xref: /freebsd/sys/dev/nvme/nvme_qpair.c (revision ba3c1f5972d7b90feb6e6da47905ff2757e0fe57)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (C) 2012-2014 Intel Corporation
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  */
28 
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
31 
32 #include <sys/param.h>
33 #include <sys/bus.h>
34 #include <sys/conf.h>
35 #include <sys/domainset.h>
36 #include <sys/proc.h>
37 
38 #include <dev/pci/pcivar.h>
39 
40 #include "nvme_private.h"
41 
42 typedef enum error_print { ERROR_PRINT_NONE, ERROR_PRINT_NO_RETRY, ERROR_PRINT_ALL } error_print_t;
43 #define DO_NOT_RETRY	1
44 
45 static void	_nvme_qpair_submit_request(struct nvme_qpair *qpair,
46 					   struct nvme_request *req);
47 static void	nvme_qpair_destroy(struct nvme_qpair *qpair);
48 
49 struct nvme_opcode_string {
50 	uint16_t	opc;
51 	const char *	str;
52 };
53 
54 static struct nvme_opcode_string admin_opcode[] = {
55 	{ NVME_OPC_DELETE_IO_SQ, "DELETE IO SQ" },
56 	{ NVME_OPC_CREATE_IO_SQ, "CREATE IO SQ" },
57 	{ NVME_OPC_GET_LOG_PAGE, "GET LOG PAGE" },
58 	{ NVME_OPC_DELETE_IO_CQ, "DELETE IO CQ" },
59 	{ NVME_OPC_CREATE_IO_CQ, "CREATE IO CQ" },
60 	{ NVME_OPC_IDENTIFY, "IDENTIFY" },
61 	{ NVME_OPC_ABORT, "ABORT" },
62 	{ NVME_OPC_SET_FEATURES, "SET FEATURES" },
63 	{ NVME_OPC_GET_FEATURES, "GET FEATURES" },
64 	{ NVME_OPC_ASYNC_EVENT_REQUEST, "ASYNC EVENT REQUEST" },
65 	{ NVME_OPC_NAMESPACE_MANAGEMENT, "NAMESPACE MANAGEMENT" },
66 	{ NVME_OPC_FIRMWARE_ACTIVATE, "FIRMWARE ACTIVATE" },
67 	{ NVME_OPC_FIRMWARE_IMAGE_DOWNLOAD, "FIRMWARE IMAGE DOWNLOAD" },
68 	{ NVME_OPC_DEVICE_SELF_TEST, "DEVICE SELF-TEST" },
69 	{ NVME_OPC_NAMESPACE_ATTACHMENT, "NAMESPACE ATTACHMENT" },
70 	{ NVME_OPC_KEEP_ALIVE, "KEEP ALIVE" },
71 	{ NVME_OPC_DIRECTIVE_SEND, "DIRECTIVE SEND" },
72 	{ NVME_OPC_DIRECTIVE_RECEIVE, "DIRECTIVE RECEIVE" },
73 	{ NVME_OPC_VIRTUALIZATION_MANAGEMENT, "VIRTUALIZATION MANAGEMENT" },
74 	{ NVME_OPC_NVME_MI_SEND, "NVME-MI SEND" },
75 	{ NVME_OPC_NVME_MI_RECEIVE, "NVME-MI RECEIVE" },
76 	{ NVME_OPC_DOORBELL_BUFFER_CONFIG, "DOORBELL BUFFER CONFIG" },
77 	{ NVME_OPC_FORMAT_NVM, "FORMAT NVM" },
78 	{ NVME_OPC_SECURITY_SEND, "SECURITY SEND" },
79 	{ NVME_OPC_SECURITY_RECEIVE, "SECURITY RECEIVE" },
80 	{ NVME_OPC_SANITIZE, "SANITIZE" },
81 	{ NVME_OPC_GET_LBA_STATUS, "GET LBA STATUS" },
82 	{ 0xFFFF, "ADMIN COMMAND" }
83 };
84 
85 static struct nvme_opcode_string io_opcode[] = {
86 	{ NVME_OPC_FLUSH, "FLUSH" },
87 	{ NVME_OPC_WRITE, "WRITE" },
88 	{ NVME_OPC_READ, "READ" },
89 	{ NVME_OPC_WRITE_UNCORRECTABLE, "WRITE UNCORRECTABLE" },
90 	{ NVME_OPC_COMPARE, "COMPARE" },
91 	{ NVME_OPC_WRITE_ZEROES, "WRITE ZEROES" },
92 	{ NVME_OPC_DATASET_MANAGEMENT, "DATASET MANAGEMENT" },
93 	{ NVME_OPC_VERIFY, "VERIFY" },
94 	{ NVME_OPC_RESERVATION_REGISTER, "RESERVATION REGISTER" },
95 	{ NVME_OPC_RESERVATION_REPORT, "RESERVATION REPORT" },
96 	{ NVME_OPC_RESERVATION_ACQUIRE, "RESERVATION ACQUIRE" },
97 	{ NVME_OPC_RESERVATION_RELEASE, "RESERVATION RELEASE" },
98 	{ 0xFFFF, "IO COMMAND" }
99 };
100 
101 static const char *
102 get_admin_opcode_string(uint16_t opc)
103 {
104 	struct nvme_opcode_string *entry;
105 
106 	entry = admin_opcode;
107 
108 	while (entry->opc != 0xFFFF) {
109 		if (entry->opc == opc)
110 			return (entry->str);
111 		entry++;
112 	}
113 	return (entry->str);
114 }
115 
116 static const char *
117 get_io_opcode_string(uint16_t opc)
118 {
119 	struct nvme_opcode_string *entry;
120 
121 	entry = io_opcode;
122 
123 	while (entry->opc != 0xFFFF) {
124 		if (entry->opc == opc)
125 			return (entry->str);
126 		entry++;
127 	}
128 	return (entry->str);
129 }
130 
131 static void
132 nvme_admin_qpair_print_command(struct nvme_qpair *qpair,
133     struct nvme_command *cmd)
134 {
135 
136 	nvme_printf(qpair->ctrlr, "%s (%02x) sqid:%d cid:%d nsid:%x "
137 	    "cdw10:%08x cdw11:%08x\n",
138 	    get_admin_opcode_string(cmd->opc), cmd->opc, qpair->id, cmd->cid,
139 	    le32toh(cmd->nsid), le32toh(cmd->cdw10), le32toh(cmd->cdw11));
140 }
141 
142 static void
143 nvme_io_qpair_print_command(struct nvme_qpair *qpair,
144     struct nvme_command *cmd)
145 {
146 
147 	switch (cmd->opc) {
148 	case NVME_OPC_WRITE:
149 	case NVME_OPC_READ:
150 	case NVME_OPC_WRITE_UNCORRECTABLE:
151 	case NVME_OPC_COMPARE:
152 	case NVME_OPC_WRITE_ZEROES:
153 	case NVME_OPC_VERIFY:
154 		nvme_printf(qpair->ctrlr, "%s sqid:%d cid:%d nsid:%d "
155 		    "lba:%llu len:%d\n",
156 		    get_io_opcode_string(cmd->opc), qpair->id, cmd->cid, le32toh(cmd->nsid),
157 		    ((unsigned long long)le32toh(cmd->cdw11) << 32) + le32toh(cmd->cdw10),
158 		    (le32toh(cmd->cdw12) & 0xFFFF) + 1);
159 		break;
160 	case NVME_OPC_FLUSH:
161 	case NVME_OPC_DATASET_MANAGEMENT:
162 	case NVME_OPC_RESERVATION_REGISTER:
163 	case NVME_OPC_RESERVATION_REPORT:
164 	case NVME_OPC_RESERVATION_ACQUIRE:
165 	case NVME_OPC_RESERVATION_RELEASE:
166 		nvme_printf(qpair->ctrlr, "%s sqid:%d cid:%d nsid:%d\n",
