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         uint16_t        opc;
  51         const char *    str;
  52 };
  53
  54 static struct nvme_opcode_string admin_opcode[] = {
  55         { NVME_OPC_DELETE_IO_SQ, "DELETE IO SQ" },
  56         { NVME_OPC_CREATE_IO_SQ, "CREATE IO SQ" },
  57         { NVME_OPC_GET_LOG_PAGE, "GET LOG PAGE" },
  58         { NVME_OPC_DELETE_IO_CQ, "DELETE IO CQ" },
  59         { NVME_OPC_CREATE_IO_CQ, "CREATE IO CQ" },
  60         { NVME_OPC_IDENTIFY, "IDENTIFY" },
  61         { NVME_OPC_ABORT, "ABORT" },
  62         { NVME_OPC_SET_FEATURES, "SET FEATURES" },
  63         { NVME_OPC_GET_FEATURES, "GET FEATURES" },
  64         { NVME_OPC_ASYNC_EVENT_REQUEST, "ASYNC EVENT REQUEST" },
  65         { NVME_OPC_FIRMWARE_ACTIVATE, "FIRMWARE ACTIVATE" },
  66         { NVME_OPC_FIRMWARE_IMAGE_DOWNLOAD, "FIRMWARE IMAGE DOWNLOAD" },
  67         { NVME_OPC_DEVICE_SELF_TEST, "DEVICE SELF-TEST" },
  68         { NVME_OPC_NAMESPACE_ATTACHMENT, "NAMESPACE ATTACHMENT" },
  69         { NVME_OPC_KEEP_ALIVE, "KEEP ALIVE" },
  70         { NVME_OPC_DIRECTIVE_SEND, "DIRECTIVE SEND" },
  71         { NVME_OPC_DIRECTIVE_RECEIVE, "DIRECTIVE RECEIVE" },
  72         { NVME_OPC_VIRTUALIZATION_MANAGEMENT, "VIRTUALIZATION MANAGEMENT" },
  73         { NVME_OPC_NVME_MI_SEND, "NVME-MI SEND" },
  74         { NVME_OPC_NVME_MI_RECEIVE, "NVME-MI RECEIVE" },
  75         { NVME_OPC_DOORBELL_BUFFER_CONFIG, "DOORBELL BUFFER CONFIG" },
  76         { NVME_OPC_FORMAT_NVM, "FORMAT NVM" },
  77         { NVME_OPC_SECURITY_SEND, "SECURITY SEND" },
  78         { NVME_OPC_SECURITY_RECEIVE, "SECURITY RECEIVE" },
  79         { NVME_OPC_SANITIZE, "SANITIZE" },
  80         { NVME_OPC_GET_LBA_STATUS, "GET LBA STATUS" },
  81         { 0xFFFF, "ADMIN COMMAND" }
  82 };
  83
  84 static struct nvme_opcode_string io_opcode[] = {
  85         { NVME_OPC_FLUSH, "FLUSH" },
  86         { NVME_OPC_WRITE, "WRITE" },
  87         { NVME_OPC_READ, "READ" },
  88         { NVME_OPC_WRITE_UNCORRECTABLE, "WRITE UNCORRECTABLE" },
  89         { NVME_OPC_COMPARE, "COMPARE" },
  90         { NVME_OPC_WRITE_ZEROES, "WRITE ZEROES" },
  91         { NVME_OPC_DATASET_MANAGEMENT, "DATASET MANAGEMENT" },
  92         { NVME_OPC_VERIFY, "VERIFY" },
  93         { NVME_OPC_RESERVATION_REGISTER, "RESERVATION REGISTER" },
  94         { NVME_OPC_RESERVATION_REPORT, "RESERVATION REPORT" },
  95         { NVME_OPC_RESERVATION_ACQUIRE, "RESERVATION ACQUIRE" },
  96         { NVME_OPC_RESERVATION_RELEASE, "RESERVATION RELEASE" },
  97         { 0xFFFF, "IO COMMAND" }
  98 };
  99
 100 static const char *
 101 get_admin_opcode_string(uint16_t opc)
 102 {
 103         struct nvme_opcode_string *entry;
 104
 105         entry = admin_opcode;
 106
 107         while (entry->opc != 0xFFFF) {
 108                 if (entry->opc == opc)
 109                         return (entry->str);
 110                 entry++;
 111         }
 112         return (entry->str);
 113 }
 114
 115 static const char *
 116 get_io_opcode_string(uint16_t opc)
 117 {
 118         struct nvme_opcode_string *entry;
 119
 120         entry = io_opcode;
 121
 122         while (entry->opc != 0xFFFF) {
 123                 if (entry->opc == opc)
 124                         return (entry->str);
 125                 entry++;
 126         }
 127         return (entry->str);
 128 }
 129
 130 static void
 131 nvme_admin_qpair_print_command(struct nvme_qpair *qpair,
 132     struct nvme_command *cmd)
 133 {
 134
 135         nvme_printf(qpair->ctrlr, "%s (%02x) sqid:%d cid:%d nsid:%x "
 136             "cdw10:%08x cdw11:%08x\n",
 137             get_admin_opcode_string(cmd->opc), cmd->opc, qpair->id, cmd->cid,
 138             le32toh(cmd->nsid), le32toh(cmd->cdw10), le32toh(cmd->cdw11));
 139 }
 140
 141 static void
 142 nvme_io_qpair_print_command(struct nvme_qpair *qpair,
 143     struct nvme_command *cmd)
 144 {
 145
 146         switch (cmd->opc) {
 147         case NVME_OPC_WRITE:
 148         case NVME_OPC_READ:
 149         case NVME_OPC_WRITE_UNCORRECTABLE:
 150         case NVME_OPC_COMPARE:
 151         case NVME_OPC_WRITE_ZEROES:
 152         case NVME_OPC_VERIFY:
 153                 nvme_printf(qpair->ctrlr, "%s sqid:%d cid:%d nsid:%d "
 154                     "lba:%llu len:%d\n",
 155                     get_io_opcode_string(cmd->opc), qpair->id, cmd->cid, le32toh(cmd->nsid),
 156                     ((unsigned long long)le32toh(cmd->cdw11) << 32) + le32toh(cmd->cdw10),
 157                     (le32toh(cmd->cdw12) & 0xFFFF) + 1);
 158                 break;
 159         case NVME_OPC_FLUSH:
 160         case NVME_OPC_DATASET_MANAGEMENT:
 161         case NVME_OPC_RESERVATION_REGISTER:
 162         case NVME_OPC_RESERVATION_REPORT:
 163         case NVME_OPC_RESERVATION_ACQUIRE:
 164         case NVME_OPC_RESERVATION_RELEASE:
 165                 nvme_printf(qpair->ctrlr, "%s sqid:%d cid:%d nsid:%d\n",
 166                     get_io_opcode_string(cmd->opc), qpair->id, cmd->cid, le32toh(cmd->nsid));
 167                 break;
 168         default:
 169                 nvme_printf(qpair->ctrlr, "%s (%02x) sqid:%d cid:%d nsid:%d\n",
 170                     get_io_opcode_string(cmd->opc), cmd->opc, qpair->id,
 171                     cmd->cid, le32toh(cmd->nsid));
 172                 break;
 173         }
 174 }
 175
 176 static void
 177 nvme_qpair_print_command(struct nvme_qpair *qpair, struct nvme_command *cmd)
 178 {
 179         if (qpair->id == 0)
 180                 nvme_admin_qpair_print_command(qpair, cmd);
 181         else
 182                 nvme_io_qpair_print_command(qpair, cmd);
 183         if (nvme_verbose_cmd_dump) {
 184                 nvme_printf(qpair->ctrlr,
 185                     "nsid:%#x rsvd2:%#x rsvd3:%#x mptr:%#jx prp1:%#jx prp2:%#jx\n",
 186                     cmd->nsid, cmd->rsvd2, cmd->rsvd3, (uintmax_t)cmd->mptr,
 187                     (uintmax_t)cmd->prp1, (uintmax_t)cmd->prp2);
 188                 nvme_printf(qpair->ctrlr,
 189                     "cdw10: %#x cdw11:%#x cdw12:%#x cdw13:%#x cdw14:%#x cdw15:%#x\n",
 190                     cmd->cdw10, cmd->cdw11, cmd->cdw12, cmd->cdw13, cmd->cdw14,
 191                     cmd->cdw15);
 192         }
 193 }
 194
 195 struct nvme_status_string {
 196         uint16_t        sc;
 197         const char *    str;
 198 };
 199
 200 static struct nvme_status_string generic_status[] = {
 201         { NVME_SC_SUCCESS, "SUCCESS" },
 202         { NVME_SC_INVALID_OPCODE, "INVALID OPCODE" },
 203         { NVME_SC_INVALID_FIELD, "INVALID_FIELD" },
 204         { NVME_SC_COMMAND_ID_CONFLICT, "COMMAND ID CONFLICT" },
 205         { NVME_SC_DATA_TRANSFER_ERROR, "DATA TRANSFER ERROR" },
 206         { NVME_SC_ABORTED_POWER_LOSS, "ABORTED - POWER LOSS" },
