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
  28 #define CREATE_TRACE_POINTS
  29 #include "trace.h"
  30
  31 #define NVME_MINORS             (1U << MINORBITS)
  32
  33 unsigned int admin_timeout = 60;
  34 module_param(admin_timeout, uint, 0644);
  35 MODULE_PARM_DESC(admin_timeout, "timeout in seconds for admin commands");
  36 EXPORT_SYMBOL_GPL(admin_timeout);
  37
  38 unsigned int nvme_io_timeout = 30;
  39 module_param_named(io_timeout, nvme_io_timeout, uint, 0644);
  40 MODULE_PARM_DESC(io_timeout, "timeout in seconds for I/O");
  41 EXPORT_SYMBOL_GPL(nvme_io_timeout);
  42
  43 static unsigned char shutdown_timeout = 5;
  44 module_param(shutdown_timeout, byte, 0644);
  45 MODULE_PARM_DESC(shutdown_timeout, "timeout in seconds for controller shutdown");
  46
  47 static u8 nvme_max_retries = 5;
  48 module_param_named(max_retries, nvme_max_retries, byte, 0644);
  49 MODULE_PARM_DESC(max_retries, "max number of retries a command may have");
  50
  51 static unsigned long default_ps_max_latency_us = 100000;
  52 module_param(default_ps_max_latency_us, ulong, 0644);
  53 MODULE_PARM_DESC(default_ps_max_latency_us,
  54                  "max power saving latency for new devices; use PM QOS to change per device");
  55
  56 static bool force_apst;
  57 module_param(force_apst, bool, 0644);
  58 MODULE_PARM_DESC(force_apst, "allow APST for newly enumerated devices even if quirked off");
  59
  60 static unsigned long apst_primary_timeout_ms = 100;
  61 module_param(apst_primary_timeout_ms, ulong, 0644);
  62 MODULE_PARM_DESC(apst_primary_timeout_ms,
  63         "primary APST timeout in ms");
  64
  65 static unsigned long apst_secondary_timeout_ms = 2000;
  66 module_param(apst_secondary_timeout_ms, ulong, 0644);
  67 MODULE_PARM_DESC(apst_secondary_timeout_ms,
  68         "secondary APST timeout in ms");
  69
  70 static unsigned long apst_primary_latency_tol_us = 15000;
  71 module_param(apst_primary_latency_tol_us, ulong, 0644);
  72 MODULE_PARM_DESC(apst_primary_latency_tol_us,
  73         "primary APST latency tolerance in us");
  74
  75 static unsigned long apst_secondary_latency_tol_us = 100000;
  76 module_param(apst_secondary_latency_tol_us, ulong, 0644);
  77 MODULE_PARM_DESC(apst_secondary_latency_tol_us,
  78         "secondary APST latency tolerance in us");
  79
  80 /*
  81  * nvme_wq - hosts nvme related works that are not reset or delete
  82  * nvme_reset_wq - hosts nvme reset works
  83  * nvme_delete_wq - hosts nvme delete works
  84  *
  85  * nvme_wq will host works such as scan, aen handling, fw activation,
  86  * keep-alive, periodic reconnects etc. nvme_reset_wq
  87  * runs reset works which also flush works hosted on nvme_wq for
  88  * serialization purposes. nvme_delete_wq host controller deletion
  89  * works which flush reset works for serialization.
  90  */
  91 struct workqueue_struct *nvme_wq;
  92 EXPORT_SYMBOL_GPL(nvme_wq);
  93
  94 struct workqueue_struct *nvme_reset_wq;
  95 EXPORT_SYMBOL_GPL(nvme_reset_wq);
  96
  97 struct workqueue_struct *nvme_delete_wq;
  98 EXPORT_SYMBOL_GPL(nvme_delete_wq);
  99
 100 static LIST_HEAD(nvme_subsystems);
 101 static DEFINE_MUTEX(nvme_subsystems_lock);
 102
 103 static DEFINE_IDA(nvme_instance_ida);
 104 static dev_t nvme_ctrl_base_chr_devt;
 105 static struct class *nvme_class;
 106 static struct class *nvme_subsys_class;
 107
 108 static DEFINE_IDA(nvme_ns_chr_minor_ida);
 109 static dev_t nvme_ns_chr_devt;
 110 static struct class *nvme_ns_chr_class;
 111
 112 static void nvme_put_subsystem(struct nvme_subsystem *subsys);
 113 static void nvme_remove_invalid_namespaces(struct nvme_ctrl *ctrl,
 114                                            unsigned nsid);
 115 static void nvme_update_keep_alive(struct nvme_ctrl *ctrl,
 116                                    struct nvme_command *cmd);
 117
 118 void nvme_queue_scan(struct nvme_ctrl *ctrl)
 119 {
 120         /*
 121          * Only new queue scan work when admin and IO queues are both alive
 122          */
 123         if (ctrl->state == NVME_CTRL_LIVE && ctrl->tagset)
 124                 queue_work(nvme_wq, &ctrl->scan_work);
 125 }
 126
 127 /*
 128  * Use this function to proceed with scheduling reset_work for a controller
 129  * that had previously been set to the resetting state. This is intended for
 130  * code paths that can't be interrupted by other reset attempts. A hot removal
 131  * may prevent this from succeeding.
 132  */
 133 int nvme_try_sched_reset(struct nvme_ctrl *ctrl)
 134 {
 135         if (ctrl->state != NVME_CTRL_RESETTING)
 136                 return -EBUSY;
 137         if (!queue_work(nvme_reset_wq, &ctrl->reset_work))
 138                 return -EBUSY;
 139         return 0;
 140 }
 141 EXPORT_SYMBOL_GPL(nvme_try_sched_reset);
 142
 143 static void nvme_failfast_work(struct work_struct *work)
 144 {
 145         struct nvme_ctrl *ctrl = container_of(to_delayed_work(work),
 146                         struct nvme_ctrl, failfast_work);
 147
 148         if (ctrl->state != NVME_CTRL_CONNECTING)
 149                 return;
 150
 151         set_bit(NVME_CTRL_FAILFAST_EXPIRED, &ctrl->flags);
 152         dev_info(ctrl->device, "failfast expired\n");
 153         nvme_kick_requeue_lists(ctrl);
 154 }
 155
 156 static inline void nvme_start_failfast_work(struct nvme_ctrl *ctrl)
 157 {
 158         if (!ctrl->opts || ctrl->opts->fast_io_fail_tmo == -1)
 159                 return;
 160
 161         schedule_delayed_work(&ctrl->failfast_work,
 162                               ctrl->opts->fast_io_fail_tmo * HZ);
 163 }
 164
 165 static inline void nvme_stop_failfast_work(struct nvme_ctrl *ctrl)
 166 {
 167         if (!ctrl->opts)
 168                 return;
 169
 170         cancel_delayed_work_sync(&ctrl->failfast_work);
 171         clear_bit(NVME_CTRL_FAILFAST_EXPIRED, &ctrl->flags);
 172 }
 173
 174
 175 int nvme_reset_ctrl(struct nvme_ctrl *ctrl)
 176 {
 177         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING))
 178                 return -EBUSY;
 179         if (!queue_work(nvme_reset_wq, &ctrl->reset_work))
 180                 return -EBUSY;
 181         return 0;
 182 }
 183 EXPORT_SYMBOL_GPL(nvme_reset_ctrl);
 184
 185 int nvme_reset_ctrl_sync(struct nvme_ctrl *ctrl)
 186 {
 187         int ret;
 188
 189         ret = nvme_reset_ctrl(ctrl);
 190         if (!ret) {
 191                 flush_work(&ctrl->reset_work);
 192                 if (ctrl->state != NVME_CTRL_LIVE)
 193                         ret = -ENETRESET;
 194         }
 195
 196         return ret;
 197 }
 198
 199 static void nvme_do_delete_ctrl(struct nvme_ctrl *ctrl)
 200 {
 201         dev_info(ctrl->device,
 202                  "Removing ctrl: NQN \"%s\"\n", nvmf_ctrl_subsysnqn(ctrl));
 203
 204         flush_work(&ctrl->reset_work);
 205         nvme_stop_ctrl(ctrl);
 206         nvme_remove_namespaces(ctrl);
 207         ctrl->ops->delete_ctrl(ctrl);
 208         nvme_uninit_ctrl(ctrl);
 209 }
 210
 211 static void nvme_delete_ctrl_work(struct work_struct *work)
 212 {
 213         struct nvme_ctrl *ctrl =
 214                 container_of(work, struct nvme_ctrl, delete_work);
 215
 216         nvme_do_delete_ctrl(ctrl);
 217 }
 218
 219 int nvme_delete_ctrl(struct nvme_ctrl *ctrl)
 220 {
 221         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_DELETING))
 222                 return -EBUSY;
 223         if (!queue_work(nvme_delete_wq, &ctrl->delete_work))
 224                 return -EBUSY;
 225         return 0;
 226 }
 227 EXPORT_SYMBOL_GPL(nvme_delete_ctrl);
 228
 229 static void nvme_delete_ctrl_sync(struct nvme_ctrl *ctrl)
 230 {
 231         /*
 232          * Keep a reference until nvme_do_delete_ctrl() complete,
 233          * since ->delete_ctrl can free the controller.
 234          */
 235         nvme_get_ctrl(ctrl);
 236         if (nvme_change_ctrl_state(ctrl, NVME_CTRL_DELETING))
 237                 nvme_do_delete_ctrl(ctrl);
 238         nvme_put_ctrl(ctrl);
 239 }
 240
 241 static blk_status_t nvme_error_status(u16 status)
 242 {
 243         switch (status & 0x7ff) {
 244         case NVME_SC_SUCCESS:
 245                 return BLK_STS_OK;
 246         case NVME_SC_CAP_EXCEEDED:
 247                 return BLK_STS_NOSPC;
 248         case NVME_SC_LBA_RANGE:
 249         case NVME_SC_CMD_INTERRUPTED:
 250         case NVME_SC_NS_NOT_READY:
 251                 return BLK_STS_TARGET;
 252         case NVME_SC_BAD_ATTRIBUTES:
 253         case NVME_SC_ONCS_NOT_SUPPORTED:
 254         case NVME_SC_INVALID_OPCODE:
 255         case NVME_SC_INVALID_FIELD:
 256         case NVME_SC_INVALID_NS:
 257                 return BLK_STS_NOTSUPP;
 258         case NVME_SC_WRITE_FAULT:
 259         case NVME_SC_READ_ERROR:
 260         case NVME_SC_UNWRITTEN_BLOCK:
 261         case NVME_SC_ACCESS_DENIED:
 262         case NVME_SC_READ_ONLY:
 263         case NVME_SC_COMPARE_FAILED:
 264                 return BLK_STS_MEDIUM;
 265         case NVME_SC_GUARD_CHECK:
 266         case NVME_SC_APPTAG_CHECK:
 267         case NVME_SC_REFTAG_CHECK:
 268         case NVME_SC_INVALID_PI:
 269                 return BLK_STS_PROTECTION;
 270         case NVME_SC_RESERVATION_CONFLICT:
 271                 return BLK_STS_NEXUS;
 272         case NVME_SC_HOST_PATH_ERROR:
 273                 return BLK_STS_TRANSPORT;
 274         case NVME_SC_ZONE_TOO_MANY_ACTIVE:
 275                 return BLK_STS_ZONE_ACTIVE_RESOURCE;
 276         case NVME_SC_ZONE_TOO_MANY_OPEN:
 277                 return BLK_STS_ZONE_OPEN_RESOURCE;
 278         default:
 279                 return BLK_STS_IOERR;
 280         }
 281 }
 282
 283 static void nvme_retry_req(struct request *req)
 284 {
 285         unsigned long delay = 0;
 286         u16 crd;
 287
 288         /* The mask and shift result must be <= 3 */
 289         crd = (nvme_req(req)->status & NVME_SC_CRD) >> 11;
 290         if (crd)
 291                 delay = nvme_req(req)->ctrl->crdt[crd - 1] * 100;
 292
 293         nvme_req(req)->retries++;
 294         blk_mq_requeue_request(req, false);
 295         blk_mq_delay_kick_requeue_list(req->q, delay);
 296 }
 297
 298 static void nvme_log_error(struct request *req)
 299 {
 300         struct nvme_ns *ns = req->q->queuedata;
 301         struct nvme_request *nr = nvme_req(req);
 302
 303         if (ns) {
 304                 pr_err_ratelimited("%s: %s(0x%x) @ LBA %llu, %llu blocks, %s (sct 0x%x / sc 0x%x) %s%s\n",
 305                        ns->disk ? ns->disk->disk_name : "?",
 306                        nvme_get_opcode_str(nr->cmd->common.opcode),
 307                        nr->cmd->common.opcode,
 308                        (unsigned long long)nvme_sect_to_lba(ns, blk_rq_pos(req)),
 309                        (unsigned long long)blk_rq_bytes(req) >> ns->lba_shift,
 310                        nvme_get_error_status_str(nr->status),
 311                        nr->status >> 8 & 7,     /* Status Code Type */
 312                        nr->status & 0xff,       /* Status Code */
 313                        nr->status & NVME_SC_MORE ? "MORE " : "",
 314                        nr->status & NVME_SC_DNR  ? "DNR "  : "");
 315                 return;
 316         }
 317
 318         pr_err_ratelimited("%s: %s(0x%x), %s (sct 0x%x / sc 0x%x) %s%s\n",
 319                            dev_name(nr->ctrl->device),
 320                            nvme_get_admin_opcode_str(nr->cmd->common.opcode),
 321                            nr->cmd->common.opcode,
 322                            nvme_get_error_status_str(nr->status),
 323                            nr->status >> 8 & 7, /* Status Code Type */
 324                            nr->status & 0xff,   /* Status Code */
 325                            nr->status & NVME_SC_MORE ? "MORE " : "",
 326                            nr->status & NVME_SC_DNR  ? "DNR "  : "");
 327 }
 328
 329 enum nvme_disposition {
 330         COMPLETE,
 331         RETRY,
 332         FAILOVER,
 333 };
 334
 335 static inline enum nvme_disposition nvme_decide_disposition(struct request *req)
 336 {
 337         if (likely(nvme_req(req)->status == 0))
 338                 return COMPLETE;
 339
 340         if (blk_noretry_request(req) ||
 341             (nvme_req(req)->status & NVME_SC_DNR) ||
 342             nvme_req(req)->retries >= nvme_max_retries)
 343                 return COMPLETE;
 344
 345         if (req->cmd_flags & REQ_NVME_MPATH) {
 346                 if (nvme_is_path_error(nvme_req(req)->status) ||
 347                     blk_queue_dying(req->q))
 348                         return FAILOVER;
 349         } else {
 350                 if (blk_queue_dying(req->q))
 351                         return COMPLETE;
 352         }
 353
 354         return RETRY;
 355 }
 356
 357 static inline void nvme_end_req_zoned(struct request *req)
 358 {
 359         if (IS_ENABLED(CONFIG_BLK_DEV_ZONED) &&
 360             req_op(req) == REQ_OP_ZONE_APPEND)
 361                 req->__sector = nvme_lba_to_sect(req->q->queuedata,
 362                         le64_to_cpu(nvme_req(req)->result.u64));
 363 }
 364
 365 static inline void nvme_end_req(struct request *req)
 366 {
 367         blk_status_t status = nvme_error_status(nvme_req(req)->status);
 368
 369         if (unlikely(nvme_req(req)->status && !(req->rq_flags & RQF_QUIET)))
 370                 nvme_log_error(req);
 371         nvme_end_req_zoned(req);
 372         nvme_trace_bio_complete(req);
 373         blk_mq_end_request(req, status);
 374 }
 375
 376 void nvme_complete_rq(struct request *req)
 377 {
 378         trace_nvme_complete_rq(req);
 379         nvme_cleanup_cmd(req);
 380
 381         if (nvme_req(req)->ctrl->kas)
 382                 nvme_req(req)->ctrl->comp_seen = true;
 383
 384         switch (nvme_decide_disposition(req)) {
 385         case COMPLETE:
 386                 nvme_end_req(req);
 387                 return;
 388         case RETRY:
 389                 nvme_retry_req(req);
 390                 return;
 391         case FAILOVER:
 392                 nvme_failover_req(req);
 393                 return;
 394         }
 395 }
 396 EXPORT_SYMBOL_GPL(nvme_complete_rq);
 397
 398 void nvme_complete_batch_req(struct request *req)
 399 {
 400         trace_nvme_complete_rq(req);
 401         nvme_cleanup_cmd(req);
 402         nvme_end_req_zoned(req);
 403 }
 404 EXPORT_SYMBOL_GPL(nvme_complete_batch_req);
 405
 406 /*
 407  * Called to unwind from ->queue_rq on a failed command submission so that the
 408  * multipathing code gets called to potentially failover to another path.
 409  * The caller needs to unwind all transport specific resource allocations and
 410  * must return propagate the return value.
 411  */
 412 blk_status_t nvme_host_path_error(struct request *req)
 413 {
 414         nvme_req(req)->status = NVME_SC_HOST_PATH_ERROR;
 415         blk_mq_set_request_complete(req);
 416         nvme_complete_rq(req);
 417         return BLK_STS_OK;
 418 }
 419 EXPORT_SYMBOL_GPL(nvme_host_path_error);
 420
 421 bool nvme_cancel_request(struct request *req, void *data, bool reserved)
 422 {
 423         dev_dbg_ratelimited(((struct nvme_ctrl *) data)->device,
 424                                 "Cancelling I/O %d", req->tag);
 425
 426         /* don't abort one completed request */
 427         if (blk_mq_request_completed(req))
 428                 return true;
 429
 430         nvme_req(req)->status = NVME_SC_HOST_ABORTED_CMD;
 431         nvme_req(req)->flags |= NVME_REQ_CANCELLED;
 432         blk_mq_complete_request(req);
 433         return true;
 434 }
 435 EXPORT_SYMBOL_GPL(nvme_cancel_request);
 436
 437 void nvme_cancel_tagset(struct nvme_ctrl *ctrl)
 438 {
 439         if (ctrl->tagset) {
 440                 blk_mq_tagset_busy_iter(ctrl->tagset,
 441                                 nvme_cancel_request, ctrl);
 442                 blk_mq_tagset_wait_completed_request(ctrl->tagset);
 443         }
 444 }
 445 EXPORT_SYMBOL_GPL(nvme_cancel_tagset);
 446
 447 void nvme_cancel_admin_tagset(struct nvme_ctrl *ctrl)
 448 {
 449         if (ctrl->admin_tagset) {
 450                 blk_mq_tagset_busy_iter(ctrl->admin_tagset,
 451                                 nvme_cancel_request, ctrl);
 452                 blk_mq_tagset_wait_completed_request(ctrl->admin_tagset);
 453         }
 454 }
 455 EXPORT_SYMBOL_GPL(nvme_cancel_admin_tagset);
 456
 457 bool nvme_change_ctrl_state(struct nvme_ctrl *ctrl,
 458                 enum nvme_ctrl_state new_state)
 459 {
 460         enum nvme_ctrl_state old_state;
 461         unsigned long flags;
 462         bool changed = false;
 463
 464         spin_lock_irqsave(&ctrl->lock, flags);
 465
 466         old_state = ctrl->state;
 467         switch (new_state) {
 468         case NVME_CTRL_LIVE:
 469                 switch (old_state) {
 470                 case NVME_CTRL_NEW:
 471                 case NVME_CTRL_RESETTING:
 472                 case NVME_CTRL_CONNECTING:
 473                         changed = true;
 474                         fallthrough;
 475                 default:
 476                         break;
 477                 }
 478                 break;
 479         case NVME_CTRL_RESETTING:
 480                 switch (old_state) {
 481                 case NVME_CTRL_NEW:
 482                 case NVME_CTRL_LIVE:
 483                         changed = true;
 484                         fallthrough;
 485                 default:
 486                         break;
 487                 }
 488                 break;
 489         case NVME_CTRL_CONNECTING:
 490                 switch (old_state) {
 491                 case NVME_CTRL_NEW:
 492                 case NVME_CTRL_RESETTING:
 493                         changed = true;
 494                         fallthrough;
 495                 default:
 496                         break;
 497                 }
 498                 break;
 499         case NVME_CTRL_DELETING:
 500                 switch (old_state) {
 501                 case NVME_CTRL_LIVE:
 502                 case NVME_CTRL_RESETTING:
 503                 case NVME_CTRL_CONNECTING:
 504                         changed = true;
 505                         fallthrough;
 506                 default:
 507                         break;
 508                 }
 509                 break;
 510         case NVME_CTRL_DELETING_NOIO:
 511                 switch (old_state) {
 512                 case NVME_CTRL_DELETING:
 513                 case NVME_CTRL_DEAD:
 514                         changed = true;
 515                         fallthrough;
 516                 default:
 517                         break;
 518                 }
 519                 break;
 520         case NVME_CTRL_DEAD:
 521                 switch (old_state) {
 522                 case NVME_CTRL_DELETING:
 523                         changed = true;
 524                         fallthrough;
 525                 default:
 526                         break;
 527                 }
 528                 break;
 529         default:
 530                 break;
 531         }
 532
 533         if (changed) {
 534                 ctrl->state = new_state;
 535                 wake_up_all(&ctrl->state_wq);
 536         }
 537
 538         spin_unlock_irqrestore(&ctrl->lock, flags);
 539         if (!changed)
 540                 return false;
 541
 542         if (ctrl->state == NVME_CTRL_LIVE) {
 543                 if (old_state == NVME_CTRL_CONNECTING)
 544                         nvme_stop_failfast_work(ctrl);
 545                 nvme_kick_requeue_lists(ctrl);
 546         } else if (ctrl->state == NVME_CTRL_CONNECTING &&
 547                 old_state == NVME_CTRL_RESETTING) {
 548                 nvme_start_failfast_work(ctrl);
 549         }
 550         return changed;
 551 }
 552 EXPORT_SYMBOL_GPL(nvme_change_ctrl_state);
 553
 554 /*
 555  * Returns true for sink states that can't ever transition back to live.
 556  */
 557 static bool nvme_state_terminal(struct nvme_ctrl *ctrl)
 558 {
 559         switch (ctrl->state) {
 560         case NVME_CTRL_NEW:
 561         case NVME_CTRL_LIVE:
 562         case NVME_CTRL_RESETTING:
 563         case NVME_CTRL_CONNECTING:
 564                 return false;
 565         case NVME_CTRL_DELETING:
 566         case NVME_CTRL_DELETING_NOIO:
 567         case NVME_CTRL_DEAD:
 568                 return true;
 569         default:
 570                 WARN_ONCE(1, "Unhandled ctrl state:%d", ctrl->state);
 571                 return true;
 572         }
 573 }
 574
 575 /*
 576  * Waits for the controller state to be resetting, or returns false if it is
 577  * not possible to ever transition to that state.
