drivers/nvme/target/admin-cmd.c

   1 /*
   2  * NVMe admin command implementation.
   3  * Copyright (c) 2015-2016 HGST, a Western Digital Company.
   4  *
   5  * This program is free software; you can redistribute it and/or modify it
   6  * under the terms and conditions of the GNU General Public License,
   7  * version 2, as published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  12  * more details.
  13  */
  14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  15 #include <linux/module.h>
  16 #include <linux/rculist.h>
  17
  18 #include <generated/utsrelease.h>
  19 #include <asm/unaligned.h>
  20 #include "nvmet.h"
  21
  22 u32 nvmet_get_log_page_len(struct nvme_command *cmd)
  23 {
  24         u32 len = le16_to_cpu(cmd->get_log_page.numdu);
  25
  26         len <<= 16;
  27         len += le16_to_cpu(cmd->get_log_page.numdl);
  28         /* NUMD is a 0's based value */
  29         len += 1;
  30         len *= sizeof(u32);
  31
  32         return len;
  33 }
  34
  35 static u16 nvmet_get_smart_log_nsid(struct nvmet_req *req,
  36                 struct nvme_smart_log *slog)
  37 {
  38         u16 status;
  39         struct nvmet_ns *ns;
  40         u64 host_reads, host_writes, data_units_read, data_units_written;
  41
  42         status = NVME_SC_SUCCESS;
  43         ns = nvmet_find_namespace(req->sq->ctrl, req->cmd->get_log_page.nsid);
  44         if (!ns) {
  45                 status = NVME_SC_INVALID_NS;
  46                 pr_err("nvmet : Could not find namespace id : %d\n",
  47                                 le32_to_cpu(req->cmd->get_log_page.nsid));
  48                 goto out;
  49         }
  50
  51         host_reads = part_stat_read(ns->bdev->bd_part, ios[READ]);
  52         data_units_read = part_stat_read(ns->bdev->bd_part, sectors[READ]);
  53         host_writes = part_stat_read(ns->bdev->bd_part, ios[WRITE]);
  54         data_units_written = part_stat_read(ns->bdev->bd_part, sectors[WRITE]);
  55
  56         put_unaligned_le64(host_reads, &slog->host_reads[0]);
  57         put_unaligned_le64(data_units_read, &slog->data_units_read[0]);
  58         put_unaligned_le64(host_writes, &slog->host_writes[0]);
  59         put_unaligned_le64(data_units_written, &slog->data_units_written[0]);
  60         nvmet_put_namespace(ns);
  61 out:
  62         return status;
  63 }
  64
  65 static u16 nvmet_get_smart_log_all(struct nvmet_req *req,
  66                 struct nvme_smart_log *slog)
  67 {
  68         u16 status;
  69         u64 host_reads = 0, host_writes = 0;
  70         u64 data_units_read = 0, data_units_written = 0;
  71         struct nvmet_ns *ns;
  72         struct nvmet_ctrl *ctrl;
  73
  74         status = NVME_SC_SUCCESS;
  75         ctrl = req->sq->ctrl;
  76
  77         rcu_read_lock();
  78         list_for_each_entry_rcu(ns, &ctrl->subsys->namespaces, dev_link) {
  79                 host_reads += part_stat_read(ns->bdev->bd_part, ios[READ]);
  80                 data_units_read +=
  81                         part_stat_read(ns->bdev->bd_part, sectors[READ]);
  82                 host_writes += part_stat_read(ns->bdev->bd_part, ios[WRITE]);
  83                 data_units_written +=
  84                         part_stat_read(ns->bdev->bd_part, sectors[WRITE]);
  85
  86         }
  87         rcu_read_unlock();
  88
  89         put_unaligned_le64(host_reads, &slog->host_reads[0]);
  90         put_unaligned_le64(data_units_read, &slog->data_units_read[0]);
  91         put_unaligned_le64(host_writes, &slog->host_writes[0]);
  92         put_unaligned_le64(data_units_written, &slog->data_units_written[0]);
  93
  94         return status;
  95 }
  96
  97 static u16 nvmet_get_smart_log(struct nvmet_req *req,
  98                 struct nvme_smart_log *slog)