 207         { NVME_SC_INTERNAL_DEVICE_ERROR, "INTERNAL DEVICE ERROR" },
 208         { NVME_SC_ABORTED_BY_REQUEST, "ABORTED - BY REQUEST" },
 209         { NVME_SC_ABORTED_SQ_DELETION, "ABORTED - SQ DELETION" },
 210         { NVME_SC_ABORTED_FAILED_FUSED, "ABORTED - FAILED FUSED" },
 211         { NVME_SC_ABORTED_MISSING_FUSED, "ABORTED - MISSING FUSED" },
 212         { NVME_SC_INVALID_NAMESPACE_OR_FORMAT, "INVALID NAMESPACE OR FORMAT" },
 213         { NVME_SC_COMMAND_SEQUENCE_ERROR, "COMMAND SEQUENCE ERROR" },
 214         { NVME_SC_INVALID_SGL_SEGMENT_DESCR, "INVALID SGL SEGMENT DESCRIPTOR" },
 215         { NVME_SC_INVALID_NUMBER_OF_SGL_DESCR, "INVALID NUMBER OF SGL DESCRIPTORS" },
 216         { NVME_SC_DATA_SGL_LENGTH_INVALID, "DATA SGL LENGTH INVALID" },
 217         { NVME_SC_METADATA_SGL_LENGTH_INVALID, "METADATA SGL LENGTH INVALID" },
 218         { NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID, "SGL DESCRIPTOR TYPE INVALID" },
 219         { NVME_SC_INVALID_USE_OF_CMB, "INVALID USE OF CONTROLLER MEMORY BUFFER" },
 220         { NVME_SC_PRP_OFFET_INVALID, "PRP OFFET INVALID" },
 221         { NVME_SC_ATOMIC_WRITE_UNIT_EXCEEDED, "ATOMIC WRITE UNIT EXCEEDED" },
 222         { NVME_SC_OPERATION_DENIED, "OPERATION DENIED" },
 223         { NVME_SC_SGL_OFFSET_INVALID, "SGL OFFSET INVALID" },
 224         { NVME_SC_HOST_ID_INCONSISTENT_FORMAT, "HOST IDENTIFIER INCONSISTENT FORMAT" },
 225         { NVME_SC_KEEP_ALIVE_TIMEOUT_EXPIRED, "KEEP ALIVE TIMEOUT EXPIRED" },
 226         { NVME_SC_KEEP_ALIVE_TIMEOUT_INVALID, "KEEP ALIVE TIMEOUT INVALID" },
 227         { NVME_SC_ABORTED_DUE_TO_PREEMPT, "COMMAND ABORTED DUE TO PREEMPT AND ABORT" },
 228         { NVME_SC_SANITIZE_FAILED, "SANITIZE FAILED" },
 229         { NVME_SC_SANITIZE_IN_PROGRESS, "SANITIZE IN PROGRESS" },
 230         { NVME_SC_SGL_DATA_BLOCK_GRAN_INVALID, "SGL_DATA_BLOCK_GRANULARITY_INVALID" },
 231         { NVME_SC_NOT_SUPPORTED_IN_CMB, "COMMAND NOT SUPPORTED FOR QUEUE IN CMB" },
 232         { NVME_SC_NAMESPACE_IS_WRITE_PROTECTED, "NAMESPACE IS WRITE PROTECTED" },
 233         { NVME_SC_COMMAND_INTERRUPTED, "COMMAND INTERRUPTED" },
 234         { NVME_SC_TRANSIENT_TRANSPORT_ERROR, "TRANSIENT TRANSPORT ERROR" },
 235
 236         { NVME_SC_LBA_OUT_OF_RANGE, "LBA OUT OF RANGE" },
 237         { NVME_SC_CAPACITY_EXCEEDED, "CAPACITY EXCEEDED" },
 238         { NVME_SC_NAMESPACE_NOT_READY, "NAMESPACE NOT READY" },
 239         { NVME_SC_RESERVATION_CONFLICT, "RESERVATION CONFLICT" },
 240         { NVME_SC_FORMAT_IN_PROGRESS, "FORMAT IN PROGRESS" },
 241         { 0xFFFF, "GENERIC" }
 242 };
 243
 244 static struct nvme_status_string command_specific_status[] = {
 245         { NVME_SC_COMPLETION_QUEUE_INVALID, "INVALID COMPLETION QUEUE" },
 246         { NVME_SC_INVALID_QUEUE_IDENTIFIER, "INVALID QUEUE IDENTIFIER" },
 247         { NVME_SC_MAXIMUM_QUEUE_SIZE_EXCEEDED, "MAX QUEUE SIZE EXCEEDED" },
 248         { NVME_SC_ABORT_COMMAND_LIMIT_EXCEEDED, "ABORT CMD LIMIT EXCEEDED" },
 249         { NVME_SC_ASYNC_EVENT_REQUEST_LIMIT_EXCEEDED, "ASYNC LIMIT EXCEEDED" },
 250         { NVME_SC_INVALID_FIRMWARE_SLOT, "INVALID FIRMWARE SLOT" },
 251         { NVME_SC_INVALID_FIRMWARE_IMAGE, "INVALID FIRMWARE IMAGE" },
 252         { NVME_SC_INVALID_INTERRUPT_VECTOR, "INVALID INTERRUPT VECTOR" },
 253         { NVME_SC_INVALID_LOG_PAGE, "INVALID LOG PAGE" },
 254         { NVME_SC_INVALID_FORMAT, "INVALID FORMAT" },
 255         { NVME_SC_FIRMWARE_REQUIRES_RESET, "FIRMWARE REQUIRES RESET" },
 256         { NVME_SC_INVALID_QUEUE_DELETION, "INVALID QUEUE DELETION" },
 257         { NVME_SC_FEATURE_NOT_SAVEABLE, "FEATURE IDENTIFIER NOT SAVEABLE" },
 258         { NVME_SC_FEATURE_NOT_CHANGEABLE, "FEATURE NOT CHANGEABLE" },
 259         { NVME_SC_FEATURE_NOT_NS_SPECIFIC, "FEATURE NOT NAMESPACE SPECIFIC" },
 260         { NVME_SC_FW_ACT_REQUIRES_NVMS_RESET, "FIRMWARE ACTIVATION REQUIRES NVM SUBSYSTEM RESET" },
 261         { NVME_SC_FW_ACT_REQUIRES_RESET, "FIRMWARE ACTIVATION REQUIRES RESET" },
 262         { NVME_SC_FW_ACT_REQUIRES_TIME, "FIRMWARE ACTIVATION REQUIRES MAXIMUM TIME VIOLATION" },
 263         { NVME_SC_FW_ACT_PROHIBITED, "FIRMWARE ACTIVATION PROHIBITED" },
 264         { NVME_SC_OVERLAPPING_RANGE, "OVERLAPPING RANGE" },
 265         { NVME_SC_NS_INSUFFICIENT_CAPACITY, "NAMESPACE INSUFFICIENT CAPACITY" },
 266         { NVME_SC_NS_ID_UNAVAILABLE, "NAMESPACE IDENTIFIER UNAVAILABLE" },
 267         { NVME_SC_NS_ALREADY_ATTACHED, "NAMESPACE ALREADY ATTACHED" },
 268         { NVME_SC_NS_IS_PRIVATE, "NAMESPACE IS PRIVATE" },
 269         { NVME_SC_NS_NOT_ATTACHED, "NS NOT ATTACHED" },
 270         { NVME_SC_THIN_PROV_NOT_SUPPORTED, "THIN PROVISIONING NOT SUPPORTED" },
 271         { NVME_SC_CTRLR_LIST_INVALID, "CONTROLLER LIST INVALID" },
 272         { NVME_SC_SELF_TEST_IN_PROGRESS, "DEVICE SELF-TEST IN PROGRESS" },
 273         { NVME_SC_BOOT_PART_WRITE_PROHIB, "BOOT PARTITION WRITE PROHIBITED" },
 274         { NVME_SC_INVALID_CTRLR_ID, "INVALID CONTROLLER IDENTIFIER" },
 275         { NVME_SC_INVALID_SEC_CTRLR_STATE, "INVALID SECONDARY CONTROLLER STATE" },
 276         { NVME_SC_INVALID_NUM_OF_CTRLR_RESRC, "INVALID NUMBER OF CONTROLLER RESOURCES" },
 277         { NVME_SC_INVALID_RESOURCE_ID, "INVALID RESOURCE IDENTIFIER" },
 278         { NVME_SC_SANITIZE_PROHIBITED_WPMRE, "SANITIZE PROHIBITED WRITE PERSISTENT MEMORY REGION ENABLED" },
 279         { NVME_SC_ANA_GROUP_ID_INVALID, "ANA GROUP IDENTIFIED INVALID" },
 280         { NVME_SC_ANA_ATTACH_FAILED, "ANA ATTACH FAILED" },
 281
 282         { NVME_SC_CONFLICTING_ATTRIBUTES, "CONFLICTING ATTRIBUTES" },
 283         { NVME_SC_INVALID_PROTECTION_INFO, "INVALID PROTECTION INFO" },
 284         { NVME_SC_ATTEMPTED_WRITE_TO_RO_PAGE, "WRITE TO RO PAGE" },
 285         { 0xFFFF, "COMMAND SPECIFIC" }
 286 };
 287
 288 static struct nvme_status_string media_error_status[] = {
 289         { NVME_SC_WRITE_FAULTS, "WRITE FAULTS" },
 290         { NVME_SC_UNRECOVERED_READ_ERROR, "UNRECOVERED READ ERROR" },
 291         { NVME_SC_GUARD_CHECK_ERROR, "GUARD CHECK ERROR" },
 292         { NVME_SC_APPLICATION_TAG_CHECK_ERROR, "APPLICATION TAG CHECK ERROR" },
 293         { NVME_SC_REFERENCE_TAG_CHECK_ERROR, "REFERENCE TAG CHECK ERROR" },
 294         { NVME_SC_COMPARE_FAILURE, "COMPARE FAILURE" },
 295         { NVME_SC_ACCESS_DENIED, "ACCESS DENIED" },
 296         { NVME_SC_DEALLOCATED_OR_UNWRITTEN, "DEALLOCATED OR UNWRITTEN LOGICAL BLOCK" },
 297         { 0xFFFF, "MEDIA ERROR" }
 298 };
 299
 300 static struct nvme_status_string path_related_status[] = {
 301         { NVME_SC_INTERNAL_PATH_ERROR, "INTERNAL PATH ERROR" },
 302         { NVME_SC_ASYMMETRIC_ACCESS_PERSISTENT_LOSS, "ASYMMETRIC ACCESS PERSISTENT LOSS" },
 303         { NVME_SC_ASYMMETRIC_ACCESS_INACCESSIBLE, "ASYMMETRIC ACCESS INACCESSIBLE" },