 578  */
 579 bool nvme_wait_reset(struct nvme_ctrl *ctrl)
 580 {
 581         wait_event(ctrl->state_wq,
 582                    nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING) ||
 583                    nvme_state_terminal(ctrl));
 584         return ctrl->state == NVME_CTRL_RESETTING;
 585 }
 586 EXPORT_SYMBOL_GPL(nvme_wait_reset);
 587
 588 static void nvme_free_ns_head(struct kref *ref)
 589 {
 590         struct nvme_ns_head *head =
 591                 container_of(ref, struct nvme_ns_head, ref);
 592
 593         nvme_mpath_remove_disk(head);
 594         ida_free(&head->subsys->ns_ida, head->instance);
 595         cleanup_srcu_struct(&head->srcu);
 596         nvme_put_subsystem(head->subsys);
 597         kfree(head);
 598 }
 599
 600 bool nvme_tryget_ns_head(struct nvme_ns_head *head)
 601 {
 602         return kref_get_unless_zero(&head->ref);
 603 }
 604
 605 void nvme_put_ns_head(struct nvme_ns_head *head)
 606 {
 607         kref_put(&head->ref, nvme_free_ns_head);
 608 }
 609
 610 static void nvme_free_ns(struct kref *kref)
 611 {
 612         struct nvme_ns *ns = container_of(kref, struct nvme_ns, kref);
 613
 614         put_disk(ns->disk);
 615         nvme_put_ns_head(ns->head);
 616         nvme_put_ctrl(ns->ctrl);
 617         kfree(ns);
 618 }
 619
 620 static inline bool nvme_get_ns(struct nvme_ns *ns)
 621 {
 622         return kref_get_unless_zero(&ns->kref);
 623 }
 624
 625 void nvme_put_ns(struct nvme_ns *ns)
 626 {
 627         kref_put(&ns->kref, nvme_free_ns);
 628 }
 629 EXPORT_SYMBOL_NS_GPL(nvme_put_ns, NVME_TARGET_PASSTHRU);
 630
 631 static inline void nvme_clear_nvme_request(struct request *req)
 632 {
 633         nvme_req(req)->status = 0;
 634         nvme_req(req)->retries = 0;
 635         nvme_req(req)->flags = 0;
 636         req->rq_flags |= RQF_DONTPREP;
 637 }
 638
 639 /* initialize a passthrough request */
 640 void nvme_init_request(struct request *req, struct nvme_command *cmd)
 641 {
 642         if (req->q->queuedata)
 643                 req->timeout = NVME_IO_TIMEOUT;
 644         else /* no queuedata implies admin queue */
 645                 req->timeout = NVME_ADMIN_TIMEOUT;
 646
 647         /* passthru commands should let the driver set the SGL flags */
 648         cmd->common.flags &= ~NVME_CMD_SGL_ALL;
 649
 650         req->cmd_flags |= REQ_FAILFAST_DRIVER;
 651         if (req->mq_hctx->type == HCTX_TYPE_POLL)
 652                 req->cmd_flags |= REQ_POLLED;
 653         nvme_clear_nvme_request(req);
 654         memcpy(nvme_req(req)->cmd, cmd, sizeof(*cmd));
 655 }
 656 EXPORT_SYMBOL_GPL(nvme_init_request);
 657
 658 /*
 659  * For something we're not in a state to send to the device the default action
 660  * is to busy it and retry it after the controller state is recovered.  However,
 661  * if the controller is deleting or if anything is marked for failfast or
 662  * nvme multipath it is immediately failed.
 663  *
 664  * Note: commands used to initialize the controller will be marked for failfast.
 665  * Note: nvme cli/ioctl commands are marked for failfast.
 666  */
 667 blk_status_t nvme_fail_nonready_command(struct nvme_ctrl *ctrl,
 668                 struct request *rq)
 669 {
 670         if (ctrl->state != NVME_CTRL_DELETING_NOIO &&
 671             ctrl->state != NVME_CTRL_DELETING &&
 672             ctrl->state != NVME_CTRL_DEAD &&
 673             !test_bit(NVME_CTRL_FAILFAST_EXPIRED, &ctrl->flags) &&
 674             !blk_noretry_request(rq) && !(rq->cmd_flags & REQ_NVME_MPATH))
 675                 return BLK_STS_RESOURCE;
 676         return nvme_host_path_error(rq);
 677 }
 678 EXPORT_SYMBOL_GPL(nvme_fail_nonready_command);
 679
 680 bool __nvme_check_ready(struct nvme_ctrl *ctrl, struct request *rq,
 681                 bool queue_live)
 682 {
 683         struct nvme_request *req = nvme_req(rq);
 684
 685         /*
 686          * currently we have a problem sending passthru commands
 687          * on the admin_q if the controller is not LIVE because we can't
 688          * make sure that they are going out after the admin connect,
 689          * controller enable and/or other commands in the initialization
 690          * sequence. until the controller will be LIVE, fail with
 691          * BLK_STS_RESOURCE so that they will be rescheduled.
 692          */
 693         if (rq->q == ctrl->admin_q && (req->flags & NVME_REQ_USERCMD))
 694                 return false;
 695
 696         if (ctrl->ops->flags & NVME_F_FABRICS) {
 697                 /*
 698                  * Only allow commands on a live queue, except for the connect
 699                  * command, which is require to set the queue live in the
 700                  * appropinquate states.
 701                  */
 702                 switch (ctrl->state) {
 703                 case NVME_CTRL_CONNECTING:
 704                         if (blk_rq_is_passthrough(rq) && nvme_is_fabrics(req->cmd) &&
 705                             req->cmd->fabrics.fctype == nvme_fabrics_type_connect)
 706                                 return true;
 707                         break;
 708                 default:
 709                         break;
 710                 case NVME_CTRL_DEAD:
 711                         return false;
 712                 }
 713         }
 714
 715         return queue_live;
 716 }
 717 EXPORT_SYMBOL_GPL(__nvme_check_ready);
 718
 719 static inline void nvme_setup_flush(struct nvme_ns *ns,
 720                 struct nvme_command *cmnd)
 721 {
 722         memset(cmnd, 0, sizeof(*cmnd));
 723         cmnd->common.opcode = nvme_cmd_flush;
 724         cmnd->common.nsid = cpu_to_le32(ns->head->ns_id);
 725 }
 726
 727 static blk_status_t nvme_setup_discard(struct nvme_ns *ns, struct request *req,
 728                 struct nvme_command *cmnd)
 729 {
 730         unsigned short segments = blk_rq_nr_discard_segments(req), n = 0;
 731         struct nvme_dsm_range *range;
 732         struct bio *bio;
 733
 734         /*
 735          * Some devices do not consider the DSM 'Number of Ranges' field when
 736          * determining how much data to DMA. Always allocate memory for maximum
 737          * number of segments to prevent device reading beyond end of buffer.
 738          */
 739         static const size_t alloc_size = sizeof(*range) * NVME_DSM_MAX_RANGES;
 740
 741         range = kzalloc(alloc_size, GFP_ATOMIC | __GFP_NOWARN);
 742         if (!range) {
 743                 /*
 744                  * If we fail allocation our range, fallback to the controller
 745                  * discard page. If that's also busy, it's safe to return
 746                  * busy, as we know we can make progress once that's freed.
 747                  */
 748                 if (test_and_set_bit_lock(0, &ns->ctrl->discard_page_busy))
 749                         return BLK_STS_RESOURCE;
 750
 751                 range = page_address(ns->ctrl->discard_page);
 752         }
 753
 754         __rq_for_each_bio(bio, req) {
 755                 u64 slba = nvme_sect_to_lba(ns, bio->bi_iter.bi_sector);
 756                 u32 nlb = bio->bi_iter.bi_size >> ns->lba_shift;
 757
 758                 if (n < segments) {
 759                         range[n].cattr = cpu_to_le32(0);
 760                         range[n].nlb = cpu_to_le32(nlb);
 761                         range[n].slba = cpu_to_le64(slba);
 762                 }
 763                 n++;
 764         }
 765
 766         if (WARN_ON_ONCE(n != segments)) {
 767                 if (virt_to_page(range) == ns->ctrl->discard_page)
 768                         clear_bit_unlock(0, &ns->ctrl->discard_page_busy);
 769                 else
 770                         kfree(range);
 771                 return BLK_STS_IOERR;
 772         }
 773
 774         memset(cmnd, 0, sizeof(*cmnd));
 775         cmnd->dsm.opcode = nvme_cmd_dsm;
 776         cmnd->dsm.nsid = cpu_to_le32(ns->head->ns_id);
 777         cmnd->dsm.nr = cpu_to_le32(segments - 1);
 778         cmnd->dsm.attributes = cpu_to_le32(NVME_DSMGMT_AD);
 779
 780         req->special_vec.bv_page = virt_to_page(range);
 781         req->special_vec.bv_offset = offset_in_page(range);
 782         req->special_vec.bv_len = alloc_size;
 783         req->rq_flags |= RQF_SPECIAL_PAYLOAD;
 784
 785         return BLK_STS_OK;
 786 }
 787
 788 static void nvme_set_ref_tag(struct nvme_ns *ns, struct nvme_command *cmnd,
 789                               struct request *req)
 790 {
 791         u32 upper, lower;
 792         u64 ref48;
 793
 794         /* both rw and write zeroes share the same reftag format */
 795         switch (ns->guard_type) {
 796         case NVME_NVM_NS_16B_GUARD:
 797                 cmnd->rw.reftag = cpu_to_le32(t10_pi_ref_tag(req));
 798                 break;
 799         case NVME_NVM_NS_64B_GUARD:
 800                 ref48 = ext_pi_ref_tag(req);
 801                 lower = lower_32_bits(ref48);
 802                 upper = upper_32_bits(ref48);
 803
 804                 cmnd->rw.reftag = cpu_to_le32(lower);
 805                 cmnd->rw.cdw3 = cpu_to_le32(upper);
 806                 break;
 807         default:
 808                 break;
 809         }
 810 }
 811
 812 static inline blk_status_t nvme_setup_write_zeroes(struct nvme_ns *ns,
 813                 struct request *req, struct nvme_command *cmnd)
 814 {
 815         memset(cmnd, 0, sizeof(*cmnd));
 816
 817         if (ns->ctrl->quirks & NVME_QUIRK_DEALLOCATE_ZEROES)
 818                 return nvme_setup_discard(ns, req, cmnd);
 819
 820         cmnd->write_zeroes.opcode = nvme_cmd_write_zeroes;
 821         cmnd->write_zeroes.nsid = cpu_to_le32(ns->head->ns_id);
 822         cmnd->write_zeroes.slba =
 823                 cpu_to_le64(nvme_sect_to_lba(ns, blk_rq_pos(req)));
 824         cmnd->write_zeroes.length =
 825                 cpu_to_le16((blk_rq_bytes(req) >> ns->lba_shift) - 1);
 826
 827         if (nvme_ns_has_pi(ns)) {
 828                 cmnd->write_zeroes.control = cpu_to_le16(NVME_RW_PRINFO_PRACT);
 829
 830                 switch (ns->pi_type) {
 831                 case NVME_NS_DPS_PI_TYPE1:
 832                 case NVME_NS_DPS_PI_TYPE2:
 833                         nvme_set_ref_tag(ns, cmnd, req);
 834                         break;
 835                 }
 836         }
 837
 838         return BLK_STS_OK;
 839 }
 840
 841 static inline blk_status_t nvme_setup_rw(struct nvme_ns *ns,
 842                 struct request *req, struct nvme_command *cmnd,
 843                 enum nvme_opcode op)
 844 {
 845         u16 control = 0;
 846         u32 dsmgmt = 0;
 847
 848         if (req->cmd_flags & REQ_FUA)
 849                 control |= NVME_RW_FUA;
 850         if (req->cmd_flags & (REQ_FAILFAST_DEV | REQ_RAHEAD))
 851                 control |= NVME_RW_LR;
 852
 853         if (req->cmd_flags & REQ_RAHEAD)
 854                 dsmgmt |= NVME_RW_DSM_FREQ_PREFETCH;
 855
 856         cmnd->rw.opcode = op;
 857         cmnd->rw.flags = 0;
 858         cmnd->rw.nsid = cpu_to_le32(ns->head->ns_id);
 859         cmnd->rw.cdw2 = 0;
 860         cmnd->rw.cdw3 = 0;
 861         cmnd->rw.metadata = 0;
 862         cmnd->rw.slba = cpu_to_le64(nvme_sect_to_lba(ns, blk_rq_pos(req)));
 863         cmnd->rw.length = cpu_to_le16((blk_rq_bytes(req) >> ns->lba_shift) - 1);
 864         cmnd->rw.reftag = 0;
 865         cmnd->rw.apptag = 0;
 866         cmnd->rw.appmask = 0;
 867
 868         if (ns->ms) {
 869                 /*
 870                  * If formated with metadata, the block layer always provides a
 871                  * metadata buffer if CONFIG_BLK_DEV_INTEGRITY is enabled.  Else
 872                  * we enable the PRACT bit for protection information or set the
 873                  * namespace capacity to zero to prevent any I/O.
 874                  */
 875                 if (!blk_integrity_rq(req)) {
 876                         if (WARN_ON_ONCE(!nvme_ns_has_pi(ns)))
 877                                 return BLK_STS_NOTSUPP;
 878                         control |= NVME_RW_PRINFO_PRACT;
 879                 }
 880
 881                 switch (ns->pi_type) {
 882                 case NVME_NS_DPS_PI_TYPE3:
 883                         control |= NVME_RW_PRINFO_PRCHK_GUARD;
 884                         break;
 885                 case NVME_NS_DPS_PI_TYPE1:
 886                 case NVME_NS_DPS_PI_TYPE2:
 887                         control |= NVME_RW_PRINFO_PRCHK_GUARD |
 888                                         NVME_RW_PRINFO_PRCHK_REF;
 889                         if (op == nvme_cmd_zone_append)
 890                                 control |= NVME_RW_APPEND_PIREMAP;
 891                         nvme_set_ref_tag(ns, cmnd, req);
 892                         break;
 893                 }
 894         }
 895
 896         cmnd->rw.control = cpu_to_le16(control);
 897         cmnd->rw.dsmgmt = cpu_to_le32(dsmgmt);
 898         return 0;
 899 }
 900
 901 void nvme_cleanup_cmd(struct request *req)
 902 {
 903         if (req->rq_flags & RQF_SPECIAL_PAYLOAD) {
 904                 struct nvme_ctrl *ctrl = nvme_req(req)->ctrl;
 905
 906                 if (req->special_vec.bv_page == ctrl->discard_page)
 907                         clear_bit_unlock(0, &ctrl->discard_page_busy);
 908                 else
 909                         kfree(bvec_virt(&req->special_vec));
 910         }
 911 }
 912 EXPORT_SYMBOL_GPL(nvme_cleanup_cmd);
 913
 914 blk_status_t nvme_setup_cmd(struct nvme_ns *ns, struct request *req)
 915 {
 916         struct nvme_command *cmd = nvme_req(req)->cmd;
 917         blk_status_t ret = BLK_STS_OK;
 918
 919         if (!(req->rq_flags & RQF_DONTPREP))
 920                 nvme_clear_nvme_request(req);
 921
 922         switch (req_op(req)) {
 923         case REQ_OP_DRV_IN:
 924         case REQ_OP_DRV_OUT:
 925                 /* these are setup prior to execution in nvme_init_request() */
 926                 break;
 927         case REQ_OP_FLUSH:
 928                 nvme_setup_flush(ns, cmd);
 929                 break;
 930         case REQ_OP_ZONE_RESET_ALL:
 931         case REQ_OP_ZONE_RESET:
 932                 ret = nvme_setup_zone_mgmt_send(ns, req, cmd, NVME_ZONE_RESET);
 933                 break;
 934         case REQ_OP_ZONE_OPEN:
 935                 ret = nvme_setup_zone_mgmt_send(ns, req, cmd, NVME_ZONE_OPEN);
 936                 break;
 937         case REQ_OP_ZONE_CLOSE:
 938                 ret = nvme_setup_zone_mgmt_send(ns, req, cmd, NVME_ZONE_CLOSE);
 939                 break;
 940         case REQ_OP_ZONE_FINISH:
 941                 ret = nvme_setup_zone_mgmt_send(ns, req, cmd, NVME_ZONE_FINISH);
 942                 break;
 943         case REQ_OP_WRITE_ZEROES:
 944                 ret = nvme_setup_write_zeroes(ns, req, cmd);
 945                 break;
 946         case REQ_OP_DISCARD:
 947                 ret = nvme_setup_discard(ns, req, cmd);
 948                 break;
 949         case REQ_OP_READ:
 950                 ret = nvme_setup_rw(ns, req, cmd, nvme_cmd_read);
 951                 break;
 952         case REQ_OP_WRITE:
 953                 ret = nvme_setup_rw(ns, req, cmd, nvme_cmd_write);
 954                 break;
 955         case REQ_OP_ZONE_APPEND:
 956                 ret = nvme_setup_rw(ns, req, cmd, nvme_cmd_zone_append);
 957                 break;
 958         default:
 959                 WARN_ON_ONCE(1);
 960                 return BLK_STS_IOERR;
 961         }
 962
 963         cmd->common.command_id = nvme_cid(req);
 964         trace_nvme_setup_cmd(req, cmd);
 965         return ret;
 966 }
 967 EXPORT_SYMBOL_GPL(nvme_setup_cmd);
 968
 969 /*
 970  * Return values:
 971  * 0:  success
 972  * >0: nvme controller's cqe status response
 973  * <0: kernel error in lieu of controller response
 974  */
 975 static int nvme_execute_rq(struct request *rq, bool at_head)
 976 {
 977         blk_status_t status;
 978
 979         status = blk_execute_rq(rq, at_head);
 980         if (nvme_req(rq)->flags & NVME_REQ_CANCELLED)
 981                 return -EINTR;
 982         if (nvme_req(rq)->status)
 983                 return nvme_req(rq)->status;
 984         return blk_status_to_errno(status);
 985 }
 986
 987 /*
 988  * Returns 0 on success.  If the result is negative, it's a Linux error code;
 989  * if the result is positive, it's an NVM Express status code
 990  */
 991 int __nvme_submit_sync_cmd(struct request_queue *q, struct nvme_command *cmd,
 992                 union nvme_result *result, void *buffer, unsigned bufflen,
 993                 unsigned timeout, int qid, int at_head,
 994                 blk_mq_req_flags_t flags)
 995 {
 996         struct request *req;
 997         int ret;
 998
 999         if (qid == NVME_QID_ANY)
1000                 req = blk_mq_alloc_request(q, nvme_req_op(cmd), flags);
1001         else
1002                 req = blk_mq_alloc_request_hctx(q, nvme_req_op(cmd), flags,
1003                                                 qid ? qid - 1 : 0);
1004
1005         if (IS_ERR(req))
1006                 return PTR_ERR(req);
1007         nvme_init_request(req, cmd);
1008
1009         if (timeout)
1010                 req->timeout = timeout;
1011
1012         if (buffer && bufflen) {
1013                 ret = blk_rq_map_kern(q, req, buffer, bufflen, GFP_KERNEL);
1014                 if (ret)
1015                         goto out;
1016         }
1017
1018         req->rq_flags |= RQF_QUIET;
1019         ret = nvme_execute_rq(req, at_head);
1020         if (result && ret >= 0)
1021                 *result = nvme_req(req)->result;
1022  out:
1023         blk_mq_free_request(req);
1024         return ret;
1025 }
1026 EXPORT_SYMBOL_GPL(__nvme_submit_sync_cmd);
1027
1028 int nvme_submit_sync_cmd(struct request_queue *q, struct nvme_command *cmd,
1029                 void *buffer, unsigned bufflen)
1030 {
1031         return __nvme_submit_sync_cmd(q, cmd, NULL, buffer, bufflen, 0,
1032                         NVME_QID_ANY, 0, 0);
1033 }
1034 EXPORT_SYMBOL_GPL(nvme_submit_sync_cmd);
1035
1036 static u32 nvme_known_admin_effects(u8 opcode)
1037 {
1038         switch (opcode) {
1039         case nvme_admin_format_nvm:
1040                 return NVME_CMD_EFFECTS_LBCC | NVME_CMD_EFFECTS_NCC |
1041                         NVME_CMD_EFFECTS_CSE_MASK;
1042         case nvme_admin_sanitize_nvm:
1043                 return NVME_CMD_EFFECTS_LBCC | NVME_CMD_EFFECTS_CSE_MASK;
1044         default:
1045                 break;
1046         }
1047         return 0;
1048 }
1049
1050 u32 nvme_command_effects(struct nvme_ctrl *ctrl, struct nvme_ns *ns, u8 opcode)
1051 {
1052         u32 effects = 0;
1053
1054         if (ns) {
1055                 if (ns->head->effects)
1056                         effects = le32_to_cpu(ns->head->effects->iocs[opcode]);
1057                 if (effects & ~(NVME_CMD_EFFECTS_CSUPP | NVME_CMD_EFFECTS_LBCC))
1058                         dev_warn_once(ctrl->device,
1059                                 "IO command:%02x has unhandled effects:%08x\n",
1060                                 opcode, effects);
1061                 return 0;
1062         }
1063
1064         if (ctrl->effects)
1065                 effects = le32_to_cpu(ctrl->effects->acs[opcode]);
1066         effects |= nvme_known_admin_effects(opcode);
1067
1068         return effects;
1069 }
1070 EXPORT_SYMBOL_NS_GPL(nvme_command_effects, NVME_TARGET_PASSTHRU);
1071
1072 static u32 nvme_passthru_start(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
1073                                u8 opcode)
1074 {
1075         u32 effects = nvme_command_effects(ctrl, ns, opcode);
1076
1077         /*
1078          * For simplicity, IO to all namespaces is quiesced even if the command
1079          * effects say only one namespace is affected.
1080          */
1081         if (effects & NVME_CMD_EFFECTS_CSE_MASK) {
1082                 mutex_lock(&ctrl->scan_lock);
1083                 mutex_lock(&ctrl->subsys->lock);
1084                 nvme_mpath_start_freeze(ctrl->subsys);
1085                 nvme_mpath_wait_freeze(ctrl->subsys);
1086                 nvme_start_freeze(ctrl);
1087                 nvme_wait_freeze(ctrl);
1088         }
1089         return effects;
1090 }
1091
1092 static void nvme_passthru_end(struct nvme_ctrl *ctrl, u32 effects,
1093                               struct nvme_command *cmd, int status)
1094 {
1095         if (effects & NVME_CMD_EFFECTS_CSE_MASK) {
1096                 nvme_unfreeze(ctrl);
1097                 nvme_mpath_unfreeze(ctrl->subsys);
1098                 mutex_unlock(&ctrl->subsys->lock);
1099                 nvme_remove_invalid_namespaces(ctrl, NVME_NSID_ALL);
1100                 mutex_unlock(&ctrl->scan_lock);
1101         }
1102         if (effects & NVME_CMD_EFFECTS_CCC)
1103                 nvme_init_ctrl_finish(ctrl);
1104         if (effects & (NVME_CMD_EFFECTS_NIC | NVME_CMD_EFFECTS_NCC)) {
1105                 nvme_queue_scan(ctrl);
1106                 flush_work(&ctrl->scan_work);
1107         }
1108
1109         switch (cmd->common.opcode) {
1110         case nvme_admin_set_features:
1111                 switch (le32_to_cpu(cmd->common.cdw10) & 0xFF) {
1112                 case NVME_FEAT_KATO:
1113                         /*
1114                          * Keep alive commands interval on the host should be
1115                          * updated when KATO is modified by Set Features
1116                          * commands.
1117                          */
1118                         if (!status)
1119                                 nvme_update_keep_alive(ctrl, cmd);
1120                         break;
1121                 default:
1122                         break;
1123                 }
1124                 break;
1125         default:
1126                 break;
1127         }
1128 }
1129
1130 int nvme_execute_passthru_rq(struct request *rq)
1131 {
1132         struct nvme_command *cmd = nvme_req(rq)->cmd;
1133         struct nvme_ctrl *ctrl = nvme_req(rq)->ctrl;
1134         struct nvme_ns *ns = rq->q->queuedata;
1135         u32 effects;
1136         int  ret;
1137
1138         effects = nvme_passthru_start(ctrl, ns, cmd->common.opcode);
1139         ret = nvme_execute_rq(rq, false);
1140         if (effects) /* nothing to be done for zero cmd effects */
1141                 nvme_passthru_end(ctrl, effects, cmd, ret);
1142
1143         return ret;
1144 }
1145 EXPORT_SYMBOL_NS_GPL(nvme_execute_passthru_rq, NVME_TARGET_PASSTHRU);
1146
1147 /*
1148  * Recommended frequency for KATO commands per NVMe 1.4 section 7.12.1:
1149  *
1150  *   The host should send Keep Alive commands at half of the Keep Alive Timeout
1151  *   accounting for transport roundtrip times [..].
