drivers/nvme/target/passthru.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * NVMe Over Fabrics Target Passthrough command implementation.
   4  *
   5  * Copyright (c) 2017-2018 Western Digital Corporation or its
   6  * affiliates.
   7  * Copyright (c) 2019-2020, Eideticom Inc.
   8  *
   9  */
  10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  11 #include <linux/module.h>
  12
  13 #include "../host/nvme.h"
  14 #include "nvmet.h"
  15
  16 MODULE_IMPORT_NS(NVME_TARGET_PASSTHRU);
  17
  18 /*
  19  * xarray to maintain one passthru subsystem per nvme controller.
  20  */
  21 static DEFINE_XARRAY(passthru_subsystems);
  22
  23 static u16 nvmet_passthru_override_id_ctrl(struct nvmet_req *req)
  24 {
  25         struct nvmet_ctrl *ctrl = req->sq->ctrl;
  26         struct nvme_ctrl *pctrl = ctrl->subsys->passthru_ctrl;
  27         u16 status = NVME_SC_SUCCESS;
  28         struct nvme_id_ctrl *id;
  29         unsigned int max_hw_sectors;
  30         int page_shift;
  31
  32         id = kzalloc(sizeof(*id), GFP_KERNEL);
  33         if (!id)
  34                 return NVME_SC_INTERNAL;
  35
  36         status = nvmet_copy_from_sgl(req, 0, id, sizeof(*id));
  37         if (status)
  38                 goto out_free;
  39
  40         id->cntlid = cpu_to_le16(ctrl->cntlid);
  41         id->ver = cpu_to_le32(ctrl->subsys->ver);
  42
  43         /*
  44          * The passthru NVMe driver may have a limit on the number of segments
  45          * which depends on the host's memory fragementation. To solve this,
  46          * ensure mdts is limited to the pages equal to the number of segments.
  47          */
  48         max_hw_sectors = min_not_zero(pctrl->max_segments << (PAGE_SHIFT - 9),
  49                                       pctrl->max_hw_sectors);
  50
  51         /*
  52          * nvmet_passthru_map_sg is limitted to using a single bio so limit
  53          * the mdts based on BIO_MAX_VECS as well
  54          */
  55         max_hw_sectors = min_not_zero(BIO_MAX_VECS << (PAGE_SHIFT - 9),
  56                                       max_hw_sectors);
  57
  58         page_shift = NVME_CAP_MPSMIN(ctrl->cap) + 12;
  59
  60         id->mdts = ilog2(max_hw_sectors) + 9 - page_shift;
  61
  62         id->acl = 3;
  63         /*
  64          * We export aerl limit for the fabrics controller, update this when
  65          * passthru based aerl support is added.
  66          */
  67         id->aerl = NVMET_ASYNC_EVENTS - 1;
  68
  69         /* emulate kas as most of the PCIe ctrl don't have a support for kas */
  70         id->kas = cpu_to_le16(NVMET_KAS);
  71
  72         /* don't support host memory buffer */
  73         id->hmpre = 0;
  74         id->hmmin = 0;
  75
  76         id->sqes = min_t(__u8, ((0x6 << 4) | 0x6), id->sqes);
  77         id->cqes = min_t(__u8, ((0x4 << 4) | 0x4), id->cqes);
  78         id->maxcmd = cpu_to_le16(NVMET_MAX_CMD);
  79
  80         /* don't support fuse commands */
  81         id->fuses = 0;
  82
  83         id->sgls = cpu_to_le32(1 << 0); /* we always support SGLs */
  84         if (ctrl->ops->flags & NVMF_KEYED_SGLS)
  85                 id->sgls |= cpu_to_le32(1 << 2);
  86         if (req->port->inline_data_size)
  87                 id->sgls |= cpu_to_le32(1 << 20);
  88
  89         /*
  90          * When passsthru controller is setup using nvme-loop transport it will
  91          * export the passthru ctrl subsysnqn (PCIe NVMe ctrl) and will fail in
  92          * the nvme/host/core.c in the nvme_init_subsystem()->nvme_active_ctrl()
  93          * code path with duplicate ctr subsynqn. In order to prevent that we
  94          * mask the passthru-ctrl subsysnqn with the target ctrl subsysnqn.
