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                 done++;
 633                 if (tr != NULL) {
 634                         nvme_qpair_complete_tracker(tr, &cpl, ERROR_PRINT_ALL);
 635                         qpair->sq_head = cpl.sqhd;
 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
 669         if (done != 0) {
 670                 bus_space_write_4(qpair->ctrlr->bus_tag, qpair->ctrlr->bus_handle,
 671                     qpair->cq_hdbl_off, qpair->cq_head);
 672         }
 673
 674         return (done != 0);
 675 }
 676
 677 static void
 678 nvme_qpair_msi_handler(void *arg)
 679 {
 680         struct nvme_qpair *qpair = arg;
 681
 682         nvme_qpair_process_completions(qpair);
 683 }
 684
 685 int
 686 nvme_qpair_construct(struct nvme_qpair *qpair,
 687     uint32_t num_entries, uint32_t num_trackers,
 688     struct nvme_controller *ctrlr)
 689 {
 690         struct nvme_tracker     *tr;
 691         size_t                  cmdsz, cplsz, prpsz, allocsz, prpmemsz;
 692         uint64_t                queuemem_phys, prpmem_phys, list_phys;
 693         uint8_t                 *queuemem, *prpmem, *prp_list;
 694         int                     i, err;
 695
 696         qpair->vector = ctrlr->msi_count > 1 ? qpair->id : 0;
 697         qpair->num_entries = num_entries;
 698         qpair->num_trackers = num_trackers;
 699         qpair->ctrlr = ctrlr;
 700
 701         mtx_init(&qpair->lock, "nvme qpair lock", NULL, MTX_DEF);
 702
 703         /* Note: NVMe PRP format is restricted to 4-byte alignment. */
 704         err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev),
 705             4, PAGE_SIZE, BUS_SPACE_MAXADDR,
 706             BUS_SPACE_MAXADDR, NULL, NULL, ctrlr->max_xfer_size,
 707             btoc(ctrlr->max_xfer_size) + 1, PAGE_SIZE, 0,
 708             NULL, NULL, &qpair->dma_tag_payload);
 709         if (err != 0) {
 710                 nvme_printf(ctrlr, "payload tag create failed %d\n", err);
 711                 goto out;
 712         }
 713
 714         /*
 715          * Each component must be page aligned, and individual PRP lists
 716          * cannot cross a page boundary.
 717          */
 718         cmdsz = qpair->num_entries * sizeof(struct nvme_command);
 719         cmdsz = roundup2(cmdsz, PAGE_SIZE);
 720         cplsz = qpair->num_entries * sizeof(struct nvme_completion);
 721         cplsz = roundup2(cplsz, PAGE_SIZE);
 722         /*
 723          * For commands requiring more than 2 PRP entries, one PRP will be
 724          * embedded in the command (prp1), and the rest of the PRP entries
 725          * will be in a list pointed to by the command (prp2).