1152  */
1153 static void nvme_queue_keep_alive_work(struct nvme_ctrl *ctrl)
1154 {
1155         queue_delayed_work(nvme_wq, &ctrl->ka_work, ctrl->kato * HZ / 2);
1156 }
1157
1158 static void nvme_keep_alive_end_io(struct request *rq, blk_status_t status)
1159 {
1160         struct nvme_ctrl *ctrl = rq->end_io_data;
1161         unsigned long flags;
1162         bool startka = false;
1163
1164         blk_mq_free_request(rq);
1165
1166         if (status) {
1167                 dev_err(ctrl->device,
1168                         "failed nvme_keep_alive_end_io error=%d\n",
1169                                 status);
1170                 return;
1171         }
1172
1173         ctrl->comp_seen = false;
1174         spin_lock_irqsave(&ctrl->lock, flags);
1175         if (ctrl->state == NVME_CTRL_LIVE ||
1176             ctrl->state == NVME_CTRL_CONNECTING)
1177                 startka = true;
1178         spin_unlock_irqrestore(&ctrl->lock, flags);
1179         if (startka)
1180                 nvme_queue_keep_alive_work(ctrl);
1181 }
1182
1183 static void nvme_keep_alive_work(struct work_struct *work)
1184 {
1185         struct nvme_ctrl *ctrl = container_of(to_delayed_work(work),
1186                         struct nvme_ctrl, ka_work);
1187         bool comp_seen = ctrl->comp_seen;
1188         struct request *rq;
1189
1190         if ((ctrl->ctratt & NVME_CTRL_ATTR_TBKAS) && comp_seen) {
1191                 dev_dbg(ctrl->device,
1192                         "reschedule traffic based keep-alive timer\n");
1193                 ctrl->comp_seen = false;
1194                 nvme_queue_keep_alive_work(ctrl);
1195                 return;
1196         }
1197
1198         rq = blk_mq_alloc_request(ctrl->admin_q, nvme_req_op(&ctrl->ka_cmd),
1199                                   BLK_MQ_REQ_RESERVED | BLK_MQ_REQ_NOWAIT);
1200         if (IS_ERR(rq)) {
1201                 /* allocation failure, reset the controller */
1202                 dev_err(ctrl->device, "keep-alive failed: %ld\n", PTR_ERR(rq));
1203                 nvme_reset_ctrl(ctrl);
1204                 return;
1205         }
1206         nvme_init_request(rq, &ctrl->ka_cmd);
1207
1208         rq->timeout = ctrl->kato * HZ;
1209         rq->end_io = nvme_keep_alive_end_io;
1210         rq->end_io_data = ctrl;
1211         rq->rq_flags |= RQF_QUIET;
1212         blk_execute_rq_nowait(rq, false);
1213 }
1214
1215 static void nvme_start_keep_alive(struct nvme_ctrl *ctrl)
1216 {
1217         if (unlikely(ctrl->kato == 0))
1218                 return;
1219
1220         nvme_queue_keep_alive_work(ctrl);
1221 }
1222
1223 void nvme_stop_keep_alive(struct nvme_ctrl *ctrl)
1224 {
1225         if (unlikely(ctrl->kato == 0))
1226                 return;
1227
1228         cancel_delayed_work_sync(&ctrl->ka_work);
1229 }
1230 EXPORT_SYMBOL_GPL(nvme_stop_keep_alive);
1231
1232 static void nvme_update_keep_alive(struct nvme_ctrl *ctrl,
1233                                    struct nvme_command *cmd)
1234 {
1235         unsigned int new_kato =
1236                 DIV_ROUND_UP(le32_to_cpu(cmd->common.cdw11), 1000);
1237
1238         dev_info(ctrl->device,
1239                  "keep alive interval updated from %u ms to %u ms\n",
1240                  ctrl->kato * 1000 / 2, new_kato * 1000 / 2);
1241
1242         nvme_stop_keep_alive(ctrl);
1243         ctrl->kato = new_kato;
1244         nvme_start_keep_alive(ctrl);
1245 }
1246
1247 /*
1248  * In NVMe 1.0 the CNS field was just a binary controller or namespace
1249  * flag, thus sending any new CNS opcodes has a big chance of not working.
1250  * Qemu unfortunately had that bug after reporting a 1.1 version compliance
1251  * (but not for any later version).
1252  */
1253 static bool nvme_ctrl_limited_cns(struct nvme_ctrl *ctrl)
1254 {
1255         if (ctrl->quirks & NVME_QUIRK_IDENTIFY_CNS)
1256                 return ctrl->vs < NVME_VS(1, 2, 0);
1257         return ctrl->vs < NVME_VS(1, 1, 0);
1258 }
1259
1260 static int nvme_identify_ctrl(struct nvme_ctrl *dev, struct nvme_id_ctrl **id)
1261 {
1262         struct nvme_command c = { };
1263         int error;
1264
1265         /* gcc-4.4.4 (at least) has issues with initializers and anon unions */
1266         c.identify.opcode = nvme_admin_identify;
1267         c.identify.cns = NVME_ID_CNS_CTRL;
1268
1269         *id = kmalloc(sizeof(struct nvme_id_ctrl), GFP_KERNEL);
1270         if (!*id)
1271                 return -ENOMEM;
1272
1273         error = nvme_submit_sync_cmd(dev->admin_q, &c, *id,
1274                         sizeof(struct nvme_id_ctrl));
1275         if (error)
1276                 kfree(*id);
1277         return error;
1278 }
1279
1280 static int nvme_process_ns_desc(struct nvme_ctrl *ctrl, struct nvme_ns_ids *ids,
1281                 struct nvme_ns_id_desc *cur, bool *csi_seen)
1282 {
1283         const char *warn_str = "ctrl returned bogus length:";
1284         void *data = cur;
1285
1286         switch (cur->nidt) {
1287         case NVME_NIDT_EUI64:
1288                 if (cur->nidl != NVME_NIDT_EUI64_LEN) {
1289                         dev_warn(ctrl->device, "%s %d for NVME_NIDT_EUI64\n",
1290                                  warn_str, cur->nidl);
1291                         return -1;
1292                 }
1293                 if (ctrl->quirks & NVME_QUIRK_BOGUS_NID)
1294                         return NVME_NIDT_EUI64_LEN;
1295                 memcpy(ids->eui64, data + sizeof(*cur), NVME_NIDT_EUI64_LEN);
1296                 return NVME_NIDT_EUI64_LEN;
1297         case NVME_NIDT_NGUID:
1298                 if (cur->nidl != NVME_NIDT_NGUID_LEN) {
1299                         dev_warn(ctrl->device, "%s %d for NVME_NIDT_NGUID\n",
1300                                  warn_str, cur->nidl);
1301                         return -1;
1302                 }
1303                 if (ctrl->quirks & NVME_QUIRK_BOGUS_NID)
1304                         return NVME_NIDT_NGUID_LEN;
1305                 memcpy(ids->nguid, data + sizeof(*cur), NVME_NIDT_NGUID_LEN);
1306                 return NVME_NIDT_NGUID_LEN;
1307         case NVME_NIDT_UUID:
1308                 if (cur->nidl != NVME_NIDT_UUID_LEN) {
1309                         dev_warn(ctrl->device, "%s %d for NVME_NIDT_UUID\n",
1310                                  warn_str, cur->nidl);
1311                         return -1;
1312                 }
1313                 if (ctrl->quirks & NVME_QUIRK_BOGUS_NID)
1314                         return NVME_NIDT_UUID_LEN;
1315                 uuid_copy(&ids->uuid, data + sizeof(*cur));
1316                 return NVME_NIDT_UUID_LEN;
1317         case NVME_NIDT_CSI:
1318                 if (cur->nidl != NVME_NIDT_CSI_LEN) {
1319                         dev_warn(ctrl->device, "%s %d for NVME_NIDT_CSI\n",
1320                                  warn_str, cur->nidl);
1321                         return -1;
1322                 }
1323                 memcpy(&ids->csi, data + sizeof(*cur), NVME_NIDT_CSI_LEN);
1324                 *csi_seen = true;
1325                 return NVME_NIDT_CSI_LEN;
1326         default:
1327                 /* Skip unknown types */
1328                 return cur->nidl;
1329         }
1330 }
1331
1332 static int nvme_identify_ns_descs(struct nvme_ctrl *ctrl, unsigned nsid,
1333                 struct nvme_ns_ids *ids)
1334 {
1335         struct nvme_command c = { };
1336         bool csi_seen = false;
1337         int status, pos, len;
1338         void *data;
1339
1340         if (ctrl->vs < NVME_VS(1, 3, 0) && !nvme_multi_css(ctrl))
1341                 return 0;
1342         if (ctrl->quirks & NVME_QUIRK_NO_NS_DESC_LIST)
1343                 return 0;
1344
1345         c.identify.opcode = nvme_admin_identify;
1346         c.identify.nsid = cpu_to_le32(nsid);
1347         c.identify.cns = NVME_ID_CNS_NS_DESC_LIST;
1348
1349         data = kzalloc(NVME_IDENTIFY_DATA_SIZE, GFP_KERNEL);
1350         if (!data)
1351                 return -ENOMEM;
1352
1353         status = nvme_submit_sync_cmd(ctrl->admin_q, &c, data,
1354                                       NVME_IDENTIFY_DATA_SIZE);
1355         if (status) {
1356                 dev_warn(ctrl->device,
1357                         "Identify Descriptors failed (nsid=%u, status=0x%x)\n",
1358                         nsid, status);
1359                 goto free_data;
1360         }
1361
1362         for (pos = 0; pos < NVME_IDENTIFY_DATA_SIZE; pos += len) {
1363                 struct nvme_ns_id_desc *cur = data + pos;
1364
1365                 if (cur->nidl == 0)
1366                         break;
1367
1368                 len = nvme_process_ns_desc(ctrl, ids, cur, &csi_seen);
1369                 if (len < 0)
1370                         break;
1371
1372                 len += sizeof(*cur);
1373         }
1374
1375         if (nvme_multi_css(ctrl) && !csi_seen) {
1376                 dev_warn(ctrl->device, "Command set not reported for nsid:%d\n",
1377                          nsid);
1378                 status = -EINVAL;
1379         }
1380
1381 free_data:
1382         kfree(data);
1383         return status;
1384 }
1385
1386 static int nvme_identify_ns(struct nvme_ctrl *ctrl, unsigned nsid,
1387                         struct nvme_ns_ids *ids, struct nvme_id_ns **id)
1388 {
1389         struct nvme_command c = { };
1390         int error;
1391
1392         /* gcc-4.4.4 (at least) has issues with initializers and anon unions */
1393         c.identify.opcode = nvme_admin_identify;
1394         c.identify.nsid = cpu_to_le32(nsid);
1395         c.identify.cns = NVME_ID_CNS_NS;
1396
1397         *id = kmalloc(sizeof(**id), GFP_KERNEL);
1398         if (!*id)
1399                 return -ENOMEM;
1400
1401         error = nvme_submit_sync_cmd(ctrl->admin_q, &c, *id, sizeof(**id));
1402         if (error) {
1403                 dev_warn(ctrl->device, "Identify namespace failed (%d)\n", error);
1404                 goto out_free_id;
1405         }
1406
1407         error = NVME_SC_INVALID_NS | NVME_SC_DNR;
1408         if ((*id)->ncap == 0) /* namespace not allocated or attached */
1409                 goto out_free_id;
1410
1411
1412         if (ctrl->quirks & NVME_QUIRK_BOGUS_NID) {
1413                 dev_info(ctrl->device,
1414                          "Ignoring bogus Namespace Identifiers\n");
1415         } else {
1416                 if (ctrl->vs >= NVME_VS(1, 1, 0) &&
1417                     !memchr_inv(ids->eui64, 0, sizeof(ids->eui64)))
1418                         memcpy(ids->eui64, (*id)->eui64, sizeof(ids->eui64));
1419                 if (ctrl->vs >= NVME_VS(1, 2, 0) &&
1420                     !memchr_inv(ids->nguid, 0, sizeof(ids->nguid)))
1421                         memcpy(ids->nguid, (*id)->nguid, sizeof(ids->nguid));
1422         }
1423
1424         return 0;
1425
1426 out_free_id:
1427         kfree(*id);
1428         return error;
1429 }
1430
1431 static int nvme_identify_ns_cs_indep(struct nvme_ctrl *ctrl, unsigned nsid,
1432                         struct nvme_id_ns_cs_indep **id)
1433 {
1434         struct nvme_command c = {
1435                 .identify.opcode        = nvme_admin_identify,
1436                 .identify.nsid          = cpu_to_le32(nsid),
1437                 .identify.cns           = NVME_ID_CNS_NS_CS_INDEP,
1438         };
1439         int ret;
1440
1441         *id = kmalloc(sizeof(**id), GFP_KERNEL);
1442         if (!*id)
1443                 return -ENOMEM;
1444
1445         ret = nvme_submit_sync_cmd(ctrl->admin_q, &c, *id, sizeof(**id));
1446         if (ret) {
1447                 dev_warn(ctrl->device,
1448                          "Identify namespace (CS independent) failed (%d)\n",
1449                          ret);
1450                 kfree(*id);
1451                 return ret;
1452         }
1453
1454         return 0;
1455 }
1456
1457 static int nvme_features(struct nvme_ctrl *dev, u8 op, unsigned int fid,
1458                 unsigned int dword11, void *buffer, size_t buflen, u32 *result)
1459 {
1460         union nvme_result res = { 0 };
1461         struct nvme_command c = { };
1462         int ret;
1463
1464         c.features.opcode = op;
1465         c.features.fid = cpu_to_le32(fid);
1466         c.features.dword11 = cpu_to_le32(dword11);
1467
1468         ret = __nvme_submit_sync_cmd(dev->admin_q, &c, &res,
1469                         buffer, buflen, 0, NVME_QID_ANY, 0, 0);
1470         if (ret >= 0 && result)
1471                 *result = le32_to_cpu(res.u32);
1472         return ret;
1473 }
1474
1475 int nvme_set_features(struct nvme_ctrl *dev, unsigned int fid,
1476                       unsigned int dword11, void *buffer, size_t buflen,
1477                       u32 *result)
1478 {
1479         return nvme_features(dev, nvme_admin_set_features, fid, dword11, buffer,
1480                              buflen, result);
1481 }
1482 EXPORT_SYMBOL_GPL(nvme_set_features);
1483
1484 int nvme_get_features(struct nvme_ctrl *dev, unsigned int fid,
1485                       unsigned int dword11, void *buffer, size_t buflen,
1486                       u32 *result)
1487 {
1488         return nvme_features(dev, nvme_admin_get_features, fid, dword11, buffer,
1489                              buflen, result);
1490 }
1491 EXPORT_SYMBOL_GPL(nvme_get_features);
1492
1493 int nvme_set_queue_count(struct nvme_ctrl *ctrl, int *count)
1494 {
1495         u32 q_count = (*count - 1) | ((*count - 1) << 16);
1496         u32 result;
1497         int status, nr_io_queues;
1498
1499         status = nvme_set_features(ctrl, NVME_FEAT_NUM_QUEUES, q_count, NULL, 0,
1500                         &result);
1501         if (status < 0)
1502                 return status;
1503
1504         /*
1505          * Degraded controllers might return an error when setting the queue
1506          * count.  We still want to be able to bring them online and offer
1507          * access to the admin queue, as that might be only way to fix them up.
1508          */
1509         if (status > 0) {
1510                 dev_err(ctrl->device, "Could not set queue count (%d)\n", status);
1511                 *count = 0;
1512         } else {
1513                 nr_io_queues = min(result & 0xffff, result >> 16) + 1;
1514                 *count = min(*count, nr_io_queues);
1515         }
1516
1517         return 0;
1518 }
1519 EXPORT_SYMBOL_GPL(nvme_set_queue_count);
1520
1521 #define NVME_AEN_SUPPORTED \
1522         (NVME_AEN_CFG_NS_ATTR | NVME_AEN_CFG_FW_ACT | \
1523          NVME_AEN_CFG_ANA_CHANGE | NVME_AEN_CFG_DISC_CHANGE)
1524
1525 static void nvme_enable_aen(struct nvme_ctrl *ctrl)
1526 {
1527         u32 result, supported_aens = ctrl->oaes & NVME_AEN_SUPPORTED;
1528         int status;
1529
1530         if (!supported_aens)
1531                 return;
1532
1533         status = nvme_set_features(ctrl, NVME_FEAT_ASYNC_EVENT, supported_aens,
1534                         NULL, 0, &result);
1535         if (status)
1536                 dev_warn(ctrl->device, "Failed to configure AEN (cfg %x)\n",
1537                          supported_aens);
1538
1539         queue_work(nvme_wq, &ctrl->async_event_work);
1540 }
1541
1542 static int nvme_ns_open(struct nvme_ns *ns)
1543 {
1544
1545         /* should never be called due to GENHD_FL_HIDDEN */
1546         if (WARN_ON_ONCE(nvme_ns_head_multipath(ns->head)))
1547                 goto fail;
1548         if (!nvme_get_ns(ns))
1549                 goto fail;
1550         if (!try_module_get(ns->ctrl->ops->module))
1551                 goto fail_put_ns;
1552
1553         return 0;
1554
1555 fail_put_ns:
1556         nvme_put_ns(ns);
1557 fail:
1558         return -ENXIO;
1559 }
1560
1561 static void nvme_ns_release(struct nvme_ns *ns)
1562 {
1563
1564         module_put(ns->ctrl->ops->module);
1565         nvme_put_ns(ns);
1566 }
1567
1568 static int nvme_open(struct block_device *bdev, fmode_t mode)
1569 {
1570         return nvme_ns_open(bdev->bd_disk->private_data);
1571 }
1572
1573 static void nvme_release(struct gendisk *disk, fmode_t mode)
1574 {
1575         nvme_ns_release(disk->private_data);
1576 }
1577
1578 int nvme_getgeo(struct block_device *bdev, struct hd_geometry *geo)
1579 {
1580         /* some standard values */
1581         geo->heads = 1 << 6;
1582         geo->sectors = 1 << 5;
1583         geo->cylinders = get_capacity(bdev->bd_disk) >> 11;
1584         return 0;
1585 }
1586
1587 #ifdef CONFIG_BLK_DEV_INTEGRITY
1588 static void nvme_init_integrity(struct gendisk *disk, struct nvme_ns *ns,
1589                                 u32 max_integrity_segments)
1590 {
1591         struct blk_integrity integrity = { };
1592
1593         switch (ns->pi_type) {
1594         case NVME_NS_DPS_PI_TYPE3:
1595                 switch (ns->guard_type) {
1596                 case NVME_NVM_NS_16B_GUARD:
1597                         integrity.profile = &t10_pi_type3_crc;
1598                         integrity.tag_size = sizeof(u16) + sizeof(u32);
1599                         integrity.flags |= BLK_INTEGRITY_DEVICE_CAPABLE;
1600                         break;
1601                 case NVME_NVM_NS_64B_GUARD:
1602                         integrity.profile = &ext_pi_type3_crc64;
1603                         integrity.tag_size = sizeof(u16) + 6;
1604                         integrity.flags |= BLK_INTEGRITY_DEVICE_CAPABLE;
1605                         break;
1606                 default:
1607                         integrity.profile = NULL;
1608                         break;
1609                 }
1610                 break;
1611         case NVME_NS_DPS_PI_TYPE1:
1612         case NVME_NS_DPS_PI_TYPE2:
1613                 switch (ns->guard_type) {
1614                 case NVME_NVM_NS_16B_GUARD:
1615                         integrity.profile = &t10_pi_type1_crc;
1616                         integrity.tag_size = sizeof(u16);
1617                         integrity.flags |= BLK_INTEGRITY_DEVICE_CAPABLE;
1618                         break;
1619                 case NVME_NVM_NS_64B_GUARD:
1620                         integrity.profile = &ext_pi_type1_crc64;
1621                         integrity.tag_size = sizeof(u16);
1622                         integrity.flags |= BLK_INTEGRITY_DEVICE_CAPABLE;
1623                         break;
1624                 default:
1625                         integrity.profile = NULL;
1626                         break;
1627                 }
1628                 break;
1629         default:
1630                 integrity.profile = NULL;
1631                 break;
1632         }
1633
1634         integrity.tuple_size = ns->ms;
1635         blk_integrity_register(disk, &integrity);
1636         blk_queue_max_integrity_segments(disk->queue, max_integrity_segments);
1637 }
1638 #else
1639 static void nvme_init_integrity(struct gendisk *disk, struct nvme_ns *ns,
1640                                 u32 max_integrity_segments)
1641 {
1642 }
1643 #endif /* CONFIG_BLK_DEV_INTEGRITY */
1644
1645 static void nvme_config_discard(struct gendisk *disk, struct nvme_ns *ns)
1646 {
1647         struct nvme_ctrl *ctrl = ns->ctrl;
1648         struct request_queue *queue = disk->queue;
1649         u32 size = queue_logical_block_size(queue);
1650
1651         if (ctrl->max_discard_sectors == 0) {
1652                 blk_queue_max_discard_sectors(queue, 0);
1653                 return;
1654         }
1655
1656         BUILD_BUG_ON(PAGE_SIZE / sizeof(struct nvme_dsm_range) <
1657                         NVME_DSM_MAX_RANGES);
1658
1659         queue->limits.discard_granularity = size;
1660
1661         /* If discard is already enabled, don't reset queue limits */
1662         if (queue->limits.max_discard_sectors)
1663                 return;
1664
1665         if (ctrl->dmrsl && ctrl->dmrsl <= nvme_sect_to_lba(ns, UINT_MAX))
1666                 ctrl->max_discard_sectors = nvme_lba_to_sect(ns, ctrl->dmrsl);
1667
1668         blk_queue_max_discard_sectors(queue, ctrl->max_discard_sectors);
1669         blk_queue_max_discard_segments(queue, ctrl->max_discard_segments);
1670
1671         if (ctrl->quirks & NVME_QUIRK_DEALLOCATE_ZEROES)
1672                 blk_queue_max_write_zeroes_sectors(queue, UINT_MAX);
1673 }
1674
1675 static bool nvme_ns_ids_equal(struct nvme_ns_ids *a, struct nvme_ns_ids *b)
1676 {
1677         return uuid_equal(&a->uuid, &b->uuid) &&
1678                 memcmp(&a->nguid, &b->nguid, sizeof(a->nguid)) == 0 &&
1679                 memcmp(&a->eui64, &b->eui64, sizeof(a->eui64)) == 0 &&
1680                 a->csi == b->csi;
1681 }
1682
1683 static int nvme_init_ms(struct nvme_ns *ns, struct nvme_id_ns *id)
1684 {
1685         bool first = id->dps & NVME_NS_DPS_PI_FIRST;
1686         unsigned lbaf = nvme_lbaf_index(id->flbas);
1687         struct nvme_ctrl *ctrl = ns->ctrl;
1688         struct nvme_command c = { };
1689         struct nvme_id_ns_nvm *nvm;
1690         int ret = 0;
1691         u32 elbaf;
1692
1693         ns->pi_size = 0;
1694         ns->ms = le16_to_cpu(id->lbaf[lbaf].ms);
1695         if (!(ctrl->ctratt & NVME_CTRL_ATTR_ELBAS)) {
1696                 ns->pi_size = sizeof(struct t10_pi_tuple);
1697                 ns->guard_type = NVME_NVM_NS_16B_GUARD;
1698                 goto set_pi;
1699         }
1700
1701         nvm = kzalloc(sizeof(*nvm), GFP_KERNEL);
1702         if (!nvm)
1703                 return -ENOMEM;
1704
1705         c.identify.opcode = nvme_admin_identify;
1706         c.identify.nsid = cpu_to_le32(ns->head->ns_id);
1707         c.identify.cns = NVME_ID_CNS_CS_NS;
1708         c.identify.csi = NVME_CSI_NVM;
1709
1710         ret = nvme_submit_sync_cmd(ns->ctrl->admin_q, &c, nvm, sizeof(*nvm));
1711         if (ret)
1712                 goto free_data;
1713
1714         elbaf = le32_to_cpu(nvm->elbaf[lbaf]);
1715
1716         /* no support for storage tag formats right now */
1717         if (nvme_elbaf_sts(elbaf))
1718                 goto free_data;
1719
1720         ns->guard_type = nvme_elbaf_guard_type(elbaf);
1721         switch (ns->guard_type) {
1722         case NVME_NVM_NS_64B_GUARD:
1723                 ns->pi_size = sizeof(struct crc64_pi_tuple);
1724                 break;
1725         case NVME_NVM_NS_16B_GUARD:
1726                 ns->pi_size = sizeof(struct t10_pi_tuple);
1727                 break;
1728         default:
1729                 break;
1730         }
1731
1732 free_data:
1733         kfree(nvm);
1734 set_pi:
1735         if (ns->pi_size && (first || ns->ms == ns->pi_size))
1736                 ns->pi_type = id->dps & NVME_NS_DPS_PI_MASK;
1737         else
1738                 ns->pi_type = 0;
1739
1740         return ret;
1741 }
1742
1743 static void nvme_configure_metadata(struct nvme_ns *ns, struct nvme_id_ns *id)
1744 {
1745         struct nvme_ctrl *ctrl = ns->ctrl;
1746
1747         if (nvme_init_ms(ns, id))
1748                 return;
1749
1750         ns->features &= ~(NVME_NS_METADATA_SUPPORTED | NVME_NS_EXT_LBAS);
1751         if (!ns->ms || !(ctrl->ops->flags & NVME_F_METADATA_SUPPORTED))
1752                 return;
1753
1754         if (ctrl->ops->flags & NVME_F_FABRICS) {
1755                 /*
1756                  * The NVMe over Fabrics specification only supports metadata as
1757                  * part of the extended data LBA.  We rely on HCA/HBA support to
1758                  * remap the separate metadata buffer from the block layer.
