		ctrlr->async_event_config |= NVME_CRIT_WARN_ST_TEMPERATURE;
858 
859 	nvme_ctrlr_cmd_set_async_event_config(ctrlr,
860 	    ctrlr->async_event_config, NULL, NULL);
861 
862 	/* aerl is a zero-based value, so we need to add 1 here. */
863 	ctrlr->num_aers = min(NVME_MAX_ASYNC_EVENTS, (ctrlr->cdata.aerl+1));
864 
865 	for (i = 0; i < ctrlr->num_aers; i++) {
866 		aer = &ctrlr->aer[i];
867 		nvme_ctrlr_construct_and_submit_aer(ctrlr, aer);
868 	}
869 }
870 
871 static void
872 nvme_ctrlr_configure_int_coalescing(struct nvme_controller *ctrlr)
873 {
874 
875 	ctrlr->int_coal_time = 0;
876 	TUNABLE_INT_FETCH("hw.nvme.int_coal_time",
877 	    &ctrlr->int_coal_time);
878 
879 	ctrlr->int_coal_threshold = 0;
880 	TUNABLE_INT_FETCH("hw.nvme.int_coal_threshold",
881 	    &ctrlr->int_coal_threshold);
882 
883 	nvme_ctrlr_cmd_set_interrupt_coalescing(ctrlr, ctrlr->int_coal_time,
884 	    ctrlr->int_coal_threshold, NULL, NULL);
885 }
886 
887 static void
888 nvme_ctrlr_hmb_free(struct nvme_controller *ctrlr)
889 {
890 	struct nvme_hmb_chunk *hmbc;
891 	int i;
892 
893 	if (ctrlr->hmb_desc_paddr) {
894 		bus_dmamap_unload(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_map);
895 		bus_dmamem_free(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_vaddr,
896 		    ctrlr->hmb_desc_map);
897 		ctrlr->hmb_desc_paddr = 0;
898 	}
899 	if (ctrlr->hmb_desc_tag) {
900 		bus_dma_tag_destroy(ctrlr->hmb_desc_tag);
901 		ctrlr->hmb_desc_tag = NULL;
902 	}
903 	for (i = 0; i < ctrlr->hmb_nchunks; i++) {
904 		hmbc = &ctrlr->hmb_chunks[i];
905 		bus_dmamap_unload(ctrlr->hmb_tag, hmbc->hmbc_map);
906 		bus_dmamem_free(ctrlr->hmb_tag, hmbc->hmbc_vaddr,
907 		    hmbc->hmbc_map);
908 	}
909 	ctrlr->hmb_nchunks = 0;
910 	if (ctrlr->hmb_tag) {
911 		bus_dma_tag_destroy(ctrlr->hmb_tag);
912 		ctrlr->hmb_tag = NULL;
913 	}
914 	if (ctrlr->hmb_chunks) {
915 		free(ctrlr->hmb_chunks, M_NVME);
916 		ctrlr->hmb_chunks = NULL;
917 	}
918 }
919 
920 static void
921 nvme_ctrlr_hmb_alloc(struct nvme_controller *ctrlr)
922 {
923 	struct nvme_hmb_chunk *hmbc;
924 	size_t pref, min, minc, size;
925 	int err, i;
926 	uint64_t max;
927 
928 	/* Limit HMB to 5% of RAM size per device by default. */
929 	max = (uint64_t)physmem * PAGE_SIZE / 20;
930 	TUNABLE_UINT64_FETCH("hw.nvme.hmb_max", &max);
931 
932 	/*
933 	 * Units of Host Memory Buffer in the Identify info are always in terms
934 	 * of 4k units.
935 	 */
936 	min = (long long unsigned)ctrlr->cdata.hmmin * NVME_HMB_UNITS;
937 	if (max == 0 || max < min)
938 		return;
939 	pref = MIN((long long unsigned)ctrlr->cdata.hmpre * NVME_HMB_UNITS, max);
940 	minc = MAX(ctrlr->cdata.hmminds * NVME_HMB_UNITS, ctrlr->page_size);
941 	if (min > 0 && ctrlr->cdata.hmmaxd > 0)
942 		minc = MAX(minc, min / ctrlr->cdata.hmmaxd);
943 	ctrlr->hmb_chunk = pref;
944 
945 again:
946 	/*
947 	 * However, the chunk sizes, number of chunks, and alignment of chunks
948 	 * are all based on the current MPS (ctrlr->page_size).
