xref: /freebsd/sys/dev/nvme/nvme_ctrlr.c (revision 0bd5d367989b3d2de0e8d8ceaa2e31d3f0d96536)
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/endian.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 int
53 nvme_ctrlr_construct_admin_qpair(struct nvme_controller *ctrlr)
54 {
55 	struct nvme_qpair	*qpair;
56 	uint32_t		num_entries;
57 	int			error;
58 
59 	qpair = &ctrlr->adminq;
60 
61 	num_entries = NVME_ADMIN_ENTRIES;
62 	TUNABLE_INT_FETCH("hw.nvme.admin_entries", &num_entries);
63 	/*
64 	 * If admin_entries was overridden to an invalid value, revert it
65 	 *  back to our default value.
66 	 */
67 	if (num_entries < NVME_MIN_ADMIN_ENTRIES ||
68 	    num_entries > NVME_MAX_ADMIN_ENTRIES) {
69 		nvme_printf(ctrlr, "invalid hw.nvme.admin_entries=%d "
70 		    "specified\n", num_entries);
71 		num_entries = NVME_ADMIN_ENTRIES;
72 	}
73 
74 	/*
75 	 * The admin queue's max xfer size is treated differently than the
76 	 *  max I/O xfer size.  16KB is sufficient here - maybe even less?
77 	 */
78 	error = nvme_qpair_construct(qpair,
79 				     0, /* qpair ID */
80 				     0, /* vector */
81 				     num_entries,
82 				     NVME_ADMIN_TRACKERS,
83 				     ctrlr);
84 	return (error);
85 }
86 
87 static int
88 nvme_ctrlr_construct_io_qpairs(struct nvme_controller *ctrlr)
89 {
90 	struct nvme_qpair	*qpair;
91 	uint32_t		cap_lo;
92 	uint16_t		mqes;
93 	int			i, error, num_entries, num_trackers, max_entries;
94 
95 	/*
96 	 * NVMe spec sets a hard limit of 64K max entries, but devices may
97 	 * specify a smaller limit, so we need to check the MQES field in the
98 	 * capabilities register. We have to cap the number of entries to the
99 	 * current stride allows for in BAR 0/1, otherwise the remainder entries
100 	 * are inaccessable. MQES should reflect this, and this is just a
101 	 * fail-safe.
102 	 */
103 	max_entries =
104 	    (rman_get_size(ctrlr->resource) - nvme_mmio_offsetof(doorbell[0])) /
105 	    (1 << (ctrlr->dstrd + 1));
106 	num_entries = NVME_IO_ENTRIES;
107 	TUNABLE_INT_FETCH("hw.nvme.io_entries", &num_entries);
108 	cap_lo = nvme_mmio_read_4(ctrlr, cap_lo);
109 	mqes = NVME_CAP_LO_MQES(cap_lo);
110 	num_entries = min(num_entries, mqes + 1);
111 	num_entries = min(num_entries, max_entries);
112 
113 	num_trackers = NVME_IO_TRACKERS;
114 	TUNABLE_INT_FETCH("hw.nvme.io_trackers", &num_trackers);
115 
116 	num_trackers = max(num_trackers, NVME_MIN_IO_TRACKERS);
117 	num_trackers = min(num_trackers, NVME_MAX_IO_TRACKERS);
118 	/*
119 	 * No need to have more trackers than entries in the submit queue.  Note
120 	 * also that for a queue size of N, we can only have (N-1) commands
121 	 * outstanding, hence the "-1" here.
122 	 */
123 	num_trackers = min(num_trackers, (num_entries-1));
124 
125 	/*
126 	 * Our best estimate for the maximum number of I/Os that we should
127 	 * normally have in flight at one time. This should be viewed as a hint,
128 	 * not a hard limit and will need to be revisited when the upper layers
129 	 * of the storage system grows multi-queue support.
130 	 */
131 	ctrlr->max_hw_pend_io = num_trackers * ctrlr->num_io_queues * 3 / 4;
132 
133 	/*
134 	 * This was calculated previously when setting up interrupts, but
135 	 *  a controller could theoretically support fewer I/O queues than
136 	 *  MSI-X vectors.  So calculate again here just to be safe.
137 	 */
138 	ctrlr->num_cpus_per_ioq = howmany(mp_ncpus, ctrlr->num_io_queues);
139 
140 	ctrlr->ioq = malloc(ctrlr->num_io_queues * sizeof(struct nvme_qpair),
141 	    M_NVME, M_ZERO | M_WAITOK);
142 
143 	for (i = 0; i < ctrlr->num_io_queues; i++) {
144 		qpair = &ctrlr->ioq[i];
145 
146 		/*
147 		 * Admin queue has ID=0. IO queues start at ID=1 -
148 		 *  hence the 'i+1' here.
149 		 *
150 		 * For I/O queues, use the controller-wide max_xfer_size
151 		 *  calculated in nvme_attach().
152 		 */
153 		error = nvme_qpair_construct(qpair,
154 				     i+1, /* qpair ID */
155 				     ctrlr->msix_enabled ? i+1 : 0, /* vector */
156 				     num_entries,
157 				     num_trackers,
158 				     ctrlr);
159 		if (error)
160 			return (error);
161 
162 		/*
163 		 * Do not bother binding interrupts if we only have one I/O
164 		 *  interrupt thread for this controller.