  99 {
 100         u16 status;
 101
 102         WARN_ON(req == NULL || slog == NULL);
 103         if (req->cmd->get_log_page.nsid == cpu_to_le32(0xFFFFFFFF))
 104                 status = nvmet_get_smart_log_all(req, slog);
 105         else
 106                 status = nvmet_get_smart_log_nsid(req, slog);
 107         return status;
 108 }
 109
 110 static void nvmet_execute_get_log_page(struct nvmet_req *req)
 111 {
 112         struct nvme_smart_log *smart_log;
 113         size_t data_len = nvmet_get_log_page_len(req->cmd);
 114         void *buf;
 115         u16 status = 0;
 116
 117         buf = kzalloc(data_len, GFP_KERNEL);
 118         if (!buf) {
 119                 status = NVME_SC_INTERNAL;
 120                 goto out;
 121         }
 122
 123         switch (req->cmd->get_log_page.lid) {
 124         case NVME_LOG_ERROR:
 125                 /*
 126                  * We currently never set the More bit in the status field,
 127                  * so all error log entries are invalid and can be zeroed out.
 128                  * This is called a minum viable implementation (TM) of this
 129                  * mandatory log page.
 130                  */
 131                 break;
 132         case NVME_LOG_SMART:
 133                 /*
 134                  * XXX: fill out actual smart log
 135                  *
 136                  * We might have a hard time coming up with useful values for
 137                  * many of the fields, and even when we have useful data
 138                  * available (e.g. units or commands read/written) those aren't
 139                  * persistent over power loss.
 140                  */
 141                 if (data_len != sizeof(*smart_log)) {
 142                         status = NVME_SC_INTERNAL;
 143                         goto err;
 144                 }
 145                 smart_log = buf;
 146                 status = nvmet_get_smart_log(req, smart_log);
 147                 if (status) {
 148                         memset(buf, '\0', data_len);
 149                         goto err;
 150                 }
 151                 break;
 152         case NVME_LOG_FW_SLOT:
 153                 /*
 154                  * We only support a single firmware slot which always is
 155                  * active, so we can zero out the whole firmware slot log and
 156                  * still claim to fully implement this mandatory log page.
 157                  */
 158                 break;
 159         default:
 160                 BUG();
 161         }
 162
 163         status = nvmet_copy_to_sgl(req, 0, buf, data_len);
 164
 165 err:
 166         kfree(buf);
 167 out:
 168         nvmet_req_complete(req, status);
 169 }
 170
 171 static void nvmet_execute_identify_ctrl(struct nvmet_req *req)
 172 {
 173         struct nvmet_ctrl *ctrl = req->sq->ctrl;
 174         struct nvme_id_ctrl *id;
 175         u16 status = 0;
 176
 177         id = kzalloc(sizeof(*id), GFP_KERNEL);
 178         if (!id) {
 179                 status = NVME_SC_INTERNAL;
 180                 goto out;
 181         }
 182
 183         /* XXX: figure out how to assign real vendors IDs. */
 184         id->vid = 0;
 185         id->ssvid = 0;
 186
 187         memset(id->sn, ' ', sizeof(id->sn));
 188         snprintf(id->sn, sizeof(id->sn), "%llx", ctrl->subsys->serial);
 189
 190         memset(id->mn, ' ', sizeof(id->mn));
 191         strncpy((char *)id->mn, "Linux", sizeof(id->mn));
 192
 193         memset(id->fr, ' ', sizeof(id->fr));
 194         strncpy((char *)id->fr, UTS_RELEASE, sizeof(id->fr));
 195
 196         id->rab = 6;
 197
 198         /*
 199          * XXX: figure out how we can assign a IEEE OUI, but until then
 200          * the safest is to leave it as zeroes.
