xref: /freebsd/sys/dev/nvme/nvme_ctrlr.c (revision 023fa553bf0000837738ce3ba118921ee24e3792)
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_log(struct nvme_controller *ctrlr, const char *type, const char *msg, ...)
60 {
61 	struct sbuf sb;
62 	va_list ap;
63 	int error;
64 
65 	if (sbuf_new(&sb, NULL, 0, SBUF_AUTOEXTEND | SBUF_NOWAIT) == NULL)
66 		return;
67 	sbuf_printf(&sb, "%s: ", device_get_nameunit(ctrlr->dev));
68 	va_start(ap, msg);
69 	sbuf_vprintf(&sb, msg, ap);
70 	va_end(ap);
71 	error = sbuf_finish(&sb);
72 	if (error == 0)
73 		printf("%s\n", sbuf_data(&sb));
74 
75 	sbuf_clear(&sb);
76 	sbuf_printf(&sb, "name=\"%s\" reason=\"", device_get_nameunit(ctrlr->dev));
77 	va_start(ap, msg);
78 	sbuf_vprintf(&sb, msg, ap);
79 	va_end(ap);
80 	sbuf_printf(&sb, "\"");
81 	error = sbuf_finish(&sb);
82 	if (error == 0)
83 		devctl_notify("nvme", "controller", type, sbuf_data(&sb));
84 	sbuf_delete(&sb);
85 }
86 
87 static int
88 nvme_ctrlr_construct_admin_qpair(struct nvme_controller *ctrlr)
89 {
90 	struct nvme_qpair	*qpair;
91 	uint32_t		num_entries;
92 	int			error;
93 
94 	qpair = &ctrlr->adminq;
95 	qpair->id = 0;
96 	qpair->cpu = CPU_FFS(&cpuset_domain[ctrlr->domain]) - 1;
97 	qpair->domain = ctrlr->domain;
98 
99 	num_entries = NVME_ADMIN_ENTRIES;
100 	TUNABLE_INT_FETCH("hw.nvme.admin_entries", &num_entries);
101 	/*
102 	 * If admin_entries was overridden to an invalid value, revert it
103 	 *  back to our default value.
104 	 */
105 	if (num_entries < NVME_MIN_ADMIN_ENTRIES ||
106 	    num_entries > NVME_MAX_ADMIN_ENTRIES) {
107 		nvme_printf(ctrlr, "invalid hw.nvme.admin_entries=%d "
108 		    "specified\n", num_entries);
109 		num_entries = NVME_ADMIN_ENTRIES;
110 	}
111 
112 	/*
113 	 * The admin queue's max xfer size is treated differently than the
114 	 *  max I/O xfer size.  16KB is sufficient here - maybe even less?
115 	 */
116 	error = nvme_qpair_construct(qpair, num_entries, NVME_ADMIN_TRACKERS,
117 	     ctrlr);
118 	return (error);
119 }
120 
121 #define QP(ctrlr, c)	((c) * (ctrlr)->num_io_queues / mp_ncpus)
122 
123 static int
124 nvme_ctrlr_construct_io_qpairs(struct nvme_controller *ctrlr)
125 {
126 	struct nvme_qpair	*qpair;
127 	uint32_t		cap_lo;
128 	uint16_t		mqes;
129 	int			c, error, i, n;
130 	int			num_entries, num_trackers, max_entries;
131 
132 	/*
133 	 * NVMe spec sets a hard limit of 64K max entries, but devices may
134 	 * specify a smaller limit, so we need to check the MQES field in the
135 	 * capabilities register. We have to cap the number of entries to the
136 	 * current stride allows for in BAR 0/1, otherwise the remainder entries
137 	 * are inaccessible. MQES should reflect this, and this is just a
138 	 * fail-safe.
139 	 */
140 	max_entries =
141 	    (rman_get_size(ctrlr->resource) - nvme_mmio_offsetof(doorbell[0])) /
142 	    (1 << (ctrlr->dstrd + 1));
143 	num_entries = NVME_IO_ENTRIES;
144 	TUNABLE_INT_FETCH("hw.nvme.io_entries", &num_entries);
145 	cap_lo = nvme_mmio_read_4(ctrlr, cap_lo);
146 	mqes = NVME_CAP_LO_MQES(cap_lo);
147 	num_entries = min(num_entries, mqes + 1);
148 	num_entries = min(num_entries, max_entries);
149 
150 	num_trackers = NVME_IO_TRACKERS;
151 	TUNABLE_INT_FETCH("hw.nvme.io_trackers", &num_trackers);
152 
153 	num_trackers = max(num_trackers, NVME_MIN_IO_TRACKERS);
154 	num_trackers = min(num_trackers, NVME_MAX_IO_TRACKERS);
155 	/*
156 	 * No need to have more trackers than entries in the submit queue.  Note
157 	 * also that for a queue size of N, we can only have (N-1) commands
158 	 * outstanding, hence the "-1" here.
159 	 */
160 	num_trackers = min(num_trackers, (num_entries-1));
161 
162 	/*
163 	 * Our best estimate for the maximum number of I/Os that we should
164 	 * normally have in flight at one time. This should be viewed as a hint,
165 	 * not a hard limit and will need to be revisited when the upper layers
166 	 * of the storage system grows multi-queue support.
167 	 */
168 	ctrlr->max_hw_pend_io = num_trackers * ctrlr->num_io_queues * 3 / 4;
169 
170 	ctrlr->ioq = malloc(ctrlr->num_io_queues * sizeof(struct nvme_qpair),
171 	    M_NVME, M_ZERO | M_WAITOK);
172 
173 	for (i = c = n = 0; i < ctrlr->num_io_queues; i++, c += n) {
174 		qpair = &ctrlr->ioq[i];
175 
176 		/*
177 		 * Admin queue has ID=0. IO queues start at ID=1 -
178 		 *  hence the 'i+1' here.
179 		 */
180 		qpair->id = i + 1;
181 		if (ctrlr->num_io_queues > 1) {
182 			/* Find number of CPUs served by this queue. */
183 			for (n = 1; QP(ctrlr, c + n) == i; n++)
184 				;
185 			/* Shuffle multiple NVMe devices between CPUs. */
186 			qpair->cpu = c + (device_get_unit(ctrlr->dev)+n/2) % n;
187 			qpair->domain = pcpu_find(qpair->cpu)->pc_domain;
188 		} else {
189 			qpair->cpu = CPU_FFS(&cpuset_domain[ctrlr->domain]) - 1;
190 			qpair->domain = ctrlr->domain;
191 		}
192 
193 		/*
194 		 * For I/O queues, use the controller-wide max_xfer_size
195 		 *  calculated in nvme_attach().
196 		 */
197 		error = nvme_qpair_construct(qpair, num_entries, num_trackers,
198 		    ctrlr);
199 		if (error)
200 			return (error);
201 
202 		/*
203 		 * Do not bother binding interrupts if we only have one I/O
204 		 *  interrupt thread for this controller.
205 		 */
206 		if (ctrlr->num_io_queues > 1)
207 			bus_bind_intr(ctrlr->dev, qpair->res, qpair->cpu);
208 	}
209 
210 	return (0);
211 }
212 
213 static void
214 nvme_ctrlr_fail(struct nvme_controller *ctrlr)
215 {
216 	int i;
217 
218 	/*
219 	 * No need to disable queues before failing them. Failing is a superet
220 	 * of disabling (though pedantically we'd abort the AERs silently with
221 	 * a different error, though when we fail, that hardly matters).
222 	 */
223 	ctrlr->is_failed = true;
224 	nvme_qpair_fail(&ctrlr->adminq);
225 	if (ctrlr->ioq != NULL) {
226 		for (i = 0; i < ctrlr->num_io_queues; i++) {
227 			nvme_qpair_fail(&ctrlr->ioq[i]);
228 		}
229 	}
230 	nvme_notify_fail_consumers(ctrlr);
231 }
232 
233 /*
234  * Wait for RDY to change.
235  *
236  * Starts sleeping for 1us and geometrically increases it the longer we wait,
237  * capped at 1ms.
238  */
239 static int
240 nvme_ctrlr_wait_for_ready(struct nvme_controller *ctrlr, int desired_val)
241 {
242 	int timeout = ticks + MSEC_2_TICKS(ctrlr->ready_timeout_in_ms);
243 	sbintime_t delta_t = SBT_1US;
244 	uint32_t csts;
245 
246 	while (1) {
247 		csts = nvme_mmio_read_4(ctrlr, csts);
248 		if (csts == NVME_GONE)		/* Hot unplug. */
249 			return (ENXIO);
250 		if (NVMEV(NVME_CSTS_REG_RDY, csts) == desired_val)
251 			break;
252 		if (timeout - ticks < 0) {
253 			nvme_printf(ctrlr, "controller ready did not become %d "
254 			    "within %d ms\n", desired_val, ctrlr->ready_timeout_in_ms);
255 			return (ENXIO);
256 		}
257 
258 		pause_sbt("nvmerdy", delta_t, 0, C_PREL(1));
259 		delta_t = min(SBT_1MS, delta_t * 3 / 2);
260 	}
261 
262 	return (0);
263 }
264 
265 static int
266 nvme_ctrlr_disable(struct nvme_controller *ctrlr)
267 {
268 	uint32_t cc;
269 	uint32_t csts;
270 	uint8_t  en, rdy;
271 	int err;
272 
273 	cc = nvme_mmio_read_4(ctrlr, cc);
274 	csts = nvme_mmio_read_4(ctrlr, csts);
275 
276 	en = NVMEV(NVME_CC_REG_EN, cc);
277 	rdy = NVMEV(NVME_CSTS_REG_RDY, csts);
278 
279 	/*
280 	 * Per 3.1.5 in NVME 1.3 spec, transitioning CC.EN from 0 to 1
281 	 * when CSTS.RDY is 1 or transitioning CC.EN from 1 to 0 when
282 	 * CSTS.RDY is 0 "has undefined results" So make sure that CSTS.RDY
283 	 * isn't the desired value. Short circuit if we're already disabled.
