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, retriable;
 391
 392         req = tr->req;
 393         error = nvme_completion_is_error(cpl);
 394         retriable = nvme_completion_is_retry(cpl);
 395         retry = error && retriable && req->retries < nvme_retry_count;
 396         if (retry)
 397                 qpair->num_retries++;
 398         if (error && req->retries >= nvme_retry_count && retriable)
 399                 qpair->num_failures++;
 400
 401         if (error && (print_on_error == ERROR_PRINT_ALL ||
 402                 (!retry && print_on_error == ERROR_PRINT_NO_RETRY))) {
 403                 nvme_qpair_print_command(qpair, &req->cmd);
 404                 nvme_qpair_print_completion(qpair, cpl);
 405         }
 406
 407         qpair->act_tr[cpl->cid] = NULL;
 408
 409         KASSERT(cpl->cid == req->cmd.cid, ("cpl cid does not match cmd cid\n"));
 410
 411         if (req->cb_fn && !retry)
 412                 req->cb_fn(req->cb_arg, cpl);
 413
 414         mtx_lock(&qpair->lock);
 415         callout_stop(&tr->timer);
 416
 417         if (retry) {
 418                 req->retries++;
 419                 nvme_qpair_submit_tracker(qpair, tr);
 420         } else {
 421                 if (req->type != NVME_REQUEST_NULL) {
 422                         bus_dmamap_sync(qpair->dma_tag_payload,
 423                             tr->payload_dma_map,
 424                             BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
 425                         bus_dmamap_unload(qpair->dma_tag_payload,
 426                             tr->payload_dma_map);
 427                 }
 428
 429                 nvme_free_request(req);
 430                 tr->req = NULL;
 431
 432                 TAILQ_REMOVE(&qpair->outstanding_tr, tr, tailq);
 433                 TAILQ_INSERT_HEAD(&qpair->free_tr, tr, tailq);
 434
 435                 /*
 436                  * If the controller is in the middle of resetting, don't
 437                  *  try to submit queued requests here - let the reset logic
 438                  *  handle that instead.
 439                  */
 440                 if (!STAILQ_EMPTY(&qpair->queued_req) &&
 441                     !qpair->ctrlr->is_resetting) {
 442                         req = STAILQ_FIRST(&qpair->queued_req);
 443                         STAILQ_REMOVE_HEAD(&qpair->queued_req, stailq);
 444                         _nvme_qpair_submit_request(qpair, req);
 445                 }
 446         }
 447
 448         mtx_unlock(&qpair->lock);
 449 }
 450
 451 static void
 452 nvme_qpair_manual_complete_tracker(struct nvme_qpair *qpair,
 453     struct nvme_tracker *tr, uint32_t sct, uint32_t sc, uint32_t dnr,
 454     error_print_t print_on_error)
 455 {
 456         struct nvme_completion  cpl;
 457
 458         memset(&cpl, 0, sizeof(cpl));
 459         cpl.sqid = qpair->id;
 460         cpl.cid = tr->cid;
 461         cpl.status |= (sct & NVME_STATUS_SCT_MASK) << NVME_STATUS_SCT_SHIFT;
 462         cpl.status |= (sc & NVME_STATUS_SC_MASK) << NVME_STATUS_SC_SHIFT;
 463         cpl.status |= (dnr & NVME_STATUS_DNR_MASK) << NVME_STATUS_DNR_SHIFT;
 464         nvme_qpair_complete_tracker(qpair, tr, &cpl, print_on_error);
 465 }
 466
 467 void
 468 nvme_qpair_manual_complete_request(struct nvme_qpair *qpair,
 469     struct nvme_request *req, uint32_t sct, uint32_t sc)
 470 {
 471         struct nvme_completion  cpl;
 472         boolean_t               error;
 473
 474         memset(&cpl, 0, sizeof(cpl));
 475         cpl.sqid = qpair->id;
 476         cpl.status |= (sct & NVME_STATUS_SCT_MASK) << NVME_STATUS_SCT_SHIFT;
 477         cpl.status |= (sc & NVME_STATUS_SC_MASK) << NVME_STATUS_SC_SHIFT;
 478
 479         error = nvme_completion_is_error(&cpl);
 480
 481         if (error) {
 482                 nvme_qpair_print_command(qpair, &req->cmd);
 483                 nvme_qpair_print_completion(qpair, &cpl);
 484         }
 485
 486         if (req->cb_fn)
 487                 req->cb_fn(req->cb_arg, &cpl);
 488
 489         nvme_free_request(req);
 490 }
 491
 492 bool
 493 nvme_qpair_process_completions(struct nvme_qpair *qpair)
 494 {
 495         struct nvme_tracker     *tr;
 496         struct nvme_completion  cpl;
 497         int done = 0;
 498         bool in_panic = dumping || SCHEDULER_STOPPED();
 499
 500         qpair->num_intr_handler_calls++;
 501
 502         /*
 503          * qpair is not enabled, likely because a controller reset is is in
 504          * progress.  Ignore the interrupt - any I/O that was associated with
 505          * this interrupt will get retried when the reset is complete.
