sys/dev/nvme/nvme_qpair.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
   3  *
   4  * Copyright (C) 2012-2014 Intel Corporation
   5  * All rights reserved.
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  * 1. Redistributions of source code must retain the above copyright
  11  *    notice, this list of conditions and the following disclaimer.
  12  * 2. Redistributions in binary form must reproduce the above copyright
  13  *    notice, this list of conditions and the following disclaimer in the
  14  *    documentation and/or other materials provided with the distribution.
  15  *
  16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  26  * SUCH DAMAGE.
  27  */
  28
  29 #include <sys/cdefs.h>
  30 __FBSDID("$FreeBSD$");
  31
  32 #include <sys/param.h>
  33 #include <sys/bus.h>
  34 #include <sys/conf.h>
  35 #include <sys/proc.h>
  36
  37 #include <dev/pci/pcivar.h>
  38
  39 #include "nvme_private.h"
  40
  41 typedef enum error_print { ERROR_PRINT_NONE, ERROR_PRINT_NO_RETRY, ERROR_PRINT_ALL } error_print_t;
  42 #define DO_NOT_RETRY    1
  43
  44 static void     _nvme_qpair_submit_request(struct nvme_qpair *qpair,
  45                                            struct nvme_request *req);
  46 static void     nvme_qpair_destroy(struct nvme_qpair *qpair);
  47
  48 struct nvme_opcode_string {
  49
  50         uint16_t        opc;
  51         const char *    str;
  52 };
  53
  54 static struct nvme_opcode_string admin_opcode[] = {
  55         { NVME_OPC_DELETE_IO_SQ, "DELETE IO SQ" },
  56         { NVME_OPC_CREATE_IO_SQ, "CREATE IO SQ" },
  57         { NVME_OPC_GET_LOG_PAGE, "GET LOG PAGE" },
  58         { NVME_OPC_DELETE_IO_CQ, "DELETE IO CQ" },
  59         { NVME_OPC_CREATE_IO_CQ, "CREATE IO CQ" },
  60         { NVME_OPC_IDENTIFY, "IDENTIFY" },
  61         { NVME_OPC_ABORT, "ABORT" },
  62         { NVME_OPC_SET_FEATURES, "SET FEATURES" },
  63         { NVME_OPC_GET_FEATURES, "GET FEATURES" },
  64         { NVME_OPC_ASYNC_EVENT_REQUEST, "ASYNC EVENT REQUEST" },
  65         { NVME_OPC_FIRMWARE_ACTIVATE, "FIRMWARE ACTIVATE" },
  66         { NVME_OPC_FIRMWARE_IMAGE_DOWNLOAD, "FIRMWARE IMAGE DOWNLOAD" },
  67         { NVME_OPC_DEVICE_SELF_TEST, "DEVICE SELF-TEST" },
  68         { NVME_OPC_NAMESPACE_ATTACHMENT, "NAMESPACE ATTACHMENT" },
  69         { NVME_OPC_KEEP_ALIVE, "KEEP ALIVE" },
  70         { NVME_OPC_DIRECTIVE_SEND, "DIRECTIVE SEND" },
  71         { NVME_OPC_DIRECTIVE_RECEIVE, "DIRECTIVE RECEIVE" },
  72         { NVME_OPC_VIRTUALIZATION_MANAGEMENT, "VIRTUALIZATION MANAGEMENT" },
  73         { NVME_OPC_NVME_MI_SEND, "NVME-MI SEND" },
  74         { NVME_OPC_NVME_MI_RECEIVE, "NVME-MI RECEIVE" },
  75         { NVME_OPC_DOORBELL_BUFFER_CONFIG, "DOORBELL BUFFER CONFIG" },
  76         { NVME_OPC_FORMAT_NVM, "FORMAT NVM" },
  77         { NVME_OPC_SECURITY_SEND, "SECURITY SEND" },
  78         { NVME_OPC_SECURITY_RECEIVE, "SECURITY RECEIVE" },
  79         { NVME_OPC_SANITIZE, "SANITIZE" },
  80         { 0xFFFF, "ADMIN COMMAND" }
  81 };
  82
  83 static struct nvme_opcode_string io_opcode[] = {
  84         { NVME_OPC_FLUSH, "FLUSH" },
  85         { NVME_OPC_WRITE, "WRITE" },
  86         { NVME_OPC_READ, "READ" },
  87         { NVME_OPC_WRITE_UNCORRECTABLE, "WRITE UNCORRECTABLE" },
  88         { NVME_OPC_COMPARE, "COMPARE" },
  89         { NVME_OPC_WRITE_ZEROES, "WRITE ZEROES" },
  90         { NVME_OPC_DATASET_MANAGEMENT, "DATASET MANAGEMENT" },
  91         { NVME_OPC_RESERVATION_REGISTER, "RESERVATION REGISTER" },
  92         { NVME_OPC_RESERVATION_REPORT, "RESERVATION REPORT" },
  93         { NVME_OPC_RESERVATION_ACQUIRE, "RESERVATION ACQUIRE" },
  94         { NVME_OPC_RESERVATION_RELEASE, "RESERVATION RELEASE" },
  95         { 0xFFFF, "IO COMMAND" }
  96 };
  97
  98 static const char *
  99 get_admin_opcode_string(uint16_t opc)
 100 {
 101         struct nvme_opcode_string *entry;
 102
 103         entry = admin_opcode;
 104
 105         while (entry->opc != 0xFFFF) {
 106                 if (entry->opc == opc)
 107                         return (entry->str);
 108                 entry++;
 109         }
 110         return (entry->str);
 111 }
 112
 113 static const char *
 114 get_io_opcode_string(uint16_t opc)
 115 {
 116         struct nvme_opcode_string *entry;
 117
 118         entry = io_opcode;
 119
 120         while (entry->opc != 0xFFFF) {
 121                 if (entry->opc == opc)
 122                         return (entry->str);
 123                 entry++;
 124         }
 125         return (entry->str);
 126 }
 127
 128
 129 static void
 130 nvme_admin_qpair_print_command(struct nvme_qpair *qpair,
 131     struct nvme_command *cmd)
 132 {
 133
 134         nvme_printf(qpair->ctrlr, "%s (%02x) sqid:%d cid:%d nsid:%x "
 135             "cdw10:%08x cdw11:%08x\n",
 136             get_admin_opcode_string(cmd->opc), cmd->opc, qpair->id, cmd->cid,
 137             le32toh(cmd->nsid), le32toh(cmd->cdw10), le32toh(cmd->cdw11));
 138 }
 139
 140 static void
 141 nvme_io_qpair_print_command(struct nvme_qpair *qpair,
 142     struct nvme_command *cmd)
 143 {
 144
 145         switch (cmd->opc) {
 146         case NVME_OPC_WRITE:
 147         case NVME_OPC_READ:
 148         case NVME_OPC_WRITE_UNCORRECTABLE:
 149         case NVME_OPC_COMPARE:
 150         case NVME_OPC_WRITE_ZEROES:
 151                 nvme_printf(qpair->ctrlr, "%s sqid:%d cid:%d nsid:%d "
 152                     "lba:%llu len:%d\n",
 153                     get_io_opcode_string(cmd->opc), qpair->id, cmd->cid, le32toh(cmd->nsid),
 154                     ((unsigned long long)le32toh(cmd->cdw11) << 32) + le32toh(cmd->cdw10),
 155                     (le32toh(cmd->cdw12) & 0xFFFF) + 1);
 156                 break;
 157         case NVME_OPC_FLUSH:
 158         case NVME_OPC_DATASET_MANAGEMENT:
 159         case NVME_OPC_RESERVATION_REGISTER:
 160         case NVME_OPC_RESERVATION_REPORT:
 161         case NVME_OPC_RESERVATION_ACQUIRE:
 162         case NVME_OPC_RESERVATION_RELEASE:
 163                 nvme_printf(qpair->ctrlr, "%s sqid:%d cid:%d nsid:%d\n",
 164                     get_io_opcode_string(cmd->opc), qpair->id, cmd->cid, le32toh(cmd->nsid));
 165                 break;
 166         default:
 167                 nvme_printf(qpair->ctrlr, "%s (%02x) sqid:%d cid:%d nsid:%d\n",
 168                     get_io_opcode_string(cmd->opc), cmd->opc, qpair->id,
 169                     cmd->cid, le32toh(cmd->nsid));
 170                 break;
 171         }
 172 }
 173
 174 static void
 175 nvme_qpair_print_command(struct nvme_qpair *qpair, struct nvme_command *cmd)
 176 {
 177         if (qpair->id == 0)
 178                 nvme_admin_qpair_print_command(qpair, cmd);
 179         else
 180                 nvme_io_qpair_print_command(qpair, cmd);
 181         if (nvme_verbose_cmd_dump) {
 182                 nvme_printf(qpair->ctrlr,
 183                     "nsid:%#x rsvd2:%#x rsvd3:%#x mptr:%#jx prp1:%#jx prp2:%#jx\n",
 184                     cmd->nsid, cmd->rsvd2, cmd->rsvd3, (uintmax_t)cmd->mptr,
 185                     (uintmax_t)cmd->prp1, (uintmax_t)cmd->prp2);
 186                 nvme_printf(qpair->ctrlr,
 187                     "cdw10: %#x cdw11:%#x cdw12:%#x cdw13:%#x cdw14:%#x cdw15:%#x\n",
 188                     cmd->cdw10, cmd->cdw11, cmd->cdw12, cmd->cdw13, cmd->cdw14,
 189                     cmd->cdw15);
 190         }
 191 }
 192
 193 struct nvme_status_string {
 194
 195         uint16_t        sc;
 196         const char *    str;
 197 };
 198
 199 static struct nvme_status_string generic_status[] = {
 200         { NVME_SC_SUCCESS, "SUCCESS" },
 201         { NVME_SC_INVALID_OPCODE, "INVALID OPCODE" },
 202         { NVME_SC_INVALID_FIELD, "INVALID_FIELD" },
 203         { NVME_SC_COMMAND_ID_CONFLICT, "COMMAND ID CONFLICT" },
 204         { NVME_SC_DATA_TRANSFER_ERROR, "DATA TRANSFER ERROR" },
 205         { NVME_SC_ABORTED_POWER_LOSS, "ABORTED - POWER LOSS" },
 206         { NVME_SC_INTERNAL_DEVICE_ERROR, "INTERNAL DEVICE ERROR" },
 207         { NVME_SC_ABORTED_BY_REQUEST, "ABORTED - BY REQUEST" },
 208         { NVME_SC_ABORTED_SQ_DELETION, "ABORTED - SQ DELETION" },
 209         { NVME_SC_ABORTED_FAILED_FUSED, "ABORTED - FAILED FUSED" },
 210         { NVME_SC_ABORTED_MISSING_FUSED, "ABORTED - MISSING FUSED" },
 211         { NVME_SC_INVALID_NAMESPACE_OR_FORMAT, "INVALID NAMESPACE OR FORMAT" },
 212         { NVME_SC_COMMAND_SEQUENCE_ERROR, "COMMAND SEQUENCE ERROR" },
 213         { NVME_SC_INVALID_SGL_SEGMENT_DESCR, "INVALID SGL SEGMENT DESCRIPTOR" },
 214         { NVME_SC_INVALID_NUMBER_OF_SGL_DESCR, "INVALID NUMBER OF SGL DESCRIPTORS" },
 215         { NVME_SC_DATA_SGL_LENGTH_INVALID, "DATA SGL LENGTH INVALID" },
 216         { NVME_SC_METADATA_SGL_LENGTH_INVALID, "METADATA SGL LENGTH INVALID" },
 217         { NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID, "SGL DESCRIPTOR TYPE INVALID" },
 218         { NVME_SC_INVALID_USE_OF_CMB, "INVALID USE OF CONTROLLER MEMORY BUFFER" },
 219         { NVME_SC_PRP_OFFET_INVALID, "PRP OFFET INVALID" },
 220         { NVME_SC_ATOMIC_WRITE_UNIT_EXCEEDED, "ATOMIC WRITE UNIT EXCEEDED" },
 221         { NVME_SC_OPERATION_DENIED, "OPERATION DENIED" },
 222         { NVME_SC_SGL_OFFSET_INVALID, "SGL OFFSET INVALID" },
 223         { NVME_SC_HOST_ID_INCONSISTENT_FORMAT, "HOST IDENTIFIER INCONSISTENT FORMAT" },
 224         { NVME_SC_KEEP_ALIVE_TIMEOUT_EXPIRED, "KEEP ALIVE TIMEOUT EXPIRED" },
 225         { NVME_SC_KEEP_ALIVE_TIMEOUT_INVALID, "KEEP ALIVE TIMEOUT INVALID" },
 226         { NVME_SC_ABORTED_DUE_TO_PREEMPT, "COMMAND ABORTED DUE TO PREEMPT AND ABORT" },
 227         { NVME_SC_SANITIZE_FAILED, "SANITIZE FAILED" },
 228         { NVME_SC_SANITIZE_IN_PROGRESS, "SANITIZE IN PROGRESS" },
 229         { NVME_SC_SGL_DATA_BLOCK_GRAN_INVALID, "SGL_DATA_BLOCK_GRANULARITY_INVALID" },
 230         { NVME_SC_NOT_SUPPORTED_IN_CMB, "COMMAND NOT SUPPORTED FOR QUEUE IN CMB" },
 231
 232         { NVME_SC_LBA_OUT_OF_RANGE, "LBA OUT OF RANGE" },
 233         { NVME_SC_CAPACITY_EXCEEDED, "CAPACITY EXCEEDED" },
 234         { NVME_SC_NAMESPACE_NOT_READY, "NAMESPACE NOT READY" },
 235         { NVME_SC_RESERVATION_CONFLICT, "RESERVATION CONFLICT" },
 236         { NVME_SC_FORMAT_IN_PROGRESS, "FORMAT IN PROGRESS" },
 237         { 0xFFFF, "GENERIC" }
 238 };
 239
 240 static struct nvme_status_string command_specific_status[] = {
 241         { NVME_SC_COMPLETION_QUEUE_INVALID, "INVALID COMPLETION QUEUE" },
 242         { NVME_SC_INVALID_QUEUE_IDENTIFIER, "INVALID QUEUE IDENTIFIER" },
 243         { NVME_SC_MAXIMUM_QUEUE_SIZE_EXCEEDED, "MAX QUEUE SIZE EXCEEDED" },
 244         { NVME_SC_ABORT_COMMAND_LIMIT_EXCEEDED, "ABORT CMD LIMIT EXCEEDED" },
