xref: /freebsd/sys/dev/nvme/nvme_ctrlr.c (revision 788ca347b816afd83b2885e0c79aeeb88649b2ab)
1 /*-
2  * Copyright (C) 2012-2014 Intel Corporation
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  *
14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24  * SUCH DAMAGE.
25  */
26 
27 #include <sys/cdefs.h>
28 __FBSDID("$FreeBSD$");
29 
30 #include <sys/param.h>
31 #include <sys/systm.h>
32 #include <sys/buf.h>
33 #include <sys/bus.h>
34 #include <sys/conf.h>
35 #include <sys/ioccom.h>
36 #include <sys/proc.h>
37 #include <sys/smp.h>
38 #include <sys/uio.h>
39 
40 #include <dev/pci/pcireg.h>
41 #include <dev/pci/pcivar.h>
42 
43 #include "nvme_private.h"
44 
45 static void nvme_ctrlr_construct_and_submit_aer(struct nvme_controller *ctrlr,
46 						struct nvme_async_event_request *aer);
47 
48 static int
49 nvme_ctrlr_allocate_bar(struct nvme_controller *ctrlr)
50 {
51 
52 	ctrlr->resource_id = PCIR_BAR(0);
53 
54 	ctrlr->resource = bus_alloc_resource(ctrlr->dev, SYS_RES_MEMORY,
55 	    &ctrlr->resource_id, 0, ~0, 1, RF_ACTIVE);
56 
57 	if(ctrlr->resource == NULL) {
58 		nvme_printf(ctrlr, "unable to allocate pci resource\n");
59 		return (ENOMEM);
60 	}
61 
62 	ctrlr->bus_tag = rman_get_bustag(ctrlr->resource);
63 	ctrlr->bus_handle = rman_get_bushandle(ctrlr->resource);
64 	ctrlr->regs = (struct nvme_registers *)ctrlr->bus_handle;
65 
66 	/*
67 	 * The NVMe spec allows for the MSI-X table to be placed behind
68 	 *  BAR 4/5, separate from the control/doorbell registers.  Always
69 	 *  try to map this bar, because it must be mapped prior to calling
70 	 *  pci_alloc_msix().  If the table isn't behind BAR 4/5,
71 	 *  bus_alloc_resource() will just return NULL which is OK.
72 	 */
73 	ctrlr->bar4_resource_id = PCIR_BAR(4);
74 	ctrlr->bar4_resource = bus_alloc_resource(ctrlr->dev, SYS_RES_MEMORY,
75 	    &ctrlr->bar4_resource_id, 0, ~0, 1, RF_ACTIVE);
76 
77 	return (0);
78 }
79 
80 static void
81 nvme_ctrlr_construct_admin_qpair(struct nvme_controller *ctrlr)
82 {
83 	struct nvme_qpair	*qpair;
84 	uint32_t		num_entries;
85 
86 	qpair = &ctrlr->adminq;
87 
88 	num_entries = NVME_ADMIN_ENTRIES;
89 	TUNABLE_INT_FETCH("hw.nvme.admin_entries", &num_entries);
90 	/*
91 	 * If admin_entries was overridden to an invalid value, revert it
92 	 *  back to our default value.
93 	 */
94 	if (num_entries < NVME_MIN_ADMIN_ENTRIES ||
95 	    num_entries > NVME_MAX_ADMIN_ENTRIES) {
96 		nvme_printf(ctrlr, "invalid hw.nvme.admin_entries=%d "
97 		    "specified\n", num_entries);
98 		num_entries = NVME_ADMIN_ENTRIES;
99 	}
100 
101 	/*
102 	 * The admin queue's max xfer size is treated differently than the
103 	 *  max I/O xfer size.  16KB is sufficient here - maybe even less?
104 	 */
105 	nvme_qpair_construct(qpair,
106 			     0, /* qpair ID */
107 			     0, /* vector */
108 			     num_entries,
109 			     NVME_ADMIN_TRACKERS,
110 			     ctrlr);
111 }
112 
113 static int
114 nvme_ctrlr_construct_io_qpairs(struct nvme_controller *ctrlr)
115 {
116 	struct nvme_qpair	*qpair;
117 	union cap_lo_register	cap_lo;
118 	int			i, num_entries, num_trackers;
119 
120 	num_entries = NVME_IO_ENTRIES;
121 	TUNABLE_INT_FETCH("hw.nvme.io_entries", &num_entries);
122 
123 	/*
124 	 * NVMe spec sets a hard limit of 64K max entries, but
125 	 *  devices may specify a smaller limit, so we need to check
126 	 *  the MQES field in the capabilities register.
127 	 */
128 	cap_lo.raw = nvme_mmio_read_4(ctrlr, cap_lo);
129 	num_entries = min(num_entries, cap_lo.bits.mqes+1);
130 
131 	num_trackers = NVME_IO_TRACKERS;
132 	TUNABLE_INT_FETCH("hw.nvme.io_trackers", &num_trackers);
133 
134 	num_trackers = max(num_trackers, NVME_MIN_IO_TRACKERS);
135 	num_trackers = min(num_trackers, NVME_MAX_IO_TRACKERS);
136 	/*
137 	 * No need to have more trackers than entries in the submit queue.
138 	 *  Note also that for a queue size of N, we can only have (N-1)
139 	 *  commands outstanding, hence the "-1" here.
140 	 */
141 	num_trackers = min(num_trackers, (num_entries-1));
142 
143 	ctrlr->ioq = malloc(ctrlr->num_io_queues * sizeof(struct nvme_qpair),
144 	    M_NVME, M_ZERO | M_WAITOK);
145 
146 	for (i = 0; i < ctrlr->num_io_queues; i++) {
147 		qpair = &ctrlr->ioq[i];
148 
149 		/*
150 		 * Admin queue has ID=0. IO queues start at ID=1 -
151 		 *  hence the 'i+1' here.
152 		 *
153 		 * For I/O queues, use the controller-wide max_xfer_size
154 		 *  calculated in nvme_attach().