167 		    get_io_opcode_string(cmd->opc), qpair->id, cmd->cid, le32toh(cmd->nsid));
168 		break;
169 	default:
170 		nvme_printf(qpair->ctrlr, "%s (%02x) sqid:%d cid:%d nsid:%d\n",
171 		    get_io_opcode_string(cmd->opc), cmd->opc, qpair->id,
172 		    cmd->cid, le32toh(cmd->nsid));
173 		break;
174 	}
175 }
176 
177 static void
178 nvme_qpair_print_command(struct nvme_qpair *qpair, struct nvme_command *cmd)
179 {
180 	if (qpair->id == 0)
181 		nvme_admin_qpair_print_command(qpair, cmd);
182 	else
183 		nvme_io_qpair_print_command(qpair, cmd);
184 	if (nvme_verbose_cmd_dump) {
185 		nvme_printf(qpair->ctrlr,
186 		    "nsid:%#x rsvd2:%#x rsvd3:%#x mptr:%#jx prp1:%#jx prp2:%#jx\n",
187 		    cmd->nsid, cmd->rsvd2, cmd->rsvd3, (uintmax_t)cmd->mptr,
188 		    (uintmax_t)cmd->prp1, (uintmax_t)cmd->prp2);
189 		nvme_printf(qpair->ctrlr,
190 		    "cdw10: %#x cdw11:%#x cdw12:%#x cdw13:%#x cdw14:%#x cdw15:%#x\n",
191 		    cmd->cdw10, cmd->cdw11, cmd->cdw12, cmd->cdw13, cmd->cdw14,
192 		    cmd->cdw15);
193 	}
194 }
195 
196 struct nvme_status_string {
197 	uint16_t	sc;
198 	const char *	str;
199 };
200 
201 static struct nvme_status_string generic_status[] = {
202 	{ NVME_SC_SUCCESS, "SUCCESS" },
203 	{ NVME_SC_INVALID_OPCODE, "INVALID OPCODE" },
204 	{ NVME_SC_INVALID_FIELD, "INVALID_FIELD" },
205 	{ NVME_SC_COMMAND_ID_CONFLICT, "COMMAND ID CONFLICT" },
206 	{ NVME_SC_DATA_TRANSFER_ERROR, "DATA TRANSFER ERROR" },
207 	{ NVME_SC_ABORTED_POWER_LOSS, "ABORTED - POWER LOSS" },
208 	{ NVME_SC_INTERNAL_DEVICE_ERROR, "INTERNAL DEVICE ERROR" },
209 	{ NVME_SC_ABORTED_BY_REQUEST, "ABORTED - BY REQUEST" },
210 	{ NVME_SC_ABORTED_SQ_DELETION, "ABORTED - SQ DELETION" },
211 	{ NVME_SC_ABORTED_FAILED_FUSED, "ABORTED - FAILED FUSED" },
212 	{ NVME_SC_ABORTED_MISSING_FUSED, "ABORTED - MISSING FUSED" },
213 	{ NVME_SC_INVALID_NAMESPACE_OR_FORMAT, "INVALID NAMESPACE OR FORMAT" },
214 	{ NVME_SC_COMMAND_SEQUENCE_ERROR, "COMMAND SEQUENCE ERROR" },
215 	{ NVME_SC_INVALID_SGL_SEGMENT_DESCR, "INVALID SGL SEGMENT DESCRIPTOR" },
216 	{ NVME_SC_INVALID_NUMBER_OF_SGL_DESCR, "INVALID NUMBER OF SGL DESCRIPTORS" },
217 	{ NVME_SC_DATA_SGL_LENGTH_INVALID, "DATA SGL LENGTH INVALID" },
218 	{ NVME_SC_METADATA_SGL_LENGTH_INVALID, "METADATA SGL LENGTH INVALID" },
219 	{ NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID, "SGL DESCRIPTOR TYPE INVALID" },
220 	{ NVME_SC_INVALID_USE_OF_CMB, "INVALID USE OF CONTROLLER MEMORY BUFFER" },
221 	{ NVME_SC_PRP_OFFET_INVALID, "PRP OFFET INVALID" },
222 	{ NVME_SC_ATOMIC_WRITE_UNIT_EXCEEDED, "ATOMIC WRITE UNIT EXCEEDED" },
223 	{ NVME_SC_OPERATION_DENIED, "OPERATION DENIED" },
224 	{ NVME_SC_SGL_OFFSET_INVALID, "SGL OFFSET INVALID" },
225 	{ NVME_SC_HOST_ID_INCONSISTENT_FORMAT, "HOST IDENTIFIER INCONSISTENT FORMAT" },
226 	{ NVME_SC_KEEP_ALIVE_TIMEOUT_EXPIRED, "KEEP ALIVE TIMEOUT EXPIRED" },
227 	{ NVME_SC_KEEP_ALIVE_TIMEOUT_INVALID, "KEEP ALIVE TIMEOUT INVALID" },
228 	{ NVME_SC_ABORTED_DUE_TO_PREEMPT, "COMMAND ABORTED DUE TO PREEMPT AND ABORT" },
229 	{ NVME_SC_SANITIZE_FAILED, "SANITIZE FAILED" },
230 	{ NVME_SC_SANITIZE_IN_PROGRESS, "SANITIZE IN PROGRESS" },
231 	{ NVME_SC_SGL_DATA_BLOCK_GRAN_INVALID, "SGL_DATA_BLOCK_GRANULARITY_INVALID" },
232 	{ NVME_SC_NOT_SUPPORTED_IN_CMB, "COMMAND NOT SUPPORTED FOR QUEUE IN CMB" },
233 	{ NVME_SC_NAMESPACE_IS_WRITE_PROTECTED, "NAMESPACE IS WRITE PROTECTED" },
234 	{ NVME_SC_COMMAND_INTERRUPTED, "COMMAND INTERRUPTED" },
235 	{ NVME_SC_TRANSIENT_TRANSPORT_ERROR, "TRANSIENT TRANSPORT ERROR" },
236 
237 	{ NVME_SC_LBA_OUT_OF_RANGE, "LBA OUT OF RANGE" },
238 	{ NVME_SC_CAPACITY_EXCEEDED, "CAPACITY EXCEEDED" },
239 	{ NVME_SC_NAMESPACE_NOT_READY, "NAMESPACE NOT READY" },
240 	{ NVME_SC_RESERVATION_CONFLICT, "RESERVATION CONFLICT" },
241 	{ NVME_SC_FORMAT_IN_PROGRESS, "FORMAT IN PROGRESS" },
242 	{ 0xFFFF, "GENERIC" }
243 };
244 
245 static struct nvme_status_string command_specific_status[] = {
246 	{ NVME_SC_COMPLETION_QUEUE_INVALID, "INVALID COMPLETION QUEUE" },
247 	{ NVME_SC_INVALID_QUEUE_IDENTIFIER, "INVALID QUEUE IDENTIFIER" },
248 	{ NVME_SC_MAXIMUM_QUEUE_SIZE_EXCEEDED, "MAX QUEUE SIZE EXCEEDED" },
249 	{ NVME_SC_ABORT_COMMAND_LIMIT_EXCEEDED, "ABORT CMD LIMIT EXCEEDED" },
250 	{ NVME_SC_ASYNC_EVENT_REQUEST_LIMIT_EXCEEDED, "ASYNC LIMIT EXCEEDED" },
251 	{ NVME_SC_INVALID_FIRMWARE_SLOT, "INVALID FIRMWARE SLOT" },
252 	{ NVME_SC_INVALID_FIRMWARE_IMAGE, "INVALID FIRMWARE IMAGE" },
253 	{ NVME_SC_INVALID_INTERRUPT_VECTOR, "INVALID INTERRUPT VECTOR" },
254 	{ NVME_SC_INVALID_LOG_PAGE, "INVALID LOG PAGE" },
255 	{ NVME_SC_INVALID_FORMAT, "INVALID FORMAT" },
256 	{ NVME_SC_FIRMWARE_REQUIRES_RESET, "FIRMWARE REQUIRES RESET" },
257 	{ NVME_SC_INVALID_QUEUE_DELETION, "INVALID QUEUE DELETION" },
258 	{ NVME_SC_FEATURE_NOT_SAVEABLE, "FEATURE IDENTIFIER NOT SAVEABLE" },
259 	{ NVME_SC_FEATURE_NOT_CHANGEABLE, "FEATURE NOT CHANGEABLE" },
260 	{ NVME_SC_FEATURE_NOT_NS_SPECIFIC, "FEATURE NOT NAMESPACE SPECIFIC" },
