xref: /freebsd/sys/dev/nvme/nvme_ctrlr.c (revision d686ce931cab72612a9e1ada9fe99d65e11a32a3)
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 #include "nvme_linux.h"
47 
48 #define B4_CHK_RDY_DELAY_MS	2300		/* work around controller bug */
49 
50 static void nvme_ctrlr_construct_and_submit_aer(struct nvme_controller *ctrlr,
51 						struct nvme_async_event_request *aer);
52 
53 static void
54 nvme_ctrlr_barrier(struct nvme_controller *ctrlr, int flags)
55 {
56 	bus_barrier(ctrlr->resource, 0, rman_get_size(ctrlr->resource), flags);
57 }
58 
59 static void
60 nvme_ctrlr_devctl_va(struct nvme_controller *ctrlr, const char *type,
61     const char *msg, va_list ap)
62 {
63 	struct sbuf sb;
64 	int error;
65 
66 	if (sbuf_new(&sb, NULL, 0, SBUF_AUTOEXTEND | SBUF_NOWAIT) == NULL)
67 		return;
68 	sbuf_printf(&sb, "name=\"%s\" ", device_get_nameunit(ctrlr->dev));
69 	sbuf_vprintf(&sb, msg, ap);
70 	error = sbuf_finish(&sb);
71 	if (error == 0)
72 		devctl_notify("nvme", "controller", type, sbuf_data(&sb));
73 	sbuf_delete(&sb);
74 }
75 
76 static void
77 nvme_ctrlr_devctl(struct nvme_controller *ctrlr, const char *type, const char *msg, ...)
78 {
79 	va_list ap;
80 
81 	va_start(ap, msg);
82 	nvme_ctrlr_devctl_va(ctrlr, type, msg, ap);
83 	va_end(ap);
84 }
85 
86 static void
87 nvme_ctrlr_devctl_log(struct nvme_controller *ctrlr, const char *type, const char *msg, ...)
88 {
89 	struct sbuf sb;
90 	va_list ap;
91 	int error;
92 
93 	if (sbuf_new(&sb, NULL, 0, SBUF_AUTOEXTEND | SBUF_NOWAIT) == NULL)
94 		return;
95 	sbuf_printf(&sb, "%s: ", device_get_nameunit(ctrlr->dev));
96 	va_start(ap, msg);
97 	sbuf_vprintf(&sb, msg, ap);
98 	va_end(ap);
99 	error = sbuf_finish(&sb);
100 	if (error == 0)
101 		printf("%s\n", sbuf_data(&sb));
102 	sbuf_delete(&sb);
103 	va_start(ap, msg);
104 	nvme_ctrlr_devctl_va(ctrlr, type, msg, ap);
105 	va_end(ap);
106 }
107 
108 static int
109 nvme_ctrlr_construct_admin_qpair(struct nvme_controller *ctrlr)
110 {
111 	struct nvme_qpair	*qpair;
112 	uint32_t		num_entries;
113 	int			error;
114 
115 	qpair = &ctrlr->adminq;
116 	qpair->id = 0;
117 	qpair->cpu = CPU_FFS(&cpuset_domain[ctrlr->domain]) - 1;
118 	qpair->domain = ctrlr->domain;
119 
120 	num_entries = NVME_ADMIN_ENTRIES;
121 	TUNABLE_INT_FETCH("hw.nvme.admin_entries", &num_entries);
122 	/*
123 	 * If admin_entries was overridden to an invalid value, revert it
124 	 *  back to our default value.
125 	 */
126 	if (num_entries < NVME_MIN_ADMIN_ENTRIES ||
127 	    num_entries > NVME_MAX_ADMIN_ENTRIES) {
128 		nvme_printf(ctrlr, "invalid hw.nvme.admin_entries=%d "
129 		    "specified\n", num_entries);
130 		num_entries = NVME_ADMIN_ENTRIES;
131 	}
132 
133 	/*
134 	 * The admin queue's max xfer size is treated differently than the
135 	 *  max I/O xfer size.  16KB is sufficient here - maybe even less?
136 	 */
137 	error = nvme_qpair_construct(qpair, num_entries, NVME_ADMIN_TRACKERS,
138 	     ctrlr);
139 	return (error);
140 }
141 
142 #define QP(ctrlr, c)	((c) * (ctrlr)->num_io_queues / mp_ncpus)
143 
144 static int
145 nvme_ctrlr_construct_io_qpairs(struct nvme_controller *ctrlr)
146 {
147 	struct nvme_qpair	*qpair;
148 	uint32_t		cap_lo;
149 	uint16_t		mqes;
150 	int			c, error, i, n;
151 	int			num_entries, num_trackers, max_entries;
152 
153 	/*
154 	 * NVMe spec sets a hard limit of 64K max entries, but devices may
155 	 * specify a smaller limit, so we need to check the MQES field in the
156 	 * capabilities register. We have to cap the number of entries to the
157 	 * current stride allows for in BAR 0/1, otherwise the remainder entries
158 	 * are inaccessible. MQES should reflect this, and this is just a
159 	 * fail-safe.
160 	 */
161 	max_entries =
162 	    (rman_get_size(ctrlr->resource) - nvme_mmio_offsetof(doorbell[0])) /
163 	    (1 << (ctrlr->dstrd + 1));
164 	num_entries = NVME_IO_ENTRIES;
165 	TUNABLE_INT_FETCH("hw.nvme.io_entries", &num_entries);
166 	cap_lo = nvme_mmio_read_4(ctrlr, cap_lo);
167 	mqes = NVME_CAP_LO_MQES(cap_lo);
168 	num_entries = min(num_entries, mqes + 1);
169 	num_entries = min(num_entries, max_entries);
170 
171 	num_trackers = NVME_IO_TRACKERS;
172 	TUNABLE_INT_FETCH("hw.nvme.io_trackers", &num_trackers);
173 
174 	num_trackers = max(num_trackers, NVME_MIN_IO_TRACKERS);
175 	num_trackers = min(num_trackers, NVME_MAX_IO_TRACKERS);
176 	/*
177 	 * No need to have more trackers than entries in the submit queue.  Note
178 	 * also that for a queue size of N, we can only have (N-1) commands
179 	 * outstanding, hence the "-1" here.
180 	 */
181 	num_trackers = min(num_trackers, (num_entries-1));
182 
183 	/*
184 	 * Our best estimate for the maximum number of I/Os that we should
185 	 * normally have in flight at one time. This should be viewed as a hint,
186 	 * not a hard limit and will need to be revisited when the upper layers
187 	 * of the storage system grows multi-queue support.
188 	 */
189 	ctrlr->max_hw_pend_io = num_trackers * ctrlr->num_io_queues * 3 / 4;
190 
191 	ctrlr->ioq = malloc(ctrlr->num_io_queues * sizeof(struct nvme_qpair),
192 	    M_NVME, M_ZERO | M_WAITOK);
193 
194 	for (i = c = n = 0; i < ctrlr->num_io_queues; i++, c += n) {
195 		qpair = &ctrlr->ioq[i];
196 
197 		/*
198 		 * Admin queue has ID=0. IO queues start at ID=1 -
199 		 *  hence the 'i+1' here.
200 		 */
201 		qpair->id = i + 1;
202 		if (ctrlr->num_io_queues > 1) {
203 			/* Find number of CPUs served by this queue. */
204 			for (n = 1; QP(ctrlr, c + n) == i; n++)
205 				;
206 			/* Shuffle multiple NVMe devices between CPUs. */
207 			qpair->cpu = c + (device_get_unit(ctrlr->dev)+n/2) % n;
208 			qpair->domain = pcpu_find(qpair->cpu)->pc_domain;
209 		} else {
210 			qpair->cpu = CPU_FFS(&cpuset_domain[ctrlr->domain]) - 1;
211 			qpair->domain = ctrlr->domain;
212 		}
213 
214 		/*
215 		 * For I/O queues, use the controller-wide max_xfer_size
216 		 *  calculated in nvme_attach().
217 		 */
218 		error = nvme_qpair_construct(qpair, num_entries, num_trackers,
219 		    ctrlr);
220 		if (error)
221 			return (error);
222 
223 		/*
224 		 * Do not bother binding interrupts if we only have one I/O
225 		 *  interrupt thread for this controller.
226 		 */
227 		if (ctrlr->num_io_queues > 1)
228 			bus_bind_intr(ctrlr->dev, qpair->res, qpair->cpu);
229 	}
230 
231 	return (0);
232 }
233 
234 static void
235 nvme_ctrlr_fail(struct nvme_controller *ctrlr, bool admin_also)
236 {
237 	int i;
238 
239 	/*
240 	 * No need to disable queues before failing them. Failing is a superet
241 	 * of disabling (though pedantically we'd abort the AERs silently with
242 	 * a different error, though when we fail, that hardly matters).
243 	 */
244 	ctrlr->is_failed = true;
245 	if (admin_also) {
246 		ctrlr->is_failed_admin = true;
247 		nvme_qpair_fail(&ctrlr->adminq);
248 	}
249 	if (ctrlr->ioq != NULL) {
250 		for (i = 0; i < ctrlr->num_io_queues; i++) {
251 			nvme_qpair_fail(&ctrlr->ioq[i]);
252 		}
253 	}
254 	nvme_notify_fail_consumers(ctrlr);
255 }
256 
257 /*
258  * Wait for RDY to change.
259  *
260  * Starts sleeping for 1us and geometrically increases it the longer we wait,
261  * capped at 1ms.