155 		 */
156 		nvme_qpair_construct(qpair,
157 				     i+1, /* qpair ID */
158 				     ctrlr->msix_enabled ? i+1 : 0, /* vector */
159 				     num_entries,
160 				     num_trackers,
161 				     ctrlr);
162 
163 		if (ctrlr->per_cpu_io_queues)
164 			bus_bind_intr(ctrlr->dev, qpair->res, i);
165 	}
166 
167 	return (0);
168 }
169 
170 static void
171 nvme_ctrlr_fail(struct nvme_controller *ctrlr)
172 {
173 	int i;
174 
175 	ctrlr->is_failed = TRUE;
176 	nvme_qpair_fail(&ctrlr->adminq);
177 	for (i = 0; i < ctrlr->num_io_queues; i++)
178 		nvme_qpair_fail(&ctrlr->ioq[i]);
179 	nvme_notify_fail_consumers(ctrlr);
180 }
181 
182 void
183 nvme_ctrlr_post_failed_request(struct nvme_controller *ctrlr,
184     struct nvme_request *req)
185 {
186 
187 	mtx_lock(&ctrlr->lock);
188 	STAILQ_INSERT_TAIL(&ctrlr->fail_req, req, stailq);
189 	mtx_unlock(&ctrlr->lock);
190 	taskqueue_enqueue(ctrlr->taskqueue, &ctrlr->fail_req_task);
191 }
192 
193 static void
194 nvme_ctrlr_fail_req_task(void *arg, int pending)
195 {
196 	struct nvme_controller	*ctrlr = arg;
197 	struct nvme_request	*req;
198 
199 	mtx_lock(&ctrlr->lock);
200 	while (!STAILQ_EMPTY(&ctrlr->fail_req)) {
201 		req = STAILQ_FIRST(&ctrlr->fail_req);
202 		STAILQ_REMOVE_HEAD(&ctrlr->fail_req, stailq);
203 		nvme_qpair_manual_complete_request(req->qpair, req,
204 		    NVME_SCT_GENERIC, NVME_SC_ABORTED_BY_REQUEST, TRUE);
205 	}
206 	mtx_unlock(&ctrlr->lock);
207 }
208 
209 static int
210 nvme_ctrlr_wait_for_ready(struct nvme_controller *ctrlr)
211 {
212 	int ms_waited;
213 	union cc_register cc;
214 	union csts_register csts;
215 
216 	cc.raw = nvme_mmio_read_4(ctrlr, cc);
217 	csts.raw = nvme_mmio_read_4(ctrlr, csts);
218 
219 	if (!cc.bits.en) {
220 		nvme_printf(ctrlr, "%s called with cc.en = 0\n", __func__);
221 		return (ENXIO);
222 	}
223 
224 	ms_waited = 0;
225 
226 	while (!csts.bits.rdy) {
227 		DELAY(1000);
228 		if (ms_waited++ > ctrlr->ready_timeout_in_ms) {
229 			nvme_printf(ctrlr, "controller did not become ready "
230 			    "within %d ms\n", ctrlr->ready_timeout_in_ms);
231 			return (ENXIO);
232 		}
233 		csts.raw = nvme_mmio_read_4(ctrlr, csts);
234 	}
235 
236 	return (0);
237 }
238 
239 static void
240 nvme_ctrlr_disable(struct nvme_controller *ctrlr)
241 {
242 	union cc_register cc;
243 	union csts_register csts;
244 
245 	cc.raw = nvme_mmio_read_4(ctrlr, cc);
246 	csts.raw = nvme_mmio_read_4(ctrlr, csts);
247 
248 	if (cc.bits.en == 1 && csts.bits.rdy == 0)
249 		nvme_ctrlr_wait_for_ready(ctrlr);
250 
251 	cc.bits.en = 0;
252 	nvme_mmio_write_4(ctrlr, cc, cc.raw);
253 	DELAY(5000);
254 }
255 
256 static int
257 nvme_ctrlr_enable(struct nvme_controller *ctrlr)
258 {
259 	union cc_register	cc;
260 	union csts_register	csts;
261 	union aqa_register	aqa;
262 
263 	cc.raw = nvme_mmio_read_4(ctrlr, cc);
264 	csts.raw = nvme_mmio_read_4(ctrlr, csts);
265 
266 	if (cc.bits.en == 1) {
267 		if (csts.bits.rdy == 1)
268 			return (0);
269 		else
270 			return (nvme_ctrlr_wait_for_ready(ctrlr));
271 	}
272 
273 	nvme_mmio_write_8(ctrlr, asq, ctrlr->adminq.cmd_bus_addr);
274 	DELAY(5000);
275 	nvme_mmio_write_8(ctrlr, acq, ctrlr->adminq.cpl_bus_addr);
276 	DELAY(5000);
277 
278 	aqa.raw = 0;
279 	/* acqs and asqs are 0-based. */
280 	aqa.bits.acqs = ctrlr->adminq.num_entries-1;
281 	aqa.bits.asqs = ctrlr->adminq.num_entries-1;
282 	nvme_mmio_write_4(ctrlr, aqa, aqa.raw);
283 	DELAY(5000);
284 
285 	cc.bits.en = 1;
286 	cc.bits.css = 0;
287 	cc.bits.ams = 0;
288 	cc.bits.shn = 0;
289 	cc.bits.iosqes = 6; /* SQ entry size == 64 == 2^6 */
290 	cc.bits.iocqes = 4; /* CQ entry size == 16 == 2^4 */
291 
292 	/* This evaluates to 0, which is according to spec. */
293 	cc.bits.mps = (PAGE_SIZE >> 13);
294 
295 	nvme_mmio_write_4(ctrlr, cc, cc.raw);
296 	DELAY(5000);
297 
298 	return (nvme_ctrlr_wait_for_ready(ctrlr));
299 }
300 
301 int
302 nvme_ctrlr_hw_reset(struct nvme_controller *ctrlr)
303 {
304 	int i;
305 
306 	nvme_admin_qpair_disable(&ctrlr->adminq);
307 	for (i = 0; i < ctrlr->num_io_queues; i++)
308 		nvme_io_qpair_disable(&ctrlr->ioq[i]);
309 
310 	DELAY(100*1000);
311 
312 	nvme_ctrlr_disable(ctrlr);
313 	return (nvme_ctrlr_enable(ctrlr));
314 }
315 
316 void
317 nvme_ctrlr_reset(struct nvme_controller *ctrlr)
318 {
319 	int cmpset;
320 
321 	cmpset = atomic_cmpset_32(&ctrlr->is_resetting, 0, 1);
322 
323 	if (cmpset == 0 || ctrlr->is_failed)
324 		/*
325 		 * Controller is already resetting or has failed.  Return
326 		 *  immediately since there is no need to kick off another
327 		 *  reset in these cases.
