sys/dev/nvme/nvme_qpair.c

   1 /*-
   2  * Copyright (C) 2012-2014 Intel Corporation
   3  * All rights reserved.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  * 1. Redistributions of source code must retain the above copyright
   9  *    notice, this list of conditions and the following disclaimer.
  10  * 2. Redistributions in binary form must reproduce the above copyright
  11  *    notice, this list of conditions and the following disclaimer in the
  12  *    documentation and/or other materials provided with the distribution.
  13  *
  14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  24  * SUCH DAMAGE.
  25  */
  26
  27 #include <sys/cdefs.h>
  28 __FBSDID("$FreeBSD$");
  29
  30 #include <sys/param.h>
  31 #include <sys/bus.h>
  32
  33 #include <dev/pci/pcivar.h>
  34
  35 #include "nvme_private.h"
  36
  37 static void     _nvme_qpair_submit_request(struct nvme_qpair *qpair,
  38                                            struct nvme_request *req);
  39 static void     nvme_qpair_destroy(struct nvme_qpair *qpair);
  40
  41 struct nvme_opcode_string {
  42
  43         uint16_t        opc;
  44         const char *    str;
  45 };
  46
  47 static struct nvme_opcode_string admin_opcode[] = {
  48         { NVME_OPC_DELETE_IO_SQ, "DELETE IO SQ" },
  49         { NVME_OPC_CREATE_IO_SQ, "CREATE IO SQ" },
  50         { NVME_OPC_GET_LOG_PAGE, "GET LOG PAGE" },
  51         { NVME_OPC_DELETE_IO_CQ, "DELETE IO CQ" },
  52         { NVME_OPC_CREATE_IO_CQ, "CREATE IO CQ" },
  53         { NVME_OPC_IDENTIFY, "IDENTIFY" },
  54         { NVME_OPC_ABORT, "ABORT" },
  55         { NVME_OPC_SET_FEATURES, "SET FEATURES" },
  56         { NVME_OPC_GET_FEATURES, "GET FEATURES" },
  57         { NVME_OPC_ASYNC_EVENT_REQUEST, "ASYNC EVENT REQUEST" },
  58         { NVME_OPC_FIRMWARE_ACTIVATE, "FIRMWARE ACTIVATE" },
  59         { NVME_OPC_FIRMWARE_IMAGE_DOWNLOAD, "FIRMWARE IMAGE DOWNLOAD" },
  60         { NVME_OPC_FORMAT_NVM, "FORMAT NVM" },
  61         { NVME_OPC_SECURITY_SEND, "SECURITY SEND" },
  62         { NVME_OPC_SECURITY_RECEIVE, "SECURITY RECEIVE" },
  63         { 0xFFFF, "ADMIN COMMAND" }
  64 };
  65
  66 static struct nvme_opcode_string io_opcode[] = {
  67         { NVME_OPC_FLUSH, "FLUSH" },
  68         { NVME_OPC_WRITE, "WRITE" },
  69         { NVME_OPC_READ, "READ" },
  70         { NVME_OPC_WRITE_UNCORRECTABLE, "WRITE UNCORRECTABLE" },
  71         { NVME_OPC_COMPARE, "COMPARE" },
  72         { NVME_OPC_DATASET_MANAGEMENT, "DATASET MANAGEMENT" },
  73         { 0xFFFF, "IO COMMAND" }
  74 };
  75
  76 static const char *
  77 get_admin_opcode_string(uint16_t opc)
  78 {
  79         struct nvme_opcode_string *entry;
  80
  81         entry = admin_opcode;
  82
  83         while (entry->opc != 0xFFFF) {
  84                 if (entry->opc == opc)
  85                         return (entry->str);
  86                 entry++;
  87         }
  88         return (entry->str);
  89 }
  90
  91 static const char *
  92 get_io_opcode_string(uint16_t opc)
  93 {
  94         struct nvme_opcode_string *entry;
  95
  96         entry = io_opcode;
  97
  98         while (entry->opc != 0xFFFF) {
  99                 if (entry->opc == opc)
 100                         return (entry->str);
 101                 entry++;
 102         }
 103         return (entry->str);
 104 }
 105
 106
 107 static void
 108 nvme_admin_qpair_print_command(struct nvme_qpair *qpair,
 109     struct nvme_command *cmd)
 110 {
 111
 112         nvme_printf(qpair->ctrlr, "%s (%02x) sqid:%d cid:%d nsid:%x "
 113             "cdw10:%08x cdw11:%08x\n",
 114             get_admin_opcode_string(cmd->opc), cmd->opc, qpair->id, cmd->cid,
 115             cmd->nsid, cmd->cdw10, cmd->cdw11);
 116 }
 117
 118 static void
 119 nvme_io_qpair_print_command(struct nvme_qpair *qpair,
 120     struct nvme_command *cmd)
 121 {
 122
 123         switch (cmd->opc) {
 124         case NVME_OPC_WRITE:
 125         case NVME_OPC_READ:
 126         case NVME_OPC_WRITE_UNCORRECTABLE:
 127         case NVME_OPC_COMPARE:
 128                 nvme_printf(qpair->ctrlr, "%s sqid:%d cid:%d nsid:%d "
 129                     "lba:%llu len:%d\n",
 130                     get_io_opcode_string(cmd->opc), qpair->id, cmd->cid,
 131                     cmd->nsid,
 132                     ((unsigned long long)cmd->cdw11 << 32) + cmd->cdw10,
 133                     (cmd->cdw12 & 0xFFFF) + 1);
 134                 break;
 135         case NVME_OPC_FLUSH:
 136         case NVME_OPC_DATASET_MANAGEMENT:
 137                 nvme_printf(qpair->ctrlr, "%s sqid:%d cid:%d nsid:%d\n",
 138                     get_io_opcode_string(cmd->opc), qpair->id, cmd->cid,
 139                     cmd->nsid);
 140                 break;
 141         default:
 142                 nvme_printf(qpair->ctrlr, "%s (%02x) sqid:%d cid:%d nsid:%d\n",
 143                     get_io_opcode_string(cmd->opc), cmd->opc, qpair->id,
 144                     cmd->cid, cmd->nsid);
 145                 break;
 146         }