 726          */
 727         prpsz = sizeof(uint64_t) * btoc(ctrlr->max_xfer_size);
 728         prpmemsz = qpair->num_trackers * prpsz;
 729         allocsz = cmdsz + cplsz + prpmemsz;
 730
 731         err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev),
 732             PAGE_SIZE, 0, BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
 733             allocsz, 1, allocsz, 0, NULL, NULL, &qpair->dma_tag);
 734         if (err != 0) {
 735                 nvme_printf(ctrlr, "tag create failed %d\n", err);
 736                 goto out;
 737         }
 738         bus_dma_tag_set_domain(qpair->dma_tag, qpair->domain);
 739
 740         if (bus_dmamem_alloc(qpair->dma_tag, (void **)&queuemem,
 741              BUS_DMA_COHERENT | BUS_DMA_NOWAIT, &qpair->queuemem_map)) {
 742                 nvme_printf(ctrlr, "failed to alloc qpair memory\n");
 743                 goto out;
 744         }
 745
 746         if (bus_dmamap_load(qpair->dma_tag, qpair->queuemem_map,
 747             queuemem, allocsz, nvme_single_map, &queuemem_phys, 0) != 0) {
 748                 nvme_printf(ctrlr, "failed to load qpair memory\n");
 749                 bus_dmamem_free(qpair->dma_tag, qpair->cmd,
 750                     qpair->queuemem_map);
 751                 goto out;
 752         }
 753
 754         qpair->num_cmds = 0;
 755         qpair->num_intr_handler_calls = 0;
 756         qpair->num_retries = 0;
 757         qpair->num_failures = 0;
 758         qpair->num_ignored = 0;
 759         qpair->cmd = (struct nvme_command *)queuemem;
 760         qpair->cpl = (struct nvme_completion *)(queuemem + cmdsz);
 761         prpmem = (uint8_t *)(queuemem + cmdsz + cplsz);
 762         qpair->cmd_bus_addr = queuemem_phys;
 763         qpair->cpl_bus_addr = queuemem_phys + cmdsz;
 764         prpmem_phys = queuemem_phys + cmdsz + cplsz;
 765
 766         callout_init(&qpair->timer, 1);
 767         qpair->timer_armed = false;
 768         qpair->recovery_state = RECOVERY_WAITING;
 769
 770         /*
 771          * Calcuate the stride of the doorbell register. Many emulators set this
 772          * value to correspond to a cache line. However, some hardware has set
 773          * it to various small values.
 774          */
 775         qpair->sq_tdbl_off = nvme_mmio_offsetof(doorbell[0]) +
 776             (qpair->id << (ctrlr->dstrd + 1));
 777         qpair->cq_hdbl_off = nvme_mmio_offsetof(doorbell[0]) +
 778             (qpair->id << (ctrlr->dstrd + 1)) + (1 << ctrlr->dstrd);
 779
 780         TAILQ_INIT(&qpair->free_tr);
 781         TAILQ_INIT(&qpair->outstanding_tr);
 782         STAILQ_INIT(&qpair->queued_req);
 783
 784         list_phys = prpmem_phys;
 785         prp_list = prpmem;
 786         for (i = 0; i < qpair->num_trackers; i++) {
 787                 if (list_phys + prpsz > prpmem_phys + prpmemsz) {
 788                         qpair->num_trackers = i;
 789                         break;
 790                 }
 791
 792                 /*
 793                  * Make sure that the PRP list for this tracker doesn't
 794                  * overflow to another page.
 795                  */
 796                 if (trunc_page(list_phys) !=
 797                     trunc_page(list_phys + prpsz - 1)) {
 798                         list_phys = roundup2(list_phys, PAGE_SIZE);
 799                         prp_list =
 800                             (uint8_t *)roundup2((uintptr_t)prp_list, PAGE_SIZE);
 801                 }
 802
 803                 tr = malloc_domainset(sizeof(*tr), M_NVME,
 804                     DOMAINSET_PREF(qpair->domain), M_ZERO | M_WAITOK);
 805                 bus_dmamap_create(qpair->dma_tag_payload, 0,
 806                     &tr->payload_dma_map);
 807                 tr->cid = i;
 808                 tr->qpair = qpair;
 809                 tr->prp = (uint64_t *)prp_list;
 810                 tr->prp_bus_addr = list_phys;
 811                 TAILQ_INSERT_HEAD(&qpair->free_tr, tr, tailq);
 812                 list_phys += prpsz;
 813                 prp_list += prpsz;
 814         }
 815
 816         if (qpair->num_trackers == 0) {
 817                 nvme_printf(ctrlr, "failed to allocate enough trackers\n");
 818                 goto out;
 819         }
 820
 821         qpair->act_tr = malloc_domainset(sizeof(struct nvme_tracker *) *
 822             qpair->num_entries, M_NVME, DOMAINSET_PREF(qpair->domain),
 823             M_ZERO | M_WAITOK);
 824
 825         if (ctrlr->msi_count > 1) {
 826                 /*
 827                  * MSI-X vector resource IDs start at 1, so we add one to
 828                  *  the queue's vector to get the corresponding rid to use.