328 		 */
329 		return;
330 
331 	taskqueue_enqueue(ctrlr->taskqueue, &ctrlr->reset_task);
332 }
333 
334 static int
335 nvme_ctrlr_identify(struct nvme_controller *ctrlr)
336 {
337 	struct nvme_completion_poll_status	status;
338 
339 	status.done = FALSE;
340 	nvme_ctrlr_cmd_identify_controller(ctrlr, &ctrlr->cdata,
341 	    nvme_completion_poll_cb, &status);
342 	while (status.done == FALSE)
343 		pause("nvme", 1);
344 	if (nvme_completion_is_error(&status.cpl)) {
345 		nvme_printf(ctrlr, "nvme_identify_controller failed!\n");
346 		return (ENXIO);
347 	}
348 
349 	/*
350 	 * Use MDTS to ensure our default max_xfer_size doesn't exceed what the
351 	 *  controller supports.
352 	 */
353 	if (ctrlr->cdata.mdts > 0)
354 		ctrlr->max_xfer_size = min(ctrlr->max_xfer_size,
355 		    ctrlr->min_page_size * (1 << (ctrlr->cdata.mdts)));
356 
357 	return (0);
358 }
359 
360 static int
361 nvme_ctrlr_set_num_qpairs(struct nvme_controller *ctrlr)
362 {
363 	struct nvme_completion_poll_status	status;
364 	int					cq_allocated, i, sq_allocated;
365 
366 	status.done = FALSE;
367 	nvme_ctrlr_cmd_set_num_queues(ctrlr, ctrlr->num_io_queues,
368 	    nvme_completion_poll_cb, &status);
369 	while (status.done == FALSE)
370 		pause("nvme", 1);
371 	if (nvme_completion_is_error(&status.cpl)) {
372 		nvme_printf(ctrlr, "nvme_set_num_queues failed!\n");
373 		return (ENXIO);
374 	}
375 
376 	/*
377 	 * Data in cdw0 is 0-based.
378 	 * Lower 16-bits indicate number of submission queues allocated.
379 	 * Upper 16-bits indicate number of completion queues allocated.
380 	 */
381 	sq_allocated = (status.cpl.cdw0 & 0xFFFF) + 1;
382 	cq_allocated = (status.cpl.cdw0 >> 16) + 1;
383 
384 	/*
385 	 * Check that the controller was able to allocate the number of
386 	 *  queues we requested.  If not, revert to one IO queue pair.
387 	 */
388 	if (sq_allocated < ctrlr->num_io_queues ||
389 	    cq_allocated < ctrlr->num_io_queues) {
390 
391 		/*
392 		 * Destroy extra IO queue pairs that were created at
393 		 *  controller construction time but are no longer
394 		 *  needed.  This will only happen when a controller
395 		 *  supports fewer queues than MSI-X vectors.  This
396 		 *  is not the normal case, but does occur with the
397 		 *  Chatham prototype board.
398 		 */
399 		for (i = 1; i < ctrlr->num_io_queues; i++)
400 			nvme_io_qpair_destroy(&ctrlr->ioq[i]);
401 
402 		ctrlr->num_io_queues = 1;
403 		ctrlr->per_cpu_io_queues = 0;
404 	}
405 
406 	return (0);
407 }
408 
409 static int
410 nvme_ctrlr_create_qpairs(struct nvme_controller *ctrlr)
411 {
412 	struct nvme_completion_poll_status	status;
413 	struct nvme_qpair			*qpair;
414 	int					i;
415 
416 	for (i = 0; i < ctrlr->num_io_queues; i++) {
417 		qpair = &ctrlr->ioq[i];
418 
419 		status.done = FALSE;
420 		nvme_ctrlr_cmd_create_io_cq(ctrlr, qpair, qpair->vector,
421 		    nvme_completion_poll_cb, &status);
422 		while (status.done == FALSE)
423 			pause("nvme", 1);
424 		if (nvme_completion_is_error(&status.cpl)) {
425 			nvme_printf(ctrlr, "nvme_create_io_cq failed!\n");
426 			return (ENXIO);
427 		}
428 
429 		status.done = FALSE;
430 		nvme_ctrlr_cmd_create_io_sq(qpair->ctrlr, qpair,
431 		    nvme_completion_poll_cb, &status);
432 		while (status.done == FALSE)
433 			pause("nvme", 1);
434 		if (nvme_completion_is_error(&status.cpl)) {
435 			nvme_printf(ctrlr, "nvme_create_io_sq failed!\n");
436 			return (ENXIO);
437 		}
438 	}
439 
440 	return (0);
441 }
442 
443 static int
444 nvme_ctrlr_construct_namespaces(struct nvme_controller *ctrlr)
445 {
446 	struct nvme_namespace	*ns;
447 	int			i, status;
448 
449 	for (i = 0; i < ctrlr->cdata.nn; i++) {
450 		ns = &ctrlr->ns[i];
451 		status = nvme_ns_construct(ns, i+1, ctrlr);
452 		if (status != 0)
453 			return (status);
454 	}
455 
456 	return (0);
457 }
458 
459 static boolean_t
460 is_log_page_id_valid(uint8_t page_id)
461 {
462 
463 	switch (page_id) {
464 	case NVME_LOG_ERROR:
465 	case NVME_LOG_HEALTH_INFORMATION:
466 	case NVME_LOG_FIRMWARE_SLOT:
467 		return (TRUE);
468 	}
469 
470 	return (FALSE);
471 }
472 
473 static uint32_t
474 nvme_ctrlr_get_log_page_size(struct nvme_controller *ctrlr, uint8_t page_id)
475 {
476 	uint32_t	log_page_size;
477 
478 	switch (page_id) {
479 	case NVME_LOG_ERROR:
480 		log_page_size = min(
481 		    sizeof(struct nvme_error_information_entry) *
482 		    ctrlr->cdata.elpe,