284 	 */
285 	if (en == 0) {
286 		/* Wait for RDY == 0 or timeout & fail */
287 		if (rdy == 0)
288 			return (0);
289 		return (nvme_ctrlr_wait_for_ready(ctrlr, 0));
290 	}
291 	if (rdy == 0) {
292 		/* EN == 1, wait for  RDY == 1 or timeout & fail */
293 		err = nvme_ctrlr_wait_for_ready(ctrlr, 1);
294 		if (err != 0)
295 			return (err);
296 	}
297 
298 	cc &= ~NVMEM(NVME_CC_REG_EN);
299 	nvme_mmio_write_4(ctrlr, cc, cc);
300 
301 	/*
302 	 * A few drives have firmware bugs that freeze the drive if we access
303 	 * the mmio too soon after we disable.
304 	 */
305 	if (ctrlr->quirks & QUIRK_DELAY_B4_CHK_RDY)
306 		pause("nvmeR", MSEC_2_TICKS(B4_CHK_RDY_DELAY_MS));
307 	return (nvme_ctrlr_wait_for_ready(ctrlr, 0));
308 }
309 
310 static int
311 nvme_ctrlr_enable(struct nvme_controller *ctrlr)
312 {
313 	uint32_t	cc;
314 	uint32_t	csts;
315 	uint32_t	aqa;
316 	uint32_t	qsize;
317 	uint8_t		en, rdy;
318 	int		err;
319 
320 	cc = nvme_mmio_read_4(ctrlr, cc);
321 	csts = nvme_mmio_read_4(ctrlr, csts);
322 
323 	en = NVMEV(NVME_CC_REG_EN, cc);
324 	rdy = NVMEV(NVME_CSTS_REG_RDY, csts);
325 
326 	/*
327 	 * See note in nvme_ctrlr_disable. Short circuit if we're already enabled.
328 	 */
329 	if (en == 1) {
330 		if (rdy == 1)
331 			return (0);
332 		return (nvme_ctrlr_wait_for_ready(ctrlr, 1));
333 	}
334 
335 	/* EN == 0 already wait for RDY == 0 or timeout & fail */
336 	err = nvme_ctrlr_wait_for_ready(ctrlr, 0);
337 	if (err != 0)
338 		return (err);
339 
340 	nvme_mmio_write_8(ctrlr, asq, ctrlr->adminq.cmd_bus_addr);
341 	nvme_mmio_write_8(ctrlr, acq, ctrlr->adminq.cpl_bus_addr);
342 
343 	/* acqs and asqs are 0-based. */
344 	qsize = ctrlr->adminq.num_entries - 1;
345 
346 	aqa = 0;
347 	aqa |= NVMEF(NVME_AQA_REG_ACQS, qsize);
348 	aqa |= NVMEF(NVME_AQA_REG_ASQS, qsize);
349 	nvme_mmio_write_4(ctrlr, aqa, aqa);
350 
351 	/* Initialization values for CC */
352 	cc = 0;
353 	cc |= NVMEF(NVME_CC_REG_EN, 1);
354 	cc |= NVMEF(NVME_CC_REG_CSS, 0);
355 	cc |= NVMEF(NVME_CC_REG_AMS, 0);
356 	cc |= NVMEF(NVME_CC_REG_SHN, 0);
357 	cc |= NVMEF(NVME_CC_REG_IOSQES, 6); /* SQ entry size == 64 == 2^6 */
358 	cc |= NVMEF(NVME_CC_REG_IOCQES, 4); /* CQ entry size == 16 == 2^4 */
359 
360 	/*
361 	 * Use the Memory Page Size selected during device initialization.  Note
362 	 * that value stored in mps is suitable to use here without adjusting by
363 	 * NVME_MPS_SHIFT.
364 	 */
365 	cc |= NVMEF(NVME_CC_REG_MPS, ctrlr->mps);
366 
367 	nvme_ctrlr_barrier(ctrlr, BUS_SPACE_BARRIER_WRITE);
368 	nvme_mmio_write_4(ctrlr, cc, cc);
369 
370 	return (nvme_ctrlr_wait_for_ready(ctrlr, 1));
371 }
372 
373 static void
374 nvme_ctrlr_disable_qpairs(struct nvme_controller *ctrlr)
375 {
376 	int i;
377 
378 	nvme_admin_qpair_disable(&ctrlr->adminq);
379 	/*
380 	 * I/O queues are not allocated before the initial HW
381 	 *  reset, so do not try to disable them.  Use is_initialized
382 	 *  to determine if this is the initial HW reset.
383 	 */
384 	if (ctrlr->is_initialized) {
385 		for (i = 0; i < ctrlr->num_io_queues; i++)
386 			nvme_io_qpair_disable(&ctrlr->ioq[i]);
387 	}
388 }
389 
390 static int
391 nvme_ctrlr_hw_reset(struct nvme_controller *ctrlr)
392 {
393 	int err;
394 
395 	TSENTER();
396 
397 	nvme_ctrlr_disable_qpairs(ctrlr);
398 
399 	err = nvme_ctrlr_disable(ctrlr);
400 	if (err != 0)
401 		goto out;
402 
403 	err = nvme_ctrlr_enable(ctrlr);
404 out:
405 
406 	TSEXIT();
407 	return (err);
408 }
409 
410 void
411 nvme_ctrlr_reset(struct nvme_controller *ctrlr)
412 {
413 	int cmpset;
414 
415 	cmpset = atomic_cmpset_32(&ctrlr->is_resetting, 0, 1);
416 
417 	if (cmpset == 0 || ctrlr->is_failed)
418 		/*
419 		 * Controller is already resetting or has failed.  Return
420 		 *  immediately since there is no need to kick off another
421 		 *  reset in these cases.
422 		 */
423 		return;
424 
425 	if (!ctrlr->is_dying)
426 		taskqueue_enqueue(ctrlr->taskqueue, &ctrlr->reset_task);
427 }
428 
429 static int
430 nvme_ctrlr_identify(struct nvme_controller *ctrlr)
431 {
432 	struct nvme_completion_poll_status	status;
433 
434 	status.done = 0;
435 	nvme_ctrlr_cmd_identify_controller(ctrlr, &ctrlr->cdata,
436 	    nvme_completion_poll_cb, &status);
437 	nvme_completion_poll(&status);
438 	if (nvme_completion_is_error(&status.cpl)) {
439 		nvme_printf(ctrlr, "nvme_identify_controller failed!\n");
440 		return (ENXIO);
441 	}
442 
443 	/* Convert data to host endian */
444 	nvme_controller_data_swapbytes(&ctrlr->cdata);
445 
446 	/*
447 	 * Use MDTS to ensure our default max_xfer_size doesn't exceed what the
448 	 *  controller supports.
449 	 */
450 	if (ctrlr->cdata.mdts > 0)
451 		ctrlr->max_xfer_size = min(ctrlr->max_xfer_size,
452 		    1 << (ctrlr->cdata.mdts + NVME_MPS_SHIFT +
453 			NVME_CAP_HI_MPSMIN(ctrlr->cap_hi)));
454 
455 	return (0);
456 }
457 
458 static int
459 nvme_ctrlr_set_num_qpairs(struct nvme_controller *ctrlr)
460 {
461 	struct nvme_completion_poll_status	status;
462 	int					cq_allocated, sq_allocated;
463 
464 	status.done = 0;
465 	nvme_ctrlr_cmd_set_num_queues(ctrlr, ctrlr->num_io_queues,
466 	    nvme_completion_poll_cb, &status);
467 	nvme_completion_poll(&status);
468 	if (nvme_completion_is_error(&status.cpl)) {
469 		nvme_printf(ctrlr, "nvme_ctrlr_set_num_qpairs failed!\n");
470 		return (ENXIO);
471 	}
472 
473 	/*
474 	 * Data in cdw0 is 0-based.
475 	 * Lower 16-bits indicate number of submission queues allocated.
476 	 * Upper 16-bits indicate number of completion queues allocated.
477 	 */
478 	sq_allocated = (status.cpl.cdw0 & 0xFFFF) + 1;
479 	cq_allocated = (status.cpl.cdw0 >> 16) + 1;
480 
481 	/*
482 	 * Controller may allocate more queues than we requested,
483 	 *  so use the minimum of the number requested and what was
484 	 *  actually allocated.
