xref: /freebsd/sys/dev/nvme/nvme_qpair.c (revision ddc0daea20280c3a06a910b72b14ffe3f624df71)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
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/proc.h>
36 
37 #include <dev/pci/pcivar.h>
38 
39 #include "nvme_private.h"
40 
41 typedef enum error_print { ERROR_PRINT_NONE, ERROR_PRINT_NO_RETRY, ERROR_PRINT_ALL } error_print_t;
42 #define DO_NOT_RETRY	1
43 
44 static void	_nvme_qpair_submit_request(struct nvme_qpair *qpair,
45 					   struct nvme_request *req);
46 static void	nvme_qpair_destroy(struct nvme_qpair *qpair);
47 
48 struct nvme_opcode_string {
49 
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_FIRMWARE_ACTIVATE, "FIRMWARE ACTIVATE" },
66 	{ NVME_OPC_FIRMWARE_IMAGE_DOWNLOAD, "FIRMWARE IMAGE DOWNLOAD" },
67 	{ NVME_OPC_DEVICE_SELF_TEST, "DEVICE SELF-TEST" },
68 	{ NVME_OPC_NAMESPACE_ATTACHMENT, "NAMESPACE ATTACHMENT" },
69 	{ NVME_OPC_KEEP_ALIVE, "KEEP ALIVE" },
70 	{ NVME_OPC_DIRECTIVE_SEND, "DIRECTIVE SEND" },
71 	{ NVME_OPC_DIRECTIVE_RECEIVE, "DIRECTIVE RECEIVE" },
72 	{ NVME_OPC_VIRTUALIZATION_MANAGEMENT, "VIRTUALIZATION MANAGEMENT" },
73 	{ NVME_OPC_NVME_MI_SEND, "NVME-MI SEND" },
74 	{ NVME_OPC_NVME_MI_RECEIVE, "NVME-MI RECEIVE" },
75 	{ NVME_OPC_DOORBELL_BUFFER_CONFIG, "DOORBELL BUFFER CONFIG" },
76 	{ NVME_OPC_FORMAT_NVM, "FORMAT NVM" },
77 	{ NVME_OPC_SECURITY_SEND, "SECURITY SEND" },
78 	{ NVME_OPC_SECURITY_RECEIVE, "SECURITY RECEIVE" },
79 	{ NVME_OPC_SANITIZE, "SANITIZE" },
80 	{ NVME_OPC_GET_LBA_STATUS, "GET LBA STATUS" },
81 	{ 0xFFFF, "ADMIN COMMAND" }
82 };
83 
84 static struct nvme_opcode_string io_opcode[] = {
85 	{ NVME_OPC_FLUSH, "FLUSH" },
86 	{ NVME_OPC_WRITE, "WRITE" },
87 	{ NVME_OPC_READ, "READ" },
88 	{ NVME_OPC_WRITE_UNCORRECTABLE, "WRITE UNCORRECTABLE" },
89 	{ NVME_OPC_COMPARE, "COMPARE" },
90 	{ NVME_OPC_WRITE_ZEROES, "WRITE ZEROES" },
91 	{ NVME_OPC_DATASET_MANAGEMENT, "DATASET MANAGEMENT" },
92 	{ NVME_OPC_VERIFY, "VERIFY" },
93 	{ NVME_OPC_RESERVATION_REGISTER, "RESERVATION REGISTER" },
94 	{ NVME_OPC_RESERVATION_REPORT, "RESERVATION REPORT" },
95 	{ NVME_OPC_RESERVATION_ACQUIRE, "RESERVATION ACQUIRE" },
96 	{ NVME_OPC_RESERVATION_RELEASE, "RESERVATION RELEASE" },
97 	{ 0xFFFF, "IO COMMAND" }
98 };
99 
100 static const char *
101 get_admin_opcode_string(uint16_t opc)
102 {
103 	struct nvme_opcode_string *entry;
104 
105 	entry = admin_opcode;
106 
107 	while (entry->opc != 0xFFFF) {
108 		if (entry->opc == opc)
109 			return (entry->str);
110 		entry++;
111 	}
112 	return (entry->str);
113 }
114 
115 static const char *
116 get_io_opcode_string(uint16_t opc)
117 {
118 	struct nvme_opcode_string *entry;
119 
120 	entry = io_opcode;
121 
122 	while (entry->opc != 0xFFFF) {
123 		if (entry->opc == opc)
124 			return (entry->str);
125 		entry++;
126 	}
127 	return (entry->str);
128 }
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 
198 	uint16_t	sc;
199 	const char *	str;
200 };
201 
202 static struct nvme_status_string generic_status[] = {
203 	{ NVME_SC_SUCCESS, "SUCCESS" },
204 	{ NVME_SC_INVALID_OPCODE, "INVALID OPCODE" },
205 	{ NVME_SC_INVALID_FIELD, "INVALID_FIELD" },
206 	{ NVME_SC_COMMAND_ID_CONFLICT, "COMMAND ID CONFLICT" },
207 	{ NVME_SC_DATA_TRANSFER_ERROR, "DATA TRANSFER ERROR" },
208 	{ NVME_SC_ABORTED_POWER_LOSS, "ABORTED - POWER LOSS" },
209 	{ NVME_SC_INTERNAL_DEVICE_ERROR, "INTERNAL DEVICE ERROR" },
210 	{ NVME_SC_ABORTED_BY_REQUEST, "ABORTED - BY REQUEST" },
211 	{ NVME_SC_ABORTED_SQ_DELETION, "ABORTED - SQ DELETION" },
212 	{ NVME_SC_ABORTED_FAILED_FUSED, "ABORTED - FAILED FUSED" },
213 	{ NVME_SC_ABORTED_MISSING_FUSED, "ABORTED - MISSING FUSED" },
214 	{ NVME_SC_INVALID_NAMESPACE_OR_FORMAT, "INVALID NAMESPACE OR FORMAT" },
215 	{ NVME_SC_COMMAND_SEQUENCE_ERROR, "COMMAND SEQUENCE ERROR" },
216 	{ NVME_SC_INVALID_SGL_SEGMENT_DESCR, "INVALID SGL SEGMENT DESCRIPTOR" },
217 	{ NVME_SC_INVALID_NUMBER_OF_SGL_DESCR, "INVALID NUMBER OF SGL DESCRIPTORS" },
218 	{ NVME_SC_DATA_SGL_LENGTH_INVALID, "DATA SGL LENGTH INVALID" },
219 	{ NVME_SC_METADATA_SGL_LENGTH_INVALID, "METADATA SGL LENGTH INVALID" },
220 	{ NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID, "SGL DESCRIPTOR TYPE INVALID" },
221 	{ NVME_SC_INVALID_USE_OF_CMB, "INVALID USE OF CONTROLLER MEMORY BUFFER" },
222 	{ NVME_SC_PRP_OFFET_INVALID, "PRP OFFET INVALID" },
223 	{ NVME_SC_ATOMIC_WRITE_UNIT_EXCEEDED, "ATOMIC WRITE UNIT EXCEEDED" },
224 	{ NVME_SC_OPERATION_DENIED, "OPERATION DENIED" },
225 	{ NVME_SC_SGL_OFFSET_INVALID, "SGL OFFSET INVALID" },
226 	{ NVME_SC_HOST_ID_INCONSISTENT_FORMAT, "HOST IDENTIFIER INCONSISTENT FORMAT" },
227 	{ NVME_SC_KEEP_ALIVE_TIMEOUT_EXPIRED, "KEEP ALIVE TIMEOUT EXPIRED" },
228 	{ NVME_SC_KEEP_ALIVE_TIMEOUT_INVALID, "KEEP ALIVE TIMEOUT INVALID" },
229 	{ NVME_SC_ABORTED_DUE_TO_PREEMPT, "COMMAND ABORTED DUE TO PREEMPT AND ABORT" },