 304         { NVME_SC_ASYMMETRIC_ACCESS_TRANSITION, "ASYMMETRIC ACCESS TRANSITION" },
 305         { NVME_SC_CONTROLLER_PATHING_ERROR, "CONTROLLER PATHING ERROR" },
 306         { NVME_SC_HOST_PATHING_ERROR, "HOST PATHING ERROR" },
 307         { NVME_SC_COMMAND_ABOTHED_BY_HOST, "COMMAND ABOTHED BY HOST" },
 308         { 0xFFFF, "PATH RELATED" },
 309 };
 310
 311 static const char *
 312 get_status_string(uint16_t sct, uint16_t sc)
 313 {
 314         struct nvme_status_string *entry;
 315
 316         switch (sct) {
 317         case NVME_SCT_GENERIC:
 318                 entry = generic_status;
 319                 break;
 320         case NVME_SCT_COMMAND_SPECIFIC:
 321                 entry = command_specific_status;
 322                 break;
 323         case NVME_SCT_MEDIA_ERROR:
 324                 entry = media_error_status;
 325                 break;
 326         case NVME_SCT_PATH_RELATED:
 327                 entry = path_related_status;
 328                 break;
 329         case NVME_SCT_VENDOR_SPECIFIC:
 330                 return ("VENDOR SPECIFIC");
 331         default:
 332                 return ("RESERVED");
 333         }
 334
 335         while (entry->sc != 0xFFFF) {
 336                 if (entry->sc == sc)
 337                         return (entry->str);
 338                 entry++;
 339         }
 340         return (entry->str);
 341 }
 342
 343 static void
 344 nvme_qpair_print_completion(struct nvme_qpair *qpair,
 345     struct nvme_completion *cpl)
 346 {
 347         uint16_t sct, sc;
 348
 349         sct = NVME_STATUS_GET_SCT(cpl->status);
 350         sc = NVME_STATUS_GET_SC(cpl->status);
 351
 352         nvme_printf(qpair->ctrlr, "%s (%02x/%02x) sqid:%d cid:%d cdw0:%x\n",
 353             get_status_string(sct, sc), sct, sc, cpl->sqid, cpl->cid,
 354             cpl->cdw0);
 355 }
 356
 357 static bool
 358 nvme_completion_is_retry(const struct nvme_completion *cpl)
 359 {
 360         uint8_t sct, sc, dnr;
 361
 362         sct = NVME_STATUS_GET_SCT(cpl->status);
 363         sc = NVME_STATUS_GET_SC(cpl->status);
 364         dnr = NVME_STATUS_GET_DNR(cpl->status); /* Do Not Retry Bit */
 365
 366         /*
 367          * TODO: spec is not clear how commands that are aborted due
 368          *  to TLER will be marked.  So for now, it seems
 369          *  NAMESPACE_NOT_READY is the only case where we should
 370          *  look at the DNR bit. Requests failed with ABORTED_BY_REQUEST
 371          *  set the DNR bit correctly since the driver controls that.
 372          */
 373         switch (sct) {
 374         case NVME_SCT_GENERIC:
 375                 switch (sc) {
 376                 case NVME_SC_ABORTED_BY_REQUEST:
 377                 case NVME_SC_NAMESPACE_NOT_READY:
 378                         if (dnr)
 379                                 return (0);
 380                         else
 381                                 return (1);
 382                 case NVME_SC_INVALID_OPCODE:
 383                 case NVME_SC_INVALID_FIELD:
 384                 case NVME_SC_COMMAND_ID_CONFLICT:
 385                 case NVME_SC_DATA_TRANSFER_ERROR:
 386                 case NVME_SC_ABORTED_POWER_LOSS:
 387                 case NVME_SC_INTERNAL_DEVICE_ERROR:
 388                 case NVME_SC_ABORTED_SQ_DELETION:
 389                 case NVME_SC_ABORTED_FAILED_FUSED:
 390                 case NVME_SC_ABORTED_MISSING_FUSED:
 391                 case NVME_SC_INVALID_NAMESPACE_OR_FORMAT:
 392                 case NVME_SC_COMMAND_SEQUENCE_ERROR:
 393                 case NVME_SC_LBA_OUT_OF_RANGE:
 394                 case NVME_SC_CAPACITY_EXCEEDED:
 395                 default:
 396                         return (0);
 397                 }
 398         case NVME_SCT_COMMAND_SPECIFIC:
 399         case NVME_SCT_MEDIA_ERROR:
 400                 return (0);
 401         case NVME_SCT_PATH_RELATED:
 402                 switch (sc) {
 403                 case NVME_SC_INTERNAL_PATH_ERROR:
 404                         if (dnr)
 405                                 return (0);
 406                         else
 407                                 return (1);
 408                 default:
 409                         return (0);
 410                 }
 411         case NVME_SCT_VENDOR_SPECIFIC:
 412         default:
 413                 return (0);
 414         }
 415 }
 416
 417 static void
 418 nvme_qpair_complete_tracker(struct nvme_tracker *tr,
 419     struct nvme_completion *cpl, error_print_t print_on_error)
 420 {
 421         struct nvme_qpair * qpair = tr->qpair;
 422         struct nvme_request     *req;
 423         bool                    retry, error, retriable;
 424
 425         req = tr->req;
 426         error = nvme_completion_is_error(cpl);
 427         retriable = nvme_completion_is_retry(cpl);
 428         retry = error && retriable && req->retries < nvme_retry_count;
 429         if (retry)
 430                 qpair->num_retries++;
 431         if (error && req->retries >= nvme_retry_count && retriable)
 432                 qpair->num_failures++;
 433
 434         if (error && (print_on_error == ERROR_PRINT_ALL ||
 435                 (!retry && print_on_error == ERROR_PRINT_NO_RETRY))) {
 436                 nvme_qpair_print_command(qpair, &req->cmd);
 437                 nvme_qpair_print_completion(qpair, cpl);
 438         }
 439
 440         qpair->act_tr[cpl->cid] = NULL;
 441
 442         KASSERT(cpl->cid == req->cmd.cid, ("cpl cid does not match cmd cid\n"));
 443
 444         if (!retry) {
 445                 if (req->type != NVME_REQUEST_NULL) {
 446                         bus_dmamap_sync(qpair->dma_tag_payload,
 447                             tr->payload_dma_map,
 448                             BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
 449                 }
 450                 if (req->cb_fn)
 451                         req->cb_fn(req->cb_arg, cpl);
 452         }
 453
 454         mtx_lock(&qpair->lock);
 455
 456         if (retry) {
 457                 req->retries++;
 458                 nvme_qpair_submit_tracker(qpair, tr);
 459         } else {
 460                 if (req->type != NVME_REQUEST_NULL) {
 461                         bus_dmamap_unload(qpair->dma_tag_payload,
 462                             tr->payload_dma_map);
 463                 }
 464
 465                 nvme_free_request(req);
 466                 tr->req = NULL;
 467
 468                 TAILQ_REMOVE(&qpair->outstanding_tr, tr, tailq);
 469                 TAILQ_INSERT_HEAD(&qpair->free_tr, tr, tailq);
 470
 471                 /*
 472                  * If the controller is in the middle of resetting, don't
 473                  *  try to submit queued requests here - let the reset logic
 474                  *  handle that instead.
