xref: /freebsd/sys/dev/nvme/nvme_ctrlr.c (revision b0d29bc47dba79f6f38e67eabadfb4b32ffd9390)
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 #include <vm/vm.h>
45 
46 #include "nvme_private.h"
47 
48 #define B4_CHK_RDY_DELAY_MS	2300		/* work around controller bug */
49 
50 static void nvme_ctrlr_construct_and_submit_aer(struct nvme_controller *ctrlr,
51 						struct nvme_async_event_request *aer);
52 
53 static int
54 nvme_ctrlr_construct_admin_qpair(struct nvme_controller *ctrlr)
55 {
56 	struct nvme_qpair	*qpair;
57 	uint32_t		num_entries;
58 	int			error;
59 
60 	qpair = &ctrlr->adminq;
61 	qpair->id = 0;
62 	qpair->cpu = CPU_FFS(&cpuset_domain[ctrlr->domain]) - 1;
63 	qpair->domain = ctrlr->domain;
64 
65 	num_entries = NVME_ADMIN_ENTRIES;
66 	TUNABLE_INT_FETCH("hw.nvme.admin_entries", &num_entries);
67 	/*
68 	 * If admin_entries was overridden to an invalid value, revert it
69 	 *  back to our default value.
70 	 */
71 	if (num_entries < NVME_MIN_ADMIN_ENTRIES ||
72 	    num_entries > NVME_MAX_ADMIN_ENTRIES) {
73 		nvme_printf(ctrlr, "invalid hw.nvme.admin_entries=%d "
74 		    "specified\n", num_entries);
75 		num_entries = NVME_ADMIN_ENTRIES;
76 	}
77 
78 	/*
79 	 * The admin queue's max xfer size is treated differently than the
80 	 *  max I/O xfer size.  16KB is sufficient here - maybe even less?
81 	 */
82 	error = nvme_qpair_construct(qpair, num_entries, NVME_ADMIN_TRACKERS,
83 	     ctrlr);
84 	return (error);
85 }
86 
87 #define QP(ctrlr, c)	((c) * (ctrlr)->num_io_queues / mp_ncpus)
88 
89 static int
90 nvme_ctrlr_construct_io_qpairs(struct nvme_controller *ctrlr)
91 {
92 	struct nvme_qpair	*qpair;
93 	uint32_t		cap_lo;
94 	uint16_t		mqes;
95 	int			c, error, i, n;
96 	int			num_entries, num_trackers, max_entries;
97 
98 	/*
99 	 * NVMe spec sets a hard limit of 64K max entries, but devices may
100 	 * specify a smaller limit, so we need to check the MQES field in the
101 	 * capabilities register. We have to cap the number of entries to the
102 	 * current stride allows for in BAR 0/1, otherwise the remainder entries
103 	 * are inaccessable. MQES should reflect this, and this is just a
104 	 * fail-safe.
105 	 */
106 	max_entries =
107 	    (rman_get_size(ctrlr->resource) - nvme_mmio_offsetof(doorbell[0])) /
108 	    (1 << (ctrlr->dstrd + 1));
109 	num_entries = NVME_IO_ENTRIES;
110 	TUNABLE_INT_FETCH("hw.nvme.io_entries", &num_entries);
111 	cap_lo = nvme_mmio_read_4(ctrlr, cap_lo);
112 	mqes = NVME_CAP_LO_MQES(cap_lo);
113 	num_entries = min(num_entries, mqes + 1);
114 	num_entries = min(num_entries, max_entries);
115 
116 	num_trackers = NVME_IO_TRACKERS;
117 	TUNABLE_INT_FETCH("hw.nvme.io_trackers", &num_trackers);
118 
119 	num_trackers = max(num_trackers, NVME_MIN_IO_TRACKERS);
120 	num_trackers = min(num_trackers, NVME_MAX_IO_TRACKERS);
121 	/*
122 	 * No need to have more trackers than entries in the submit queue.  Note
123 	 * also that for a queue size of N, we can only have (N-1) commands
124 	 * outstanding, hence the "-1" here.
125 	 */
126 	num_trackers = min(num_trackers, (num_entries-1));
127 
128 	/*
129 	 * Our best estimate for the maximum number of I/Os that we should
130 	 * normally have in flight at one time. This should be viewed as a hint,
131 	 * not a hard limit and will need to be revisited when the upper layers
132 	 * of the storage system grows multi-queue support.
133 	 */
134 	ctrlr->max_hw_pend_io = num_trackers * ctrlr->num_io_queues * 3 / 4;
135 
136 	ctrlr->ioq = malloc(ctrlr->num_io_queues * sizeof(struct nvme_qpair),
137 	    M_NVME, M_ZERO | M_WAITOK);
138 
139 	for (i = c = n = 0; i < ctrlr->num_io_queues; i++, c += n) {
140 		qpair = &ctrlr->ioq[i];
141 
142 		/*
143 		 * Admin queue has ID=0. IO queues start at ID=1 -
144 		 *  hence the 'i+1' here.
145 		 */
146 		qpair->id = i + 1;
147 		if (ctrlr->num_io_queues > 1) {
148 			/* Find number of CPUs served by this queue. */
149 			for (n = 1; QP(ctrlr, c + n) == i; n++)
150 				;
151 			/* Shuffle multiple NVMe devices between CPUs. */
152 			qpair->cpu = c + (device_get_unit(ctrlr->dev)+n/2) % n;
153 			qpair->domain = pcpu_find(qpair->cpu)->pc_domain;
154 		} else {
155 			qpair->cpu = CPU_FFS(&cpuset_domain[ctrlr->domain]) - 1;
156 			qpair->domain = ctrlr->domain;
157 		}
158 
159 		/*
160 		 * For I/O queues, use the controller-wide max_xfer_size
161 		 *  calculated in nvme_attach().
162 		 */
163 		error = nvme_qpair_construct(qpair, num_entries, num_trackers,
164 		    ctrlr);
165 		if (error)
166 			return (error);
167 
168 		/*
169 		 * Do not bother binding interrupts if we only have one I/O
170 		 *  interrupt thread for this controller.
171 		 */
172 		if (ctrlr->num_io_queues > 1)
173 			bus_bind_intr(ctrlr->dev, qpair->res, qpair->cpu);
174 	}
175 
176 	return (0);
177 }
178 
179 static void
180 nvme_ctrlr_fail(struct nvme_controller *ctrlr)
181 {
182 	int i;
183 
184 	ctrlr->is_failed = true;
185 	nvme_admin_qpair_disable(&ctrlr->adminq);
186 	nvme_qpair_fail(&ctrlr->adminq);
187 	if (ctrlr->ioq != NULL) {
188 		for (i = 0; i < ctrlr->num_io_queues; i++) {
189 			nvme_io_qpair_disable(&ctrlr->ioq[i]);
190 			nvme_qpair_fail(&ctrlr->ioq[i]);
191 		}
192 	}
193 	nvme_notify_fail_consumers(ctrlr);
194 }
195 
196 void
197 nvme_ctrlr_post_failed_request(struct nvme_controller *ctrlr,
198     struct nvme_request *req)
199 {
200 
201 	mtx_lock(&ctrlr->lock);
202 	STAILQ_INSERT_TAIL(&ctrlr->fail_req, req, stailq);
203 	mtx_unlock(&ctrlr->lock);
204 	taskqueue_enqueue(ctrlr->taskqueue, &ctrlr->fail_req_task);
205 }
206 
207 static void
208 nvme_ctrlr_fail_req_task(void *arg, int pending)
209 {
210 	struct nvme_controller	*ctrlr = arg;
211 	struct nvme_request	*req;
212 
213 	mtx_lock(&ctrlr->lock);
214 	while ((req = STAILQ_FIRST(&ctrlr->fail_req)) != NULL) {
215 		STAILQ_REMOVE_HEAD(&ctrlr->fail_req, stailq);
216 		mtx_unlock(&ctrlr->lock);
217 		nvme_qpair_manual_complete_request(req->qpair, req,
218 		    NVME_SCT_GENERIC, NVME_SC_ABORTED_BY_REQUEST);
219 		mtx_lock(&ctrlr->lock);
220 	}
221 	mtx_unlock(&ctrlr->lock);
222 }
223 
224 static int
225 nvme_ctrlr_wait_for_ready(struct nvme_controller *ctrlr, int desired_val)
226 {
227 	int ms_waited;
228 	uint32_t csts;
229 
230 	ms_waited = 0;
231 	while (1) {
232 		csts = nvme_mmio_read_4(ctrlr, csts);
233 		if (csts == 0xffffffff)		/* Hot unplug. */
234 			return (ENXIO);
235 		if (((csts >> NVME_CSTS_REG_RDY_SHIFT) & NVME_CSTS_REG_RDY_MASK)
236 		    == desired_val)
237 			break;
238 		if (ms_waited++ > ctrlr->ready_timeout_in_ms) {
239 			nvme_printf(ctrlr, "controller ready did not become %d "
240 			    "within %d ms\n", desired_val, ctrlr->ready_timeout_in_ms);
241 			return (ENXIO);
242 		}
243 		DELAY(1000);
244 	}
245 
246 	return (0);
247 }
248 
249 static int
250 nvme_ctrlr_disable(struct nvme_controller *ctrlr)
251 {
252 	uint32_t cc;
253 	uint32_t csts;
254 	uint8_t  en, rdy;
255 	int err;
256 
257 	cc = nvme_mmio_read_4(ctrlr, cc);
258 	csts = nvme_mmio_read_4(ctrlr, csts);
259 
260 	en = (cc >> NVME_CC_REG_EN_SHIFT) & NVME_CC_REG_EN_MASK;
261 	rdy = (csts >> NVME_CSTS_REG_RDY_SHIFT) & NVME_CSTS_REG_RDY_MASK;
262 
263 	/*
264 	 * Per 3.1.5 in NVME 1.3 spec, transitioning CC.EN from 0 to 1
265 	 * when CSTS.RDY is 1 or transitioning CC.EN from 1 to 0 when
266 	 * CSTS.RDY is 0 "has undefined results" So make sure that CSTS.RDY
267 	 * isn't the desired value. Short circuit if we're already disabled.