261 	{ NVME_SC_FW_ACT_REQUIRES_NVMS_RESET, "FIRMWARE ACTIVATION REQUIRES NVM SUBSYSTEM RESET" },
262 	{ NVME_SC_FW_ACT_REQUIRES_RESET, "FIRMWARE ACTIVATION REQUIRES RESET" },
263 	{ NVME_SC_FW_ACT_REQUIRES_TIME, "FIRMWARE ACTIVATION REQUIRES MAXIMUM TIME VIOLATION" },
264 	{ NVME_SC_FW_ACT_PROHIBITED, "FIRMWARE ACTIVATION PROHIBITED" },
265 	{ NVME_SC_OVERLAPPING_RANGE, "OVERLAPPING RANGE" },
266 	{ NVME_SC_NS_INSUFFICIENT_CAPACITY, "NAMESPACE INSUFFICIENT CAPACITY" },
267 	{ NVME_SC_NS_ID_UNAVAILABLE, "NAMESPACE IDENTIFIER UNAVAILABLE" },
268 	{ NVME_SC_NS_ALREADY_ATTACHED, "NAMESPACE ALREADY ATTACHED" },
269 	{ NVME_SC_NS_IS_PRIVATE, "NAMESPACE IS PRIVATE" },
270 	{ NVME_SC_NS_NOT_ATTACHED, "NS NOT ATTACHED" },
271 	{ NVME_SC_THIN_PROV_NOT_SUPPORTED, "THIN PROVISIONING NOT SUPPORTED" },
272 	{ NVME_SC_CTRLR_LIST_INVALID, "CONTROLLER LIST INVALID" },
273 	{ NVME_SC_SELF_TEST_IN_PROGRESS, "DEVICE SELF-TEST IN PROGRESS" },
274 	{ NVME_SC_BOOT_PART_WRITE_PROHIB, "BOOT PARTITION WRITE PROHIBITED" },
275 	{ NVME_SC_INVALID_CTRLR_ID, "INVALID CONTROLLER IDENTIFIER" },
276 	{ NVME_SC_INVALID_SEC_CTRLR_STATE, "INVALID SECONDARY CONTROLLER STATE" },
277 	{ NVME_SC_INVALID_NUM_OF_CTRLR_RESRC, "INVALID NUMBER OF CONTROLLER RESOURCES" },
278 	{ NVME_SC_INVALID_RESOURCE_ID, "INVALID RESOURCE IDENTIFIER" },
279 	{ NVME_SC_SANITIZE_PROHIBITED_WPMRE, "SANITIZE PROHIBITED WRITE PERSISTENT MEMORY REGION ENABLED" },
280 	{ NVME_SC_ANA_GROUP_ID_INVALID, "ANA GROUP IDENTIFIED INVALID" },
281 	{ NVME_SC_ANA_ATTACH_FAILED, "ANA ATTACH FAILED" },
282 
283 	{ NVME_SC_CONFLICTING_ATTRIBUTES, "CONFLICTING ATTRIBUTES" },
284 	{ NVME_SC_INVALID_PROTECTION_INFO, "INVALID PROTECTION INFO" },
285 	{ NVME_SC_ATTEMPTED_WRITE_TO_RO_PAGE, "WRITE TO RO PAGE" },
286 	{ 0xFFFF, "COMMAND SPECIFIC" }
287 };
288 
289 static struct nvme_status_string media_error_status[] = {
290 	{ NVME_SC_WRITE_FAULTS, "WRITE FAULTS" },
291 	{ NVME_SC_UNRECOVERED_READ_ERROR, "UNRECOVERED READ ERROR" },
292 	{ NVME_SC_GUARD_CHECK_ERROR, "GUARD CHECK ERROR" },
293 	{ NVME_SC_APPLICATION_TAG_CHECK_ERROR, "APPLICATION TAG CHECK ERROR" },
294 	{ NVME_SC_REFERENCE_TAG_CHECK_ERROR, "REFERENCE TAG CHECK ERROR" },
295 	{ NVME_SC_COMPARE_FAILURE, "COMPARE FAILURE" },
296 	{ NVME_SC_ACCESS_DENIED, "ACCESS DENIED" },
297 	{ NVME_SC_DEALLOCATED_OR_UNWRITTEN, "DEALLOCATED OR UNWRITTEN LOGICAL BLOCK" },
298 	{ 0xFFFF, "MEDIA ERROR" }
299 };
300 
301 static struct nvme_status_string path_related_status[] = {
302 	{ NVME_SC_INTERNAL_PATH_ERROR, "INTERNAL PATH ERROR" },
303 	{ NVME_SC_ASYMMETRIC_ACCESS_PERSISTENT_LOSS, "ASYMMETRIC ACCESS PERSISTENT LOSS" },
304 	{ NVME_SC_ASYMMETRIC_ACCESS_INACCESSIBLE, "ASYMMETRIC ACCESS INACCESSIBLE" },
305 	{ NVME_SC_ASYMMETRIC_ACCESS_TRANSITION, "ASYMMETRIC ACCESS TRANSITION" },
306 	{ NVME_SC_CONTROLLER_PATHING_ERROR, "CONTROLLER PATHING ERROR" },
307 	{ NVME_SC_HOST_PATHING_ERROR, "HOST PATHING ERROR" },
308 	{ NVME_SC_COMMAND_ABORTED_BY_HOST, "COMMAND ABORTED BY HOST" },
309 	{ 0xFFFF, "PATH RELATED" },
310 };
311 
312 static const char *
313 get_status_string(uint16_t sct, uint16_t sc)
314 {
315 	struct nvme_status_string *entry;
316 
317 	switch (sct) {
318 	case NVME_SCT_GENERIC:
319 		entry = generic_status;
320 		break;
321 	case NVME_SCT_COMMAND_SPECIFIC:
322 		entry = command_specific_status;
323 		break;
324 	case NVME_SCT_MEDIA_ERROR:
325 		entry = media_error_status;
326 		break;
327 	case NVME_SCT_PATH_RELATED:
328 		entry = path_related_status;
329 		break;
330 	case NVME_SCT_VENDOR_SPECIFIC:
331 		return ("VENDOR SPECIFIC");
332 	default:
333 		return ("RESERVED");
334 	}
335 
336 	while (entry->sc != 0xFFFF) {
337 		if (entry->sc == sc)
338 			return (entry->str);
339 		entry++;
340 	}
341 	return (entry->str);
342 }
343 
344 static void
345 nvme_qpair_print_completion(struct nvme_qpair *qpair,
346     struct nvme_completion *cpl)
347 {
348 	uint8_t sct, sc, crd, m, dnr;
349 
350 	sct = NVME_STATUS_GET_SCT(cpl->status);
351 	sc = NVME_STATUS_GET_SC(cpl->status);
352 	crd = NVME_STATUS_GET_CRD(cpl->status);
353 	m = NVME_STATUS_GET_M(cpl->status);
354 	dnr = NVME_STATUS_GET_DNR(cpl->status);
355 
356 	nvme_printf(qpair->ctrlr, "%s (%02x/%02x) crd:%x m:%x dnr:%x "
357 	    "sqid:%d cid:%d cdw0:%x\n",
358 	    get_status_string(sct, sc), sct, sc, crd, m, dnr,
359 	    cpl->sqid, cpl->cid, cpl->cdw0);
360 }
361 
362 static bool
363 nvme_completion_is_retry(const struct nvme_completion *cpl)
364 {
365 	uint8_t sct, sc, dnr;
366 
367 	sct = NVME_STATUS_GET_SCT(cpl->status);
368 	sc = NVME_STATUS_GET_SC(cpl->status);
369 	dnr = NVME_STATUS_GET_DNR(cpl->status);	/* Do Not Retry Bit */
370 
371 	/*
372 	 * TODO: spec is not clear how commands that are aborted due
373 	 *  to TLER will be marked.  So for now, it seems
374 	 *  NAMESPACE_NOT_READY is the only case where we should
375 	 *  look at the DNR bit. Requests failed with ABORTED_BY_REQUEST
376 	 *  set the DNR bit correctly since the driver controls that.