 147 }
 148
 149 static void
 150 nvme_qpair_print_command(struct nvme_qpair *qpair, struct nvme_command *cmd)
 151 {
 152         if (qpair->id == 0)
 153                 nvme_admin_qpair_print_command(qpair, cmd);
 154         else
 155                 nvme_io_qpair_print_command(qpair, cmd);
 156 }
 157
 158 struct nvme_status_string {
 159
 160         uint16_t        sc;
 161         const char *    str;
 162 };
 163
 164 static struct nvme_status_string generic_status[] = {
 165         { NVME_SC_SUCCESS, "SUCCESS" },
 166         { NVME_SC_INVALID_OPCODE, "INVALID OPCODE" },
 167         { NVME_SC_INVALID_FIELD, "INVALID_FIELD" },
 168         { NVME_SC_COMMAND_ID_CONFLICT, "COMMAND ID CONFLICT" },
 169         { NVME_SC_DATA_TRANSFER_ERROR, "DATA TRANSFER ERROR" },
 170         { NVME_SC_ABORTED_POWER_LOSS, "ABORTED - POWER LOSS" },
 171         { NVME_SC_INTERNAL_DEVICE_ERROR, "INTERNAL DEVICE ERROR" },
 172         { NVME_SC_ABORTED_BY_REQUEST, "ABORTED - BY REQUEST" },
 173         { NVME_SC_ABORTED_SQ_DELETION, "ABORTED - SQ DELETION" },
 174         { NVME_SC_ABORTED_FAILED_FUSED, "ABORTED - FAILED FUSED" },
 175         { NVME_SC_ABORTED_MISSING_FUSED, "ABORTED - MISSING FUSED" },
 176         { NVME_SC_INVALID_NAMESPACE_OR_FORMAT, "INVALID NAMESPACE OR FORMAT" },
 177         { NVME_SC_COMMAND_SEQUENCE_ERROR, "COMMAND SEQUENCE ERROR" },
 178         { NVME_SC_LBA_OUT_OF_RANGE, "LBA OUT OF RANGE" },
 179         { NVME_SC_CAPACITY_EXCEEDED, "CAPACITY EXCEEDED" },
 180         { NVME_SC_NAMESPACE_NOT_READY, "NAMESPACE NOT READY" },
 181         { 0xFFFF, "GENERIC" }
 182 };
 183
 184 static struct nvme_status_string command_specific_status[] = {
 185         { NVME_SC_COMPLETION_QUEUE_INVALID, "INVALID COMPLETION QUEUE" },
 186         { NVME_SC_INVALID_QUEUE_IDENTIFIER, "INVALID QUEUE IDENTIFIER" },
 187         { NVME_SC_MAXIMUM_QUEUE_SIZE_EXCEEDED, "MAX QUEUE SIZE EXCEEDED" },
 188         { NVME_SC_ABORT_COMMAND_LIMIT_EXCEEDED, "ABORT CMD LIMIT EXCEEDED" },
 189         { NVME_SC_ASYNC_EVENT_REQUEST_LIMIT_EXCEEDED, "ASYNC LIMIT EXCEEDED" },
 190         { NVME_SC_INVALID_FIRMWARE_SLOT, "INVALID FIRMWARE SLOT" },
 191         { NVME_SC_INVALID_FIRMWARE_IMAGE, "INVALID FIRMWARE IMAGE" },
 192         { NVME_SC_INVALID_INTERRUPT_VECTOR, "INVALID INTERRUPT VECTOR" },
 193         { NVME_SC_INVALID_LOG_PAGE, "INVALID LOG PAGE" },
 194         { NVME_SC_INVALID_FORMAT, "INVALID FORMAT" },
 195         { NVME_SC_FIRMWARE_REQUIRES_RESET, "FIRMWARE REQUIRES RESET" },
 196         { NVME_SC_CONFLICTING_ATTRIBUTES, "CONFLICTING ATTRIBUTES" },
 197         { NVME_SC_INVALID_PROTECTION_INFO, "INVALID PROTECTION INFO" },
 198         { NVME_SC_ATTEMPTED_WRITE_TO_RO_PAGE, "WRITE TO RO PAGE" },
 199         { 0xFFFF, "COMMAND SPECIFIC" }
 200 };
 201
 202 static struct nvme_status_string media_error_status[] = {
 203         { NVME_SC_WRITE_FAULTS, "WRITE FAULTS" },
 204         { NVME_SC_UNRECOVERED_READ_ERROR, "UNRECOVERED READ ERROR" },
 205         { NVME_SC_GUARD_CHECK_ERROR, "GUARD CHECK ERROR" },
 206         { NVME_SC_APPLICATION_TAG_CHECK_ERROR, "APPLICATION TAG CHECK ERROR" },
 207         { NVME_SC_REFERENCE_TAG_CHECK_ERROR, "REFERENCE TAG CHECK ERROR" },
 208         { NVME_SC_COMPARE_FAILURE, "COMPARE FAILURE" },
 209         { NVME_SC_ACCESS_DENIED, "ACCESS DENIED" },
 210         { 0xFFFF, "MEDIA ERROR" }
 211 };
 212
 213 static const char *
 214 get_status_string(uint16_t sct, uint16_t sc)
 215 {
 216         struct nvme_status_string *entry;
 217
 218         switch (sct) {
 219         case NVME_SCT_GENERIC:
 220                 entry = generic_status;
 221                 break;
 222         case NVME_SCT_COMMAND_SPECIFIC:
 223                 entry = command_specific_status;
 224                 break;
 225         case NVME_SCT_MEDIA_ERROR:
 226                 entry = media_error_status;
 227                 break;
 228         case NVME_SCT_VENDOR_SPECIFIC:
 229                 return ("VENDOR SPECIFIC");
 230         default:
 231                 return ("RESERVED");
 232         }
 233
 234         while (entry->sc != 0xFFFF) {
 235                 if (entry->sc == sc)
 236                         return (entry->str);
 237                 entry++;
 238         }
 239         return (entry->str);
 240 }
 241
 242 static void
 243 nvme_qpair_print_completion(struct nvme_qpair *qpair,
 244     struct nvme_completion *cpl)
 245 {
 246         nvme_printf(qpair->ctrlr, "%s (%02x/%02x) sqid:%d cid:%d cdw0:%x\n",
 247             get_status_string(cpl->status.sct, cpl->status.sc),
 248             cpl->status.sct, cpl->status.sc, cpl->sqid, cpl->cid, cpl->cdw0);
 249 }
 250
 251 static boolean_t
 252 nvme_completion_is_retry(const struct nvme_completion *cpl)
 253 {
 254         /*
 255          * TODO: spec is not clear how commands that are aborted due
 256          *  to TLER will be marked.  So for now, it seems
 257          *  NAMESPACE_NOT_READY is the only case where we should
 258          *  look at the DNR bit.
