xref: /freebsd/sys/dev/nvme/nvme_ctrlr.c (revision 7b32f4f0a7fe9b1b2f5a3905ca15f656713255ad)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (C) 2012-2016 Intel Corporation
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  */
28 
29 #include "opt_nvme.h"
30 
31 #include <sys/param.h>
32 #include <sys/systm.h>
33 #include <sys/buf.h>
34 #include <sys/bus.h>
35 #include <sys/conf.h>
36 #include <sys/disk.h>
37 #include <sys/ioccom.h>
38 #include <sys/proc.h>
39 #include <sys/smp.h>
40 #include <sys/uio.h>
41 #include <sys/sbuf.h>
42 #include <sys/endian.h>
43 #include <sys/stdarg.h>
44 #include <vm/vm.h>
45 #include <vm/vm_page.h>
46 #include <vm/vm_extern.h>
47 #include <vm/vm_map.h>
48 
49 #include "nvme_private.h"
50 #include "nvme_linux.h"
51 
52 #define B4_CHK_RDY_DELAY_MS	2300		/* work around controller bug */
53 
54 static void nvme_ctrlr_construct_and_submit_aer(struct nvme_controller *ctrlr,
55     struct nvme_async_event_request *aer);
56 
57 static void
nvme_ctrlr_barrier(struct nvme_controller * ctrlr,int flags)58 nvme_ctrlr_barrier(struct nvme_controller *ctrlr, int flags)
59 {
60 	bus_barrier(ctrlr->resource, 0, rman_get_size(ctrlr->resource), flags);
61 }
62 
63 static void
nvme_ctrlr_devctl_va(struct nvme_controller * ctrlr,const char * type,const char * msg,va_list ap)64 nvme_ctrlr_devctl_va(struct nvme_controller *ctrlr, const char *type,
65     const char *msg, va_list ap)
66 {
67 	struct sbuf sb;
68 	int error;
69 
70 	if (sbuf_new(&sb, NULL, 0, SBUF_AUTOEXTEND | SBUF_NOWAIT) == NULL)
71 		return;
72 	sbuf_printf(&sb, "name=\"%s\" ", device_get_nameunit(ctrlr->dev));
73 	sbuf_vprintf(&sb, msg, ap);
74 	error = sbuf_finish(&sb);
75 	if (error == 0)
76 		devctl_notify("nvme", "controller", type, sbuf_data(&sb));
77 	sbuf_delete(&sb);
78 }
79 
80 static void
nvme_ctrlr_devctl(struct nvme_controller * ctrlr,const char * type,const char * msg,...)81 nvme_ctrlr_devctl(struct nvme_controller *ctrlr, const char *type, const char *msg, ...)
82 {
83 	va_list ap;
84 
85 	va_start(ap, msg);
86 	nvme_ctrlr_devctl_va(ctrlr, type, msg, ap);
87 	va_end(ap);
88 }
89 
90 static void
nvme_ctrlr_devctl_log(struct nvme_controller * ctrlr,const char * type,const char * msg,...)91 nvme_ctrlr_devctl_log(struct nvme_controller *ctrlr, const char *type, const char *msg, ...)
92 {
93 	struct sbuf sb;
94 	va_list ap;
95 	int error;
96 
97 	if (sbuf_new(&sb, NULL, 0, SBUF_AUTOEXTEND | SBUF_NOWAIT) == NULL)
98 		return;
99 	sbuf_printf(&sb, "%s: ", device_get_nameunit(ctrlr->dev));
100 	va_start(ap, msg);
101 	sbuf_vprintf(&sb, msg, ap);
102 	va_end(ap);
103 	error = sbuf_finish(&sb);
104 	if (error == 0)
105 		printf("%s\n", sbuf_data(&sb));
106 	sbuf_delete(&sb);
107 	va_start(ap, msg);
108 	nvme_ctrlr_devctl_va(ctrlr, type, msg, ap);
109 	va_end(ap);
110 }
111 
112 static int
nvme_ctrlr_construct_admin_qpair(struct nvme_controller * ctrlr)113 nvme_ctrlr_construct_admin_qpair(struct nvme_controller *ctrlr)
114 {
115 	struct nvme_qpair	*qpair;
116 	uint32_t		num_entries;
117 	int			error;
118 
119 	qpair = &ctrlr->adminq;
120 	qpair->id = 0;
121 	qpair->cpu = CPU_FFS(&cpuset_domain[ctrlr->domain]) - 1;
122 	qpair->domain = ctrlr->domain;
123 
124 	num_entries = NVME_ADMIN_ENTRIES;
125 	TUNABLE_INT_FETCH("hw.nvme.admin_entries", &num_entries);
126 	/*
127 	 * If admin_entries was overridden to an invalid value, revert it
128 	 *  back to our default value.
129 	 */
130 	if (num_entries < NVME_MIN_ADMIN_ENTRIES ||
131 	    num_entries > NVME_MAX_ADMIN_ENTRIES) {
132 		nvme_printf(ctrlr, "invalid hw.nvme.admin_entries=%d "
133 		    "specified\n", num_entries);
134 		num_entries = NVME_ADMIN_ENTRIES;
135 	}
136 
137 	/*
138 	 * The admin queue's max xfer size is treated differently than the
139 	 *  max I/O xfer size.  16KB is sufficient here - maybe even less?
140 	 */
141 	error = nvme_qpair_construct(qpair, num_entries, NVME_ADMIN_TRACKERS,
142 	     ctrlr);
143 	return (error);
144 }
145 
146 #define QP(ctrlr, c)	((c) * (ctrlr)->num_io_queues / mp_ncpus)
147 
148 static int
nvme_ctrlr_construct_io_qpairs(struct nvme_controller * ctrlr)149 nvme_ctrlr_construct_io_qpairs(struct nvme_controller *ctrlr)
150 {
151 	struct nvme_qpair	*qpair;
152 	uint32_t		cap_lo;
153 	uint16_t		mqes;
154 	int			c, error, i, n;
155 	int			num_entries, num_trackers, max_entries;
156 
157 	/*
158 	 * NVMe spec sets a hard limit of 64K max entries, but devices may
159 	 * specify a smaller limit, so we need to check the MQES field in the
160 	 * capabilities register. We have to cap the number of entries to the
161 	 * current stride allows for in BAR 0/1, otherwise the remainder entries
162 	 * are inaccessible. MQES should reflect this, and this is just a
163 	 * fail-safe.
164 	 */
165 	max_entries =
166 	    (rman_get_size(ctrlr->resource) - nvme_mmio_offsetof(doorbell[0])) /
167 	    (1 << (ctrlr->dstrd + 1));
168 	num_entries = NVME_IO_ENTRIES;
169 	TUNABLE_INT_FETCH("hw.nvme.io_entries", &num_entries);
170 	cap_lo = nvme_mmio_read_4(ctrlr, cap_lo);
171 	mqes = NVME_CAP_LO_MQES(cap_lo);
172 	num_entries = min(num_entries, mqes + 1);
173 	num_entries = min(num_entries, max_entries);
174 
175 	num_trackers = NVME_IO_TRACKERS;
176 	TUNABLE_INT_FETCH("hw.nvme.io_trackers", &num_trackers);
177 
178 	num_trackers = max(num_trackers, NVME_MIN_IO_TRACKERS);
179 	num_trackers = min(num_trackers, NVME_MAX_IO_TRACKERS);
180 	/*
181 	 * No need to have more trackers than entries in the submit queue.  Note
182 	 * also that for a queue size of N, we can only have (N-1) commands
183 	 * outstanding, hence the "-1" here.
184 	 */
185 	num_trackers = min(num_trackers, (num_entries-1));
186 
187 	/*
188 	 * Our best estimate for the maximum number of I/Os that we should
189 	 * normally have in flight at one time. This should be viewed as a hint,
190 	 * not a hard limit and will need to be revisited when the upper layers
191 	 * of the storage system grows multi-queue support.
192 	 */
193 	ctrlr->max_hw_pend_io = num_trackers * ctrlr->num_io_queues * 3 / 4;
194 
195 	ctrlr->ioq = malloc(ctrlr->num_io_queues * sizeof(struct nvme_qpair),
196 	    M_NVME, M_ZERO | M_WAITOK);
197 
198 	for (i = c = n = 0; i < ctrlr->num_io_queues; i++, c += n) {
199 		qpair = &ctrlr->ioq[i];
200 
201 		/*
202 		 * Admin queue has ID=0. IO queues start at ID=1 -
203 		 *  hence the 'i+1' here.
204 		 */
205 		qpair->id = i + 1;
206 		if (ctrlr->num_io_queues > 1) {
207 			/* Find number of CPUs served by this queue. */
208 			for (n = 1; QP(ctrlr, c + n) == i; n++)
209 				;
210 			/* Shuffle multiple NVMe devices between CPUs. */
211 			qpair->cpu = c + (device_get_unit(ctrlr->dev)+n/2) % n;
212 			qpair->domain = pcpu_find(qpair->cpu)->pc_domain;
213 		} else {
214 			qpair->cpu = CPU_FFS(&cpuset_domain[ctrlr->domain]) - 1;
215 			qpair->domain = ctrlr->domain;
216 		}
217 
218 		/*
219 		 * For I/O queues, use the controller-wide max_xfer_size
220 		 *  calculated in nvme_attach().
221 		 */
222 		error = nvme_qpair_construct(qpair, num_entries, num_trackers,
223 		    ctrlr);
224 		if (error)
225 			return (error);
226 
227 		/*
228 		 * Do not bother binding interrupts if we only have one I/O
229 		 *  interrupt thread for this controller.
230 		 */
231 		if (ctrlr->num_io_queues > 1)
232 			bus_bind_intr(ctrlr->dev, qpair->res, qpair->cpu);
233 	}
234 
235 	return (0);
236 }
237 
238 static void
nvme_ctrlr_fail(struct nvme_controller * ctrlr,bool admin_also)239 nvme_ctrlr_fail(struct nvme_controller *ctrlr, bool admin_also)
240 {
241 	int i;
242 
243 	/*
244 	 * No need to disable queues before failing them. Failing is a superet
245 	 * of disabling (though pedantically we'd abort the AERs silently with
246 	 * a different error, though when we fail, that hardly matters).
247 	 */
248 	ctrlr->is_failed = true;
249 	if (admin_also) {
250 		ctrlr->is_failed_admin = true;
251 		nvme_qpair_fail(&ctrlr->adminq);
252 	}
253 	if (ctrlr->ioq != NULL) {
254 		for (i = 0; i < ctrlr->num_io_queues; i++) {
255 			nvme_qpair_fail(&ctrlr->ioq[i]);
256 		}
257 	}
258 	nvme_notify_fail_consumers(ctrlr);
259 }
260 
261 /*
262  * Wait for RDY to change.
263  *
264  * Starts sleeping for 1us and geometrically increases it the longer we wait,
265  * capped at 1ms.