230 	{ NVME_SC_SANITIZE_FAILED, "SANITIZE FAILED" },
231 	{ NVME_SC_SANITIZE_IN_PROGRESS, "SANITIZE IN PROGRESS" },
232 	{ NVME_SC_SGL_DATA_BLOCK_GRAN_INVALID, "SGL_DATA_BLOCK_GRANULARITY_INVALID" },
233 	{ NVME_SC_NOT_SUPPORTED_IN_CMB, "COMMAND NOT SUPPORTED FOR QUEUE IN CMB" },
234 	{ NVME_SC_NAMESPACE_IS_WRITE_PROTECTED, "NAMESPACE IS WRITE PROTECTED" },
235 	{ NVME_SC_COMMAND_INTERRUPTED, "COMMAND INTERRUPTED" },
236 	{ NVME_SC_TRANSIENT_TRANSPORT_ERROR, "TRANSIENT TRANSPORT ERROR" },
237 
238 	{ NVME_SC_LBA_OUT_OF_RANGE, "LBA OUT OF RANGE" },
239 	{ NVME_SC_CAPACITY_EXCEEDED, "CAPACITY EXCEEDED" },
240 	{ NVME_SC_NAMESPACE_NOT_READY, "NAMESPACE NOT READY" },
241 	{ NVME_SC_RESERVATION_CONFLICT, "RESERVATION CONFLICT" },
242 	{ NVME_SC_FORMAT_IN_PROGRESS, "FORMAT IN PROGRESS" },
243 	{ 0xFFFF, "GENERIC" }
244 };
245 
246 static struct nvme_status_string command_specific_status[] = {
247 	{ NVME_SC_COMPLETION_QUEUE_INVALID, "INVALID COMPLETION QUEUE" },
248 	{ NVME_SC_INVALID_QUEUE_IDENTIFIER, "INVALID QUEUE IDENTIFIER" },
249 	{ NVME_SC_MAXIMUM_QUEUE_SIZE_EXCEEDED, "MAX QUEUE SIZE EXCEEDED" },
250 	{ NVME_SC_ABORT_COMMAND_LIMIT_EXCEEDED, "ABORT CMD LIMIT EXCEEDED" },
251 	{ NVME_SC_ASYNC_EVENT_REQUEST_LIMIT_EXCEEDED, "ASYNC LIMIT EXCEEDED" },
252 	{ NVME_SC_INVALID_FIRMWARE_SLOT, "INVALID FIRMWARE SLOT" },
253 	{ NVME_SC_INVALID_FIRMWARE_IMAGE, "INVALID FIRMWARE IMAGE" },
254 	{ NVME_SC_INVALID_INTERRUPT_VECTOR, "INVALID INTERRUPT VECTOR" },
255 	{ NVME_SC_INVALID_LOG_PAGE, "INVALID LOG PAGE" },
256 	{ NVME_SC_INVALID_FORMAT, "INVALID FORMAT" },
257 	{ NVME_SC_FIRMWARE_REQUIRES_RESET, "FIRMWARE REQUIRES RESET" },
258 	{ NVME_SC_INVALID_QUEUE_DELETION, "INVALID QUEUE DELETION" },
259 	{ NVME_SC_FEATURE_NOT_SAVEABLE, "FEATURE IDENTIFIER NOT SAVEABLE" },
260 	{ NVME_SC_FEATURE_NOT_CHANGEABLE, "FEATURE NOT CHANGEABLE" },
261 	{ NVME_SC_FEATURE_NOT_NS_SPECIFIC, "FEATURE NOT NAMESPACE SPECIFIC" },
262 	{ NVME_SC_FW_ACT_REQUIRES_NVMS_RESET, "FIRMWARE ACTIVATION REQUIRES NVM SUBSYSTEM RESET" },
263 	{ NVME_SC_FW_ACT_REQUIRES_RESET, "FIRMWARE ACTIVATION REQUIRES RESET" },
264 	{ NVME_SC_FW_ACT_REQUIRES_TIME, "FIRMWARE ACTIVATION REQUIRES MAXIMUM TIME VIOLATION" },
265 	{ NVME_SC_FW_ACT_PROHIBITED, "FIRMWARE ACTIVATION PROHIBITED" },
266 	{ NVME_SC_OVERLAPPING_RANGE, "OVERLAPPING RANGE" },
267 	{ NVME_SC_NS_INSUFFICIENT_CAPACITY, "NAMESPACE INSUFFICIENT CAPACITY" },
268 	{ NVME_SC_NS_ID_UNAVAILABLE, "NAMESPACE IDENTIFIER UNAVAILABLE" },
269 	{ NVME_SC_NS_ALREADY_ATTACHED, "NAMESPACE ALREADY ATTACHED" },
270 	{ NVME_SC_NS_IS_PRIVATE, "NAMESPACE IS PRIVATE" },
271 	{ NVME_SC_NS_NOT_ATTACHED, "NS NOT ATTACHED" },
272 	{ NVME_SC_THIN_PROV_NOT_SUPPORTED, "THIN PROVISIONING NOT SUPPORTED" },
273 	{ NVME_SC_CTRLR_LIST_INVALID, "CONTROLLER LIST INVALID" },
274 	{ NVME_SC_SELT_TEST_IN_PROGRESS, "DEVICE SELT-TEST IN PROGRESS" },
275 	{ NVME_SC_BOOT_PART_WRITE_PROHIB, "BOOT PARTITION WRITE PROHIBITED" },
276 	{ NVME_SC_INVALID_CTRLR_ID, "INVALID CONTROLLER IDENTIFIER" },
277 	{ NVME_SC_INVALID_SEC_CTRLR_STATE, "INVALID SECONDARY CONTROLLER STATE" },
278 	{ NVME_SC_INVALID_NUM_OF_CTRLR_RESRC, "INVALID NUMBER OF CONTROLLER RESOURCES" },
279 	{ NVME_SC_INVALID_RESOURCE_ID, "INVALID RESOURCE IDENTIFIER" },
280 	{ NVME_SC_SANITIZE_PROHIBITED_WPMRE, "SANITIZE PROHIBITED WRITE PERSISTENT MEMORY REGION ENABLED" },
281 	{ NVME_SC_ANA_GROUP_ID_INVALID, "ANA GROUP IDENTIFIED INVALID" },
282 	{ NVME_SC_ANA_ATTACH_FAILED, "ANA ATTACH FAILED" },
283 
284 	{ NVME_SC_CONFLICTING_ATTRIBUTES, "CONFLICTING ATTRIBUTES" },
285 	{ NVME_SC_INVALID_PROTECTION_INFO, "INVALID PROTECTION INFO" },
286 	{ NVME_SC_ATTEMPTED_WRITE_TO_RO_PAGE, "WRITE TO RO PAGE" },
287 	{ 0xFFFF, "COMMAND SPECIFIC" }
288 };
289 
290 static struct nvme_status_string media_error_status[] = {
291 	{ NVME_SC_WRITE_FAULTS, "WRITE FAULTS" },
292 	{ NVME_SC_UNRECOVERED_READ_ERROR, "UNRECOVERED READ ERROR" },
293 	{ NVME_SC_GUARD_CHECK_ERROR, "GUARD CHECK ERROR" },
294 	{ NVME_SC_APPLICATION_TAG_CHECK_ERROR, "APPLICATION TAG CHECK ERROR" },
295 	{ NVME_SC_REFERENCE_TAG_CHECK_ERROR, "REFERENCE TAG CHECK ERROR" },
296 	{ NVME_SC_COMPARE_FAILURE, "COMPARE FAILURE" },
297 	{ NVME_SC_ACCESS_DENIED, "ACCESS DENIED" },
298 	{ NVME_SC_DEALLOCATED_OR_UNWRITTEN, "DEALLOCATED OR UNWRITTEN LOGICAL BLOCK" },
299 	{ 0xFFFF, "MEDIA ERROR" }
300 };
301 
302 static struct nvme_status_string path_related_status[] = {
303 	{ NVME_SC_INTERNAL_PATH_ERROR, "INTERNAL PATH ERROR" },
304 	{ NVME_SC_ASYMMETRIC_ACCESS_PERSISTENT_LOSS, "ASYMMETRIC ACCESS PERSISTENT LOSS" },
305 	{ NVME_SC_ASYMMETRIC_ACCESS_INACCESSIBLE, "ASYMMETRIC ACCESS INACCESSIBLE" },
306 	{ NVME_SC_ASYMMETRIC_ACCESS_TRANSITION, "ASYMMETRIC ACCESS TRANSITION" },
307 	{ NVME_SC_CONTROLLER_PATHING_ERROR, "CONTROLLER PATHING ERROR" },
308 	{ NVME_SC_HOST_PATHING_ERROR, "HOST PATHING ERROR" },