1759                  */
1760                 if (WARN_ON_ONCE(!(id->flbas & NVME_NS_FLBAS_META_EXT)))
1761                         return;
1762
1763                 ns->features |= NVME_NS_EXT_LBAS;
1764
1765                 /*
1766                  * The current fabrics transport drivers support namespace
1767                  * metadata formats only if nvme_ns_has_pi() returns true.
1768                  * Suppress support for all other formats so the namespace will
1769                  * have a 0 capacity and not be usable through the block stack.
1770                  *
1771                  * Note, this check will need to be modified if any drivers
1772                  * gain the ability to use other metadata formats.
1773                  */
1774                 if (ctrl->max_integrity_segments && nvme_ns_has_pi(ns))
1775                         ns->features |= NVME_NS_METADATA_SUPPORTED;
1776         } else {
1777                 /*
1778                  * For PCIe controllers, we can't easily remap the separate
1779                  * metadata buffer from the block layer and thus require a
1780                  * separate metadata buffer for block layer metadata/PI support.
1781                  * We allow extended LBAs for the passthrough interface, though.
1782                  */
1783                 if (id->flbas & NVME_NS_FLBAS_META_EXT)
1784                         ns->features |= NVME_NS_EXT_LBAS;
1785                 else
1786                         ns->features |= NVME_NS_METADATA_SUPPORTED;
1787         }
1788 }
1789
1790 static void nvme_set_queue_limits(struct nvme_ctrl *ctrl,
1791                 struct request_queue *q)
1792 {
1793         bool vwc = ctrl->vwc & NVME_CTRL_VWC_PRESENT;
1794
1795         if (ctrl->max_hw_sectors) {
1796                 u32 max_segments =
1797                         (ctrl->max_hw_sectors / (NVME_CTRL_PAGE_SIZE >> 9)) + 1;
1798
1799                 max_segments = min_not_zero(max_segments, ctrl->max_segments);
1800                 blk_queue_max_hw_sectors(q, ctrl->max_hw_sectors);
1801                 blk_queue_max_segments(q, min_t(u32, max_segments, USHRT_MAX));
1802         }
1803         blk_queue_virt_boundary(q, NVME_CTRL_PAGE_SIZE - 1);
1804         blk_queue_dma_alignment(q, 3);
1805         blk_queue_write_cache(q, vwc, vwc);
1806 }
1807
1808 static void nvme_update_disk_info(struct gendisk *disk,
1809                 struct nvme_ns *ns, struct nvme_id_ns *id)
1810 {
1811         sector_t capacity = nvme_lba_to_sect(ns, le64_to_cpu(id->nsze));
1812         unsigned short bs = 1 << ns->lba_shift;
1813         u32 atomic_bs, phys_bs, io_opt = 0;
1814
1815         /*
1816          * The block layer can't support LBA sizes larger than the page size
1817          * yet, so catch this early and don't allow block I/O.
1818          */
1819         if (ns->lba_shift > PAGE_SHIFT) {
1820                 capacity = 0;
1821                 bs = (1 << 9);
1822         }
1823
1824         blk_integrity_unregister(disk);
1825
1826         atomic_bs = phys_bs = bs;
1827         if (id->nabo == 0) {
1828                 /*
1829                  * Bit 1 indicates whether NAWUPF is defined for this namespace
1830                  * and whether it should be used instead of AWUPF. If NAWUPF ==
1831                  * 0 then AWUPF must be used instead.
1832                  */
1833                 if (id->nsfeat & NVME_NS_FEAT_ATOMICS && id->nawupf)
1834                         atomic_bs = (1 + le16_to_cpu(id->nawupf)) * bs;
1835                 else
1836                         atomic_bs = (1 + ns->ctrl->subsys->awupf) * bs;
1837         }
1838
1839         if (id->nsfeat & NVME_NS_FEAT_IO_OPT) {
1840                 /* NPWG = Namespace Preferred Write Granularity */
1841                 phys_bs = bs * (1 + le16_to_cpu(id->npwg));
1842                 /* NOWS = Namespace Optimal Write Size */
1843                 io_opt = bs * (1 + le16_to_cpu(id->nows));
1844         }
1845
1846         blk_queue_logical_block_size(disk->queue, bs);
1847         /*
1848          * Linux filesystems assume writing a single physical block is
1849          * an atomic operation. Hence limit the physical block size to the
1850          * value of the Atomic Write Unit Power Fail parameter.
1851          */
1852         blk_queue_physical_block_size(disk->queue, min(phys_bs, atomic_bs));
1853         blk_queue_io_min(disk->queue, phys_bs);
1854         blk_queue_io_opt(disk->queue, io_opt);
1855
1856         /*
1857          * Register a metadata profile for PI, or the plain non-integrity NVMe
1858          * metadata masquerading as Type 0 if supported, otherwise reject block
1859          * I/O to namespaces with metadata except when the namespace supports
1860          * PI, as it can strip/insert in that case.
1861          */
1862         if (ns->ms) {
1863                 if (IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY) &&
1864                     (ns->features & NVME_NS_METADATA_SUPPORTED))
1865                         nvme_init_integrity(disk, ns,
1866                                             ns->ctrl->max_integrity_segments);
1867                 else if (!nvme_ns_has_pi(ns))
1868                         capacity = 0;
1869         }
1870
1871         set_capacity_and_notify(disk, capacity);
1872
1873         nvme_config_discard(disk, ns);
1874         blk_queue_max_write_zeroes_sectors(disk->queue,
1875                                            ns->ctrl->max_zeroes_sectors);
1876 }
1877
1878 static inline bool nvme_first_scan(struct gendisk *disk)
1879 {
1880         /* nvme_alloc_ns() scans the disk prior to adding it */
1881         return !disk_live(disk);
1882 }
1883
1884 static void nvme_set_chunk_sectors(struct nvme_ns *ns, struct nvme_id_ns *id)
1885 {
1886         struct nvme_ctrl *ctrl = ns->ctrl;
1887         u32 iob;
1888
1889         if ((ctrl->quirks & NVME_QUIRK_STRIPE_SIZE) &&
1890             is_power_of_2(ctrl->max_hw_sectors))
1891                 iob = ctrl->max_hw_sectors;
1892         else
1893                 iob = nvme_lba_to_sect(ns, le16_to_cpu(id->noiob));
1894
1895         if (!iob)
1896                 return;
1897
1898         if (!is_power_of_2(iob)) {
1899                 if (nvme_first_scan(ns->disk))
1900                         pr_warn("%s: ignoring unaligned IO boundary:%u\n",
1901                                 ns->disk->disk_name, iob);
1902                 return;
1903         }
1904
1905         if (blk_queue_is_zoned(ns->disk->queue)) {
1906                 if (nvme_first_scan(ns->disk))
1907                         pr_warn("%s: ignoring zoned namespace IO boundary\n",
1908                                 ns->disk->disk_name);
1909                 return;
1910         }
1911
1912         blk_queue_chunk_sectors(ns->queue, iob);
1913 }
1914
1915 static int nvme_update_ns_info(struct nvme_ns *ns, struct nvme_id_ns *id)
1916 {
1917         unsigned lbaf = nvme_lbaf_index(id->flbas);
1918         int ret;
1919
1920         blk_mq_freeze_queue(ns->disk->queue);
1921         ns->lba_shift = id->lbaf[lbaf].ds;
1922         nvme_set_queue_limits(ns->ctrl, ns->queue);
1923
1924         nvme_configure_metadata(ns, id);
1925         nvme_set_chunk_sectors(ns, id);
1926         nvme_update_disk_info(ns->disk, ns, id);
1927
1928         if (ns->head->ids.csi == NVME_CSI_ZNS) {
1929                 ret = nvme_update_zone_info(ns, lbaf);
1930                 if (ret)
1931                         goto out_unfreeze;
1932         }
1933
1934         set_disk_ro(ns->disk, (id->nsattr & NVME_NS_ATTR_RO) ||
1935                 test_bit(NVME_NS_FORCE_RO, &ns->flags));
1936         set_bit(NVME_NS_READY, &ns->flags);
1937         blk_mq_unfreeze_queue(ns->disk->queue);
1938
1939         if (blk_queue_is_zoned(ns->queue)) {
1940                 ret = nvme_revalidate_zones(ns);
1941                 if (ret && !nvme_first_scan(ns->disk))
1942                         return ret;
1943         }
1944
1945         if (nvme_ns_head_multipath(ns->head)) {
1946                 blk_mq_freeze_queue(ns->head->disk->queue);
1947                 nvme_update_disk_info(ns->head->disk, ns, id);
1948                 set_disk_ro(ns->head->disk,
1949                             (id->nsattr & NVME_NS_ATTR_RO) ||
1950                                     test_bit(NVME_NS_FORCE_RO, &ns->flags));
1951                 nvme_mpath_revalidate_paths(ns);
1952                 blk_stack_limits(&ns->head->disk->queue->limits,
1953                                  &ns->queue->limits, 0);
1954                 disk_update_readahead(ns->head->disk);
1955                 blk_mq_unfreeze_queue(ns->head->disk->queue);
1956         }
1957         return 0;
1958
1959 out_unfreeze:
1960         /*
1961          * If probing fails due an unsupported feature, hide the block device,
1962          * but still allow other access.
1963          */
1964         if (ret == -ENODEV) {
1965                 ns->disk->flags |= GENHD_FL_HIDDEN;
1966                 set_bit(NVME_NS_READY, &ns->flags);
1967                 ret = 0;
1968         }
1969         blk_mq_unfreeze_queue(ns->disk->queue);
1970         return ret;
1971 }
1972
1973 static char nvme_pr_type(enum pr_type type)
1974 {
1975         switch (type) {
1976         case PR_WRITE_EXCLUSIVE:
1977                 return 1;
1978         case PR_EXCLUSIVE_ACCESS:
1979                 return 2;
1980         case PR_WRITE_EXCLUSIVE_REG_ONLY:
1981                 return 3;
1982         case PR_EXCLUSIVE_ACCESS_REG_ONLY:
1983                 return 4;
1984         case PR_WRITE_EXCLUSIVE_ALL_REGS:
1985                 return 5;
1986         case PR_EXCLUSIVE_ACCESS_ALL_REGS:
1987                 return 6;
1988         default:
1989                 return 0;
1990         }
1991 }
1992
1993 static int nvme_send_ns_head_pr_command(struct block_device *bdev,
1994                 struct nvme_command *c, u8 data[16])
1995 {
1996         struct nvme_ns_head *head = bdev->bd_disk->private_data;
1997         int srcu_idx = srcu_read_lock(&head->srcu);
1998         struct nvme_ns *ns = nvme_find_path(head);
1999         int ret = -EWOULDBLOCK;
2000
2001         if (ns) {
2002                 c->common.nsid = cpu_to_le32(ns->head->ns_id);
2003                 ret = nvme_submit_sync_cmd(ns->queue, c, data, 16);
2004         }
2005         srcu_read_unlock(&head->srcu, srcu_idx);
2006         return ret;
2007 }
2008
2009 static int nvme_send_ns_pr_command(struct nvme_ns *ns, struct nvme_command *c,
2010                 u8 data[16])
2011 {
2012         c->common.nsid = cpu_to_le32(ns->head->ns_id);
2013         return nvme_submit_sync_cmd(ns->queue, c, data, 16);
2014 }
2015
2016 static int nvme_pr_command(struct block_device *bdev, u32 cdw10,
2017                                 u64 key, u64 sa_key, u8 op)
2018 {
2019         struct nvme_command c = { };
2020         u8 data[16] = { 0, };
2021
2022         put_unaligned_le64(key, &data[0]);
2023         put_unaligned_le64(sa_key, &data[8]);
2024
2025         c.common.opcode = op;
2026         c.common.cdw10 = cpu_to_le32(cdw10);
2027
2028         if (IS_ENABLED(CONFIG_NVME_MULTIPATH) &&
2029             bdev->bd_disk->fops == &nvme_ns_head_ops)
2030                 return nvme_send_ns_head_pr_command(bdev, &c, data);
2031         return nvme_send_ns_pr_command(bdev->bd_disk->private_data, &c, data);
2032 }
2033
2034 static int nvme_pr_register(struct block_device *bdev, u64 old,
2035                 u64 new, unsigned flags)
2036 {
2037         u32 cdw10;
2038
2039         if (flags & ~PR_FL_IGNORE_KEY)
2040                 return -EOPNOTSUPP;
2041
2042         cdw10 = old ? 2 : 0;
2043         cdw10 |= (flags & PR_FL_IGNORE_KEY) ? 1 << 3 : 0;
2044         cdw10 |= (1 << 30) | (1 << 31); /* PTPL=1 */
2045         return nvme_pr_command(bdev, cdw10, old, new, nvme_cmd_resv_register);
2046 }
2047
2048 static int nvme_pr_reserve(struct block_device *bdev, u64 key,
2049                 enum pr_type type, unsigned flags)
2050 {
2051         u32 cdw10;
2052
2053         if (flags & ~PR_FL_IGNORE_KEY)
2054                 return -EOPNOTSUPP;
2055
2056         cdw10 = nvme_pr_type(type) << 8;
2057         cdw10 |= ((flags & PR_FL_IGNORE_KEY) ? 1 << 3 : 0);
2058         return nvme_pr_command(bdev, cdw10, key, 0, nvme_cmd_resv_acquire);
2059 }
2060
2061 static int nvme_pr_preempt(struct block_device *bdev, u64 old, u64 new,
2062                 enum pr_type type, bool abort)
2063 {
2064         u32 cdw10 = nvme_pr_type(type) << 8 | (abort ? 2 : 1);
2065
2066         return nvme_pr_command(bdev, cdw10, old, new, nvme_cmd_resv_acquire);
2067 }
2068
2069 static int nvme_pr_clear(struct block_device *bdev, u64 key)
2070 {
2071         u32 cdw10 = 1 | (key ? 1 << 3 : 0);
2072
2073         return nvme_pr_command(bdev, cdw10, key, 0, nvme_cmd_resv_register);
2074 }
2075
2076 static int nvme_pr_release(struct block_device *bdev, u64 key, enum pr_type type)
2077 {
2078         u32 cdw10 = nvme_pr_type(type) << 8 | (key ? 1 << 3 : 0);
2079
2080         return nvme_pr_command(bdev, cdw10, key, 0, nvme_cmd_resv_release);
2081 }
2082
2083 const struct pr_ops nvme_pr_ops = {
2084         .pr_register    = nvme_pr_register,
2085         .pr_reserve     = nvme_pr_reserve,
2086         .pr_release     = nvme_pr_release,
2087         .pr_preempt     = nvme_pr_preempt,
2088         .pr_clear       = nvme_pr_clear,
2089 };
2090
2091 #ifdef CONFIG_BLK_SED_OPAL
2092 int nvme_sec_submit(void *data, u16 spsp, u8 secp, void *buffer, size_t len,
2093                 bool send)
2094 {
2095         struct nvme_ctrl *ctrl = data;
2096         struct nvme_command cmd = { };
2097
2098         if (send)
2099                 cmd.common.opcode = nvme_admin_security_send;
2100         else
2101                 cmd.common.opcode = nvme_admin_security_recv;
2102         cmd.common.nsid = 0;
2103         cmd.common.cdw10 = cpu_to_le32(((u32)secp) << 24 | ((u32)spsp) << 8);
2104         cmd.common.cdw11 = cpu_to_le32(len);
2105
2106         return __nvme_submit_sync_cmd(ctrl->admin_q, &cmd, NULL, buffer, len, 0,
2107                         NVME_QID_ANY, 1, 0);
2108 }
2109 EXPORT_SYMBOL_GPL(nvme_sec_submit);
2110 #endif /* CONFIG_BLK_SED_OPAL */
2111
2112 #ifdef CONFIG_BLK_DEV_ZONED
2113 static int nvme_report_zones(struct gendisk *disk, sector_t sector,
2114                 unsigned int nr_zones, report_zones_cb cb, void *data)
2115 {
2116         return nvme_ns_report_zones(disk->private_data, sector, nr_zones, cb,
2117                         data);
2118 }
2119 #else
2120 #define nvme_report_zones       NULL
2121 #endif /* CONFIG_BLK_DEV_ZONED */
2122
2123 static const struct block_device_operations nvme_bdev_ops = {
2124         .owner          = THIS_MODULE,
2125         .ioctl          = nvme_ioctl,
2126         .open           = nvme_open,
2127         .release        = nvme_release,
2128         .getgeo         = nvme_getgeo,
2129         .report_zones   = nvme_report_zones,
2130         .pr_ops         = &nvme_pr_ops,
2131 };
2132
2133 static int nvme_wait_ready(struct nvme_ctrl *ctrl, u32 timeout, bool enabled)
2134 {
2135         unsigned long timeout_jiffies = ((timeout + 1) * HZ / 2) + jiffies;
2136         u32 csts, bit = enabled ? NVME_CSTS_RDY : 0;
2137         int ret;
2138
2139         while ((ret = ctrl->ops->reg_read32(ctrl, NVME_REG_CSTS, &csts)) == 0) {
2140                 if (csts == ~0)
2141                         return -ENODEV;
2142                 if ((csts & NVME_CSTS_RDY) == bit)
2143                         break;
2144
2145                 usleep_range(1000, 2000);
2146                 if (fatal_signal_pending(current))
2147                         return -EINTR;
2148                 if (time_after(jiffies, timeout_jiffies)) {
2149                         dev_err(ctrl->device,
2150                                 "Device not ready; aborting %s, CSTS=0x%x\n",
2151                                 enabled ? "initialisation" : "reset", csts);
2152                         return -ENODEV;
2153                 }
2154         }
2155
2156         return ret;
2157 }
2158
2159 /*
2160  * If the device has been passed off to us in an enabled state, just clear
2161  * the enabled bit.  The spec says we should set the 'shutdown notification
2162  * bits', but doing so may cause the device to complete commands to the
2163  * admin queue ... and we don't know what memory that might be pointing at!
2164  */
2165 int nvme_disable_ctrl(struct nvme_ctrl *ctrl)
2166 {
2167         int ret;
2168
2169         ctrl->ctrl_config &= ~NVME_CC_SHN_MASK;
2170         ctrl->ctrl_config &= ~NVME_CC_ENABLE;
2171
2172         ret = ctrl->ops->reg_write32(ctrl, NVME_REG_CC, ctrl->ctrl_config);
2173         if (ret)
2174                 return ret;
2175
2176         if (ctrl->quirks & NVME_QUIRK_DELAY_BEFORE_CHK_RDY)
2177                 msleep(NVME_QUIRK_DELAY_AMOUNT);
2178
2179         return nvme_wait_ready(ctrl, NVME_CAP_TIMEOUT(ctrl->cap), false);
2180 }
2181 EXPORT_SYMBOL_GPL(nvme_disable_ctrl);
2182
2183 int nvme_enable_ctrl(struct nvme_ctrl *ctrl)
2184 {
2185         unsigned dev_page_min;
2186         u32 timeout;
2187         int ret;
2188
2189         ret = ctrl->ops->reg_read64(ctrl, NVME_REG_CAP, &ctrl->cap);
2190         if (ret) {
2191                 dev_err(ctrl->device, "Reading CAP failed (%d)\n", ret);
2192                 return ret;
2193         }
2194         dev_page_min = NVME_CAP_MPSMIN(ctrl->cap) + 12;
2195
2196         if (NVME_CTRL_PAGE_SHIFT < dev_page_min) {
2197                 dev_err(ctrl->device,
2198                         "Minimum device page size %u too large for host (%u)\n",
2199                         1 << dev_page_min, 1 << NVME_CTRL_PAGE_SHIFT);
2200                 return -ENODEV;
2201         }
2202
2203         if (NVME_CAP_CSS(ctrl->cap) & NVME_CAP_CSS_CSI)
2204                 ctrl->ctrl_config = NVME_CC_CSS_CSI;
2205         else
2206                 ctrl->ctrl_config = NVME_CC_CSS_NVM;
2207
2208         if (ctrl->cap & NVME_CAP_CRMS_CRWMS) {
2209                 u32 crto;
2210
2211                 ret = ctrl->ops->reg_read32(ctrl, NVME_REG_CRTO, &crto);
2212                 if (ret) {
2213                         dev_err(ctrl->device, "Reading CRTO failed (%d)\n",
2214                                 ret);
2215                         return ret;
2216                 }
2217
2218                 if (ctrl->cap & NVME_CAP_CRMS_CRIMS) {
2219                         ctrl->ctrl_config |= NVME_CC_CRIME;
2220                         timeout = NVME_CRTO_CRIMT(crto);
2221                 } else {
2222                         timeout = NVME_CRTO_CRWMT(crto);
2223                 }
2224         } else {
2225                 timeout = NVME_CAP_TIMEOUT(ctrl->cap);
2226         }
2227
2228         ctrl->ctrl_config |= (NVME_CTRL_PAGE_SHIFT - 12) << NVME_CC_MPS_SHIFT;
2229         ctrl->ctrl_config |= NVME_CC_AMS_RR | NVME_CC_SHN_NONE;
2230         ctrl->ctrl_config |= NVME_CC_IOSQES | NVME_CC_IOCQES;
2231         ret = ctrl->ops->reg_write32(ctrl, NVME_REG_CC, ctrl->ctrl_config);
2232         if (ret)
2233                 return ret;
2234
2235         /* Flush write to device (required if transport is PCI) */
2236         ret = ctrl->ops->reg_read32(ctrl, NVME_REG_CC, &ctrl->ctrl_config);
2237         if (ret)
2238                 return ret;
2239
2240         ctrl->ctrl_config |= NVME_CC_ENABLE;
2241         ret = ctrl->ops->reg_write32(ctrl, NVME_REG_CC, ctrl->ctrl_config);
2242         if (ret)
2243                 return ret;
2244         return nvme_wait_ready(ctrl, timeout, true);
2245 }
2246 EXPORT_SYMBOL_GPL(nvme_enable_ctrl);
2247
2248 int nvme_shutdown_ctrl(struct nvme_ctrl *ctrl)
2249 {
2250         unsigned long timeout = jiffies + (ctrl->shutdown_timeout * HZ);
2251         u32 csts;
2252         int ret;
2253
2254         ctrl->ctrl_config &= ~NVME_CC_SHN_MASK;
2255         ctrl->ctrl_config |= NVME_CC_SHN_NORMAL;
2256
2257         ret = ctrl->ops->reg_write32(ctrl, NVME_REG_CC, ctrl->ctrl_config);
2258         if (ret)
2259                 return ret;
2260
2261         while ((ret = ctrl->ops->reg_read32(ctrl, NVME_REG_CSTS, &csts)) == 0) {
2262                 if ((csts & NVME_CSTS_SHST_MASK) == NVME_CSTS_SHST_CMPLT)
2263                         break;
2264
2265                 msleep(100);
2266                 if (fatal_signal_pending(current))
2267                         return -EINTR;
2268                 if (time_after(jiffies, timeout)) {
2269                         dev_err(ctrl->device,
2270                                 "Device shutdown incomplete; abort shutdown\n");
2271                         return -ENODEV;
2272                 }
2273         }
2274
2275         return ret;
2276 }
2277 EXPORT_SYMBOL_GPL(nvme_shutdown_ctrl);
2278
2279 static int nvme_configure_timestamp(struct nvme_ctrl *ctrl)
2280 {
2281         __le64 ts;
2282         int ret;
2283
2284         if (!(ctrl->oncs & NVME_CTRL_ONCS_TIMESTAMP))
2285                 return 0;
2286
2287         ts = cpu_to_le64(ktime_to_ms(ktime_get_real()));
2288         ret = nvme_set_features(ctrl, NVME_FEAT_TIMESTAMP, 0, &ts, sizeof(ts),
2289                         NULL);
2290         if (ret)
2291                 dev_warn_once(ctrl->device,
2292                         "could not set timestamp (%d)\n", ret);
2293         return ret;
2294 }
2295
2296 static int nvme_configure_host_options(struct nvme_ctrl *ctrl)
2297 {
2298         struct nvme_feat_host_behavior *host;
2299         u8 acre = 0, lbafee = 0;
2300         int ret;
2301
2302         /* Don't bother enabling the feature if retry delay is not reported */
2303         if (ctrl->crdt[0])
2304                 acre = NVME_ENABLE_ACRE;
2305         if (ctrl->ctratt & NVME_CTRL_ATTR_ELBAS)
2306                 lbafee = NVME_ENABLE_LBAFEE;
2307
2308         if (!acre && !lbafee)
2309                 return 0;
2310
2311         host = kzalloc(sizeof(*host), GFP_KERNEL);
2312         if (!host)
2313                 return 0;
2314
2315         host->acre = acre;
2316         host->lbafee = lbafee;
2317         ret = nvme_set_features(ctrl, NVME_FEAT_HOST_BEHAVIOR, 0,
2318                                 host, sizeof(*host), NULL);
2319         kfree(host);
2320         return ret;
2321 }
2322
2323 /*
2324  * The function checks whether the given total (exlat + enlat) latency of
2325  * a power state allows the latter to be used as an APST transition target.