266  */
267 static int
nvme_ctrlr_wait_for_ready(struct nvme_controller * ctrlr,int desired_val)268 nvme_ctrlr_wait_for_ready(struct nvme_controller *ctrlr, int desired_val)
269 {
270 	int timeout = ticks + MSEC_2_TICKS(ctrlr->ready_timeout_in_ms);
271 	sbintime_t delta_t = SBT_1US;
272 	uint32_t csts;
273 
274 	while (1) {
275 		csts = nvme_mmio_read_4(ctrlr, csts);
276 		if (csts == NVME_GONE)		/* Hot unplug. */
277 			return (ENXIO);
278 		if (NVMEV(NVME_CSTS_REG_RDY, csts) == desired_val)
279 			break;
280 		if (timeout - ticks < 0) {
281 			nvme_printf(ctrlr, "controller ready did not become %d "
282 			    "within %d ms\n", desired_val, ctrlr->ready_timeout_in_ms);
283 			return (ENXIO);
284 		}
285 
286 		pause_sbt("nvmerdy", delta_t, 0, C_PREL(1));
287 		delta_t = min(SBT_1MS, delta_t * 3 / 2);
288 	}
289 
290 	return (0);
291 }
292 
293 static int
nvme_ctrlr_disable(struct nvme_controller * ctrlr)294 nvme_ctrlr_disable(struct nvme_controller *ctrlr)
295 {
296 	uint32_t cc;
297 	uint32_t csts;
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 = NVMEV(NVME_CC_REG_EN, cc);
305 	rdy = NVMEV(NVME_CSTS_REG_RDY, csts);
306 
307 	/*
308 	 * Per 3.1.5 in NVME 1.3 spec, transitioning CC.EN from 0 to 1
309 	 * when CSTS.RDY is 1 or transitioning CC.EN from 1 to 0 when
310 	 * CSTS.RDY is 0 "has undefined results" So make sure that CSTS.RDY
311 	 * isn't the desired value. Short circuit if we're already disabled.
312 	 */
313 	if (en == 0) {
314 		/* Wait for RDY == 0 or timeout & fail */
315 		if (rdy == 0)
316 			return (0);
317 		return (nvme_ctrlr_wait_for_ready(ctrlr, 0));
318 	}
319 	if (rdy == 0) {
320 		/* EN == 1, wait for  RDY == 1 or timeout & fail */
321 		err = nvme_ctrlr_wait_for_ready(ctrlr, 1);
322 		if (err != 0)
323 			return (err);
324 	}
325 
326 	cc &= ~NVMEM(NVME_CC_REG_EN);
327 	nvme_mmio_write_4(ctrlr, cc, cc);
328 
329 	/*
330 	 * A few drives have firmware bugs that freeze the drive if we access
331 	 * the mmio too soon after we disable.
332 	 */
333 	if (ctrlr->quirks & QUIRK_DELAY_B4_CHK_RDY)
334 		pause("nvmeR", MSEC_2_TICKS(B4_CHK_RDY_DELAY_MS));
335 	return (nvme_ctrlr_wait_for_ready(ctrlr, 0));
336 }
337 
338 static int
nvme_ctrlr_enable(struct nvme_controller * ctrlr)339 nvme_ctrlr_enable(struct nvme_controller *ctrlr)
340 {
341 	uint32_t	cc;
342 	uint32_t	csts;
343 	uint32_t	aqa;
344 	uint32_t	qsize;
345 	uint8_t		en, rdy;
346 	int		err;
347 
348 	cc = nvme_mmio_read_4(ctrlr, cc);
349 	csts = nvme_mmio_read_4(ctrlr, csts);
350 
351 	en = NVMEV(NVME_CC_REG_EN, cc);
352 	rdy = NVMEV(NVME_CSTS_REG_RDY, csts);
353 
354 	/*
355 	 * See note in nvme_ctrlr_disable. Short circuit if we're already enabled.
356 	 */
357 	if (en == 1) {
358 		if (rdy == 1)
359 			return (0);
360 		return (nvme_ctrlr_wait_for_ready(ctrlr, 1));
361 	}
362 
363 	/* EN == 0 already wait for RDY == 0 or timeout & fail */
364 	err = nvme_ctrlr_wait_for_ready(ctrlr, 0);
365 	if (err != 0)
366 		return (err);
367 
368 	nvme_mmio_write_8(ctrlr, asq, ctrlr->adminq.cmd_bus_addr);
369 	nvme_mmio_write_8(ctrlr, acq, ctrlr->adminq.cpl_bus_addr);
370 
371 	/* acqs and asqs are 0-based. */
372 	qsize = ctrlr->adminq.num_entries - 1;
373 
374 	aqa = 0;
375 	aqa |= NVMEF(NVME_AQA_REG_ACQS, qsize);
376 	aqa |= NVMEF(NVME_AQA_REG_ASQS, qsize);
377 	nvme_mmio_write_4(ctrlr, aqa, aqa);
378 
379 	/* Initialization values for CC */
380 	cc = 0;
381 	cc |= NVMEF(NVME_CC_REG_EN, 1);
382 	cc |= NVMEF(NVME_CC_REG_CSS, 0);
383 	cc |= NVMEF(NVME_CC_REG_AMS, 0);
384 	cc |= NVMEF(NVME_CC_REG_SHN, 0);
385 	cc |= NVMEF(NVME_CC_REG_IOSQES, 6); /* SQ entry size == 64 == 2^6 */
386 	cc |= NVMEF(NVME_CC_REG_IOCQES, 4); /* CQ entry size == 16 == 2^4 */
387 
388 	/*
389 	 * Use the Memory Page Size selected during device initialization.  Note
390 	 * that value stored in mps is suitable to use here without adjusting by
391 	 * NVME_MPS_SHIFT.
392 	 */
393 	cc |= NVMEF(NVME_CC_REG_MPS, ctrlr->mps);
394 
395 	nvme_ctrlr_barrier(ctrlr, BUS_SPACE_BARRIER_WRITE);
396 	nvme_mmio_write_4(ctrlr, cc, cc);
397 
398 	return (nvme_ctrlr_wait_for_ready(ctrlr, 1));
399 }
400 
401 static void
nvme_ctrlr_disable_qpairs(struct nvme_controller * ctrlr)402 nvme_ctrlr_disable_qpairs(struct nvme_controller *ctrlr)
403 {
404 	int i;
405 
406 	nvme_admin_qpair_disable(&ctrlr->adminq);
407 	/*
408 	 * I/O queues are not allocated before the initial HW
409 	 *  reset, so do not try to disable them.  Use is_initialized
410 	 *  to determine if this is the initial HW reset.
411 	 */
412 	if (ctrlr->is_initialized) {
413 		for (i = 0; i < ctrlr->num_io_queues; i++)
414 			nvme_io_qpair_disable(&ctrlr->ioq[i]);
415 	}
416 }
417 
418 static int
nvme_ctrlr_hw_reset(struct nvme_controller * ctrlr)419 nvme_ctrlr_hw_reset(struct nvme_controller *ctrlr)
420 {
421 	int err;
422 
423 	TSENTER();
424 
425 	ctrlr->is_failed_admin = true;
426 	nvme_ctrlr_disable_qpairs(ctrlr);
427 
428 	err = nvme_ctrlr_disable(ctrlr);
429 	if (err != 0)
430 		goto out;
431 
432 	err = nvme_ctrlr_enable(ctrlr);
433 out:
434 	if (err == 0)
435 		ctrlr->is_failed_admin = false;
436 
437 	TSEXIT();
438 	return (err);
439 }
440 
441 void
nvme_ctrlr_reset(struct nvme_controller * ctrlr)442 nvme_ctrlr_reset(struct nvme_controller *ctrlr)
443 {
444 	int cmpset;
445 
446 	cmpset = atomic_cmpset_32(&ctrlr->is_resetting, 0, 1);
447 
448 	if (cmpset == 0)
449 		/*
450 		 * Controller is already resetting.  Return immediately since
451 		 * there is no need to kick off another reset.
452 		 */
453 		return;
454 
455 	if (!ctrlr->is_dying)
456 		taskqueue_enqueue(ctrlr->taskqueue, &ctrlr->reset_task);
457 }
458 
459 static int
nvme_ctrlr_identify(struct nvme_controller * ctrlr)460 nvme_ctrlr_identify(struct nvme_controller *ctrlr)
461 {
462 	struct nvme_completion_poll_status	status;
463 
464 	status.done = 0;
465 	nvme_ctrlr_cmd_identify_controller(ctrlr, &ctrlr->cdata,
466 	    nvme_completion_poll_cb, &status);
467 	nvme_completion_poll(&status);
468 	if (nvme_completion_is_error(&status.cpl)) {
469 		nvme_printf(ctrlr, "nvme_identify_controller failed!\n");
470 		return (ENXIO);
471 	}
472 
473 	/* Convert data to host endian */
474 	nvme_controller_data_swapbytes(&ctrlr->cdata);
475 
476 	/*
477 	 * Use MDTS to ensure our default max_xfer_size doesn't exceed what the
478 	 *  controller supports.
479 	 */
480 	if (ctrlr->cdata.mdts > 0)
481 		ctrlr->max_xfer_size = min(ctrlr->max_xfer_size,
482 		    1 << (ctrlr->cdata.mdts + NVME_MPS_SHIFT +
483 			NVME_CAP_HI_MPSMIN(ctrlr->cap_hi)));
484 
485 	return (0);
486 }
487 
488 static int
nvme_ctrlr_set_num_qpairs(struct nvme_controller * ctrlr)489 nvme_ctrlr_set_num_qpairs(struct nvme_controller *ctrlr)
490 {
491 	struct nvme_completion_poll_status	status;
492 	int					cq_allocated, sq_allocated;
493 
494 	status.done = 0;
495 	nvme_ctrlr_cmd_set_num_queues(ctrlr, ctrlr->num_io_queues,
496 	    nvme_completion_poll_cb, &status);
497 	nvme_completion_poll(&status);
498 	if (nvme_completion_is_error(&status.cpl)) {
499 		nvme_printf(ctrlr, "nvme_ctrlr_set_num_qpairs failed!\n");
500 		return (ENXIO);
501 	}
502 
503 	/*
504 	 * Data in cdw0 is 0-based.
505 	 * Lower 16-bits indicate number of submission queues allocated.
506 	 * Upper 16-bits indicate number of completion queues allocated.
507 	 */
508 	sq_allocated = (status.cpl.cdw0 & 0xFFFF) + 1;
509 	cq_allocated = (status.cpl.cdw0 >> 16) + 1;
510 
511 	/*
512 	 * Controller may allocate more queues than we requested,
513 	 *  so use the minimum of the number requested and what was
514 	 *  actually allocated.