165 		 */
166 		if (ctrlr->num_io_queues > 1)
167 			bus_bind_intr(ctrlr->dev, qpair->res,
168 			    i * ctrlr->num_cpus_per_ioq);
169 	}
170 
171 	return (0);
172 }
173 
174 static void
175 nvme_ctrlr_fail(struct nvme_controller *ctrlr)
176 {
177 	int i;
178 
179 	ctrlr->is_failed = TRUE;
180 	nvme_admin_qpair_disable(&ctrlr->adminq);
181 	nvme_qpair_fail(&ctrlr->adminq);
182 	if (ctrlr->ioq != NULL) {
183 		for (i = 0; i < ctrlr->num_io_queues; i++) {
184 			nvme_io_qpair_disable(&ctrlr->ioq[i]);
185 			nvme_qpair_fail(&ctrlr->ioq[i]);
186 		}
187 	}
188 	nvme_notify_fail_consumers(ctrlr);
189 }
190 
191 void
192 nvme_ctrlr_post_failed_request(struct nvme_controller *ctrlr,
193     struct nvme_request *req)
194 {
195 
196 	mtx_lock(&ctrlr->lock);
197 	STAILQ_INSERT_TAIL(&ctrlr->fail_req, req, stailq);
198 	mtx_unlock(&ctrlr->lock);
199 	taskqueue_enqueue(ctrlr->taskqueue, &ctrlr->fail_req_task);
200 }
201 
202 static void
203 nvme_ctrlr_fail_req_task(void *arg, int pending)
204 {
205 	struct nvme_controller	*ctrlr = arg;
206 	struct nvme_request	*req;
207 
208 	mtx_lock(&ctrlr->lock);
209 	while ((req = STAILQ_FIRST(&ctrlr->fail_req)) != NULL) {
210 		STAILQ_REMOVE_HEAD(&ctrlr->fail_req, stailq);
211 		mtx_unlock(&ctrlr->lock);
212 		nvme_qpair_manual_complete_request(req->qpair, req,
213 		    NVME_SCT_GENERIC, NVME_SC_ABORTED_BY_REQUEST);
214 		mtx_lock(&ctrlr->lock);
215 	}
216 	mtx_unlock(&ctrlr->lock);
217 }
218 
219 static int
220 nvme_ctrlr_wait_for_ready(struct nvme_controller *ctrlr, int desired_val)
221 {
222 	int ms_waited;
223 	uint32_t csts;
224 
225 	ms_waited = 0;
226 	while (1) {
227 		csts = nvme_mmio_read_4(ctrlr, csts);
228 		if (csts == 0xffffffff)		/* Hot unplug. */
229 			return (ENXIO);
230 		if (((csts >> NVME_CSTS_REG_RDY_SHIFT) & NVME_CSTS_REG_RDY_MASK)
231 		    == desired_val)
232 			break;
233 		if (ms_waited++ > ctrlr->ready_timeout_in_ms) {
234 			nvme_printf(ctrlr, "controller ready did not become %d "
235 			    "within %d ms\n", desired_val, ctrlr->ready_timeout_in_ms);
236 			return (ENXIO);
237 		}
238 		DELAY(1000);
239 	}
240 
241 	return (0);
242 }
243 
244 static int
245 nvme_ctrlr_disable(struct nvme_controller *ctrlr)
246 {
247 	uint32_t cc;
248 	uint32_t csts;
249 	uint8_t  en, rdy;
250 	int err;
251 
252 	cc = nvme_mmio_read_4(ctrlr, cc);
253 	csts = nvme_mmio_read_4(ctrlr, csts);
254 
255 	en = (cc >> NVME_CC_REG_EN_SHIFT) & NVME_CC_REG_EN_MASK;
256 	rdy = (csts >> NVME_CSTS_REG_RDY_SHIFT) & NVME_CSTS_REG_RDY_MASK;
257 
258 	/*
259 	 * Per 3.1.5 in NVME 1.3 spec, transitioning CC.EN from 0 to 1
260 	 * when CSTS.RDY is 1 or transitioning CC.EN from 1 to 0 when
261 	 * CSTS.RDY is 0 "has undefined results" So make sure that CSTS.RDY
262 	 * isn't the desired value. Short circuit if we're already disabled.
263 	 */
264 	if (en == 1) {
265 		if (rdy == 0) {
266 			/* EN == 1, wait for  RDY == 1 or fail */
267 			err = nvme_ctrlr_wait_for_ready(ctrlr, 1);
268 			if (err != 0)
269 				return (err);
270 		}
271 	} else {
272 		/* EN == 0 already wait for RDY == 0 */
273 		if (rdy == 0)
274 			return (0);
275 		else
276 			return (nvme_ctrlr_wait_for_ready(ctrlr, 0));
277 	}
278 
279 	cc &= ~NVME_CC_REG_EN_MASK;
280 	nvme_mmio_write_4(ctrlr, cc, cc);
281 	/*
282 	 * Some drives have issues with accessing the mmio after we
283 	 * disable, so delay for a bit after we write the bit to
284 	 * cope with these issues.
285 	 */
286 	if (ctrlr->quirks & QUIRK_DELAY_B4_CHK_RDY)
287 		pause("nvmeR", B4_CHK_RDY_DELAY_MS * hz / 1000);
288 	return (nvme_ctrlr_wait_for_ready(ctrlr, 0));
289 }
290 
291 static int
292 nvme_ctrlr_enable(struct nvme_controller *ctrlr)
293 {
294 	uint32_t	cc;
295 	uint32_t	csts;
296 	uint32_t	aqa;
297 	uint32_t	qsize;
298 	uint8_t		en, rdy;
299 	int		err;
300 
301 	cc = nvme_mmio_read_4(ctrlr, cc);
302 	csts = nvme_mmio_read_4(ctrlr, csts);
303 
304 	en = (cc >> NVME_CC_REG_EN_SHIFT) & NVME_CC_REG_EN_MASK;
305 	rdy = (csts >> NVME_CSTS_REG_RDY_SHIFT) & NVME_CSTS_REG_RDY_MASK;
306 
307 	/*
308 	 * See note in nvme_ctrlr_disable. Short circuit if we're already enabled.
309 	 */
310 	if (en == 1) {
311 		if (rdy == 1)
312 			return (0);
313 		else
314 			return (nvme_ctrlr_wait_for_ready(ctrlr, 1));
315 	} else {
316 		/* EN == 0 already wait for RDY == 0 or fail */
317 		err = nvme_ctrlr_wait_for_ready(ctrlr, 0);
318 		if (err != 0)
319 			return (err);
320 	}
321 
322 	nvme_mmio_write_8(ctrlr, asq, ctrlr->adminq.cmd_bus_addr);
323 	DELAY(5000);
324 	nvme_mmio_write_8(ctrlr, acq, ctrlr->adminq.cpl_bus_addr);
325 	DELAY(5000);
326 
327 	/* acqs and asqs are 0-based. */
328 	qsize = ctrlr->adminq.num_entries - 1;
329 
330 	aqa = 0;
331 	aqa = (qsize & NVME_AQA_REG_ACQS_MASK) << NVME_AQA_REG_ACQS_SHIFT;
332 	aqa |= (qsize & NVME_AQA_REG_ASQS_MASK) << NVME_AQA_REG_ASQS_SHIFT;
333 	nvme_mmio_write_4(ctrlr, aqa, aqa);
334 	DELAY(5000);
335 
336 	/* Initialization values for CC */
337 	cc = 0;
338 	cc |= 1 << NVME_CC_REG_EN_SHIFT;
339 	cc |= 0 << NVME_CC_REG_CSS_SHIFT;
340 	cc |= 0 << NVME_CC_REG_AMS_SHIFT;
341 	cc |= 0 << NVME_CC_REG_SHN_SHIFT;
342 	cc |= 6 << NVME_CC_REG_IOSQES_SHIFT; /* SQ entry size == 64 == 2^6 */
343 	cc |= 4 << NVME_CC_REG_IOCQES_SHIFT; /* CQ entry size == 16 == 2^4 */
344 
345 	/* This evaluates to 0, which is according to spec. */
346 	cc |= (PAGE_SIZE >> 13) << NVME_CC_REG_MPS_SHIFT;
347 
348 	nvme_mmio_write_4(ctrlr, cc, cc);
349 
350 	return (nvme_ctrlr_wait_for_ready(ctrlr, 1));
351 }
352 
353 static void
354 nvme_ctrlr_disable_qpairs(struct nvme_controller *ctrlr)
355 {
356 	int i;
357 
358 	nvme_admin_qpair_disable(&ctrlr->adminq);
359 	/*
360 	 * I/O queues are not allocated before the initial HW
361 	 *  reset, so do not try to disable them.  Use is_initialized
362 	 *  to determine if this is the initial HW reset.
