drivers/nvme/host/core.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * NVM Express device driver
   4  * Copyright (c) 2011-2014, Intel Corporation.
   5  */
   6
   7 #include <linux/blkdev.h>
   8 #include <linux/blk-mq.h>
   9 #include <linux/blk-integrity.h>
  10 #include <linux/compat.h>
  11 #include <linux/delay.h>
  12 #include <linux/errno.h>
  13 #include <linux/hdreg.h>
  14 #include <linux/kernel.h>
  15 #include <linux/module.h>
  16 #include <linux/backing-dev.h>
  17 #include <linux/slab.h>
  18 #include <linux/types.h>
  19 #include <linux/pr.h>
  20 #include <linux/ptrace.h>
  21 #include <linux/nvme_ioctl.h>
  22 #include <linux/pm_qos.h>
  23 #include <asm/unaligned.h>
  24
  25 #include "nvme.h"
  26 #include "fabrics.h"
  27 #include <linux/nvme-auth.h>
  28
  29 #define CREATE_TRACE_POINTS
  30 #include "trace.h"
  31
  32 #define NVME_MINORS             (1U << MINORBITS)
  33
  34 struct nvme_ns_info {
  35         struct nvme_ns_ids ids;
  36         u32 nsid;
  37         __le32 anagrpid;
  38         bool is_shared;
  39         bool is_readonly;
  40         bool is_ready;
  41         bool is_removed;
  42 };
  43
  44 unsigned int admin_timeout = 60;
  45 module_param(admin_timeout, uint, 0644);
  46 MODULE_PARM_DESC(admin_timeout, "timeout in seconds for admin commands");
  47 EXPORT_SYMBOL_GPL(admin_timeout);
  48
  49 unsigned int nvme_io_timeout = 30;
  50 module_param_named(io_timeout, nvme_io_timeout, uint, 0644);
  51 MODULE_PARM_DESC(io_timeout, "timeout in seconds for I/O");
  52 EXPORT_SYMBOL_GPL(nvme_io_timeout);
  53
  54 static unsigned char shutdown_timeout = 5;
  55 module_param(shutdown_timeout, byte, 0644);
  56 MODULE_PARM_DESC(shutdown_timeout, "timeout in seconds for controller shutdown");
  57
  58 static u8 nvme_max_retries = 5;
  59 module_param_named(max_retries, nvme_max_retries, byte, 0644);
  60 MODULE_PARM_DESC(max_retries, "max number of retries a command may have");
  61
  62 static unsigned long default_ps_max_latency_us = 100000;
  63 module_param(default_ps_max_latency_us, ulong, 0644);
  64 MODULE_PARM_DESC(default_ps_max_latency_us,
  65                  "max power saving latency for new devices; use PM QOS to change per device");
  66
  67 static bool force_apst;
  68 module_param(force_apst, bool, 0644);
  69 MODULE_PARM_DESC(force_apst, "allow APST for newly enumerated devices even if quirked off");
  70
  71 static unsigned long apst_primary_timeout_ms = 100;
  72 module_param(apst_primary_timeout_ms, ulong, 0644);
  73 MODULE_PARM_DESC(apst_primary_timeout_ms,
  74         "primary APST timeout in ms");
  75
  76 static unsigned long apst_secondary_timeout_ms = 2000;
  77 module_param(apst_secondary_timeout_ms, ulong, 0644);
  78 MODULE_PARM_DESC(apst_secondary_timeout_ms,
  79         "secondary APST timeout in ms");
  80
  81 static unsigned long apst_primary_latency_tol_us = 15000;
  82 module_param(apst_primary_latency_tol_us, ulong, 0644);
  83 MODULE_PARM_DESC(apst_primary_latency_tol_us,
  84         "primary APST latency tolerance in us");
  85
  86 static unsigned long apst_secondary_latency_tol_us = 100000;
  87 module_param(apst_secondary_latency_tol_us, ulong, 0644);
  88 MODULE_PARM_DESC(apst_secondary_latency_tol_us,
  89         "secondary APST latency tolerance in us");
  90
  91 /*
  92  * nvme_wq - hosts nvme related works that are not reset or delete
  93  * nvme_reset_wq - hosts nvme reset works
  94  * nvme_delete_wq - hosts nvme delete works
  95  *
  96  * nvme_wq will host works such as scan, aen handling, fw activation,
  97  * keep-alive, periodic reconnects etc. nvme_reset_wq
  98  * runs reset works which also flush works hosted on nvme_wq for
  99  * serialization purposes. nvme_delete_wq host controller deletion
 100  * works which flush reset works for serialization.
 101  */
 102 struct workqueue_struct *nvme_wq;
 103 EXPORT_SYMBOL_GPL(nvme_wq);
 104
 105 struct workqueue_struct *nvme_reset_wq;
 106 EXPORT_SYMBOL_GPL(nvme_reset_wq);
 107
 108 struct workqueue_struct *nvme_delete_wq;
 109 EXPORT_SYMBOL_GPL(nvme_delete_wq);
 110
 111 static LIST_HEAD(nvme_subsystems);
 112 static DEFINE_MUTEX(nvme_subsystems_lock);
 113
 114 static DEFINE_IDA(nvme_instance_ida);
 115 static dev_t nvme_ctrl_base_chr_devt;
 116 static struct class *nvme_class;
 117 static struct class *nvme_subsys_class;
 118
 119 static DEFINE_IDA(nvme_ns_chr_minor_ida);
 120 static dev_t nvme_ns_chr_devt;
 121 static struct class *nvme_ns_chr_class;
 122
 123 static void nvme_put_subsystem(struct nvme_subsystem *subsys);
 124 static void nvme_remove_invalid_namespaces(struct nvme_ctrl *ctrl,
 125                                            unsigned nsid);
 126 static void nvme_update_keep_alive(struct nvme_ctrl *ctrl,
 127                                    struct nvme_command *cmd);
 128
 129 void nvme_queue_scan(struct nvme_ctrl *ctrl)
 130 {
 131         /*
 132          * Only new queue scan work when admin and IO queues are both alive
 133          */
 134         if (ctrl->state == NVME_CTRL_LIVE && ctrl->tagset)
 135                 queue_work(nvme_wq, &ctrl->scan_work);
 136 }
 137
 138 /*
 139  * Use this function to proceed with scheduling reset_work for a controller
 140  * that had previously been set to the resetting state. This is intended for
 141  * code paths that can't be interrupted by other reset attempts. A hot removal
 142  * may prevent this from succeeding.
 143  */
 144 int nvme_try_sched_reset(struct nvme_ctrl *ctrl)
 145 {
 146         if (ctrl->state != NVME_CTRL_RESETTING)
 147                 return -EBUSY;
 148         if (!queue_work(nvme_reset_wq, &ctrl->reset_work))
 149                 return -EBUSY;
 150         return 0;
 151 }
 152 EXPORT_SYMBOL_GPL(nvme_try_sched_reset);
 153
 154 static void nvme_failfast_work(struct work_struct *work)
 155 {
 156         struct nvme_ctrl *ctrl = container_of(to_delayed_work(work),
 157                         struct nvme_ctrl, failfast_work);
 158
 159         if (ctrl->state != NVME_CTRL_CONNECTING)
 160                 return;
 161
 162         set_bit(NVME_CTRL_FAILFAST_EXPIRED, &ctrl->flags);
 163         dev_info(ctrl->device, "failfast expired\n");
 164         nvme_kick_requeue_lists(ctrl);
 165 }
 166
 167 static inline void nvme_start_failfast_work(struct nvme_ctrl *ctrl)
 168 {
 169         if (!ctrl->opts || ctrl->opts->fast_io_fail_tmo == -1)
 170                 return;
 171
 172         schedule_delayed_work(&ctrl->failfast_work,
 173                               ctrl->opts->fast_io_fail_tmo * HZ);
 174 }
 175
 176 static inline void nvme_stop_failfast_work(struct nvme_ctrl *ctrl)
 177 {
 178         if (!ctrl->opts)
 179                 return;
 180
 181         cancel_delayed_work_sync(&ctrl->failfast_work);
 182         clear_bit(NVME_CTRL_FAILFAST_EXPIRED, &ctrl->flags);
 183 }
 184
 185
 186 int nvme_reset_ctrl(struct nvme_ctrl *ctrl)
 187 {
 188         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING))
 189                 return -EBUSY;
 190         if (!queue_work(nvme_reset_wq, &ctrl->reset_work))
 191                 return -EBUSY;
 192         return 0;
 193 }
 194 EXPORT_SYMBOL_GPL(nvme_reset_ctrl);
 195
 196 int nvme_reset_ctrl_sync(struct nvme_ctrl *ctrl)
 197 {
 198         int ret;
 199
 200         ret = nvme_reset_ctrl(ctrl);
 201         if (!ret) {
 202                 flush_work(&ctrl->reset_work);
 203                 if (ctrl->state != NVME_CTRL_LIVE)
 204                         ret = -ENETRESET;
 205         }
 206
 207         return ret;
 208 }
 209
 210 static void nvme_do_delete_ctrl(struct nvme_ctrl *ctrl)
 211 {
 212         dev_info(ctrl->device,
 213                  "Removing ctrl: NQN \"%s\"\n", nvmf_ctrl_subsysnqn(ctrl));
 214
 215         flush_work(&ctrl->reset_work);
 216         nvme_stop_ctrl(ctrl);
 217         nvme_remove_namespaces(ctrl);
 218         ctrl->ops->delete_ctrl(ctrl);
 219         nvme_uninit_ctrl(ctrl);
 220 }
 221
 222 static void nvme_delete_ctrl_work(struct work_struct *work)
 223 {
 224         struct nvme_ctrl *ctrl =
 225                 container_of(work, struct nvme_ctrl, delete_work);
 226
 227         nvme_do_delete_ctrl(ctrl);
 228 }
 229
 230 int nvme_delete_ctrl(struct nvme_ctrl *ctrl)
 231 {
 232         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_DELETING))
 233                 return -EBUSY;
 234         if (!queue_work(nvme_delete_wq, &ctrl->delete_work))
 235                 return -EBUSY;
 236         return 0;
 237 }
 238 EXPORT_SYMBOL_GPL(nvme_delete_ctrl);
 239
 240 void nvme_delete_ctrl_sync(struct nvme_ctrl *ctrl)
 241 {
 242         /*
 243          * Keep a reference until nvme_do_delete_ctrl() complete,
 244          * since ->delete_ctrl can free the controller.
 245          */
 246         nvme_get_ctrl(ctrl);
 247         if (nvme_change_ctrl_state(ctrl, NVME_CTRL_DELETING))
 248                 nvme_do_delete_ctrl(ctrl);
 249         nvme_put_ctrl(ctrl);
 250 }
 251
 252 static blk_status_t nvme_error_status(u16 status)
 253 {
 254         switch (status & 0x7ff) {
 255         case NVME_SC_SUCCESS:
 256                 return BLK_STS_OK;
 257         case NVME_SC_CAP_EXCEEDED:
 258                 return BLK_STS_NOSPC;
 259         case NVME_SC_LBA_RANGE:
 260         case NVME_SC_CMD_INTERRUPTED:
 261         case NVME_SC_NS_NOT_READY:
 262                 return BLK_STS_TARGET;
 263         case NVME_SC_BAD_ATTRIBUTES:
 264         case NVME_SC_ONCS_NOT_SUPPORTED:
 265         case NVME_SC_INVALID_OPCODE:
 266         case NVME_SC_INVALID_FIELD:
 267         case NVME_SC_INVALID_NS:
 268                 return BLK_STS_NOTSUPP;
 269         case NVME_SC_WRITE_FAULT:
 270         case NVME_SC_READ_ERROR:
 271         case NVME_SC_UNWRITTEN_BLOCK:
 272         case NVME_SC_ACCESS_DENIED:
 273         case NVME_SC_READ_ONLY:
 274         case NVME_SC_COMPARE_FAILED:
 275                 return BLK_STS_MEDIUM;
 276         case NVME_SC_GUARD_CHECK:
 277         case NVME_SC_APPTAG_CHECK:
 278         case NVME_SC_REFTAG_CHECK:
 279         case NVME_SC_INVALID_PI:
 280                 return BLK_STS_PROTECTION;
 281         case NVME_SC_RESERVATION_CONFLICT:
 282                 return BLK_STS_RESV_CONFLICT;
 283         case NVME_SC_HOST_PATH_ERROR:
 284                 return BLK_STS_TRANSPORT;
 285         case NVME_SC_ZONE_TOO_MANY_ACTIVE:
 286                 return BLK_STS_ZONE_ACTIVE_RESOURCE;
 287         case NVME_SC_ZONE_TOO_MANY_OPEN:
 288                 return BLK_STS_ZONE_OPEN_RESOURCE;
 289         default:
 290                 return BLK_STS_IOERR;
 291         }
 292 }
 293
 294 static void nvme_retry_req(struct request *req)
 295 {
 296         unsigned long delay = 0;
 297         u16 crd;
 298
 299         /* The mask and shift result must be <= 3 */
 300         crd = (nvme_req(req)->status & NVME_SC_CRD) >> 11;
 301         if (crd)
 302                 delay = nvme_req(req)->ctrl->crdt[crd - 1] * 100;
 303
 304         nvme_req(req)->retries++;
 305         blk_mq_requeue_request(req, false);
 306         blk_mq_delay_kick_requeue_list(req->q, delay);
 307 }
 308
 309 static void nvme_log_error(struct request *req)
 310 {
 311         struct nvme_ns *ns = req->q->queuedata;
 312         struct nvme_request *nr = nvme_req(req);
 313
 314         if (ns) {
 315                 pr_err_ratelimited("%s: %s(0x%x) @ LBA %llu, %llu blocks, %s (sct 0x%x / sc 0x%x) %s%s\n",
 316                        ns->disk ? ns->disk->disk_name : "?",
 317                        nvme_get_opcode_str(nr->cmd->common.opcode),
 318                        nr->cmd->common.opcode,
 319                        (unsigned long long)nvme_sect_to_lba(ns, blk_rq_pos(req)),
 320                        (unsigned long long)blk_rq_bytes(req) >> ns->lba_shift,
 321                        nvme_get_error_status_str(nr->status),
 322                        nr->status >> 8 & 7,     /* Status Code Type */
 323                        nr->status & 0xff,       /* Status Code */
 324                        nr->status & NVME_SC_MORE ? "MORE " : "",
 325                        nr->status & NVME_SC_DNR  ? "DNR "  : "");
 326                 return;
 327         }
 328
 329         pr_err_ratelimited("%s: %s(0x%x), %s (sct 0x%x / sc 0x%x) %s%s\n",
 330                            dev_name(nr->ctrl->device),
 331                            nvme_get_admin_opcode_str(nr->cmd->common.opcode),
 332                            nr->cmd->common.opcode,
 333                            nvme_get_error_status_str(nr->status),
 334                            nr->status >> 8 & 7, /* Status Code Type */
 335                            nr->status & 0xff,   /* Status Code */
 336                            nr->status & NVME_SC_MORE ? "MORE " : "",
 337                            nr->status & NVME_SC_DNR  ? "DNR "  : "");
 338 }
 339
 340 enum nvme_disposition {
 341         COMPLETE,
 342         RETRY,
 343         FAILOVER,
 344         AUTHENTICATE,
 345 };
 346
 347 static inline enum nvme_disposition nvme_decide_disposition(struct request *req)
 348 {
 349         if (likely(nvme_req(req)->status == 0))
 350                 return COMPLETE;
 351
 352         if ((nvme_req(req)->status & 0x7ff) == NVME_SC_AUTH_REQUIRED)
 353                 return AUTHENTICATE;
 354
 355         if (blk_noretry_request(req) ||
 356             (nvme_req(req)->status & NVME_SC_DNR) ||
 357             nvme_req(req)->retries >= nvme_max_retries)
 358                 return COMPLETE;
 359
 360         if (req->cmd_flags & REQ_NVME_MPATH) {
 361                 if (nvme_is_path_error(nvme_req(req)->status) ||
 362                     blk_queue_dying(req->q))
 363                         return FAILOVER;
 364         } else {
 365                 if (blk_queue_dying(req->q))
 366                         return COMPLETE;
 367         }
 368
 369         return RETRY;
 370 }
 371
 372 static inline void nvme_end_req_zoned(struct request *req)
 373 {
 374         if (IS_ENABLED(CONFIG_BLK_DEV_ZONED) &&
 375             req_op(req) == REQ_OP_ZONE_APPEND)
 376                 req->__sector = nvme_lba_to_sect(req->q->queuedata,
 377                         le64_to_cpu(nvme_req(req)->result.u64));
 378 }
 379
 380 static inline void nvme_end_req(struct request *req)
 381 {
 382         blk_status_t status = nvme_error_status(nvme_req(req)->status);
 383
 384         if (unlikely(nvme_req(req)->status && !(req->rq_flags & RQF_QUIET)))
 385                 nvme_log_error(req);
 386         nvme_end_req_zoned(req);
 387         nvme_trace_bio_complete(req);
 388         if (req->cmd_flags & REQ_NVME_MPATH)
 389                 nvme_mpath_end_request(req);
 390         blk_mq_end_request(req, status);
 391 }
 392
 393 void nvme_complete_rq(struct request *req)
 394 {
 395         struct nvme_ctrl *ctrl = nvme_req(req)->ctrl;
 396
 397         trace_nvme_complete_rq(req);
 398         nvme_cleanup_cmd(req);
 399
 400         /*
 401          * Completions of long-running commands should not be able to
 402          * defer sending of periodic keep alives, since the controller
 403          * may have completed processing such commands a long time ago
 404          * (arbitrarily close to command submission time).
 405          * req->deadline - req->timeout is the command submission time
 406          * in jiffies.
 407          */
 408         if (ctrl->kas &&
 409             req->deadline - req->timeout >= ctrl->ka_last_check_time)
 410                 ctrl->comp_seen = true;
 411
 412         switch (nvme_decide_disposition(req)) {
 413         case COMPLETE:
 414                 nvme_end_req(req);
 415                 return;
 416         case RETRY:
 417                 nvme_retry_req(req);
 418                 return;
 419         case FAILOVER:
 420                 nvme_failover_req(req);
 421                 return;
 422         case AUTHENTICATE:
 423 #ifdef CONFIG_NVME_AUTH
 424                 queue_work(nvme_wq, &ctrl->dhchap_auth_work);
 425                 nvme_retry_req(req);
 426 #else
 427                 nvme_end_req(req);
 428 #endif
 429                 return;
 430         }
 431 }
 432 EXPORT_SYMBOL_GPL(nvme_complete_rq);
 433
 434 void nvme_complete_batch_req(struct request *req)
 435 {
 436         trace_nvme_complete_rq(req);
 437         nvme_cleanup_cmd(req);
 438         nvme_end_req_zoned(req);
 439 }
 440 EXPORT_SYMBOL_GPL(nvme_complete_batch_req);
 441
 442 /*
 443  * Called to unwind from ->queue_rq on a failed command submission so that the
 444  * multipathing code gets called to potentially failover to another path.
 445  * The caller needs to unwind all transport specific resource allocations and
 446  * must return propagate the return value.
 447  */
 448 blk_status_t nvme_host_path_error(struct request *req)
 449 {
 450         nvme_req(req)->status = NVME_SC_HOST_PATH_ERROR;
 451         blk_mq_set_request_complete(req);
 452         nvme_complete_rq(req);
 453         return BLK_STS_OK;
 454 }
 455 EXPORT_SYMBOL_GPL(nvme_host_path_error);
 456
 457 bool nvme_cancel_request(struct request *req, void *data)
 458 {
 459         dev_dbg_ratelimited(((struct nvme_ctrl *) data)->device,
 460                                 "Cancelling I/O %d", req->tag);
 461
 462         /* don't abort one completed or idle request */
 463         if (blk_mq_rq_state(req) != MQ_RQ_IN_FLIGHT)
 464                 return true;
 465
 466         nvme_req(req)->status = NVME_SC_HOST_ABORTED_CMD;
 467         nvme_req(req)->flags |= NVME_REQ_CANCELLED;
 468         blk_mq_complete_request(req);
 469         return true;
 470 }
 471 EXPORT_SYMBOL_GPL(nvme_cancel_request);
 472
 473 void nvme_cancel_tagset(struct nvme_ctrl *ctrl)
 474 {
 475         if (ctrl->tagset) {
 476                 blk_mq_tagset_busy_iter(ctrl->tagset,
 477                                 nvme_cancel_request, ctrl);
 478                 blk_mq_tagset_wait_completed_request(ctrl->tagset);
 479         }
 480 }
 481 EXPORT_SYMBOL_GPL(nvme_cancel_tagset);
 482
 483 void nvme_cancel_admin_tagset(struct nvme_ctrl *ctrl)
 484 {
 485         if (ctrl->admin_tagset) {
 486                 blk_mq_tagset_busy_iter(ctrl->admin_tagset,
 487                                 nvme_cancel_request, ctrl);
 488                 blk_mq_tagset_wait_completed_request(ctrl->admin_tagset);
 489         }
 490 }
 491 EXPORT_SYMBOL_GPL(nvme_cancel_admin_tagset);
 492
 493 bool nvme_change_ctrl_state(struct nvme_ctrl *ctrl,
 494                 enum nvme_ctrl_state new_state)
 495 {
 496         enum nvme_ctrl_state old_state;
 497         unsigned long flags;
 498         bool changed = false;
 499
 500         spin_lock_irqsave(&ctrl->lock, flags);
 501
 502         old_state = ctrl->state;
 503         switch (new_state) {
 504         case NVME_CTRL_LIVE:
 505                 switch (old_state) {
 506                 case NVME_CTRL_NEW:
 507                 case NVME_CTRL_RESETTING:
 508                 case NVME_CTRL_CONNECTING:
 509                         changed = true;
 510                         fallthrough;
 511                 default:
 512                         break;
 513                 }
 514                 break;
 515         case NVME_CTRL_RESETTING:
 516                 switch (old_state) {
 517                 case NVME_CTRL_NEW:
 518                 case NVME_CTRL_LIVE:
 519                         changed = true;
 520                         fallthrough;
 521                 default:
 522                         break;
 523                 }
 524                 break;
 525         case NVME_CTRL_CONNECTING:
 526                 switch (old_state) {
 527                 case NVME_CTRL_NEW:
 528                 case NVME_CTRL_RESETTING:
 529                         changed = true;
 530                         fallthrough;
 531                 default:
 532                         break;
 533                 }
 534                 break;
 535         case NVME_CTRL_DELETING:
 536                 switch (old_state) {
 537                 case NVME_CTRL_LIVE:
 538                 case NVME_CTRL_RESETTING:
 539                 case NVME_CTRL_CONNECTING:
 540                         changed = true;
 541                         fallthrough;
 542                 default:
 543                         break;
 544                 }
 545                 break;
 546         case NVME_CTRL_DELETING_NOIO:
 547                 switch (old_state) {
 548                 case NVME_CTRL_DELETING:
 549                 case NVME_CTRL_DEAD:
 550                         changed = true;
 551                         fallthrough;
 552                 default:
 553                         break;
 554                 }
 555                 break;
 556         case NVME_CTRL_DEAD:
 557                 switch (old_state) {
 558                 case NVME_CTRL_DELETING:
 559                         changed = true;
 560                         fallthrough;
 561                 default:
 562                         break;
 563                 }
 564                 break;
 565         default:
 566                 break;
 567         }
 568
 569         if (changed) {
 570                 ctrl->state = new_state;
 571                 wake_up_all(&ctrl->state_wq);
 572         }
 573
 574         spin_unlock_irqrestore(&ctrl->lock, flags);
 575         if (!changed)
 576                 return false;
 577
 578         if (ctrl->state == NVME_CTRL_LIVE) {
 579                 if (old_state == NVME_CTRL_CONNECTING)
 580                         nvme_stop_failfast_work(ctrl);
 581                 nvme_kick_requeue_lists(ctrl);
 582         } else if (ctrl->state == NVME_CTRL_CONNECTING &&
 583                 old_state == NVME_CTRL_RESETTING) {
 584                 nvme_start_failfast_work(ctrl);
 585         }
 586         return changed;
 587 }
 588 EXPORT_SYMBOL_GPL(nvme_change_ctrl_state);
 589
 590 /*
 591  * Returns true for sink states that can't ever transition back to live.
 592  */
 593 static bool nvme_state_terminal(struct nvme_ctrl *ctrl)
 594 {
 595         switch (ctrl->state) {
 596         case NVME_CTRL_NEW:
 597         case NVME_CTRL_LIVE:
 598         case NVME_CTRL_RESETTING:
 599         case NVME_CTRL_CONNECTING:
 600                 return false;
 601         case NVME_CTRL_DELETING:
 602         case NVME_CTRL_DELETING_NOIO:
 603         case NVME_CTRL_DEAD:
 604                 return true;
 605         default:
 606                 WARN_ONCE(1, "Unhandled ctrl state:%d", ctrl->state);
 607                 return true;
 608         }
 609 }
 610
 611 /*
 612  * Waits for the controller state to be resetting, or returns false if it is
 613  * not possible to ever transition to that state.