949 	 */
950 	ctrlr->hmb_chunk = roundup2(ctrlr->hmb_chunk, ctrlr->page_size);
951 	ctrlr->hmb_nchunks = howmany(pref, ctrlr->hmb_chunk);
952 	if (ctrlr->cdata.hmmaxd > 0 && ctrlr->hmb_nchunks > ctrlr->cdata.hmmaxd)
953 		ctrlr->hmb_nchunks = ctrlr->cdata.hmmaxd;
954 	ctrlr->hmb_chunks = malloc(sizeof(struct nvme_hmb_chunk) *
955 	    ctrlr->hmb_nchunks, M_NVME, M_WAITOK);
956 	err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev),
957 	    ctrlr->page_size, 0, BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
958 	    ctrlr->hmb_chunk, 1, ctrlr->hmb_chunk, 0, NULL, NULL, &ctrlr->hmb_tag);
959 	if (err != 0) {
960 		nvme_printf(ctrlr, "HMB tag create failed %d\n", err);
961 		nvme_ctrlr_hmb_free(ctrlr);
962 		return;
963 	}
964 
965 	for (i = 0; i < ctrlr->hmb_nchunks; i++) {
966 		hmbc = &ctrlr->hmb_chunks[i];
967 		if (bus_dmamem_alloc(ctrlr->hmb_tag,
968 		    (void **)&hmbc->hmbc_vaddr, BUS_DMA_NOWAIT,
969 		    &hmbc->hmbc_map)) {
970 			nvme_printf(ctrlr, "failed to alloc HMB\n");
971 			break;
972 		}
973 		if (bus_dmamap_load(ctrlr->hmb_tag, hmbc->hmbc_map,
974 		    hmbc->hmbc_vaddr, ctrlr->hmb_chunk, nvme_single_map,
975 		    &hmbc->hmbc_paddr, BUS_DMA_NOWAIT) != 0) {
976 			bus_dmamem_free(ctrlr->hmb_tag, hmbc->hmbc_vaddr,
977 			    hmbc->hmbc_map);
978 			nvme_printf(ctrlr, "failed to load HMB\n");
979 			break;
980 		}
981 		bus_dmamap_sync(ctrlr->hmb_tag, hmbc->hmbc_map,
982 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
983 	}
984 
985 	if (i < ctrlr->hmb_nchunks && i * ctrlr->hmb_chunk < min &&
986 	    ctrlr->hmb_chunk / 2 >= minc) {
987 		ctrlr->hmb_nchunks = i;
988 		nvme_ctrlr_hmb_free(ctrlr);
989 		ctrlr->hmb_chunk /= 2;
990 		goto again;
991 	}
992 	ctrlr->hmb_nchunks = i;
993 	if (ctrlr->hmb_nchunks * ctrlr->hmb_chunk < min) {
994 		nvme_ctrlr_hmb_free(ctrlr);
995 		return;
996 	}
997 
998 	size = sizeof(struct nvme_hmb_desc) * ctrlr->hmb_nchunks;
999 	err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev),
1000 	    16, 0, BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
1001 	    size, 1, size, 0, NULL, NULL, &ctrlr->hmb_desc_tag);
1002 	if (err != 0) {
1003 		nvme_printf(ctrlr, "HMB desc tag create failed %d\n", err);
1004 		nvme_ctrlr_hmb_free(ctrlr);
1005 		return;
1006 	}
1007 	if (bus_dmamem_alloc(ctrlr->hmb_desc_tag,
1008 	    (void **)&ctrlr->hmb_desc_vaddr, BUS_DMA_WAITOK,
1009 	    &ctrlr->hmb_desc_map)) {
1010 		nvme_printf(ctrlr, "failed to alloc HMB desc\n");
1011 		nvme_ctrlr_hmb_free(ctrlr);
1012 		return;
1013 	}
1014 	if (bus_dmamap_load(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_map,
1015 	    ctrlr->hmb_desc_vaddr, size, nvme_single_map,
1016 	    &ctrlr->hmb_desc_paddr, BUS_DMA_NOWAIT) != 0) {
1017 		bus_dmamem_free(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_vaddr,
1018 		    ctrlr->hmb_desc_map);
1019 		nvme_printf(ctrlr, "failed to load HMB desc\n");
1020 		nvme_ctrlr_hmb_free(ctrlr);
1021 		return;
1022 	}
1023 
1024 	for (i = 0; i < ctrlr->hmb_nchunks; i++) {
1025 		memset(&ctrlr->hmb_desc_vaddr[i], 0,
1026 		    sizeof(struct nvme_hmb_desc));
1027 		ctrlr->hmb_desc_vaddr[i].addr =
1028 		    htole64(ctrlr->hmb_chunks[i].hmbc_paddr);
1029 		ctrlr->hmb_desc_vaddr[i].size = htole32(ctrlr->hmb_chunk / ctrlr->page_size);
1030 	}
1031 	bus_dmamap_sync(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_map,
1032 	    BUS_DMASYNC_PREWRITE);
1033 
1034 	nvme_printf(ctrlr, "Allocated %lluMB host memory buffer\n",
1035 	    (long long unsigned)ctrlr->hmb_nchunks * ctrlr->hmb_chunk
1036 	    / 1024 / 1024);
1037 }
1038 
1039 static void
1040 nvme_ctrlr_hmb_enable(struct nvme_controller *ctrlr, bool enable, bool memret)
1041 {
1042 	struct nvme_completion_poll_status	status;
1043 	uint32_t cdw11;
1044 
1045 	cdw11 = 0;
1046 	if (enable)
1047 		cdw11 |= 1;
1048 	if (memret)
1049 		cdw11 |= 2;
1050 	status.done = 0;
1051 	nvme_ctrlr_cmd_set_feature(ctrlr, NVME_FEAT_HOST_MEMORY_BUFFER, cdw11,
1052 	    ctrlr->hmb_nchunks * ctrlr->hmb_chunk / ctrlr->page_size,
1053 	    ctrlr->hmb_desc_paddr, ctrlr->hmb_desc_paddr >> 32,
1054 	    ctrlr->hmb_nchunks, NULL, 0,
1055 	    nvme_completion_poll_cb, &status);
1056 	nvme_completion_poll(&status);
1057 	if (nvme_completion_is_error(&status.cpl))
1058 		nvme_printf(ctrlr, "nvme_ctrlr_hmb_enable failed!\n");
1059 }
1060 
1061 static void
1062 nvme_ctrlr_start(void *ctrlr_arg, bool resetting)
1063 {
1064 	struct nvme_controller *ctrlr = ctrlr_arg;
1065 	uint32_t old_num_io_queues;
1066 	int i;
1067 
1068 	TSENTER();
1069 
1070 	/*
1071 	 * Only reset adminq here when we are restarting the
1072 	 *  controller after a reset.  During initialization,
1073 	 *  we have already submitted admin commands to get
1074 	 *  the number of I/O queues supported, so cannot reset
1075 	 *  the adminq again here.