363 	 */
364 	if (ctrlr->is_initialized) {
365 		for (i = 0; i < ctrlr->num_io_queues; i++)
366 			nvme_io_qpair_disable(&ctrlr->ioq[i]);
367 	}
368 }
369 
370 int
371 nvme_ctrlr_hw_reset(struct nvme_controller *ctrlr)
372 {
373 	int err;
374 
375 	nvme_ctrlr_disable_qpairs(ctrlr);
376 
377 	DELAY(100*1000);
378 
379 	err = nvme_ctrlr_disable(ctrlr);
380 	if (err != 0)
381 		return err;
382 	return (nvme_ctrlr_enable(ctrlr));
383 }
384 
385 void
386 nvme_ctrlr_reset(struct nvme_controller *ctrlr)
387 {
388 	int cmpset;
389 
390 	cmpset = atomic_cmpset_32(&ctrlr->is_resetting, 0, 1);
391 
392 	if (cmpset == 0 || ctrlr->is_failed)
393 		/*
394 		 * Controller is already resetting or has failed.  Return
395 		 *  immediately since there is no need to kick off another
396 		 *  reset in these cases.
397 		 */
398 		return;
399 
400 	taskqueue_enqueue(ctrlr->taskqueue, &ctrlr->reset_task);
401 }
402 
403 static int
404 nvme_ctrlr_identify(struct nvme_controller *ctrlr)
405 {
406 	struct nvme_completion_poll_status	status;
407 
408 	status.done = 0;
409 	nvme_ctrlr_cmd_identify_controller(ctrlr, &ctrlr->cdata,
410 	    nvme_completion_poll_cb, &status);
411 	nvme_completion_poll(&status);
412 	if (nvme_completion_is_error(&status.cpl)) {
413 		nvme_printf(ctrlr, "nvme_identify_controller failed!\n");
414 		return (ENXIO);
415 	}
416 
417 	/* Convert data to host endian */
418 	nvme_controller_data_swapbytes(&ctrlr->cdata);
419 
420 	/*
421 	 * Use MDTS to ensure our default max_xfer_size doesn't exceed what the
422 	 *  controller supports.
423 	 */
424 	if (ctrlr->cdata.mdts > 0)
425 		ctrlr->max_xfer_size = min(ctrlr->max_xfer_size,
426 		    ctrlr->min_page_size * (1 << (ctrlr->cdata.mdts)));
427 
428 	return (0);
429 }
430 
431 static int
432 nvme_ctrlr_set_num_qpairs(struct nvme_controller *ctrlr)
433 {
434 	struct nvme_completion_poll_status	status;
435 	int					cq_allocated, sq_allocated;
436 
437 	status.done = 0;
438 	nvme_ctrlr_cmd_set_num_queues(ctrlr, ctrlr->num_io_queues,
439 	    nvme_completion_poll_cb, &status);
440 	nvme_completion_poll(&status);
441 	if (nvme_completion_is_error(&status.cpl)) {
442 		nvme_printf(ctrlr, "nvme_ctrlr_set_num_qpairs failed!\n");
443 		return (ENXIO);
444 	}
445 
446 	/*
447 	 * Data in cdw0 is 0-based.
448 	 * Lower 16-bits indicate number of submission queues allocated.
449 	 * Upper 16-bits indicate number of completion queues allocated.
450 	 */
451 	sq_allocated = (status.cpl.cdw0 & 0xFFFF) + 1;
452 	cq_allocated = (status.cpl.cdw0 >> 16) + 1;
453 
454 	/*
455 	 * Controller may allocate more queues than we requested,
456 	 *  so use the minimum of the number requested and what was
457 	 *  actually allocated.
458 	 */
459 	ctrlr->num_io_queues = min(ctrlr->num_io_queues, sq_allocated);
460 	ctrlr->num_io_queues = min(ctrlr->num_io_queues, cq_allocated);
461 
462 	return (0);
463 }
464 
465 static int
466 nvme_ctrlr_create_qpairs(struct nvme_controller *ctrlr)
467 {
468 	struct nvme_completion_poll_status	status;
469 	struct nvme_qpair			*qpair;
470 	int					i;
471 
472 	for (i = 0; i < ctrlr->num_io_queues; i++) {
473 		qpair = &ctrlr->ioq[i];
474 
475 		status.done = 0;
476 		nvme_ctrlr_cmd_create_io_cq(ctrlr, qpair, qpair->vector,
477 		    nvme_completion_poll_cb, &status);
478 		nvme_completion_poll(&status);
479 		if (nvme_completion_is_error(&status.cpl)) {
480 			nvme_printf(ctrlr, "nvme_create_io_cq failed!\n");
481 			return (ENXIO);
482 		}
483 
484 		status.done = 0;
485 		nvme_ctrlr_cmd_create_io_sq(qpair->ctrlr, qpair,
486 		    nvme_completion_poll_cb, &status);
487 		nvme_completion_poll(&status);
488 		if (nvme_completion_is_error(&status.cpl)) {
489 			nvme_printf(ctrlr, "nvme_create_io_sq failed!\n");
490 			return (ENXIO);
491 		}
492 	}
493 
494 	return (0);
495 }
496 
497 static int
498 nvme_ctrlr_delete_qpairs(struct nvme_controller *ctrlr)
499 {
500 	struct nvme_completion_poll_status	status;
501 	struct nvme_qpair			*qpair;
502 
503 	for (int i = 0; i < ctrlr->num_io_queues; i++) {
504 		qpair = &ctrlr->ioq[i];
505 
506 		status.done = 0;
507 		nvme_ctrlr_cmd_delete_io_sq(ctrlr, qpair,
508 		    nvme_completion_poll_cb, &status);
509 		nvme_completion_poll(&status);
510 		if (nvme_completion_is_error(&status.cpl)) {
511 			nvme_printf(ctrlr, "nvme_destroy_io_sq failed!\n");
512 			return (ENXIO);
513 		}
514 
515 		status.done = 0;
516 		nvme_ctrlr_cmd_delete_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_destroy_io_cq failed!\n");
521 			return (ENXIO);
522 		}
523 	}
524 
525 	return (0);
526 }
527 
528 static int
529 nvme_ctrlr_construct_namespaces(struct nvme_controller *ctrlr)
530 {
531 	struct nvme_namespace	*ns;
532 	uint32_t 		i;
533 
534 	for (i = 0; i < min(ctrlr->cdata.nn, NVME_MAX_NAMESPACES); i++) {
535 		ns = &ctrlr->ns[i];
536 		nvme_ns_construct(ns, i+1, ctrlr);
537 	}
538 
539 	return (0);
540 }
541 
542 static boolean_t
543 is_log_page_id_valid(uint8_t page_id)
544 {
545 
546 	switch (page_id) {
547 	case NVME_LOG_ERROR:
548 	case NVME_LOG_HEALTH_INFORMATION:
549 	case NVME_LOG_FIRMWARE_SLOT:
550 	case NVME_LOG_CHANGED_NAMESPACE:
551 	case NVME_LOG_COMMAND_EFFECT:
552 	case NVME_LOG_RES_NOTIFICATION:
553 	case NVME_LOG_SANITIZE_STATUS:
554 		return (TRUE);
555 	}
556 
557 	return (FALSE);
558 }
559 
560 static uint32_t
561 nvme_ctrlr_get_log_page_size(struct nvme_controller *ctrlr, uint8_t page_id)
562 {
563 	uint32_t	log_page_size;
564 
565 	switch (page_id) {
566 	case NVME_LOG_ERROR:
567 		log_page_size = min(
568 		    sizeof(struct nvme_error_information_entry) *
569 		    (ctrlr->cdata.elpe + 1), NVME_MAX_AER_LOG_SIZE);
570 		break;
571 	case NVME_LOG_HEALTH_INFORMATION:
572 		log_page_size = sizeof(struct nvme_health_information_page);
573 		break;
574 	case NVME_LOG_FIRMWARE_SLOT:
575 		log_page_size = sizeof(struct nvme_firmware_page);
576 		break;
577 	case NVME_LOG_CHANGED_NAMESPACE:
578 		log_page_size = sizeof(struct nvme_ns_list);
579 		break;
580 	case NVME_LOG_COMMAND_EFFECT:
581 		log_page_size = sizeof(struct nvme_command_effects_page);
582 		break;
583 	case NVME_LOG_RES_NOTIFICATION:
584 		log_page_size = sizeof(struct nvme_res_notification_page);
585 		break;
586 	case NVME_LOG_SANITIZE_STATUS:
587 		log_page_size = sizeof(struct nvme_sanitize_status_page);
588 		break;
589 	default:
590 		log_page_size = 0;
591 		break;
592 	}
593 
594 	return (log_page_size);
595 }
596 
597 static void
598 nvme_ctrlr_log_critical_warnings(struct nvme_controller *ctrlr,
599     uint8_t state)
600 {
601 
602 	if (state & NVME_CRIT_WARN_ST_AVAILABLE_SPARE)
603 		nvme_printf(ctrlr, "available spare space below threshold\n");
604 
605 	if (state & NVME_CRIT_WARN_ST_TEMPERATURE)
606 		nvme_printf(ctrlr, "temperature above threshold\n");
607 
608 	if (state & NVME_CRIT_WARN_ST_DEVICE_RELIABILITY)
609 		nvme_printf(ctrlr, "device reliability degraded\n");
610 
611 	if (state & NVME_CRIT_WARN_ST_READ_ONLY)
612 		nvme_printf(ctrlr, "media placed in read only mode\n");
613 
614 	if (state & NVME_CRIT_WARN_ST_VOLATILE_MEMORY_BACKUP)
615 		nvme_printf(ctrlr, "volatile memory backup device failed\n");
616 
617 	if (state & NVME_CRIT_WARN_ST_RESERVED_MASK)
618 		nvme_printf(ctrlr,
619 		    "unknown critical warning(s): state = 0x%02x\n", state);
620 }
621 
622 static void
623 nvme_ctrlr_async_event_log_page_cb(void *arg, const struct nvme_completion *cpl)
624 {
625 	struct nvme_async_event_request		*aer = arg;
626 	struct nvme_health_information_page	*health_info;
627 	struct nvme_ns_list			*nsl;
628 	struct nvme_error_information_entry	*err;
629 	int i;
630 
631 	/*
632 	 * If the log page fetch for some reason completed with an error,
633 	 *  don't pass log page data to the consumers.  In practice, this case
634 	 *  should never happen.
635 	 */
636 	if (nvme_completion_is_error(cpl))
637 		nvme_notify_async_consumers(aer->ctrlr, &aer->cpl,
638 		    aer->log_page_id, NULL, 0);
639 	else {
640 		/* Convert data to host endian */
641 		switch (aer->log_page_id) {
642 		case NVME_LOG_ERROR:
643 			err = (struct nvme_error_information_entry *)aer->log_page_buffer;
644 			for (i = 0; i < (aer->ctrlr->cdata.elpe + 1); i++)
645 				nvme_error_information_entry_swapbytes(err++);
646 			break;
647 		case NVME_LOG_HEALTH_INFORMATION:
648 			nvme_health_information_page_swapbytes(
649 			    (struct nvme_health_information_page *)aer->log_page_buffer);
650 			break;
651 		case NVME_LOG_FIRMWARE_SLOT:
652 			nvme_firmware_page_swapbytes(
653 			    (struct nvme_firmware_page *)aer->log_page_buffer);
654 			break;
655 		case NVME_LOG_CHANGED_NAMESPACE:
656 			nvme_ns_list_swapbytes(
657 			    (struct nvme_ns_list *)aer->log_page_buffer);
658 			break;
659 		case NVME_LOG_COMMAND_EFFECT:
660 			nvme_command_effects_page_swapbytes(
661 			    (struct nvme_command_effects_page *)aer->log_page_buffer);
662 			break;
663 		case NVME_LOG_RES_NOTIFICATION:
664 			nvme_res_notification_page_swapbytes(
665 			    (struct nvme_res_notification_page *)aer->log_page_buffer);
666 			break;
667 		case NVME_LOG_SANITIZE_STATUS:
668 			nvme_sanitize_status_page_swapbytes(
669 			    (struct nvme_sanitize_status_page *)aer->log_page_buffer);
670 			break;
671 		case INTEL_LOG_TEMP_STATS:
672 			intel_log_temp_stats_swapbytes(
673 			    (struct intel_log_temp_stats *)aer->log_page_buffer);
674 			break;
675 		default:
676 			break;
677 		}
678 
679 		if (aer->log_page_id == NVME_LOG_HEALTH_INFORMATION) {
680 			health_info = (struct nvme_health_information_page *)
681 			    aer->log_page_buffer;
682 			nvme_ctrlr_log_critical_warnings(aer->ctrlr,
683 			    health_info->critical_warning);
684 			/*
685 			 * Critical warnings reported through the
686 			 *  SMART/health log page are persistent, so
687 			 *  clear the associated bits in the async event
688 			 *  config so that we do not receive repeated
689 			 *  notifications for the same event.
690 			 */
691 			aer->ctrlr->async_event_config &=
692 			    ~health_info->critical_warning;
693 			nvme_ctrlr_cmd_set_async_event_config(aer->ctrlr,
694 			    aer->ctrlr->async_event_config, NULL, NULL);
695 		} else if (aer->log_page_id == NVME_LOG_CHANGED_NAMESPACE &&
696 		    !nvme_use_nvd) {
697 			nsl = (struct nvme_ns_list *)aer->log_page_buffer;
698 			for (i = 0; i < nitems(nsl->ns) && nsl->ns[i] != 0; i++) {
699 				if (nsl->ns[i] > NVME_MAX_NAMESPACES)
700 					break;
701 				nvme_notify_ns(aer->ctrlr, nsl->ns[i]);
702 			}
703 		}
704 
705 
706 		/*
707 		 * Pass the cpl data from the original async event completion,
708 		 *  not the log page fetch.
709 		 */
710 		nvme_notify_async_consumers(aer->ctrlr, &aer->cpl,
711 		    aer->log_page_id, aer->log_page_buffer, aer->log_page_size);
712 	}
713 
714 	/*
715 	 * Repost another asynchronous event request to replace the one
716 	 *  that just completed.
717 	 */
718 	nvme_ctrlr_construct_and_submit_aer(aer->ctrlr, aer);
719 }
720 
721 static void
722 nvme_ctrlr_async_event_cb(void *arg, const struct nvme_completion *cpl)
723 {
724 	struct nvme_async_event_request	*aer = arg;
725 
726 	if (nvme_completion_is_error(cpl)) {
727 		/*
728 		 *  Do not retry failed async event requests.  This avoids
729 		 *  infinite loops where a new async event request is submitted
730 		 *  to replace the one just failed, only to fail again and
731 		 *  perpetuate the loop.