262  */
263 static int
264 nvme_ctrlr_wait_for_ready(struct nvme_controller *ctrlr, int desired_val)
265 {
266 	int timeout = ticks + MSEC_2_TICKS(ctrlr->ready_timeout_in_ms);
267 	sbintime_t delta_t = SBT_1US;
268 	uint32_t csts;
269 
270 	while (1) {
271 		csts = nvme_mmio_read_4(ctrlr, csts);
272 		if (csts == NVME_GONE)		/* Hot unplug. */
273 			return (ENXIO);
274 		if (NVMEV(NVME_CSTS_REG_RDY, csts) == desired_val)
275 			break;
276 		if (timeout - ticks < 0) {
277 			nvme_printf(ctrlr, "controller ready did not become %d "
278 			    "within %d ms\n", desired_val, ctrlr->ready_timeout_in_ms);
279 			return (ENXIO);
280 		}
281 
282 		pause_sbt("nvmerdy", delta_t, 0, C_PREL(1));
283 		delta_t = min(SBT_1MS, delta_t * 3 / 2);
284 	}
285 
286 	return (0);
287 }
288 
289 static int
290 nvme_ctrlr_disable(struct nvme_controller *ctrlr)
291 {
292 	uint32_t cc;
293 	uint32_t csts;
294 	uint8_t  en, rdy;
295 	int err;
296 
297 	cc = nvme_mmio_read_4(ctrlr, cc);
298 	csts = nvme_mmio_read_4(ctrlr, csts);
299 
300 	en = NVMEV(NVME_CC_REG_EN, cc);
301 	rdy = NVMEV(NVME_CSTS_REG_RDY, csts);
302 
303 	/*
304 	 * Per 3.1.5 in NVME 1.3 spec, transitioning CC.EN from 0 to 1
305 	 * when CSTS.RDY is 1 or transitioning CC.EN from 1 to 0 when
306 	 * CSTS.RDY is 0 "has undefined results" So make sure that CSTS.RDY
307 	 * isn't the desired value. Short circuit if we're already disabled.
308 	 */
309 	if (en == 0) {
310 		/* Wait for RDY == 0 or timeout & fail */
311 		if (rdy == 0)
312 			return (0);
313 		return (nvme_ctrlr_wait_for_ready(ctrlr, 0));
314 	}
315 	if (rdy == 0) {
316 		/* EN == 1, wait for  RDY == 1 or timeout & fail */
317 		err = nvme_ctrlr_wait_for_ready(ctrlr, 1);
318 		if (err != 0)
319 			return (err);
320 	}
321 
322 	cc &= ~NVMEM(NVME_CC_REG_EN);
323 	nvme_mmio_write_4(ctrlr, cc, cc);
324 
325 	/*
326 	 * A few drives have firmware bugs that freeze the drive if we access
327 	 * the mmio too soon after we disable.
328 	 */
329 	if (ctrlr->quirks & QUIRK_DELAY_B4_CHK_RDY)
330 		pause("nvmeR", MSEC_2_TICKS(B4_CHK_RDY_DELAY_MS));
331 	return (nvme_ctrlr_wait_for_ready(ctrlr, 0));
332 }
333 
334 static int
335 nvme_ctrlr_enable(struct nvme_controller *ctrlr)
336 {
337 	uint32_t	cc;
338 	uint32_t	csts;
339 	uint32_t	aqa;
340 	uint32_t	qsize;
341 	uint8_t		en, rdy;
342 	int		err;
343 
344 	cc = nvme_mmio_read_4(ctrlr, cc);
345 	csts = nvme_mmio_read_4(ctrlr, csts);
346 
347 	en = NVMEV(NVME_CC_REG_EN, cc);
348 	rdy = NVMEV(NVME_CSTS_REG_RDY, csts);
349 
350 	/*
351 	 * See note in nvme_ctrlr_disable. Short circuit if we're already enabled.
352 	 */
353 	if (en == 1) {
354 		if (rdy == 1)
355 			return (0);
356 		return (nvme_ctrlr_wait_for_ready(ctrlr, 1));
357 	}
358 
359 	/* EN == 0 already wait for RDY == 0 or timeout & fail */
360 	err = nvme_ctrlr_wait_for_ready(ctrlr, 0);
361 	if (err != 0)
362 		return (err);
363 
364 	nvme_mmio_write_8(ctrlr, asq, ctrlr->adminq.cmd_bus_addr);
365 	nvme_mmio_write_8(ctrlr, acq, ctrlr->adminq.cpl_bus_addr);
366 
367 	/* acqs and asqs are 0-based. */
368 	qsize = ctrlr->adminq.num_entries - 1;
369 
370 	aqa = 0;
371 	aqa |= NVMEF(NVME_AQA_REG_ACQS, qsize);
372 	aqa |= NVMEF(NVME_AQA_REG_ASQS, qsize);
373 	nvme_mmio_write_4(ctrlr, aqa, aqa);
374 
375 	/* Initialization values for CC */
376 	cc = 0;
377 	cc |= NVMEF(NVME_CC_REG_EN, 1);
378 	cc |= NVMEF(NVME_CC_REG_CSS, 0);
379 	cc |= NVMEF(NVME_CC_REG_AMS, 0);
380 	cc |= NVMEF(NVME_CC_REG_SHN, 0);
381 	cc |= NVMEF(NVME_CC_REG_IOSQES, 6); /* SQ entry size == 64 == 2^6 */
382 	cc |= NVMEF(NVME_CC_REG_IOCQES, 4); /* CQ entry size == 16 == 2^4 */
383 
384 	/*
385 	 * Use the Memory Page Size selected during device initialization.  Note
386 	 * that value stored in mps is suitable to use here without adjusting by
387 	 * NVME_MPS_SHIFT.
388 	 */
389 	cc |= NVMEF(NVME_CC_REG_MPS, ctrlr->mps);
390 
391 	nvme_ctrlr_barrier(ctrlr, BUS_SPACE_BARRIER_WRITE);
392 	nvme_mmio_write_4(ctrlr, cc, cc);
393 
394 	return (nvme_ctrlr_wait_for_ready(ctrlr, 1));
395 }
396 
397 static void
398 nvme_ctrlr_disable_qpairs(struct nvme_controller *ctrlr)
399 {
400 	int i;
401 
402 	nvme_admin_qpair_disable(&ctrlr->adminq);
403 	/*
404 	 * I/O queues are not allocated before the initial HW
405 	 *  reset, so do not try to disable them.  Use is_initialized
406 	 *  to determine if this is the initial HW reset.
407 	 */
408 	if (ctrlr->is_initialized) {
409 		for (i = 0; i < ctrlr->num_io_queues; i++)
410 			nvme_io_qpair_disable(&ctrlr->ioq[i]);
411 	}
412 }
413 
414 static int
415 nvme_ctrlr_hw_reset(struct nvme_controller *ctrlr)
416 {
417 	int err;
418 
419 	TSENTER();
420 
421 	ctrlr->is_failed_admin = true;
422 	nvme_ctrlr_disable_qpairs(ctrlr);
423 
424 	err = nvme_ctrlr_disable(ctrlr);
425 	if (err != 0)
426 		goto out;
427 
428 	err = nvme_ctrlr_enable(ctrlr);
429 out:
430 	if (err == 0)
431 		ctrlr->is_failed_admin = false;
432 
433 	TSEXIT();
434 	return (err);
435 }
436 
437 void
438 nvme_ctrlr_reset(struct nvme_controller *ctrlr)
439 {
440 	int cmpset;
441 
442 	cmpset = atomic_cmpset_32(&ctrlr->is_resetting, 0, 1);
443 
444 	if (cmpset == 0)
445 		/*
446 		 * Controller is already resetting.  Return immediately since
447 		 * there is no need to kick off another reset.
448 		 */
449 		return;
450 
451 	if (!ctrlr->is_dying)
452 		taskqueue_enqueue(ctrlr->taskqueue, &ctrlr->reset_task);
453 }
454 
455 static int
456 nvme_ctrlr_identify(struct nvme_controller *ctrlr)
457 {
458 	struct nvme_completion_poll_status	status;
459 
460 	status.done = 0;
461 	nvme_ctrlr_cmd_identify_controller(ctrlr, &ctrlr->cdata,
462 	    nvme_completion_poll_cb, &status);
463 	nvme_completion_poll(&status);
464 	if (nvme_completion_is_error(&status.cpl)) {
465 		nvme_printf(ctrlr, "nvme_identify_controller failed!\n");
466 		return (ENXIO);
467 	}
468 
469 	/* Convert data to host endian */
470 	nvme_controller_data_swapbytes(&ctrlr->cdata);
471 
472 	/*
473 	 * Use MDTS to ensure our default max_xfer_size doesn't exceed what the
474 	 *  controller supports.
475 	 */
476 	if (ctrlr->cdata.mdts > 0)
477 		ctrlr->max_xfer_size = min(ctrlr->max_xfer_size,
478 		    1 << (ctrlr->cdata.mdts + NVME_MPS_SHIFT +
479 			NVME_CAP_HI_MPSMIN(ctrlr->cap_hi)));
480 
481 	return (0);
482 }
483 
484 static int
485 nvme_ctrlr_set_num_qpairs(struct nvme_controller *ctrlr)
486 {
487 	struct nvme_completion_poll_status	status;
488 	int					cq_allocated, sq_allocated;
489 
490 	status.done = 0;
491 	nvme_ctrlr_cmd_set_num_queues(ctrlr, ctrlr->num_io_queues,
492 	    nvme_completion_poll_cb, &status);
493 	nvme_completion_poll(&status);
494 	if (nvme_completion_is_error(&status.cpl)) {
495 		nvme_printf(ctrlr, "nvme_ctrlr_set_num_qpairs failed!\n");
496 		return (ENXIO);
497 	}
498 
499 	/*
500 	 * Data in cdw0 is 0-based.
501 	 * Lower 16-bits indicate number of submission queues allocated.
502 	 * Upper 16-bits indicate number of completion queues allocated.
503 	 */
504 	sq_allocated = (status.cpl.cdw0 & 0xFFFF) + 1;
505 	cq_allocated = (status.cpl.cdw0 >> 16) + 1;
506 
507 	/*
508 	 * Controller may allocate more queues than we requested,
509 	 *  so use the minimum of the number requested and what was
510 	 *  actually allocated.