309 	{ NVME_SC_COMMAND_ABOTHED_BY_HOST, "COMMAND ABOTHED BY HOST" },
310 	{ 0xFFFF, "PATH RELATED" },
311 };
312 
313 static const char *
314 get_status_string(uint16_t sct, uint16_t sc)
315 {
316 	struct nvme_status_string *entry;
317 
318 	switch (sct) {
319 	case NVME_SCT_GENERIC:
320 		entry = generic_status;
321 		break;
322 	case NVME_SCT_COMMAND_SPECIFIC:
323 		entry = command_specific_status;
324 		break;
325 	case NVME_SCT_MEDIA_ERROR:
326 		entry = media_error_status;
327 		break;
328 	case NVME_SCT_PATH_RELATED:
329 		entry = path_related_status;
330 		break;
331 	case NVME_SCT_VENDOR_SPECIFIC:
332 		return ("VENDOR SPECIFIC");
333 	default:
334 		return ("RESERVED");
335 	}
336 
337 	while (entry->sc != 0xFFFF) {
338 		if (entry->sc == sc)
339 			return (entry->str);
340 		entry++;
341 	}
342 	return (entry->str);
343 }
344 
345 static void
346 nvme_qpair_print_completion(struct nvme_qpair *qpair,
347     struct nvme_completion *cpl)
348 {
349 	uint16_t sct, sc;
350 
351 	sct = NVME_STATUS_GET_SCT(cpl->status);
352 	sc = NVME_STATUS_GET_SC(cpl->status);
353 
354 	nvme_printf(qpair->ctrlr, "%s (%02x/%02x) sqid:%d cid:%d cdw0:%x\n",
355 	    get_status_string(sct, sc), sct, sc, cpl->sqid, cpl->cid,
356 	    cpl->cdw0);
357 }
358 
359 static boolean_t
360 nvme_completion_is_retry(const struct nvme_completion *cpl)
361 {
362 	uint8_t sct, sc, dnr;
363 
364 	sct = NVME_STATUS_GET_SCT(cpl->status);
365 	sc = NVME_STATUS_GET_SC(cpl->status);
366 	dnr = NVME_STATUS_GET_DNR(cpl->status);	/* Do Not Retry Bit */
367 
368 	/*
369 	 * TODO: spec is not clear how commands that are aborted due
370 	 *  to TLER will be marked.  So for now, it seems
371 	 *  NAMESPACE_NOT_READY is the only case where we should
372 	 *  look at the DNR bit. Requests failed with ABORTED_BY_REQUEST
373 	 *  set the DNR bit correctly since the driver controls that.
374 	 */
375 	switch (sct) {
376 	case NVME_SCT_GENERIC:
377 		switch (sc) {
378 		case NVME_SC_ABORTED_BY_REQUEST:
379 		case NVME_SC_NAMESPACE_NOT_READY:
380 			if (dnr)
381 				return (0);
382 			else
383 				return (1);
384 		case NVME_SC_INVALID_OPCODE:
385 		case NVME_SC_INVALID_FIELD:
386 		case NVME_SC_COMMAND_ID_CONFLICT:
387 		case NVME_SC_DATA_TRANSFER_ERROR:
388 		case NVME_SC_ABORTED_POWER_LOSS:
389 		case NVME_SC_INTERNAL_DEVICE_ERROR:
390 		case NVME_SC_ABORTED_SQ_DELETION:
391 		case NVME_SC_ABORTED_FAILED_FUSED:
392 		case NVME_SC_ABORTED_MISSING_FUSED:
393 		case NVME_SC_INVALID_NAMESPACE_OR_FORMAT:
394 		case NVME_SC_COMMAND_SEQUENCE_ERROR:
395 		case NVME_SC_LBA_OUT_OF_RANGE:
396 		case NVME_SC_CAPACITY_EXCEEDED:
397 		default:
398 			return (0);
399 		}
400 	case NVME_SCT_COMMAND_SPECIFIC:
401 	case NVME_SCT_MEDIA_ERROR:
402 		return (0);
403 	case NVME_SCT_PATH_RELATED:
404 		switch (sc) {
405 		case NVME_SC_INTERNAL_PATH_ERROR:
406 			if (dnr)
407 				return (0);
408 			else
409 				return (1);
410 		default:
411 			return (0);
412 		}
413 	case NVME_SCT_VENDOR_SPECIFIC:
414 	default:
415 		return (0);
416 	}
417 }
418 
419 static void
420 nvme_qpair_complete_tracker(struct nvme_qpair *qpair, struct nvme_tracker *tr,
421     struct nvme_completion *cpl, error_print_t print_on_error)
422 {
423 	struct nvme_request	*req;
424 	boolean_t		retry, error, retriable;
425 
426 	req = tr->req;
427 	error = nvme_completion_is_error(cpl);
428 	retriable = nvme_completion_is_retry(cpl);
429 	retry = error && retriable && req->retries < nvme_retry_count;
430 	if (retry)
431 		qpair->num_retries++;
432 	if (error && req->retries >= nvme_retry_count && retriable)
433 		qpair->num_failures++;
434 
435 	if (error && (print_on_error == ERROR_PRINT_ALL ||
436 		(!retry && print_on_error == ERROR_PRINT_NO_RETRY))) {
437 		nvme_qpair_print_command(qpair, &req->cmd);
438 		nvme_qpair_print_completion(qpair, cpl);
439 	}
440 
441 	qpair->act_tr[cpl->cid] = NULL;
442 
443 	KASSERT(cpl->cid == req->cmd.cid, ("cpl cid does not match cmd cid\n"));
444 
445 	if (req->cb_fn && !retry)
446 		req->cb_fn(req->cb_arg, cpl);
447 
448 	mtx_lock(&qpair->lock);
449 	callout_stop(&tr->timer);
450 
451 	if (retry) {
452 		req->retries++;
453 		nvme_qpair_submit_tracker(qpair, tr);
454 	} else {
455 		if (req->type != NVME_REQUEST_NULL) {
456 			bus_dmamap_sync(qpair->dma_tag_payload,
457 			    tr->payload_dma_map,
458 			    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
459 			bus_dmamap_unload(qpair->dma_tag_payload,
460 			    tr->payload_dma_map);
461 		}
462 
463 		nvme_free_request(req);
464 		tr->req = NULL;
465 
466 		TAILQ_REMOVE(&qpair->outstanding_tr, tr, tailq);
467 		TAILQ_INSERT_HEAD(&qpair->free_tr, tr, tailq);
468 
469 		/*
470 		 * If the controller is in the middle of resetting, don't
471 		 *  try to submit queued requests here - let the reset logic
472 		 *  handle that instead.