268 	 */
269 	if (en == 1) {
270 		if (rdy == 0) {
271 			/* EN == 1, wait for  RDY == 1 or fail */
272 			err = nvme_ctrlr_wait_for_ready(ctrlr, 1);
273 			if (err != 0)
274 				return (err);
275 		}
276 	} else {
277 		/* EN == 0 already wait for RDY == 0 */
278 		if (rdy == 0)
279 			return (0);
280 		else
281 			return (nvme_ctrlr_wait_for_ready(ctrlr, 0));
282 	}
283 
284 	cc &= ~NVME_CC_REG_EN_MASK;
285 	nvme_mmio_write_4(ctrlr, cc, cc);
286 	/*
287 	 * Some drives have issues with accessing the mmio after we
288 	 * disable, so delay for a bit after we write the bit to
289 	 * cope with these issues.
290 	 */
291 	if (ctrlr->quirks & QUIRK_DELAY_B4_CHK_RDY)
292 		pause("nvmeR", B4_CHK_RDY_DELAY_MS * hz / 1000);
293 	return (nvme_ctrlr_wait_for_ready(ctrlr, 0));
294 }
295 
296 static int
297 nvme_ctrlr_enable(struct nvme_controller *ctrlr)
298 {
299 	uint32_t	cc;
300 	uint32_t	csts;
301 	uint32_t	aqa;
302 	uint32_t	qsize;
303 	uint8_t		en, rdy;
304 	int		err;
305 
306 	cc = nvme_mmio_read_4(ctrlr, cc);
307 	csts = nvme_mmio_read_4(ctrlr, csts);
308 
309 	en = (cc >> NVME_CC_REG_EN_SHIFT) & NVME_CC_REG_EN_MASK;
310 	rdy = (csts >> NVME_CSTS_REG_RDY_SHIFT) & NVME_CSTS_REG_RDY_MASK;
311 
312 	/*
313 	 * See note in nvme_ctrlr_disable. Short circuit if we're already enabled.
314 	 */
315 	if (en == 1) {
316 		if (rdy == 1)
317 			return (0);
318 		else
319 			return (nvme_ctrlr_wait_for_ready(ctrlr, 1));
320 	} else {
321 		/* EN == 0 already wait for RDY == 0 or fail */
322 		err = nvme_ctrlr_wait_for_ready(ctrlr, 0);
323 		if (err != 0)
324 			return (err);
325 	}
326 
327 	nvme_mmio_write_8(ctrlr, asq, ctrlr->adminq.cmd_bus_addr);
328 	DELAY(5000);
329 	nvme_mmio_write_8(ctrlr, acq, ctrlr->adminq.cpl_bus_addr);
330 	DELAY(5000);
331 
332 	/* acqs and asqs are 0-based. */
333 	qsize = ctrlr->adminq.num_entries - 1;
334 
335 	aqa = 0;
336 	aqa = (qsize & NVME_AQA_REG_ACQS_MASK) << NVME_AQA_REG_ACQS_SHIFT;
337 	aqa |= (qsize & NVME_AQA_REG_ASQS_MASK) << NVME_AQA_REG_ASQS_SHIFT;
338 	nvme_mmio_write_4(ctrlr, aqa, aqa);
339 	DELAY(5000);
340 
341 	/* Initialization values for CC */
342 	cc = 0;
343 	cc |= 1 << NVME_CC_REG_EN_SHIFT;
344 	cc |= 0 << NVME_CC_REG_CSS_SHIFT;
345 	cc |= 0 << NVME_CC_REG_AMS_SHIFT;
346 	cc |= 0 << NVME_CC_REG_SHN_SHIFT;
347 	cc |= 6 << NVME_CC_REG_IOSQES_SHIFT; /* SQ entry size == 64 == 2^6 */
348 	cc |= 4 << NVME_CC_REG_IOCQES_SHIFT; /* CQ entry size == 16 == 2^4 */
349 
350 	/* This evaluates to 0, which is according to spec. */
351 	cc |= (PAGE_SIZE >> 13) << NVME_CC_REG_MPS_SHIFT;
352 
353 	nvme_mmio_write_4(ctrlr, cc, cc);
354 
355 	return (nvme_ctrlr_wait_for_ready(ctrlr, 1));
356 }
357 
358 static void
359 nvme_ctrlr_disable_qpairs(struct nvme_controller *ctrlr)
360 {
361 	int i;
362 
363 	nvme_admin_qpair_disable(&ctrlr->adminq);
364 	/*
365 	 * I/O queues are not allocated before the initial HW
366 	 *  reset, so do not try to disable them.  Use is_initialized
367 	 *  to determine if this is the initial HW reset.
368 	 */
369 	if (ctrlr->is_initialized) {
370 		for (i = 0; i < ctrlr->num_io_queues; i++)
371 			nvme_io_qpair_disable(&ctrlr->ioq[i]);
372 	}
373 }
374 
375 int
376 nvme_ctrlr_hw_reset(struct nvme_controller *ctrlr)
377 {
378 	int err;
379 
380 	nvme_ctrlr_disable_qpairs(ctrlr);
381 
382 	DELAY(100*1000);
383 
384 	err = nvme_ctrlr_disable(ctrlr);
385 	if (err != 0)
386 		return err;
387 	return (nvme_ctrlr_enable(ctrlr));
388 }
389 
390 void
391 nvme_ctrlr_reset(struct nvme_controller *ctrlr)
392 {
393 	int cmpset;
394 
395 	cmpset = atomic_cmpset_32(&ctrlr->is_resetting, 0, 1);
396 
397 	if (cmpset == 0 || ctrlr->is_failed)
398 		/*
399 		 * Controller is already resetting or has failed.  Return
400 		 *  immediately since there is no need to kick off another
401 		 *  reset in these cases.
402 		 */
403 		return;
404 
405 	taskqueue_enqueue(ctrlr->taskqueue, &ctrlr->reset_task);
406 }
407 
408 static int
409 nvme_ctrlr_identify(struct nvme_controller *ctrlr)
410 {
411 	struct nvme_completion_poll_status	status;
412 
413 	status.done = 0;
414 	nvme_ctrlr_cmd_identify_controller(ctrlr, &ctrlr->cdata,
415 	    nvme_completion_poll_cb, &status);
416 	nvme_completion_poll(&status);
417 	if (nvme_completion_is_error(&status.cpl)) {
418 		nvme_printf(ctrlr, "nvme_identify_controller failed!\n");
419 		return (ENXIO);
420 	}
421 
422 	/* Convert data to host endian */
423 	nvme_controller_data_swapbytes(&ctrlr->cdata);
424 
425 	/*
426 	 * Use MDTS to ensure our default max_xfer_size doesn't exceed what the
427 	 *  controller supports.
428 	 */
429 	if (ctrlr->cdata.mdts > 0)
430 		ctrlr->max_xfer_size = min(ctrlr->max_xfer_size,
431 		    ctrlr->min_page_size * (1 << (ctrlr->cdata.mdts)));
432 
433 	return (0);
434 }
435 
436 static int
437 nvme_ctrlr_set_num_qpairs(struct nvme_controller *ctrlr)
438 {
439 	struct nvme_completion_poll_status	status;
440 	int					cq_allocated, sq_allocated;
441 
442 	status.done = 0;
443 	nvme_ctrlr_cmd_set_num_queues(ctrlr, ctrlr->num_io_queues,
444 	    nvme_completion_poll_cb, &status);
445 	nvme_completion_poll(&status);
446 	if (nvme_completion_is_error(&status.cpl)) {
447 		nvme_printf(ctrlr, "nvme_ctrlr_set_num_qpairs failed!\n");
448 		return (ENXIO);
449 	}
450 
451 	/*
452 	 * Data in cdw0 is 0-based.