 475                  */
 476                 if (!STAILQ_EMPTY(&qpair->queued_req) &&
 477                     !qpair->ctrlr->is_resetting) {
 478                         req = STAILQ_FIRST(&qpair->queued_req);
 479                         STAILQ_REMOVE_HEAD(&qpair->queued_req, stailq);
 480                         _nvme_qpair_submit_request(qpair, req);
 481                 }
 482         }
 483
 484         mtx_unlock(&qpair->lock);
 485 }
 486
 487 static void
 488 nvme_qpair_manual_complete_tracker(
 489     struct nvme_tracker *tr, uint32_t sct, uint32_t sc, uint32_t dnr,
 490     error_print_t print_on_error)
 491 {
 492         struct nvme_completion  cpl;
 493
 494         memset(&cpl, 0, sizeof(cpl));
 495
 496         struct nvme_qpair * qpair = tr->qpair;
 497
 498         cpl.sqid = qpair->id;
 499         cpl.cid = tr->cid;
 500         cpl.status |= (sct & NVME_STATUS_SCT_MASK) << NVME_STATUS_SCT_SHIFT;
 501         cpl.status |= (sc & NVME_STATUS_SC_MASK) << NVME_STATUS_SC_SHIFT;
 502         cpl.status |= (dnr & NVME_STATUS_DNR_MASK) << NVME_STATUS_DNR_SHIFT;
 503         nvme_qpair_complete_tracker(tr, &cpl, print_on_error);
 504 }
 505
 506 void
 507 nvme_qpair_manual_complete_request(struct nvme_qpair *qpair,
 508     struct nvme_request *req, uint32_t sct, uint32_t sc)
 509 {
 510         struct nvme_completion  cpl;
 511         bool                    error;
 512
 513         memset(&cpl, 0, sizeof(cpl));
 514         cpl.sqid = qpair->id;
 515         cpl.status |= (sct & NVME_STATUS_SCT_MASK) << NVME_STATUS_SCT_SHIFT;
 516         cpl.status |= (sc & NVME_STATUS_SC_MASK) << NVME_STATUS_SC_SHIFT;
 517
 518         error = nvme_completion_is_error(&cpl);
 519
 520         if (error) {
 521                 nvme_qpair_print_command(qpair, &req->cmd);
 522                 nvme_qpair_print_completion(qpair, &cpl);
 523         }
 524
 525         if (req->cb_fn)
 526                 req->cb_fn(req->cb_arg, &cpl);
 527
 528         nvme_free_request(req);
 529 }
 530
 531 bool
 532 nvme_qpair_process_completions(struct nvme_qpair *qpair)
 533 {
 534         struct nvme_tracker     *tr;
 535         struct nvme_completion  cpl;
 536         int done = 0;
 537         bool in_panic = dumping || SCHEDULER_STOPPED();
 538
 539         /*
 540          * qpair is not enabled, likely because a controller reset is is in
 541          * progress.  Ignore the interrupt - any I/O that was associated with
 542          * this interrupt will get retried when the reset is complete. Any
 543          * pending completions for when we're in startup will be completed
 544          * as soon as initialization is complete and we start sending commands
 545          * to the device.
 546          */
 547         if (qpair->recovery_state != RECOVERY_NONE) {
 548                 qpair->num_ignored++;
 549                 return (false);
 550         }
 551
 552         /*
 553          * Sanity check initialization. After we reset the hardware, the phase
 554          * is defined to be 1. So if we get here with zero prior calls and the
 555          * phase is 0, it means that we've lost a race between the
 556          * initialization and the ISR running. With the phase wrong, we'll
 557          * process a bunch of completions that aren't really completions leading
 558          * to a KASSERT below.
 559          */
 560         KASSERT(!(qpair->num_intr_handler_calls == 0 && qpair->phase == 0),
 561             ("%s: Phase wrong for first interrupt call.",
 562                 device_get_nameunit(qpair->ctrlr->dev)));
 563
 564         qpair->num_intr_handler_calls++;
 565
 566         bus_dmamap_sync(qpair->dma_tag, qpair->queuemem_map,
 567             BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
 568         /*
 569          * A panic can stop the CPU this routine is running on at any point.  If
 570          * we're called during a panic, complete the sq_head wrap protocol for
 571          * the case where we are interrupted just after the increment at 1
 572          * below, but before we can reset cq_head to zero at 2. Also cope with
 573          * the case where we do the zero at 2, but may or may not have done the
 574          * phase adjustment at step 3. The panic machinery flushes all pending
 575          * memory writes, so we can make these strong ordering assumptions
 576          * that would otherwise be unwise if we were racing in real time.
 577          */
 578         if (__predict_false(in_panic)) {
 579                 if (qpair->cq_head == qpair->num_entries) {
 580                         /*
 581                          * Here we know that we need to zero cq_head and then negate
 582                          * the phase, which hasn't been assigned if cq_head isn't
 583                          * zero due to the atomic_store_rel.
 584                          */
 585                         qpair->cq_head = 0;
 586                         qpair->phase = !qpair->phase;
 587                 } else if (qpair->cq_head == 0) {
 588                         /*
 589                          * In this case, we know that the assignment at 2
 590                          * happened below, but we don't know if it 3 happened or
 591                          * not. To do this, we look at the last completion
 592                          * entry and set the phase to the opposite phase
 593                          * that it has. This gets us back in sync
 594                          */
 595                         cpl = qpair->cpl[qpair->num_entries - 1];
 596                         nvme_completion_swapbytes(&cpl);
 597                         qpair->phase = !NVME_STATUS_GET_P(cpl.status);
 598                 }
 599         }
 600
 601         while (1) {
 602                 uint16_t status;
 603
 604                 /*
 605                  * We need to do this dance to avoid a race between the host and
 606                  * the device where the device overtakes the host while the host
 607                  * is reading this record, leaving the status field 'new' and
 608                  * the sqhd and cid fields potentially stale. If the phase
 609                  * doesn't match, that means status hasn't yet been updated and
 610                  * we'll get any pending changes next time. It also means that
 611                  * the phase must be the same the second time. We have to sync
 612                  * before reading to ensure any bouncing completes.