 614  */
 615 bool nvme_wait_reset(struct nvme_ctrl *ctrl)
 616 {
 617         wait_event(ctrl->state_wq,
 618                    nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING) ||
 619                    nvme_state_terminal(ctrl));
 620         return ctrl->state == NVME_CTRL_RESETTING;
 621 }
 622 EXPORT_SYMBOL_GPL(nvme_wait_reset);
 623
 624 static void nvme_free_ns_head(struct kref *ref)
 625 {
 626         struct nvme_ns_head *head =
 627                 container_of(ref, struct nvme_ns_head, ref);
 628
 629         nvme_mpath_remove_disk(head);
 630         ida_free(&head->subsys->ns_ida, head->instance);
 631         cleanup_srcu_struct(&head->srcu);
 632         nvme_put_subsystem(head->subsys);
 633         kfree(head);
 634 }
 635
 636 bool nvme_tryget_ns_head(struct nvme_ns_head *head)
 637 {
 638         return kref_get_unless_zero(&head->ref);
 639 }
 640
 641 void nvme_put_ns_head(struct nvme_ns_head *head)
 642 {
 643         kref_put(&head->ref, nvme_free_ns_head);
 644 }
 645
 646 static void nvme_free_ns(struct kref *kref)
 647 {
 648         struct nvme_ns *ns = container_of(kref, struct nvme_ns, kref);
 649
 650         put_disk(ns->disk);
 651         nvme_put_ns_head(ns->head);
 652         nvme_put_ctrl(ns->ctrl);
 653         kfree(ns);
 654 }
 655
 656 static inline bool nvme_get_ns(struct nvme_ns *ns)
 657 {
 658         return kref_get_unless_zero(&ns->kref);
 659 }
 660
 661 void nvme_put_ns(struct nvme_ns *ns)
 662 {
 663         kref_put(&ns->kref, nvme_free_ns);
 664 }
 665 EXPORT_SYMBOL_NS_GPL(nvme_put_ns, NVME_TARGET_PASSTHRU);
 666
 667 static inline void nvme_clear_nvme_request(struct request *req)
 668 {
 669         nvme_req(req)->status = 0;
 670         nvme_req(req)->retries = 0;
 671         nvme_req(req)->flags = 0;
 672         req->rq_flags |= RQF_DONTPREP;
 673 }
 674
 675 /* initialize a passthrough request */
 676 void nvme_init_request(struct request *req, struct nvme_command *cmd)
 677 {
 678         if (req->q->queuedata)
 679                 req->timeout = NVME_IO_TIMEOUT;
 680         else /* no queuedata implies admin queue */
 681                 req->timeout = NVME_ADMIN_TIMEOUT;
 682
 683         /* passthru commands should let the driver set the SGL flags */
 684         cmd->common.flags &= ~NVME_CMD_SGL_ALL;
 685
 686         req->cmd_flags |= REQ_FAILFAST_DRIVER;
 687         if (req->mq_hctx->type == HCTX_TYPE_POLL)
 688                 req->cmd_flags |= REQ_POLLED;
 689         nvme_clear_nvme_request(req);
 690         req->rq_flags |= RQF_QUIET;
 691         memcpy(nvme_req(req)->cmd, cmd, sizeof(*cmd));
 692 }
 693 EXPORT_SYMBOL_GPL(nvme_init_request);
 694
 695 /*
 696  * For something we're not in a state to send to the device the default action
 697  * is to busy it and retry it after the controller state is recovered.  However,
 698  * if the controller is deleting or if anything is marked for failfast or
 699  * nvme multipath it is immediately failed.
 700  *
 701  * Note: commands used to initialize the controller will be marked for failfast.
 702  * Note: nvme cli/ioctl commands are marked for failfast.
 703  */
 704 blk_status_t nvme_fail_nonready_command(struct nvme_ctrl *ctrl,
 705                 struct request *rq)
 706 {
 707         if (ctrl->state != NVME_CTRL_DELETING_NOIO &&
 708             ctrl->state != NVME_CTRL_DELETING &&
 709             ctrl->state != NVME_CTRL_DEAD &&
 710             !test_bit(NVME_CTRL_FAILFAST_EXPIRED, &ctrl->flags) &&
 711             !blk_noretry_request(rq) && !(rq->cmd_flags & REQ_NVME_MPATH))
 712                 return BLK_STS_RESOURCE;
 713         return nvme_host_path_error(rq);
 714 }
 715 EXPORT_SYMBOL_GPL(nvme_fail_nonready_command);
 716
 717 bool __nvme_check_ready(struct nvme_ctrl *ctrl, struct request *rq,
 718                 bool queue_live)
 719 {
 720         struct nvme_request *req = nvme_req(rq);
 721
 722         /*
 723          * currently we have a problem sending passthru commands
 724          * on the admin_q if the controller is not LIVE because we can't
 725          * make sure that they are going out after the admin connect,
 726          * controller enable and/or other commands in the initialization
 727          * sequence. until the controller will be LIVE, fail with
 728          * BLK_STS_RESOURCE so that they will be rescheduled.
 729          */
 730         if (rq->q == ctrl->admin_q && (req->flags & NVME_REQ_USERCMD))
 731                 return false;
 732
 733         if (ctrl->ops->flags & NVME_F_FABRICS) {
 734                 /*
 735                  * Only allow commands on a live queue, except for the connect
 736                  * command, which is require to set the queue live in the
 737                  * appropinquate states.
 738                  */
 739                 switch (ctrl->state) {
 740                 case NVME_CTRL_CONNECTING:
 741                         if (blk_rq_is_passthrough(rq) && nvme_is_fabrics(req->cmd) &&
 742                             (req->cmd->fabrics.fctype == nvme_fabrics_type_connect ||
 743                              req->cmd->fabrics.fctype == nvme_fabrics_type_auth_send ||
 744                              req->cmd->fabrics.fctype == nvme_fabrics_type_auth_receive))
 745                                 return true;
 746                         break;
 747                 default:
 748                         break;
 749                 case NVME_CTRL_DEAD:
 750                         return false;
 751                 }
 752         }
 753
 754         return queue_live;
 755 }
 756 EXPORT_SYMBOL_GPL(__nvme_check_ready);
 757
 758 static inline void nvme_setup_flush(struct nvme_ns *ns,
 759                 struct nvme_command *cmnd)
 760 {
 761         memset(cmnd, 0, sizeof(*cmnd));
 762         cmnd->common.opcode = nvme_cmd_flush;
 763         cmnd->common.nsid = cpu_to_le32(ns->head->ns_id);
 764 }
 765
 766 static blk_status_t nvme_setup_discard(struct nvme_ns *ns, struct request *req,
 767                 struct nvme_command *cmnd)
 768 {
 769         unsigned short segments = blk_rq_nr_discard_segments(req), n = 0;
 770         struct nvme_dsm_range *range;
 771         struct bio *bio;
 772
 773         /*
 774          * Some devices do not consider the DSM 'Number of Ranges' field when
 775          * determining how much data to DMA. Always allocate memory for maximum
 776          * number of segments to prevent device reading beyond end of buffer.
 777          */
 778         static const size_t alloc_size = sizeof(*range) * NVME_DSM_MAX_RANGES;
 779
 780         range = kzalloc(alloc_size, GFP_ATOMIC | __GFP_NOWARN);
 781         if (!range) {
 782                 /*
 783                  * If we fail allocation our range, fallback to the controller
 784                  * discard page. If that's also busy, it's safe to return
 785                  * busy, as we know we can make progress once that's freed.
 786                  */
 787                 if (test_and_set_bit_lock(0, &ns->ctrl->discard_page_busy))
 788                         return BLK_STS_RESOURCE;
 789
 790                 range = page_address(ns->ctrl->discard_page);
 791         }
 792
 793         if (queue_max_discard_segments(req->q) == 1) {
 794                 u64 slba = nvme_sect_to_lba(ns, blk_rq_pos(req));
 795                 u32 nlb = blk_rq_sectors(req) >> (ns->lba_shift - 9);
 796
 797                 range[0].cattr = cpu_to_le32(0);
 798                 range[0].nlb = cpu_to_le32(nlb);
 799                 range[0].slba = cpu_to_le64(slba);
 800                 n = 1;
 801         } else {
 802                 __rq_for_each_bio(bio, req) {
 803                         u64 slba = nvme_sect_to_lba(ns, bio->bi_iter.bi_sector);
 804                         u32 nlb = bio->bi_iter.bi_size >> ns->lba_shift;
 805
 806                         if (n < segments) {
 807                                 range[n].cattr = cpu_to_le32(0);
 808                                 range[n].nlb = cpu_to_le32(nlb);
 809                                 range[n].slba = cpu_to_le64(slba);
 810                         }
 811                         n++;
 812                 }
 813         }
 814
 815         if (WARN_ON_ONCE(n != segments)) {
 816                 if (virt_to_page(range) == ns->ctrl->discard_page)
 817                         clear_bit_unlock(0, &ns->ctrl->discard_page_busy);
 818                 else
 819                         kfree(range);
 820                 return BLK_STS_IOERR;
 821         }
 822
 823         memset(cmnd, 0, sizeof(*cmnd));
 824         cmnd->dsm.opcode = nvme_cmd_dsm;
 825         cmnd->dsm.nsid = cpu_to_le32(ns->head->ns_id);
 826         cmnd->dsm.nr = cpu_to_le32(segments - 1);
 827         cmnd->dsm.attributes = cpu_to_le32(NVME_DSMGMT_AD);
 828
 829         bvec_set_virt(&req->special_vec, range, alloc_size);
 830         req->rq_flags |= RQF_SPECIAL_PAYLOAD;
 831
 832         return BLK_STS_OK;
 833 }
 834
 835 static void nvme_set_ref_tag(struct nvme_ns *ns, struct nvme_command *cmnd,
 836                               struct request *req)
 837 {
 838         u32 upper, lower;
 839         u64 ref48;
 840
 841         /* both rw and write zeroes share the same reftag format */
 842         switch (ns->guard_type) {
 843         case NVME_NVM_NS_16B_GUARD:
 844                 cmnd->rw.reftag = cpu_to_le32(t10_pi_ref_tag(req));
 845                 break;
 846         case NVME_NVM_NS_64B_GUARD:
 847                 ref48 = ext_pi_ref_tag(req);
 848                 lower = lower_32_bits(ref48);
 849                 upper = upper_32_bits(ref48);
 850
 851                 cmnd->rw.reftag = cpu_to_le32(lower);
 852                 cmnd->rw.cdw3 = cpu_to_le32(upper);
 853                 break;
 854         default:
 855                 break;
 856         }
 857 }
 858
 859 static inline blk_status_t nvme_setup_write_zeroes(struct nvme_ns *ns,
 860                 struct request *req, struct nvme_command *cmnd)
 861 {
 862         memset(cmnd, 0, sizeof(*cmnd));
 863
 864         if (ns->ctrl->quirks & NVME_QUIRK_DEALLOCATE_ZEROES)
 865                 return nvme_setup_discard(ns, req, cmnd);
 866
 867         cmnd->write_zeroes.opcode = nvme_cmd_write_zeroes;
 868         cmnd->write_zeroes.nsid = cpu_to_le32(ns->head->ns_id);
 869         cmnd->write_zeroes.slba =
 870                 cpu_to_le64(nvme_sect_to_lba(ns, blk_rq_pos(req)));
 871         cmnd->write_zeroes.length =
 872                 cpu_to_le16((blk_rq_bytes(req) >> ns->lba_shift) - 1);
 873
 874         if (!(req->cmd_flags & REQ_NOUNMAP) && (ns->features & NVME_NS_DEAC))
 875                 cmnd->write_zeroes.control |= cpu_to_le16(NVME_WZ_DEAC);
 876
 877         if (nvme_ns_has_pi(ns)) {
 878                 cmnd->write_zeroes.control |= cpu_to_le16(NVME_RW_PRINFO_PRACT);
 879
 880                 switch (ns->pi_type) {
 881                 case NVME_NS_DPS_PI_TYPE1:
 882                 case NVME_NS_DPS_PI_TYPE2:
 883                         nvme_set_ref_tag(ns, cmnd, req);
 884                         break;
 885                 }
 886         }
 887
 888         return BLK_STS_OK;
 889 }
 890
 891 static inline blk_status_t nvme_setup_rw(struct nvme_ns *ns,
 892                 struct request *req, struct nvme_command *cmnd,
 893                 enum nvme_opcode op)
 894 {
 895         u16 control = 0;
 896         u32 dsmgmt = 0;
 897
 898         if (req->cmd_flags & REQ_FUA)
 899                 control |= NVME_RW_FUA;
 900         if (req->cmd_flags & (REQ_FAILFAST_DEV | REQ_RAHEAD))
 901                 control |= NVME_RW_LR;
 902
 903         if (req->cmd_flags & REQ_RAHEAD)
 904                 dsmgmt |= NVME_RW_DSM_FREQ_PREFETCH;
 905
 906         cmnd->rw.opcode = op;
 907         cmnd->rw.flags = 0;
 908         cmnd->rw.nsid = cpu_to_le32(ns->head->ns_id);
 909         cmnd->rw.cdw2 = 0;
 910         cmnd->rw.cdw3 = 0;
 911         cmnd->rw.metadata = 0;
 912         cmnd->rw.slba = cpu_to_le64(nvme_sect_to_lba(ns, blk_rq_pos(req)));
 913         cmnd->rw.length = cpu_to_le16((blk_rq_bytes(req) >> ns->lba_shift) - 1);
 914         cmnd->rw.reftag = 0;
 915         cmnd->rw.apptag = 0;
 916         cmnd->rw.appmask = 0;
 917
 918         if (ns->ms) {
 919                 /*
 920                  * If formated with metadata, the block layer always provides a
 921                  * metadata buffer if CONFIG_BLK_DEV_INTEGRITY is enabled.  Else
 922                  * we enable the PRACT bit for protection information or set the
 923                  * namespace capacity to zero to prevent any I/O.
 924                  */
 925                 if (!blk_integrity_rq(req)) {
 926                         if (WARN_ON_ONCE(!nvme_ns_has_pi(ns)))
 927                                 return BLK_STS_NOTSUPP;
 928                         control |= NVME_RW_PRINFO_PRACT;
 929                 }
 930
 931                 switch (ns->pi_type) {
 932                 case NVME_NS_DPS_PI_TYPE3:
 933                         control |= NVME_RW_PRINFO_PRCHK_GUARD;
 934                         break;
 935                 case NVME_NS_DPS_PI_TYPE1:
 936                 case NVME_NS_DPS_PI_TYPE2:
 937                         control |= NVME_RW_PRINFO_PRCHK_GUARD |
 938                                         NVME_RW_PRINFO_PRCHK_REF;
 939                         if (op == nvme_cmd_zone_append)
 940                                 control |= NVME_RW_APPEND_PIREMAP;
 941                         nvme_set_ref_tag(ns, cmnd, req);
 942                         break;
 943                 }
 944         }
 945
 946         cmnd->rw.control = cpu_to_le16(control);
 947         cmnd->rw.dsmgmt = cpu_to_le32(dsmgmt);
 948         return 0;
 949 }
 950
 951 void nvme_cleanup_cmd(struct request *req)
 952 {
 953         if (req->rq_flags & RQF_SPECIAL_PAYLOAD) {
 954                 struct nvme_ctrl *ctrl = nvme_req(req)->ctrl;
 955
 956                 if (req->special_vec.bv_page == ctrl->discard_page)
 957                         clear_bit_unlock(0, &ctrl->discard_page_busy);
 958                 else
 959                         kfree(bvec_virt(&req->special_vec));
 960         }
 961 }
 962 EXPORT_SYMBOL_GPL(nvme_cleanup_cmd);
 963
 964 blk_status_t nvme_setup_cmd(struct nvme_ns *ns, struct request *req)
 965 {
 966         struct nvme_command *cmd = nvme_req(req)->cmd;
 967         blk_status_t ret = BLK_STS_OK;
 968
 969         if (!(req->rq_flags & RQF_DONTPREP))
 970                 nvme_clear_nvme_request(req);
 971
 972         switch (req_op(req)) {
 973         case REQ_OP_DRV_IN:
 974         case REQ_OP_DRV_OUT:
 975                 /* these are setup prior to execution in nvme_init_request() */
 976                 break;
 977         case REQ_OP_FLUSH:
 978                 nvme_setup_flush(ns, cmd);
 979                 break;
 980         case REQ_OP_ZONE_RESET_ALL:
 981         case REQ_OP_ZONE_RESET:
 982                 ret = nvme_setup_zone_mgmt_send(ns, req, cmd, NVME_ZONE_RESET);
 983                 break;
 984         case REQ_OP_ZONE_OPEN:
 985                 ret = nvme_setup_zone_mgmt_send(ns, req, cmd, NVME_ZONE_OPEN);
 986                 break;
 987         case REQ_OP_ZONE_CLOSE:
 988                 ret = nvme_setup_zone_mgmt_send(ns, req, cmd, NVME_ZONE_CLOSE);
 989                 break;
 990         case REQ_OP_ZONE_FINISH:
 991                 ret = nvme_setup_zone_mgmt_send(ns, req, cmd, NVME_ZONE_FINISH);
 992                 break;
 993         case REQ_OP_WRITE_ZEROES:
 994                 ret = nvme_setup_write_zeroes(ns, req, cmd);
 995                 break;
 996         case REQ_OP_DISCARD:
 997                 ret = nvme_setup_discard(ns, req, cmd);
 998                 break;
 999         case REQ_OP_READ:
1000                 ret = nvme_setup_rw(ns, req, cmd, nvme_cmd_read);
1001                 break;
1002         case REQ_OP_WRITE:
1003                 ret = nvme_setup_rw(ns, req, cmd, nvme_cmd_write);
1004                 break;
1005         case REQ_OP_ZONE_APPEND:
1006                 ret = nvme_setup_rw(ns, req, cmd, nvme_cmd_zone_append);
1007                 break;
1008         default:
1009                 WARN_ON_ONCE(1);
1010                 return BLK_STS_IOERR;
1011         }
1012
1013         cmd->common.command_id = nvme_cid(req);
1014         trace_nvme_setup_cmd(req, cmd);
1015         return ret;
1016 }
1017 EXPORT_SYMBOL_GPL(nvme_setup_cmd);
1018
1019 /*
1020  * Return values:
1021  * 0:  success
1022  * >0: nvme controller's cqe status response
1023  * <0: kernel error in lieu of controller response
1024  */
1025 int nvme_execute_rq(struct request *rq, bool at_head)
1026 {
1027         blk_status_t status;
1028
1029         status = blk_execute_rq(rq, at_head);
1030         if (nvme_req(rq)->flags & NVME_REQ_CANCELLED)
1031                 return -EINTR;
1032         if (nvme_req(rq)->status)
1033                 return nvme_req(rq)->status;
1034         return blk_status_to_errno(status);
1035 }
1036 EXPORT_SYMBOL_NS_GPL(nvme_execute_rq, NVME_TARGET_PASSTHRU);
1037
1038 /*
1039  * Returns 0 on success.  If the result is negative, it's a Linux error code;
1040  * if the result is positive, it's an NVM Express status code
1041  */
1042 int __nvme_submit_sync_cmd(struct request_queue *q, struct nvme_command *cmd,
1043                 union nvme_result *result, void *buffer, unsigned bufflen,
1044                 int qid, int at_head, blk_mq_req_flags_t flags)
1045 {
1046         struct request *req;
1047         int ret;
1048
1049         if (qid == NVME_QID_ANY)
1050                 req = blk_mq_alloc_request(q, nvme_req_op(cmd), flags);
1051         else
1052                 req = blk_mq_alloc_request_hctx(q, nvme_req_op(cmd), flags,
1053                                                 qid - 1);
1054
1055         if (IS_ERR(req))
1056                 return PTR_ERR(req);
1057         nvme_init_request(req, cmd);
1058
1059         if (buffer && bufflen) {
1060                 ret = blk_rq_map_kern(q, req, buffer, bufflen, GFP_KERNEL);
1061                 if (ret)
1062                         goto out;
1063         }
1064
1065         ret = nvme_execute_rq(req, at_head);
1066         if (result && ret >= 0)
1067                 *result = nvme_req(req)->result;
1068  out:
1069         blk_mq_free_request(req);
1070         return ret;
1071 }
1072 EXPORT_SYMBOL_GPL(__nvme_submit_sync_cmd);
1073
1074 int nvme_submit_sync_cmd(struct request_queue *q, struct nvme_command *cmd,
1075                 void *buffer, unsigned bufflen)
1076 {
1077         return __nvme_submit_sync_cmd(q, cmd, NULL, buffer, bufflen,
1078                         NVME_QID_ANY, 0, 0);
1079 }
1080 EXPORT_SYMBOL_GPL(nvme_submit_sync_cmd);
1081
1082 u32 nvme_command_effects(struct nvme_ctrl *ctrl, struct nvme_ns *ns, u8 opcode)
1083 {
1084         u32 effects = 0;
1085
1086         if (ns) {
1087                 effects = le32_to_cpu(ns->head->effects->iocs[opcode]);
1088                 if (effects & ~(NVME_CMD_EFFECTS_CSUPP | NVME_CMD_EFFECTS_LBCC))
1089                         dev_warn_once(ctrl->device,
1090                                 "IO command:%02x has unusual effects:%08x\n",
1091                                 opcode, effects);
1092
1093                 /*
1094                  * NVME_CMD_EFFECTS_CSE_MASK causes a freeze all I/O queues,
1095                  * which would deadlock when done on an I/O command.  Note that
1096                  * We already warn about an unusual effect above.
1097                  */
1098                 effects &= ~NVME_CMD_EFFECTS_CSE_MASK;
1099         } else {
1100                 effects = le32_to_cpu(ctrl->effects->acs[opcode]);
1101         }
1102
1103         return effects;
1104 }
1105 EXPORT_SYMBOL_NS_GPL(nvme_command_effects, NVME_TARGET_PASSTHRU);
1106
1107 u32 nvme_passthru_start(struct nvme_ctrl *ctrl, struct nvme_ns *ns, u8 opcode)
1108 {
1109         u32 effects = nvme_command_effects(ctrl, ns, opcode);
1110
1111         /*
1112          * For simplicity, IO to all namespaces is quiesced even if the command
1113          * effects say only one namespace is affected.
1114          */
1115         if (effects & NVME_CMD_EFFECTS_CSE_MASK) {
1116                 mutex_lock(&ctrl->scan_lock);
1117                 mutex_lock(&ctrl->subsys->lock);
1118                 nvme_mpath_start_freeze(ctrl->subsys);
1119                 nvme_mpath_wait_freeze(ctrl->subsys);
1120                 nvme_start_freeze(ctrl);
1121                 nvme_wait_freeze(ctrl);
1122         }
1123         return effects;
1124 }
1125 EXPORT_SYMBOL_NS_GPL(nvme_passthru_start, NVME_TARGET_PASSTHRU);
1126
1127 void nvme_passthru_end(struct nvme_ctrl *ctrl, struct nvme_ns *ns, u32 effects,
1128                        struct nvme_command *cmd, int status)
1129 {
1130         if (effects & NVME_CMD_EFFECTS_CSE_MASK) {
1131                 nvme_unfreeze(ctrl);
1132                 nvme_mpath_unfreeze(ctrl->subsys);
1133                 mutex_unlock(&ctrl->subsys->lock);
1134                 mutex_unlock(&ctrl->scan_lock);
1135         }
1136         if (effects & NVME_CMD_EFFECTS_CCC) {
1137                 if (!test_and_set_bit(NVME_CTRL_DIRTY_CAPABILITY,
1138                                       &ctrl->flags)) {
1139                         dev_info(ctrl->device,
1140 "controller capabilities changed, reset may be required to take effect.\n");
1141                 }
1142         }
1143         if (effects & (NVME_CMD_EFFECTS_NIC | NVME_CMD_EFFECTS_NCC)) {
1144                 nvme_queue_scan(ctrl);
1145                 flush_work(&ctrl->scan_work);
1146         }
1147         if (ns)
1148                 return;
1149
1150         switch (cmd->common.opcode) {
1151         case nvme_admin_set_features:
1152                 switch (le32_to_cpu(cmd->common.cdw10) & 0xFF) {
1153                 case NVME_FEAT_KATO:
1154                         /*
1155                          * Keep alive commands interval on the host should be
1156                          * updated when KATO is modified by Set Features
1157                          * commands.
1158                          */
1159                         if (!status)
1160                                 nvme_update_keep_alive(ctrl, cmd);
1161                         break;
1162                 default:
1163                         break;
1164                 }
1165                 break;
1166         default:
1167                 break;
1168         }
1169 }
1170 EXPORT_SYMBOL_NS_GPL(nvme_passthru_end, NVME_TARGET_PASSTHRU);
1171
1172 /*
1173  * Recommended frequency for KATO commands per NVMe 1.4 section 7.12.1:
1174  *
1175  *   The host should send Keep Alive commands at half of the Keep Alive Timeout
1176  *   accounting for transport roundtrip times [..].
1177  */
1178 static unsigned long nvme_keep_alive_work_period(struct nvme_ctrl *ctrl)
1179 {
1180         unsigned long delay = ctrl->kato * HZ / 2;
1181
1182         /*
1183          * When using Traffic Based Keep Alive, we need to run
1184          * nvme_keep_alive_work at twice the normal frequency, as one
1185          * command completion can postpone sending a keep alive command
1186          * by up to twice the delay between runs.