2326  * It does so by comparing the latency to the primary and secondary latency
2327  * tolerances defined by module params. If there's a match, the corresponding
2328  * timeout value is returned and the matching tolerance index (1 or 2) is
2329  * reported.
2330  */
2331 static bool nvme_apst_get_transition_time(u64 total_latency,
2332                 u64 *transition_time, unsigned *last_index)
2333 {
2334         if (total_latency <= apst_primary_latency_tol_us) {
2335                 if (*last_index == 1)
2336                         return false;
2337                 *last_index = 1;
2338                 *transition_time = apst_primary_timeout_ms;
2339                 return true;
2340         }
2341         if (apst_secondary_timeout_ms &&
2342                 total_latency <= apst_secondary_latency_tol_us) {
2343                 if (*last_index <= 2)
2344                         return false;
2345                 *last_index = 2;
2346                 *transition_time = apst_secondary_timeout_ms;
2347                 return true;
2348         }
2349         return false;
2350 }
2351
2352 /*
2353  * APST (Autonomous Power State Transition) lets us program a table of power
2354  * state transitions that the controller will perform automatically.
2355  *
2356  * Depending on module params, one of the two supported techniques will be used:
2357  *
2358  * - If the parameters provide explicit timeouts and tolerances, they will be
2359  *   used to build a table with up to 2 non-operational states to transition to.
2360  *   The default parameter values were selected based on the values used by
2361  *   Microsoft's and Intel's NVMe drivers. Yet, since we don't implement dynamic
2362  *   regeneration of the APST table in the event of switching between external
2363  *   and battery power, the timeouts and tolerances reflect a compromise
2364  *   between values used by Microsoft for AC and battery scenarios.
2365  * - If not, we'll configure the table with a simple heuristic: we are willing
2366  *   to spend at most 2% of the time transitioning between power states.
2367  *   Therefore, when running in any given state, we will enter the next
2368  *   lower-power non-operational state after waiting 50 * (enlat + exlat)
2369  *   microseconds, as long as that state's exit latency is under the requested
2370  *   maximum latency.
2371  *
2372  * We will not autonomously enter any non-operational state for which the total
2373  * latency exceeds ps_max_latency_us.
2374  *
2375  * Users can set ps_max_latency_us to zero to turn off APST.
2376  */
2377 static int nvme_configure_apst(struct nvme_ctrl *ctrl)
2378 {
2379         struct nvme_feat_auto_pst *table;
2380         unsigned apste = 0;
2381         u64 max_lat_us = 0;
2382         __le64 target = 0;
2383         int max_ps = -1;
2384         int state;
2385         int ret;
2386         unsigned last_lt_index = UINT_MAX;
2387
2388         /*
2389          * If APST isn't supported or if we haven't been initialized yet,
2390          * then don't do anything.
2391          */
2392         if (!ctrl->apsta)
2393                 return 0;
2394
2395         if (ctrl->npss > 31) {
2396                 dev_warn(ctrl->device, "NPSS is invalid; not using APST\n");
2397                 return 0;
2398         }
2399
2400         table = kzalloc(sizeof(*table), GFP_KERNEL);
2401         if (!table)
2402                 return 0;
2403
2404         if (!ctrl->apst_enabled || ctrl->ps_max_latency_us == 0) {
2405                 /* Turn off APST. */
2406                 dev_dbg(ctrl->device, "APST disabled\n");
2407                 goto done;
2408         }
2409
2410         /*
2411          * Walk through all states from lowest- to highest-power.
2412          * According to the spec, lower-numbered states use more power.  NPSS,
2413          * despite the name, is the index of the lowest-power state, not the
2414          * number of states.
2415          */
2416         for (state = (int)ctrl->npss; state >= 0; state--) {
2417                 u64 total_latency_us, exit_latency_us, transition_ms;
2418
2419                 if (target)
2420                         table->entries[state] = target;
2421
2422                 /*
2423                  * Don't allow transitions to the deepest state if it's quirked
2424                  * off.
2425                  */
2426                 if (state == ctrl->npss &&
2427                     (ctrl->quirks & NVME_QUIRK_NO_DEEPEST_PS))
2428                         continue;
2429
2430                 /*
2431                  * Is this state a useful non-operational state for higher-power
2432                  * states to autonomously transition to?
2433                  */
2434                 if (!(ctrl->psd[state].flags & NVME_PS_FLAGS_NON_OP_STATE))
2435                         continue;
2436
2437                 exit_latency_us = (u64)le32_to_cpu(ctrl->psd[state].exit_lat);
2438                 if (exit_latency_us > ctrl->ps_max_latency_us)
2439                         continue;
2440
2441                 total_latency_us = exit_latency_us +
2442                         le32_to_cpu(ctrl->psd[state].entry_lat);
2443
2444                 /*
2445                  * This state is good. It can be used as the APST idle target
2446                  * for higher power states.
2447                  */
2448                 if (apst_primary_timeout_ms && apst_primary_latency_tol_us) {
2449                         if (!nvme_apst_get_transition_time(total_latency_us,
2450                                         &transition_ms, &last_lt_index))
2451                                 continue;
2452                 } else {
2453                         transition_ms = total_latency_us + 19;
2454                         do_div(transition_ms, 20);
2455                         if (transition_ms > (1 << 24) - 1)
2456                                 transition_ms = (1 << 24) - 1;
2457                 }
2458
2459                 target = cpu_to_le64((state << 3) | (transition_ms << 8));
2460                 if (max_ps == -1)
2461                         max_ps = state;
2462                 if (total_latency_us > max_lat_us)
2463                         max_lat_us = total_latency_us;
2464         }
2465
2466         if (max_ps == -1)
2467                 dev_dbg(ctrl->device, "APST enabled but no non-operational states are available\n");
2468         else
2469                 dev_dbg(ctrl->device, "APST enabled: max PS = %d, max round-trip latency = %lluus, table = %*phN\n",
2470                         max_ps, max_lat_us, (int)sizeof(*table), table);
2471         apste = 1;
2472
2473 done:
2474         ret = nvme_set_features(ctrl, NVME_FEAT_AUTO_PST, apste,
2475                                 table, sizeof(*table), NULL);
2476         if (ret)
2477                 dev_err(ctrl->device, "failed to set APST feature (%d)\n", ret);
2478         kfree(table);
2479         return ret;
2480 }
2481
2482 static void nvme_set_latency_tolerance(struct device *dev, s32 val)
2483 {
2484         struct nvme_ctrl *ctrl = dev_get_drvdata(dev);
2485         u64 latency;
2486
2487         switch (val) {
2488         case PM_QOS_LATENCY_TOLERANCE_NO_CONSTRAINT:
2489         case PM_QOS_LATENCY_ANY:
2490                 latency = U64_MAX;
2491                 break;
2492
2493         default:
2494                 latency = val;
2495         }
2496
2497         if (ctrl->ps_max_latency_us != latency) {
2498                 ctrl->ps_max_latency_us = latency;
2499                 if (ctrl->state == NVME_CTRL_LIVE)
2500                         nvme_configure_apst(ctrl);
2501         }
2502 }
2503
2504 struct nvme_core_quirk_entry {
2505         /*
2506          * NVMe model and firmware strings are padded with spaces.  For
2507          * simplicity, strings in the quirk table are padded with NULLs
2508          * instead.
2509          */
2510         u16 vid;
2511         const char *mn;
2512         const char *fr;
2513         unsigned long quirks;
2514 };
2515
2516 static const struct nvme_core_quirk_entry core_quirks[] = {
2517         {
2518                 /*
2519                  * This Toshiba device seems to die using any APST states.  See:
2520                  * https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1678184/comments/11
2521                  */
2522                 .vid = 0x1179,
2523                 .mn = "THNSF5256GPUK TOSHIBA",
2524                 .quirks = NVME_QUIRK_NO_APST,
2525         },
2526         {
2527                 /*
2528                  * This LiteON CL1-3D*-Q11 firmware version has a race
2529                  * condition associated with actions related to suspend to idle
2530                  * LiteON has resolved the problem in future firmware
2531                  */
2532                 .vid = 0x14a4,
2533                 .fr = "22301111",
2534                 .quirks = NVME_QUIRK_SIMPLE_SUSPEND,
2535         },
2536         {
2537                 /*
2538                  * This Kioxia CD6-V Series / HPE PE8030 device times out and
2539                  * aborts I/O during any load, but more easily reproducible
2540                  * with discards (fstrim).
2541                  *
2542                  * The device is left in a state where it is also not possible
2543                  * to use "nvme set-feature" to disable APST, but booting with
2544                  * nvme_core.default_ps_max_latency=0 works.
2545                  */
2546                 .vid = 0x1e0f,
2547                 .mn = "KCD6XVUL6T40",
2548                 .quirks = NVME_QUIRK_NO_APST,
2549         },
2550         {
2551                 /*
2552                  * The external Samsung X5 SSD fails initialization without a
2553                  * delay before checking if it is ready and has a whole set of
2554                  * other problems.  To make this even more interesting, it
2555                  * shares the PCI ID with internal Samsung 970 Evo Plus that
2556                  * does not need or want these quirks.
2557                  */
2558                 .vid = 0x144d,
2559                 .mn = "Samsung Portable SSD X5",
2560                 .quirks = NVME_QUIRK_DELAY_BEFORE_CHK_RDY |
2561                           NVME_QUIRK_NO_DEEPEST_PS |
2562                           NVME_QUIRK_IGNORE_DEV_SUBNQN,
2563         }
2564 };
2565
2566 /* match is null-terminated but idstr is space-padded. */
2567 static bool string_matches(const char *idstr, const char *match, size_t len)
2568 {
2569         size_t matchlen;
2570
2571         if (!match)
2572                 return true;
2573
2574         matchlen = strlen(match);
2575         WARN_ON_ONCE(matchlen > len);
2576
2577         if (memcmp(idstr, match, matchlen))
2578                 return false;
2579
2580         for (; matchlen < len; matchlen++)
2581                 if (idstr[matchlen] != ' ')
2582                         return false;
2583
2584         return true;
2585 }
2586
2587 static bool quirk_matches(const struct nvme_id_ctrl *id,
2588                           const struct nvme_core_quirk_entry *q)
2589 {
2590         return q->vid == le16_to_cpu(id->vid) &&
2591                 string_matches(id->mn, q->mn, sizeof(id->mn)) &&
2592                 string_matches(id->fr, q->fr, sizeof(id->fr));
2593 }
2594
2595 static void nvme_init_subnqn(struct nvme_subsystem *subsys, struct nvme_ctrl *ctrl,
2596                 struct nvme_id_ctrl *id)
2597 {
2598         size_t nqnlen;
2599         int off;
2600
2601         if(!(ctrl->quirks & NVME_QUIRK_IGNORE_DEV_SUBNQN)) {
2602                 nqnlen = strnlen(id->subnqn, NVMF_NQN_SIZE);
2603                 if (nqnlen > 0 && nqnlen < NVMF_NQN_SIZE) {
2604                         strlcpy(subsys->subnqn, id->subnqn, NVMF_NQN_SIZE);
2605                         return;
2606                 }
2607
2608                 if (ctrl->vs >= NVME_VS(1, 2, 1))
2609                         dev_warn(ctrl->device, "missing or invalid SUBNQN field.\n");
2610         }
2611
2612         /* Generate a "fake" NQN per Figure 254 in NVMe 1.3 + ECN 001 */
2613         off = snprintf(subsys->subnqn, NVMF_NQN_SIZE,
2614                         "nqn.2014.08.org.nvmexpress:%04x%04x",
2615                         le16_to_cpu(id->vid), le16_to_cpu(id->ssvid));
2616         memcpy(subsys->subnqn + off, id->sn, sizeof(id->sn));
2617         off += sizeof(id->sn);
2618         memcpy(subsys->subnqn + off, id->mn, sizeof(id->mn));
2619         off += sizeof(id->mn);
2620         memset(subsys->subnqn + off, 0, sizeof(subsys->subnqn) - off);
2621 }
2622
2623 static void nvme_release_subsystem(struct device *dev)
2624 {
2625         struct nvme_subsystem *subsys =
2626                 container_of(dev, struct nvme_subsystem, dev);
2627
2628         if (subsys->instance >= 0)
2629                 ida_free(&nvme_instance_ida, subsys->instance);
2630         kfree(subsys);
2631 }
2632
2633 static void nvme_destroy_subsystem(struct kref *ref)
2634 {
2635         struct nvme_subsystem *subsys =
2636                         container_of(ref, struct nvme_subsystem, ref);
2637
2638         mutex_lock(&nvme_subsystems_lock);
2639         list_del(&subsys->entry);
2640         mutex_unlock(&nvme_subsystems_lock);
2641
2642         ida_destroy(&subsys->ns_ida);
2643         device_del(&subsys->dev);
2644         put_device(&subsys->dev);
2645 }
2646
2647 static void nvme_put_subsystem(struct nvme_subsystem *subsys)
2648 {
2649         kref_put(&subsys->ref, nvme_destroy_subsystem);
2650 }
2651
2652 static struct nvme_subsystem *__nvme_find_get_subsystem(const char *subsysnqn)
2653 {
2654         struct nvme_subsystem *subsys;
2655
2656         lockdep_assert_held(&nvme_subsystems_lock);
2657
2658         /*
2659          * Fail matches for discovery subsystems. This results
2660          * in each discovery controller bound to a unique subsystem.
2661          * This avoids issues with validating controller values
2662          * that can only be true when there is a single unique subsystem.
2663          * There may be multiple and completely independent entities
2664          * that provide discovery controllers.
2665          */
2666         if (!strcmp(subsysnqn, NVME_DISC_SUBSYS_NAME))
2667                 return NULL;
2668
2669         list_for_each_entry(subsys, &nvme_subsystems, entry) {
2670                 if (strcmp(subsys->subnqn, subsysnqn))
2671                         continue;
2672                 if (!kref_get_unless_zero(&subsys->ref))
2673                         continue;
2674                 return subsys;
2675         }
2676
2677         return NULL;
2678 }
2679
2680 #define SUBSYS_ATTR_RO(_name, _mode, _show)                     \
2681         struct device_attribute subsys_attr_##_name = \
2682                 __ATTR(_name, _mode, _show, NULL)
2683
2684 static ssize_t nvme_subsys_show_nqn(struct device *dev,
2685                                     struct device_attribute *attr,
2686                                     char *buf)
2687 {
2688         struct nvme_subsystem *subsys =
2689                 container_of(dev, struct nvme_subsystem, dev);
2690
2691         return sysfs_emit(buf, "%s\n", subsys->subnqn);
2692 }
2693 static SUBSYS_ATTR_RO(subsysnqn, S_IRUGO, nvme_subsys_show_nqn);
2694
2695 static ssize_t nvme_subsys_show_type(struct device *dev,
2696                                     struct device_attribute *attr,
2697                                     char *buf)
2698 {
2699         struct nvme_subsystem *subsys =
2700                 container_of(dev, struct nvme_subsystem, dev);
2701
2702         switch (subsys->subtype) {
2703         case NVME_NQN_DISC:
2704                 return sysfs_emit(buf, "discovery\n");
2705         case NVME_NQN_NVME:
2706                 return sysfs_emit(buf, "nvm\n");
2707         default:
2708                 return sysfs_emit(buf, "reserved\n");
2709         }
2710 }
2711 static SUBSYS_ATTR_RO(subsystype, S_IRUGO, nvme_subsys_show_type);
2712
2713 #define nvme_subsys_show_str_function(field)                            \
2714 static ssize_t subsys_##field##_show(struct device *dev,                \
2715                             struct device_attribute *attr, char *buf)   \
2716 {                                                                       \
2717         struct nvme_subsystem *subsys =                                 \
2718                 container_of(dev, struct nvme_subsystem, dev);          \
2719         return sysfs_emit(buf, "%.*s\n",                                \
2720                            (int)sizeof(subsys->field), subsys->field);  \
2721 }                                                                       \
2722 static SUBSYS_ATTR_RO(field, S_IRUGO, subsys_##field##_show);
2723
2724 nvme_subsys_show_str_function(model);
2725 nvme_subsys_show_str_function(serial);
2726 nvme_subsys_show_str_function(firmware_rev);
2727
2728 static struct attribute *nvme_subsys_attrs[] = {
2729         &subsys_attr_model.attr,
2730         &subsys_attr_serial.attr,
2731         &subsys_attr_firmware_rev.attr,
2732         &subsys_attr_subsysnqn.attr,
2733         &subsys_attr_subsystype.attr,
2734 #ifdef CONFIG_NVME_MULTIPATH
2735         &subsys_attr_iopolicy.attr,
2736 #endif
2737         NULL,
2738 };
2739
2740 static const struct attribute_group nvme_subsys_attrs_group = {
2741         .attrs = nvme_subsys_attrs,
2742 };
2743
2744 static const struct attribute_group *nvme_subsys_attrs_groups[] = {
2745         &nvme_subsys_attrs_group,
2746         NULL,
2747 };
2748
2749 static inline bool nvme_discovery_ctrl(struct nvme_ctrl *ctrl)
2750 {
2751         return ctrl->opts && ctrl->opts->discovery_nqn;
2752 }
2753
2754 static bool nvme_validate_cntlid(struct nvme_subsystem *subsys,
2755                 struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id)
2756 {
2757         struct nvme_ctrl *tmp;
2758
2759         lockdep_assert_held(&nvme_subsystems_lock);
2760
2761         list_for_each_entry(tmp, &subsys->ctrls, subsys_entry) {
2762                 if (nvme_state_terminal(tmp))
2763                         continue;
2764
2765                 if (tmp->cntlid == ctrl->cntlid) {
2766                         dev_err(ctrl->device,
2767                                 "Duplicate cntlid %u with %s, subsys %s, rejecting\n",
2768                                 ctrl->cntlid, dev_name(tmp->device),
2769                                 subsys->subnqn);
2770                         return false;
2771                 }
2772
2773                 if ((id->cmic & NVME_CTRL_CMIC_MULTI_CTRL) ||
2774                     nvme_discovery_ctrl(ctrl))
2775                         continue;
2776
2777                 dev_err(ctrl->device,
2778                         "Subsystem does not support multiple controllers\n");
2779                 return false;
2780         }
2781
2782         return true;
2783 }
2784
2785 static int nvme_init_subsystem(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id)
2786 {
2787         struct nvme_subsystem *subsys, *found;
2788         int ret;
2789
2790         subsys = kzalloc(sizeof(*subsys), GFP_KERNEL);
2791         if (!subsys)
2792                 return -ENOMEM;
2793
2794         subsys->instance = -1;
2795         mutex_init(&subsys->lock);
2796         kref_init(&subsys->ref);
2797         INIT_LIST_HEAD(&subsys->ctrls);
2798         INIT_LIST_HEAD(&subsys->nsheads);
2799         nvme_init_subnqn(subsys, ctrl, id);
2800         memcpy(subsys->serial, id->sn, sizeof(subsys->serial));
2801         memcpy(subsys->model, id->mn, sizeof(subsys->model));
2802         memcpy(subsys->firmware_rev, id->fr, sizeof(subsys->firmware_rev));
2803         subsys->vendor_id = le16_to_cpu(id->vid);
2804         subsys->cmic = id->cmic;
2805
2806         /* Versions prior to 1.4 don't necessarily report a valid type */
2807         if (id->cntrltype == NVME_CTRL_DISC ||
2808             !strcmp(subsys->subnqn, NVME_DISC_SUBSYS_NAME))
2809                 subsys->subtype = NVME_NQN_DISC;
2810         else
2811                 subsys->subtype = NVME_NQN_NVME;
2812
2813         if (nvme_discovery_ctrl(ctrl) && subsys->subtype != NVME_NQN_DISC) {
2814                 dev_err(ctrl->device,
2815                         "Subsystem %s is not a discovery controller",
2816                         subsys->subnqn);
2817                 kfree(subsys);
2818                 return -EINVAL;
2819         }
2820         subsys->awupf = le16_to_cpu(id->awupf);
2821         nvme_mpath_default_iopolicy(subsys);
2822
2823         subsys->dev.class = nvme_subsys_class;
2824         subsys->dev.release = nvme_release_subsystem;
2825         subsys->dev.groups = nvme_subsys_attrs_groups;
2826         dev_set_name(&subsys->dev, "nvme-subsys%d", ctrl->instance);
2827         device_initialize(&subsys->dev);
2828
2829         mutex_lock(&nvme_subsystems_lock);
2830         found = __nvme_find_get_subsystem(subsys->subnqn);
2831         if (found) {
2832                 put_device(&subsys->dev);
2833                 subsys = found;
2834
2835                 if (!nvme_validate_cntlid(subsys, ctrl, id)) {
2836                         ret = -EINVAL;
2837                         goto out_put_subsystem;
2838                 }
2839         } else {
2840                 ret = device_add(&subsys->dev);
2841                 if (ret) {
2842                         dev_err(ctrl->device,
2843                                 "failed to register subsystem device.\n");
2844                         put_device(&subsys->dev);
2845                         goto out_unlock;
2846                 }
2847                 ida_init(&subsys->ns_ida);
2848                 list_add_tail(&subsys->entry, &nvme_subsystems);
2849         }
2850
2851         ret = sysfs_create_link(&subsys->dev.kobj, &ctrl->device->kobj,
2852                                 dev_name(ctrl->device));
2853         if (ret) {
2854                 dev_err(ctrl->device,
2855                         "failed to create sysfs link from subsystem.\n");
2856                 goto out_put_subsystem;
2857         }
2858
2859         if (!found)
2860                 subsys->instance = ctrl->instance;
2861         ctrl->subsys = subsys;
2862         list_add_tail(&ctrl->subsys_entry, &subsys->ctrls);
2863         mutex_unlock(&nvme_subsystems_lock);
2864         return 0;
2865
2866 out_put_subsystem:
2867         nvme_put_subsystem(subsys);
2868 out_unlock:
2869         mutex_unlock(&nvme_subsystems_lock);
2870         return ret;
2871 }
2872
2873 int nvme_get_log(struct nvme_ctrl *ctrl, u32 nsid, u8 log_page, u8 lsp, u8 csi,
2874                 void *log, size_t size, u64 offset)
2875 {
2876         struct nvme_command c = { };
2877         u32 dwlen = nvme_bytes_to_numd(size);
2878
2879         c.get_log_page.opcode = nvme_admin_get_log_page;
2880         c.get_log_page.nsid = cpu_to_le32(nsid);
2881         c.get_log_page.lid = log_page;
2882         c.get_log_page.lsp = lsp;
2883         c.get_log_page.numdl = cpu_to_le16(dwlen & ((1 << 16) - 1));
2884         c.get_log_page.numdu = cpu_to_le16(dwlen >> 16);
2885         c.get_log_page.lpol = cpu_to_le32(lower_32_bits(offset));
2886         c.get_log_page.lpou = cpu_to_le32(upper_32_bits(offset));
2887         c.get_log_page.csi = csi;
2888
2889         return nvme_submit_sync_cmd(ctrl->admin_q, &c, log, size);
2890 }
2891
2892 static int nvme_get_effects_log(struct nvme_ctrl *ctrl, u8 csi,
2893                                 struct nvme_effects_log **log)
2894 {
2895         struct nvme_effects_log *cel = xa_load(&ctrl->cels, csi);
2896         int ret;
2897
2898         if (cel)
2899                 goto out;
2900
2901         cel = kzalloc(sizeof(*cel), GFP_KERNEL);
2902         if (!cel)
2903                 return -ENOMEM;
2904
2905         ret = nvme_get_log(ctrl, 0x00, NVME_LOG_CMD_EFFECTS, 0, csi,
2906                         cel, sizeof(*cel), 0);
2907         if (ret) {
2908                 kfree(cel);
2909                 return ret;
2910         }
2911
2912         xa_store(&ctrl->cels, csi, cel, GFP_KERNEL);
2913 out:
2914         *log = cel;
2915         return 0;
2916 }
2917
2918 static inline u32 nvme_mps_to_sectors(struct nvme_ctrl *ctrl, u32 units)
2919 {
2920         u32 page_shift = NVME_CAP_MPSMIN(ctrl->cap) + 12, val;
2921
2922         if (check_shl_overflow(1U, units + page_shift - 9, &val))
2923                 return UINT_MAX;
2924         return val;
2925 }
2926
2927 static int nvme_init_non_mdts_limits(struct nvme_ctrl *ctrl)
2928 {
2929         struct nvme_command c = { };
2930         struct nvme_id_ctrl_nvm *id;
2931         int ret;
2932
2933         if (ctrl->oncs & NVME_CTRL_ONCS_DSM) {
2934                 ctrl->max_discard_sectors = UINT_MAX;
2935                 ctrl->max_discard_segments = NVME_DSM_MAX_RANGES;
2936         } else {
2937                 ctrl->max_discard_sectors = 0;
2938                 ctrl->max_discard_segments = 0;
2939         }
2940
2941         /*
2942          * Even though NVMe spec explicitly states that MDTS is not applicable
2943          * to the write-zeroes, we are cautious and limit the size to the
2944          * controllers max_hw_sectors value, which is based on the MDTS field
2945          * and possibly other limiting factors.