485 	 */
486 	ctrlr->num_io_queues = min(ctrlr->num_io_queues, sq_allocated);
487 	ctrlr->num_io_queues = min(ctrlr->num_io_queues, cq_allocated);
488 	if (ctrlr->num_io_queues > vm_ndomains)
489 		ctrlr->num_io_queues -= ctrlr->num_io_queues % vm_ndomains;
490 
491 	return (0);
492 }
493 
494 static int
495 nvme_ctrlr_create_qpairs(struct nvme_controller *ctrlr)
496 {
497 	struct nvme_completion_poll_status	status;
498 	struct nvme_qpair			*qpair;
499 	int					i;
500 
501 	for (i = 0; i < ctrlr->num_io_queues; i++) {
502 		qpair = &ctrlr->ioq[i];
503 
504 		status.done = 0;
505 		nvme_ctrlr_cmd_create_io_cq(ctrlr, qpair,
506 		    nvme_completion_poll_cb, &status);
507 		nvme_completion_poll(&status);
508 		if (nvme_completion_is_error(&status.cpl)) {
509 			nvme_printf(ctrlr, "nvme_create_io_cq failed!\n");
510 			return (ENXIO);
511 		}
512 
513 		status.done = 0;
514 		nvme_ctrlr_cmd_create_io_sq(ctrlr, qpair,
515 		    nvme_completion_poll_cb, &status);
516 		nvme_completion_poll(&status);
517 		if (nvme_completion_is_error(&status.cpl)) {
518 			nvme_printf(ctrlr, "nvme_create_io_sq failed!\n");
519 			return (ENXIO);
520 		}
521 	}
522 
523 	return (0);
524 }
525 
526 static int
527 nvme_ctrlr_delete_qpairs(struct nvme_controller *ctrlr)
528 {
529 	struct nvme_completion_poll_status	status;
530 	struct nvme_qpair			*qpair;
531 
532 	for (int i = 0; i < ctrlr->num_io_queues; i++) {
533 		qpair = &ctrlr->ioq[i];
534 
535 		status.done = 0;
536 		nvme_ctrlr_cmd_delete_io_sq(ctrlr, qpair,
537 		    nvme_completion_poll_cb, &status);
538 		nvme_completion_poll(&status);
539 		if (nvme_completion_is_error(&status.cpl)) {
540 			nvme_printf(ctrlr, "nvme_destroy_io_sq failed!\n");
541 			return (ENXIO);
542 		}
543 
544 		status.done = 0;
545 		nvme_ctrlr_cmd_delete_io_cq(ctrlr, qpair,
546 		    nvme_completion_poll_cb, &status);
547 		nvme_completion_poll(&status);
548 		if (nvme_completion_is_error(&status.cpl)) {
549 			nvme_printf(ctrlr, "nvme_destroy_io_cq failed!\n");
550 			return (ENXIO);
551 		}
552 	}
553 
554 	return (0);
555 }
556 
557 static int
558 nvme_ctrlr_construct_namespaces(struct nvme_controller *ctrlr)
559 {
560 	struct nvme_namespace	*ns;
561 	uint32_t 		i;
562 
563 	for (i = 0; i < min(ctrlr->cdata.nn, NVME_MAX_NAMESPACES); i++) {
564 		ns = &ctrlr->ns[i];
565 		nvme_ns_construct(ns, i+1, ctrlr);
566 	}
567 
568 	return (0);
569 }
570 
571 static bool
572 is_log_page_id_valid(uint8_t page_id)
573 {
574 
575 	switch (page_id) {
576 	case NVME_LOG_ERROR:
577 	case NVME_LOG_HEALTH_INFORMATION:
578 	case NVME_LOG_FIRMWARE_SLOT:
579 	case NVME_LOG_CHANGED_NAMESPACE:
580 	case NVME_LOG_COMMAND_EFFECT:
581 	case NVME_LOG_RES_NOTIFICATION:
582 	case NVME_LOG_SANITIZE_STATUS:
583 		return (true);
584 	}
585 
586 	return (false);
587 }
588 
589 static uint32_t
590 nvme_ctrlr_get_log_page_size(struct nvme_controller *ctrlr, uint8_t page_id)
591 {
592 	uint32_t	log_page_size;
593 
594 	switch (page_id) {
595 	case NVME_LOG_ERROR:
596 		log_page_size = min(
597 		    sizeof(struct nvme_error_information_entry) *
598 		    (ctrlr->cdata.elpe + 1), NVME_MAX_AER_LOG_SIZE);
599 		break;
600 	case NVME_LOG_HEALTH_INFORMATION:
601 		log_page_size = sizeof(struct nvme_health_information_page);
602 		break;
603 	case NVME_LOG_FIRMWARE_SLOT:
604 		log_page_size = sizeof(struct nvme_firmware_page);
605 		break;
606 	case NVME_LOG_CHANGED_NAMESPACE:
607 		log_page_size = sizeof(struct nvme_ns_list);
608 		break;
609 	case NVME_LOG_COMMAND_EFFECT:
610 		log_page_size = sizeof(struct nvme_command_effects_page);
611 		break;
612 	case NVME_LOG_RES_NOTIFICATION:
613 		log_page_size = sizeof(struct nvme_res_notification_page);
614 		break;
615 	case NVME_LOG_SANITIZE_STATUS:
616 		log_page_size = sizeof(struct nvme_sanitize_status_page);
617 		break;
618 	default:
619 		log_page_size = 0;
620 		break;
621 	}
622 
623 	return (log_page_size);
624 }
625 
626 static void
627 nvme_ctrlr_log_critical_warnings(struct nvme_controller *ctrlr,
628     uint8_t state)
629 {
630 
631 	if (state & NVME_CRIT_WARN_ST_AVAILABLE_SPARE)
632 		nvme_ctrlr_devctl_log(ctrlr, "critical",
633 		    "available spare space below threshold");
634 
635 	if (state & NVME_CRIT_WARN_ST_TEMPERATURE)
636 		nvme_ctrlr_devctl_log(ctrlr, "critical",
637 		    "temperature above threshold");
638 
639 	if (state & NVME_CRIT_WARN_ST_DEVICE_RELIABILITY)
640 		nvme_ctrlr_devctl_log(ctrlr, "critical",
641 		    "device reliability degraded");
642 
643 	if (state & NVME_CRIT_WARN_ST_READ_ONLY)
644 		nvme_ctrlr_devctl_log(ctrlr, "critical",
645 		    "media placed in read only mode");
646 
647 	if (state & NVME_CRIT_WARN_ST_VOLATILE_MEMORY_BACKUP)
648 		nvme_ctrlr_devctl_log(ctrlr, "critical",
649 		    "volatile memory backup device failed");
650 
651 	if (state & NVME_CRIT_WARN_ST_RESERVED_MASK)
652 		nvme_ctrlr_devctl_log(ctrlr, "critical",
653 		    "unknown critical warning(s): state = 0x%02x",
654 		    state & NVME_CRIT_WARN_ST_RESERVED_MASK);
655 }
656 
657 static void
658 nvme_ctrlr_async_event_log_page_cb(void *arg, const struct nvme_completion *cpl)
659 {
660 	struct nvme_async_event_request		*aer = arg;
661 	struct nvme_health_information_page	*health_info;
662 	struct nvme_ns_list			*nsl;
663 	struct nvme_error_information_entry	*err;
664 	int i;
665 
666 	/*
667 	 * If the log page fetch for some reason completed with an error,
668 	 *  don't pass log page data to the consumers.  In practice, this case
669 	 *  should never happen.
670 	 */
671 	if (nvme_completion_is_error(cpl))
672 		nvme_notify_async_consumers(aer->ctrlr, &aer->cpl,
673 		    aer->log_page_id, NULL, 0);
674 	else {
675 		/* Convert data to host endian */
676 		switch (aer->log_page_id) {
677 		case NVME_LOG_ERROR:
678 			err = (struct nvme_error_information_entry *)aer->log_page_buffer;
679 			for (i = 0; i < (aer->ctrlr->cdata.elpe + 1); i++)
680 				nvme_error_information_entry_swapbytes(err++);
681 			break;
682 		case NVME_LOG_HEALTH_INFORMATION:
683 			nvme_health_information_page_swapbytes(
684 			    (struct nvme_health_information_page *)aer->log_page_buffer);
685 			break;
686 		case NVME_LOG_CHANGED_NAMESPACE:
687 			nvme_ns_list_swapbytes(
688 			    (struct nvme_ns_list *)aer->log_page_buffer);
689 			break;
690 		case NVME_LOG_COMMAND_EFFECT:
691 			nvme_command_effects_page_swapbytes(
692 			    (struct nvme_command_effects_page *)aer->log_page_buffer);
693 			break;
694 		case NVME_LOG_RES_NOTIFICATION:
695 			nvme_res_notification_page_swapbytes(
696 			    (struct nvme_res_notification_page *)aer->log_page_buffer);
697 			break;
698 		case NVME_LOG_SANITIZE_STATUS:
699 			nvme_sanitize_status_page_swapbytes(
700 			    (struct nvme_sanitize_status_page *)aer->log_page_buffer);
701 			break;
702 		case INTEL_LOG_TEMP_STATS:
703 			intel_log_temp_stats_swapbytes(
704 			    (struct intel_log_temp_stats *)aer->log_page_buffer);
705 			break;
706 		default:
707 			break;
708 		}
709 
710 		if (aer->log_page_id == NVME_LOG_HEALTH_INFORMATION) {
711 			health_info = (struct nvme_health_information_page *)
712 			    aer->log_page_buffer;
713 			nvme_ctrlr_log_critical_warnings(aer->ctrlr,
714 			    health_info->critical_warning);
715 			/*
716 			 * Critical warnings reported through the
717 			 *  SMART/health log page are persistent, so
718 			 *  clear the associated bits in the async event
719 			 *  config so that we do not receive repeated
720 			 *  notifications for the same event.