 245         { NVME_SC_ASYNC_EVENT_REQUEST_LIMIT_EXCEEDED, "ASYNC LIMIT EXCEEDED" },
 246         { NVME_SC_INVALID_FIRMWARE_SLOT, "INVALID FIRMWARE SLOT" },
 247         { NVME_SC_INVALID_FIRMWARE_IMAGE, "INVALID FIRMWARE IMAGE" },
 248         { NVME_SC_INVALID_INTERRUPT_VECTOR, "INVALID INTERRUPT VECTOR" },
 249         { NVME_SC_INVALID_LOG_PAGE, "INVALID LOG PAGE" },
 250         { NVME_SC_INVALID_FORMAT, "INVALID FORMAT" },
 251         { NVME_SC_FIRMWARE_REQUIRES_RESET, "FIRMWARE REQUIRES RESET" },
 252         { NVME_SC_INVALID_QUEUE_DELETION, "INVALID QUEUE DELETION" },
 253         { NVME_SC_FEATURE_NOT_SAVEABLE, "FEATURE IDENTIFIER NOT SAVEABLE" },
 254         { NVME_SC_FEATURE_NOT_CHANGEABLE, "FEATURE NOT CHANGEABLE" },
 255         { NVME_SC_FEATURE_NOT_NS_SPECIFIC, "FEATURE NOT NAMESPACE SPECIFIC" },
 256         { NVME_SC_FW_ACT_REQUIRES_NVMS_RESET, "FIRMWARE ACTIVATION REQUIRES NVM SUBSYSTEM RESET" },
 257         { NVME_SC_FW_ACT_REQUIRES_RESET, "FIRMWARE ACTIVATION REQUIRES RESET" },
 258         { NVME_SC_FW_ACT_REQUIRES_TIME, "FIRMWARE ACTIVATION REQUIRES MAXIMUM TIME VIOLATION" },
 259         { NVME_SC_FW_ACT_PROHIBITED, "FIRMWARE ACTIVATION PROHIBITED" },
 260         { NVME_SC_OVERLAPPING_RANGE, "OVERLAPPING RANGE" },
 261         { NVME_SC_NS_INSUFFICIENT_CAPACITY, "NAMESPACE INSUFFICIENT CAPACITY" },
 262         { NVME_SC_NS_ID_UNAVAILABLE, "NAMESPACE IDENTIFIER UNAVAILABLE" },
 263         { NVME_SC_NS_ALREADY_ATTACHED, "NAMESPACE ALREADY ATTACHED" },
 264         { NVME_SC_NS_IS_PRIVATE, "NAMESPACE IS PRIVATE" },
 265         { NVME_SC_NS_NOT_ATTACHED, "NS NOT ATTACHED" },
 266         { NVME_SC_THIN_PROV_NOT_SUPPORTED, "THIN PROVISIONING NOT SUPPORTED" },
 267         { NVME_SC_CTRLR_LIST_INVALID, "CONTROLLER LIST INVALID" },
 268         { NVME_SC_SELT_TEST_IN_PROGRESS, "DEVICE SELT-TEST IN PROGRESS" },
 269         { NVME_SC_BOOT_PART_WRITE_PROHIB, "BOOT PARTITION WRITE PROHIBITED" },
 270         { NVME_SC_INVALID_CTRLR_ID, "INVALID CONTROLLER IDENTIFIER" },
 271         { NVME_SC_INVALID_SEC_CTRLR_STATE, "INVALID SECONDARY CONTROLLER STATE" },
 272         { NVME_SC_INVALID_NUM_OF_CTRLR_RESRC, "INVALID NUMBER OF CONTROLLER RESOURCES" },
 273         { NVME_SC_INVALID_RESOURCE_ID, "INVALID RESOURCE IDENTIFIER" },
 274
 275         { NVME_SC_CONFLICTING_ATTRIBUTES, "CONFLICTING ATTRIBUTES" },
 276         { NVME_SC_INVALID_PROTECTION_INFO, "INVALID PROTECTION INFO" },
 277         { NVME_SC_ATTEMPTED_WRITE_TO_RO_PAGE, "WRITE TO RO PAGE" },
 278         { 0xFFFF, "COMMAND SPECIFIC" }
 279 };
 280
 281 static struct nvme_status_string media_error_status[] = {
 282         { NVME_SC_WRITE_FAULTS, "WRITE FAULTS" },
 283         { NVME_SC_UNRECOVERED_READ_ERROR, "UNRECOVERED READ ERROR" },
 284         { NVME_SC_GUARD_CHECK_ERROR, "GUARD CHECK ERROR" },
 285         { NVME_SC_APPLICATION_TAG_CHECK_ERROR, "APPLICATION TAG CHECK ERROR" },
 286         { NVME_SC_REFERENCE_TAG_CHECK_ERROR, "REFERENCE TAG CHECK ERROR" },
 287         { NVME_SC_COMPARE_FAILURE, "COMPARE FAILURE" },
 288         { NVME_SC_ACCESS_DENIED, "ACCESS DENIED" },
 289         { NVME_SC_DEALLOCATED_OR_UNWRITTEN, "DEALLOCATED OR UNWRITTEN LOGICAL BLOCK" },
 290         { 0xFFFF, "MEDIA ERROR" }
 291 };
 292
 293 static const char *
 294 get_status_string(uint16_t sct, uint16_t sc)
 295 {
 296         struct nvme_status_string *entry;
 297
 298         switch (sct) {
 299         case NVME_SCT_GENERIC:
 300                 entry = generic_status;
 301                 break;
 302         case NVME_SCT_COMMAND_SPECIFIC:
 303                 entry = command_specific_status;
 304                 break;
 305         case NVME_SCT_MEDIA_ERROR:
 306                 entry = media_error_status;
 307                 break;
 308         case NVME_SCT_VENDOR_SPECIFIC:
 309                 return ("VENDOR SPECIFIC");
 310         default:
 311                 return ("RESERVED");
 312         }
 313
 314         while (entry->sc != 0xFFFF) {
 315                 if (entry->sc == sc)
 316                         return (entry->str);
 317                 entry++;
 318         }
 319         return (entry->str);
 320 }
 321
 322 static void
 323 nvme_qpair_print_completion(struct nvme_qpair *qpair,
 324     struct nvme_completion *cpl)
 325 {
 326         uint16_t sct, sc;
 327
 328         sct = NVME_STATUS_GET_SCT(cpl->status);
 329         sc = NVME_STATUS_GET_SC(cpl->status);
 330
 331         nvme_printf(qpair->ctrlr, "%s (%02x/%02x) sqid:%d cid:%d cdw0:%x\n",
 332             get_status_string(sct, sc), sct, sc, cpl->sqid, cpl->cid,
 333             cpl->cdw0);
 334 }
 335
 336 static boolean_t
 337 nvme_completion_is_retry(const struct nvme_completion *cpl)
 338 {
 339         uint8_t sct, sc, dnr;
 340
 341         sct = NVME_STATUS_GET_SCT(cpl->status);
 342         sc = NVME_STATUS_GET_SC(cpl->status);
 343         dnr = NVME_STATUS_GET_DNR(cpl->status); /* Do Not Retry Bit */
 344
 345         /*
 346          * TODO: spec is not clear how commands that are aborted due
 347          *  to TLER will be marked.  So for now, it seems
 348          *  NAMESPACE_NOT_READY is the only case where we should
 349          *  look at the DNR bit. Requests failed with ABORTED_BY_REQUEST
 350          *  set the DNR bit correctly since the driver controls that.