453 	 * Lower 16-bits indicate number of submission queues allocated.
454 	 * Upper 16-bits indicate number of completion queues allocated.
455 	 */
456 	sq_allocated = (status.cpl.cdw0 & 0xFFFF) + 1;
457 	cq_allocated = (status.cpl.cdw0 >> 16) + 1;
458 
459 	/*
460 	 * Controller may allocate more queues than we requested,
461 	 *  so use the minimum of the number requested and what was
462 	 *  actually allocated.
463 	 */
464 	ctrlr->num_io_queues = min(ctrlr->num_io_queues, sq_allocated);
465 	ctrlr->num_io_queues = min(ctrlr->num_io_queues, cq_allocated);
466 	if (ctrlr->num_io_queues > vm_ndomains)
467 		ctrlr->num_io_queues -= ctrlr->num_io_queues % vm_ndomains;
468 
469 	return (0);
470 }
471 
472 static int
473 nvme_ctrlr_create_qpairs(struct nvme_controller *ctrlr)
474 {
475 	struct nvme_completion_poll_status	status;
476 	struct nvme_qpair			*qpair;
477 	int					i;
478 
479 	for (i = 0; i < ctrlr->num_io_queues; i++) {
480 		qpair = &ctrlr->ioq[i];
481 
482 		status.done = 0;
483 		nvme_ctrlr_cmd_create_io_cq(ctrlr, qpair,
484 		    nvme_completion_poll_cb, &status);
485 		nvme_completion_poll(&status);
486 		if (nvme_completion_is_error(&status.cpl)) {
487 			nvme_printf(ctrlr, "nvme_create_io_cq failed!\n");
488 			return (ENXIO);
489 		}
490 
491 		status.done = 0;
492 		nvme_ctrlr_cmd_create_io_sq(qpair->ctrlr, qpair,
493 		    nvme_completion_poll_cb, &status);
494 		nvme_completion_poll(&status);
495 		if (nvme_completion_is_error(&status.cpl)) {
496 			nvme_printf(ctrlr, "nvme_create_io_sq failed!\n");
497 			return (ENXIO);
498 		}
499 	}
500 
501 	return (0);
502 }
503 
504 static int
505 nvme_ctrlr_delete_qpairs(struct nvme_controller *ctrlr)
506 {
507 	struct nvme_completion_poll_status	status;
508 	struct nvme_qpair			*qpair;
509 
510 	for (int i = 0; i < ctrlr->num_io_queues; i++) {
511 		qpair = &ctrlr->ioq[i];
512 
513 		status.done = 0;
514 		nvme_ctrlr_cmd_delete_io_sq(ctrlr, qpair,
515 		    nvme_completion_poll_cb, &status);
516 		nvme_completion_poll(&status);
517 		if (nvme_completion_is_error(&status.cpl)) {
518 			nvme_printf(ctrlr, "nvme_destroy_io_sq failed!\n");
519 			return (ENXIO);
520 		}
521 
522 		status.done = 0;
523 		nvme_ctrlr_cmd_delete_io_cq(ctrlr, qpair,
524 		    nvme_completion_poll_cb, &status);
525 		nvme_completion_poll(&status);
526 		if (nvme_completion_is_error(&status.cpl)) {
527 			nvme_printf(ctrlr, "nvme_destroy_io_cq failed!\n");
528 			return (ENXIO);
529 		}
530 	}
531 
532 	return (0);
533 }
534 
535 static int
536 nvme_ctrlr_construct_namespaces(struct nvme_controller *ctrlr)
537 {
538 	struct nvme_namespace	*ns;
539 	uint32_t 		i;
540 
541 	for (i = 0; i < min(ctrlr->cdata.nn, NVME_MAX_NAMESPACES); i++) {
542 		ns = &ctrlr->ns[i];
543 		nvme_ns_construct(ns, i+1, ctrlr);
544 	}
545 
546 	return (0);
547 }
548 
549 static bool
550 is_log_page_id_valid(uint8_t page_id)
551 {
552 
553 	switch (page_id) {
554 	case NVME_LOG_ERROR:
555 	case NVME_LOG_HEALTH_INFORMATION:
556 	case NVME_LOG_FIRMWARE_SLOT:
557 	case NVME_LOG_CHANGED_NAMESPACE:
558 	case NVME_LOG_COMMAND_EFFECT:
559 	case NVME_LOG_RES_NOTIFICATION:
560 	case NVME_LOG_SANITIZE_STATUS:
561 		return (true);
562 	}
563 
564 	return (false);
565 }
566 
567 static uint32_t
568 nvme_ctrlr_get_log_page_size(struct nvme_controller *ctrlr, uint8_t page_id)
569 {
570 	uint32_t	log_page_size;
571 
572 	switch (page_id) {
573 	case NVME_LOG_ERROR:
574 		log_page_size = min(
575 		    sizeof(struct nvme_error_information_entry) *
576 		    (ctrlr->cdata.elpe + 1), NVME_MAX_AER_LOG_SIZE);
577 		break;
578 	case NVME_LOG_HEALTH_INFORMATION:
579 		log_page_size = sizeof(struct nvme_health_information_page);
580 		break;
581 	case NVME_LOG_FIRMWARE_SLOT:
582 		log_page_size = sizeof(struct nvme_firmware_page);
583 		break;
584 	case NVME_LOG_CHANGED_NAMESPACE:
585 		log_page_size = sizeof(struct nvme_ns_list);
586 		break;
587 	case NVME_LOG_COMMAND_EFFECT:
588 		log_page_size = sizeof(struct nvme_command_effects_page);
589 		break;
590 	case NVME_LOG_RES_NOTIFICATION:
591 		log_page_size = sizeof(struct nvme_res_notification_page);
592 		break;
593 	case NVME_LOG_SANITIZE_STATUS:
594 		log_page_size = sizeof(struct nvme_sanitize_status_page);
595 		break;
596 	default:
597 		log_page_size = 0;
598 		break;
599 	}
600 
601 	return (log_page_size);
602 }
603 
604 static void
605 nvme_ctrlr_log_critical_warnings(struct nvme_controller *ctrlr,
606     uint8_t state)
607 {
608 
609 	if (state & NVME_CRIT_WARN_ST_AVAILABLE_SPARE)
610 		nvme_printf(ctrlr, "available spare space below threshold\n");
611 
612 	if (state & NVME_CRIT_WARN_ST_TEMPERATURE)
613 		nvme_printf(ctrlr, "temperature above threshold\n");
614 
615 	if (state & NVME_CRIT_WARN_ST_DEVICE_RELIABILITY)
616 		nvme_printf(ctrlr, "device reliability degraded\n");
617 
618 	if (state & NVME_CRIT_WARN_ST_READ_ONLY)
619 		nvme_printf(ctrlr, "media placed in read only mode\n");
620 
621 	if (state & NVME_CRIT_WARN_ST_VOLATILE_MEMORY_BACKUP)
622 		nvme_printf(ctrlr, "volatile memory backup device failed\n");
623 
624 	if (state & NVME_CRIT_WARN_ST_RESERVED_MASK)
625 		nvme_printf(ctrlr,
626 		    "unknown critical warning(s): state = 0x%02x\n", state);
627 }
628 
629 static void
630 nvme_ctrlr_async_event_log_page_cb(void *arg, const struct nvme_completion *cpl)
631 {
632 	struct nvme_async_event_request		*aer = arg;
633 	struct nvme_health_information_page	*health_info;
634 	struct nvme_ns_list			*nsl;
635 	struct nvme_error_information_entry	*err;
636 	int i;
637 
638 	/*
639 	 * If the log page fetch for some reason completed with an error,
640 	 *  don't pass log page data to the consumers.  In practice, this case
641 	 *  should never happen.