1076 	 */
1077 	if (resetting) {
1078 		nvme_qpair_reset(&ctrlr->adminq);
1079 		nvme_admin_qpair_enable(&ctrlr->adminq);
1080 	}
1081 
1082 	if (ctrlr->ioq != NULL) {
1083 		for (i = 0; i < ctrlr->num_io_queues; i++)
1084 			nvme_qpair_reset(&ctrlr->ioq[i]);
1085 	}
1086 
1087 	/*
1088 	 * If it was a reset on initialization command timeout, just
1089 	 * return here, letting initialization code fail gracefully.
1090 	 */
1091 	if (resetting && !ctrlr->is_initialized)
1092 		return;
1093 
1094 	if (resetting && nvme_ctrlr_identify(ctrlr) != 0) {
1095 		nvme_ctrlr_fail(ctrlr);
1096 		return;
1097 	}
1098 
1099 	/*
1100 	 * The number of qpairs are determined during controller initialization,
1101 	 *  including using NVMe SET_FEATURES/NUMBER_OF_QUEUES to determine the
1102 	 *  HW limit.  We call SET_FEATURES again here so that it gets called
1103 	 *  after any reset for controllers that depend on the driver to
1104 	 *  explicit specify how many queues it will use.  This value should
1105 	 *  never change between resets, so panic if somehow that does happen.
1106 	 */
1107 	if (resetting) {
1108 		old_num_io_queues = ctrlr->num_io_queues;
1109 		if (nvme_ctrlr_set_num_qpairs(ctrlr) != 0) {
1110 			nvme_ctrlr_fail(ctrlr);
1111 			return;
1112 		}
1113 
1114 		if (old_num_io_queues != ctrlr->num_io_queues) {
1115 			panic("num_io_queues changed from %u to %u",
1116 			      old_num_io_queues, ctrlr->num_io_queues);
1117 		}
1118 	}
1119 
1120 	if (ctrlr->cdata.hmpre > 0 && ctrlr->hmb_nchunks == 0) {
1121 		nvme_ctrlr_hmb_alloc(ctrlr);
1122 		if (ctrlr->hmb_nchunks > 0)
1123 			nvme_ctrlr_hmb_enable(ctrlr, true, false);
1124 	} else if (ctrlr->hmb_nchunks > 0)
1125 		nvme_ctrlr_hmb_enable(ctrlr, true, true);
1126 
1127 	if (nvme_ctrlr_create_qpairs(ctrlr) != 0) {
1128 		nvme_ctrlr_fail(ctrlr);
1129 		return;
1130 	}
1131 
1132 	if (nvme_ctrlr_construct_namespaces(ctrlr) != 0) {
1133 		nvme_ctrlr_fail(ctrlr);
1134 		return;
1135 	}
1136 
1137 	nvme_ctrlr_configure_aer(ctrlr);
1138 	nvme_ctrlr_configure_int_coalescing(ctrlr);
1139 
1140 	for (i = 0; i < ctrlr->num_io_queues; i++)
1141 		nvme_io_qpair_enable(&ctrlr->ioq[i]);
1142 	TSEXIT();
1143 }
1144 
1145 void
1146 nvme_ctrlr_start_config_hook(void *arg)
1147 {
1148 	struct nvme_controller *ctrlr = arg;
1149 
1150 	TSENTER();
1151 
1152 	if (nvme_ctrlr_hw_reset(ctrlr) != 0) {
1153 fail:
1154 		nvme_ctrlr_fail(ctrlr);
1155 		config_intrhook_disestablish(&ctrlr->config_hook);
1156 		return;
1157 	}
1158 
1159 	nvme_qpair_reset(&ctrlr->adminq);
1160 	nvme_admin_qpair_enable(&ctrlr->adminq);
1161 
1162 	if (nvme_ctrlr_identify(ctrlr) == 0 &&
1163 	    nvme_ctrlr_set_num_qpairs(ctrlr) == 0 &&
1164 	    nvme_ctrlr_construct_io_qpairs(ctrlr) == 0)
1165 		nvme_ctrlr_start(ctrlr, false);
1166 	else
1167 		goto fail;
1168 
1169 	nvme_sysctl_initialize_ctrlr(ctrlr);
1170 	config_intrhook_disestablish(&ctrlr->config_hook);
1171 
1172 	ctrlr->is_initialized = 1;
1173 	nvme_notify_new_controller(ctrlr);
1174 	TSEXIT();
1175 }
1176 
1177 static void
1178 nvme_ctrlr_reset_task(void *arg, int pending)
1179 {
1180 	struct nvme_controller	*ctrlr = arg;
1181 	int			status;
1182 
1183 	nvme_ctrlr_devctl_log(ctrlr, "RESET", "event=\"start\"");
1184 	status = nvme_ctrlr_hw_reset(ctrlr);
1185 	if (status == 0) {
1186 		nvme_ctrlr_devctl_log(ctrlr, "RESET", "event=\"success\"");
1187 		nvme_ctrlr_start(ctrlr, true);
1188 	} else {
1189 		nvme_ctrlr_devctl_log(ctrlr, "RESET", "event=\"timed_out\"");
1190 		nvme_ctrlr_fail(ctrlr);
1191 	}
1192 
1193 	atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
1194 }
1195 
1196 /*
1197  * Poll all the queues enabled on the device for completion.