511 	 */
512 	ctrlr->num_io_queues = min(ctrlr->num_io_queues, sq_allocated);
513 	ctrlr->num_io_queues = min(ctrlr->num_io_queues, cq_allocated);
514 	if (ctrlr->num_io_queues > vm_ndomains)
515 		ctrlr->num_io_queues -= ctrlr->num_io_queues % vm_ndomains;
516 
517 	return (0);
518 }
519 
520 static int
521 nvme_ctrlr_create_qpairs(struct nvme_controller *ctrlr)
522 {
523 	struct nvme_completion_poll_status	status;
524 	struct nvme_qpair			*qpair;
525 	int					i;
526 
527 	for (i = 0; i < ctrlr->num_io_queues; i++) {
528 		qpair = &ctrlr->ioq[i];
529 
530 		status.done = 0;
531 		nvme_ctrlr_cmd_create_io_cq(ctrlr, qpair,
532 		    nvme_completion_poll_cb, &status);
533 		nvme_completion_poll(&status);
534 		if (nvme_completion_is_error(&status.cpl)) {
535 			nvme_printf(ctrlr, "nvme_create_io_cq failed!\n");
536 			return (ENXIO);
537 		}
538 
539 		status.done = 0;
540 		nvme_ctrlr_cmd_create_io_sq(ctrlr, qpair,
541 		    nvme_completion_poll_cb, &status);
542 		nvme_completion_poll(&status);
543 		if (nvme_completion_is_error(&status.cpl)) {
544 			nvme_printf(ctrlr, "nvme_create_io_sq failed!\n");
545 			return (ENXIO);
546 		}
547 	}
548 
549 	return (0);
550 }
551 
552 static int
553 nvme_ctrlr_delete_qpairs(struct nvme_controller *ctrlr)
554 {
555 	struct nvme_completion_poll_status	status;
556 	struct nvme_qpair			*qpair;
557 
558 	for (int i = 0; i < ctrlr->num_io_queues; i++) {
559 		qpair = &ctrlr->ioq[i];
560 
561 		status.done = 0;
562 		nvme_ctrlr_cmd_delete_io_sq(ctrlr, qpair,
563 		    nvme_completion_poll_cb, &status);
564 		nvme_completion_poll(&status);
565 		if (nvme_completion_is_error(&status.cpl)) {
566 			nvme_printf(ctrlr, "nvme_destroy_io_sq failed!\n");
567 			return (ENXIO);
568 		}
569 
570 		status.done = 0;
571 		nvme_ctrlr_cmd_delete_io_cq(ctrlr, qpair,
572 		    nvme_completion_poll_cb, &status);
573 		nvme_completion_poll(&status);
574 		if (nvme_completion_is_error(&status.cpl)) {
575 			nvme_printf(ctrlr, "nvme_destroy_io_cq failed!\n");
576 			return (ENXIO);
577 		}
578 	}
579 
580 	return (0);
581 }
582 
583 static int
584 nvme_ctrlr_construct_namespaces(struct nvme_controller *ctrlr)
585 {
586 	struct nvme_namespace	*ns;
587 	uint32_t 		i;
588 
589 	for (i = 0; i < min(ctrlr->cdata.nn, NVME_MAX_NAMESPACES); i++) {
590 		ns = &ctrlr->ns[i];
591 		nvme_ns_construct(ns, i+1, ctrlr);
592 	}
593 
594 	return (0);
595 }
596 
597 static bool
598 is_log_page_id_valid(uint8_t page_id)
599 {
600 
601 	switch (page_id) {
602 	case NVME_LOG_ERROR:
603 	case NVME_LOG_HEALTH_INFORMATION:
604 	case NVME_LOG_FIRMWARE_SLOT:
605 	case NVME_LOG_CHANGED_NAMESPACE:
606 	case NVME_LOG_COMMAND_EFFECT:
607 	case NVME_LOG_RES_NOTIFICATION:
608 	case NVME_LOG_SANITIZE_STATUS:
609 		return (true);
610 	}
611 
612 	return (false);
613 }
614 
615 static uint32_t
616 nvme_ctrlr_get_log_page_size(struct nvme_controller *ctrlr, uint8_t page_id)
617 {
618 	uint32_t	log_page_size;
619 
620 	switch (page_id) {
621 	case NVME_LOG_ERROR:
622 		log_page_size = min(
623 		    sizeof(struct nvme_error_information_entry) *
624 		    (ctrlr->cdata.elpe + 1), NVME_MAX_AER_LOG_SIZE);
625 		break;
626 	case NVME_LOG_HEALTH_INFORMATION:
627 		log_page_size = sizeof(struct nvme_health_information_page);
628 		break;
629 	case NVME_LOG_FIRMWARE_SLOT:
630 		log_page_size = sizeof(struct nvme_firmware_page);
631 		break;
632 	case NVME_LOG_CHANGED_NAMESPACE:
633 		log_page_size = sizeof(struct nvme_ns_list);
634 		break;
635 	case NVME_LOG_COMMAND_EFFECT:
636 		log_page_size = sizeof(struct nvme_command_effects_page);
637 		break;
638 	case NVME_LOG_RES_NOTIFICATION:
639 		log_page_size = sizeof(struct nvme_res_notification_page);
640 		break;
641 	case NVME_LOG_SANITIZE_STATUS:
642 		log_page_size = sizeof(struct nvme_sanitize_status_page);
643 		break;
644 	default:
645 		log_page_size = 0;
646 		break;
647 	}
648 
649 	return (log_page_size);
650 }
651 
652 static void
653 nvme_ctrlr_log_critical_warnings(struct nvme_controller *ctrlr,
654     uint8_t state)
655 {
656 
657 	if (state & NVME_CRIT_WARN_ST_AVAILABLE_SPARE)
658 		nvme_printf(ctrlr, "SMART WARNING: available spare space below threshold\n");
659 
660 	if (state & NVME_CRIT_WARN_ST_TEMPERATURE)
661 		nvme_printf(ctrlr, "SMART WARNING: temperature above threshold\n");
662 
663 	if (state & NVME_CRIT_WARN_ST_DEVICE_RELIABILITY)
664 		nvme_printf(ctrlr, "SMART WARNING: device reliability degraded\n");
665 
666 	if (state & NVME_CRIT_WARN_ST_READ_ONLY)
667 		nvme_printf(ctrlr, "SMART WARNING: media placed in read only mode\n");
668 
669 	if (state & NVME_CRIT_WARN_ST_VOLATILE_MEMORY_BACKUP)
670 		nvme_printf(ctrlr, "SMART WARNING: volatile memory backup device failed\n");
671 
672 	if (state & NVME_CRIT_WARN_ST_PERSISTENT_MEMORY_REGION)
673 		nvme_printf(ctrlr, "SMART WARNING: persistent memory read only or unreliable\n");
674 
675 	if (state & NVME_CRIT_WARN_ST_RESERVED_MASK)
676 		nvme_printf(ctrlr, "SMART WARNING: unknown critical warning(s): state = 0x%02x\n",
677 		    state & NVME_CRIT_WARN_ST_RESERVED_MASK);
678 
679 	nvme_ctrlr_devctl(ctrlr, "critical", "SMART_ERROR", "state=0x%02x", state);
680 }
681 
682 static void
683 nvme_ctrlr_async_event_log_page_cb(void *arg, const struct nvme_completion *cpl)
684 {
685 	struct nvme_async_event_request		*aer = arg;
686 	struct nvme_health_information_page	*health_info;
687 	struct nvme_ns_list			*nsl;
688 	struct nvme_error_information_entry	*err;
689 	int i;
690 
691 	/*
692 	 * If the log page fetch for some reason completed with an error,
693 	 *  don't pass log page data to the consumers.  In practice, this case
694 	 *  should never happen.
695 	 */
696 	if (nvme_completion_is_error(cpl))
697 		nvme_notify_async_consumers(aer->ctrlr, &aer->cpl,
698 		    aer->log_page_id, NULL, 0);
699 	else {
700 		/* Convert data to host endian */
701 		switch (aer->log_page_id) {
702 		case NVME_LOG_ERROR:
703 			err = (struct nvme_error_information_entry *)aer->log_page_buffer;
704 			for (i = 0; i < (aer->ctrlr->cdata.elpe + 1); i++)
705 				nvme_error_information_entry_swapbytes(err++);
706 			break;
707 		case NVME_LOG_HEALTH_INFORMATION:
708 			nvme_health_information_page_swapbytes(
709 			    (struct nvme_health_information_page *)aer->log_page_buffer);
710 			break;
711 		case NVME_LOG_CHANGED_NAMESPACE:
712 			nvme_ns_list_swapbytes(
713 			    (struct nvme_ns_list *)aer->log_page_buffer);
714 			break;
715 		case NVME_LOG_COMMAND_EFFECT:
716 			nvme_command_effects_page_swapbytes(
717 			    (struct nvme_command_effects_page *)aer->log_page_buffer);
718 			break;
719 		case NVME_LOG_RES_NOTIFICATION:
720 			nvme_res_notification_page_swapbytes(
721 			    (struct nvme_res_notification_page *)aer->log_page_buffer);
722 			break;
723 		case NVME_LOG_SANITIZE_STATUS:
724 			nvme_sanitize_status_page_swapbytes(
725 			    (struct nvme_sanitize_status_page *)aer->log_page_buffer);
726 			break;
727 		default:
728 			break;
729 		}
730 
731 		if (aer->log_page_id == NVME_LOG_HEALTH_INFORMATION) {
732 			health_info = (struct nvme_health_information_page *)
733 			    aer->log_page_buffer;
734 			nvme_ctrlr_log_critical_warnings(aer->ctrlr,
735 			    health_info->critical_warning);
736 			/*
737 			 * Critical warnings reported through the
738 			 *  SMART/health log page are persistent, so
739 			 *  clear the associated bits in the async event
740 			 *  config so that we do not receive repeated
741 			 *  notifications for the same event.
742 			 */
743 			aer->ctrlr->async_event_config &=
744 			    ~health_info->critical_warning;
745 			nvme_ctrlr_cmd_set_async_event_config(aer->ctrlr,
746 			    aer->ctrlr->async_event_config, NULL, NULL);
747 		} else if (aer->log_page_id == NVME_LOG_CHANGED_NAMESPACE &&
748 		    !nvme_use_nvd) {
749 			nsl = (struct nvme_ns_list *)aer->log_page_buffer;
750 			for (i = 0; i < nitems(nsl->ns) && nsl->ns[i] != 0; i++) {
751 				if (nsl->ns[i] > NVME_MAX_NAMESPACES)
752 					break;
753 				nvme_notify_ns(aer->ctrlr, nsl->ns[i]);
754 			}
755 		}
756 
757 		/*
758 		 * Pass the cpl data from the original async event completion,
759 		 *  not the log page fetch.