 829                  */
 830                 qpair->rid = qpair->vector + 1;
 831
 832                 qpair->res = bus_alloc_resource_any(ctrlr->dev, SYS_RES_IRQ,
 833                     &qpair->rid, RF_ACTIVE);
 834                 if (qpair->res == NULL) {
 835                         nvme_printf(ctrlr, "unable to allocate MSI\n");
 836                         goto out;
 837                 }
 838                 if (bus_setup_intr(ctrlr->dev, qpair->res,
 839                     INTR_TYPE_MISC | INTR_MPSAFE, NULL,
 840                     nvme_qpair_msi_handler, qpair, &qpair->tag) != 0) {
 841                         nvme_printf(ctrlr, "unable to setup MSI\n");
 842                         goto out;
 843                 }
 844                 if (qpair->id == 0) {
 845                         bus_describe_intr(ctrlr->dev, qpair->res, qpair->tag,
 846                             "admin");
 847                 } else {
 848                         bus_describe_intr(ctrlr->dev, qpair->res, qpair->tag,
 849                             "io%d", qpair->id - 1);
 850                 }
 851         }
 852
 853         return (0);
 854
 855 out:
 856         nvme_qpair_destroy(qpair);
 857         return (ENOMEM);
 858 }
 859
 860 static void
 861 nvme_qpair_destroy(struct nvme_qpair *qpair)
 862 {
 863         struct nvme_tracker     *tr;
 864
 865         callout_drain(&qpair->timer);
 866
 867         if (qpair->tag) {
 868                 bus_teardown_intr(qpair->ctrlr->dev, qpair->res, qpair->tag);
 869                 qpair->tag = NULL;
 870         }
 871
 872         if (qpair->act_tr) {
 873                 free(qpair->act_tr, M_NVME);
 874                 qpair->act_tr = NULL;
 875         }
 876
 877         while (!TAILQ_EMPTY(&qpair->free_tr)) {
 878                 tr = TAILQ_FIRST(&qpair->free_tr);
 879                 TAILQ_REMOVE(&qpair->free_tr, tr, tailq);
 880                 bus_dmamap_destroy(qpair->dma_tag_payload,
 881                     tr->payload_dma_map);
 882                 free(tr, M_NVME);
 883         }
 884
 885         if (qpair->cmd != NULL) {
 886                 bus_dmamap_unload(qpair->dma_tag, qpair->queuemem_map);
 887                 bus_dmamem_free(qpair->dma_tag, qpair->cmd,
 888                     qpair->queuemem_map);
 889                 qpair->cmd = NULL;
 890         }
 891
 892         if (qpair->dma_tag) {
 893                 bus_dma_tag_destroy(qpair->dma_tag);
 894                 qpair->dma_tag = NULL;
 895         }
 896
 897         if (qpair->dma_tag_payload) {
 898                 bus_dma_tag_destroy(qpair->dma_tag_payload);
 899                 qpair->dma_tag_payload = NULL;
 900         }
 901
 902         if (mtx_initialized(&qpair->lock))
 903                 mtx_destroy(&qpair->lock);
 904
 905         if (qpair->res) {
 906                 bus_release_resource(qpair->ctrlr->dev, SYS_RES_IRQ,
 907                     rman_get_rid(qpair->res), qpair->res);
 908                 qpair->res = NULL;
 909         }
 910 }
 911
 912 static void
 913 nvme_admin_qpair_abort_aers(struct nvme_qpair *qpair)
 914 {
 915         struct nvme_tracker     *tr;
 916
 917         tr = TAILQ_FIRST(&qpair->outstanding_tr);
 918         while (tr != NULL) {
 919                 if (tr->req->cmd.opc == NVME_OPC_ASYNC_EVENT_REQUEST) {
 920                         nvme_qpair_manual_complete_tracker(tr,
 921                             NVME_SCT_GENERIC, NVME_SC_ABORTED_SQ_DELETION, 0,
 922                             ERROR_PRINT_NONE);
 923                         tr = TAILQ_FIRST(&qpair->outstanding_tr);
 924                 } else {
 925                         tr = TAILQ_NEXT(tr, tailq);
 926                 }
 927         }
 928 }
 929
 930 void
 931 nvme_admin_qpair_destroy(struct nvme_qpair *qpair)
 932 {
 933
 934         nvme_admin_qpair_abort_aers(qpair);