732 		 */
733 		return;
734 	}
735 
736 	/* Associated log page is in bits 23:16 of completion entry dw0. */
737 	aer->log_page_id = (cpl->cdw0 & 0xFF0000) >> 16;
738 
739 	nvme_printf(aer->ctrlr, "async event occurred (type 0x%x, info 0x%02x,"
740 	    " page 0x%02x)\n", (cpl->cdw0 & 0x07), (cpl->cdw0 & 0xFF00) >> 8,
741 	    aer->log_page_id);
742 
743 	if (is_log_page_id_valid(aer->log_page_id)) {
744 		aer->log_page_size = nvme_ctrlr_get_log_page_size(aer->ctrlr,
745 		    aer->log_page_id);
746 		memcpy(&aer->cpl, cpl, sizeof(*cpl));
747 		nvme_ctrlr_cmd_get_log_page(aer->ctrlr, aer->log_page_id,
748 		    NVME_GLOBAL_NAMESPACE_TAG, aer->log_page_buffer,
749 		    aer->log_page_size, nvme_ctrlr_async_event_log_page_cb,
750 		    aer);
751 		/* Wait to notify consumers until after log page is fetched. */
752 	} else {
753 		nvme_notify_async_consumers(aer->ctrlr, cpl, aer->log_page_id,
754 		    NULL, 0);
755 
756 		/*
757 		 * Repost another asynchronous event request to replace the one
758 		 *  that just completed.
759 		 */
760 		nvme_ctrlr_construct_and_submit_aer(aer->ctrlr, aer);
761 	}
762 }
763 
764 static void
765 nvme_ctrlr_construct_and_submit_aer(struct nvme_controller *ctrlr,
766     struct nvme_async_event_request *aer)
767 {
768 	struct nvme_request *req;
769 
770 	aer->ctrlr = ctrlr;
771 	req = nvme_allocate_request_null(nvme_ctrlr_async_event_cb, aer);
772 	aer->req = req;
773 
774 	/*
775 	 * Disable timeout here, since asynchronous event requests should by
776 	 *  nature never be timed out.
777 	 */
778 	req->timeout = FALSE;
779 	req->cmd.opc = NVME_OPC_ASYNC_EVENT_REQUEST;
780 	nvme_ctrlr_submit_admin_request(ctrlr, req);
781 }
782 
783 static void
784 nvme_ctrlr_configure_aer(struct nvme_controller *ctrlr)
785 {
786 	struct nvme_completion_poll_status	status;
787 	struct nvme_async_event_request		*aer;
788 	uint32_t				i;
789 
790 	ctrlr->async_event_config = NVME_CRIT_WARN_ST_AVAILABLE_SPARE |
791 	    NVME_CRIT_WARN_ST_DEVICE_RELIABILITY |
792 	    NVME_CRIT_WARN_ST_READ_ONLY |
793 	    NVME_CRIT_WARN_ST_VOLATILE_MEMORY_BACKUP;
794 	if (ctrlr->cdata.ver >= NVME_REV(1, 2))
795 		ctrlr->async_event_config |= 0x300;
796 
797 	status.done = 0;
798 	nvme_ctrlr_cmd_get_feature(ctrlr, NVME_FEAT_TEMPERATURE_THRESHOLD,
799 	    0, NULL, 0, nvme_completion_poll_cb, &status);
800 	nvme_completion_poll(&status);
801 	if (nvme_completion_is_error(&status.cpl) ||
802 	    (status.cpl.cdw0 & 0xFFFF) == 0xFFFF ||
803 	    (status.cpl.cdw0 & 0xFFFF) == 0x0000) {
804 		nvme_printf(ctrlr, "temperature threshold not supported\n");
805 	} else
806 		ctrlr->async_event_config |= NVME_CRIT_WARN_ST_TEMPERATURE;
807 
808 	nvme_ctrlr_cmd_set_async_event_config(ctrlr,
809 	    ctrlr->async_event_config, NULL, NULL);
810 
811 	/* aerl is a zero-based value, so we need to add 1 here. */
812 	ctrlr->num_aers = min(NVME_MAX_ASYNC_EVENTS, (ctrlr->cdata.aerl+1));
813 
814 	for (i = 0; i < ctrlr->num_aers; i++) {
815 		aer = &ctrlr->aer[i];
816 		nvme_ctrlr_construct_and_submit_aer(ctrlr, aer);
817 	}
818 }
819 
820 static void
821 nvme_ctrlr_configure_int_coalescing(struct nvme_controller *ctrlr)
822 {
823 
824 	ctrlr->int_coal_time = 0;
825 	TUNABLE_INT_FETCH("hw.nvme.int_coal_time",
826 	    &ctrlr->int_coal_time);
827 
828 	ctrlr->int_coal_threshold = 0;
829 	TUNABLE_INT_FETCH("hw.nvme.int_coal_threshold",
830 	    &ctrlr->int_coal_threshold);
831 
832 	nvme_ctrlr_cmd_set_interrupt_coalescing(ctrlr, ctrlr->int_coal_time,
833 	    ctrlr->int_coal_threshold, NULL, NULL);
834 }
835 
836 static void
837 nvme_ctrlr_start(void *ctrlr_arg, bool resetting)
838 {
839 	struct nvme_controller *ctrlr = ctrlr_arg;
840 	uint32_t old_num_io_queues;
841 	int i;
842 
843 	/*
844 	 * Only reset adminq here when we are restarting the
845 	 *  controller after a reset.  During initialization,
846 	 *  we have already submitted admin commands to get
847 	 *  the number of I/O queues supported, so cannot reset
848 	 *  the adminq again here.
849 	 */
850 	if (resetting)
851 		nvme_qpair_reset(&ctrlr->adminq);
852 
853 	for (i = 0; i < ctrlr->num_io_queues; i++)
854 		nvme_qpair_reset(&ctrlr->ioq[i]);
855 
856 	nvme_admin_qpair_enable(&ctrlr->adminq);
857 
858 	if (nvme_ctrlr_identify(ctrlr) != 0) {
859 		nvme_ctrlr_fail(ctrlr);
860 		return;
861 	}
862 
863 	/*
864 	 * The number of qpairs are determined during controller initialization,
865 	 *  including using NVMe SET_FEATURES/NUMBER_OF_QUEUES to determine the
866 	 *  HW limit.  We call SET_FEATURES again here so that it gets called
867 	 *  after any reset for controllers that depend on the driver to
868 	 *  explicit specify how many queues it will use.  This value should
869 	 *  never change between resets, so panic if somehow that does happen.