 351          */
 352         switch (sct) {
 353         case NVME_SCT_GENERIC:
 354                 switch (sc) {
 355                 case NVME_SC_ABORTED_BY_REQUEST:
 356                 case NVME_SC_NAMESPACE_NOT_READY:
 357                         if (dnr)
 358                                 return (0);
 359                         else
 360                                 return (1);
 361                 case NVME_SC_INVALID_OPCODE:
 362                 case NVME_SC_INVALID_FIELD:
 363                 case NVME_SC_COMMAND_ID_CONFLICT:
 364                 case NVME_SC_DATA_TRANSFER_ERROR:
 365                 case NVME_SC_ABORTED_POWER_LOSS:
 366                 case NVME_SC_INTERNAL_DEVICE_ERROR:
 367                 case NVME_SC_ABORTED_SQ_DELETION:
 368                 case NVME_SC_ABORTED_FAILED_FUSED:
 369                 case NVME_SC_ABORTED_MISSING_FUSED:
 370                 case NVME_SC_INVALID_NAMESPACE_OR_FORMAT:
 371                 case NVME_SC_COMMAND_SEQUENCE_ERROR:
 372                 case NVME_SC_LBA_OUT_OF_RANGE:
 373                 case NVME_SC_CAPACITY_EXCEEDED:
 374                 default:
 375                         return (0);
 376                 }
 377         case NVME_SCT_COMMAND_SPECIFIC:
 378         case NVME_SCT_MEDIA_ERROR:
 379         case NVME_SCT_VENDOR_SPECIFIC:
 380         default:
 381                 return (0);
 382         }
 383 }
 384
 385 static void
 386 nvme_qpair_complete_tracker(struct nvme_qpair *qpair, struct nvme_tracker *tr,
 387     struct nvme_completion *cpl, error_print_t print_on_error)
 388 {
 389         struct nvme_request     *req;
 390         boolean_t               retry, error;
 391
 392         req = tr->req;
 393         error = nvme_completion_is_error(cpl);
 394         retry = error && nvme_completion_is_retry(cpl) &&
 395            req->retries < nvme_retry_count;
 396         if (retry)
 397                 qpair->num_retries++;
 398
 399         if (error && (print_on_error == ERROR_PRINT_ALL ||
 400                 (!retry && print_on_error == ERROR_PRINT_NO_RETRY))) {
 401                 nvme_qpair_print_command(qpair, &req->cmd);
 402                 nvme_qpair_print_completion(qpair, cpl);
 403         }
 404
 405         qpair->act_tr[cpl->cid] = NULL;
 406
 407         KASSERT(cpl->cid == req->cmd.cid, ("cpl cid does not match cmd cid\n"));
 408
 409         if (req->cb_fn && !retry)
 410                 req->cb_fn(req->cb_arg, cpl);
 411
 412         mtx_lock(&qpair->lock);
 413         callout_stop(&tr->timer);
 414
 415         if (retry) {
 416                 req->retries++;
 417                 nvme_qpair_submit_tracker(qpair, tr);
 418         } else {
 419                 if (req->type != NVME_REQUEST_NULL) {
 420                         bus_dmamap_sync(qpair->dma_tag_payload,
 421                             tr->payload_dma_map,
 422                             BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
 423                         bus_dmamap_unload(qpair->dma_tag_payload,
 424                             tr->payload_dma_map);
 425                 }
 426
 427                 nvme_free_request(req);
 428                 tr->req = NULL;
 429
 430                 TAILQ_REMOVE(&qpair->outstanding_tr, tr, tailq);
 431                 TAILQ_INSERT_HEAD(&qpair->free_tr, tr, tailq);
 432
 433                 /*
 434                  * If the controller is in the middle of resetting, don't
 435                  *  try to submit queued requests here - let the reset logic
 436                  *  handle that instead.
 437                  */
 438                 if (!STAILQ_EMPTY(&qpair->queued_req) &&
 439                     !qpair->ctrlr->is_resetting) {
 440                         req = STAILQ_FIRST(&qpair->queued_req);
 441                         STAILQ_REMOVE_HEAD(&qpair->queued_req, stailq);
 442                         _nvme_qpair_submit_request(qpair, req);
 443                 }
 444         }
 445
 446         mtx_unlock(&qpair->lock);
 447 }
 448
 449 static void
 450 nvme_qpair_manual_complete_tracker(struct nvme_qpair *qpair,
 451     struct nvme_tracker *tr, uint32_t sct, uint32_t sc, uint32_t dnr,
 452     error_print_t print_on_error)
 453 {
 454         struct nvme_completion  cpl;
 455
 456         memset(&cpl, 0, sizeof(cpl));
 457         cpl.sqid = qpair->id;
 458         cpl.cid = tr->cid;
 459         cpl.status |= (sct & NVME_STATUS_SCT_MASK) << NVME_STATUS_SCT_SHIFT;
 460         cpl.status |= (sc & NVME_STATUS_SC_MASK) << NVME_STATUS_SC_SHIFT;
 461         cpl.status |= (dnr & NVME_STATUS_DNR_MASK) << NVME_STATUS_DNR_SHIFT;
 462         nvme_qpair_complete_tracker(qpair, tr, &cpl, print_on_error);
 463 }
 464
 465 void
 466 nvme_qpair_manual_complete_request(struct nvme_qpair *qpair,
 467     struct nvme_request *req, uint32_t sct, uint32_t sc)
 468 {
 469         struct nvme_completion  cpl;
 470         boolean_t               error;
 471
 472         memset(&cpl, 0, sizeof(cpl));
 473         cpl.sqid = qpair->id;
 474         cpl.status |= (sct & NVME_STATUS_SCT_MASK) << NVME_STATUS_SCT_SHIFT;
 475         cpl.status |= (sc & NVME_STATUS_SC_MASK) << NVME_STATUS_SC_SHIFT;
 476
 477         error = nvme_completion_is_error(&cpl);
 478
 479         if (error) {
 480                 nvme_qpair_print_command(qpair, &req->cmd);
 481                 nvme_qpair_print_completion(qpair, &cpl);
 482         }
 483
 484         if (req->cb_fn)
 485                 req->cb_fn(req->cb_arg, &cpl);
 486
 487         nvme_free_request(req);
 488 }
 489
 490 bool
 491 nvme_qpair_process_completions(struct nvme_qpair *qpair)
 492 {
 493         struct nvme_tracker     *tr;
 494         struct nvme_completion  cpl;
 495         int done = 0;
 496         bool in_panic = dumping || SCHEDULER_STOPPED();
 497
 498         qpair->num_intr_handler_calls++;
 499
 500         /*
 501          * qpair is not enabled, likely because a controller reset is is in
 502          * progress.  Ignore the interrupt - any I/O that was associated with
 503          * this interrupt will get retried when the reset is complete.
 504          */
 505         if (!qpair->is_enabled)
 506                 return (false);
 507
 508         /*
 509          * A panic can stop the CPU this routine is running on at any point.  If
 510          * we're called during a panic, complete the sq_head wrap protocol for
 511          * the case where we are interrupted just after the increment at 1
 512          * below, but before we can reset cq_head to zero at 2. Also cope with
 513          * the case where we do the zero at 2, but may or may not have done the
 514          * phase adjustment at step 3. The panic machinery flushes all pending
 515          * memory writes, so we can make these strong ordering assumptions
 516          * that would otherwise be unwise if we were racing in real time.
 517          */
 518         if (__predict_false(in_panic)) {
 519                 if (qpair->cq_head == qpair->num_entries) {
 520                         /*
 521                          * Here we know that we need to zero cq_head and then negate
 522                          * the phase, which hasn't been assigned if cq_head isn't
 523                          * zero due to the atomic_store_rel.