 613                  */
 614                 status = le16toh(qpair->cpl[qpair->cq_head].status);
 615                 if (NVME_STATUS_GET_P(status) != qpair->phase)
 616                         break;
 617
 618                 bus_dmamap_sync(qpair->dma_tag, qpair->queuemem_map,
 619                     BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
 620                 cpl = qpair->cpl[qpair->cq_head];
 621                 nvme_completion_swapbytes(&cpl);
 622
 623                 KASSERT(
 624                     NVME_STATUS_GET_P(status) == NVME_STATUS_GET_P(cpl.status),
 625                     ("Phase unexpectedly inconsistent"));
 626
 627                 if (cpl.cid < qpair->num_trackers)
 628                         tr = qpair->act_tr[cpl.cid];
 629                 else
 630                         tr = NULL;
 631
 632                 if (tr != NULL) {
 633                         nvme_qpair_complete_tracker(tr, &cpl, ERROR_PRINT_ALL);
 634                         qpair->sq_head = cpl.sqhd;
 635                         done++;
 636                 } else if (!in_panic) {
 637                         /*
 638                          * A missing tracker is normally an error.  However, a
 639                          * panic can stop the CPU this routine is running on
 640                          * after completing an I/O but before updating
 641                          * qpair->cq_head at 1 below.  Later, we re-enter this
 642                          * routine to poll I/O associated with the kernel
 643                          * dump. We find that the tr has been set to null before
 644                          * calling the completion routine.  If it hasn't
 645                          * completed (or it triggers a panic), then '1' below
 646                          * won't have updated cq_head. Rather than panic again,
 647                          * ignore this condition because it's not unexpected.
 648                          */
 649                         nvme_printf(qpair->ctrlr,
 650                             "cpl (cid = %u) does not map to outstanding cmd\n",
 651                                 cpl.cid);
 652                         /* nvme_dump_completion expects device endianess */
 653                         nvme_dump_completion(&qpair->cpl[qpair->cq_head]);
 654                         KASSERT(0, ("received completion for unknown cmd"));
 655                 }
 656
 657                 /*
 658                  * There's a number of races with the following (see above) when
 659                  * the system panics. We compensate for each one of them by
 660                  * using the atomic store to force strong ordering (at least when
 661                  * viewed in the aftermath of a panic).
 662                  */
 663                 if (++qpair->cq_head == qpair->num_entries) {           /* 1 */
 664                         atomic_store_rel_int(&qpair->cq_head, 0);       /* 2 */
 665                         qpair->phase = !qpair->phase;                   /* 3 */
 666                 }
 667
 668                 bus_space_write_4(qpair->ctrlr->bus_tag, qpair->ctrlr->bus_handle,
 669                     qpair->cq_hdbl_off, qpair->cq_head);
 670         }
 671         return (done != 0);
 672 }
 673
 674 static void
 675 nvme_qpair_msi_handler(void *arg)
 676 {
 677         struct nvme_qpair *qpair = arg;
 678
 679         nvme_qpair_process_completions(qpair);
 680 }
 681
 682 int
 683 nvme_qpair_construct(struct nvme_qpair *qpair,
 684     uint32_t num_entries, uint32_t num_trackers,
 685     struct nvme_controller *ctrlr)
 686 {
 687         struct nvme_tracker     *tr;
 688         size_t                  cmdsz, cplsz, prpsz, allocsz, prpmemsz;
 689         uint64_t                queuemem_phys, prpmem_phys, list_phys;
 690         uint8_t                 *queuemem, *prpmem, *prp_list;
 691         int                     i, err;
 692
 693         qpair->vector = ctrlr->msi_count > 1 ? qpair->id : 0;
 694         qpair->num_entries = num_entries;
 695         qpair->num_trackers = num_trackers;
 696         qpair->ctrlr = ctrlr;
 697
 698         mtx_init(&qpair->lock, "nvme qpair lock", NULL, MTX_DEF);
 699
 700         /* Note: NVMe PRP format is restricted to 4-byte alignment. */
 701         err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev),
 702             4, PAGE_SIZE, BUS_SPACE_MAXADDR,
 703             BUS_SPACE_MAXADDR, NULL, NULL, ctrlr->max_xfer_size,
 704             btoc(ctrlr->max_xfer_size) + 1, PAGE_SIZE, 0,
 705             NULL, NULL, &qpair->dma_tag_payload);
 706         if (err != 0) {
 707                 nvme_printf(ctrlr, "payload tag create failed %d\n", err);
 708                 goto out;
 709         }
 710
 711         /*
 712          * Each component must be page aligned, and individual PRP lists
 713          * cannot cross a page boundary.
 714          */
 715         cmdsz = qpair->num_entries * sizeof(struct nvme_command);
 716         cmdsz = roundup2(cmdsz, PAGE_SIZE);
 717         cplsz = qpair->num_entries * sizeof(struct nvme_completion);
 718         cplsz = roundup2(cplsz, PAGE_SIZE);
 719         /*
 720          * For commands requiring more than 2 PRP entries, one PRP will be
 721          * embedded in the command (prp1), and the rest of the PRP entries
 722          * will be in a list pointed to by the command (prp2).
 723          */
 724         prpsz = sizeof(uint64_t) * btoc(ctrlr->max_xfer_size);
 725         prpmemsz = qpair->num_trackers * prpsz;
 726         allocsz = cmdsz + cplsz + prpmemsz;
 727
 728         err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev),
 729             PAGE_SIZE, 0, BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
 730             allocsz, 1, allocsz, 0, NULL, NULL, &qpair->dma_tag);
 731         if (err != 0) {
 732                 nvme_printf(ctrlr, "tag create failed %d\n", err);
 733                 goto out;
 734         }
 735         bus_dma_tag_set_domain(qpair->dma_tag, qpair->domain);
 736
 737         if (bus_dmamem_alloc(qpair->dma_tag, (void **)&queuemem,
 738              BUS_DMA_COHERENT | BUS_DMA_NOWAIT, &qpair->queuemem_map)) {
 739                 nvme_printf(ctrlr, "failed to alloc qpair memory\n");
 740                 goto out;
 741         }
 742
 743         if (bus_dmamap_load(qpair->dma_tag, qpair->queuemem_map,
 744             queuemem, allocsz, nvme_single_map, &queuemem_phys, 0) != 0) {
 745                 nvme_printf(ctrlr, "failed to load qpair memory\n");
 746                 bus_dmamem_free(qpair->dma_tag, qpair->cmd,
 747                     qpair->queuemem_map);
 748                 goto out;
 749         }
 750
 751         qpair->num_cmds = 0;
 752         qpair->num_intr_handler_calls = 0;
 753         qpair->num_retries = 0;
 754         qpair->num_failures = 0;
 755         qpair->num_ignored = 0;
 756         qpair->cmd = (struct nvme_command *)queuemem;
 757         qpair->cpl = (struct nvme_completion *)(queuemem + cmdsz);
 758         prpmem = (uint8_t *)(queuemem + cmdsz + cplsz);
 759         qpair->cmd_bus_addr = queuemem_phys;
 760         qpair->cpl_bus_addr = queuemem_phys + cmdsz;
 761         prpmem_phys = queuemem_phys + cmdsz + cplsz;
 762
 763         callout_init(&qpair->timer, 1);
 764         qpair->timer_armed = false;
 765         qpair->recovery_state = RECOVERY_WAITING;
 766
 767         /*
 768          * Calcuate the stride of the doorbell register. Many emulators set this
 769          * value to correspond to a cache line. However, some hardware has set
 770          * it to various small values.
 771          */
 772         qpair->sq_tdbl_off = nvme_mmio_offsetof(doorbell[0]) +
 773             (qpair->id << (ctrlr->dstrd + 1));
 774         qpair->cq_hdbl_off = nvme_mmio_offsetof(doorbell[0]) +
 775             (qpair->id << (ctrlr->dstrd + 1)) + (1 << ctrlr->dstrd);
 776
 777         TAILQ_INIT(&qpair->free_tr);
 778         TAILQ_INIT(&qpair->outstanding_tr);
 779         STAILQ_INIT(&qpair->queued_req);
 780
 781         list_phys = prpmem_phys;
 782         prp_list = prpmem;
 783         for (i = 0; i < qpair->num_trackers; i++) {
 784                 if (list_phys + prpsz > prpmem_phys + prpmemsz) {
 785                         qpair->num_trackers = i;
 786                         break;
 787                 }
 788
 789                 /*
 790                  * Make sure that the PRP list for this tracker doesn't
 791                  * overflow to another page.
