xref: /freebsd/sys/dev/nvme/nvme_ctrlr.c (revision efe014e6b177ba1b0d70f7df751f5066d4eafe9b)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (C) 2012-2016 Intel Corporation
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  */
28 
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
31 
32 #include "opt_cam.h"
33 
34 #include <sys/param.h>
35 #include <sys/systm.h>
36 #include <sys/buf.h>
37 #include <sys/bus.h>
38 #include <sys/conf.h>
39 #include <sys/ioccom.h>
40 #include <sys/proc.h>
41 #include <sys/smp.h>
42 #include <sys/uio.h>
43 #include <sys/sbuf.h>
44 #include <sys/endian.h>
45 #include <machine/stdarg.h>
46 #include <vm/vm.h>
47 
48 #include "nvme_private.h"
49 
50 #define B4_CHK_RDY_DELAY_MS	2300		/* work around controller bug */
51 
52 static void nvme_ctrlr_construct_and_submit_aer(struct nvme_controller *ctrlr,
53 						struct nvme_async_event_request *aer);
54 
55 static void
56 nvme_ctrlr_devctl_log(struct nvme_controller *ctrlr, const char *type, const char *msg, ...)
57 {
58 	struct sbuf sb;
59 	va_list ap;
60 	int error;
61 
62 	if (sbuf_new(&sb, NULL, 0, SBUF_AUTOEXTEND | SBUF_NOWAIT) == NULL)
63 		return;
64 	sbuf_printf(&sb, "%s: ", device_get_nameunit(ctrlr->dev));
65 	va_start(ap, msg);
66 	sbuf_vprintf(&sb, msg, ap);
67 	va_end(ap);
68 	error = sbuf_finish(&sb);
69 	if (error == 0)
70 		printf("%s\n", sbuf_data(&sb));
71 
72 	sbuf_clear(&sb);
73 	sbuf_printf(&sb, "name=\"%s\" reason=\"", device_get_nameunit(ctrlr->dev));
74 	va_start(ap, msg);
75 	sbuf_vprintf(&sb, msg, ap);
76 	va_end(ap);
77 	sbuf_printf(&sb, "\"");
78 	error = sbuf_finish(&sb);
79 	if (error == 0)
80 		devctl_notify("nvme", "controller", type, sbuf_data(&sb));
81 	sbuf_delete(&sb);
82 }
83 
84 static int
85 nvme_ctrlr_construct_admin_qpair(struct nvme_controller *ctrlr)
86 {
87 	struct nvme_qpair	*qpair;
88 	uint32_t		num_entries;
89 	int			error;
90 
91 	qpair = &ctrlr->adminq;
92 	qpair->id = 0;
93 	qpair->cpu = CPU_FFS(&cpuset_domain[ctrlr->domain]) - 1;
94 	qpair->domain = ctrlr->domain;
95 
96 	num_entries = NVME_ADMIN_ENTRIES;
97 	TUNABLE_INT_FETCH("hw.nvme.admin_entries", &num_entries);
98 	/*
99 	 * If admin_entries was overridden to an invalid value, revert it
100 	 *  back to our default value.
101 	 */
102 	if (num_entries < NVME_MIN_ADMIN_ENTRIES ||
103 	    num_entries > NVME_MAX_ADMIN_ENTRIES) {
104 		nvme_printf(ctrlr, "invalid hw.nvme.admin_entries=%d "
105 		    "specified\n", num_entries);
106 		num_entries = NVME_ADMIN_ENTRIES;
107 	}
108 
109 	/*
110 	 * The admin queue's max xfer size is treated differently than the
111 	 *  max I/O xfer size.  16KB is sufficient here - maybe even less?
112 	 */
113 	error = nvme_qpair_construct(qpair, num_entries, NVME_ADMIN_TRACKERS,
114 	     ctrlr);
115 	return (error);
116 }
117 
118 #define QP(ctrlr, c)	((c) * (ctrlr)->num_io_queues / mp_ncpus)
119 
120 static int
121 nvme_ctrlr_construct_io_qpairs(struct nvme_controller *ctrlr)
122 {
123 	struct nvme_qpair	*qpair;
124 	uint32_t		cap_lo;
125 	uint16_t		mqes;
126 	int			c, error, i, n;
127 	int			num_entries, num_trackers, max_entries;
128 
129 	/*
130 	 * NVMe spec sets a hard limit of 64K max entries, but devices may
131 	 * specify a smaller limit, so we need to check the MQES field in the
132 	 * capabilities register. We have to cap the number of entries to the
133 	 * current stride allows for in BAR 0/1, otherwise the remainder entries
134 	 * are inaccessable. MQES should reflect this, and this is just a
135 	 * fail-safe.
136 	 */
137 	max_entries =
138 	    (rman_get_size(ctrlr->resource) - nvme_mmio_offsetof(doorbell[0])) /
139 	    (1 << (ctrlr->dstrd + 1));
140 	num_entries = NVME_IO_ENTRIES;
141 	TUNABLE_INT_FETCH("hw.nvme.io_entries", &num_entries);
142 	cap_lo = nvme_mmio_read_4(ctrlr, cap_lo);
143 	mqes = NVME_CAP_LO_MQES(cap_lo);
144 	num_entries = min(num_entries, mqes + 1);
145 	num_entries = min(num_entries, max_entries);
146 
147 	num_trackers = NVME_IO_TRACKERS;
148 	TUNABLE_INT_FETCH("hw.nvme.io_trackers", &num_trackers);
149 
150 	num_trackers = max(num_trackers, NVME_MIN_IO_TRACKERS);
151 	num_trackers = min(num_trackers, NVME_MAX_IO_TRACKERS);
152 	/*
153 	 * No need to have more trackers than entries in the submit queue.  Note
154 	 * also that for a queue size of N, we can only have (N-1) commands
155 	 * outstanding, hence the "-1" here.
156 	 */
157 	num_trackers = min(num_trackers, (num_entries-1));
158 
159 	/*
160 	 * Our best estimate for the maximum number of I/Os that we should
161 	 * normally have in flight at one time. This should be viewed as a hint,
162 	 * not a hard limit and will need to be revisited when the upper layers
163 	 * of the storage system grows multi-queue support.
164 	 */
165 	ctrlr->max_hw_pend_io = num_trackers * ctrlr->num_io_queues * 3 / 4;
166 
167 	ctrlr->ioq = malloc(ctrlr->num_io_queues * sizeof(struct nvme_qpair),
168 	    M_NVME, M_ZERO | M_WAITOK);
169 
170 	for (i = c = n = 0; i < ctrlr->num_io_queues; i++, c += n) {
171 		qpair = &ctrlr->ioq[i];
172 
173 		/*
174 		 * Admin queue has ID=0. IO queues start at ID=1 -
175 		 *  hence the 'i+1' here.
176 		 */
177 		qpair->id = i + 1;
178 		if (ctrlr->num_io_queues > 1) {
179 			/* Find number of CPUs served by this queue. */
180 			for (n = 1; QP(ctrlr, c + n) == i; n++)
181 				;
182 			/* Shuffle multiple NVMe devices between CPUs. */
183 			qpair->cpu = c + (device_get_unit(ctrlr->dev)+n/2) % n;
184 			qpair->domain = pcpu_find(qpair->cpu)->pc_domain;
185 		} else {
186 			qpair->cpu = CPU_FFS(&cpuset_domain[ctrlr->domain]) - 1;
187 			qpair->domain = ctrlr->domain;
188 		}
189 
190 		/*
191 		 * For I/O queues, use the controller-wide max_xfer_size
192 		 *  calculated in nvme_attach().
193 		 */
194 		error = nvme_qpair_construct(qpair, num_entries, num_trackers,
195 		    ctrlr);
196 		if (error)
197 			return (error);
198 
199 		/*
200 		 * Do not bother binding interrupts if we only have one I/O
201 		 *  interrupt thread for this controller.