721 			 */
722 			aer->ctrlr->async_event_config &=
723 			    ~health_info->critical_warning;
724 			nvme_ctrlr_cmd_set_async_event_config(aer->ctrlr,
725 			    aer->ctrlr->async_event_config, NULL, NULL);
726 		} else if (aer->log_page_id == NVME_LOG_CHANGED_NAMESPACE &&
727 		    !nvme_use_nvd) {
728 			nsl = (struct nvme_ns_list *)aer->log_page_buffer;
729 			for (i = 0; i < nitems(nsl->ns) && nsl->ns[i] != 0; i++) {
730 				if (nsl->ns[i] > NVME_MAX_NAMESPACES)
731 					break;
732 				nvme_notify_ns(aer->ctrlr, nsl->ns[i]);
733 			}
734 		}
735 
736 		/*
737 		 * Pass the cpl data from the original async event completion,
738 		 *  not the log page fetch.
739 		 */
740 		nvme_notify_async_consumers(aer->ctrlr, &aer->cpl,
741 		    aer->log_page_id, aer->log_page_buffer, aer->log_page_size);
742 	}
743 
744 	/*
745 	 * Repost another asynchronous event request to replace the one
746 	 *  that just completed.
747 	 */
748 	nvme_ctrlr_construct_and_submit_aer(aer->ctrlr, aer);
749 }
750 
751 static void
752 nvme_ctrlr_async_event_cb(void *arg, const struct nvme_completion *cpl)
753 {
754 	struct nvme_async_event_request	*aer = arg;
755 
756 	if (nvme_completion_is_error(cpl)) {
757 		/*
758 		 *  Do not retry failed async event requests.  This avoids
759 		 *  infinite loops where a new async event request is submitted
760 		 *  to replace the one just failed, only to fail again and
761 		 *  perpetuate the loop.
762 		 */
763 		return;
764 	}
765 
766 	/* Associated log page is in bits 23:16 of completion entry dw0. */
767 	aer->log_page_id = (cpl->cdw0 & 0xFF0000) >> 16;
768 
769 	nvme_printf(aer->ctrlr, "async event occurred (type 0x%x, info 0x%02x,"
770 	    " page 0x%02x)\n", (cpl->cdw0 & 0x07), (cpl->cdw0 & 0xFF00) >> 8,
771 	    aer->log_page_id);
772 
773 	if (is_log_page_id_valid(aer->log_page_id)) {
774 		aer->log_page_size = nvme_ctrlr_get_log_page_size(aer->ctrlr,
775 		    aer->log_page_id);
776 		memcpy(&aer->cpl, cpl, sizeof(*cpl));
777 		nvme_ctrlr_cmd_get_log_page(aer->ctrlr, aer->log_page_id,
778 		    NVME_GLOBAL_NAMESPACE_TAG, aer->log_page_buffer,
779 		    aer->log_page_size, nvme_ctrlr_async_event_log_page_cb,
780 		    aer);
781 		/* Wait to notify consumers until after log page is fetched. */
782 	} else {
783 		nvme_notify_async_consumers(aer->ctrlr, cpl, aer->log_page_id,
784 		    NULL, 0);
785 
786 		/*
787 		 * Repost another asynchronous event request to replace the one
788 		 *  that just completed.
789 		 */
790 		nvme_ctrlr_construct_and_submit_aer(aer->ctrlr, aer);
791 	}
792 }
793 
794 static void
795 nvme_ctrlr_construct_and_submit_aer(struct nvme_controller *ctrlr,
796     struct nvme_async_event_request *aer)
797 {
798 	struct nvme_request *req;
799 
800 	aer->ctrlr = ctrlr;
801 	req = nvme_allocate_request_null(nvme_ctrlr_async_event_cb, aer);
802 	aer->req = req;
803 
804 	/*
805 	 * Disable timeout here, since asynchronous event requests should by
806 	 *  nature never be timed out.
807 	 */
808 	req->timeout = false;
809 	req->cmd.opc = NVME_OPC_ASYNC_EVENT_REQUEST;
810 	nvme_ctrlr_submit_admin_request(ctrlr, req);
811 }
812 
813 static void
814 nvme_ctrlr_configure_aer(struct nvme_controller *ctrlr)
815 {
816 	struct nvme_completion_poll_status	status;
817 	struct nvme_async_event_request		*aer;
818 	uint32_t				i;
819 
820 	ctrlr->async_event_config = NVME_CRIT_WARN_ST_AVAILABLE_SPARE |
821 	    NVME_CRIT_WARN_ST_DEVICE_RELIABILITY |
822 	    NVME_CRIT_WARN_ST_READ_ONLY |
823 	    NVME_CRIT_WARN_ST_VOLATILE_MEMORY_BACKUP;
824 	if (ctrlr->cdata.ver >= NVME_REV(1, 2))
825 		ctrlr->async_event_config |=
826 		    ctrlr->cdata.oaes & (NVME_ASYNC_EVENT_NS_ATTRIBUTE |
827 			NVME_ASYNC_EVENT_FW_ACTIVATE);
828 
829 	status.done = 0;
830 	nvme_ctrlr_cmd_get_feature(ctrlr, NVME_FEAT_TEMPERATURE_THRESHOLD,
831 	    0, NULL, 0, nvme_completion_poll_cb, &status);
832 	nvme_completion_poll(&status);
833 	if (nvme_completion_is_error(&status.cpl) ||
834 	    (status.cpl.cdw0 & 0xFFFF) == 0xFFFF ||
835 	    (status.cpl.cdw0 & 0xFFFF) == 0x0000) {
836 		nvme_printf(ctrlr, "temperature threshold not supported\n");
837 	} else
838 		ctrlr->async_event_config |= NVME_CRIT_WARN_ST_TEMPERATURE;
839 
840 	nvme_ctrlr_cmd_set_async_event_config(ctrlr,
841 	    ctrlr->async_event_config, NULL, NULL);
842 
843 	/* aerl is a zero-based value, so we need to add 1 here. */
844 	ctrlr->num_aers = min(NVME_MAX_ASYNC_EVENTS, (ctrlr->cdata.aerl+1));
845 
846 	for (i = 0; i < ctrlr->num_aers; i++) {
847 		aer = &ctrlr->aer[i];
848 		nvme_ctrlr_construct_and_submit_aer(ctrlr, aer);
849 	}
850 }
851 
852 static void
853 nvme_ctrlr_configure_int_coalescing(struct nvme_controller *ctrlr)
854 {
855 
856 	ctrlr->int_coal_time = 0;
857 	TUNABLE_INT_FETCH("hw.nvme.int_coal_time",
858 	    &ctrlr->int_coal_time);
859 
860 	ctrlr->int_coal_threshold = 0;
861 	TUNABLE_INT_FETCH("hw.nvme.int_coal_threshold",
862 	    &ctrlr->int_coal_threshold);
863 
864 	nvme_ctrlr_cmd_set_interrupt_coalescing(ctrlr, ctrlr->int_coal_time,
865 	    ctrlr->int_coal_threshold, NULL, NULL);
866 }
867 
868 static void
869 nvme_ctrlr_hmb_free(struct nvme_controller *ctrlr)
870 {
871 	struct nvme_hmb_chunk *hmbc;
872 	int i;
873 
874 	if (ctrlr->hmb_desc_paddr) {
875 		bus_dmamap_unload(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_map);
876 		bus_dmamem_free(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_vaddr,
877 		    ctrlr->hmb_desc_map);
878 		ctrlr->hmb_desc_paddr = 0;
879 	}
880 	if (ctrlr->hmb_desc_tag) {
881 		bus_dma_tag_destroy(ctrlr->hmb_desc_tag);
882 		ctrlr->hmb_desc_tag = NULL;
883 	}
884 	for (i = 0; i < ctrlr->hmb_nchunks; i++) {
885 		hmbc = &ctrlr->hmb_chunks[i];
886 		bus_dmamap_unload(ctrlr->hmb_tag, hmbc->hmbc_map);
887 		bus_dmamem_free(ctrlr->hmb_tag, hmbc->hmbc_vaddr,
888 		    hmbc->hmbc_map);
889 	}
890 	ctrlr->hmb_nchunks = 0;
891 	if (ctrlr->hmb_tag) {
892 		bus_dma_tag_destroy(ctrlr->hmb_tag);
893 		ctrlr->hmb_tag = NULL;
894 	}
895 	if (ctrlr->hmb_chunks) {
896 		free(ctrlr->hmb_chunks, M_NVME);
897 		ctrlr->hmb_chunks = NULL;
898 	}
899 }
900 
901 static void
902 nvme_ctrlr_hmb_alloc(struct nvme_controller *ctrlr)
903 {
904 	struct nvme_hmb_chunk *hmbc;
905 	size_t pref, min, minc, size;
906 	int err, i;
907 	uint64_t max;
908 
909 	/* Limit HMB to 5% of RAM size per device by default. */
910 	max = (uint64_t)physmem * PAGE_SIZE / 20;
911 	TUNABLE_UINT64_FETCH("hw.nvme.hmb_max", &max);
912 
913 	/*
914 	 * Units of Host Memory Buffer in the Identify info are always in terms
915 	 * of 4k units.
