sys/dev/nvme/nvme_qpair.c

   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         { 0xFFFF, "ADMIN COMMAND" }
  81 };
  82
  83 static struct nvme_opcode_string io_opcode[] = {
  84         { NVME_OPC_FLUSH, "FLUSH" },
  85         { NVME_OPC_WRITE, "WRITE" },
  86         { NVME_OPC_READ, "READ" },
  87         { NVME_OPC_WRITE_UNCORRECTABLE, "WRITE UNCORRECTABLE" },
  88         { NVME_OPC_COMPARE, "COMPARE" },
  89         { NVME_OPC_WRITE_ZEROES, "WRITE ZEROES" },
  90         { NVME_OPC_DATASET_MANAGEMENT, "DATASET MANAGEMENT" },
  91         { NVME_OPC_RESERVATION_REGISTER, "RESERVATION REGISTER" },
  92         { NVME_OPC_RESERVATION_REPORT, "RESERVATION REPORT" },
  93         { NVME_OPC_RESERVATION_ACQUIRE, "RESERVATION ACQUIRE" },
  94         { NVME_OPC_RESERVATION_RELEASE, "RESERVATION RELEASE" },
  95         { 0xFFFF, "IO COMMAND" }
  96 };
  97
  98 static const char *
  99 get_admin_opcode_string(uint16_t opc)
 100 {
 101         struct nvme_opcode_string *entry;
 102
 103         entry = admin_opcode;
 104
 105         while (entry->opc != 0xFFFF) {
 106                 if (entry->opc == opc)
 107                         return (entry->str);
 108                 entry++;
 109         }
 110         return (entry->str);
 111 }
 112
 113 static const char *
 114 get_io_opcode_string(uint16_t opc)
 115 {
 116         struct nvme_opcode_string *entry;
 117
 118         entry = io_opcode;
 119
 120         while (entry->opc != 0xFFFF) {
 121                 if (entry->opc == opc)
 122                         return (entry->str);
 123                 entry++;
 124         }
 125         return (entry->str);
 126 }
 127
 128
 129 static void
 130 nvme_admin_qpair_print_command(struct nvme_qpair *qpair,
 131     struct nvme_command *cmd)
 132 {
 133
 134         nvme_printf(qpair->ctrlr, "%s (%02x) sqid:%d cid:%d nsid:%x "
 135             "cdw10:%08x cdw11:%08x\n",
 136             get_admin_opcode_string(cmd->opc), cmd->opc, qpair->id, cmd->cid,
 137             le32toh(cmd->nsid), le32toh(cmd->cdw10), le32toh(cmd->cdw11));
 138 }
 139
 140 static void
 141 nvme_io_qpair_print_command(struct nvme_qpair *qpair,
 142     struct nvme_command *cmd)
 143 {
 144
 145         switch (cmd->opc) {
 146         case NVME_OPC_WRITE:
 147         case NVME_OPC_READ:
 148         case NVME_OPC_WRITE_UNCORRECTABLE:
 149         case NVME_OPC_COMPARE:
 150         case NVME_OPC_WRITE_ZEROES:
 151                 nvme_printf(qpair->ctrlr, "%s sqid:%d cid:%d nsid:%d "
 152                     "lba:%llu len:%d\n",
 153                     get_io_opcode_string(cmd->opc), qpair->id, cmd->cid, le32toh(cmd->nsid),
 154                     ((unsigned long long)le32toh(cmd->cdw11) << 32) + le32toh(cmd->cdw10),
 155                     (le32toh(cmd->cdw12) & 0xFFFF) + 1);
 156                 break;
 157         case NVME_OPC_FLUSH:
 158         case NVME_OPC_DATASET_MANAGEMENT:
 159         case NVME_OPC_RESERVATION_REGISTER:
 160         case NVME_OPC_RESERVATION_REPORT:
 161         case NVME_OPC_RESERVATION_ACQUIRE:
 162         case NVME_OPC_RESERVATION_RELEASE:
 163                 nvme_printf(qpair->ctrlr, "%s sqid:%d cid:%d nsid:%d\n",
 164                     get_io_opcode_string(cmd->opc), qpair->id, cmd->cid, le32toh(cmd->nsid));
 165                 break;
 166         default:
 167                 nvme_printf(qpair->ctrlr, "%s (%02x) sqid:%d cid:%d nsid:%d\n",
 168                     get_io_opcode_string(cmd->opc), cmd->opc, qpair->id,
 169                     cmd->cid, le32toh(cmd->nsid));
 170                 break;
 171         }
 172 }
 173
 174 static void
 175 nvme_qpair_print_command(struct nvme_qpair *qpair, struct nvme_command *cmd)
 176 {
 177         if (qpair->id == 0)
 178                 nvme_admin_qpair_print_command(qpair, cmd);
 179         else
 180                 nvme_io_qpair_print_command(qpair, cmd);
 181 }
 182
 183 struct nvme_status_string {
 184
 185         uint16_t        sc;
 186         const char *    str;
 187 };
 188
 189 static struct nvme_status_string generic_status[] = {
 190         { NVME_SC_SUCCESS, "SUCCESS" },
 191         { NVME_SC_INVALID_OPCODE, "INVALID OPCODE" },
 192         { NVME_SC_INVALID_FIELD, "INVALID_FIELD" },
 193         { NVME_SC_COMMAND_ID_CONFLICT, "COMMAND ID CONFLICT" },
 194         { NVME_SC_DATA_TRANSFER_ERROR, "DATA TRANSFER ERROR" },
 195         { NVME_SC_ABORTED_POWER_LOSS, "ABORTED - POWER LOSS" },
 196         { NVME_SC_INTERNAL_DEVICE_ERROR, "INTERNAL DEVICE ERROR" },
 197         { NVME_SC_ABORTED_BY_REQUEST, "ABORTED - BY REQUEST" },
 198         { NVME_SC_ABORTED_SQ_DELETION, "ABORTED - SQ DELETION" },
 199         { NVME_SC_ABORTED_FAILED_FUSED, "ABORTED - FAILED FUSED" },
 200         { NVME_SC_ABORTED_MISSING_FUSED, "ABORTED - MISSING FUSED" },
 201         { NVME_SC_INVALID_NAMESPACE_OR_FORMAT, "INVALID NAMESPACE OR FORMAT" },
 202         { NVME_SC_COMMAND_SEQUENCE_ERROR, "COMMAND SEQUENCE ERROR" },
 203         { NVME_SC_INVALID_SGL_SEGMENT_DESCR, "INVALID SGL SEGMENT DESCRIPTOR" },
 204         { NVME_SC_INVALID_NUMBER_OF_SGL_DESCR, "INVALID NUMBER OF SGL DESCRIPTORS" },
 205         { NVME_SC_DATA_SGL_LENGTH_INVALID, "DATA SGL LENGTH INVALID" },
 206         { NVME_SC_METADATA_SGL_LENGTH_INVALID, "METADATA SGL LENGTH INVALID" },