473 		 */
474 		if (!STAILQ_EMPTY(&qpair->queued_req) &&
475 		    !qpair->ctrlr->is_resetting) {
476 			req = STAILQ_FIRST(&qpair->queued_req);
477 			STAILQ_REMOVE_HEAD(&qpair->queued_req, stailq);
478 			_nvme_qpair_submit_request(qpair, req);
479 		}
480 	}
481 
482 	mtx_unlock(&qpair->lock);
483 }
484 
485 static void
486 nvme_qpair_manual_complete_tracker(struct nvme_qpair *qpair,
487     struct nvme_tracker *tr, uint32_t sct, uint32_t sc, uint32_t dnr,
488     error_print_t print_on_error)
489 {
490 	struct nvme_completion	cpl;
491 
492 	memset(&cpl, 0, sizeof(cpl));
493 	cpl.sqid = qpair->id;
494 	cpl.cid = tr->cid;
495 	cpl.status |= (sct & NVME_STATUS_SCT_MASK) << NVME_STATUS_SCT_SHIFT;
496 	cpl.status |= (sc & NVME_STATUS_SC_MASK) << NVME_STATUS_SC_SHIFT;
497 	cpl.status |= (dnr & NVME_STATUS_DNR_MASK) << NVME_STATUS_DNR_SHIFT;
498 	nvme_qpair_complete_tracker(qpair, tr, &cpl, print_on_error);
499 }
500 
501 void
502 nvme_qpair_manual_complete_request(struct nvme_qpair *qpair,
503     struct nvme_request *req, uint32_t sct, uint32_t sc)
504 {
505 	struct nvme_completion	cpl;
506 	boolean_t		error;
507 
508 	memset(&cpl, 0, sizeof(cpl));
509 	cpl.sqid = qpair->id;
510 	cpl.status |= (sct & NVME_STATUS_SCT_MASK) << NVME_STATUS_SCT_SHIFT;
511 	cpl.status |= (sc & NVME_STATUS_SC_MASK) << NVME_STATUS_SC_SHIFT;
512 
513 	error = nvme_completion_is_error(&cpl);
514 
515 	if (error) {
516 		nvme_qpair_print_command(qpair, &req->cmd);
517 		nvme_qpair_print_completion(qpair, &cpl);
518 	}
519 
520 	if (req->cb_fn)
521 		req->cb_fn(req->cb_arg, &cpl);
522 
523 	nvme_free_request(req);
524 }
525 
526 bool
527 nvme_qpair_process_completions(struct nvme_qpair *qpair)
528 {
529 	struct nvme_tracker	*tr;
530 	struct nvme_completion	cpl;
531 	int done = 0;
532 	bool in_panic = dumping || SCHEDULER_STOPPED();
533 
534 	qpair->num_intr_handler_calls++;
535 
536 	/*
537 	 * qpair is not enabled, likely because a controller reset is is in
538 	 * progress.  Ignore the interrupt - any I/O that was associated with
539 	 * this interrupt will get retried when the reset is complete.
540 	 */
541 	if (!qpair->is_enabled)
542 		return (false);
543 
544 	/*
545 	 * A panic can stop the CPU this routine is running on at any point.  If
546 	 * we're called during a panic, complete the sq_head wrap protocol for
547 	 * the case where we are interrupted just after the increment at 1
548 	 * below, but before we can reset cq_head to zero at 2. Also cope with
549 	 * the case where we do the zero at 2, but may or may not have done the
550 	 * phase adjustment at step 3. The panic machinery flushes all pending
551 	 * memory writes, so we can make these strong ordering assumptions
552 	 * that would otherwise be unwise if we were racing in real time.
553 	 */
554 	if (__predict_false(in_panic)) {
555 		if (qpair->cq_head == qpair->num_entries) {
556 			/*
557 			 * Here we know that we need to zero cq_head and then negate
558 			 * the phase, which hasn't been assigned if cq_head isn't
559 			 * zero due to the atomic_store_rel.
560 			 */
561 			qpair->cq_head = 0;
562 			qpair->phase = !qpair->phase;
563 		} else if (qpair->cq_head == 0) {
564 			/*
565 			 * In this case, we know that the assignment at 2
566 			 * happened below, but we don't know if it 3 happened or
567 			 * not. To do this, we look at the last completion
568 			 * entry and set the phase to the opposite phase
569 			 * that it has. This gets us back in sync
570 			 */
571 			cpl = qpair->cpl[qpair->num_entries - 1];
572 			nvme_completion_swapbytes(&cpl);
573 			qpair->phase = !NVME_STATUS_GET_P(cpl.status);
574 		}
575 	}
576 
577 	bus_dmamap_sync(qpair->dma_tag, qpair->queuemem_map,
578 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
579 	while (1) {
580 		cpl = qpair->cpl[qpair->cq_head];
581 
582 		/* Convert to host endian */
583 		nvme_completion_swapbytes(&cpl);
584 
585 		if (NVME_STATUS_GET_P(cpl.status) != qpair->phase)
586 			break;
587 
588 		tr = qpair->act_tr[cpl.cid];
589 
590 		if (tr != NULL) {
591 			nvme_qpair_complete_tracker(qpair, tr, &cpl, ERROR_PRINT_ALL);
592 			qpair->sq_head = cpl.sqhd;
593 			done++;
594 		} else if (!in_panic) {
595 			/*
596 			 * A missing tracker is normally an error.  However, a
597 			 * panic can stop the CPU this routine is running on
598 			 * after completing an I/O but before updating
599 			 * qpair->cq_head at 1 below.  Later, we re-enter this
600 			 * routine to poll I/O associated with the kernel
601 			 * dump. We find that the tr has been set to null before
602 			 * calling the completion routine.  If it hasn't
603 			 * completed (or it triggers a panic), then '1' below
604 			 * won't have updated cq_head. Rather than panic again,
605 			 * ignore this condition because it's not unexpected.