202 		 */
203 		if (ctrlr->num_io_queues > 1)
204 			bus_bind_intr(ctrlr->dev, qpair->res, qpair->cpu);
205 	}
206 
207 	return (0);
208 }
209 
210 static void
211 nvme_ctrlr_fail(struct nvme_controller *ctrlr)
212 {
213 	int i;
214 
215 	ctrlr->is_failed = true;
216 	nvme_admin_qpair_disable(&ctrlr->adminq);
217 	nvme_qpair_fail(&ctrlr->adminq);
218 	if (ctrlr->ioq != NULL) {
219 		for (i = 0; i < ctrlr->num_io_queues; i++) {
220 			nvme_io_qpair_disable(&ctrlr->ioq[i]);
221 			nvme_qpair_fail(&ctrlr->ioq[i]);
222 		}
223 	}
224 	nvme_notify_fail_consumers(ctrlr);
225 }
226 
227 void
228 nvme_ctrlr_post_failed_request(struct nvme_controller *ctrlr,
229     struct nvme_request *req)
230 {
231 
232 	mtx_lock(&ctrlr->lock);
233 	STAILQ_INSERT_TAIL(&ctrlr->fail_req, req, stailq);
234 	mtx_unlock(&ctrlr->lock);
235 	taskqueue_enqueue(ctrlr->taskqueue, &ctrlr->fail_req_task);
236 }
237 
238 static void
239 nvme_ctrlr_fail_req_task(void *arg, int pending)
240 {
241 	struct nvme_controller	*ctrlr = arg;
242 	struct nvme_request	*req;
243 
244 	mtx_lock(&ctrlr->lock);
245 	while ((req = STAILQ_FIRST(&ctrlr->fail_req)) != NULL) {
246 		STAILQ_REMOVE_HEAD(&ctrlr->fail_req, stailq);
247 		mtx_unlock(&ctrlr->lock);
248 		nvme_qpair_manual_complete_request(req->qpair, req,
249 		    NVME_SCT_GENERIC, NVME_SC_ABORTED_BY_REQUEST);
250 		mtx_lock(&ctrlr->lock);
251 	}
252 	mtx_unlock(&ctrlr->lock);
253 }
254 
255 static int
256 nvme_ctrlr_wait_for_ready(struct nvme_controller *ctrlr, int desired_val)
257 {
258 	int timeout = ticks + (uint64_t)ctrlr->ready_timeout_in_ms * hz / 1000;
259 	uint32_t csts;
260 
261 	while (1) {
262 		csts = nvme_mmio_read_4(ctrlr, csts);
263 		if (csts == NVME_GONE)		/* Hot unplug. */
264 			return (ENXIO);
265 		if (((csts >> NVME_CSTS_REG_RDY_SHIFT) & NVME_CSTS_REG_RDY_MASK)
266 		    == desired_val)
267 			break;
268 		if (timeout - ticks < 0) {
269 			nvme_printf(ctrlr, "controller ready did not become %d "
270 			    "within %d ms\n", desired_val, ctrlr->ready_timeout_in_ms);
271 			return (ENXIO);
272 		}
273 		pause("nvmerdy", 1);
274 	}
275 
276 	return (0);
277 }
278 
279 static int
280 nvme_ctrlr_disable(struct nvme_controller *ctrlr)
281 {
282 	uint32_t cc;
283 	uint32_t csts;
284 	uint8_t  en, rdy;
285 	int err;
286 
287 	cc = nvme_mmio_read_4(ctrlr, cc);
288 	csts = nvme_mmio_read_4(ctrlr, csts);
289 
290 	en = (cc >> NVME_CC_REG_EN_SHIFT) & NVME_CC_REG_EN_MASK;
291 	rdy = (csts >> NVME_CSTS_REG_RDY_SHIFT) & NVME_CSTS_REG_RDY_MASK;
292 
293 	/*
294 	 * Per 3.1.5 in NVME 1.3 spec, transitioning CC.EN from 0 to 1
295 	 * when CSTS.RDY is 1 or transitioning CC.EN from 1 to 0 when
296 	 * CSTS.RDY is 0 "has undefined results" So make sure that CSTS.RDY
297 	 * isn't the desired value. Short circuit if we're already disabled.
298 	 */
299 	if (en == 1) {
300 		if (rdy == 0) {
301 			/* EN == 1, wait for  RDY == 1 or fail */
302 			err = nvme_ctrlr_wait_for_ready(ctrlr, 1);
303 			if (err != 0)
304 				return (err);
305 		}
306 	} else {
307 		/* EN == 0 already wait for RDY == 0 */
308 		if (rdy == 0)
309 			return (0);
310 		else
311 			return (nvme_ctrlr_wait_for_ready(ctrlr, 0));
312 	}
313 
314 	cc &= ~NVME_CC_REG_EN_MASK;
315 	nvme_mmio_write_4(ctrlr, cc, cc);
316 	/*
317 	 * Some drives have issues with accessing the mmio after we
318 	 * disable, so delay for a bit after we write the bit to
319 	 * cope with these issues.
320 	 */
321 	if (ctrlr->quirks & QUIRK_DELAY_B4_CHK_RDY)
322 		pause("nvmeR", B4_CHK_RDY_DELAY_MS * hz / 1000);
323 	return (nvme_ctrlr_wait_for_ready(ctrlr, 0));
324 }
325 
326 static int
327 nvme_ctrlr_enable(struct nvme_controller *ctrlr)
328 {
329 	uint32_t	cc;
330 	uint32_t	csts;
331 	uint32_t	aqa;
332 	uint32_t	qsize;
333 	uint8_t		en, rdy;
334 	int		err;
335 
336 	cc = nvme_mmio_read_4(ctrlr, cc);
337 	csts = nvme_mmio_read_4(ctrlr, csts);
338 
339 	en = (cc >> NVME_CC_REG_EN_SHIFT) & NVME_CC_REG_EN_MASK;
340 	rdy = (csts >> NVME_CSTS_REG_RDY_SHIFT) & NVME_CSTS_REG_RDY_MASK;
341 
342 	/*
343 	 * See note in nvme_ctrlr_disable. Short circuit if we're already enabled.
344 	 */
345 	if (en == 1) {
346 		if (rdy == 1)
347 			return (0);
348 		else
349 			return (nvme_ctrlr_wait_for_ready(ctrlr, 1));
350 	} else {
351 		/* EN == 0 already wait for RDY == 0 or fail */
352 		err = nvme_ctrlr_wait_for_ready(ctrlr, 0);
353 		if (err != 0)
354 			return (err);
355 	}
356 
357 	nvme_mmio_write_8(ctrlr, asq, ctrlr->adminq.cmd_bus_addr);
358 	DELAY(5000);
359 	nvme_mmio_write_8(ctrlr, acq, ctrlr->adminq.cpl_bus_addr);
360 	DELAY(5000);
361 
362 	/* acqs and asqs are 0-based. */
363 	qsize = ctrlr->adminq.num_entries - 1;
364 
365 	aqa = 0;
366 	aqa = (qsize & NVME_AQA_REG_ACQS_MASK) << NVME_AQA_REG_ACQS_SHIFT;
367 	aqa |= (qsize & NVME_AQA_REG_ASQS_MASK) << NVME_AQA_REG_ASQS_SHIFT;
368 	nvme_mmio_write_4(ctrlr, aqa, aqa);
369 	DELAY(5000);
370 
371 	/* Initialization values for CC */
372 	cc = 0;
373 	cc |= 1 << NVME_CC_REG_EN_SHIFT;
374 	cc |= 0 << NVME_CC_REG_CSS_SHIFT;
375 	cc |= 0 << NVME_CC_REG_AMS_SHIFT;
376 	cc |= 0 << NVME_CC_REG_SHN_SHIFT;
377 	cc |= 6 << NVME_CC_REG_IOSQES_SHIFT; /* SQ entry size == 64 == 2^6 */
378 	cc |= 4 << NVME_CC_REG_IOCQES_SHIFT; /* CQ entry size == 16 == 2^4 */
379 
380 	/* This evaluates to 0, which is according to spec. */
381 	cc |= (PAGE_SIZE >> 13) << NVME_CC_REG_MPS_SHIFT;
382 
383 	nvme_mmio_write_4(ctrlr, cc, cc);
384 
385 	return (nvme_ctrlr_wait_for_ready(ctrlr, 1));
386 }
387 
388 static void
389 nvme_ctrlr_disable_qpairs(struct nvme_controller *ctrlr)
390 {
391 	int i;
392 
393 	nvme_admin_qpair_disable(&ctrlr->adminq);
394 	/*
395 	 * I/O queues are not allocated before the initial HW
396 	 *  reset, so do not try to disable them.  Use is_initialized
397 	 *  to determine if this is the initial HW reset.
398 	 */
399 	if (ctrlr->is_initialized) {
400 		for (i = 0; i < ctrlr->num_io_queues; i++)
401 			nvme_io_qpair_disable(&ctrlr->ioq[i]);
402 	}
403 }
404 
405 static int
406 nvme_ctrlr_hw_reset(struct nvme_controller *ctrlr)
407 {
408 	int err;
409 
410 	TSENTER();
411 	nvme_ctrlr_disable_qpairs(ctrlr);
412 
413 	pause("nvmehwreset", hz / 10);
414 
415 	err = nvme_ctrlr_disable(ctrlr);
416 	if (err != 0)
417 		return err;
418 	err = nvme_ctrlr_enable(ctrlr);
419 	TSEXIT();
420 	return (err);
421 }
422 
423 void
424 nvme_ctrlr_reset(struct nvme_controller *ctrlr)
425 {
426 	int cmpset;
427 
428 	cmpset = atomic_cmpset_32(&ctrlr->is_resetting, 0, 1);
429 
430 	if (cmpset == 0 || ctrlr->is_failed)
431 		/*
432 		 * Controller is already resetting or has failed.  Return
433 		 *  immediately since there is no need to kick off another
434 		 *  reset in these cases.
435 		 */
436 		return;
437 
438 	taskqueue_enqueue(ctrlr->taskqueue, &ctrlr->reset_task);
439 }
440 
441 static int
442 nvme_ctrlr_identify(struct nvme_controller *ctrlr)
443 {
444 	struct nvme_completion_poll_status	status;
445 
446 	status.done = 0;
447 	nvme_ctrlr_cmd_identify_controller(ctrlr, &ctrlr->cdata,
448 	    nvme_completion_poll_cb, &status);
449 	nvme_completion_poll(&status);
450 	if (nvme_completion_is_error(&status.cpl)) {
451 		nvme_printf(ctrlr, "nvme_identify_controller failed!\n");
452 		return (ENXIO);
453 	}
454 
455 	/* Convert data to host endian */
456 	nvme_controller_data_swapbytes(&ctrlr->cdata);
457 
458 	/*
459 	 * Use MDTS to ensure our default max_xfer_size doesn't exceed what the
460 	 *  controller supports.