1187          */
1188         if (ctrl->ctratt & NVME_CTRL_ATTR_TBKAS)
1189                 delay /= 2;
1190         return delay;
1191 }
1192
1193 static void nvme_queue_keep_alive_work(struct nvme_ctrl *ctrl)
1194 {
1195         queue_delayed_work(nvme_wq, &ctrl->ka_work,
1196                            nvme_keep_alive_work_period(ctrl));
1197 }
1198
1199 static enum rq_end_io_ret nvme_keep_alive_end_io(struct request *rq,
1200                                                  blk_status_t status)
1201 {
1202         struct nvme_ctrl *ctrl = rq->end_io_data;
1203         unsigned long flags;
1204         bool startka = false;
1205         unsigned long rtt = jiffies - (rq->deadline - rq->timeout);
1206         unsigned long delay = nvme_keep_alive_work_period(ctrl);
1207
1208         /*
1209          * Subtract off the keepalive RTT so nvme_keep_alive_work runs
1210          * at the desired frequency.
1211          */
1212         if (rtt <= delay) {
1213                 delay -= rtt;
1214         } else {
1215                 dev_warn(ctrl->device, "long keepalive RTT (%u ms)\n",
1216                          jiffies_to_msecs(rtt));
1217                 delay = 0;
1218         }
1219
1220         blk_mq_free_request(rq);
1221
1222         if (status) {
1223                 dev_err(ctrl->device,
1224                         "failed nvme_keep_alive_end_io error=%d\n",
1225                                 status);
1226                 return RQ_END_IO_NONE;
1227         }
1228
1229         ctrl->ka_last_check_time = jiffies;
1230         ctrl->comp_seen = false;
1231         spin_lock_irqsave(&ctrl->lock, flags);
1232         if (ctrl->state == NVME_CTRL_LIVE ||
1233             ctrl->state == NVME_CTRL_CONNECTING)
1234                 startka = true;
1235         spin_unlock_irqrestore(&ctrl->lock, flags);
1236         if (startka)
1237                 queue_delayed_work(nvme_wq, &ctrl->ka_work, delay);
1238         return RQ_END_IO_NONE;
1239 }
1240
1241 static void nvme_keep_alive_work(struct work_struct *work)
1242 {
1243         struct nvme_ctrl *ctrl = container_of(to_delayed_work(work),
1244                         struct nvme_ctrl, ka_work);
1245         bool comp_seen = ctrl->comp_seen;
1246         struct request *rq;
1247
1248         ctrl->ka_last_check_time = jiffies;
1249
1250         if ((ctrl->ctratt & NVME_CTRL_ATTR_TBKAS) && comp_seen) {
1251                 dev_dbg(ctrl->device,
1252                         "reschedule traffic based keep-alive timer\n");
1253                 ctrl->comp_seen = false;
1254                 nvme_queue_keep_alive_work(ctrl);
1255                 return;
1256         }
1257
1258         rq = blk_mq_alloc_request(ctrl->admin_q, nvme_req_op(&ctrl->ka_cmd),
1259                                   BLK_MQ_REQ_RESERVED | BLK_MQ_REQ_NOWAIT);
1260         if (IS_ERR(rq)) {
1261                 /* allocation failure, reset the controller */
1262                 dev_err(ctrl->device, "keep-alive failed: %ld\n", PTR_ERR(rq));
1263                 nvme_reset_ctrl(ctrl);
1264                 return;
1265         }
1266         nvme_init_request(rq, &ctrl->ka_cmd);
1267
1268         rq->timeout = ctrl->kato * HZ;
1269         rq->end_io = nvme_keep_alive_end_io;
1270         rq->end_io_data = ctrl;
1271         blk_execute_rq_nowait(rq, false);
1272 }
1273
1274 static void nvme_start_keep_alive(struct nvme_ctrl *ctrl)
1275 {
1276         if (unlikely(ctrl->kato == 0))
1277                 return;
1278
1279         nvme_queue_keep_alive_work(ctrl);
1280 }
1281
1282 void nvme_stop_keep_alive(struct nvme_ctrl *ctrl)
1283 {
1284         if (unlikely(ctrl->kato == 0))
1285                 return;
1286
1287         cancel_delayed_work_sync(&ctrl->ka_work);
1288 }
1289 EXPORT_SYMBOL_GPL(nvme_stop_keep_alive);
1290
1291 static void nvme_update_keep_alive(struct nvme_ctrl *ctrl,
1292                                    struct nvme_command *cmd)
1293 {
1294         unsigned int new_kato =
1295                 DIV_ROUND_UP(le32_to_cpu(cmd->common.cdw11), 1000);
1296
1297         dev_info(ctrl->device,
1298                  "keep alive interval updated from %u ms to %u ms\n",
1299                  ctrl->kato * 1000 / 2, new_kato * 1000 / 2);
1300
1301         nvme_stop_keep_alive(ctrl);
1302         ctrl->kato = new_kato;
1303         nvme_start_keep_alive(ctrl);
1304 }
1305
1306 /*
1307  * In NVMe 1.0 the CNS field was just a binary controller or namespace
1308  * flag, thus sending any new CNS opcodes has a big chance of not working.
1309  * Qemu unfortunately had that bug after reporting a 1.1 version compliance
1310  * (but not for any later version).
1311  */
1312 static bool nvme_ctrl_limited_cns(struct nvme_ctrl *ctrl)
1313 {
1314         if (ctrl->quirks & NVME_QUIRK_IDENTIFY_CNS)
1315                 return ctrl->vs < NVME_VS(1, 2, 0);
1316         return ctrl->vs < NVME_VS(1, 1, 0);
1317 }
1318
1319 static int nvme_identify_ctrl(struct nvme_ctrl *dev, struct nvme_id_ctrl **id)
1320 {
1321         struct nvme_command c = { };
1322         int error;
1323
1324         /* gcc-4.4.4 (at least) has issues with initializers and anon unions */
1325         c.identify.opcode = nvme_admin_identify;
1326         c.identify.cns = NVME_ID_CNS_CTRL;
1327
1328         *id = kmalloc(sizeof(struct nvme_id_ctrl), GFP_KERNEL);
1329         if (!*id)
1330                 return -ENOMEM;
1331
1332         error = nvme_submit_sync_cmd(dev->admin_q, &c, *id,
1333                         sizeof(struct nvme_id_ctrl));
1334         if (error)
1335                 kfree(*id);
1336         return error;
1337 }
1338
1339 static int nvme_process_ns_desc(struct nvme_ctrl *ctrl, struct nvme_ns_ids *ids,
1340                 struct nvme_ns_id_desc *cur, bool *csi_seen)
1341 {
1342         const char *warn_str = "ctrl returned bogus length:";
1343         void *data = cur;
1344
1345         switch (cur->nidt) {
1346         case NVME_NIDT_EUI64:
1347                 if (cur->nidl != NVME_NIDT_EUI64_LEN) {
1348                         dev_warn(ctrl->device, "%s %d for NVME_NIDT_EUI64\n",
1349                                  warn_str, cur->nidl);
1350                         return -1;
1351                 }
1352                 if (ctrl->quirks & NVME_QUIRK_BOGUS_NID)
1353                         return NVME_NIDT_EUI64_LEN;
1354                 memcpy(ids->eui64, data + sizeof(*cur), NVME_NIDT_EUI64_LEN);
1355                 return NVME_NIDT_EUI64_LEN;
1356         case NVME_NIDT_NGUID:
1357                 if (cur->nidl != NVME_NIDT_NGUID_LEN) {
1358                         dev_warn(ctrl->device, "%s %d for NVME_NIDT_NGUID\n",
1359                                  warn_str, cur->nidl);
1360                         return -1;
1361                 }
1362                 if (ctrl->quirks & NVME_QUIRK_BOGUS_NID)
1363                         return NVME_NIDT_NGUID_LEN;
1364                 memcpy(ids->nguid, data + sizeof(*cur), NVME_NIDT_NGUID_LEN);
1365                 return NVME_NIDT_NGUID_LEN;
1366         case NVME_NIDT_UUID:
1367                 if (cur->nidl != NVME_NIDT_UUID_LEN) {
1368                         dev_warn(ctrl->device, "%s %d for NVME_NIDT_UUID\n",
1369                                  warn_str, cur->nidl);
1370                         return -1;
1371                 }
1372                 if (ctrl->quirks & NVME_QUIRK_BOGUS_NID)
1373                         return NVME_NIDT_UUID_LEN;
1374                 uuid_copy(&ids->uuid, data + sizeof(*cur));
1375                 return NVME_NIDT_UUID_LEN;
1376         case NVME_NIDT_CSI:
1377                 if (cur->nidl != NVME_NIDT_CSI_LEN) {
1378                         dev_warn(ctrl->device, "%s %d for NVME_NIDT_CSI\n",
1379                                  warn_str, cur->nidl);
1380                         return -1;
1381                 }
1382                 memcpy(&ids->csi, data + sizeof(*cur), NVME_NIDT_CSI_LEN);
1383                 *csi_seen = true;
1384                 return NVME_NIDT_CSI_LEN;
1385         default:
1386                 /* Skip unknown types */
1387                 return cur->nidl;
1388         }
1389 }
1390
1391 static int nvme_identify_ns_descs(struct nvme_ctrl *ctrl,
1392                 struct nvme_ns_info *info)
1393 {
1394         struct nvme_command c = { };
1395         bool csi_seen = false;
1396         int status, pos, len;
1397         void *data;
1398
1399         if (ctrl->vs < NVME_VS(1, 3, 0) && !nvme_multi_css(ctrl))
1400                 return 0;
1401         if (ctrl->quirks & NVME_QUIRK_NO_NS_DESC_LIST)
1402                 return 0;
1403
1404         c.identify.opcode = nvme_admin_identify;
1405         c.identify.nsid = cpu_to_le32(info->nsid);
1406         c.identify.cns = NVME_ID_CNS_NS_DESC_LIST;
1407
1408         data = kzalloc(NVME_IDENTIFY_DATA_SIZE, GFP_KERNEL);
1409         if (!data)
1410                 return -ENOMEM;
1411
1412         status = nvme_submit_sync_cmd(ctrl->admin_q, &c, data,
1413                                       NVME_IDENTIFY_DATA_SIZE);
1414         if (status) {
1415                 dev_warn(ctrl->device,
1416                         "Identify Descriptors failed (nsid=%u, status=0x%x)\n",
1417                         info->nsid, status);
1418                 goto free_data;
1419         }
1420
1421         for (pos = 0; pos < NVME_IDENTIFY_DATA_SIZE; pos += len) {
1422                 struct nvme_ns_id_desc *cur = data + pos;
1423
1424                 if (cur->nidl == 0)
1425                         break;
1426
1427                 len = nvme_process_ns_desc(ctrl, &info->ids, cur, &csi_seen);
1428                 if (len < 0)
1429                         break;
1430
1431                 len += sizeof(*cur);
1432         }
1433
1434         if (nvme_multi_css(ctrl) && !csi_seen) {
1435                 dev_warn(ctrl->device, "Command set not reported for nsid:%d\n",
1436                          info->nsid);
1437                 status = -EINVAL;
1438         }
1439
1440 free_data:
1441         kfree(data);
1442         return status;
1443 }
1444
1445 static int nvme_identify_ns(struct nvme_ctrl *ctrl, unsigned nsid,
1446                         struct nvme_id_ns **id)
1447 {
1448         struct nvme_command c = { };
1449         int error;
1450
1451         /* gcc-4.4.4 (at least) has issues with initializers and anon unions */
1452         c.identify.opcode = nvme_admin_identify;
1453         c.identify.nsid = cpu_to_le32(nsid);
1454         c.identify.cns = NVME_ID_CNS_NS;
1455
1456         *id = kmalloc(sizeof(**id), GFP_KERNEL);
1457         if (!*id)
1458                 return -ENOMEM;
1459
1460         error = nvme_submit_sync_cmd(ctrl->admin_q, &c, *id, sizeof(**id));
1461         if (error) {
1462                 dev_warn(ctrl->device, "Identify namespace failed (%d)\n", error);
1463                 kfree(*id);
1464         }
1465         return error;
1466 }
1467
1468 static int nvme_ns_info_from_identify(struct nvme_ctrl *ctrl,
1469                 struct nvme_ns_info *info)
1470 {
1471         struct nvme_ns_ids *ids = &info->ids;
1472         struct nvme_id_ns *id;
1473         int ret;
1474
1475         ret = nvme_identify_ns(ctrl, info->nsid, &id);
1476         if (ret)
1477                 return ret;
1478
1479         if (id->ncap == 0) {
1480                 /* namespace not allocated or attached */
1481                 info->is_removed = true;
1482                 return -ENODEV;
1483         }
1484
1485         info->anagrpid = id->anagrpid;
1486         info->is_shared = id->nmic & NVME_NS_NMIC_SHARED;
1487         info->is_readonly = id->nsattr & NVME_NS_ATTR_RO;
1488         info->is_ready = true;
1489         if (ctrl->quirks & NVME_QUIRK_BOGUS_NID) {
1490                 dev_info(ctrl->device,
1491                          "Ignoring bogus Namespace Identifiers\n");
1492         } else {
1493                 if (ctrl->vs >= NVME_VS(1, 1, 0) &&
1494                     !memchr_inv(ids->eui64, 0, sizeof(ids->eui64)))
1495                         memcpy(ids->eui64, id->eui64, sizeof(ids->eui64));
1496                 if (ctrl->vs >= NVME_VS(1, 2, 0) &&
1497                     !memchr_inv(ids->nguid, 0, sizeof(ids->nguid)))
1498                         memcpy(ids->nguid, id->nguid, sizeof(ids->nguid));
1499         }
1500         kfree(id);
1501         return 0;
1502 }
1503
1504 static int nvme_ns_info_from_id_cs_indep(struct nvme_ctrl *ctrl,
1505                 struct nvme_ns_info *info)
1506 {
1507         struct nvme_id_ns_cs_indep *id;
1508         struct nvme_command c = {
1509                 .identify.opcode        = nvme_admin_identify,
1510                 .identify.nsid          = cpu_to_le32(info->nsid),
1511                 .identify.cns           = NVME_ID_CNS_NS_CS_INDEP,
1512         };
1513         int ret;
1514
1515         id = kmalloc(sizeof(*id), GFP_KERNEL);
1516         if (!id)
1517                 return -ENOMEM;
1518
1519         ret = nvme_submit_sync_cmd(ctrl->admin_q, &c, id, sizeof(*id));
1520         if (!ret) {
1521                 info->anagrpid = id->anagrpid;
1522                 info->is_shared = id->nmic & NVME_NS_NMIC_SHARED;
1523                 info->is_readonly = id->nsattr & NVME_NS_ATTR_RO;
1524                 info->is_ready = id->nstat & NVME_NSTAT_NRDY;
1525         }
1526         kfree(id);
1527         return ret;
1528 }
1529
1530 static int nvme_features(struct nvme_ctrl *dev, u8 op, unsigned int fid,
1531                 unsigned int dword11, void *buffer, size_t buflen, u32 *result)
1532 {
1533         union nvme_result res = { 0 };
1534         struct nvme_command c = { };
1535         int ret;
1536
1537         c.features.opcode = op;
1538         c.features.fid = cpu_to_le32(fid);
1539         c.features.dword11 = cpu_to_le32(dword11);
1540
1541         ret = __nvme_submit_sync_cmd(dev->admin_q, &c, &res,
1542                         buffer, buflen, NVME_QID_ANY, 0, 0);
1543         if (ret >= 0 && result)
1544                 *result = le32_to_cpu(res.u32);
1545         return ret;
1546 }
1547
1548 int nvme_set_features(struct nvme_ctrl *dev, unsigned int fid,
1549                       unsigned int dword11, void *buffer, size_t buflen,
1550                       u32 *result)
1551 {
1552         return nvme_features(dev, nvme_admin_set_features, fid, dword11, buffer,
1553                              buflen, result);
1554 }
1555 EXPORT_SYMBOL_GPL(nvme_set_features);
1556
1557 int nvme_get_features(struct nvme_ctrl *dev, unsigned int fid,
1558                       unsigned int dword11, void *buffer, size_t buflen,
1559                       u32 *result)
1560 {
1561         return nvme_features(dev, nvme_admin_get_features, fid, dword11, buffer,
1562                              buflen, result);
1563 }
1564 EXPORT_SYMBOL_GPL(nvme_get_features);
1565
1566 int nvme_set_queue_count(struct nvme_ctrl *ctrl, int *count)
1567 {
1568         u32 q_count = (*count - 1) | ((*count - 1) << 16);
1569         u32 result;
1570         int status, nr_io_queues;
1571
1572         status = nvme_set_features(ctrl, NVME_FEAT_NUM_QUEUES, q_count, NULL, 0,
1573                         &result);
1574         if (status < 0)
1575                 return status;
1576
1577         /*
1578          * Degraded controllers might return an error when setting the queue
1579          * count.  We still want to be able to bring them online and offer
1580          * access to the admin queue, as that might be only way to fix them up.
1581          */
1582         if (status > 0) {
1583                 dev_err(ctrl->device, "Could not set queue count (%d)\n", status);
1584                 *count = 0;
1585         } else {
1586                 nr_io_queues = min(result & 0xffff, result >> 16) + 1;
1587                 *count = min(*count, nr_io_queues);
1588         }
1589
1590         return 0;
1591 }
1592 EXPORT_SYMBOL_GPL(nvme_set_queue_count);
1593
1594 #define NVME_AEN_SUPPORTED \
1595         (NVME_AEN_CFG_NS_ATTR | NVME_AEN_CFG_FW_ACT | \
1596          NVME_AEN_CFG_ANA_CHANGE | NVME_AEN_CFG_DISC_CHANGE)
1597
1598 static void nvme_enable_aen(struct nvme_ctrl *ctrl)
1599 {
1600         u32 result, supported_aens = ctrl->oaes & NVME_AEN_SUPPORTED;
1601         int status;
1602
1603         if (!supported_aens)
1604                 return;
1605
1606         status = nvme_set_features(ctrl, NVME_FEAT_ASYNC_EVENT, supported_aens,
1607                         NULL, 0, &result);
1608         if (status)
1609                 dev_warn(ctrl->device, "Failed to configure AEN (cfg %x)\n",
1610                          supported_aens);
1611
1612         queue_work(nvme_wq, &ctrl->async_event_work);
1613 }
1614
1615 static int nvme_ns_open(struct nvme_ns *ns)
1616 {
1617
1618         /* should never be called due to GENHD_FL_HIDDEN */
1619         if (WARN_ON_ONCE(nvme_ns_head_multipath(ns->head)))
1620                 goto fail;
1621         if (!nvme_get_ns(ns))
1622                 goto fail;
1623         if (!try_module_get(ns->ctrl->ops->module))
1624                 goto fail_put_ns;
1625
1626         return 0;
1627
1628 fail_put_ns:
1629         nvme_put_ns(ns);
1630 fail:
1631         return -ENXIO;
1632 }
1633
1634 static void nvme_ns_release(struct nvme_ns *ns)
1635 {
1636
1637         module_put(ns->ctrl->ops->module);
1638         nvme_put_ns(ns);
1639 }
1640
1641 static int nvme_open(struct gendisk *disk, blk_mode_t mode)
1642 {
1643         return nvme_ns_open(disk->private_data);
1644 }
1645
1646 static void nvme_release(struct gendisk *disk)
1647 {
1648         nvme_ns_release(disk->private_data);
1649 }
1650
1651 int nvme_getgeo(struct block_device *bdev, struct hd_geometry *geo)
1652 {
1653         /* some standard values */
1654         geo->heads = 1 << 6;
1655         geo->sectors = 1 << 5;
1656         geo->cylinders = get_capacity(bdev->bd_disk) >> 11;
1657         return 0;
1658 }
1659
1660 #ifdef CONFIG_BLK_DEV_INTEGRITY
1661 static void nvme_init_integrity(struct gendisk *disk, struct nvme_ns *ns,
1662                                 u32 max_integrity_segments)
1663 {
1664         struct blk_integrity integrity = { };
1665
1666         switch (ns->pi_type) {
1667         case NVME_NS_DPS_PI_TYPE3:
1668                 switch (ns->guard_type) {
1669                 case NVME_NVM_NS_16B_GUARD:
1670                         integrity.profile = &t10_pi_type3_crc;
1671                         integrity.tag_size = sizeof(u16) + sizeof(u32);
1672                         integrity.flags |= BLK_INTEGRITY_DEVICE_CAPABLE;
1673                         break;
1674                 case NVME_NVM_NS_64B_GUARD:
1675                         integrity.profile = &ext_pi_type3_crc64;
1676                         integrity.tag_size = sizeof(u16) + 6;
1677                         integrity.flags |= BLK_INTEGRITY_DEVICE_CAPABLE;
1678                         break;
1679                 default:
1680                         integrity.profile = NULL;
1681                         break;
1682                 }
1683                 break;
1684         case NVME_NS_DPS_PI_TYPE1:
1685         case NVME_NS_DPS_PI_TYPE2:
1686                 switch (ns->guard_type) {
1687                 case NVME_NVM_NS_16B_GUARD:
1688                         integrity.profile = &t10_pi_type1_crc;
1689                         integrity.tag_size = sizeof(u16);
1690                         integrity.flags |= BLK_INTEGRITY_DEVICE_CAPABLE;
1691                         break;
1692                 case NVME_NVM_NS_64B_GUARD:
1693                         integrity.profile = &ext_pi_type1_crc64;
1694                         integrity.tag_size = sizeof(u16);
1695                         integrity.flags |= BLK_INTEGRITY_DEVICE_CAPABLE;
1696                         break;
1697                 default:
1698                         integrity.profile = NULL;
1699                         break;
1700                 }
1701                 break;
1702         default:
1703                 integrity.profile = NULL;
1704                 break;
1705         }
1706
1707         integrity.tuple_size = ns->ms;
1708         blk_integrity_register(disk, &integrity);
1709         blk_queue_max_integrity_segments(disk->queue, max_integrity_segments);
1710 }
1711 #else
1712 static void nvme_init_integrity(struct gendisk *disk, struct nvme_ns *ns,
1713                                 u32 max_integrity_segments)
1714 {
1715 }
1716 #endif /* CONFIG_BLK_DEV_INTEGRITY */
1717
1718 static void nvme_config_discard(struct gendisk *disk, struct nvme_ns *ns)
1719 {
1720         struct nvme_ctrl *ctrl = ns->ctrl;
1721         struct request_queue *queue = disk->queue;
1722         u32 size = queue_logical_block_size(queue);
1723
1724         if (ctrl->dmrsl && ctrl->dmrsl <= nvme_sect_to_lba(ns, UINT_MAX))
1725                 ctrl->max_discard_sectors = nvme_lba_to_sect(ns, ctrl->dmrsl);
1726
1727         if (ctrl->max_discard_sectors == 0) {
1728                 blk_queue_max_discard_sectors(queue, 0);
1729                 return;
1730         }
1731
1732         BUILD_BUG_ON(PAGE_SIZE / sizeof(struct nvme_dsm_range) <
1733                         NVME_DSM_MAX_RANGES);
1734
1735         queue->limits.discard_granularity = size;
1736
1737         /* If discard is already enabled, don't reset queue limits */
1738         if (queue->limits.max_discard_sectors)
1739                 return;
1740
1741         blk_queue_max_discard_sectors(queue, ctrl->max_discard_sectors);
1742         blk_queue_max_discard_segments(queue, ctrl->max_discard_segments);
1743
1744         if (ctrl->quirks & NVME_QUIRK_DEALLOCATE_ZEROES)
1745                 blk_queue_max_write_zeroes_sectors(queue, UINT_MAX);
1746 }
1747
1748 static bool nvme_ns_ids_equal(struct nvme_ns_ids *a, struct nvme_ns_ids *b)
1749 {
1750         return uuid_equal(&a->uuid, &b->uuid) &&
1751                 memcmp(&a->nguid, &b->nguid, sizeof(a->nguid)) == 0 &&
1752                 memcmp(&a->eui64, &b->eui64, sizeof(a->eui64)) == 0 &&
1753                 a->csi == b->csi;
1754 }
1755
1756 static int nvme_init_ms(struct nvme_ns *ns, struct nvme_id_ns *id)
1757 {
1758         bool first = id->dps & NVME_NS_DPS_PI_FIRST;
1759         unsigned lbaf = nvme_lbaf_index(id->flbas);
1760         struct nvme_ctrl *ctrl = ns->ctrl;
1761         struct nvme_command c = { };
1762         struct nvme_id_ns_nvm *nvm;
1763         int ret = 0;
1764         u32 elbaf;
1765
1766         ns->pi_size = 0;
1767         ns->ms = le16_to_cpu(id->lbaf[lbaf].ms);
1768         if (!(ctrl->ctratt & NVME_CTRL_ATTR_ELBAS)) {
1769                 ns->pi_size = sizeof(struct t10_pi_tuple);
1770                 ns->guard_type = NVME_NVM_NS_16B_GUARD;
1771                 goto set_pi;
1772         }
1773
1774         nvm = kzalloc(sizeof(*nvm), GFP_KERNEL);
1775         if (!nvm)
1776                 return -ENOMEM;
1777
1778         c.identify.opcode = nvme_admin_identify;
1779         c.identify.nsid = cpu_to_le32(ns->head->ns_id);
1780         c.identify.cns = NVME_ID_CNS_CS_NS;
1781         c.identify.csi = NVME_CSI_NVM;
1782
1783         ret = nvme_submit_sync_cmd(ns->ctrl->admin_q, &c, nvm, sizeof(*nvm));
1784         if (ret)
1785                 goto free_data;
1786
1787         elbaf = le32_to_cpu(nvm->elbaf[lbaf]);
1788
1789         /* no support for storage tag formats right now */
1790         if (nvme_elbaf_sts(elbaf))
1791                 goto free_data;
1792
1793         ns->guard_type = nvme_elbaf_guard_type(elbaf);
1794         switch (ns->guard_type) {
1795         case NVME_NVM_NS_64B_GUARD:
1796                 ns->pi_size = sizeof(struct crc64_pi_tuple);
1797                 break;
1798         case NVME_NVM_NS_16B_GUARD:
1799                 ns->pi_size = sizeof(struct t10_pi_tuple);
1800                 break;
1801         default:
1802                 break;
1803         }
1804
1805 free_data:
1806         kfree(nvm);
1807 set_pi:
1808         if (ns->pi_size && (first || ns->ms == ns->pi_size))
1809                 ns->pi_type = id->dps & NVME_NS_DPS_PI_MASK;
1810         else
1811                 ns->pi_type = 0;
1812
1813         return ret;
1814 }
1815
1816 static void nvme_configure_metadata(struct nvme_ns *ns, struct nvme_id_ns *id)
1817 {
1818         struct nvme_ctrl *ctrl = ns->ctrl;
1819
1820         if (nvme_init_ms(ns, id))
1821                 return;
1822
1823         ns->features &= ~(NVME_NS_METADATA_SUPPORTED | NVME_NS_EXT_LBAS);
1824         if (!ns->ms || !(ctrl->ops->flags & NVME_F_METADATA_SUPPORTED))
1825                 return;
1826
1827         if (ctrl->ops->flags & NVME_F_FABRICS) {
1828                 /*
1829                  * The NVMe over Fabrics specification only supports metadata as
1830                  * part of the extended data LBA.  We rely on HCA/HBA support to
1831                  * remap the separate metadata buffer from the block layer.