  95          */
  96         memcpy(id->subnqn, ctrl->subsysnqn, sizeof(id->subnqn));
  97
  98         /* use fabric id-ctrl values */
  99         id->ioccsz = cpu_to_le32((sizeof(struct nvme_command) +
 100                                 req->port->inline_data_size) / 16);
 101         id->iorcsz = cpu_to_le32(sizeof(struct nvme_completion) / 16);
 102
 103         id->msdbd = ctrl->ops->msdbd;
 104
 105         /* Support multipath connections with fabrics */
 106         id->cmic |= 1 << 1;
 107
 108         /* Disable reservations, see nvmet_parse_passthru_io_cmd() */
 109         id->oncs &= cpu_to_le16(~NVME_CTRL_ONCS_RESERVATIONS);
 110
 111         status = nvmet_copy_to_sgl(req, 0, id, sizeof(struct nvme_id_ctrl));
 112
 113 out_free:
 114         kfree(id);
 115         return status;
 116 }
 117
 118 static u16 nvmet_passthru_override_id_ns(struct nvmet_req *req)
 119 {
 120         u16 status = NVME_SC_SUCCESS;
 121         struct nvme_id_ns *id;
 122         int i;
 123
 124         id = kzalloc(sizeof(*id), GFP_KERNEL);
 125         if (!id)
 126                 return NVME_SC_INTERNAL;
 127
 128         status = nvmet_copy_from_sgl(req, 0, id, sizeof(struct nvme_id_ns));
 129         if (status)
 130                 goto out_free;
 131
 132         for (i = 0; i < (id->nlbaf + 1); i++)
 133                 if (id->lbaf[i].ms)
 134                         memset(&id->lbaf[i], 0, sizeof(id->lbaf[i]));
 135
 136         id->flbas = id->flbas & ~(1 << 4);
 137
 138         /*
 139          * Presently the NVMEof target code does not support sending
 140          * metadata, so we must disable it here. This should be updated
 141          * once target starts supporting metadata.
 142          */
 143         id->mc = 0;
 144
 145         status = nvmet_copy_to_sgl(req, 0, id, sizeof(*id));
 146
 147 out_free:
 148         kfree(id);
 149         return status;
 150 }
 151
 152 static void nvmet_passthru_execute_cmd_work(struct work_struct *w)
 153 {
 154         struct nvmet_req *req = container_of(w, struct nvmet_req, p.work);
 155         struct request *rq = req->p.rq;
 156         int status;
 157
 158         status = nvme_execute_passthru_rq(rq);
 159
 160         if (status == NVME_SC_SUCCESS &&
 161             req->cmd->common.opcode == nvme_admin_identify) {
 162                 switch (req->cmd->identify.cns) {
 163                 case NVME_ID_CNS_CTRL:
 164                         nvmet_passthru_override_id_ctrl(req);
 165                         break;
 166                 case NVME_ID_CNS_NS:
 167                         nvmet_passthru_override_id_ns(req);
 168                         break;
 169                 }
 170         } else if (status < 0)
 171                 status = NVME_SC_INTERNAL;
 172
 173         req->cqe->result = nvme_req(rq)->result;
 174         nvmet_req_complete(req, status);
 175         blk_mq_free_request(rq);
 176 }
 177
 178 static void nvmet_passthru_req_done(struct request *rq,
 179                                     blk_status_t blk_status)
 180 {
 181         struct nvmet_req *req = rq->end_io_data;
 182
 183         req->cqe->result = nvme_req(rq)->result;
 184         nvmet_req_complete(req, nvme_req(rq)->status);
 185         blk_mq_free_request(rq);
 186 }
 187
 188 static int nvmet_passthru_map_sg(struct nvmet_req *req, struct request *rq)
 189 {
 190         struct scatterlist *sg;
 191         struct bio *bio;
 192         int i;
 193
 194         if (req->sg_cnt > BIO_MAX_VECS)
 195                 return -EINVAL;
 196
 197         if (nvmet_use_inline_bvec(req)) {
 198                 bio = &req->p.inline_bio;
 199                 bio_init(bio, req->inline_bvec, ARRAY_SIZE(req->inline_bvec));