461 	 */
462 	if (ctrlr->cdata.mdts > 0)
463 		ctrlr->max_xfer_size = min(ctrlr->max_xfer_size,
464 		    ctrlr->min_page_size * (1 << (ctrlr->cdata.mdts)));
465 
466 	return (0);
467 }
468 
469 static int
470 nvme_ctrlr_set_num_qpairs(struct nvme_controller *ctrlr)
471 {
472 	struct nvme_completion_poll_status	status;
473 	int					cq_allocated, sq_allocated;
474 
475 	status.done = 0;
476 	nvme_ctrlr_cmd_set_num_queues(ctrlr, ctrlr->num_io_queues,
477 	    nvme_completion_poll_cb, &status);
478 	nvme_completion_poll(&status);
479 	if (nvme_completion_is_error(&status.cpl)) {
480 		nvme_printf(ctrlr, "nvme_ctrlr_set_num_qpairs failed!\n");
481 		return (ENXIO);
482 	}
483 
484 	/*
485 	 * Data in cdw0 is 0-based.
486 	 * Lower 16-bits indicate number of submission queues allocated.
487 	 * Upper 16-bits indicate number of completion queues allocated.
488 	 */
489 	sq_allocated = (status.cpl.cdw0 & 0xFFFF) + 1;
490 	cq_allocated = (status.cpl.cdw0 >> 16) + 1;
491 
492 	/*
493 	 * Controller may allocate more queues than we requested,
494 	 *  so use the minimum of the number requested and what was
495 	 *  actually allocated.
496 	 */
497 	ctrlr->num_io_queues = min(ctrlr->num_io_queues, sq_allocated);
498 	ctrlr->num_io_queues = min(ctrlr->num_io_queues, cq_allocated);
499 	if (ctrlr->num_io_queues > vm_ndomains)
500 		ctrlr->num_io_queues -= ctrlr->num_io_queues % vm_ndomains;
501 
502 	return (0);
503 }
504 
505 static int
506 nvme_ctrlr_create_qpairs(struct nvme_controller *ctrlr)
507 {
508 	struct nvme_completion_poll_status	status;
509 	struct nvme_qpair			*qpair;
510 	int					i;
511 
512 	for (i = 0; i < ctrlr->num_io_queues; i++) {
513 		qpair = &ctrlr->ioq[i];
514 
515 		status.done = 0;
516 		nvme_ctrlr_cmd_create_io_cq(ctrlr, qpair,
517 		    nvme_completion_poll_cb, &status);
518 		nvme_completion_poll(&status);
519 		if (nvme_completion_is_error(&status.cpl)) {
520 			nvme_printf(ctrlr, "nvme_create_io_cq failed!\n");
521 			return (ENXIO);
522 		}
523 
524 		status.done = 0;
525 		nvme_ctrlr_cmd_create_io_sq(ctrlr, qpair,
526 		    nvme_completion_poll_cb, &status);
527 		nvme_completion_poll(&status);
528 		if (nvme_completion_is_error(&status.cpl)) {
529 			nvme_printf(ctrlr, "nvme_create_io_sq failed!\n");
530 			return (ENXIO);
531 		}
532 	}
533 
534 	return (0);
535 }
536 
537 static int
538 nvme_ctrlr_delete_qpairs(struct nvme_controller *ctrlr)
539 {
540 	struct nvme_completion_poll_status	status;
541 	struct nvme_qpair			*qpair;
542 
543 	for (int i = 0; i < ctrlr->num_io_queues; i++) {
544 		qpair = &ctrlr->ioq[i];
545 
546 		status.done = 0;
547 		nvme_ctrlr_cmd_delete_io_sq(ctrlr, qpair,
548 		    nvme_completion_poll_cb, &status);
549 		nvme_completion_poll(&status);
550 		if (nvme_completion_is_error(&status.cpl)) {
551 			nvme_printf(ctrlr, "nvme_destroy_io_sq failed!\n");
552 			return (ENXIO);
553 		}
554 
555 		status.done = 0;
556 		nvme_ctrlr_cmd_delete_io_cq(ctrlr, qpair,
557 		    nvme_completion_poll_cb, &status);
558 		nvme_completion_poll(&status);
559 		if (nvme_completion_is_error(&status.cpl)) {
560 			nvme_printf(ctrlr, "nvme_destroy_io_cq failed!\n");
561 			return (ENXIO);
562 		}
563 	}
564 
565 	return (0);
566 }
567 
568 static int
569 nvme_ctrlr_construct_namespaces(struct nvme_controller *ctrlr)
570 {
571 	struct nvme_namespace	*ns;
572 	uint32_t 		i;
573 
574 	for (i = 0; i < min(ctrlr->cdata.nn, NVME_MAX_NAMESPACES); i++) {
575 		ns = &ctrlr->ns[i];
576 		nvme_ns_construct(ns, i+1, ctrlr);
577 	}
578 
579 	return (0);
580 }
581 
582 static bool
583 is_log_page_id_valid(uint8_t page_id)
584 {
585 
586 	switch (page_id) {
587 	case NVME_LOG_ERROR:
588 	case NVME_LOG_HEALTH_INFORMATION:
589 	case NVME_LOG_FIRMWARE_SLOT:
590 	case NVME_LOG_CHANGED_NAMESPACE:
591 	case NVME_LOG_COMMAND_EFFECT:
592 	case NVME_LOG_RES_NOTIFICATION:
593 	case NVME_LOG_SANITIZE_STATUS:
594 		return (true);
595 	}
596 
597 	return (false);
598 }
599 
600 static uint32_t
601 nvme_ctrlr_get_log_page_size(struct nvme_controller *ctrlr, uint8_t page_id)
602 {
603 	uint32_t	log_page_size;
604 
605 	switch (page_id) {
606 	case NVME_LOG_ERROR:
607 		log_page_size = min(
608 		    sizeof(struct nvme_error_information_entry) *
609 		    (ctrlr->cdata.elpe + 1), NVME_MAX_AER_LOG_SIZE);
610 		break;
611 	case NVME_LOG_HEALTH_INFORMATION:
612 		log_page_size = sizeof(struct nvme_health_information_page);
613 		break;
614 	case NVME_LOG_FIRMWARE_SLOT:
615 		log_page_size = sizeof(struct nvme_firmware_page);
616 		break;
617 	case NVME_LOG_CHANGED_NAMESPACE:
618 		log_page_size = sizeof(struct nvme_ns_list);
619 		break;
620 	case NVME_LOG_COMMAND_EFFECT:
621 		log_page_size = sizeof(struct nvme_command_effects_page);
622 		break;
623 	case NVME_LOG_RES_NOTIFICATION:
624 		log_page_size = sizeof(struct nvme_res_notification_page);
625 		break;
626 	case NVME_LOG_SANITIZE_STATUS:
627 		log_page_size = sizeof(struct nvme_sanitize_status_page);
628 		break;
629 	default:
630 		log_page_size = 0;
631 		break;
632 	}
633 
634 	return (log_page_size);
635 }
636 
637 static void
638 nvme_ctrlr_log_critical_warnings(struct nvme_controller *ctrlr,
639     uint8_t state)
640 {
641 
642 	if (state & NVME_CRIT_WARN_ST_AVAILABLE_SPARE)
643 		nvme_ctrlr_devctl_log(ctrlr, "critical",
644 		    "available spare space below threshold");
645 
646 	if (state & NVME_CRIT_WARN_ST_TEMPERATURE)
647 		nvme_ctrlr_devctl_log(ctrlr, "critical",
648 		    "temperature above threshold");
649 
650 	if (state & NVME_CRIT_WARN_ST_DEVICE_RELIABILITY)
651 		nvme_ctrlr_devctl_log(ctrlr, "critical",
652 		    "device reliability degraded");
653 
654 	if (state & NVME_CRIT_WARN_ST_READ_ONLY)
655 		nvme_ctrlr_devctl_log(ctrlr, "critical",
656 		    "media placed in read only mode");
657 
658 	if (state & NVME_CRIT_WARN_ST_VOLATILE_MEMORY_BACKUP)
659 		nvme_ctrlr_devctl_log(ctrlr, "critical",
660 		    "volatile memory backup device failed");
661 
662 	if (state & NVME_CRIT_WARN_ST_RESERVED_MASK)
663 		nvme_ctrlr_devctl_log(ctrlr, "critical",
664 		    "unknown critical warning(s): state = 0x%02x", state);
665 }
666 
667 static void
668 nvme_ctrlr_async_event_log_page_cb(void *arg, const struct nvme_completion *cpl)
669 {
670 	struct nvme_async_event_request		*aer = arg;
671 	struct nvme_health_information_page	*health_info;
672 	struct nvme_ns_list			*nsl;
673 	struct nvme_error_information_entry	*err;
674 	int i;
675 
676 	/*
677 	 * If the log page fetch for some reason completed with an error,
678 	 *  don't pass log page data to the consumers.  In practice, this case
679 	 *  should never happen.