 524                          */
 525                         qpair->cq_head = 0;
 526                         qpair->phase = !qpair->phase;
 527                 } else if (qpair->cq_head == 0) {
 528                         /*
 529                          * In this case, we know that the assignment at 2
 530                          * happened below, but we don't know if it 3 happened or
 531                          * not. To do this, we look at the last completion
 532                          * entry and set the phase to the opposite phase
 533                          * that it has. This gets us back in sync
 534                          */
 535                         cpl = qpair->cpl[qpair->num_entries - 1];
 536                         nvme_completion_swapbytes(&cpl);
 537                         qpair->phase = !NVME_STATUS_GET_P(cpl.status);
 538                 }
 539         }
 540
 541         bus_dmamap_sync(qpair->dma_tag, qpair->queuemem_map,
 542             BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
 543         while (1) {
 544                 cpl = qpair->cpl[qpair->cq_head];
 545
 546                 /* Convert to host endian */
 547                 nvme_completion_swapbytes(&cpl);
 548
 549                 if (NVME_STATUS_GET_P(cpl.status) != qpair->phase)
 550                         break;
 551
 552                 tr = qpair->act_tr[cpl.cid];
 553
 554                 if (tr != NULL) {
 555                         nvme_qpair_complete_tracker(qpair, tr, &cpl, ERROR_PRINT_ALL);
 556                         qpair->sq_head = cpl.sqhd;
 557                         done++;
 558                 } else if (!in_panic) {
 559                         /*
 560                          * A missing tracker is normally an error.  However, a
 561                          * panic can stop the CPU this routine is running on
 562                          * after completing an I/O but before updating
 563                          * qpair->cq_head at 1 below.  Later, we re-enter this
 564                          * routine to poll I/O associated with the kernel
 565                          * dump. We find that the tr has been set to null before
 566                          * calling the completion routine.  If it hasn't
 567                          * completed (or it triggers a panic), then '1' below
 568                          * won't have updated cq_head. Rather than panic again,
 569                          * ignore this condition because it's not unexpected.
 570                          */
 571                         nvme_printf(qpair->ctrlr,
 572                             "cpl does not map to outstanding cmd\n");
 573                         /* nvme_dump_completion expects device endianess */
 574                         nvme_dump_completion(&qpair->cpl[qpair->cq_head]);
 575                         KASSERT(0, ("received completion for unknown cmd"));
 576                 }
 577
 578                 /*
 579                  * There's a number of races with the following (see above) when
 580                  * the system panics. We compensate for each one of them by
 581                  * using the atomic store to force strong ordering (at least when
 582                  * viewed in the aftermath of a panic).
 583                  */
 584                 if (++qpair->cq_head == qpair->num_entries) {           /* 1 */
 585                         atomic_store_rel_int(&qpair->cq_head, 0);       /* 2 */
 586                         qpair->phase = !qpair->phase;                   /* 3 */
 587                 }
 588
 589                 nvme_mmio_write_4(qpair->ctrlr, doorbell[qpair->id].cq_hdbl,
 590                     qpair->cq_head);
 591         }
 592         return (done != 0);
 593 }
 594
 595 static void
 596 nvme_qpair_msix_handler(void *arg)
 597 {
 598         struct nvme_qpair *qpair = arg;
 599
 600         nvme_qpair_process_completions(qpair);
 601 }
 602
 603 int
 604 nvme_qpair_construct(struct nvme_qpair *qpair, uint32_t id,
 605     uint16_t vector, uint32_t num_entries, uint32_t num_trackers,
 606     struct nvme_controller *ctrlr)
 607 {
 608         struct nvme_tracker     *tr;
 609         size_t                  cmdsz, cplsz, prpsz, allocsz, prpmemsz;
 610         uint64_t                queuemem_phys, prpmem_phys, list_phys;
 611         uint8_t                 *queuemem, *prpmem, *prp_list;
 612         int                     i, err;
 613
 614         qpair->id = id;
 615         qpair->vector = vector;
 616         qpair->num_entries = num_entries;
 617         qpair->num_trackers = num_trackers;
 618         qpair->ctrlr = ctrlr;
 619
 620         if (ctrlr->msix_enabled) {
 621
 622                 /*
 623                  * MSI-X vector resource IDs start at 1, so we add one to
 624                  *  the queue's vector to get the corresponding rid to use.
 625                  */
 626                 qpair->rid = vector + 1;
 627
 628                 qpair->res = bus_alloc_resource_any(ctrlr->dev, SYS_RES_IRQ,
 629                     &qpair->rid, RF_ACTIVE);
 630                 bus_setup_intr(ctrlr->dev, qpair->res,
 631                     INTR_TYPE_MISC | INTR_MPSAFE, NULL,
 632                     nvme_qpair_msix_handler, qpair, &qpair->tag);
 633                 if (id == 0) {
 634                         bus_describe_intr(ctrlr->dev, qpair->res, qpair->tag,
 635                             "admin");
 636                 } else {
 637                         bus_describe_intr(ctrlr->dev, qpair->res, qpair->tag,
 638                             "io%d", id - 1);
 639                 }
 640         }
 641
 642         mtx_init(&qpair->lock, "nvme qpair lock", NULL, MTX_DEF);
 643
 644         /* Note: NVMe PRP format is restricted to 4-byte alignment. */
 645         err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev),
 646             4, PAGE_SIZE, BUS_SPACE_MAXADDR,
 647             BUS_SPACE_MAXADDR, NULL, NULL, NVME_MAX_XFER_SIZE,
 648             (NVME_MAX_XFER_SIZE/PAGE_SIZE)+1, PAGE_SIZE, 0,
 649             NULL, NULL, &qpair->dma_tag_payload);
 650         if (err != 0) {
 651                 nvme_printf(ctrlr, "payload tag create failed %d\n", err);
 652                 goto out;
 653         }
 654
 655         /*
 656          * Each component must be page aligned, and individual PRP lists
 657          * cannot cross a page boundary.
 658          */
 659         cmdsz = qpair->num_entries * sizeof(struct nvme_command);
 660         cmdsz = roundup2(cmdsz, PAGE_SIZE);
 661         cplsz = qpair->num_entries * sizeof(struct nvme_completion);
 662         cplsz = roundup2(cplsz, PAGE_SIZE);
 663         prpsz = sizeof(uint64_t) * NVME_MAX_PRP_LIST_ENTRIES;;
 664         prpmemsz = qpair->num_trackers * prpsz;
 665         allocsz = cmdsz + cplsz + prpmemsz;
 666
 667         err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev),
 668             PAGE_SIZE, 0, BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
 669             allocsz, 1, allocsz, 0, NULL, NULL, &qpair->dma_tag);
 670         if (err != 0) {
 671                 nvme_printf(ctrlr, "tag create failed %d\n", err);
 672                 goto out;
 673         }
 674
 675         if (bus_dmamem_alloc(qpair->dma_tag, (void **)&queuemem,
 676             BUS_DMA_NOWAIT, &qpair->queuemem_map)) {
 677                 nvme_printf(ctrlr, "failed to alloc qpair memory\n");
 678                 goto out;
 679         }
 680
 681         if (bus_dmamap_load(qpair->dma_tag, qpair->queuemem_map,
 682             queuemem, allocsz, nvme_single_map, &queuemem_phys, 0) != 0) {
 683                 nvme_printf(ctrlr, "failed to load qpair memory\n");
 684                 goto out;
 685         }
 686
 687         qpair->num_cmds = 0;
 688         qpair->num_intr_handler_calls = 0;
 689         qpair->num_retries = 0;
 690         qpair->cmd = (struct nvme_command *)queuemem;
 691         qpair->cpl = (struct nvme_completion *)(queuemem + cmdsz);
 692         prpmem = (uint8_t *)(queuemem + cmdsz + cplsz);
 693         qpair->cmd_bus_addr = queuemem_phys;
 694         qpair->cpl_bus_addr = queuemem_phys + cmdsz;
 695         prpmem_phys = queuemem_phys + cmdsz + cplsz;
 696
 697         qpair->sq_tdbl_off = nvme_mmio_offsetof(doorbell[id].sq_tdbl);
 698         qpair->cq_hdbl_off = nvme_mmio_offsetof(doorbell[id].cq_hdbl);
 699
 700         TAILQ_INIT(&qpair->free_tr);
 701         TAILQ_INIT(&qpair->outstanding_tr);
 702         STAILQ_INIT(&qpair->queued_req);
 703
 704         list_phys = prpmem_phys;
 705         prp_list = prpmem;
 706         for (i = 0; i < qpair->num_trackers; i++) {
 707
 708                 if (list_phys + prpsz > prpmem_phys + prpmemsz) {
 709                         qpair->num_trackers = i;
 710                         break;
 711                 }
 712
 713                 /*
 714                  * Make sure that the PRP list for this tracker doesn't
 715                  * overflow to another page.