642 	 */
643 	if (nvme_completion_is_error(cpl))
644 		nvme_notify_async_consumers(aer->ctrlr, &aer->cpl,
645 		    aer->log_page_id, NULL, 0);
646 	else {
647 		/* Convert data to host endian */
648 		switch (aer->log_page_id) {
649 		case NVME_LOG_ERROR:
650 			err = (struct nvme_error_information_entry *)aer->log_page_buffer;
651 			for (i = 0; i < (aer->ctrlr->cdata.elpe + 1); i++)
652 				nvme_error_information_entry_swapbytes(err++);
653 			break;
654 		case NVME_LOG_HEALTH_INFORMATION:
655 			nvme_health_information_page_swapbytes(
656 			    (struct nvme_health_information_page *)aer->log_page_buffer);
657 			break;
658 		case NVME_LOG_FIRMWARE_SLOT:
659 			nvme_firmware_page_swapbytes(
660 			    (struct nvme_firmware_page *)aer->log_page_buffer);
661 			break;
662 		case NVME_LOG_CHANGED_NAMESPACE:
663 			nvme_ns_list_swapbytes(
664 			    (struct nvme_ns_list *)aer->log_page_buffer);
665 			break;
666 		case NVME_LOG_COMMAND_EFFECT:
667 			nvme_command_effects_page_swapbytes(
668 			    (struct nvme_command_effects_page *)aer->log_page_buffer);
669 			break;
670 		case NVME_LOG_RES_NOTIFICATION:
671 			nvme_res_notification_page_swapbytes(
672 			    (struct nvme_res_notification_page *)aer->log_page_buffer);
673 			break;
674 		case NVME_LOG_SANITIZE_STATUS:
675 			nvme_sanitize_status_page_swapbytes(
676 			    (struct nvme_sanitize_status_page *)aer->log_page_buffer);
677 			break;
678 		case INTEL_LOG_TEMP_STATS:
679 			intel_log_temp_stats_swapbytes(
680 			    (struct intel_log_temp_stats *)aer->log_page_buffer);
681 			break;
682 		default:
683 			break;
684 		}
685 
686 		if (aer->log_page_id == NVME_LOG_HEALTH_INFORMATION) {
687 			health_info = (struct nvme_health_information_page *)
688 			    aer->log_page_buffer;
689 			nvme_ctrlr_log_critical_warnings(aer->ctrlr,
690 			    health_info->critical_warning);
691 			/*
692 			 * Critical warnings reported through the
693 			 *  SMART/health log page are persistent, so
694 			 *  clear the associated bits in the async event
695 			 *  config so that we do not receive repeated
696 			 *  notifications for the same event.
697 			 */
698 			aer->ctrlr->async_event_config &=
699 			    ~health_info->critical_warning;
700 			nvme_ctrlr_cmd_set_async_event_config(aer->ctrlr,
701 			    aer->ctrlr->async_event_config, NULL, NULL);
702 		} else if (aer->log_page_id == NVME_LOG_CHANGED_NAMESPACE &&
703 		    !nvme_use_nvd) {
704 			nsl = (struct nvme_ns_list *)aer->log_page_buffer;
705 			for (i = 0; i < nitems(nsl->ns) && nsl->ns[i] != 0; i++) {
706 				if (nsl->ns[i] > NVME_MAX_NAMESPACES)
707 					break;
708 				nvme_notify_ns(aer->ctrlr, nsl->ns[i]);
709 			}
710 		}
711 
712 
713 		/*
714 		 * Pass the cpl data from the original async event completion,
715 		 *  not the log page fetch.
716 		 */
717 		nvme_notify_async_consumers(aer->ctrlr, &aer->cpl,
718 		    aer->log_page_id, aer->log_page_buffer, aer->log_page_size);
719 	}
720 
721 	/*
722 	 * Repost another asynchronous event request to replace the one
723 	 *  that just completed.
724 	 */
725 	nvme_ctrlr_construct_and_submit_aer(aer->ctrlr, aer);
726 }
727 
728 static void
729 nvme_ctrlr_async_event_cb(void *arg, const struct nvme_completion *cpl)
730 {
731 	struct nvme_async_event_request	*aer = arg;
732 
733 	if (nvme_completion_is_error(cpl)) {
734 		/*
735 		 *  Do not retry failed async event requests.  This avoids
736 		 *  infinite loops where a new async event request is submitted
737 		 *  to replace the one just failed, only to fail again and
738 		 *  perpetuate the loop.
739 		 */
740 		return;
741 	}
742 
743 	/* Associated log page is in bits 23:16 of completion entry dw0. */
744 	aer->log_page_id = (cpl->cdw0 & 0xFF0000) >> 16;
745 
746 	nvme_printf(aer->ctrlr, "async event occurred (type 0x%x, info 0x%02x,"
747 	    " page 0x%02x)\n", (cpl->cdw0 & 0x07), (cpl->cdw0 & 0xFF00) >> 8,
748 	    aer->log_page_id);
749 
750 	if (is_log_page_id_valid(aer->log_page_id)) {
751 		aer->log_page_size = nvme_ctrlr_get_log_page_size(aer->ctrlr,
752 		    aer->log_page_id);
753 		memcpy(&aer->cpl, cpl, sizeof(*cpl));
754 		nvme_ctrlr_cmd_get_log_page(aer->ctrlr, aer->log_page_id,
755 		    NVME_GLOBAL_NAMESPACE_TAG, aer->log_page_buffer,
756 		    aer->log_page_size, nvme_ctrlr_async_event_log_page_cb,
757 		    aer);
758 		/* Wait to notify consumers until after log page is fetched. */
759 	} else {
760 		nvme_notify_async_consumers(aer->ctrlr, cpl, aer->log_page_id,
761 		    NULL, 0);
762 
763 		/*
764 		 * Repost another asynchronous event request to replace the one
765 		 *  that just completed.
766 		 */
767 		nvme_ctrlr_construct_and_submit_aer(aer->ctrlr, aer);
768 	}
769 }
770 
771 static void
772 nvme_ctrlr_construct_and_submit_aer(struct nvme_controller *ctrlr,
773     struct nvme_async_event_request *aer)
774 {
775 	struct nvme_request *req;
776 
777 	aer->ctrlr = ctrlr;
778 	req = nvme_allocate_request_null(nvme_ctrlr_async_event_cb, aer);
779 	aer->req = req;
780 
781 	/*
782 	 * Disable timeout here, since asynchronous event requests should by
783 	 *  nature never be timed out.
784 	 */
785 	req->timeout = false;
786 	req->cmd.opc = NVME_OPC_ASYNC_EVENT_REQUEST;
787 	nvme_ctrlr_submit_admin_request(ctrlr, req);
788 }
789 
790 static void
791 nvme_ctrlr_configure_aer(struct nvme_controller *ctrlr)
792 {
793 	struct nvme_completion_poll_status	status;
794 	struct nvme_async_event_request		*aer;
795 	uint32_t				i;
796 
797 	ctrlr->async_event_config = NVME_CRIT_WARN_ST_AVAILABLE_SPARE |
798 	    NVME_CRIT_WARN_ST_DEVICE_RELIABILITY |
799 	    NVME_CRIT_WARN_ST_READ_ONLY |
800 	    NVME_CRIT_WARN_ST_VOLATILE_MEMORY_BACKUP;
801 	if (ctrlr->cdata.ver >= NVME_REV(1, 2))
802 		ctrlr->async_event_config |= 0x300;
803 
804 	status.done = 0;
805 	nvme_ctrlr_cmd_get_feature(ctrlr, NVME_FEAT_TEMPERATURE_THRESHOLD,
806 	    0, NULL, 0, nvme_completion_poll_cb, &status);
807 	nvme_completion_poll(&status);
808 	if (nvme_completion_is_error(&status.cpl) ||
809 	    (status.cpl.cdw0 & 0xFFFF) == 0xFFFF ||
810 	    (status.cpl.cdw0 & 0xFFFF) == 0x0000) {
811 		nvme_printf(ctrlr, "temperature threshold not supported\n");
812 	} else
813 		ctrlr->async_event_config |= NVME_CRIT_WARN_ST_TEMPERATURE;
814 
815 	nvme_ctrlr_cmd_set_async_event_config(ctrlr,
816 	    ctrlr->async_event_config, NULL, NULL);
817 
818 	/* aerl is a zero-based value, so we need to add 1 here. */
819 	ctrlr->num_aers = min(NVME_MAX_ASYNC_EVENTS, (ctrlr->cdata.aerl+1));
820 
821 	for (i = 0; i < ctrlr->num_aers; i++) {
822 		aer = &ctrlr->aer[i];
823 		nvme_ctrlr_construct_and_submit_aer(ctrlr, aer);
824 	}
825 }
826 
827 static void
828 nvme_ctrlr_configure_int_coalescing(struct nvme_controller *ctrlr)