 207         { NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID, "SGL DESCRIPTOR TYPE INVALID" },
 208         { NVME_SC_INVALID_USE_OF_CMB, "INVALID USE OF CONTROLLER MEMORY BUFFER" },
 209         { NVME_SC_PRP_OFFET_INVALID, "PRP OFFET INVALID" },
 210         { NVME_SC_ATOMIC_WRITE_UNIT_EXCEEDED, "ATOMIC WRITE UNIT EXCEEDED" },
 211         { NVME_SC_OPERATION_DENIED, "OPERATION DENIED" },
 212         { NVME_SC_SGL_OFFSET_INVALID, "SGL OFFSET INVALID" },
 213         { NVME_SC_HOST_ID_INCONSISTENT_FORMAT, "HOST IDENTIFIER INCONSISTENT FORMAT" },
 214         { NVME_SC_KEEP_ALIVE_TIMEOUT_EXPIRED, "KEEP ALIVE TIMEOUT EXPIRED" },
 215         { NVME_SC_KEEP_ALIVE_TIMEOUT_INVALID, "KEEP ALIVE TIMEOUT INVALID" },
 216         { NVME_SC_ABORTED_DUE_TO_PREEMPT, "COMMAND ABORTED DUE TO PREEMPT AND ABORT" },
 217         { NVME_SC_SANITIZE_FAILED, "SANITIZE FAILED" },
 218         { NVME_SC_SANITIZE_IN_PROGRESS, "SANITIZE IN PROGRESS" },
 219         { NVME_SC_SGL_DATA_BLOCK_GRAN_INVALID, "SGL_DATA_BLOCK_GRANULARITY_INVALID" },
 220         { NVME_SC_NOT_SUPPORTED_IN_CMB, "COMMAND NOT SUPPORTED FOR QUEUE IN CMB" },
 221
 222         { NVME_SC_LBA_OUT_OF_RANGE, "LBA OUT OF RANGE" },
 223         { NVME_SC_CAPACITY_EXCEEDED, "CAPACITY EXCEEDED" },
 224         { NVME_SC_NAMESPACE_NOT_READY, "NAMESPACE NOT READY" },
 225         { NVME_SC_RESERVATION_CONFLICT, "RESERVATION CONFLICT" },
 226         { NVME_SC_FORMAT_IN_PROGRESS, "FORMAT IN PROGRESS" },
 227         { 0xFFFF, "GENERIC" }
 228 };
 229
 230 static struct nvme_status_string command_specific_status[] = {
 231         { NVME_SC_COMPLETION_QUEUE_INVALID, "INVALID COMPLETION QUEUE" },
 232         { NVME_SC_INVALID_QUEUE_IDENTIFIER, "INVALID QUEUE IDENTIFIER" },
 233         { NVME_SC_MAXIMUM_QUEUE_SIZE_EXCEEDED, "MAX QUEUE SIZE EXCEEDED" },
 234         { NVME_SC_ABORT_COMMAND_LIMIT_EXCEEDED, "ABORT CMD LIMIT EXCEEDED" },
 235         { NVME_SC_ASYNC_EVENT_REQUEST_LIMIT_EXCEEDED, "ASYNC LIMIT EXCEEDED" },
 236         { NVME_SC_INVALID_FIRMWARE_SLOT, "INVALID FIRMWARE SLOT" },
 237         { NVME_SC_INVALID_FIRMWARE_IMAGE, "INVALID FIRMWARE IMAGE" },
 238         { NVME_SC_INVALID_INTERRUPT_VECTOR, "INVALID INTERRUPT VECTOR" },
 239         { NVME_SC_INVALID_LOG_PAGE, "INVALID LOG PAGE" },
 240         { NVME_SC_INVALID_FORMAT, "INVALID FORMAT" },
 241         { NVME_SC_FIRMWARE_REQUIRES_RESET, "FIRMWARE REQUIRES RESET" },
 242         { NVME_SC_INVALID_QUEUE_DELETION, "INVALID QUEUE DELETION" },
 243         { NVME_SC_FEATURE_NOT_SAVEABLE, "FEATURE IDENTIFIER NOT SAVEABLE" },
 244         { NVME_SC_FEATURE_NOT_CHANGEABLE, "FEATURE NOT CHANGEABLE" },
 245         { NVME_SC_FEATURE_NOT_NS_SPECIFIC, "FEATURE NOT NAMESPACE SPECIFIC" },
 246         { NVME_SC_FW_ACT_REQUIRES_NVMS_RESET, "FIRMWARE ACTIVATION REQUIRES NVM SUBSYSTEM RESET" },
 247         { NVME_SC_FW_ACT_REQUIRES_RESET, "FIRMWARE ACTIVATION REQUIRES RESET" },
 248         { NVME_SC_FW_ACT_REQUIRES_TIME, "FIRMWARE ACTIVATION REQUIRES MAXIMUM TIME VIOLATION" },
 249         { NVME_SC_FW_ACT_PROHIBITED, "FIRMWARE ACTIVATION PROHIBITED" },
 250         { NVME_SC_OVERLAPPING_RANGE, "OVERLAPPING RANGE" },
 251         { NVME_SC_NS_INSUFFICIENT_CAPACITY, "NAMESPACE INSUFFICIENT CAPACITY" },
 252         { NVME_SC_NS_ID_UNAVAILABLE, "NAMESPACE IDENTIFIER UNAVAILABLE" },
 253         { NVME_SC_NS_ALREADY_ATTACHED, "NAMESPACE ALREADY ATTACHED" },
 254         { NVME_SC_NS_IS_PRIVATE, "NAMESPACE IS PRIVATE" },
 255         { NVME_SC_NS_NOT_ATTACHED, "NS NOT ATTACHED" },
 256         { NVME_SC_THIN_PROV_NOT_SUPPORTED, "THIN PROVISIONING NOT SUPPORTED" },
 257         { NVME_SC_CTRLR_LIST_INVALID, "CONTROLLER LIST INVALID" },
 258         { NVME_SC_SELT_TEST_IN_PROGRESS, "DEVICE SELT-TEST IN PROGRESS" },
 259         { NVME_SC_BOOT_PART_WRITE_PROHIB, "BOOT PARTITION WRITE PROHIBITED" },
 260         { NVME_SC_INVALID_CTRLR_ID, "INVALID CONTROLLER IDENTIFIER" },
 261         { NVME_SC_INVALID_SEC_CTRLR_STATE, "INVALID SECONDARY CONTROLLER STATE" },
 262         { NVME_SC_INVALID_NUM_OF_CTRLR_RESRC, "INVALID NUMBER OF CONTROLLER RESOURCES" },
 263         { NVME_SC_INVALID_RESOURCE_ID, "INVALID RESOURCE IDENTIFIER" },
 264
 265         { NVME_SC_CONFLICTING_ATTRIBUTES, "CONFLICTING ATTRIBUTES" },
 266         { NVME_SC_INVALID_PROTECTION_INFO, "INVALID PROTECTION INFO" },
 267         { NVME_SC_ATTEMPTED_WRITE_TO_RO_PAGE, "WRITE TO RO PAGE" },
 268         { 0xFFFF, "COMMAND SPECIFIC" }
 269 };
 270
 271 static struct nvme_status_string media_error_status[] = {
 272         { NVME_SC_WRITE_FAULTS, "WRITE FAULTS" },
 273         { NVME_SC_UNRECOVERED_READ_ERROR, "UNRECOVERED READ ERROR" },
 274         { NVME_SC_GUARD_CHECK_ERROR, "GUARD CHECK ERROR" },
 275         { NVME_SC_APPLICATION_TAG_CHECK_ERROR, "APPLICATION TAG CHECK ERROR" },
 276         { NVME_SC_REFERENCE_TAG_CHECK_ERROR, "REFERENCE TAG CHECK ERROR" },
 277         { NVME_SC_COMPARE_FAILURE, "COMPARE FAILURE" },
 278         { NVME_SC_ACCESS_DENIED, "ACCESS DENIED" },
 279         { NVME_SC_DEALLOCATED_OR_UNWRITTEN, "DEALLOCATED OR UNWRITTEN LOGICAL BLOCK" },
 280         { 0xFFFF, "MEDIA ERROR" }
 281 };
 282
 283 static const char *
 284 get_status_string(uint16_t sct, uint16_t sc)
 285 {
 286         struct nvme_status_string *entry;
 287
 288         switch (sct) {
 289         case NVME_SCT_GENERIC:
 290                 entry = generic_status;