1832                  */
1833                 if (WARN_ON_ONCE(!(id->flbas & NVME_NS_FLBAS_META_EXT)))
1834                         return;
1835
1836                 ns->features |= NVME_NS_EXT_LBAS;
1837
1838                 /*
1839                  * The current fabrics transport drivers support namespace
1840                  * metadata formats only if nvme_ns_has_pi() returns true.
1841                  * Suppress support for all other formats so the namespace will
1842                  * have a 0 capacity and not be usable through the block stack.
1843                  *
1844                  * Note, this check will need to be modified if any drivers
1845                  * gain the ability to use other metadata formats.
1846                  */
1847                 if (ctrl->max_integrity_segments && nvme_ns_has_pi(ns))
1848                         ns->features |= NVME_NS_METADATA_SUPPORTED;
1849         } else {
1850                 /*
1851                  * For PCIe controllers, we can't easily remap the separate
1852                  * metadata buffer from the block layer and thus require a
1853                  * separate metadata buffer for block layer metadata/PI support.
1854                  * We allow extended LBAs for the passthrough interface, though.
1855                  */
1856                 if (id->flbas & NVME_NS_FLBAS_META_EXT)
1857                         ns->features |= NVME_NS_EXT_LBAS;
1858                 else
1859                         ns->features |= NVME_NS_METADATA_SUPPORTED;
1860         }
1861 }
1862
1863 static void nvme_set_queue_limits(struct nvme_ctrl *ctrl,
1864                 struct request_queue *q)
1865 {
1866         bool vwc = ctrl->vwc & NVME_CTRL_VWC_PRESENT;
1867
1868         if (ctrl->max_hw_sectors) {
1869                 u32 max_segments =
1870                         (ctrl->max_hw_sectors / (NVME_CTRL_PAGE_SIZE >> 9)) + 1;
1871
1872                 max_segments = min_not_zero(max_segments, ctrl->max_segments);
1873                 blk_queue_max_hw_sectors(q, ctrl->max_hw_sectors);
1874                 blk_queue_max_segments(q, min_t(u32, max_segments, USHRT_MAX));
1875         }
1876         blk_queue_virt_boundary(q, NVME_CTRL_PAGE_SIZE - 1);
1877         blk_queue_dma_alignment(q, 3);
1878         blk_queue_write_cache(q, vwc, vwc);
1879 }
1880
1881 static void nvme_update_disk_info(struct gendisk *disk,
1882                 struct nvme_ns *ns, struct nvme_id_ns *id)
1883 {
1884         sector_t capacity = nvme_lba_to_sect(ns, le64_to_cpu(id->nsze));
1885         u32 bs = 1U << ns->lba_shift;
1886         u32 atomic_bs, phys_bs, io_opt = 0;
1887
1888         /*
1889          * The block layer can't support LBA sizes larger than the page size
1890          * yet, so catch this early and don't allow block I/O.
1891          */
1892         if (ns->lba_shift > PAGE_SHIFT) {
1893                 capacity = 0;
1894                 bs = (1 << 9);
1895         }
1896
1897         blk_integrity_unregister(disk);
1898
1899         atomic_bs = phys_bs = bs;
1900         if (id->nabo == 0) {
1901                 /*
1902                  * Bit 1 indicates whether NAWUPF is defined for this namespace
1903                  * and whether it should be used instead of AWUPF. If NAWUPF ==
1904                  * 0 then AWUPF must be used instead.
1905                  */
1906                 if (id->nsfeat & NVME_NS_FEAT_ATOMICS && id->nawupf)
1907                         atomic_bs = (1 + le16_to_cpu(id->nawupf)) * bs;
1908                 else
1909                         atomic_bs = (1 + ns->ctrl->subsys->awupf) * bs;
1910         }
1911
1912         if (id->nsfeat & NVME_NS_FEAT_IO_OPT) {
1913                 /* NPWG = Namespace Preferred Write Granularity */
1914                 phys_bs = bs * (1 + le16_to_cpu(id->npwg));
1915                 /* NOWS = Namespace Optimal Write Size */
1916                 io_opt = bs * (1 + le16_to_cpu(id->nows));
1917         }
1918
1919         blk_queue_logical_block_size(disk->queue, bs);
1920         /*
1921          * Linux filesystems assume writing a single physical block is
1922          * an atomic operation. Hence limit the physical block size to the
1923          * value of the Atomic Write Unit Power Fail parameter.
1924          */
1925         blk_queue_physical_block_size(disk->queue, min(phys_bs, atomic_bs));
1926         blk_queue_io_min(disk->queue, phys_bs);
1927         blk_queue_io_opt(disk->queue, io_opt);
1928
1929         /*
1930          * Register a metadata profile for PI, or the plain non-integrity NVMe
1931          * metadata masquerading as Type 0 if supported, otherwise reject block
1932          * I/O to namespaces with metadata except when the namespace supports
1933          * PI, as it can strip/insert in that case.
1934          */
1935         if (ns->ms) {
1936                 if (IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY) &&
1937                     (ns->features & NVME_NS_METADATA_SUPPORTED))
1938                         nvme_init_integrity(disk, ns,
1939                                             ns->ctrl->max_integrity_segments);
1940                 else if (!nvme_ns_has_pi(ns))
1941                         capacity = 0;
1942         }
1943
1944         set_capacity_and_notify(disk, capacity);
1945
1946         nvme_config_discard(disk, ns);
1947         blk_queue_max_write_zeroes_sectors(disk->queue,
1948                                            ns->ctrl->max_zeroes_sectors);
1949 }
1950
1951 static bool nvme_ns_is_readonly(struct nvme_ns *ns, struct nvme_ns_info *info)
1952 {
1953         return info->is_readonly || test_bit(NVME_NS_FORCE_RO, &ns->flags);
1954 }
1955
1956 static inline bool nvme_first_scan(struct gendisk *disk)
1957 {
1958         /* nvme_alloc_ns() scans the disk prior to adding it */
1959         return !disk_live(disk);
1960 }
1961
1962 static void nvme_set_chunk_sectors(struct nvme_ns *ns, struct nvme_id_ns *id)
1963 {
1964         struct nvme_ctrl *ctrl = ns->ctrl;
1965         u32 iob;
1966
1967         if ((ctrl->quirks & NVME_QUIRK_STRIPE_SIZE) &&
1968             is_power_of_2(ctrl->max_hw_sectors))
1969                 iob = ctrl->max_hw_sectors;
1970         else
1971                 iob = nvme_lba_to_sect(ns, le16_to_cpu(id->noiob));
1972
1973         if (!iob)
1974                 return;
1975
1976         if (!is_power_of_2(iob)) {
1977                 if (nvme_first_scan(ns->disk))
1978                         pr_warn("%s: ignoring unaligned IO boundary:%u\n",
1979                                 ns->disk->disk_name, iob);
1980                 return;
1981         }
1982
1983         if (blk_queue_is_zoned(ns->disk->queue)) {
1984                 if (nvme_first_scan(ns->disk))
1985                         pr_warn("%s: ignoring zoned namespace IO boundary\n",
1986                                 ns->disk->disk_name);
1987                 return;
1988         }
1989
1990         blk_queue_chunk_sectors(ns->queue, iob);
1991 }
1992
1993 static int nvme_update_ns_info_generic(struct nvme_ns *ns,
1994                 struct nvme_ns_info *info)
1995 {
1996         blk_mq_freeze_queue(ns->disk->queue);
1997         nvme_set_queue_limits(ns->ctrl, ns->queue);
1998         set_disk_ro(ns->disk, nvme_ns_is_readonly(ns, info));
1999         blk_mq_unfreeze_queue(ns->disk->queue);
2000
2001         if (nvme_ns_head_multipath(ns->head)) {
2002                 blk_mq_freeze_queue(ns->head->disk->queue);
2003                 set_disk_ro(ns->head->disk, nvme_ns_is_readonly(ns, info));
2004                 nvme_mpath_revalidate_paths(ns);
2005                 blk_stack_limits(&ns->head->disk->queue->limits,
2006                                  &ns->queue->limits, 0);
2007                 ns->head->disk->flags |= GENHD_FL_HIDDEN;
2008                 blk_mq_unfreeze_queue(ns->head->disk->queue);
2009         }
2010
2011         /* Hide the block-interface for these devices */
2012         ns->disk->flags |= GENHD_FL_HIDDEN;
2013         set_bit(NVME_NS_READY, &ns->flags);
2014
2015         return 0;
2016 }
2017
2018 static int nvme_update_ns_info_block(struct nvme_ns *ns,
2019                 struct nvme_ns_info *info)
2020 {
2021         struct nvme_id_ns *id;
2022         unsigned lbaf;
2023         int ret;
2024
2025         ret = nvme_identify_ns(ns->ctrl, info->nsid, &id);
2026         if (ret)
2027                 return ret;
2028
2029         blk_mq_freeze_queue(ns->disk->queue);
2030         lbaf = nvme_lbaf_index(id->flbas);
2031         ns->lba_shift = id->lbaf[lbaf].ds;
2032         nvme_set_queue_limits(ns->ctrl, ns->queue);
2033
2034         nvme_configure_metadata(ns, id);
2035         nvme_set_chunk_sectors(ns, id);
2036         nvme_update_disk_info(ns->disk, ns, id);
2037
2038         if (ns->head->ids.csi == NVME_CSI_ZNS) {
2039                 ret = nvme_update_zone_info(ns, lbaf);
2040                 if (ret) {
2041                         blk_mq_unfreeze_queue(ns->disk->queue);
2042                         goto out;
2043                 }
2044         }
2045
2046         /*
2047          * Only set the DEAC bit if the device guarantees that reads from
2048          * deallocated data return zeroes.  While the DEAC bit does not
2049          * require that, it must be a no-op if reads from deallocated data
2050          * do not return zeroes.
2051          */
2052         if ((id->dlfeat & 0x7) == 0x1 && (id->dlfeat & (1 << 3)))
2053                 ns->features |= NVME_NS_DEAC;
2054         set_disk_ro(ns->disk, nvme_ns_is_readonly(ns, info));
2055         set_bit(NVME_NS_READY, &ns->flags);
2056         blk_mq_unfreeze_queue(ns->disk->queue);
2057
2058         if (blk_queue_is_zoned(ns->queue)) {
2059                 ret = nvme_revalidate_zones(ns);
2060                 if (ret && !nvme_first_scan(ns->disk))
2061                         goto out;
2062         }
2063
2064         if (nvme_ns_head_multipath(ns->head)) {
2065                 blk_mq_freeze_queue(ns->head->disk->queue);
2066                 nvme_update_disk_info(ns->head->disk, ns, id);
2067                 set_disk_ro(ns->head->disk, nvme_ns_is_readonly(ns, info));
2068                 nvme_mpath_revalidate_paths(ns);
2069                 blk_stack_limits(&ns->head->disk->queue->limits,
2070                                  &ns->queue->limits, 0);
2071                 disk_update_readahead(ns->head->disk);
2072                 blk_mq_unfreeze_queue(ns->head->disk->queue);
2073         }
2074
2075         ret = 0;
2076 out:
2077         /*
2078          * If probing fails due an unsupported feature, hide the block device,
2079          * but still allow other access.
2080          */
2081         if (ret == -ENODEV) {
2082                 ns->disk->flags |= GENHD_FL_HIDDEN;
2083                 set_bit(NVME_NS_READY, &ns->flags);
2084                 ret = 0;
2085         }
2086         kfree(id);
2087         return ret;
2088 }
2089
2090 static int nvme_update_ns_info(struct nvme_ns *ns, struct nvme_ns_info *info)
2091 {
2092         switch (info->ids.csi) {
2093         case NVME_CSI_ZNS:
2094                 if (!IS_ENABLED(CONFIG_BLK_DEV_ZONED)) {
2095                         dev_info(ns->ctrl->device,
2096         "block device for nsid %u not supported without CONFIG_BLK_DEV_ZONED\n",
2097                                 info->nsid);
2098                         return nvme_update_ns_info_generic(ns, info);
2099                 }
2100                 return nvme_update_ns_info_block(ns, info);
2101         case NVME_CSI_NVM:
2102                 return nvme_update_ns_info_block(ns, info);
2103         default:
2104                 dev_info(ns->ctrl->device,
2105                         "block device for nsid %u not supported (csi %u)\n",
2106                         info->nsid, info->ids.csi);
2107                 return nvme_update_ns_info_generic(ns, info);
2108         }
2109 }
2110
2111 #ifdef CONFIG_BLK_SED_OPAL
2112 static int nvme_sec_submit(void *data, u16 spsp, u8 secp, void *buffer, size_t len,
2113                 bool send)
2114 {
2115         struct nvme_ctrl *ctrl = data;
2116         struct nvme_command cmd = { };
2117
2118         if (send)
2119                 cmd.common.opcode = nvme_admin_security_send;
2120         else
2121                 cmd.common.opcode = nvme_admin_security_recv;
2122         cmd.common.nsid = 0;
2123         cmd.common.cdw10 = cpu_to_le32(((u32)secp) << 24 | ((u32)spsp) << 8);
2124         cmd.common.cdw11 = cpu_to_le32(len);
2125
2126         return __nvme_submit_sync_cmd(ctrl->admin_q, &cmd, NULL, buffer, len,
2127                         NVME_QID_ANY, 1, 0);
2128 }
2129
2130 static void nvme_configure_opal(struct nvme_ctrl *ctrl, bool was_suspended)
2131 {
2132         if (ctrl->oacs & NVME_CTRL_OACS_SEC_SUPP) {
2133                 if (!ctrl->opal_dev)
2134                         ctrl->opal_dev = init_opal_dev(ctrl, &nvme_sec_submit);
2135                 else if (was_suspended)
2136                         opal_unlock_from_suspend(ctrl->opal_dev);
2137         } else {
2138                 free_opal_dev(ctrl->opal_dev);
2139                 ctrl->opal_dev = NULL;
2140         }
2141 }
2142 #else
2143 static void nvme_configure_opal(struct nvme_ctrl *ctrl, bool was_suspended)
2144 {
2145 }
2146 #endif /* CONFIG_BLK_SED_OPAL */
2147
2148 #ifdef CONFIG_BLK_DEV_ZONED
2149 static int nvme_report_zones(struct gendisk *disk, sector_t sector,
2150                 unsigned int nr_zones, report_zones_cb cb, void *data)
2151 {
2152         return nvme_ns_report_zones(disk->private_data, sector, nr_zones, cb,
2153                         data);
2154 }
2155 #else
2156 #define nvme_report_zones       NULL
2157 #endif /* CONFIG_BLK_DEV_ZONED */
2158
2159 const struct block_device_operations nvme_bdev_ops = {
2160         .owner          = THIS_MODULE,
2161         .ioctl          = nvme_ioctl,
2162         .compat_ioctl   = blkdev_compat_ptr_ioctl,
2163         .open           = nvme_open,
2164         .release        = nvme_release,
2165         .getgeo         = nvme_getgeo,
2166         .report_zones   = nvme_report_zones,
2167         .pr_ops         = &nvme_pr_ops,
2168 };
2169
2170 static int nvme_wait_ready(struct nvme_ctrl *ctrl, u32 mask, u32 val,
2171                 u32 timeout, const char *op)
2172 {
2173         unsigned long timeout_jiffies = jiffies + timeout * HZ;
2174         u32 csts;
2175         int ret;
2176
2177         while ((ret = ctrl->ops->reg_read32(ctrl, NVME_REG_CSTS, &csts)) == 0) {
2178                 if (csts == ~0)
2179                         return -ENODEV;
2180                 if ((csts & mask) == val)
2181                         break;
2182
2183                 usleep_range(1000, 2000);
2184                 if (fatal_signal_pending(current))
2185                         return -EINTR;
2186                 if (time_after(jiffies, timeout_jiffies)) {
2187                         dev_err(ctrl->device,
2188                                 "Device not ready; aborting %s, CSTS=0x%x\n",
2189                                 op, csts);
2190                         return -ENODEV;
2191                 }
2192         }
2193
2194         return ret;
2195 }
2196
2197 int nvme_disable_ctrl(struct nvme_ctrl *ctrl, bool shutdown)
2198 {
2199         int ret;
2200
2201         ctrl->ctrl_config &= ~NVME_CC_SHN_MASK;
2202         if (shutdown)
2203                 ctrl->ctrl_config |= NVME_CC_SHN_NORMAL;
2204         else
2205                 ctrl->ctrl_config &= ~NVME_CC_ENABLE;
2206
2207         ret = ctrl->ops->reg_write32(ctrl, NVME_REG_CC, ctrl->ctrl_config);
2208         if (ret)
2209                 return ret;
2210
2211         if (shutdown) {
2212                 return nvme_wait_ready(ctrl, NVME_CSTS_SHST_MASK,
2213                                        NVME_CSTS_SHST_CMPLT,
2214                                        ctrl->shutdown_timeout, "shutdown");
2215         }
2216         if (ctrl->quirks & NVME_QUIRK_DELAY_BEFORE_CHK_RDY)
2217                 msleep(NVME_QUIRK_DELAY_AMOUNT);
2218         return nvme_wait_ready(ctrl, NVME_CSTS_RDY, 0,
2219                                (NVME_CAP_TIMEOUT(ctrl->cap) + 1) / 2, "reset");
2220 }
2221 EXPORT_SYMBOL_GPL(nvme_disable_ctrl);
2222
2223 int nvme_enable_ctrl(struct nvme_ctrl *ctrl)
2224 {
2225         unsigned dev_page_min;
2226         u32 timeout;
2227         int ret;
2228
2229         ret = ctrl->ops->reg_read64(ctrl, NVME_REG_CAP, &ctrl->cap);
2230         if (ret) {
2231                 dev_err(ctrl->device, "Reading CAP failed (%d)\n", ret);
2232                 return ret;
2233         }
2234         dev_page_min = NVME_CAP_MPSMIN(ctrl->cap) + 12;
2235
2236         if (NVME_CTRL_PAGE_SHIFT < dev_page_min) {
2237                 dev_err(ctrl->device,
2238                         "Minimum device page size %u too large for host (%u)\n",
2239                         1 << dev_page_min, 1 << NVME_CTRL_PAGE_SHIFT);
2240                 return -ENODEV;
2241         }
2242
2243         if (NVME_CAP_CSS(ctrl->cap) & NVME_CAP_CSS_CSI)
2244                 ctrl->ctrl_config = NVME_CC_CSS_CSI;
2245         else
2246                 ctrl->ctrl_config = NVME_CC_CSS_NVM;
2247
2248         if (ctrl->cap & NVME_CAP_CRMS_CRWMS && ctrl->cap & NVME_CAP_CRMS_CRIMS)
2249                 ctrl->ctrl_config |= NVME_CC_CRIME;
2250
2251         ctrl->ctrl_config |= (NVME_CTRL_PAGE_SHIFT - 12) << NVME_CC_MPS_SHIFT;
2252         ctrl->ctrl_config |= NVME_CC_AMS_RR | NVME_CC_SHN_NONE;
2253         ctrl->ctrl_config |= NVME_CC_IOSQES | NVME_CC_IOCQES;
2254         ret = ctrl->ops->reg_write32(ctrl, NVME_REG_CC, ctrl->ctrl_config);
2255         if (ret)
2256                 return ret;
2257
2258         /* Flush write to device (required if transport is PCI) */
2259         ret = ctrl->ops->reg_read32(ctrl, NVME_REG_CC, &ctrl->ctrl_config);
2260         if (ret)
2261                 return ret;
2262
2263         /* CAP value may change after initial CC write */
2264         ret = ctrl->ops->reg_read64(ctrl, NVME_REG_CAP, &ctrl->cap);
2265         if (ret)
2266                 return ret;
2267
2268         timeout = NVME_CAP_TIMEOUT(ctrl->cap);
2269         if (ctrl->cap & NVME_CAP_CRMS_CRWMS) {
2270                 u32 crto, ready_timeout;
2271
2272                 ret = ctrl->ops->reg_read32(ctrl, NVME_REG_CRTO, &crto);
2273                 if (ret) {
2274                         dev_err(ctrl->device, "Reading CRTO failed (%d)\n",
2275                                 ret);
2276                         return ret;
2277                 }
2278
2279                 /*
2280                  * CRTO should always be greater or equal to CAP.TO, but some
2281                  * devices are known to get this wrong. Use the larger of the
2282                  * two values.
2283                  */
2284                 if (ctrl->ctrl_config & NVME_CC_CRIME)
2285                         ready_timeout = NVME_CRTO_CRIMT(crto);
2286                 else
2287                         ready_timeout = NVME_CRTO_CRWMT(crto);
2288
2289                 if (ready_timeout < timeout)
2290                         dev_warn_once(ctrl->device, "bad crto:%x cap:%llx\n",
2291                                       crto, ctrl->cap);
2292                 else
2293                         timeout = ready_timeout;
2294         }
2295
2296         ctrl->ctrl_config |= NVME_CC_ENABLE;
2297         ret = ctrl->ops->reg_write32(ctrl, NVME_REG_CC, ctrl->ctrl_config);
2298         if (ret)
2299                 return ret;
2300         return nvme_wait_ready(ctrl, NVME_CSTS_RDY, NVME_CSTS_RDY,
2301                                (timeout + 1) / 2, "initialisation");
2302 }
2303 EXPORT_SYMBOL_GPL(nvme_enable_ctrl);
2304
2305 static int nvme_configure_timestamp(struct nvme_ctrl *ctrl)
2306 {
2307         __le64 ts;
2308         int ret;
2309
2310         if (!(ctrl->oncs & NVME_CTRL_ONCS_TIMESTAMP))
2311                 return 0;
2312
2313         ts = cpu_to_le64(ktime_to_ms(ktime_get_real()));
2314         ret = nvme_set_features(ctrl, NVME_FEAT_TIMESTAMP, 0, &ts, sizeof(ts),
2315                         NULL);
2316         if (ret)
2317                 dev_warn_once(ctrl->device,
2318                         "could not set timestamp (%d)\n", ret);
2319         return ret;
2320 }
2321
2322 static int nvme_configure_host_options(struct nvme_ctrl *ctrl)
2323 {
2324         struct nvme_feat_host_behavior *host;
2325         u8 acre = 0, lbafee = 0;
2326         int ret;
2327
2328         /* Don't bother enabling the feature if retry delay is not reported */
2329         if (ctrl->crdt[0])
2330                 acre = NVME_ENABLE_ACRE;
2331         if (ctrl->ctratt & NVME_CTRL_ATTR_ELBAS)
2332                 lbafee = NVME_ENABLE_LBAFEE;
2333
2334         if (!acre && !lbafee)
2335                 return 0;
2336
2337         host = kzalloc(sizeof(*host), GFP_KERNEL);
2338         if (!host)
2339                 return 0;
2340
2341         host->acre = acre;
2342         host->lbafee = lbafee;
2343         ret = nvme_set_features(ctrl, NVME_FEAT_HOST_BEHAVIOR, 0,
2344                                 host, sizeof(*host), NULL);
2345         kfree(host);
2346         return ret;
2347 }
2348
2349 /*
2350  * The function checks whether the given total (exlat + enlat) latency of
2351  * a power state allows the latter to be used as an APST transition target.
2352  * It does so by comparing the latency to the primary and secondary latency
2353  * tolerances defined by module params. If there's a match, the corresponding
2354  * timeout value is returned and the matching tolerance index (1 or 2) is
2355  * reported.