916 	 */
917 	min = (long long unsigned)ctrlr->cdata.hmmin * NVME_HMB_UNITS;
918 	if (max == 0 || max < min)
919 		return;
920 	pref = MIN((long long unsigned)ctrlr->cdata.hmpre * NVME_HMB_UNITS, max);
921 	minc = MAX(ctrlr->cdata.hmminds * NVME_HMB_UNITS, ctrlr->page_size);
922 	if (min > 0 && ctrlr->cdata.hmmaxd > 0)
923 		minc = MAX(minc, min / ctrlr->cdata.hmmaxd);
924 	ctrlr->hmb_chunk = pref;
925 
926 again:
927 	/*
928 	 * However, the chunk sizes, number of chunks, and alignment of chunks
929 	 * are all based on the current MPS (ctrlr->page_size).
930 	 */
931 	ctrlr->hmb_chunk = roundup2(ctrlr->hmb_chunk, ctrlr->page_size);
932 	ctrlr->hmb_nchunks = howmany(pref, ctrlr->hmb_chunk);
933 	if (ctrlr->cdata.hmmaxd > 0 && ctrlr->hmb_nchunks > ctrlr->cdata.hmmaxd)
934 		ctrlr->hmb_nchunks = ctrlr->cdata.hmmaxd;
935 	ctrlr->hmb_chunks = malloc(sizeof(struct nvme_hmb_chunk) *
936 	    ctrlr->hmb_nchunks, M_NVME, M_WAITOK);
937 	err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev),
938 	    ctrlr->page_size, 0, BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
939 	    ctrlr->hmb_chunk, 1, ctrlr->hmb_chunk, 0, NULL, NULL, &ctrlr->hmb_tag);
940 	if (err != 0) {
941 		nvme_printf(ctrlr, "HMB tag create failed %d\n", err);
942 		nvme_ctrlr_hmb_free(ctrlr);
943 		return;
944 	}
945 
946 	for (i = 0; i < ctrlr->hmb_nchunks; i++) {
947 		hmbc = &ctrlr->hmb_chunks[i];
948 		if (bus_dmamem_alloc(ctrlr->hmb_tag,
949 		    (void **)&hmbc->hmbc_vaddr, BUS_DMA_NOWAIT,
950 		    &hmbc->hmbc_map)) {
951 			nvme_printf(ctrlr, "failed to alloc HMB\n");
952 			break;
953 		}
954 		if (bus_dmamap_load(ctrlr->hmb_tag, hmbc->hmbc_map,
955 		    hmbc->hmbc_vaddr, ctrlr->hmb_chunk, nvme_single_map,
956 		    &hmbc->hmbc_paddr, BUS_DMA_NOWAIT) != 0) {
957 			bus_dmamem_free(ctrlr->hmb_tag, hmbc->hmbc_vaddr,
958 			    hmbc->hmbc_map);
959 			nvme_printf(ctrlr, "failed to load HMB\n");
960 			break;
961 		}
962 		bus_dmamap_sync(ctrlr->hmb_tag, hmbc->hmbc_map,
963 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
964 	}
965 
966 	if (i < ctrlr->hmb_nchunks && i * ctrlr->hmb_chunk < min &&
967 	    ctrlr->hmb_chunk / 2 >= minc) {
968 		ctrlr->hmb_nchunks = i;
969 		nvme_ctrlr_hmb_free(ctrlr);
970 		ctrlr->hmb_chunk /= 2;
971 		goto again;
972 	}
973 	ctrlr->hmb_nchunks = i;
974 	if (ctrlr->hmb_nchunks * ctrlr->hmb_chunk < min) {
975 		nvme_ctrlr_hmb_free(ctrlr);
976 		return;
977 	}
978 
979 	size = sizeof(struct nvme_hmb_desc) * ctrlr->hmb_nchunks;
980 	err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev),
981 	    16, 0, BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
982 	    size, 1, size, 0, NULL, NULL, &ctrlr->hmb_desc_tag);
983 	if (err != 0) {
984 		nvme_printf(ctrlr, "HMB desc tag create failed %d\n", err);
985 		nvme_ctrlr_hmb_free(ctrlr);
986 		return;
987 	}
988 	if (bus_dmamem_alloc(ctrlr->hmb_desc_tag,
989 	    (void **)&ctrlr->hmb_desc_vaddr, BUS_DMA_WAITOK,
990 	    &ctrlr->hmb_desc_map)) {
991 		nvme_printf(ctrlr, "failed to alloc HMB desc\n");
992 		nvme_ctrlr_hmb_free(ctrlr);
993 		return;
994 	}
995 	if (bus_dmamap_load(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_map,
996 	    ctrlr->hmb_desc_vaddr, size, nvme_single_map,
997 	    &ctrlr->hmb_desc_paddr, BUS_DMA_NOWAIT) != 0) {
998 		bus_dmamem_free(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_vaddr,
999 		    ctrlr->hmb_desc_map);
1000 		nvme_printf(ctrlr, "failed to load HMB desc\n");
1001 		nvme_ctrlr_hmb_free(ctrlr);
1002 		return;
1003 	}
1004 
1005 	for (i = 0; i < ctrlr->hmb_nchunks; i++) {
1006 		memset(&ctrlr->hmb_desc_vaddr[i], 0,
1007 		    sizeof(struct nvme_hmb_desc));
1008 		ctrlr->hmb_desc_vaddr[i].addr =
1009 		    htole64(ctrlr->hmb_chunks[i].hmbc_paddr);
1010 		ctrlr->hmb_desc_vaddr[i].size = htole32(ctrlr->hmb_chunk / ctrlr->page_size);
1011 	}
1012 	bus_dmamap_sync(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_map,
1013 	    BUS_DMASYNC_PREWRITE);
1014 
1015 	nvme_printf(ctrlr, "Allocated %lluMB host memory buffer\n",
1016 	    (long long unsigned)ctrlr->hmb_nchunks * ctrlr->hmb_chunk
1017 	    / 1024 / 1024);
1018 }
1019 
1020 static void
1021 nvme_ctrlr_hmb_enable(struct nvme_controller *ctrlr, bool enable, bool memret)
1022 {
1023 	struct nvme_completion_poll_status	status;
1024 	uint32_t cdw11;
1025 
1026 	cdw11 = 0;
1027 	if (enable)
1028 		cdw11 |= 1;
1029 	if (memret)
1030 		cdw11 |= 2;
1031 	status.done = 0;
1032 	nvme_ctrlr_cmd_set_feature(ctrlr, NVME_FEAT_HOST_MEMORY_BUFFER, cdw11,
1033 	    ctrlr->hmb_nchunks * ctrlr->hmb_chunk / ctrlr->page_size,
1034 	    ctrlr->hmb_desc_paddr, ctrlr->hmb_desc_paddr >> 32,
1035 	    ctrlr->hmb_nchunks, NULL, 0,
1036 	    nvme_completion_poll_cb, &status);
1037 	nvme_completion_poll(&status);
1038 	if (nvme_completion_is_error(&status.cpl))
1039 		nvme_printf(ctrlr, "nvme_ctrlr_hmb_enable failed!\n");
1040 }
1041 
1042 static void
1043 nvme_ctrlr_start(void *ctrlr_arg, bool resetting)
1044 {
1045 	struct nvme_controller *ctrlr = ctrlr_arg;
1046 	uint32_t old_num_io_queues;
1047 	int i;
1048 
1049 	TSENTER();
1050 
1051 	/*
1052 	 * Only reset adminq here when we are restarting the
1053 	 *  controller after a reset.  During initialization,
1054 	 *  we have already submitted admin commands to get
1055 	 *  the number of I/O queues supported, so cannot reset
1056 	 *  the adminq again here.
1057 	 */
1058 	if (resetting) {
1059 		nvme_qpair_reset(&ctrlr->adminq);
1060 		nvme_admin_qpair_enable(&ctrlr->adminq);
1061 	}
1062 
1063 	if (ctrlr->ioq != NULL) {
1064 		for (i = 0; i < ctrlr->num_io_queues; i++)
1065 			nvme_qpair_reset(&ctrlr->ioq[i]);
1066 	}
1067 
1068 	/*
1069 	 * If it was a reset on initialization command timeout, just
1070 	 * return here, letting initialization code fail gracefully.
1071 	 */
1072 	if (resetting && !ctrlr->is_initialized)
1073 		return;
1074 
1075 	if (resetting && nvme_ctrlr_identify(ctrlr) != 0) {
1076 		nvme_ctrlr_fail(ctrlr);
1077 		return;
1078 	}
1079 
1080 	/*
1081 	 * The number of qpairs are determined during controller initialization,
1082 	 *  including using NVMe SET_FEATURES/NUMBER_OF_QUEUES to determine the
1083 	 *  HW limit.  We call SET_FEATURES again here so that it gets called
1084 	 *  after any reset for controllers that depend on the driver to
1085 	 *  explicit specify how many queues it will use.  This value should
1086 	 *  never change between resets, so panic if somehow that does happen.