870 	 */
871 	if (resetting) {
872 		old_num_io_queues = ctrlr->num_io_queues;
873 		if (nvme_ctrlr_set_num_qpairs(ctrlr) != 0) {
874 			nvme_ctrlr_fail(ctrlr);
875 			return;
876 		}
877 
878 		if (old_num_io_queues != ctrlr->num_io_queues) {
879 			panic("num_io_queues changed from %u to %u",
880 			      old_num_io_queues, ctrlr->num_io_queues);
881 		}
882 	}
883 
884 	if (nvme_ctrlr_create_qpairs(ctrlr) != 0) {
885 		nvme_ctrlr_fail(ctrlr);
886 		return;
887 	}
888 
889 	if (nvme_ctrlr_construct_namespaces(ctrlr) != 0) {
890 		nvme_ctrlr_fail(ctrlr);
891 		return;
892 	}
893 
894 	nvme_ctrlr_configure_aer(ctrlr);
895 	nvme_ctrlr_configure_int_coalescing(ctrlr);
896 
897 	for (i = 0; i < ctrlr->num_io_queues; i++)
898 		nvme_io_qpair_enable(&ctrlr->ioq[i]);
899 }
900 
901 void
902 nvme_ctrlr_start_config_hook(void *arg)
903 {
904 	struct nvme_controller *ctrlr = arg;
905 
906 	nvme_qpair_reset(&ctrlr->adminq);
907 	nvme_admin_qpair_enable(&ctrlr->adminq);
908 
909 	if (nvme_ctrlr_set_num_qpairs(ctrlr) == 0 &&
910 	    nvme_ctrlr_construct_io_qpairs(ctrlr) == 0)
911 		nvme_ctrlr_start(ctrlr, false);
912 	else
913 		nvme_ctrlr_fail(ctrlr);
914 
915 	nvme_sysctl_initialize_ctrlr(ctrlr);
916 	config_intrhook_disestablish(&ctrlr->config_hook);
917 
918 	ctrlr->is_initialized = 1;
919 	nvme_notify_new_controller(ctrlr);
920 }
921 
922 static void
923 nvme_ctrlr_reset_task(void *arg, int pending)
924 {
925 	struct nvme_controller	*ctrlr = arg;
926 	int			status;
927 
928 	nvme_printf(ctrlr, "resetting controller\n");
929 	status = nvme_ctrlr_hw_reset(ctrlr);
930 	/*
931 	 * Use pause instead of DELAY, so that we yield to any nvme interrupt
932 	 *  handlers on this CPU that were blocked on a qpair lock. We want
933 	 *  all nvme interrupts completed before proceeding with restarting the
934 	 *  controller.
935 	 *
936 	 * XXX - any way to guarantee the interrupt handlers have quiesced?
937 	 */
938 	pause("nvmereset", hz / 10);
939 	if (status == 0)
940 		nvme_ctrlr_start(ctrlr, true);
941 	else
942 		nvme_ctrlr_fail(ctrlr);
943 
944 	atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
945 }
946 
947 /*
948  * Poll all the queues enabled on the device for completion.
949  */
950 void
951 nvme_ctrlr_poll(struct nvme_controller *ctrlr)
952 {
953 	int i;
954 
955 	nvme_qpair_process_completions(&ctrlr->adminq);
956 
957 	for (i = 0; i < ctrlr->num_io_queues; i++)
958 		if (ctrlr->ioq && ctrlr->ioq[i].cpl)
959 			nvme_qpair_process_completions(&ctrlr->ioq[i]);
960 }
961 
962 /*
963  * Poll the single-vector interrupt case: num_io_queues will be 1 and
964  * there's only a single vector. While we're polling, we mask further
965  * interrupts in the controller.
966  */
967 void
968 nvme_ctrlr_intx_handler(void *arg)
969 {
970 	struct nvme_controller *ctrlr = arg;
971 
972 	nvme_mmio_write_4(ctrlr, intms, 1);
973 	nvme_ctrlr_poll(ctrlr);
974 	nvme_mmio_write_4(ctrlr, intmc, 1);
975 }
976 
977 static void
978 nvme_pt_done(void *arg, const struct nvme_completion *cpl)
979 {
980 	struct nvme_pt_command *pt = arg;
981 	struct mtx *mtx = pt->driver_lock;
982 	uint16_t status;
983 
984 	bzero(&pt->cpl, sizeof(pt->cpl));
985 	pt->cpl.cdw0 = cpl->cdw0;
986 
987 	status = cpl->status;
988 	status &= ~NVME_STATUS_P_MASK;
989 	pt->cpl.status = status;
990 
991 	mtx_lock(mtx);
992 	pt->driver_lock = NULL;
993 	wakeup(pt);
994 	mtx_unlock(mtx);
995 }
996 
997 int
998 nvme_ctrlr_passthrough_cmd(struct nvme_controller *ctrlr,
999     struct nvme_pt_command *pt, uint32_t nsid, int is_user_buffer,
1000     int is_admin_cmd)
1001 {
1002 	struct nvme_request	*req;
1003 	struct mtx		*mtx;
1004 	struct buf		*buf = NULL;
1005 	int			ret = 0;
1006 	vm_offset_t		addr, end;
1007 
1008 	if (pt->len > 0) {
1009 		/*
1010 		 * vmapbuf calls vm_fault_quick_hold_pages which only maps full
1011 		 * pages. Ensure this request has fewer than MAXPHYS bytes when
1012 		 * extended to full pages.
1013 		 */
1014 		addr = (vm_offset_t)pt->buf;
1015 		end = round_page(addr + pt->len);
1016 		addr = trunc_page(addr);
1017 		if (end - addr > MAXPHYS)
1018 			return EIO;
1019 
1020 		if (pt->len > ctrlr->max_xfer_size) {
1021 			nvme_printf(ctrlr, "pt->len (%d) "
1022 			    "exceeds max_xfer_size (%d)\n", pt->len,
1023 			    ctrlr->max_xfer_size);
1024 			return EIO;
1025 		}
1026 		if (is_user_buffer) {
1027 			/*
1028 			 * Ensure the user buffer is wired for the duration of
1029 			 *  this pass-through command.