 201          */
 202
 203         /* we support multiple ports and multiples hosts: */
 204         id->cmic = (1 << 0) | (1 << 1);
 205
 206         /* no limit on data transfer sizes for now */
 207         id->mdts = 0;
 208         id->cntlid = cpu_to_le16(ctrl->cntlid);
 209         id->ver = cpu_to_le32(ctrl->subsys->ver);
 210
 211         /* XXX: figure out what to do about RTD3R/RTD3 */
 212         id->oaes = cpu_to_le32(1 << 8);
 213         id->ctratt = cpu_to_le32(1 << 0);
 214
 215         id->oacs = 0;
 216
 217         /*
 218          * We don't really have a practical limit on the number of abort
 219          * comands.  But we don't do anything useful for abort either, so
 220          * no point in allowing more abort commands than the spec requires.
 221          */
 222         id->acl = 3;
 223
 224         id->aerl = NVMET_ASYNC_EVENTS - 1;
 225
 226         /* first slot is read-only, only one slot supported */
 227         id->frmw = (1 << 0) | (1 << 1);
 228         id->lpa = (1 << 0) | (1 << 2);
 229         id->elpe = NVMET_ERROR_LOG_SLOTS - 1;
 230         id->npss = 0;
 231
 232         /* We support keep-alive timeout in granularity of seconds */
 233         id->kas = cpu_to_le16(NVMET_KAS);
 234
 235         id->sqes = (0x6 << 4) | 0x6;
 236         id->cqes = (0x4 << 4) | 0x4;
 237
 238         /* no enforcement soft-limit for maxcmd - pick arbitrary high value */
 239         id->maxcmd = cpu_to_le16(NVMET_MAX_CMD);
 240
 241         id->nn = cpu_to_le32(ctrl->subsys->max_nsid);
 242         id->oncs = cpu_to_le16(NVME_CTRL_ONCS_DSM |
 243                         NVME_CTRL_ONCS_WRITE_ZEROES);
 244
 245         /* XXX: don't report vwc if the underlying device is write through */
 246         id->vwc = NVME_CTRL_VWC_PRESENT;
 247
 248         /*
 249          * We can't support atomic writes bigger than a LBA without support
 250          * from the backend device.
 251          */
 252         id->awun = 0;
 253         id->awupf = 0;
 254
 255         id->sgls = cpu_to_le32(1 << 0); /* we always support SGLs */
 256         if (ctrl->ops->has_keyed_sgls)
 257                 id->sgls |= cpu_to_le32(1 << 2);
 258         if (ctrl->ops->sqe_inline_size)
 259                 id->sgls |= cpu_to_le32(1 << 20);
 260
 261         strcpy(id->subnqn, ctrl->subsys->subsysnqn);
 262
 263         /* Max command capsule size is sqe + single page of in-capsule data */
 264         id->ioccsz = cpu_to_le32((sizeof(struct nvme_command) +
 265                                   ctrl->ops->sqe_inline_size) / 16);
 266         /* Max response capsule size is cqe */
 267         id->iorcsz = cpu_to_le32(sizeof(struct nvme_completion) / 16);
 268
 269         id->msdbd = ctrl->ops->msdbd;
 270
 271         /*
 272          * Meh, we don't really support any power state.  Fake up the same
 273          * values that qemu does.
 274          */
 275         id->psd[0].max_power = cpu_to_le16(0x9c4);
 276         id->psd[0].entry_lat = cpu_to_le32(0x10);
 277         id->psd[0].exit_lat = cpu_to_le32(0x4);
 278
 279         status = nvmet_copy_to_sgl(req, 0, id, sizeof(*id));
 280
 281         kfree(id);
 282 out:
 283         nvmet_req_complete(req, status);
 284 }
 285
 286 static void nvmet_execute_identify_ns(struct nvmet_req *req)
 287 {
 288         struct nvmet_ns *ns;
 289         struct nvme_id_ns *id;
 290         u16 status = 0;
 291
 292         ns = nvmet_find_namespace(req->sq->ctrl, req->cmd->identify.nsid);
 293         if (!ns) {
 294                 status = NVME_SC_INVALID_NS | NVME_SC_DNR;
 295                 goto out;
 296         }
 297
 298         id = kzalloc(sizeof(*id), GFP_KERNEL);
 299         if (!id) {
 300                 status = NVME_SC_INTERNAL;
 301                 goto out_put_ns;
 302         }
 303
 304         /*
 305          * nuse = ncap = nsze isn't aways true, but we have no way to find
 306          * that out from the underlying device.