 792                  */
 793                 if (trunc_page(list_phys) !=
 794                     trunc_page(list_phys + prpsz - 1)) {
 795                         list_phys = roundup2(list_phys, PAGE_SIZE);
 796                         prp_list =
 797                             (uint8_t *)roundup2((uintptr_t)prp_list, PAGE_SIZE);
 798                 }
 799
 800                 tr = malloc_domainset(sizeof(*tr), M_NVME,
 801                     DOMAINSET_PREF(qpair->domain), M_ZERO | M_WAITOK);
 802                 bus_dmamap_create(qpair->dma_tag_payload, 0,
 803                     &tr->payload_dma_map);
 804                 tr->cid = i;
 805                 tr->qpair = qpair;
 806                 tr->prp = (uint64_t *)prp_list;
 807                 tr->prp_bus_addr = list_phys;
 808                 TAILQ_INSERT_HEAD(&qpair->free_tr, tr, tailq);
 809                 list_phys += prpsz;
 810                 prp_list += prpsz;
 811         }
 812
 813         if (qpair->num_trackers == 0) {
 814                 nvme_printf(ctrlr, "failed to allocate enough trackers\n");
 815                 goto out;
 816         }
 817
 818         qpair->act_tr = malloc_domainset(sizeof(struct nvme_tracker *) *
 819             qpair->num_entries, M_NVME, DOMAINSET_PREF(qpair->domain),
 820             M_ZERO | M_WAITOK);
 821
 822         if (ctrlr->msi_count > 1) {
 823                 /*
 824                  * MSI-X vector resource IDs start at 1, so we add one to
 825                  *  the queue's vector to get the corresponding rid to use.
 826                  */
 827                 qpair->rid = qpair->vector + 1;
 828
 829                 qpair->res = bus_alloc_resource_any(ctrlr->dev, SYS_RES_IRQ,
 830                     &qpair->rid, RF_ACTIVE);
 831                 if (qpair->res == NULL) {
 832                         nvme_printf(ctrlr, "unable to allocate MSI\n");
 833                         goto out;
 834                 }
 835                 if (bus_setup_intr(ctrlr->dev, qpair->res,
 836                     INTR_TYPE_MISC | INTR_MPSAFE, NULL,
 837                     nvme_qpair_msi_handler, qpair, &qpair->tag) != 0) {
 838                         nvme_printf(ctrlr, "unable to setup MSI\n");
 839                         goto out;
 840                 }
 841                 if (qpair->id == 0) {
 842                         bus_describe_intr(ctrlr->dev, qpair->res, qpair->tag,
 843                             "admin");
 844                 } else {
 845                         bus_describe_intr(ctrlr->dev, qpair->res, qpair->tag,
 846                             "io%d", qpair->id - 1);
 847                 }
 848         }
 849
 850         return (0);
 851
 852 out:
 853         nvme_qpair_destroy(qpair);
 854         return (ENOMEM);
 855 }
 856
 857 static void
 858 nvme_qpair_destroy(struct nvme_qpair *qpair)
 859 {
 860         struct nvme_tracker     *tr;
 861
 862         callout_drain(&qpair->timer);
 863
 864         if (qpair->tag) {
 865                 bus_teardown_intr(qpair->ctrlr->dev, qpair->res, qpair->tag);
 866                 qpair->tag = NULL;
 867         }
 868
 869         if (qpair->act_tr) {
 870                 free(qpair->act_tr, M_NVME);
 871                 qpair->act_tr = NULL;
 872         }
 873
 874         while (!TAILQ_EMPTY(&qpair->free_tr)) {
 875                 tr = TAILQ_FIRST(&qpair->free_tr);
 876                 TAILQ_REMOVE(&qpair->free_tr, tr, tailq);
 877                 bus_dmamap_destroy(qpair->dma_tag_payload,
 878                     tr->payload_dma_map);
 879                 free(tr, M_NVME);
 880         }
 881
 882         if (qpair->cmd != NULL) {
 883                 bus_dmamap_unload(qpair->dma_tag, qpair->queuemem_map);
 884                 bus_dmamem_free(qpair->dma_tag, qpair->cmd,
 885                     qpair->queuemem_map);
 886                 qpair->cmd = NULL;
 887         }
 888
 889         if (qpair->dma_tag) {
 890                 bus_dma_tag_destroy(qpair->dma_tag);
 891                 qpair->dma_tag = NULL;
 892         }
 893
 894         if (qpair->dma_tag_payload) {
 895                 bus_dma_tag_destroy(qpair->dma_tag_payload);
 896                 qpair->dma_tag_payload = NULL;
 897         }
 898
 899         if (mtx_initialized(&qpair->lock))
 900                 mtx_destroy(&qpair->lock);
 901
 902         if (qpair->res) {
 903                 bus_release_resource(qpair->ctrlr->dev, SYS_RES_IRQ,
 904                     rman_get_rid(qpair->res), qpair->res);
 905                 qpair->res = NULL;
 906         }
 907 }
 908
 909 static void
 910 nvme_admin_qpair_abort_aers(struct nvme_qpair *qpair)
 911 {
 912         struct nvme_tracker     *tr;
 913
 914         tr = TAILQ_FIRST(&qpair->outstanding_tr);
 915         while (tr != NULL) {
 916                 if (tr->req->cmd.opc == NVME_OPC_ASYNC_EVENT_REQUEST) {
 917                         nvme_qpair_manual_complete_tracker(tr,
 918                             NVME_SCT_GENERIC, NVME_SC_ABORTED_SQ_DELETION, 0,
 919                             ERROR_PRINT_NONE);
 920                         tr = TAILQ_FIRST(&qpair->outstanding_tr);
 921                 } else {
 922                         tr = TAILQ_NEXT(tr, tailq);
 923                 }
 924         }
 925 }
 926
 927 void
 928 nvme_admin_qpair_destroy(struct nvme_qpair *qpair)
 929 {
 930
 931         nvme_admin_qpair_abort_aers(qpair);
 932         nvme_qpair_destroy(qpair);
 933 }
 934
 935 void
 936 nvme_io_qpair_destroy(struct nvme_qpair *qpair)
 937 {
 938
 939         nvme_qpair_destroy(qpair);
 940 }
 941
 942 static void
 943 nvme_qpair_timeout(void *arg)
 944 {
 945         struct nvme_qpair       *qpair = arg;
 946         struct nvme_controller  *ctrlr = qpair->ctrlr;
 947         struct nvme_tracker     *tr;
 948         struct nvme_tracker     *tr_temp;
 949         sbintime_t              now;
 950         bool                    idle;
 951         uint32_t                csts;
 952         uint8_t                 cfs;
 953
 954         mtx_lock(&qpair->lock);
 955         idle = TAILQ_EMPTY(&qpair->outstanding_tr);
 956 again:
 957         switch (qpair->recovery_state) {
 958         case RECOVERY_NONE:
 959                 if (idle)
 960                         break;
 961                 now = getsbinuptime();
 962                 TAILQ_FOREACH_SAFE(tr, &qpair->outstanding_tr, tailq, tr_temp) {
 963                         if (now > tr->deadline && tr->deadline != 0) {
 964                                 /*
 965                                  * We're now passed our earliest deadline. We
 966                                  * need to do expensive things to cope, but next
 967                                  * time. Flag that and close the door to any
 968                                  * further processing.
 969                                  */
 970                                 qpair->recovery_state = RECOVERY_START;
 971                                 nvme_printf(ctrlr, "RECOVERY_START %jd vs %jd\n",
 972                                     (uintmax_t)now, (uintmax_t)tr->deadline);
 973                                 break;
 974                         }
 975                 }
 976                 break;
 977         case RECOVERY_START:
 978                 /*
 979                  * Read csts to get value of cfs - controller fatal status.
 980                  * If no fatal status, try to call the completion routine, and
 981                  * if completes transactions, report a missed interrupt and
 982                  * return (this may need to be rate limited). Otherwise, if
 983                  * aborts are enabled and the controller is not reporting
 984                  * fatal status, abort the command. Otherwise, just reset the
 985                  * controller and hope for the best.