 506          */
 507         if (!qpair->is_enabled)
 508                 return (false);
 509
 510         /*
 511          * A panic can stop the CPU this routine is running on at any point.  If
 512          * we're called during a panic, complete the sq_head wrap protocol for
 513          * the case where we are interrupted just after the increment at 1
 514          * below, but before we can reset cq_head to zero at 2. Also cope with
 515          * the case where we do the zero at 2, but may or may not have done the
 516          * phase adjustment at step 3. The panic machinery flushes all pending
 517          * memory writes, so we can make these strong ordering assumptions
 518          * that would otherwise be unwise if we were racing in real time.
 519          */
 520         if (__predict_false(in_panic)) {
 521                 if (qpair->cq_head == qpair->num_entries) {
 522                         /*
 523                          * Here we know that we need to zero cq_head and then negate
 524                          * the phase, which hasn't been assigned if cq_head isn't
 525                          * zero due to the atomic_store_rel.
 526                          */
 527                         qpair->cq_head = 0;
 528                         qpair->phase = !qpair->phase;
 529                 } else if (qpair->cq_head == 0) {
 530                         /*
 531                          * In this case, we know that the assignment at 2
 532                          * happened below, but we don't know if it 3 happened or
 533                          * not. To do this, we look at the last completion
 534                          * entry and set the phase to the opposite phase
 535                          * that it has. This gets us back in sync
 536                          */
 537                         cpl = qpair->cpl[qpair->num_entries - 1];
 538                         nvme_completion_swapbytes(&cpl);
 539                         qpair->phase = !NVME_STATUS_GET_P(cpl.status);
 540                 }
 541         }
 542
 543         bus_dmamap_sync(qpair->dma_tag, qpair->queuemem_map,
 544             BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
 545         while (1) {
 546                 cpl = qpair->cpl[qpair->cq_head];
 547
 548                 /* Convert to host endian */
 549                 nvme_completion_swapbytes(&cpl);
 550
 551                 if (NVME_STATUS_GET_P(cpl.status) != qpair->phase)
 552                         break;
 553
 554                 tr = qpair->act_tr[cpl.cid];
 555
 556                 if (tr != NULL) {
 557                         nvme_qpair_complete_tracker(qpair, tr, &cpl, ERROR_PRINT_ALL);
 558                         qpair->sq_head = cpl.sqhd;
 559                         done++;
 560                 } else if (!in_panic) {
 561                         /*
 562                          * A missing tracker is normally an error.  However, a
 563                          * panic can stop the CPU this routine is running on
 564                          * after completing an I/O but before updating
 565                          * qpair->cq_head at 1 below.  Later, we re-enter this
 566                          * routine to poll I/O associated with the kernel
 567                          * dump. We find that the tr has been set to null before
 568                          * calling the completion routine.  If it hasn't
 569                          * completed (or it triggers a panic), then '1' below
 570                          * won't have updated cq_head. Rather than panic again,
 571                          * ignore this condition because it's not unexpected.
 572                          */
 573                         nvme_printf(qpair->ctrlr,
 574                             "cpl does not map to outstanding cmd\n");
 575                         /* nvme_dump_completion expects device endianess */
 576                         nvme_dump_completion(&qpair->cpl[qpair->cq_head]);
 577                         KASSERT(0, ("received completion for unknown cmd"));
 578                 }
 579
 580                 /*
 581                  * There's a number of races with the following (see above) when
 582                  * the system panics. We compensate for each one of them by
 583                  * using the atomic store to force strong ordering (at least when
 584                  * viewed in the aftermath of a panic).