760 		 */
761 		nvme_notify_async_consumers(aer->ctrlr, &aer->cpl,
762 		    aer->log_page_id, aer->log_page_buffer, aer->log_page_size);
763 	}
764 
765 	/*
766 	 * Repost another asynchronous event request to replace the one
767 	 *  that just completed.
768 	 */
769 	nvme_ctrlr_construct_and_submit_aer(aer->ctrlr, aer);
770 }
771 
772 static void
773 nvme_ctrlr_async_event_cb(void *arg, const struct nvme_completion *cpl)
774 {
775 	struct nvme_async_event_request	*aer = arg;
776 
777 	if (nvme_completion_is_error(cpl)) {
778 		/*
779 		 *  Do not retry failed async event requests.  This avoids
780 		 *  infinite loops where a new async event request is submitted
781 		 *  to replace the one just failed, only to fail again and
782 		 *  perpetuate the loop.
783 		 */
784 		return;
785 	}
786 
787 	/* Associated log page is in bits 23:16 of completion entry dw0. */
788 	aer->log_page_id = NVMEV(NVME_ASYNC_EVENT_LOG_PAGE_ID, cpl->cdw0);
789 
790 	nvme_printf(aer->ctrlr, "async event occurred (type 0x%x, info 0x%02x,"
791 	    " page 0x%02x)\n", NVMEV(NVME_ASYNC_EVENT_TYPE, cpl->cdw0),
792 	    NVMEV(NVME_ASYNC_EVENT_INFO, cpl->cdw0),
793 	    aer->log_page_id);
794 
795 	if (is_log_page_id_valid(aer->log_page_id)) {
796 		aer->log_page_size = nvme_ctrlr_get_log_page_size(aer->ctrlr,
797 		    aer->log_page_id);
798 		memcpy(&aer->cpl, cpl, sizeof(*cpl));
799 		nvme_ctrlr_cmd_get_log_page(aer->ctrlr, aer->log_page_id,
800 		    NVME_GLOBAL_NAMESPACE_TAG, aer->log_page_buffer,
801 		    aer->log_page_size, nvme_ctrlr_async_event_log_page_cb,
802 		    aer);
803 		/* Wait to notify consumers until after log page is fetched. */
804 	} else {
805 		nvme_notify_async_consumers(aer->ctrlr, cpl, aer->log_page_id,
806 		    NULL, 0);
807 
808 		/*
809 		 * Repost another asynchronous event request to replace the one
810 		 *  that just completed.
811 		 */
812 		nvme_ctrlr_construct_and_submit_aer(aer->ctrlr, aer);
813 	}
814 }
815 
816 static void
817 nvme_ctrlr_construct_and_submit_aer(struct nvme_controller *ctrlr,
818     struct nvme_async_event_request *aer)
819 {
820 	struct nvme_request *req;
821 
822 	aer->ctrlr = ctrlr;
823 	req = nvme_allocate_request_null(nvme_ctrlr_async_event_cb, aer);
824 	aer->req = req;
825 
826 	/*
827 	 * Disable timeout here, since asynchronous event requests should by
828 	 *  nature never be timed out.
829 	 */
830 	req->timeout = false;
831 	req->cmd.opc = NVME_OPC_ASYNC_EVENT_REQUEST;
832 	nvme_ctrlr_submit_admin_request(ctrlr, req);
833 }
834 
835 static void
836 nvme_ctrlr_configure_aer(struct nvme_controller *ctrlr)
837 {
838 	struct nvme_completion_poll_status	status;
839 	struct nvme_async_event_request		*aer;
840 	uint32_t				i;
841 
842 	ctrlr->async_event_config = NVME_CRIT_WARN_ST_AVAILABLE_SPARE |
843 	    NVME_CRIT_WARN_ST_DEVICE_RELIABILITY |
844 	    NVME_CRIT_WARN_ST_READ_ONLY |
845 	    NVME_CRIT_WARN_ST_VOLATILE_MEMORY_BACKUP;
846 	if (ctrlr->cdata.ver >= NVME_REV(1, 2))
847 		ctrlr->async_event_config |=
848 		    ctrlr->cdata.oaes & (NVME_ASYNC_EVENT_NS_ATTRIBUTE |
849 			NVME_ASYNC_EVENT_FW_ACTIVATE);
850 
851 	status.done = 0;
852 	nvme_ctrlr_cmd_get_feature(ctrlr, NVME_FEAT_TEMPERATURE_THRESHOLD,
853 	    0, NULL, 0, nvme_completion_poll_cb, &status);
854 	nvme_completion_poll(&status);
855 	if (nvme_completion_is_error(&status.cpl) ||
856 	    (status.cpl.cdw0 & 0xFFFF) == 0xFFFF ||
857 	    (status.cpl.cdw0 & 0xFFFF) == 0x0000) {
858 		nvme_printf(ctrlr, "temperature threshold not supported\n");
859 	} else
860 		ctrlr->async_event_config |= NVME_CRIT_WARN_ST_TEMPERATURE;
861 
862 	nvme_ctrlr_cmd_set_async_event_config(ctrlr,
863 	    ctrlr->async_event_config, NULL, NULL);
864 
865 	/* aerl is a zero-based value, so we need to add 1 here. */
866 	ctrlr->num_aers = min(NVME_MAX_ASYNC_EVENTS, (ctrlr->cdata.aerl+1));
867 
868 	for (i = 0; i < ctrlr->num_aers; i++) {
869 		aer = &ctrlr->aer[i];
870 		nvme_ctrlr_construct_and_submit_aer(ctrlr, aer);
871 	}
872 }
873 
874 static void
875 nvme_ctrlr_configure_int_coalescing(struct nvme_controller *ctrlr)
876 {
877 
878 	ctrlr->int_coal_time = 0;
879 	TUNABLE_INT_FETCH("hw.nvme.int_coal_time",
880 	    &ctrlr->int_coal_time);
881 
882 	ctrlr->int_coal_threshold = 0;
883 	TUNABLE_INT_FETCH("hw.nvme.int_coal_threshold",
884 	    &ctrlr->int_coal_threshold);
885 
886 	nvme_ctrlr_cmd_set_interrupt_coalescing(ctrlr, ctrlr->int_coal_time,
887 	    ctrlr->int_coal_threshold, NULL, NULL);
888 }
889 
890 static void
891 nvme_ctrlr_hmb_free(struct nvme_controller *ctrlr)
892 {
893 	struct nvme_hmb_chunk *hmbc;
894 	int i;
895 
896 	if (ctrlr->hmb_desc_paddr) {
897 		bus_dmamap_unload(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_map);
898 		bus_dmamem_free(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_vaddr,
899 		    ctrlr->hmb_desc_map);
900 		ctrlr->hmb_desc_paddr = 0;
901 	}
902 	if (ctrlr->hmb_desc_tag) {
903 		bus_dma_tag_destroy(ctrlr->hmb_desc_tag);
904 		ctrlr->hmb_desc_tag = NULL;
905 	}
906 	for (i = 0; i < ctrlr->hmb_nchunks; i++) {
907 		hmbc = &ctrlr->hmb_chunks[i];
908 		bus_dmamap_unload(ctrlr->hmb_tag, hmbc->hmbc_map);
909 		bus_dmamem_free(ctrlr->hmb_tag, hmbc->hmbc_vaddr,
910 		    hmbc->hmbc_map);
911 	}
912 	ctrlr->hmb_nchunks = 0;
913 	if (ctrlr->hmb_tag) {
914 		bus_dma_tag_destroy(ctrlr->hmb_tag);
915 		ctrlr->hmb_tag = NULL;
916 	}
917 	if (ctrlr->hmb_chunks) {
918 		free(ctrlr->hmb_chunks, M_NVME);
919 		ctrlr->hmb_chunks = NULL;
920 	}
921 }
922 
923 static void
924 nvme_ctrlr_hmb_alloc(struct nvme_controller *ctrlr)
925 {
926 	struct nvme_hmb_chunk *hmbc;
927 	size_t pref, min, minc, size;
928 	int err, i;
929 	uint64_t max;
930 
931 	/* Limit HMB to 5% of RAM size per device by default. */
932 	max = (uint64_t)physmem * PAGE_SIZE / 20;
933 	TUNABLE_UINT64_FETCH("hw.nvme.hmb_max", &max);
934 
935 	/*
936 	 * Units of Host Memory Buffer in the Identify info are always in terms
937 	 * of 4k units.
938 	 */
939 	min = (long long unsigned)ctrlr->cdata.hmmin * NVME_HMB_UNITS;
940 	if (max == 0 || max < min)
941 		return;
942 	pref = MIN((long long unsigned)ctrlr->cdata.hmpre * NVME_HMB_UNITS, max);
943 	minc = MAX(ctrlr->cdata.hmminds * NVME_HMB_UNITS, ctrlr->page_size);
944 	if (min > 0 && ctrlr->cdata.hmmaxd > 0)
945 		minc = MAX(minc, min / ctrlr->cdata.hmmaxd);
946 	ctrlr->hmb_chunk = pref;
947 
948 again:
949 	/*
950 	 * However, the chunk sizes, number of chunks, and alignment of chunks
951 	 * are all based on the current MPS (ctrlr->page_size).