 935         nvme_qpair_destroy(qpair);
 936 }
 937
 938 void
 939 nvme_io_qpair_destroy(struct nvme_qpair *qpair)
 940 {
 941
 942         nvme_qpair_destroy(qpair);
 943 }
 944
 945 static void
 946 nvme_qpair_timeout(void *arg)
 947 {
 948         struct nvme_qpair       *qpair = arg;
 949         struct nvme_controller  *ctrlr = qpair->ctrlr;
 950         struct nvme_tracker     *tr;
 951         sbintime_t              now;
 952         bool                    idle;
 953         uint32_t                csts;
 954         uint8_t                 cfs;
 955
 956         mtx_lock(&qpair->lock);
 957         idle = TAILQ_EMPTY(&qpair->outstanding_tr);
 958 again:
 959         switch (qpair->recovery_state) {
 960         case RECOVERY_NONE:
 961                 if (idle)
 962                         break;
 963                 now = getsbinuptime();
 964                 idle = true;
 965                 TAILQ_FOREACH(tr, &qpair->outstanding_tr, tailq) {
 966                         if (tr->deadline == SBT_MAX)
 967                                 continue;
 968                         idle = false;
 969                         if (now > tr->deadline) {
 970                                 /*
 971                                  * We're now passed our earliest deadline. We
 972                                  * need to do expensive things to cope, but next
 973                                  * time. Flag that and close the door to any
 974                                  * further processing.
 975                                  */
 976                                 qpair->recovery_state = RECOVERY_START;
 977                                 nvme_printf(ctrlr, "RECOVERY_START %jd vs %jd\n",
 978                                     (uintmax_t)now, (uintmax_t)tr->deadline);
 979                                 break;
 980                         }
 981                 }
 982                 break;
 983         case RECOVERY_START:
 984                 /*
 985                  * Read csts to get value of cfs - controller fatal status.
 986                  * If no fatal status, try to call the completion routine, and
 987                  * if completes transactions, report a missed interrupt and
 988                  * return (this may need to be rate limited). Otherwise, if
 989                  * aborts are enabled and the controller is not reporting
 990                  * fatal status, abort the command. Otherwise, just reset the
 991                  * controller and hope for the best.
 992                  */
 993                 csts = nvme_mmio_read_4(ctrlr, csts);
 994                 cfs = (csts >> NVME_CSTS_REG_CFS_SHIFT) & NVME_CSTS_REG_CFS_MASK;
 995                 if (cfs) {
 996                         nvme_printf(ctrlr, "Controller in fatal status, resetting\n");
 997                         qpair->recovery_state = RECOVERY_RESET;
 998                         goto again;
 999                 }
1000                 mtx_unlock(&qpair->lock);
1001                 if (nvme_qpair_process_completions(qpair)) {
1002                         nvme_printf(ctrlr, "Completions present in output without an interrupt\n");
1003                         qpair->recovery_state = RECOVERY_NONE;
1004                 } else {
1005                         nvme_printf(ctrlr, "timeout with nothing complete, resetting\n");
1006                         qpair->recovery_state = RECOVERY_RESET;
1007                         mtx_lock(&qpair->lock);
1008                         goto again;
1009                 }
1010                 mtx_lock(&qpair->lock);
1011                 break;
1012         case RECOVERY_RESET:
1013                 /*
1014                  * If we get here due to a possible surprise hot-unplug event,
1015                  * then we let nvme_ctrlr_reset confirm and fail the
1016                  * controller.