 585                  */
 586                 if (++qpair->cq_head == qpair->num_entries) {           /* 1 */
 587                         atomic_store_rel_int(&qpair->cq_head, 0);       /* 2 */
 588                         qpair->phase = !qpair->phase;                   /* 3 */
 589                 }
 590
 591                 nvme_mmio_write_4(qpair->ctrlr, doorbell[qpair->id].cq_hdbl,
 592                     qpair->cq_head);
 593         }
 594         return (done != 0);
 595 }
 596
 597 static void
 598 nvme_qpair_msix_handler(void *arg)
 599 {
 600         struct nvme_qpair *qpair = arg;
 601
 602         nvme_qpair_process_completions(qpair);
 603 }
 604
 605 int
 606 nvme_qpair_construct(struct nvme_qpair *qpair, uint32_t id,
 607     uint16_t vector, uint32_t num_entries, uint32_t num_trackers,
 608     struct nvme_controller *ctrlr)
 609 {
 610         struct nvme_tracker     *tr;
 611         size_t                  cmdsz, cplsz, prpsz, allocsz, prpmemsz;
 612         uint64_t                queuemem_phys, prpmem_phys, list_phys;
 613         uint8_t                 *queuemem, *prpmem, *prp_list;
 614         int                     i, err;
 615
 616         qpair->id = id;
 617         qpair->vector = vector;
 618         qpair->num_entries = num_entries;
 619         qpair->num_trackers = num_trackers;
 620         qpair->ctrlr = ctrlr;
 621
 622         if (ctrlr->msix_enabled) {
 623
 624                 /*
 625                  * MSI-X vector resource IDs start at 1, so we add one to
 626                  *  the queue's vector to get the corresponding rid to use.
 627                  */
 628                 qpair->rid = vector + 1;
 629
 630                 qpair->res = bus_alloc_resource_any(ctrlr->dev, SYS_RES_IRQ,
 631                     &qpair->rid, RF_ACTIVE);
 632                 bus_setup_intr(ctrlr->dev, qpair->res,
 633                     INTR_TYPE_MISC | INTR_MPSAFE, NULL,
 634                     nvme_qpair_msix_handler, qpair, &qpair->tag);
 635                 if (id == 0) {
 636                         bus_describe_intr(ctrlr->dev, qpair->res, qpair->tag,
 637                             "admin");
 638                 } else {
 639                         bus_describe_intr(ctrlr->dev, qpair->res, qpair->tag,
 640                             "io%d", id - 1);
 641                 }
 642         }
 643
 644         mtx_init(&qpair->lock, "nvme qpair lock", NULL, MTX_DEF);
 645
 646         /* Note: NVMe PRP format is restricted to 4-byte alignment. */
 647         err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev),
 648             4, PAGE_SIZE, BUS_SPACE_MAXADDR,
 649             BUS_SPACE_MAXADDR, NULL, NULL, NVME_MAX_XFER_SIZE,
 650             (NVME_MAX_XFER_SIZE/PAGE_SIZE)+1, PAGE_SIZE, 0,
 651             NULL, NULL, &qpair->dma_tag_payload);
 652         if (err != 0) {
 653                 nvme_printf(ctrlr, "payload tag create failed %d\n", err);
 654                 goto out;
 655         }
 656
 657         /*
 658          * Each component must be page aligned, and individual PRP lists
 659          * cannot cross a page boundary.
 660          */
 661         cmdsz = qpair->num_entries * sizeof(struct nvme_command);
 662         cmdsz = roundup2(cmdsz, PAGE_SIZE);
 663         cplsz = qpair->num_entries * sizeof(struct nvme_completion);
 664         cplsz = roundup2(cplsz, PAGE_SIZE);
 665         prpsz = sizeof(uint64_t) * NVME_MAX_PRP_LIST_ENTRIES;;
 666         prpmemsz = qpair->num_trackers * prpsz;
 667         allocsz = cmdsz + cplsz + prpmemsz;
 668
 669         err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev),
 670             PAGE_SIZE, 0, BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
 671             allocsz, 1, allocsz, 0, NULL, NULL, &qpair->dma_tag);
 672         if (err != 0) {
 673                 nvme_printf(ctrlr, "tag create failed %d\n", err);
 674                 goto out;
 675         }
 676
 677         if (bus_dmamem_alloc(qpair->dma_tag, (void **)&queuemem,
 678             BUS_DMA_NOWAIT, &qpair->queuemem_map)) {
 679                 nvme_printf(ctrlr, "failed to alloc qpair memory\n");
 680                 goto out;
 681         }
 682
 683         if (bus_dmamap_load(qpair->dma_tag, qpair->queuemem_map,
 684             queuemem, allocsz, nvme_single_map, &queuemem_phys, 0) != 0) {
 685                 nvme_printf(ctrlr, "failed to load qpair memory\n");
 686                 goto out;
 687         }
 688
 689         qpair->num_cmds = 0;
 690         qpair->num_intr_handler_calls = 0;
 691         qpair->num_retries = 0;
 692         qpair->num_failures = 0;
 693         qpair->cmd = (struct nvme_command *)queuemem;
 694         qpair->cpl = (struct nvme_completion *)(queuemem + cmdsz);
 695         prpmem = (uint8_t *)(queuemem + cmdsz + cplsz);
 696         qpair->cmd_bus_addr = queuemem_phys;
 697         qpair->cpl_bus_addr = queuemem_phys + cmdsz;
 698         prpmem_phys = queuemem_phys + cmdsz + cplsz;
 699
 700         qpair->sq_tdbl_off = nvme_mmio_offsetof(doorbell[id].sq_tdbl);
 701         qpair->cq_hdbl_off = nvme_mmio_offsetof(doorbell[id].cq_hdbl);
 702
 703         TAILQ_INIT(&qpair->free_tr);
 704         TAILQ_INIT(&qpair->outstanding_tr);
 705         STAILQ_INIT(&qpair->queued_req);
 706
 707         list_phys = prpmem_phys;
 708         prp_list = prpmem;
 709         for (i = 0; i < qpair->num_trackers; i++) {
 710
 711                 if (list_phys + prpsz > prpmem_phys + prpmemsz) {
 712                         qpair->num_trackers = i;
 713                         break;
 714                 }
 715
 716                 /*
 717                  * Make sure that the PRP list for this tracker doesn't
 718                  * overflow to another page.