 291                 break;
 292         case NVME_SCT_COMMAND_SPECIFIC:
 293                 entry = command_specific_status;
 294                 break;
 295         case NVME_SCT_MEDIA_ERROR:
 296                 entry = media_error_status;
 297                 break;
 298         case NVME_SCT_VENDOR_SPECIFIC:
 299                 return ("VENDOR SPECIFIC");
 300         default:
 301                 return ("RESERVED");
 302         }
 303
 304         while (entry->sc != 0xFFFF) {
 305                 if (entry->sc == sc)
 306                         return (entry->str);
 307                 entry++;
 308         }
 309         return (entry->str);
 310 }
 311
 312 static void
 313 nvme_qpair_print_completion(struct nvme_qpair *qpair,
 314     struct nvme_completion *cpl)
 315 {
 316         uint16_t sct, sc;
 317
 318         sct = NVME_STATUS_GET_SCT(cpl->status);
 319         sc = NVME_STATUS_GET_SC(cpl->status);
 320
 321         nvme_printf(qpair->ctrlr, "%s (%02x/%02x) sqid:%d cid:%d cdw0:%x\n",
 322             get_status_string(sct, sc), sct, sc, cpl->sqid, cpl->cid,
 323             cpl->cdw0);
 324 }
 325
 326 static boolean_t
 327 nvme_completion_is_retry(const struct nvme_completion *cpl)
 328 {
 329         uint8_t sct, sc, dnr;
 330
 331         sct = NVME_STATUS_GET_SCT(cpl->status);
 332         sc = NVME_STATUS_GET_SC(cpl->status);
 333         dnr = NVME_STATUS_GET_DNR(cpl->status); /* Do Not Retry Bit */
 334
 335         /*
 336          * TODO: spec is not clear how commands that are aborted due
 337          *  to TLER will be marked.  So for now, it seems
 338          *  NAMESPACE_NOT_READY is the only case where we should
 339          *  look at the DNR bit. Requests failed with ABORTED_BY_REQUEST
 340          *  set the DNR bit correctly since the driver controls that.
 341          */
 342         switch (sct) {
 343         case NVME_SCT_GENERIC:
 344                 switch (sc) {
 345                 case NVME_SC_ABORTED_BY_REQUEST:
 346                 case NVME_SC_NAMESPACE_NOT_READY:
 347                         if (dnr)
 348                                 return (0);
 349                         else
 350                                 return (1);
 351                 case NVME_SC_INVALID_OPCODE:
 352                 case NVME_SC_INVALID_FIELD:
 353                 case NVME_SC_COMMAND_ID_CONFLICT:
 354                 case NVME_SC_DATA_TRANSFER_ERROR:
 355                 case NVME_SC_ABORTED_POWER_LOSS:
 356                 case NVME_SC_INTERNAL_DEVICE_ERROR:
 357                 case NVME_SC_ABORTED_SQ_DELETION:
 358                 case NVME_SC_ABORTED_FAILED_FUSED:
 359                 case NVME_SC_ABORTED_MISSING_FUSED:
 360                 case NVME_SC_INVALID_NAMESPACE_OR_FORMAT:
 361                 case NVME_SC_COMMAND_SEQUENCE_ERROR:
 362                 case NVME_SC_LBA_OUT_OF_RANGE:
 363                 case NVME_SC_CAPACITY_EXCEEDED:
 364                 default:
 365                         return (0);
 366                 }
 367         case NVME_SCT_COMMAND_SPECIFIC:
 368         case NVME_SCT_MEDIA_ERROR:
 369         case NVME_SCT_VENDOR_SPECIFIC:
 370         default:
 371                 return (0);
 372         }
 373 }
 374
 375 static void
 376 nvme_qpair_complete_tracker(struct nvme_qpair *qpair, struct nvme_tracker *tr,
 377     struct nvme_completion *cpl, error_print_t print_on_error)
 378 {
 379         struct nvme_request     *req;
 380         boolean_t               retry, error;
 381
 382         req = tr->req;
 383         error = nvme_completion_is_error(cpl);
 384         retry = error && nvme_completion_is_retry(cpl) &&
 385            req->retries < nvme_retry_count;
 386
 387         if (error && (print_on_error == ERROR_PRINT_ALL ||
 388                 (!retry && print_on_error == ERROR_PRINT_NO_RETRY))) {
 389                 nvme_qpair_print_command(qpair, &req->cmd);
 390                 nvme_qpair_print_completion(qpair, cpl);
 391         }
 392
 393         qpair->act_tr[cpl->cid] = NULL;
 394
 395         KASSERT(cpl->cid == req->cmd.cid, ("cpl cid does not match cmd cid\n"));
 396
 397         if (req->cb_fn && !retry)
 398                 req->cb_fn(req->cb_arg, cpl);
 399
 400         mtx_lock(&qpair->lock);
 401         callout_stop(&tr->timer);
 402
 403         if (retry) {
 404                 req->retries++;
 405                 nvme_qpair_submit_tracker(qpair, tr);
 406         } else {
 407                 if (req->type != NVME_REQUEST_NULL) {
 408                         bus_dmamap_sync(qpair->dma_tag_payload,
 409                             tr->payload_dma_map,
 410                             BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
 411                         bus_dmamap_unload(qpair->dma_tag_payload,
 412                             tr->payload_dma_map);
 413                 }
 414
 415                 nvme_free_request(req);
 416                 tr->req = NULL;
 417
 418                 TAILQ_REMOVE(&qpair->outstanding_tr, tr, tailq);
 419                 TAILQ_INSERT_HEAD(&qpair->free_tr, tr, tailq);
 420
 421                 /*
 422                  * If the controller is in the middle of resetting, don't
 423                  *  try to submit queued requests here - let the reset logic
 424                  *  handle that instead.