1017                  */
1018                 nvme_printf(ctrlr, "Resetting controller due to a timeout%s.\n",
1019                     (csts == 0xffffffff) ? " and possible hot unplug" :
1020                     (cfs ? " and fatal error status" : ""));
1021                 nvme_printf(ctrlr, "RECOVERY_WAITING\n");
1022                 qpair->recovery_state = RECOVERY_WAITING;
1023                 nvme_ctrlr_reset(ctrlr);
1024                 break;
1025         case RECOVERY_WAITING:
1026                 nvme_printf(ctrlr, "waiting\n");
1027                 break;
1028         }
1029
1030         /*
1031          * Rearm the timeout.
1032          */
1033         if (!idle) {
1034                 callout_schedule_sbt(&qpair->timer, SBT_1S / 2, SBT_1S / 2, 0);
1035         } else {
1036                 qpair->timer_armed = false;
1037         }
1038         mtx_unlock(&qpair->lock);
1039 }
1040
1041 /*
1042  * Submit the tracker to the hardware. Must already be in the
1043  * outstanding queue when called.
1044  */
1045 void
1046 nvme_qpair_submit_tracker(struct nvme_qpair *qpair, struct nvme_tracker *tr)
1047 {
1048         struct nvme_request     *req;
1049         struct nvme_controller  *ctrlr;
1050         int timeout;
1051
1052         mtx_assert(&qpair->lock, MA_OWNED);
1053
1054         req = tr->req;
1055         req->cmd.cid = tr->cid;
1056         qpair->act_tr[tr->cid] = tr;
1057         ctrlr = qpair->ctrlr;
1058
1059         if (req->timeout) {
1060                 if (req->cb_fn == nvme_completion_poll_cb)
1061                         timeout = 1;
1062                 else
1063                         timeout = ctrlr->timeout_period;
1064                 tr->deadline = getsbinuptime() + timeout * SBT_1S;
1065                 if (!qpair->timer_armed) {
1066                         qpair->timer_armed = true;
1067                         callout_reset_sbt_on(&qpair->timer, SBT_1S / 2, SBT_1S / 2,
1068                             nvme_qpair_timeout, qpair, qpair->cpu, 0);
1069                 }
1070         } else
1071                 tr->deadline = SBT_MAX;
1072
1073         /* Copy the command from the tracker to the submission queue. */
1074         memcpy(&qpair->cmd[qpair->sq_tail], &req->cmd, sizeof(req->cmd));
1075
1076         if (++qpair->sq_tail == qpair->num_entries)
1077                 qpair->sq_tail = 0;
1078
1079         bus_dmamap_sync(qpair->dma_tag, qpair->queuemem_map,
1080             BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1081         bus_space_write_4(qpair->ctrlr->bus_tag, qpair->ctrlr->bus_handle,
1082             qpair->sq_tdbl_off, qpair->sq_tail);
1083         qpair->num_cmds++;
1084 }
1085
1086 static void
1087 nvme_payload_map(void *arg, bus_dma_segment_t *seg, int nseg, int error)
1088 {
1089         struct nvme_tracker     *tr = arg;
1090         uint32_t                cur_nseg;
1091
1092         /*
1093          * If the mapping operation failed, return immediately.  The caller
1094          *  is responsible for detecting the error status and failing the
1095          *  tracker manually.
1096          */
1097         if (error != 0) {
1098                 nvme_printf(tr->qpair->ctrlr,
1099                     "nvme_payload_map err %d\n", error);
1100                 return;
1101         }
1102
1103         /*
1104          * Note that we specified PAGE_SIZE for alignment and max
1105          *  segment size when creating the bus dma tags.  So here
1106          *  we can safely just transfer each segment to its
1107          *  associated PRP entry.
1108          */
1109         tr->req->cmd.prp1 = htole64(seg[0].ds_addr);
1110
1111         if (nseg == 2) {
1112                 tr->req->cmd.prp2 = htole64(seg[1].ds_addr);
1113         } else if (nseg > 2) {
1114                 cur_nseg = 1;
1115                 tr->req->cmd.prp2 = htole64((uint64_t)tr->prp_bus_addr);
1116                 while (cur_nseg < nseg) {
1117                         tr->prp[cur_nseg-1] =
1118                             htole64((uint64_t)seg[cur_nseg].ds_addr);
1119                         cur_nseg++;
1120                 }
1121         } else {
1122                 /*
1123                  * prp2 should not be used by the controller
1124                  *  since there is only one segment, but set
1125                  *  to 0 just to be safe.