829 {
830 
831 	ctrlr->int_coal_time = 0;
832 	TUNABLE_INT_FETCH("hw.nvme.int_coal_time",
833 	    &ctrlr->int_coal_time);
834 
835 	ctrlr->int_coal_threshold = 0;
836 	TUNABLE_INT_FETCH("hw.nvme.int_coal_threshold",
837 	    &ctrlr->int_coal_threshold);
838 
839 	nvme_ctrlr_cmd_set_interrupt_coalescing(ctrlr, ctrlr->int_coal_time,
840 	    ctrlr->int_coal_threshold, NULL, NULL);
841 }
842 
843 static void
844 nvme_ctrlr_hmb_free(struct nvme_controller *ctrlr)
845 {
846 	struct nvme_hmb_chunk *hmbc;
847 	int i;
848 
849 	if (ctrlr->hmb_desc_paddr) {
850 		bus_dmamap_unload(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_map);
851 		bus_dmamem_free(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_vaddr,
852 		    ctrlr->hmb_desc_map);
853 		ctrlr->hmb_desc_paddr = 0;
854 	}
855 	if (ctrlr->hmb_desc_tag) {
856 		bus_dma_tag_destroy(ctrlr->hmb_desc_tag);
857 		ctrlr->hmb_desc_tag = NULL;
858 	}
859 	for (i = 0; i < ctrlr->hmb_nchunks; i++) {
860 		hmbc = &ctrlr->hmb_chunks[i];
861 		bus_dmamap_unload(ctrlr->hmb_tag, hmbc->hmbc_map);
862 		bus_dmamem_free(ctrlr->hmb_tag, hmbc->hmbc_vaddr,
863 		    hmbc->hmbc_map);
864 	}
865 	ctrlr->hmb_nchunks = 0;
866 	if (ctrlr->hmb_tag) {
867 		bus_dma_tag_destroy(ctrlr->hmb_tag);
868 		ctrlr->hmb_tag = NULL;
869 	}
870 	if (ctrlr->hmb_chunks) {
871 		free(ctrlr->hmb_chunks, M_NVME);
872 		ctrlr->hmb_chunks = NULL;
873 	}
874 }
875 
876 static void
877 nvme_ctrlr_hmb_alloc(struct nvme_controller *ctrlr)
878 {
879 	struct nvme_hmb_chunk *hmbc;
880 	size_t pref, min, minc, size;
881 	int err, i;
882 	uint64_t max;
883 
884 	/* Limit HMB to 5% of RAM size per device by default. */
885 	max = (uint64_t)physmem * PAGE_SIZE / 20;
886 	TUNABLE_UINT64_FETCH("hw.nvme.hmb_max", &max);
887 
888 	min = (long long unsigned)ctrlr->cdata.hmmin * 4096;
889 	if (max == 0 || max < min)
890 		return;
891 	pref = MIN((long long unsigned)ctrlr->cdata.hmpre * 4096, max);
892 	minc = MAX(ctrlr->cdata.hmminds * 4096, PAGE_SIZE);
893 	if (min > 0 && ctrlr->cdata.hmmaxd > 0)
894 		minc = MAX(minc, min / ctrlr->cdata.hmmaxd);
895 	ctrlr->hmb_chunk = pref;
896 
897 again:
898 	ctrlr->hmb_chunk = roundup2(ctrlr->hmb_chunk, PAGE_SIZE);
899 	ctrlr->hmb_nchunks = howmany(pref, ctrlr->hmb_chunk);
900 	if (ctrlr->cdata.hmmaxd > 0 && ctrlr->hmb_nchunks > ctrlr->cdata.hmmaxd)
901 		ctrlr->hmb_nchunks = ctrlr->cdata.hmmaxd;
902 	ctrlr->hmb_chunks = malloc(sizeof(struct nvme_hmb_chunk) *
903 	    ctrlr->hmb_nchunks, M_NVME, M_WAITOK);
904 	err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev),
905 	    PAGE_SIZE, 0, BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
906 	    ctrlr->hmb_chunk, 1, ctrlr->hmb_chunk, 0, NULL, NULL, &ctrlr->hmb_tag);
907 	if (err != 0) {
908 		nvme_printf(ctrlr, "HMB tag create failed %d\n", err);
909 		nvme_ctrlr_hmb_free(ctrlr);
910 		return;
911 	}
912 
913 	for (i = 0; i < ctrlr->hmb_nchunks; i++) {
914 		hmbc = &ctrlr->hmb_chunks[i];
915 		if (bus_dmamem_alloc(ctrlr->hmb_tag,
916 		    (void **)&hmbc->hmbc_vaddr, BUS_DMA_NOWAIT,
917 		    &hmbc->hmbc_map)) {
918 			nvme_printf(ctrlr, "failed to alloc HMB\n");
919 			break;
920 		}
921 		if (bus_dmamap_load(ctrlr->hmb_tag, hmbc->hmbc_map,
922 		    hmbc->hmbc_vaddr, ctrlr->hmb_chunk, nvme_single_map,
923 		    &hmbc->hmbc_paddr, BUS_DMA_NOWAIT) != 0) {
924 			bus_dmamem_free(ctrlr->hmb_tag, hmbc->hmbc_vaddr,
925 			    hmbc->hmbc_map);
926 			nvme_printf(ctrlr, "failed to load HMB\n");
927 			break;
928 		}
929 		bus_dmamap_sync(ctrlr->hmb_tag, hmbc->hmbc_map,
930 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
931 	}
932 
933 	if (i < ctrlr->hmb_nchunks && i * ctrlr->hmb_chunk < min &&
934 	    ctrlr->hmb_chunk / 2 >= minc) {
935 		ctrlr->hmb_nchunks = i;
936 		nvme_ctrlr_hmb_free(ctrlr);
937 		ctrlr->hmb_chunk /= 2;
938 		goto again;
939 	}
940 	ctrlr->hmb_nchunks = i;
941 	if (ctrlr->hmb_nchunks * ctrlr->hmb_chunk < min) {
942 		nvme_ctrlr_hmb_free(ctrlr);
943 		return;
944 	}
945 
946 	size = sizeof(struct nvme_hmb_desc) * ctrlr->hmb_nchunks;
947 	err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev),
948 	    16, 0, BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
949 	    size, 1, size, 0, NULL, NULL, &ctrlr->hmb_desc_tag);
950 	if (err != 0) {
951 		nvme_printf(ctrlr, "HMB desc tag create failed %d\n", err);
952 		nvme_ctrlr_hmb_free(ctrlr);
953 		return;
954 	}
955 	if (bus_dmamem_alloc(ctrlr->hmb_desc_tag,
956 	    (void **)&ctrlr->hmb_desc_vaddr, BUS_DMA_WAITOK,
957 	    &ctrlr->hmb_desc_map)) {
958 		nvme_printf(ctrlr, "failed to alloc HMB desc\n");
959 		nvme_ctrlr_hmb_free(ctrlr);
960 		return;
961 	}
962 	if (bus_dmamap_load(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_map,
963 	    ctrlr->hmb_desc_vaddr, size, nvme_single_map,
964 	    &ctrlr->hmb_desc_paddr, BUS_DMA_NOWAIT) != 0) {
965 		bus_dmamem_free(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_vaddr,
966 		    ctrlr->hmb_desc_map);
967 		nvme_printf(ctrlr, "failed to load HMB desc\n");
968 		nvme_ctrlr_hmb_free(ctrlr);
969 		return;
970 	}
971 
972 	for (i = 0; i < ctrlr->hmb_nchunks; i++) {
973 		ctrlr->hmb_desc_vaddr[i].addr =
974 		    htole64(ctrlr->hmb_chunks[i].hmbc_paddr);
975 		ctrlr->hmb_desc_vaddr[i].size = htole32(ctrlr->hmb_chunk / 4096);
976 	}
977 	bus_dmamap_sync(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_map,
978 	    BUS_DMASYNC_PREWRITE);
979 
980 	nvme_printf(ctrlr, "Allocated %lluMB host memory buffer\n",
981 	    (long long unsigned)ctrlr->hmb_nchunks * ctrlr->hmb_chunk
982 	    / 1024 / 1024);
983 }
984 
985 static void
986 nvme_ctrlr_hmb_enable(struct nvme_controller *ctrlr, bool enable, bool memret)
987 {
988 	struct nvme_completion_poll_status	status;
989 	uint32_t cdw11;
990 
991 	cdw11 = 0;
992 	if (enable)
993 		cdw11 |= 1;
994 	if (memret)
995 		cdw11 |= 2;
996 	status.done = 0;
997 	nvme_ctrlr_cmd_set_feature(ctrlr, NVME_FEAT_HOST_MEMORY_BUFFER, cdw11,
998 	    ctrlr->hmb_nchunks * ctrlr->hmb_chunk / 4096, ctrlr->hmb_desc_paddr,
999 	    ctrlr->hmb_desc_paddr >> 32, ctrlr->hmb_nchunks, NULL, 0,
1000 	    nvme_completion_poll_cb, &status);
1001 	nvme_completion_poll(&status);
1002 	if (nvme_completion_is_error(&status.cpl))
1003 		nvme_printf(ctrlr, "nvme_ctrlr_hmb_enable failed!\n");
1004 }
1005 
1006 static void
1007 nvme_ctrlr_start(void *ctrlr_arg, bool resetting)
1008 {
1009 	struct nvme_controller *ctrlr = ctrlr_arg;
1010 	uint32_t old_num_io_queues;
1011 	int i;
1012 
1013 	/*
1014 	 * Only reset adminq here when we are restarting the
1015 	 *  controller after a reset.  During initialization,
1016 	 *  we have already submitted admin commands to get
1017 	 *  the number of I/O queues supported, so cannot reset
1018 	 *  the adminq again here.