1030 			 */
1031 			PHOLD(curproc);
1032 			buf = uma_zalloc(pbuf_zone, M_WAITOK);
1033 			buf->b_data = pt->buf;
1034 			buf->b_bufsize = pt->len;
1035 			buf->b_iocmd = pt->is_read ? BIO_READ : BIO_WRITE;
1036 			if (vmapbuf(buf, 1) < 0) {
1037 				ret = EFAULT;
1038 				goto err;
1039 			}
1040 			req = nvme_allocate_request_vaddr(buf->b_data, pt->len,
1041 			    nvme_pt_done, pt);
1042 		} else
1043 			req = nvme_allocate_request_vaddr(pt->buf, pt->len,
1044 			    nvme_pt_done, pt);
1045 	} else
1046 		req = nvme_allocate_request_null(nvme_pt_done, pt);
1047 
1048 	/* Assume user space already converted to little-endian */
1049 	req->cmd.opc = pt->cmd.opc;
1050 	req->cmd.fuse = pt->cmd.fuse;
1051 	req->cmd.rsvd2 = pt->cmd.rsvd2;
1052 	req->cmd.rsvd3 = pt->cmd.rsvd3;
1053 	req->cmd.cdw10 = pt->cmd.cdw10;
1054 	req->cmd.cdw11 = pt->cmd.cdw11;
1055 	req->cmd.cdw12 = pt->cmd.cdw12;
1056 	req->cmd.cdw13 = pt->cmd.cdw13;
1057 	req->cmd.cdw14 = pt->cmd.cdw14;
1058 	req->cmd.cdw15 = pt->cmd.cdw15;
1059 
1060 	req->cmd.nsid = htole32(nsid);
1061 
1062 	mtx = mtx_pool_find(mtxpool_sleep, pt);
1063 	pt->driver_lock = mtx;
1064 
1065 	if (is_admin_cmd)
1066 		nvme_ctrlr_submit_admin_request(ctrlr, req);
1067 	else
1068 		nvme_ctrlr_submit_io_request(ctrlr, req);
1069 
1070 	mtx_lock(mtx);
1071 	while (pt->driver_lock != NULL)
1072 		mtx_sleep(pt, mtx, PRIBIO, "nvme_pt", 0);
1073 	mtx_unlock(mtx);
1074 
1075 err:
1076 	if (buf != NULL) {
1077 		uma_zfree(pbuf_zone, buf);
1078 		PRELE(curproc);
1079 	}
1080 
1081 	return (ret);
1082 }
1083 
1084 static int
1085 nvme_ctrlr_ioctl(struct cdev *cdev, u_long cmd, caddr_t arg, int flag,
1086     struct thread *td)
1087 {
1088 	struct nvme_controller			*ctrlr;
1089 	struct nvme_pt_command			*pt;
1090 
1091 	ctrlr = cdev->si_drv1;
1092 
1093 	switch (cmd) {
1094 	case NVME_RESET_CONTROLLER:
1095 		nvme_ctrlr_reset(ctrlr);
1096 		break;
1097 	case NVME_PASSTHROUGH_CMD:
1098 		pt = (struct nvme_pt_command *)arg;
1099 		return (nvme_ctrlr_passthrough_cmd(ctrlr, pt, le32toh(pt->cmd.nsid),
1100 		    1 /* is_user_buffer */, 1 /* is_admin_cmd */));
1101 	case NVME_GET_NSID:
1102 	{
1103 		struct nvme_get_nsid *gnsid = (struct nvme_get_nsid *)arg;
1104 		strncpy(gnsid->cdev, device_get_nameunit(ctrlr->dev),
1105 		    sizeof(gnsid->cdev));
1106 		gnsid->nsid = 0;
1107 		break;
1108 	}
1109 	default:
1110 		return (ENOTTY);
1111 	}
1112 
1113 	return (0);
1114 }
1115 
1116 static struct cdevsw nvme_ctrlr_cdevsw = {
1117 	.d_version =	D_VERSION,
1118 	.d_flags =	0,
1119 	.d_ioctl =	nvme_ctrlr_ioctl
1120 };
1121 
1122 int
1123 nvme_ctrlr_construct(struct nvme_controller *ctrlr, device_t dev)
1124 {
1125 	struct make_dev_args	md_args;
1126 	uint32_t	cap_lo;
1127 	uint32_t	cap_hi;
1128 	uint32_t	to;
1129 	uint8_t		mpsmin;
1130 	int		status, timeout_period;
1131 
1132 	ctrlr->dev = dev;
1133 
1134 	mtx_init(&ctrlr->lock, "nvme ctrlr lock", NULL, MTX_DEF);
1135 
1136 	cap_hi = nvme_mmio_read_4(ctrlr, cap_hi);
1137 	ctrlr->dstrd = NVME_CAP_HI_DSTRD(cap_hi) + 2;
1138 
1139 	mpsmin = NVME_CAP_HI_MPSMIN(cap_hi);
1140 	ctrlr->min_page_size = 1 << (12 + mpsmin);
1141 
1142 	/* Get ready timeout value from controller, in units of 500ms. */
1143 	cap_lo = nvme_mmio_read_4(ctrlr, cap_lo);
1144 	to = NVME_CAP_LO_TO(cap_lo) + 1;
1145 	ctrlr->ready_timeout_in_ms = to * 500;
1146 
1147 	timeout_period = NVME_DEFAULT_TIMEOUT_PERIOD;
1148 	TUNABLE_INT_FETCH("hw.nvme.timeout_period", &timeout_period);
1149 	timeout_period = min(timeout_period, NVME_MAX_TIMEOUT_PERIOD);
1150 	timeout_period = max(timeout_period, NVME_MIN_TIMEOUT_PERIOD);
1151 	ctrlr->timeout_period = timeout_period;
1152 
1153 	nvme_retry_count = NVME_DEFAULT_RETRY_COUNT;
1154 	TUNABLE_INT_FETCH("hw.nvme.retry_count", &nvme_retry_count);
1155 
1156 	ctrlr->enable_aborts = 0;
1157 	TUNABLE_INT_FETCH("hw.nvme.enable_aborts", &ctrlr->enable_aborts);
1158 
1159 	ctrlr->max_xfer_size = NVME_MAX_XFER_SIZE;
1160 	if (nvme_ctrlr_construct_admin_qpair(ctrlr) != 0)
1161 		return (ENXIO);
1162 
1163 	ctrlr->taskqueue = taskqueue_create("nvme_taskq", M_WAITOK,
1164 	    taskqueue_thread_enqueue, &ctrlr->taskqueue);
1165 	taskqueue_start_threads(&ctrlr->taskqueue, 1, PI_DISK, "nvme taskq");
1166 
1167 	ctrlr->is_resetting = 0;
1168 	ctrlr->is_initialized = 0;
1169 	ctrlr->notification_sent = 0;
1170 	TASK_INIT(&ctrlr->reset_task, 0, nvme_ctrlr_reset_task, ctrlr);
1171 	TASK_INIT(&ctrlr->fail_req_task, 0, nvme_ctrlr_fail_req_task, ctrlr);
1172 	STAILQ_INIT(&ctrlr->fail_req);
1173 	ctrlr->is_failed = FALSE;
1174 
1175 	make_dev_args_init(&md_args);
1176 	md_args.mda_devsw = &nvme_ctrlr_cdevsw;
1177 	md_args.mda_uid = UID_ROOT;
1178 	md_args.mda_gid = GID_WHEEL;
1179 	md_args.mda_mode = 0600;
1180 	md_args.mda_unit = device_get_unit(dev);
1181 	md_args.mda_si_drv1 = (void *)ctrlr;
1182 	status = make_dev_s(&md_args, &ctrlr->cdev, "nvme%d",
1183 	    device_get_unit(dev));
1184 	if (status != 0)
1185 		return (ENXIO);
1186 
1187 	return (0);
1188 }
1189 
1190 void
1191 nvme_ctrlr_destruct(struct nvme_controller *ctrlr, device_t dev)
1192 {
1193 	int	gone, i;
1194 
1195 	if (ctrlr->resource == NULL)
1196 		goto nores;
1197 
1198 	/*
1199 	 * Check whether it is a hot unplug or a clean driver detach.
1200 	 * If device is not there any more, skip any shutdown commands.
1201 	 */
1202 	gone = (nvme_mmio_read_4(ctrlr, csts) == 0xffffffff);
1203 	if (gone)
1204 		nvme_ctrlr_fail(ctrlr);
1205 	else
1206 		nvme_notify_fail_consumers(ctrlr);
1207 
1208 	for (i = 0; i < NVME_MAX_NAMESPACES; i++)
1209 		nvme_ns_destruct(&ctrlr->ns[i]);
1210 
1211 	if (ctrlr->cdev)
1212 		destroy_dev(ctrlr->cdev);
1213 
1214 	if (ctrlr->is_initialized) {
1215 		if (!gone)
1216 			nvme_ctrlr_delete_qpairs(ctrlr);
1217 		for (i = 0; i < ctrlr->num_io_queues; i++)
1218 			nvme_io_qpair_destroy(&ctrlr->ioq[i]);
1219 		free(ctrlr->ioq, M_NVME);
1220 		nvme_admin_qpair_destroy(&ctrlr->adminq);
1221 	}
1222 
1223 	/*
1224 	 *  Notify the controller of a shutdown, even though this is due to
1225 	 *   a driver unload, not a system shutdown (this path is not invoked
1226 	 *   during shutdown).  This ensures the controller receives a
1227 	 *   shutdown notification in case the system is shutdown before
1228 	 *   reloading the driver.