377 	 */
378 	switch (sct) {
379 	case NVME_SCT_GENERIC:
380 		switch (sc) {
381 		case NVME_SC_ABORTED_BY_REQUEST:
382 		case NVME_SC_NAMESPACE_NOT_READY:
383 			if (dnr)
384 				return (0);
385 			else
386 				return (1);
387 		case NVME_SC_INVALID_OPCODE:
388 		case NVME_SC_INVALID_FIELD:
389 		case NVME_SC_COMMAND_ID_CONFLICT:
390 		case NVME_SC_DATA_TRANSFER_ERROR:
391 		case NVME_SC_ABORTED_POWER_LOSS:
392 		case NVME_SC_INTERNAL_DEVICE_ERROR:
393 		case NVME_SC_ABORTED_SQ_DELETION:
394 		case NVME_SC_ABORTED_FAILED_FUSED:
395 		case NVME_SC_ABORTED_MISSING_FUSED:
396 		case NVME_SC_INVALID_NAMESPACE_OR_FORMAT:
397 		case NVME_SC_COMMAND_SEQUENCE_ERROR:
398 		case NVME_SC_LBA_OUT_OF_RANGE:
399 		case NVME_SC_CAPACITY_EXCEEDED:
400 		default:
401 			return (0);
402 		}
403 	case NVME_SCT_COMMAND_SPECIFIC:
404 	case NVME_SCT_MEDIA_ERROR:
405 		return (0);
406 	case NVME_SCT_PATH_RELATED:
407 		switch (sc) {
408 		case NVME_SC_INTERNAL_PATH_ERROR:
409 			if (dnr)
410 				return (0);
411 			else
412 				return (1);
413 		default:
414 			return (0);
415 		}
416 	case NVME_SCT_VENDOR_SPECIFIC:
417 	default:
418 		return (0);
419 	}
420 }
421 
422 static void
423 nvme_qpair_complete_tracker(struct nvme_tracker *tr,
424     struct nvme_completion *cpl, error_print_t print_on_error)
425 {
426 	struct nvme_qpair * qpair = tr->qpair;
427 	struct nvme_request	*req;
428 	bool			retry, error, retriable;
429 
430 	req = tr->req;
431 	error = nvme_completion_is_error(cpl);
432 	retriable = nvme_completion_is_retry(cpl);
433 	retry = error && retriable && req->retries < nvme_retry_count;
434 	if (retry)
435 		qpair->num_retries++;
436 	if (error && req->retries >= nvme_retry_count && retriable)
437 		qpair->num_failures++;
438 
439 	if (error && (print_on_error == ERROR_PRINT_ALL ||
440 		(!retry && print_on_error == ERROR_PRINT_NO_RETRY))) {
441 		nvme_qpair_print_command(qpair, &req->cmd);
442 		nvme_qpair_print_completion(qpair, cpl);
443 	}
444 
445 	qpair->act_tr[cpl->cid] = NULL;
446 
447 	KASSERT(cpl->cid == req->cmd.cid, ("cpl cid does not match cmd cid\n"));
448 
449 	if (!retry) {
450 		if (req->payload_valid) {
451 			bus_dmamap_sync(qpair->dma_tag_payload,
452 			    tr->payload_dma_map,
453 			    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
454 		}
455 		if (req->cb_fn)
456 			req->cb_fn(req->cb_arg, cpl);
457 	}
458 
459 	mtx_lock(&qpair->lock);
460 
461 	if (retry) {
462 		req->retries++;
463 		nvme_qpair_submit_tracker(qpair, tr);
464 	} else {
465 		if (req->payload_valid) {
466 			bus_dmamap_unload(qpair->dma_tag_payload,
467 			    tr->payload_dma_map);
468 		}
469 
470 		nvme_free_request(req);
471 		tr->req = NULL;
472 
473 		TAILQ_REMOVE(&qpair->outstanding_tr, tr, tailq);
474 		TAILQ_INSERT_HEAD(&qpair->free_tr, tr, tailq);
475 
476 		/*
477 		 * If the controller is in the middle of resetting, don't
478 		 *  try to submit queued requests here - let the reset logic
479 		 *  handle that instead.
480 		 */
481 		if (!STAILQ_EMPTY(&qpair->queued_req) &&
482 		    !qpair->ctrlr->is_resetting) {
483 			req = STAILQ_FIRST(&qpair->queued_req);
484 			STAILQ_REMOVE_HEAD(&qpair->queued_req, stailq);
485 			_nvme_qpair_submit_request(qpair, req);
486 		}
487 	}
488 
489 	mtx_unlock(&qpair->lock);
490 }
491 
492 static void
493 nvme_qpair_manual_complete_tracker(
494     struct nvme_tracker *tr, uint32_t sct, uint32_t sc, uint32_t dnr,
495     error_print_t print_on_error)
496 {
497 	struct nvme_completion	cpl;
498 
499 	memset(&cpl, 0, sizeof(cpl));
500 
501 	struct nvme_qpair * qpair = tr->qpair;
502 
503 	cpl.sqid = qpair->id;
504 	cpl.cid = tr->cid;
505 	cpl.status |= (sct & NVME_STATUS_SCT_MASK) << NVME_STATUS_SCT_SHIFT;
506 	cpl.status |= (sc & NVME_STATUS_SC_MASK) << NVME_STATUS_SC_SHIFT;
507 	cpl.status |= (dnr & NVME_STATUS_DNR_MASK) << NVME_STATUS_DNR_SHIFT;
508 	nvme_qpair_complete_tracker(tr, &cpl, print_on_error);
509 }
510 
511 void
512 nvme_qpair_manual_complete_request(struct nvme_qpair *qpair,
513     struct nvme_request *req, uint32_t sct, uint32_t sc)
514 {
515 	struct nvme_completion	cpl;
516 	bool			error;
517 
518 	memset(&cpl, 0, sizeof(cpl));
519 	cpl.sqid = qpair->id;
520 	cpl.status |= (sct & NVME_STATUS_SCT_MASK) << NVME_STATUS_SCT_SHIFT;
521 	cpl.status |= (sc & NVME_STATUS_SC_MASK) << NVME_STATUS_SC_SHIFT;
522 
523 	error = nvme_completion_is_error(&cpl);
524 
525 	if (error) {
526 		nvme_qpair_print_command(qpair, &req->cmd);
527 		nvme_qpair_print_completion(qpair, &cpl);
528 	}
529 
530 	if (req->cb_fn)
531 		req->cb_fn(req->cb_arg, &cpl);
532 
533 	nvme_free_request(req);
534 }
535 
536 bool
537 nvme_qpair_process_completions(struct nvme_qpair *qpair)
538 {
539 	struct nvme_tracker	*tr;
540 	struct nvme_completion	cpl;
541 	int done = 0;
542 	bool in_panic = dumping || SCHEDULER_STOPPED();
543 
544 	/*
545 	 * qpair is not enabled, likely because a controller reset is in
546 	 * progress.  Ignore the interrupt - any I/O that was associated with
547 	 * this interrupt will get retried when the reset is complete. Any
548 	 * pending completions for when we're in startup will be completed
549 	 * as soon as initialization is complete and we start sending commands
550 	 * to the device.
551 	 */
552 	if (qpair->recovery_state != RECOVERY_NONE) {
553 		qpair->num_ignored++;
554 		return (false);
555 	}
556 
557 	/*
558 	 * Sanity check initialization. After we reset the hardware, the phase
559 	 * is defined to be 1. So if we get here with zero prior calls and the
560 	 * phase is 0, it means that we've lost a race between the
561 	 * initialization and the ISR running. With the phase wrong, we'll
562 	 * process a bunch of completions that aren't really completions leading
563 	 * to a KASSERT below.
564 	 */
565 	KASSERT(!(qpair->num_intr_handler_calls == 0 && qpair->phase == 0),
566 	    ("%s: Phase wrong for first interrupt call.",
567 		device_get_nameunit(qpair->ctrlr->dev)));
568 
569 	qpair->num_intr_handler_calls++;
570 
571 	bus_dmamap_sync(qpair->dma_tag, qpair->queuemem_map,
572 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
573 	/*
574 	 * A panic can stop the CPU this routine is running on at any point.  If
575 	 * we're called during a panic, complete the sq_head wrap protocol for
576 	 * the case where we are interrupted just after the increment at 1
577 	 * below, but before we can reset cq_head to zero at 2. Also cope with
578 	 * the case where we do the zero at 2, but may or may not have done the
579 	 * phase adjustment at step 3. The panic machinery flushes all pending
580 	 * memory writes, so we can make these strong ordering assumptions
581 	 * that would otherwise be unwise if we were racing in real time.
582 	 */
583 	if (__predict_false(in_panic)) {
584 		if (qpair->cq_head == qpair->num_entries) {
585 			/*
586 			 * Here we know that we need to zero cq_head and then negate
587 			 * the phase, which hasn't been assigned if cq_head isn't
588 			 * zero due to the atomic_store_rel.