 719                  */
 720                 if (trunc_page(list_phys) !=
 721                     trunc_page(list_phys + prpsz - 1)) {
 722                         list_phys = roundup2(list_phys, PAGE_SIZE);
 723                         prp_list =
 724                             (uint8_t *)roundup2((uintptr_t)prp_list, PAGE_SIZE);
 725                 }
 726
 727                 tr = malloc(sizeof(*tr), M_NVME, M_ZERO | M_WAITOK);
 728                 bus_dmamap_create(qpair->dma_tag_payload, 0,
 729                     &tr->payload_dma_map);
 730                 callout_init(&tr->timer, 1);
 731                 tr->cid = i;
 732                 tr->qpair = qpair;
 733                 tr->prp = (uint64_t *)prp_list;
 734                 tr->prp_bus_addr = list_phys;
 735                 TAILQ_INSERT_HEAD(&qpair->free_tr, tr, tailq);
 736                 list_phys += prpsz;
 737                 prp_list += prpsz;
 738         }
 739
 740         if (qpair->num_trackers == 0) {
 741                 nvme_printf(ctrlr, "failed to allocate enough trackers\n");
 742                 goto out;
 743         }
 744
 745         qpair->act_tr = malloc(sizeof(struct nvme_tracker *) *
 746             qpair->num_entries, M_NVME, M_ZERO | M_WAITOK);
 747         return (0);
 748
 749 out:
 750         nvme_qpair_destroy(qpair);
 751         return (ENOMEM);
 752 }
 753
 754 static void
 755 nvme_qpair_destroy(struct nvme_qpair *qpair)
 756 {
 757         struct nvme_tracker     *tr;
 758
 759         if (qpair->tag)
 760                 bus_teardown_intr(qpair->ctrlr->dev, qpair->res, qpair->tag);
 761
 762         if (mtx_initialized(&qpair->lock))
 763                 mtx_destroy(&qpair->lock);
 764
 765         if (qpair->res)
 766                 bus_release_resource(qpair->ctrlr->dev, SYS_RES_IRQ,
 767                     rman_get_rid(qpair->res), qpair->res);
 768
 769         if (qpair->cmd != NULL) {
 770                 bus_dmamap_unload(qpair->dma_tag, qpair->queuemem_map);
 771                 bus_dmamem_free(qpair->dma_tag, qpair->cmd,
 772                     qpair->queuemem_map);
 773         }
 774
 775         if (qpair->act_tr)
 776                 free(qpair->act_tr, M_NVME);
 777
 778         while (!TAILQ_EMPTY(&qpair->free_tr)) {
 779                 tr = TAILQ_FIRST(&qpair->free_tr);
 780                 TAILQ_REMOVE(&qpair->free_tr, tr, tailq);
 781                 bus_dmamap_destroy(qpair->dma_tag_payload,
 782                     tr->payload_dma_map);
 783                 free(tr, M_NVME);
 784         }
 785
 786         if (qpair->dma_tag)
 787                 bus_dma_tag_destroy(qpair->dma_tag);
 788
 789         if (qpair->dma_tag_payload)
 790                 bus_dma_tag_destroy(qpair->dma_tag_payload);
 791 }
 792
 793 static void
 794 nvme_admin_qpair_abort_aers(struct nvme_qpair *qpair)
 795 {
 796         struct nvme_tracker     *tr;
 797
 798         tr = TAILQ_FIRST(&qpair->outstanding_tr);
 799         while (tr != NULL) {
 800                 if (tr->req->cmd.opc == NVME_OPC_ASYNC_EVENT_REQUEST) {
 801                         nvme_qpair_manual_complete_tracker(qpair, tr,
 802                             NVME_SCT_GENERIC, NVME_SC_ABORTED_SQ_DELETION, 0,
 803                             ERROR_PRINT_NONE);
 804                         tr = TAILQ_FIRST(&qpair->outstanding_tr);
 805                 } else {
 806                         tr = TAILQ_NEXT(tr, tailq);
 807                 }
 808         }
 809 }
 810
 811 void
 812 nvme_admin_qpair_destroy(struct nvme_qpair *qpair)
 813 {
 814
 815         nvme_admin_qpair_abort_aers(qpair);
 816         nvme_qpair_destroy(qpair);
 817 }
 818
 819 void
 820 nvme_io_qpair_destroy(struct nvme_qpair *qpair)
 821 {
 822
 823         nvme_qpair_destroy(qpair);
 824 }
 825
 826 static void
