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         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_PAGES as well
  54          */
  55         max_hw_sectors = min_not_zero(BIO_MAX_PAGES << (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         u16 status;
 157
 158         nvme_execute_passthru_rq(rq);
 159
 160         status = nvme_req(rq)->status;
 161         if (status == NVME_SC_SUCCESS &&
 162             req->cmd->common.opcode == nvme_admin_identify) {
 163                 switch (req->cmd->identify.cns) {
 164                 case NVME_ID_CNS_CTRL:
 165                         nvmet_passthru_override_id_ctrl(req);
 166                         break;
 167                 case NVME_ID_CNS_NS:
 168                         nvmet_passthru_override_id_ns(req);
 169                         break;
 170                 }
 171         }
 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         int op_flags = 0;
 192         struct bio *bio;
 193         int i, ret;
 194
 195         if (req->sg_cnt > BIO_MAX_PAGES)
 196                 return -EINVAL;
 197
 198         if (req->cmd->common.opcode == nvme_cmd_flush)
 199                 op_flags = REQ_FUA;
 200         else if (nvme_is_write(req->cmd))
 201                 op_flags = REQ_SYNC | REQ_IDLE;
 202
 203         bio = bio_alloc(GFP_KERNEL, req->sg_cnt);
 204         bio->bi_end_io = bio_put;
 205         bio->bi_opf = req_op(rq) | op_flags;
 206
 207         for_each_sg(req->sg, sg, req->sg_cnt, i) {
 208                 if (bio_add_pc_page(rq->q, bio, sg_page(sg), sg->length,
 209                                     sg->offset) < sg->length) {
 210                         bio_put(bio);
 211                         return -EINVAL;
 212                 }
 213         }
 214
 215         ret = blk_rq_append_bio(rq, &bio);
 216         if (unlikely(ret)) {
 217                 bio_put(bio);
 218                 return ret;
 219         }
 220
 221         return 0;
 222 }
 223
 224 static void nvmet_passthru_execute_cmd(struct nvmet_req *req)
 225 {
 226         struct nvme_ctrl *ctrl = nvmet_req_passthru_ctrl(req);
 227         struct request_queue *q = ctrl->admin_q;
 228         struct nvme_ns *ns = NULL;
 229         struct request *rq = NULL;
 230         u32 effects;
 231         u16 status;
 232         int ret;
 233
 234         if (likely(req->sq->qid != 0)) {
 235                 u32 nsid = le32_to_cpu(req->cmd->common.nsid);
 236
 237                 ns = nvme_find_get_ns(ctrl, nsid);
 238                 if (unlikely(!ns)) {
 239                         pr_err("failed to get passthru ns nsid:%u\n", nsid);
 240                         status = NVME_SC_INVALID_NS | NVME_SC_DNR;
 241                         goto out;
 242                 }
 243
 244                 q = ns->queue;
 245         }
 246
 247         rq = nvme_alloc_request(q, req->cmd, 0);
 248         if (IS_ERR(rq)) {
 249                 status = NVME_SC_INTERNAL;
 250                 goto out_put_ns;
 251         }
 252
 253         if (req->sg_cnt) {
 254                 ret = nvmet_passthru_map_sg(req, rq);
 255                 if (unlikely(ret)) {
 256                         status = NVME_SC_INTERNAL;
 257                         goto out_put_req;
 258                 }
 259         }
 260
 261         /*
 262          * If there are effects for the command we are about to execute, or
 263          * an end_req function we need to use nvme_execute_passthru_rq()
 264          * synchronously in a work item seeing the end_req function and
 265          * nvme_passthru_end() can't be called in the request done callback
 266          * which is typically in interrupt context.
 267          */
 268         effects = nvme_command_effects(ctrl, ns, req->cmd->common.opcode);
 269         if (req->p.use_workqueue || effects) {
 270                 INIT_WORK(&req->p.work, nvmet_passthru_execute_cmd_work);
 271                 req->p.rq = rq;
 272                 schedule_work(&req->p.work);
 273         } else {
 274                 rq->end_io_data = req;
 275                 blk_execute_rq_nowait(rq->q, ns ? ns->disk : NULL, rq, 0,
 276                                       nvmet_passthru_req_done);
 277         }
 278
 279         if (ns)
 280                 nvme_put_ns(ns);
 281
 282         return;
 283
 284 out_put_req:
 285         blk_mq_free_request(rq);
 286 out_put_ns:
 287         if (ns)
 288                 nvme_put_ns(ns);
 289 out:
 290         nvmet_req_complete(req, status);
 291 }
 292
 293 /*
 294  * We need to emulate set host behaviour to ensure that any requested
 295  * behaviour of the target's host matches the requested behaviour
 296  * of the device's host and fail otherwise.