 425                  */
 426                 if (!STAILQ_EMPTY(&qpair->queued_req) &&
 427                     !qpair->ctrlr->is_resetting) {
 428                         req = STAILQ_FIRST(&qpair->queued_req);
 429                         STAILQ_REMOVE_HEAD(&qpair->queued_req, stailq);
 430                         _nvme_qpair_submit_request(qpair, req);
 431                 }
 432         }
 433
 434         mtx_unlock(&qpair->lock);
 435 }
 436
 437 static void
 438 nvme_qpair_manual_complete_tracker(struct nvme_qpair *qpair,
 439     struct nvme_tracker *tr, uint32_t sct, uint32_t sc, uint32_t dnr,
 440     error_print_t print_on_error)
 441 {
 442         struct nvme_completion  cpl;
 443
 444         memset(&cpl, 0, sizeof(cpl));
 445         cpl.sqid = qpair->id;
 446         cpl.cid = tr->cid;
 447         cpl.status |= (sct & NVME_STATUS_SCT_MASK) << NVME_STATUS_SCT_SHIFT;
 448         cpl.status |= (sc & NVME_STATUS_SC_MASK) << NVME_STATUS_SC_SHIFT;
 449         cpl.status |= (dnr & NVME_STATUS_DNR_MASK) << NVME_STATUS_DNR_SHIFT;
 450         nvme_qpair_complete_tracker(qpair, tr, &cpl, print_on_error);
 451 }
 452
 453 void
 454 nvme_qpair_manual_complete_request(struct nvme_qpair *qpair,
 455     struct nvme_request *req, uint32_t sct, uint32_t sc)
 456 {
 457         struct nvme_completion  cpl;
 458         boolean_t               error;
 459
 460         memset(&cpl, 0, sizeof(cpl));
 461         cpl.sqid = qpair->id;
 462         cpl.status |= (sct & NVME_STATUS_SCT_MASK) << NVME_STATUS_SCT_SHIFT;
 463         cpl.status |= (sc & NVME_STATUS_SC_MASK) << NVME_STATUS_SC_SHIFT;
 464
 465         error = nvme_completion_is_error(&cpl);
 466
 467         if (error) {
 468                 nvme_qpair_print_command(qpair, &req->cmd);
 469                 nvme_qpair_print_completion(qpair, &cpl);
 470         }
 471
 472         if (req->cb_fn)
 473                 req->cb_fn(req->cb_arg, &cpl);
 474
 475         nvme_free_request(req);
 476 }
 477
 478 bool
 479 nvme_qpair_process_completions(struct nvme_qpair *qpair)
 480 {
 481         struct nvme_tracker     *tr;
 482         struct nvme_completion  cpl;
 483         int done = 0;
 484         bool in_panic = dumping || SCHEDULER_STOPPED();
 485
 486         qpair->num_intr_handler_calls++;
 487
 488         /*
 489          * qpair is not enabled, likely because a controller reset is is in
 490          * progress.  Ignore the interrupt - any I/O that was associated with
 491          * this interrupt will get retried when the reset is complete.
 492          */
 493         if (!qpair->is_enabled)
 494                 return (false);
 495
 496         /*
 497          * A panic can stop the CPU this routine is running on at any point.  If
 498          * we're called during a panic, complete the sq_head wrap protocol for
 499          * the case where we are interrupted just after the increment at 1
 500          * below, but before we can reset cq_head to zero at 2. Also cope with
 501          * the case where we do the zero at 2, but may or may not have done the
 502          * phase adjustment at step 3. The panic machinery flushes all pending
 503          * memory writes, so we can make these strong ordering assumptions
 504          * that would otherwise be unwise if we were racing in real time.
 505          */
 506         if (__predict_false(in_panic)) {
 507                 if (qpair->cq_head == qpair->num_entries) {
 508                         /*
 509                          * Here we know that we need to zero cq_head and then negate
 510                          * the phase, which hasn't been assigned if cq_head isn't
 511                          * zero due to the atomic_store_rel.
 512                          */
 513                         qpair->cq_head = 0;
 514                         qpair->phase = !qpair->phase;
 515                 } else if (qpair->cq_head == 0) {
 516                         /*
 517                          * In this case, we know that the assignment at 2
 518                          * happened below, but we don't know if it 3 happened or
 519                          * not. To do this, we look at the last completion
 520                          * entry and set the phase to the opposite phase
 521                          * that it has. This gets us back in sync
 522                          */
 523                         cpl = qpair->cpl[qpair->num_entries - 1];
 524                         nvme_completion_swapbytes(&cpl);
 525                         qpair->phase = !NVME_STATUS_GET_P(cpl.status);
 526                 }
 527         }
 528
 529         bus_dmamap_sync(qpair->dma_tag, qpair->queuemem_map,
 530             BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
 531         while (1) {
 532                 cpl = qpair->cpl[qpair->cq_head];
 533
 534                 /* Convert to host endian */
 535                 nvme_completion_swapbytes(&cpl);
 536
 537                 if (NVME_STATUS_GET_P(cpl.status) != qpair->phase)
 538                         break;
 539
 540                 tr = qpair->act_tr[cpl.cid];
 541
 542                 if (tr != NULL) {
 543                         nvme_qpair_complete_tracker(qpair, tr, &cpl, ERROR_PRINT_ALL);
 544                         qpair->sq_head = cpl.sqhd;
 545                         done++;
 546                 } else if (!in_panic) {
 547                         /*
 548                          * A missing tracker is normally an error.  However, a
 549                          * panic can stop the CPU this routine is running on
 550                          * after completing an I/O but before updating
 551                          * qpair->cq_head at 1 below.  Later, we re-enter this
 552                          * routine to poll I/O associated with the kernel
 553                          * dump. We find that the tr has been set to null before
 554                          * calling the completion routine.  If it hasn't
 555                          * completed (or it triggers a panic), then '1' below
 556                          * won't have updated cq_head. Rather than panic again,
 557                          * ignore this condition because it's not unexpected.
 558                          */
 559                         nvme_printf(qpair->ctrlr,
 560                             "cpl does not map to outstanding cmd\n");
 561                         /* nvme_dump_completion expects device endianess */
 562                         nvme_dump_completion(&qpair->cpl[qpair->cq_head]);
 563                         KASSERT(0, ("received completion for unknown cmd"));
 564                 }
 565
 566                 /*
 567                  * There's a number of races with the following (see above) when
 568                  * the system panics. We compensate for each one of them by
 569                  * using the atomic store to force strong ordering (at least when
 570                  * viewed in the aftermath of a panic).