952 	 */
953 	ctrlr->hmb_chunk = roundup2(ctrlr->hmb_chunk, ctrlr->page_size);
954 	ctrlr->hmb_nchunks = howmany(pref, ctrlr->hmb_chunk);
955 	if (ctrlr->cdata.hmmaxd > 0 && ctrlr->hmb_nchunks > ctrlr->cdata.hmmaxd)
956 		ctrlr->hmb_nchunks = ctrlr->cdata.hmmaxd;
957 	ctrlr->hmb_chunks = malloc(sizeof(struct nvme_hmb_chunk) *
958 	    ctrlr->hmb_nchunks, M_NVME, M_WAITOK);
959 	err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev),
960 	    ctrlr->page_size, 0, BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
961 	    ctrlr->hmb_chunk, 1, ctrlr->hmb_chunk, 0, NULL, NULL, &ctrlr->hmb_tag);
962 	if (err != 0) {
963 		nvme_printf(ctrlr, "HMB tag create failed %d\n", err);
964 		nvme_ctrlr_hmb_free(ctrlr);
965 		return;
966 	}
967 
968 	for (i = 0; i < ctrlr->hmb_nchunks; i++) {
969 		hmbc = &ctrlr->hmb_chunks[i];
970 		if (bus_dmamem_alloc(ctrlr->hmb_tag,
971 		    (void **)&hmbc->hmbc_vaddr, BUS_DMA_NOWAIT,
972 		    &hmbc->hmbc_map)) {
973 			nvme_printf(ctrlr, "failed to alloc HMB\n");
974 			break;
975 		}
976 		if (bus_dmamap_load(ctrlr->hmb_tag, hmbc->hmbc_map,
977 		    hmbc->hmbc_vaddr, ctrlr->hmb_chunk, nvme_single_map,
978 		    &hmbc->hmbc_paddr, BUS_DMA_NOWAIT) != 0) {
979 			bus_dmamem_free(ctrlr->hmb_tag, hmbc->hmbc_vaddr,
980 			    hmbc->hmbc_map);
981 			nvme_printf(ctrlr, "failed to load HMB\n");
982 			break;
983 		}
984 		bus_dmamap_sync(ctrlr->hmb_tag, hmbc->hmbc_map,
985 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
986 	}
987 
988 	if (i < ctrlr->hmb_nchunks && i * ctrlr->hmb_chunk < min &&
989 	    ctrlr->hmb_chunk / 2 >= minc) {
990 		ctrlr->hmb_nchunks = i;
991 		nvme_ctrlr_hmb_free(ctrlr);
992 		ctrlr->hmb_chunk /= 2;
993 		goto again;
994 	}
995 	ctrlr->hmb_nchunks = i;
996 	if (ctrlr->hmb_nchunks * ctrlr->hmb_chunk < min) {
997 		nvme_ctrlr_hmb_free(ctrlr);
998 		return;
999 	}
1000 
1001 	size = sizeof(struct nvme_hmb_desc) * ctrlr->hmb_nchunks;
1002 	err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev),
1003 	    16, 0, BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
1004 	    size, 1, size, 0, NULL, NULL, &ctrlr->hmb_desc_tag);
1005 	if (err != 0) {
1006 		nvme_printf(ctrlr, "HMB desc tag create failed %d\n", err);
1007 		nvme_ctrlr_hmb_free(ctrlr);
1008 		return;
1009 	}
1010 	if (bus_dmamem_alloc(ctrlr->hmb_desc_tag,
1011 	    (void **)&ctrlr->hmb_desc_vaddr, BUS_DMA_WAITOK,
1012 	    &ctrlr->hmb_desc_map)) {
1013 		nvme_printf(ctrlr, "failed to alloc HMB desc\n");
1014 		nvme_ctrlr_hmb_free(ctrlr);
1015 		return;
1016 	}
1017 	if (bus_dmamap_load(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_map,
1018 	    ctrlr->hmb_desc_vaddr, size, nvme_single_map,
1019 	    &ctrlr->hmb_desc_paddr, BUS_DMA_NOWAIT) != 0) {
1020 		bus_dmamem_free(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_vaddr,
1021 		    ctrlr->hmb_desc_map);
1022 		nvme_printf(ctrlr, "failed to load HMB desc\n");
1023 		nvme_ctrlr_hmb_free(ctrlr);
1024 		return;
1025 	}
1026 
1027 	for (i = 0; i < ctrlr->hmb_nchunks; i++) {
1028 		memset(&ctrlr->hmb_desc_vaddr[i], 0,
1029 		    sizeof(struct nvme_hmb_desc));
1030 		ctrlr->hmb_desc_vaddr[i].addr =
1031 		    htole64(ctrlr->hmb_chunks[i].hmbc_paddr);
1032 		ctrlr->hmb_desc_vaddr[i].size = htole32(ctrlr->hmb_chunk / ctrlr->page_size);
1033 	}
1034 	bus_dmamap_sync(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_map,
1035 	    BUS_DMASYNC_PREWRITE);
1036 
1037 	nvme_printf(ctrlr, "Allocated %lluMB host memory buffer\n",
1038 	    (long long unsigned)ctrlr->hmb_nchunks * ctrlr->hmb_chunk
1039 	    / 1024 / 1024);
1040 }
1041 
1042 static void
1043 nvme_ctrlr_hmb_enable(struct nvme_controller *ctrlr, bool enable, bool memret)
1044 {
1045 	struct nvme_completion_poll_status	status;
1046 	uint32_t cdw11;
1047 
1048 	cdw11 = 0;
1049 	if (enable)
1050 		cdw11 |= 1;
1051 	if (memret)
1052 		cdw11 |= 2;
1053 	status.done = 0;
1054 	nvme_ctrlr_cmd_set_feature(ctrlr, NVME_FEAT_HOST_MEMORY_BUFFER, cdw11,
1055 	    ctrlr->hmb_nchunks * ctrlr->hmb_chunk / ctrlr->page_size,
1056 	    ctrlr->hmb_desc_paddr, ctrlr->hmb_desc_paddr >> 32,
1057 	    ctrlr->hmb_nchunks, NULL, 0,
1058 	    nvme_completion_poll_cb, &status);
1059 	nvme_completion_poll(&status);
1060 	if (nvme_completion_is_error(&status.cpl))
1061 		nvme_printf(ctrlr, "nvme_ctrlr_hmb_enable failed!\n");
1062 }
1063 
1064 static void
1065 nvme_ctrlr_start(void *ctrlr_arg, bool resetting)
1066 {
1067 	struct nvme_controller *ctrlr = ctrlr_arg;
1068 	uint32_t old_num_io_queues;
1069 	int i;
1070 
1071 	TSENTER();
1072 
1073 	/*
1074 	 * Only reset adminq here when we are restarting the
1075 	 *  controller after a reset.  During initialization,
1076 	 *  we have already submitted admin commands to get
1077 	 *  the number of I/O queues supported, so cannot reset
1078 	 *  the adminq again here.
1079 	 */
1080 	if (resetting) {
1081 		nvme_qpair_reset(&ctrlr->adminq);
1082 		nvme_admin_qpair_enable(&ctrlr->adminq);
1083 	}
1084 
1085 	if (ctrlr->ioq != NULL) {
1086 		for (i = 0; i < ctrlr->num_io_queues; i++)
1087 			nvme_qpair_reset(&ctrlr->ioq[i]);
1088 	}
1089 
1090 	/*
1091 	 * If it was a reset on initialization command timeout, just
1092 	 * return here, letting initialization code fail gracefully.
1093 	 */
1094 	if (resetting && !ctrlr->is_initialized)
1095 		return;
1096 
1097 	if (resetting && nvme_ctrlr_identify(ctrlr) != 0) {
1098 		nvme_ctrlr_fail(ctrlr, false);
1099 		return;
1100 	}
1101 
1102 	/*
1103 	 * The number of qpairs are determined during controller initialization,
1104 	 *  including using NVMe SET_FEATURES/NUMBER_OF_QUEUES to determine the
1105 	 *  HW limit.  We call SET_FEATURES again here so that it gets called
1106 	 *  after any reset for controllers that depend on the driver to
1107 	 *  explicit specify how many queues it will use.  This value should
1108 	 *  never change between resets, so panic if somehow that does happen.
1109 	 */
1110 	if (resetting) {
1111 		old_num_io_queues = ctrlr->num_io_queues;
1112 		if (nvme_ctrlr_set_num_qpairs(ctrlr) != 0) {
1113 			nvme_ctrlr_fail(ctrlr, false);
1114 			return;
1115 		}
1116 
1117 		if (old_num_io_queues != ctrlr->num_io_queues) {
1118 			panic("num_io_queues changed from %u to %u",
1119 			      old_num_io_queues, ctrlr->num_io_queues);
1120 		}
1121 	}
1122 
1123 	if (ctrlr->cdata.hmpre > 0 && ctrlr->hmb_nchunks == 0) {
1124 		nvme_ctrlr_hmb_alloc(ctrlr);
1125 		if (ctrlr->hmb_nchunks > 0)
1126 			nvme_ctrlr_hmb_enable(ctrlr, true, false);
1127 	} else if (ctrlr->hmb_nchunks > 0)
1128 		nvme_ctrlr_hmb_enable(ctrlr, true, true);
1129 
1130 	if (nvme_ctrlr_create_qpairs(ctrlr) != 0) {
1131 		nvme_ctrlr_fail(ctrlr, false);
1132 		return;
1133 	}
1134 
1135 	if (nvme_ctrlr_construct_namespaces(ctrlr) != 0) {
1136 		nvme_ctrlr_fail(ctrlr, false);
1137 		return;
1138 	}
1139 
1140 	nvme_ctrlr_configure_aer(ctrlr);
1141 	nvme_ctrlr_configure_int_coalescing(ctrlr);
1142 
1143 	for (i = 0; i < ctrlr->num_io_queues; i++)
1144 		nvme_io_qpair_enable(&ctrlr->ioq[i]);
1145 	TSEXIT();
1146 }
1147 
1148 void
1149 nvme_ctrlr_start_config_hook(void *arg)
1150 {
1151 	struct nvme_controller *ctrlr = arg;
1152 
1153 	TSENTER();
1154 
1155 	if (nvme_ctrlr_hw_reset(ctrlr) != 0 || ctrlr->fail_on_reset != 0) {
1156 		nvme_ctrlr_fail(ctrlr, true);
1157 		config_intrhook_disestablish(&ctrlr->config_hook);
1158 		return;
1159 	}
1160 
1161 	nvme_qpair_reset(&ctrlr->adminq);
1162 	nvme_admin_qpair_enable(&ctrlr->adminq);
1163 
1164 	if (nvme_ctrlr_identify(ctrlr) == 0 &&
1165 	    nvme_ctrlr_set_num_qpairs(ctrlr) == 0 &&
1166 	    nvme_ctrlr_construct_io_qpairs(ctrlr) == 0)
1167 		nvme_ctrlr_start(ctrlr, false);
1168 	else
1169 		nvme_ctrlr_fail(ctrlr, false);
1170 
1171 	nvme_sysctl_initialize_ctrlr(ctrlr);
1172 	config_intrhook_disestablish(&ctrlr->config_hook);
1173 
1174 	if (!ctrlr->is_failed) {
1175 		ctrlr->is_initialized = true;
1176 		nvme_notify_new_controller(ctrlr);
1177 	}
1178 	TSEXIT();
1179 }
1180 
1181 static void
1182 nvme_ctrlr_reset_task(void *arg, int pending)
1183 {
1184 	struct nvme_controller	*ctrlr = arg;
1185 	int			status;
1186 
1187 	nvme_ctrlr_devctl_log(ctrlr, "RESET", "event=\"start\"");
1188 	status = nvme_ctrlr_hw_reset(ctrlr);
1189 	if (status == 0) {
1190 		nvme_ctrlr_devctl_log(ctrlr, "RESET", "event=\"success\"");
1191 		nvme_ctrlr_start(ctrlr, true);
1192 	} else {
1193 		nvme_ctrlr_devctl_log(ctrlr, "RESET", "event=\"timed_out\"");
1194 		nvme_ctrlr_fail(ctrlr, true);
1195 	}
1196 
1197 	atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
1198 }
1199 
1200 /*
1201  * Poll all the queues enabled on the device for completion.
