xref: /freebsd/sys/dev/nvme/nvme_qpair.c (revision 1f4bcc459a76b7aa664f3fd557684cd0ba6da352)
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/bus.h>
32 
33 #include <dev/pci/pcivar.h>
34 
35 #include "nvme_private.h"
36 
37 static void	_nvme_qpair_submit_request(struct nvme_qpair *qpair,
38 					   struct nvme_request *req);
39 
40 struct nvme_opcode_string {
41 
42 	uint16_t	opc;
43 	const char *	str;
44 };
45 
46 static struct nvme_opcode_string admin_opcode[] = {
47 	{ NVME_OPC_DELETE_IO_SQ, "DELETE IO SQ" },
48 	{ NVME_OPC_CREATE_IO_SQ, "CREATE IO SQ" },
49 	{ NVME_OPC_GET_LOG_PAGE, "GET LOG PAGE" },
50 	{ NVME_OPC_DELETE_IO_CQ, "DELETE IO CQ" },
51 	{ NVME_OPC_CREATE_IO_CQ, "CREATE IO CQ" },
52 	{ NVME_OPC_IDENTIFY, "IDENTIFY" },
53 	{ NVME_OPC_ABORT, "ABORT" },
54 	{ NVME_OPC_SET_FEATURES, "SET FEATURES" },
55 	{ NVME_OPC_GET_FEATURES, "GET FEATURES" },
56 	{ NVME_OPC_ASYNC_EVENT_REQUEST, "ASYNC EVENT REQUEST" },
57 	{ NVME_OPC_FIRMWARE_ACTIVATE, "FIRMWARE ACTIVATE" },
58 	{ NVME_OPC_FIRMWARE_IMAGE_DOWNLOAD, "FIRMWARE IMAGE DOWNLOAD" },
59 	{ NVME_OPC_FORMAT_NVM, "FORMAT NVM" },
60 	{ NVME_OPC_SECURITY_SEND, "SECURITY SEND" },
61 	{ NVME_OPC_SECURITY_RECEIVE, "SECURITY RECEIVE" },
62 	{ 0xFFFF, "ADMIN COMMAND" }
63 };
64 
65 static struct nvme_opcode_string io_opcode[] = {
66 	{ NVME_OPC_FLUSH, "FLUSH" },
67 	{ NVME_OPC_WRITE, "WRITE" },
68 	{ NVME_OPC_READ, "READ" },
69 	{ NVME_OPC_WRITE_UNCORRECTABLE, "WRITE UNCORRECTABLE" },
70 	{ NVME_OPC_COMPARE, "COMPARE" },
71 	{ NVME_OPC_DATASET_MANAGEMENT, "DATASET MANAGEMENT" },
72 	{ 0xFFFF, "IO COMMAND" }
73 };
74 
75 static const char *
76 get_admin_opcode_string(uint16_t opc)
77 {
78 	struct nvme_opcode_string *entry;
79 
80 	entry = admin_opcode;
81 
82 	while (entry->opc != 0xFFFF) {
83 		if (entry->opc == opc)
84 			return (entry->str);
85 		entry++;
86 	}
87 	return (entry->str);
88 }
89 
90 static const char *
91 get_io_opcode_string(uint16_t opc)
92 {
93 	struct nvme_opcode_string *entry;
94 
95 	entry = io_opcode;
96 
97 	while (entry->opc != 0xFFFF) {
98 		if (entry->opc == opc)
99 			return (entry->str);
100 		entry++;
101 	}
102 	return (entry->str);
103 }
104 
105 
106 static void
107 nvme_admin_qpair_print_command(struct nvme_qpair *qpair,
108     struct nvme_command *cmd)
109 {
110 
111 	nvme_printf(qpair->ctrlr, "%s (%02x) sqid:%d cid:%d nsid:%x "
112 	    "cdw10:%08x cdw11:%08x\n",
113 	    get_admin_opcode_string(cmd->opc), cmd->opc, qpair->id, cmd->cid,
114 	    cmd->nsid, cmd->cdw10, cmd->cdw11);
115 }
116 
117 static void
118 nvme_io_qpair_print_command(struct nvme_qpair *qpair,
119     struct nvme_command *cmd)
120 {
121 
122 	switch (cmd->opc) {
123 	case NVME_OPC_WRITE:
124 	case NVME_OPC_READ:
125 	case NVME_OPC_WRITE_UNCORRECTABLE:
126 	case NVME_OPC_COMPARE:
127 		nvme_printf(qpair->ctrlr, "%s sqid:%d cid:%d nsid:%d "
128 		    "lba:%llu len:%d\n",
129 		    get_io_opcode_string(cmd->opc), qpair->id, cmd->cid,
130 		    cmd->nsid,
131 		    ((unsigned long long)cmd->cdw11 << 32) + cmd->cdw10,
132 		    (cmd->cdw12 & 0xFFFF) + 1);
133 		break;
134 	case NVME_OPC_FLUSH:
135 	case NVME_OPC_DATASET_MANAGEMENT:
136 		nvme_printf(qpair->ctrlr, "%s sqid:%d cid:%d nsid:%d\n",
137 		    get_io_opcode_string(cmd->opc), qpair->id, cmd->cid,
138 		    cmd->nsid);
139 		break;
140 	default:
141 		nvme_printf(qpair->ctrlr, "%s (%02x) sqid:%d cid:%d nsid:%d\n",
142 		    get_io_opcode_string(cmd->opc), cmd->opc, qpair->id,
143 		    cmd->cid, cmd->nsid);
144 		break;
145 	}
146 }
147 
148 static void
149 nvme_qpair_print_command(struct nvme_qpair *qpair, struct nvme_command *cmd)
150 {
151 	if (qpair->id == 0)
152 		nvme_admin_qpair_print_command(qpair, cmd);
153 	else
154 		nvme_io_qpair_print_command(qpair, cmd);
155 }
156 
157 struct nvme_status_string {
158 
159 	uint16_t	sc;
160 	const char *	str;
161 };
162 
163 static struct nvme_status_string generic_status[] = {