483 		    NVME_MAX_AER_LOG_SIZE);
484 		break;
485 	case NVME_LOG_HEALTH_INFORMATION:
486 		log_page_size = sizeof(struct nvme_health_information_page);
487 		break;
488 	case NVME_LOG_FIRMWARE_SLOT:
489 		log_page_size = sizeof(struct nvme_firmware_page);
490 		break;
491 	default:
492 		log_page_size = 0;
493 		break;
494 	}
495 
496 	return (log_page_size);
497 }
498 
499 static void
500 nvme_ctrlr_log_critical_warnings(struct nvme_controller *ctrlr,
501     union nvme_critical_warning_state state)
502 {
503 
504 	if (state.bits.available_spare == 1)
505 		nvme_printf(ctrlr, "available spare space below threshold\n");
506 
507 	if (state.bits.temperature == 1)
508 		nvme_printf(ctrlr, "temperature above threshold\n");
509 
510 	if (state.bits.device_reliability == 1)
511 		nvme_printf(ctrlr, "device reliability degraded\n");
512 
513 	if (state.bits.read_only == 1)
514 		nvme_printf(ctrlr, "media placed in read only mode\n");
515 
516 	if (state.bits.volatile_memory_backup == 1)
517 		nvme_printf(ctrlr, "volatile memory backup device failed\n");
518 
519 	if (state.bits.reserved != 0)
520 		nvme_printf(ctrlr,
521 		    "unknown critical warning(s): state = 0x%02x\n", state.raw);
522 }
523 
524 static void
525 nvme_ctrlr_async_event_log_page_cb(void *arg, const struct nvme_completion *cpl)
526 {
527 	struct nvme_async_event_request		*aer = arg;
528 	struct nvme_health_information_page	*health_info;
529 
530 	/*
531 	 * If the log page fetch for some reason completed with an error,
532 	 *  don't pass log page data to the consumers.  In practice, this case
533 	 *  should never happen.
534 	 */
535 	if (nvme_completion_is_error(cpl))
536 		nvme_notify_async_consumers(aer->ctrlr, &aer->cpl,
537 		    aer->log_page_id, NULL, 0);
538 	else {
539 		if (aer->log_page_id == NVME_LOG_HEALTH_INFORMATION) {
540 			health_info = (struct nvme_health_information_page *)
541 			    aer->log_page_buffer;
542 			nvme_ctrlr_log_critical_warnings(aer->ctrlr,
543 			    health_info->critical_warning);
544 			/*
545 			 * Critical warnings reported through the
546 			 *  SMART/health log page are persistent, so
547 			 *  clear the associated bits in the async event
548 			 *  config so that we do not receive repeated
549 			 *  notifications for the same event.
550 			 */
551 			aer->ctrlr->async_event_config.raw &=
552 			    ~health_info->critical_warning.raw;
553 			nvme_ctrlr_cmd_set_async_event_config(aer->ctrlr,
554 			    aer->ctrlr->async_event_config, NULL, NULL);
555 		}
556 
557 
558 		/*
559 		 * Pass the cpl data from the original async event completion,
560 		 *  not the log page fetch.
561 		 */
562 		nvme_notify_async_consumers(aer->ctrlr, &aer->cpl,
563 		    aer->log_page_id, aer->log_page_buffer, aer->log_page_size);
564 	}
565 
566 	/*
567 	 * Repost another asynchronous event request to replace the one
568 	 *  that just completed.
569 	 */
570 	nvme_ctrlr_construct_and_submit_aer(aer->ctrlr, aer);
571 }
572 
573 static void
574 nvme_ctrlr_async_event_cb(void *arg, const struct nvme_completion *cpl)
575 {
576 	struct nvme_async_event_request	*aer = arg;
577 
578 	if (nvme_completion_is_error(cpl)) {
579 		/*
580 		 *  Do not retry failed async event requests.  This avoids
581 		 *  infinite loops where a new async event request is submitted
582 		 *  to replace the one just failed, only to fail again and
583 		 *  perpetuate the loop.
584 		 */
585 		return;
586 	}
587 
588 	/* Associated log page is in bits 23:16 of completion entry dw0. */
589 	aer->log_page_id = (cpl->cdw0 & 0xFF0000) >> 16;
590 
591 	nvme_printf(aer->ctrlr, "async event occurred (log page id=0x%x)\n",
592 	    aer->log_page_id);
593 
594 	if (is_log_page_id_valid(aer->log_page_id)) {
595 		aer->log_page_size = nvme_ctrlr_get_log_page_size(aer->ctrlr,
596 		    aer->log_page_id);
597 		memcpy(&aer->cpl, cpl, sizeof(*cpl));
598 		nvme_ctrlr_cmd_get_log_page(aer->ctrlr, aer->log_page_id,
599 		    NVME_GLOBAL_NAMESPACE_TAG, aer->log_page_buffer,
600 		    aer->log_page_size, nvme_ctrlr_async_event_log_page_cb,
601 		    aer);
602 		/* Wait to notify consumers until after log page is fetched. */
603 	} else {
604 		nvme_notify_async_consumers(aer->ctrlr, cpl, aer->log_page_id,
605 		    NULL, 0);
606 
607 		/*
608 		 * Repost another asynchronous event request to replace the one
609 		 *  that just completed.