 297  */
 298 static void nvmet_passthru_set_host_behaviour(struct nvmet_req *req)
 299 {
 300         struct nvme_ctrl *ctrl = nvmet_req_passthru_ctrl(req);
 301         struct nvme_feat_host_behavior *host;
 302         u16 status = NVME_SC_INTERNAL;
 303         int ret;
 304
 305         host = kzalloc(sizeof(*host) * 2, GFP_KERNEL);
 306         if (!host)
 307                 goto out_complete_req;
 308
 309         ret = nvme_get_features(ctrl, NVME_FEAT_HOST_BEHAVIOR, 0,
 310                                 host, sizeof(*host), NULL);
 311         if (ret)
 312                 goto out_free_host;
 313
 314         status = nvmet_copy_from_sgl(req, 0, &host[1], sizeof(*host));
 315         if (status)
 316                 goto out_free_host;
 317
 318         if (memcmp(&host[0], &host[1], sizeof(host[0]))) {
 319                 pr_warn("target host has requested different behaviour from the local host\n");
 320                 status = NVME_SC_INTERNAL;
 321         }
 322
 323 out_free_host:
 324         kfree(host);
 325 out_complete_req:
 326         nvmet_req_complete(req, status);
 327 }
 328
 329 static u16 nvmet_setup_passthru_command(struct nvmet_req *req)
 330 {
 331         req->p.use_workqueue = false;
 332         req->execute = nvmet_passthru_execute_cmd;
 333         return NVME_SC_SUCCESS;
 334 }
 335
 336 u16 nvmet_parse_passthru_io_cmd(struct nvmet_req *req)
 337 {
 338         /* Reject any commands with non-sgl flags set (ie. fused commands) */
 339         if (req->cmd->common.flags & ~NVME_CMD_SGL_ALL)
 340                 return NVME_SC_INVALID_FIELD;
 341
 342         switch (req->cmd->common.opcode) {
 343         case nvme_cmd_resv_register:
 344         case nvme_cmd_resv_report:
 345         case nvme_cmd_resv_acquire:
 346         case nvme_cmd_resv_release:
 347                 /*
 348                  * Reservations cannot be supported properly because the
 349                  * underlying device has no way of differentiating different
 350                  * hosts that connect via fabrics. This could potentially be
 351                  * emulated in the future if regular targets grow support for
 352                  * this feature.
 353                  */
 354                 return NVME_SC_INVALID_OPCODE | NVME_SC_DNR;
 355         }
 356
 357         return nvmet_setup_passthru_command(req);
 358 }
 359
 360 /*
 361  * Only features that are emulated or specifically allowed in the list  are
 362  * passed down to the controller. This function implements the allow list for
 363  * both get and set features.