1198  */
1199 void
1200 nvme_ctrlr_poll(struct nvme_controller *ctrlr)
1201 {
1202 	int i;
1203 
1204 	nvme_qpair_process_completions(&ctrlr->adminq);
1205 
1206 	for (i = 0; i < ctrlr->num_io_queues; i++)
1207 		if (ctrlr->ioq && ctrlr->ioq[i].cpl)
1208 			nvme_qpair_process_completions(&ctrlr->ioq[i]);
1209 }
1210 
1211 /*
1212  * Poll the single-vector interrupt case: num_io_queues will be 1 and
1213  * there's only a single vector. While we're polling, we mask further
1214  * interrupts in the controller.
1215  */
1216 void
1217 nvme_ctrlr_shared_handler(void *arg)
1218 {
1219 	struct nvme_controller *ctrlr = arg;
1220 
1221 	nvme_mmio_write_4(ctrlr, intms, 1);
1222 	nvme_ctrlr_poll(ctrlr);
1223 	nvme_mmio_write_4(ctrlr, intmc, 1);
1224 }
1225 
1226 static void
1227 nvme_pt_done(void *arg, const struct nvme_completion *cpl)
1228 {
1229 	struct nvme_pt_command *pt = arg;
1230 	struct mtx *mtx = pt->driver_lock;
1231 	uint16_t status;
1232 
1233 	bzero(&pt->cpl, sizeof(pt->cpl));
1234 	pt->cpl.cdw0 = cpl->cdw0;
1235 
1236 	status = cpl->status;
1237 	status &= ~NVMEM(NVME_STATUS_P);
1238 	pt->cpl.status = status;
1239 
1240 	mtx_lock(mtx);
1241 	pt->driver_lock = NULL;
1242 	wakeup(pt);
1243 	mtx_unlock(mtx);
1244 }
1245 
1246 int
1247 nvme_ctrlr_passthrough_cmd(struct nvme_controller *ctrlr,
1248     struct nvme_pt_command *pt, uint32_t nsid, int is_user_buffer,
1249     int is_admin_cmd)
1250 {
1251 	struct nvme_request	*req;
1252 	struct mtx		*mtx;
1253 	struct buf		*buf = NULL;
1254 	int			ret = 0;
1255 
1256 	if (pt->len > 0) {
1257 		if (pt->len > ctrlr->max_xfer_size) {
1258 			nvme_printf(ctrlr, "pt->len (%d) "
1259 			    "exceeds max_xfer_size (%d)\n", pt->len,
1260 			    ctrlr->max_xfer_size);
1261 			return EIO;
1262 		}
1263 		if (is_user_buffer) {
1264 			/*
1265 			 * Ensure the user buffer is wired for the duration of
1266 			 *  this pass-through command.