610 		 */
611 		nvme_ctrlr_construct_and_submit_aer(aer->ctrlr, aer);
612 	}
613 }
614 
615 static void
616 nvme_ctrlr_construct_and_submit_aer(struct nvme_controller *ctrlr,
617     struct nvme_async_event_request *aer)
618 {
619 	struct nvme_request *req;
620 
621 	aer->ctrlr = ctrlr;
622 	req = nvme_allocate_request_null(nvme_ctrlr_async_event_cb, aer);
623 	aer->req = req;
624 
625 	/*
626 	 * Disable timeout here, since asynchronous event requests should by
627 	 *  nature never be timed out.
628 	 */
629 	req->timeout = FALSE;
630 	req->cmd.opc = NVME_OPC_ASYNC_EVENT_REQUEST;
631 	nvme_ctrlr_submit_admin_request(ctrlr, req);
632 }
633 
634 static void
635 nvme_ctrlr_configure_aer(struct nvme_controller *ctrlr)
636 {
637 	struct nvme_completion_poll_status	status;
638 	struct nvme_async_event_request		*aer;
639 	uint32_t				i;
640 
641 	ctrlr->async_event_config.raw = 0xFF;
642 	ctrlr->async_event_config.bits.reserved = 0;
643 
644 	status.done = FALSE;
645 	nvme_ctrlr_cmd_get_feature(ctrlr, NVME_FEAT_TEMPERATURE_THRESHOLD,
646 	    0, NULL, 0, nvme_completion_poll_cb, &status);
647 	while (status.done == FALSE)
648 		pause("nvme", 1);
649 	if (nvme_completion_is_error(&status.cpl) ||
650 	    (status.cpl.cdw0 & 0xFFFF) == 0xFFFF ||
651 	    (status.cpl.cdw0 & 0xFFFF) == 0x0000) {
652 		nvme_printf(ctrlr, "temperature threshold not supported\n");
653 		ctrlr->async_event_config.bits.temperature = 0;
654 	}
655 
656 	nvme_ctrlr_cmd_set_async_event_config(ctrlr,
657 	    ctrlr->async_event_config, NULL, NULL);
658 
659 	/* aerl is a zero-based value, so we need to add 1 here. */
660 	ctrlr->num_aers = min(NVME_MAX_ASYNC_EVENTS, (ctrlr->cdata.aerl+1));
661 
662 	for (i = 0; i < ctrlr->num_aers; i++) {
663 		aer = &ctrlr->aer[i];
664 		nvme_ctrlr_construct_and_submit_aer(ctrlr, aer);
665 	}
666 }
667 
668 static void
669 nvme_ctrlr_configure_int_coalescing(struct nvme_controller *ctrlr)
670 {
671 
672 	ctrlr->int_coal_time = 0;
673 	TUNABLE_INT_FETCH("hw.nvme.int_coal_time",
674 	    &ctrlr->int_coal_time);
675 
676 	ctrlr->int_coal_threshold = 0;
677 	TUNABLE_INT_FETCH("hw.nvme.int_coal_threshold",
678 	    &ctrlr->int_coal_threshold);
679 
680 	nvme_ctrlr_cmd_set_interrupt_coalescing(ctrlr, ctrlr->int_coal_time,
681 	    ctrlr->int_coal_threshold, NULL, NULL);
682 }
683 
684 static void
685 nvme_ctrlr_start(void *ctrlr_arg)
686 {
687 	struct nvme_controller *ctrlr = ctrlr_arg;
688 	int i;
689 
690 	nvme_qpair_reset(&ctrlr->adminq);
691 	for (i = 0; i < ctrlr->num_io_queues; i++)
692 		nvme_qpair_reset(&ctrlr->ioq[i]);
693 
694 	nvme_admin_qpair_enable(&ctrlr->adminq);
695 
696 	if (nvme_ctrlr_identify(ctrlr) != 0) {
697 		nvme_ctrlr_fail(ctrlr);
698 		return;
699 	}
700 
701 	if (nvme_ctrlr_set_num_qpairs(ctrlr) != 0) {
702 		nvme_ctrlr_fail(ctrlr);
703 		return;
704 	}
705 
706 	if (nvme_ctrlr_create_qpairs(ctrlr) != 0) {
707 		nvme_ctrlr_fail(ctrlr);
708 		return;
709 	}
710 
711 	if (nvme_ctrlr_construct_namespaces(ctrlr) != 0) {
712 		nvme_ctrlr_fail(ctrlr);
713 		return;
714 	}
715 
716 	nvme_ctrlr_configure_aer(ctrlr);
717 	nvme_ctrlr_configure_int_coalescing(ctrlr);
718 
719 	for (i = 0; i < ctrlr->num_io_queues; i++)
720 		nvme_io_qpair_enable(&ctrlr->ioq[i]);
721 }
722 
723 void
724 nvme_ctrlr_start_config_hook(void *arg)
725 {
726 	struct nvme_controller *ctrlr = arg;
727 
728 	nvme_ctrlr_start(ctrlr);
729 	config_intrhook_disestablish(&ctrlr->config_hook);
730 
731 	ctrlr->is_initialized = 1;
732 	nvme_notify_new_controller(ctrlr);
733 }
734 
735 static void
736 nvme_ctrlr_reset_task(void *arg, int pending)
737 {
738 	struct nvme_controller	*ctrlr = arg;
739 	int			status;
740 
741 	nvme_printf(ctrlr, "resetting controller\n");
742 	status = nvme_ctrlr_hw_reset(ctrlr);
743 	/*
744 	 * Use pause instead of DELAY, so that we yield to any nvme interrupt
745 	 *  handlers on this CPU that were blocked on a qpair lock. We want
746 	 *  all nvme interrupts completed before proceeding with restarting the
747 	 *  controller.
748 	 *
749 	 * XXX - any way to guarantee the interrupt handlers have quiesced?