589 			 */
590 			qpair->cq_head = 0;
591 			qpair->phase = !qpair->phase;
592 		} else if (qpair->cq_head == 0) {
593 			/*
594 			 * In this case, we know that the assignment at 2
595 			 * happened below, but we don't know if it 3 happened or
596 			 * not. To do this, we look at the last completion
597 			 * entry and set the phase to the opposite phase
598 			 * that it has. This gets us back in sync
599 			 */
600 			cpl = qpair->cpl[qpair->num_entries - 1];
601 			nvme_completion_swapbytes(&cpl);
602 			qpair->phase = !NVME_STATUS_GET_P(cpl.status);
603 		}
604 	}
605 
606 	while (1) {
607 		uint16_t status;
608 
609 		/*
610 		 * We need to do this dance to avoid a race between the host and
611 		 * the device where the device overtakes the host while the host
612 		 * is reading this record, leaving the status field 'new' and
613 		 * the sqhd and cid fields potentially stale. If the phase
614 		 * doesn't match, that means status hasn't yet been updated and
615 		 * we'll get any pending changes next time. It also means that
616 		 * the phase must be the same the second time. We have to sync
617 		 * before reading to ensure any bouncing completes.
618 		 */
619 		status = le16toh(qpair->cpl[qpair->cq_head].status);
620 		if (NVME_STATUS_GET_P(status) != qpair->phase)
621 			break;
622 
623 		bus_dmamap_sync(qpair->dma_tag, qpair->queuemem_map,
624 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
625 		cpl = qpair->cpl[qpair->cq_head];
626 		nvme_completion_swapbytes(&cpl);
627 
628 		KASSERT(
629 		    NVME_STATUS_GET_P(status) == NVME_STATUS_GET_P(cpl.status),
630 		    ("Phase unexpectedly inconsistent"));
631 
632 		if (cpl.cid < qpair->num_trackers)
633 			tr = qpair->act_tr[cpl.cid];
634 		else
635 			tr = NULL;
636 
637 		done++;
638 		if (tr != NULL) {
639 			nvme_qpair_complete_tracker(tr, &cpl, ERROR_PRINT_ALL);
640 			qpair->sq_head = cpl.sqhd;
641 		} else if (!in_panic) {
642 			/*
643 			 * A missing tracker is normally an error.  However, a
644 			 * panic can stop the CPU this routine is running on
645 			 * after completing an I/O but before updating
646 			 * qpair->cq_head at 1 below.  Later, we re-enter this
647 			 * routine to poll I/O associated with the kernel
648 			 * dump. We find that the tr has been set to null before
649 			 * calling the completion routine.  If it hasn't
650 			 * completed (or it triggers a panic), then '1' below
651 			 * won't have updated cq_head. Rather than panic again,
652 			 * ignore this condition because it's not unexpected.
653 			 */
654 			nvme_printf(qpair->ctrlr,
655 			    "cpl (cid = %u) does not map to outstanding cmd\n",
656 				cpl.cid);
657 			/* nvme_dump_completion expects device endianess */
658 			nvme_dump_completion(&qpair->cpl[qpair->cq_head]);
659 			KASSERT(0, ("received completion for unknown cmd"));
660 		}
661 
662 		/*
663 		 * There's a number of races with the following (see above) when
664 		 * the system panics. We compensate for each one of them by
665 		 * using the atomic store to force strong ordering (at least when
666 		 * viewed in the aftermath of a panic).
667 		 */
668 		if (++qpair->cq_head == qpair->num_entries) {		/* 1 */
669 			atomic_store_rel_int(&qpair->cq_head, 0);	/* 2 */
670 			qpair->phase = !qpair->phase;			/* 3 */
671 		}
672 	}
673 
674 	if (done != 0) {
675 		bus_space_write_4(qpair->ctrlr->bus_tag, qpair->ctrlr->bus_handle,
676 		    qpair->cq_hdbl_off, qpair->cq_head);
677 	}
678 
679 	return (done != 0);
680 }
681 
682 static void
683 nvme_qpair_msi_handler(void *arg)
684 {
685 	struct nvme_qpair *qpair = arg;
686 
687 	nvme_qpair_process_completions(qpair);
688 }
689 
690 int
691 nvme_qpair_construct(struct nvme_qpair *qpair,
692     uint32_t num_entries, uint32_t num_trackers,
693     struct nvme_controller *ctrlr)
694 {
695 	struct nvme_tracker	*tr;
696 	size_t			cmdsz, cplsz, prpsz, allocsz, prpmemsz;
697 	uint64_t		queuemem_phys, prpmem_phys, list_phys;
698 	uint8_t			*queuemem, *prpmem, *prp_list;
699 	int			i, err;
700 
701 	qpair->vector = ctrlr->msi_count > 1 ? qpair->id : 0;
702 	qpair->num_entries = num_entries;
703 	qpair->num_trackers = num_trackers;
704 	qpair->ctrlr = ctrlr;
705 
706 	mtx_init(&qpair->lock, "nvme qpair lock", NULL, MTX_DEF);
707 
708 	/* Note: NVMe PRP format is restricted to 4-byte alignment. */
709 	err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev),
710 	    4, ctrlr->page_size, BUS_SPACE_MAXADDR,
711 	    BUS_SPACE_MAXADDR, NULL, NULL, ctrlr->max_xfer_size,
712 	    howmany(ctrlr->max_xfer_size, ctrlr->page_size) + 1,
713 	    ctrlr->page_size, 0,
714 	    NULL, NULL, &qpair->dma_tag_payload);
715 	if (err != 0) {
716 		nvme_printf(ctrlr, "payload tag create failed %d\n", err);
717 		goto out;
718 	}
719 
720 	/*
721 	 * Each component must be page aligned, and individual PRP lists
722 	 * cannot cross a page boundary.
723 	 */
724 	cmdsz = qpair->num_entries * sizeof(struct nvme_command);
725 	cmdsz = roundup2(cmdsz, ctrlr->page_size);
726 	cplsz = qpair->num_entries * sizeof(struct nvme_completion);
727 	cplsz = roundup2(cplsz, ctrlr->page_size);
728 	/*
729 	 * For commands requiring more than 2 PRP entries, one PRP will be
730 	 * embedded in the command (prp1), and the rest of the PRP entries
731 	 * will be in a list pointed to by the command (prp2).
732 	 */
733 	prpsz = sizeof(uint64_t) *
734 	    howmany(ctrlr->max_xfer_size, ctrlr->page_size);
735 	prpmemsz = qpair->num_trackers * prpsz;
736 	allocsz = cmdsz + cplsz + prpmemsz;
737 
738 	err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev),
739 	    ctrlr->page_size, 0, BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
740 	    allocsz, 1, allocsz, 0, NULL, NULL, &qpair->dma_tag);
741 	if (err != 0) {
742 		nvme_printf(ctrlr, "tag create failed %d\n", err);
743 		goto out;
744 	}
745 	bus_dma_tag_set_domain(qpair->dma_tag, qpair->domain);
746 
747 	if (bus_dmamem_alloc(qpair->dma_tag, (void **)&queuemem,
748 	     BUS_DMA_COHERENT | BUS_DMA_NOWAIT, &qpair->queuemem_map)) {
749 		nvme_printf(ctrlr, "failed to alloc qpair memory\n");
750 		goto out;
751 	}
752 
753 	if (bus_dmamap_load(qpair->dma_tag, qpair->queuemem_map,
754 	    queuemem, allocsz, nvme_single_map, &queuemem_phys, 0) != 0) {
755 		nvme_printf(ctrlr, "failed to load qpair memory\n");
756 		bus_dmamem_free(qpair->dma_tag, qpair->cmd,
757 		    qpair->queuemem_map);
758 		goto out;
759 	}
760 
761 	qpair->num_cmds = 0;
762 	qpair->num_intr_handler_calls = 0;
763 	qpair->num_retries = 0;
764 	qpair->num_failures = 0;
765 	qpair->num_ignored = 0;
766 	qpair->cmd = (struct nvme_command *)queuemem;
767 	qpair->cpl = (struct nvme_completion *)(queuemem + cmdsz);
768 	prpmem = (uint8_t *)(queuemem + cmdsz + cplsz);
769 	qpair->cmd_bus_addr = queuemem_phys;
770 	qpair->cpl_bus_addr = queuemem_phys + cmdsz;
771 	prpmem_phys = queuemem_phys + cmdsz + cplsz;
772 
773 	callout_init(&qpair->timer, 1);
774 	qpair->timer_armed = false;
775 	qpair->recovery_state = RECOVERY_WAITING;
776 
777 	/*
778 	 * Calcuate the stride of the doorbell register. Many emulators set this
779 	 * value to correspond to a cache line. However, some hardware has set
780 	 * it to various small values.