 571                  */
 572                 if (++qpair->cq_head == qpair->num_entries) {           /* 1 */
 573                         atomic_store_rel_int(&qpair->cq_head, 0);       /* 2 */
 574                         qpair->phase = !qpair->phase;                   /* 3 */
 575                 }
 576
 577                 nvme_mmio_write_4(qpair->ctrlr, doorbell[qpair->id].cq_hdbl,
 578                     qpair->cq_head);
 579         }
 580         return (done != 0);
 581 }
 582
 583 static void
 584 nvme_qpair_msix_handler(void *arg)
 585 {
 586         struct nvme_qpair *qpair = arg;
 587
 588         nvme_qpair_process_completions(qpair);
 589 }
 590
 591 int
 592 nvme_qpair_construct(struct nvme_qpair *qpair, uint32_t id,
 593     uint16_t vector, uint32_t num_entries, uint32_t num_trackers,
 594     struct nvme_controller *ctrlr)
 595 {
 596         struct nvme_tracker     *tr;
 597         size_t                  cmdsz, cplsz, prpsz, allocsz, prpmemsz;
 598         uint64_t                queuemem_phys, prpmem_phys, list_phys;
 599         uint8_t                 *queuemem, *prpmem, *prp_list;
 600         int                     i, err;
 601
 602         qpair->id = id;
 603         qpair->vector = vector;
 604         qpair->num_entries = num_entries;
 605         qpair->num_trackers = num_trackers;
 606         qpair->ctrlr = ctrlr;
 607
 608         if (ctrlr->msix_enabled) {
 609
 610                 /*
 611                  * MSI-X vector resource IDs start at 1, so we add one to
 612                  *  the queue's vector to get the corresponding rid to use.
 613                  */
 614                 qpair->rid = vector + 1;
 615
 616                 qpair->res = bus_alloc_resource_any(ctrlr->dev, SYS_RES_IRQ,
 617                     &qpair->rid, RF_ACTIVE);
 618                 bus_setup_intr(ctrlr->dev, qpair->res,
 619                     INTR_TYPE_MISC | INTR_MPSAFE, NULL,
 620                     nvme_qpair_msix_handler, qpair, &qpair->tag);
 621                 if (id == 0) {
 622                         bus_describe_intr(ctrlr->dev, qpair->res, qpair->tag,
 623                             "admin");
 624                 } else {
 625                         bus_describe_intr(ctrlr->dev, qpair->res, qpair->tag,
 626                             "io%d", id - 1);
 627                 }
 628         }
 629
 630         mtx_init(&qpair->lock, "nvme qpair lock", NULL, MTX_DEF);
 631
 632         /* Note: NVMe PRP format is restricted to 4-byte alignment. */
 633         err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev),
 634             4, PAGE_SIZE, BUS_SPACE_MAXADDR,
 635             BUS_SPACE_MAXADDR, NULL, NULL, NVME_MAX_XFER_SIZE,
 636             (NVME_MAX_XFER_SIZE/PAGE_SIZE)+1, PAGE_SIZE, 0,
 637             NULL, NULL, &qpair->dma_tag_payload);
 638         if (err != 0) {
 639                 nvme_printf(ctrlr, "payload tag create failed %d\n", err);
 640                 goto out;
 641         }
 642
 643         /*
 644          * Each component must be page aligned, and individual PRP lists
 645          * cannot cross a page boundary.
 646          */
 647         cmdsz = qpair->num_entries * sizeof(struct nvme_command);
 648         cmdsz = roundup2(cmdsz, PAGE_SIZE);
 649         cplsz = qpair->num_entries * sizeof(struct nvme_completion);
 650         cplsz = roundup2(cplsz, PAGE_SIZE);
 651         prpsz = sizeof(uint64_t) * NVME_MAX_PRP_LIST_ENTRIES;;
 652         prpmemsz = qpair->num_trackers * prpsz;
 653         allocsz = cmdsz + cplsz + prpmemsz;
 654
 655         err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev),
 656             PAGE_SIZE, 0, BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
 657             allocsz, 1, allocsz, 0, NULL, NULL, &qpair->dma_tag);
 658         if (err != 0) {
 659                 nvme_printf(ctrlr, "tag create failed %d\n", err);
 660                 goto out;
 661         }
 662
 663         if (bus_dmamem_alloc(qpair->dma_tag, (void **)&queuemem,
 664             BUS_DMA_NOWAIT, &qpair->queuemem_map)) {
 665                 nvme_printf(ctrlr, "failed to alloc qpair memory\n");
 666                 goto out;
 667         }
 668
 669         if (bus_dmamap_load(qpair->dma_tag, qpair->queuemem_map,
 670             queuemem, allocsz, nvme_single_map, &queuemem_phys, 0) != 0) {
 671                 nvme_printf(ctrlr, "failed to load qpair memory\n");
 672                 goto out;
 673         }
 674
 675         qpair->num_cmds = 0;
 676         qpair->num_intr_handler_calls = 0;
 677         qpair->cmd = (struct nvme_command *)queuemem;
 678         qpair->cpl = (struct nvme_completion *)(queuemem + cmdsz);
 679         prpmem = (uint8_t *)(queuemem + cmdsz + cplsz);
 680         qpair->cmd_bus_addr = queuemem_phys;
 681         qpair->cpl_bus_addr = queuemem_phys + cmdsz;
 682         prpmem_phys = queuemem_phys + cmdsz + cplsz;
 683
 684         qpair->sq_tdbl_off = nvme_mmio_offsetof(doorbell[id].sq_tdbl);
 685         qpair->cq_hdbl_off = nvme_mmio_offsetof(doorbell[id].cq_hdbl);
 686
 687         TAILQ_INIT(&qpair->free_tr);
 688         TAILQ_INIT(&qpair->outstanding_tr);
 689         STAILQ_INIT(&qpair->queued_req);
 690
 691         list_phys = prpmem_phys;
 692         prp_list = prpmem;
 693         for (i = 0; i < qpair->num_trackers; i++) {
 694
 695                 if (list_phys + prpsz > prpmem_phys + prpmemsz) {
 696                         qpair->num_trackers = i;
 697                         break;
 698                 }
 699
 700                 /*
 701                  * Make sure that the PRP list for this tracker doesn't
 702                  * overflow to another page.