 259          */
 260         switch (cpl->status.sct) {
 261         case NVME_SCT_GENERIC:
 262                 switch (cpl->status.sc) {
 263                 case NVME_SC_ABORTED_BY_REQUEST:
 264                 case NVME_SC_NAMESPACE_NOT_READY:
 265                         if (cpl->status.dnr)
 266                                 return (0);
 267                         else
 268                                 return (1);
 269                 case NVME_SC_INVALID_OPCODE:
 270                 case NVME_SC_INVALID_FIELD:
 271                 case NVME_SC_COMMAND_ID_CONFLICT:
 272                 case NVME_SC_DATA_TRANSFER_ERROR:
 273                 case NVME_SC_ABORTED_POWER_LOSS:
 274                 case NVME_SC_INTERNAL_DEVICE_ERROR:
 275                 case NVME_SC_ABORTED_SQ_DELETION:
 276                 case NVME_SC_ABORTED_FAILED_FUSED:
 277                 case NVME_SC_ABORTED_MISSING_FUSED:
 278                 case NVME_SC_INVALID_NAMESPACE_OR_FORMAT:
 279                 case NVME_SC_COMMAND_SEQUENCE_ERROR:
 280                 case NVME_SC_LBA_OUT_OF_RANGE:
 281                 case NVME_SC_CAPACITY_EXCEEDED:
 282                 default:
 283                         return (0);
 284                 }
 285         case NVME_SCT_COMMAND_SPECIFIC:
 286         case NVME_SCT_MEDIA_ERROR:
 287         case NVME_SCT_VENDOR_SPECIFIC:
 288         default:
 289                 return (0);
 290         }
 291 }
 292
 293 static void
 294 nvme_qpair_complete_tracker(struct nvme_qpair *qpair, struct nvme_tracker *tr,
 295     struct nvme_completion *cpl, boolean_t print_on_error)
 296 {
 297         struct nvme_request     *req;
 298         boolean_t               retry, error;
 299
 300         req = tr->req;
 301         error = nvme_completion_is_error(cpl);
 302         retry = error && nvme_completion_is_retry(cpl) &&
 303            req->retries < nvme_retry_count;
 304
 305         if (error && print_on_error) {
 306                 nvme_qpair_print_command(qpair, &req->cmd);
 307                 nvme_qpair_print_completion(qpair, cpl);
 308         }
 309
 310         qpair->act_tr[cpl->cid] = NULL;
 311
 312         KASSERT(cpl->cid == req->cmd.cid, ("cpl cid does not match cmd cid\n"));
 313
 314         if (req->cb_fn && !retry)
 315                 req->cb_fn(req->cb_arg, cpl);
 316
 317         mtx_lock(&qpair->lock);
 318         callout_stop(&tr->timer);
 319
 320         if (retry) {
 321                 req->retries++;
 322                 nvme_qpair_submit_tracker(qpair, tr);
 323         } else {
 324                 if (req->type != NVME_REQUEST_NULL)
 325                         bus_dmamap_unload(qpair->dma_tag_payload,
 326                             tr->payload_dma_map);
 327
 328                 nvme_free_request(req);
 329                 tr->req = NULL;
 330
 331                 TAILQ_REMOVE(&qpair->outstanding_tr, tr, tailq);
 332                 TAILQ_INSERT_HEAD(&qpair->free_tr, tr, tailq);
 333
 334                 /*
 335                  * If the controller is in the middle of resetting, don't
 336                  *  try to submit queued requests here - let the reset logic
 337                  *  handle that instead.