1019 	 */
1020 	if (resetting)
1021 		nvme_qpair_reset(&ctrlr->adminq);
1022 
1023 	for (i = 0; i < ctrlr->num_io_queues; i++)
1024 		nvme_qpair_reset(&ctrlr->ioq[i]);
1025 
1026 	nvme_admin_qpair_enable(&ctrlr->adminq);
1027 
1028 	if (nvme_ctrlr_identify(ctrlr) != 0) {
1029 		nvme_ctrlr_fail(ctrlr);
1030 		return;
1031 	}
1032 
1033 	/*
1034 	 * The number of qpairs are determined during controller initialization,
1035 	 *  including using NVMe SET_FEATURES/NUMBER_OF_QUEUES to determine the
1036 	 *  HW limit.  We call SET_FEATURES again here so that it gets called
1037 	 *  after any reset for controllers that depend on the driver to
1038 	 *  explicit specify how many queues it will use.  This value should
1039 	 *  never change between resets, so panic if somehow that does happen.
1040 	 */
1041 	if (resetting) {
1042 		old_num_io_queues = ctrlr->num_io_queues;
1043 		if (nvme_ctrlr_set_num_qpairs(ctrlr) != 0) {
1044 			nvme_ctrlr_fail(ctrlr);
1045 			return;
1046 		}
1047 
1048 		if (old_num_io_queues != ctrlr->num_io_queues) {
1049 			panic("num_io_queues changed from %u to %u",
1050 			      old_num_io_queues, ctrlr->num_io_queues);
1051 		}
1052 	}
1053 
1054 	if (ctrlr->cdata.hmpre > 0 && ctrlr->hmb_nchunks == 0) {
1055 		nvme_ctrlr_hmb_alloc(ctrlr);
1056 		if (ctrlr->hmb_nchunks > 0)
1057 			nvme_ctrlr_hmb_enable(ctrlr, true, false);
1058 	} else if (ctrlr->hmb_nchunks > 0)
1059 		nvme_ctrlr_hmb_enable(ctrlr, true, true);
1060 
1061 	if (nvme_ctrlr_create_qpairs(ctrlr) != 0) {
1062 		nvme_ctrlr_fail(ctrlr);
1063 		return;
1064 	}
1065 
1066 	if (nvme_ctrlr_construct_namespaces(ctrlr) != 0) {
1067 		nvme_ctrlr_fail(ctrlr);
1068 		return;
1069 	}
1070 
1071 	nvme_ctrlr_configure_aer(ctrlr);
1072 	nvme_ctrlr_configure_int_coalescing(ctrlr);
1073 
1074 	for (i = 0; i < ctrlr->num_io_queues; i++)
1075 		nvme_io_qpair_enable(&ctrlr->ioq[i]);
1076 }
1077 
1078 void
1079 nvme_ctrlr_start_config_hook(void *arg)
1080 {
1081 	struct nvme_controller *ctrlr = arg;
1082 	int status;
1083 
1084 	/*
1085 	 * Reset controller twice to ensure we do a transition from cc.en==1 to
1086 	 * cc.en==0.  This is because we don't really know what status the
1087 	 * controller was left in when boot handed off to OS.  Linux doesn't do
1088 	 * this, however. If we adopt that policy, see also nvme_ctrlr_resume().
1089 	 */
1090 	status = nvme_ctrlr_hw_reset(ctrlr);
1091 	if (status != 0) {
1092 		nvme_ctrlr_fail(ctrlr);
1093 		return;
1094 	}
1095 
1096 	status = nvme_ctrlr_hw_reset(ctrlr);
1097 	if (status != 0) {
1098 		nvme_ctrlr_fail(ctrlr);
1099 		return;
1100 	}
1101 
1102 	nvme_qpair_reset(&ctrlr->adminq);
1103 	nvme_admin_qpair_enable(&ctrlr->adminq);
1104 
1105 	if (nvme_ctrlr_set_num_qpairs(ctrlr) == 0 &&
1106 	    nvme_ctrlr_construct_io_qpairs(ctrlr) == 0)
1107 		nvme_ctrlr_start(ctrlr, false);
1108 	else
1109 		nvme_ctrlr_fail(ctrlr);
1110 
1111 	nvme_sysctl_initialize_ctrlr(ctrlr);
1112 	config_intrhook_disestablish(&ctrlr->config_hook);
1113 
1114 	ctrlr->is_initialized = 1;
1115 	nvme_notify_new_controller(ctrlr);
1116 }
1117 
1118 static void
1119 nvme_ctrlr_reset_task(void *arg, int pending)
1120 {
1121 	struct nvme_controller	*ctrlr = arg;
1122 	int			status;
1123 
1124 	nvme_printf(ctrlr, "resetting controller\n");
1125 	status = nvme_ctrlr_hw_reset(ctrlr);
1126 	/*
1127 	 * Use pause instead of DELAY, so that we yield to any nvme interrupt
1128 	 *  handlers on this CPU that were blocked on a qpair lock. We want
1129 	 *  all nvme interrupts completed before proceeding with restarting the
1130 	 *  controller.
1131 	 *
1132 	 * XXX - any way to guarantee the interrupt handlers have quiesced?
1133 	 */
1134 	pause("nvmereset", hz / 10);
1135 	if (status == 0)
1136 		nvme_ctrlr_start(ctrlr, true);
1137 	else
1138 		nvme_ctrlr_fail(ctrlr);
1139 
1140 	atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
1141 }
1142 
1143 /*
1144  * Poll all the queues enabled on the device for completion.
1145  */
1146 void
1147 nvme_ctrlr_poll(struct nvme_controller *ctrlr)
1148 {
1149 	int i;
1150 
1151 	nvme_qpair_process_completions(&ctrlr->adminq);
1152 
1153 	for (i = 0; i < ctrlr->num_io_queues; i++)
1154 		if (ctrlr->ioq && ctrlr->ioq[i].cpl)
1155 			nvme_qpair_process_completions(&ctrlr->ioq[i]);
1156 }
1157 
1158 /*
1159  * Poll the single-vector interrupt case: num_io_queues will be 1 and
1160  * there's only a single vector. While we're polling, we mask further
1161  * interrupts in the controller.
1162  */
1163 void
1164 nvme_ctrlr_intx_handler(void *arg)
1165 {
1166 	struct nvme_controller *ctrlr = arg;
1167 
1168 	nvme_mmio_write_4(ctrlr, intms, 1);
1169 	nvme_ctrlr_poll(ctrlr);
1170 	nvme_mmio_write_4(ctrlr, intmc, 1);
1171 }
1172 
1173 static void
1174 nvme_pt_done(void *arg, const struct nvme_completion *cpl)
1175 {
1176 	struct nvme_pt_command *pt = arg;
1177 	struct mtx *mtx = pt->driver_lock;
1178 	uint16_t status;
1179 
1180 	bzero(&pt->cpl, sizeof(pt->cpl));
1181 	pt->cpl.cdw0 = cpl->cdw0;
1182 
1183 	status = cpl->status;
1184 	status &= ~NVME_STATUS_P_MASK;
1185 	pt->cpl.status = status;
1186 
1187 	mtx_lock(mtx);
1188 	pt->driver_lock = NULL;
1189 	wakeup(pt);
1190 	mtx_unlock(mtx);
1191 }
1192 
1193 int
1194 nvme_ctrlr_passthrough_cmd(struct nvme_controller *ctrlr,
1195     struct nvme_pt_command *pt, uint32_t nsid, int is_user_buffer,
1196     int is_admin_cmd)
1197 {
1198 	struct nvme_request	*req;
1199 	struct mtx		*mtx;
1200 	struct buf		*buf = NULL;
1201 	int			ret = 0;
1202 	vm_offset_t		addr, end;
1203 
1204 	if (pt->len > 0) {
1205 		/*
1206 		 * vmapbuf calls vm_fault_quick_hold_pages which only maps full
1207 		 * pages. Ensure this request has fewer than MAXPHYS bytes when
1208 		 * extended to full pages.