1202  */
1203 void
1204 nvme_ctrlr_poll(struct nvme_controller *ctrlr)
1205 {
1206 	int i;
1207 
1208 	nvme_qpair_process_completions(&ctrlr->adminq);
1209 
1210 	for (i = 0; i < ctrlr->num_io_queues; i++)
1211 		if (ctrlr->ioq && ctrlr->ioq[i].cpl)
1212 			nvme_qpair_process_completions(&ctrlr->ioq[i]);
1213 }
1214 
1215 /*
1216  * Poll the single-vector interrupt case: num_io_queues will be 1 and
1217  * there's only a single vector. While we're polling, we mask further
1218  * interrupts in the controller.
1219  */
1220 void
1221 nvme_ctrlr_shared_handler(void *arg)
1222 {
1223 	struct nvme_controller *ctrlr = arg;
1224 
1225 	nvme_mmio_write_4(ctrlr, intms, 1);
1226 	nvme_ctrlr_poll(ctrlr);
1227 	nvme_mmio_write_4(ctrlr, intmc, 1);
1228 }
1229 
1230 static void
1231 nvme_pt_done(void *arg, const struct nvme_completion *cpl)
1232 {
1233 	struct nvme_pt_command *pt = arg;
1234 	struct mtx *mtx = pt->driver_lock;
1235 	uint16_t status;
1236 
1237 	bzero(&pt->cpl, sizeof(pt->cpl));
1238 	pt->cpl.cdw0 = cpl->cdw0;
1239 
1240 	status = cpl->status;
1241 	status &= ~NVMEM(NVME_STATUS_P);
1242 	pt->cpl.status = status;
1243 
1244 	mtx_lock(mtx);
1245 	pt->driver_lock = NULL;
1246 	wakeup(pt);
1247 	mtx_unlock(mtx);
1248 }
1249 
1250 int
1251 nvme_ctrlr_passthrough_cmd(struct nvme_controller *ctrlr,
1252     struct nvme_pt_command *pt, uint32_t nsid, int is_user_buffer,
1253     int is_admin_cmd)
1254 {
1255 	struct nvme_request	*req;
1256 	struct mtx		*mtx;
1257 	struct buf		*buf = NULL;
1258 	int			ret = 0;
1259 
1260 	if (pt->len > 0) {
1261 		if (pt->len > ctrlr->max_xfer_size) {
1262 			nvme_printf(ctrlr, "pt->len (%d) "
1263 			    "exceeds max_xfer_size (%d)\n", pt->len,
1264 			    ctrlr->max_xfer_size);
1265 			return EIO;
1266 		}
1267 		if (is_user_buffer) {
1268 			buf = uma_zalloc(pbuf_zone, M_WAITOK);
1269 			buf->b_iocmd = pt->is_read ? BIO_READ : BIO_WRITE;
1270 			if (vmapbuf(buf, pt->buf, pt->len, 1) < 0) {
1271 				ret = EFAULT;
1272 				goto err;
1273 			}
1274 			req = nvme_allocate_request_vaddr(buf->b_data, pt->len,
1275 			    nvme_pt_done, pt);
1276 		} else
1277 			req = nvme_allocate_request_vaddr(pt->buf, pt->len,
1278 			    nvme_pt_done, pt);
1279 	} else
1280 		req = nvme_allocate_request_null(nvme_pt_done, pt);
1281 
1282 	/* Assume user space already converted to little-endian */
1283 	req->cmd.opc = pt->cmd.opc;
1284 	req->cmd.fuse = pt->cmd.fuse;
1285 	req->cmd.rsvd2 = pt->cmd.rsvd2;
1286 	req->cmd.rsvd3 = pt->cmd.rsvd3;
1287 	req->cmd.cdw10 = pt->cmd.cdw10;
1288 	req->cmd.cdw11 = pt->cmd.cdw11;
1289 	req->cmd.cdw12 = pt->cmd.cdw12;
1290 	req->cmd.cdw13 = pt->cmd.cdw13;
1291 	req->cmd.cdw14 = pt->cmd.cdw14;
1292 	req->cmd.cdw15 = pt->cmd.cdw15;
1293 
1294 	req->cmd.nsid = htole32(nsid);
1295 
1296 	mtx = mtx_pool_find(mtxpool_sleep, pt);
1297 	pt->driver_lock = mtx;
1298 
1299 	if (is_admin_cmd)
1300 		nvme_ctrlr_submit_admin_request(ctrlr, req);
1301 	else
1302 		nvme_ctrlr_submit_io_request(ctrlr, req);
1303 
1304 	mtx_lock(mtx);
1305 	while (pt->driver_lock != NULL)
1306 		mtx_sleep(pt, mtx, PRIBIO, "nvme_pt", 0);
1307 	mtx_unlock(mtx);
1308 
1309 	if (buf != NULL) {
1310 		vunmapbuf(buf);
1311 err:
1312 		uma_zfree(pbuf_zone, buf);
1313 	}
1314 
1315 	return (ret);
1316 }
1317 
1318 static void
1319 nvme_npc_done(void *arg, const struct nvme_completion *cpl)
1320 {
1321 	struct nvme_passthru_cmd *npc = arg;
1322 	struct mtx *mtx = (void *)(uintptr_t)npc->metadata;
1323 
1324 	npc->result = cpl->cdw0;	/* cpl in host order by now */
1325 	mtx_lock(mtx);
1326 	npc->metadata = 0;
1327 	wakeup(npc);
1328 	mtx_unlock(mtx);
1329 }
1330 
1331 /* XXX refactor? */
1332 
1333 int
1334 nvme_ctrlr_linux_passthru_cmd(struct nvme_controller *ctrlr,
1335     struct nvme_passthru_cmd *npc, uint32_t nsid, bool is_user, bool is_admin)
1336 {
1337 	struct nvme_request	*req;
1338 	struct mtx		*mtx;
1339 	struct buf		*buf = NULL;
1340 	int			ret = 0;
1341 
1342 	/*
1343 	 * We don't support metadata.