1267 			 */
1268 			PHOLD(curproc);
1269 			buf = uma_zalloc(pbuf_zone, M_WAITOK);
1270 			buf->b_iocmd = pt->is_read ? BIO_READ : BIO_WRITE;
1271 			if (vmapbuf(buf, pt->buf, pt->len, 1) < 0) {
1272 				ret = EFAULT;
1273 				goto err;
1274 			}
1275 			req = nvme_allocate_request_vaddr(buf->b_data, pt->len,
1276 			    nvme_pt_done, pt);
1277 		} else
1278 			req = nvme_allocate_request_vaddr(pt->buf, pt->len,
1279 			    nvme_pt_done, pt);
1280 	} else
1281 		req = nvme_allocate_request_null(nvme_pt_done, pt);
1282 
1283 	/* Assume user space already converted to little-endian */
1284 	req->cmd.opc = pt->cmd.opc;
1285 	req->cmd.fuse = pt->cmd.fuse;
1286 	req->cmd.rsvd2 = pt->cmd.rsvd2;
1287 	req->cmd.rsvd3 = pt->cmd.rsvd3;
1288 	req->cmd.cdw10 = pt->cmd.cdw10;
1289 	req->cmd.cdw11 = pt->cmd.cdw11;
1290 	req->cmd.cdw12 = pt->cmd.cdw12;
1291 	req->cmd.cdw13 = pt->cmd.cdw13;
1292 	req->cmd.cdw14 = pt->cmd.cdw14;
1293 	req->cmd.cdw15 = pt->cmd.cdw15;
1294 
1295 	req->cmd.nsid = htole32(nsid);
1296 
1297 	mtx = mtx_pool_find(mtxpool_sleep, pt);
1298 	pt->driver_lock = mtx;
1299 
1300 	if (is_admin_cmd)
1301 		nvme_ctrlr_submit_admin_request(ctrlr, req);
1302 	else
1303 		nvme_ctrlr_submit_io_request(ctrlr, req);
1304 
1305 	mtx_lock(mtx);
1306 	while (pt->driver_lock != NULL)
1307 		mtx_sleep(pt, mtx, PRIBIO, "nvme_pt", 0);
1308 	mtx_unlock(mtx);
1309 
1310 	if (buf != NULL) {
1311 		vunmapbuf(buf);
1312 err:
1313 		uma_zfree(pbuf_zone, buf);
1314 		PRELE(curproc);
1315 	}
1316 
1317 	return (ret);
1318 }
1319 
1320 static int
1321 nvme_ctrlr_ioctl(struct cdev *cdev, u_long cmd, caddr_t arg, int flag,
1322     struct thread *td)
1323 {
1324 	struct nvme_controller			*ctrlr;
1325 	struct nvme_pt_command			*pt;
1326 
1327 	ctrlr = cdev->si_drv1;
1328 
1329 	switch (cmd) {
1330 	case NVME_RESET_CONTROLLER:
1331 		nvme_ctrlr_reset(ctrlr);
1332 		break;
1333 	case NVME_PASSTHROUGH_CMD:
1334 		pt = (struct nvme_pt_command *)arg;
1335 		return (nvme_ctrlr_passthrough_cmd(ctrlr, pt, le32toh(pt->cmd.nsid),
1336 		    1 /* is_user_buffer */, 1 /* is_admin_cmd */));
1337 	case NVME_GET_NSID:
1338 	{
1339 		struct nvme_get_nsid *gnsid = (struct nvme_get_nsid *)arg;
1340 		strncpy(gnsid->cdev, device_get_nameunit(ctrlr->dev),
1341 		    sizeof(gnsid->cdev));
1342 		gnsid->cdev[sizeof(gnsid->cdev) - 1] = '\0';
1343 		gnsid->nsid = 0;
1344 		break;
1345 	}
1346 	case NVME_GET_MAX_XFER_SIZE:
1347 		*(uint64_t *)arg = ctrlr->max_xfer_size;
1348 		break;
1349 	default:
1350 		return (ENOTTY);
1351 	}
1352 
1353 	return (0);
1354 }
1355 
1356 static struct cdevsw nvme_ctrlr_cdevsw = {
1357 	.d_version =	D_VERSION,
1358 	.d_flags =	0,
1359 	.d_ioctl =	nvme_ctrlr_ioctl
1360 };
1361 
1362 int
1363 nvme_ctrlr_construct(struct nvme_controller *ctrlr, device_t dev)
1364 {
1365 	struct make_dev_args	md_args;
1366 	uint32_t	cap_lo;
1367 	uint32_t	cap_hi;
1368 	uint32_t	to, vs, pmrcap;
1369 	int		status, timeout_period;
1370 
1371 	ctrlr->dev = dev;
1372 
1373 	mtx_init(&ctrlr->lock, "nvme ctrlr lock", NULL, MTX_DEF);
1374 	if (bus_get_domain(dev, &ctrlr->domain) != 0)
1375 		ctrlr->domain = 0;
1376 
1377 	ctrlr->cap_lo = cap_lo = nvme_mmio_read_4(ctrlr, cap_lo);
1378 	if (bootverbose) {
1379 		device_printf(dev, "CapLo: 0x%08x: MQES %u%s%s%s%s, TO %u\n",
1380 		    cap_lo, NVME_CAP_LO_MQES(cap_lo),
1381 		    NVME_CAP_LO_CQR(cap_lo) ? ", CQR" : "",
1382 		    NVME_CAP_LO_AMS(cap_lo) ? ", AMS" : "",
1383 		    (NVME_CAP_LO_AMS(cap_lo) & 0x1) ? " WRRwUPC" : "",
1384 		    (NVME_CAP_LO_AMS(cap_lo) & 0x2) ? " VS" : "",
1385 		    NVME_CAP_LO_TO(cap_lo));