 200         } else {
 201                 bio = bio_alloc(GFP_KERNEL, bio_max_segs(req->sg_cnt));
 202                 bio->bi_end_io = bio_put;
 203         }
 204         bio->bi_opf = req_op(rq);
 205
 206         for_each_sg(req->sg, sg, req->sg_cnt, i) {
 207                 if (bio_add_pc_page(rq->q, bio, sg_page(sg), sg->length,
 208                                     sg->offset) < sg->length) {
 209                         nvmet_req_bio_put(req, bio);
 210                         return -EINVAL;
 211                 }
 212         }
 213
 214         blk_rq_bio_prep(rq, bio, req->sg_cnt);
 215
 216         return 0;
 217 }
 218
 219 static void nvmet_passthru_execute_cmd(struct nvmet_req *req)
 220 {
 221         struct nvme_ctrl *ctrl = nvmet_req_passthru_ctrl(req);
 222         struct request_queue *q = ctrl->admin_q;
 223         struct nvme_ns *ns = NULL;
 224         struct request *rq = NULL;
 225         unsigned int timeout;
 226         u32 effects;
 227         u16 status;
 228         int ret;
 229
 230         if (likely(req->sq->qid != 0)) {
 231                 u32 nsid = le32_to_cpu(req->cmd->common.nsid);
 232
 233                 ns = nvme_find_get_ns(ctrl, nsid);
 234                 if (unlikely(!ns)) {
 235                         pr_err("failed to get passthru ns nsid:%u\n", nsid);
 236                         status = NVME_SC_INVALID_NS | NVME_SC_DNR;
 237                         goto out;
 238                 }
 239
 240                 q = ns->queue;
 241                 timeout = nvmet_req_subsys(req)->io_timeout;
 242         } else {
 243                 timeout = nvmet_req_subsys(req)->admin_timeout;
 244         }
 245
 246         rq = nvme_alloc_request(q, req->cmd, 0);
 247         if (IS_ERR(rq)) {
 248                 status = NVME_SC_INTERNAL;
 249                 goto out_put_ns;
 250         }
 251
 252         if (timeout)
 253                 rq->timeout = timeout;
 254
 255         if (req->sg_cnt) {
 256                 ret = nvmet_passthru_map_sg(req, rq);
 257                 if (unlikely(ret)) {
 258                         status = NVME_SC_INTERNAL;
 259                         goto out_put_req;
 260                 }
 261         }
 262
 263         /*
 264          * If there are effects for the command we are about to execute, or
 265          * an end_req function we need to use nvme_execute_passthru_rq()
 266          * synchronously in a work item seeing the end_req function and
 267          * nvme_passthru_end() can't be called in the request done callback
 268          * which is typically in interrupt context.
 269          */
 270         effects = nvme_command_effects(ctrl, ns, req->cmd->common.opcode);
 271         if (req->p.use_workqueue || effects) {
 272                 INIT_WORK(&req->p.work, nvmet_passthru_execute_cmd_work);
 273                 req->p.rq = rq;
 274                 schedule_work(&req->p.work);
 275         } else {
 276                 rq->end_io_data = req;
 277                 blk_execute_rq_nowait(ns ? ns->disk : NULL, rq, 0,
 278                                       nvmet_passthru_req_done);
 279         }
 280
 281         if (ns)
 282                 nvme_put_ns(ns);
 283
 284         return;
 285
 286 out_put_req:
 287         blk_mq_free_request(rq);
 288 out_put_ns:
 289         if (ns)
 290                 nvme_put_ns(ns);
 291 out:
 292         nvmet_req_complete(req, status);
 293 }
 294
 295 /*
 296  * We need to emulate set host behaviour to ensure that any requested
 297  * behaviour of the target's host matches the requested behaviour
 298  * of the device's host and fail otherwise.