 716                  */
 717                 if (trunc_page(list_phys) !=
 718                     trunc_page(list_phys + prpsz - 1)) {
 719                         list_phys = roundup2(list_phys, PAGE_SIZE);
 720                         prp_list =
 721                             (uint8_t *)roundup2((uintptr_t)prp_list, PAGE_SIZE);
 722                 }
 723
 724                 tr = malloc(sizeof(*tr), M_NVME, M_ZERO | M_WAITOK);
 725                 bus_dmamap_create(qpair->dma_tag_payload, 0,
 726                     &tr->payload_dma_map);
 727                 callout_init(&tr->timer, 1);
 728                 tr->cid = i;
 729                 tr->qpair = qpair;
 730                 tr->prp = (uint64_t *)prp_list;
 731                 tr->prp_bus_addr = list_phys;
 732                 TAILQ_INSERT_HEAD(&qpair->free_tr, tr, tailq);
 733                 list_phys += prpsz;
 734                 prp_list += prpsz;
 735         }
 736
 737         if (qpair->num_trackers == 0) {
 738                 nvme_printf(ctrlr, "failed to allocate enough trackers\n");
 739                 goto out;
 740         }
 741
 742         qpair->act_tr = malloc(sizeof(struct nvme_tracker *) *
 743             qpair->num_entries, M_NVME, M_ZERO | M_WAITOK);
 744         return (0);
 745
 746 out:
 747         nvme_qpair_destroy(qpair);
 748         return (ENOMEM);
 749 }
 750
 751 static void
 752 nvme_qpair_destroy(struct nvme_qpair *qpair)
 753 {
 754         struct nvme_tracker     *tr;
 755
 756         if (qpair->tag)
 757                 bus_teardown_intr(qpair->ctrlr->dev, qpair->res, qpair->tag);
 758
 759         if (mtx_initialized(&qpair->lock))
 760                 mtx_destroy(&qpair->lock);
 761
 762         if (qpair->res)
 763                 bus_release_resource(qpair->ctrlr->dev, SYS_RES_IRQ,
 764                     rman_get_rid(qpair->res), qpair->res);
 765
 766         if (qpair->cmd != NULL) {
 767                 bus_dmamap_unload(qpair->dma_tag, qpair->queuemem_map);
 768                 bus_dmamem_free(qpair->dma_tag, qpair->cmd,
 769                     qpair->queuemem_map);
 770         }
 771
 772         if (qpair->act_tr)
 773                 free(qpair->act_tr, M_NVME);
 774
 775         while (!TAILQ_EMPTY(&qpair->free_tr)) {
 776                 tr = TAILQ_FIRST(&qpair->free_tr);
 777                 TAILQ_REMOVE(&qpair->free_tr, tr, tailq);
 778                 bus_dmamap_destroy(qpair->dma_tag_payload,
 779                     tr->payload_dma_map);
 780                 free(tr, M_NVME);
 781         }
 782
 783         if (qpair->dma_tag)
 784                 bus_dma_tag_destroy(qpair->dma_tag);
 785
 786         if (qpair->dma_tag_payload)
 787                 bus_dma_tag_destroy(qpair->dma_tag_payload);
 788 }
 789
 790 static void
 791 nvme_admin_qpair_abort_aers(struct nvme_qpair *qpair)
 792 {
 793         struct nvme_tracker     *tr;
 794
 795         tr = TAILQ_FIRST(&qpair->outstanding_tr);
 796         while (tr != NULL) {
 797                 if (tr->req->cmd.opc == NVME_OPC_ASYNC_EVENT_REQUEST) {
 798                         nvme_qpair_manual_complete_tracker(qpair, tr,
 799                             NVME_SCT_GENERIC, NVME_SC_ABORTED_SQ_DELETION, 0,
 800                             ERROR_PRINT_NONE);
 801                         tr = TAILQ_FIRST(&qpair->outstanding_tr);
 802                 } else {
 803                         tr = TAILQ_NEXT(tr, tailq);
 804                 }
 805         }
 806 }
 807
 808 void
 809 nvme_admin_qpair_destroy(struct nvme_qpair *qpair)
 810 {
 811
 812         nvme_admin_qpair_abort_aers(qpair);
 813         nvme_qpair_destroy(qpair);
 814 }
 815
 816 void
 817 nvme_io_qpair_destroy(struct nvme_qpair *qpair)
 818 {
 819
 820         nvme_qpair_destroy(qpair);
 821 }
 822
 823 static void
 824 nvme_abort_complete(void *arg, const struct nvme_completion *status)
 825 {
 826         struct nvme_tracker     *tr = arg;
 827
 828         /*
 829          * If cdw0 == 1, the controller was not able to abort the command
 830          *  we requested.  We still need to check the active tracker array,
 831          *  to cover race where I/O timed out at same time controller was
 832          *  completing the I/O.
 833          */
 834         if (status->cdw0 == 1 && tr->qpair->act_tr[tr->cid] != NULL) {
 835                 /*
 836                  * An I/O has timed out, and the controller was unable to
 837                  *  abort it for some reason.  Construct a fake completion
 838                  *  status, and then complete the I/O's tracker manually.
 839                  */
 840                 nvme_printf(tr->qpair->ctrlr,
 841                     "abort command failed, aborting command manually\n");
 842                 nvme_qpair_manual_complete_tracker(tr->qpair, tr,
 843                     NVME_SCT_GENERIC, NVME_SC_ABORTED_BY_REQUEST, 0, ERROR_PRINT_ALL);
 844         }
 845 }
 846
 847 static void
 848 nvme_timeout(void *arg)
 849 {
 850         struct nvme_tracker     *tr = arg;
 851         struct nvme_qpair       *qpair = tr->qpair;
 852         struct nvme_controller  *ctrlr = qpair->ctrlr;
 853         uint32_t                csts;
 854         uint8_t                 cfs;
 855
 856         /*
 857          * Read csts to get value of cfs - controller fatal status.
 858          * If no fatal status, try to call the completion routine, and
 859          * if completes transactions, report a missed interrupt and
 860          * return (this may need to be rate limited). Otherwise, if
 861          * aborts are enabled and the controller is not reporting
 862          * fatal status, abort the command. Otherwise, just reset the
 863          * controller and hope for the best.