1087 	 */
1088 	if (resetting) {
1089 		old_num_io_queues = ctrlr->num_io_queues;
1090 		if (nvme_ctrlr_set_num_qpairs(ctrlr) != 0) {
1091 			nvme_ctrlr_fail(ctrlr);
1092 			return;
1093 		}
1094 
1095 		if (old_num_io_queues != ctrlr->num_io_queues) {
1096 			panic("num_io_queues changed from %u to %u",
1097 			      old_num_io_queues, ctrlr->num_io_queues);
1098 		}
1099 	}
1100 
1101 	if (ctrlr->cdata.hmpre > 0 && ctrlr->hmb_nchunks == 0) {
1102 		nvme_ctrlr_hmb_alloc(ctrlr);
1103 		if (ctrlr->hmb_nchunks > 0)
1104 			nvme_ctrlr_hmb_enable(ctrlr, true, false);
1105 	} else if (ctrlr->hmb_nchunks > 0)
1106 		nvme_ctrlr_hmb_enable(ctrlr, true, true);
1107 
1108 	if (nvme_ctrlr_create_qpairs(ctrlr) != 0) {
1109 		nvme_ctrlr_fail(ctrlr);
1110 		return;
1111 	}
1112 
1113 	if (nvme_ctrlr_construct_namespaces(ctrlr) != 0) {
1114 		nvme_ctrlr_fail(ctrlr);
1115 		return;
1116 	}
1117 
1118 	nvme_ctrlr_configure_aer(ctrlr);
1119 	nvme_ctrlr_configure_int_coalescing(ctrlr);
1120 
1121 	for (i = 0; i < ctrlr->num_io_queues; i++)
1122 		nvme_io_qpair_enable(&ctrlr->ioq[i]);
1123 	TSEXIT();
1124 }
1125 
1126 void
1127 nvme_ctrlr_start_config_hook(void *arg)
1128 {
1129 	struct nvme_controller *ctrlr = arg;
1130 
1131 	TSENTER();
1132 
1133 	if (nvme_ctrlr_hw_reset(ctrlr) != 0) {
1134 fail:
1135 		nvme_ctrlr_fail(ctrlr);
1136 		config_intrhook_disestablish(&ctrlr->config_hook);
1137 		return;
1138 	}
1139 
1140 	nvme_qpair_reset(&ctrlr->adminq);
1141 	nvme_admin_qpair_enable(&ctrlr->adminq);
1142 
1143 	if (nvme_ctrlr_identify(ctrlr) == 0 &&
1144 	    nvme_ctrlr_set_num_qpairs(ctrlr) == 0 &&
1145 	    nvme_ctrlr_construct_io_qpairs(ctrlr) == 0)
1146 		nvme_ctrlr_start(ctrlr, false);
1147 	else
1148 		goto fail;
1149 
1150 	nvme_sysctl_initialize_ctrlr(ctrlr);
1151 	config_intrhook_disestablish(&ctrlr->config_hook);
1152 
1153 	ctrlr->is_initialized = 1;
1154 	nvme_notify_new_controller(ctrlr);
1155 	TSEXIT();
1156 }
1157 
1158 static void
1159 nvme_ctrlr_reset_task(void *arg, int pending)
1160 {
1161 	struct nvme_controller	*ctrlr = arg;
1162 	int			status;
1163 
1164 	nvme_ctrlr_devctl_log(ctrlr, "RESET", "resetting controller");
1165 	status = nvme_ctrlr_hw_reset(ctrlr);
1166 	if (status == 0)
1167 		nvme_ctrlr_start(ctrlr, true);
1168 	else
1169 		nvme_ctrlr_fail(ctrlr);
1170 
1171 	atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
1172 }
1173 
1174 /*
1175  * Poll all the queues enabled on the device for completion.
1176  */
1177 void
1178 nvme_ctrlr_poll(struct nvme_controller *ctrlr)
1179 {
1180 	int i;
1181 
1182 	nvme_qpair_process_completions(&ctrlr->adminq);
1183 
1184 	for (i = 0; i < ctrlr->num_io_queues; i++)
1185 		if (ctrlr->ioq && ctrlr->ioq[i].cpl)
1186 			nvme_qpair_process_completions(&ctrlr->ioq[i]);
1187 }
1188 
1189 /*
1190  * Poll the single-vector interrupt case: num_io_queues will be 1 and
1191  * there's only a single vector. While we're polling, we mask further
1192  * interrupts in the controller.
1193  */
1194 void
1195 nvme_ctrlr_shared_handler(void *arg)
1196 {
1197 	struct nvme_controller *ctrlr = arg;
1198 
1199 	nvme_mmio_write_4(ctrlr, intms, 1);
1200 	nvme_ctrlr_poll(ctrlr);
1201 	nvme_mmio_write_4(ctrlr, intmc, 1);
1202 }
1203 
1204 static void
1205 nvme_pt_done(void *arg, const struct nvme_completion *cpl)
1206 {
1207 	struct nvme_pt_command *pt = arg;
1208 	struct mtx *mtx = pt->driver_lock;
1209 	uint16_t status;
1210 
1211 	bzero(&pt->cpl, sizeof(pt->cpl));
1212 	pt->cpl.cdw0 = cpl->cdw0;
1213 
1214 	status = cpl->status;
1215 	status &= ~NVMEM(NVME_STATUS_P);
1216 	pt->cpl.status = status;
1217 
1218 	mtx_lock(mtx);
1219 	pt->driver_lock = NULL;
1220 	wakeup(pt);
1221 	mtx_unlock(mtx);
1222 }
1223 
1224 int
1225 nvme_ctrlr_passthrough_cmd(struct nvme_controller *ctrlr,
1226     struct nvme_pt_command *pt, uint32_t nsid, int is_user_buffer,
1227     int is_admin_cmd)
1228 {
1229 	struct nvme_request	*req;
1230 	struct mtx		*mtx;
1231 	struct buf		*buf = NULL;
1232 	int			ret = 0;
1233 
1234 	if (pt->len > 0) {
1235 		if (pt->len > ctrlr->max_xfer_size) {
1236 			nvme_printf(ctrlr, "pt->len (%d) "
1237 			    "exceeds max_xfer_size (%d)\n", pt->len,
1238 			    ctrlr->max_xfer_size);
1239 			return EIO;
1240 		}
1241 		if (is_user_buffer) {
1242 			/*
1243 			 * Ensure the user buffer is wired for the duration of
1244 			 *  this pass-through command.
1245 			 */
1246 			PHOLD(curproc);
1247 			buf = uma_zalloc(pbuf_zone, M_WAITOK);
1248 			buf->b_iocmd = pt->is_read ? BIO_READ : BIO_WRITE;
1249 			if (vmapbuf(buf, pt->buf, pt->len, 1) < 0) {
1250 				ret = EFAULT;
1251 				goto err;
1252 			}
1253 			req = nvme_allocate_request_vaddr(buf->b_data, pt->len,
1254 			    nvme_pt_done, pt);
1255 		} else
1256 			req = nvme_allocate_request_vaddr(pt->buf, pt->len,
1257 			    nvme_pt_done, pt);
1258 	} else
1259 		req = nvme_allocate_request_null(nvme_pt_done, pt);
1260 
1261 	/* Assume user space already converted to little-endian */
1262 	req->cmd.opc = pt->cmd.opc;
1263 	req->cmd.fuse = pt->cmd.fuse;
1264 	req->cmd.rsvd2 = pt->cmd.rsvd2;
1265 	req->cmd.rsvd3 = pt->cmd.rsvd3;
1266 	req->cmd.cdw10 = pt->cmd.cdw10;
1267 	req->cmd.cdw11 = pt->cmd.cdw11;
1268 	req->cmd.cdw12 = pt->cmd.cdw12;
1269 	req->cmd.cdw13 = pt->cmd.cdw13;
1270 	req->cmd.cdw14 = pt->cmd.cdw14;
1271 	req->cmd.cdw15 = pt->cmd.cdw15;
1272 
1273 	req->cmd.nsid = htole32(nsid);
1274 
1275 	mtx = mtx_pool_find(mtxpool_sleep, pt);
1276 	pt->driver_lock = mtx;
1277 
1278 	if (is_admin_cmd)
1279 		nvme_ctrlr_submit_admin_request(ctrlr, req);
1280 	else
1281 		nvme_ctrlr_submit_io_request(ctrlr, req);
1282 
1283 	mtx_lock(mtx);
1284 	while (pt->driver_lock != NULL)
1285 		mtx_sleep(pt, mtx, PRIBIO, "nvme_pt", 0);
1286 	mtx_unlock(mtx);
1287 
1288 	if (buf != NULL) {