1344 	 */
1345 	if (npc->metadata != 0 || npc->metadata_len != 0)
1346 		return (EIO);
1347 
1348 	if (npc->data_len > 0 && npc->addr != 0) {
1349 		if (npc->data_len > ctrlr->max_xfer_size) {
1350 			nvme_printf(ctrlr,
1351 			    "npc->data_len (%d) exceeds max_xfer_size (%d)\n",
1352 			    npc->data_len, ctrlr->max_xfer_size);
1353 			return (EIO);
1354 		}
1355 		/* We only support data out or data in commands, but not both at once. */
1356 		if ((npc->opcode & 0x3) == 0 || (npc->opcode & 0x3) == 3)
1357 			return (EINVAL);
1358 		if (is_user) {
1359 			buf = uma_zalloc(pbuf_zone, M_WAITOK);
1360 			buf->b_iocmd = npc->opcode & 1 ? BIO_WRITE : BIO_READ;
1361 			if (vmapbuf(buf, (void *)(uintptr_t)npc->addr,
1362 			    npc->data_len, 1) < 0) {
1363 				ret = EFAULT;
1364 				goto err;
1365 			}
1366 			req = nvme_allocate_request_vaddr(buf->b_data, npc->data_len,
1367 			    nvme_npc_done, npc);
1368 		} else
1369 			req = nvme_allocate_request_vaddr(
1370 			    (void *)(uintptr_t)npc->addr, npc->data_len,
1371 			    nvme_npc_done, npc);
1372 	} else
1373 		req = nvme_allocate_request_null(nvme_npc_done, npc);
1374 
1375 	req->cmd.opc = npc->opcode;
1376 	req->cmd.fuse = npc->flags;
1377 	req->cmd.rsvd2 = htole16(npc->cdw2);
1378 	req->cmd.rsvd3 = htole16(npc->cdw3);
1379 	req->cmd.cdw10 = htole32(npc->cdw10);
1380 	req->cmd.cdw11 = htole32(npc->cdw11);
1381 	req->cmd.cdw12 = htole32(npc->cdw12);
1382 	req->cmd.cdw13 = htole32(npc->cdw13);
1383 	req->cmd.cdw14 = htole32(npc->cdw14);
1384 	req->cmd.cdw15 = htole32(npc->cdw15);
1385 
1386 	req->cmd.nsid = htole32(nsid);
1387 
1388 	mtx = mtx_pool_find(mtxpool_sleep, npc);
1389 	npc->metadata = (uintptr_t) mtx;
1390 
1391 	/* XXX no timeout passed down */
1392 	if (is_admin)
1393 		nvme_ctrlr_submit_admin_request(ctrlr, req);
1394 	else
1395 		nvme_ctrlr_submit_io_request(ctrlr, req);
1396 
1397 	mtx_lock(mtx);
1398 	while (npc->metadata != 0)
1399 		mtx_sleep(npc, mtx, PRIBIO, "nvme_npc", 0);
1400 	mtx_unlock(mtx);
1401 
1402 	if (buf != NULL) {
1403 		vunmapbuf(buf);
1404 err:
1405 		uma_zfree(pbuf_zone, buf);
1406 	}
1407 
1408 	return (ret);
1409 }
1410 
1411 static int
1412 nvme_ctrlr_ioctl(struct cdev *cdev, u_long cmd, caddr_t arg, int flag,
1413     struct thread *td)
1414 {
1415 	struct nvme_controller			*ctrlr;
1416 	struct nvme_pt_command			*pt;
1417 
1418 	ctrlr = cdev->si_drv1;
1419 
1420 	switch (cmd) {
1421 	case NVME_IOCTL_RESET: /* Linux compat */
1422 	case NVME_RESET_CONTROLLER:
1423 		nvme_ctrlr_reset(ctrlr);
1424 		break;
1425 	case NVME_PASSTHROUGH_CMD:
1426 		pt = (struct nvme_pt_command *)arg;
1427 		return (nvme_ctrlr_passthrough_cmd(ctrlr, pt, le32toh(pt->cmd.nsid),
1428 		    1 /* is_user_buffer */, 1 /* is_admin_cmd */));
1429 	case NVME_GET_NSID:
1430 	{
1431 		struct nvme_get_nsid *gnsid = (struct nvme_get_nsid *)arg;
1432 		strlcpy(gnsid->cdev, device_get_nameunit(ctrlr->dev),
1433 		    sizeof(gnsid->cdev));
1434 		gnsid->nsid = 0;
1435 		break;
1436 	}
1437 	case NVME_GET_MAX_XFER_SIZE:
1438 		*(uint64_t *)arg = ctrlr->max_xfer_size;
1439 		break;
1440 	/* Linux Compatible (see nvme_linux.h) */
1441 	case NVME_IOCTL_ID:
1442 		td->td_retval[0] = 0xfffffffful;
1443 		return (0);
1444 
1445 	case NVME_IOCTL_ADMIN_CMD:
1446 	case NVME_IOCTL_IO_CMD: {
1447 		struct nvme_passthru_cmd *npc = (struct nvme_passthru_cmd *)arg;
1448 
1449 		return (nvme_ctrlr_linux_passthru_cmd(ctrlr, npc, npc->nsid, true,
1450 		    cmd == NVME_IOCTL_ADMIN_CMD));
1451 	}
1452 
1453 	default:
1454 		return (ENOTTY);
1455 	}
1456 
1457 	return (0);
1458 }
1459 
1460 static struct cdevsw nvme_ctrlr_cdevsw = {
1461 	.d_version =	D_VERSION,
1462 	.d_flags =	0,
1463 	.d_ioctl =	nvme_ctrlr_ioctl
1464 };
1465 
1466 int
1467 nvme_ctrlr_construct(struct nvme_controller *ctrlr, device_t dev)
1468 {
1469 	struct make_dev_args	md_args;
1470 	uint32_t	cap_lo;
1471 	uint32_t	cap_hi;
1472 	uint32_t	to, vs, pmrcap;
1473 	int		status, timeout_period;
1474 
1475 	ctrlr->dev = dev;
1476 
1477 	mtx_init(&ctrlr->lock, "nvme ctrlr lock", NULL, MTX_DEF);
1478 	if (bus_get_domain(dev, &ctrlr->domain) != 0)
1479 		ctrlr->domain = 0;
1480 
1481 	ctrlr->cap_lo = cap_lo = nvme_mmio_read_4(ctrlr, cap_lo);
1482 	if (bootverbose) {
1483 		device_printf(dev, "CapLo: 0x%08x: MQES %u%s%s%s%s, TO %u\n",
1484 		    cap_lo, NVME_CAP_LO_MQES(cap_lo),
1485 		    NVME_CAP_LO_CQR(cap_lo) ? ", CQR" : "",
1486 		    NVME_CAP_LO_AMS(cap_lo) ? ", AMS" : "",
1487 		    (NVME_CAP_LO_AMS(cap_lo) & 0x1) ? " WRRwUPC" : "",
1488 		    (NVME_CAP_LO_AMS(cap_lo) & 0x2) ? " VS" : "",
1489 		    NVME_CAP_LO_TO(cap_lo));
1490 	}
1491 	ctrlr->cap_hi = cap_hi = nvme_mmio_read_4(ctrlr, cap_hi);
1492 	if (bootverbose) {
1493 		device_printf(dev, "CapHi: 0x%08x: DSTRD %u%s, CSS %x%s, "
1494 		    "CPS %x, MPSMIN %u, MPSMAX %u%s%s%s%s%s\n", cap_hi,
1495 		    NVME_CAP_HI_DSTRD(cap_hi),
1496 		    NVME_CAP_HI_NSSRS(cap_hi) ? ", NSSRS" : "",
1497 		    NVME_CAP_HI_CSS(cap_hi),
1498 		    NVME_CAP_HI_BPS(cap_hi) ? ", BPS" : "",
1499 		    NVME_CAP_HI_CPS(cap_hi),
1500 		    NVME_CAP_HI_MPSMIN(cap_hi),
1501 		    NVME_CAP_HI_MPSMAX(cap_hi),
1502 		    NVME_CAP_HI_PMRS(cap_hi) ? ", PMRS" : "",
1503 		    NVME_CAP_HI_CMBS(cap_hi) ? ", CMBS" : "",
1504 		    NVME_CAP_HI_NSSS(cap_hi) ? ", NSSS" : "",
1505 		    NVME_CAP_HI_CRWMS(cap_hi) ? ", CRWMS" : "",
1506 		    NVME_CAP_HI_CRIMS(cap_hi) ? ", CRIMS" : "");
1507 	}
1508 	if (bootverbose) {
1509 		vs = nvme_mmio_read_4(ctrlr, vs);
1510 		device_printf(dev, "Version: 0x%08x: %d.%d\n", vs,
1511 		    NVME_MAJOR(vs), NVME_MINOR(vs));
1512 	}
1513 	if (bootverbose && NVME_CAP_HI_PMRS(cap_hi)) {
1514 		pmrcap = nvme_mmio_read_4(ctrlr, pmrcap);
1515 		device_printf(dev, "PMRCap: 0x%08x: BIR %u%s%s, PMRTU %u, "
1516 		    "PMRWBM %x, PMRTO %u%s\n", pmrcap,
1517 		    NVME_PMRCAP_BIR(pmrcap),
1518 		    NVME_PMRCAP_RDS(pmrcap) ? ", RDS" : "",
1519 		    NVME_PMRCAP_WDS(pmrcap) ? ", WDS" : "",
1520 		    NVME_PMRCAP_PMRTU(pmrcap),
1521 		    NVME_PMRCAP_PMRWBM(pmrcap),
1522 		    NVME_PMRCAP_PMRTO(pmrcap),
1523 		    NVME_PMRCAP_CMSS(pmrcap) ? ", CMSS" : "");
1524 	}
1525 
1526 	ctrlr->dstrd = NVME_CAP_HI_DSTRD(cap_hi) + 2;
1527 
1528 	ctrlr->mps = NVME_CAP_HI_MPSMIN(cap_hi);
1529 	ctrlr->page_size = 1 << (NVME_MPS_SHIFT + ctrlr->mps);
1530 
1531 	/* Get ready timeout value from controller, in units of 500ms. */
1532 	to = NVME_CAP_LO_TO(cap_lo) + 1;
1533 	ctrlr->ready_timeout_in_ms = to * 500;
1534 
1535 	timeout_period = NVME_ADMIN_TIMEOUT_PERIOD;
1536 	TUNABLE_INT_FETCH("hw.nvme.admin_timeout_period", &timeout_period);
1537 	timeout_period = min(timeout_period, NVME_MAX_TIMEOUT_PERIOD);
1538 	timeout_period = max(timeout_period, NVME_MIN_TIMEOUT_PERIOD);
1539 	ctrlr->admin_timeout_period = timeout_period;
1540 
1541 	timeout_period = NVME_DEFAULT_TIMEOUT_PERIOD;
1542 	TUNABLE_INT_FETCH("hw.nvme.timeout_period", &timeout_period);
1543 	timeout_period = min(timeout_period, NVME_MAX_TIMEOUT_PERIOD);
1544 	timeout_period = max(timeout_period, NVME_MIN_TIMEOUT_PERIOD);
1545 	ctrlr->timeout_period = timeout_period;
1546 
1547 	nvme_retry_count = NVME_DEFAULT_RETRY_COUNT;
1548 	TUNABLE_INT_FETCH("hw.nvme.retry_count", &nvme_retry_count);
1549 
1550 	ctrlr->enable_aborts = 0;
1551 	TUNABLE_INT_FETCH("hw.nvme.enable_aborts", &ctrlr->enable_aborts);
1552 
1553 	ctrlr->alignment_splits = counter_u64_alloc(M_WAITOK);
1554 
1555 	/* Cap transfers by the maximum addressable by page-sized PRP (4KB pages -> 2MB). */
1556 	ctrlr->max_xfer_size = MIN(maxphys, (ctrlr->page_size / 8 * ctrlr->page_size));
1557 	if (nvme_ctrlr_construct_admin_qpair(ctrlr) != 0)
1558 		return (ENXIO);
1559 
1560 	/*
1561 	 * Create 2 threads for the taskqueue. The reset thread will block when
1562 	 * it detects that the controller has failed until all I/O has been
1563 	 * failed up the stack. The fail_req task needs to be able to run in
1564 	 * this case to finish the request failure for some cases.
1565 	 *
1566 	 * We could partially solve this race by draining the failed requeust
1567 	 * queue before proceding to free the sim, though nothing would stop
1568 	 * new I/O from coming in after we do that drain, but before we reach
1569 	 * cam_sim_free, so this big hammer is used instead.