606 			 */
607 			nvme_printf(qpair->ctrlr,
608 			    "cpl does not map to outstanding cmd\n");
609 			/* nvme_dump_completion expects device endianess */
610 			nvme_dump_completion(&qpair->cpl[qpair->cq_head]);
611 			KASSERT(0, ("received completion for unknown cmd"));
612 		}
613 
614 		/*
615 		 * There's a number of races with the following (see above) when
616 		 * the system panics. We compensate for each one of them by
617 		 * using the atomic store to force strong ordering (at least when
618 		 * viewed in the aftermath of a panic).
619 		 */
620 		if (++qpair->cq_head == qpair->num_entries) {		/* 1 */
621 			atomic_store_rel_int(&qpair->cq_head, 0);	/* 2 */
622 			qpair->phase = !qpair->phase;			/* 3 */
623 		}
624 
625 		nvme_mmio_write_4(qpair->ctrlr, doorbell[qpair->id].cq_hdbl,
626 		    qpair->cq_head);
627 	}
628 	return (done != 0);
629 }
630 
631 static void
632 nvme_qpair_msix_handler(void *arg)
633 {
634 	struct nvme_qpair *qpair = arg;
635 
636 	nvme_qpair_process_completions(qpair);
637 }
638 
639 int
640 nvme_qpair_construct(struct nvme_qpair *qpair, uint32_t id,
641     uint16_t vector, uint32_t num_entries, uint32_t num_trackers,
642     struct nvme_controller *ctrlr)
643 {
644 	struct nvme_tracker	*tr;
645 	size_t			cmdsz, cplsz, prpsz, allocsz, prpmemsz;
646 	uint64_t		queuemem_phys, prpmem_phys, list_phys;
647 	uint8_t			*queuemem, *prpmem, *prp_list;
648 	int			i, err;
649 
650 	qpair->id = id;
651 	qpair->vector = vector;
652 	qpair->num_entries = num_entries;
653 	qpair->num_trackers = num_trackers;
654 	qpair->ctrlr = ctrlr;
655 
656 	if (ctrlr->msix_enabled) {
657 
658 		/*
659 		 * MSI-X vector resource IDs start at 1, so we add one to
660 		 *  the queue's vector to get the corresponding rid to use.
661 		 */
662 		qpair->rid = vector + 1;
663 
664 		qpair->res = bus_alloc_resource_any(ctrlr->dev, SYS_RES_IRQ,
665 		    &qpair->rid, RF_ACTIVE);
666 		bus_setup_intr(ctrlr->dev, qpair->res,
667 		    INTR_TYPE_MISC | INTR_MPSAFE, NULL,
668 		    nvme_qpair_msix_handler, qpair, &qpair->tag);
669 		if (id == 0) {
670 			bus_describe_intr(ctrlr->dev, qpair->res, qpair->tag,
671 			    "admin");
672 		} else {
673 			bus_describe_intr(ctrlr->dev, qpair->res, qpair->tag,
674 			    "io%d", id - 1);
675 		}
676 	}
677 
678 	mtx_init(&qpair->lock, "nvme qpair lock", NULL, MTX_DEF);
679 
680 	/* Note: NVMe PRP format is restricted to 4-byte alignment. */
681 	err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev),
682 	    4, PAGE_SIZE, BUS_SPACE_MAXADDR,
683 	    BUS_SPACE_MAXADDR, NULL, NULL, NVME_MAX_XFER_SIZE,
684 	    (NVME_MAX_XFER_SIZE/PAGE_SIZE)+1, PAGE_SIZE, 0,
685 	    NULL, NULL, &qpair->dma_tag_payload);
686 	if (err != 0) {
687 		nvme_printf(ctrlr, "payload tag create failed %d\n", err);
688 		goto out;
689 	}
690 
691 	/*
692 	 * Each component must be page aligned, and individual PRP lists
693 	 * cannot cross a page boundary.
694 	 */
695 	cmdsz = qpair->num_entries * sizeof(struct nvme_command);
696 	cmdsz = roundup2(cmdsz, PAGE_SIZE);
697 	cplsz = qpair->num_entries * sizeof(struct nvme_completion);
698 	cplsz = roundup2(cplsz, PAGE_SIZE);
699 	prpsz = sizeof(uint64_t) * NVME_MAX_PRP_LIST_ENTRIES;;
700 	prpmemsz = qpair->num_trackers * prpsz;
701 	allocsz = cmdsz + cplsz + prpmemsz;
702 
703 	err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev),
704 	    PAGE_SIZE, 0, BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
705 	    allocsz, 1, allocsz, 0, NULL, NULL, &qpair->dma_tag);
706 	if (err != 0) {
707 		nvme_printf(ctrlr, "tag create failed %d\n", err);
708 		goto out;
709 	}
710 
711 	if (bus_dmamem_alloc(qpair->dma_tag, (void **)&queuemem,
712 	    BUS_DMA_NOWAIT, &qpair->queuemem_map)) {
713 		nvme_printf(ctrlr, "failed to alloc qpair memory\n");
714 		goto out;
715 	}
716 
717 	if (bus_dmamap_load(qpair->dma_tag, qpair->queuemem_map,
718 	    queuemem, allocsz, nvme_single_map, &queuemem_phys, 0) != 0) {
719 		nvme_printf(ctrlr, "failed to load qpair memory\n");
720 		goto out;
721 	}
722 
723 	qpair->num_cmds = 0;
724 	qpair->num_intr_handler_calls = 0;
725 	qpair->num_retries = 0;
726 	qpair->num_failures = 0;
727 	qpair->cmd = (struct nvme_command *)queuemem;
728 	qpair->cpl = (struct nvme_completion *)(queuemem + cmdsz);
729 	prpmem = (uint8_t *)(queuemem + cmdsz + cplsz);
730 	qpair->cmd_bus_addr = queuemem_phys;
731 	qpair->cpl_bus_addr = queuemem_phys + cmdsz;
732 	prpmem_phys = queuemem_phys + cmdsz + cplsz;
733 
734 	qpair->sq_tdbl_off = nvme_mmio_offsetof(doorbell[id].sq_tdbl);
735 	qpair->cq_hdbl_off = nvme_mmio_offsetof(doorbell[id].cq_hdbl);
736 
737 	TAILQ_INIT(&qpair->free_tr);
738 	TAILQ_INIT(&qpair->outstanding_tr);
739 	STAILQ_INIT(&qpair->queued_req);
740 
741 	list_phys = prpmem_phys;
742 	prp_list = prpmem;
743 	for (i = 0; i < qpair->num_trackers; i++) {
744 
745 		if (list_phys + prpsz > prpmem_phys + prpmemsz) {
746 			qpair->num_trackers = i;
747 			break;
748 		}
749 
750 		/*
751 		 * Make sure that the PRP list for this tracker doesn't
752 		 * overflow to another page.