 703                  */
 704                 if (trunc_page(list_phys) !=
 705                     trunc_page(list_phys + prpsz - 1)) {
 706                         list_phys = roundup2(list_phys, PAGE_SIZE);
 707                         prp_list =
 708                             (uint8_t *)roundup2((uintptr_t)prp_list, PAGE_SIZE);
 709                 }
 710
 711                 tr = malloc(sizeof(*tr), M_NVME, M_ZERO | M_WAITOK);
 712                 bus_dmamap_create(qpair->dma_tag_payload, 0,
 713                     &tr->payload_dma_map);
 714                 callout_init(&tr->timer, 1);
 715                 tr->cid = i;
 716                 tr->qpair = qpair;
 717                 tr->prp = (uint64_t *)prp_list;
 718                 tr->prp_bus_addr = list_phys;
 719                 TAILQ_INSERT_HEAD(&qpair->free_tr, tr, tailq);
 720                 list_phys += prpsz;
 721                 prp_list += prpsz;
 722         }
 723
 724         if (qpair->num_trackers == 0) {
 725                 nvme_printf(ctrlr, "failed to allocate enough trackers\n");
 726                 goto out;
 727         }
 728
 729         qpair->act_tr = malloc(sizeof(struct nvme_tracker *) *
 730             qpair->num_entries, M_NVME, M_ZERO | M_WAITOK);
 731         return (0);
 732
 733 out:
 734         nvme_qpair_destroy(qpair);
 735         return (ENOMEM);
 736 }
 737
 738 static void
 739 nvme_qpair_destroy(struct nvme_qpair *qpair)
 740 {
 741         struct nvme_tracker     *tr;
 742
 743         if (qpair->tag)
 744                 bus_teardown_intr(qpair->ctrlr->dev, qpair->res, qpair->tag);
 745
 746         if (mtx_initialized(&qpair->lock))
 747                 mtx_destroy(&qpair->lock);
 748
 749         if (qpair->res)
 750                 bus_release_resource(qpair->ctrlr->dev, SYS_RES_IRQ,
 751                     rman_get_rid(qpair->res), qpair->res);
 752
 753         if (qpair->cmd != NULL) {
 754                 bus_dmamap_unload(qpair->dma_tag, qpair->queuemem_map);
 755                 bus_dmamem_free(qpair->dma_tag, qpair->cmd,
 756                     qpair->queuemem_map);
 757         }
 758
 759         if (qpair->act_tr)
 760                 free(qpair->act_tr, M_NVME);
 761
 762         while (!TAILQ_EMPTY(&qpair->free_tr)) {
 763                 tr = TAILQ_FIRST(&qpair->free_tr);
 764                 TAILQ_REMOVE(&qpair->free_tr, tr, tailq);
 765                 bus_dmamap_destroy(qpair->dma_tag_payload,
 766                     tr->payload_dma_map);
 767                 free(tr, M_NVME);
 768         }
 769
 770         if (qpair->dma_tag)
 771                 bus_dma_tag_destroy(qpair->dma_tag);
 772
 773         if (qpair->dma_tag_payload)
 774                 bus_dma_tag_destroy(qpair->dma_tag_payload);
 775 }
 776
 777 static void
 778 nvme_admin_qpair_abort_aers(struct nvme_qpair *qpair)
 779 {
 780         struct nvme_tracker     *tr;
 781
 782         tr = TAILQ_FIRST(&qpair->outstanding_tr);
 783         while (tr != NULL) {
 784                 if (tr->req->cmd.opc == NVME_OPC_ASYNC_EVENT_REQUEST) {
 785                         nvme_qpair_manual_complete_tracker(qpair, tr,
 786                             NVME_SCT_GENERIC, NVME_SC_ABORTED_SQ_DELETION, 0,
 787                             ERROR_PRINT_NONE);
 788                         tr = TAILQ_FIRST(&qpair->outstanding_tr);
 789                 } else {
 790                         tr = TAILQ_NEXT(tr, tailq);
 791                 }
 792         }
 793 }
 794
 795 void
 796 nvme_admin_qpair_destroy(struct nvme_qpair *qpair)
 797 {
 798
 799         nvme_admin_qpair_abort_aers(qpair);
 800         nvme_qpair_destroy(qpair);
 801 }
 802
 803 void
 804 nvme_io_qpair_destroy(struct nvme_qpair *qpair)
 805 {
 806
 807         nvme_qpair_destroy(qpair);
 808 }
 809
 810 static void
 811 nvme_abort_complete(void *arg, const struct nvme_completion *status)
 812 {
 813         struct nvme_tracker     *tr = arg;
 814
 815         /*
 816          * If cdw0 == 1, the controller was not able to abort the command
 817          *  we requested.  We still need to check the active tracker array,
 818          *  to cover race where I/O timed out at same time controller was
 819          *  completing the I/O.
 820          */
 821         if (status->cdw0 == 1 && tr->qpair->act_tr[tr->cid] != NULL) {
 822                 /*
 823                  * An I/O has timed out, and the controller was unable to
 824                  *  abort it for some reason.  Construct a fake completion
 825                  *  status, and then complete the I/O's tracker manually.
 826                  */
 827                 nvme_printf(tr->qpair->ctrlr,
 828                     "abort command failed, aborting command manually\n");
 829                 nvme_qpair_manual_complete_tracker(tr->qpair, tr,
 830                     NVME_SCT_GENERIC, NVME_SC_ABORTED_BY_REQUEST, 0, ERROR_PRINT_ALL);
 831         }
 832 }
 833
 834 static void
 835 nvme_timeout(void *arg)
 836 {
 837         struct nvme_tracker     *tr = arg;
 838         struct nvme_qpair       *qpair = tr->qpair;
 839         struct nvme_controller  *ctrlr = qpair->ctrlr;
 840         uint32_t                csts;
 841         uint8_t                 cfs;
 842
 843         /*
 844          * Read csts to get value of cfs - controller fatal status.
 845          * If no fatal status, try to call the completion routine, and
 846          * if completes transactions, report a missed interrupt and
 847          * return (this may need to be rate limited). Otherwise, if
 848          * aborts are enabled and the controller is not reporting
 849          * fatal status, abort the command. Otherwise, just reset the
 850          * controller and hope for the best.
 851          */
 852         csts = nvme_mmio_read_4(ctrlr, csts);
 853         cfs = (csts >> NVME_CSTS_REG_CFS_SHIFT) & NVME_CSTS_REG_CFS_MASK;
 854         if (cfs == 0 && nvme_qpair_process_completions(qpair)) {
 855                 nvme_printf(ctrlr, "Missing interrupt\n");
 856                 return;
 857         }
 858         if (ctrlr->enable_aborts && cfs == 0) {
 859                 nvme_printf(ctrlr, "Aborting command due to a timeout.\n");
 860                 nvme_ctrlr_cmd_abort(ctrlr, tr->cid, qpair->id,
 861                     nvme_abort_complete, tr);
 862         } else {
 863                 nvme_printf(ctrlr, "Resetting controller due to a timeout%s.\n",
 864                     cfs ? " and fatal error status" : "");
 865                 nvme_ctrlr_reset(ctrlr);
 866         }
 867 }
 868
 869 void
 870 nvme_qpair_submit_tracker(struct nvme_qpair *qpair, struct nvme_tracker *tr)
 871 {
 872         struct nvme_request     *req;
 873         struct nvme_controller  *ctrlr;
 874
 875         mtx_assert(&qpair->lock, MA_OWNED);
 876
 877         req = tr->req;
 878         req->cmd.cid = tr->cid;
 879         qpair->act_tr[tr->cid] = tr;
 880         ctrlr = qpair->ctrlr;
 881
 882         if (req->timeout)
 883                 callout_reset_curcpu(&tr->timer, ctrlr->timeout_period * hz,
 884                     nvme_timeout, tr);
 885
 886         /* Copy the command from the tracker to the submission queue. */
 887         memcpy(&qpair->cmd[qpair->sq_tail], &req->cmd, sizeof(req->cmd));
 888
 889         if (++qpair->sq_tail == qpair->num_entries)
 890                 qpair->sq_tail = 0;
 891
 892         bus_dmamap_sync(qpair->dma_tag, qpair->queuemem_map,
 893             BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
 894 #ifndef __powerpc__
 895         /*
 896          * powerpc's bus_dmamap_sync() already includes a heavyweight sync, but
 897          * no other archs do.