1570 	 */
1571 	ctrlr->taskqueue = taskqueue_create("nvme_taskq", M_WAITOK,
1572 	    taskqueue_thread_enqueue, &ctrlr->taskqueue);
1573 	taskqueue_start_threads(&ctrlr->taskqueue, 2, PI_DISK, "nvme taskq");
1574 
1575 	ctrlr->is_resetting = 0;
1576 	ctrlr->is_initialized = false;
1577 	ctrlr->notification_sent = 0;
1578 	TASK_INIT(&ctrlr->reset_task, 0, nvme_ctrlr_reset_task, ctrlr);
1579 	STAILQ_INIT(&ctrlr->fail_req);
1580 	ctrlr->is_failed = false;
1581 
1582 	make_dev_args_init(&md_args);
1583 	md_args.mda_devsw = &nvme_ctrlr_cdevsw;
1584 	md_args.mda_uid = UID_ROOT;
1585 	md_args.mda_gid = GID_WHEEL;
1586 	md_args.mda_mode = 0600;
1587 	md_args.mda_unit = device_get_unit(dev);
1588 	md_args.mda_si_drv1 = (void *)ctrlr;
1589 	status = make_dev_s(&md_args, &ctrlr->cdev, "%s",
1590 	    device_get_nameunit(dev));
1591 	if (status != 0)
1592 		return (ENXIO);
1593 
1594 	return (0);
1595 }
1596 
1597 /*
1598  * Called on detach, or on error on attach. The nvme_controller won't be used
1599  * again once we return, so we have to tear everything down (so nothing
1600  * references this, no callbacks, etc), but don't need to reset all the state
1601  * since nvme_controller will be freed soon.
1602  */
1603 void
1604 nvme_ctrlr_destruct(struct nvme_controller *ctrlr, device_t dev)
1605 {
1606 	int	gone, i;
1607 
1608 	ctrlr->is_dying = true;
1609 
1610 	if (ctrlr->resource == NULL)
1611 		goto nores;
1612 	if (!mtx_initialized(&ctrlr->adminq.lock))
1613 		goto noadminq;
1614 
1615 	/*
1616 	 * Check whether it is a hot unplug or a clean driver detach.
1617 	 * If device is not there any more, skip any shutdown commands.
1618 	 */
1619 	gone = (nvme_mmio_read_4(ctrlr, csts) == NVME_GONE);
1620 	if (gone)
1621 		nvme_ctrlr_fail(ctrlr, true);
1622 	else
1623 		nvme_notify_fail_consumers(ctrlr);
1624 
1625 	for (i = 0; i < NVME_MAX_NAMESPACES; i++)
1626 		nvme_ns_destruct(&ctrlr->ns[i]);
1627 
1628 	if (ctrlr->cdev)
1629 		destroy_dev(ctrlr->cdev);
1630 
1631 	if (ctrlr->is_initialized) {
1632 		if (!gone) {
1633 			if (ctrlr->hmb_nchunks > 0)
1634 				nvme_ctrlr_hmb_enable(ctrlr, false, false);
1635 			nvme_ctrlr_delete_qpairs(ctrlr);
1636 		}
1637 		nvme_ctrlr_hmb_free(ctrlr);
1638 	}
1639 	if (ctrlr->ioq != NULL) {
1640 		for (i = 0; i < ctrlr->num_io_queues; i++)
1641 			nvme_io_qpair_destroy(&ctrlr->ioq[i]);
1642 		free(ctrlr->ioq, M_NVME);
1643 	}
1644 	nvme_admin_qpair_destroy(&ctrlr->adminq);
1645 
1646 	/*
1647 	 *  Notify the controller of a shutdown, even though this is due to
1648 	 *   a driver unload, not a system shutdown (this path is not invoked
1649 	 *   during shutdown).  This ensures the controller receives a
1650 	 *   shutdown notification in case the system is shutdown before
1651 	 *   reloading the driver.
1652 	 */
1653 	if (!gone)
1654 		nvme_ctrlr_shutdown(ctrlr);
1655 
1656 	if (!gone)
1657 		nvme_ctrlr_disable(ctrlr);
1658 
1659 noadminq:
1660 	if (ctrlr->taskqueue)
1661 		taskqueue_free(ctrlr->taskqueue);
1662 
1663 	if (ctrlr->tag)
1664 		bus_teardown_intr(ctrlr->dev, ctrlr->res, ctrlr->tag);
1665 
1666 	if (ctrlr->res)
1667 		bus_release_resource(ctrlr->dev, SYS_RES_IRQ,
1668 		    rman_get_rid(ctrlr->res), ctrlr->res);
1669 
1670 	if (ctrlr->bar4_resource != NULL) {
1671 		bus_release_resource(dev, SYS_RES_MEMORY,
1672 		    ctrlr->bar4_resource_id, ctrlr->bar4_resource);
1673 	}
1674 
1675 	bus_release_resource(dev, SYS_RES_MEMORY,
1676 	    ctrlr->resource_id, ctrlr->resource);
1677 
1678 nores:
1679 	if (ctrlr->alignment_splits)
1680 		counter_u64_free(ctrlr->alignment_splits);
1681 
1682 	mtx_destroy(&ctrlr->lock);
1683 }
1684 
1685 void
1686 nvme_ctrlr_shutdown(struct nvme_controller *ctrlr)
1687 {
1688 	uint32_t	cc;
1689 	uint32_t	csts;
1690 	int		timeout;
1691 
1692 	cc = nvme_mmio_read_4(ctrlr, cc);
1693 	cc &= ~NVMEM(NVME_CC_REG_SHN);
1694 	cc |= NVMEF(NVME_CC_REG_SHN, NVME_SHN_NORMAL);
1695 	nvme_mmio_write_4(ctrlr, cc, cc);
1696 
1697 	timeout = ticks + (ctrlr->cdata.rtd3e == 0 ? 5 * hz :
1698 	    ((uint64_t)ctrlr->cdata.rtd3e * hz + 999999) / 1000000);
1699 	while (1) {
1700 		csts = nvme_mmio_read_4(ctrlr, csts);
1701 		if (csts == NVME_GONE)		/* Hot unplug. */
1702 			break;
1703 		if (NVME_CSTS_GET_SHST(csts) == NVME_SHST_COMPLETE)
1704 			break;
1705 		if (timeout - ticks < 0) {
1706 			nvme_printf(ctrlr, "shutdown timeout\n");
1707 			break;
1708 		}
1709 		pause("nvmeshut", 1);
1710 	}
1711 }
1712 
1713 void
1714 nvme_ctrlr_submit_admin_request(struct nvme_controller *ctrlr,
1715     struct nvme_request *req)
1716 {
1717 
1718 	nvme_qpair_submit_request(&ctrlr->adminq, req);
1719 }
1720 
1721 void
1722 nvme_ctrlr_submit_io_request(struct nvme_controller *ctrlr,
1723     struct nvme_request *req)
1724 {
1725 	struct nvme_qpair       *qpair;
1726 
1727 	qpair = &ctrlr->ioq[QP(ctrlr, curcpu)];
1728 	nvme_qpair_submit_request(qpair, req);
1729 }
1730 
1731 device_t
1732 nvme_ctrlr_get_device(struct nvme_controller *ctrlr)
1733 {
1734 
1735 	return (ctrlr->dev);
1736 }
1737 
1738 const struct nvme_controller_data *
1739 nvme_ctrlr_get_data(struct nvme_controller *ctrlr)
1740 {
1741 
1742 	return (&ctrlr->cdata);
1743 }
1744 
1745 int
1746 nvme_ctrlr_suspend(struct nvme_controller *ctrlr)
1747 {
1748 	int to = hz;
1749 
1750 	/*
1751 	 * Can't touch failed controllers, so it's already suspended. User will
1752 	 * need to do an explicit reset to bring it back, if that's even
1753 	 * possible.
1754 	 */
1755 	if (ctrlr->is_failed)
1756 		return (0);
1757 
1758 	/*
1759 	 * We don't want the reset taskqueue running, since it does similar
1760 	 * things, so prevent it from running after we start. Wait for any reset
1761 	 * that may have been started to complete. The reset process we follow
1762 	 * will ensure that any new I/O will queue and be given to the hardware
1763 	 * after we resume (though there should be none).
1764 	 */
1765 	while (atomic_cmpset_32(&ctrlr->is_resetting, 0, 1) == 0 && to-- > 0)
1766 		pause("nvmesusp", 1);
1767 	if (to <= 0) {
1768 		nvme_printf(ctrlr,
1769 		    "Competing reset task didn't finish. Try again later.\n");
1770 		return (EWOULDBLOCK);
1771 	}
1772 
1773 	if (ctrlr->hmb_nchunks > 0)
1774 		nvme_ctrlr_hmb_enable(ctrlr, false, false);
1775 
1776 	/*
1777 	 * Per Section 7.6.2 of NVMe spec 1.4, to properly suspend, we need to
1778 	 * delete the hardware I/O queues, and then shutdown. This properly
1779 	 * flushes any metadata the drive may have stored so it can survive
1780 	 * having its power removed and prevents the unsafe shutdown count from
1781 	 * incriminating. Once we delete the qpairs, we have to disable them
1782 	 * before shutting down.
1783 	 */
1784 	nvme_ctrlr_delete_qpairs(ctrlr);
1785 	nvme_ctrlr_disable_qpairs(ctrlr);
1786 	nvme_ctrlr_shutdown(ctrlr);
1787 
1788 	return (0);
1789 }
1790 
1791 int
1792 nvme_ctrlr_resume(struct nvme_controller *ctrlr)
1793 {
1794 
1795 	/*
1796 	 * Can't touch failed controllers, so nothing to do to resume.
1797 	 */
1798 	if (ctrlr->is_failed)
1799 		return (0);
1800 
1801 	if (nvme_ctrlr_hw_reset(ctrlr) != 0)
1802 		goto fail;
1803 
1804 	/*
1805 	 * Now that we've reset the hardware, we can restart the controller. Any
1806 	 * I/O that was pending is requeued. Any admin commands are aborted with
1807 	 * an error. Once we've restarted, stop flagging the controller as being
1808 	 * in the reset phase.
1809 	 */
1810 	nvme_ctrlr_start(ctrlr, true);
1811 	(void)atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
1812 
1813 	return (0);
1814 fail:
1815 	/*
1816 	 * Since we can't bring the controller out of reset, announce and fail
1817 	 * the controller. However, we have to return success for the resume
1818 	 * itself, due to questionable APIs.
1819 	 */
1820 	nvme_printf(ctrlr, "Failed to reset on resume, failing.\n");
1821 	nvme_ctrlr_fail(ctrlr, true);
1822 	(void)atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
1823 	return (0);
1824 }
1825