1209 		 */
1210 		addr = (vm_offset_t)pt->buf;
1211 		end = round_page(addr + pt->len);
1212 		addr = trunc_page(addr);
1213 		if (end - addr > MAXPHYS)
1214 			return EIO;
1215 
1216 		if (pt->len > ctrlr->max_xfer_size) {
1217 			nvme_printf(ctrlr, "pt->len (%d) "
1218 			    "exceeds max_xfer_size (%d)\n", pt->len,
1219 			    ctrlr->max_xfer_size);
1220 			return EIO;
1221 		}
1222 		if (is_user_buffer) {
1223 			/*
1224 			 * Ensure the user buffer is wired for the duration of
1225 			 *  this pass-through command.
1226 			 */
1227 			PHOLD(curproc);
1228 			buf = uma_zalloc(pbuf_zone, M_WAITOK);
1229 			buf->b_data = pt->buf;
1230 			buf->b_bufsize = pt->len;
1231 			buf->b_iocmd = pt->is_read ? BIO_READ : BIO_WRITE;
1232 			if (vmapbuf(buf, 1) < 0) {
1233 				ret = EFAULT;
1234 				goto err;
1235 			}
1236 			req = nvme_allocate_request_vaddr(buf->b_data, pt->len,
1237 			    nvme_pt_done, pt);
1238 		} else
1239 			req = nvme_allocate_request_vaddr(pt->buf, pt->len,
1240 			    nvme_pt_done, pt);
1241 	} else
1242 		req = nvme_allocate_request_null(nvme_pt_done, pt);
1243 
1244 	/* Assume user space already converted to little-endian */
1245 	req->cmd.opc = pt->cmd.opc;
1246 	req->cmd.fuse = pt->cmd.fuse;
1247 	req->cmd.rsvd2 = pt->cmd.rsvd2;
1248 	req->cmd.rsvd3 = pt->cmd.rsvd3;
1249 	req->cmd.cdw10 = pt->cmd.cdw10;
1250 	req->cmd.cdw11 = pt->cmd.cdw11;
1251 	req->cmd.cdw12 = pt->cmd.cdw12;
1252 	req->cmd.cdw13 = pt->cmd.cdw13;
1253 	req->cmd.cdw14 = pt->cmd.cdw14;
1254 	req->cmd.cdw15 = pt->cmd.cdw15;
1255 
1256 	req->cmd.nsid = htole32(nsid);
1257 
1258 	mtx = mtx_pool_find(mtxpool_sleep, pt);
1259 	pt->driver_lock = mtx;
1260 
1261 	if (is_admin_cmd)
1262 		nvme_ctrlr_submit_admin_request(ctrlr, req);
1263 	else
1264 		nvme_ctrlr_submit_io_request(ctrlr, req);
1265 
1266 	mtx_lock(mtx);
1267 	while (pt->driver_lock != NULL)
1268 		mtx_sleep(pt, mtx, PRIBIO, "nvme_pt", 0);
1269 	mtx_unlock(mtx);
1270 
1271 err:
1272 	if (buf != NULL) {
1273 		uma_zfree(pbuf_zone, buf);
1274 		PRELE(curproc);
1275 	}
1276 
1277 	return (ret);
1278 }
1279 
1280 static int
1281 nvme_ctrlr_ioctl(struct cdev *cdev, u_long cmd, caddr_t arg, int flag,
1282     struct thread *td)
1283 {
1284 	struct nvme_controller			*ctrlr;
1285 	struct nvme_pt_command			*pt;
1286 
1287 	ctrlr = cdev->si_drv1;
1288 
1289 	switch (cmd) {
1290 	case NVME_RESET_CONTROLLER:
1291 		nvme_ctrlr_reset(ctrlr);
1292 		break;
1293 	case NVME_PASSTHROUGH_CMD:
1294 		pt = (struct nvme_pt_command *)arg;
1295 		return (nvme_ctrlr_passthrough_cmd(ctrlr, pt, le32toh(pt->cmd.nsid),
1296 		    1 /* is_user_buffer */, 1 /* is_admin_cmd */));
1297 	case NVME_GET_NSID:
1298 	{
1299 		struct nvme_get_nsid *gnsid = (struct nvme_get_nsid *)arg;
1300 		strncpy(gnsid->cdev, device_get_nameunit(ctrlr->dev),
1301 		    sizeof(gnsid->cdev));
1302 		gnsid->nsid = 0;
1303 		break;
1304 	}
1305 	default:
1306 		return (ENOTTY);
1307 	}
1308 
1309 	return (0);
1310 }
1311 
1312 static struct cdevsw nvme_ctrlr_cdevsw = {
1313 	.d_version =	D_VERSION,
1314 	.d_flags =	0,
1315 	.d_ioctl =	nvme_ctrlr_ioctl
1316 };
1317 
1318 int
1319 nvme_ctrlr_construct(struct nvme_controller *ctrlr, device_t dev)
1320 {
1321 	struct make_dev_args	md_args;
1322 	uint32_t	cap_lo;
1323 	uint32_t	cap_hi;
1324 	uint32_t	to;
1325 	uint8_t		mpsmin;
1326 	int		status, timeout_period;
1327 
1328 	ctrlr->dev = dev;
1329 
1330 	mtx_init(&ctrlr->lock, "nvme ctrlr lock", NULL, MTX_DEF);
1331 	if (bus_get_domain(dev, &ctrlr->domain) != 0)
1332 		ctrlr->domain = 0;
1333 
1334 	cap_hi = nvme_mmio_read_4(ctrlr, cap_hi);
1335 	ctrlr->dstrd = NVME_CAP_HI_DSTRD(cap_hi) + 2;
1336 
1337 	mpsmin = NVME_CAP_HI_MPSMIN(cap_hi);
1338 	ctrlr->min_page_size = 1 << (12 + mpsmin);
1339 
1340 	/* Get ready timeout value from controller, in units of 500ms. */
1341 	cap_lo = nvme_mmio_read_4(ctrlr, cap_lo);
1342 	to = NVME_CAP_LO_TO(cap_lo) + 1;
1343 	ctrlr->ready_timeout_in_ms = to * 500;
1344 
1345 	timeout_period = NVME_DEFAULT_TIMEOUT_PERIOD;
1346 	TUNABLE_INT_FETCH("hw.nvme.timeout_period", &timeout_period);
1347 	timeout_period = min(timeout_period, NVME_MAX_TIMEOUT_PERIOD);
1348 	timeout_period = max(timeout_period, NVME_MIN_TIMEOUT_PERIOD);
1349 	ctrlr->timeout_period = timeout_period;
1350 
1351 	nvme_retry_count = NVME_DEFAULT_RETRY_COUNT;
1352 	TUNABLE_INT_FETCH("hw.nvme.retry_count", &nvme_retry_count);
1353 
1354 	ctrlr->enable_aborts = 0;
1355 	TUNABLE_INT_FETCH("hw.nvme.enable_aborts", &ctrlr->enable_aborts);
1356 
1357 	ctrlr->max_xfer_size = NVME_MAX_XFER_SIZE;
1358 	if (nvme_ctrlr_construct_admin_qpair(ctrlr) != 0)
1359 		return (ENXIO);
1360 
1361 	ctrlr->taskqueue = taskqueue_create("nvme_taskq", M_WAITOK,
1362 	    taskqueue_thread_enqueue, &ctrlr->taskqueue);
1363 	taskqueue_start_threads(&ctrlr->taskqueue, 1, PI_DISK, "nvme taskq");
1364 
1365 	ctrlr->is_resetting = 0;
1366 	ctrlr->is_initialized = 0;
1367 	ctrlr->notification_sent = 0;
1368 	TASK_INIT(&ctrlr->reset_task, 0, nvme_ctrlr_reset_task, ctrlr);
1369 	TASK_INIT(&ctrlr->fail_req_task, 0, nvme_ctrlr_fail_req_task, ctrlr);
1370 	STAILQ_INIT(&ctrlr->fail_req);
1371 	ctrlr->is_failed = false;
1372 
1373 	make_dev_args_init(&md_args);
1374 	md_args.mda_devsw = &nvme_ctrlr_cdevsw;
1375 	md_args.mda_uid = UID_ROOT;
1376 	md_args.mda_gid = GID_WHEEL;
1377 	md_args.mda_mode = 0600;
1378 	md_args.mda_unit = device_get_unit(dev);
1379 	md_args.mda_si_drv1 = (void *)ctrlr;
1380 	status = make_dev_s(&md_args, &ctrlr->cdev, "nvme%d",
1381 	    device_get_unit(dev));
1382 	if (status != 0)
1383 		return (ENXIO);
1384 
1385 	return (0);
1386 }
1387 
1388 void
1389 nvme_ctrlr_destruct(struct nvme_controller *ctrlr, device_t dev)
1390 {
1391 	int	gone, i;
1392 
1393 	if (ctrlr->resource == NULL)
1394 		goto nores;
1395 
1396 	/*
1397 	 * Check whether it is a hot unplug or a clean driver detach.