515 	 */
516 	ctrlr->num_io_queues = min(ctrlr->num_io_queues, sq_allocated);
517 	ctrlr->num_io_queues = min(ctrlr->num_io_queues, cq_allocated);
518 	if (ctrlr->num_io_queues > vm_ndomains)
519 		ctrlr->num_io_queues -= ctrlr->num_io_queues % vm_ndomains;
520 
521 	return (0);
522 }
523 
524 static int
nvme_ctrlr_create_qpairs(struct nvme_controller * ctrlr)525 nvme_ctrlr_create_qpairs(struct nvme_controller *ctrlr)
526 {
527 	struct nvme_completion_poll_status	status;
528 	struct nvme_qpair			*qpair;
529 	int					i;
530 
531 	for (i = 0; i < ctrlr->num_io_queues; i++) {
532 		qpair = &ctrlr->ioq[i];
533 
534 		status.done = 0;
535 		nvme_ctrlr_cmd_create_io_cq(ctrlr, qpair,
536 		    nvme_completion_poll_cb, &status);
537 		nvme_completion_poll(&status);
538 		if (nvme_completion_is_error(&status.cpl)) {
539 			nvme_printf(ctrlr, "nvme_create_io_cq failed!\n");
540 			return (ENXIO);
541 		}
542 
543 		status.done = 0;
544 		nvme_ctrlr_cmd_create_io_sq(ctrlr, qpair,
545 		    nvme_completion_poll_cb, &status);
546 		nvme_completion_poll(&status);
547 		if (nvme_completion_is_error(&status.cpl)) {
548 			nvme_printf(ctrlr, "nvme_create_io_sq failed!\n");
549 			return (ENXIO);
550 		}
551 	}
552 
553 	return (0);
554 }
555 
556 static int
nvme_ctrlr_delete_qpairs(struct nvme_controller * ctrlr)557 nvme_ctrlr_delete_qpairs(struct nvme_controller *ctrlr)
558 {
559 	struct nvme_completion_poll_status	status;
560 	struct nvme_qpair			*qpair;
561 
562 	for (int i = 0; i < ctrlr->num_io_queues; i++) {
563 		qpair = &ctrlr->ioq[i];
564 
565 		status.done = 0;
566 		nvme_ctrlr_cmd_delete_io_sq(ctrlr, qpair,
567 		    nvme_completion_poll_cb, &status);
568 		nvme_completion_poll(&status);
569 		if (nvme_completion_is_error(&status.cpl)) {
570 			nvme_printf(ctrlr, "nvme_destroy_io_sq failed!\n");
571 			return (ENXIO);
572 		}
573 
574 		status.done = 0;
575 		nvme_ctrlr_cmd_delete_io_cq(ctrlr, qpair,
576 		    nvme_completion_poll_cb, &status);
577 		nvme_completion_poll(&status);
578 		if (nvme_completion_is_error(&status.cpl)) {
579 			nvme_printf(ctrlr, "nvme_destroy_io_cq failed!\n");
580 			return (ENXIO);
581 		}
582 	}
583 
584 	return (0);
585 }
586 
587 static int
nvme_ctrlr_construct_namespaces(struct nvme_controller * ctrlr)588 nvme_ctrlr_construct_namespaces(struct nvme_controller *ctrlr)
589 {
590 	struct nvme_namespace	*ns;
591 	uint32_t 		i;
592 
593 	for (i = 0; i < min(ctrlr->cdata.nn, NVME_MAX_NAMESPACES); i++) {
594 		ns = &ctrlr->ns[i];
595 		nvme_ns_construct(ns, i+1, ctrlr);
596 	}
597 
598 	return (0);
599 }
600 
601 static bool
is_log_page_id_valid(uint8_t page_id)602 is_log_page_id_valid(uint8_t page_id)
603 {
604 	switch (page_id) {
605 	case NVME_LOG_ERROR:
606 	case NVME_LOG_HEALTH_INFORMATION:
607 	case NVME_LOG_FIRMWARE_SLOT:
608 	case NVME_LOG_CHANGED_NAMESPACE:
609 	case NVME_LOG_COMMAND_EFFECT:
610 	case NVME_LOG_RES_NOTIFICATION:
611 	case NVME_LOG_SANITIZE_STATUS:
612 		return (true);
613 	}
614 
615 	return (false);
616 }
617 
618 static uint32_t
nvme_ctrlr_get_log_page_size(struct nvme_controller * ctrlr,uint8_t page_id)619 nvme_ctrlr_get_log_page_size(struct nvme_controller *ctrlr, uint8_t page_id)
620 {
621 	uint32_t	log_page_size;
622 
623 	switch (page_id) {
624 	case NVME_LOG_ERROR:
625 		log_page_size = min(
626 		    sizeof(struct nvme_error_information_entry) *
627 		    (ctrlr->cdata.elpe + 1), NVME_MAX_AER_LOG_SIZE);
628 		break;
629 	case NVME_LOG_HEALTH_INFORMATION:
630 		log_page_size = sizeof(struct nvme_health_information_page);
631 		break;
632 	case NVME_LOG_FIRMWARE_SLOT:
633 		log_page_size = sizeof(struct nvme_firmware_page);
634 		break;
635 	case NVME_LOG_CHANGED_NAMESPACE:
636 		log_page_size = sizeof(struct nvme_ns_list);
637 		break;
638 	case NVME_LOG_COMMAND_EFFECT:
639 		log_page_size = sizeof(struct nvme_command_effects_page);
640 		break;
641 	case NVME_LOG_RES_NOTIFICATION:
642 		log_page_size = sizeof(struct nvme_res_notification_page);
643 		break;
644 	case NVME_LOG_SANITIZE_STATUS:
645 		log_page_size = sizeof(struct nvme_sanitize_status_page);
646 		break;
647 	default:
648 		log_page_size = 0;
649 		break;
650 	}
651 
652 	return (log_page_size);
653 }
654 
655 static void
nvme_ctrlr_log_critical_warnings(struct nvme_controller * ctrlr,uint8_t state)656 nvme_ctrlr_log_critical_warnings(struct nvme_controller *ctrlr,
657     uint8_t state)
658 {
659 	if (state & NVME_CRIT_WARN_ST_AVAILABLE_SPARE)
660 		nvme_printf(ctrlr, "SMART WARNING: available spare space below threshold\n");
661 
662 	if (state & NVME_CRIT_WARN_ST_TEMPERATURE)
663 		nvme_printf(ctrlr, "SMART WARNING: temperature above threshold\n");
664 
665 	if (state & NVME_CRIT_WARN_ST_DEVICE_RELIABILITY)
666 		nvme_printf(ctrlr, "SMART WARNING: device reliability degraded\n");
667 
668 	if (state & NVME_CRIT_WARN_ST_READ_ONLY)
669 		nvme_printf(ctrlr, "SMART WARNING: media placed in read only mode\n");
670 
671 	if (state & NVME_CRIT_WARN_ST_VOLATILE_MEMORY_BACKUP)
672 		nvme_printf(ctrlr, "SMART WARNING: volatile memory backup device failed\n");
673 
674 	if (state & NVME_CRIT_WARN_ST_PERSISTENT_MEMORY_REGION)
675 		nvme_printf(ctrlr, "SMART WARNING: persistent memory read only or unreliable\n");
676 
677 	if (state & NVME_CRIT_WARN_ST_RESERVED_MASK)
678 		nvme_printf(ctrlr, "SMART WARNING: unknown critical warning(s): state = 0x%02x\n",
679 		    state & NVME_CRIT_WARN_ST_RESERVED_MASK);
680 
681 	nvme_ctrlr_devctl(ctrlr, "critical", "SMART_ERROR", "state=0x%02x", state);
682 }
683 
684 static void
nvme_ctrlr_async_event_cb(void * arg,const struct nvme_completion * cpl)685 nvme_ctrlr_async_event_cb(void *arg, const struct nvme_completion *cpl)
686 {
687 	struct nvme_async_event_request	*aer = arg;
688 
689 	if (nvme_completion_is_error(cpl)) {
690 		/*
691 		 *  Do not retry failed async event requests.  This avoids
692 		 *  infinite loops where a new async event request is submitted
693 		 *  to replace the one just failed, only to fail again and
694 		 *  perpetuate the loop.
695 		 */
696 		return;
697 	}
698 
699 	/*
700 	 * Save the completion status and associated log page is in bits 23:16
701 	 * of completion entry dw0. Print a message and queue it for further
702 	 * processing.
703 	 */
704 	memcpy(&aer->cpl, cpl, sizeof(*cpl));
705 	aer->log_page_id = NVMEV(NVME_ASYNC_EVENT_LOG_PAGE_ID, cpl->cdw0);
706 	nvme_printf(aer->ctrlr, "async event occurred (type 0x%x, info 0x%02x,"
707 	    " page 0x%02x)\n", NVMEV(NVME_ASYNC_EVENT_TYPE, cpl->cdw0),
708 	    NVMEV(NVME_ASYNC_EVENT_INFO, cpl->cdw0),
709 	    aer->log_page_id);
710 	taskqueue_enqueue(aer->ctrlr->taskqueue, &aer->task);
711 }
712 
713 static void
nvme_ctrlr_construct_and_submit_aer(struct nvme_controller * ctrlr,struct nvme_async_event_request * aer)714 nvme_ctrlr_construct_and_submit_aer(struct nvme_controller *ctrlr,
715     struct nvme_async_event_request *aer)
716 {
717 	struct nvme_request *req;
718 
719 	/*
720 	 * We're racing the reset thread, so let that process submit this again.
721 	 * XXX does this really solve that race? And is that race even possible
722 	 * since we only reset when we've no theard from the card in a long
723 	 * time. Why would we get an AER in the middle of that just before we
724 	 * kick off the reset?
725 	 */
726 	if (ctrlr->is_resetting)
727 		return;
728 
729 	aer->ctrlr = ctrlr;
730 	req = nvme_allocate_request_null(M_WAITOK, nvme_ctrlr_async_event_cb,
731 	    aer);
732 	aer->req = req;
733 	aer->log_page_id = 0;		/* Not a valid page */
734 
735 	/*
736 	 * Disable timeout here, since asynchronous event requests should by
737 	 *  nature never be timed out.