1126                  */
1127                 tr->req->cmd.prp2 = 0;
1128         }
1129
1130         bus_dmamap_sync(tr->qpair->dma_tag_payload, tr->payload_dma_map,
1131             BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1132         nvme_qpair_submit_tracker(tr->qpair, tr);
1133 }
1134
1135 static void
1136 _nvme_qpair_submit_request(struct nvme_qpair *qpair, struct nvme_request *req)
1137 {
1138         struct nvme_tracker     *tr;
1139         int                     err = 0;
1140
1141         mtx_assert(&qpair->lock, MA_OWNED);
1142
1143         tr = TAILQ_FIRST(&qpair->free_tr);
1144         req->qpair = qpair;
1145
1146         if (tr == NULL || qpair->recovery_state != RECOVERY_NONE) {
1147                 /*
1148                  * No tracker is available, or the qpair is disabled due to
1149                  *  an in-progress controller-level reset or controller
1150                  *  failure.
1151                  */
1152
1153                 if (qpair->ctrlr->is_failed) {
1154                         /*
1155                          * The controller has failed, so fail the request.
1156                          */
1157                         nvme_qpair_manual_complete_request(qpair, req,
1158                             NVME_SCT_GENERIC, NVME_SC_ABORTED_BY_REQUEST);
1159                 } else {
1160                         /*
1161                          * Put the request on the qpair's request queue to be
1162                          *  processed when a tracker frees up via a command
1163                          *  completion or when the controller reset is
1164                          *  completed.
1165                          */
1166                         STAILQ_INSERT_TAIL(&qpair->queued_req, req, stailq);
1167                 }
1168                 return;
1169         }
1170
1171         TAILQ_REMOVE(&qpair->free_tr, tr, tailq);
1172         TAILQ_INSERT_TAIL(&qpair->outstanding_tr, tr, tailq);
1173         if (!qpair->timer_armed)
1174                 tr->deadline = SBT_MAX;
1175         tr->req = req;
1176
1177         switch (req->type) {
1178         case NVME_REQUEST_VADDR:
1179                 KASSERT(req->payload_size <= qpair->ctrlr->max_xfer_size,
1180                     ("payload_size (%d) exceeds max_xfer_size (%d)\n",
1181                     req->payload_size, qpair->ctrlr->max_xfer_size));
1182                 err = bus_dmamap_load(tr->qpair->dma_tag_payload,
1183                     tr->payload_dma_map, req->u.payload, req->payload_size,
1184                     nvme_payload_map, tr, 0);
1185                 if (err != 0)
1186                         nvme_printf(qpair->ctrlr,
1187                             "bus_dmamap_load returned 0x%x!\n", err);
1188                 break;
1189         case NVME_REQUEST_NULL:
1190                 nvme_qpair_submit_tracker(tr->qpair, tr);
1191                 break;
1192         case NVME_REQUEST_BIO:
1193                 KASSERT(req->u.bio->bio_bcount <= qpair->ctrlr->max_xfer_size,
1194                     ("bio->bio_bcount (%jd) exceeds max_xfer_size (%d)\n",
1195                     (intmax_t)req->u.bio->bio_bcount,
1196                     qpair->ctrlr->max_xfer_size));
1197                 err = bus_dmamap_load_bio(tr->qpair->dma_tag_payload,
1198                     tr->payload_dma_map, req->u.bio, nvme_payload_map, tr, 0);
1199                 if (err != 0)
1200                         nvme_printf(qpair->ctrlr,
1201                             "bus_dmamap_load_bio returned 0x%x!\n", err);
1202                 break;
1203         case NVME_REQUEST_CCB:
1204                 err = bus_dmamap_load_ccb(tr->qpair->dma_tag_payload,
1205                     tr->payload_dma_map, req->u.payload,
1206                     nvme_payload_map, tr, 0);
1207                 if (err != 0)
1208                         nvme_printf(qpair->ctrlr,
1209                             "bus_dmamap_load_ccb returned 0x%x!\n", err);
1210                 break;
1211         default:
1212                 panic("unknown nvme request type 0x%x\n", req->type);
1213                 break;
1214         }
1215
1216         if (err != 0) {
1217                 /*
1218                  * The dmamap operation failed, so we manually fail the
1219                  *  tracker here with DATA_TRANSFER_ERROR status.