 299  */
 300 static void nvmet_passthru_set_host_behaviour(struct nvmet_req *req)
 301 {
 302         struct nvme_ctrl *ctrl = nvmet_req_passthru_ctrl(req);
 303         struct nvme_feat_host_behavior *host;
 304         u16 status = NVME_SC_INTERNAL;
 305         int ret;
 306
 307         host = kzalloc(sizeof(*host) * 2, GFP_KERNEL);
 308         if (!host)
 309                 goto out_complete_req;
 310
 311         ret = nvme_get_features(ctrl, NVME_FEAT_HOST_BEHAVIOR, 0,
 312                                 host, sizeof(*host), NULL);
 313         if (ret)
 314                 goto out_free_host;
 315
 316         status = nvmet_copy_from_sgl(req, 0, &host[1], sizeof(*host));
 317         if (status)
 318                 goto out_free_host;
 319
 320         if (memcmp(&host[0], &host[1], sizeof(host[0]))) {
 321                 pr_warn("target host has requested different behaviour from the local host\n");
 322                 status = NVME_SC_INTERNAL;
 323         }
 324
 325 out_free_host:
 326         kfree(host);
 327 out_complete_req:
 328         nvmet_req_complete(req, status);
 329 }
 330
 331 static u16 nvmet_setup_passthru_command(struct nvmet_req *req)
 332 {
 333         req->p.use_workqueue = false;
 334         req->execute = nvmet_passthru_execute_cmd;
 335         return NVME_SC_SUCCESS;
 336 }
 337
 338 u16 nvmet_parse_passthru_io_cmd(struct nvmet_req *req)
 339 {
 340         /* Reject any commands with non-sgl flags set (ie. fused commands) */
 341         if (req->cmd->common.flags & ~NVME_CMD_SGL_ALL)
 342                 return NVME_SC_INVALID_FIELD;
 343
 344         switch (req->cmd->common.opcode) {
 345         case nvme_cmd_resv_register:
 346         case nvme_cmd_resv_report:
 347         case nvme_cmd_resv_acquire:
 348         case nvme_cmd_resv_release:
 349                 /*
 350                  * Reservations cannot be supported properly because the
 351                  * underlying device has no way of differentiating different
 352                  * hosts that connect via fabrics. This could potentially be
 353                  * emulated in the future if regular targets grow support for
 354                  * this feature.
 355                  */
 356                 return NVME_SC_INVALID_OPCODE | NVME_SC_DNR;
 357         }
 358
 359         return nvmet_setup_passthru_command(req);
 360 }
 361
 362 /*
 363  * Only features that are emulated or specifically allowed in the list  are
 364  * passed down to the controller. This function implements the allow list for
 365  * both get and set features.
 366  */
 367 static u16 nvmet_passthru_get_set_features(struct nvmet_req *req)
 368 {
 369         switch (le32_to_cpu(req->cmd->features.fid)) {
 370         case NVME_FEAT_ARBITRATION:
 371         case NVME_FEAT_POWER_MGMT:
 372         case NVME_FEAT_LBA_RANGE:
 373         case NVME_FEAT_TEMP_THRESH:
 374         case NVME_FEAT_ERR_RECOVERY:
 375         case NVME_FEAT_VOLATILE_WC:
 376         case NVME_FEAT_WRITE_ATOMIC:
 377         case NVME_FEAT_AUTO_PST:
 378         case NVME_FEAT_TIMESTAMP:
 379         case NVME_FEAT_HCTM:
 380         case NVME_FEAT_NOPSC:
 381         case NVME_FEAT_RRL:
 382         case NVME_FEAT_PLM_CONFIG:
 383         case NVME_FEAT_PLM_WINDOW:
 384         case NVME_FEAT_HOST_BEHAVIOR:
 385         case NVME_FEAT_SANITIZE:
 386         case NVME_FEAT_VENDOR_START ... NVME_FEAT_VENDOR_END:
 387                 return nvmet_setup_passthru_command(req);
 388
 389         case NVME_FEAT_ASYNC_EVENT:
 390                 /* There is no support for forwarding ASYNC events */
 391         case NVME_FEAT_IRQ_COALESCE:
 392         case NVME_FEAT_IRQ_CONFIG:
 393                 /* The IRQ settings will not apply to the target controller */
 394         case NVME_FEAT_HOST_MEM_BUF:
 395                 /*
 396                  * Any HMB that's set will not be passed through and will
 397                  * not work as expected