738 	 */
739 	req->timeout = false;
740 	req->cmd.opc = NVME_OPC_ASYNC_EVENT_REQUEST;
741 	nvme_ctrlr_submit_admin_request(ctrlr, req);
742 }
743 
744 static void
nvme_ctrlr_configure_aer(struct nvme_controller * ctrlr)745 nvme_ctrlr_configure_aer(struct nvme_controller *ctrlr)
746 {
747 	struct nvme_completion_poll_status	status;
748 	struct nvme_async_event_request		*aer;
749 	uint32_t				i;
750 
751 	ctrlr->async_event_config = NVME_CRIT_WARN_ST_AVAILABLE_SPARE |
752 	    NVME_CRIT_WARN_ST_DEVICE_RELIABILITY |
753 	    NVME_CRIT_WARN_ST_READ_ONLY |
754 	    NVME_CRIT_WARN_ST_VOLATILE_MEMORY_BACKUP;
755 	if (ctrlr->cdata.ver >= NVME_REV(1, 2))
756 		ctrlr->async_event_config |=
757 		    ctrlr->cdata.oaes & (NVME_ASYNC_EVENT_NS_ATTRIBUTE |
758 			NVME_ASYNC_EVENT_FW_ACTIVATE);
759 
760 	status.done = 0;
761 	nvme_ctrlr_cmd_get_feature(ctrlr, NVME_FEAT_TEMPERATURE_THRESHOLD,
762 	    0, NULL, 0, nvme_completion_poll_cb, &status);
763 	nvme_completion_poll(&status);
764 	if (nvme_completion_is_error(&status.cpl) ||
765 	    (status.cpl.cdw0 & 0xFFFF) == 0xFFFF ||
766 	    (status.cpl.cdw0 & 0xFFFF) == 0x0000) {
767 		nvme_printf(ctrlr, "temperature threshold not supported\n");
768 	} else
769 		ctrlr->async_event_config |= NVME_CRIT_WARN_ST_TEMPERATURE;
770 
771 	nvme_ctrlr_cmd_set_async_event_config(ctrlr,
772 	    ctrlr->async_event_config, NULL, NULL);
773 
774 	/* aerl is a zero-based value, so we need to add 1 here. */
775 	ctrlr->num_aers = min(NVME_MAX_ASYNC_EVENTS, (ctrlr->cdata.aerl+1));
776 
777 	for (i = 0; i < ctrlr->num_aers; i++) {
778 		aer = &ctrlr->aer[i];
779 		nvme_ctrlr_construct_and_submit_aer(ctrlr, aer);
780 	}
781 }
782 
783 static void
nvme_ctrlr_configure_int_coalescing(struct nvme_controller * ctrlr)784 nvme_ctrlr_configure_int_coalescing(struct nvme_controller *ctrlr)
785 {
786 	ctrlr->int_coal_time = 0;
787 	TUNABLE_INT_FETCH("hw.nvme.int_coal_time",
788 	    &ctrlr->int_coal_time);
789 
790 	ctrlr->int_coal_threshold = 0;
791 	TUNABLE_INT_FETCH("hw.nvme.int_coal_threshold",
792 	    &ctrlr->int_coal_threshold);
793 
794 	nvme_ctrlr_cmd_set_interrupt_coalescing(ctrlr, ctrlr->int_coal_time,
795 	    ctrlr->int_coal_threshold, NULL, NULL);
796 }
797 
798 static void
nvme_ctrlr_hmb_free(struct nvme_controller * ctrlr)799 nvme_ctrlr_hmb_free(struct nvme_controller *ctrlr)
800 {
801 	struct nvme_hmb_chunk *hmbc;
802 	int i;
803 
804 	if (ctrlr->hmb_desc_paddr) {
805 		bus_dmamap_unload(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_map);
806 		bus_dmamem_free(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_vaddr,
807 		    ctrlr->hmb_desc_map);
808 		ctrlr->hmb_desc_paddr = 0;
809 	}
810 	if (ctrlr->hmb_desc_tag) {
811 		bus_dma_tag_destroy(ctrlr->hmb_desc_tag);
812 		ctrlr->hmb_desc_tag = NULL;
813 	}
814 	for (i = 0; i < ctrlr->hmb_nchunks; i++) {
815 		hmbc = &ctrlr->hmb_chunks[i];
816 		bus_dmamap_unload(ctrlr->hmb_tag, hmbc->hmbc_map);
817 		bus_dmamem_free(ctrlr->hmb_tag, hmbc->hmbc_vaddr,
818 		    hmbc->hmbc_map);
819 	}
820 	ctrlr->hmb_nchunks = 0;
821 	if (ctrlr->hmb_tag) {
822 		bus_dma_tag_destroy(ctrlr->hmb_tag);
823 		ctrlr->hmb_tag = NULL;
824 	}
825 	if (ctrlr->hmb_chunks) {
826 		free(ctrlr->hmb_chunks, M_NVME);
827 		ctrlr->hmb_chunks = NULL;
828 	}
829 }
830 
831 static void
nvme_ctrlr_hmb_alloc(struct nvme_controller * ctrlr)832 nvme_ctrlr_hmb_alloc(struct nvme_controller *ctrlr)
833 {
834 	struct nvme_hmb_chunk *hmbc;
835 	size_t pref, min, minc, size;
836 	int err, i;
837 	uint64_t max;
838 
839 	/* Limit HMB to 5% of RAM size per device by default. */
840 	max = (uint64_t)physmem * PAGE_SIZE / 20;
841 	TUNABLE_UINT64_FETCH("hw.nvme.hmb_max", &max);
842 
843 	/*
844 	 * Units of Host Memory Buffer in the Identify info are always in terms
845 	 * of 4k units.
846 	 */
847 	min = (long long unsigned)ctrlr->cdata.hmmin * NVME_HMB_UNITS;
848 	if (max == 0 || max < min)
849 		return;
850 	pref = MIN((long long unsigned)ctrlr->cdata.hmpre * NVME_HMB_UNITS, max);
851 	minc = MAX(ctrlr->cdata.hmminds * NVME_HMB_UNITS, ctrlr->page_size);
852 	if (min > 0 && ctrlr->cdata.hmmaxd > 0)
853 		minc = MAX(minc, min / ctrlr->cdata.hmmaxd);
854 	ctrlr->hmb_chunk = pref;
855 
856 again:
857 	/*
858 	 * However, the chunk sizes, number of chunks, and alignment of chunks
859 	 * are all based on the current MPS (ctrlr->page_size).
860 	 */
861 	ctrlr->hmb_chunk = roundup2(ctrlr->hmb_chunk, ctrlr->page_size);
862 	ctrlr->hmb_nchunks = howmany(pref, ctrlr->hmb_chunk);
863 	if (ctrlr->cdata.hmmaxd > 0 && ctrlr->hmb_nchunks > ctrlr->cdata.hmmaxd)
864 		ctrlr->hmb_nchunks = ctrlr->cdata.hmmaxd;
865 	ctrlr->hmb_chunks = malloc(sizeof(struct nvme_hmb_chunk) *
866 	    ctrlr->hmb_nchunks, M_NVME, M_WAITOK);
867 	err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev),
868 	    ctrlr->page_size, 0, BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
869 	    ctrlr->hmb_chunk, 1, ctrlr->hmb_chunk, 0, NULL, NULL, &ctrlr->hmb_tag);
870 	if (err != 0) {
871 		nvme_printf(ctrlr, "HMB tag create failed %d\n", err);
872 		nvme_ctrlr_hmb_free(ctrlr);
873 		return;
874 	}
875 
876 	for (i = 0; i < ctrlr->hmb_nchunks; i++) {
877 		hmbc = &ctrlr->hmb_chunks[i];
878 		if (bus_dmamem_alloc(ctrlr->hmb_tag,
879 		    (void **)&hmbc->hmbc_vaddr, BUS_DMA_NOWAIT,
880 		    &hmbc->hmbc_map)) {
881 			nvme_printf(ctrlr, "failed to alloc HMB\n");
882 			break;
883 		}
884 		if (bus_dmamap_load(ctrlr->hmb_tag, hmbc->hmbc_map,
885 		    hmbc->hmbc_vaddr, ctrlr->hmb_chunk, nvme_single_map,
886 		    &hmbc->hmbc_paddr, BUS_DMA_NOWAIT) != 0) {
887 			bus_dmamem_free(ctrlr->hmb_tag, hmbc->hmbc_vaddr,
888 			    hmbc->hmbc_map);
889 			nvme_printf(ctrlr, "failed to load HMB\n");
890 			break;
891 		}
892 		bus_dmamap_sync(ctrlr->hmb_tag, hmbc->hmbc_map,
893 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
894 	}
895 
896 	if (i < ctrlr->hmb_nchunks && i * ctrlr->hmb_chunk < min &&
897 	    ctrlr->hmb_chunk / 2 >= minc) {
898 		ctrlr->hmb_nchunks = i;
899 		nvme_ctrlr_hmb_free(ctrlr);
900 		ctrlr->hmb_chunk /= 2;
901 		goto again;
902 	}
903 	ctrlr->hmb_nchunks = i;
904 	if (ctrlr->hmb_nchunks * ctrlr->hmb_chunk < min) {
905 		nvme_ctrlr_hmb_free(ctrlr);
906 		return;
907 	}
908 
909 	size = sizeof(struct nvme_hmb_desc) * ctrlr->hmb_nchunks;
910 	err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev),
911 	    16, 0, BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
912 	    size, 1, size, 0, NULL, NULL, &ctrlr->hmb_desc_tag);
913 	if (err != 0) {
914 		nvme_printf(ctrlr, "HMB desc tag create failed %d\n", err);
915 		nvme_ctrlr_hmb_free(ctrlr);
916 		return;
917 	}
918 	if (bus_dmamem_alloc(ctrlr->hmb_desc_tag,
919 	    (void **)&ctrlr->hmb_desc_vaddr, BUS_DMA_WAITOK,
920 	    &ctrlr->hmb_desc_map)) {
921 		nvme_printf(ctrlr, "failed to alloc HMB desc\n");
922 		nvme_ctrlr_hmb_free(ctrlr);
923 		return;
924 	}
925 	if (bus_dmamap_load(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_map,
926 	    ctrlr->hmb_desc_vaddr, size, nvme_single_map,
927 	    &ctrlr->hmb_desc_paddr, BUS_DMA_NOWAIT) != 0) {
928 		bus_dmamem_free(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_vaddr,
929 		    ctrlr->hmb_desc_map);
930 		nvme_printf(ctrlr, "failed to load HMB desc\n");
931 		nvme_ctrlr_hmb_free(ctrlr);
932 		return;
933 	}
934 
935 	for (i = 0; i < ctrlr->hmb_nchunks; i++) {
936 		memset(&ctrlr->hmb_desc_vaddr[i], 0,
937 		    sizeof(struct nvme_hmb_desc));
938 		ctrlr->hmb_desc_vaddr[i].addr =
939 		    htole64(ctrlr->hmb_chunks[i].hmbc_paddr);
940 		ctrlr->hmb_desc_vaddr[i].size = htole32(ctrlr->hmb_chunk / ctrlr->page_size);
941 	}
942 	bus_dmamap_sync(ctrlr->hmb_desc_tag, ctrlr->hmb_desc_map,
943 	    BUS_DMASYNC_PREWRITE);
944 
945 	nvme_printf(ctrlr, "Allocated %lluMB host memory buffer\n",
946 	    (long long unsigned)ctrlr->hmb_nchunks * ctrlr->hmb_chunk
947 	    / 1024 / 1024);
948 }
949 
950 static void
nvme_ctrlr_hmb_enable(struct nvme_controller * ctrlr,bool enable,bool memret)951 nvme_ctrlr_hmb_enable(struct nvme_controller *ctrlr, bool enable, bool memret)
952 {
953 	struct nvme_completion_poll_status	status;
954 	uint32_t cdw11;
955 
956 	cdw11 = 0;
957 	if (enable)
958 		cdw11 |= 1;
959 	if (memret)
960 		cdw11 |= 2;
961 	status.done = 0;
962 	nvme_ctrlr_cmd_set_feature(ctrlr, NVME_FEAT_HOST_MEMORY_BUFFER, cdw11,
963 	    ctrlr->hmb_nchunks * ctrlr->hmb_chunk / ctrlr->page_size,
964 	    ctrlr->hmb_desc_paddr, ctrlr->hmb_desc_paddr >> 32,
965 	    ctrlr->hmb_nchunks, NULL, 0,
966 	    nvme_completion_poll_cb, &status);
967 	nvme_completion_poll(&status);
968 	if (nvme_completion_is_error(&status.cpl))
969 		nvme_printf(ctrlr, "nvme_ctrlr_hmb_enable failed!\n");
970 }
971 
972 static void
nvme_ctrlr_start(void * ctrlr_arg,bool resetting)973 nvme_ctrlr_start(void *ctrlr_arg, bool resetting)
974 {
975 	struct nvme_controller *ctrlr = ctrlr_arg;
976 	uint32_t old_num_io_queues;
977 	int i;
978 
979 	TSENTER();
980 
981 	/*
982 	 * Only reset adminq here when we are restarting the
983 	 *  controller after a reset.  During initialization,
984 	 *  we have already submitted admin commands to get
985 	 *  the number of I/O queues supported, so cannot reset
986 	 *  the adminq again here.
987 	 */
988 	if (resetting) {
989 		nvme_qpair_reset(&ctrlr->adminq);
990 		nvme_admin_qpair_enable(&ctrlr->adminq);
991 	}
992 
993 	if (ctrlr->ioq != NULL) {
994 		for (i = 0; i < ctrlr->num_io_queues; i++)
995 			nvme_qpair_reset(&ctrlr->ioq[i]);
996 	}
997 
998 	/*
999 	 * If it was a reset on initialization command timeout, just
1000 	 * return here, letting initialization code fail gracefully.