1386 	}
1387 	ctrlr->cap_hi = cap_hi = nvme_mmio_read_4(ctrlr, cap_hi);
1388 	if (bootverbose) {
1389 		device_printf(dev, "CapHi: 0x%08x: DSTRD %u%s, CSS %x%s, "
1390 		    "CPS %x, MPSMIN %u, MPSMAX %u%s%s%s%s%s\n", cap_hi,
1391 		    NVME_CAP_HI_DSTRD(cap_hi),
1392 		    NVME_CAP_HI_NSSRS(cap_hi) ? ", NSSRS" : "",
1393 		    NVME_CAP_HI_CSS(cap_hi),
1394 		    NVME_CAP_HI_BPS(cap_hi) ? ", BPS" : "",
1395 		    NVME_CAP_HI_CPS(cap_hi),
1396 		    NVME_CAP_HI_MPSMIN(cap_hi),
1397 		    NVME_CAP_HI_MPSMAX(cap_hi),
1398 		    NVME_CAP_HI_PMRS(cap_hi) ? ", PMRS" : "",
1399 		    NVME_CAP_HI_CMBS(cap_hi) ? ", CMBS" : "",
1400 		    NVME_CAP_HI_NSSS(cap_hi) ? ", NSSS" : "",
1401 		    NVME_CAP_HI_CRWMS(cap_hi) ? ", CRWMS" : "",
1402 		    NVME_CAP_HI_CRIMS(cap_hi) ? ", CRIMS" : "");
1403 	}
1404 	if (bootverbose) {
1405 		vs = nvme_mmio_read_4(ctrlr, vs);
1406 		device_printf(dev, "Version: 0x%08x: %d.%d\n", vs,
1407 		    NVME_MAJOR(vs), NVME_MINOR(vs));
1408 	}
1409 	if (bootverbose && NVME_CAP_HI_PMRS(cap_hi)) {
1410 		pmrcap = nvme_mmio_read_4(ctrlr, pmrcap);
1411 		device_printf(dev, "PMRCap: 0x%08x: BIR %u%s%s, PMRTU %u, "
1412 		    "PMRWBM %x, PMRTO %u%s\n", pmrcap,
1413 		    NVME_PMRCAP_BIR(pmrcap),
1414 		    NVME_PMRCAP_RDS(pmrcap) ? ", RDS" : "",
1415 		    NVME_PMRCAP_WDS(pmrcap) ? ", WDS" : "",
1416 		    NVME_PMRCAP_PMRTU(pmrcap),
1417 		    NVME_PMRCAP_PMRWBM(pmrcap),
1418 		    NVME_PMRCAP_PMRTO(pmrcap),
1419 		    NVME_PMRCAP_CMSS(pmrcap) ? ", CMSS" : "");
1420 	}
1421 
1422 	ctrlr->dstrd = NVME_CAP_HI_DSTRD(cap_hi) + 2;
1423 
1424 	ctrlr->mps = NVME_CAP_HI_MPSMIN(cap_hi);
1425 	ctrlr->page_size = 1 << (NVME_MPS_SHIFT + ctrlr->mps);
1426 
1427 	/* Get ready timeout value from controller, in units of 500ms. */
1428 	to = NVME_CAP_LO_TO(cap_lo) + 1;
1429 	ctrlr->ready_timeout_in_ms = to * 500;
1430 
1431 	timeout_period = NVME_ADMIN_TIMEOUT_PERIOD;
1432 	TUNABLE_INT_FETCH("hw.nvme.admin_timeout_period", &timeout_period);
1433 	timeout_period = min(timeout_period, NVME_MAX_TIMEOUT_PERIOD);
1434 	timeout_period = max(timeout_period, NVME_MIN_TIMEOUT_PERIOD);
1435 	ctrlr->admin_timeout_period = timeout_period;
1436 
1437 	timeout_period = NVME_DEFAULT_TIMEOUT_PERIOD;
1438 	TUNABLE_INT_FETCH("hw.nvme.timeout_period", &timeout_period);
1439 	timeout_period = min(timeout_period, NVME_MAX_TIMEOUT_PERIOD);
1440 	timeout_period = max(timeout_period, NVME_MIN_TIMEOUT_PERIOD);
1441 	ctrlr->timeout_period = timeout_period;
1442 
1443 	nvme_retry_count = NVME_DEFAULT_RETRY_COUNT;
1444 	TUNABLE_INT_FETCH("hw.nvme.retry_count", &nvme_retry_count);
1445 
1446 	ctrlr->enable_aborts = 0;
1447 	TUNABLE_INT_FETCH("hw.nvme.enable_aborts", &ctrlr->enable_aborts);
1448 
1449 	/* Cap transfers by the maximum addressable by page-sized PRP (4KB pages -> 2MB). */
1450 	ctrlr->max_xfer_size = MIN(maxphys, (ctrlr->page_size / 8 * ctrlr->page_size));
1451 	if (nvme_ctrlr_construct_admin_qpair(ctrlr) != 0)
1452 		return (ENXIO);
1453 
1454 	/*
1455 	 * Create 2 threads for the taskqueue. The reset thread will block when
1456 	 * it detects that the controller has failed until all I/O has been
1457 	 * failed up the stack. The fail_req task needs to be able to run in
1458 	 * this case to finish the request failure for some cases.
1459 	 *
1460 	 * We could partially solve this race by draining the failed requeust
1461 	 * queue before proceding to free the sim, though nothing would stop
1462 	 * new I/O from coming in after we do that drain, but before we reach
1463 	 * cam_sim_free, so this big hammer is used instead.