 307          */
 308         id->ncap = id->nuse = id->nsze =
 309                 cpu_to_le64(ns->size >> ns->blksize_shift);
 310
 311         /*
 312          * We just provide a single LBA format that matches what the
 313          * underlying device reports.
 314          */
 315         id->nlbaf = 0;
 316         id->flbas = 0;
 317
 318         /*
 319          * Our namespace might always be shared.  Not just with other
 320          * controllers, but also with any other user of the block device.
 321          */
 322         id->nmic = (1 << 0);
 323
 324         memcpy(&id->nguid, &ns->nguid, sizeof(uuid_le));
 325
 326         id->lbaf[0].ds = ns->blksize_shift;
 327
 328         status = nvmet_copy_to_sgl(req, 0, id, sizeof(*id));
 329
 330         kfree(id);
 331 out_put_ns:
 332         nvmet_put_namespace(ns);
 333 out:
 334         nvmet_req_complete(req, status);
 335 }
 336
 337 static void nvmet_execute_identify_nslist(struct nvmet_req *req)
 338 {
 339         static const int buf_size = NVME_IDENTIFY_DATA_SIZE;
 340         struct nvmet_ctrl *ctrl = req->sq->ctrl;
 341         struct nvmet_ns *ns;
 342         u32 min_nsid = le32_to_cpu(req->cmd->identify.nsid);
 343         __le32 *list;
 344         u16 status = 0;
 345         int i = 0;
 346
 347         list = kzalloc(buf_size, GFP_KERNEL);
 348         if (!list) {
 349                 status = NVME_SC_INTERNAL;
 350                 goto out;
 351         }
 352
 353         rcu_read_lock();
 354         list_for_each_entry_rcu(ns, &ctrl->subsys->namespaces, dev_link) {
 355                 if (ns->nsid <= min_nsid)
 356                         continue;
 357                 list[i++] = cpu_to_le32(ns->nsid);
 358                 if (i == buf_size / sizeof(__le32))
 359                         break;
 360         }
 361         rcu_read_unlock();
 362
 363         status = nvmet_copy_to_sgl(req, 0, list, buf_size);
 364
 365         kfree(list);
 366 out:
 367         nvmet_req_complete(req, status);
 368 }
 369
 370 static u16 nvmet_copy_ns_identifier(struct nvmet_req *req, u8 type, u8 len,
 371                                     void *id, off_t *off)
 372 {
 373         struct nvme_ns_id_desc desc = {
 374                 .nidt = type,
 375                 .nidl = len,
 376         };
 377         u16 status;
 378
 379         status = nvmet_copy_to_sgl(req, *off, &desc, sizeof(desc));
 380         if (status)
 381                 return status;
 382         *off += sizeof(desc);
 383
 384         status = nvmet_copy_to_sgl(req, *off, id, len);
 385         if (status)
 386                 return status;
 387         *off += len;
 388
 389         return 0;
 390 }
 391
 392 static void nvmet_execute_identify_desclist(struct nvmet_req *req)
 393 {
 394         struct nvmet_ns *ns;
 395         u16 status = 0;
 396         off_t off = 0;
 397
 398         ns = nvmet_find_namespace(req->sq->ctrl, req->cmd->identify.nsid);
 399         if (!ns) {
 400                 status = NVME_SC_INVALID_NS | NVME_SC_DNR;
 401                 goto out;
 402         }
 403
 404         if (memchr_inv(&ns->uuid, 0, sizeof(ns->uuid))) {
 405                 status = nvmet_copy_ns_identifier(req, NVME_NIDT_UUID,
 406                                                   NVME_NIDT_UUID_LEN,
 407                                                   &ns->uuid, &off);
 408                 if (status)
 409                         goto out_put_ns;
 410         }
 411         if (memchr_inv(ns->nguid, 0, sizeof(ns->nguid))) {
 412                 status = nvmet_copy_ns_identifier(req, NVME_NIDT_NGUID,
 413                                                   NVME_NIDT_NGUID_LEN,
 414                                                   &ns->nguid, &off);
 415                 if (status)
 416                         goto out_put_ns;
 417         }
 418
 419         if (sg_zero_buffer(req->sg, req->sg_cnt, NVME_IDENTIFY_DATA_SIZE - off,
 420                         off) != NVME_IDENTIFY_DATA_SIZE - off)
 421                 status = NVME_SC_INTERNAL | NVME_SC_DNR;
 422 out_put_ns:
 423         nvmet_put_namespace(ns);
 424 out:
 425         nvmet_req_complete(req, status);
 426 }
 427
 428 /*
 429  * A "mimimum viable" abort implementation: the command is mandatory in the
 430  * spec, but we are not required to do any useful work.  We couldn't really
 431  * do a useful abort, so don't bother even with waiting for the command
 432  * to be exectuted and return immediately telling the command to abort
 433  * wasn't found.