1001 	 */
1002 	if (resetting && !ctrlr->is_initialized)
1003 		return;
1004 
1005 	if (resetting && nvme_ctrlr_identify(ctrlr) != 0) {
1006 		nvme_ctrlr_fail(ctrlr, false);
1007 		return;
1008 	}
1009 
1010 	/*
1011 	 * The number of qpairs are determined during controller initialization,
1012 	 *  including using NVMe SET_FEATURES/NUMBER_OF_QUEUES to determine the
1013 	 *  HW limit.  We call SET_FEATURES again here so that it gets called
1014 	 *  after any reset for controllers that depend on the driver to
1015 	 *  explicit specify how many queues it will use.  This value should
1016 	 *  never change between resets, so panic if somehow that does happen.
1017 	 */
1018 	if (resetting) {
1019 		old_num_io_queues = ctrlr->num_io_queues;
1020 		if (nvme_ctrlr_set_num_qpairs(ctrlr) != 0) {
1021 			nvme_ctrlr_fail(ctrlr, false);
1022 			return;
1023 		}
1024 
1025 		if (old_num_io_queues != ctrlr->num_io_queues) {
1026 			panic("num_io_queues changed from %u to %u",
1027 			      old_num_io_queues, ctrlr->num_io_queues);
1028 		}
1029 	}
1030 
1031 	if (ctrlr->cdata.hmpre > 0 && ctrlr->hmb_nchunks == 0) {
1032 		nvme_ctrlr_hmb_alloc(ctrlr);
1033 		if (ctrlr->hmb_nchunks > 0)
1034 			nvme_ctrlr_hmb_enable(ctrlr, true, false);
1035 	} else if (ctrlr->hmb_nchunks > 0)
1036 		nvme_ctrlr_hmb_enable(ctrlr, true, true);
1037 
1038 	if (nvme_ctrlr_create_qpairs(ctrlr) != 0) {
1039 		nvme_ctrlr_fail(ctrlr, false);
1040 		return;
1041 	}
1042 
1043 	if (nvme_ctrlr_construct_namespaces(ctrlr) != 0) {
1044 		nvme_ctrlr_fail(ctrlr, false);
1045 		return;
1046 	}
1047 
1048 	nvme_ctrlr_configure_aer(ctrlr);
1049 	nvme_ctrlr_configure_int_coalescing(ctrlr);
1050 
1051 	for (i = 0; i < ctrlr->num_io_queues; i++)
1052 		nvme_io_qpair_enable(&ctrlr->ioq[i]);
1053 	TSEXIT();
1054 }
1055 
1056 void
nvme_ctrlr_start_config_hook(void * arg)1057 nvme_ctrlr_start_config_hook(void *arg)
1058 {
1059 	struct nvme_controller *ctrlr = arg;
1060 
1061 	TSENTER();
1062 
1063 	if (nvme_ctrlr_hw_reset(ctrlr) != 0 || ctrlr->fail_on_reset != 0) {
1064 		nvme_ctrlr_fail(ctrlr, true);
1065 		config_intrhook_disestablish(&ctrlr->config_hook);
1066 		return;
1067 	}
1068 
1069 	nvme_qpair_reset(&ctrlr->adminq);
1070 	nvme_admin_qpair_enable(&ctrlr->adminq);
1071 
1072 	if (nvme_ctrlr_identify(ctrlr) == 0 &&
1073 	    nvme_ctrlr_set_num_qpairs(ctrlr) == 0 &&
1074 	    nvme_ctrlr_construct_io_qpairs(ctrlr) == 0)
1075 		nvme_ctrlr_start(ctrlr, false);
1076 	else
1077 		nvme_ctrlr_fail(ctrlr, false);
1078 
1079 	nvme_sysctl_initialize_ctrlr(ctrlr);
1080 	config_intrhook_disestablish(&ctrlr->config_hook);
1081 
1082 	if (!ctrlr->is_failed) {
1083 		ctrlr->is_initialized = true;
1084 		nvme_notify_new_controller(ctrlr);
1085 	}
1086 	TSEXIT();
1087 }
1088 
1089 static void
nvme_ctrlr_reset_task(void * arg,int pending)1090 nvme_ctrlr_reset_task(void *arg, int pending)
1091 {
1092 	struct nvme_controller	*ctrlr = arg;
1093 	int			status;
1094 
1095 	nvme_ctrlr_devctl_log(ctrlr, "RESET", "event=\"start\"");
1096 	status = nvme_ctrlr_hw_reset(ctrlr);
1097 	if (status == 0) {
1098 		nvme_ctrlr_devctl_log(ctrlr, "RESET", "event=\"success\"");
1099 		nvme_ctrlr_start(ctrlr, true);
1100 	} else {
1101 		nvme_ctrlr_devctl_log(ctrlr, "RESET", "event=\"timed_out\"");
1102 		nvme_ctrlr_fail(ctrlr, true);
1103 	}
1104 
1105 	atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
1106 }
1107 
1108 static void
nvme_ctrlr_aer_done(void * arg,const struct nvme_completion * cpl)1109 nvme_ctrlr_aer_done(void *arg,  const struct nvme_completion *cpl)
1110 {
1111 	struct nvme_async_event_request	*aer = arg;
1112 
1113 	mtx_lock(&aer->mtx);
1114 	if (nvme_completion_is_error(cpl))
1115 		aer->log_page_size = (uint32_t)-1;
1116 	else
1117 		aer->log_page_size = nvme_ctrlr_get_log_page_size(
1118 		    aer->ctrlr, aer->log_page_id);
1119 	wakeup(aer);
1120 	mtx_unlock(&aer->mtx);
1121 }
1122 
1123 static void
nvme_ctrlr_aer_task(void * arg,int pending)1124 nvme_ctrlr_aer_task(void *arg, int pending)
1125 {
1126 	struct nvme_async_event_request	*aer = arg;
1127 	struct nvme_controller	*ctrlr = aer->ctrlr;
1128 	uint32_t len;
1129 
1130 	/*
1131 	 * We're resetting, so just punt.
1132 	 */
1133 	if (ctrlr->is_resetting)
1134 		return;
1135 
1136 	if (!is_log_page_id_valid(aer->log_page_id)) {
1137 		/*
1138 		 * Repost another asynchronous event request to replace the one
1139 		 * that just completed.
1140 		 */
1141 		nvme_notify_async_consumers(ctrlr, &aer->cpl, aer->log_page_id,
1142 		    NULL, 0);
1143 		nvme_ctrlr_construct_and_submit_aer(ctrlr, aer);
1144 		goto out;
1145 	}
1146 
1147 	aer->log_page_size = 0;
1148 	len = nvme_ctrlr_get_log_page_size(aer->ctrlr, aer->log_page_id);
1149 	nvme_ctrlr_cmd_get_log_page(aer->ctrlr, aer->log_page_id,
1150 	    NVME_GLOBAL_NAMESPACE_TAG, aer->log_page_buffer, len,
1151 	    nvme_ctrlr_aer_done, aer);
1152 	mtx_lock(&aer->mtx);
1153 	while (aer->log_page_size == 0)
1154 		mtx_sleep(aer, &aer->mtx, PRIBIO, "nvme_pt", 0);
1155 	mtx_unlock(&aer->mtx);
1156 
1157 	if (aer->log_page_size != (uint32_t)-1) {
1158 		/*
1159 		 * If the log page fetch for some reason completed with an
1160 		 * error, don't pass log page data to the consumers.  In
1161 		 * practice, this case should never happen.
1162 		 */
1163 		nvme_notify_async_consumers(aer->ctrlr, &aer->cpl,
1164 		    aer->log_page_id, NULL, 0);
1165 		goto out;
1166 	}
1167 
1168 	/* Convert data to host endian */
1169 	switch (aer->log_page_id) {
1170 	case NVME_LOG_ERROR: {
1171 		struct nvme_error_information_entry *err =
1172 		    (struct nvme_error_information_entry *)aer->log_page_buffer;
1173 		for (int i = 0; i < (aer->ctrlr->cdata.elpe + 1); i++)
1174 			nvme_error_information_entry_swapbytes(err++);
1175 		break;
1176 	}
1177 	case NVME_LOG_HEALTH_INFORMATION:
1178 		nvme_health_information_page_swapbytes(
1179 			(struct nvme_health_information_page *)aer->log_page_buffer);
1180 		break;
1181 	case NVME_LOG_CHANGED_NAMESPACE:
1182 		nvme_ns_list_swapbytes(
1183 			(struct nvme_ns_list *)aer->log_page_buffer);
1184 		break;
1185 	case NVME_LOG_COMMAND_EFFECT:
1186 		nvme_command_effects_page_swapbytes(
1187 			(struct nvme_command_effects_page *)aer->log_page_buffer);
1188 		break;
1189 	case NVME_LOG_RES_NOTIFICATION:
1190 		nvme_res_notification_page_swapbytes(
1191 			(struct nvme_res_notification_page *)aer->log_page_buffer);
1192 		break;
1193 	case NVME_LOG_SANITIZE_STATUS:
1194 		nvme_sanitize_status_page_swapbytes(
1195 			(struct nvme_sanitize_status_page *)aer->log_page_buffer);
1196 		break;
1197 	default:
1198 		break;
1199 	}
1200 
1201 	if (aer->log_page_id == NVME_LOG_HEALTH_INFORMATION) {
1202 		struct nvme_health_information_page *health_info =
1203 		    (struct nvme_health_information_page *)aer->log_page_buffer;
1204 
1205 		/*
1206 		 * Critical warnings reported through the SMART/health log page
1207 		 * are persistent, so clear the associated bits in the async
1208 		 * event config so that we do not receive repeated notifications
1209 		 * for the same event.
1210 		 */
1211 		nvme_ctrlr_log_critical_warnings(aer->ctrlr,
1212 		    health_info->critical_warning);
1213 		aer->ctrlr->async_event_config &=
1214 		    ~health_info->critical_warning;
1215 		nvme_ctrlr_cmd_set_async_event_config(aer->ctrlr,
1216 		    aer->ctrlr->async_event_config, NULL, NULL);
1217 	} else if (aer->log_page_id == NVME_LOG_CHANGED_NAMESPACE) {
1218 		struct nvme_ns_list *nsl =
1219 		    (struct nvme_ns_list *)aer->log_page_buffer;
1220 		for (int i = 0; i < nitems(nsl->ns) && nsl->ns[i] != 0; i++) {
1221 			if (nsl->ns[i] > NVME_MAX_NAMESPACES)
1222 				break;
1223 			nvme_notify_ns(aer->ctrlr, nsl->ns[i]);
1224 		}
1225 	}
1226 
1227 	/*
1228 	 * Pass the cpl data from the original async event completion, not the
1229 	 * log page fetch.
1230 	 */
1231 	nvme_notify_async_consumers(aer->ctrlr, &aer->cpl,
1232 	    aer->log_page_id, aer->log_page_buffer, aer->log_page_size);
1233 
1234 	/*
1235 	 * Repost another asynchronous event request to replace the one
1236 	 *  that just completed.
1237 	 */
1238 out:
1239 	nvme_ctrlr_construct_and_submit_aer(ctrlr, aer);
1240 }
1241 
1242 /*
1243  * Poll all the queues enabled on the device for completion.