 898          */
 899         wmb();
 900 #endif
 901
 902         nvme_mmio_write_4(qpair->ctrlr, doorbell[qpair->id].sq_tdbl,
 903             qpair->sq_tail);
 904
 905         qpair->num_cmds++;
 906 }
 907
 908 static void
 909 nvme_payload_map(void *arg, bus_dma_segment_t *seg, int nseg, int error)
 910 {
 911         struct nvme_tracker     *tr = arg;
 912         uint32_t                cur_nseg;
 913
 914         /*
 915          * If the mapping operation failed, return immediately.  The caller
 916          *  is responsible for detecting the error status and failing the
 917          *  tracker manually.
 918          */
 919         if (error != 0) {
 920                 nvme_printf(tr->qpair->ctrlr,
 921                     "nvme_payload_map err %d\n", error);
 922                 return;
 923         }
 924
 925         /*
 926          * Note that we specified PAGE_SIZE for alignment and max
 927          *  segment size when creating the bus dma tags.  So here
 928          *  we can safely just transfer each segment to its
 929          *  associated PRP entry.
 930          */
 931         tr->req->cmd.prp1 = htole64(seg[0].ds_addr);
 932
 933         if (nseg == 2) {
 934                 tr->req->cmd.prp2 = htole64(seg[1].ds_addr);
 935         } else if (nseg > 2) {
 936                 cur_nseg = 1;
 937                 tr->req->cmd.prp2 = htole64((uint64_t)tr->prp_bus_addr);
 938                 while (cur_nseg < nseg) {
 939                         tr->prp[cur_nseg-1] =
 940                             htole64((uint64_t)seg[cur_nseg].ds_addr);
 941                         cur_nseg++;
 942                 }
 943         } else {
 944                 /*
 945                  * prp2 should not be used by the controller
 946                  *  since there is only one segment, but set
 947                  *  to 0 just to be safe.
 948                  */
 949                 tr->req->cmd.prp2 = 0;
 950         }
 951
 952         bus_dmamap_sync(tr->qpair->dma_tag_payload, tr->payload_dma_map,
 953             BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
 954         nvme_qpair_submit_tracker(tr->qpair, tr);
 955 }
 956
 957 static void
 958 _nvme_qpair_submit_request(struct nvme_qpair *qpair, struct nvme_request *req)
 959 {
 960         struct nvme_tracker     *tr;
 961         int                     err = 0;
 962
 963         mtx_assert(&qpair->lock, MA_OWNED);
 964
 965         tr = TAILQ_FIRST(&qpair->free_tr);
 966         req->qpair = qpair;
 967
 968         if (tr == NULL || !qpair->is_enabled) {
 969                 /*
 970                  * No tracker is available, or the qpair is disabled due to
 971                  *  an in-progress controller-level reset or controller
 972                  *  failure.
 973                  */
 974
 975                 if (qpair->ctrlr->is_failed) {
 976                         /*
 977                          * The controller has failed.  Post the request to a
 978                          *  task where it will be aborted, so that we do not
 979                          *  invoke the request's callback in the context
 980                          *  of the submission.
 981                          */
 982                         nvme_ctrlr_post_failed_request(qpair->ctrlr, req);
 983                 } else {
 984                         /*
 985                          * Put the request on the qpair's request queue to be
 986                          *  processed when a tracker frees up via a command
 987                          *  completion or when the controller reset is
 988                          *  completed.
 989                          */
 990                         STAILQ_INSERT_TAIL(&qpair->queued_req, req, stailq);
 991                 }
 992                 return;
 993         }
 994
 995         TAILQ_REMOVE(&qpair->free_tr, tr, tailq);
 996         TAILQ_INSERT_TAIL(&qpair->outstanding_tr, tr, tailq);
 997         tr->req = req;
 998
 999         switch (req->type) {
1000         case NVME_REQUEST_VADDR:
1001                 KASSERT(req->payload_size <= qpair->ctrlr->max_xfer_size,
1002                     ("payload_size (%d) exceeds max_xfer_size (%d)\n",
1003                     req->payload_size, qpair->ctrlr->max_xfer_size));
1004                 err = bus_dmamap_load(tr->qpair->dma_tag_payload,
1005                     tr->payload_dma_map, req->u.payload, req->payload_size,
1006                     nvme_payload_map, tr, 0);
1007                 if (err != 0)
1008                         nvme_printf(qpair->ctrlr,
1009                             "bus_dmamap_load returned 0x%x!\n", err);
1010                 break;
1011         case NVME_REQUEST_NULL:
1012                 nvme_qpair_submit_tracker(tr->qpair, tr);
1013                 break;
1014         case NVME_REQUEST_BIO:
1015                 KASSERT(req->u.bio->bio_bcount <= qpair->ctrlr->max_xfer_size,
1016                     ("bio->bio_bcount (%jd) exceeds max_xfer_size (%d)\n",
1017                     (intmax_t)req->u.bio->bio_bcount,
1018                     qpair->ctrlr->max_xfer_size));
1019                 err = bus_dmamap_load_bio(tr->qpair->dma_tag_payload,
1020                     tr->payload_dma_map, req->u.bio, nvme_payload_map, tr, 0);
1021                 if (err != 0)
1022                         nvme_printf(qpair->ctrlr,
1023                             "bus_dmamap_load_bio returned 0x%x!\n", err);
1024                 break;
1025         case NVME_REQUEST_CCB:
1026                 err = bus_dmamap_load_ccb(tr->qpair->dma_tag_payload,
1027                     tr->payload_dma_map, req->u.payload,
1028                     nvme_payload_map, tr, 0);
1029                 if (err != 0)
1030                         nvme_printf(qpair->ctrlr,
1031                             "bus_dmamap_load_ccb returned 0x%x!\n", err);
1032                 break;
1033         default:
1034                 panic("unknown nvme request type 0x%x\n", req->type);
1035                 break;
1036         }
1037
1038         if (err != 0) {
1039                 /*
1040                  * The dmamap operation failed, so we manually fail the
1041                  *  tracker here with DATA_TRANSFER_ERROR status.
1042                  *
1043                  * nvme_qpair_manual_complete_tracker must not be called
1044                  *  with the qpair lock held.