750 	 */
751 	pause("nvmereset", hz / 10);
752 	if (status == 0)
753 		nvme_ctrlr_start(ctrlr);
754 	else
755 		nvme_ctrlr_fail(ctrlr);
756 
757 	atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
758 }
759 
760 static void
761 nvme_ctrlr_intx_handler(void *arg)
762 {
763 	struct nvme_controller *ctrlr = arg;
764 
765 	nvme_mmio_write_4(ctrlr, intms, 1);
766 
767 	nvme_qpair_process_completions(&ctrlr->adminq);
768 
769 	if (ctrlr->ioq[0].cpl)
770 		nvme_qpair_process_completions(&ctrlr->ioq[0]);
771 
772 	nvme_mmio_write_4(ctrlr, intmc, 1);
773 }
774 
775 static int
776 nvme_ctrlr_configure_intx(struct nvme_controller *ctrlr)
777 {
778 
779 	ctrlr->num_io_queues = 1;
780 	ctrlr->per_cpu_io_queues = 0;
781 	ctrlr->rid = 0;
782 	ctrlr->res = bus_alloc_resource_any(ctrlr->dev, SYS_RES_IRQ,
783 	    &ctrlr->rid, RF_SHAREABLE | RF_ACTIVE);
784 
785 	if (ctrlr->res == NULL) {
786 		nvme_printf(ctrlr, "unable to allocate shared IRQ\n");
787 		return (ENOMEM);
788 	}
789 
790 	bus_setup_intr(ctrlr->dev, ctrlr->res,
791 	    INTR_TYPE_MISC | INTR_MPSAFE, NULL, nvme_ctrlr_intx_handler,
792 	    ctrlr, &ctrlr->tag);
793 
794 	if (ctrlr->tag == NULL) {
795 		nvme_printf(ctrlr, "unable to setup intx handler\n");
796 		return (ENOMEM);
797 	}
798 
799 	return (0);
800 }
801 
802 static void
803 nvme_pt_done(void *arg, const struct nvme_completion *cpl)
804 {
805 	struct nvme_pt_command *pt = arg;
806 
807 	bzero(&pt->cpl, sizeof(pt->cpl));
808 	pt->cpl.cdw0 = cpl->cdw0;
809 	pt->cpl.status = cpl->status;
810 	pt->cpl.status.p = 0;
811 
812 	mtx_lock(pt->driver_lock);
813 	wakeup(pt);
814 	mtx_unlock(pt->driver_lock);
815 }
816 
817 int
818 nvme_ctrlr_passthrough_cmd(struct nvme_controller *ctrlr,
819     struct nvme_pt_command *pt, uint32_t nsid, int is_user_buffer,
820     int is_admin_cmd)
821 {
822 	struct nvme_request	*req;
823 	struct mtx		*mtx;
824 	struct buf		*buf = NULL;
825 	int			ret = 0;
826 
827 	if (pt->len > 0) {
828 		if (pt->len > ctrlr->max_xfer_size) {
829 			nvme_printf(ctrlr, "pt->len (%d) "
830 			    "exceeds max_xfer_size (%d)\n", pt->len,
831 			    ctrlr->max_xfer_size);
832 			return EIO;
833 		}
834 		if (is_user_buffer) {
835 			/*
836 			 * Ensure the user buffer is wired for the duration of
837 			 *  this passthrough command.
838 			 */
839 			PHOLD(curproc);
840 			buf = getpbuf(NULL);
841 			buf->b_saveaddr = buf->b_data;
842 			buf->b_data = pt->buf;
843 			buf->b_bufsize = pt->len;
844 			buf->b_iocmd = pt->is_read ? BIO_READ : BIO_WRITE;
845 #ifdef NVME_UNMAPPED_BIO_SUPPORT
846 			if (vmapbuf(buf, 1) < 0) {
847 #else
848 			if (vmapbuf(buf) < 0) {
849 #endif
850 				ret = EFAULT;
851 				goto err;
852 			}
853 			req = nvme_allocate_request_vaddr(buf->b_data, pt->len,
854 			    nvme_pt_done, pt);
855 		} else
856 			req = nvme_allocate_request_vaddr(pt->buf, pt->len,
857 			    nvme_pt_done, pt);
858 	} else
859 		req = nvme_allocate_request_null(nvme_pt_done, pt);
860 
861 	req->cmd.opc	= pt->cmd.opc;
862 	req->cmd.cdw10	= pt->cmd.cdw10;
863 	req->cmd.cdw11	= pt->cmd.cdw11;
864 	req->cmd.cdw12	= pt->cmd.cdw12;
865 	req->cmd.cdw13	= pt->cmd.cdw13;
866 	req->cmd.cdw14	= pt->cmd.cdw14;
867 	req->cmd.cdw15	= pt->cmd.cdw15;
868 
869 	req->cmd.nsid = nsid;
870 
871 	if (is_admin_cmd)
872 		mtx = &ctrlr->lock;
873 	else
874 		mtx = &ctrlr->ns[nsid-1].lock;
875 
876 	mtx_lock(mtx);
877 	pt->driver_lock = mtx;
878 
879 	if (is_admin_cmd)
880 		nvme_ctrlr_submit_admin_request(ctrlr, req);
881 	else
882 		nvme_ctrlr_submit_io_request(ctrlr, req);
883 
884 	mtx_sleep(pt, mtx, PRIBIO, "nvme_pt", 0);
885 	mtx_unlock(mtx);
886 
887 	pt->driver_lock = NULL;
888 
889 err:
890 	if (buf != NULL) {
891 		relpbuf(buf, NULL);
892 		PRELE(curproc);
893 	}
894 
895 	return (ret);
896 }
897 
898 static int
899 nvme_ctrlr_ioctl(struct cdev *cdev, u_long cmd, caddr_t arg, int flag,
900     struct thread *td)
901 {
902 	struct nvme_controller			*ctrlr;
903 	struct nvme_pt_command			*pt;
904 
905 	ctrlr = cdev->si_drv1;
906 
907 	switch (cmd) {
908 	case NVME_RESET_CONTROLLER:
909 		nvme_ctrlr_reset(ctrlr);
910 		break;
911 	case NVME_PASSTHROUGH_CMD:
912 		pt = (struct nvme_pt_command *)arg;
913 		return (nvme_ctrlr_passthrough_cmd(ctrlr, pt, pt->cmd.nsid,
914 		    1 /* is_user_buffer */, 1 /* is_admin_cmd */));
915 	default:
916 		return (ENOTTY);
917 	}
918 
919 	return (0);
920 }
921 
922 static struct cdevsw nvme_ctrlr_cdevsw = {
923 	.d_version =	D_VERSION,
924 	.d_flags =	0,
925 	.d_ioctl =	nvme_ctrlr_ioctl
926 };
927 
928 int
929 nvme_ctrlr_construct(struct nvme_controller *ctrlr, device_t dev)
930 {
931 	union cap_lo_register	cap_lo;
932 	union cap_hi_register	cap_hi;
933 	int			i, num_vectors, per_cpu_io_queues, rid;
934 	int			status, timeout_period;
935 
936 	ctrlr->dev = dev;
937 
938 	mtx_init(&ctrlr->lock, "nvme ctrlr lock", NULL, MTX_DEF);
939 
940 	status = nvme_ctrlr_allocate_bar(ctrlr);
941 
942 	if (status != 0)
943 		return (status);
944 
945 	/*
946 	 * Software emulators may set the doorbell stride to something
947 	 *  other than zero, but this driver is not set up to handle that.
