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