753 		 */
754 		if (trunc_page(list_phys) !=
755 		    trunc_page(list_phys + prpsz - 1)) {
756 			list_phys = roundup2(list_phys, PAGE_SIZE);
757 			prp_list =
758 			    (uint8_t *)roundup2((uintptr_t)prp_list, PAGE_SIZE);
759 		}
760 
761 		tr = malloc(sizeof(*tr), M_NVME, M_ZERO | M_WAITOK);
762 		bus_dmamap_create(qpair->dma_tag_payload, 0,
763 		    &tr->payload_dma_map);
764 		callout_init(&tr->timer, 1);
765 		tr->cid = i;
766 		tr->qpair = qpair;
767 		tr->prp = (uint64_t *)prp_list;
768 		tr->prp_bus_addr = list_phys;
769 		TAILQ_INSERT_HEAD(&qpair->free_tr, tr, tailq);
770 		list_phys += prpsz;
771 		prp_list += prpsz;
772 	}
773 
774 	if (qpair->num_trackers == 0) {
775 		nvme_printf(ctrlr, "failed to allocate enough trackers\n");
776 		goto out;
777 	}
778 
779 	qpair->act_tr = malloc(sizeof(struct nvme_tracker *) *
780 	    qpair->num_entries, M_NVME, M_ZERO | M_WAITOK);
781 	return (0);
782 
783 out:
784 	nvme_qpair_destroy(qpair);
785 	return (ENOMEM);
786 }
787 
788 static void
789 nvme_qpair_destroy(struct nvme_qpair *qpair)
790 {
791 	struct nvme_tracker	*tr;
792 
793 	if (qpair->tag)
794 		bus_teardown_intr(qpair->ctrlr->dev, qpair->res, qpair->tag);
795 
796 	if (mtx_initialized(&qpair->lock))
797 		mtx_destroy(&qpair->lock);
798 
799 	if (qpair->res)
800 		bus_release_resource(qpair->ctrlr->dev, SYS_RES_IRQ,
801 		    rman_get_rid(qpair->res), qpair->res);
802 
803 	if (qpair->cmd != NULL) {
804 		bus_dmamap_unload(qpair->dma_tag, qpair->queuemem_map);
805 		bus_dmamem_free(qpair->dma_tag, qpair->cmd,
806 		    qpair->queuemem_map);
807 	}
808 
809 	if (qpair->act_tr)
810 		free(qpair->act_tr, M_NVME);
811 
812 	while (!TAILQ_EMPTY(&qpair->free_tr)) {
813 		tr = TAILQ_FIRST(&qpair->free_tr);
814 		TAILQ_REMOVE(&qpair->free_tr, tr, tailq);
815 		bus_dmamap_destroy(qpair->dma_tag_payload,
816 		    tr->payload_dma_map);
817 		free(tr, M_NVME);
818 	}
819 
820 	if (qpair->dma_tag)
821 		bus_dma_tag_destroy(qpair->dma_tag);
822 
823 	if (qpair->dma_tag_payload)
824 		bus_dma_tag_destroy(qpair->dma_tag_payload);
825 }
826 
827 static void
828 nvme_admin_qpair_abort_aers(struct nvme_qpair *qpair)
829 {
830 	struct nvme_tracker	*tr;
831 
832 	tr = TAILQ_FIRST(&qpair->outstanding_tr);
833 	while (tr != NULL) {
834 		if (tr->req->cmd.opc == NVME_OPC_ASYNC_EVENT_REQUEST) {
835 			nvme_qpair_manual_complete_tracker(qpair, tr,
836 			    NVME_SCT_GENERIC, NVME_SC_ABORTED_SQ_DELETION, 0,
837 			    ERROR_PRINT_NONE);
838 			tr = TAILQ_FIRST(&qpair->outstanding_tr);
839 		} else {
840 			tr = TAILQ_NEXT(tr, tailq);
841 		}
842 	}
843 }
844 
845 void
846 nvme_admin_qpair_destroy(struct nvme_qpair *qpair)
847 {
848 
849 	nvme_admin_qpair_abort_aers(qpair);
850 	nvme_qpair_destroy(qpair);
851 }
852 
853 void
854 nvme_io_qpair_destroy(struct nvme_qpair *qpair)
855 {
856 
857 	nvme_qpair_destroy(qpair);
858 }
859 
860 static void
861 nvme_abort_complete(void *arg, const struct nvme_completion *status)
862 {
863 	struct nvme_tracker	*tr = arg;
864 
865 	/*
866 	 * If cdw0 == 1, the controller was not able to abort the command
867 	 *  we requested.  We still need to check the active tracker array,
868 	 *  to cover race where I/O timed out at same time controller was
869 	 *  completing the I/O.
870 	 */
871 	if (status->cdw0 == 1 && tr->qpair->act_tr[tr->cid] != NULL) {
872 		/*
873 		 * An I/O has timed out, and the controller was unable to
874 		 *  abort it for some reason.  Construct a fake completion
875 		 *  status, and then complete the I/O's tracker manually.
876 		 */
877 		nvme_printf(tr->qpair->ctrlr,
878 		    "abort command failed, aborting command manually\n");
879 		nvme_qpair_manual_complete_tracker(tr->qpair, tr,
880 		    NVME_SCT_GENERIC, NVME_SC_ABORTED_BY_REQUEST, 0, ERROR_PRINT_ALL);
881 	}
882 }
883 
884 static void
885 nvme_timeout(void *arg)
886 {
887 	struct nvme_tracker	*tr = arg;
888 	struct nvme_qpair	*qpair = tr->qpair;
889 	struct nvme_controller	*ctrlr = qpair->ctrlr;
890 	uint32_t		csts;
891 	uint8_t			cfs;
892 
893 	/*
894 	 * Read csts to get value of cfs - controller fatal status.
895 	 * If no fatal status, try to call the completion routine, and
896 	 * if completes transactions, report a missed interrupt and
897 	 * return (this may need to be rate limited). Otherwise, if
898 	 * aborts are enabled and the controller is not reporting
899 	 * fatal status, abort the command. Otherwise, just reset the
900 	 * controller and hope for the best.
901 	 */
902 	csts = nvme_mmio_read_4(ctrlr, csts);
903 	cfs = (csts >> NVME_CSTS_REG_CFS_SHIFT) & NVME_CSTS_REG_CFS_MASK;
904 	if (cfs == 0 && nvme_qpair_process_completions(qpair)) {
905 		nvme_printf(ctrlr, "Missing interrupt\n");
906 		return;
907 	}
908 	if (ctrlr->enable_aborts && cfs == 0) {
909 		nvme_printf(ctrlr, "Aborting command due to a timeout.\n");
910 		nvme_ctrlr_cmd_abort(ctrlr, tr->cid, qpair->id,
911 		    nvme_abort_complete, tr);
912 	} else {
913 		nvme_printf(ctrlr, "Resetting controller due to a timeout%s.\n",
914 		    (csts == 0xffffffff) ? " and possible hot unplug" :
915 		    (cfs ? " and fatal error status" : ""));
916 		nvme_ctrlr_reset(ctrlr);
917 	}
918 }
919 
920 void
921 nvme_qpair_submit_tracker(struct nvme_qpair *qpair, struct nvme_tracker *tr)
922 {
923 	struct nvme_request	*req;
924 	struct nvme_controller	*ctrlr;
925 
926 	mtx_assert(&qpair->lock, MA_OWNED);
927 
928 	req = tr->req;
929 	req->cmd.cid = tr->cid;
930 	qpair->act_tr[tr->cid] = tr;
931 	ctrlr = qpair->ctrlr;
932 
933 	if (req->timeout)
934 		callout_reset_curcpu(&tr->timer, ctrlr->timeout_period * hz,
935 		    nvme_timeout, tr);
936 
937 	/* Copy the command from the tracker to the submission queue. */
938 	memcpy(&qpair->cmd[qpair->sq_tail], &req->cmd, sizeof(req->cmd));
939 
940 	if (++qpair->sq_tail == qpair->num_entries)
941 		qpair->sq_tail = 0;
942 
943 	bus_dmamap_sync(qpair->dma_tag, qpair->queuemem_map,
944 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
945 #ifndef __powerpc__
946 	/*
947 	 * powerpc's bus_dmamap_sync() already includes a heavyweight sync, but
948 	 * no other archs do.