2946          */
2947         if ((ctrl->oncs & NVME_CTRL_ONCS_WRITE_ZEROES) &&
2948             !(ctrl->quirks & NVME_QUIRK_DISABLE_WRITE_ZEROES))
2949                 ctrl->max_zeroes_sectors = ctrl->max_hw_sectors;
2950         else
2951                 ctrl->max_zeroes_sectors = 0;
2952
2953         if (nvme_ctrl_limited_cns(ctrl))
2954                 return 0;
2955
2956         id = kzalloc(sizeof(*id), GFP_KERNEL);
2957         if (!id)
2958                 return 0;
2959
2960         c.identify.opcode = nvme_admin_identify;
2961         c.identify.cns = NVME_ID_CNS_CS_CTRL;
2962         c.identify.csi = NVME_CSI_NVM;
2963
2964         ret = nvme_submit_sync_cmd(ctrl->admin_q, &c, id, sizeof(*id));
2965         if (ret)
2966                 goto free_data;
2967
2968         if (id->dmrl)
2969                 ctrl->max_discard_segments = id->dmrl;
2970         ctrl->dmrsl = le32_to_cpu(id->dmrsl);
2971         if (id->wzsl)
2972                 ctrl->max_zeroes_sectors = nvme_mps_to_sectors(ctrl, id->wzsl);
2973
2974 free_data:
2975         kfree(id);
2976         return ret;
2977 }
2978
2979 static int nvme_init_identify(struct nvme_ctrl *ctrl)
2980 {
2981         struct nvme_id_ctrl *id;
2982         u32 max_hw_sectors;
2983         bool prev_apst_enabled;
2984         int ret;
2985
2986         ret = nvme_identify_ctrl(ctrl, &id);
2987         if (ret) {
2988                 dev_err(ctrl->device, "Identify Controller failed (%d)\n", ret);
2989                 return -EIO;
2990         }
2991
2992         if (id->lpa & NVME_CTRL_LPA_CMD_EFFECTS_LOG) {
2993                 ret = nvme_get_effects_log(ctrl, NVME_CSI_NVM, &ctrl->effects);
2994                 if (ret < 0)
2995                         goto out_free;
2996         }
2997
2998         if (!(ctrl->ops->flags & NVME_F_FABRICS))
2999                 ctrl->cntlid = le16_to_cpu(id->cntlid);
3000
3001         if (!ctrl->identified) {
3002                 unsigned int i;
3003
3004                 ret = nvme_init_subsystem(ctrl, id);
3005                 if (ret)
3006                         goto out_free;
3007
3008                 /*
3009                  * Check for quirks.  Quirk can depend on firmware version,
3010                  * so, in principle, the set of quirks present can change
3011                  * across a reset.  As a possible future enhancement, we
3012                  * could re-scan for quirks every time we reinitialize
3013                  * the device, but we'd have to make sure that the driver
3014                  * behaves intelligently if the quirks change.
3015                  */
3016                 for (i = 0; i < ARRAY_SIZE(core_quirks); i++) {
3017                         if (quirk_matches(id, &core_quirks[i]))
3018                                 ctrl->quirks |= core_quirks[i].quirks;
3019                 }
3020         }
3021
3022         if (force_apst && (ctrl->quirks & NVME_QUIRK_NO_DEEPEST_PS)) {
3023                 dev_warn(ctrl->device, "forcibly allowing all power states due to nvme_core.force_apst -- use at your own risk\n");
3024                 ctrl->quirks &= ~NVME_QUIRK_NO_DEEPEST_PS;
3025         }
3026
3027         ctrl->crdt[0] = le16_to_cpu(id->crdt1);
3028         ctrl->crdt[1] = le16_to_cpu(id->crdt2);
3029         ctrl->crdt[2] = le16_to_cpu(id->crdt3);
3030
3031         ctrl->oacs = le16_to_cpu(id->oacs);
3032         ctrl->oncs = le16_to_cpu(id->oncs);
3033         ctrl->mtfa = le16_to_cpu(id->mtfa);
3034         ctrl->oaes = le32_to_cpu(id->oaes);
3035         ctrl->wctemp = le16_to_cpu(id->wctemp);
3036         ctrl->cctemp = le16_to_cpu(id->cctemp);
3037
3038         atomic_set(&ctrl->abort_limit, id->acl + 1);
3039         ctrl->vwc = id->vwc;
3040         if (id->mdts)
3041                 max_hw_sectors = nvme_mps_to_sectors(ctrl, id->mdts);
3042         else
3043                 max_hw_sectors = UINT_MAX;
3044         ctrl->max_hw_sectors =
3045                 min_not_zero(ctrl->max_hw_sectors, max_hw_sectors);
3046
3047         nvme_set_queue_limits(ctrl, ctrl->admin_q);
3048         ctrl->sgls = le32_to_cpu(id->sgls);
3049         ctrl->kas = le16_to_cpu(id->kas);
3050         ctrl->max_namespaces = le32_to_cpu(id->mnan);
3051         ctrl->ctratt = le32_to_cpu(id->ctratt);
3052
3053         ctrl->cntrltype = id->cntrltype;
3054         ctrl->dctype = id->dctype;
3055
3056         if (id->rtd3e) {
3057                 /* us -> s */
3058                 u32 transition_time = le32_to_cpu(id->rtd3e) / USEC_PER_SEC;
3059
3060                 ctrl->shutdown_timeout = clamp_t(unsigned int, transition_time,
3061                                                  shutdown_timeout, 60);
3062
3063                 if (ctrl->shutdown_timeout != shutdown_timeout)
3064                         dev_info(ctrl->device,
3065                                  "Shutdown timeout set to %u seconds\n",
3066                                  ctrl->shutdown_timeout);
3067         } else
3068                 ctrl->shutdown_timeout = shutdown_timeout;
3069
3070         ctrl->npss = id->npss;
3071         ctrl->apsta = id->apsta;
3072         prev_apst_enabled = ctrl->apst_enabled;
3073         if (ctrl->quirks & NVME_QUIRK_NO_APST) {
3074                 if (force_apst && id->apsta) {
3075                         dev_warn(ctrl->device, "forcibly allowing APST due to nvme_core.force_apst -- use at your own risk\n");
3076                         ctrl->apst_enabled = true;
3077                 } else {
3078                         ctrl->apst_enabled = false;
3079                 }
3080         } else {
3081                 ctrl->apst_enabled = id->apsta;
3082         }
3083         memcpy(ctrl->psd, id->psd, sizeof(ctrl->psd));
3084
3085         if (ctrl->ops->flags & NVME_F_FABRICS) {
3086                 ctrl->icdoff = le16_to_cpu(id->icdoff);
3087                 ctrl->ioccsz = le32_to_cpu(id->ioccsz);
3088                 ctrl->iorcsz = le32_to_cpu(id->iorcsz);
3089                 ctrl->maxcmd = le16_to_cpu(id->maxcmd);
3090
3091                 /*
3092                  * In fabrics we need to verify the cntlid matches the
3093                  * admin connect
3094                  */
3095                 if (ctrl->cntlid != le16_to_cpu(id->cntlid)) {
3096                         dev_err(ctrl->device,
3097                                 "Mismatching cntlid: Connect %u vs Identify "
3098                                 "%u, rejecting\n",
3099                                 ctrl->cntlid, le16_to_cpu(id->cntlid));
3100                         ret = -EINVAL;
3101                         goto out_free;
3102                 }
3103
3104                 if (!nvme_discovery_ctrl(ctrl) && !ctrl->kas) {
3105                         dev_err(ctrl->device,
3106                                 "keep-alive support is mandatory for fabrics\n");
3107                         ret = -EINVAL;
3108                         goto out_free;
3109                 }
3110         } else {
3111                 ctrl->hmpre = le32_to_cpu(id->hmpre);
3112                 ctrl->hmmin = le32_to_cpu(id->hmmin);
3113                 ctrl->hmminds = le32_to_cpu(id->hmminds);
3114                 ctrl->hmmaxd = le16_to_cpu(id->hmmaxd);
3115         }
3116
3117         ret = nvme_mpath_init_identify(ctrl, id);
3118         if (ret < 0)
3119                 goto out_free;
3120
3121         if (ctrl->apst_enabled && !prev_apst_enabled)
3122                 dev_pm_qos_expose_latency_tolerance(ctrl->device);
3123         else if (!ctrl->apst_enabled && prev_apst_enabled)
3124                 dev_pm_qos_hide_latency_tolerance(ctrl->device);
3125
3126 out_free:
3127         kfree(id);
3128         return ret;
3129 }
3130
3131 /*
3132  * Initialize the cached copies of the Identify data and various controller
3133  * register in our nvme_ctrl structure.  This should be called as soon as
3134  * the admin queue is fully up and running.
3135  */
3136 int nvme_init_ctrl_finish(struct nvme_ctrl *ctrl)
3137 {
3138         int ret;
3139
3140         ret = ctrl->ops->reg_read32(ctrl, NVME_REG_VS, &ctrl->vs);
3141         if (ret) {
3142                 dev_err(ctrl->device, "Reading VS failed (%d)\n", ret);
3143                 return ret;
3144         }
3145
3146         ctrl->sqsize = min_t(u16, NVME_CAP_MQES(ctrl->cap), ctrl->sqsize);
3147
3148         if (ctrl->vs >= NVME_VS(1, 1, 0))
3149                 ctrl->subsystem = NVME_CAP_NSSRC(ctrl->cap);
3150
3151         ret = nvme_init_identify(ctrl);
3152         if (ret)
3153                 return ret;
3154
3155         ret = nvme_configure_apst(ctrl);
3156         if (ret < 0)
3157                 return ret;
3158
3159         ret = nvme_configure_timestamp(ctrl);
3160         if (ret < 0)
3161                 return ret;
3162
3163         ret = nvme_configure_host_options(ctrl);
3164         if (ret < 0)
3165                 return ret;
3166
3167         if (!ctrl->identified && !nvme_discovery_ctrl(ctrl)) {
3168                 ret = nvme_hwmon_init(ctrl);
3169                 if (ret < 0)
3170                         return ret;
3171         }
3172
3173         ctrl->identified = true;
3174
3175         return 0;
3176 }
3177 EXPORT_SYMBOL_GPL(nvme_init_ctrl_finish);
3178
3179 static int nvme_dev_open(struct inode *inode, struct file *file)
3180 {
3181         struct nvme_ctrl *ctrl =
3182                 container_of(inode->i_cdev, struct nvme_ctrl, cdev);
3183
3184         switch (ctrl->state) {
3185         case NVME_CTRL_LIVE:
3186                 break;
3187         default:
3188                 return -EWOULDBLOCK;
3189         }
3190
3191         nvme_get_ctrl(ctrl);
3192         if (!try_module_get(ctrl->ops->module)) {
3193                 nvme_put_ctrl(ctrl);
3194                 return -EINVAL;
3195         }
3196
3197         file->private_data = ctrl;
3198         return 0;
3199 }
3200
3201 static int nvme_dev_release(struct inode *inode, struct file *file)
3202 {
3203         struct nvme_ctrl *ctrl =
3204                 container_of(inode->i_cdev, struct nvme_ctrl, cdev);
3205
3206         module_put(ctrl->ops->module);
3207         nvme_put_ctrl(ctrl);
3208         return 0;
3209 }
3210
3211 static const struct file_operations nvme_dev_fops = {
3212         .owner          = THIS_MODULE,
3213         .open           = nvme_dev_open,
3214         .release        = nvme_dev_release,
3215         .unlocked_ioctl = nvme_dev_ioctl,
3216         .compat_ioctl   = compat_ptr_ioctl,
3217         .uring_cmd      = nvme_dev_uring_cmd,
3218 };
3219
3220 static ssize_t nvme_sysfs_reset(struct device *dev,
3221                                 struct device_attribute *attr, const char *buf,
3222                                 size_t count)
3223 {
3224         struct nvme_ctrl *ctrl = dev_get_drvdata(dev);
3225         int ret;
3226
3227         ret = nvme_reset_ctrl_sync(ctrl);
3228         if (ret < 0)
3229                 return ret;
3230         return count;
3231 }
3232 static DEVICE_ATTR(reset_controller, S_IWUSR, NULL, nvme_sysfs_reset);
3233
3234 static ssize_t nvme_sysfs_rescan(struct device *dev,
3235                                 struct device_attribute *attr, const char *buf,
3236                                 size_t count)
3237 {
3238         struct nvme_ctrl *ctrl = dev_get_drvdata(dev);
3239
3240         nvme_queue_scan(ctrl);
3241         return count;
3242 }
3243 static DEVICE_ATTR(rescan_controller, S_IWUSR, NULL, nvme_sysfs_rescan);
3244
3245 static inline struct nvme_ns_head *dev_to_ns_head(struct device *dev)
3246 {
3247         struct gendisk *disk = dev_to_disk(dev);
3248
3249         if (disk->fops == &nvme_bdev_ops)
3250                 return nvme_get_ns_from_dev(dev)->head;
3251         else
3252                 return disk->private_data;
3253 }
3254
3255 static ssize_t wwid_show(struct device *dev, struct device_attribute *attr,
3256                 char *buf)
3257 {
3258         struct nvme_ns_head *head = dev_to_ns_head(dev);
3259         struct nvme_ns_ids *ids = &head->ids;
3260         struct nvme_subsystem *subsys = head->subsys;
3261         int serial_len = sizeof(subsys->serial);
3262         int model_len = sizeof(subsys->model);
3263
3264         if (!uuid_is_null(&ids->uuid))
3265                 return sysfs_emit(buf, "uuid.%pU\n", &ids->uuid);
3266
3267         if (memchr_inv(ids->nguid, 0, sizeof(ids->nguid)))
3268                 return sysfs_emit(buf, "eui.%16phN\n", ids->nguid);
3269
3270         if (memchr_inv(ids->eui64, 0, sizeof(ids->eui64)))
3271                 return sysfs_emit(buf, "eui.%8phN\n", ids->eui64);
3272
3273         while (serial_len > 0 && (subsys->serial[serial_len - 1] == ' ' ||
3274                                   subsys->serial[serial_len - 1] == '\0'))
3275                 serial_len--;
3276         while (model_len > 0 && (subsys->model[model_len - 1] == ' ' ||
3277                                  subsys->model[model_len - 1] == '\0'))
3278                 model_len--;
3279
3280         return sysfs_emit(buf, "nvme.%04x-%*phN-%*phN-%08x\n", subsys->vendor_id,
3281                 serial_len, subsys->serial, model_len, subsys->model,
3282                 head->ns_id);
3283 }
3284 static DEVICE_ATTR_RO(wwid);
3285
3286 static ssize_t nguid_show(struct device *dev, struct device_attribute *attr,
3287                 char *buf)
3288 {
3289         return sysfs_emit(buf, "%pU\n", dev_to_ns_head(dev)->ids.nguid);
3290 }
3291 static DEVICE_ATTR_RO(nguid);
3292
3293 static ssize_t uuid_show(struct device *dev, struct device_attribute *attr,
3294                 char *buf)
3295 {
3296         struct nvme_ns_ids *ids = &dev_to_ns_head(dev)->ids;
3297
3298         /* For backward compatibility expose the NGUID to userspace if
3299          * we have no UUID set
3300          */
3301         if (uuid_is_null(&ids->uuid)) {
3302                 dev_warn_ratelimited(dev,
3303                         "No UUID available providing old NGUID\n");
3304                 return sysfs_emit(buf, "%pU\n", ids->nguid);
3305         }
3306         return sysfs_emit(buf, "%pU\n", &ids->uuid);
3307 }
3308 static DEVICE_ATTR_RO(uuid);
3309
3310 static ssize_t eui_show(struct device *dev, struct device_attribute *attr,
3311                 char *buf)
3312 {
3313         return sysfs_emit(buf, "%8ph\n", dev_to_ns_head(dev)->ids.eui64);
3314 }
3315 static DEVICE_ATTR_RO(eui);
3316
3317 static ssize_t nsid_show(struct device *dev, struct device_attribute *attr,
3318                 char *buf)
3319 {
3320         return sysfs_emit(buf, "%d\n", dev_to_ns_head(dev)->ns_id);
3321 }
3322 static DEVICE_ATTR_RO(nsid);
3323
3324 static struct attribute *nvme_ns_id_attrs[] = {
3325         &dev_attr_wwid.attr,
3326         &dev_attr_uuid.attr,
3327         &dev_attr_nguid.attr,
3328         &dev_attr_eui.attr,
3329         &dev_attr_nsid.attr,
3330 #ifdef CONFIG_NVME_MULTIPATH
3331         &dev_attr_ana_grpid.attr,
3332         &dev_attr_ana_state.attr,
3333 #endif
3334         NULL,
3335 };
3336
3337 static umode_t nvme_ns_id_attrs_are_visible(struct kobject *kobj,
3338                 struct attribute *a, int n)
3339 {
3340         struct device *dev = container_of(kobj, struct device, kobj);
3341         struct nvme_ns_ids *ids = &dev_to_ns_head(dev)->ids;
3342
3343         if (a == &dev_attr_uuid.attr) {
3344                 if (uuid_is_null(&ids->uuid) &&
3345                     !memchr_inv(ids->nguid, 0, sizeof(ids->nguid)))
3346                         return 0;
3347         }
3348         if (a == &dev_attr_nguid.attr) {
3349                 if (!memchr_inv(ids->nguid, 0, sizeof(ids->nguid)))
3350                         return 0;
3351         }
3352         if (a == &dev_attr_eui.attr) {
3353                 if (!memchr_inv(ids->eui64, 0, sizeof(ids->eui64)))
3354                         return 0;
3355         }
3356 #ifdef CONFIG_NVME_MULTIPATH
3357         if (a == &dev_attr_ana_grpid.attr || a == &dev_attr_ana_state.attr) {
3358                 if (dev_to_disk(dev)->fops != &nvme_bdev_ops) /* per-path attr */
3359                         return 0;
3360                 if (!nvme_ctrl_use_ana(nvme_get_ns_from_dev(dev)->ctrl))
3361                         return 0;
3362         }
3363 #endif
3364         return a->mode;
3365 }
3366
3367 static const struct attribute_group nvme_ns_id_attr_group = {
3368         .attrs          = nvme_ns_id_attrs,
3369         .is_visible     = nvme_ns_id_attrs_are_visible,
3370 };
3371
3372 const struct attribute_group *nvme_ns_id_attr_groups[] = {
3373         &nvme_ns_id_attr_group,
3374         NULL,
3375 };
3376
3377 #define nvme_show_str_function(field)                                           \
3378 static ssize_t  field##_show(struct device *dev,                                \
3379                             struct device_attribute *attr, char *buf)           \
3380 {                                                                               \
3381         struct nvme_ctrl *ctrl = dev_get_drvdata(dev);                          \
3382         return sysfs_emit(buf, "%.*s\n",                                        \
3383                 (int)sizeof(ctrl->subsys->field), ctrl->subsys->field);         \
3384 }                                                                               \
3385 static DEVICE_ATTR(field, S_IRUGO, field##_show, NULL);
3386
3387 nvme_show_str_function(model);
3388 nvme_show_str_function(serial);
3389 nvme_show_str_function(firmware_rev);
3390
3391 #define nvme_show_int_function(field)                                           \
3392 static ssize_t  field##_show(struct device *dev,                                \
3393                             struct device_attribute *attr, char *buf)           \
3394 {                                                                               \
3395         struct nvme_ctrl *ctrl = dev_get_drvdata(dev);                          \
3396         return sysfs_emit(buf, "%d\n", ctrl->field);                            \
3397 }                                                                               \
3398 static DEVICE_ATTR(field, S_IRUGO, field##_show, NULL);
3399
3400 nvme_show_int_function(cntlid);
3401 nvme_show_int_function(numa_node);
3402 nvme_show_int_function(queue_count);
3403 nvme_show_int_function(sqsize);
3404 nvme_show_int_function(kato);
3405
3406 static ssize_t nvme_sysfs_delete(struct device *dev,
3407                                 struct device_attribute *attr, const char *buf,
3408                                 size_t count)
3409 {
3410         struct nvme_ctrl *ctrl = dev_get_drvdata(dev);
3411
3412         if (device_remove_file_self(dev, attr))
3413                 nvme_delete_ctrl_sync(ctrl);
3414         return count;
3415 }
3416 static DEVICE_ATTR(delete_controller, S_IWUSR, NULL, nvme_sysfs_delete);
3417
3418 static ssize_t nvme_sysfs_show_transport(struct device *dev,
3419                                          struct device_attribute *attr,
3420                                          char *buf)
3421 {
3422         struct nvme_ctrl *ctrl = dev_get_drvdata(dev);
3423
3424         return sysfs_emit(buf, "%s\n", ctrl->ops->name);
3425 }
3426 static DEVICE_ATTR(transport, S_IRUGO, nvme_sysfs_show_transport, NULL);
3427
3428 static ssize_t nvme_sysfs_show_state(struct device *dev,
3429                                      struct device_attribute *attr,
3430                                      char *buf)
3431 {
3432         struct nvme_ctrl *ctrl = dev_get_drvdata(dev);
3433         static const char *const state_name[] = {
3434                 [NVME_CTRL_NEW]         = "new",
3435                 [NVME_CTRL_LIVE]        = "live",
3436                 [NVME_CTRL_RESETTING]   = "resetting",
3437                 [NVME_CTRL_CONNECTING]  = "connecting",
3438                 [NVME_CTRL_DELETING]    = "deleting",
3439                 [NVME_CTRL_DELETING_NOIO]= "deleting (no IO)",
3440                 [NVME_CTRL_DEAD]        = "dead",
3441         };
3442
3443         if ((unsigned)ctrl->state < ARRAY_SIZE(state_name) &&
3444             state_name[ctrl->state])
3445                 return sysfs_emit(buf, "%s\n", state_name[ctrl->state]);
3446
3447         return sysfs_emit(buf, "unknown state\n");
3448 }
3449
3450 static DEVICE_ATTR(state, S_IRUGO, nvme_sysfs_show_state, NULL);
3451
3452 static ssize_t nvme_sysfs_show_subsysnqn(struct device *dev,
3453                                          struct device_attribute *attr,
3454                                          char *buf)
3455 {
3456         struct nvme_ctrl *ctrl = dev_get_drvdata(dev);
3457
3458         return sysfs_emit(buf, "%s\n", ctrl->subsys->subnqn);
3459 }
3460 static DEVICE_ATTR(subsysnqn, S_IRUGO, nvme_sysfs_show_subsysnqn, NULL);
3461
3462 static ssize_t nvme_sysfs_show_hostnqn(struct device *dev,
3463                                         struct device_attribute *attr,
3464                                         char *buf)
3465 {
3466         struct nvme_ctrl *ctrl = dev_get_drvdata(dev);