 338                  */
 339                 if (!STAILQ_EMPTY(&qpair->queued_req) &&
 340                     !qpair->ctrlr->is_resetting) {
 341                         req = STAILQ_FIRST(&qpair->queued_req);
 342                         STAILQ_REMOVE_HEAD(&qpair->queued_req, stailq);
 343                         _nvme_qpair_submit_request(qpair, req);
 344                 }
 345         }
 346
 347         mtx_unlock(&qpair->lock);
 348 }
 349
 350 static void
 351 nvme_qpair_manual_complete_tracker(struct nvme_qpair *qpair,
 352     struct nvme_tracker *tr, uint32_t sct, uint32_t sc, uint32_t dnr,
 353     boolean_t print_on_error)
 354 {
 355         struct nvme_completion  cpl;
 356
 357         memset(&cpl, 0, sizeof(cpl));
 358         cpl.sqid = qpair->id;
 359         cpl.cid = tr->cid;
 360         cpl.status.sct = sct;
 361         cpl.status.sc = sc;
 362         cpl.status.dnr = dnr;
 363         nvme_qpair_complete_tracker(qpair, tr, &cpl, print_on_error);
 364 }
 365
 366 void
 367 nvme_qpair_manual_complete_request(struct nvme_qpair *qpair,
 368     struct nvme_request *req, uint32_t sct, uint32_t sc,
 369     boolean_t print_on_error)
 370 {
 371         struct nvme_completion  cpl;
 372         boolean_t               error;
 373
 374         memset(&cpl, 0, sizeof(cpl));
 375         cpl.sqid = qpair->id;
 376         cpl.status.sct = sct;
 377         cpl.status.sc = sc;
 378
 379         error = nvme_completion_is_error(&cpl);
 380
 381         if (error && print_on_error) {
 382                 nvme_qpair_print_command(qpair, &req->cmd);
 383                 nvme_qpair_print_completion(qpair, &cpl);
 384         }
 385
 386         if (req->cb_fn)
 387                 req->cb_fn(req->cb_arg, &cpl);
 388
 389         nvme_free_request(req);
 390 }
 391
 392 void
 393 nvme_qpair_process_completions(struct nvme_qpair *qpair)
 394 {
 395         struct nvme_tracker     *tr;
 396         struct nvme_completion  *cpl;
 397
 398         qpair->num_intr_handler_calls++;
 399
 400         if (!qpair->is_enabled)
 401                 /*
 402                  * qpair is not enabled, likely because a controller reset is
 403                  *  is in progress.  Ignore the interrupt - any I/O that was
 404                  *  associated with this interrupt will get retried when the
 405                  *  reset is complete.
 406                  */
 407                 return;
 408
 409         while (1) {
 410                 cpl = &qpair->cpl[qpair->cq_head];
 411
 412                 if (cpl->status.p != qpair->phase)
 413                         break;
 414
 415                 tr = qpair->act_tr[cpl->cid];
 416
 417                 if (tr != NULL) {
 418                         nvme_qpair_complete_tracker(qpair, tr, cpl, TRUE);
 419                         qpair->sq_head = cpl->sqhd;
 420                 } else {
 421                         nvme_printf(qpair->ctrlr,
 422                             "cpl does not map to outstanding cmd\n");
 423                         nvme_dump_completion(cpl);
 424                         KASSERT(0, ("received completion for unknown cmd\n"));
 425                 }
 426
 427                 if (++qpair->cq_head == qpair->num_entries) {
 428                         qpair->cq_head = 0;
 429                         qpair->phase = !qpair->phase;
 430                 }
 431
 432                 nvme_mmio_write_4(qpair->ctrlr, doorbell[qpair->id].cq_hdbl,
 433                     qpair->cq_head);
 434         }
 435 }
 436
 437 static void
 438 nvme_qpair_msix_handler(void *arg)
 439 {
 440         struct nvme_qpair *qpair = arg;
 441
 442         nvme_qpair_process_completions(qpair);
 443 }
 444
 445 int
 446 nvme_qpair_construct(struct nvme_qpair *qpair, uint32_t id,
 447     uint16_t vector, uint32_t num_entries, uint32_t num_trackers,
 448     struct nvme_controller *ctrlr)
 449 {
 450         struct nvme_tracker     *tr;
 451         size_t                  cmdsz, cplsz, prpsz, allocsz, prpmemsz;
 452         uint64_t                queuemem_phys, prpmem_phys, list_phys;
 453         uint8_t                 *queuemem, *prpmem, *prp_list;
 454         int                     i, err;
 455
 456         qpair->id = id;
 457         qpair->vector = vector;
 458         qpair->num_entries = num_entries;
 459         qpair->num_trackers = num_trackers;
 460         qpair->ctrlr = ctrlr;
 461
 462         if (ctrlr->msix_enabled) {
 463
 464                 /*
 465                  * MSI-X vector resource IDs start at 1, so we add one to
 466                  *  the queue's vector to get the corresponding rid to use.
 467                  */
 468                 qpair->rid = vector + 1;
 469
 470                 qpair->res = bus_alloc_resource_any(ctrlr->dev, SYS_RES_IRQ,
 471                     &qpair->rid, RF_ACTIVE);
 472                 bus_setup_intr(ctrlr->dev, qpair->res,
 473                     INTR_TYPE_MISC | INTR_MPSAFE, NULL,
 474                     nvme_qpair_msix_handler, qpair, &qpair->tag);
 475         }
 476
 477         mtx_init(&qpair->lock, "nvme qpair lock", NULL, MTX_DEF);
 478
 479         /* Note: NVMe PRP format is restricted to 4-byte alignment. */
 480         err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev),
 481             4, PAGE_SIZE, BUS_SPACE_MAXADDR,
 482             BUS_SPACE_MAXADDR, NULL, NULL, NVME_MAX_XFER_SIZE,
 483             (NVME_MAX_XFER_SIZE/PAGE_SIZE)+1, PAGE_SIZE, 0,
 484             NULL, NULL, &qpair->dma_tag_payload);
 485         if (err != 0) {
 486                 nvme_printf(ctrlr, "payload tag create failed %d\n", err);
 487                 goto out;
 488         }
 489
 490         /*
 491          * Each component must be page aligned, and individual PRP lists
 492          * cannot cross a page boundary.