948 	 */
949 	cap_hi.raw = nvme_mmio_read_4(ctrlr, cap_hi);
950 	if (cap_hi.bits.dstrd != 0)
951 		return (ENXIO);
952 
953 	ctrlr->min_page_size = 1 << (12 + cap_hi.bits.mpsmin);
954 
955 	/* Get ready timeout value from controller, in units of 500ms. */
956 	cap_lo.raw = nvme_mmio_read_4(ctrlr, cap_lo);
957 	ctrlr->ready_timeout_in_ms = cap_lo.bits.to * 500;
958 
959 	timeout_period = NVME_DEFAULT_TIMEOUT_PERIOD;
960 	TUNABLE_INT_FETCH("hw.nvme.timeout_period", &timeout_period);
961 	timeout_period = min(timeout_period, NVME_MAX_TIMEOUT_PERIOD);
962 	timeout_period = max(timeout_period, NVME_MIN_TIMEOUT_PERIOD);
963 	ctrlr->timeout_period = timeout_period;
964 
965 	nvme_retry_count = NVME_DEFAULT_RETRY_COUNT;
966 	TUNABLE_INT_FETCH("hw.nvme.retry_count", &nvme_retry_count);
967 
968 	per_cpu_io_queues = 1;
969 	TUNABLE_INT_FETCH("hw.nvme.per_cpu_io_queues", &per_cpu_io_queues);
970 	ctrlr->per_cpu_io_queues = per_cpu_io_queues ? TRUE : FALSE;
971 
972 	if (ctrlr->per_cpu_io_queues)
973 		ctrlr->num_io_queues = mp_ncpus;
974 	else
975 		ctrlr->num_io_queues = 1;
976 
977 	ctrlr->force_intx = 0;
978 	TUNABLE_INT_FETCH("hw.nvme.force_intx", &ctrlr->force_intx);
979 
980 	ctrlr->enable_aborts = 0;
981 	TUNABLE_INT_FETCH("hw.nvme.enable_aborts", &ctrlr->enable_aborts);
982 
983 	ctrlr->msix_enabled = 1;
984 
985 	if (ctrlr->force_intx) {
986 		ctrlr->msix_enabled = 0;
987 		goto intx;
988 	}
989 
990 	/* One vector per IO queue, plus one vector for admin queue. */
991 	num_vectors = ctrlr->num_io_queues + 1;
992 
993 	/*
994 	 * If we cannot even allocate 2 vectors (one for admin, one for
995 	 *  I/O), then revert to INTx.
996 	 */
997 	if (pci_msix_count(dev) < 2) {
998 		ctrlr->msix_enabled = 0;
999 		goto intx;
1000 	} else if (pci_msix_count(dev) < num_vectors) {
1001 		ctrlr->per_cpu_io_queues = FALSE;
1002 		ctrlr->num_io_queues = 1;
1003 		num_vectors = 2; /* one for admin, one for I/O */
1004 	}
1005 
1006 	if (pci_alloc_msix(dev, &num_vectors) != 0) {
1007 		ctrlr->msix_enabled = 0;
1008 		goto intx;
1009 	}
1010 
1011 	/*
1012 	 * On earlier FreeBSD releases, there are reports that
1013 	 *  pci_alloc_msix() can return successfully with all vectors
1014 	 *  requested, but a subsequent bus_alloc_resource_any()
1015 	 *  for one of those vectors fails.  This issue occurs more
1016 	 *  readily with multiple devices using per-CPU vectors.
1017 	 * To workaround this issue, try to allocate the resources now,
1018 	 *  and fall back to INTx if we cannot allocate all of them.
1019 	 *  This issue cannot be reproduced on more recent versions of
1020 	 *  FreeBSD which have increased the maximum number of MSI-X
1021 	 *  vectors, but adding the workaround makes it easier for
1022 	 *  vendors wishing to import this driver into kernels based on
1023 	 *  older versions of FreeBSD.