1398 	 * If device is not there any more, skip any shutdown commands.
1399 	 */
1400 	gone = (nvme_mmio_read_4(ctrlr, csts) == 0xffffffff);
1401 	if (gone)
1402 		nvme_ctrlr_fail(ctrlr);
1403 	else
1404 		nvme_notify_fail_consumers(ctrlr);
1405 
1406 	for (i = 0; i < NVME_MAX_NAMESPACES; i++)
1407 		nvme_ns_destruct(&ctrlr->ns[i]);
1408 
1409 	if (ctrlr->cdev)
1410 		destroy_dev(ctrlr->cdev);
1411 
1412 	if (ctrlr->is_initialized) {
1413 		if (!gone) {
1414 			if (ctrlr->hmb_nchunks > 0)
1415 				nvme_ctrlr_hmb_enable(ctrlr, false, false);
1416 			nvme_ctrlr_delete_qpairs(ctrlr);
1417 		}
1418 		for (i = 0; i < ctrlr->num_io_queues; i++)
1419 			nvme_io_qpair_destroy(&ctrlr->ioq[i]);
1420 		free(ctrlr->ioq, M_NVME);
1421 		nvme_ctrlr_hmb_free(ctrlr);
1422 		nvme_admin_qpair_destroy(&ctrlr->adminq);
1423 	}
1424 
1425 	/*
1426 	 *  Notify the controller of a shutdown, even though this is due to
1427 	 *   a driver unload, not a system shutdown (this path is not invoked
1428 	 *   during shutdown).  This ensures the controller receives a
1429 	 *   shutdown notification in case the system is shutdown before
1430 	 *   reloading the driver.
1431 	 */
1432 	if (!gone)
1433 		nvme_ctrlr_shutdown(ctrlr);
1434 
1435 	if (!gone)
1436 		nvme_ctrlr_disable(ctrlr);
1437 
1438 	if (ctrlr->taskqueue)
1439 		taskqueue_free(ctrlr->taskqueue);
1440 
1441 	if (ctrlr->tag)
1442 		bus_teardown_intr(ctrlr->dev, ctrlr->res, ctrlr->tag);
1443 
1444 	if (ctrlr->res)
1445 		bus_release_resource(ctrlr->dev, SYS_RES_IRQ,
1446 		    rman_get_rid(ctrlr->res), ctrlr->res);
1447 
1448 	if (ctrlr->bar4_resource != NULL) {
1449 		bus_release_resource(dev, SYS_RES_MEMORY,
1450 		    ctrlr->bar4_resource_id, ctrlr->bar4_resource);
1451 	}
1452 
1453 	bus_release_resource(dev, SYS_RES_MEMORY,
1454 	    ctrlr->resource_id, ctrlr->resource);
1455 
1456 nores:
1457 	mtx_destroy(&ctrlr->lock);
1458 }
1459 
1460 void
1461 nvme_ctrlr_shutdown(struct nvme_controller *ctrlr)
1462 {
1463 	uint32_t	cc;
1464 	uint32_t	csts;
1465 	int		ticks = 0;
1466 
1467 	cc = nvme_mmio_read_4(ctrlr, cc);
1468 	cc &= ~(NVME_CC_REG_SHN_MASK << NVME_CC_REG_SHN_SHIFT);
1469 	cc |= NVME_SHN_NORMAL << NVME_CC_REG_SHN_SHIFT;
1470 	nvme_mmio_write_4(ctrlr, cc, cc);
1471 
1472 	while (1) {
1473 		csts = nvme_mmio_read_4(ctrlr, csts);
1474 		if (csts == 0xffffffff)		/* Hot unplug. */
1475 			break;
1476 		if (NVME_CSTS_GET_SHST(csts) == NVME_SHST_COMPLETE)
1477 			break;
1478 		if (ticks++ > 5*hz) {
1479 			nvme_printf(ctrlr, "did not complete shutdown within"
1480 			    " 5 seconds of notification\n");
1481 			break;
1482 		}
1483 		pause("nvme shn", 1);
1484 	}
1485 }
1486 
1487 void
1488 nvme_ctrlr_submit_admin_request(struct nvme_controller *ctrlr,
1489     struct nvme_request *req)
1490 {
1491 
1492 	nvme_qpair_submit_request(&ctrlr->adminq, req);
1493 }
1494 
1495 void
1496 nvme_ctrlr_submit_io_request(struct nvme_controller *ctrlr,
1497     struct nvme_request *req)
1498 {
1499 	struct nvme_qpair       *qpair;
1500 
1501 	qpair = &ctrlr->ioq[QP(ctrlr, curcpu)];
1502 	nvme_qpair_submit_request(qpair, req);
1503 }
1504 
1505 device_t
1506 nvme_ctrlr_get_device(struct nvme_controller *ctrlr)
1507 {
1508 
1509 	return (ctrlr->dev);
1510 }
1511 
1512 const struct nvme_controller_data *
1513 nvme_ctrlr_get_data(struct nvme_controller *ctrlr)
1514 {
1515 
1516 	return (&ctrlr->cdata);
1517 }
1518 
1519 int
1520 nvme_ctrlr_suspend(struct nvme_controller *ctrlr)
1521 {
1522 	int to = hz;
1523 
1524 	/*
1525 	 * Can't touch failed controllers, so it's already suspended.
1526 	 */
1527 	if (ctrlr->is_failed)
1528 		return (0);
1529 
1530 	/*
1531 	 * We don't want the reset taskqueue running, since it does similar
1532 	 * things, so prevent it from running after we start. Wait for any reset
1533 	 * that may have been started to complete. The reset process we follow
1534 	 * will ensure that any new I/O will queue and be given to the hardware
1535 	 * after we resume (though there should be none).
1536 	 */
1537 	while (atomic_cmpset_32(&ctrlr->is_resetting, 0, 1) == 0 && to-- > 0)
1538 		pause("nvmesusp", 1);
1539 	if (to <= 0) {
1540 		nvme_printf(ctrlr,
1541 		    "Competing reset task didn't finish. Try again later.\n");
1542 		return (EWOULDBLOCK);
1543 	}
1544 
1545 	if (ctrlr->hmb_nchunks > 0)
1546 		nvme_ctrlr_hmb_enable(ctrlr, false, false);
1547 
1548 	/*
1549 	 * Per Section 7.6.2 of NVMe spec 1.4, to properly suspend, we need to
1550 	 * delete the hardware I/O queues, and then shutdown. This properly
1551 	 * flushes any metadata the drive may have stored so it can survive
1552 	 * having its power removed and prevents the unsafe shutdown count from
1553 	 * incriminating. Once we delete the qpairs, we have to disable them
1554 	 * before shutting down. The delay is out of paranoia in
1555 	 * nvme_ctrlr_hw_reset, and is repeated here (though we should have no
1556 	 * pending I/O that the delay copes with).
1557 	 */
1558 	nvme_ctrlr_delete_qpairs(ctrlr);
1559 	nvme_ctrlr_disable_qpairs(ctrlr);
1560 	DELAY(100*1000);
1561 	nvme_ctrlr_shutdown(ctrlr);
1562 
1563 	return (0);
1564 }
1565 
1566 int
1567 nvme_ctrlr_resume(struct nvme_controller *ctrlr)
1568 {
1569 
1570 	/*
1571 	 * Can't touch failed controllers, so nothing to do to resume.
1572 	 */
1573 	if (ctrlr->is_failed)
1574 		return (0);
1575 
1576 	/*
1577 	 * Have to reset the hardware twice, just like we do on attach. See
1578 	 * nmve_attach() for why.
1579 	 */
1580 	if (nvme_ctrlr_hw_reset(ctrlr) != 0)
1581 		goto fail;
1582 	if (nvme_ctrlr_hw_reset(ctrlr) != 0)
1583 		goto fail;
1584 
1585 	/*
1586 	 * Now that we're reset the hardware, we can restart the controller. Any
1587 	 * I/O that was pending is requeued. Any admin commands are aborted with
1588 	 * an error. Once we've restarted, take the controller out of reset.
1589 	 */
1590 	nvme_ctrlr_start(ctrlr, true);
1591 	atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
1592 
1593 	return (0);
1594 fail:
1595 	/*
1596 	 * Since we can't bring the controller out of reset, announce and fail
1597 	 * the controller. However, we have to return success for the resume
1598 	 * itself, due to questionable APIs.
1599 	 */
1600 	nvme_printf(ctrlr, "Failed to reset on resume, failing.\n");
1601 	nvme_ctrlr_fail(ctrlr);
1602 	atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
1603 	return (0);
1604 }
1605