 827 nvme_abort_complete(void *arg, const struct nvme_completion *status)
 828 {
 829         struct nvme_tracker     *tr = arg;
 830
 831         /*
 832          * If cdw0 == 1, the controller was not able to abort the command
 833          *  we requested.  We still need to check the active tracker array,
 834          *  to cover race where I/O timed out at same time controller was
 835          *  completing the I/O.
 836          */
 837         if (status->cdw0 == 1 && tr->qpair->act_tr[tr->cid] != NULL) {
 838                 /*
 839                  * An I/O has timed out, and the controller was unable to
 840                  *  abort it for some reason.  Construct a fake completion
 841                  *  status, and then complete the I/O's tracker manually.
 842                  */
 843                 nvme_printf(tr->qpair->ctrlr,
 844                     "abort command failed, aborting command manually\n");
 845                 nvme_qpair_manual_complete_tracker(tr->qpair, tr,
 846                     NVME_SCT_GENERIC, NVME_SC_ABORTED_BY_REQUEST, 0, ERROR_PRINT_ALL);
 847         }
 848 }
 849
 850 static void
 851 nvme_timeout(void *arg)
 852 {
 853         struct nvme_tracker     *tr = arg;
 854         struct nvme_qpair       *qpair = tr->qpair;
 855         struct nvme_controller  *ctrlr = qpair->ctrlr;
 856         uint32_t                csts;
 857         uint8_t                 cfs;
 858
 859         /*
 860          * Read csts to get value of cfs - controller fatal status.
 861          * If no fatal status, try to call the completion routine, and
 862          * if completes transactions, report a missed interrupt and
 863          * return (this may need to be rate limited). Otherwise, if
 864          * aborts are enabled and the controller is not reporting
 865          * fatal status, abort the command. Otherwise, just reset the
 866          * controller and hope for the best.
 867          */
 868         csts = nvme_mmio_read_4(ctrlr, csts);
 869         cfs = (csts >> NVME_CSTS_REG_CFS_SHIFT) & NVME_CSTS_REG_CFS_MASK;
 870         if (cfs == 0 && nvme_qpair_process_completions(qpair)) {
 871                 nvme_printf(ctrlr, "Missing interrupt\n");
 872                 return;
 873         }
 874         if (ctrlr->enable_aborts && cfs == 0) {
 875                 nvme_printf(ctrlr, "Aborting command due to a timeout.\n");
 876                 nvme_ctrlr_cmd_abort(ctrlr, tr->cid, qpair->id,
 877                     nvme_abort_complete, tr);
 878         } else {
 879                 nvme_printf(ctrlr, "Resetting controller due to a timeout%s.\n",
 880                     cfs ? " and fatal error status" : "");
 881                 nvme_ctrlr_reset(ctrlr);
 882         }
 883 }
 884
 885 void
 886 nvme_qpair_submit_tracker(struct nvme_qpair *qpair, struct nvme_tracker *tr)
 887 {
 888         struct nvme_request     *req;
 889         struct nvme_controller  *ctrlr;
 890
 891         mtx_assert(&qpair->lock, MA_OWNED);
 892
 893         req = tr->req;
 894         req->cmd.cid = tr->cid;
 895         qpair->act_tr[tr->cid] = tr;
 896         ctrlr = qpair->ctrlr;
 897
 898         if (req->timeout)
 899                 callout_reset_curcpu(&tr->timer, ctrlr->timeout_period * hz,
 900                     nvme_timeout, tr);
 901
 902         /* Copy the command from the tracker to the submission queue. */
 903         memcpy(&qpair->cmd[qpair->sq_tail], &req->cmd, sizeof(req->cmd));
 904
 905         if (++qpair->sq_tail == qpair->num_entries)
 906                 qpair->sq_tail = 0;
 907
 908         bus_dmamap_sync(qpair->dma_tag, qpair->queuemem_map,
 909             BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
 910 #ifndef __powerpc__
 911         /*
 912          * powerpc's bus_dmamap_sync() already includes a heavyweight sync, but
 913          * no other archs do.