1289 		vunmapbuf(buf);
1290 err:
1291 		uma_zfree(pbuf_zone, buf);
1292 		PRELE(curproc);
1293 	}
1294 
1295 	return (ret);
1296 }
1297 
1298 static int
1299 nvme_ctrlr_ioctl(struct cdev *cdev, u_long cmd, caddr_t arg, int flag,
1300     struct thread *td)
1301 {
1302 	struct nvme_controller			*ctrlr;
1303 	struct nvme_pt_command			*pt;
1304 
1305 	ctrlr = cdev->si_drv1;
1306 
1307 	switch (cmd) {
1308 	case NVME_RESET_CONTROLLER:
1309 		nvme_ctrlr_reset(ctrlr);
1310 		break;
1311 	case NVME_PASSTHROUGH_CMD:
1312 		pt = (struct nvme_pt_command *)arg;
1313 		return (nvme_ctrlr_passthrough_cmd(ctrlr, pt, le32toh(pt->cmd.nsid),
1314 		    1 /* is_user_buffer */, 1 /* is_admin_cmd */));
1315 	case NVME_GET_NSID:
1316 	{
1317 		struct nvme_get_nsid *gnsid = (struct nvme_get_nsid *)arg;
1318 		strncpy(gnsid->cdev, device_get_nameunit(ctrlr->dev),
1319 		    sizeof(gnsid->cdev));
1320 		gnsid->cdev[sizeof(gnsid->cdev) - 1] = '\0';
1321 		gnsid->nsid = 0;
1322 		break;
1323 	}
1324 	case NVME_GET_MAX_XFER_SIZE:
1325 		*(uint64_t *)arg = ctrlr->max_xfer_size;
1326 		break;
1327 	default:
1328 		return (ENOTTY);
1329 	}
1330 
1331 	return (0);
1332 }
1333 
1334 static struct cdevsw nvme_ctrlr_cdevsw = {
1335 	.d_version =	D_VERSION,
1336 	.d_flags =	0,
1337 	.d_ioctl =	nvme_ctrlr_ioctl
1338 };
1339 
1340 int
1341 nvme_ctrlr_construct(struct nvme_controller *ctrlr, device_t dev)
1342 {
1343 	struct make_dev_args	md_args;
1344 	uint32_t	cap_lo;
1345 	uint32_t	cap_hi;
1346 	uint32_t	to, vs, pmrcap;
1347 	int		status, timeout_period;
1348 
1349 	ctrlr->dev = dev;
1350 
1351 	mtx_init(&ctrlr->lock, "nvme ctrlr lock", NULL, MTX_DEF);
1352 	if (bus_get_domain(dev, &ctrlr->domain) != 0)
1353 		ctrlr->domain = 0;
1354 
1355 	ctrlr->cap_lo = cap_lo = nvme_mmio_read_4(ctrlr, cap_lo);
1356 	if (bootverbose) {
1357 		device_printf(dev, "CapLo: 0x%08x: MQES %u%s%s%s%s, TO %u\n",
1358 		    cap_lo, NVME_CAP_LO_MQES(cap_lo),
1359 		    NVME_CAP_LO_CQR(cap_lo) ? ", CQR" : "",
1360 		    NVME_CAP_LO_AMS(cap_lo) ? ", AMS" : "",
1361 		    (NVME_CAP_LO_AMS(cap_lo) & 0x1) ? " WRRwUPC" : "",
1362 		    (NVME_CAP_LO_AMS(cap_lo) & 0x2) ? " VS" : "",
1363 		    NVME_CAP_LO_TO(cap_lo));
1364 	}
1365 	ctrlr->cap_hi = cap_hi = nvme_mmio_read_4(ctrlr, cap_hi);
1366 	if (bootverbose) {
1367 		device_printf(dev, "CapHi: 0x%08x: DSTRD %u%s, CSS %x%s, "
1368 		    "CPS %x, MPSMIN %u, MPSMAX %u%s%s%s%s%s\n", cap_hi,
1369 		    NVME_CAP_HI_DSTRD(cap_hi),
1370 		    NVME_CAP_HI_NSSRS(cap_hi) ? ", NSSRS" : "",
1371 		    NVME_CAP_HI_CSS(cap_hi),
1372 		    NVME_CAP_HI_BPS(cap_hi) ? ", BPS" : "",
1373 		    NVME_CAP_HI_CPS(cap_hi),
1374 		    NVME_CAP_HI_MPSMIN(cap_hi),
1375 		    NVME_CAP_HI_MPSMAX(cap_hi),
1376 		    NVME_CAP_HI_PMRS(cap_hi) ? ", PMRS" : "",
1377 		    NVME_CAP_HI_CMBS(cap_hi) ? ", CMBS" : "",
1378 		    NVME_CAP_HI_NSSS(cap_hi) ? ", NSSS" : "",
1379 		    NVME_CAP_HI_CRWMS(cap_hi) ? ", CRWMS" : "",
1380 		    NVME_CAP_HI_CRIMS(cap_hi) ? ", CRIMS" : "");
1381 	}
1382 	if (bootverbose) {
1383 		vs = nvme_mmio_read_4(ctrlr, vs);
1384 		device_printf(dev, "Version: 0x%08x: %d.%d\n", vs,
1385 		    NVME_MAJOR(vs), NVME_MINOR(vs));
1386 	}
1387 	if (bootverbose && NVME_CAP_HI_PMRS(cap_hi)) {
1388 		pmrcap = nvme_mmio_read_4(ctrlr, pmrcap);
1389 		device_printf(dev, "PMRCap: 0x%08x: BIR %u%s%s, PMRTU %u, "
1390 		    "PMRWBM %x, PMRTO %u%s\n", pmrcap,
1391 		    NVME_PMRCAP_BIR(pmrcap),
1392 		    NVME_PMRCAP_RDS(pmrcap) ? ", RDS" : "",
1393 		    NVME_PMRCAP_WDS(pmrcap) ? ", WDS" : "",
1394 		    NVME_PMRCAP_PMRTU(pmrcap),
1395 		    NVME_PMRCAP_PMRWBM(pmrcap),
1396 		    NVME_PMRCAP_PMRTO(pmrcap),
1397 		    NVME_PMRCAP_CMSS(pmrcap) ? ", CMSS" : "");
1398 	}
1399 
1400 	ctrlr->dstrd = NVME_CAP_HI_DSTRD(cap_hi) + 2;
1401 
1402 	ctrlr->mps = NVME_CAP_HI_MPSMIN(cap_hi);
1403 	ctrlr->page_size = 1 << (NVME_MPS_SHIFT + ctrlr->mps);
1404 
1405 	/* Get ready timeout value from controller, in units of 500ms. */
1406 	to = NVME_CAP_LO_TO(cap_lo) + 1;
1407 	ctrlr->ready_timeout_in_ms = to * 500;
1408 
1409 	timeout_period = NVME_ADMIN_TIMEOUT_PERIOD;
1410 	TUNABLE_INT_FETCH("hw.nvme.admin_timeout_period", &timeout_period);
1411 	timeout_period = min(timeout_period, NVME_MAX_TIMEOUT_PERIOD);
1412 	timeout_period = max(timeout_period, NVME_MIN_TIMEOUT_PERIOD);
1413 	ctrlr->admin_timeout_period = timeout_period;
1414 
1415 	timeout_period = NVME_DEFAULT_TIMEOUT_PERIOD;
1416 	TUNABLE_INT_FETCH("hw.nvme.timeout_period", &timeout_period);
1417 	timeout_period = min(timeout_period, NVME_MAX_TIMEOUT_PERIOD);
1418 	timeout_period = max(timeout_period, NVME_MIN_TIMEOUT_PERIOD);
1419 	ctrlr->timeout_period = timeout_period;
1420 
1421 	nvme_retry_count = NVME_DEFAULT_RETRY_COUNT;
1422 	TUNABLE_INT_FETCH("hw.nvme.retry_count", &nvme_retry_count);
1423 
1424 	ctrlr->enable_aborts = 0;
1425 	TUNABLE_INT_FETCH("hw.nvme.enable_aborts", &ctrlr->enable_aborts);
1426 
1427 	/* Cap transfers by the maximum addressable by page-sized PRP (4KB pages -> 2MB). */
1428 	ctrlr->max_xfer_size = MIN(maxphys, (ctrlr->page_size / 8 * ctrlr->page_size));
1429 	if (nvme_ctrlr_construct_admin_qpair(ctrlr) != 0)
1430 		return (ENXIO);
1431 
1432 	/*
1433 	 * Create 2 threads for the taskqueue. The reset thread will block when
1434 	 * it detects that the controller has failed until all I/O has been
1435 	 * failed up the stack. The fail_req task needs to be able to run in
1436 	 * this case to finish the request failure for some cases.
1437 	 *
1438 	 * We could partially solve this race by draining the failed requeust
1439 	 * queue before proceding to free the sim, though nothing would stop
1440 	 * new I/O from coming in after we do that drain, but before we reach
1441 	 * cam_sim_free, so this big hammer is used instead.