3467
3468         return sysfs_emit(buf, "%s\n", ctrl->opts->host->nqn);
3469 }
3470 static DEVICE_ATTR(hostnqn, S_IRUGO, nvme_sysfs_show_hostnqn, NULL);
3471
3472 static ssize_t nvme_sysfs_show_hostid(struct device *dev,
3473                                         struct device_attribute *attr,
3474                                         char *buf)
3475 {
3476         struct nvme_ctrl *ctrl = dev_get_drvdata(dev);
3477
3478         return sysfs_emit(buf, "%pU\n", &ctrl->opts->host->id);
3479 }
3480 static DEVICE_ATTR(hostid, S_IRUGO, nvme_sysfs_show_hostid, NULL);
3481
3482 static ssize_t nvme_sysfs_show_address(struct device *dev,
3483                                          struct device_attribute *attr,
3484                                          char *buf)
3485 {
3486         struct nvme_ctrl *ctrl = dev_get_drvdata(dev);
3487
3488         return ctrl->ops->get_address(ctrl, buf, PAGE_SIZE);
3489 }
3490 static DEVICE_ATTR(address, S_IRUGO, nvme_sysfs_show_address, NULL);
3491
3492 static ssize_t nvme_ctrl_loss_tmo_show(struct device *dev,
3493                 struct device_attribute *attr, char *buf)
3494 {
3495         struct nvme_ctrl *ctrl = dev_get_drvdata(dev);
3496         struct nvmf_ctrl_options *opts = ctrl->opts;
3497
3498         if (ctrl->opts->max_reconnects == -1)
3499                 return sysfs_emit(buf, "off\n");
3500         return sysfs_emit(buf, "%d\n",
3501                           opts->max_reconnects * opts->reconnect_delay);
3502 }
3503
3504 static ssize_t nvme_ctrl_loss_tmo_store(struct device *dev,
3505                 struct device_attribute *attr, const char *buf, size_t count)
3506 {
3507         struct nvme_ctrl *ctrl = dev_get_drvdata(dev);
3508         struct nvmf_ctrl_options *opts = ctrl->opts;
3509         int ctrl_loss_tmo, err;
3510
3511         err = kstrtoint(buf, 10, &ctrl_loss_tmo);
3512         if (err)
3513                 return -EINVAL;
3514
3515         if (ctrl_loss_tmo < 0)
3516                 opts->max_reconnects = -1;
3517         else
3518                 opts->max_reconnects = DIV_ROUND_UP(ctrl_loss_tmo,
3519                                                 opts->reconnect_delay);
3520         return count;
3521 }
3522 static DEVICE_ATTR(ctrl_loss_tmo, S_IRUGO | S_IWUSR,
3523         nvme_ctrl_loss_tmo_show, nvme_ctrl_loss_tmo_store);
3524
3525 static ssize_t nvme_ctrl_reconnect_delay_show(struct device *dev,
3526                 struct device_attribute *attr, char *buf)
3527 {
3528         struct nvme_ctrl *ctrl = dev_get_drvdata(dev);
3529
3530         if (ctrl->opts->reconnect_delay == -1)
3531                 return sysfs_emit(buf, "off\n");
3532         return sysfs_emit(buf, "%d\n", ctrl->opts->reconnect_delay);
3533 }
3534
3535 static ssize_t nvme_ctrl_reconnect_delay_store(struct device *dev,
3536                 struct device_attribute *attr, const char *buf, size_t count)
3537 {
3538         struct nvme_ctrl *ctrl = dev_get_drvdata(dev);
3539         unsigned int v;
3540         int err;
3541
3542         err = kstrtou32(buf, 10, &v);
3543         if (err)
3544                 return err;
3545
3546         ctrl->opts->reconnect_delay = v;
3547         return count;
3548 }
3549 static DEVICE_ATTR(reconnect_delay, S_IRUGO | S_IWUSR,
3550         nvme_ctrl_reconnect_delay_show, nvme_ctrl_reconnect_delay_store);
3551
3552 static ssize_t nvme_ctrl_fast_io_fail_tmo_show(struct device *dev,
3553                 struct device_attribute *attr, char *buf)
3554 {
3555         struct nvme_ctrl *ctrl = dev_get_drvdata(dev);
3556
3557         if (ctrl->opts->fast_io_fail_tmo == -1)
3558                 return sysfs_emit(buf, "off\n");
3559         return sysfs_emit(buf, "%d\n", ctrl->opts->fast_io_fail_tmo);
3560 }
3561
3562 static ssize_t nvme_ctrl_fast_io_fail_tmo_store(struct device *dev,
3563                 struct device_attribute *attr, const char *buf, size_t count)
3564 {
3565         struct nvme_ctrl *ctrl = dev_get_drvdata(dev);
3566         struct nvmf_ctrl_options *opts = ctrl->opts;
3567         int fast_io_fail_tmo, err;
3568
3569         err = kstrtoint(buf, 10, &fast_io_fail_tmo);
3570         if (err)
3571                 return -EINVAL;
3572
3573         if (fast_io_fail_tmo < 0)
3574                 opts->fast_io_fail_tmo = -1;
3575         else
3576                 opts->fast_io_fail_tmo = fast_io_fail_tmo;
3577         return count;
3578 }
3579 static DEVICE_ATTR(fast_io_fail_tmo, S_IRUGO | S_IWUSR,
3580         nvme_ctrl_fast_io_fail_tmo_show, nvme_ctrl_fast_io_fail_tmo_store);
3581
3582 static ssize_t cntrltype_show(struct device *dev,
3583                               struct device_attribute *attr, char *buf)
3584 {
3585         static const char * const type[] = {
3586                 [NVME_CTRL_IO] = "io\n",
3587                 [NVME_CTRL_DISC] = "discovery\n",
3588                 [NVME_CTRL_ADMIN] = "admin\n",
3589         };
3590         struct nvme_ctrl *ctrl = dev_get_drvdata(dev);
3591
3592         if (ctrl->cntrltype > NVME_CTRL_ADMIN || !type[ctrl->cntrltype])
3593                 return sysfs_emit(buf, "reserved\n");
3594
3595         return sysfs_emit(buf, type[ctrl->cntrltype]);
3596 }
3597 static DEVICE_ATTR_RO(cntrltype);
3598
3599 static ssize_t dctype_show(struct device *dev,
3600                            struct device_attribute *attr, char *buf)
3601 {
3602         static const char * const type[] = {
3603                 [NVME_DCTYPE_NOT_REPORTED] = "none\n",
3604                 [NVME_DCTYPE_DDC] = "ddc\n",
3605                 [NVME_DCTYPE_CDC] = "cdc\n",
3606         };
3607         struct nvme_ctrl *ctrl = dev_get_drvdata(dev);
3608
3609         if (ctrl->dctype > NVME_DCTYPE_CDC || !type[ctrl->dctype])
3610                 return sysfs_emit(buf, "reserved\n");
3611
3612         return sysfs_emit(buf, type[ctrl->dctype]);
3613 }
3614 static DEVICE_ATTR_RO(dctype);
3615
3616 static struct attribute *nvme_dev_attrs[] = {
3617         &dev_attr_reset_controller.attr,
3618         &dev_attr_rescan_controller.attr,
3619         &dev_attr_model.attr,
3620         &dev_attr_serial.attr,
3621         &dev_attr_firmware_rev.attr,
3622         &dev_attr_cntlid.attr,
3623         &dev_attr_delete_controller.attr,
3624         &dev_attr_transport.attr,
3625         &dev_attr_subsysnqn.attr,
3626         &dev_attr_address.attr,
3627         &dev_attr_state.attr,
3628         &dev_attr_numa_node.attr,
3629         &dev_attr_queue_count.attr,
3630         &dev_attr_sqsize.attr,
3631         &dev_attr_hostnqn.attr,
3632         &dev_attr_hostid.attr,
3633         &dev_attr_ctrl_loss_tmo.attr,
3634         &dev_attr_reconnect_delay.attr,
3635         &dev_attr_fast_io_fail_tmo.attr,
3636         &dev_attr_kato.attr,
3637         &dev_attr_cntrltype.attr,
3638         &dev_attr_dctype.attr,
3639         NULL
3640 };
3641
3642 static umode_t nvme_dev_attrs_are_visible(struct kobject *kobj,
3643                 struct attribute *a, int n)
3644 {
3645         struct device *dev = container_of(kobj, struct device, kobj);
3646         struct nvme_ctrl *ctrl = dev_get_drvdata(dev);
3647
3648         if (a == &dev_attr_delete_controller.attr && !ctrl->ops->delete_ctrl)
3649                 return 0;
3650         if (a == &dev_attr_address.attr && !ctrl->ops->get_address)
3651                 return 0;
3652         if (a == &dev_attr_hostnqn.attr && !ctrl->opts)
3653                 return 0;
3654         if (a == &dev_attr_hostid.attr && !ctrl->opts)
3655                 return 0;
3656         if (a == &dev_attr_ctrl_loss_tmo.attr && !ctrl->opts)
3657                 return 0;
3658         if (a == &dev_attr_reconnect_delay.attr && !ctrl->opts)
3659                 return 0;
3660         if (a == &dev_attr_fast_io_fail_tmo.attr && !ctrl->opts)
3661                 return 0;
3662
3663         return a->mode;
3664 }
3665
3666 static const struct attribute_group nvme_dev_attrs_group = {
3667         .attrs          = nvme_dev_attrs,
3668         .is_visible     = nvme_dev_attrs_are_visible,
3669 };
3670
3671 static const struct attribute_group *nvme_dev_attr_groups[] = {
3672         &nvme_dev_attrs_group,
3673         NULL,
3674 };
3675
3676 static struct nvme_ns_head *nvme_find_ns_head(struct nvme_ctrl *ctrl,
3677                 unsigned nsid)
3678 {
3679         struct nvme_ns_head *h;
3680
3681         lockdep_assert_held(&ctrl->subsys->lock);
3682
3683         list_for_each_entry(h, &ctrl->subsys->nsheads, entry) {
3684                 /*
3685                  * Private namespaces can share NSIDs under some conditions.
3686                  * In that case we can't use the same ns_head for namespaces
3687                  * with the same NSID.
3688                  */
3689                 if (h->ns_id != nsid || !nvme_is_unique_nsid(ctrl, h))
3690                         continue;
3691                 if (!list_empty(&h->list) && nvme_tryget_ns_head(h))
3692                         return h;
3693         }
3694
3695         return NULL;
3696 }
3697
3698 static int nvme_subsys_check_duplicate_ids(struct nvme_subsystem *subsys,
3699                 struct nvme_ns_ids *ids)
3700 {
3701         bool has_uuid = !uuid_is_null(&ids->uuid);
3702         bool has_nguid = memchr_inv(ids->nguid, 0, sizeof(ids->nguid));
3703         bool has_eui64 = memchr_inv(ids->eui64, 0, sizeof(ids->eui64));
3704         struct nvme_ns_head *h;
3705
3706         lockdep_assert_held(&subsys->lock);
3707
3708         list_for_each_entry(h, &subsys->nsheads, entry) {
3709                 if (has_uuid && uuid_equal(&ids->uuid, &h->ids.uuid))
3710                         return -EINVAL;
3711                 if (has_nguid &&
3712                     memcmp(&ids->nguid, &h->ids.nguid, sizeof(ids->nguid)) == 0)
3713                         return -EINVAL;
3714                 if (has_eui64 &&
3715                     memcmp(&ids->eui64, &h->ids.eui64, sizeof(ids->eui64)) == 0)
3716                         return -EINVAL;
3717         }
3718
3719         return 0;
3720 }
3721
3722 static void nvme_cdev_rel(struct device *dev)
3723 {
3724         ida_free(&nvme_ns_chr_minor_ida, MINOR(dev->devt));
3725 }
3726
3727 void nvme_cdev_del(struct cdev *cdev, struct device *cdev_device)
3728 {
3729         cdev_device_del(cdev, cdev_device);
3730         put_device(cdev_device);
3731 }
3732
3733 int nvme_cdev_add(struct cdev *cdev, struct device *cdev_device,
3734                 const struct file_operations *fops, struct module *owner)
3735 {
3736         int minor, ret;
3737
3738         minor = ida_alloc(&nvme_ns_chr_minor_ida, GFP_KERNEL);
3739         if (minor < 0)
3740                 return minor;
3741         cdev_device->devt = MKDEV(MAJOR(nvme_ns_chr_devt), minor);
3742         cdev_device->class = nvme_ns_chr_class;
3743         cdev_device->release = nvme_cdev_rel;
3744         device_initialize(cdev_device);
3745         cdev_init(cdev, fops);
3746         cdev->owner = owner;
3747         ret = cdev_device_add(cdev, cdev_device);
3748         if (ret)
3749                 put_device(cdev_device);
3750
3751         return ret;
3752 }
3753
3754 static int nvme_ns_chr_open(struct inode *inode, struct file *file)
3755 {
3756         return nvme_ns_open(container_of(inode->i_cdev, struct nvme_ns, cdev));
3757 }
3758
3759 static int nvme_ns_chr_release(struct inode *inode, struct file *file)
3760 {
3761         nvme_ns_release(container_of(inode->i_cdev, struct nvme_ns, cdev));
3762         return 0;
3763 }
3764
3765 static const struct file_operations nvme_ns_chr_fops = {
3766         .owner          = THIS_MODULE,
3767         .open           = nvme_ns_chr_open,
3768         .release        = nvme_ns_chr_release,
3769         .unlocked_ioctl = nvme_ns_chr_ioctl,
3770         .compat_ioctl   = compat_ptr_ioctl,
3771         .uring_cmd      = nvme_ns_chr_uring_cmd,
3772 };
3773
3774 static int nvme_add_ns_cdev(struct nvme_ns *ns)
3775 {
3776         int ret;
3777
3778         ns->cdev_device.parent = ns->ctrl->device;
3779         ret = dev_set_name(&ns->cdev_device, "ng%dn%d",
3780                            ns->ctrl->instance, ns->head->instance);
3781         if (ret)
3782                 return ret;
3783
3784         return nvme_cdev_add(&ns->cdev, &ns->cdev_device, &nvme_ns_chr_fops,
3785                              ns->ctrl->ops->module);
3786 }
3787
3788 static struct nvme_ns_head *nvme_alloc_ns_head(struct nvme_ctrl *ctrl,
3789                 unsigned nsid, struct nvme_ns_ids *ids, bool is_shared)
3790 {
3791         struct nvme_ns_head *head;
3792         size_t size = sizeof(*head);
3793         int ret = -ENOMEM;
3794
3795 #ifdef CONFIG_NVME_MULTIPATH
3796         size += num_possible_nodes() * sizeof(struct nvme_ns *);
3797 #endif
3798
3799         head = kzalloc(size, GFP_KERNEL);
3800         if (!head)
3801                 goto out;
3802         ret = ida_alloc_min(&ctrl->subsys->ns_ida, 1, GFP_KERNEL);
3803         if (ret < 0)
3804                 goto out_free_head;
3805         head->instance = ret;
3806         INIT_LIST_HEAD(&head->list);
3807         ret = init_srcu_struct(&head->srcu);
3808         if (ret)
3809                 goto out_ida_remove;
3810         head->subsys = ctrl->subsys;
3811         head->ns_id = nsid;
3812         head->ids = *ids;
3813         head->shared = is_shared;
3814         kref_init(&head->ref);
3815
3816         if (head->ids.csi) {
3817                 ret = nvme_get_effects_log(ctrl, head->ids.csi, &head->effects);
3818                 if (ret)
3819                         goto out_cleanup_srcu;
3820         } else
3821                 head->effects = ctrl->effects;
3822
3823         ret = nvme_mpath_alloc_disk(ctrl, head);
3824         if (ret)
3825                 goto out_cleanup_srcu;
3826
3827         list_add_tail(&head->entry, &ctrl->subsys->nsheads);
3828
3829         kref_get(&ctrl->subsys->ref);
3830
3831         return head;
3832 out_cleanup_srcu:
3833         cleanup_srcu_struct(&head->srcu);
3834 out_ida_remove:
3835         ida_free(&ctrl->subsys->ns_ida, head->instance);
3836 out_free_head:
3837         kfree(head);
3838 out:
3839         if (ret > 0)
3840                 ret = blk_status_to_errno(nvme_error_status(ret));
3841         return ERR_PTR(ret);
3842 }
3843
3844 static int nvme_global_check_duplicate_ids(struct nvme_subsystem *this,
3845                 struct nvme_ns_ids *ids)
3846 {
3847         struct nvme_subsystem *s;
3848         int ret = 0;
3849
3850         /*
3851          * Note that this check is racy as we try to avoid holding the global
3852          * lock over the whole ns_head creation.  But it is only intended as
3853          * a sanity check anyway.
3854          */
3855         mutex_lock(&nvme_subsystems_lock);
3856         list_for_each_entry(s, &nvme_subsystems, entry) {
3857                 if (s == this)
3858                         continue;
3859                 mutex_lock(&s->lock);
3860                 ret = nvme_subsys_check_duplicate_ids(s, ids);
3861                 mutex_unlock(&s->lock);
3862                 if (ret)
3863                         break;
3864         }
3865         mutex_unlock(&nvme_subsystems_lock);
3866
3867         return ret;
3868 }
3869
3870 static int nvme_init_ns_head(struct nvme_ns *ns, unsigned nsid,
3871                 struct nvme_ns_ids *ids, bool is_shared)
3872 {
3873         struct nvme_ctrl *ctrl = ns->ctrl;
3874         struct nvme_ns_head *head = NULL;
3875         int ret;
3876
3877         ret = nvme_global_check_duplicate_ids(ctrl->subsys, ids);
3878         if (ret) {
3879                 dev_err(ctrl->device,
3880                         "globally duplicate IDs for nsid %d\n", nsid);
3881                 nvme_print_device_info(ctrl);
3882                 return ret;
3883         }
3884
3885         mutex_lock(&ctrl->subsys->lock);
3886         head = nvme_find_ns_head(ctrl, nsid);
3887         if (!head) {
3888                 ret = nvme_subsys_check_duplicate_ids(ctrl->subsys, ids);
3889                 if (ret) {
3890                         dev_err(ctrl->device,
3891                                 "duplicate IDs in subsystem for nsid %d\n",
3892                                 nsid);
3893                         goto out_unlock;
3894                 }
3895                 head = nvme_alloc_ns_head(ctrl, nsid, ids, is_shared);
3896                 if (IS_ERR(head)) {
3897                         ret = PTR_ERR(head);
3898                         goto out_unlock;
3899                 }
3900         } else {
3901                 ret = -EINVAL;
3902                 if (!is_shared || !head->shared) {
3903                         dev_err(ctrl->device,
3904                                 "Duplicate unshared namespace %d\n", nsid);
3905                         goto out_put_ns_head;
3906                 }
3907                 if (!nvme_ns_ids_equal(&head->ids, ids)) {
3908                         dev_err(ctrl->device,
3909                                 "IDs don't match for shared namespace %d\n",
3910                                         nsid);
3911                         goto out_put_ns_head;
3912                 }
3913
3914                 if (!multipath && !list_empty(&head->list)) {
3915                         dev_warn(ctrl->device,
3916                                 "Found shared namespace %d, but multipathing not supported.\n",
3917                                 nsid);
3918                         dev_warn_once(ctrl->device,
3919                                 "Support for shared namespaces without CONFIG_NVME_MULTIPATH is deprecated and will be removed in Linux 6.0\n.");
3920                 }
3921         }
3922
3923         list_add_tail_rcu(&ns->siblings, &head->list);
3924         ns->head = head;
3925         mutex_unlock(&ctrl->subsys->lock);
3926         return 0;
3927
3928 out_put_ns_head:
3929         nvme_put_ns_head(head);
3930 out_unlock:
3931         mutex_unlock(&ctrl->subsys->lock);
3932         return ret;
3933 }
3934
3935 struct nvme_ns *nvme_find_get_ns(struct nvme_ctrl *ctrl, unsigned nsid)
3936 {
3937         struct nvme_ns *ns, *ret = NULL;
3938
3939         down_read(&ctrl->namespaces_rwsem);
3940         list_for_each_entry(ns, &ctrl->namespaces, list) {
3941                 if (ns->head->ns_id == nsid) {
3942                         if (!nvme_get_ns(ns))
3943                                 continue;
3944                         ret = ns;
3945                         break;
3946                 }
3947                 if (ns->head->ns_id > nsid)
3948                         break;
3949         }
3950         up_read(&ctrl->namespaces_rwsem);
3951         return ret;
3952 }
3953 EXPORT_SYMBOL_NS_GPL(nvme_find_get_ns, NVME_TARGET_PASSTHRU);
3954
3955 /*
3956  * Add the namespace to the controller list while keeping the list ordered.
3957  */
3958 static void nvme_ns_add_to_ctrl_list(struct nvme_ns *ns)
3959 {
3960         struct nvme_ns *tmp;
3961
3962         list_for_each_entry_reverse(tmp, &ns->ctrl->namespaces, list) {
3963                 if (tmp->head->ns_id < ns->head->ns_id) {
3964                         list_add(&ns->list, &tmp->list);
3965                         return;
3966                 }
3967         }
3968         list_add(&ns->list, &ns->ctrl->namespaces);
3969 }
3970
3971 static void nvme_alloc_ns(struct nvme_ctrl *ctrl, unsigned nsid,
3972                 struct nvme_ns_ids *ids)
3973 {
3974         struct nvme_ns *ns;
3975         struct gendisk *disk;
3976         struct nvme_id_ns *id;
3977         int node = ctrl->numa_node;
3978
3979         if (nvme_identify_ns(ctrl, nsid, ids, &id))
3980                 return;
3981
3982         ns = kzalloc_node(sizeof(*ns), GFP_KERNEL, node);
3983         if (!ns)
3984                 goto out_free_id;
3985
3986         disk = blk_mq_alloc_disk(ctrl->tagset, ns);
3987         if (IS_ERR(disk))
3988                 goto out_free_ns;
3989         disk->fops = &nvme_bdev_ops;
3990         disk->private_data = ns;
3991
3992         ns->disk = disk;
3993         ns->queue = disk->queue;
3994
3995         if (ctrl->opts && ctrl->opts->data_digest)
3996                 blk_queue_flag_set(QUEUE_FLAG_STABLE_WRITES, ns->queue);
3997
3998         blk_queue_flag_set(QUEUE_FLAG_NONROT, ns->queue);
3999         if (ctrl->ops->flags & NVME_F_PCI_P2PDMA)
4000                 blk_queue_flag_set(QUEUE_FLAG_PCI_P2PDMA, ns->queue);
4001
4002         ns->ctrl = ctrl;
4003         kref_init(&ns->kref);
4004
4005         if (nvme_init_ns_head(ns, nsid, ids, id->nmic & NVME_NS_NMIC_SHARED))
4006                 goto out_cleanup_disk;
4007
4008         /*
4009          * If multipathing is enabled, the device name for all disks and not
4010          * just those that represent shared namespaces needs to be based on the
4011          * subsystem instance.  Using the controller instance for private
4012          * namespaces could lead to naming collisions between shared and private
4013          * namespaces if they don't use a common numbering scheme.