 914          */
 915         wmb();
 916 #endif
 917
 918         nvme_mmio_write_4(qpair->ctrlr, doorbell[qpair->id].sq_tdbl,
 919             qpair->sq_tail);
 920
 921         qpair->num_cmds++;
 922 }
 923
 924 static void
 925 nvme_payload_map(void *arg, bus_dma_segment_t *seg, int nseg, int error)
 926 {
 927         struct nvme_tracker     *tr = arg;
 928         uint32_t                cur_nseg;
 929
 930         /*
 931          * If the mapping operation failed, return immediately.  The caller
 932          *  is responsible for detecting the error status and failing the
 933          *  tracker manually.
 934          */
 935         if (error != 0) {
 936                 nvme_printf(tr->qpair->ctrlr,
 937                     "nvme_payload_map err %d\n", error);
 938                 return;
 939         }
 940
 941         /*
 942          * Note that we specified PAGE_SIZE for alignment and max
 943          *  segment size when creating the bus dma tags.  So here
 944          *  we can safely just transfer each segment to its
 945          *  associated PRP entry.
 946          */
 947         tr->req->cmd.prp1 = htole64(seg[0].ds_addr);
 948
 949         if (nseg == 2) {
 950                 tr->req->cmd.prp2 = htole64(seg[1].ds_addr);
 951         } else if (nseg > 2) {
 952                 cur_nseg = 1;
 953                 tr->req->cmd.prp2 = htole64((uint64_t)tr->prp_bus_addr);
 954                 while (cur_nseg < nseg) {
 955                         tr->prp[cur_nseg-1] =
 956                             htole64((uint64_t)seg[cur_nseg].ds_addr);
 957                         cur_nseg++;
 958                 }
 959         } else {
 960                 /*
 961                  * prp2 should not be used by the controller
 962                  *  since there is only one segment, but set
 963                  *  to 0 just to be safe.
 964                  */
 965                 tr->req->cmd.prp2 = 0;
 966         }
 967
 968         bus_dmamap_sync(tr->qpair->dma_tag_payload, tr->payload_dma_map,
 969             BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
 970         nvme_qpair_submit_tracker(tr->qpair, tr);
 971 }
 972
 973 static void
 974 _nvme_qpair_submit_request(struct nvme_qpair *qpair, struct nvme_request *req)
 975 {
 976         struct nvme_tracker     *tr;
 977         int                     err = 0;
 978
 979         mtx_assert(&qpair->lock, MA_OWNED);
 980
 981         tr = TAILQ_FIRST(&qpair->free_tr);
 982         req->qpair = qpair;
 983
 984         if (tr == NULL || !qpair->is_enabled) {
 985                 /*
 986                  * No tracker is available, or the qpair is disabled due to
 987                  *  an in-progress controller-level reset or controller
 988                  *  failure.
 989                  */
 990
 991                 if (qpair->ctrlr->is_failed) {
 992                         /*
 993                          * The controller has failed.  Post the request to a
 994                          *  task where it will be aborted, so that we do not
 995                          *  invoke the request's callback in the context
 996                          *  of the submission.
 997                          */
 998                         nvme_ctrlr_post_failed_request(qpair->ctrlr, req);
 999                 } else {
1000                         /*
1001                          * Put the request on the qpair's request queue to be
1002                          *  processed when a tracker frees up via a command
1003                          *  completion or when the controller reset is
1004                          *  completed.