781 	 */
782 	qpair->sq_tdbl_off = nvme_mmio_offsetof(doorbell[0]) +
783 	    (qpair->id << (ctrlr->dstrd + 1));
784 	qpair->cq_hdbl_off = nvme_mmio_offsetof(doorbell[0]) +
785 	    (qpair->id << (ctrlr->dstrd + 1)) + (1 << ctrlr->dstrd);
786 
787 	TAILQ_INIT(&qpair->free_tr);
788 	TAILQ_INIT(&qpair->outstanding_tr);
789 	STAILQ_INIT(&qpair->queued_req);
790 
791 	list_phys = prpmem_phys;
792 	prp_list = prpmem;
793 	for (i = 0; i < qpair->num_trackers; i++) {
794 		if (list_phys + prpsz > prpmem_phys + prpmemsz) {
795 			qpair->num_trackers = i;
796 			break;
797 		}
798 
799 		/*
800 		 * Make sure that the PRP list for this tracker doesn't
801 		 * overflow to another nvme page.
802 		 */
803 		if (trunc_page(list_phys) !=
804 		    trunc_page(list_phys + prpsz - 1)) {
805 			list_phys = roundup2(list_phys, ctrlr->page_size);
806 			prp_list =
807 			    (uint8_t *)roundup2((uintptr_t)prp_list, ctrlr->page_size);
808 		}
809 
810 		tr = malloc_domainset(sizeof(*tr), M_NVME,
811 		    DOMAINSET_PREF(qpair->domain), M_ZERO | M_WAITOK);
812 		bus_dmamap_create(qpair->dma_tag_payload, 0,
813 		    &tr->payload_dma_map);
814 		tr->cid = i;
815 		tr->qpair = qpair;
816 		tr->prp = (uint64_t *)prp_list;
817 		tr->prp_bus_addr = list_phys;
818 		TAILQ_INSERT_HEAD(&qpair->free_tr, tr, tailq);
819 		list_phys += prpsz;
820 		prp_list += prpsz;
821 	}
822 
823 	if (qpair->num_trackers == 0) {
824 		nvme_printf(ctrlr, "failed to allocate enough trackers\n");
825 		goto out;
826 	}
827 
828 	qpair->act_tr = malloc_domainset(sizeof(struct nvme_tracker *) *
829 	    qpair->num_entries, M_NVME, DOMAINSET_PREF(qpair->domain),
830 	    M_ZERO | M_WAITOK);
831 
832 	if (ctrlr->msi_count > 1) {
833 		/*
834 		 * MSI-X vector resource IDs start at 1, so we add one to
835 		 *  the queue's vector to get the corresponding rid to use.
836 		 */
837 		qpair->rid = qpair->vector + 1;
838 
839 		qpair->res = bus_alloc_resource_any(ctrlr->dev, SYS_RES_IRQ,
840 		    &qpair->rid, RF_ACTIVE);
841 		if (qpair->res == NULL) {
842 			nvme_printf(ctrlr, "unable to allocate MSI\n");
843 			goto out;
844 		}
845 		if (bus_setup_intr(ctrlr->dev, qpair->res,
846 		    INTR_TYPE_MISC | INTR_MPSAFE, NULL,
847 		    nvme_qpair_msi_handler, qpair, &qpair->tag) != 0) {
848 			nvme_printf(ctrlr, "unable to setup MSI\n");
849 			goto out;
850 		}
851 		if (qpair->id == 0) {
852 			bus_describe_intr(ctrlr->dev, qpair->res, qpair->tag,
853 			    "admin");
854 		} else {
855 			bus_describe_intr(ctrlr->dev, qpair->res, qpair->tag,
856 			    "io%d", qpair->id - 1);
857 		}
858 	}
859 
860 	return (0);
861 
862 out:
863 	nvme_qpair_destroy(qpair);
864 	return (ENOMEM);
865 }
866 
867 static void
868 nvme_qpair_destroy(struct nvme_qpair *qpair)
869 {
870 	struct nvme_tracker	*tr;
871 
872 	callout_drain(&qpair->timer);
873 
874 	if (qpair->tag) {
875 		bus_teardown_intr(qpair->ctrlr->dev, qpair->res, qpair->tag);
876 		qpair->tag = NULL;
877 	}
878 
879 	if (qpair->act_tr) {
880 		free(qpair->act_tr, M_NVME);
881 		qpair->act_tr = NULL;
882 	}
883 
884 	while (!TAILQ_EMPTY(&qpair->free_tr)) {
885 		tr = TAILQ_FIRST(&qpair->free_tr);
886 		TAILQ_REMOVE(&qpair->free_tr, tr, tailq);
887 		bus_dmamap_destroy(qpair->dma_tag_payload,
888 		    tr->payload_dma_map);
889 		free(tr, M_NVME);
890 	}
891 
892 	if (qpair->cmd != NULL) {
893 		bus_dmamap_unload(qpair->dma_tag, qpair->queuemem_map);
894 		bus_dmamem_free(qpair->dma_tag, qpair->cmd,
895 		    qpair->queuemem_map);
896 		qpair->cmd = NULL;
897 	}
898 
899 	if (qpair->dma_tag) {
900 		bus_dma_tag_destroy(qpair->dma_tag);
901 		qpair->dma_tag = NULL;
902 	}
903 
904 	if (qpair->dma_tag_payload) {
905 		bus_dma_tag_destroy(qpair->dma_tag_payload);
906 		qpair->dma_tag_payload = NULL;
907 	}
908 
909 	if (mtx_initialized(&qpair->lock))
910 		mtx_destroy(&qpair->lock);
911 
912 	if (qpair->res) {
913 		bus_release_resource(qpair->ctrlr->dev, SYS_RES_IRQ,
914 		    rman_get_rid(qpair->res), qpair->res);
915 		qpair->res = NULL;
916 	}
917 }
918 
919 static void
920 nvme_admin_qpair_abort_aers(struct nvme_qpair *qpair)
921 {
922 	struct nvme_tracker	*tr;
923 
924 	tr = TAILQ_FIRST(&qpair->outstanding_tr);
925 	while (tr != NULL) {
926 		if (tr->req->cmd.opc == NVME_OPC_ASYNC_EVENT_REQUEST) {
927 			nvme_qpair_manual_complete_tracker(tr,
928 			    NVME_SCT_GENERIC, NVME_SC_ABORTED_SQ_DELETION, 0,
929 			    ERROR_PRINT_NONE);
930 			tr = TAILQ_FIRST(&qpair->outstanding_tr);
931 		} else {
932 			tr = TAILQ_NEXT(tr, tailq);
933 		}
934 	}
935 }
936 
937 void
938 nvme_admin_qpair_destroy(struct nvme_qpair *qpair)
939 {
940 
941 	nvme_admin_qpair_abort_aers(qpair);
942 	nvme_qpair_destroy(qpair);
943 }
944 
945 void
946 nvme_io_qpair_destroy(struct nvme_qpair *qpair)
947 {
948 
949 	nvme_qpair_destroy(qpair);
950 }
951 
952 static void
953 nvme_qpair_timeout(void *arg)
954 {
955 	struct nvme_qpair	*qpair = arg;
956 	struct nvme_controller	*ctrlr = qpair->ctrlr;
957 	struct nvme_tracker	*tr;
958 	sbintime_t		now;
959 	bool			idle;
960 	uint32_t		csts;
961 	uint8_t			cfs;
962 
963 	mtx_lock(&qpair->lock);
964 	idle = TAILQ_EMPTY(&qpair->outstanding_tr);
965 again:
966 	switch (qpair->recovery_state) {
967 	case RECOVERY_NONE:
968 		if (idle)
969 			break;
970 		now = getsbinuptime();
971 		idle = true;
972 		TAILQ_FOREACH(tr, &qpair->outstanding_tr, tailq) {
973 			if (tr->deadline == SBT_MAX)
974 				continue;
975 			idle = false;
976 			if (now > tr->deadline) {
977 				/*
978 				 * We're now passed our earliest deadline. We
979 				 * need to do expensive things to cope, but next
980 				 * time. Flag that and close the door to any
981 				 * further processing.