680 	 */
681 	if (nvme_completion_is_error(cpl))
682 		nvme_notify_async_consumers(aer->ctrlr, &aer->cpl,
683 		    aer->log_page_id, NULL, 0);
684 	else {
685 		/* Convert data to host endian */
686 		switch (aer->log_page_id) {
687 		case NVME_LOG_ERROR:
688 			err = (struct nvme_error_information_entry *)aer->log_page_buffer;
689 			for (i = 0; i < (aer->ctrlr->cdata.elpe + 1); i++)
690 				nvme_error_information_entry_swapbytes(err++);
691 			break;
692 		case NVME_LOG_HEALTH_INFORMATION:
693 			nvme_health_information_page_swapbytes(
694 			    (struct nvme_health_information_page *)aer->log_page_buffer);
695 			break;
696 		case NVME_LOG_FIRMWARE_SLOT:
697 			nvme_firmware_page_swapbytes(
698 			    (struct nvme_firmware_page *)aer->log_page_buffer);
699 			break;
700 		case NVME_LOG_CHANGED_NAMESPACE:
701 			nvme_ns_list_swapbytes(
702 			    (struct nvme_ns_list *)aer->log_page_buffer);
703 			break;
704 		case NVME_LOG_COMMAND_EFFECT:
705 			nvme_command_effects_page_swapbytes(
706 			    (struct nvme_command_effects_page *)aer->log_page_buffer);
707 			break;
708 		case NVME_LOG_RES_NOTIFICATION:
709 			nvme_res_notification_page_swapbytes(
710 			    (struct nvme_res_notification_page *)aer->log_page_buffer);
711 			break;
712 		case NVME_LOG_SANITIZE_STATUS:
713 			nvme_sanitize_status_page_swapbytes(
714 			    (struct nvme_sanitize_status_page *)aer->log_page_buffer);
715 			break;
716 		case INTEL_LOG_TEMP_STATS:
717 			intel_log_temp_stats_swapbytes(
718 			    (struct intel_log_temp_stats *)aer->log_page_buffer);
719 			break;
720 		default:
721 			break;
722 		}
723 
724 		if (aer->log_page_id == NVME_LOG_HEALTH_INFORMATION) {
725 			health_info = (struct nvme_health_information_page *)
726 			    aer->log_page_buffer;
727 			nvme_ctrlr_log_critical_warnings(aer->ctrlr,
728 			    health_info->critical_warning);
729 			/*
730 			 * Critical warnings reported through the
731 			 *  SMART/health log page are persistent, so
732 			 *  clear the associated bits in the async event
733 			 *  config so that we do not receive repeated
734 			 *  notifications for the same event.
735 			 */
736 			aer->ctrlr->async_event_config &=
737 			    ~health_info->critical_warning;
738 			nvme_ctrlr_cmd_set_async_event_config(aer->ctrlr,
739 			    aer->ctrlr->async_event_config, NULL, NULL);
740 		} else if (aer->log_page_id == NVME_LOG_CHANGED_NAMESPACE &&
741 		    !nvme_use_nvd) {
742 			nsl = (struct nvme_ns_list *)aer->log_page_buffer;
743 			for (i = 0; i < nitems(nsl->ns) && nsl->ns[i] != 0; i++) {
744 				if (nsl->ns[i] > NVME_MAX_NAMESPACES)
745 					break;
746 				nvme_notify_ns(aer->ctrlr, nsl->ns[i]);
747 			}
748 		}
749 
750 		/*
751 		 * Pass the cpl data from the original async event completion,
752 		 *  not the log page fetch.
753 		 */
754 		nvme_notify_async_consumers(aer->ctrlr, &aer->cpl,
755 		    aer->log_page_id, aer->log_page_buffer, aer->log_page_size);
756 	}
757 
758 	/*
759 	 * Repost another asynchronous event request to replace the one
760 	 *  that just completed.
761 	 */
762 	nvme_ctrlr_construct_and_submit_aer(aer->ctrlr, aer);
763 }
764 
765 static void
766 nvme_ctrlr_async_event_cb(void *arg, const struct nvme_completion *cpl)
767 {
768 	struct nvme_async_event_request	*aer = arg;
769 
770 	if (nvme_completion_is_error(cpl)) {
771 		/*
772 		 *  Do not retry failed async event requests.  This avoids
773 		 *  infinite loops where a new async event request is submitted
774 		 *  to replace the one just failed, only to fail again and
775 		 *  perpetuate the loop.
776 		 */
777 		return;
778 	}
779 
780 	/* Associated log page is in bits 23:16 of completion entry dw0. */
781 	aer->log_page_id = (cpl->cdw0 & 0xFF0000) >> 16;
782 
783 	nvme_printf(aer->ctrlr, "async event occurred (type 0x%x, info 0x%02x,"
784 	    " page 0x%02x)\n", (cpl->cdw0 & 0x07), (cpl->cdw0 & 0xFF00) >> 8,
785 	    aer->log_page_id);
786 
787 	if (is_log_page_id_valid(aer->log_page_id)) {
788 		aer->log_page_size = nvme_ctrlr_get_log_page_size(aer->ctrlr,
789 		    aer->log_page_id);
790 		memcpy(&aer->cpl, cpl, sizeof(*cpl));
791 		nvme_ctrlr_cmd_get_log_page(aer->ctrlr, aer->log_page_id,
792 		    NVME_GLOBAL_NAMESPACE_TAG, aer->log_page_buffer,
793 		    aer->log_page_size, nvme_ctrlr_async_event_log_page_cb,
794 		    aer);
795 		/* Wait to notify consumers until after log page is fetched. */
796 	} else {
797 		nvme_notify_async_consumers(aer->ctrlr, cpl, aer->log_page_id,
798 		    NULL, 0);
799 
800 		/*
801 		 * Repost another asynchronous event request to replace the one
802 		 *  that just completed.
803 		 */
804 		nvme_ctrlr_construct_and_submit_aer(aer->ctrlr, aer);
805 	}
806 }
807 
808 static void
809 nvme_ctrlr_construct_and_submit_aer(struct nvme_controller *ctrlr,
810     struct nvme_async_event_request *aer)
811 {
812 	struct nvme_request *req;
813 
814 	aer->ctrlr = ctrlr;
815 	req = nvme_allocate_request_null(nvme_ctrlr_async_event_cb, aer);
816 	aer->req = req;
817 
818 	/*
819 	 * Disable timeout here, since asynchronous event requests should by
820 	 *  nature never be timed out.
821 	 */
822 	req->timeout = false;
823 	req->cmd.opc = NVME_OPC_ASYNC_EVENT_REQUEST;
824 	nvme_ctrlr_submit_admin_request(ctrlr, req);
825 }
826 
827 static void
828 nvme_ctrlr_configure_aer(struct nvme_controller *ctrlr)
829 {
830 	struct nvme_completion_poll_status	status;
831 	struct nvme_async_event_request		*aer;
832 	uint32_t				i;
833 
834 	ctrlr->async_event_config = NVME_CRIT_WARN_ST_AVAILABLE_SPARE |
835 	    NVME_CRIT_WARN_ST_DEVICE_RELIABILITY |
836 	    NVME_CRIT_WARN_ST_READ_ONLY |
837 	    NVME_CRIT_WARN_ST_VOLATILE_MEMORY_BACKUP;
838 	if (ctrlr->cdata.ver >= NVME_REV(1, 2))
839 		ctrlr->async_event_config |= NVME_ASYNC_EVENT_NS_ATTRIBUTE |
840 		    NVME_ASYNC_EVENT_FW_ACTIVATE;
841 
842 	status.done = 0;
843 	nvme_ctrlr_cmd_get_feature(ctrlr, NVME_FEAT_TEMPERATURE_THRESHOLD,
844 	    0, NULL, 0, nvme_completion_poll_cb, &status);
845 	nvme_completion_poll(&status);
846 	if (nvme_completion_is_error(&status.cpl) ||
847 	    (status.cpl.cdw0 & 0xFFFF) == 0xFFFF ||
848 	    (status.cpl.cdw0 & 0xFFFF) == 0x0000) {
849 		nvme_printf(ctrlr, "temperature threshold not supported\n");
850 	} else
851 		ctrlr->async_event_config |= NVME_CRIT_WARN_ST_TEMPERATURE;
852 
853 	nvme_ctrlr_cmd_set_async_event_config(ctrlr,
854 	    ctrlr->async_event_config, NULL, NULL);
855 
856 	/* aerl is a zero-based value, so we need to add 1 here. */
857 	ctrlr->num_aers = min(NVME_MAX_ASYNC_EVENTS, (ctrlr->cdata.aerl+1));
858 
859 	for (i = 0; i < ctrlr->num_aers; i++) {
860 		aer = &ctrlr->aer[i];
861 		nvme_ctrlr_construct_and_submit_aer(ctrlr, aer);
862 	}
863 }
864 
865 static void
866 nvme_ctrlr_configure_int_coalescing(struct nvme_controller *ctrlr)
867 {
868 
869 	ctrlr->int_coal_time = 0;
870 	TUNABLE_INT_FETCH("hw.nvme.int_coal_time",
871 	    &ctrlr->int_coal_time);
872 
873 	ctrlr->int_coal_threshold = 0;
874 	TUNABLE_INT_FETCH("hw.nvme.int_coal_threshold",
875 	    &ctrlr->int_coal_threshold);
876 
877 	nvme_ctrlr_cmd_set_interrupt_coalescing(ctrlr, ctrlr->int_coal_time,
878 	    ctrlr->int_coal_threshold, NULL, NULL);
879 }
880 
881 static void
882 nvme_ctrlr_hmb_free(struct nvme_controller *ctrlr)
883 {
884 	struct nvme_hmb_chunk *hmbc;
885 	int i;
886 