164 	{ NVME_SC_SUCCESS, "SUCCESS" },
165 	{ NVME_SC_INVALID_OPCODE, "INVALID OPCODE" },
166 	{ NVME_SC_INVALID_FIELD, "INVALID_FIELD" },
167 	{ NVME_SC_COMMAND_ID_CONFLICT, "COMMAND ID CONFLICT" },
168 	{ NVME_SC_DATA_TRANSFER_ERROR, "DATA TRANSFER ERROR" },
169 	{ NVME_SC_ABORTED_POWER_LOSS, "ABORTED - POWER LOSS" },
170 	{ NVME_SC_INTERNAL_DEVICE_ERROR, "INTERNAL DEVICE ERROR" },
171 	{ NVME_SC_ABORTED_BY_REQUEST, "ABORTED - BY REQUEST" },
172 	{ NVME_SC_ABORTED_SQ_DELETION, "ABORTED - SQ DELETION" },
173 	{ NVME_SC_ABORTED_FAILED_FUSED, "ABORTED - FAILED FUSED" },
174 	{ NVME_SC_ABORTED_MISSING_FUSED, "ABORTED - MISSING FUSED" },
175 	{ NVME_SC_INVALID_NAMESPACE_OR_FORMAT, "INVALID NAMESPACE OR FORMAT" },
176 	{ NVME_SC_COMMAND_SEQUENCE_ERROR, "COMMAND SEQUENCE ERROR" },
177 	{ NVME_SC_LBA_OUT_OF_RANGE, "LBA OUT OF RANGE" },
178 	{ NVME_SC_CAPACITY_EXCEEDED, "CAPACITY EXCEEDED" },
179 	{ NVME_SC_NAMESPACE_NOT_READY, "NAMESPACE NOT READY" },
180 	{ 0xFFFF, "GENERIC" }
181 };
182 
183 static struct nvme_status_string command_specific_status[] = {
184 	{ NVME_SC_COMPLETION_QUEUE_INVALID, "INVALID COMPLETION QUEUE" },
185 	{ NVME_SC_INVALID_QUEUE_IDENTIFIER, "INVALID QUEUE IDENTIFIER" },
186 	{ NVME_SC_MAXIMUM_QUEUE_SIZE_EXCEEDED, "MAX QUEUE SIZE EXCEEDED" },
187 	{ NVME_SC_ABORT_COMMAND_LIMIT_EXCEEDED, "ABORT CMD LIMIT EXCEEDED" },
188 	{ NVME_SC_ASYNC_EVENT_REQUEST_LIMIT_EXCEEDED, "ASYNC LIMIT EXCEEDED" },
189 	{ NVME_SC_INVALID_FIRMWARE_SLOT, "INVALID FIRMWARE SLOT" },
190 	{ NVME_SC_INVALID_FIRMWARE_IMAGE, "INVALID FIRMWARE IMAGE" },
191 	{ NVME_SC_INVALID_INTERRUPT_VECTOR, "INVALID INTERRUPT VECTOR" },
192 	{ NVME_SC_INVALID_LOG_PAGE, "INVALID LOG PAGE" },
193 	{ NVME_SC_INVALID_FORMAT, "INVALID FORMAT" },
194 	{ NVME_SC_FIRMWARE_REQUIRES_RESET, "FIRMWARE REQUIRES RESET" },
195 	{ NVME_SC_CONFLICTING_ATTRIBUTES, "CONFLICTING ATTRIBUTES" },
196 	{ NVME_SC_INVALID_PROTECTION_INFO, "INVALID PROTECTION INFO" },
197 	{ NVME_SC_ATTEMPTED_WRITE_TO_RO_PAGE, "WRITE TO RO PAGE" },
198 	{ 0xFFFF, "COMMAND SPECIFIC" }
199 };
200 
201 static struct nvme_status_string media_error_status[] = {
202 	{ NVME_SC_WRITE_FAULTS, "WRITE FAULTS" },
203 	{ NVME_SC_UNRECOVERED_READ_ERROR, "UNRECOVERED READ ERROR" },
204 	{ NVME_SC_GUARD_CHECK_ERROR, "GUARD CHECK ERROR" },
205 	{ NVME_SC_APPLICATION_TAG_CHECK_ERROR, "APPLICATION TAG CHECK ERROR" },
206 	{ NVME_SC_REFERENCE_TAG_CHECK_ERROR, "REFERENCE TAG CHECK ERROR" },
207 	{ NVME_SC_COMPARE_FAILURE, "COMPARE FAILURE" },
208 	{ NVME_SC_ACCESS_DENIED, "ACCESS DENIED" },
209 	{ 0xFFFF, "MEDIA ERROR" }
210 };
211 
212 static const char *
213 get_status_string(uint16_t sct, uint16_t sc)
214 {
215 	struct nvme_status_string *entry;
216 
217 	switch (sct) {
218 	case NVME_SCT_GENERIC:
219 		entry = generic_status;
220 		break;
221 	case NVME_SCT_COMMAND_SPECIFIC:
222 		entry = command_specific_status;
223 		break;
224 	case NVME_SCT_MEDIA_ERROR:
225 		entry = media_error_status;
226 		break;
227 	case NVME_SCT_VENDOR_SPECIFIC:
228 		return ("VENDOR SPECIFIC");
229 	default:
230 		return ("RESERVED");
231 	}
232 
233 	while (entry->sc != 0xFFFF) {
234 		if (entry->sc == sc)
235 			return (entry->str);
236 		entry++;
237 	}
238 	return (entry->str);
239 }
240 
241 static void
242 nvme_qpair_print_completion(struct nvme_qpair *qpair,
243     struct nvme_completion *cpl)
244 {
245 	nvme_printf(qpair->ctrlr, "%s (%02x/%02x) sqid:%d cid:%d cdw0:%x\n",
246 	    get_status_string(cpl->status.sct, cpl->status.sc),
247 	    cpl->status.sct, cpl->status.sc, cpl->sqid, cpl->cid, cpl->cdw0);
248 }
249 
250 static boolean_t
251 nvme_completion_is_retry(const struct nvme_completion *cpl)
252 {
253 	/*
254 	 * TODO: spec is not clear how commands that are aborted due
255 	 *  to TLER will be marked.  So for now, it seems
256 	 *  NAMESPACE_NOT_READY is the only case where we should
257 	 *  look at the DNR bit.