1005                          */
1006                         STAILQ_INSERT_TAIL(&qpair->queued_req, req, stailq);
1007                 }
1008                 return;
1009         }
1010
1011         TAILQ_REMOVE(&qpair->free_tr, tr, tailq);
1012         TAILQ_INSERT_TAIL(&qpair->outstanding_tr, tr, tailq);
1013         tr->req = req;
1014
1015         switch (req->type) {
1016         case NVME_REQUEST_VADDR:
1017                 KASSERT(req->payload_size <= qpair->ctrlr->max_xfer_size,
1018                     ("payload_size (%d) exceeds max_xfer_size (%d)\n",
1019                     req->payload_size, qpair->ctrlr->max_xfer_size));
1020                 err = bus_dmamap_load(tr->qpair->dma_tag_payload,
1021                     tr->payload_dma_map, req->u.payload, req->payload_size,
1022                     nvme_payload_map, tr, 0);
1023                 if (err != 0)
1024                         nvme_printf(qpair->ctrlr,
1025                             "bus_dmamap_load returned 0x%x!\n", err);
1026                 break;
1027         case NVME_REQUEST_NULL:
1028                 nvme_qpair_submit_tracker(tr->qpair, tr);
1029                 break;
1030         case NVME_REQUEST_BIO:
1031                 KASSERT(req->u.bio->bio_bcount <= qpair->ctrlr->max_xfer_size,
1032                     ("bio->bio_bcount (%jd) exceeds max_xfer_size (%d)\n",
1033                     (intmax_t)req->u.bio->bio_bcount,
1034                     qpair->ctrlr->max_xfer_size));
1035                 err = bus_dmamap_load_bio(tr->qpair->dma_tag_payload,
1036                     tr->payload_dma_map, req->u.bio, nvme_payload_map, tr, 0);
1037                 if (err != 0)
1038                         nvme_printf(qpair->ctrlr,
1039                             "bus_dmamap_load_bio returned 0x%x!\n", err);
1040                 break;
1041         case NVME_REQUEST_CCB:
1042                 err = bus_dmamap_load_ccb(tr->qpair->dma_tag_payload,
1043                     tr->payload_dma_map, req->u.payload,
1044                     nvme_payload_map, tr, 0);
1045                 if (err != 0)
1046                         nvme_printf(qpair->ctrlr,
1047                             "bus_dmamap_load_ccb returned 0x%x!\n", err);
1048                 break;
1049         default:
1050                 panic("unknown nvme request type 0x%x\n", req->type);
1051                 break;
1052         }
1053
1054         if (err != 0) {
1055                 /*
1056                  * The dmamap operation failed, so we manually fail the
1057                  *  tracker here with DATA_TRANSFER_ERROR status.
1058                  *
1059                  * nvme_qpair_manual_complete_tracker must not be called
1060                  *  with the qpair lock held.
1061                  */
1062                 mtx_unlock(&qpair->lock);
1063                 nvme_qpair_manual_complete_tracker(qpair, tr, NVME_SCT_GENERIC,
1064                     NVME_SC_DATA_TRANSFER_ERROR, DO_NOT_RETRY, ERROR_PRINT_ALL);
1065                 mtx_lock(&qpair->lock);
1066         }
1067 }
1068
1069 void
1070 nvme_qpair_submit_request(struct nvme_qpair *qpair, struct nvme_request *req)
1071 {
1072
1073         mtx_lock(&qpair->lock);
1074         _nvme_qpair_submit_request(qpair, req);
1075         mtx_unlock(&qpair->lock);
1076 }
1077
1078 static void
1079 nvme_qpair_enable(struct nvme_qpair *qpair)
1080 {
1081
1082         qpair->is_enabled = TRUE;
1083 }
1084
1085 void
1086 nvme_qpair_reset(struct nvme_qpair *qpair)
1087 {
1088
1089         qpair->sq_head = qpair->sq_tail = qpair->cq_head = 0;
1090
1091         /*
1092          * First time through the completion queue, HW will set phase
1093          *  bit on completions to 1.  So set this to 1 here, indicating
1094          *  we're looking for a 1 to know which entries have completed.
1095          *  we'll toggle the bit each time when the completion queue
1096          *  rolls over.
1097          */
1098         qpair->phase = 1;
1099
1100         memset(qpair->cmd, 0,
1101             qpair->num_entries * sizeof(struct nvme_command));
1102         memset(qpair->cpl, 0,
1103             qpair->num_entries * sizeof(struct nvme_completion));
1104 }
1105
1106 void
1107 nvme_admin_qpair_enable(struct nvme_qpair *qpair)
1108 {
1109         struct nvme_tracker             *tr;
1110         struct nvme_tracker             *tr_temp;
1111
1112         /*
1113          * Manually abort each outstanding admin command.  Do not retry
1114          *  admin commands found here, since they will be left over from
1115          *  a controller reset and its likely the context in which the
1116          *  command was issued no longer applies.