887 	if (ctrlr->hmb_desc_paddr) {
888 		bus_dmamap_unload(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_map);
889 		bus_dmamem_free(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_vaddr,
890 		    ctrlr->hmb_desc_map);
891 		ctrlr->hmb_desc_paddr = 0;
892 	}
893 	if (ctrlr->hmb_desc_tag) {
894 		bus_dma_tag_destroy(ctrlr->hmb_desc_tag);
895 		ctrlr->hmb_desc_tag = NULL;
896 	}
897 	for (i = 0; i < ctrlr->hmb_nchunks; i++) {
898 		hmbc = &ctrlr->hmb_chunks[i];
899 		bus_dmamap_unload(ctrlr->hmb_tag, hmbc->hmbc_map);
900 		bus_dmamem_free(ctrlr->hmb_tag, hmbc->hmbc_vaddr,
901 		    hmbc->hmbc_map);
902 	}
903 	ctrlr->hmb_nchunks = 0;
904 	if (ctrlr->hmb_tag) {
905 		bus_dma_tag_destroy(ctrlr->hmb_tag);
906 		ctrlr->hmb_tag = NULL;
907 	}
908 	if (ctrlr->hmb_chunks) {
909 		free(ctrlr->hmb_chunks, M_NVME);
910 		ctrlr->hmb_chunks = NULL;
911 	}
912 }
913 
914 static void
915 nvme_ctrlr_hmb_alloc(struct nvme_controller *ctrlr)
916 {
917 	struct nvme_hmb_chunk *hmbc;
918 	size_t pref, min, minc, size;
919 	int err, i;
920 	uint64_t max;
921 
922 	/* Limit HMB to 5% of RAM size per device by default. */
923 	max = (uint64_t)physmem * PAGE_SIZE / 20;
924 	TUNABLE_UINT64_FETCH("hw.nvme.hmb_max", &max);
925 
926 	min = (long long unsigned)ctrlr->cdata.hmmin * 4096;
927 	if (max == 0 || max < min)
928 		return;
929 	pref = MIN((long long unsigned)ctrlr->cdata.hmpre * 4096, max);
930 	minc = MAX(ctrlr->cdata.hmminds * 4096, PAGE_SIZE);
931 	if (min > 0 && ctrlr->cdata.hmmaxd > 0)
932 		minc = MAX(minc, min / ctrlr->cdata.hmmaxd);
933 	ctrlr->hmb_chunk = pref;
934 
935 again:
936 	ctrlr->hmb_chunk = roundup2(ctrlr->hmb_chunk, PAGE_SIZE);
937 	ctrlr->hmb_nchunks = howmany(pref, ctrlr->hmb_chunk);
938 	if (ctrlr->cdata.hmmaxd > 0 && ctrlr->hmb_nchunks > ctrlr->cdata.hmmaxd)
939 		ctrlr->hmb_nchunks = ctrlr->cdata.hmmaxd;
940 	ctrlr->hmb_chunks = malloc(sizeof(struct nvme_hmb_chunk) *
941 	    ctrlr->hmb_nchunks, M_NVME, M_WAITOK);
942 	err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev),
943 	    PAGE_SIZE, 0, BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
944 	    ctrlr->hmb_chunk, 1, ctrlr->hmb_chunk, 0, NULL, NULL, &ctrlr->hmb_tag);
945 	if (err != 0) {
946 		nvme_printf(ctrlr, "HMB tag create failed %d\n", err);
947 		nvme_ctrlr_hmb_free(ctrlr);
948 		return;
949 	}
950 
951 	for (i = 0; i < ctrlr->hmb_nchunks; i++) {
952 		hmbc = &ctrlr->hmb_chunks[i];
953 		if (bus_dmamem_alloc(ctrlr->hmb_tag,
954 		    (void **)&hmbc->hmbc_vaddr, BUS_DMA_NOWAIT,
955 		    &hmbc->hmbc_map)) {
956 			nvme_printf(ctrlr, "failed to alloc HMB\n");
957 			break;
958 		}
959 		if (bus_dmamap_load(ctrlr->hmb_tag, hmbc->hmbc_map,
960 		    hmbc->hmbc_vaddr, ctrlr->hmb_chunk, nvme_single_map,
961 		    &hmbc->hmbc_paddr, BUS_DMA_NOWAIT) != 0) {
962 			bus_dmamem_free(ctrlr->hmb_tag, hmbc->hmbc_vaddr,
963 			    hmbc->hmbc_map);
964 			nvme_printf(ctrlr, "failed to load HMB\n");
965 			break;
966 		}
967 		bus_dmamap_sync(ctrlr->hmb_tag, hmbc->hmbc_map,
968 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
969 	}
970 
971 	if (i < ctrlr->hmb_nchunks && i * ctrlr->hmb_chunk < min &&
972 	    ctrlr->hmb_chunk / 2 >= minc) {
973 		ctrlr->hmb_nchunks = i;
974 		nvme_ctrlr_hmb_free(ctrlr);
975 		ctrlr->hmb_chunk /= 2;
976 		goto again;
977 	}
978 	ctrlr->hmb_nchunks = i;
979 	if (ctrlr->hmb_nchunks * ctrlr->hmb_chunk < min) {
980 		nvme_ctrlr_hmb_free(ctrlr);
981 		return;
982 	}
983 
984 	size = sizeof(struct nvme_hmb_desc) * ctrlr->hmb_nchunks;
985 	err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev),
986 	    16, 0, BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
987 	    size, 1, size, 0, NULL, NULL, &ctrlr->hmb_desc_tag);
988 	if (err != 0) {
989 		nvme_printf(ctrlr, "HMB desc tag create failed %d\n", err);
990 		nvme_ctrlr_hmb_free(ctrlr);
991 		return;
992 	}
993 	if (bus_dmamem_alloc(ctrlr->hmb_desc_tag,
994 	    (void **)&ctrlr->hmb_desc_vaddr, BUS_DMA_WAITOK,
995 	    &ctrlr->hmb_desc_map)) {
996 		nvme_printf(ctrlr, "failed to alloc HMB desc\n");
997 		nvme_ctrlr_hmb_free(ctrlr);
998 		return;
999 	}
1000 	if (bus_dmamap_load(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_map,
1001 	    ctrlr->hmb_desc_vaddr, size, nvme_single_map,
1002 	    &ctrlr->hmb_desc_paddr, BUS_DMA_NOWAIT) != 0) {
1003 		bus_dmamem_free(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_vaddr,
1004 		    ctrlr->hmb_desc_map);
1005 		nvme_printf(ctrlr, "failed to load HMB desc\n");
1006 		nvme_ctrlr_hmb_free(ctrlr);
1007 		return;
1008 	}
1009 
1010 	for (i = 0; i < ctrlr->hmb_nchunks; i++) {
1011 		ctrlr->hmb_desc_vaddr[i].addr =
1012 		    htole64(ctrlr->hmb_chunks[i].hmbc_paddr);
1013 		ctrlr->hmb_desc_vaddr[i].size = htole32(ctrlr->hmb_chunk / 4096);
1014 	}
1015 	bus_dmamap_sync(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_map,
1016 	    BUS_DMASYNC_PREWRITE);
1017 
1018 	nvme_printf(ctrlr, "Allocated %lluMB host memory buffer\n",
1019 	    (long long unsigned)ctrlr->hmb_nchunks * ctrlr->hmb_chunk
1020 	    / 1024 / 1024);
1021 }
1022 
1023 static void
1024 nvme_ctrlr_hmb_enable(struct nvme_controller *ctrlr, bool enable, bool memret)
1025 {
1026 	struct nvme_completion_poll_status	status;
1027 	uint32_t cdw11;
1028 
1029 	cdw11 = 0;
1030 	if (enable)
1031 		cdw11 |= 1;
1032 	if (memret)
1033 		cdw11 |= 2;
1034 	status.done = 0;
1035 	nvme_ctrlr_cmd_set_feature(ctrlr, NVME_FEAT_HOST_MEMORY_BUFFER, cdw11,
1036 	    ctrlr->hmb_nchunks * ctrlr->hmb_chunk / 4096, ctrlr->hmb_desc_paddr,
1037 	    ctrlr->hmb_desc_paddr >> 32, ctrlr->hmb_nchunks, NULL, 0,
1038 	    nvme_completion_poll_cb, &status);
1039 	nvme_completion_poll(&status);
1040 	if (nvme_completion_is_error(&status.cpl))
1041 		nvme_printf(ctrlr, "nvme_ctrlr_hmb_enable failed!\n");
1042 }
1043 
1044 static void
1045 nvme_ctrlr_start(void *ctrlr_arg, bool resetting)
1046 {
1047 	struct nvme_controller *ctrlr = ctrlr_arg;
1048 	uint32_t old_num_io_queues;
1049 	int i;
1050 
1051 	TSENTER();
1052 
1053 	/*
1054 	 * Only reset adminq here when we are restarting the
1055 	 *  controller after a reset.  During initialization,
1056 	 *  we have already submitted admin commands to get
1057 	 *  the number of I/O queues supported, so cannot reset
1058 	 *  the adminq again here.
1059 	 */
1060 	if (resetting) {
1061 		nvme_qpair_reset(&ctrlr->adminq);
1062 		nvme_admin_qpair_enable(&ctrlr->adminq);
1063 	}
1064 
1065 	if (ctrlr->ioq != NULL) {
1066 		for (i = 0; i < ctrlr->num_io_queues; i++)
1067 			nvme_qpair_reset(&ctrlr->ioq[i]);
1068 	}
1069 
1070 	/*
1071 	 * If it was a reset on initialization command timeout, just
1072 	 * return here, letting initialization code fail gracefully.
1073 	 */
1074 	if (resetting && !ctrlr->is_initialized)
1075 		return;
1076 
1077 	if (resetting && nvme_ctrlr_identify(ctrlr) != 0) {
1078 		nvme_ctrlr_fail(ctrlr);
1079 		return;
1080 	}
1081 
1082 	/*
1083 	 * The number of qpairs are determined during controller initialization,
1084 	 *  including using NVMe SET_FEATURES/NUMBER_OF_QUEUES to determine the
1085 	 *  HW limit.  We call SET_FEATURES again here so that it gets called
1086 	 *  after any reset for controllers that depend on the driver to
1087 	 *  explicit specify how many queues it will use.  This value should
1088 	 *  never change between resets, so panic if somehow that does happen.