258 	 */
259 	switch (cpl->status.sct) {
260 	case NVME_SCT_GENERIC:
261 		switch (cpl->status.sc) {
262 		case NVME_SC_ABORTED_BY_REQUEST:
263 		case NVME_SC_NAMESPACE_NOT_READY:
264 			if (cpl->status.dnr)
265 				return (0);
266 			else
267 				return (1);
268 		case NVME_SC_INVALID_OPCODE:
269 		case NVME_SC_INVALID_FIELD:
270 		case NVME_SC_COMMAND_ID_CONFLICT:
271 		case NVME_SC_DATA_TRANSFER_ERROR:
272 		case NVME_SC_ABORTED_POWER_LOSS:
273 		case NVME_SC_INTERNAL_DEVICE_ERROR:
274 		case NVME_SC_ABORTED_SQ_DELETION:
275 		case NVME_SC_ABORTED_FAILED_FUSED:
276 		case NVME_SC_ABORTED_MISSING_FUSED:
277 		case NVME_SC_INVALID_NAMESPACE_OR_FORMAT:
278 		case NVME_SC_COMMAND_SEQUENCE_ERROR:
279 		case NVME_SC_LBA_OUT_OF_RANGE:
280 		case NVME_SC_CAPACITY_EXCEEDED:
281 		default:
282 			return (0);
283 		}
284 	case NVME_SCT_COMMAND_SPECIFIC:
285 	case NVME_SCT_MEDIA_ERROR:
286 	case NVME_SCT_VENDOR_SPECIFIC:
287 	default:
288 		return (0);
289 	}
290 }
291 
292 static void
293 nvme_qpair_construct_tracker(struct nvme_qpair *qpair, struct nvme_tracker *tr,
294     uint16_t cid)
295 {
296 
297 	bus_dmamap_create(qpair->dma_tag_payload, 0, &tr->payload_dma_map);
298 	bus_dmamap_create(qpair->dma_tag, 0, &tr->prp_dma_map);
299 
300 	bus_dmamap_load(qpair->dma_tag, tr->prp_dma_map, tr->prp,
301 	    sizeof(tr->prp), nvme_single_map, &tr->prp_bus_addr, 0);
302 
303 	callout_init(&tr->timer, 1);
304 	tr->cid = cid;
305 	tr->qpair = qpair;
306 }
307 
308 static void
309 nvme_qpair_complete_tracker(struct nvme_qpair *qpair, struct nvme_tracker *tr,
310     struct nvme_completion *cpl, boolean_t print_on_error)
311 {
312 	struct nvme_request	*req;
313 	boolean_t		retry, error;
314 
315 	req = tr->req;
316 	error = nvme_completion_is_error(cpl);
317 	retry = error && nvme_completion_is_retry(cpl) &&
318 	   req->retries < nvme_retry_count;
319 
320 	if (error && print_on_error) {
321 		nvme_qpair_print_command(qpair, &req->cmd);
322 		nvme_qpair_print_completion(qpair, cpl);
323 	}
324 
325 	qpair->act_tr[cpl->cid] = NULL;
326 
327 	KASSERT(cpl->cid == req->cmd.cid, ("cpl cid does not match cmd cid\n"));
328 
329 	if (req->cb_fn && !retry)
330 		req->cb_fn(req->cb_arg, cpl);
331 
332 	mtx_lock(&qpair->lock);
333 	callout_stop(&tr->timer);
334 
335 	if (retry) {
336 		req->retries++;
337 		nvme_qpair_submit_tracker(qpair, tr);
338 	} else {
339 		if (req->type != NVME_REQUEST_NULL)
340 			bus_dmamap_unload(qpair->dma_tag_payload,
341 			    tr->payload_dma_map);
342 
343 		nvme_free_request(req);
344 		tr->req = NULL;
345 
346 		TAILQ_REMOVE(&qpair->outstanding_tr, tr, tailq);
347 		TAILQ_INSERT_HEAD(&qpair->free_tr, tr, tailq);
348 
349 		/*
350 		 * If the controller is in the middle of resetting, don't
351 		 *  try to submit queued requests here - let the reset logic
352 		 *  handle that instead.