 864          */
 865         csts = nvme_mmio_read_4(ctrlr, csts);
 866         cfs = (csts >> NVME_CSTS_REG_CFS_SHIFT) & NVME_CSTS_REG_CFS_MASK;
 867         if (cfs == 0 && nvme_qpair_process_completions(qpair)) {
 868                 nvme_printf(ctrlr, "Missing interrupt\n");
 869                 return;
 870         }
 871         if (ctrlr->enable_aborts && cfs == 0) {
 872                 nvme_printf(ctrlr, "Aborting command due to a timeout.\n");
 873                 nvme_ctrlr_cmd_abort(ctrlr, tr->cid, qpair->id,
 874                     nvme_abort_complete, tr);
 875         } else {
 876                 nvme_printf(ctrlr, "Resetting controller due to a timeout%s.\n",
 877                     cfs ? " and fatal error status" : "");
 878                 nvme_ctrlr_reset(ctrlr);
 879         }
 880 }
 881
 882 void
 883 nvme_qpair_submit_tracker(struct nvme_qpair *qpair, struct nvme_tracker *tr)
 884 {
 885         struct nvme_request     *req;
 886         struct nvme_controller  *ctrlr;
 887
 888         mtx_assert(&qpair->lock, MA_OWNED);
 889
 890         req = tr->req;
 891         req->cmd.cid = tr->cid;
 892         qpair->act_tr[tr->cid] = tr;
 893         ctrlr = qpair->ctrlr;
 894
 895         if (req->timeout)
 896                 callout_reset_curcpu(&tr->timer, ctrlr->timeout_period * hz,
 897                     nvme_timeout, tr);
 898
 899         /* Copy the command from the tracker to the submission queue. */
 900         memcpy(&qpair->cmd[qpair->sq_tail], &req->cmd, sizeof(req->cmd));
 901
 902         if (++qpair->sq_tail == qpair->num_entries)
 903                 qpair->sq_tail = 0;
 904
 905         bus_dmamap_sync(qpair->dma_tag, qpair->queuemem_map,
 906             BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
 907 #ifndef __powerpc__
 908         /*
 909          * powerpc's bus_dmamap_sync() already includes a heavyweight sync, but
 910          * no other archs do.
 911          */
 912         wmb();
 913 #endif
 914
 915         nvme_mmio_write_4(qpair->ctrlr, doorbell[qpair->id].sq_tdbl,
 916             qpair->sq_tail);
 917
 918         qpair->num_cmds++;
 919 }
 920
 921 static void
 922 nvme_payload_map(void *arg, bus_dma_segment_t *seg, int nseg, int error)
 923 {
 924         struct nvme_tracker     *tr = arg;
 925         uint32_t                cur_nseg;
 926
 927         /*
 928          * If the mapping operation failed, return immediately.  The caller
 929          *  is responsible for detecting the error status and failing the
 930          *  tracker manually.
 931          */
 932         if (error != 0) {
 933                 nvme_printf(tr->qpair->ctrlr,
 934                     "nvme_payload_map err %d\n", error);
 935                 return;
 936         }
 937
 938         /*
 939          * Note that we specified PAGE_SIZE for alignment and max
 940          *  segment size when creating the bus dma tags.  So here
 941          *  we can safely just transfer each segment to its
 942          *  associated PRP entry.
 943          */
 944         tr->req->cmd.prp1 = htole64(seg[0].ds_addr);
 945
 946         if (nseg == 2) {
 947                 tr->req->cmd.prp2 = htole64(seg[1].ds_addr);
 948         } else if (nseg > 2) {
 949                 cur_nseg = 1;
 950                 tr->req->cmd.prp2 = htole64((uint64_t)tr->prp_bus_addr);
 951                 while (cur_nseg < nseg) {
 952                         tr->prp[cur_nseg-1] =
 953                             htole64((uint64_t)seg[cur_nseg].ds_addr);
 954                         cur_nseg++;
 955                 }
 956         } else {
 957                 /*
 958                  * prp2 should not be used by the controller
 959                  *  since there is only one segment, but set
 960                  *  to 0 just to be safe.
 961                  */
 962                 tr->req->cmd.prp2 = 0;
 963         }
 964
 965         bus_dmamap_sync(tr->qpair->dma_tag_payload, tr->payload_dma_map,
 966             BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
 967         nvme_qpair_submit_tracker(tr->qpair, tr);
 968 }
 969
 970 static void
 971 _nvme_qpair_submit_request(struct nvme_qpair *qpair, struct nvme_request *req)
 972 {
 973         struct nvme_tracker     *tr;
 974         int                     err = 0;
 975
 976         mtx_assert(&qpair->lock, MA_OWNED);
 977
 978         tr = TAILQ_FIRST(&qpair->free_tr);
 979         req->qpair = qpair;
 980
 981         if (tr == NULL || !qpair->is_enabled) {
 982                 /*
 983                  * No tracker is available, or the qpair is disabled due to
 984                  *  an in-progress controller-level reset or controller
 985                  *  failure.
 986                  */
 987
 988                 if (qpair->ctrlr->is_failed) {
 989                         /*
 990                          * The controller has failed.  Post the request to a
 991                          *  task where it will be aborted, so that we do not
 992                          *  invoke the request's callback in the context
 993                          *  of the submission.
 994                          */
 995                         nvme_ctrlr_post_failed_request(qpair->ctrlr, req);
 996                 } else {
 997                         /*
 998                          * Put the request on the qpair's request queue to be
 999                          *  processed when a tracker frees up via a command
1000                          *  completion or when the controller reset is
1001                          *  completed.
1002                          */
1003                         STAILQ_INSERT_TAIL(&qpair->queued_req, req, stailq);
1004                 }
1005                 return;
1006         }
1007
1008         TAILQ_REMOVE(&qpair->free_tr, tr, tailq);
1009         TAILQ_INSERT_TAIL(&qpair->outstanding_tr, tr, tailq);
1010         tr->req = req;
1011
1012         switch (req->type) {
1013         case NVME_REQUEST_VADDR:
1014                 KASSERT(req->payload_size <= qpair->ctrlr->max_xfer_size,
1015                     ("payload_size (%d) exceeds max_xfer_size (%d)\n",
1016                     req->payload_size, qpair->ctrlr->max_xfer_size));
1017                 err = bus_dmamap_load(tr->qpair->dma_tag_payload,
1018                     tr->payload_dma_map, req->u.payload, req->payload_size,
1019                     nvme_payload_map, tr, 0);
1020                 if (err != 0)
1021                         nvme_printf(qpair->ctrlr,
1022                             "bus_dmamap_load returned 0x%x!\n", err);
1023                 break;
1024         case NVME_REQUEST_NULL:
1025                 nvme_qpair_submit_tracker(tr->qpair, tr);
1026                 break;
1027         case NVME_REQUEST_BIO:
1028                 KASSERT(req->u.bio->bio_bcount <= qpair->ctrlr->max_xfer_size,
1029                     ("bio->bio_bcount (%jd) exceeds max_xfer_size (%d)\n",
1030                     (intmax_t)req->u.bio->bio_bcount,
1031                     qpair->ctrlr->max_xfer_size));
1032                 err = bus_dmamap_load_bio(tr->qpair->dma_tag_payload,
1033                     tr->payload_dma_map, req->u.bio, nvme_payload_map, tr, 0);
1034                 if (err != 0)
1035                         nvme_printf(qpair->ctrlr,
1036                             "bus_dmamap_load_bio returned 0x%x!\n", err);
1037                 break;
1038         case NVME_REQUEST_CCB:
1039                 err = bus_dmamap_load_ccb(tr->qpair->dma_tag_payload,
1040                     tr->payload_dma_map, req->u.payload,
1041                     nvme_payload_map, tr, 0);
1042                 if (err != 0)
1043                         nvme_printf(qpair->ctrlr,
1044                             "bus_dmamap_load_ccb returned 0x%x!\n", err);
1045                 break;
1046         default:
1047                 panic("unknown nvme request type 0x%x\n", req->type);
1048                 break;
1049         }
1050
1051         if (err != 0) {
1052                 /*
1053                  * The dmamap operation failed, so we manually fail the
1054                  *  tracker here with DATA_TRANSFER_ERROR status.