1229 	 */
1230 	if (!gone)
1231 		nvme_ctrlr_shutdown(ctrlr);
1232 
1233 	if (!gone)
1234 		nvme_ctrlr_disable(ctrlr);
1235 
1236 	if (ctrlr->taskqueue)
1237 		taskqueue_free(ctrlr->taskqueue);
1238 
1239 	if (ctrlr->tag)
1240 		bus_teardown_intr(ctrlr->dev, ctrlr->res, ctrlr->tag);
1241 
1242 	if (ctrlr->res)
1243 		bus_release_resource(ctrlr->dev, SYS_RES_IRQ,
1244 		    rman_get_rid(ctrlr->res), ctrlr->res);
1245 
1246 	if (ctrlr->bar4_resource != NULL) {
1247 		bus_release_resource(dev, SYS_RES_MEMORY,
1248 		    ctrlr->bar4_resource_id, ctrlr->bar4_resource);
1249 	}
1250 
1251 	bus_release_resource(dev, SYS_RES_MEMORY,
1252 	    ctrlr->resource_id, ctrlr->resource);
1253 
1254 nores:
1255 	mtx_destroy(&ctrlr->lock);
1256 }
1257 
1258 void
1259 nvme_ctrlr_shutdown(struct nvme_controller *ctrlr)
1260 {
1261 	uint32_t	cc;
1262 	uint32_t	csts;
1263 	int		ticks = 0;
1264 
1265 	cc = nvme_mmio_read_4(ctrlr, cc);
1266 	cc &= ~(NVME_CC_REG_SHN_MASK << NVME_CC_REG_SHN_SHIFT);
1267 	cc |= NVME_SHN_NORMAL << NVME_CC_REG_SHN_SHIFT;
1268 	nvme_mmio_write_4(ctrlr, cc, cc);
1269 
1270 	while (1) {
1271 		csts = nvme_mmio_read_4(ctrlr, csts);
1272 		if (csts == 0xffffffff)		/* Hot unplug. */
1273 			break;
1274 		if (NVME_CSTS_GET_SHST(csts) == NVME_SHST_COMPLETE)
1275 			break;
1276 		if (ticks++ > 5*hz) {
1277 			nvme_printf(ctrlr, "did not complete shutdown within"
1278 			    " 5 seconds of notification\n");
1279 			break;
1280 		}
1281 		pause("nvme shn", 1);
1282 	}
1283 }
1284 
1285 void
1286 nvme_ctrlr_submit_admin_request(struct nvme_controller *ctrlr,
1287     struct nvme_request *req)
1288 {
1289 
1290 	nvme_qpair_submit_request(&ctrlr->adminq, req);
1291 }
1292 
1293 void
1294 nvme_ctrlr_submit_io_request(struct nvme_controller *ctrlr,
1295     struct nvme_request *req)
1296 {
1297 	struct nvme_qpair       *qpair;
1298 
1299 	qpair = &ctrlr->ioq[curcpu / ctrlr->num_cpus_per_ioq];
1300 	nvme_qpair_submit_request(qpair, req);
1301 }
1302 
1303 device_t
1304 nvme_ctrlr_get_device(struct nvme_controller *ctrlr)
1305 {
1306 
1307 	return (ctrlr->dev);
1308 }
1309 
1310 const struct nvme_controller_data *
1311 nvme_ctrlr_get_data(struct nvme_controller *ctrlr)
1312 {
1313 
1314 	return (&ctrlr->cdata);
1315 }
1316 
1317 int
1318 nvme_ctrlr_suspend(struct nvme_controller *ctrlr)
1319 {
1320 	int to = hz;
1321 
1322 	/*
1323 	 * Can't touch failed controllers, so it's already suspended.
1324 	 */
1325 	if (ctrlr->is_failed)
1326 		return (0);
1327 
1328 	/*
1329 	 * We don't want the reset taskqueue running, since it does similar
1330 	 * things, so prevent it from running after we start. Wait for any reset
1331 	 * that may have been started to complete. The reset process we follow
1332 	 * will ensure that any new I/O will queue and be given to the hardware
1333 	 * after we resume (though there should be none).
1334 	 */
1335 	while (atomic_cmpset_32(&ctrlr->is_resetting, 0, 1) == 0 && to-- > 0)
1336 		pause("nvmesusp", 1);
1337 	if (to <= 0) {
1338 		nvme_printf(ctrlr,
1339 		    "Competing reset task didn't finish. Try again later.\n");
1340 		return (EWOULDBLOCK);
1341 	}
1342 
1343 	/*
1344 	 * Per Section 7.6.2 of NVMe spec 1.4, to properly suspend, we need to
1345 	 * delete the hardware I/O queues, and then shutdown. This properly
1346 	 * flushes any metadata the drive may have stored so it can survive
1347 	 * having its power removed and prevents the unsafe shutdown count from
1348 	 * incriminating. Once we delete the qpairs, we have to disable them
1349 	 * before shutting down. The delay is out of paranoia in
1350 	 * nvme_ctrlr_hw_reset, and is repeated here (though we should have no
1351 	 * pending I/O that the delay copes with).
1352 	 */
1353 	nvme_ctrlr_delete_qpairs(ctrlr);
1354 	nvme_ctrlr_disable_qpairs(ctrlr);
1355 	DELAY(100*1000);
1356 	nvme_ctrlr_shutdown(ctrlr);
1357 
1358 	return (0);
1359 }
1360 
1361 int
1362 nvme_ctrlr_resume(struct nvme_controller *ctrlr)
1363 {
1364 
1365 	/*
1366 	 * Can't touch failed controllers, so nothing to do to resume.
1367 	 */
1368 	if (ctrlr->is_failed)
1369 		return (0);
1370 
1371 	/*
1372 	 * Have to reset the hardware twice, just like we do on attach. See
1373 	 * nmve_attach() for why.
1374 	 */
1375 	if (nvme_ctrlr_hw_reset(ctrlr) != 0)
1376 		goto fail;
1377 	if (nvme_ctrlr_hw_reset(ctrlr) != 0)
1378 		goto fail;
1379 
1380 	/*
1381 	 * Now that we're reset the hardware, we can restart the controller. Any
1382 	 * I/O that was pending is requeued. Any admin commands are aborted with
1383 	 * an error. Once we've restarted, take the controller out of reset.
1384 	 */
1385 	nvme_ctrlr_start(ctrlr, true);
1386 	atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
1387 
1388 	return (0);
1389 fail:
1390 	/*
1391 	 * Since we can't bring the controller out of reset, announce and fail
1392 	 * the controller. However, we have to return success for the resume
1393 	 * itself, due to questionable APIs.
1394 	 */
1395 	nvme_printf(ctrlr, "Failed to reset on resume, failing.\n");
1396 	nvme_ctrlr_fail(ctrlr);
1397 	atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
1398 	return (0);
1399 }
1400