1464 	 */
1465 	ctrlr->taskqueue = taskqueue_create("nvme_taskq", M_WAITOK,
1466 	    taskqueue_thread_enqueue, &ctrlr->taskqueue);
1467 	taskqueue_start_threads(&ctrlr->taskqueue, 2, PI_DISK, "nvme taskq");
1468 
1469 	ctrlr->is_resetting = 0;
1470 	ctrlr->is_initialized = 0;
1471 	ctrlr->notification_sent = 0;
1472 	TASK_INIT(&ctrlr->reset_task, 0, nvme_ctrlr_reset_task, ctrlr);
1473 	STAILQ_INIT(&ctrlr->fail_req);
1474 	ctrlr->is_failed = false;
1475 
1476 	make_dev_args_init(&md_args);
1477 	md_args.mda_devsw = &nvme_ctrlr_cdevsw;
1478 	md_args.mda_uid = UID_ROOT;
1479 	md_args.mda_gid = GID_WHEEL;
1480 	md_args.mda_mode = 0600;
1481 	md_args.mda_unit = device_get_unit(dev);
1482 	md_args.mda_si_drv1 = (void *)ctrlr;
1483 	status = make_dev_s(&md_args, &ctrlr->cdev, "nvme%d",
1484 	    device_get_unit(dev));
1485 	if (status != 0)
1486 		return (ENXIO);
1487 
1488 	return (0);
1489 }
1490 
1491 void
1492 nvme_ctrlr_destruct(struct nvme_controller *ctrlr, device_t dev)
1493 {
1494 	int	gone, i;
1495 
1496 	ctrlr->is_dying = true;
1497 
1498 	if (ctrlr->resource == NULL)
1499 		goto nores;
1500 	if (!mtx_initialized(&ctrlr->adminq.lock))
1501 		goto noadminq;
1502 
1503 	/*
1504 	 * Check whether it is a hot unplug or a clean driver detach.
1505 	 * If device is not there any more, skip any shutdown commands.
1506 	 */
1507 	gone = (nvme_mmio_read_4(ctrlr, csts) == NVME_GONE);
1508 	if (gone)
1509 		nvme_ctrlr_fail(ctrlr);
1510 	else
1511 		nvme_notify_fail_consumers(ctrlr);
1512 
1513 	for (i = 0; i < NVME_MAX_NAMESPACES; i++)
1514 		nvme_ns_destruct(&ctrlr->ns[i]);
1515 
1516 	if (ctrlr->cdev)
1517 		destroy_dev(ctrlr->cdev);
1518 
1519 	if (ctrlr->is_initialized) {
1520 		if (!gone) {
1521 			if (ctrlr->hmb_nchunks > 0)
1522 				nvme_ctrlr_hmb_enable(ctrlr, false, false);
1523 			nvme_ctrlr_delete_qpairs(ctrlr);
1524 		}
1525 		nvme_ctrlr_hmb_free(ctrlr);
1526 	}
1527 	if (ctrlr->ioq != NULL) {
1528 		for (i = 0; i < ctrlr->num_io_queues; i++)
1529 			nvme_io_qpair_destroy(&ctrlr->ioq[i]);
1530 		free(ctrlr->ioq, M_NVME);
1531 	}
1532 	nvme_admin_qpair_destroy(&ctrlr->adminq);
1533 
1534 	/*
1535 	 *  Notify the controller of a shutdown, even though this is due to
1536 	 *   a driver unload, not a system shutdown (this path is not invoked
1537 	 *   during shutdown).  This ensures the controller receives a
1538 	 *   shutdown notification in case the system is shutdown before
1539 	 *   reloading the driver.
1540 	 */
1541 	if (!gone)
1542 		nvme_ctrlr_shutdown(ctrlr);
1543 
1544 	if (!gone)
1545 		nvme_ctrlr_disable(ctrlr);
1546 
1547 noadminq:
1548 	if (ctrlr->taskqueue)
1549 		taskqueue_free(ctrlr->taskqueue);
1550 
1551 	if (ctrlr->tag)
1552 		bus_teardown_intr(ctrlr->dev, ctrlr->res, ctrlr->tag);
1553 
1554 	if (ctrlr->res)
1555 		bus_release_resource(ctrlr->dev, SYS_RES_IRQ,
1556 		    rman_get_rid(ctrlr->res), ctrlr->res);
1557 
1558 	if (ctrlr->bar4_resource != NULL) {
1559 		bus_release_resource(dev, SYS_RES_MEMORY,
1560 		    ctrlr->bar4_resource_id, ctrlr->bar4_resource);
1561 	}
1562 
1563 	bus_release_resource(dev, SYS_RES_MEMORY,
1564 	    ctrlr->resource_id, ctrlr->resource);
1565 
1566 nores:
1567 	mtx_destroy(&ctrlr->lock);
1568 }
1569 
1570 void
1571 nvme_ctrlr_shutdown(struct nvme_controller *ctrlr)
1572 {
1573 	uint32_t	cc;
1574 	uint32_t	csts;
1575 	int		timeout;
1576 
1577 	cc = nvme_mmio_read_4(ctrlr, cc);
1578 	cc &= ~NVMEM(NVME_CC_REG_SHN);
1579 	cc |= NVMEF(NVME_CC_REG_SHN, NVME_SHN_NORMAL);
1580 	nvme_mmio_write_4(ctrlr, cc, cc);
1581 
1582 	timeout = ticks + (ctrlr->cdata.rtd3e == 0 ? 5 * hz :
1583 	    ((uint64_t)ctrlr->cdata.rtd3e * hz + 999999) / 1000000);
1584 	while (1) {