949 	 */
950 	wmb();
951 #endif
952 
953 	nvme_mmio_write_4(qpair->ctrlr, doorbell[qpair->id].sq_tdbl,
954 	    qpair->sq_tail);
955 
956 	qpair->num_cmds++;
957 }
958 
959 static void
960 nvme_payload_map(void *arg, bus_dma_segment_t *seg, int nseg, int error)
961 {
962 	struct nvme_tracker 	*tr = arg;
963 	uint32_t		cur_nseg;
964 
965 	/*
966 	 * If the mapping operation failed, return immediately.  The caller
967 	 *  is responsible for detecting the error status and failing the
968 	 *  tracker manually.
969 	 */
970 	if (error != 0) {
971 		nvme_printf(tr->qpair->ctrlr,
972 		    "nvme_payload_map err %d\n", error);
973 		return;
974 	}
975 
976 	/*
977 	 * Note that we specified PAGE_SIZE for alignment and max
978 	 *  segment size when creating the bus dma tags.  So here
979 	 *  we can safely just transfer each segment to its
980 	 *  associated PRP entry.
981 	 */
982 	tr->req->cmd.prp1 = htole64(seg[0].ds_addr);
983 
984 	if (nseg == 2) {
985 		tr->req->cmd.prp2 = htole64(seg[1].ds_addr);
986 	} else if (nseg > 2) {
987 		cur_nseg = 1;
988 		tr->req->cmd.prp2 = htole64((uint64_t)tr->prp_bus_addr);
989 		while (cur_nseg < nseg) {
990 			tr->prp[cur_nseg-1] =
991 			    htole64((uint64_t)seg[cur_nseg].ds_addr);
992 			cur_nseg++;
993 		}
994 	} else {
995 		/*
996 		 * prp2 should not be used by the controller
997 		 *  since there is only one segment, but set
998 		 *  to 0 just to be safe.
999 		 */
1000 		tr->req->cmd.prp2 = 0;
1001 	}
1002 
1003 	bus_dmamap_sync(tr->qpair->dma_tag_payload, tr->payload_dma_map,
1004 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1005 	nvme_qpair_submit_tracker(tr->qpair, tr);
1006 }
1007 
1008 static void
1009 _nvme_qpair_submit_request(struct nvme_qpair *qpair, struct nvme_request *req)
1010 {
1011 	struct nvme_tracker	*tr;
1012 	int			err = 0;
1013 
1014 	mtx_assert(&qpair->lock, MA_OWNED);
1015 
1016 	tr = TAILQ_FIRST(&qpair->free_tr);
1017 	req->qpair = qpair;
1018 
1019 	if (tr == NULL || !qpair->is_enabled) {
1020 		/*
1021 		 * No tracker is available, or the qpair is disabled due to
1022 		 *  an in-progress controller-level reset or controller
1023 		 *  failure.
1024 		 */
1025 
1026 		if (qpair->ctrlr->is_failed) {
1027 			/*
1028 			 * The controller has failed.  Post the request to a
1029 			 *  task where it will be aborted, so that we do not
1030 			 *  invoke the request's callback in the context
1031 			 *  of the submission.
1032 			 */
1033 			nvme_ctrlr_post_failed_request(qpair->ctrlr, req);
1034 		} else {
1035 			/*
1036 			 * Put the request on the qpair's request queue to be
1037 			 *  processed when a tracker frees up via a command
1038 			 *  completion or when the controller reset is
1039 			 *  completed.
1040 			 */
1041 			STAILQ_INSERT_TAIL(&qpair->queued_req, req, stailq);
1042 		}
1043 		return;
1044 	}
1045 
1046 	TAILQ_REMOVE(&qpair->free_tr, tr, tailq);
1047 	TAILQ_INSERT_TAIL(&qpair->outstanding_tr, tr, tailq);
1048 	tr->req = req;
1049 
1050 	switch (req->type) {
1051 	case NVME_REQUEST_VADDR:
1052 		KASSERT(req->payload_size <= qpair->ctrlr->max_xfer_size,
1053 		    ("payload_size (%d) exceeds max_xfer_size (%d)\n",
1054 		    req->payload_size, qpair->ctrlr->max_xfer_size));
1055 		err = bus_dmamap_load(tr->qpair->dma_tag_payload,
1056 		    tr->payload_dma_map, req->u.payload, req->payload_size,
1057 		    nvme_payload_map, tr, 0);
1058 		if (err != 0)
1059 			nvme_printf(qpair->ctrlr,
1060 			    "bus_dmamap_load returned 0x%x!\n", err);
1061 		break;
1062 	case NVME_REQUEST_NULL:
1063 		nvme_qpair_submit_tracker(tr->qpair, tr);
1064 		break;
1065 	case NVME_REQUEST_BIO:
1066 		KASSERT(req->u.bio->bio_bcount <= qpair->ctrlr->max_xfer_size,
1067 		    ("bio->bio_bcount (%jd) exceeds max_xfer_size (%d)\n",
1068 		    (intmax_t)req->u.bio->bio_bcount,
1069 		    qpair->ctrlr->max_xfer_size));
1070 		err = bus_dmamap_load_bio(tr->qpair->dma_tag_payload,
1071 		    tr->payload_dma_map, req->u.bio, nvme_payload_map, tr, 0);
1072 		if (err != 0)
1073 			nvme_printf(qpair->ctrlr,
1074 			    "bus_dmamap_load_bio returned 0x%x!\n", err);
1075 		break;
1076 	case NVME_REQUEST_CCB:
1077 		err = bus_dmamap_load_ccb(tr->qpair->dma_tag_payload,
1078 		    tr->payload_dma_map, req->u.payload,
1079 		    nvme_payload_map, tr, 0);
1080 		if (err != 0)
1081 			nvme_printf(qpair->ctrlr,
1082 			    "bus_dmamap_load_ccb returned 0x%x!\n", err);
1083 		break;
1084 	default:
1085 		panic("unknown nvme request type 0x%x\n", req->type);
1086 		break;
1087 	}
1088 
1089 	if (err != 0) {
1090 		/*
1091 		 * The dmamap operation failed, so we manually fail the
1092 		 *  tracker here with DATA_TRANSFER_ERROR status.
1093 		 *
1094 		 * nvme_qpair_manual_complete_tracker must not be called
1095 		 *  with the qpair lock held.