1220                  *
1221                  * nvme_qpair_manual_complete_tracker must not be called
1222                  *  with the qpair lock held.
1223                  */
1224                 mtx_unlock(&qpair->lock);
1225                 nvme_qpair_manual_complete_tracker(tr, NVME_SCT_GENERIC,
1226                     NVME_SC_DATA_TRANSFER_ERROR, DO_NOT_RETRY, ERROR_PRINT_ALL);
1227                 mtx_lock(&qpair->lock);
1228         }
1229 }
1230
1231 void
1232 nvme_qpair_submit_request(struct nvme_qpair *qpair, struct nvme_request *req)
1233 {
1234
1235         mtx_lock(&qpair->lock);
1236         _nvme_qpair_submit_request(qpair, req);
1237         mtx_unlock(&qpair->lock);
1238 }
1239
1240 static void
1241 nvme_qpair_enable(struct nvme_qpair *qpair)
1242 {
1243         mtx_assert(&qpair->lock, MA_OWNED);
1244
1245         qpair->recovery_state = RECOVERY_NONE;
1246 }
1247
1248 void
1249 nvme_qpair_reset(struct nvme_qpair *qpair)
1250 {
1251
1252         qpair->sq_head = qpair->sq_tail = qpair->cq_head = 0;
1253
1254         /*
1255          * First time through the completion queue, HW will set phase
1256          *  bit on completions to 1.  So set this to 1 here, indicating
1257          *  we're looking for a 1 to know which entries have completed.
1258          *  we'll toggle the bit each time when the completion queue
1259          *  rolls over.
1260          */
1261         qpair->phase = 1;
1262
1263         memset(qpair->cmd, 0,
1264             qpair->num_entries * sizeof(struct nvme_command));
1265         memset(qpair->cpl, 0,
1266             qpair->num_entries * sizeof(struct nvme_completion));
1267 }
1268
1269 void
1270 nvme_admin_qpair_enable(struct nvme_qpair *qpair)
1271 {
1272         struct nvme_tracker             *tr;
1273         struct nvme_tracker             *tr_temp;
1274
1275         /*
1276          * Manually abort each outstanding admin command.  Do not retry
1277          *  admin commands found here, since they will be left over from
1278          *  a controller reset and its likely the context in which the
1279          *  command was issued no longer applies.
1280          */
1281         TAILQ_FOREACH_SAFE(tr, &qpair->outstanding_tr, tailq, tr_temp) {
1282                 nvme_printf(qpair->ctrlr,
1283                     "aborting outstanding admin command\n");
1284                 nvme_qpair_manual_complete_tracker(tr, NVME_SCT_GENERIC,
1285                     NVME_SC_ABORTED_BY_REQUEST, DO_NOT_RETRY, ERROR_PRINT_ALL);
1286         }
1287
1288         mtx_lock(&qpair->lock);
1289         nvme_qpair_enable(qpair);
1290         mtx_unlock(&qpair->lock);
1291 }
1292
1293 void
1294 nvme_io_qpair_enable(struct nvme_qpair *qpair)
1295 {
1296         STAILQ_HEAD(, nvme_request)     temp;
1297         struct nvme_tracker             *tr;
1298         struct nvme_tracker             *tr_temp;
1299         struct nvme_request             *req;
1300
1301         /*
1302          * Manually abort each outstanding I/O.  This normally results in a
1303          *  retry, unless the retry count on the associated request has
1304          *  reached its limit.