 398                  */
 399         case NVME_FEAT_SW_PROGRESS:
 400                 /*
 401                  * The Pre-Boot Software Load Count doesn't make much
 402                  * sense for a target to export
 403                  */
 404         case NVME_FEAT_RESV_MASK:
 405         case NVME_FEAT_RESV_PERSIST:
 406                 /* No reservations, see nvmet_parse_passthru_io_cmd() */
 407         default:
 408                 return NVME_SC_INVALID_OPCODE | NVME_SC_DNR;
 409         }
 410 }
 411
 412 u16 nvmet_parse_passthru_admin_cmd(struct nvmet_req *req)
 413 {
 414         /* Reject any commands with non-sgl flags set (ie. fused commands) */
 415         if (req->cmd->common.flags & ~NVME_CMD_SGL_ALL)
 416                 return NVME_SC_INVALID_FIELD;
 417
 418         /*
 419          * Passthru all vendor specific commands
 420          */
 421         if (req->cmd->common.opcode >= nvme_admin_vendor_start)
 422                 return nvmet_setup_passthru_command(req);
 423
 424         switch (req->cmd->common.opcode) {
 425         case nvme_admin_async_event:
 426                 req->execute = nvmet_execute_async_event;
 427                 return NVME_SC_SUCCESS;
 428         case nvme_admin_keep_alive:
 429                 /*
 430                  * Most PCIe ctrls don't support keep alive cmd, we route keep
 431                  * alive to the non-passthru mode. In future please change this
 432                  * code when PCIe ctrls with keep alive support available.
 433                  */
 434                 req->execute = nvmet_execute_keep_alive;
 435                 return NVME_SC_SUCCESS;
 436         case nvme_admin_set_features:
 437                 switch (le32_to_cpu(req->cmd->features.fid)) {
 438                 case NVME_FEAT_ASYNC_EVENT:
 439                 case NVME_FEAT_KATO:
 440                 case NVME_FEAT_NUM_QUEUES:
 441                 case NVME_FEAT_HOST_ID:
 442                         req->execute = nvmet_execute_set_features;
 443                         return NVME_SC_SUCCESS;
 444                 case NVME_FEAT_HOST_BEHAVIOR:
 445                         req->execute = nvmet_passthru_set_host_behaviour;
 446                         return NVME_SC_SUCCESS;
 447                 default:
 448                         return nvmet_passthru_get_set_features(req);
 449                 }
 450                 break;
 451         case nvme_admin_get_features:
 452                 switch (le32_to_cpu(req->cmd->features.fid)) {
 453                 case NVME_FEAT_ASYNC_EVENT:
 454                 case NVME_FEAT_KATO:
 455                 case NVME_FEAT_NUM_QUEUES:
 456                 case NVME_FEAT_HOST_ID:
 457                         req->execute = nvmet_execute_get_features;
 458                         return NVME_SC_SUCCESS;
 459                 default:
 460                         return nvmet_passthru_get_set_features(req);
 461                 }
 462                 break;
 463         case nvme_admin_identify:
 464                 switch (req->cmd->identify.cns) {
 465                 case NVME_ID_CNS_CTRL:
 466                         req->execute = nvmet_passthru_execute_cmd;
 467                         req->p.use_workqueue = true;
 468                         return NVME_SC_SUCCESS;
 469                 case NVME_ID_CNS_CS_CTRL:
 470                         switch (req->cmd->identify.csi) {
 471                         case NVME_CSI_ZNS:
 472                                 req->execute = nvmet_passthru_execute_cmd;
 473                                 req->p.use_workqueue = true;
 474                                 return NVME_SC_SUCCESS;
 475                         }
 476                         return NVME_SC_INVALID_OPCODE | NVME_SC_DNR;
 477                 case NVME_ID_CNS_NS:
 478                         req->execute = nvmet_passthru_execute_cmd;
 479                         req->p.use_workqueue = true;