1089 	 */
1090 	if (resetting) {
1091 		old_num_io_queues = ctrlr->num_io_queues;
1092 		if (nvme_ctrlr_set_num_qpairs(ctrlr) != 0) {
1093 			nvme_ctrlr_fail(ctrlr);
1094 			return;
1095 		}
1096 
1097 		if (old_num_io_queues != ctrlr->num_io_queues) {
1098 			panic("num_io_queues changed from %u to %u",
1099 			      old_num_io_queues, ctrlr->num_io_queues);
1100 		}
1101 	}
1102 
1103 	if (ctrlr->cdata.hmpre > 0 && ctrlr->hmb_nchunks == 0) {
1104 		nvme_ctrlr_hmb_alloc(ctrlr);
1105 		if (ctrlr->hmb_nchunks > 0)
1106 			nvme_ctrlr_hmb_enable(ctrlr, true, false);
1107 	} else if (ctrlr->hmb_nchunks > 0)
1108 		nvme_ctrlr_hmb_enable(ctrlr, true, true);
1109 
1110 	if (nvme_ctrlr_create_qpairs(ctrlr) != 0) {
1111 		nvme_ctrlr_fail(ctrlr);
1112 		return;
1113 	}
1114 
1115 	if (nvme_ctrlr_construct_namespaces(ctrlr) != 0) {
1116 		nvme_ctrlr_fail(ctrlr);
1117 		return;
1118 	}
1119 
1120 	nvme_ctrlr_configure_aer(ctrlr);
1121 	nvme_ctrlr_configure_int_coalescing(ctrlr);
1122 
1123 	for (i = 0; i < ctrlr->num_io_queues; i++)
1124 		nvme_io_qpair_enable(&ctrlr->ioq[i]);
1125 	TSEXIT();
1126 }
1127 
1128 void
1129 nvme_ctrlr_start_config_hook(void *arg)
1130 {
1131 	struct nvme_controller *ctrlr = arg;
1132 
1133 	TSENTER();
1134 
1135 	/*
1136 	 * Reset controller twice to ensure we do a transition from cc.en==1 to
1137 	 * cc.en==0.  This is because we don't really know what status the
1138 	 * controller was left in when boot handed off to OS.  Linux doesn't do
1139 	 * this, however. If we adopt that policy, see also nvme_ctrlr_resume().
1140 	 */
1141 	if (nvme_ctrlr_hw_reset(ctrlr) != 0) {
1142 fail:
1143 		nvme_ctrlr_fail(ctrlr);
1144 		config_intrhook_disestablish(&ctrlr->config_hook);
1145 		return;
1146 	}
1147 
1148 	if (nvme_ctrlr_hw_reset(ctrlr) != 0)
1149 		goto fail;
1150 
1151 	nvme_qpair_reset(&ctrlr->adminq);
1152 	nvme_admin_qpair_enable(&ctrlr->adminq);
1153 
1154 	if (nvme_ctrlr_identify(ctrlr) == 0 &&
1155 	    nvme_ctrlr_set_num_qpairs(ctrlr) == 0 &&
1156 	    nvme_ctrlr_construct_io_qpairs(ctrlr) == 0)
1157 		nvme_ctrlr_start(ctrlr, false);
1158 	else
1159 		goto fail;
1160 
1161 	nvme_sysctl_initialize_ctrlr(ctrlr);
1162 	config_intrhook_disestablish(&ctrlr->config_hook);
1163 
1164 	ctrlr->is_initialized = 1;
1165 	nvme_notify_new_controller(ctrlr);
1166 	TSEXIT();
1167 }
1168 
1169 static void
1170 nvme_ctrlr_reset_task(void *arg, int pending)
1171 {
1172 	struct nvme_controller	*ctrlr = arg;
1173 	int			status;
1174 
1175 	nvme_ctrlr_devctl_log(ctrlr, "RESET", "resetting controller");
1176 	status = nvme_ctrlr_hw_reset(ctrlr);
1177 	/*
1178 	 * Use pause instead of DELAY, so that we yield to any nvme interrupt
1179 	 *  handlers on this CPU that were blocked on a qpair lock. We want
1180 	 *  all nvme interrupts completed before proceeding with restarting the
1181 	 *  controller.
1182 	 *
1183 	 * XXX - any way to guarantee the interrupt handlers have quiesced?
1184 	 */
1185 	pause("nvmereset", hz / 10);
1186 	if (status == 0)
1187 		nvme_ctrlr_start(ctrlr, true);
1188 	else
1189 		nvme_ctrlr_fail(ctrlr);
1190 
1191 	atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
1192 }
1193 
1194 /*
1195  * Poll all the queues enabled on the device for completion.
1196  */
1197 void
1198 nvme_ctrlr_poll(struct nvme_controller *ctrlr)
1199 {
1200 	int i;
1201 
1202 	nvme_qpair_process_completions(&ctrlr->adminq);
1203 
1204 	for (i = 0; i < ctrlr->num_io_queues; i++)
1205 		if (ctrlr->ioq && ctrlr->ioq[i].cpl)
1206 			nvme_qpair_process_completions(&ctrlr->ioq[i]);
1207 }
1208 
1209 /*
1210  * Poll the single-vector interrupt case: num_io_queues will be 1 and
1211  * there's only a single vector. While we're polling, we mask further
1212  * interrupts in the controller.
1213  */
1214 void
1215 nvme_ctrlr_shared_handler(void *arg)
1216 {
1217 	struct nvme_controller *ctrlr = arg;
1218 
1219 	nvme_mmio_write_4(ctrlr, intms, 1);
1220 	nvme_ctrlr_poll(ctrlr);
1221 	nvme_mmio_write_4(ctrlr, intmc, 1);
1222 }
1223 
1224 static void
1225 nvme_pt_done(void *arg, const struct nvme_completion *cpl)
1226 {
1227 	struct nvme_pt_command *pt = arg;
1228 	struct mtx *mtx = pt->driver_lock;
1229 	uint16_t status;
1230 
1231 	bzero(&pt->cpl, sizeof(pt->cpl));
1232 	pt->cpl.cdw0 = cpl->cdw0;
1233 
1234 	status = cpl->status;
1235 	status &= ~NVME_STATUS_P_MASK;
1236 	pt->cpl.status = status;
1237 
1238 	mtx_lock(mtx);
1239 	pt->driver_lock = NULL;
1240 	wakeup(pt);
1241 	mtx_unlock(mtx);
1242 }
1243 
1244 int
1245 nvme_ctrlr_passthrough_cmd(struct nvme_controller *ctrlr,
1246     struct nvme_pt_command *pt, uint32_t nsid, int is_user_buffer,
1247     int is_admin_cmd)
1248 {
1249 	struct nvme_request	*req;
1250 	struct mtx		*mtx;
1251 	struct buf		*buf = NULL;
1252 	int			ret = 0;
1253 
1254 	if (pt->len > 0) {
1255 		if (pt->len > ctrlr->max_xfer_size) {
1256 			nvme_printf(ctrlr, "pt->len (%d) "
1257 			    "exceeds max_xfer_size (%d)\n", pt->len,
1258 			    ctrlr->max_xfer_size);
1259 			return EIO;
1260 		}
1261 		if (is_user_buffer) {
1262 			/*
1263 			 * Ensure the user buffer is wired for the duration of
1264 			 *  this pass-through command.