 493          */
 494         cmdsz = qpair->num_entries * sizeof(struct nvme_command);
 495         cmdsz = roundup2(cmdsz, PAGE_SIZE);
 496         cplsz = qpair->num_entries * sizeof(struct nvme_completion);
 497         cplsz = roundup2(cplsz, PAGE_SIZE);
 498         prpsz = sizeof(uint64_t) * NVME_MAX_PRP_LIST_ENTRIES;;
 499         prpmemsz = qpair->num_trackers * prpsz;
 500         allocsz = cmdsz + cplsz + prpmemsz;
 501
 502         err = bus_dma_tag_create(bus_get_dma_tag(ctrlr->dev),
 503             PAGE_SIZE, 0, BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
 504             allocsz, 1, allocsz, 0, NULL, NULL, &qpair->dma_tag);
 505         if (err != 0) {
 506                 nvme_printf(ctrlr, "tag create failed %d\n", err);
 507                 goto out;
 508         }
 509
 510         if (bus_dmamem_alloc(qpair->dma_tag, (void **)&queuemem,
 511             BUS_DMA_NOWAIT, &qpair->queuemem_map)) {
 512                 nvme_printf(ctrlr, "failed to alloc qpair memory\n");
 513                 goto out;
 514         }
 515
 516         if (bus_dmamap_load(qpair->dma_tag, qpair->queuemem_map,
 517             queuemem, allocsz, nvme_single_map, &queuemem_phys, 0) != 0) {
 518                 nvme_printf(ctrlr, "failed to load qpair memory\n");
 519                 goto out;
 520         }
 521
 522         qpair->num_cmds = 0;
 523         qpair->num_intr_handler_calls = 0;
 524         qpair->cmd = (struct nvme_command *)queuemem;
 525         qpair->cpl = (struct nvme_completion *)(queuemem + cmdsz);
 526         prpmem = (uint8_t *)(queuemem + cmdsz + cplsz);
 527         qpair->cmd_bus_addr = queuemem_phys;
 528         qpair->cpl_bus_addr = queuemem_phys + cmdsz;
 529         prpmem_phys = queuemem_phys + cmdsz + cplsz;
 530
 531         qpair->sq_tdbl_off = nvme_mmio_offsetof(doorbell[id].sq_tdbl);
 532         qpair->cq_hdbl_off = nvme_mmio_offsetof(doorbell[id].cq_hdbl);
 533
 534         TAILQ_INIT(&qpair->free_tr);
 535         TAILQ_INIT(&qpair->outstanding_tr);
 536         STAILQ_INIT(&qpair->queued_req);
 537
 538         list_phys = prpmem_phys;
 539         prp_list = prpmem;
 540         for (i = 0; i < qpair->num_trackers; i++) {
 541
 542                 if (list_phys + prpsz > prpmem_phys + prpmemsz) {
 543                         qpair->num_trackers = i;
 544                         break;
 545                 }
 546
 547                 /*
 548                  * Make sure that the PRP list for this tracker doesn't
 549                  * overflow to another page.
 550                  */
 551                 if (trunc_page(list_phys) !=
 552                     trunc_page(list_phys + prpsz - 1)) {
 553                         list_phys = roundup2(list_phys, PAGE_SIZE);
 554                         prp_list =
 555                             (uint8_t *)roundup2((uintptr_t)prp_list, PAGE_SIZE);
 556                 }
 557
 558                 tr = malloc(sizeof(*tr), M_NVME, M_ZERO | M_WAITOK);
 559                 bus_dmamap_create(qpair->dma_tag_payload, 0,
 560                     &tr->payload_dma_map);
 561                 callout_init(&tr->timer, 1);
 562                 tr->cid = i;
 563                 tr->qpair = qpair;
 564                 tr->prp = (uint64_t *)prp_list;
 565                 tr->prp_bus_addr = list_phys;
 566                 TAILQ_INSERT_HEAD(&qpair->free_tr, tr, tailq);
 567                 list_phys += prpsz;
 568                 prp_list += prpsz;
 569         }
 570
 571         if (qpair->num_trackers == 0) {
 572                 nvme_printf(ctrlr, "failed to allocate enough trackers\n");
 573                 goto out;
 574         }
 575
 576         qpair->act_tr = malloc(sizeof(struct nvme_tracker *) *
 577             qpair->num_entries, M_NVME, M_ZERO | M_WAITOK);
 578         return (0);
 579
 580 out:
 581         nvme_qpair_destroy(qpair);
 582         return (ENOMEM);
 583 }
 584
 585 static void
 586 nvme_qpair_destroy(struct nvme_qpair *qpair)
 587 {
 588         struct nvme_tracker     *tr;
 589
 590         if (qpair->tag)
 591                 bus_teardown_intr(qpair->ctrlr->dev, qpair->res, qpair->tag);
 592
 593         if (mtx_initialized(&qpair->lock))
 594                 mtx_destroy(&qpair->lock);
 595
 596         if (qpair->res)
 597                 bus_release_resource(qpair->ctrlr->dev, SYS_RES_IRQ,
 598                     rman_get_rid(qpair->res), qpair->res);
 599
 600         if (qpair->cmd != NULL) {
 601                 bus_dmamap_unload(qpair->dma_tag, qpair->queuemem_map);
 602                 bus_dmamem_free(qpair->dma_tag, qpair->cmd,
 603                     qpair->queuemem_map);
 604         }
 605
 606         if (qpair->dma_tag)
 607                 bus_dma_tag_destroy(qpair->dma_tag);
 608
 609         if (qpair->dma_tag_payload)
 610                 bus_dma_tag_destroy(qpair->dma_tag_payload);
 611
 612         if (qpair->act_tr)
 613                 free(qpair->act_tr, M_NVME);
 614
 615         while (!TAILQ_EMPTY(&qpair->free_tr)) {
 616                 tr = TAILQ_FIRST(&qpair->free_tr);
 617                 TAILQ_REMOVE(&qpair->free_tr, tr, tailq);
 618                 bus_dmamap_destroy(qpair->dma_tag, tr->payload_dma_map);
 619                 free(tr, M_NVME);
 620         }
 621 }
 622
 623 static void
 624 nvme_admin_qpair_abort_aers(struct nvme_qpair *qpair)
 625 {
 626         struct nvme_tracker     *tr;
 627
 628         tr = TAILQ_FIRST(&qpair->outstanding_tr);
 629         while (tr != NULL) {
 630                 if (tr->req->cmd.opc == NVME_OPC_ASYNC_EVENT_REQUEST) {
 631                         nvme_qpair_manual_complete_tracker(qpair, tr,
 632                             NVME_SCT_GENERIC, NVME_SC_ABORTED_SQ_DELETION, 0,
 633                             FALSE);
 634                         tr = TAILQ_FIRST(&qpair->outstanding_tr);
 635                 } else {
 636                         tr = TAILQ_NEXT(tr, tailq);
 637                 }
 638         }
 639 }
 640
 641 void
 642 nvme_admin_qpair_destroy(struct nvme_qpair *qpair)
 643 {
 644
 645         nvme_admin_qpair_abort_aers(qpair);
 646         nvme_qpair_destroy(qpair);
 647 }
 648
 649 void
 650 nvme_io_qpair_destroy(struct nvme_qpair *qpair)
 651 {
 652
 653         nvme_qpair_destroy(qpair);
 654 }
 655
 656 static void
 657 nvme_abort_complete(void *arg, const struct nvme_completion *status)
 658 {
 659         struct nvme_tracker     *tr = arg;
 660
 661         /*
 662          * If cdw0 == 1, the controller was not able to abort the command
 663          *  we requested.  We still need to check the active tracker array,
 664          *  to cover race where I/O timed out at same time controller was
 665          *  completing the I/O.