1244  */
1245 void
nvme_ctrlr_poll(struct nvme_controller * ctrlr)1246 nvme_ctrlr_poll(struct nvme_controller *ctrlr)
1247 {
1248 	int i;
1249 
1250 	nvme_qpair_process_completions(&ctrlr->adminq);
1251 
1252 	for (i = 0; i < ctrlr->num_io_queues; i++)
1253 		if (ctrlr->ioq && ctrlr->ioq[i].cpl)
1254 			nvme_qpair_process_completions(&ctrlr->ioq[i]);
1255 }
1256 
1257 /*
1258  * Copy the NVME device's serial number to the provided buffer, which must be
1259  * at least DISK_IDENT_SIZE bytes large.
1260  */
1261 void
nvme_ctrlr_get_ident(const struct nvme_controller * ctrlr,uint8_t * sn)1262 nvme_ctrlr_get_ident(const struct nvme_controller *ctrlr, uint8_t *sn)
1263 {
1264 	_Static_assert(NVME_SERIAL_NUMBER_LENGTH < DISK_IDENT_SIZE,
1265 		"NVME serial number too big for disk ident");
1266 
1267 	memmove(sn, ctrlr->cdata.sn, NVME_SERIAL_NUMBER_LENGTH);
1268 	sn[NVME_SERIAL_NUMBER_LENGTH] = '\0';
1269 	for (int i = 0; sn[i] != '\0'; i++) {
1270 		if (sn[i] < 0x20 || sn[i] >= 0x80)
1271 			sn[i] = ' ';
1272 	}
1273 }
1274 
1275 /*
1276  * Poll the single-vector interrupt case: num_io_queues will be 1 and
1277  * there's only a single vector. While we're polling, we mask further
1278  * interrupts in the controller.
1279  */
1280 void
nvme_ctrlr_shared_handler(void * arg)1281 nvme_ctrlr_shared_handler(void *arg)
1282 {
1283 	struct nvme_controller *ctrlr = arg;
1284 
1285 	nvme_mmio_write_4(ctrlr, intms, 1);
1286 	nvme_ctrlr_poll(ctrlr);
1287 	nvme_mmio_write_4(ctrlr, intmc, 1);
1288 }
1289 
1290 #define NVME_MAX_PAGES  (int)(1024 / sizeof(vm_page_t))
1291 
1292 static int
nvme_user_ioctl_req(vm_offset_t addr,size_t len,bool is_read,vm_page_t * upages,int max_pages,int * npagesp,struct nvme_request ** req,nvme_cb_fn_t cb_fn,void * cb_arg)1293 nvme_user_ioctl_req(vm_offset_t addr, size_t len, bool is_read,
1294     vm_page_t *upages, int max_pages, int *npagesp, struct nvme_request **req,
1295     nvme_cb_fn_t cb_fn, void *cb_arg)
1296 {
1297 	vm_prot_t prot = VM_PROT_READ;
1298 	int err;
1299 
1300 	if (is_read)
1301 		prot |= VM_PROT_WRITE;	/* Device will write to host memory */
1302 	err = vm_fault_hold_pages(&curproc->p_vmspace->vm_map,
1303 	    addr, len, prot, upages, max_pages, npagesp);
1304 	if (err != 0)
1305 		return (err);
1306 	*req = nvme_allocate_request_null(M_WAITOK, cb_fn, cb_arg);
1307 	(*req)->payload = memdesc_vmpages(upages, len, addr & PAGE_MASK);
1308 	(*req)->payload_valid = true;
1309 	return (0);
1310 }
1311 
1312 static void
nvme_user_ioctl_free(vm_page_t * pages,int npage)1313 nvme_user_ioctl_free(vm_page_t *pages, int npage)
1314 {
1315 	vm_page_unhold_pages(pages, npage);
1316 }
1317 
1318 static void
nvme_pt_done(void * arg,const struct nvme_completion * cpl)1319 nvme_pt_done(void *arg, const struct nvme_completion *cpl)
1320 {
1321 	struct nvme_pt_command *pt = arg;
1322 	struct mtx *mtx = pt->driver_lock;
1323 	uint16_t status;
1324 
1325 	bzero(&pt->cpl, sizeof(pt->cpl));
1326 	pt->cpl.cdw0 = cpl->cdw0;
1327 
1328 	status = cpl->status;
1329 	status &= ~NVMEM(NVME_STATUS_P);
1330 	pt->cpl.status = status;
1331 
1332 	mtx_lock(mtx);
1333 	pt->driver_lock = NULL;
1334 	wakeup(pt);
1335 	mtx_unlock(mtx);
1336 }
1337 
1338 int
nvme_ctrlr_passthrough_cmd(struct nvme_controller * ctrlr,struct nvme_pt_command * pt,uint32_t nsid,int is_user,int is_admin_cmd)1339 nvme_ctrlr_passthrough_cmd(struct nvme_controller *ctrlr,
1340     struct nvme_pt_command *pt, uint32_t nsid, int is_user,
1341     int is_admin_cmd)
1342 {
1343 	struct nvme_request *req;
1344 	struct mtx *mtx;
1345 	int ret = 0;
1346 	int npages = 0;
1347 	vm_page_t upages[NVME_MAX_PAGES];
1348 
1349 	if (pt->len > 0) {
1350 		if (pt->len > ctrlr->max_xfer_size) {
1351 			nvme_printf(ctrlr,
1352 			    "len (%d) exceeds max_xfer_size (%d)\n",
1353 			    pt->len, ctrlr->max_xfer_size);
1354 			return (EIO);
1355 		}
1356 		if (is_user) {
1357 			ret = nvme_user_ioctl_req((vm_offset_t)pt->buf, pt->len,
1358 			    pt->is_read, upages, nitems(upages), &npages, &req,
1359 			    nvme_pt_done, pt);
1360 			if (ret != 0)
1361 				return (ret);
1362 		} else
1363 			req = nvme_allocate_request_vaddr(pt->buf, pt->len,
1364 			    M_WAITOK, nvme_pt_done, pt);
1365 	} else
1366 		req = nvme_allocate_request_null(M_WAITOK, nvme_pt_done, pt);
1367 
1368 	/* Assume user space already converted to little-endian */
1369 	req->cmd.opc = pt->cmd.opc;
1370 	req->cmd.fuse = pt->cmd.fuse;
1371 	req->cmd.rsvd2 = pt->cmd.rsvd2;
1372 	req->cmd.rsvd3 = pt->cmd.rsvd3;
1373 	req->cmd.cdw10 = pt->cmd.cdw10;
1374 	req->cmd.cdw11 = pt->cmd.cdw11;
1375 	req->cmd.cdw12 = pt->cmd.cdw12;
1376 	req->cmd.cdw13 = pt->cmd.cdw13;
1377 	req->cmd.cdw14 = pt->cmd.cdw14;
1378 	req->cmd.cdw15 = pt->cmd.cdw15;
1379 
1380 	req->cmd.nsid = htole32(nsid);
1381 
1382 	mtx = mtx_pool_find(mtxpool_sleep, pt);
1383 	pt->driver_lock = mtx;
1384 
1385 	if (is_admin_cmd)
1386 		nvme_ctrlr_submit_admin_request(ctrlr, req);
1387 	else
1388 		nvme_ctrlr_submit_io_request(ctrlr, req);
1389 
1390 	mtx_lock(mtx);
1391 	while (pt->driver_lock != NULL)
1392 		mtx_sleep(pt, mtx, PRIBIO, "nvme_pt", 0);
1393 	mtx_unlock(mtx);
1394 
1395 	if (npages > 0)
1396 		nvme_user_ioctl_free(upages, npages);
1397 
1398 	return (ret);
1399 }
1400 
1401 static void
nvme_npc_done(void * arg,const struct nvme_completion * cpl)1402 nvme_npc_done(void *arg, const struct nvme_completion *cpl)
1403 {
1404 	struct nvme_passthru_cmd *npc = arg;
1405 	struct mtx *mtx = (void *)(uintptr_t)npc->metadata;
1406 
1407 	npc->result = cpl->cdw0;	/* cpl in host order by now */
1408 	mtx_lock(mtx);
1409 	npc->metadata = 0;
1410 	wakeup(npc);
1411 	mtx_unlock(mtx);
1412 }
1413 
1414 /* XXX refactor? */
1415 
1416 int
nvme_ctrlr_linux_passthru_cmd(struct nvme_controller * ctrlr,struct nvme_passthru_cmd * npc,uint32_t nsid,bool is_user,bool is_admin)1417 nvme_ctrlr_linux_passthru_cmd(struct nvme_controller *ctrlr,
1418     struct nvme_passthru_cmd *npc, uint32_t nsid, bool is_user, bool is_admin)
1419 {
1420 	struct nvme_request	*req;
1421 	struct mtx		*mtx;
1422 	int			ret = 0;
1423 	int			npages = 0;
1424 	vm_page_t		upages[NVME_MAX_PAGES];
1425 
1426 	/*
1427 	 * We don't support metadata.