982 				 */
983 				qpair->recovery_state = RECOVERY_START;
984 				nvme_printf(ctrlr, "RECOVERY_START %jd vs %jd\n",
985 				    (uintmax_t)now, (uintmax_t)tr->deadline);
986 				break;
987 			}
988 		}
989 		break;
990 	case RECOVERY_START:
991 		/*
992 		 * Read csts to get value of cfs - controller fatal status.
993 		 * If no fatal status, try to call the completion routine, and
994 		 * if completes transactions, report a missed interrupt and
995 		 * return (this may need to be rate limited). Otherwise, if
996 		 * aborts are enabled and the controller is not reporting
997 		 * fatal status, abort the command. Otherwise, just reset the
998 		 * controller and hope for the best.
999 		 */
1000 		csts = nvme_mmio_read_4(ctrlr, csts);
1001 		cfs = (csts >> NVME_CSTS_REG_CFS_SHIFT) & NVME_CSTS_REG_CFS_MASK;
1002 		if (cfs) {
1003 			nvme_printf(ctrlr, "Controller in fatal status, resetting\n");
1004 			qpair->recovery_state = RECOVERY_RESET;
1005 			goto again;
1006 		}
1007 		mtx_unlock(&qpair->lock);
1008 		if (nvme_qpair_process_completions(qpair)) {
1009 			nvme_printf(ctrlr, "Completions present in output without an interrupt\n");
1010 			qpair->recovery_state = RECOVERY_NONE;
1011 		} else {
1012 			nvme_printf(ctrlr, "timeout with nothing complete, resetting\n");
1013 			qpair->recovery_state = RECOVERY_RESET;
1014 			mtx_lock(&qpair->lock);
1015 			goto again;
1016 		}
1017 		mtx_lock(&qpair->lock);
1018 		break;
1019 	case RECOVERY_RESET:
1020 		/*
1021 		 * If we get here due to a possible surprise hot-unplug event,
1022 		 * then we let nvme_ctrlr_reset confirm and fail the
1023 		 * controller.
1024 		 */
1025 		nvme_printf(ctrlr, "Resetting controller due to a timeout%s.\n",
1026 		    (csts == 0xffffffff) ? " and possible hot unplug" :
1027 		    (cfs ? " and fatal error status" : ""));
1028 		nvme_printf(ctrlr, "RECOVERY_WAITING\n");
1029 		qpair->recovery_state = RECOVERY_WAITING;
1030 		nvme_ctrlr_reset(ctrlr);
1031 		break;
1032 	case RECOVERY_WAITING:
1033 		nvme_printf(ctrlr, "waiting\n");
1034 		break;
1035 	}
1036 
1037 	/*
1038 	 * Rearm the timeout.
1039 	 */
1040 	if (!idle) {
1041 		callout_schedule_sbt(&qpair->timer, SBT_1S / 2, SBT_1S / 2, 0);
1042 	} else {
1043 		qpair->timer_armed = false;
1044 	}
1045 	mtx_unlock(&qpair->lock);
1046 }
1047 
1048 /*
1049  * Submit the tracker to the hardware. Must already be in the
1050  * outstanding queue when called.
1051  */
1052 void
1053 nvme_qpair_submit_tracker(struct nvme_qpair *qpair, struct nvme_tracker *tr)
1054 {
1055 	struct nvme_request	*req;
1056 	struct nvme_controller	*ctrlr;
1057 	int timeout;
1058 
1059 	mtx_assert(&qpair->lock, MA_OWNED);
1060 
1061 	req = tr->req;
1062 	req->cmd.cid = tr->cid;
1063 	qpair->act_tr[tr->cid] = tr;
1064 	ctrlr = qpair->ctrlr;
1065 
1066 	if (req->timeout) {
1067 		if (req->cb_fn == nvme_completion_poll_cb)
1068 			timeout = 1;
1069 		else
1070 			timeout = ctrlr->timeout_period;
1071 		tr->deadline = getsbinuptime() + timeout * SBT_1S;
1072 		if (!qpair->timer_armed) {
1073 			qpair->timer_armed = true;
1074 			callout_reset_sbt_on(&qpair->timer, SBT_1S / 2, SBT_1S / 2,
1075 			    nvme_qpair_timeout, qpair, qpair->cpu, 0);
1076 		}
1077 	} else
1078 		tr->deadline = SBT_MAX;
1079 
1080 	/* Copy the command from the tracker to the submission queue. */
1081 	memcpy(&qpair->cmd[qpair->sq_tail], &req->cmd, sizeof(req->cmd));
1082 
1083 	if (++qpair->sq_tail == qpair->num_entries)
1084 		qpair->sq_tail = 0;
1085 
1086 	bus_dmamap_sync(qpair->dma_tag, qpair->queuemem_map,
1087 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1088 	bus_space_write_4(qpair->ctrlr->bus_tag, qpair->ctrlr->bus_handle,
1089 	    qpair->sq_tdbl_off, qpair->sq_tail);
1090 	qpair->num_cmds++;
1091 }
1092 
1093 static void
1094 nvme_payload_map(void *arg, bus_dma_segment_t *seg, int nseg, int error)
1095 {
1096 	struct nvme_tracker 	*tr = arg;
1097 	uint32_t		cur_nseg;
1098 
1099 	/*
1100 	 * If the mapping operation failed, return immediately.  The caller
1101 	 *  is responsible for detecting the error status and failing the
1102 	 *  tracker manually.
1103 	 */
1104 	if (error != 0) {
1105 		nvme_printf(tr->qpair->ctrlr,
1106 		    "nvme_payload_map err %d\n", error);
1107 		return;
1108 	}
1109 
1110 	/*
1111 	 * Note that we specified ctrlr->page_size for alignment and max
1112 	 * segment size when creating the bus dma tags.  So here we can safely
1113 	 * just transfer each segment to its associated PRP entry.
1114 	 */
1115 	tr->req->cmd.prp1 = htole64(seg[0].ds_addr);
1116 
1117 	if (nseg == 2) {
1118 		tr->req->cmd.prp2 = htole64(seg[1].ds_addr);
1119 	} else if (nseg > 2) {
1120 		cur_nseg = 1;
1121 		tr->req->cmd.prp2 = htole64((uint64_t)tr->prp_bus_addr);
1122 		while (cur_nseg < nseg) {
1123 			tr->prp[cur_nseg-1] =
1124 			    htole64((uint64_t)seg[cur_nseg].ds_addr);
1125 			cur_nseg++;
1126 		}
1127 	} else {
1128 		/*
1129 		 * prp2 should not be used by the controller
1130 		 *  since there is only one segment, but set
1131 		 *  to 0 just to be safe.
1132 		 */
1133 		tr->req->cmd.prp2 = 0;
1134 	}
1135 
1136 	bus_dmamap_sync(tr->qpair->dma_tag_payload, tr->payload_dma_map,
1137 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1138 	nvme_qpair_submit_tracker(tr->qpair, tr);
1139 }
1140 
1141 static void
1142 _nvme_qpair_submit_request(struct nvme_qpair *qpair, struct nvme_request *req)
1143 {
1144 	struct nvme_tracker	*tr;
1145 	int			err = 0;
1146 
1147 	mtx_assert(&qpair->lock, MA_OWNED);
1148 
1149 	tr = TAILQ_FIRST(&qpair->free_tr);
1150 	req->qpair = qpair;
1151 
1152 	if (tr == NULL || qpair->recovery_state != RECOVERY_NONE) {
1153 		/*
1154 		 * No tracker is available, or the qpair is disabled due to
1155 		 *  an in-progress controller-level reset or controller
1156 		 *  failure.
1157 		 */
1158 
1159 		if (qpair->ctrlr->is_failed) {
1160 			/*
1161 			 * The controller has failed, so fail the request.
1162 			 */
1163 			nvme_qpair_manual_complete_request(qpair, req,
1164 			    NVME_SCT_GENERIC, NVME_SC_ABORTED_BY_REQUEST);
1165 		} else {
1166 			/*
1167 			 * Put the request on the qpair's request queue to be
1168 			 *  processed when a tracker frees up via a command
1169 			 *  completion or when the controller reset is
1170 			 *  completed.