1117          */
1118         TAILQ_FOREACH_SAFE(tr, &qpair->outstanding_tr, tailq, tr_temp) {
1119                 nvme_printf(qpair->ctrlr,
1120                     "aborting outstanding admin command\n");
1121                 nvme_qpair_manual_complete_tracker(qpair, tr, NVME_SCT_GENERIC,
1122                     NVME_SC_ABORTED_BY_REQUEST, DO_NOT_RETRY, ERROR_PRINT_ALL);
1123         }
1124
1125         nvme_qpair_enable(qpair);
1126 }
1127
1128 void
1129 nvme_io_qpair_enable(struct nvme_qpair *qpair)
1130 {
1131         STAILQ_HEAD(, nvme_request)     temp;
1132         struct nvme_tracker             *tr;
1133         struct nvme_tracker             *tr_temp;
1134         struct nvme_request             *req;
1135
1136         /*
1137          * Manually abort each outstanding I/O.  This normally results in a
1138          *  retry, unless the retry count on the associated request has
1139          *  reached its limit.
1140          */
1141         TAILQ_FOREACH_SAFE(tr, &qpair->outstanding_tr, tailq, tr_temp) {
1142                 nvme_printf(qpair->ctrlr, "aborting outstanding i/o\n");
1143                 nvme_qpair_manual_complete_tracker(qpair, tr, NVME_SCT_GENERIC,
1144                     NVME_SC_ABORTED_BY_REQUEST, 0, ERROR_PRINT_NO_RETRY);
1145         }
1146
1147         mtx_lock(&qpair->lock);
1148
1149         nvme_qpair_enable(qpair);
1150
1151         STAILQ_INIT(&temp);
1152         STAILQ_SWAP(&qpair->queued_req, &temp, nvme_request);
1153
1154         while (!STAILQ_EMPTY(&temp)) {
1155                 req = STAILQ_FIRST(&temp);
1156                 STAILQ_REMOVE_HEAD(&temp, stailq);
1157                 nvme_printf(qpair->ctrlr, "resubmitting queued i/o\n");
1158                 nvme_qpair_print_command(qpair, &req->cmd);
1159                 _nvme_qpair_submit_request(qpair, req);
1160         }
1161
1162         mtx_unlock(&qpair->lock);
1163 }
1164
1165 static void
1166 nvme_qpair_disable(struct nvme_qpair *qpair)
1167 {
1168         struct nvme_tracker *tr;
1169
1170         qpair->is_enabled = FALSE;
1171         mtx_lock(&qpair->lock);
1172         TAILQ_FOREACH(tr, &qpair->outstanding_tr, tailq)
1173                 callout_stop(&tr->timer);
1174         mtx_unlock(&qpair->lock);
1175 }
1176
1177 void
1178 nvme_admin_qpair_disable(struct nvme_qpair *qpair)
1179 {
1180
1181         nvme_qpair_disable(qpair);
1182         nvme_admin_qpair_abort_aers(qpair);
1183 }
1184
1185 void
1186 nvme_io_qpair_disable(struct nvme_qpair *qpair)
1187 {
1188
1189         nvme_qpair_disable(qpair);
1190 }
1191
1192 void
1193 nvme_qpair_fail(struct nvme_qpair *qpair)
1194 {
1195         struct nvme_tracker             *tr;
1196         struct nvme_request             *req;
1197
1198         if (!mtx_initialized(&qpair->lock))
1199                 return;
1200
1201         mtx_lock(&qpair->lock);
1202
1203         while (!STAILQ_EMPTY(&qpair->queued_req)) {
1204                 req = STAILQ_FIRST(&qpair->queued_req);
1205                 STAILQ_REMOVE_HEAD(&qpair->queued_req, stailq);
1206                 nvme_printf(qpair->ctrlr, "failing queued i/o\n");
1207                 mtx_unlock(&qpair->lock);
1208                 nvme_qpair_manual_complete_request(qpair, req, NVME_SCT_GENERIC,
1209                     NVME_SC_ABORTED_BY_REQUEST);
1210                 mtx_lock(&qpair->lock);
1211         }
1212
1213         /* Manually abort each outstanding I/O. */
1214         while (!TAILQ_EMPTY(&qpair->outstanding_tr)) {
1215                 tr = TAILQ_FIRST(&qpair->outstanding_tr);
1216                 /*
1217                  * Do not remove the tracker.  The abort_tracker path will
1218                  *  do that for us.
1219                  */
1220                 nvme_printf(qpair->ctrlr, "failing outstanding i/o\n");
1221                 mtx_unlock(&qpair->lock);
1222                 nvme_qpair_manual_complete_tracker(qpair, tr, NVME_SCT_GENERIC,
1223                     NVME_SC_ABORTED_BY_REQUEST, DO_NOT_RETRY, ERROR_PRINT_ALL);
1224                 mtx_lock(&qpair->lock);
1225         }
1226
1227         mtx_unlock(&qpair->lock);
1228 }
1229