2356  */
2357 static bool nvme_apst_get_transition_time(u64 total_latency,
2358                 u64 *transition_time, unsigned *last_index)
2359 {
2360         if (total_latency <= apst_primary_latency_tol_us) {
2361                 if (*last_index == 1)
2362                         return false;
2363                 *last_index = 1;
2364                 *transition_time = apst_primary_timeout_ms;
2365                 return true;
2366         }
2367         if (apst_secondary_timeout_ms &&
2368                 total_latency <= apst_secondary_latency_tol_us) {
2369                 if (*last_index <= 2)
2370                         return false;
2371                 *last_index = 2;
2372                 *transition_time = apst_secondary_timeout_ms;
2373                 return true;
2374         }
2375         return false;
2376 }
2377
2378 /*
2379  * APST (Autonomous Power State Transition) lets us program a table of power
2380  * state transitions that the controller will perform automatically.
2381  *
2382  * Depending on module params, one of the two supported techniques will be used:
2383  *
2384  * - If the parameters provide explicit timeouts and tolerances, they will be
2385  *   used to build a table with up to 2 non-operational states to transition to.
2386  *   The default parameter values were selected based on the values used by
2387  *   Microsoft's and Intel's NVMe drivers. Yet, since we don't implement dynamic
2388  *   regeneration of the APST table in the event of switching between external
2389  *   and battery power, the timeouts and tolerances reflect a compromise
2390  *   between values used by Microsoft for AC and battery scenarios.
2391  * - If not, we'll configure the table with a simple heuristic: we are willing
2392  *   to spend at most 2% of the time transitioning between power states.
2393  *   Therefore, when running in any given state, we will enter the next
2394  *   lower-power non-operational state after waiting 50 * (enlat + exlat)
2395  *   microseconds, as long as that state's exit latency is under the requested
2396  *   maximum latency.
2397  *
2398  * We will not autonomously enter any non-operational state for which the total
2399  * latency exceeds ps_max_latency_us.
2400  *
2401  * Users can set ps_max_latency_us to zero to turn off APST.
2402  */
2403 static int nvme_configure_apst(struct nvme_ctrl *ctrl)
2404 {
2405         struct nvme_feat_auto_pst *table;
2406         unsigned apste = 0;
2407         u64 max_lat_us = 0;
2408         __le64 target = 0;
2409         int max_ps = -1;
2410         int state;
2411         int ret;
2412         unsigned last_lt_index = UINT_MAX;
2413
2414         /*
2415          * If APST isn't supported or if we haven't been initialized yet,
2416          * then don't do anything.
2417          */
2418         if (!ctrl->apsta)
2419                 return 0;
2420
2421         if (ctrl->npss > 31) {
2422                 dev_warn(ctrl->device, "NPSS is invalid; not using APST\n");
2423                 return 0;
2424         }
2425
2426         table = kzalloc(sizeof(*table), GFP_KERNEL);
2427         if (!table)
2428                 return 0;
2429
2430         if (!ctrl->apst_enabled || ctrl->ps_max_latency_us == 0) {
2431                 /* Turn off APST. */
2432                 dev_dbg(ctrl->device, "APST disabled\n");
2433                 goto done;
2434         }
2435
2436         /*
2437          * Walk through all states from lowest- to highest-power.
2438          * According to the spec, lower-numbered states use more power.  NPSS,
2439          * despite the name, is the index of the lowest-power state, not the
2440          * number of states.
2441          */
2442         for (state = (int)ctrl->npss; state >= 0; state--) {
2443                 u64 total_latency_us, exit_latency_us, transition_ms;
2444
2445                 if (target)
2446                         table->entries[state] = target;
2447
2448                 /*
2449                  * Don't allow transitions to the deepest state if it's quirked
2450                  * off.
2451                  */
2452                 if (state == ctrl->npss &&
2453                     (ctrl->quirks & NVME_QUIRK_NO_DEEPEST_PS))
2454                         continue;
2455
2456                 /*
2457                  * Is this state a useful non-operational state for higher-power
2458                  * states to autonomously transition to?
2459                  */
2460                 if (!(ctrl->psd[state].flags & NVME_PS_FLAGS_NON_OP_STATE))
2461                         continue;
2462
2463                 exit_latency_us = (u64)le32_to_cpu(ctrl->psd[state].exit_lat);
2464                 if (exit_latency_us > ctrl->ps_max_latency_us)
2465                         continue;
2466
2467                 total_latency_us = exit_latency_us +
2468                         le32_to_cpu(ctrl->psd[state].entry_lat);
2469
2470                 /*
2471                  * This state is good. It can be used as the APST idle target
2472                  * for higher power states.
2473                  */
2474                 if (apst_primary_timeout_ms && apst_primary_latency_tol_us) {
2475                         if (!nvme_apst_get_transition_time(total_latency_us,
2476                                         &transition_ms, &last_lt_index))
2477                                 continue;
2478                 } else {
2479                         transition_ms = total_latency_us + 19;
2480                         do_div(transition_ms, 20);
2481                         if (transition_ms > (1 << 24) - 1)
2482                                 transition_ms = (1 << 24) - 1;
2483                 }
2484
2485                 target = cpu_to_le64((state << 3) | (transition_ms << 8));
2486                 if (max_ps == -1)
2487                         max_ps = state;
2488                 if (total_latency_us > max_lat_us)
2489                         max_lat_us = total_latency_us;
2490         }
2491
2492         if (max_ps == -1)
2493                 dev_dbg(ctrl->device, "APST enabled but no non-operational states are available\n");
2494         else
2495                 dev_dbg(ctrl->device, "APST enabled: max PS = %d, max round-trip latency = %lluus, table = %*phN\n",
2496                         max_ps, max_lat_us, (int)sizeof(*table), table);
2497         apste = 1;
2498
2499 done:
2500         ret = nvme_set_features(ctrl, NVME_FEAT_AUTO_PST, apste,
2501                                 table, sizeof(*table), NULL);
2502         if (ret)
2503                 dev_err(ctrl->device, "failed to set APST feature (%d)\n", ret);
2504         kfree(table);
2505         return ret;
2506 }
2507
2508 static void nvme_set_latency_tolerance(struct device *dev, s32 val)
2509 {
2510         struct nvme_ctrl *ctrl = dev_get_drvdata(dev);
2511         u64 latency;
2512
2513         switch (val) {
2514         case PM_QOS_LATENCY_TOLERANCE_NO_CONSTRAINT:
2515         case PM_QOS_LATENCY_ANY:
2516                 latency = U64_MAX;
2517                 break;
2518
2519         default:
2520                 latency = val;
2521         }
2522
2523         if (ctrl->ps_max_latency_us != latency) {
2524                 ctrl->ps_max_latency_us = latency;
2525                 if (ctrl->state == NVME_CTRL_LIVE)
2526                         nvme_configure_apst(ctrl);
2527         }
2528 }
2529
2530 struct nvme_core_quirk_entry {
2531         /*
2532          * NVMe model and firmware strings are padded with spaces.  For
2533          * simplicity, strings in the quirk table are padded with NULLs
2534          * instead.
2535          */
2536         u16 vid;
2537         const char *mn;
2538         const char *fr;
2539         unsigned long quirks;
2540 };
2541
2542 static const struct nvme_core_quirk_entry core_quirks[] = {
2543         {
2544                 /*
2545                  * This Toshiba device seems to die using any APST states.  See:
2546                  * https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1678184/comments/11
2547                  */
2548                 .vid = 0x1179,
2549                 .mn = "THNSF5256GPUK TOSHIBA",
2550                 .quirks = NVME_QUIRK_NO_APST,
2551         },
2552         {
2553                 /*
2554                  * This LiteON CL1-3D*-Q11 firmware version has a race
2555                  * condition associated with actions related to suspend to idle
2556                  * LiteON has resolved the problem in future firmware
2557                  */
2558                 .vid = 0x14a4,
2559                 .fr = "22301111",
2560                 .quirks = NVME_QUIRK_SIMPLE_SUSPEND,
2561         },
2562         {
2563                 /*
2564                  * This Kioxia CD6-V Series / HPE PE8030 device times out and
2565                  * aborts I/O during any load, but more easily reproducible
2566                  * with discards (fstrim).
2567                  *
2568                  * The device is left in a state where it is also not possible
2569                  * to use "nvme set-feature" to disable APST, but booting with
2570                  * nvme_core.default_ps_max_latency=0 works.
2571                  */
2572                 .vid = 0x1e0f,
2573                 .mn = "KCD6XVUL6T40",
2574                 .quirks = NVME_QUIRK_NO_APST,
2575         },
2576         {
2577                 /*
2578                  * The external Samsung X5 SSD fails initialization without a
2579                  * delay before checking if it is ready and has a whole set of
2580                  * other problems.  To make this even more interesting, it
2581                  * shares the PCI ID with internal Samsung 970 Evo Plus that
2582                  * does not need or want these quirks.
2583                  */
2584                 .vid = 0x144d,
2585                 .mn = "Samsung Portable SSD X5",
2586                 .quirks = NVME_QUIRK_DELAY_BEFORE_CHK_RDY |
2587                           NVME_QUIRK_NO_DEEPEST_PS |
2588                           NVME_QUIRK_IGNORE_DEV_SUBNQN,
2589         }
2590 };
2591
2592 /* match is null-terminated but idstr is space-padded. */
2593 static bool string_matches(const char *idstr, const char *match, size_t len)
2594 {
2595         size_t matchlen;
2596
2597         if (!match)
2598                 return true;
2599
2600         matchlen = strlen(match);
2601         WARN_ON_ONCE(matchlen > len);
2602
2603         if (memcmp(idstr, match, matchlen))
2604                 return false;
2605
2606         for (; matchlen < len; matchlen++)
2607                 if (idstr[matchlen] != ' ')
2608                         return false;
2609
2610         return true;
2611 }
2612
2613 static bool quirk_matches(const struct nvme_id_ctrl *id,
2614                           const struct nvme_core_quirk_entry *q)
2615 {
2616         return q->vid == le16_to_cpu(id->vid) &&
2617                 string_matches(id->mn, q->mn, sizeof(id->mn)) &&
2618                 string_matches(id->fr, q->fr, sizeof(id->fr));
2619 }
2620
2621 static void nvme_init_subnqn(struct nvme_subsystem *subsys, struct nvme_ctrl *ctrl,
2622                 struct nvme_id_ctrl *id)
2623 {
2624         size_t nqnlen;
2625         int off;
2626
2627         if(!(ctrl->quirks & NVME_QUIRK_IGNORE_DEV_SUBNQN)) {
2628                 nqnlen = strnlen(id->subnqn, NVMF_NQN_SIZE);
2629                 if (nqnlen > 0 && nqnlen < NVMF_NQN_SIZE) {
2630                         strscpy(subsys->subnqn, id->subnqn, NVMF_NQN_SIZE);
2631                         return;
2632                 }
2633
2634                 if (ctrl->vs >= NVME_VS(1, 2, 1))
2635                         dev_warn(ctrl->device, "missing or invalid SUBNQN field.\n");
2636         }
2637
2638         /*
2639          * Generate a "fake" NQN similar to the one in Section 4.5 of the NVMe
2640          * Base Specification 2.0.  It is slightly different from the format
2641          * specified there due to historic reasons, and we can't change it now.
2642          */
2643         off = snprintf(subsys->subnqn, NVMF_NQN_SIZE,
2644                         "nqn.2014.08.org.nvmexpress:%04x%04x",
2645                         le16_to_cpu(id->vid), le16_to_cpu(id->ssvid));
2646         memcpy(subsys->subnqn + off, id->sn, sizeof(id->sn));
2647         off += sizeof(id->sn);
2648         memcpy(subsys->subnqn + off, id->mn, sizeof(id->mn));
2649         off += sizeof(id->mn);
2650         memset(subsys->subnqn + off, 0, sizeof(subsys->subnqn) - off);
2651 }
2652
2653 static void nvme_release_subsystem(struct device *dev)
2654 {
2655         struct nvme_subsystem *subsys =
2656                 container_of(dev, struct nvme_subsystem, dev);
2657
2658         if (subsys->instance >= 0)
2659                 ida_free(&nvme_instance_ida, subsys->instance);
2660         kfree(subsys);
2661 }
2662
2663 static void nvme_destroy_subsystem(struct kref *ref)
2664 {
2665         struct nvme_subsystem *subsys =
2666                         container_of(ref, struct nvme_subsystem, ref);
2667
2668         mutex_lock(&nvme_subsystems_lock);
2669         list_del(&subsys->entry);
2670         mutex_unlock(&nvme_subsystems_lock);
2671
2672         ida_destroy(&subsys->ns_ida);
2673         device_del(&subsys->dev);
2674         put_device(&subsys->dev);
2675 }
2676
2677 static void nvme_put_subsystem(struct nvme_subsystem *subsys)
2678 {
2679         kref_put(&subsys->ref, nvme_destroy_subsystem);
2680 }
2681
2682 static struct nvme_subsystem *__nvme_find_get_subsystem(const char *subsysnqn)
2683 {
2684         struct nvme_subsystem *subsys;
2685
2686         lockdep_assert_held(&nvme_subsystems_lock);
2687
2688         /*
2689          * Fail matches for discovery subsystems. This results
2690          * in each discovery controller bound to a unique subsystem.
2691          * This avoids issues with validating controller values
2692          * that can only be true when there is a single unique subsystem.
2693          * There may be multiple and completely independent entities
2694          * that provide discovery controllers.
2695          */
2696         if (!strcmp(subsysnqn, NVME_DISC_SUBSYS_NAME))
2697                 return NULL;
2698
2699         list_for_each_entry(subsys, &nvme_subsystems, entry) {
2700                 if (strcmp(subsys->subnqn, subsysnqn))
2701                         continue;
2702                 if (!kref_get_unless_zero(&subsys->ref))
2703                         continue;
2704                 return subsys;
2705         }
2706
2707         return NULL;
2708 }
2709
2710 static inline bool nvme_discovery_ctrl(struct nvme_ctrl *ctrl)
2711 {
2712         return ctrl->opts && ctrl->opts->discovery_nqn;
2713 }
2714
2715 static bool nvme_validate_cntlid(struct nvme_subsystem *subsys,
2716                 struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id)
2717 {
2718         struct nvme_ctrl *tmp;
2719
2720         lockdep_assert_held(&nvme_subsystems_lock);
2721
2722         list_for_each_entry(tmp, &subsys->ctrls, subsys_entry) {
2723                 if (nvme_state_terminal(tmp))
2724                         continue;
2725
2726                 if (tmp->cntlid == ctrl->cntlid) {
2727                         dev_err(ctrl->device,
2728                                 "Duplicate cntlid %u with %s, subsys %s, rejecting\n",
2729                                 ctrl->cntlid, dev_name(tmp->device),
2730                                 subsys->subnqn);
2731                         return false;
2732                 }
2733
2734                 if ((id->cmic & NVME_CTRL_CMIC_MULTI_CTRL) ||
2735                     nvme_discovery_ctrl(ctrl))
2736                         continue;
2737
2738                 dev_err(ctrl->device,
2739                         "Subsystem does not support multiple controllers\n");
2740                 return false;
2741         }
2742
2743         return true;
2744 }
2745
2746 static int nvme_init_subsystem(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id)
2747 {
2748         struct nvme_subsystem *subsys, *found;
2749         int ret;
2750
2751         subsys = kzalloc(sizeof(*subsys), GFP_KERNEL);
2752         if (!subsys)
2753                 return -ENOMEM;
2754
2755         subsys->instance = -1;
2756         mutex_init(&subsys->lock);
2757         kref_init(&subsys->ref);
2758         INIT_LIST_HEAD(&subsys->ctrls);
2759         INIT_LIST_HEAD(&subsys->nsheads);
2760         nvme_init_subnqn(subsys, ctrl, id);
2761         memcpy(subsys->serial, id->sn, sizeof(subsys->serial));
2762         memcpy(subsys->model, id->mn, sizeof(subsys->model));
2763         subsys->vendor_id = le16_to_cpu(id->vid);
2764         subsys->cmic = id->cmic;
2765
2766         /* Versions prior to 1.4 don't necessarily report a valid type */
2767         if (id->cntrltype == NVME_CTRL_DISC ||
2768             !strcmp(subsys->subnqn, NVME_DISC_SUBSYS_NAME))
2769                 subsys->subtype = NVME_NQN_DISC;
2770         else
2771                 subsys->subtype = NVME_NQN_NVME;
2772
2773         if (nvme_discovery_ctrl(ctrl) && subsys->subtype != NVME_NQN_DISC) {
2774                 dev_err(ctrl->device,
2775                         "Subsystem %s is not a discovery controller",
2776                         subsys->subnqn);
2777                 kfree(subsys);
2778                 return -EINVAL;
2779         }
2780         subsys->awupf = le16_to_cpu(id->awupf);
2781         nvme_mpath_default_iopolicy(subsys);
2782
2783         subsys->dev.class = nvme_subsys_class;
2784         subsys->dev.release = nvme_release_subsystem;
2785         subsys->dev.groups = nvme_subsys_attrs_groups;
2786         dev_set_name(&subsys->dev, "nvme-subsys%d", ctrl->instance);
2787         device_initialize(&subsys->dev);
2788
2789         mutex_lock(&nvme_subsystems_lock);
2790         found = __nvme_find_get_subsystem(subsys->subnqn);
2791         if (found) {
2792                 put_device(&subsys->dev);
2793                 subsys = found;
2794
2795                 if (!nvme_validate_cntlid(subsys, ctrl, id)) {
2796                         ret = -EINVAL;
2797                         goto out_put_subsystem;
2798                 }
2799         } else {
2800                 ret = device_add(&subsys->dev);
2801                 if (ret) {
2802                         dev_err(ctrl->device,
2803                                 "failed to register subsystem device.\n");
2804                         put_device(&subsys->dev);
2805                         goto out_unlock;
2806                 }
2807                 ida_init(&subsys->ns_ida);
2808                 list_add_tail(&subsys->entry, &nvme_subsystems);
2809         }
2810
2811         ret = sysfs_create_link(&subsys->dev.kobj, &ctrl->device->kobj,
2812                                 dev_name(ctrl->device));
2813         if (ret) {
2814                 dev_err(ctrl->device,
2815                         "failed to create sysfs link from subsystem.\n");
2816                 goto out_put_subsystem;
2817         }
2818
2819         if (!found)
2820                 subsys->instance = ctrl->instance;
2821         ctrl->subsys = subsys;
2822         list_add_tail(&ctrl->subsys_entry, &subsys->ctrls);
2823         mutex_unlock(&nvme_subsystems_lock);
2824         return 0;
2825
2826 out_put_subsystem:
2827         nvme_put_subsystem(subsys);
2828 out_unlock:
2829         mutex_unlock(&nvme_subsystems_lock);
2830         return ret;
2831 }
2832
2833 int nvme_get_log(struct nvme_ctrl *ctrl, u32 nsid, u8 log_page, u8 lsp, u8 csi,
2834                 void *log, size_t size, u64 offset)
2835 {
2836         struct nvme_command c = { };
2837         u32 dwlen = nvme_bytes_to_numd(size);
2838
2839         c.get_log_page.opcode = nvme_admin_get_log_page;
2840         c.get_log_page.nsid = cpu_to_le32(nsid);
2841         c.get_log_page.lid = log_page;
2842         c.get_log_page.lsp = lsp;
2843         c.get_log_page.numdl = cpu_to_le16(dwlen & ((1 << 16) - 1));
2844         c.get_log_page.numdu = cpu_to_le16(dwlen >> 16);
2845         c.get_log_page.lpol = cpu_to_le32(lower_32_bits(offset));
2846         c.get_log_page.lpou = cpu_to_le32(upper_32_bits(offset));
2847         c.get_log_page.csi = csi;
2848
2849         return nvme_submit_sync_cmd(ctrl->admin_q, &c, log, size);
2850 }
2851
2852 static int nvme_get_effects_log(struct nvme_ctrl *ctrl, u8 csi,
2853                                 struct nvme_effects_log **log)
2854 {
2855         struct nvme_effects_log *cel = xa_load(&ctrl->cels, csi);
2856         int ret;
2857
2858         if (cel)
2859                 goto out;
2860
2861         cel = kzalloc(sizeof(*cel), GFP_KERNEL);
2862         if (!cel)
2863                 return -ENOMEM;
2864
2865         ret = nvme_get_log(ctrl, 0x00, NVME_LOG_CMD_EFFECTS, 0, csi,
2866                         cel, sizeof(*cel), 0);
2867         if (ret) {
2868                 kfree(cel);
2869                 return ret;
2870         }
2871
2872         xa_store(&ctrl->cels, csi, cel, GFP_KERNEL);
2873 out:
2874         *log = cel;
2875         return 0;
2876 }
2877
2878 static inline u32 nvme_mps_to_sectors(struct nvme_ctrl *ctrl, u32 units)
2879 {
2880         u32 page_shift = NVME_CAP_MPSMIN(ctrl->cap) + 12, val;
2881
2882         if (check_shl_overflow(1U, units + page_shift - 9, &val))
2883                 return UINT_MAX;
2884         return val;
2885 }
2886
2887 static int nvme_init_non_mdts_limits(struct nvme_ctrl *ctrl)
2888 {
2889         struct nvme_command c = { };
2890         struct nvme_id_ctrl_nvm *id;
2891         int ret;
2892
2893         if (ctrl->oncs & NVME_CTRL_ONCS_DSM) {
2894                 ctrl->max_discard_sectors = UINT_MAX;
2895                 ctrl->max_discard_segments = NVME_DSM_MAX_RANGES;
2896         } else {
2897                 ctrl->max_discard_sectors = 0;
2898                 ctrl->max_discard_segments = 0;
2899         }
2900
2901         /*
2902          * Even though NVMe spec explicitly states that MDTS is not applicable
2903          * to the write-zeroes, we are cautious and limit the size to the
2904          * controllers max_hw_sectors value, which is based on the MDTS field
2905          * and possibly other limiting factors.
2906          */
2907         if ((ctrl->oncs & NVME_CTRL_ONCS_WRITE_ZEROES) &&
2908             !(ctrl->quirks & NVME_QUIRK_DISABLE_WRITE_ZEROES))
2909                 ctrl->max_zeroes_sectors = ctrl->max_hw_sectors;
2910         else
2911                 ctrl->max_zeroes_sectors = 0;
2912
2913         if (ctrl->subsys->subtype != NVME_NQN_NVME ||
2914             nvme_ctrl_limited_cns(ctrl) ||
2915             test_bit(NVME_CTRL_SKIP_ID_CNS_CS, &ctrl->flags))
2916                 return 0;
2917
2918         id = kzalloc(sizeof(*id), GFP_KERNEL);
2919         if (!id)
2920                 return -ENOMEM;
2921
2922         c.identify.opcode = nvme_admin_identify;
2923         c.identify.cns = NVME_ID_CNS_CS_CTRL;
2924         c.identify.csi = NVME_CSI_NVM;
2925
2926         ret = nvme_submit_sync_cmd(ctrl->admin_q, &c, id, sizeof(*id));
2927         if (ret)
2928                 goto free_data;
2929
2930         if (id->dmrl)
2931                 ctrl->max_discard_segments = id->dmrl;
2932         ctrl->dmrsl = le32_to_cpu(id->dmrsl);
2933         if (id->wzsl)
2934                 ctrl->max_zeroes_sectors = nvme_mps_to_sectors(ctrl, id->wzsl);
2935
2936 free_data:
2937         if (ret > 0)
2938                 set_bit(NVME_CTRL_SKIP_ID_CNS_CS, &ctrl->flags);
2939         kfree(id);
2940         return ret;
2941 }
2942
2943 static void nvme_init_known_nvm_effects(struct nvme_ctrl *ctrl)
2944 {
2945         struct nvme_effects_log *log = ctrl->effects;
2946
2947         log->acs[nvme_admin_format_nvm] |= cpu_to_le32(NVME_CMD_EFFECTS_LBCC |
2948                                                 NVME_CMD_EFFECTS_NCC |
2949                                                 NVME_CMD_EFFECTS_CSE_MASK);
2950         log->acs[nvme_admin_sanitize_nvm] |= cpu_to_le32(NVME_CMD_EFFECTS_LBCC |
2951                                                 NVME_CMD_EFFECTS_CSE_MASK);
2952
2953         /*
2954          * The spec says the result of a security receive command depends on
2955          * the previous security send command. As such, many vendors log this
2956          * command as one to submitted only when no other commands to the same
2957          * namespace are outstanding. The intention is to tell the host to
2958          * prevent mixing security send and receive.
2959          *
2960          * This driver can only enforce such exclusive access against IO
2961          * queues, though. We are not readily able to enforce such a rule for
2962          * two commands to the admin queue, which is the only queue that
2963          * matters for this command.
2964          *
2965          * Rather than blindly freezing the IO queues for this effect that
2966          * doesn't even apply to IO, mask it off.
