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