 986                  */
 987                 csts = nvme_mmio_read_4(ctrlr, csts);
 988                 cfs = (csts >> NVME_CSTS_REG_CFS_SHIFT) & NVME_CSTS_REG_CFS_MASK;
 989                 if (cfs) {
 990                         nvme_printf(ctrlr, "Controller in fatal status, resetting\n");
 991                         qpair->recovery_state = RECOVERY_RESET;
 992                         goto again;
 993                 }
 994                 mtx_unlock(&qpair->lock);
 995                 if (nvme_qpair_process_completions(qpair)) {
 996                         nvme_printf(ctrlr, "Completions present in output without an interrupt\n");
 997                         qpair->recovery_state = RECOVERY_NONE;
 998                 } else {
 999                         nvme_printf(ctrlr, "timeout with nothing complete, resetting\n");
1000                         qpair->recovery_state = RECOVERY_RESET;
1001                         mtx_lock(&qpair->lock);
1002                         goto again;
1003                 }
1004                 mtx_lock(&qpair->lock);
1005                 break;
1006         case RECOVERY_RESET:
1007                 /*
1008                  * If we get here due to a possible surprise hot-unplug event,
1009                  * then we let nvme_ctrlr_reset confirm and fail the
1010                  * controller.
1011                  */
1012                 nvme_printf(ctrlr, "Resetting controller due to a timeout%s.\n",
1013                     cfs ? " and fatal error status" : "");
1014                 nvme_printf(ctrlr, "RECOVERY_WAITING\n");
1015                 qpair->recovery_state = RECOVERY_WAITING;
1016                 nvme_ctrlr_reset(ctrlr);
1017                 break;
1018         case RECOVERY_WAITING:
1019                 nvme_printf(ctrlr, "waiting\n");
1020                 break;
1021         }
1022
1023         /*
1024          * Rearm the timeout.
1025          */
1026         if (!idle) {
1027                 callout_schedule(&qpair->timer, hz / 2);
1028         } else {
1029                 qpair->timer_armed = false;
1030         }
1031         mtx_unlock(&qpair->lock);
1032 }
1033
1034 /*
1035  * Submit the tracker to the hardware. Must already be in the
1036  * outstanding queue when called.
1037  */
1038 void
1039 nvme_qpair_submit_tracker(struct nvme_qpair *qpair, struct nvme_tracker *tr)
1040 {
1041         struct nvme_request     *req;
1042         struct nvme_controller  *ctrlr;
1043         int timeout;
1044
1045         mtx_assert(&qpair->lock, MA_OWNED);
1046
1047         req = tr->req;
1048         req->cmd.cid = tr->cid;
1049         qpair->act_tr[tr->cid] = tr;
1050         ctrlr = qpair->ctrlr;
1051
1052         if (req->timeout) {
1053                 if (req->cb_fn == nvme_completion_poll_cb)
1054                         timeout = 1;
1055                 else
1056                         timeout = ctrlr->timeout_period;
1057                 tr->deadline = getsbinuptime() + timeout * SBT_1S;
1058                 if (!qpair->timer_armed) {
1059                         qpair->timer_armed = true;
1060                         callout_reset_on(&qpair->timer, hz / 2,
1061                             nvme_qpair_timeout, qpair, qpair->cpu);
1062                 }
1063         } else
1064                 tr->deadline = SBT_MAX;
1065
1066         /* Copy the command from the tracker to the submission queue. */
1067         memcpy(&qpair->cmd[qpair->sq_tail], &req->cmd, sizeof(req->cmd));
1068
1069         if (++qpair->sq_tail == qpair->num_entries)
1070                 qpair->sq_tail = 0;
1071
1072         bus_dmamap_sync(qpair->dma_tag, qpair->queuemem_map,
1073             BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1074         bus_space_write_4(qpair->ctrlr->bus_tag, qpair->ctrlr->bus_handle,
1075             qpair->sq_tdbl_off, qpair->sq_tail);
1076         qpair->num_cmds++;
1077 }
1078
1079 static void
1080 nvme_payload_map(void *arg, bus_dma_segment_t *seg, int nseg, int error)
1081 {
1082         struct nvme_tracker     *tr = arg;
1083         uint32_t                cur_nseg;
1084
1085         /*
1086          * If the mapping operation failed, return immediately.  The caller
1087          *  is responsible for detecting the error status and failing the
1088          *  tracker manually.
1089          */
1090         if (error != 0) {
1091                 nvme_printf(tr->qpair->ctrlr,
1092                     "nvme_payload_map err %d\n", error);
1093                 return;
1094         }
1095
1096         /*
1097          * Note that we specified PAGE_SIZE for alignment and max
1098          *  segment size when creating the bus dma tags.  So here
1099          *  we can safely just transfer each segment to its
1100          *  associated PRP entry.
1101          */
1102         tr->req->cmd.prp1 = htole64(seg[0].ds_addr);
1103
1104         if (nseg == 2) {
1105                 tr->req->cmd.prp2 = htole64(seg[1].ds_addr);
1106         } else if (nseg > 2) {
1107                 cur_nseg = 1;
1108                 tr->req->cmd.prp2 = htole64((uint64_t)tr->prp_bus_addr);
1109                 while (cur_nseg < nseg) {
1110                         tr->prp[cur_nseg-1] =
1111                             htole64((uint64_t)seg[cur_nseg].ds_addr);
1112                         cur_nseg++;
1113                 }
1114         } else {
1115                 /*
1116                  * prp2 should not be used by the controller
1117                  *  since there is only one segment, but set
1118                  *  to 0 just to be safe.
1119                  */
1120                 tr->req->cmd.prp2 = 0;
1121         }
1122
1123         bus_dmamap_sync(tr->qpair->dma_tag_payload, tr->payload_dma_map,
1124             BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1125         nvme_qpair_submit_tracker(tr->qpair, tr);
1126 }
1127
1128 static void
1129 _nvme_qpair_submit_request(struct nvme_qpair *qpair, struct nvme_request *req)
1130 {
1131         struct nvme_tracker     *tr;
1132         int                     err = 0;
1133
1134         mtx_assert(&qpair->lock, MA_OWNED);
1135
1136         tr = TAILQ_FIRST(&qpair->free_tr);
1137         req->qpair = qpair;
1138
1139         if (tr == NULL || qpair->recovery_state != RECOVERY_NONE) {
1140                 /*
1141                  * No tracker is available, or the qpair is disabled due to
1142                  *  an in-progress controller-level reset or controller
1143                  *  failure.
1144                  */
1145
1146                 if (qpair->ctrlr->is_failed) {
1147                         /*
1148                          * The controller has failed, so fail the request.
1149                          */
1150                         nvme_qpair_manual_complete_request(qpair, req,
1151                             NVME_SCT_GENERIC, NVME_SC_ABORTED_BY_REQUEST);
1152                 } else {
1153                         /*
1154                          * Put the request on the qpair's request queue to be
1155                          *  processed when a tracker frees up via a command
1156                          *  completion or when the controller reset is
1157                          *  completed.
1158                          */
1159                         STAILQ_INSERT_TAIL(&qpair->queued_req, req, stailq);
1160                 }
1161                 return;
1162         }
1163
1164         TAILQ_REMOVE(&qpair->free_tr, tr, tailq);
1165         TAILQ_INSERT_TAIL(&qpair->outstanding_tr, tr, tailq);
1166         if (!qpair->timer_armed)
1167                 tr->deadline = SBT_MAX;
1168         tr->req = req;
1169
1170         switch (req->type) {
1171         case NVME_REQUEST_VADDR:
1172                 KASSERT(req->payload_size <= qpair->ctrlr->max_xfer_size,
1173                     ("payload_size (%d) exceeds max_xfer_size (%d)\n",
1174                     req->payload_size, qpair->ctrlr->max_xfer_size));
1175                 err = bus_dmamap_load(tr->qpair->dma_tag_payload,
1176                     tr->payload_dma_map, req->u.payload, req->payload_size,
1177                     nvme_payload_map, tr, 0);
1178                 if (err != 0)
1179                         nvme_printf(qpair->ctrlr,
1180                             "bus_dmamap_load returned 0x%x!\n", err);
1181                 break;
1182         case NVME_REQUEST_NULL:
1183                 nvme_qpair_submit_tracker(tr->qpair, tr);
1184                 break;
1185         case NVME_REQUEST_BIO:
1186                 KASSERT(req->u.bio->bio_bcount <= qpair->ctrlr->max_xfer_size,
1187                     ("bio->bio_bcount (%jd) exceeds max_xfer_size (%d)\n",
1188                     (intmax_t)req->u.bio->bio_bcount,
1189                     qpair->ctrlr->max_xfer_size));
1190                 err = bus_dmamap_load_bio(tr->qpair->dma_tag_payload,
1191                     tr->payload_dma_map, req->u.bio, nvme_payload_map, tr, 0);
1192                 if (err != 0)
1193                         nvme_printf(qpair->ctrlr,
1194                             "bus_dmamap_load_bio returned 0x%x!\n", err);
1195                 break;
1196         case NVME_REQUEST_CCB:
1197                 err = bus_dmamap_load_ccb(tr->qpair->dma_tag_payload,
1198                     tr->payload_dma_map, req->u.payload,
1199                     nvme_payload_map, tr, 0);
1200                 if (err != 0)
1201                         nvme_printf(qpair->ctrlr,
1202                             "bus_dmamap_load_ccb returned 0x%x!\n", err);
1203                 break;
1204         default:
1205                 panic("unknown nvme request type 0x%x\n", req->type);
1206                 break;
1207         }
1208
1209         if (err != 0) {
1210                 /*
1211                  * The dmamap operation failed, so we manually fail the
1212                  *  tracker here with DATA_TRANSFER_ERROR status.