1305          */
1306         TAILQ_FOREACH_SAFE(tr, &qpair->outstanding_tr, tailq, tr_temp) {
1307                 nvme_printf(qpair->ctrlr, "aborting outstanding i/o\n");
1308                 nvme_qpair_manual_complete_tracker(tr, NVME_SCT_GENERIC,
1309                     NVME_SC_ABORTED_BY_REQUEST, 0, ERROR_PRINT_NO_RETRY);
1310         }
1311
1312         mtx_lock(&qpair->lock);
1313
1314         nvme_qpair_enable(qpair);
1315
1316         STAILQ_INIT(&temp);
1317         STAILQ_SWAP(&qpair->queued_req, &temp, nvme_request);
1318
1319         while (!STAILQ_EMPTY(&temp)) {
1320                 req = STAILQ_FIRST(&temp);
1321                 STAILQ_REMOVE_HEAD(&temp, stailq);
1322                 nvme_printf(qpair->ctrlr, "resubmitting queued i/o\n");
1323                 nvme_qpair_print_command(qpair, &req->cmd);
1324                 _nvme_qpair_submit_request(qpair, req);
1325         }
1326
1327         mtx_unlock(&qpair->lock);
1328 }
1329
1330 static void
1331 nvme_qpair_disable(struct nvme_qpair *qpair)
1332 {
1333         struct nvme_tracker     *tr, *tr_temp;
1334
1335         mtx_lock(&qpair->lock);
1336         qpair->recovery_state = RECOVERY_WAITING;
1337         TAILQ_FOREACH_SAFE(tr, &qpair->outstanding_tr, tailq, tr_temp) {
1338                 tr->deadline = SBT_MAX;
1339         }
1340         mtx_unlock(&qpair->lock);
1341 }
1342
1343 void
1344 nvme_admin_qpair_disable(struct nvme_qpair *qpair)
1345 {
1346
1347         nvme_qpair_disable(qpair);
1348         nvme_admin_qpair_abort_aers(qpair);
1349 }
1350
1351 void
1352 nvme_io_qpair_disable(struct nvme_qpair *qpair)
1353 {
1354
1355         nvme_qpair_disable(qpair);
1356 }
1357
1358 void
1359 nvme_qpair_fail(struct nvme_qpair *qpair)
1360 {
1361         struct nvme_tracker             *tr;
1362         struct nvme_request             *req;
1363
1364         if (!mtx_initialized(&qpair->lock))
1365                 return;
1366
1367         mtx_lock(&qpair->lock);
1368
1369         while (!STAILQ_EMPTY(&qpair->queued_req)) {
1370                 req = STAILQ_FIRST(&qpair->queued_req);
1371                 STAILQ_REMOVE_HEAD(&qpair->queued_req, stailq);
1372                 nvme_printf(qpair->ctrlr, "failing queued i/o\n");
1373                 mtx_unlock(&qpair->lock);
1374                 nvme_qpair_manual_complete_request(qpair, req, NVME_SCT_GENERIC,
1375                     NVME_SC_ABORTED_BY_REQUEST);
1376                 mtx_lock(&qpair->lock);
1377         }
1378
1379         /* Manually abort each outstanding I/O. */
1380         while (!TAILQ_EMPTY(&qpair->outstanding_tr)) {
1381                 tr = TAILQ_FIRST(&qpair->outstanding_tr);
1382                 /*
1383                  * Do not remove the tracker.  The abort_tracker path will
1384                  *  do that for us.
1385                  */
1386                 nvme_printf(qpair->ctrlr, "failing outstanding i/o\n");
1387                 mtx_unlock(&qpair->lock);
1388                 nvme_qpair_manual_complete_tracker(tr, NVME_SCT_GENERIC,
1389                     NVME_SC_ABORTED_BY_REQUEST, DO_NOT_RETRY, ERROR_PRINT_ALL);
1390                 mtx_lock(&qpair->lock);
1391         }
1392
1393         mtx_unlock(&qpair->lock);
1394 }