2967          */
2968         log->acs[nvme_admin_security_recv] &= cpu_to_le32(~NVME_CMD_EFFECTS_CSE_MASK);
2969
2970         log->iocs[nvme_cmd_write] |= cpu_to_le32(NVME_CMD_EFFECTS_LBCC);
2971         log->iocs[nvme_cmd_write_zeroes] |= cpu_to_le32(NVME_CMD_EFFECTS_LBCC);
2972         log->iocs[nvme_cmd_write_uncor] |= cpu_to_le32(NVME_CMD_EFFECTS_LBCC);
2973 }
2974
2975 static int nvme_init_effects(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id)
2976 {
2977         int ret = 0;
2978
2979         if (ctrl->effects)
2980                 return 0;
2981
2982         if (id->lpa & NVME_CTRL_LPA_CMD_EFFECTS_LOG) {
2983                 ret = nvme_get_effects_log(ctrl, NVME_CSI_NVM, &ctrl->effects);
2984                 if (ret < 0)
2985                         return ret;
2986         }
2987
2988         if (!ctrl->effects) {
2989                 ctrl->effects = kzalloc(sizeof(*ctrl->effects), GFP_KERNEL);
2990                 if (!ctrl->effects)
2991                         return -ENOMEM;
2992                 xa_store(&ctrl->cels, NVME_CSI_NVM, ctrl->effects, GFP_KERNEL);
2993         }
2994
2995         nvme_init_known_nvm_effects(ctrl);
2996         return 0;
2997 }
2998
2999 static int nvme_init_identify(struct nvme_ctrl *ctrl)
3000 {
3001         struct nvme_id_ctrl *id;
3002         u32 max_hw_sectors;
3003         bool prev_apst_enabled;
3004         int ret;
3005
3006         ret = nvme_identify_ctrl(ctrl, &id);
3007         if (ret) {
3008                 dev_err(ctrl->device, "Identify Controller failed (%d)\n", ret);
3009                 return -EIO;
3010         }
3011
3012         if (!(ctrl->ops->flags & NVME_F_FABRICS))
3013                 ctrl->cntlid = le16_to_cpu(id->cntlid);
3014
3015         if (!ctrl->identified) {
3016                 unsigned int i;
3017
3018                 /*
3019                  * Check for quirks.  Quirk can depend on firmware version,
3020                  * so, in principle, the set of quirks present can change
3021                  * across a reset.  As a possible future enhancement, we
3022                  * could re-scan for quirks every time we reinitialize
3023                  * the device, but we'd have to make sure that the driver
3024                  * behaves intelligently if the quirks change.
3025                  */
3026                 for (i = 0; i < ARRAY_SIZE(core_quirks); i++) {
3027                         if (quirk_matches(id, &core_quirks[i]))
3028                                 ctrl->quirks |= core_quirks[i].quirks;
3029                 }
3030
3031                 ret = nvme_init_subsystem(ctrl, id);
3032                 if (ret)
3033                         goto out_free;
3034
3035                 ret = nvme_init_effects(ctrl, id);
3036                 if (ret)
3037                         goto out_free;
3038         }
3039         memcpy(ctrl->subsys->firmware_rev, id->fr,
3040                sizeof(ctrl->subsys->firmware_rev));
3041
3042         if (force_apst && (ctrl->quirks & NVME_QUIRK_NO_DEEPEST_PS)) {
3043                 dev_warn(ctrl->device, "forcibly allowing all power states due to nvme_core.force_apst -- use at your own risk\n");
3044                 ctrl->quirks &= ~NVME_QUIRK_NO_DEEPEST_PS;
3045         }
3046
3047         ctrl->crdt[0] = le16_to_cpu(id->crdt1);
3048         ctrl->crdt[1] = le16_to_cpu(id->crdt2);
3049         ctrl->crdt[2] = le16_to_cpu(id->crdt3);
3050
3051         ctrl->oacs = le16_to_cpu(id->oacs);
3052         ctrl->oncs = le16_to_cpu(id->oncs);
3053         ctrl->mtfa = le16_to_cpu(id->mtfa);
3054         ctrl->oaes = le32_to_cpu(id->oaes);
3055         ctrl->wctemp = le16_to_cpu(id->wctemp);
3056         ctrl->cctemp = le16_to_cpu(id->cctemp);
3057
3058         atomic_set(&ctrl->abort_limit, id->acl + 1);
3059         ctrl->vwc = id->vwc;
3060         if (id->mdts)
3061                 max_hw_sectors = nvme_mps_to_sectors(ctrl, id->mdts);
3062         else
3063                 max_hw_sectors = UINT_MAX;
3064         ctrl->max_hw_sectors =
3065                 min_not_zero(ctrl->max_hw_sectors, max_hw_sectors);
3066
3067         nvme_set_queue_limits(ctrl, ctrl->admin_q);
3068         ctrl->sgls = le32_to_cpu(id->sgls);
3069         ctrl->kas = le16_to_cpu(id->kas);
3070         ctrl->max_namespaces = le32_to_cpu(id->mnan);
3071         ctrl->ctratt = le32_to_cpu(id->ctratt);
3072
3073         ctrl->cntrltype = id->cntrltype;
3074         ctrl->dctype = id->dctype;
3075
3076         if (id->rtd3e) {
3077                 /* us -> s */
3078                 u32 transition_time = le32_to_cpu(id->rtd3e) / USEC_PER_SEC;
3079
3080                 ctrl->shutdown_timeout = clamp_t(unsigned int, transition_time,
3081                                                  shutdown_timeout, 60);
3082
3083                 if (ctrl->shutdown_timeout != shutdown_timeout)
3084                         dev_info(ctrl->device,
3085                                  "Shutdown timeout set to %u seconds\n",
3086                                  ctrl->shutdown_timeout);
3087         } else
3088                 ctrl->shutdown_timeout = shutdown_timeout;
3089
3090         ctrl->npss = id->npss;
3091         ctrl->apsta = id->apsta;
3092         prev_apst_enabled = ctrl->apst_enabled;
3093         if (ctrl->quirks & NVME_QUIRK_NO_APST) {
3094                 if (force_apst && id->apsta) {
3095                         dev_warn(ctrl->device, "forcibly allowing APST due to nvme_core.force_apst -- use at your own risk\n");
3096                         ctrl->apst_enabled = true;
3097                 } else {
3098                         ctrl->apst_enabled = false;
3099                 }
3100         } else {
3101                 ctrl->apst_enabled = id->apsta;
3102         }
3103         memcpy(ctrl->psd, id->psd, sizeof(ctrl->psd));
3104
3105         if (ctrl->ops->flags & NVME_F_FABRICS) {
3106                 ctrl->icdoff = le16_to_cpu(id->icdoff);
3107                 ctrl->ioccsz = le32_to_cpu(id->ioccsz);
3108                 ctrl->iorcsz = le32_to_cpu(id->iorcsz);
3109                 ctrl->maxcmd = le16_to_cpu(id->maxcmd);
3110
3111                 /*
3112                  * In fabrics we need to verify the cntlid matches the
3113                  * admin connect
3114                  */
3115                 if (ctrl->cntlid != le16_to_cpu(id->cntlid)) {
3116                         dev_err(ctrl->device,
3117                                 "Mismatching cntlid: Connect %u vs Identify "
3118                                 "%u, rejecting\n",
3119                                 ctrl->cntlid, le16_to_cpu(id->cntlid));
3120                         ret = -EINVAL;
3121                         goto out_free;
3122                 }
3123
3124                 if (!nvme_discovery_ctrl(ctrl) && !ctrl->kas) {
3125                         dev_err(ctrl->device,
3126                                 "keep-alive support is mandatory for fabrics\n");
3127                         ret = -EINVAL;
3128                         goto out_free;
3129                 }
3130         } else {
3131                 ctrl->hmpre = le32_to_cpu(id->hmpre);
3132                 ctrl->hmmin = le32_to_cpu(id->hmmin);
3133                 ctrl->hmminds = le32_to_cpu(id->hmminds);
3134                 ctrl->hmmaxd = le16_to_cpu(id->hmmaxd);
3135         }
3136
3137         ret = nvme_mpath_init_identify(ctrl, id);
3138         if (ret < 0)
3139                 goto out_free;
3140
3141         if (ctrl->apst_enabled && !prev_apst_enabled)
3142                 dev_pm_qos_expose_latency_tolerance(ctrl->device);
3143         else if (!ctrl->apst_enabled && prev_apst_enabled)
3144                 dev_pm_qos_hide_latency_tolerance(ctrl->device);
3145
3146 out_free:
3147         kfree(id);
3148         return ret;
3149 }
3150
3151 /*
3152  * Initialize the cached copies of the Identify data and various controller
3153  * register in our nvme_ctrl structure.  This should be called as soon as
3154  * the admin queue is fully up and running.
3155  */
3156 int nvme_init_ctrl_finish(struct nvme_ctrl *ctrl, bool was_suspended)
3157 {
3158         int ret;
3159
3160         ret = ctrl->ops->reg_read32(ctrl, NVME_REG_VS, &ctrl->vs);
3161         if (ret) {
3162                 dev_err(ctrl->device, "Reading VS failed (%d)\n", ret);
3163                 return ret;
3164         }
3165
3166         ctrl->sqsize = min_t(u16, NVME_CAP_MQES(ctrl->cap), ctrl->sqsize);
3167
3168         if (ctrl->vs >= NVME_VS(1, 1, 0))
3169                 ctrl->subsystem = NVME_CAP_NSSRC(ctrl->cap);
3170
3171         ret = nvme_init_identify(ctrl);
3172         if (ret)
3173                 return ret;
3174
3175         ret = nvme_configure_apst(ctrl);
3176         if (ret < 0)
3177                 return ret;
3178
3179         ret = nvme_configure_timestamp(ctrl);
3180         if (ret < 0)
3181                 return ret;
3182
3183         ret = nvme_configure_host_options(ctrl);
3184         if (ret < 0)
3185                 return ret;
3186
3187         nvme_configure_opal(ctrl, was_suspended);
3188
3189         if (!ctrl->identified && !nvme_discovery_ctrl(ctrl)) {
3190                 /*
3191                  * Do not return errors unless we are in a controller reset,
3192                  * the controller works perfectly fine without hwmon.
3193                  */
3194                 ret = nvme_hwmon_init(ctrl);
3195                 if (ret == -EINTR)
3196                         return ret;
3197         }
3198
3199         clear_bit(NVME_CTRL_DIRTY_CAPABILITY, &ctrl->flags);
3200         ctrl->identified = true;
3201
3202         return 0;
3203 }
3204 EXPORT_SYMBOL_GPL(nvme_init_ctrl_finish);
3205
3206 static int nvme_dev_open(struct inode *inode, struct file *file)
3207 {
3208         struct nvme_ctrl *ctrl =
3209                 container_of(inode->i_cdev, struct nvme_ctrl, cdev);
3210
3211         switch (ctrl->state) {
3212         case NVME_CTRL_LIVE:
3213                 break;
3214         default:
3215                 return -EWOULDBLOCK;
3216         }
3217
3218         nvme_get_ctrl(ctrl);
3219         if (!try_module_get(ctrl->ops->module)) {
3220                 nvme_put_ctrl(ctrl);
3221                 return -EINVAL;
3222         }
3223
3224         file->private_data = ctrl;
3225         return 0;
3226 }
3227
3228 static int nvme_dev_release(struct inode *inode, struct file *file)
3229 {
3230         struct nvme_ctrl *ctrl =
3231                 container_of(inode->i_cdev, struct nvme_ctrl, cdev);
3232
3233         module_put(ctrl->ops->module);
3234         nvme_put_ctrl(ctrl);
3235         return 0;
3236 }
3237
3238 static const struct file_operations nvme_dev_fops = {
3239         .owner          = THIS_MODULE,
3240         .open           = nvme_dev_open,
3241         .release        = nvme_dev_release,
3242         .unlocked_ioctl = nvme_dev_ioctl,
3243         .compat_ioctl   = compat_ptr_ioctl,
3244         .uring_cmd      = nvme_dev_uring_cmd,
3245 };
3246
3247 static struct nvme_ns_head *nvme_find_ns_head(struct nvme_ctrl *ctrl,
3248                 unsigned nsid)
3249 {
3250         struct nvme_ns_head *h;
3251
3252         lockdep_assert_held(&ctrl->subsys->lock);
3253
3254         list_for_each_entry(h, &ctrl->subsys->nsheads, entry) {
3255                 /*
3256                  * Private namespaces can share NSIDs under some conditions.
3257                  * In that case we can't use the same ns_head for namespaces
3258                  * with the same NSID.
3259                  */
3260                 if (h->ns_id != nsid || !nvme_is_unique_nsid(ctrl, h))
3261                         continue;
3262                 if (!list_empty(&h->list) && nvme_tryget_ns_head(h))
3263                         return h;
3264         }
3265
3266         return NULL;
3267 }
3268
3269 static int nvme_subsys_check_duplicate_ids(struct nvme_subsystem *subsys,
3270                 struct nvme_ns_ids *ids)
3271 {
3272         bool has_uuid = !uuid_is_null(&ids->uuid);
3273         bool has_nguid = memchr_inv(ids->nguid, 0, sizeof(ids->nguid));
3274         bool has_eui64 = memchr_inv(ids->eui64, 0, sizeof(ids->eui64));
3275         struct nvme_ns_head *h;
3276
3277         lockdep_assert_held(&subsys->lock);
3278
3279         list_for_each_entry(h, &subsys->nsheads, entry) {
3280                 if (has_uuid && uuid_equal(&ids->uuid, &h->ids.uuid))
3281                         return -EINVAL;
3282                 if (has_nguid &&
3283                     memcmp(&ids->nguid, &h->ids.nguid, sizeof(ids->nguid)) == 0)
3284                         return -EINVAL;
3285                 if (has_eui64 &&
3286                     memcmp(&ids->eui64, &h->ids.eui64, sizeof(ids->eui64)) == 0)
3287                         return -EINVAL;
3288         }
3289
3290         return 0;
3291 }
3292
3293 static void nvme_cdev_rel(struct device *dev)
3294 {
3295         ida_free(&nvme_ns_chr_minor_ida, MINOR(dev->devt));
3296 }
3297
3298 void nvme_cdev_del(struct cdev *cdev, struct device *cdev_device)
3299 {
3300         cdev_device_del(cdev, cdev_device);
3301         put_device(cdev_device);
3302 }
3303
3304 int nvme_cdev_add(struct cdev *cdev, struct device *cdev_device,
3305                 const struct file_operations *fops, struct module *owner)
3306 {
3307         int minor, ret;
3308
3309         minor = ida_alloc(&nvme_ns_chr_minor_ida, GFP_KERNEL);
3310         if (minor < 0)
3311                 return minor;
3312         cdev_device->devt = MKDEV(MAJOR(nvme_ns_chr_devt), minor);
3313         cdev_device->class = nvme_ns_chr_class;
3314         cdev_device->release = nvme_cdev_rel;
3315         device_initialize(cdev_device);
3316         cdev_init(cdev, fops);
3317         cdev->owner = owner;
3318         ret = cdev_device_add(cdev, cdev_device);
3319         if (ret)
3320                 put_device(cdev_device);
3321
3322         return ret;
3323 }
3324
3325 static int nvme_ns_chr_open(struct inode *inode, struct file *file)
3326 {
3327         return nvme_ns_open(container_of(inode->i_cdev, struct nvme_ns, cdev));
3328 }
3329
3330 static int nvme_ns_chr_release(struct inode *inode, struct file *file)
3331 {
3332         nvme_ns_release(container_of(inode->i_cdev, struct nvme_ns, cdev));
3333         return 0;
3334 }
3335
3336 static const struct file_operations nvme_ns_chr_fops = {
3337         .owner          = THIS_MODULE,
3338         .open           = nvme_ns_chr_open,
3339         .release        = nvme_ns_chr_release,
3340         .unlocked_ioctl = nvme_ns_chr_ioctl,
3341         .compat_ioctl   = compat_ptr_ioctl,
3342         .uring_cmd      = nvme_ns_chr_uring_cmd,
3343         .uring_cmd_iopoll = nvme_ns_chr_uring_cmd_iopoll,
3344 };
3345
3346 static int nvme_add_ns_cdev(struct nvme_ns *ns)
3347 {
3348         int ret;
3349
3350         ns->cdev_device.parent = ns->ctrl->device;
3351         ret = dev_set_name(&ns->cdev_device, "ng%dn%d",
3352                            ns->ctrl->instance, ns->head->instance);
3353         if (ret)
3354                 return ret;
3355
3356         return nvme_cdev_add(&ns->cdev, &ns->cdev_device, &nvme_ns_chr_fops,
3357                              ns->ctrl->ops->module);
3358 }
3359
3360 static struct nvme_ns_head *nvme_alloc_ns_head(struct nvme_ctrl *ctrl,
3361                 struct nvme_ns_info *info)
3362 {
3363         struct nvme_ns_head *head;
3364         size_t size = sizeof(*head);
3365         int ret = -ENOMEM;
3366
3367 #ifdef CONFIG_NVME_MULTIPATH
3368         size += num_possible_nodes() * sizeof(struct nvme_ns *);
3369 #endif
3370
3371         head = kzalloc(size, GFP_KERNEL);
3372         if (!head)
3373                 goto out;
3374         ret = ida_alloc_min(&ctrl->subsys->ns_ida, 1, GFP_KERNEL);
3375         if (ret < 0)
3376                 goto out_free_head;
3377         head->instance = ret;
3378         INIT_LIST_HEAD(&head->list);
3379         ret = init_srcu_struct(&head->srcu);
3380         if (ret)
3381                 goto out_ida_remove;
3382         head->subsys = ctrl->subsys;
3383         head->ns_id = info->nsid;
3384         head->ids = info->ids;
3385         head->shared = info->is_shared;
3386         kref_init(&head->ref);
3387
3388         if (head->ids.csi) {
3389                 ret = nvme_get_effects_log(ctrl, head->ids.csi, &head->effects);
3390                 if (ret)
3391                         goto out_cleanup_srcu;
3392         } else
3393                 head->effects = ctrl->effects;
3394
3395         ret = nvme_mpath_alloc_disk(ctrl, head);
3396         if (ret)
3397                 goto out_cleanup_srcu;
3398
3399         list_add_tail(&head->entry, &ctrl->subsys->nsheads);
3400
3401         kref_get(&ctrl->subsys->ref);
3402
3403         return head;
3404 out_cleanup_srcu:
3405         cleanup_srcu_struct(&head->srcu);
3406 out_ida_remove:
3407         ida_free(&ctrl->subsys->ns_ida, head->instance);
3408 out_free_head:
3409         kfree(head);
3410 out:
3411         if (ret > 0)
3412                 ret = blk_status_to_errno(nvme_error_status(ret));
3413         return ERR_PTR(ret);
3414 }
3415
3416 static int nvme_global_check_duplicate_ids(struct nvme_subsystem *this,
3417                 struct nvme_ns_ids *ids)
3418 {
3419         struct nvme_subsystem *s;
3420         int ret = 0;
3421
3422         /*
3423          * Note that this check is racy as we try to avoid holding the global
3424          * lock over the whole ns_head creation.  But it is only intended as
3425          * a sanity check anyway.
3426          */
3427         mutex_lock(&nvme_subsystems_lock);
3428         list_for_each_entry(s, &nvme_subsystems, entry) {
3429                 if (s == this)
3430                         continue;
3431                 mutex_lock(&s->lock);
3432                 ret = nvme_subsys_check_duplicate_ids(s, ids);
3433                 mutex_unlock(&s->lock);
3434                 if (ret)
3435                         break;
3436         }
3437         mutex_unlock(&nvme_subsystems_lock);
3438
3439         return ret;
3440 }
3441
3442 static int nvme_init_ns_head(struct nvme_ns *ns, struct nvme_ns_info *info)
3443 {
3444         struct nvme_ctrl *ctrl = ns->ctrl;
3445         struct nvme_ns_head *head = NULL;
3446         int ret;
3447
3448         ret = nvme_global_check_duplicate_ids(ctrl->subsys, &info->ids);
3449         if (ret) {
3450                 /*
3451                  * We've found two different namespaces on two different
3452                  * subsystems that report the same ID.  This is pretty nasty
3453                  * for anything that actually requires unique device
3454                  * identification.  In the kernel we need this for multipathing,
3455                  * and in user space the /dev/disk/by-id/ links rely on it.
3456                  *
3457                  * If the device also claims to be multi-path capable back off
3458                  * here now and refuse the probe the second device as this is a
3459                  * recipe for data corruption.  If not this is probably a
3460                  * cheap consumer device if on the PCIe bus, so let the user
3461                  * proceed and use the shiny toy, but warn that with changing
3462                  * probing order (which due to our async probing could just be
3463                  * device taking longer to startup) the other device could show
3464                  * up at any time.
3465                  */
3466                 nvme_print_device_info(ctrl);
3467                 if ((ns->ctrl->ops->flags & NVME_F_FABRICS) || /* !PCIe */
3468                     ((ns->ctrl->subsys->cmic & NVME_CTRL_CMIC_MULTI_CTRL) &&
3469                      info->is_shared)) {
3470                         dev_err(ctrl->device,
3471                                 "ignoring nsid %d because of duplicate IDs\n",
3472                                 info->nsid);
3473                         return ret;
3474                 }
3475
3476                 dev_err(ctrl->device,
3477                         "clearing duplicate IDs for nsid %d\n", info->nsid);
3478                 dev_err(ctrl->device,
3479                         "use of /dev/disk/by-id/ may cause data corruption\n");
3480                 memset(&info->ids.nguid, 0, sizeof(info->ids.nguid));
3481                 memset(&info->ids.uuid, 0, sizeof(info->ids.uuid));
3482                 memset(&info->ids.eui64, 0, sizeof(info->ids.eui64));
3483                 ctrl->quirks |= NVME_QUIRK_BOGUS_NID;
3484         }
3485
3486         mutex_lock(&ctrl->subsys->lock);
3487         head = nvme_find_ns_head(ctrl, info->nsid);
3488         if (!head) {
3489                 ret = nvme_subsys_check_duplicate_ids(ctrl->subsys, &info->ids);
3490                 if (ret) {
3491                         dev_err(ctrl->device,
3492                                 "duplicate IDs in subsystem for nsid %d\n",
3493                                 info->nsid);
3494                         goto out_unlock;
3495                 }
3496                 head = nvme_alloc_ns_head(ctrl, info);
3497                 if (IS_ERR(head)) {
3498                         ret = PTR_ERR(head);
3499                         goto out_unlock;
3500                 }
3501         } else {
3502                 ret = -EINVAL;
3503                 if (!info->is_shared || !head->shared) {
3504                         dev_err(ctrl->device,
3505                                 "Duplicate unshared namespace %d\n",
3506                                 info->nsid);
3507                         goto out_put_ns_head;
3508                 }
3509                 if (!nvme_ns_ids_equal(&head->ids, &info->ids)) {
3510                         dev_err(ctrl->device,
3511                                 "IDs don't match for shared namespace %d\n",
3512                                         info->nsid);
3513                         goto out_put_ns_head;
3514                 }
3515
3516                 if (!multipath) {
3517                         dev_warn(ctrl->device,
3518                                 "Found shared namespace %d, but multipathing not supported.\n",
3519                                 info->nsid);
3520                         dev_warn_once(ctrl->device,
3521                                 "Support for shared namespaces without CONFIG_NVME_MULTIPATH is deprecated and will be removed in Linux 6.0\n.");
3522                 }
3523         }
3524
3525         list_add_tail_rcu(&ns->siblings, &head->list);
3526         ns->head = head;
3527         mutex_unlock(&ctrl->subsys->lock);
3528         return 0;
3529
3530 out_put_ns_head:
3531         nvme_put_ns_head(head);
3532 out_unlock:
3533         mutex_unlock(&ctrl->subsys->lock);
3534         return ret;
3535 }
3536
3537 struct nvme_ns *nvme_find_get_ns(struct nvme_ctrl *ctrl, unsigned nsid)
3538 {
3539         struct nvme_ns *ns, *ret = NULL;
3540
3541         down_read(&ctrl->namespaces_rwsem);
3542         list_for_each_entry(ns, &ctrl->namespaces, list) {
3543                 if (ns->head->ns_id == nsid) {
3544                         if (!nvme_get_ns(ns))
3545                                 continue;
3546                         ret = ns;
3547                         break;
3548                 }
3549                 if (ns->head->ns_id > nsid)
3550                         break;
3551         }
3552         up_read(&ctrl->namespaces_rwsem);
3553         return ret;
3554 }
3555 EXPORT_SYMBOL_NS_GPL(nvme_find_get_ns, NVME_TARGET_PASSTHRU);
3556
3557 /*
3558  * Add the namespace to the controller list while keeping the list ordered.
3559  */
3560 static void nvme_ns_add_to_ctrl_list(struct nvme_ns *ns)
3561 {
3562         struct nvme_ns *tmp;
3563
3564         list_for_each_entry_reverse(tmp, &ns->ctrl->namespaces, list) {
3565                 if (tmp->head->ns_id < ns->head->ns_id) {
3566                         list_add(&ns->list, &tmp->list);
3567                         return;
3568                 }
3569         }
3570         list_add(&ns->list, &ns->ctrl->namespaces);
3571 }
3572
3573 static void nvme_alloc_ns(struct nvme_ctrl *ctrl, struct nvme_ns_info *info)
3574 {
3575         struct nvme_ns *ns;
3576         struct gendisk *disk;
3577         int node = ctrl->numa_node;
3578
3579         ns = kzalloc_node(sizeof(*ns), GFP_KERNEL, node);
3580         if (!ns)
3581                 return;
3582
3583         disk = blk_mq_alloc_disk(ctrl->tagset, ns);
3584         if (IS_ERR(disk))
3585                 goto out_free_ns;
3586         disk->fops = &nvme_bdev_ops;
3587         disk->private_data = ns;
3588
3589         ns->disk = disk;
3590         ns->queue = disk->queue;
3591
3592         if (ctrl->opts && ctrl->opts->data_digest)
3593                 blk_queue_flag_set(QUEUE_FLAG_STABLE_WRITES, ns->queue);
3594
3595         blk_queue_flag_set(QUEUE_FLAG_NONROT, ns->queue);
3596         if (ctrl->ops->supports_pci_p2pdma &&
3597             ctrl->ops->supports_pci_p2pdma(ctrl))
3598                 blk_queue_flag_set(QUEUE_FLAG_PCI_P2PDMA, ns->queue);
3599
3600         ns->ctrl = ctrl;
3601         kref_init(&ns->kref);
3602
3603         if (nvme_init_ns_head(ns, info))
3604                 goto out_cleanup_disk;
3605
3606         /*
3607          * If multipathing is enabled, the device name for all disks and not
3608          * just those that represent shared namespaces needs to be based on the
3609          * subsystem instance.  Using the controller instance for private
3610          * namespaces could lead to naming collisions between shared and private
3611          * namespaces if they don't use a common numbering scheme.