 666          */
 667         if (status->cdw0 == 1 && tr->qpair->act_tr[tr->cid] != NULL) {
 668                 /*
 669                  * An I/O has timed out, and the controller was unable to
 670                  *  abort it for some reason.  Construct a fake completion
 671                  *  status, and then complete the I/O's tracker manually.
 672                  */
 673                 nvme_printf(tr->qpair->ctrlr,
 674                     "abort command failed, aborting command manually\n");
 675                 nvme_qpair_manual_complete_tracker(tr->qpair, tr,
 676                     NVME_SCT_GENERIC, NVME_SC_ABORTED_BY_REQUEST, 0, TRUE);
 677         }
 678 }
 679
 680 static void
 681 nvme_timeout(void *arg)
 682 {
 683         struct nvme_tracker     *tr = arg;
 684         struct nvme_qpair       *qpair = tr->qpair;
 685         struct nvme_controller  *ctrlr = qpair->ctrlr;
 686         union csts_register     csts;
 687
 688         /* Read csts to get value of cfs - controller fatal status. */
 689         csts.raw = nvme_mmio_read_4(ctrlr, csts);
 690
 691         if (ctrlr->enable_aborts && csts.bits.cfs == 0) {
 692                 /*
 693                  * If aborts are enabled, only use them if the controller is
 694                  *  not reporting fatal status.
 695                  */
 696                 nvme_ctrlr_cmd_abort(ctrlr, tr->cid, qpair->id,
 697                     nvme_abort_complete, tr);
 698         } else
 699                 nvme_ctrlr_reset(ctrlr);
 700 }
 701
 702 void
 703 nvme_qpair_submit_tracker(struct nvme_qpair *qpair, struct nvme_tracker *tr)
 704 {
 705         struct nvme_request     *req;
 706         struct nvme_controller  *ctrlr;
 707
 708         mtx_assert(&qpair->lock, MA_OWNED);
 709
 710         req = tr->req;
 711         req->cmd.cid = tr->cid;
 712         qpair->act_tr[tr->cid] = tr;
 713         ctrlr = qpair->ctrlr;
 714
 715         if (req->timeout)
 716 #if __FreeBSD_version >= 800030
 717                 callout_reset_curcpu(&tr->timer, ctrlr->timeout_period * hz,
 718                     nvme_timeout, tr);
 719 #else
 720                 callout_reset(&tr->timer, ctrlr->timeout_period * hz,
 721                     nvme_timeout, tr);
 722 #endif
 723
 724         /* Copy the command from the tracker to the submission queue. */
 725         memcpy(&qpair->cmd[qpair->sq_tail], &req->cmd, sizeof(req->cmd));
 726
 727         if (++qpair->sq_tail == qpair->num_entries)
 728                 qpair->sq_tail = 0;
 729
 730         wmb();
 731         nvme_mmio_write_4(qpair->ctrlr, doorbell[qpair->id].sq_tdbl,
 732             qpair->sq_tail);
 733
 734         qpair->num_cmds++;
 735 }
 736
 737 static void
 738 nvme_payload_map(void *arg, bus_dma_segment_t *seg, int nseg, int error)
 739 {
 740         struct nvme_tracker     *tr = arg;
 741         uint32_t                cur_nseg;
 742
 743         /*
 744          * If the mapping operation failed, return immediately.  The caller
 745          *  is responsible for detecting the error status and failing the
 746          *  tracker manually.
 747          */
 748         if (error != 0) {
 749                 nvme_printf(tr->qpair->ctrlr,
 750                     "nvme_payload_map err %d\n", error);
 751                 return;
 752         }
 753
 754         /*
 755          * Note that we specified PAGE_SIZE for alignment and max
 756          *  segment size when creating the bus dma tags.  So here
 757          *  we can safely just transfer each segment to its
 758          *  associated PRP entry.
 759          */
 760         tr->req->cmd.prp1 = seg[0].ds_addr;
 761
 762         if (nseg == 2) {
 763                 tr->req->cmd.prp2 = seg[1].ds_addr;
 764         } else if (nseg > 2) {
 765                 cur_nseg = 1;
 766                 tr->req->cmd.prp2 = (uint64_t)tr->prp_bus_addr;
 767                 while (cur_nseg < nseg) {
 768                         tr->prp[cur_nseg-1] =
 769                             (uint64_t)seg[cur_nseg].ds_addr;
 770                         cur_nseg++;
 771                 }
 772         } else {
 773                 /*
 774                  * prp2 should not be used by the controller
 775                  *  since there is only one segment, but set
 776                  *  to 0 just to be safe.