1213                  *
1214                  * nvme_qpair_manual_complete_tracker must not be called
1215                  *  with the qpair lock held.
1216                  */
1217                 mtx_unlock(&qpair->lock);
1218                 nvme_qpair_manual_complete_tracker(tr, NVME_SCT_GENERIC,
1219                     NVME_SC_DATA_TRANSFER_ERROR, DO_NOT_RETRY, ERROR_PRINT_ALL);
1220                 mtx_lock(&qpair->lock);
1221         }
1222 }
1223
1224 void
1225 nvme_qpair_submit_request(struct nvme_qpair *qpair, struct nvme_request *req)
1226 {
1227
1228         mtx_lock(&qpair->lock);
1229         _nvme_qpair_submit_request(qpair, req);
1230         mtx_unlock(&qpair->lock);
1231 }
1232
1233 static void
1234 nvme_qpair_enable(struct nvme_qpair *qpair)
1235 {
1236         mtx_assert(&qpair->lock, MA_OWNED);
1237
1238         qpair->recovery_state = RECOVERY_NONE;
1239 }
1240
1241 void
1242 nvme_qpair_reset(struct nvme_qpair *qpair)
1243 {
1244
1245         qpair->sq_head = qpair->sq_tail = qpair->cq_head = 0;
1246
1247         /*
1248          * First time through the completion queue, HW will set phase
1249          *  bit on completions to 1.  So set this to 1 here, indicating
1250          *  we're looking for a 1 to know which entries have completed.
1251          *  we'll toggle the bit each time when the completion queue
1252          *  rolls over.
1253          */
1254         qpair->phase = 1;
1255
1256         memset(qpair->cmd, 0,
1257             qpair->num_entries * sizeof(struct nvme_command));
1258         memset(qpair->cpl, 0,
1259             qpair->num_entries * sizeof(struct nvme_completion));
1260 }
1261
1262 void
1263 nvme_admin_qpair_enable(struct nvme_qpair *qpair)
1264 {
1265         struct nvme_tracker             *tr;
1266         struct nvme_tracker             *tr_temp;
1267
1268         /*
1269          * Manually abort each outstanding admin command.  Do not retry
1270          *  admin commands found here, since they will be left over from
1271          *  a controller reset and its likely the context in which the
1272          *  command was issued no longer applies.
1273          */
1274         TAILQ_FOREACH_SAFE(tr, &qpair->outstanding_tr, tailq, tr_temp) {
1275                 nvme_printf(qpair->ctrlr,
1276                     "aborting outstanding admin command\n");
1277                 nvme_qpair_manual_complete_tracker(tr, NVME_SCT_GENERIC,
1278                     NVME_SC_ABORTED_BY_REQUEST, DO_NOT_RETRY, ERROR_PRINT_ALL);
1279         }
1280
1281         mtx_lock(&qpair->lock);
1282         nvme_qpair_enable(qpair);
1283         mtx_unlock(&qpair->lock);
1284 }
1285
1286 void
1287 nvme_io_qpair_enable(struct nvme_qpair *qpair)
1288 {
1289         STAILQ_HEAD(, nvme_request)     temp;
1290         struct nvme_tracker             *tr;
1291         struct nvme_tracker             *tr_temp;
1292         struct nvme_request             *req;
1293
1294         /*
1295          * Manually abort each outstanding I/O.  This normally results in a
1296          *  retry, unless the retry count on the associated request has
1297          *  reached its limit.
1298          */
1299         TAILQ_FOREACH_SAFE(tr, &qpair->outstanding_tr, tailq, tr_temp) {
1300                 nvme_printf(qpair->ctrlr, "aborting outstanding i/o\n");
1301                 nvme_qpair_manual_complete_tracker(tr, NVME_SCT_GENERIC,
1302                     NVME_SC_ABORTED_BY_REQUEST, 0, ERROR_PRINT_NO_RETRY);
1303         }
1304
1305         mtx_lock(&qpair->lock);
1306
1307         nvme_qpair_enable(qpair);
1308
1309         STAILQ_INIT(&temp);
1310         STAILQ_SWAP(&qpair->queued_req, &temp, nvme_request);
1311
1312         while (!STAILQ_EMPTY(&temp)) {
1313                 req = STAILQ_FIRST(&temp);
1314                 STAILQ_REMOVE_HEAD(&temp, stailq);
1315                 nvme_printf(qpair->ctrlr, "resubmitting queued i/o\n");
1316                 nvme_qpair_print_command(qpair, &req->cmd);
1317                 _nvme_qpair_submit_request(qpair, req);
1318         }
1319
1320         mtx_unlock(&qpair->lock);
1321 }
1322
1323 static void
1324 nvme_qpair_disable(struct nvme_qpair *qpair)
1325 {
1326         struct nvme_tracker     *tr, *tr_temp;
1327
1328         mtx_lock(&qpair->lock);
1329         qpair->recovery_state = RECOVERY_WAITING;
1330         TAILQ_FOREACH_SAFE(tr, &qpair->outstanding_tr, tailq, tr_temp) {
1331                 tr->deadline = SBT_MAX;
1332         }
1333         mtx_unlock(&qpair->lock);
1334 }
1335
1336 void
1337 nvme_admin_qpair_disable(struct nvme_qpair *qpair)
1338 {
1339
1340         nvme_qpair_disable(qpair);
1341         nvme_admin_qpair_abort_aers(qpair);
1342 }
1343
1344 void
1345 nvme_io_qpair_disable(struct nvme_qpair *qpair)
1346 {
1347
1348         nvme_qpair_disable(qpair);
1349 }
1350
1351 void
1352 nvme_qpair_fail(struct nvme_qpair *qpair)
1353 {
1354         struct nvme_tracker             *tr;
1355         struct nvme_request             *req;
1356
1357         if (!mtx_initialized(&qpair->lock))
1358                 return;
1359
1360         mtx_lock(&qpair->lock);
1361
1362         while (!STAILQ_EMPTY(&qpair->queued_req)) {
1363                 req = STAILQ_FIRST(&qpair->queued_req);
1364                 STAILQ_REMOVE_HEAD(&qpair->queued_req, stailq);
1365                 nvme_printf(qpair->ctrlr, "failing queued i/o\n");
1366                 mtx_unlock(&qpair->lock);
1367                 nvme_qpair_manual_complete_request(qpair, req, NVME_SCT_GENERIC,
1368                     NVME_SC_ABORTED_BY_REQUEST);
1369                 mtx_lock(&qpair->lock);
1370         }
1371
1372         /* Manually abort each outstanding I/O. */
1373         while (!TAILQ_EMPTY(&qpair->outstanding_tr)) {
1374                 tr = TAILQ_FIRST(&qpair->outstanding_tr);
1375                 /*
1376                  * Do not remove the tracker.  The abort_tracker path will
1377                  *  do that for us.
1378                  */
1379                 nvme_printf(qpair->ctrlr, "failing outstanding i/o\n");
1380                 mtx_unlock(&qpair->lock);
1381                 nvme_qpair_manual_complete_tracker(tr, NVME_SCT_GENERIC,
1382                     NVME_SC_ABORTED_BY_REQUEST, DO_NOT_RETRY, ERROR_PRINT_ALL);
1383                 mtx_lock(&qpair->lock);
1384         }
1385
1386         mtx_unlock(&qpair->lock);
1387 }