353 		 */
354 		if (!STAILQ_EMPTY(&qpair->queued_req) &&
355 		    !qpair->ctrlr->is_resetting) {
356 			req = STAILQ_FIRST(&qpair->queued_req);
357 			STAILQ_REMOVE_HEAD(&qpair->queued_req, stailq);
358 			_nvme_qpair_submit_request(qpair, req);
359 		}
360 	}
361 
362 	mtx_unlock(&qpair->lock);
363 }
364 
365 static void
366 nvme_qpair_manual_complete_tracker(struct nvme_qpair *qpair,
367     struct nvme_tracker *tr, uint32_t sct, uint32_t sc, uint32_t dnr,
368     boolean_t print_on_error)
369 {
370 	struct nvme_completion	cpl;
371 
372 	memset(&cpl, 0, sizeof(cpl));
373 	cpl.sqid = qpair->id;
374 	cpl.cid = tr->cid;
375 	cpl.status.sct = sct;
376 	cpl.status.sc = sc;
377 	cpl.status.dnr = dnr;
378 	nvme_qpair_complete_tracker(qpair, tr, &cpl, print_on_error);
379 }
380 
381 void
382 nvme_qpair_manual_complete_request(struct nvme_qpair *qpair,
383     struct nvme_request *req, uint32_t sct, uint32_t sc,
384     boolean_t print_on_error)
385 {
386 	struct nvme_completion	cpl;
387 	boolean_t		error;
388 
389 	memset(&cpl, 0, sizeof(cpl));
390 	cpl.sqid = qpair->id;
391 	cpl.status.sct = sct;
392 	cpl.status.sc = sc;
393 
394 	error = nvme_completion_is_error(&cpl);
395 
396 	if (error && print_on_error) {
397 		nvme_qpair_print_command(qpair, &req->cmd);
398 		nvme_qpair_print_completion(qpair, &cpl);
399 	}
400 
401 	if (req->cb_fn)
402 		req->cb_fn(req->cb_arg, &cpl);
403 
404 	nvme_free_request(req);
405 }
406 
407 void
408 nvme_qpair_process_completions(struct nvme_qpair *qpair)
409 {
410 	struct nvme_tracker	*tr;
411 	struct nvme_completion	*cpl;
412 
413 	qpair->num_intr_handler_calls++;
414 
415 	if (!qpair->is_enabled)
416 		/*
417 		 * qpair is not enabled, likely because a controller reset is
418 		 *  is in progress.  Ignore the interrupt - any I/O that was
419 		 *  associated with this interrupt will get retried when the
420 		 *  reset is complete.
421 		 */
422 		return;
423 
424 	while (1) {
425 		cpl = &qpair->cpl[qpair->cq_head];
426 
427 		if (cpl->status.p != qpair->phase)
428 			break;
429 
430 		tr = qpair->act_tr[cpl->cid];
431 
432 		if (tr != NULL) {
433 			nvme_qpair_complete_tracker(qpair, tr, cpl, TRUE);
434 			qpair->sq_head = cpl->sqhd;
435 		} else {
436 			nvme_printf(qpair->ctrlr,
437 			    "cpl does not map to outstanding cmd\n");
438 			nvme_dump_completion(cpl);
439 			KASSERT(0, ("received completion for unknown cmd\n"));
440 		}
441 
442 		if (++qpair->cq_head == qpair->num_entries) {
443 			qpair->cq_head = 0;
444 			qpair->phase = !qpair->phase;
445 		}
446 
447 		nvme_mmio_write_4(qpair->ctrlr, doorbell[qpair->id].cq_hdbl,
448 		    qpair->cq_head);
449 	}
450 }
451 
452 static void
453 nvme_qpair_msix_handler(void *arg)
454 {
455 	struct nvme_qpair *qpair = arg;
456 
457 	nvme_qpair_process_completions(qpair);
458 }
459 
460 void
461 nvme_qpair_construct(struct nvme_qpair *qpair, uint32_t id,
462     uint16_t vector, uint32_t num_entries, uint32_t num_trackers,
463     struct nvme_controller *ctrlr)
464 {
465 	struct nvme_tracker	*tr;
466 	uint32_t		i;
467 	int			err;
468 
469 	qpair->id = id;
470 	qpair->vector = vector;
471 	qpair->num_entries = num_entries;
472 	qpair->num_trackers = num_trackers;
473 	qpair->ctrlr = ctrlr;
474 
475 	if (ctrlr->msix_enabled) {
476 
477 		/*
478 		 * MSI-X vector resource IDs start at 1, so we add one to
479 		 *  the queue's vector to get the corresponding rid to use.