1024 	 */
1025 	for (i = 0; i < num_vectors; i++) {
1026 		rid = i + 1;
1027 		ctrlr->msi_res[i] = bus_alloc_resource_any(ctrlr->dev,
1028 		    SYS_RES_IRQ, &rid, RF_ACTIVE);
1029 
1030 		if (ctrlr->msi_res[i] == NULL) {
1031 			ctrlr->msix_enabled = 0;
1032 			while (i > 0) {
1033 				i--;
1034 				bus_release_resource(ctrlr->dev,
1035 				    SYS_RES_IRQ,
1036 				    rman_get_rid(ctrlr->msi_res[i]),
1037 				    ctrlr->msi_res[i]);
1038 			}
1039 			pci_release_msi(dev);
1040 			nvme_printf(ctrlr, "could not obtain all MSI-X "
1041 			    "resources, reverting to intx\n");
1042 			break;
1043 		}
1044 	}
1045 
1046 intx:
1047 
1048 	if (!ctrlr->msix_enabled)
1049 		nvme_ctrlr_configure_intx(ctrlr);
1050 
1051 	ctrlr->max_xfer_size = NVME_MAX_XFER_SIZE;
1052 	nvme_ctrlr_construct_admin_qpair(ctrlr);
1053 	status = nvme_ctrlr_construct_io_qpairs(ctrlr);
1054 
1055 	if (status != 0)
1056 		return (status);
1057 
1058 	ctrlr->cdev = make_dev(&nvme_ctrlr_cdevsw, device_get_unit(dev),
1059 	    UID_ROOT, GID_WHEEL, 0600, "nvme%d", device_get_unit(dev));
1060 
1061 	if (ctrlr->cdev == NULL)
1062 		return (ENXIO);
1063 
1064 	ctrlr->cdev->si_drv1 = (void *)ctrlr;
1065 
1066 	ctrlr->taskqueue = taskqueue_create("nvme_taskq", M_WAITOK,
1067 	    taskqueue_thread_enqueue, &ctrlr->taskqueue);
1068 	taskqueue_start_threads(&ctrlr->taskqueue, 1, PI_DISK, "nvme taskq");
1069 
1070 	ctrlr->is_resetting = 0;
1071 	ctrlr->is_initialized = 0;
1072 	ctrlr->notification_sent = 0;
1073 	TASK_INIT(&ctrlr->reset_task, 0, nvme_ctrlr_reset_task, ctrlr);
1074 
1075 	TASK_INIT(&ctrlr->fail_req_task, 0, nvme_ctrlr_fail_req_task, ctrlr);
1076 	STAILQ_INIT(&ctrlr->fail_req);
1077 	ctrlr->is_failed = FALSE;
1078 
1079 	return (0);
1080 }
1081 
1082 void
1083 nvme_ctrlr_destruct(struct nvme_controller *ctrlr, device_t dev)
1084 {
1085 	int				i;
1086 
1087 	/*
1088 	 *  Notify the controller of a shutdown, even though this is due to
1089 	 *   a driver unload, not a system shutdown (this path is not invoked
1090 	 *   during shutdown).  This ensures the controller receives a
1091 	 *   shutdown notification in case the system is shutdown before
1092 	 *   reloading the driver.
1093 	 */
1094 	nvme_ctrlr_shutdown(ctrlr);
1095 
1096 	nvme_ctrlr_disable(ctrlr);
1097 	taskqueue_free(ctrlr->taskqueue);
1098 
1099 	for (i = 0; i < NVME_MAX_NAMESPACES; i++)
1100 		nvme_ns_destruct(&ctrlr->ns[i]);
1101 
1102 	if (ctrlr->cdev)
1103 		destroy_dev(ctrlr->cdev);
1104 
1105 	for (i = 0; i < ctrlr->num_io_queues; i++) {
1106 		nvme_io_qpair_destroy(&ctrlr->ioq[i]);
1107 	}
1108 
1109 	free(ctrlr->ioq, M_NVME);
1110 
1111 	nvme_admin_qpair_destroy(&ctrlr->adminq);
1112 
1113 	if (ctrlr->resource != NULL) {
1114 		bus_release_resource(dev, SYS_RES_MEMORY,
1115 		    ctrlr->resource_id, ctrlr->resource);
1116 	}
1117 
1118 	if (ctrlr->bar4_resource != NULL) {
1119 		bus_release_resource(dev, SYS_RES_MEMORY,
1120 		    ctrlr->bar4_resource_id, ctrlr->bar4_resource);
1121 	}
1122 
1123 	if (ctrlr->tag)
1124 		bus_teardown_intr(ctrlr->dev, ctrlr->res, ctrlr->tag);
1125 
1126 	if (ctrlr->res)
1127 		bus_release_resource(ctrlr->dev, SYS_RES_IRQ,
1128 		    rman_get_rid(ctrlr->res), ctrlr->res);
1129 
1130 	if (ctrlr->msix_enabled)
1131 		pci_release_msi(dev);
1132 }
1133 
1134 void
1135 nvme_ctrlr_shutdown(struct nvme_controller *ctrlr)
1136 {
1137 	union cc_register	cc;
1138 	union csts_register	csts;
1139 	int			ticks = 0;
1140 
1141 	cc.raw = nvme_mmio_read_4(ctrlr, cc);
1142 	cc.bits.shn = NVME_SHN_NORMAL;
1143 	nvme_mmio_write_4(ctrlr, cc, cc.raw);
1144 	csts.raw = nvme_mmio_read_4(ctrlr, csts);
1145 	while ((csts.bits.shst != NVME_SHST_COMPLETE) && (ticks++ < 5*hz)) {
1146 		pause("nvme shn", 1);
1147 		csts.raw = nvme_mmio_read_4(ctrlr, csts);
1148 	}
1149 	if (csts.bits.shst != NVME_SHST_COMPLETE)
1150 		nvme_printf(ctrlr, "did not complete shutdown within 5 seconds "
1151 		    "of notification\n");
1152 }
1153 
1154 void
1155 nvme_ctrlr_submit_admin_request(struct nvme_controller *ctrlr,
1156     struct nvme_request *req)
1157 {
1158 
1159 	nvme_qpair_submit_request(&ctrlr->adminq, req);
1160 }
1161 
1162 void
1163 nvme_ctrlr_submit_io_request(struct nvme_controller *ctrlr,
1164     struct nvme_request *req)
1165 {
1166 	struct nvme_qpair       *qpair;
1167 
1168 	if (ctrlr->per_cpu_io_queues)
1169 		qpair = &ctrlr->ioq[curcpu];
1170 	else
1171 		qpair = &ctrlr->ioq[0];
1172 
1173 	nvme_qpair_submit_request(qpair, req);
1174 }
1175 
1176 device_t
1177 nvme_ctrlr_get_device(struct nvme_controller *ctrlr)
1178 {
1179 
1180 	return (ctrlr->dev);
1181 }
1182 
1183 const struct nvme_controller_data *
1184 nvme_ctrlr_get_data(struct nvme_controller *ctrlr)
1185 {
1186 
1187 	return (&ctrlr->cdata);
1188 }
1189