1171 			 */
1172 			STAILQ_INSERT_TAIL(&qpair->queued_req, req, stailq);
1173 		}
1174 		return;
1175 	}
1176 
1177 	TAILQ_REMOVE(&qpair->free_tr, tr, tailq);
1178 	TAILQ_INSERT_TAIL(&qpair->outstanding_tr, tr, tailq);
1179 	tr->deadline = SBT_MAX;
1180 	tr->req = req;
1181 
1182 	if (!req->payload_valid) {
1183 		nvme_qpair_submit_tracker(tr->qpair, tr);
1184 		return;
1185 	}
1186 
1187 	err = bus_dmamap_load_mem(tr->qpair->dma_tag_payload,
1188 	    tr->payload_dma_map, &req->payload, nvme_payload_map, tr, 0);
1189 	if (err != 0) {
1190 		/*
1191 		 * The dmamap operation failed, so we manually fail the
1192 		 *  tracker here with DATA_TRANSFER_ERROR status.
1193 		 *
1194 		 * nvme_qpair_manual_complete_tracker must not be called
1195 		 *  with the qpair lock held.
1196 		 */
1197 		nvme_printf(qpair->ctrlr,
1198 		    "bus_dmamap_load_mem returned 0x%x!\n", err);
1199 		mtx_unlock(&qpair->lock);
1200 		nvme_qpair_manual_complete_tracker(tr, NVME_SCT_GENERIC,
1201 		    NVME_SC_DATA_TRANSFER_ERROR, DO_NOT_RETRY, ERROR_PRINT_ALL);
1202 		mtx_lock(&qpair->lock);
1203 	}
1204 }
1205 
1206 void
1207 nvme_qpair_submit_request(struct nvme_qpair *qpair, struct nvme_request *req)
1208 {
1209 
1210 	mtx_lock(&qpair->lock);
1211 	_nvme_qpair_submit_request(qpair, req);
1212 	mtx_unlock(&qpair->lock);
1213 }
1214 
1215 static void
1216 nvme_qpair_enable(struct nvme_qpair *qpair)
1217 {
1218 	mtx_assert(&qpair->lock, MA_OWNED);
1219 
1220 	qpair->recovery_state = RECOVERY_NONE;
1221 }
1222 
1223 void
1224 nvme_qpair_reset(struct nvme_qpair *qpair)
1225 {
1226 
1227 	qpair->sq_head = qpair->sq_tail = qpair->cq_head = 0;
1228 
1229 	/*
1230 	 * First time through the completion queue, HW will set phase
1231 	 *  bit on completions to 1.  So set this to 1 here, indicating
1232 	 *  we're looking for a 1 to know which entries have completed.
1233 	 *  we'll toggle the bit each time when the completion queue
1234 	 *  rolls over.
1235 	 */
1236 	qpair->phase = 1;
1237 
1238 	memset(qpair->cmd, 0,
1239 	    qpair->num_entries * sizeof(struct nvme_command));
1240 	memset(qpair->cpl, 0,
1241 	    qpair->num_entries * sizeof(struct nvme_completion));
1242 }
1243 
1244 void
1245 nvme_admin_qpair_enable(struct nvme_qpair *qpair)
1246 {
1247 	struct nvme_tracker		*tr;
1248 	struct nvme_tracker		*tr_temp;
1249 
1250 	/*
1251 	 * Manually abort each outstanding admin command.  Do not retry
1252 	 *  admin commands found here, since they will be left over from
1253 	 *  a controller reset and its likely the context in which the
1254 	 *  command was issued no longer applies.
1255 	 */
1256 	TAILQ_FOREACH_SAFE(tr, &qpair->outstanding_tr, tailq, tr_temp) {
1257 		nvme_printf(qpair->ctrlr,
1258 		    "aborting outstanding admin command\n");
1259 		nvme_qpair_manual_complete_tracker(tr, NVME_SCT_GENERIC,
1260 		    NVME_SC_ABORTED_BY_REQUEST, DO_NOT_RETRY, ERROR_PRINT_ALL);
1261 	}
1262 
1263 	mtx_lock(&qpair->lock);
1264 	nvme_qpair_enable(qpair);
1265 	mtx_unlock(&qpair->lock);
1266 }
1267 
1268 void
1269 nvme_io_qpair_enable(struct nvme_qpair *qpair)
1270 {
1271 	STAILQ_HEAD(, nvme_request)	temp;
1272 	struct nvme_tracker		*tr;
1273 	struct nvme_tracker		*tr_temp;
1274 	struct nvme_request		*req;
1275 
1276 	/*
1277 	 * Manually abort each outstanding I/O.  This normally results in a
1278 	 *  retry, unless the retry count on the associated request has
1279 	 *  reached its limit.
1280 	 */
1281 	TAILQ_FOREACH_SAFE(tr, &qpair->outstanding_tr, tailq, tr_temp) {
1282 		nvme_printf(qpair->ctrlr, "aborting outstanding i/o\n");
1283 		nvme_qpair_manual_complete_tracker(tr, NVME_SCT_GENERIC,
1284 		    NVME_SC_ABORTED_BY_REQUEST, 0, ERROR_PRINT_NO_RETRY);
1285 	}
1286 
1287 	mtx_lock(&qpair->lock);
1288 
1289 	nvme_qpair_enable(qpair);
1290 
1291 	STAILQ_INIT(&temp);
1292 	STAILQ_SWAP(&qpair->queued_req, &temp, nvme_request);
1293 
1294 	while (!STAILQ_EMPTY(&temp)) {
1295 		req = STAILQ_FIRST(&temp);
1296 		STAILQ_REMOVE_HEAD(&temp, stailq);
1297 		nvme_printf(qpair->ctrlr, "resubmitting queued i/o\n");
1298 		nvme_qpair_print_command(qpair, &req->cmd);
1299 		_nvme_qpair_submit_request(qpair, req);
1300 	}
1301 
1302 	mtx_unlock(&qpair->lock);
1303 }
1304 
1305 static void
1306 nvme_qpair_disable(struct nvme_qpair *qpair)
1307 {
1308 	struct nvme_tracker	*tr, *tr_temp;
1309 
1310 	mtx_lock(&qpair->lock);
1311 	qpair->recovery_state = RECOVERY_WAITING;
1312 	TAILQ_FOREACH_SAFE(tr, &qpair->outstanding_tr, tailq, tr_temp) {
1313 		tr->deadline = SBT_MAX;
1314 	}
1315 	mtx_unlock(&qpair->lock);
1316 }
1317 
1318 void
1319 nvme_admin_qpair_disable(struct nvme_qpair *qpair)
1320 {
1321 
1322 	nvme_qpair_disable(qpair);
1323 	nvme_admin_qpair_abort_aers(qpair);
1324 }
1325 
1326 void
1327 nvme_io_qpair_disable(struct nvme_qpair *qpair)
1328 {
1329 
1330 	nvme_qpair_disable(qpair);
1331 }
1332 
1333 void
1334 nvme_qpair_fail(struct nvme_qpair *qpair)
1335 {
1336 	struct nvme_tracker		*tr;
1337 	struct nvme_request		*req;
1338 
1339 	if (!mtx_initialized(&qpair->lock))
1340 		return;
1341 
1342 	mtx_lock(&qpair->lock);
1343 
1344 	while (!STAILQ_EMPTY(&qpair->queued_req)) {
1345 		req = STAILQ_FIRST(&qpair->queued_req);
1346 		STAILQ_REMOVE_HEAD(&qpair->queued_req, stailq);
1347 		nvme_printf(qpair->ctrlr, "failing queued i/o\n");
1348 		mtx_unlock(&qpair->lock);
1349 		nvme_qpair_manual_complete_request(qpair, req, NVME_SCT_GENERIC,
1350 		    NVME_SC_ABORTED_BY_REQUEST);
1351 		mtx_lock(&qpair->lock);
1352 	}
1353 
1354 	/* Manually abort each outstanding I/O. */
1355 	while (!TAILQ_EMPTY(&qpair->outstanding_tr)) {
1356 		tr = TAILQ_FIRST(&qpair->outstanding_tr);
1357 		/*
1358 		 * Do not remove the tracker.  The abort_tracker path will
1359 		 *  do that for us.
1360 		 */
1361 		nvme_printf(qpair->ctrlr, "failing outstanding i/o\n");
1362 		mtx_unlock(&qpair->lock);
1363 		nvme_qpair_manual_complete_tracker(tr, NVME_SCT_GENERIC,
1364 		    NVME_SC_ABORTED_BY_REQUEST, DO_NOT_RETRY, ERROR_PRINT_ALL);
1365 		mtx_lock(&qpair->lock);
1366 	}
1367 
1368 	mtx_unlock(&qpair->lock);
1369 }
1370