480 		 */
481 		qpair->rid = vector + 1;
482 
483 		qpair->res = bus_alloc_resource_any(ctrlr->dev, SYS_RES_IRQ,
484 		    &qpair->rid, RF_ACTIVE);
485 		bus_setup_intr(ctrlr->dev, qpair->res,
486 		    INTR_TYPE_MISC | INTR_MPSAFE, NULL,
487 		    nvme_qpair_msix_handler, qpair, &qpair->tag);
488 	}
489 
490 	mtx_init(&qpair->lock, "nvme qpair lock", NULL, MTX_DEF);
491 
492 	/* Note: NVMe PRP format is restricted to 4-byte alignment. */
493 	err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev),
494 	    4, PAGE_SIZE, BUS_SPACE_MAXADDR,
495 	    BUS_SPACE_MAXADDR, NULL, NULL, NVME_MAX_XFER_SIZE,
496 	    (NVME_MAX_XFER_SIZE/PAGE_SIZE)+1, PAGE_SIZE, 0,
497 	    NULL, NULL, &qpair->dma_tag_payload);
498 	if (err != 0)
499 		nvme_printf(ctrlr, "payload tag create failed %d\n", err);
500 
501 	err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev),
502 	    4, 0, BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
503 	    BUS_SPACE_MAXSIZE, 1, BUS_SPACE_MAXSIZE, 0,
504 	    NULL, NULL, &qpair->dma_tag);
505 	if (err != 0)
506 		nvme_printf(ctrlr, "tag create failed %d\n", err);
507 
508 	qpair->num_cmds = 0;
509 	qpair->num_intr_handler_calls = 0;
510 
511 	qpair->cmd = contigmalloc(qpair->num_entries *
512 	    sizeof(struct nvme_command), M_NVME, M_ZERO,
513 	    0, BUS_SPACE_MAXADDR, PAGE_SIZE, 0);
514 	qpair->cpl = contigmalloc(qpair->num_entries *
515 	    sizeof(struct nvme_completion), M_NVME, M_ZERO,
516 	    0, BUS_SPACE_MAXADDR, PAGE_SIZE, 0);
517 
518 	err = bus_dmamap_create(qpair->dma_tag, 0, &qpair->cmd_dma_map);
519 	if (err != 0)
520 		nvme_printf(ctrlr, "cmd_dma_map create failed %d\n", err);
521 
522 	err = bus_dmamap_create(qpair->dma_tag, 0, &qpair->cpl_dma_map);
523 	if (err != 0)
524 		nvme_printf(ctrlr, "cpl_dma_map create failed %d\n", err);
525 
526 	bus_dmamap_load(qpair->dma_tag, qpair->cmd_dma_map,
527 	    qpair->cmd, qpair->num_entries * sizeof(struct nvme_command),
528 	    nvme_single_map, &qpair->cmd_bus_addr, 0);
529 	bus_dmamap_load(qpair->dma_tag, qpair->cpl_dma_map,
530 	    qpair->cpl, qpair->num_entries * sizeof(struct nvme_completion),
531 	    nvme_single_map, &qpair->cpl_bus_addr, 0);
532 
533 	qpair->sq_tdbl_off = nvme_mmio_offsetof(doorbell[id].sq_tdbl);
534 	qpair->cq_hdbl_off = nvme_mmio_offsetof(doorbell[id].cq_hdbl);
535 
536 	TAILQ_INIT(&qpair->free_tr);
537 	TAILQ_INIT(&qpair->outstanding_tr);
538 	STAILQ_INIT(&qpair->queued_req);
539 
540 	for (i = 0; i < qpair->num_trackers; i++) {
541 		tr = malloc(sizeof(*tr), M_NVME, M_ZERO | M_WAITOK);
542 		nvme_qpair_construct_tracker(qpair, tr, i);
543 		TAILQ_INSERT_HEAD(&qpair->free_tr, tr, tailq);
544 	}
545 
546 	qpair->act_tr = malloc(sizeof(struct nvme_tracker *) * qpair->num_entries,
547 	    M_NVME, M_ZERO | M_WAITOK);
548 }
549 
550 static void
551 nvme_qpair_destroy(struct nvme_qpair *qpair)
552 {
553 	struct nvme_tracker	*tr;
554 
555 	if (qpair->tag)
556 		bus_teardown_intr(qpair->ctrlr->dev, qpair->res, qpair->tag);
557 
558 	if (qpair->res)
559 		bus_release_resource(qpair->ctrlr->dev, SYS_RES_IRQ,
560 		    rman_get_rid(qpair->res), qpair->res);
561 
562 	if (qpair->cmd) {
563 		bus_dmamap_unload(qpair->dma_tag, qpair->cmd_dma_map);
564 		bus_dmamap_destroy(qpair->dma_tag, qpair->cmd_dma_map);
565 		contigfree(qpair->cmd,
566 		    qpair->num_entries * sizeof(struct nvme_command), M_NVME);
567 	}
568 
569 	if (qpair->cpl) {
570 		bus_dmamap_unload(qpair->dma_tag, qpair->cpl_dma_map);
571 		bus_dmamap_destroy(qpair->dma_tag, qpair->cpl_dma_map);
572 		contigfree(qpair->cpl,
573 		    qpair->num_entries * sizeof(struct nvme_completion),
574 		    M_NVME);
575 	}
576 
577 	if (qpair->dma_tag)
578 		bus_dma_tag_destroy(qpair->dma_tag);
579 
580 	if (qpair->dma_tag_payload)
581 		bus_dma_tag_destroy(qpair->dma_tag_payload);
582 
583 	if (qpair->act_tr)
584 		free(qpair->act_tr, M_NVME);
585 
586 	while (!TAILQ_EMPTY(&qpair->free_tr)) {
587 		tr = TAILQ_FIRST(&qpair->free_tr);
588 		TAILQ_REMOVE(&qpair->free_tr, tr, tailq);
589 		bus_dmamap_destroy(qpair->dma_tag, tr->payload_dma_map);
590 		bus_dmamap_destroy(qpair->dma_tag, tr->prp_dma_map);
591 		free(tr, M_NVME);
592 	}
593 }
594 
595 static void
596 nvme_admin_qpair_abort_aers(struct nvme_qpair *qpair)
597 {
598 	struct nvme_tracker	*tr;
599 
600 	tr = TAILQ_FIRST(&qpair->outstanding_tr);
601 	while (tr != NULL) {
602 		if (tr->req->cmd.opc == NVME_OPC_ASYNC_EVENT_REQUEST) {
603 			nvme_qpair_manual_complete_tracker(qpair, tr,
604 			    NVME_SCT_GENERIC, NVME_SC_ABORTED_SQ_DELETION, 0,
605 			    FALSE);
606 			tr = TAILQ_FIRST(&qpair->outstanding_tr);
607 		} else {
608 			tr = TAILQ_NEXT(tr, tailq);
609 		}
610 	}
611 }
612 
613 void
614 nvme_admin_qpair_destroy(struct nvme_qpair *qpair)
615 {
616 
617 	nvme_admin_qpair_abort_aers(qpair);
618 	nvme_qpair_destroy(qpair);
619 }
620 
621 void
622 nvme_io_qpair_destroy(struct nvme_qpair *qpair)
623 {
624 
625 	nvme_qpair_destroy(qpair);
626 }
627 
628 static void
629 nvme_abort_complete(void *arg, const struct nvme_completion *status)
630 {
631 	struct nvme_tracker	*tr = arg;
632 
633 	/*
634 	 * If cdw0 == 1, the controller was not able to abort the command
635 	 *  we requested.  We still need to check the active tracker array,
636 	 *  to cover race where I/O timed out at same time controller was
637 	 *  completing the I/O.