1265 			 */
1266 			PHOLD(curproc);
1267 			buf = uma_zalloc(pbuf_zone, M_WAITOK);
1268 			buf->b_iocmd = pt->is_read ? BIO_READ : BIO_WRITE;
1269 			if (vmapbuf(buf, pt->buf, pt->len, 1) < 0) {
1270 				ret = EFAULT;
1271 				goto err;
1272 			}
1273 			req = nvme_allocate_request_vaddr(buf->b_data, pt->len,
1274 			    nvme_pt_done, pt);
1275 		} else
1276 			req = nvme_allocate_request_vaddr(pt->buf, pt->len,
1277 			    nvme_pt_done, pt);
1278 	} else
1279 		req = nvme_allocate_request_null(nvme_pt_done, pt);
1280 
1281 	/* Assume user space already converted to little-endian */
1282 	req->cmd.opc = pt->cmd.opc;
1283 	req->cmd.fuse = pt->cmd.fuse;
1284 	req->cmd.rsvd2 = pt->cmd.rsvd2;
1285 	req->cmd.rsvd3 = pt->cmd.rsvd3;
1286 	req->cmd.cdw10 = pt->cmd.cdw10;
1287 	req->cmd.cdw11 = pt->cmd.cdw11;
1288 	req->cmd.cdw12 = pt->cmd.cdw12;
1289 	req->cmd.cdw13 = pt->cmd.cdw13;
1290 	req->cmd.cdw14 = pt->cmd.cdw14;
1291 	req->cmd.cdw15 = pt->cmd.cdw15;
1292 
1293 	req->cmd.nsid = htole32(nsid);
1294 
1295 	mtx = mtx_pool_find(mtxpool_sleep, pt);
1296 	pt->driver_lock = mtx;
1297 
1298 	if (is_admin_cmd)
1299 		nvme_ctrlr_submit_admin_request(ctrlr, req);
1300 	else
1301 		nvme_ctrlr_submit_io_request(ctrlr, req);
1302 
1303 	mtx_lock(mtx);
1304 	while (pt->driver_lock != NULL)
1305 		mtx_sleep(pt, mtx, PRIBIO, "nvme_pt", 0);
1306 	mtx_unlock(mtx);
1307 
1308 err:
1309 	if (buf != NULL) {
1310 		uma_zfree(pbuf_zone, buf);
1311 		PRELE(curproc);
1312 	}
1313 
1314 	return (ret);
1315 }
1316 
1317 static int
1318 nvme_ctrlr_ioctl(struct cdev *cdev, u_long cmd, caddr_t arg, int flag,
1319     struct thread *td)
1320 {
1321 	struct nvme_controller			*ctrlr;
1322 	struct nvme_pt_command			*pt;
1323 
1324 	ctrlr = cdev->si_drv1;
1325 
1326 	switch (cmd) {
1327 	case NVME_RESET_CONTROLLER:
1328 		nvme_ctrlr_reset(ctrlr);
1329 		break;
1330 	case NVME_PASSTHROUGH_CMD:
1331 		pt = (struct nvme_pt_command *)arg;
1332 		return (nvme_ctrlr_passthrough_cmd(ctrlr, pt, le32toh(pt->cmd.nsid),
1333 		    1 /* is_user_buffer */, 1 /* is_admin_cmd */));
1334 	case NVME_GET_NSID:
1335 	{
1336 		struct nvme_get_nsid *gnsid = (struct nvme_get_nsid *)arg;
1337 		strncpy(gnsid->cdev, device_get_nameunit(ctrlr->dev),
1338 		    sizeof(gnsid->cdev));
1339 		gnsid->cdev[sizeof(gnsid->cdev) - 1] = '\0';
1340 		gnsid->nsid = 0;
1341 		break;
1342 	}
1343 	case NVME_GET_MAX_XFER_SIZE:
1344 		*(uint64_t *)arg = ctrlr->max_xfer_size;
1345 		break;
1346 	default:
1347 		return (ENOTTY);
1348 	}
1349 
1350 	return (0);
1351 }
1352 
1353 static struct cdevsw nvme_ctrlr_cdevsw = {
1354 	.d_version =	D_VERSION,
1355 	.d_flags =	0,
1356 	.d_ioctl =	nvme_ctrlr_ioctl
1357 };
1358 
1359 int
1360 nvme_ctrlr_construct(struct nvme_controller *ctrlr, device_t dev)
1361 {
1362 	struct make_dev_args	md_args;
1363 	uint32_t	cap_lo;
1364 	uint32_t	cap_hi;
1365 	uint32_t	to, vs, pmrcap;
1366 	uint8_t		mpsmin;
1367 	int		status, timeout_period;
1368 
1369 	ctrlr->dev = dev;
1370 
1371 	mtx_init(&ctrlr->lock, "nvme ctrlr lock", NULL, MTX_DEF);
1372 	if (bus_get_domain(dev, &ctrlr->domain) != 0)
1373 		ctrlr->domain = 0;
1374 
1375 	cap_lo = nvme_mmio_read_4(ctrlr, cap_lo);
1376 	if (bootverbose) {
1377 		device_printf(dev, "CapLo: 0x%08x: MQES %u%s%s%s%s, TO %u\n",
1378 		    cap_lo, NVME_CAP_LO_MQES(cap_lo),
1379 		    NVME_CAP_LO_CQR(cap_lo) ? ", CQR" : "",
1380 		    NVME_CAP_LO_AMS(cap_lo) ? ", AMS" : "",
1381 		    (NVME_CAP_LO_AMS(cap_lo) & 0x1) ? " WRRwUPC" : "",
1382 		    (NVME_CAP_LO_AMS(cap_lo) & 0x2) ? " VS" : "",
1383 		    NVME_CAP_LO_TO(cap_lo));
1384 	}
1385 	cap_hi = nvme_mmio_read_4(ctrlr, cap_hi);
1386 	if (bootverbose) {
1387 		device_printf(dev, "CapHi: 0x%08x: DSTRD %u%s, CSS %x%s, "
1388 		    "MPSMIN %u, MPSMAX %u%s%s\n", cap_hi,
1389 		    NVME_CAP_HI_DSTRD(cap_hi),
1390 		    NVME_CAP_HI_NSSRS(cap_hi) ? ", NSSRS" : "",
1391 		    NVME_CAP_HI_CSS(cap_hi),
1392 		    NVME_CAP_HI_BPS(cap_hi) ? ", BPS" : "",
1393 		    NVME_CAP_HI_MPSMIN(cap_hi),
1394 		    NVME_CAP_HI_MPSMAX(cap_hi),
1395 		    NVME_CAP_HI_PMRS(cap_hi) ? ", PMRS" : "",
1396 		    NVME_CAP_HI_CMBS(cap_hi) ? ", CMBS" : "");
1397 	}
1398 	if (bootverbose) {
1399 		vs = nvme_mmio_read_4(ctrlr, vs);
1400 		device_printf(dev, "Version: 0x%08x: %d.%d\n", vs,
1401 		    NVME_MAJOR(vs), NVME_MINOR(vs));
1402 	}
1403 	if (bootverbose && NVME_CAP_HI_PMRS(cap_hi)) {
1404 		pmrcap = nvme_mmio_read_4(ctrlr, pmrcap);
1405 		device_printf(dev, "PMRCap: 0x%08x: BIR %u%s%s, PMRTU %u, "
1406 		    "PMRWBM %x, PMRTO %u%s\n", pmrcap,
1407 		    NVME_PMRCAP_BIR(pmrcap),
1408 		    NVME_PMRCAP_RDS(pmrcap) ? ", RDS" : "",
1409 		    NVME_PMRCAP_WDS(pmrcap) ? ", WDS" : "",
1410 		    NVME_PMRCAP_PMRTU(pmrcap),
1411 		    NVME_PMRCAP_PMRWBM(pmrcap),
1412 		    NVME_PMRCAP_PMRTO(pmrcap),
1413 		    NVME_PMRCAP_CMSS(pmrcap) ? ", CMSS" : "");
1414 	}
1415 
1416 	ctrlr->dstrd = NVME_CAP_HI_DSTRD(cap_hi) + 2;
1417 
1418 	mpsmin = NVME_CAP_HI_MPSMIN(cap_hi);
1419 	ctrlr->min_page_size = 1 << (12 + mpsmin);
1420 
1421 	/* Get ready timeout value from controller, in units of 500ms. */
1422 	to = NVME_CAP_LO_TO(cap_lo) + 1;
1423 	ctrlr->ready_timeout_in_ms = to * 500;
1424 
1425 	timeout_period = NVME_DEFAULT_TIMEOUT_PERIOD;
1426 	TUNABLE_INT_FETCH("hw.nvme.timeout_period", &timeout_period);
1427 	timeout_period = min(timeout_period, NVME_MAX_TIMEOUT_PERIOD);
1428 	timeout_period = max(timeout_period, NVME_MIN_TIMEOUT_PERIOD);
1429 	ctrlr->timeout_period = timeout_period;
1430 
1431 	nvme_retry_count = NVME_DEFAULT_RETRY_COUNT;
1432 	TUNABLE_INT_FETCH("hw.nvme.retry_count", &nvme_retry_count);
1433 
1434 	ctrlr->enable_aborts = 0;
1435 	TUNABLE_INT_FETCH("hw.nvme.enable_aborts", &ctrlr->enable_aborts);
1436 
1437 	ctrlr->max_xfer_size = NVME_MAX_XFER_SIZE;
1438 	if (nvme_ctrlr_construct_admin_qpair(ctrlr) != 0)
1439 		return (ENXIO);
1440 
1441 	/*
1442 	 * Create 2 threads for the taskqueue. The reset thread will block when
1443 	 * it detects that the controller has failed until all I/O has been
1444 	 * failed up the stack. The fail_req task needs to be able to run in
1445 	 * this case to finish the request failure for some cases.
1446 	 *
1447 	 * We could partially solve this race by draining the failed requeust
1448 	 * queue before proceding to free the sim, though nothing would stop
1449 	 * new I/O from coming in after we do that drain, but before we reach
1450 	 * cam_sim_free, so this big hammer is used instead.
1451 	 */
1452 	ctrlr->taskqueue = taskqueue_create("nvme_taskq", M_WAITOK,
1453 	    taskqueue_thread_enqueue, &ctrlr->taskqueue);
1454 	taskqueue_start_threads(&ctrlr->taskqueue, 2, PI_DISK, "nvme taskq");
1455 
1456 	ctrlr->is_resetting = 0;
1457 	ctrlr->is_initialized = 0;
1458 	ctrlr->notification_sent = 0;
1459 	TASK_INIT(&ctrlr->reset_task, 0, nvme_ctrlr_reset_task, ctrlr);
1460 	TASK_INIT(&ctrlr->fail_req_task, 0, nvme_ctrlr_fail_req_task, ctrlr);
1461 	STAILQ_INIT(&ctrlr->fail_req);
1462 	ctrlr->is_failed = false;
1463 
1464 	make_dev_args_init(&md_args);
1465 	md_args.mda_devsw = &nvme_ctrlr_cdevsw;
1466 	md_args.mda_uid = UID_ROOT;
1467 	md_args.mda_gid = GID_WHEEL;
1468 	md_args.mda_mode = 0600;
1469 	md_args.mda_unit = device_get_unit(dev);
1470 	md_args.mda_si_drv1 = (void *)ctrlr;
1471 	status = make_dev_s(&md_args, &ctrlr->cdev, "nvme%d",
1472 	    device_get_unit(dev));
1473 	if (status != 0)
1474 		return (ENXIO);
1475 
1476 	return (0);
1477 }
1478 
1479 void
1480 nvme_ctrlr_destruct(struct nvme_controller *ctrlr, device_t dev)
1481 {
1482 	int	gone, i;
1483 
1484 	if (ctrlr->resource == NULL)
1485 		goto nores;
1486 	if (!mtx_initialized(&ctrlr->adminq.lock))
1487 		goto noadminq;
1488 
1489 	/*
1490 	 * Check whether it is a hot unplug or a clean driver detach.