1428 	 */
1429 	if (npc->metadata != 0 || npc->metadata_len != 0)
1430 		return (EIO);
1431 
1432 	if (npc->data_len > 0 && npc->addr != 0) {
1433 		if (npc->data_len > ctrlr->max_xfer_size) {
1434 			nvme_printf(ctrlr,
1435 			    "data_len (%d) exceeds max_xfer_size (%d)\n",
1436 			    npc->data_len, ctrlr->max_xfer_size);
1437 			return (EIO);
1438 		}
1439 		if (is_user) {
1440 			ret = nvme_user_ioctl_req(npc->addr, npc->data_len,
1441 			    npc->opcode & 0x1, upages, nitems(upages), &npages,
1442 			    &req, nvme_npc_done, npc);
1443 			if (ret != 0)
1444 				return (ret);
1445 		} else
1446 			req = nvme_allocate_request_vaddr(
1447 			    (void *)(uintptr_t)npc->addr, npc->data_len,
1448 			    M_WAITOK, nvme_npc_done, npc);
1449 	} else
1450 		req = nvme_allocate_request_null(M_WAITOK, nvme_npc_done, npc);
1451 
1452 	req->cmd.opc = npc->opcode;
1453 	req->cmd.fuse = npc->flags;
1454 	req->cmd.rsvd2 = htole32(npc->cdw2);
1455 	req->cmd.rsvd3 = htole32(npc->cdw3);
1456 	req->cmd.cdw10 = htole32(npc->cdw10);
1457 	req->cmd.cdw11 = htole32(npc->cdw11);
1458 	req->cmd.cdw12 = htole32(npc->cdw12);
1459 	req->cmd.cdw13 = htole32(npc->cdw13);
1460 	req->cmd.cdw14 = htole32(npc->cdw14);
1461 	req->cmd.cdw15 = htole32(npc->cdw15);
1462 
1463 	req->cmd.nsid = htole32(nsid);
1464 
1465 	mtx = mtx_pool_find(mtxpool_sleep, npc);
1466 	npc->metadata = (uintptr_t) mtx;
1467 
1468 	/* XXX no timeout passed down */
1469 	if (is_admin)
1470 		nvme_ctrlr_submit_admin_request(ctrlr, req);
1471 	else
1472 		nvme_ctrlr_submit_io_request(ctrlr, req);
1473 
1474 	mtx_lock(mtx);
1475 	while (npc->metadata != 0)
1476 		mtx_sleep(npc, mtx, PRIBIO, "nvme_npc", 0);
1477 	mtx_unlock(mtx);
1478 
1479 	if (npages > 0)
1480 		nvme_user_ioctl_free(upages, npages);
1481 
1482 	return (ret);
1483 }
1484 
1485 static int
nvme_ctrlr_ioctl(struct cdev * cdev,u_long cmd,caddr_t arg,int flag,struct thread * td)1486 nvme_ctrlr_ioctl(struct cdev *cdev, u_long cmd, caddr_t arg, int flag,
1487     struct thread *td)
1488 {
1489 	struct nvme_controller			*ctrlr;
1490 	struct nvme_pt_command			*pt;
1491 
1492 	ctrlr = cdev->si_drv1;
1493 
1494 	switch (cmd) {
1495 	case NVME_IOCTL_RESET: /* Linux compat */
1496 	case NVME_RESET_CONTROLLER:
1497 		nvme_ctrlr_reset(ctrlr);
1498 		break;
1499 	case NVME_PASSTHROUGH_CMD:
1500 		pt = (struct nvme_pt_command *)arg;
1501 		return (nvme_ctrlr_passthrough_cmd(ctrlr, pt, le32toh(pt->cmd.nsid),
1502 		    1 /* is_user_buffer */, 1 /* is_admin_cmd */));
1503 	case NVME_GET_NSID:
1504 	{
1505 		struct nvme_get_nsid *gnsid = (struct nvme_get_nsid *)arg;
1506 		strlcpy(gnsid->cdev, device_get_nameunit(ctrlr->dev),
1507 		    sizeof(gnsid->cdev));
1508 		gnsid->nsid = 0;
1509 		break;
1510 	}
1511 	case NVME_GET_MAX_XFER_SIZE:
1512 		*(uint64_t *)arg = ctrlr->max_xfer_size;
1513 		break;
1514 	case NVME_GET_CONTROLLER_DATA:
1515 		memcpy(arg, &ctrlr->cdata, sizeof(ctrlr->cdata));
1516 		break;
1517 	case DIOCGIDENT: {
1518 		uint8_t *sn = arg;
1519 		nvme_ctrlr_get_ident(ctrlr, sn);
1520 		break;
1521 	}
1522 	/* Linux Compatible (see nvme_linux.h) */
1523 	case NVME_IOCTL_ID:
1524 		td->td_retval[0] = 0xfffffffful;
1525 		return (0);
1526 
1527 	case NVME_IOCTL_ADMIN_CMD:
1528 	case NVME_IOCTL_IO_CMD: {
1529 		struct nvme_passthru_cmd *npc = (struct nvme_passthru_cmd *)arg;
1530 
1531 		return (nvme_ctrlr_linux_passthru_cmd(ctrlr, npc, npc->nsid, true,
1532 		    cmd == NVME_IOCTL_ADMIN_CMD));
1533 	}
1534 
1535 	default:
1536 		return (ENOTTY);
1537 	}
1538 
1539 	return (0);
1540 }
1541 
1542 static struct cdevsw nvme_ctrlr_cdevsw = {
1543 	.d_version =	D_VERSION,
1544 	.d_flags =	0,
1545 	.d_ioctl =	nvme_ctrlr_ioctl
1546 };
1547 
1548 int
nvme_ctrlr_construct(struct nvme_controller * ctrlr,device_t dev)1549 nvme_ctrlr_construct(struct nvme_controller *ctrlr, device_t dev)
1550 {
1551 	struct make_dev_args	md_args;
1552 	uint32_t	cap_lo;
1553 	uint32_t	cap_hi;
1554 	uint32_t	to, vs, pmrcap;
1555 	int		status, timeout_period;
1556 
1557 	ctrlr->dev = dev;
1558 
1559 	mtx_init(&ctrlr->lock, "nvme ctrlr lock", NULL, MTX_DEF);
1560 	if (bus_get_domain(dev, &ctrlr->domain) != 0)
1561 		ctrlr->domain = 0;
1562 
1563 	ctrlr->cap_lo = cap_lo = nvme_mmio_read_4(ctrlr, cap_lo);
1564 	if (bootverbose) {
1565 		device_printf(dev, "CapLo: 0x%08x: MQES %u%s%s%s%s, TO %u\n",
1566 		    cap_lo, NVME_CAP_LO_MQES(cap_lo),
1567 		    NVME_CAP_LO_CQR(cap_lo) ? ", CQR" : "",
1568 		    NVME_CAP_LO_AMS(cap_lo) ? ", AMS" : "",
1569 		    (NVME_CAP_LO_AMS(cap_lo) & 0x1) ? " WRRwUPC" : "",
1570 		    (NVME_CAP_LO_AMS(cap_lo) & 0x2) ? " VS" : "",
1571 		    NVME_CAP_LO_TO(cap_lo));
1572 	}
1573 	ctrlr->cap_hi = cap_hi = nvme_mmio_read_4(ctrlr, cap_hi);
1574 	if (bootverbose) {
1575 		device_printf(dev, "CapHi: 0x%08x: DSTRD %u%s, CSS %x%s, "
1576 		    "CPS %x, MPSMIN %u, MPSMAX %u%s%s%s%s%s\n", cap_hi,
1577 		    NVME_CAP_HI_DSTRD(cap_hi),
1578 		    NVME_CAP_HI_NSSRS(cap_hi) ? ", NSSRS" : "",
1579 		    NVME_CAP_HI_CSS(cap_hi),
1580 		    NVME_CAP_HI_BPS(cap_hi) ? ", BPS" : "",
1581 		    NVME_CAP_HI_CPS(cap_hi),
1582 		    NVME_CAP_HI_MPSMIN(cap_hi),
1583 		    NVME_CAP_HI_MPSMAX(cap_hi),
1584 		    NVME_CAP_HI_PMRS(cap_hi) ? ", PMRS" : "",
1585 		    NVME_CAP_HI_CMBS(cap_hi) ? ", CMBS" : "",
1586 		    NVME_CAP_HI_NSSS(cap_hi) ? ", NSSS" : "",
1587 		    NVME_CAP_HI_CRWMS(cap_hi) ? ", CRWMS" : "",
1588 		    NVME_CAP_HI_CRIMS(cap_hi) ? ", CRIMS" : "");
1589 	}
1590 	if (bootverbose) {
1591 		vs = nvme_mmio_read_4(ctrlr, vs);
1592 		device_printf(dev, "Version: 0x%08x: %d.%d\n", vs,
1593 		    NVME_MAJOR(vs), NVME_MINOR(vs));
1594 	}
1595 	if (bootverbose && NVME_CAP_HI_PMRS(cap_hi)) {
1596 		pmrcap = nvme_mmio_read_4(ctrlr, pmrcap);
1597 		device_printf(dev, "PMRCap: 0x%08x: BIR %u%s%s, PMRTU %u, "
1598 		    "PMRWBM %x, PMRTO %u%s\n", pmrcap,
1599 		    NVME_PMRCAP_BIR(pmrcap),
1600 		    NVME_PMRCAP_RDS(pmrcap) ? ", RDS" : "",
1601 		    NVME_PMRCAP_WDS(pmrcap) ? ", WDS" : "",
1602 		    NVME_PMRCAP_PMRTU(pmrcap),
1603 		    NVME_PMRCAP_PMRWBM(pmrcap),
1604 		    NVME_PMRCAP_PMRTO(pmrcap),
1605 		    NVME_PMRCAP_CMSS(pmrcap) ? ", CMSS" : "");
1606 	}
1607 
1608 	ctrlr->dstrd = NVME_CAP_HI_DSTRD(cap_hi) + 2;
1609 
1610 	ctrlr->mps = NVME_CAP_HI_MPSMIN(cap_hi);
1611 	ctrlr->page_size = 1 << (NVME_MPS_SHIFT + ctrlr->mps);
1612 
1613 	/* Get ready timeout value from controller, in units of 500ms. */
1614 	to = NVME_CAP_LO_TO(cap_lo) + 1;
1615 	ctrlr->ready_timeout_in_ms = to * 500;
1616 
1617 	timeout_period = NVME_ADMIN_TIMEOUT_PERIOD;
1618 	TUNABLE_INT_FETCH("hw.nvme.admin_timeout_period", &timeout_period);
1619 	timeout_period = min(timeout_period, NVME_MAX_TIMEOUT_PERIOD);
1620 	timeout_period = max(timeout_period, NVME_MIN_TIMEOUT_PERIOD);
1621 	ctrlr->admin_timeout_period = timeout_period;
1622 
1623 	timeout_period = NVME_DEFAULT_TIMEOUT_PERIOD;
1624 	TUNABLE_INT_FETCH("hw.nvme.timeout_period", &timeout_period);
1625 	timeout_period = min(timeout_period, NVME_MAX_TIMEOUT_PERIOD);
1626 	timeout_period = max(timeout_period, NVME_MIN_TIMEOUT_PERIOD);
1627 	ctrlr->timeout_period = timeout_period;
1628 
1629 	nvme_retry_count = NVME_DEFAULT_RETRY_COUNT;
1630 	TUNABLE_INT_FETCH("hw.nvme.retry_count", &nvme_retry_count);
1631 
1632 	ctrlr->enable_aborts = 0;
1633 	TUNABLE_INT_FETCH("hw.nvme.enable_aborts", &ctrlr->enable_aborts);
1634 
1635 	ctrlr->alignment_splits = counter_u64_alloc(M_WAITOK);
1636 
1637 	/* Cap transfers by the maximum addressable by page-sized PRP (4KB pages -> 2MB). */
1638 	ctrlr->max_xfer_size = MIN(maxphys, (ctrlr->page_size / 8 * ctrlr->page_size));
1639 	if (nvme_ctrlr_construct_admin_qpair(ctrlr) != 0)
1640 		return (ENXIO);
1641 
1642 	/*
1643 	 * Create 2 threads for the taskqueue. The reset thread will block when
1644 	 * it detects that the controller has failed until all I/O has been
1645 	 * failed up the stack. The second thread is used for AER events, which
1646 	 * can block, but only briefly for memory and log page fetching.
1647 	 */
1648 	ctrlr->taskqueue = taskqueue_create("nvme_taskq", M_WAITOK,
1649 	    taskqueue_thread_enqueue, &ctrlr->taskqueue);
1650 	taskqueue_start_threads(&ctrlr->taskqueue, 2, PI_DISK, "nvme taskq");
1651 
1652 	ctrlr->is_resetting = 0;
1653 	ctrlr->is_initialized = false;
1654 	ctrlr->notification_sent = 0;
1655 	TASK_INIT(&ctrlr->reset_task, 0, nvme_ctrlr_reset_task, ctrlr);
1656 	for (int i = 0; i < NVME_MAX_ASYNC_EVENTS; i++) {
1657 		struct nvme_async_event_request *aer = &ctrlr->aer[i];
1658 
1659 		TASK_INIT(&aer->task, 0, nvme_ctrlr_aer_task, aer);
1660 		mtx_init(&aer->mtx, "AER mutex", NULL, MTX_DEF);
1661 	}
1662 	ctrlr->is_failed = false;
1663 
1664 	make_dev_args_init(&md_args);
1665 	md_args.mda_devsw = &nvme_ctrlr_cdevsw;
1666 	md_args.mda_uid = UID_ROOT;
1667 	md_args.mda_gid = GID_WHEEL;
1668 	md_args.mda_mode = 0600;
1669 	md_args.mda_unit = device_get_unit(dev);
1670 	md_args.mda_si_drv1 = (void *)ctrlr;
1671 	status = make_dev_s(&md_args, &ctrlr->cdev, "%s",
1672 	    device_get_nameunit(dev));
1673 	if (status != 0)
1674 		return (ENXIO);
1675 
1676 	return (0);
1677 }
1678 
1679 /*
1680  * Called on detach, or on error on attach. The nvme_controller won't be used
1681  * again once we return, so we have to tear everything down (so nothing
1682  * references this, no callbacks, etc), but don't need to reset all the state
1683  * since nvme_controller will be freed soon.