638 	 */
639 	if (status->cdw0 == 1 && tr->qpair->act_tr[tr->cid] != NULL) {
640 		/*
641 		 * An I/O has timed out, and the controller was unable to
642 		 *  abort it for some reason.  Construct a fake completion
643 		 *  status, and then complete the I/O's tracker manually.
644 		 */
645 		nvme_printf(tr->qpair->ctrlr,
646 		    "abort command failed, aborting command manually\n");
647 		nvme_qpair_manual_complete_tracker(tr->qpair, tr,
648 		    NVME_SCT_GENERIC, NVME_SC_ABORTED_BY_REQUEST, 0, TRUE);
649 	}
650 }
651 
652 static void
653 nvme_timeout(void *arg)
654 {
655 	struct nvme_tracker	*tr = arg;
656 	struct nvme_qpair	*qpair = tr->qpair;
657 	struct nvme_controller	*ctrlr = qpair->ctrlr;
658 	union csts_register	csts;
659 
660 	/* Read csts to get value of cfs - controller fatal status. */
661 	csts.raw = nvme_mmio_read_4(ctrlr, csts);
662 
663 	if (ctrlr->enable_aborts && csts.bits.cfs == 0) {
664 		/*
665 		 * If aborts are enabled, only use them if the controller is
666 		 *  not reporting fatal status.
667 		 */
668 		nvme_ctrlr_cmd_abort(ctrlr, tr->cid, qpair->id,
669 		    nvme_abort_complete, tr);
670 	} else
671 		nvme_ctrlr_reset(ctrlr);
672 }
673 
674 void
675 nvme_qpair_submit_tracker(struct nvme_qpair *qpair, struct nvme_tracker *tr)
676 {
677 	struct nvme_request	*req;
678 	struct nvme_controller	*ctrlr;
679 
680 	mtx_assert(&qpair->lock, MA_OWNED);
681 
682 	req = tr->req;
683 	req->cmd.cid = tr->cid;
684 	qpair->act_tr[tr->cid] = tr;
685 	ctrlr = qpair->ctrlr;
686 
687 	if (req->timeout)
688 #if __FreeBSD_version >= 800030
689 		callout_reset_curcpu(&tr->timer, ctrlr->timeout_period * hz,
690 		    nvme_timeout, tr);
691 #else
692 		callout_reset(&tr->timer, ctrlr->timeout_period * hz,
693 		    nvme_timeout, tr);
694 #endif
695 
696 	/* Copy the command from the tracker to the submission queue. */
697 	memcpy(&qpair->cmd[qpair->sq_tail], &req->cmd, sizeof(req->cmd));
698 
699 	if (++qpair->sq_tail == qpair->num_entries)
700 		qpair->sq_tail = 0;
701 
702 	wmb();
703 	nvme_mmio_write_4(qpair->ctrlr, doorbell[qpair->id].sq_tdbl,
704 	    qpair->sq_tail);
705 
706 	qpair->num_cmds++;
707 }
708 
709 static void
710 nvme_payload_map(void *arg, bus_dma_segment_t *seg, int nseg, int error)
711 {
712 	struct nvme_tracker 	*tr = arg;
713 	uint32_t		cur_nseg;
714 
715 	/*
716 	 * If the mapping operation failed, return immediately.  The caller
717 	 *  is responsible for detecting the error status and failing the
718 	 *  tracker manually.
719 	 */
720 	if (error != 0) {
721 		nvme_printf(tr->qpair->ctrlr,
722 		    "nvme_payload_map err %d\n", error);
723 		return;
724 	}
725 
726 	/*
727 	 * Note that we specified PAGE_SIZE for alignment and max
728 	 *  segment size when creating the bus dma tags.  So here
729 	 *  we can safely just transfer each segment to its
730 	 *  associated PRP entry.