1491 	 * If device is not there any more, skip any shutdown commands.
1492 	 */
1493 	gone = (nvme_mmio_read_4(ctrlr, csts) == NVME_GONE);
1494 	if (gone)
1495 		nvme_ctrlr_fail(ctrlr);
1496 	else
1497 		nvme_notify_fail_consumers(ctrlr);
1498 
1499 	for (i = 0; i < NVME_MAX_NAMESPACES; i++)
1500 		nvme_ns_destruct(&ctrlr->ns[i]);
1501 
1502 	if (ctrlr->cdev)
1503 		destroy_dev(ctrlr->cdev);
1504 
1505 	if (ctrlr->is_initialized) {
1506 		if (!gone) {
1507 			if (ctrlr->hmb_nchunks > 0)
1508 				nvme_ctrlr_hmb_enable(ctrlr, false, false);
1509 			nvme_ctrlr_delete_qpairs(ctrlr);
1510 		}
1511 		nvme_ctrlr_hmb_free(ctrlr);
1512 	}
1513 	if (ctrlr->ioq != NULL) {
1514 		for (i = 0; i < ctrlr->num_io_queues; i++)
1515 			nvme_io_qpair_destroy(&ctrlr->ioq[i]);
1516 		free(ctrlr->ioq, M_NVME);
1517 	}
1518 	nvme_admin_qpair_destroy(&ctrlr->adminq);
1519 
1520 	/*
1521 	 *  Notify the controller of a shutdown, even though this is due to
1522 	 *   a driver unload, not a system shutdown (this path is not invoked
1523 	 *   during shutdown).  This ensures the controller receives a
1524 	 *   shutdown notification in case the system is shutdown before
1525 	 *   reloading the driver.
1526 	 */
1527 	if (!gone)
1528 		nvme_ctrlr_shutdown(ctrlr);
1529 
1530 	if (!gone)
1531 		nvme_ctrlr_disable(ctrlr);
1532 
1533 noadminq:
1534 	if (ctrlr->taskqueue)
1535 		taskqueue_free(ctrlr->taskqueue);
1536 
1537 	if (ctrlr->tag)
1538 		bus_teardown_intr(ctrlr->dev, ctrlr->res, ctrlr->tag);
1539 
1540 	if (ctrlr->res)
1541 		bus_release_resource(ctrlr->dev, SYS_RES_IRQ,
1542 		    rman_get_rid(ctrlr->res), ctrlr->res);
1543 
1544 	if (ctrlr->bar4_resource != NULL) {
1545 		bus_release_resource(dev, SYS_RES_MEMORY,
1546 		    ctrlr->bar4_resource_id, ctrlr->bar4_resource);
1547 	}
1548 
1549 	bus_release_resource(dev, SYS_RES_MEMORY,
1550 	    ctrlr->resource_id, ctrlr->resource);
1551 
1552 nores:
1553 	mtx_destroy(&ctrlr->lock);
1554 }
1555 
1556 void
1557 nvme_ctrlr_shutdown(struct nvme_controller *ctrlr)
1558 {
1559 	uint32_t	cc;
1560 	uint32_t	csts;
1561 	int		timeout;
1562 
1563 	cc = nvme_mmio_read_4(ctrlr, cc);
1564 	cc &= ~(NVME_CC_REG_SHN_MASK << NVME_CC_REG_SHN_SHIFT);
1565 	cc |= NVME_SHN_NORMAL << NVME_CC_REG_SHN_SHIFT;
1566 	nvme_mmio_write_4(ctrlr, cc, cc);
1567 
1568 	timeout = ticks + (ctrlr->cdata.rtd3e == 0 ? 5 * hz :
1569 	    ((uint64_t)ctrlr->cdata.rtd3e * hz + 999999) / 1000000);
1570 	while (1) {
1571 		csts = nvme_mmio_read_4(ctrlr, csts);
1572 		if (csts == NVME_GONE)		/* Hot unplug. */
1573 			break;
1574 		if (NVME_CSTS_GET_SHST(csts) == NVME_SHST_COMPLETE)
1575 			break;
1576 		if (timeout - ticks < 0) {
1577 			nvme_printf(ctrlr, "shutdown timeout\n");
1578 			break;
1579 		}
1580 		pause("nvmeshut", 1);
1581 	}
1582 }
1583 
1584 void
1585 nvme_ctrlr_submit_admin_request(struct nvme_controller *ctrlr,
1586     struct nvme_request *req)
1587 {
1588 
1589 	nvme_qpair_submit_request(&ctrlr->adminq, req);
1590 }
1591 
1592 void
1593 nvme_ctrlr_submit_io_request(struct nvme_controller *ctrlr,
1594     struct nvme_request *req)
1595 {
1596 	struct nvme_qpair       *qpair;
1597 
1598 	qpair = &ctrlr->ioq[QP(ctrlr, curcpu)];
1599 	nvme_qpair_submit_request(qpair, req);
1600 }
1601 
1602 device_t
1603 nvme_ctrlr_get_device(struct nvme_controller *ctrlr)
1604 {
1605 
1606 	return (ctrlr->dev);
1607 }
1608 
1609 const struct nvme_controller_data *
1610 nvme_ctrlr_get_data(struct nvme_controller *ctrlr)
1611 {
1612 
1613 	return (&ctrlr->cdata);
1614 }
1615 
1616 int
1617 nvme_ctrlr_suspend(struct nvme_controller *ctrlr)
1618 {
1619 	int to = hz;
1620 
1621 	/*
1622 	 * Can't touch failed controllers, so it's already suspended.
1623 	 */
1624 	if (ctrlr->is_failed)
1625 		return (0);
1626 
1627 	/*
1628 	 * We don't want the reset taskqueue running, since it does similar
1629 	 * things, so prevent it from running after we start. Wait for any reset
1630 	 * that may have been started to complete. The reset process we follow
1631 	 * will ensure that any new I/O will queue and be given to the hardware
1632 	 * after we resume (though there should be none).
1633 	 */
1634 	while (atomic_cmpset_32(&ctrlr->is_resetting, 0, 1) == 0 && to-- > 0)
1635 		pause("nvmesusp", 1);
1636 	if (to <= 0) {
1637 		nvme_printf(ctrlr,
1638 		    "Competing reset task didn't finish. Try again later.\n");
1639 		return (EWOULDBLOCK);
1640 	}
1641 
1642 	if (ctrlr->hmb_nchunks > 0)
1643 		nvme_ctrlr_hmb_enable(ctrlr, false, false);
1644 
1645 	/*
1646 	 * Per Section 7.6.2 of NVMe spec 1.4, to properly suspend, we need to
1647 	 * delete the hardware I/O queues, and then shutdown. This properly
1648 	 * flushes any metadata the drive may have stored so it can survive
1649 	 * having its power removed and prevents the unsafe shutdown count from
1650 	 * incriminating. Once we delete the qpairs, we have to disable them
1651 	 * before shutting down. The delay is out of paranoia in
1652 	 * nvme_ctrlr_hw_reset, and is repeated here (though we should have no
1653 	 * pending I/O that the delay copes with).
1654 	 */
1655 	nvme_ctrlr_delete_qpairs(ctrlr);
1656 	nvme_ctrlr_disable_qpairs(ctrlr);
1657 	pause("nvmesusp", hz / 10);
1658 	nvme_ctrlr_shutdown(ctrlr);
1659 
1660 	return (0);
1661 }
1662 
1663 int
1664 nvme_ctrlr_resume(struct nvme_controller *ctrlr)
1665 {
1666 
1667 	/*
1668 	 * Can't touch failed controllers, so nothing to do to resume.
1669 	 */
1670 	if (ctrlr->is_failed)
1671 		return (0);
1672 
1673 	/*
1674 	 * Have to reset the hardware twice, just like we do on attach. See
1675 	 * nmve_attach() for why.
1676 	 */
1677 	if (nvme_ctrlr_hw_reset(ctrlr) != 0)
1678 		goto fail;
1679 	if (nvme_ctrlr_hw_reset(ctrlr) != 0)
1680 		goto fail;
1681 
1682 	/*
1683 	 * Now that we've reset the hardware, we can restart the controller. Any
1684 	 * I/O that was pending is requeued. Any admin commands are aborted with
1685 	 * an error. Once we've restarted, take the controller out of reset.
1686 	 */
1687 	nvme_ctrlr_start(ctrlr, true);
1688 	(void)atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
1689 
1690 	return (0);
1691 fail:
1692 	/*
1693 	 * Since we can't bring the controller out of reset, announce and fail
1694 	 * the controller. However, we have to return success for the resume
1695 	 * itself, due to questionable APIs.
1696 	 */
1697 	nvme_printf(ctrlr, "Failed to reset on resume, failing.\n");
1698 	nvme_ctrlr_fail(ctrlr);
1699 	(void)atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
1700 	return (0);
1701 }
1702