 434  */
 435 static void nvmet_execute_abort(struct nvmet_req *req)
 436 {
 437         nvmet_set_result(req, 1);
 438         nvmet_req_complete(req, 0);
 439 }
 440
 441 static void nvmet_execute_set_features(struct nvmet_req *req)
 442 {
 443         struct nvmet_subsys *subsys = req->sq->ctrl->subsys;
 444         u32 cdw10 = le32_to_cpu(req->cmd->common.cdw10[0]);
 445         u32 val32;
 446         u16 status = 0;
 447
 448         switch (cdw10 & 0xf) {
 449         case NVME_FEAT_NUM_QUEUES:
 450                 nvmet_set_result(req,
 451                         (subsys->max_qid - 1) | ((subsys->max_qid - 1) << 16));
 452                 break;
 453         case NVME_FEAT_KATO:
 454                 val32 = le32_to_cpu(req->cmd->common.cdw10[1]);
 455                 req->sq->ctrl->kato = DIV_ROUND_UP(val32, 1000);
 456                 nvmet_set_result(req, req->sq->ctrl->kato);
 457                 break;
 458         default:
 459                 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 460                 break;
 461         }
 462
 463         nvmet_req_complete(req, status);
 464 }
 465
 466 static void nvmet_execute_get_features(struct nvmet_req *req)
 467 {
 468         struct nvmet_subsys *subsys = req->sq->ctrl->subsys;
 469         u32 cdw10 = le32_to_cpu(req->cmd->common.cdw10[0]);
 470         u16 status = 0;
 471
 472         switch (cdw10 & 0xf) {
 473         /*
 474          * These features are mandatory in the spec, but we don't
 475          * have a useful way to implement them.  We'll eventually
 476          * need to come up with some fake values for these.