3612          *
3613          * If multipathing is not enabled, disk names must use the controller
3614          * instance as shared namespaces will show up as multiple block
3615          * devices.
3616          */
3617         if (nvme_ns_head_multipath(ns->head)) {
3618                 sprintf(disk->disk_name, "nvme%dc%dn%d", ctrl->subsys->instance,
3619                         ctrl->instance, ns->head->instance);
3620                 disk->flags |= GENHD_FL_HIDDEN;
3621         } else if (multipath) {
3622                 sprintf(disk->disk_name, "nvme%dn%d", ctrl->subsys->instance,
3623                         ns->head->instance);
3624         } else {
3625                 sprintf(disk->disk_name, "nvme%dn%d", ctrl->instance,
3626                         ns->head->instance);
3627         }
3628
3629         if (nvme_update_ns_info(ns, info))
3630                 goto out_unlink_ns;
3631
3632         down_write(&ctrl->namespaces_rwsem);
3633         nvme_ns_add_to_ctrl_list(ns);
3634         up_write(&ctrl->namespaces_rwsem);
3635         nvme_get_ctrl(ctrl);
3636
3637         if (device_add_disk(ctrl->device, ns->disk, nvme_ns_id_attr_groups))
3638                 goto out_cleanup_ns_from_list;
3639
3640         if (!nvme_ns_head_multipath(ns->head))
3641                 nvme_add_ns_cdev(ns);
3642
3643         nvme_mpath_add_disk(ns, info->anagrpid);
3644         nvme_fault_inject_init(&ns->fault_inject, ns->disk->disk_name);
3645
3646         return;
3647
3648  out_cleanup_ns_from_list:
3649         nvme_put_ctrl(ctrl);
3650         down_write(&ctrl->namespaces_rwsem);
3651         list_del_init(&ns->list);
3652         up_write(&ctrl->namespaces_rwsem);
3653  out_unlink_ns:
3654         mutex_lock(&ctrl->subsys->lock);
3655         list_del_rcu(&ns->siblings);
3656         if (list_empty(&ns->head->list))
3657                 list_del_init(&ns->head->entry);
3658         mutex_unlock(&ctrl->subsys->lock);
3659         nvme_put_ns_head(ns->head);
3660  out_cleanup_disk:
3661         put_disk(disk);
3662  out_free_ns:
3663         kfree(ns);
3664 }
3665
3666 static void nvme_ns_remove(struct nvme_ns *ns)
3667 {
3668         bool last_path = false;
3669
3670         if (test_and_set_bit(NVME_NS_REMOVING, &ns->flags))
3671                 return;
3672
3673         clear_bit(NVME_NS_READY, &ns->flags);
3674         set_capacity(ns->disk, 0);
3675         nvme_fault_inject_fini(&ns->fault_inject);
3676
3677         /*
3678          * Ensure that !NVME_NS_READY is seen by other threads to prevent
3679          * this ns going back into current_path.
3680          */
3681         synchronize_srcu(&ns->head->srcu);
3682
3683         /* wait for concurrent submissions */
3684         if (nvme_mpath_clear_current_path(ns))
3685                 synchronize_srcu(&ns->head->srcu);
3686
3687         mutex_lock(&ns->ctrl->subsys->lock);
3688         list_del_rcu(&ns->siblings);
3689         if (list_empty(&ns->head->list)) {
3690                 list_del_init(&ns->head->entry);
3691                 last_path = true;
3692         }
3693         mutex_unlock(&ns->ctrl->subsys->lock);
3694
3695         /* guarantee not available in head->list */
3696         synchronize_srcu(&ns->head->srcu);
3697
3698         if (!nvme_ns_head_multipath(ns->head))
3699                 nvme_cdev_del(&ns->cdev, &ns->cdev_device);
3700         del_gendisk(ns->disk);
3701
3702         down_write(&ns->ctrl->namespaces_rwsem);
3703         list_del_init(&ns->list);
3704         up_write(&ns->ctrl->namespaces_rwsem);
3705
3706         if (last_path)
3707                 nvme_mpath_shutdown_disk(ns->head);
3708         nvme_put_ns(ns);
3709 }
3710
3711 static void nvme_ns_remove_by_nsid(struct nvme_ctrl *ctrl, u32 nsid)
3712 {
3713         struct nvme_ns *ns = nvme_find_get_ns(ctrl, nsid);
3714
3715         if (ns) {
3716                 nvme_ns_remove(ns);
3717                 nvme_put_ns(ns);
3718         }
3719 }
3720
3721 static void nvme_validate_ns(struct nvme_ns *ns, struct nvme_ns_info *info)
3722 {
3723         int ret = NVME_SC_INVALID_NS | NVME_SC_DNR;
3724
3725         if (!nvme_ns_ids_equal(&ns->head->ids, &info->ids)) {
3726                 dev_err(ns->ctrl->device,
3727                         "identifiers changed for nsid %d\n", ns->head->ns_id);
3728                 goto out;
3729         }
3730
3731         ret = nvme_update_ns_info(ns, info);
3732 out:
3733         /*
3734          * Only remove the namespace if we got a fatal error back from the
3735          * device, otherwise ignore the error and just move on.
3736          *
3737          * TODO: we should probably schedule a delayed retry here.
3738          */
3739         if (ret > 0 && (ret & NVME_SC_DNR))
3740                 nvme_ns_remove(ns);
3741 }
3742
3743 static void nvme_scan_ns(struct nvme_ctrl *ctrl, unsigned nsid)
3744 {
3745         struct nvme_ns_info info = { .nsid = nsid };
3746         struct nvme_ns *ns;
3747         int ret;
3748
3749         if (nvme_identify_ns_descs(ctrl, &info))
3750                 return;
3751
3752         if (info.ids.csi != NVME_CSI_NVM && !nvme_multi_css(ctrl)) {
3753                 dev_warn(ctrl->device,
3754                         "command set not reported for nsid: %d\n", nsid);
3755                 return;
3756         }
3757
3758         /*
3759          * If available try to use the Command Set Idependent Identify Namespace
3760          * data structure to find all the generic information that is needed to
3761          * set up a namespace.  If not fall back to the legacy version.
3762          */
3763         if ((ctrl->cap & NVME_CAP_CRMS_CRIMS) ||
3764             (info.ids.csi != NVME_CSI_NVM && info.ids.csi != NVME_CSI_ZNS))
3765                 ret = nvme_ns_info_from_id_cs_indep(ctrl, &info);
3766         else
3767                 ret = nvme_ns_info_from_identify(ctrl, &info);
3768
3769         if (info.is_removed)
3770                 nvme_ns_remove_by_nsid(ctrl, nsid);
3771
3772         /*
3773          * Ignore the namespace if it is not ready. We will get an AEN once it
3774          * becomes ready and restart the scan.
3775          */
3776         if (ret || !info.is_ready)
3777                 return;
3778
3779         ns = nvme_find_get_ns(ctrl, nsid);
3780         if (ns) {
3781                 nvme_validate_ns(ns, &info);
3782                 nvme_put_ns(ns);
3783         } else {
3784                 nvme_alloc_ns(ctrl, &info);
3785         }
3786 }
3787
3788 static void nvme_remove_invalid_namespaces(struct nvme_ctrl *ctrl,
3789                                         unsigned nsid)
3790 {
3791         struct nvme_ns *ns, *next;
3792         LIST_HEAD(rm_list);
3793
3794         down_write(&ctrl->namespaces_rwsem);
3795         list_for_each_entry_safe(ns, next, &ctrl->namespaces, list) {
3796                 if (ns->head->ns_id > nsid)
3797                         list_move_tail(&ns->list, &rm_list);
3798         }
3799         up_write(&ctrl->namespaces_rwsem);
3800
3801         list_for_each_entry_safe(ns, next, &rm_list, list)
3802                 nvme_ns_remove(ns);
3803
3804 }
3805
3806 static int nvme_scan_ns_list(struct nvme_ctrl *ctrl)
3807 {
3808         const int nr_entries = NVME_IDENTIFY_DATA_SIZE / sizeof(__le32);
3809         __le32 *ns_list;
3810         u32 prev = 0;
3811         int ret = 0, i;
3812
3813         ns_list = kzalloc(NVME_IDENTIFY_DATA_SIZE, GFP_KERNEL);
3814         if (!ns_list)
3815                 return -ENOMEM;
3816
3817         for (;;) {
3818                 struct nvme_command cmd = {
3819                         .identify.opcode        = nvme_admin_identify,
3820                         .identify.cns           = NVME_ID_CNS_NS_ACTIVE_LIST,
3821                         .identify.nsid          = cpu_to_le32(prev),
3822                 };
3823
3824                 ret = nvme_submit_sync_cmd(ctrl->admin_q, &cmd, ns_list,
3825                                             NVME_IDENTIFY_DATA_SIZE);
3826                 if (ret) {
3827                         dev_warn(ctrl->device,
3828                                 "Identify NS List failed (status=0x%x)\n", ret);
3829                         goto free;
3830                 }
3831
3832                 for (i = 0; i < nr_entries; i++) {
3833                         u32 nsid = le32_to_cpu(ns_list[i]);
3834
3835                         if (!nsid)      /* end of the list? */
3836                                 goto out;
3837                         nvme_scan_ns(ctrl, nsid);
3838                         while (++prev < nsid)
3839                                 nvme_ns_remove_by_nsid(ctrl, prev);
3840                 }
3841         }
3842  out:
3843         nvme_remove_invalid_namespaces(ctrl, prev);
3844  free:
3845         kfree(ns_list);
3846         return ret;
3847 }
3848
3849 static void nvme_scan_ns_sequential(struct nvme_ctrl *ctrl)
3850 {
3851         struct nvme_id_ctrl *id;
3852         u32 nn, i;
3853
3854         if (nvme_identify_ctrl(ctrl, &id))
3855                 return;
3856         nn = le32_to_cpu(id->nn);
3857         kfree(id);
3858
3859         for (i = 1; i <= nn; i++)
3860                 nvme_scan_ns(ctrl, i);
3861
3862         nvme_remove_invalid_namespaces(ctrl, nn);
3863 }
3864
3865 static void nvme_clear_changed_ns_log(struct nvme_ctrl *ctrl)
3866 {
3867         size_t log_size = NVME_MAX_CHANGED_NAMESPACES * sizeof(__le32);
3868         __le32 *log;
3869         int error;
3870
3871         log = kzalloc(log_size, GFP_KERNEL);
3872         if (!log)
3873                 return;
3874
3875         /*
3876          * We need to read the log to clear the AEN, but we don't want to rely
3877          * on it for the changed namespace information as userspace could have
3878          * raced with us in reading the log page, which could cause us to miss
3879          * updates.
3880          */
3881         error = nvme_get_log(ctrl, NVME_NSID_ALL, NVME_LOG_CHANGED_NS, 0,
3882                         NVME_CSI_NVM, log, log_size, 0);
3883         if (error)
3884                 dev_warn(ctrl->device,
3885                         "reading changed ns log failed: %d\n", error);
3886
3887         kfree(log);
3888 }
3889
3890 static void nvme_scan_work(struct work_struct *work)
3891 {
3892         struct nvme_ctrl *ctrl =
3893                 container_of(work, struct nvme_ctrl, scan_work);
3894         int ret;
3895
3896         /* No tagset on a live ctrl means IO queues could not created */
3897         if (ctrl->state != NVME_CTRL_LIVE || !ctrl->tagset)
3898                 return;
3899
3900         /*
3901          * Identify controller limits can change at controller reset due to
3902          * new firmware download, even though it is not common we cannot ignore
3903          * such scenario. Controller's non-mdts limits are reported in the unit
3904          * of logical blocks that is dependent on the format of attached
3905          * namespace. Hence re-read the limits at the time of ns allocation.
3906          */
3907         ret = nvme_init_non_mdts_limits(ctrl);
3908         if (ret < 0) {
3909                 dev_warn(ctrl->device,
3910                         "reading non-mdts-limits failed: %d\n", ret);
3911                 return;
3912         }
3913
3914         if (test_and_clear_bit(NVME_AER_NOTICE_NS_CHANGED, &ctrl->events)) {
3915                 dev_info(ctrl->device, "rescanning namespaces.\n");
3916                 nvme_clear_changed_ns_log(ctrl);
3917         }
3918
3919         mutex_lock(&ctrl->scan_lock);
3920         if (nvme_ctrl_limited_cns(ctrl)) {
3921                 nvme_scan_ns_sequential(ctrl);
3922         } else {
3923                 /*
3924                  * Fall back to sequential scan if DNR is set to handle broken
3925                  * devices which should support Identify NS List (as per the VS
3926                  * they report) but don't actually support it.
3927                  */
3928                 ret = nvme_scan_ns_list(ctrl);
3929                 if (ret > 0 && ret & NVME_SC_DNR)
3930                         nvme_scan_ns_sequential(ctrl);
3931         }
3932         mutex_unlock(&ctrl->scan_lock);
3933 }
3934
3935 /*
3936  * This function iterates the namespace list unlocked to allow recovery from
3937  * controller failure. It is up to the caller to ensure the namespace list is
3938  * not modified by scan work while this function is executing.
3939  */
3940 void nvme_remove_namespaces(struct nvme_ctrl *ctrl)
3941 {
3942         struct nvme_ns *ns, *next;
3943         LIST_HEAD(ns_list);
3944
3945         /*
3946          * make sure to requeue I/O to all namespaces as these
3947          * might result from the scan itself and must complete
3948          * for the scan_work to make progress
3949          */
3950         nvme_mpath_clear_ctrl_paths(ctrl);
3951
3952         /*
3953          * Unquiesce io queues so any pending IO won't hang, especially
3954          * those submitted from scan work
3955          */
3956         nvme_unquiesce_io_queues(ctrl);
3957
3958         /* prevent racing with ns scanning */
3959         flush_work(&ctrl->scan_work);
3960
3961         /*
3962          * The dead states indicates the controller was not gracefully
3963          * disconnected. In that case, we won't be able to flush any data while
3964          * removing the namespaces' disks; fail all the queues now to avoid
3965          * potentially having to clean up the failed sync later.
3966          */
3967         if (ctrl->state == NVME_CTRL_DEAD)
3968                 nvme_mark_namespaces_dead(ctrl);
3969
3970         /* this is a no-op when called from the controller reset handler */
3971         nvme_change_ctrl_state(ctrl, NVME_CTRL_DELETING_NOIO);
3972
3973         down_write(&ctrl->namespaces_rwsem);
3974         list_splice_init(&ctrl->namespaces, &ns_list);
3975         up_write(&ctrl->namespaces_rwsem);
3976
3977         list_for_each_entry_safe(ns, next, &ns_list, list)
3978                 nvme_ns_remove(ns);
3979 }
3980 EXPORT_SYMBOL_GPL(nvme_remove_namespaces);
3981
3982 static int nvme_class_uevent(const struct device *dev, struct kobj_uevent_env *env)
3983 {
3984         const struct nvme_ctrl *ctrl =
3985                 container_of(dev, struct nvme_ctrl, ctrl_device);
3986         struct nvmf_ctrl_options *opts = ctrl->opts;
3987         int ret;
3988
3989         ret = add_uevent_var(env, "NVME_TRTYPE=%s", ctrl->ops->name);
3990         if (ret)
3991                 return ret;
3992
3993         if (opts) {
3994                 ret = add_uevent_var(env, "NVME_TRADDR=%s", opts->traddr);
3995                 if (ret)
3996                         return ret;
3997
3998                 ret = add_uevent_var(env, "NVME_TRSVCID=%s",
3999                                 opts->trsvcid ?: "none");
4000                 if (ret)
4001                         return ret;
4002
4003                 ret = add_uevent_var(env, "NVME_HOST_TRADDR=%s",
4004                                 opts->host_traddr ?: "none");
4005                 if (ret)
4006                         return ret;
4007
4008                 ret = add_uevent_var(env, "NVME_HOST_IFACE=%s",
4009                                 opts->host_iface ?: "none");
4010         }
4011         return ret;
4012 }
4013
4014 static void nvme_change_uevent(struct nvme_ctrl *ctrl, char *envdata)
4015 {
4016         char *envp[2] = { envdata, NULL };
4017
4018         kobject_uevent_env(&ctrl->device->kobj, KOBJ_CHANGE, envp);
4019 }
4020
4021 static void nvme_aen_uevent(struct nvme_ctrl *ctrl)
4022 {
4023         char *envp[2] = { NULL, NULL };
4024         u32 aen_result = ctrl->aen_result;
4025
4026         ctrl->aen_result = 0;
4027         if (!aen_result)
4028                 return;
4029
4030         envp[0] = kasprintf(GFP_KERNEL, "NVME_AEN=%#08x", aen_result);
4031         if (!envp[0])
4032                 return;
4033         kobject_uevent_env(&ctrl->device->kobj, KOBJ_CHANGE, envp);
4034         kfree(envp[0]);
4035 }
4036
4037 static void nvme_async_event_work(struct work_struct *work)
4038 {
4039         struct nvme_ctrl *ctrl =
4040                 container_of(work, struct nvme_ctrl, async_event_work);
4041
4042         nvme_aen_uevent(ctrl);
4043
4044         /*
4045          * The transport drivers must guarantee AER submission here is safe by
4046          * flushing ctrl async_event_work after changing the controller state
4047          * from LIVE and before freeing the admin queue.
4048         */
4049         if (ctrl->state == NVME_CTRL_LIVE)
4050                 ctrl->ops->submit_async_event(ctrl);
4051 }
4052
4053 static bool nvme_ctrl_pp_status(struct nvme_ctrl *ctrl)
4054 {
4055
4056         u32 csts;
4057
4058         if (ctrl->ops->reg_read32(ctrl, NVME_REG_CSTS, &csts))
4059                 return false;
4060
4061         if (csts == ~0)
4062                 return false;
4063
4064         return ((ctrl->ctrl_config & NVME_CC_ENABLE) && (csts & NVME_CSTS_PP));
4065 }
4066
4067 static void nvme_get_fw_slot_info(struct nvme_ctrl *ctrl)
4068 {
4069         struct nvme_fw_slot_info_log *log;
4070
4071         log = kmalloc(sizeof(*log), GFP_KERNEL);
4072         if (!log)
4073                 return;
4074
4075         if (nvme_get_log(ctrl, NVME_NSID_ALL, NVME_LOG_FW_SLOT, 0, NVME_CSI_NVM,
4076                         log, sizeof(*log), 0))
4077                 dev_warn(ctrl->device, "Get FW SLOT INFO log error\n");
4078         kfree(log);
4079 }
4080
4081 static void nvme_fw_act_work(struct work_struct *work)
4082 {
4083         struct nvme_ctrl *ctrl = container_of(work,
4084                                 struct nvme_ctrl, fw_act_work);
4085         unsigned long fw_act_timeout;
4086
4087         if (ctrl->mtfa)
4088                 fw_act_timeout = jiffies +
4089                                 msecs_to_jiffies(ctrl->mtfa * 100);
4090         else
4091                 fw_act_timeout = jiffies +
4092                                 msecs_to_jiffies(admin_timeout * 1000);
4093
4094         nvme_quiesce_io_queues(ctrl);
4095         while (nvme_ctrl_pp_status(ctrl)) {
4096                 if (time_after(jiffies, fw_act_timeout)) {
4097                         dev_warn(ctrl->device,
4098                                 "Fw activation timeout, reset controller\n");
4099                         nvme_try_sched_reset(ctrl);
4100                         return;
4101                 }
4102                 msleep(100);
4103         }
4104
4105         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_LIVE))
4106                 return;
4107
4108         nvme_unquiesce_io_queues(ctrl);
4109         /* read FW slot information to clear the AER */
4110         nvme_get_fw_slot_info(ctrl);
4111
4112         queue_work(nvme_wq, &ctrl->async_event_work);
4113 }
4114
4115 static u32 nvme_aer_type(u32 result)
4116 {
4117         return result & 0x7;
4118 }
4119
4120 static u32 nvme_aer_subtype(u32 result)
4121 {
4122         return (result & 0xff00) >> 8;
4123 }
4124
4125 static bool nvme_handle_aen_notice(struct nvme_ctrl *ctrl, u32 result)
4126 {
4127         u32 aer_notice_type = nvme_aer_subtype(result);
4128         bool requeue = true;
4129
4130         switch (aer_notice_type) {
4131         case NVME_AER_NOTICE_NS_CHANGED:
4132                 set_bit(NVME_AER_NOTICE_NS_CHANGED, &ctrl->events);
4133                 nvme_queue_scan(ctrl);
4134                 break;
4135         case NVME_AER_NOTICE_FW_ACT_STARTING:
4136                 /*
4137                  * We are (ab)using the RESETTING state to prevent subsequent
4138                  * recovery actions from interfering with the controller's
4139                  * firmware activation.
4140                  */
4141                 if (nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING)) {
4142                         nvme_auth_stop(ctrl);
4143                         requeue = false;
4144                         queue_work(nvme_wq, &ctrl->fw_act_work);
4145                 }
4146                 break;
4147 #ifdef CONFIG_NVME_MULTIPATH
4148         case NVME_AER_NOTICE_ANA:
4149                 if (!ctrl->ana_log_buf)
4150                         break;
4151                 queue_work(nvme_wq, &ctrl->ana_work);
4152                 break;
4153 #endif
4154         case NVME_AER_NOTICE_DISC_CHANGED:
4155                 ctrl->aen_result = result;
4156                 break;
4157         default:
4158                 dev_warn(ctrl->device, "async event result %08x\n", result);
4159         }
4160         return requeue;
4161 }
4162
4163 static void nvme_handle_aer_persistent_error(struct nvme_ctrl *ctrl)
4164 {
4165         dev_warn(ctrl->device, "resetting controller due to AER\n");
4166         nvme_reset_ctrl(ctrl);
4167 }
4168
4169 void nvme_complete_async_event(struct nvme_ctrl *ctrl, __le16 status,
4170                 volatile union nvme_result *res)
4171 {
4172         u32 result = le32_to_cpu(res->u32);
4173         u32 aer_type = nvme_aer_type(result);
4174         u32 aer_subtype = nvme_aer_subtype(result);
4175         bool requeue = true;
4176
4177         if (le16_to_cpu(status) >> 1 != NVME_SC_SUCCESS)
4178                 return;
4179
4180         trace_nvme_async_event(ctrl, result);
4181         switch (aer_type) {
4182         case NVME_AER_NOTICE:
4183                 requeue = nvme_handle_aen_notice(ctrl, result);
4184                 break;
4185         case NVME_AER_ERROR:
4186                 /*
4187                  * For a persistent internal error, don't run async_event_work
4188                  * to submit a new AER. The controller reset will do it.