1045                  */
1046                 mtx_unlock(&qpair->lock);
1047                 nvme_qpair_manual_complete_tracker(qpair, tr, NVME_SCT_GENERIC,
1048                     NVME_SC_DATA_TRANSFER_ERROR, DO_NOT_RETRY, ERROR_PRINT_ALL);
1049                 mtx_lock(&qpair->lock);
1050         }
1051 }
1052
1053 void
1054 nvme_qpair_submit_request(struct nvme_qpair *qpair, struct nvme_request *req)
1055 {
1056
1057         mtx_lock(&qpair->lock);
1058         _nvme_qpair_submit_request(qpair, req);
1059         mtx_unlock(&qpair->lock);
1060 }
1061
1062 static void
1063 nvme_qpair_enable(struct nvme_qpair *qpair)
1064 {
1065
1066         qpair->is_enabled = TRUE;
1067 }
1068
1069 void
1070 nvme_qpair_reset(struct nvme_qpair *qpair)
1071 {
1072
1073         qpair->sq_head = qpair->sq_tail = qpair->cq_head = 0;
1074
1075         /*
1076          * First time through the completion queue, HW will set phase
1077          *  bit on completions to 1.  So set this to 1 here, indicating
1078          *  we're looking for a 1 to know which entries have completed.
1079          *  we'll toggle the bit each time when the completion queue
1080          *  rolls over.
1081          */
1082         qpair->phase = 1;
1083
1084         memset(qpair->cmd, 0,
1085             qpair->num_entries * sizeof(struct nvme_command));
1086         memset(qpair->cpl, 0,
1087             qpair->num_entries * sizeof(struct nvme_completion));
1088 }
1089
1090 void
1091 nvme_admin_qpair_enable(struct nvme_qpair *qpair)
1092 {
1093         struct nvme_tracker             *tr;
1094         struct nvme_tracker             *tr_temp;
1095
1096         /*
1097          * Manually abort each outstanding admin command.  Do not retry
1098          *  admin commands found here, since they will be left over from
1099          *  a controller reset and its likely the context in which the
1100          *  command was issued no longer applies.
1101          */
1102         TAILQ_FOREACH_SAFE(tr, &qpair->outstanding_tr, tailq, tr_temp) {
1103                 nvme_printf(qpair->ctrlr,
1104                     "aborting outstanding admin command\n");
1105                 nvme_qpair_manual_complete_tracker(qpair, tr, NVME_SCT_GENERIC,
1106                     NVME_SC_ABORTED_BY_REQUEST, DO_NOT_RETRY, ERROR_PRINT_ALL);
1107         }
1108
1109         nvme_qpair_enable(qpair);
1110 }
1111
1112 void
1113 nvme_io_qpair_enable(struct nvme_qpair *qpair)
1114 {
1115         STAILQ_HEAD(, nvme_request)     temp;
1116         struct nvme_tracker             *tr;
1117         struct nvme_tracker             *tr_temp;
1118         struct nvme_request             *req;
1119
1120         /*
1121          * Manually abort each outstanding I/O.  This normally results in a
1122          *  retry, unless the retry count on the associated request has
1123          *  reached its limit.
1124          */
1125         TAILQ_FOREACH_SAFE(tr, &qpair->outstanding_tr, tailq, tr_temp) {
1126                 nvme_printf(qpair->ctrlr, "aborting outstanding i/o\n");
1127                 nvme_qpair_manual_complete_tracker(qpair, tr, NVME_SCT_GENERIC,
1128                     NVME_SC_ABORTED_BY_REQUEST, 0, ERROR_PRINT_NO_RETRY);
1129         }
1130
1131         mtx_lock(&qpair->lock);
1132
1133         nvme_qpair_enable(qpair);
1134
1135         STAILQ_INIT(&temp);
1136         STAILQ_SWAP(&qpair->queued_req, &temp, nvme_request);
1137
1138         while (!STAILQ_EMPTY(&temp)) {
1139                 req = STAILQ_FIRST(&temp);
1140                 STAILQ_REMOVE_HEAD(&temp, stailq);
1141                 nvme_printf(qpair->ctrlr, "resubmitting queued i/o\n");
1142                 nvme_qpair_print_command(qpair, &req->cmd);
1143                 _nvme_qpair_submit_request(qpair, req);
1144         }
1145
1146         mtx_unlock(&qpair->lock);
1147 }
1148
1149 static void
1150 nvme_qpair_disable(struct nvme_qpair *qpair)
1151 {
1152         struct nvme_tracker *tr;
1153
1154         qpair->is_enabled = FALSE;
1155         mtx_lock(&qpair->lock);
1156         TAILQ_FOREACH(tr, &qpair->outstanding_tr, tailq)
1157                 callout_stop(&tr->timer);
1158         mtx_unlock(&qpair->lock);
1159 }
1160
1161 void
1162 nvme_admin_qpair_disable(struct nvme_qpair *qpair)
1163 {
1164
1165         nvme_qpair_disable(qpair);
1166         nvme_admin_qpair_abort_aers(qpair);
1167 }
1168
1169 void
1170 nvme_io_qpair_disable(struct nvme_qpair *qpair)
1171 {
1172
1173         nvme_qpair_disable(qpair);
1174 }
1175
1176 void
1177 nvme_qpair_fail(struct nvme_qpair *qpair)
1178 {
1179         struct nvme_tracker             *tr;
1180         struct nvme_request             *req;
1181
1182         if (!mtx_initialized(&qpair->lock))
1183                 return;
1184
1185         mtx_lock(&qpair->lock);
1186
1187         while (!STAILQ_EMPTY(&qpair->queued_req)) {
1188                 req = STAILQ_FIRST(&qpair->queued_req);
1189                 STAILQ_REMOVE_HEAD(&qpair->queued_req, stailq);
1190                 nvme_printf(qpair->ctrlr, "failing queued i/o\n");
1191                 mtx_unlock(&qpair->lock);
1192                 nvme_qpair_manual_complete_request(qpair, req, NVME_SCT_GENERIC,
1193                     NVME_SC_ABORTED_BY_REQUEST);
1194                 mtx_lock(&qpair->lock);
1195         }
1196
1197         /* Manually abort each outstanding I/O. */
1198         while (!TAILQ_EMPTY(&qpair->outstanding_tr)) {
1199                 tr = TAILQ_FIRST(&qpair->outstanding_tr);
1200                 /*
1201                  * Do not remove the tracker.  The abort_tracker path will
1202                  *  do that for us.
1203                  */
1204                 nvme_printf(qpair->ctrlr, "failing outstanding i/o\n");
1205                 mtx_unlock(&qpair->lock);
1206                 nvme_qpair_manual_complete_tracker(qpair, tr, NVME_SCT_GENERIC,
1207                     NVME_SC_ABORTED_BY_REQUEST, DO_NOT_RETRY, ERROR_PRINT_ALL);
1208                 mtx_lock(&qpair->lock);
1209         }
1210
1211         mtx_unlock(&qpair->lock);
1212 }
1213