731 	 */
732 	tr->req->cmd.prp1 = seg[0].ds_addr;
733 
734 	if (nseg == 2) {
735 		tr->req->cmd.prp2 = seg[1].ds_addr;
736 	} else if (nseg > 2) {
737 		cur_nseg = 1;
738 		tr->req->cmd.prp2 = (uint64_t)tr->prp_bus_addr;
739 		while (cur_nseg < nseg) {
740 			tr->prp[cur_nseg-1] =
741 			    (uint64_t)seg[cur_nseg].ds_addr;
742 			cur_nseg++;
743 		}
744 	} else {
745 		/*
746 		 * prp2 should not be used by the controller
747 		 *  since there is only one segment, but set
748 		 *  to 0 just to be safe.
749 		 */
750 		tr->req->cmd.prp2 = 0;
751 	}
752 
753 	nvme_qpair_submit_tracker(tr->qpair, tr);
754 }
755 
756 static void
757 _nvme_qpair_submit_request(struct nvme_qpair *qpair, struct nvme_request *req)
758 {
759 	struct nvme_tracker	*tr;
760 	int			err = 0;
761 
762 	mtx_assert(&qpair->lock, MA_OWNED);
763 
764 	tr = TAILQ_FIRST(&qpair->free_tr);
765 	req->qpair = qpair;
766 
767 	if (tr == NULL || !qpair->is_enabled) {
768 		/*
769 		 * No tracker is available, or the qpair is disabled due to
770 		 *  an in-progress controller-level reset or controller
771 		 *  failure.
772 		 */
773 
774 		if (qpair->ctrlr->is_failed) {
775 			/*
776 			 * The controller has failed.  Post the request to a
777 			 *  task where it will be aborted, so that we do not
778 			 *  invoke the request's callback in the context
779 			 *  of the submission.
780 			 */
781 			nvme_ctrlr_post_failed_request(qpair->ctrlr, req);
782 		} else {
783 			/*
784 			 * Put the request on the qpair's request queue to be
785 			 *  processed when a tracker frees up via a command
786 			 *  completion or when the controller reset is
787 			 *  completed.
788 			 */
789 			STAILQ_INSERT_TAIL(&qpair->queued_req, req, stailq);
790 		}
791 		return;
792 	}
793 
794 	TAILQ_REMOVE(&qpair->free_tr, tr, tailq);
795 	TAILQ_INSERT_TAIL(&qpair->outstanding_tr, tr, tailq);
796 	tr->req = req;
797 
798 	switch (req->type) {
799 	case NVME_REQUEST_VADDR:
800 		KASSERT(req->payload_size <= qpair->ctrlr->max_xfer_size,
801 		    ("payload_size (%d) exceeds max_xfer_size (%d)\n",
802 		    req->payload_size, qpair->ctrlr->max_xfer_size));
803 		err = bus_dmamap_load(tr->qpair->dma_tag_payload,
804 		    tr->payload_dma_map, req->u.payload, req->payload_size,
805 		    nvme_payload_map, tr, 0);
806 		if (err != 0)
807 			nvme_printf(qpair->ctrlr,
808 			    "bus_dmamap_load returned 0x%x!\n", err);
809 		break;
810 	case NVME_REQUEST_NULL:
811 		nvme_qpair_submit_tracker(tr->qpair, tr);
812 		break;
813 #ifdef NVME_UNMAPPED_BIO_SUPPORT
814 	case NVME_REQUEST_BIO:
815 		KASSERT(req->u.bio->bio_bcount <= qpair->ctrlr->max_xfer_size,
816 		    ("bio->bio_bcount (%jd) exceeds max_xfer_size (%d)\n",
817 		    (intmax_t)req->u.bio->bio_bcount,
818 		    qpair->ctrlr->max_xfer_size));
819 		err = bus_dmamap_load_bio(tr->qpair->dma_tag_payload,
820 		    tr->payload_dma_map, req->u.bio, nvme_payload_map, tr, 0);
821 		if (err != 0)
822 			nvme_printf(qpair->ctrlr,
823 			    "bus_dmamap_load_bio returned 0x%x!\n", err);
824 		break;
825 #endif
826 	default:
827 		panic("unknown nvme request type 0x%x\n", req->type);
828 		break;
829 	}
830 
831 	if (err != 0) {
832 		/*
833 		 * The dmamap operation failed, so we manually fail the
834 		 *  tracker here with DATA_TRANSFER_ERROR status.
835 		 *
836 		 * nvme_qpair_manual_complete_tracker must not be called
837 		 *  with the qpair lock held.
838 		 */
839 		mtx_unlock(&qpair->lock);
840 		nvme_qpair_manual_complete_tracker(qpair, tr, NVME_SCT_GENERIC,
841 		    NVME_SC_DATA_TRANSFER_ERROR, 1 /* do not retry */, TRUE);
842 		mtx_lock(&qpair->lock);
843 	}
844 }
845 
846 void
847 nvme_qpair_submit_request(struct nvme_qpair *qpair, struct nvme_request *req)
848 {
849 
850 	mtx_lock(&qpair->lock);
851 	_nvme_qpair_submit_request(qpair, req);
852 	mtx_unlock(&qpair->lock);
853 }
854 
855 static void
856 nvme_qpair_enable(struct nvme_qpair *qpair)
857 {
858 
859 	qpair->is_enabled = TRUE;
860 }
861 
862 void
863 nvme_qpair_reset(struct nvme_qpair *qpair)
864 {
865 
866 	qpair->sq_head = qpair->sq_tail = qpair->cq_head = 0;
867 
868 	/*
869 	 * First time through the completion queue, HW will set phase
870 	 *  bit on completions to 1.  So set this to 1 here, indicating
871 	 *  we're looking for a 1 to know which entries have completed.