1585 		csts = nvme_mmio_read_4(ctrlr, csts);
1586 		if (csts == NVME_GONE)		/* Hot unplug. */
1587 			break;
1588 		if (NVME_CSTS_GET_SHST(csts) == NVME_SHST_COMPLETE)
1589 			break;
1590 		if (timeout - ticks < 0) {
1591 			nvme_printf(ctrlr, "shutdown timeout\n");
1592 			break;
1593 		}
1594 		pause("nvmeshut", 1);
1595 	}
1596 }
1597 
1598 void
1599 nvme_ctrlr_submit_admin_request(struct nvme_controller *ctrlr,
1600     struct nvme_request *req)
1601 {
1602 
1603 	nvme_qpair_submit_request(&ctrlr->adminq, req);
1604 }
1605 
1606 void
1607 nvme_ctrlr_submit_io_request(struct nvme_controller *ctrlr,
1608     struct nvme_request *req)
1609 {
1610 	struct nvme_qpair       *qpair;
1611 
1612 	qpair = &ctrlr->ioq[QP(ctrlr, curcpu)];
1613 	nvme_qpair_submit_request(qpair, req);
1614 }
1615 
1616 device_t
1617 nvme_ctrlr_get_device(struct nvme_controller *ctrlr)
1618 {
1619 
1620 	return (ctrlr->dev);
1621 }
1622 
1623 const struct nvme_controller_data *
1624 nvme_ctrlr_get_data(struct nvme_controller *ctrlr)
1625 {
1626 
1627 	return (&ctrlr->cdata);
1628 }
1629 
1630 int
1631 nvme_ctrlr_suspend(struct nvme_controller *ctrlr)
1632 {
1633 	int to = hz;
1634 
1635 	/*
1636 	 * Can't touch failed controllers, so it's already suspended.
1637 	 */
1638 	if (ctrlr->is_failed)
1639 		return (0);
1640 
1641 	/*
1642 	 * We don't want the reset taskqueue running, since it does similar
1643 	 * things, so prevent it from running after we start. Wait for any reset
1644 	 * that may have been started to complete. The reset process we follow
1645 	 * will ensure that any new I/O will queue and be given to the hardware
1646 	 * after we resume (though there should be none).
1647 	 */
1648 	while (atomic_cmpset_32(&ctrlr->is_resetting, 0, 1) == 0 && to-- > 0)
1649 		pause("nvmesusp", 1);
1650 	if (to <= 0) {
1651 		nvme_printf(ctrlr,
1652 		    "Competing reset task didn't finish. Try again later.\n");
1653 		return (EWOULDBLOCK);
1654 	}
1655 
1656 	if (ctrlr->hmb_nchunks > 0)
1657 		nvme_ctrlr_hmb_enable(ctrlr, false, false);
1658 
1659 	/*
1660 	 * Per Section 7.6.2 of NVMe spec 1.4, to properly suspend, we need to
1661 	 * delete the hardware I/O queues, and then shutdown. This properly
1662 	 * flushes any metadata the drive may have stored so it can survive
1663 	 * having its power removed and prevents the unsafe shutdown count from
1664 	 * incriminating. Once we delete the qpairs, we have to disable them
1665 	 * before shutting down.
1666 	 */
1667 	nvme_ctrlr_delete_qpairs(ctrlr);
1668 	nvme_ctrlr_disable_qpairs(ctrlr);
1669 	nvme_ctrlr_shutdown(ctrlr);
1670 
1671 	return (0);
1672 }
1673 
1674 int
1675 nvme_ctrlr_resume(struct nvme_controller *ctrlr)
1676 {
1677 
1678 	/*
1679 	 * Can't touch failed controllers, so nothing to do to resume.
1680 	 */
1681 	if (ctrlr->is_failed)
1682 		return (0);
1683 
1684 	if (nvme_ctrlr_hw_reset(ctrlr) != 0)
1685 		goto fail;
1686 
1687 	/*
1688 	 * Now that we've reset the hardware, we can restart the controller. Any
1689 	 * I/O that was pending is requeued. Any admin commands are aborted with
1690 	 * an error. Once we've restarted, take the controller out of reset.
1691 	 */
1692 	nvme_ctrlr_start(ctrlr, true);
1693 	(void)atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
1694 
1695 	return (0);
1696 fail:
1697 	/*
1698 	 * Since we can't bring the controller out of reset, announce and fail
1699 	 * the controller. However, we have to return success for the resume
1700 	 * itself, due to questionable APIs.
1701 	 */
1702 	nvme_printf(ctrlr, "Failed to reset on resume, failing.\n");
1703 	nvme_ctrlr_fail(ctrlr);
1704 	(void)atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
1705 	return (0);
1706 }
1707