1096 		 */
1097 		mtx_unlock(&qpair->lock);
1098 		nvme_qpair_manual_complete_tracker(qpair, tr, NVME_SCT_GENERIC,
1099 		    NVME_SC_DATA_TRANSFER_ERROR, DO_NOT_RETRY, ERROR_PRINT_ALL);
1100 		mtx_lock(&qpair->lock);
1101 	}
1102 }
1103 
1104 void
1105 nvme_qpair_submit_request(struct nvme_qpair *qpair, struct nvme_request *req)
1106 {
1107 
1108 	mtx_lock(&qpair->lock);
1109 	_nvme_qpair_submit_request(qpair, req);
1110 	mtx_unlock(&qpair->lock);
1111 }
1112 
1113 static void
1114 nvme_qpair_enable(struct nvme_qpair *qpair)
1115 {
1116 
1117 	qpair->is_enabled = TRUE;
1118 }
1119 
1120 void
1121 nvme_qpair_reset(struct nvme_qpair *qpair)
1122 {
1123 
1124 	qpair->sq_head = qpair->sq_tail = qpair->cq_head = 0;
1125 
1126 	/*
1127 	 * First time through the completion queue, HW will set phase
1128 	 *  bit on completions to 1.  So set this to 1 here, indicating
1129 	 *  we're looking for a 1 to know which entries have completed.
1130 	 *  we'll toggle the bit each time when the completion queue
1131 	 *  rolls over.
1132 	 */
1133 	qpair->phase = 1;
1134 
1135 	memset(qpair->cmd, 0,
1136 	    qpair->num_entries * sizeof(struct nvme_command));
1137 	memset(qpair->cpl, 0,
1138 	    qpair->num_entries * sizeof(struct nvme_completion));
1139 }
1140 
1141 void
1142 nvme_admin_qpair_enable(struct nvme_qpair *qpair)
1143 {
1144 	struct nvme_tracker		*tr;
1145 	struct nvme_tracker		*tr_temp;
1146 
1147 	/*
1148 	 * Manually abort each outstanding admin command.  Do not retry
1149 	 *  admin commands found here, since they will be left over from
1150 	 *  a controller reset and its likely the context in which the
1151 	 *  command was issued no longer applies.
1152 	 */
1153 	TAILQ_FOREACH_SAFE(tr, &qpair->outstanding_tr, tailq, tr_temp) {
1154 		nvme_printf(qpair->ctrlr,
1155 		    "aborting outstanding admin command\n");
1156 		nvme_qpair_manual_complete_tracker(qpair, tr, NVME_SCT_GENERIC,
1157 		    NVME_SC_ABORTED_BY_REQUEST, DO_NOT_RETRY, ERROR_PRINT_ALL);
1158 	}
1159 
1160 	nvme_qpair_enable(qpair);
1161 }
1162 
1163 void
1164 nvme_io_qpair_enable(struct nvme_qpair *qpair)
1165 {
1166 	STAILQ_HEAD(, nvme_request)	temp;
1167 	struct nvme_tracker		*tr;
1168 	struct nvme_tracker		*tr_temp;
1169 	struct nvme_request		*req;
1170 
1171 	/*
1172 	 * Manually abort each outstanding I/O.  This normally results in a
1173 	 *  retry, unless the retry count on the associated request has
1174 	 *  reached its limit.
1175 	 */
1176 	TAILQ_FOREACH_SAFE(tr, &qpair->outstanding_tr, tailq, tr_temp) {
1177 		nvme_printf(qpair->ctrlr, "aborting outstanding i/o\n");
1178 		nvme_qpair_manual_complete_tracker(qpair, tr, NVME_SCT_GENERIC,
1179 		    NVME_SC_ABORTED_BY_REQUEST, 0, ERROR_PRINT_NO_RETRY);
1180 	}
1181 
1182 	mtx_lock(&qpair->lock);
1183 
1184 	nvme_qpair_enable(qpair);
1185 
1186 	STAILQ_INIT(&temp);
1187 	STAILQ_SWAP(&qpair->queued_req, &temp, nvme_request);
1188 
1189 	while (!STAILQ_EMPTY(&temp)) {
1190 		req = STAILQ_FIRST(&temp);
1191 		STAILQ_REMOVE_HEAD(&temp, stailq);
1192 		nvme_printf(qpair->ctrlr, "resubmitting queued i/o\n");
1193 		nvme_qpair_print_command(qpair, &req->cmd);
1194 		_nvme_qpair_submit_request(qpair, req);
1195 	}
1196 
1197 	mtx_unlock(&qpair->lock);
1198 }
1199 
1200 static void
1201 nvme_qpair_disable(struct nvme_qpair *qpair)
1202 {
1203 	struct nvme_tracker *tr;
1204 
1205 	qpair->is_enabled = FALSE;
1206 	mtx_lock(&qpair->lock);
1207 	TAILQ_FOREACH(tr, &qpair->outstanding_tr, tailq)
1208 		callout_stop(&tr->timer);
1209 	mtx_unlock(&qpair->lock);
1210 }
1211 
1212 void
1213 nvme_admin_qpair_disable(struct nvme_qpair *qpair)
1214 {
1215 
1216 	nvme_qpair_disable(qpair);
1217 	nvme_admin_qpair_abort_aers(qpair);
1218 }
1219 
1220 void
1221 nvme_io_qpair_disable(struct nvme_qpair *qpair)
1222 {
1223 
1224 	nvme_qpair_disable(qpair);
1225 }
1226 
1227 void
1228 nvme_qpair_fail(struct nvme_qpair *qpair)
1229 {
1230 	struct nvme_tracker		*tr;
1231 	struct nvme_request		*req;
1232 
1233 	if (!mtx_initialized(&qpair->lock))
1234 		return;
1235 
1236 	mtx_lock(&qpair->lock);
1237 
1238 	while (!STAILQ_EMPTY(&qpair->queued_req)) {
1239 		req = STAILQ_FIRST(&qpair->queued_req);
1240 		STAILQ_REMOVE_HEAD(&qpair->queued_req, stailq);
1241 		nvme_printf(qpair->ctrlr, "failing queued i/o\n");
1242 		mtx_unlock(&qpair->lock);
1243 		nvme_qpair_manual_complete_request(qpair, req, NVME_SCT_GENERIC,
1244 		    NVME_SC_ABORTED_BY_REQUEST);
1245 		mtx_lock(&qpair->lock);
1246 	}
1247 
1248 	/* Manually abort each outstanding I/O. */
1249 	while (!TAILQ_EMPTY(&qpair->outstanding_tr)) {
1250 		tr = TAILQ_FIRST(&qpair->outstanding_tr);
1251 		/*
1252 		 * Do not remove the tracker.  The abort_tracker path will
1253 		 *  do that for us.
1254 		 */
1255 		nvme_printf(qpair->ctrlr, "failing outstanding i/o\n");
1256 		mtx_unlock(&qpair->lock);
1257 		nvme_qpair_manual_complete_tracker(qpair, tr, NVME_SCT_GENERIC,
1258 		    NVME_SC_ABORTED_BY_REQUEST, DO_NOT_RETRY, ERROR_PRINT_ALL);
1259 		mtx_lock(&qpair->lock);
1260 	}
1261 
1262 	mtx_unlock(&qpair->lock);
1263 }
1264 
1265