 477          */
 478 #if 0
 479         case NVME_FEAT_ARBITRATION:
 480                 break;
 481         case NVME_FEAT_POWER_MGMT:
 482                 break;
 483         case NVME_FEAT_TEMP_THRESH:
 484                 break;
 485         case NVME_FEAT_ERR_RECOVERY:
 486                 break;
 487         case NVME_FEAT_IRQ_COALESCE:
 488                 break;
 489         case NVME_FEAT_IRQ_CONFIG:
 490                 break;
 491         case NVME_FEAT_WRITE_ATOMIC:
 492                 break;
 493         case NVME_FEAT_ASYNC_EVENT:
 494                 break;
 495 #endif
 496         case NVME_FEAT_VOLATILE_WC:
 497                 nvmet_set_result(req, 1);
 498                 break;
 499         case NVME_FEAT_NUM_QUEUES:
 500                 nvmet_set_result(req,
 501                         (subsys->max_qid-1) | ((subsys->max_qid-1) << 16));
 502                 break;
 503         case NVME_FEAT_KATO:
 504                 nvmet_set_result(req, req->sq->ctrl->kato * 1000);
 505                 break;
 506         default:
 507                 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 508                 break;
 509         }
 510
 511         nvmet_req_complete(req, status);
 512 }
 513
 514 static void nvmet_execute_async_event(struct nvmet_req *req)
 515 {
 516         struct nvmet_ctrl *ctrl = req->sq->ctrl;
 517
 518         mutex_lock(&ctrl->lock);
 519         if (ctrl->nr_async_event_cmds >= NVMET_ASYNC_EVENTS) {
 520                 mutex_unlock(&ctrl->lock);
 521                 nvmet_req_complete(req, NVME_SC_ASYNC_LIMIT | NVME_SC_DNR);
 522                 return;
 523         }
 524         ctrl->async_event_cmds[ctrl->nr_async_event_cmds++] = req;
 525         mutex_unlock(&ctrl->lock);
 526
 527         schedule_work(&ctrl->async_event_work);
 528 }
 529
 530 static void nvmet_execute_keep_alive(struct nvmet_req *req)
 531 {
 532         struct nvmet_ctrl *ctrl = req->sq->ctrl;
 533
 534         pr_debug("ctrl %d update keep-alive timer for %d secs\n",
 535                 ctrl->cntlid, ctrl->kato);
 536
 537         mod_delayed_work(system_wq, &ctrl->ka_work, ctrl->kato * HZ);
 538         nvmet_req_complete(req, 0);
 539 }
 540
 541 u16 nvmet_parse_admin_cmd(struct nvmet_req *req)
 542 {
 543         struct nvme_command *cmd = req->cmd;
 544         u16 ret;
 545
 546         req->ns = NULL;
 547
 548         ret = nvmet_check_ctrl_status(req, cmd);
 549         if (unlikely(ret))
 550                 return ret;
 551
 552         switch (cmd->common.opcode) {
 553         case nvme_admin_get_log_page:
 554                 req->data_len = nvmet_get_log_page_len(cmd);
 555
 556                 switch (cmd->get_log_page.lid) {
 557                 case NVME_LOG_ERROR:
 558                 case NVME_LOG_SMART:
 559                 case NVME_LOG_FW_SLOT:
 560                         req->execute = nvmet_execute_get_log_page;
 561                         return 0;
 562                 }
 563                 break;
 564         case nvme_admin_identify:
 565                 req->data_len = NVME_IDENTIFY_DATA_SIZE;
 566                 switch (cmd->identify.cns) {
 567                 case NVME_ID_CNS_NS:
 568                         req->execute = nvmet_execute_identify_ns;
 569                         return 0;
 570                 case NVME_ID_CNS_CTRL:
 571                         req->execute = nvmet_execute_identify_ctrl;
 572                         return 0;
 573                 case NVME_ID_CNS_NS_ACTIVE_LIST:
 574                         req->execute = nvmet_execute_identify_nslist;
 575                         return 0;
 576                 case NVME_ID_CNS_NS_DESC_LIST:
 577                         req->execute = nvmet_execute_identify_desclist;
 578                         return 0;
 579                 }
 580                 break;
 581         case nvme_admin_abort_cmd:
 582                 req->execute = nvmet_execute_abort;
 583                 req->data_len = 0;
 584                 return 0;
 585         case nvme_admin_set_features:
 586                 req->execute = nvmet_execute_set_features;
 587                 req->data_len = 0;
 588                 return 0;
 589         case nvme_admin_get_features:
 590                 req->execute = nvmet_execute_get_features;
 591                 req->data_len = 0;
 592                 return 0;
 593         case nvme_admin_async_event:
 594                 req->execute = nvmet_execute_async_event;
 595                 req->data_len = 0;
 596                 return 0;
 597         case nvme_admin_keep_alive:
 598                 req->execute = nvmet_execute_keep_alive;
 599                 req->data_len = 0;
 600                 return 0;
 601         }
 602
 603         pr_err("unhandled cmd %d on qid %d\n", cmd->common.opcode,
 604                req->sq->qid);
 605         return NVME_SC_INVALID_OPCODE | NVME_SC_DNR;
 606 }