872 	 *  we'll toggle the bit each time when the completion queue
873 	 *  rolls over.
874 	 */
875 	qpair->phase = 1;
876 
877 	memset(qpair->cmd, 0,
878 	    qpair->num_entries * sizeof(struct nvme_command));
879 	memset(qpair->cpl, 0,
880 	    qpair->num_entries * sizeof(struct nvme_completion));
881 }
882 
883 void
884 nvme_admin_qpair_enable(struct nvme_qpair *qpair)
885 {
886 	struct nvme_tracker		*tr;
887 	struct nvme_tracker		*tr_temp;
888 
889 	/*
890 	 * Manually abort each outstanding admin command.  Do not retry
891 	 *  admin commands found here, since they will be left over from
892 	 *  a controller reset and its likely the context in which the
893 	 *  command was issued no longer applies.
894 	 */
895 	TAILQ_FOREACH_SAFE(tr, &qpair->outstanding_tr, tailq, tr_temp) {
896 		nvme_printf(qpair->ctrlr,
897 		    "aborting outstanding admin command\n");
898 		nvme_qpair_manual_complete_tracker(qpair, tr, NVME_SCT_GENERIC,
899 		    NVME_SC_ABORTED_BY_REQUEST, 1 /* do not retry */, TRUE);
900 	}
901 
902 	nvme_qpair_enable(qpair);
903 }
904 
905 void
906 nvme_io_qpair_enable(struct nvme_qpair *qpair)
907 {
908 	STAILQ_HEAD(, nvme_request)	temp;
909 	struct nvme_tracker		*tr;
910 	struct nvme_tracker		*tr_temp;
911 	struct nvme_request		*req;
912 
913 	/*
914 	 * Manually abort each outstanding I/O.  This normally results in a
915 	 *  retry, unless the retry count on the associated request has
916 	 *  reached its limit.
917 	 */
918 	TAILQ_FOREACH_SAFE(tr, &qpair->outstanding_tr, tailq, tr_temp) {
919 		nvme_printf(qpair->ctrlr, "aborting outstanding i/o\n");
920 		nvme_qpair_manual_complete_tracker(qpair, tr, NVME_SCT_GENERIC,
921 		    NVME_SC_ABORTED_BY_REQUEST, 0, TRUE);
922 	}
923 
924 	mtx_lock(&qpair->lock);
925 
926 	nvme_qpair_enable(qpair);
927 
928 	STAILQ_INIT(&temp);
929 	STAILQ_SWAP(&qpair->queued_req, &temp, nvme_request);
930 
931 	while (!STAILQ_EMPTY(&temp)) {
932 		req = STAILQ_FIRST(&temp);
933 		STAILQ_REMOVE_HEAD(&temp, stailq);
934 		nvme_printf(qpair->ctrlr, "resubmitting queued i/o\n");
935 		nvme_qpair_print_command(qpair, &req->cmd);
936 		_nvme_qpair_submit_request(qpair, req);
937 	}
938 
939 	mtx_unlock(&qpair->lock);
940 }
941 
942 static void
943 nvme_qpair_disable(struct nvme_qpair *qpair)
944 {
945 	struct nvme_tracker *tr;
946 
947 	qpair->is_enabled = FALSE;
948 	mtx_lock(&qpair->lock);
949 	TAILQ_FOREACH(tr, &qpair->outstanding_tr, tailq)
950 		callout_stop(&tr->timer);
951 	mtx_unlock(&qpair->lock);
952 }
953 
954 void
955 nvme_admin_qpair_disable(struct nvme_qpair *qpair)
956 {
957 
958 	nvme_qpair_disable(qpair);
959 	nvme_admin_qpair_abort_aers(qpair);
960 }
961 
962 void
963 nvme_io_qpair_disable(struct nvme_qpair *qpair)
964 {
965 
966 	nvme_qpair_disable(qpair);
967 }
968 
969 void
970 nvme_qpair_fail(struct nvme_qpair *qpair)
971 {
972 	struct nvme_tracker		*tr;
973 	struct nvme_request		*req;
974 
975 	mtx_lock(&qpair->lock);
976 
977 	while (!STAILQ_EMPTY(&qpair->queued_req)) {
978 		req = STAILQ_FIRST(&qpair->queued_req);
979 		STAILQ_REMOVE_HEAD(&qpair->queued_req, stailq);
980 		nvme_printf(qpair->ctrlr, "failing queued i/o\n");
981 		mtx_unlock(&qpair->lock);
982 		nvme_qpair_manual_complete_request(qpair, req, NVME_SCT_GENERIC,
983 		    NVME_SC_ABORTED_BY_REQUEST, TRUE);
984 		mtx_lock(&qpair->lock);
985 	}
986 
987 	/* Manually abort each outstanding I/O. */
988 	while (!TAILQ_EMPTY(&qpair->outstanding_tr)) {
989 		tr = TAILQ_FIRST(&qpair->outstanding_tr);
990 		/*
991 		 * Do not remove the tracker.  The abort_tracker path will
992 		 *  do that for us.
993 		 */
994 		nvme_printf(qpair->ctrlr, "failing outstanding i/o\n");
995 		mtx_unlock(&qpair->lock);
996 		nvme_qpair_manual_complete_tracker(qpair, tr, NVME_SCT_GENERIC,
997 		    NVME_SC_ABORTED_BY_REQUEST, 1 /* do not retry */, TRUE);
998 		mtx_lock(&qpair->lock);
999 	}
1000 
1001 	mtx_unlock(&qpair->lock);
1002 }
1003 
1004