4189                  */
4190                 if (aer_subtype == NVME_AER_ERROR_PERSIST_INT_ERR) {
4191                         nvme_handle_aer_persistent_error(ctrl);
4192                         return;
4193                 }
4194                 fallthrough;
4195         case NVME_AER_SMART:
4196         case NVME_AER_CSS:
4197         case NVME_AER_VS:
4198                 ctrl->aen_result = result;
4199                 break;
4200         default:
4201                 break;
4202         }
4203
4204         if (requeue)
4205                 queue_work(nvme_wq, &ctrl->async_event_work);
4206 }
4207 EXPORT_SYMBOL_GPL(nvme_complete_async_event);
4208
4209 int nvme_alloc_admin_tag_set(struct nvme_ctrl *ctrl, struct blk_mq_tag_set *set,
4210                 const struct blk_mq_ops *ops, unsigned int cmd_size)
4211 {
4212         int ret;
4213
4214         memset(set, 0, sizeof(*set));
4215         set->ops = ops;
4216         set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
4217         if (ctrl->ops->flags & NVME_F_FABRICS)
4218                 set->reserved_tags = NVMF_RESERVED_TAGS;
4219         set->numa_node = ctrl->numa_node;
4220         set->flags = BLK_MQ_F_NO_SCHED;
4221         if (ctrl->ops->flags & NVME_F_BLOCKING)
4222                 set->flags |= BLK_MQ_F_BLOCKING;
4223         set->cmd_size = cmd_size;
4224         set->driver_data = ctrl;
4225         set->nr_hw_queues = 1;
4226         set->timeout = NVME_ADMIN_TIMEOUT;
4227         ret = blk_mq_alloc_tag_set(set);
4228         if (ret)
4229                 return ret;
4230
4231         ctrl->admin_q = blk_mq_init_queue(set);
4232         if (IS_ERR(ctrl->admin_q)) {
4233                 ret = PTR_ERR(ctrl->admin_q);
4234                 goto out_free_tagset;
4235         }
4236
4237         if (ctrl->ops->flags & NVME_F_FABRICS) {
4238                 ctrl->fabrics_q = blk_mq_init_queue(set);
4239                 if (IS_ERR(ctrl->fabrics_q)) {
4240                         ret = PTR_ERR(ctrl->fabrics_q);
4241                         goto out_cleanup_admin_q;
4242                 }
4243         }
4244
4245         ctrl->admin_tagset = set;
4246         return 0;
4247
4248 out_cleanup_admin_q:
4249         blk_mq_destroy_queue(ctrl->admin_q);
4250         blk_put_queue(ctrl->admin_q);
4251 out_free_tagset:
4252         blk_mq_free_tag_set(set);
4253         ctrl->admin_q = NULL;
4254         ctrl->fabrics_q = NULL;
4255         return ret;
4256 }
4257 EXPORT_SYMBOL_GPL(nvme_alloc_admin_tag_set);
4258
4259 void nvme_remove_admin_tag_set(struct nvme_ctrl *ctrl)
4260 {
4261         blk_mq_destroy_queue(ctrl->admin_q);
4262         blk_put_queue(ctrl->admin_q);
4263         if (ctrl->ops->flags & NVME_F_FABRICS) {
4264                 blk_mq_destroy_queue(ctrl->fabrics_q);
4265                 blk_put_queue(ctrl->fabrics_q);
4266         }
4267         blk_mq_free_tag_set(ctrl->admin_tagset);
4268 }
4269 EXPORT_SYMBOL_GPL(nvme_remove_admin_tag_set);
4270
4271 int nvme_alloc_io_tag_set(struct nvme_ctrl *ctrl, struct blk_mq_tag_set *set,
4272                 const struct blk_mq_ops *ops, unsigned int nr_maps,
4273                 unsigned int cmd_size)
4274 {
4275         int ret;
4276
4277         memset(set, 0, sizeof(*set));
4278         set->ops = ops;
4279         set->queue_depth = min_t(unsigned, ctrl->sqsize, BLK_MQ_MAX_DEPTH - 1);
4280         /*
4281          * Some Apple controllers requires tags to be unique across admin and
4282          * the (only) I/O queue, so reserve the first 32 tags of the I/O queue.
4283          */
4284         if (ctrl->quirks & NVME_QUIRK_SHARED_TAGS)
4285                 set->reserved_tags = NVME_AQ_DEPTH;
4286         else if (ctrl->ops->flags & NVME_F_FABRICS)
4287                 set->reserved_tags = NVMF_RESERVED_TAGS;
4288         set->numa_node = ctrl->numa_node;
4289         set->flags = BLK_MQ_F_SHOULD_MERGE;
4290         if (ctrl->ops->flags & NVME_F_BLOCKING)
4291                 set->flags |= BLK_MQ_F_BLOCKING;
4292         set->cmd_size = cmd_size,
4293         set->driver_data = ctrl;
4294         set->nr_hw_queues = ctrl->queue_count - 1;
4295         set->timeout = NVME_IO_TIMEOUT;
4296         set->nr_maps = nr_maps;
4297         ret = blk_mq_alloc_tag_set(set);
4298         if (ret)
4299                 return ret;
4300
4301         if (ctrl->ops->flags & NVME_F_FABRICS) {
4302                 ctrl->connect_q = blk_mq_init_queue(set);
4303                 if (IS_ERR(ctrl->connect_q)) {
4304                         ret = PTR_ERR(ctrl->connect_q);
4305                         goto out_free_tag_set;
4306                 }
4307                 blk_queue_flag_set(QUEUE_FLAG_SKIP_TAGSET_QUIESCE,
4308                                    ctrl->connect_q);
4309         }
4310
4311         ctrl->tagset = set;
4312         return 0;
4313
4314 out_free_tag_set:
4315         blk_mq_free_tag_set(set);
4316         ctrl->connect_q = NULL;
4317         return ret;
4318 }
4319 EXPORT_SYMBOL_GPL(nvme_alloc_io_tag_set);
4320
4321 void nvme_remove_io_tag_set(struct nvme_ctrl *ctrl)
4322 {
4323         if (ctrl->ops->flags & NVME_F_FABRICS) {
4324                 blk_mq_destroy_queue(ctrl->connect_q);
4325                 blk_put_queue(ctrl->connect_q);
4326         }
4327         blk_mq_free_tag_set(ctrl->tagset);
4328 }
4329 EXPORT_SYMBOL_GPL(nvme_remove_io_tag_set);
4330
4331 void nvme_stop_ctrl(struct nvme_ctrl *ctrl)
4332 {
4333         nvme_mpath_stop(ctrl);
4334         nvme_auth_stop(ctrl);
4335         nvme_stop_keep_alive(ctrl);
4336         nvme_stop_failfast_work(ctrl);
4337         flush_work(&ctrl->async_event_work);
4338         cancel_work_sync(&ctrl->fw_act_work);
4339         if (ctrl->ops->stop_ctrl)
4340                 ctrl->ops->stop_ctrl(ctrl);
4341 }
4342 EXPORT_SYMBOL_GPL(nvme_stop_ctrl);
4343
4344 void nvme_start_ctrl(struct nvme_ctrl *ctrl)
4345 {
4346         nvme_start_keep_alive(ctrl);
4347
4348         nvme_enable_aen(ctrl);
4349
4350         /*
4351          * persistent discovery controllers need to send indication to userspace
4352          * to re-read the discovery log page to learn about possible changes
4353          * that were missed. We identify persistent discovery controllers by
4354          * checking that they started once before, hence are reconnecting back.
4355          */
4356         if (test_bit(NVME_CTRL_STARTED_ONCE, &ctrl->flags) &&
4357             nvme_discovery_ctrl(ctrl))
4358                 nvme_change_uevent(ctrl, "NVME_EVENT=rediscover");
4359
4360         if (ctrl->queue_count > 1) {
4361                 nvme_queue_scan(ctrl);
4362                 nvme_unquiesce_io_queues(ctrl);
4363                 nvme_mpath_update(ctrl);
4364         }
4365
4366         nvme_change_uevent(ctrl, "NVME_EVENT=connected");
4367         set_bit(NVME_CTRL_STARTED_ONCE, &ctrl->flags);
4368 }
4369 EXPORT_SYMBOL_GPL(nvme_start_ctrl);
4370
4371 void nvme_uninit_ctrl(struct nvme_ctrl *ctrl)
4372 {
4373         nvme_hwmon_exit(ctrl);
4374         nvme_fault_inject_fini(&ctrl->fault_inject);
4375         dev_pm_qos_hide_latency_tolerance(ctrl->device);
4376         cdev_device_del(&ctrl->cdev, ctrl->device);
4377         nvme_put_ctrl(ctrl);
4378 }
4379 EXPORT_SYMBOL_GPL(nvme_uninit_ctrl);
4380
4381 static void nvme_free_cels(struct nvme_ctrl *ctrl)
4382 {
4383         struct nvme_effects_log *cel;
4384         unsigned long i;
4385
4386         xa_for_each(&ctrl->cels, i, cel) {
4387                 xa_erase(&ctrl->cels, i);
4388                 kfree(cel);
4389         }
4390
4391         xa_destroy(&ctrl->cels);
4392 }
4393
4394 static void nvme_free_ctrl(struct device *dev)
4395 {
4396         struct nvme_ctrl *ctrl =
4397                 container_of(dev, struct nvme_ctrl, ctrl_device);
4398         struct nvme_subsystem *subsys = ctrl->subsys;
4399
4400         if (!subsys || ctrl->instance != subsys->instance)
4401                 ida_free(&nvme_instance_ida, ctrl->instance);
4402
4403         nvme_free_cels(ctrl);
4404         nvme_mpath_uninit(ctrl);
4405         nvme_auth_stop(ctrl);
4406         nvme_auth_free(ctrl);
4407         __free_page(ctrl->discard_page);
4408         free_opal_dev(ctrl->opal_dev);
4409
4410         if (subsys) {
4411                 mutex_lock(&nvme_subsystems_lock);
4412                 list_del(&ctrl->subsys_entry);
4413                 sysfs_remove_link(&subsys->dev.kobj, dev_name(ctrl->device));
4414                 mutex_unlock(&nvme_subsystems_lock);
4415         }
4416
4417         ctrl->ops->free_ctrl(ctrl);
4418
4419         if (subsys)
4420                 nvme_put_subsystem(subsys);
4421 }
4422
4423 /*
4424  * Initialize a NVMe controller structures.  This needs to be called during
4425  * earliest initialization so that we have the initialized structured around
4426  * during probing.
4427  */
4428 int nvme_init_ctrl(struct nvme_ctrl *ctrl, struct device *dev,
4429                 const struct nvme_ctrl_ops *ops, unsigned long quirks)
4430 {
4431         int ret;
4432
4433         ctrl->state = NVME_CTRL_NEW;
4434         clear_bit(NVME_CTRL_FAILFAST_EXPIRED, &ctrl->flags);
4435         spin_lock_init(&ctrl->lock);
4436         mutex_init(&ctrl->scan_lock);
4437         INIT_LIST_HEAD(&ctrl->namespaces);
4438         xa_init(&ctrl->cels);
4439         init_rwsem(&ctrl->namespaces_rwsem);
4440         ctrl->dev = dev;
4441         ctrl->ops = ops;
4442         ctrl->quirks = quirks;
4443         ctrl->numa_node = NUMA_NO_NODE;
4444         INIT_WORK(&ctrl->scan_work, nvme_scan_work);
4445         INIT_WORK(&ctrl->async_event_work, nvme_async_event_work);
4446         INIT_WORK(&ctrl->fw_act_work, nvme_fw_act_work);
4447         INIT_WORK(&ctrl->delete_work, nvme_delete_ctrl_work);
4448         init_waitqueue_head(&ctrl->state_wq);
4449
4450         INIT_DELAYED_WORK(&ctrl->ka_work, nvme_keep_alive_work);
4451         INIT_DELAYED_WORK(&ctrl->failfast_work, nvme_failfast_work);
4452         memset(&ctrl->ka_cmd, 0, sizeof(ctrl->ka_cmd));
4453         ctrl->ka_cmd.common.opcode = nvme_admin_keep_alive;
4454
4455         BUILD_BUG_ON(NVME_DSM_MAX_RANGES * sizeof(struct nvme_dsm_range) >
4456                         PAGE_SIZE);
4457         ctrl->discard_page = alloc_page(GFP_KERNEL);
4458         if (!ctrl->discard_page) {
4459                 ret = -ENOMEM;
4460                 goto out;
4461         }
4462
4463         ret = ida_alloc(&nvme_instance_ida, GFP_KERNEL);
4464         if (ret < 0)
4465                 goto out;
4466         ctrl->instance = ret;
4467
4468         device_initialize(&ctrl->ctrl_device);
4469         ctrl->device = &ctrl->ctrl_device;
4470         ctrl->device->devt = MKDEV(MAJOR(nvme_ctrl_base_chr_devt),
4471                         ctrl->instance);
4472         ctrl->device->class = nvme_class;
4473         ctrl->device->parent = ctrl->dev;
4474         if (ops->dev_attr_groups)
4475                 ctrl->device->groups = ops->dev_attr_groups;
4476         else
4477                 ctrl->device->groups = nvme_dev_attr_groups;
4478         ctrl->device->release = nvme_free_ctrl;
4479         dev_set_drvdata(ctrl->device, ctrl);
4480         ret = dev_set_name(ctrl->device, "nvme%d", ctrl->instance);
4481         if (ret)
4482                 goto out_release_instance;
4483
4484         nvme_get_ctrl(ctrl);
4485         cdev_init(&ctrl->cdev, &nvme_dev_fops);
4486         ctrl->cdev.owner = ops->module;
4487         ret = cdev_device_add(&ctrl->cdev, ctrl->device);
4488         if (ret)
4489                 goto out_free_name;
4490
4491         /*
4492          * Initialize latency tolerance controls.  The sysfs files won't
4493          * be visible to userspace unless the device actually supports APST.
4494          */
4495         ctrl->device->power.set_latency_tolerance = nvme_set_latency_tolerance;
4496         dev_pm_qos_update_user_latency_tolerance(ctrl->device,
4497                 min(default_ps_max_latency_us, (unsigned long)S32_MAX));
4498
4499         nvme_fault_inject_init(&ctrl->fault_inject, dev_name(ctrl->device));
4500         nvme_mpath_init_ctrl(ctrl);
4501         ret = nvme_auth_init_ctrl(ctrl);
4502         if (ret)
4503                 goto out_free_cdev;
4504
4505         return 0;
4506 out_free_cdev:
4507         nvme_fault_inject_fini(&ctrl->fault_inject);
4508         dev_pm_qos_hide_latency_tolerance(ctrl->device);
4509         cdev_device_del(&ctrl->cdev, ctrl->device);
4510 out_free_name:
4511         nvme_put_ctrl(ctrl);
4512         kfree_const(ctrl->device->kobj.name);
4513 out_release_instance:
4514         ida_free(&nvme_instance_ida, ctrl->instance);
4515 out:
4516         if (ctrl->discard_page)
4517                 __free_page(ctrl->discard_page);
4518         return ret;
4519 }
4520 EXPORT_SYMBOL_GPL(nvme_init_ctrl);
4521
4522 /* let I/O to all namespaces fail in preparation for surprise removal */
4523 void nvme_mark_namespaces_dead(struct nvme_ctrl *ctrl)
4524 {
4525         struct nvme_ns *ns;
4526
4527         down_read(&ctrl->namespaces_rwsem);
4528         list_for_each_entry(ns, &ctrl->namespaces, list)
4529                 blk_mark_disk_dead(ns->disk);
4530         up_read(&ctrl->namespaces_rwsem);
4531 }
4532 EXPORT_SYMBOL_GPL(nvme_mark_namespaces_dead);
4533
4534 void nvme_unfreeze(struct nvme_ctrl *ctrl)
4535 {
4536         struct nvme_ns *ns;
4537
4538         down_read(&ctrl->namespaces_rwsem);
4539         list_for_each_entry(ns, &ctrl->namespaces, list)
4540                 blk_mq_unfreeze_queue(ns->queue);
4541         up_read(&ctrl->namespaces_rwsem);
4542 }
4543 EXPORT_SYMBOL_GPL(nvme_unfreeze);
4544
4545 int nvme_wait_freeze_timeout(struct nvme_ctrl *ctrl, long timeout)
4546 {
4547         struct nvme_ns *ns;
4548
4549         down_read(&ctrl->namespaces_rwsem);
4550         list_for_each_entry(ns, &ctrl->namespaces, list) {
4551                 timeout = blk_mq_freeze_queue_wait_timeout(ns->queue, timeout);
4552                 if (timeout <= 0)
4553                         break;
4554         }
4555         up_read(&ctrl->namespaces_rwsem);
4556         return timeout;
4557 }
4558 EXPORT_SYMBOL_GPL(nvme_wait_freeze_timeout);
4559
4560 void nvme_wait_freeze(struct nvme_ctrl *ctrl)
4561 {
4562         struct nvme_ns *ns;
4563
4564         down_read(&ctrl->namespaces_rwsem);
4565         list_for_each_entry(ns, &ctrl->namespaces, list)
4566                 blk_mq_freeze_queue_wait(ns->queue);
4567         up_read(&ctrl->namespaces_rwsem);
4568 }
4569 EXPORT_SYMBOL_GPL(nvme_wait_freeze);
4570
4571 void nvme_start_freeze(struct nvme_ctrl *ctrl)
4572 {
4573         struct nvme_ns *ns;
4574
4575         down_read(&ctrl->namespaces_rwsem);
4576         list_for_each_entry(ns, &ctrl->namespaces, list)
4577                 blk_freeze_queue_start(ns->queue);
4578         up_read(&ctrl->namespaces_rwsem);
4579 }
4580 EXPORT_SYMBOL_GPL(nvme_start_freeze);
4581
4582 void nvme_quiesce_io_queues(struct nvme_ctrl *ctrl)
4583 {
4584         if (!ctrl->tagset)
4585                 return;
4586         if (!test_and_set_bit(NVME_CTRL_STOPPED, &ctrl->flags))
4587                 blk_mq_quiesce_tagset(ctrl->tagset);
4588         else
4589                 blk_mq_wait_quiesce_done(ctrl->tagset);
4590 }
4591 EXPORT_SYMBOL_GPL(nvme_quiesce_io_queues);
4592
4593 void nvme_unquiesce_io_queues(struct nvme_ctrl *ctrl)
4594 {
4595         if (!ctrl->tagset)
4596                 return;
4597         if (test_and_clear_bit(NVME_CTRL_STOPPED, &ctrl->flags))
4598                 blk_mq_unquiesce_tagset(ctrl->tagset);
4599 }
4600 EXPORT_SYMBOL_GPL(nvme_unquiesce_io_queues);
4601
4602 void nvme_quiesce_admin_queue(struct nvme_ctrl *ctrl)
4603 {
4604         if (!test_and_set_bit(NVME_CTRL_ADMIN_Q_STOPPED, &ctrl->flags))
4605                 blk_mq_quiesce_queue(ctrl->admin_q);
4606         else
4607                 blk_mq_wait_quiesce_done(ctrl->admin_q->tag_set);
4608 }
4609 EXPORT_SYMBOL_GPL(nvme_quiesce_admin_queue);
4610
4611 void nvme_unquiesce_admin_queue(struct nvme_ctrl *ctrl)
4612 {
4613         if (test_and_clear_bit(NVME_CTRL_ADMIN_Q_STOPPED, &ctrl->flags))
4614                 blk_mq_unquiesce_queue(ctrl->admin_q);
4615 }
4616 EXPORT_SYMBOL_GPL(nvme_unquiesce_admin_queue);
4617
4618 void nvme_sync_io_queues(struct nvme_ctrl *ctrl)
4619 {
4620         struct nvme_ns *ns;
4621
4622         down_read(&ctrl->namespaces_rwsem);
4623         list_for_each_entry(ns, &ctrl->namespaces, list)
4624                 blk_sync_queue(ns->queue);
4625         up_read(&ctrl->namespaces_rwsem);
4626 }
4627 EXPORT_SYMBOL_GPL(nvme_sync_io_queues);
4628
4629 void nvme_sync_queues(struct nvme_ctrl *ctrl)
4630 {
4631         nvme_sync_io_queues(ctrl);
4632         if (ctrl->admin_q)
4633                 blk_sync_queue(ctrl->admin_q);
4634 }
4635 EXPORT_SYMBOL_GPL(nvme_sync_queues);
4636
4637 struct nvme_ctrl *nvme_ctrl_from_file(struct file *file)
4638 {
4639         if (file->f_op != &nvme_dev_fops)
4640                 return NULL;
4641         return file->private_data;
4642 }
4643 EXPORT_SYMBOL_NS_GPL(nvme_ctrl_from_file, NVME_TARGET_PASSTHRU);
4644
4645 /*
4646  * Check we didn't inadvertently grow the command structure sizes:
4647  */
4648 static inline void _nvme_check_size(void)
4649 {
4650         BUILD_BUG_ON(sizeof(struct nvme_common_command) != 64);
4651         BUILD_BUG_ON(sizeof(struct nvme_rw_command) != 64);
4652         BUILD_BUG_ON(sizeof(struct nvme_identify) != 64);
4653         BUILD_BUG_ON(sizeof(struct nvme_features) != 64);
4654         BUILD_BUG_ON(sizeof(struct nvme_download_firmware) != 64);
4655         BUILD_BUG_ON(sizeof(struct nvme_format_cmd) != 64);
4656         BUILD_BUG_ON(sizeof(struct nvme_dsm_cmd) != 64);
4657         BUILD_BUG_ON(sizeof(struct nvme_write_zeroes_cmd) != 64);
4658         BUILD_BUG_ON(sizeof(struct nvme_abort_cmd) != 64);
4659         BUILD_BUG_ON(sizeof(struct nvme_get_log_page_command) != 64);
4660         BUILD_BUG_ON(sizeof(struct nvme_command) != 64);
4661         BUILD_BUG_ON(sizeof(struct nvme_id_ctrl) != NVME_IDENTIFY_DATA_SIZE);
4662         BUILD_BUG_ON(sizeof(struct nvme_id_ns) != NVME_IDENTIFY_DATA_SIZE);
4663         BUILD_BUG_ON(sizeof(struct nvme_id_ns_cs_indep) !=
4664                         NVME_IDENTIFY_DATA_SIZE);
4665         BUILD_BUG_ON(sizeof(struct nvme_id_ns_zns) != NVME_IDENTIFY_DATA_SIZE);
4666         BUILD_BUG_ON(sizeof(struct nvme_id_ns_nvm) != NVME_IDENTIFY_DATA_SIZE);
4667         BUILD_BUG_ON(sizeof(struct nvme_id_ctrl_zns) != NVME_IDENTIFY_DATA_SIZE);
4668         BUILD_BUG_ON(sizeof(struct nvme_id_ctrl_nvm) != NVME_IDENTIFY_DATA_SIZE);
4669         BUILD_BUG_ON(sizeof(struct nvme_lba_range_type) != 64);
4670         BUILD_BUG_ON(sizeof(struct nvme_smart_log) != 512);
4671         BUILD_BUG_ON(sizeof(struct nvme_dbbuf) != 64);
4672         BUILD_BUG_ON(sizeof(struct nvme_directive_cmd) != 64);
4673         BUILD_BUG_ON(sizeof(struct nvme_feat_host_behavior) != 512);
4674 }
4675
4676
4677 static int __init nvme_core_init(void)
4678 {
4679         int result = -ENOMEM;
4680
4681         _nvme_check_size();
4682
4683         nvme_wq = alloc_workqueue("nvme-wq",
4684                         WQ_UNBOUND | WQ_MEM_RECLAIM | WQ_SYSFS, 0);
4685         if (!nvme_wq)
4686                 goto out;
4687
4688         nvme_reset_wq = alloc_workqueue("nvme-reset-wq",
4689                         WQ_UNBOUND | WQ_MEM_RECLAIM | WQ_SYSFS, 0);
4690         if (!nvme_reset_wq)
4691                 goto destroy_wq;
4692
4693         nvme_delete_wq = alloc_workqueue("nvme-delete-wq",
4694                         WQ_UNBOUND | WQ_MEM_RECLAIM | WQ_SYSFS, 0);
4695         if (!nvme_delete_wq)
4696                 goto destroy_reset_wq;
4697
4698         result = alloc_chrdev_region(&nvme_ctrl_base_chr_devt, 0,
4699                         NVME_MINORS, "nvme");
4700         if (result < 0)
4701                 goto destroy_delete_wq;
4702
4703         nvme_class = class_create("nvme");
4704         if (IS_ERR(nvme_class)) {
4705                 result = PTR_ERR(nvme_class);
4706                 goto unregister_chrdev;
4707         }
4708         nvme_class->dev_uevent = nvme_class_uevent;
4709
4710         nvme_subsys_class = class_create("nvme-subsystem");
4711         if (IS_ERR(nvme_subsys_class)) {
4712                 result = PTR_ERR(nvme_subsys_class);
4713                 goto destroy_class;
4714         }
4715
4716         result = alloc_chrdev_region(&nvme_ns_chr_devt, 0, NVME_MINORS,
4717                                      "nvme-generic");
4718         if (result < 0)
4719                 goto destroy_subsys_class;
4720
4721         nvme_ns_chr_class = class_create("nvme-generic");
4722         if (IS_ERR(nvme_ns_chr_class)) {
4723                 result = PTR_ERR(nvme_ns_chr_class);
4724                 goto unregister_generic_ns;
4725         }
4726
4727         result = nvme_init_auth();
4728         if (result)
4729                 goto destroy_ns_chr;
4730         return 0;
4731
4732 destroy_ns_chr:
4733         class_destroy(nvme_ns_chr_class);
4734 unregister_generic_ns:
4735         unregister_chrdev_region(nvme_ns_chr_devt, NVME_MINORS);
4736 destroy_subsys_class:
4737         class_destroy(nvme_subsys_class);
4738 destroy_class:
4739         class_destroy(nvme_class);
4740 unregister_chrdev:
4741         unregister_chrdev_region(nvme_ctrl_base_chr_devt, NVME_MINORS);
4742 destroy_delete_wq:
4743         destroy_workqueue(nvme_delete_wq);
4744 destroy_reset_wq:
4745         destroy_workqueue(nvme_reset_wq);
4746 destroy_wq:
4747         destroy_workqueue(nvme_wq);
4748 out:
4749         return result;
4750 }
4751
4752 static void __exit nvme_core_exit(void)
4753 {
4754         nvme_exit_auth();
4755         class_destroy(nvme_ns_chr_class);
4756         class_destroy(nvme_subsys_class);
4757         class_destroy(nvme_class);
4758         unregister_chrdev_region(nvme_ns_chr_devt, NVME_MINORS);
4759         unregister_chrdev_region(nvme_ctrl_base_chr_devt, NVME_MINORS);
4760         destroy_workqueue(nvme_delete_wq);
4761         destroy_workqueue(nvme_reset_wq);
4762         destroy_workqueue(nvme_wq);
4763         ida_destroy(&nvme_ns_chr_minor_ida);
4764         ida_destroy(&nvme_instance_ida);
4765 }
4766
4767 MODULE_LICENSE("GPL");
4768 MODULE_VERSION("1.0");
4769 module_init(nvme_core_init);
4770 module_exit(nvme_core_exit);