1442 	 */
1443 	ctrlr->taskqueue = taskqueue_create("nvme_taskq", M_WAITOK,
1444 	    taskqueue_thread_enqueue, &ctrlr->taskqueue);
1445 	taskqueue_start_threads(&ctrlr->taskqueue, 2, PI_DISK, "nvme taskq");
1446 
1447 	ctrlr->is_resetting = 0;
1448 	ctrlr->is_initialized = 0;
1449 	ctrlr->notification_sent = 0;
1450 	TASK_INIT(&ctrlr->reset_task, 0, nvme_ctrlr_reset_task, ctrlr);
1451 	STAILQ_INIT(&ctrlr->fail_req);
1452 	ctrlr->is_failed = false;
1453 
1454 	make_dev_args_init(&md_args);
1455 	md_args.mda_devsw = &nvme_ctrlr_cdevsw;
1456 	md_args.mda_uid = UID_ROOT;
1457 	md_args.mda_gid = GID_WHEEL;
1458 	md_args.mda_mode = 0600;
1459 	md_args.mda_unit = device_get_unit(dev);
1460 	md_args.mda_si_drv1 = (void *)ctrlr;
1461 	status = make_dev_s(&md_args, &ctrlr->cdev, "nvme%d",
1462 	    device_get_unit(dev));
1463 	if (status != 0)
1464 		return (ENXIO);
1465 
1466 	return (0);
1467 }
1468 
1469 void
1470 nvme_ctrlr_destruct(struct nvme_controller *ctrlr, device_t dev)
1471 {
1472 	int	gone, i;
1473 
1474 	ctrlr->is_dying = true;
1475 
1476 	if (ctrlr->resource == NULL)
1477 		goto nores;
1478 	if (!mtx_initialized(&ctrlr->adminq.lock))
1479 		goto noadminq;
1480 
1481 	/*
1482 	 * Check whether it is a hot unplug or a clean driver detach.
1483 	 * If device is not there any more, skip any shutdown commands.
1484 	 */
1485 	gone = (nvme_mmio_read_4(ctrlr, csts) == NVME_GONE);
1486 	if (gone)
1487 		nvme_ctrlr_fail(ctrlr);
1488 	else
1489 		nvme_notify_fail_consumers(ctrlr);
1490 
1491 	for (i = 0; i < NVME_MAX_NAMESPACES; i++)
1492 		nvme_ns_destruct(&ctrlr->ns[i]);
1493 
1494 	if (ctrlr->cdev)
1495 		destroy_dev(ctrlr->cdev);
1496 
1497 	if (ctrlr->is_initialized) {
1498 		if (!gone) {
1499 			if (ctrlr->hmb_nchunks > 0)
1500 				nvme_ctrlr_hmb_enable(ctrlr, false, false);
1501 			nvme_ctrlr_delete_qpairs(ctrlr);
1502 		}
1503 		nvme_ctrlr_hmb_free(ctrlr);
1504 	}
1505 	if (ctrlr->ioq != NULL) {
1506 		for (i = 0; i < ctrlr->num_io_queues; i++)
1507 			nvme_io_qpair_destroy(&ctrlr->ioq[i]);
1508 		free(ctrlr->ioq, M_NVME);
1509 	}
1510 	nvme_admin_qpair_destroy(&ctrlr->adminq);
1511 
1512 	/*
1513 	 *  Notify the controller of a shutdown, even though this is due to
1514 	 *   a driver unload, not a system shutdown (this path is not invoked
1515 	 *   during shutdown).  This ensures the controller receives a
1516 	 *   shutdown notification in case the system is shutdown before
1517 	 *   reloading the driver.
1518 	 */
1519 	if (!gone)
1520 		nvme_ctrlr_shutdown(ctrlr);
1521 
1522 	if (!gone)
1523 		nvme_ctrlr_disable(ctrlr);
1524 
1525 noadminq:
1526 	if (ctrlr->taskqueue)
1527 		taskqueue_free(ctrlr->taskqueue);
1528 
1529 	if (ctrlr->tag)
1530 		bus_teardown_intr(ctrlr->dev, ctrlr->res, ctrlr->tag);
1531 
1532 	if (ctrlr->res)
1533 		bus_release_resource(ctrlr->dev, SYS_RES_IRQ,
1534 		    rman_get_rid(ctrlr->res), ctrlr->res);
1535 
1536 	if (ctrlr->bar4_resource != NULL) {
1537 		bus_release_resource(dev, SYS_RES_MEMORY,
1538 		    ctrlr->bar4_resource_id, ctrlr->bar4_resource);
1539 	}
1540 
1541 	bus_release_resource(dev, SYS_RES_MEMORY,
1542 	    ctrlr->resource_id, ctrlr->resource);
1543 
1544 nores:
1545 	mtx_destroy(&ctrlr->lock);
1546 }
1547 
1548 void
1549 nvme_ctrlr_shutdown(struct nvme_controller *ctrlr)
1550 {
1551 	uint32_t	cc;
1552 	uint32_t	csts;
1553 	int		timeout;
1554 
1555 	cc = nvme_mmio_read_4(ctrlr, cc);
1556 	cc &= ~NVMEM(NVME_CC_REG_SHN);
1557 	cc |= NVMEF(NVME_CC_REG_SHN, NVME_SHN_NORMAL);
1558 	nvme_mmio_write_4(ctrlr, cc, cc);
1559 
1560 	timeout = ticks + (ctrlr->cdata.rtd3e == 0 ? 5 * hz :
1561 	    ((uint64_t)ctrlr->cdata.rtd3e * hz + 999999) / 1000000);
1562 	while (1) {
1563 		csts = nvme_mmio_read_4(ctrlr, csts);
1564 		if (csts == NVME_GONE)		/* Hot unplug. */
1565 			break;
1566 		if (NVME_CSTS_GET_SHST(csts) == NVME_SHST_COMPLETE)
1567 			break;
1568 		if (timeout - ticks < 0) {
1569 			nvme_printf(ctrlr, "shutdown timeout\n");
1570 			break;
1571 		}
1572 		pause("nvmeshut", 1);
1573 	}
1574 }
1575 
1576 void
1577 nvme_ctrlr_submit_admin_request(struct nvme_controller *ctrlr,
1578     struct nvme_request *req)
1579 {
1580 
1581 	nvme_qpair_submit_request(&ctrlr->adminq, req);
1582 }
1583 
1584 void
1585 nvme_ctrlr_submit_io_request(struct nvme_controller *ctrlr,
1586     struct nvme_request *req)
1587 {
1588 	struct nvme_qpair       *qpair;
1589 
1590 	qpair = &ctrlr->ioq[QP(ctrlr, curcpu)];
1591 	nvme_qpair_submit_request(qpair, req);
1592 }
1593 
1594 device_t
1595 nvme_ctrlr_get_device(struct nvme_controller *ctrlr)
1596 {
1597 
1598 	return (ctrlr->dev);
1599 }
1600 
1601 const struct nvme_controller_data *
1602 nvme_ctrlr_get_data(struct nvme_controller *ctrlr)
1603 {
1604 
1605 	return (&ctrlr->cdata);
1606 }
1607 
1608 int
1609 nvme_ctrlr_suspend(struct nvme_controller *ctrlr)
1610 {
1611 	int to = hz;
1612 
1613 	/*
1614 	 * Can't touch failed controllers, so it's already suspended.
1615 	 */
1616 	if (ctrlr->is_failed)
1617 		return (0);
1618 
1619 	/*
1620 	 * We don't want the reset taskqueue running, since it does similar
1621 	 * things, so prevent it from running after we start. Wait for any reset
1622 	 * that may have been started to complete. The reset process we follow
1623 	 * will ensure that any new I/O will queue and be given to the hardware
1624 	 * after we resume (though there should be none).
1625 	 */
1626 	while (atomic_cmpset_32(&ctrlr->is_resetting, 0, 1) == 0 && to-- > 0)
1627 		pause("nvmesusp", 1);
1628 	if (to <= 0) {
1629 		nvme_printf(ctrlr,
1630 		    "Competing reset task didn't finish. Try again later.\n");
1631 		return (EWOULDBLOCK);
1632 	}
1633 
1634 	if (ctrlr->hmb_nchunks > 0)
1635 		nvme_ctrlr_hmb_enable(ctrlr, false, false);
1636 
1637 	/*
1638 	 * Per Section 7.6.2 of NVMe spec 1.4, to properly suspend, we need to
1639 	 * delete the hardware I/O queues, and then shutdown. This properly
1640 	 * flushes any metadata the drive may have stored so it can survive
1641 	 * having its power removed and prevents the unsafe shutdown count from
1642 	 * incriminating. Once we delete the qpairs, we have to disable them
1643 	 * before shutting down.
1644 	 */
1645 	nvme_ctrlr_delete_qpairs(ctrlr);
1646 	nvme_ctrlr_disable_qpairs(ctrlr);
1647 	nvme_ctrlr_shutdown(ctrlr);
1648 
1649 	return (0);
1650 }
1651 
1652 int
1653 nvme_ctrlr_resume(struct nvme_controller *ctrlr)
1654 {
1655 
1656 	/*
1657 	 * Can't touch failed controllers, so nothing to do to resume.
1658 	 */
1659 	if (ctrlr->is_failed)
1660 		return (0);
1661 
1662 	if (nvme_ctrlr_hw_reset(ctrlr) != 0)
1663 		goto fail;
1664 
1665 	/*
1666 	 * Now that we've reset the hardware, we can restart the controller. Any
1667 	 * I/O that was pending is requeued. Any admin commands are aborted with
1668 	 * an error. Once we've restarted, take the controller out of reset.
1669 	 */
1670 	nvme_ctrlr_start(ctrlr, true);
1671 	(void)atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
1672 
1673 	return (0);
1674 fail:
1675 	/*
1676 	 * Since we can't bring the controller out of reset, announce and fail
1677 	 * the controller. However, we have to return success for the resume
1678 	 * itself, due to questionable APIs.
1679 	 */
1680 	nvme_printf(ctrlr, "Failed to reset on resume, failing.\n");
1681 	nvme_ctrlr_fail(ctrlr);
1682 	(void)atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
1683 	return (0);
1684 }
1685