 777                  */
 778                 tr->req->cmd.prp2 = 0;
 779         }
 780
 781         nvme_qpair_submit_tracker(tr->qpair, tr);
 782 }
 783
 784 static void
 785 _nvme_qpair_submit_request(struct nvme_qpair *qpair, struct nvme_request *req)
 786 {
 787         struct nvme_tracker     *tr;
 788         int                     err = 0;
 789
 790         mtx_assert(&qpair->lock, MA_OWNED);
 791
 792         tr = TAILQ_FIRST(&qpair->free_tr);
 793         req->qpair = qpair;
 794
 795         if (tr == NULL || !qpair->is_enabled) {
 796                 /*
 797                  * No tracker is available, or the qpair is disabled due to
 798                  *  an in-progress controller-level reset or controller
 799                  *  failure.
 800                  */
 801
 802                 if (qpair->ctrlr->is_failed) {
 803                         /*
 804                          * The controller has failed.  Post the request to a
 805                          *  task where it will be aborted, so that we do not
 806                          *  invoke the request's callback in the context
 807                          *  of the submission.
 808                          */
 809                         nvme_ctrlr_post_failed_request(qpair->ctrlr, req);
 810                 } else {
 811                         /*
 812                          * Put the request on the qpair's request queue to be
 813                          *  processed when a tracker frees up via a command
 814                          *  completion or when the controller reset is
 815                          *  completed.
 816                          */
 817                         STAILQ_INSERT_TAIL(&qpair->queued_req, req, stailq);
 818                 }
 819                 return;
 820         }
 821
 822         TAILQ_REMOVE(&qpair->free_tr, tr, tailq);
 823         TAILQ_INSERT_TAIL(&qpair->outstanding_tr, tr, tailq);
 824         tr->req = req;
 825
 826         switch (req->type) {
 827         case NVME_REQUEST_VADDR:
 828                 KASSERT(req->payload_size <= qpair->ctrlr->max_xfer_size,
 829                     ("payload_size (%d) exceeds max_xfer_size (%d)\n",
 830                     req->payload_size, qpair->ctrlr->max_xfer_size));
 831                 err = bus_dmamap_load(tr->qpair->dma_tag_payload,
 832                     tr->payload_dma_map, req->u.payload, req->payload_size,
 833                     nvme_payload_map, tr, 0);
 834                 if (err != 0)
 835                         nvme_printf(qpair->ctrlr,
 836                             "bus_dmamap_load returned 0x%x!\n", err);
 837                 break;
 838         case NVME_REQUEST_NULL:
 839                 nvme_qpair_submit_tracker(tr->qpair, tr);
 840                 break;
 841 #ifdef NVME_UNMAPPED_BIO_SUPPORT
 842         case NVME_REQUEST_BIO:
 843                 KASSERT(req->u.bio->bio_bcount <= qpair->ctrlr->max_xfer_size,
 844                     ("bio->bio_bcount (%jd) exceeds max_xfer_size (%d)\n",
 845                     (intmax_t)req->u.bio->bio_bcount,
 846                     qpair->ctrlr->max_xfer_size));
 847                 err = bus_dmamap_load_bio(tr->qpair->dma_tag_payload,
 848                     tr->payload_dma_map, req->u.bio, nvme_payload_map, tr, 0);
 849                 if (err != 0)
 850                         nvme_printf(qpair->ctrlr,
 851                             "bus_dmamap_load_bio returned 0x%x!\n", err);
 852                 break;
 853 #endif
 854         case NVME_REQUEST_CCB:
 855                 err = bus_dmamap_load_ccb(tr->qpair->dma_tag_payload,
 856                     tr->payload_dma_map, req->u.payload,
 857                     nvme_payload_map, tr, 0);
 858                 if (err != 0)
 859                         nvme_printf(qpair->ctrlr,
 860                             "bus_dmamap_load_ccb returned 0x%x!\n", err);
 861                 break;
 862         default:
 863                 panic("unknown nvme request type 0x%x\n", req->type);
 864                 break;
 865         }
 866
 867         if (err != 0) {
 868                 /*
 869                  * The dmamap operation failed, so we manually fail the
 870                  *  tracker here with DATA_TRANSFER_ERROR status.
 871                  *
 872                  * nvme_qpair_manual_complete_tracker must not be called
 873                  *  with the qpair lock held.
 874                  */
 875                 mtx_unlock(&qpair->lock);
 876                 nvme_qpair_manual_complete_tracker(qpair, tr, NVME_SCT_GENERIC,
 877                     NVME_SC_DATA_TRANSFER_ERROR, 1 /* do not retry */, TRUE);
 878                 mtx_lock(&qpair->lock);
 879         }
 880 }
 881
 882 void
 883 nvme_qpair_submit_request(struct nvme_qpair *qpair, struct nvme_request *req)
 884 {
 885
 886         mtx_lock(&qpair->lock);
 887         _nvme_qpair_submit_request(qpair, req);
 888         mtx_unlock(&qpair->lock);
 889 }
 890
 891 static void
 892 nvme_qpair_enable(struct nvme_qpair *qpair)
 893 {
 894
 895         qpair->is_enabled = TRUE;
 896 }
 897
 898 void
 899 nvme_qpair_reset(struct nvme_qpair *qpair)
 900 {
 901
 902         qpair->sq_head = qpair->sq_tail = qpair->cq_head = 0;
 903
 904         /*
 905          * First time through the completion queue, HW will set phase
 906          *  bit on completions to 1.  So set this to 1 here, indicating
 907          *  we're looking for a 1 to know which entries have completed.