 480                         return NVME_SC_SUCCESS;
 481                 case NVME_ID_CNS_CS_NS:
 482                         switch (req->cmd->identify.csi) {
 483                         case NVME_CSI_ZNS:
 484                                 req->execute = nvmet_passthru_execute_cmd;
 485                                 req->p.use_workqueue = true;
 486                                 return NVME_SC_SUCCESS;
 487                         }
 488                         return NVME_SC_INVALID_OPCODE | NVME_SC_DNR;
 489                 default:
 490                         return nvmet_setup_passthru_command(req);
 491                 }
 492         case nvme_admin_get_log_page:
 493                 return nvmet_setup_passthru_command(req);
 494         default:
 495                 /* Reject commands not in the allowlist above */
 496                 return nvmet_report_invalid_opcode(req);
 497         }
 498 }
 499
 500 int nvmet_passthru_ctrl_enable(struct nvmet_subsys *subsys)
 501 {
 502         struct nvme_ctrl *ctrl;
 503         struct file *file;
 504         int ret = -EINVAL;
 505         void *old;
 506
 507         mutex_lock(&subsys->lock);
 508         if (!subsys->passthru_ctrl_path)
 509                 goto out_unlock;
 510         if (subsys->passthru_ctrl)
 511                 goto out_unlock;
 512
 513         if (subsys->nr_namespaces) {
 514                 pr_info("cannot enable both passthru and regular namespaces for a single subsystem");
 515                 goto out_unlock;
 516         }
 517
 518         file = filp_open(subsys->passthru_ctrl_path, O_RDWR, 0);
 519         if (IS_ERR(file)) {
 520                 ret = PTR_ERR(file);
 521                 goto out_unlock;
 522         }
 523
 524         ctrl = nvme_ctrl_from_file(file);
 525         if (!ctrl) {
 526                 pr_err("failed to open nvme controller %s\n",
 527                        subsys->passthru_ctrl_path);
 528
 529                 goto out_put_file;
 530         }
 531
 532         old = xa_cmpxchg(&passthru_subsystems, ctrl->cntlid, NULL,
 533                          subsys, GFP_KERNEL);
 534         if (xa_is_err(old)) {
 535                 ret = xa_err(old);
 536                 goto out_put_file;
 537         }
 538
 539         if (old)
 540                 goto out_put_file;
 541
 542         subsys->passthru_ctrl = ctrl;
 543         subsys->ver = ctrl->vs;
 544
 545         if (subsys->ver < NVME_VS(1, 2, 1)) {
 546                 pr_warn("nvme controller version is too old: %llu.%llu.%llu, advertising 1.2.1\n",
 547                         NVME_MAJOR(subsys->ver), NVME_MINOR(subsys->ver),
 548                         NVME_TERTIARY(subsys->ver));
 549                 subsys->ver = NVME_VS(1, 2, 1);
 550         }
 551         nvme_get_ctrl(ctrl);
 552         __module_get(subsys->passthru_ctrl->ops->module);
 553         ret = 0;
 554
 555 out_put_file:
 556         filp_close(file, NULL);
 557 out_unlock:
 558         mutex_unlock(&subsys->lock);
 559         return ret;
 560 }
 561
 562 static void __nvmet_passthru_ctrl_disable(struct nvmet_subsys *subsys)
 563 {
 564         if (subsys->passthru_ctrl) {
 565                 xa_erase(&passthru_subsystems, subsys->passthru_ctrl->cntlid);
 566                 module_put(subsys->passthru_ctrl->ops->module);
 567                 nvme_put_ctrl(subsys->passthru_ctrl);
 568         }
 569         subsys->passthru_ctrl = NULL;
 570         subsys->ver = NVMET_DEFAULT_VS;
 571 }
 572
 573 void nvmet_passthru_ctrl_disable(struct nvmet_subsys *subsys)
 574 {
 575         mutex_lock(&subsys->lock);
 576         __nvmet_passthru_ctrl_disable(subsys);
 577         mutex_unlock(&subsys->lock);
 578 }
 579
 580 void nvmet_passthru_subsys_free(struct nvmet_subsys *subsys)
 581 {
 582         mutex_lock(&subsys->lock);
 583         __nvmet_passthru_ctrl_disable(subsys);
 584         mutex_unlock(&subsys->lock);
 585         kfree(subsys->passthru_ctrl_path);
 586 }