 364  */
 365 static u16 nvmet_passthru_get_set_features(struct nvmet_req *req)
 366 {
 367         switch (le32_to_cpu(req->cmd->features.fid)) {
 368         case NVME_FEAT_ARBITRATION:
 369         case NVME_FEAT_POWER_MGMT:
 370         case NVME_FEAT_LBA_RANGE:
 371         case NVME_FEAT_TEMP_THRESH:
 372         case NVME_FEAT_ERR_RECOVERY:
 373         case NVME_FEAT_VOLATILE_WC:
 374         case NVME_FEAT_WRITE_ATOMIC:
 375         case NVME_FEAT_AUTO_PST:
 376         case NVME_FEAT_TIMESTAMP:
 377         case NVME_FEAT_HCTM:
 378         case NVME_FEAT_NOPSC:
 379         case NVME_FEAT_RRL:
 380         case NVME_FEAT_PLM_CONFIG:
 381         case NVME_FEAT_PLM_WINDOW:
 382         case NVME_FEAT_HOST_BEHAVIOR:
 383         case NVME_FEAT_SANITIZE:
 384         case NVME_FEAT_VENDOR_START ... NVME_FEAT_VENDOR_END:
 385                 return nvmet_setup_passthru_command(req);
 386
 387         case NVME_FEAT_ASYNC_EVENT:
 388                 /* There is no support for forwarding ASYNC events */
 389         case NVME_FEAT_IRQ_COALESCE:
 390         case NVME_FEAT_IRQ_CONFIG:
 391                 /* The IRQ settings will not apply to the target controller */
 392         case NVME_FEAT_HOST_MEM_BUF:
 393                 /*
 394                  * Any HMB that's set will not be passed through and will
 395                  * not work as expected
 396                  */
 397         case NVME_FEAT_SW_PROGRESS:
 398                 /*
 399                  * The Pre-Boot Software Load Count doesn't make much
 400                  * sense for a target to export
 401                  */
 402         case NVME_FEAT_RESV_MASK:
 403         case NVME_FEAT_RESV_PERSIST:
 404                 /* No reservations, see nvmet_parse_passthru_io_cmd() */
 405         default:
 406                 return NVME_SC_INVALID_OPCODE | NVME_SC_DNR;
 407         }
 408 }
 409
 410 u16 nvmet_parse_passthru_admin_cmd(struct nvmet_req *req)
 411 {
 412         /* Reject any commands with non-sgl flags set (ie. fused commands) */
 413         if (req->cmd->common.flags & ~NVME_CMD_SGL_ALL)
 414                 return NVME_SC_INVALID_FIELD;
 415
 416         /*
 417          * Passthru all vendor specific commands
 418          */
 419         if (req->cmd->common.opcode >= nvme_admin_vendor_start)
 420                 return nvmet_setup_passthru_command(req);
 421
 422         switch (req->cmd->common.opcode) {
 423         case nvme_admin_async_event:
 424                 req->execute = nvmet_execute_async_event;
 425                 return NVME_SC_SUCCESS;
 426         case nvme_admin_keep_alive:
 427                 /*
 428                  * Most PCIe ctrls don't support keep alive cmd, we route keep
 429                  * alive to the non-passthru mode. In future please change this
 430                  * code when PCIe ctrls with keep alive support available.
 431                  */
 432                 req->execute = nvmet_execute_keep_alive;
 433                 return NVME_SC_SUCCESS;
 434         case nvme_admin_set_features:
 435                 switch (le32_to_cpu(req->cmd->features.fid)) {
 436                 case NVME_FEAT_ASYNC_EVENT:
 437                 case NVME_FEAT_KATO:
 438                 case NVME_FEAT_NUM_QUEUES:
 439                 case NVME_FEAT_HOST_ID:
 440                         req->execute = nvmet_execute_set_features;
 441                         return NVME_SC_SUCCESS;
 442                 case NVME_FEAT_HOST_BEHAVIOR:
 443                         req->execute = nvmet_passthru_set_host_behaviour;
 444                         return NVME_SC_SUCCESS;
 445                 default:
 446                         return nvmet_passthru_get_set_features(req);
 447                 }
 448                 break;
 449         case nvme_admin_get_features:
 450                 switch (le32_to_cpu(req->cmd->features.fid)) {
 451                 case NVME_FEAT_ASYNC_EVENT:
 452                 case NVME_FEAT_KATO:
 453                 case NVME_FEAT_NUM_QUEUES:
 454                 case NVME_FEAT_HOST_ID:
 455                         req->execute = nvmet_execute_get_features;
 456                         return NVME_SC_SUCCESS;
 457                 default:
 458                         return nvmet_passthru_get_set_features(req);
 459                 }
 460                 break;
 461         case nvme_admin_identify:
 462                 switch (req->cmd->identify.cns) {
 463                 case NVME_ID_CNS_CTRL:
 464                         req->execute = nvmet_passthru_execute_cmd;
 465                         req->p.use_workqueue = true;
 466                         return NVME_SC_SUCCESS;