1055                  *
1056                  * nvme_qpair_manual_complete_tracker must not be called
1057                  *  with the qpair lock held.
1058                  */
1059                 mtx_unlock(&qpair->lock);
1060                 nvme_qpair_manual_complete_tracker(qpair, tr, NVME_SCT_GENERIC,
1061                     NVME_SC_DATA_TRANSFER_ERROR, DO_NOT_RETRY, ERROR_PRINT_ALL);
1062                 mtx_lock(&qpair->lock);
1063         }
1064 }
1065
1066 void
1067 nvme_qpair_submit_request(struct nvme_qpair *qpair, struct nvme_request *req)
1068 {
1069
1070         mtx_lock(&qpair->lock);
1071         _nvme_qpair_submit_request(qpair, req);
1072         mtx_unlock(&qpair->lock);
1073 }
1074
1075 static void
1076 nvme_qpair_enable(struct nvme_qpair *qpair)
1077 {
1078
1079         qpair->is_enabled = TRUE;
1080 }
1081
1082 void
1083 nvme_qpair_reset(struct nvme_qpair *qpair)
1084 {
1085
1086         qpair->sq_head = qpair->sq_tail = qpair->cq_head = 0;
1087
1088         /*
1089          * First time through the completion queue, HW will set phase
1090          *  bit on completions to 1.  So set this to 1 here, indicating
1091          *  we're looking for a 1 to know which entries have completed.
1092          *  we'll toggle the bit each time when the completion queue
1093          *  rolls over.
1094          */
1095         qpair->phase = 1;
1096
1097         memset(qpair->cmd, 0,
1098             qpair->num_entries * sizeof(struct nvme_command));
1099         memset(qpair->cpl, 0,
1100             qpair->num_entries * sizeof(struct nvme_completion));
1101 }
1102
1103 void
1104 nvme_admin_qpair_enable(struct nvme_qpair *qpair)
1105 {
1106         struct nvme_tracker             *tr;
1107         struct nvme_tracker             *tr_temp;
1108
1109         /*
1110          * Manually abort each outstanding admin command.  Do not retry
1111          *  admin commands found here, since they will be left over from
1112          *  a controller reset and its likely the context in which the
1113          *  command was issued no longer applies.
1114          */
1115         TAILQ_FOREACH_SAFE(tr, &qpair->outstanding_tr, tailq, tr_temp) {
1116                 nvme_printf(qpair->ctrlr,
1117                     "aborting outstanding admin command\n");
1118                 nvme_qpair_manual_complete_tracker(qpair, tr, NVME_SCT_GENERIC,
1119                     NVME_SC_ABORTED_BY_REQUEST, DO_NOT_RETRY, ERROR_PRINT_ALL);
1120         }
1121
1122         nvme_qpair_enable(qpair);
1123 }
1124
1125 void
1126 nvme_io_qpair_enable(struct nvme_qpair *qpair)
1127 {
1128         STAILQ_HEAD(, nvme_request)     temp;
1129         struct nvme_tracker             *tr;
1130         struct nvme_tracker             *tr_temp;
1131         struct nvme_request             *req;
1132
1133         /*
1134          * Manually abort each outstanding I/O.  This normally results in a
1135          *  retry, unless the retry count on the associated request has
1136          *  reached its limit.
1137          */
1138         TAILQ_FOREACH_SAFE(tr, &qpair->outstanding_tr, tailq, tr_temp) {
1139                 nvme_printf(qpair->ctrlr, "aborting outstanding i/o\n");
1140                 nvme_qpair_manual_complete_tracker(qpair, tr, NVME_SCT_GENERIC,
1141                     NVME_SC_ABORTED_BY_REQUEST, 0, ERROR_PRINT_NO_RETRY);
1142         }
1143
1144         mtx_lock(&qpair->lock);
1145
1146         nvme_qpair_enable(qpair);
1147
1148         STAILQ_INIT(&temp);
1149         STAILQ_SWAP(&qpair->queued_req, &temp, nvme_request);
1150
1151         while (!STAILQ_EMPTY(&temp)) {
1152                 req = STAILQ_FIRST(&temp);
1153                 STAILQ_REMOVE_HEAD(&temp, stailq);
1154                 nvme_printf(qpair->ctrlr, "resubmitting queued i/o\n");
1155                 nvme_qpair_print_command(qpair, &req->cmd);
1156                 _nvme_qpair_submit_request(qpair, req);
1157         }
1158
1159         mtx_unlock(&qpair->lock);
1160 }
1161
1162 static void
1163 nvme_qpair_disable(struct nvme_qpair *qpair)
1164 {
1165         struct nvme_tracker *tr;
1166
1167         qpair->is_enabled = FALSE;
1168         mtx_lock(&qpair->lock);
1169         TAILQ_FOREACH(tr, &qpair->outstanding_tr, tailq)
1170                 callout_stop(&tr->timer);
1171         mtx_unlock(&qpair->lock);
1172 }
1173
1174 void
1175 nvme_admin_qpair_disable(struct nvme_qpair *qpair)
1176 {
1177
1178         nvme_qpair_disable(qpair);
1179         nvme_admin_qpair_abort_aers(qpair);
1180 }
1181
1182 void
1183 nvme_io_qpair_disable(struct nvme_qpair *qpair)
1184 {
1185
1186         nvme_qpair_disable(qpair);
1187 }
1188
1189 void
1190 nvme_qpair_fail(struct nvme_qpair *qpair)
1191 {
1192         struct nvme_tracker             *tr;
1193         struct nvme_request             *req;
1194
1195         if (!mtx_initialized(&qpair->lock))
1196                 return;
1197
1198         mtx_lock(&qpair->lock);
1199
1200         while (!STAILQ_EMPTY(&qpair->queued_req)) {
1201                 req = STAILQ_FIRST(&qpair->queued_req);
1202                 STAILQ_REMOVE_HEAD(&qpair->queued_req, stailq);
1203                 nvme_printf(qpair->ctrlr, "failing queued i/o\n");
1204                 mtx_unlock(&qpair->lock);
1205                 nvme_qpair_manual_complete_request(qpair, req, NVME_SCT_GENERIC,
1206                     NVME_SC_ABORTED_BY_REQUEST);
1207                 mtx_lock(&qpair->lock);
1208         }
1209
1210         /* Manually abort each outstanding I/O. */
1211         while (!TAILQ_EMPTY(&qpair->outstanding_tr)) {
1212                 tr = TAILQ_FIRST(&qpair->outstanding_tr);
1213                 /*
1214                  * Do not remove the tracker.  The abort_tracker path will
1215                  *  do that for us.
1216                  */
1217                 nvme_printf(qpair->ctrlr, "failing outstanding i/o\n");
1218                 mtx_unlock(&qpair->lock);
1219                 nvme_qpair_manual_complete_tracker(qpair, tr, NVME_SCT_GENERIC,
1220                     NVME_SC_ABORTED_BY_REQUEST, DO_NOT_RETRY, ERROR_PRINT_ALL);
1221                 mtx_lock(&qpair->lock);
1222         }
1223
1224         mtx_unlock(&qpair->lock);
1225 }
1226