1684  */
1685 void
nvme_ctrlr_destruct(struct nvme_controller * ctrlr,device_t dev)1686 nvme_ctrlr_destruct(struct nvme_controller *ctrlr, device_t dev)
1687 {
1688 	int	i;
1689 	bool	gone;
1690 
1691 	ctrlr->is_dying = true;
1692 
1693 	if (ctrlr->resource == NULL)
1694 		goto nores;
1695 	if (!mtx_initialized(&ctrlr->adminq.lock))
1696 		goto noadminq;
1697 
1698 	/*
1699 	 * Check whether it is a hot unplug or a clean driver detach.  If device
1700 	 * is not there any more, skip any shutdown commands.  Some hotplug
1701 	 * bridges will return zeros instead of ff's when the device is
1702 	 * departing, so ask the bridge if the device is gone. Some systems can
1703 	 * remove the drive w/o the bridge knowing its gone (they don't really
1704 	 * do hotplug), so failsafe with detecting all ff's (impossible with
1705 	 * this hardware) as the device being gone.
1706 	 */
1707 	gone = bus_child_present(dev) == 0 ||
1708 	    (nvme_mmio_read_4(ctrlr, csts) == NVME_GONE);
1709 	if (gone)
1710 		nvme_ctrlr_fail(ctrlr, true);
1711 	else
1712 		nvme_notify_fail_consumers(ctrlr);
1713 
1714 	for (i = 0; i < NVME_MAX_NAMESPACES; i++)
1715 		nvme_ns_destruct(&ctrlr->ns[i]);
1716 
1717 	if (ctrlr->cdev)
1718 		destroy_dev(ctrlr->cdev);
1719 
1720 	if (ctrlr->is_initialized) {
1721 		if (!gone) {
1722 			if (ctrlr->hmb_nchunks > 0)
1723 				nvme_ctrlr_hmb_enable(ctrlr, false, false);
1724 			nvme_ctrlr_delete_qpairs(ctrlr);
1725 		}
1726 		nvme_ctrlr_hmb_free(ctrlr);
1727 	}
1728 	if (ctrlr->ioq != NULL) {
1729 		for (i = 0; i < ctrlr->num_io_queues; i++)
1730 			nvme_io_qpair_destroy(&ctrlr->ioq[i]);
1731 		free(ctrlr->ioq, M_NVME);
1732 	}
1733 	nvme_admin_qpair_destroy(&ctrlr->adminq);
1734 
1735 	/*
1736 	 * Notify the controller of a shutdown, even though this is due to a
1737 	 * driver unload, not a system shutdown (this path is not invoked uring
1738 	 * shutdown).  This ensures the controller receives a shutdown
1739 	 * notification in case the system is shutdown before reloading the
1740 	 * driver. Some NVMe drives need this to flush their cache to stable
1741 	 * media and consider it a safe shutdown in SMART stats.
1742 	 */
1743 	if (!gone) {
1744 		nvme_ctrlr_shutdown(ctrlr);
1745 		nvme_ctrlr_disable(ctrlr);
1746 	}
1747 
1748 noadminq:
1749 	if (ctrlr->taskqueue) {
1750 		taskqueue_free(ctrlr->taskqueue);
1751 		for (int i = 0; i < NVME_MAX_ASYNC_EVENTS; i++) {
1752 			struct nvme_async_event_request *aer = &ctrlr->aer[i];
1753 
1754 			mtx_destroy(&aer->mtx);
1755 		}
1756 	}
1757 
1758 	if (ctrlr->tag)
1759 		bus_teardown_intr(ctrlr->dev, ctrlr->res, ctrlr->tag);
1760 
1761 	if (ctrlr->res)
1762 		bus_release_resource(ctrlr->dev, SYS_RES_IRQ,
1763 		    rman_get_rid(ctrlr->res), ctrlr->res);
1764 
1765 	if (ctrlr->msix_table_resource != NULL) {
1766 		bus_release_resource(dev, SYS_RES_MEMORY,
1767 		    ctrlr->msix_table_resource_id, ctrlr->msix_table_resource);
1768 	}
1769 
1770 	if (ctrlr->msix_pba_resource != NULL) {
1771 		bus_release_resource(dev, SYS_RES_MEMORY,
1772 		    ctrlr->msix_pba_resource_id, ctrlr->msix_pba_resource);
1773 	}
1774 
1775 	bus_release_resource(dev, SYS_RES_MEMORY,
1776 	    ctrlr->resource_id, ctrlr->resource);
1777 
1778 nores:
1779 	if (ctrlr->alignment_splits)
1780 		counter_u64_free(ctrlr->alignment_splits);
1781 
1782 	mtx_destroy(&ctrlr->lock);
1783 }
1784 
1785 void
nvme_ctrlr_shutdown(struct nvme_controller * ctrlr)1786 nvme_ctrlr_shutdown(struct nvme_controller *ctrlr)
1787 {
1788 	uint32_t	cc;
1789 	uint32_t	csts;
1790 	int		timeout;
1791 
1792 	cc = nvme_mmio_read_4(ctrlr, cc);
1793 	cc &= ~NVMEM(NVME_CC_REG_SHN);
1794 	cc |= NVMEF(NVME_CC_REG_SHN, NVME_SHN_NORMAL);
1795 	nvme_mmio_write_4(ctrlr, cc, cc);
1796 
1797 	timeout = ticks + (ctrlr->cdata.rtd3e == 0 ? 5 * hz :
1798 	    ((uint64_t)ctrlr->cdata.rtd3e * hz + 999999) / 1000000);
1799 	while (1) {
1800 		csts = nvme_mmio_read_4(ctrlr, csts);
1801 		if (csts == NVME_GONE)		/* Hot unplug. */
1802 			break;
1803 		if (NVME_CSTS_GET_SHST(csts) == NVME_SHST_COMPLETE)
1804 			break;
1805 		if (timeout - ticks < 0) {
1806 			nvme_printf(ctrlr, "shutdown timeout\n");
1807 			break;
1808 		}
1809 		pause("nvmeshut", 1);
1810 	}
1811 }
1812 
1813 void
nvme_ctrlr_submit_admin_request(struct nvme_controller * ctrlr,struct nvme_request * req)1814 nvme_ctrlr_submit_admin_request(struct nvme_controller *ctrlr,
1815     struct nvme_request *req)
1816 {
1817 	nvme_qpair_submit_request(&ctrlr->adminq, req);
1818 }
1819 
1820 void
nvme_ctrlr_submit_io_request(struct nvme_controller * ctrlr,struct nvme_request * req)1821 nvme_ctrlr_submit_io_request(struct nvme_controller *ctrlr,
1822     struct nvme_request *req)
1823 {
1824 	struct nvme_qpair       *qpair;
1825 
1826 	qpair = &ctrlr->ioq[QP(ctrlr, curcpu)];
1827 	nvme_qpair_submit_request(qpair, req);
1828 }
1829 
1830 device_t
nvme_ctrlr_get_device(struct nvme_controller * ctrlr)1831 nvme_ctrlr_get_device(struct nvme_controller *ctrlr)
1832 {
1833 	return (ctrlr->dev);
1834 }
1835 
1836 const struct nvme_controller_data *
nvme_ctrlr_get_data(struct nvme_controller * ctrlr)1837 nvme_ctrlr_get_data(struct nvme_controller *ctrlr)
1838 {
1839 	return (&ctrlr->cdata);
1840 }
1841 
1842 int
nvme_ctrlr_suspend(struct nvme_controller * ctrlr)1843 nvme_ctrlr_suspend(struct nvme_controller *ctrlr)
1844 {
1845 	int to = hz;
1846 
1847 	/*
1848 	 * Can't touch failed controllers, so it's already suspended. User will
1849 	 * need to do an explicit reset to bring it back, if that's even
1850 	 * possible.
1851 	 */
1852 	if (ctrlr->is_failed)
1853 		return (0);
1854 
1855 	/*
1856 	 * We don't want the reset taskqueue running, since it does similar
1857 	 * things, so prevent it from running after we start. Wait for any reset
1858 	 * that may have been started to complete. The reset process we follow
1859 	 * will ensure that any new I/O will queue and be given to the hardware
1860 	 * after we resume (though there should be none).
1861 	 */
1862 	while (atomic_cmpset_32(&ctrlr->is_resetting, 0, 1) == 0 && to-- > 0)
1863 		pause("nvmesusp", 1);
1864 	if (to <= 0) {
1865 		nvme_printf(ctrlr,
1866 		    "Competing reset task didn't finish. Try again later.\n");
1867 		return (EWOULDBLOCK);
1868 	}
1869 
1870 	if (ctrlr->hmb_nchunks > 0)
1871 		nvme_ctrlr_hmb_enable(ctrlr, false, false);
1872 
1873 	/*
1874 	 * Per Section 7.6.2 of NVMe spec 1.4, to properly suspend, we need to
1875 	 * delete the hardware I/O queues, and then shutdown. This properly
1876 	 * flushes any metadata the drive may have stored so it can survive
1877 	 * having its power removed and prevents the unsafe shutdown count from
1878 	 * incriminating. Once we delete the qpairs, we have to disable them
1879 	 * before shutting down.
1880 	 */
1881 	nvme_ctrlr_delete_qpairs(ctrlr);
1882 	nvme_ctrlr_disable_qpairs(ctrlr);
1883 	nvme_ctrlr_shutdown(ctrlr);
1884 
1885 	return (0);
1886 }
1887 
1888 int
nvme_ctrlr_resume(struct nvme_controller * ctrlr)1889 nvme_ctrlr_resume(struct nvme_controller *ctrlr)
1890 {
1891 	/*
1892 	 * Can't touch failed controllers, so nothing to do to resume.
1893 	 */
1894 	if (ctrlr->is_failed)
1895 		return (0);
1896 
1897 	if (nvme_ctrlr_hw_reset(ctrlr) != 0)
1898 		goto fail;
1899 
1900 	/*
1901 	 * Now that we've reset the hardware, we can restart the controller. Any
1902 	 * I/O that was pending is requeued. Any admin commands are aborted with
1903 	 * an error. Once we've restarted, stop flagging the controller as being
1904 	 * in the reset phase.
1905 	 */
1906 	nvme_ctrlr_start(ctrlr, true);
1907 	(void)atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
1908 
1909 	return (0);
1910 fail:
1911 	/*
1912 	 * Since we can't bring the controller out of reset, announce and fail
1913 	 * the controller. However, we have to return success for the resume
1914 	 * itself, due to questionable APIs.
1915 	 */
1916 	nvme_printf(ctrlr, "Failed to reset on resume, failing.\n");
1917 	nvme_ctrlr_fail(ctrlr, true);
1918 	(void)atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
1919 	return (0);
1920 }
1921