 467                 case NVME_ID_CNS_CS_CTRL:
 468                         switch (req->cmd->identify.csi) {
 469                         case NVME_CSI_ZNS:
 470                                 req->execute = nvmet_passthru_execute_cmd;
 471                                 req->p.use_workqueue = true;
 472                                 return NVME_SC_SUCCESS;
 473                         }
 474                         return NVME_SC_INVALID_OPCODE | NVME_SC_DNR;
 475                 case NVME_ID_CNS_NS:
 476                         req->execute = nvmet_passthru_execute_cmd;
 477                         req->p.use_workqueue = true;
 478                         return NVME_SC_SUCCESS;
 479                 case NVME_ID_CNS_CS_NS:
 480                         switch (req->cmd->identify.csi) {
 481                         case NVME_CSI_ZNS:
 482                                 req->execute = nvmet_passthru_execute_cmd;
 483                                 req->p.use_workqueue = true;
 484                                 return NVME_SC_SUCCESS;
 485                         }
 486                         return NVME_SC_INVALID_OPCODE | NVME_SC_DNR;
 487                 default:
 488                         return nvmet_setup_passthru_command(req);
 489                 }
 490         case nvme_admin_get_log_page:
 491                 return nvmet_setup_passthru_command(req);
 492         default:
 493                 /* Reject commands not in the allowlist above */
 494                 return NVME_SC_INVALID_OPCODE | NVME_SC_DNR;
 495         }
 496 }
 497
 498 int nvmet_passthru_ctrl_enable(struct nvmet_subsys *subsys)
 499 {
 500         struct nvme_ctrl *ctrl;
 501         struct file *file;
 502         int ret = -EINVAL;
 503         void *old;
 504
 505         mutex_lock(&subsys->lock);
 506         if (!subsys->passthru_ctrl_path)
 507                 goto out_unlock;
 508         if (subsys->passthru_ctrl)
 509                 goto out_unlock;
 510
 511         if (subsys->nr_namespaces) {
 512                 pr_info("cannot enable both passthru and regular namespaces for a single subsystem");
 513                 goto out_unlock;
 514         }
 515
 516         file = filp_open(subsys->passthru_ctrl_path, O_RDWR, 0);
 517         if (IS_ERR(file)) {
 518                 ret = PTR_ERR(file);
 519                 goto out_unlock;
 520         }
 521
 522         ctrl = nvme_ctrl_from_file(file);
 523         if (!ctrl) {
 524                 pr_err("failed to open nvme controller %s\n",
 525                        subsys->passthru_ctrl_path);
 526
 527                 goto out_put_file;
 528         }
 529
 530         old = xa_cmpxchg(&passthru_subsystems, ctrl->cntlid, NULL,
 531                          subsys, GFP_KERNEL);
 532         if (xa_is_err(old)) {
 533                 ret = xa_err(old);
 534                 goto out_put_file;
 535         }
 536
 537         if (old)
 538                 goto out_put_file;
 539
 540         subsys->passthru_ctrl = ctrl;
 541         subsys->ver = ctrl->vs;
 542
 543         if (subsys->ver < NVME_VS(1, 2, 1)) {
 544                 pr_warn("nvme controller version is too old: %llu.%llu.%llu, advertising 1.2.1\n",
 545                         NVME_MAJOR(subsys->ver), NVME_MINOR(subsys->ver),
 546                         NVME_TERTIARY(subsys->ver));
 547                 subsys->ver = NVME_VS(1, 2, 1);
 548         }
 549         nvme_get_ctrl(ctrl);
 550         __module_get(subsys->passthru_ctrl->ops->module);
 551         ret = 0;
 552
 553 out_put_file:
 554         filp_close(file, NULL);
 555 out_unlock:
 556         mutex_unlock(&subsys->lock);
 557         return ret;
 558 }
 559
 560 static void __nvmet_passthru_ctrl_disable(struct nvmet_subsys *subsys)
 561 {
 562         if (subsys->passthru_ctrl) {
 563                 xa_erase(&passthru_subsystems, subsys->passthru_ctrl->cntlid);
 564                 module_put(subsys->passthru_ctrl->ops->module);
 565                 nvme_put_ctrl(subsys->passthru_ctrl);
 566         }
 567         subsys->passthru_ctrl = NULL;
 568         subsys->ver = NVMET_DEFAULT_VS;
 569 }
 570
 571 void nvmet_passthru_ctrl_disable(struct nvmet_subsys *subsys)
 572 {
 573         mutex_lock(&subsys->lock);
 574         __nvmet_passthru_ctrl_disable(subsys);
 575         mutex_unlock(&subsys->lock);
 576 }
 577
 578 void nvmet_passthru_subsys_free(struct nvmet_subsys *subsys)
 579 {
 580         mutex_lock(&subsys->lock);
 581         __nvmet_passthru_ctrl_disable(subsys);
 582         mutex_unlock(&subsys->lock);
 583         kfree(subsys->passthru_ctrl_path);
 584 }