 908          *  we'll toggle the bit each time when the completion queue
 909          *  rolls over.
 910          */
 911         qpair->phase = 1;
 912
 913         memset(qpair->cmd, 0,
 914             qpair->num_entries * sizeof(struct nvme_command));
 915         memset(qpair->cpl, 0,
 916             qpair->num_entries * sizeof(struct nvme_completion));
 917 }
 918
 919 void
 920 nvme_admin_qpair_enable(struct nvme_qpair *qpair)
 921 {
 922         struct nvme_tracker             *tr;
 923         struct nvme_tracker             *tr_temp;
 924
 925         /*
 926          * Manually abort each outstanding admin command.  Do not retry
 927          *  admin commands found here, since they will be left over from
 928          *  a controller reset and its likely the context in which the
 929          *  command was issued no longer applies.
 930          */
 931         TAILQ_FOREACH_SAFE(tr, &qpair->outstanding_tr, tailq, tr_temp) {
 932                 nvme_printf(qpair->ctrlr,
 933                     "aborting outstanding admin command\n");
 934                 nvme_qpair_manual_complete_tracker(qpair, tr, NVME_SCT_GENERIC,
 935                     NVME_SC_ABORTED_BY_REQUEST, 1 /* do not retry */, TRUE);
 936         }
 937
 938         nvme_qpair_enable(qpair);
 939 }
 940
 941 void
 942 nvme_io_qpair_enable(struct nvme_qpair *qpair)
 943 {
 944         STAILQ_HEAD(, nvme_request)     temp;
 945         struct nvme_tracker             *tr;
 946         struct nvme_tracker             *tr_temp;
 947         struct nvme_request             *req;
 948
 949         /*
 950          * Manually abort each outstanding I/O.  This normally results in a
 951          *  retry, unless the retry count on the associated request has
 952          *  reached its limit.
 953          */
 954         TAILQ_FOREACH_SAFE(tr, &qpair->outstanding_tr, tailq, tr_temp) {
 955                 nvme_printf(qpair->ctrlr, "aborting outstanding i/o\n");
 956                 nvme_qpair_manual_complete_tracker(qpair, tr, NVME_SCT_GENERIC,
 957                     NVME_SC_ABORTED_BY_REQUEST, 0, TRUE);
 958         }
 959
 960         mtx_lock(&qpair->lock);
 961
 962         nvme_qpair_enable(qpair);
 963
 964         STAILQ_INIT(&temp);
 965         STAILQ_SWAP(&qpair->queued_req, &temp, nvme_request);
 966
 967         while (!STAILQ_EMPTY(&temp)) {
 968                 req = STAILQ_FIRST(&temp);
 969                 STAILQ_REMOVE_HEAD(&temp, stailq);
 970                 nvme_printf(qpair->ctrlr, "resubmitting queued i/o\n");
 971                 nvme_qpair_print_command(qpair, &req->cmd);
 972                 _nvme_qpair_submit_request(qpair, req);
 973         }
 974
 975         mtx_unlock(&qpair->lock);
 976 }
 977
 978 static void
 979 nvme_qpair_disable(struct nvme_qpair *qpair)
 980 {
 981         struct nvme_tracker *tr;
 982
 983         qpair->is_enabled = FALSE;
 984         mtx_lock(&qpair->lock);
 985         TAILQ_FOREACH(tr, &qpair->outstanding_tr, tailq)
 986                 callout_stop(&tr->timer);
 987         mtx_unlock(&qpair->lock);
 988 }
 989
 990 void
 991 nvme_admin_qpair_disable(struct nvme_qpair *qpair)
 992 {
 993
 994         nvme_qpair_disable(qpair);
 995         nvme_admin_qpair_abort_aers(qpair);
 996 }
 997
 998 void
 999 nvme_io_qpair_disable(struct nvme_qpair *qpair)
1000 {
1001
1002         nvme_qpair_disable(qpair);
1003 }
1004
1005 void
1006 nvme_qpair_fail(struct nvme_qpair *qpair)
1007 {
1008         struct nvme_tracker             *tr;
1009         struct nvme_request             *req;
1010
1011         if (!mtx_initialized(&qpair->lock))
1012                 return;
1013
1014         mtx_lock(&qpair->lock);
1015
1016         while (!STAILQ_EMPTY(&qpair->queued_req)) {
1017                 req = STAILQ_FIRST(&qpair->queued_req);
1018                 STAILQ_REMOVE_HEAD(&qpair->queued_req, stailq);
1019                 nvme_printf(qpair->ctrlr, "failing queued i/o\n");
1020                 mtx_unlock(&qpair->lock);
1021                 nvme_qpair_manual_complete_request(qpair, req, NVME_SCT_GENERIC,
1022                     NVME_SC_ABORTED_BY_REQUEST, TRUE);
1023                 mtx_lock(&qpair->lock);
1024         }
1025
1026         /* Manually abort each outstanding I/O. */
1027         while (!TAILQ_EMPTY(&qpair->outstanding_tr)) {
1028                 tr = TAILQ_FIRST(&qpair->outstanding_tr);
1029                 /*
1030                  * Do not remove the tracker.  The abort_tracker path will
1031                  *  do that for us.
1032                  */
1033                 nvme_printf(qpair->ctrlr, "failing outstanding i/o\n");
1034                 mtx_unlock(&qpair->lock);
1035                 nvme_qpair_manual_complete_tracker(qpair, tr, NVME_SCT_GENERIC,
1036                     NVME_SC_ABORTED_BY_REQUEST, 1 /* do not retry */, TRUE);
1037                 mtx_lock(&qpair->lock);
1038         }
1039
1040         mtx_unlock(&qpair->lock);
1041 }
1042