4014          *
4015          * If multipathing is not enabled, disk names must use the controller
4016          * instance as shared namespaces will show up as multiple block
4017          * devices.
4018          */
4019         if (ns->head->disk) {
4020                 sprintf(disk->disk_name, "nvme%dc%dn%d", ctrl->subsys->instance,
4021                         ctrl->instance, ns->head->instance);
4022                 disk->flags |= GENHD_FL_HIDDEN;
4023         } else if (multipath) {
4024                 sprintf(disk->disk_name, "nvme%dn%d", ctrl->subsys->instance,
4025                         ns->head->instance);
4026         } else {
4027                 sprintf(disk->disk_name, "nvme%dn%d", ctrl->instance,
4028                         ns->head->instance);
4029         }
4030
4031         if (nvme_update_ns_info(ns, id))
4032                 goto out_unlink_ns;
4033
4034         down_write(&ctrl->namespaces_rwsem);
4035         nvme_ns_add_to_ctrl_list(ns);
4036         up_write(&ctrl->namespaces_rwsem);
4037         nvme_get_ctrl(ctrl);
4038
4039         if (device_add_disk(ctrl->device, ns->disk, nvme_ns_id_attr_groups))
4040                 goto out_cleanup_ns_from_list;
4041
4042         if (!nvme_ns_head_multipath(ns->head))
4043                 nvme_add_ns_cdev(ns);
4044
4045         nvme_mpath_add_disk(ns, id);
4046         nvme_fault_inject_init(&ns->fault_inject, ns->disk->disk_name);
4047         kfree(id);
4048
4049         return;
4050
4051  out_cleanup_ns_from_list:
4052         nvme_put_ctrl(ctrl);
4053         down_write(&ctrl->namespaces_rwsem);
4054         list_del_init(&ns->list);
4055         up_write(&ctrl->namespaces_rwsem);
4056  out_unlink_ns:
4057         mutex_lock(&ctrl->subsys->lock);
4058         list_del_rcu(&ns->siblings);
4059         if (list_empty(&ns->head->list))
4060                 list_del_init(&ns->head->entry);
4061         mutex_unlock(&ctrl->subsys->lock);
4062         nvme_put_ns_head(ns->head);
4063  out_cleanup_disk:
4064         blk_cleanup_disk(disk);
4065  out_free_ns:
4066         kfree(ns);
4067  out_free_id:
4068         kfree(id);
4069 }
4070
4071 static void nvme_ns_remove(struct nvme_ns *ns)
4072 {
4073         bool last_path = false;
4074
4075         if (test_and_set_bit(NVME_NS_REMOVING, &ns->flags))
4076                 return;
4077
4078         clear_bit(NVME_NS_READY, &ns->flags);
4079         set_capacity(ns->disk, 0);
4080         nvme_fault_inject_fini(&ns->fault_inject);
4081
4082         /*
4083          * Ensure that !NVME_NS_READY is seen by other threads to prevent
4084          * this ns going back into current_path.
4085          */
4086         synchronize_srcu(&ns->head->srcu);
4087
4088         /* wait for concurrent submissions */
4089         if (nvme_mpath_clear_current_path(ns))
4090                 synchronize_srcu(&ns->head->srcu);
4091
4092         mutex_lock(&ns->ctrl->subsys->lock);
4093         list_del_rcu(&ns->siblings);
4094         if (list_empty(&ns->head->list)) {
4095                 list_del_init(&ns->head->entry);
4096                 last_path = true;
4097         }
4098         mutex_unlock(&ns->ctrl->subsys->lock);
4099
4100         /* guarantee not available in head->list */
4101         synchronize_rcu();
4102
4103         if (!nvme_ns_head_multipath(ns->head))
4104                 nvme_cdev_del(&ns->cdev, &ns->cdev_device);
4105         del_gendisk(ns->disk);
4106         blk_cleanup_queue(ns->queue);
4107
4108         down_write(&ns->ctrl->namespaces_rwsem);
4109         list_del_init(&ns->list);
4110         up_write(&ns->ctrl->namespaces_rwsem);
4111
4112         if (last_path)
4113                 nvme_mpath_shutdown_disk(ns->head);
4114         nvme_put_ns(ns);
4115 }
4116
4117 static void nvme_ns_remove_by_nsid(struct nvme_ctrl *ctrl, u32 nsid)
4118 {
4119         struct nvme_ns *ns = nvme_find_get_ns(ctrl, nsid);
4120
4121         if (ns) {
4122                 nvme_ns_remove(ns);
4123                 nvme_put_ns(ns);
4124         }
4125 }
4126
4127 static void nvme_validate_ns(struct nvme_ns *ns, struct nvme_ns_ids *ids)
4128 {
4129         struct nvme_id_ns *id;
4130         int ret = NVME_SC_INVALID_NS | NVME_SC_DNR;
4131
4132         if (test_bit(NVME_NS_DEAD, &ns->flags))
4133                 goto out;
4134
4135         ret = nvme_identify_ns(ns->ctrl, ns->head->ns_id, ids, &id);
4136         if (ret)
4137                 goto out;
4138
4139         ret = NVME_SC_INVALID_NS | NVME_SC_DNR;
4140         if (!nvme_ns_ids_equal(&ns->head->ids, ids)) {
4141                 dev_err(ns->ctrl->device,
4142                         "identifiers changed for nsid %d\n", ns->head->ns_id);
4143                 goto out_free_id;
4144         }
4145
4146         ret = nvme_update_ns_info(ns, id);
4147
4148 out_free_id:
4149         kfree(id);
4150 out:
4151         /*
4152          * Only remove the namespace if we got a fatal error back from the
4153          * device, otherwise ignore the error and just move on.
4154          *
4155          * TODO: we should probably schedule a delayed retry here.
4156          */
4157         if (ret > 0 && (ret & NVME_SC_DNR))
4158                 nvme_ns_remove(ns);
4159 }
4160
4161 static void nvme_validate_or_alloc_ns(struct nvme_ctrl *ctrl, unsigned nsid)
4162 {
4163         struct nvme_ns_ids ids = { };
4164         struct nvme_id_ns_cs_indep *id;
4165         struct nvme_ns *ns;
4166         bool ready = true;
4167
4168         if (nvme_identify_ns_descs(ctrl, nsid, &ids))
4169                 return;
4170
4171         /*
4172          * Check if the namespace is ready.  If not ignore it, we will get an
4173          * AEN once it becomes ready and restart the scan.
4174          */
4175         if ((ctrl->cap & NVME_CAP_CRMS_CRIMS) &&
4176             !nvme_identify_ns_cs_indep(ctrl, nsid, &id)) {
4177                 ready = id->nstat & NVME_NSTAT_NRDY;
4178                 kfree(id);
4179         }
4180
4181         if (!ready)
4182                 return;
4183
4184         ns = nvme_find_get_ns(ctrl, nsid);
4185         if (ns) {
4186                 nvme_validate_ns(ns, &ids);
4187                 nvme_put_ns(ns);
4188                 return;
4189         }
4190
4191         switch (ids.csi) {
4192         case NVME_CSI_NVM:
4193                 nvme_alloc_ns(ctrl, nsid, &ids);
4194                 break;
4195         case NVME_CSI_ZNS:
4196                 if (!IS_ENABLED(CONFIG_BLK_DEV_ZONED)) {
4197                         dev_warn(ctrl->device,
4198                                 "nsid %u not supported without CONFIG_BLK_DEV_ZONED\n",
4199                                 nsid);
4200                         break;
4201                 }
4202                 if (!nvme_multi_css(ctrl)) {
4203                         dev_warn(ctrl->device,
4204                                 "command set not reported for nsid: %d\n",
4205                                 nsid);
4206                         break;
4207                 }
4208                 nvme_alloc_ns(ctrl, nsid, &ids);
4209                 break;
4210         default:
4211                 dev_warn(ctrl->device, "unknown csi %u for nsid %u\n",
4212                         ids.csi, nsid);
4213                 break;
4214         }
4215 }
4216
4217 static void nvme_remove_invalid_namespaces(struct nvme_ctrl *ctrl,
4218                                         unsigned nsid)
4219 {
4220         struct nvme_ns *ns, *next;
4221         LIST_HEAD(rm_list);
4222
4223         down_write(&ctrl->namespaces_rwsem);
4224         list_for_each_entry_safe(ns, next, &ctrl->namespaces, list) {
4225                 if (ns->head->ns_id > nsid || test_bit(NVME_NS_DEAD, &ns->flags))
4226                         list_move_tail(&ns->list, &rm_list);
4227         }
4228         up_write(&ctrl->namespaces_rwsem);
4229
4230         list_for_each_entry_safe(ns, next, &rm_list, list)
4231                 nvme_ns_remove(ns);
4232
4233 }
4234
4235 static int nvme_scan_ns_list(struct nvme_ctrl *ctrl)
4236 {
4237         const int nr_entries = NVME_IDENTIFY_DATA_SIZE / sizeof(__le32);
4238         __le32 *ns_list;
4239         u32 prev = 0;
4240         int ret = 0, i;
4241
4242         if (nvme_ctrl_limited_cns(ctrl))
4243                 return -EOPNOTSUPP;
4244
4245         ns_list = kzalloc(NVME_IDENTIFY_DATA_SIZE, GFP_KERNEL);
4246         if (!ns_list)
4247                 return -ENOMEM;
4248
4249         for (;;) {
4250                 struct nvme_command cmd = {
4251                         .identify.opcode        = nvme_admin_identify,
4252                         .identify.cns           = NVME_ID_CNS_NS_ACTIVE_LIST,
4253                         .identify.nsid          = cpu_to_le32(prev),
4254                 };
4255
4256                 ret = nvme_submit_sync_cmd(ctrl->admin_q, &cmd, ns_list,
4257                                             NVME_IDENTIFY_DATA_SIZE);
4258                 if (ret) {
4259                         dev_warn(ctrl->device,
4260                                 "Identify NS List failed (status=0x%x)\n", ret);
4261                         goto free;
4262                 }
4263
4264                 for (i = 0; i < nr_entries; i++) {
4265                         u32 nsid = le32_to_cpu(ns_list[i]);
4266
4267                         if (!nsid)      /* end of the list? */
4268                                 goto out;
4269                         nvme_validate_or_alloc_ns(ctrl, nsid);
4270                         while (++prev < nsid)
4271                                 nvme_ns_remove_by_nsid(ctrl, prev);
4272                 }
4273         }
4274  out:
4275         nvme_remove_invalid_namespaces(ctrl, prev);
4276  free:
4277         kfree(ns_list);
4278         return ret;
4279 }
4280
4281 static void nvme_scan_ns_sequential(struct nvme_ctrl *ctrl)
4282 {
4283         struct nvme_id_ctrl *id;
4284         u32 nn, i;
4285
4286         if (nvme_identify_ctrl(ctrl, &id))
4287                 return;
4288         nn = le32_to_cpu(id->nn);
4289         kfree(id);
4290
4291         for (i = 1; i <= nn; i++)
4292                 nvme_validate_or_alloc_ns(ctrl, i);
4293
4294         nvme_remove_invalid_namespaces(ctrl, nn);
4295 }
4296
4297 static void nvme_clear_changed_ns_log(struct nvme_ctrl *ctrl)
4298 {
4299         size_t log_size = NVME_MAX_CHANGED_NAMESPACES * sizeof(__le32);
4300         __le32 *log;
4301         int error;
4302
4303         log = kzalloc(log_size, GFP_KERNEL);
4304         if (!log)
4305                 return;
4306
4307         /*
4308          * We need to read the log to clear the AEN, but we don't want to rely
4309          * on it for the changed namespace information as userspace could have
4310          * raced with us in reading the log page, which could cause us to miss
4311          * updates.
4312          */
4313         error = nvme_get_log(ctrl, NVME_NSID_ALL, NVME_LOG_CHANGED_NS, 0,
4314                         NVME_CSI_NVM, log, log_size, 0);
4315         if (error)
4316                 dev_warn(ctrl->device,
4317                         "reading changed ns log failed: %d\n", error);
4318
4319         kfree(log);
4320 }
4321
4322 static void nvme_scan_work(struct work_struct *work)
4323 {
4324         struct nvme_ctrl *ctrl =
4325                 container_of(work, struct nvme_ctrl, scan_work);
4326         int ret;
4327
4328         /* No tagset on a live ctrl means IO queues could not created */
4329         if (ctrl->state != NVME_CTRL_LIVE || !ctrl->tagset)
4330                 return;
4331
4332         /*
4333          * Identify controller limits can change at controller reset due to
4334          * new firmware download, even though it is not common we cannot ignore
4335          * such scenario. Controller's non-mdts limits are reported in the unit
4336          * of logical blocks that is dependent on the format of attached
4337          * namespace. Hence re-read the limits at the time of ns allocation.
4338          */
4339         ret = nvme_init_non_mdts_limits(ctrl);
4340         if (ret < 0) {
4341                 dev_warn(ctrl->device,
4342                         "reading non-mdts-limits failed: %d\n", ret);
4343                 return;
4344         }
4345
4346         if (test_and_clear_bit(NVME_AER_NOTICE_NS_CHANGED, &ctrl->events)) {
4347                 dev_info(ctrl->device, "rescanning namespaces.\n");
4348                 nvme_clear_changed_ns_log(ctrl);
4349         }
4350
4351         mutex_lock(&ctrl->scan_lock);
4352         if (nvme_scan_ns_list(ctrl) != 0)
4353                 nvme_scan_ns_sequential(ctrl);
4354         mutex_unlock(&ctrl->scan_lock);
4355 }
4356
4357 /*
4358  * This function iterates the namespace list unlocked to allow recovery from
4359  * controller failure. It is up to the caller to ensure the namespace list is
4360  * not modified by scan work while this function is executing.
4361  */
4362 void nvme_remove_namespaces(struct nvme_ctrl *ctrl)
4363 {
4364         struct nvme_ns *ns, *next;
4365         LIST_HEAD(ns_list);
4366
4367         /*
4368          * make sure to requeue I/O to all namespaces as these
4369          * might result from the scan itself and must complete
4370          * for the scan_work to make progress
4371          */
4372         nvme_mpath_clear_ctrl_paths(ctrl);
4373
4374         /* prevent racing with ns scanning */
4375         flush_work(&ctrl->scan_work);
4376
4377         /*
4378          * The dead states indicates the controller was not gracefully
4379          * disconnected. In that case, we won't be able to flush any data while
4380          * removing the namespaces' disks; fail all the queues now to avoid
4381          * potentially having to clean up the failed sync later.
4382          */
4383         if (ctrl->state == NVME_CTRL_DEAD)
4384                 nvme_kill_queues(ctrl);
4385
4386         /* this is a no-op when called from the controller reset handler */
4387         nvme_change_ctrl_state(ctrl, NVME_CTRL_DELETING_NOIO);
4388
4389         down_write(&ctrl->namespaces_rwsem);
4390         list_splice_init(&ctrl->namespaces, &ns_list);
4391         up_write(&ctrl->namespaces_rwsem);
4392
4393         list_for_each_entry_safe(ns, next, &ns_list, list)
4394                 nvme_ns_remove(ns);
4395 }
4396 EXPORT_SYMBOL_GPL(nvme_remove_namespaces);
4397
4398 static int nvme_class_uevent(struct device *dev, struct kobj_uevent_env *env)
4399 {
4400         struct nvme_ctrl *ctrl =
4401                 container_of(dev, struct nvme_ctrl, ctrl_device);
4402         struct nvmf_ctrl_options *opts = ctrl->opts;
4403         int ret;
4404
4405         ret = add_uevent_var(env, "NVME_TRTYPE=%s", ctrl->ops->name);
4406         if (ret)
4407                 return ret;
4408
4409         if (opts) {
4410                 ret = add_uevent_var(env, "NVME_TRADDR=%s", opts->traddr);
4411                 if (ret)
4412                         return ret;
4413
4414                 ret = add_uevent_var(env, "NVME_TRSVCID=%s",
4415                                 opts->trsvcid ?: "none");
4416                 if (ret)
4417                         return ret;
4418
4419                 ret = add_uevent_var(env, "NVME_HOST_TRADDR=%s",
4420                                 opts->host_traddr ?: "none");
4421                 if (ret)
4422                         return ret;
4423
4424                 ret = add_uevent_var(env, "NVME_HOST_IFACE=%s",
4425                                 opts->host_iface ?: "none");
4426         }
4427         return ret;
4428 }
4429
4430 static void nvme_change_uevent(struct nvme_ctrl *ctrl, char *envdata)
4431 {
4432         char *envp[2] = { envdata, NULL };
4433
4434         kobject_uevent_env(&ctrl->device->kobj, KOBJ_CHANGE, envp);
4435 }
4436
4437 static void nvme_aen_uevent(struct nvme_ctrl *ctrl)
4438 {
4439         char *envp[2] = { NULL, NULL };
4440         u32 aen_result = ctrl->aen_result;
4441
4442         ctrl->aen_result = 0;
4443         if (!aen_result)
4444                 return;
4445
4446         envp[0] = kasprintf(GFP_KERNEL, "NVME_AEN=%#08x", aen_result);
4447         if (!envp[0])
4448                 return;
4449         kobject_uevent_env(&ctrl->device->kobj, KOBJ_CHANGE, envp);
4450         kfree(envp[0]);
4451 }
4452
4453 static void nvme_async_event_work(struct work_struct *work)
4454 {
4455         struct nvme_ctrl *ctrl =
4456                 container_of(work, struct nvme_ctrl, async_event_work);
4457
4458         nvme_aen_uevent(ctrl);
4459
4460         /*
4461          * The transport drivers must guarantee AER submission here is safe by
4462          * flushing ctrl async_event_work after changing the controller state
4463          * from LIVE and before freeing the admin queue.
4464         */
4465         if (ctrl->state == NVME_CTRL_LIVE)
4466                 ctrl->ops->submit_async_event(ctrl);
4467 }
4468
4469 static bool nvme_ctrl_pp_status(struct nvme_ctrl *ctrl)
4470 {
4471
4472         u32 csts;
4473
4474         if (ctrl->ops->reg_read32(ctrl, NVME_REG_CSTS, &csts))
4475                 return false;
4476
4477         if (csts == ~0)
4478                 return false;
4479
4480         return ((ctrl->ctrl_config & NVME_CC_ENABLE) && (csts & NVME_CSTS_PP));
4481 }
4482
4483 static void nvme_get_fw_slot_info(struct nvme_ctrl *ctrl)
4484 {
4485         struct nvme_fw_slot_info_log *log;
4486
4487         log = kmalloc(sizeof(*log), GFP_KERNEL);
4488         if (!log)
4489                 return;
4490
4491         if (nvme_get_log(ctrl, NVME_NSID_ALL, NVME_LOG_FW_SLOT, 0, NVME_CSI_NVM,
4492                         log, sizeof(*log), 0))
4493                 dev_warn(ctrl->device, "Get FW SLOT INFO log error\n");
4494         kfree(log);
4495 }
4496
4497 static void nvme_fw_act_work(struct work_struct *work)
4498 {
4499         struct nvme_ctrl *ctrl = container_of(work,
4500                                 struct nvme_ctrl, fw_act_work);
4501         unsigned long fw_act_timeout;
4502
4503         if (ctrl->mtfa)
4504                 fw_act_timeout = jiffies +
4505                                 msecs_to_jiffies(ctrl->mtfa * 100);
4506         else
4507                 fw_act_timeout = jiffies +
4508                                 msecs_to_jiffies(admin_timeout * 1000);
4509
4510         nvme_stop_queues(ctrl);
4511         while (nvme_ctrl_pp_status(ctrl)) {
4512                 if (time_after(jiffies, fw_act_timeout)) {
4513                         dev_warn(ctrl->device,
4514                                 "Fw activation timeout, reset controller\n");
4515                         nvme_try_sched_reset(ctrl);
4516                         return;
4517                 }
4518                 msleep(100);
4519         }
4520
4521         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_LIVE))
4522                 return;
4523
4524         nvme_start_queues(ctrl);
4525         /* read FW slot information to clear the AER */
4526         nvme_get_fw_slot_info(ctrl);
4527 }
4528
4529 static void nvme_handle_aen_notice(struct nvme_ctrl *ctrl, u32 result)
4530 {
4531         u32 aer_notice_type = (result & 0xff00) >> 8;
4532
4533         trace_nvme_async_event(ctrl, aer_notice_type);
4534
4535         switch (aer_notice_type) {
4536         case NVME_AER_NOTICE_NS_CHANGED:
4537                 set_bit(NVME_AER_NOTICE_NS_CHANGED, &ctrl->events);
4538                 nvme_queue_scan(ctrl);
4539                 break;
4540         case NVME_AER_NOTICE_FW_ACT_STARTING:
4541                 /*
4542                  * We are (ab)using the RESETTING state to prevent subsequent
4543                  * recovery actions from interfering with the controller's
4544                  * firmware activation.
4545                  */
4546                 if (nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING))
4547                         queue_work(nvme_wq, &ctrl->fw_act_work);
4548                 break;
4549 #ifdef CONFIG_NVME_MULTIPATH
4550         case NVME_AER_NOTICE_ANA:
4551                 if (!ctrl->ana_log_buf)
4552                         break;
4553                 queue_work(nvme_wq, &ctrl->ana_work);
4554                 break;
4555 #endif
4556         case NVME_AER_NOTICE_DISC_CHANGED:
4557                 ctrl->aen_result = result;
4558                 break;
4559         default:
4560                 dev_warn(ctrl->device, "async event result %08x\n", result);
4561         }
4562 }
4563
4564 void nvme_complete_async_event(struct nvme_ctrl *ctrl, __le16 status,
4565                 volatile union nvme_result *res)
4566 {
4567         u32 result = le32_to_cpu(res->u32);
4568         u32 aer_type = result & 0x07;
4569
4570         if (le16_to_cpu(status) >> 1 != NVME_SC_SUCCESS)
4571                 return;
4572
4573         switch (aer_type) {
4574         case NVME_AER_NOTICE:
4575                 nvme_handle_aen_notice(ctrl, result);
4576                 break;
4577         case NVME_AER_ERROR:
4578         case NVME_AER_SMART:
4579         case NVME_AER_CSS:
4580         case NVME_AER_VS:
4581                 trace_nvme_async_event(ctrl, aer_type);
4582                 ctrl->aen_result = result;
4583                 break;
4584         default:
4585                 break;
4586         }
4587         queue_work(nvme_wq, &ctrl->async_event_work);
4588 }
4589 EXPORT_SYMBOL_GPL(nvme_complete_async_event);
4590
4591 void nvme_stop_ctrl(struct nvme_ctrl *ctrl)
4592 {
4593         nvme_mpath_stop(ctrl);
4594         nvme_stop_keep_alive(ctrl);
4595         nvme_stop_failfast_work(ctrl);
4596         flush_work(&ctrl->async_event_work);
4597         cancel_work_sync(&ctrl->fw_act_work);
4598         if (ctrl->ops->stop_ctrl)
4599                 ctrl->ops->stop_ctrl(ctrl);
4600 }
4601 EXPORT_SYMBOL_GPL(nvme_stop_ctrl);
4602
4603 void nvme_start_ctrl(struct nvme_ctrl *ctrl)
4604 {
4605         nvme_start_keep_alive(ctrl);
4606
4607         nvme_enable_aen(ctrl);
4608
4609         if (ctrl->queue_count > 1) {
4610                 nvme_queue_scan(ctrl);
4611                 nvme_start_queues(ctrl);
4612                 nvme_mpath_update(ctrl);
4613         }
4614
4615         nvme_change_uevent(ctrl, "NVME_EVENT=connected");
4616 }
4617 EXPORT_SYMBOL_GPL(nvme_start_ctrl);
4618
4619 void nvme_uninit_ctrl(struct nvme_ctrl *ctrl)
4620 {
4621         nvme_hwmon_exit(ctrl);
4622         nvme_fault_inject_fini(&ctrl->fault_inject);
4623         dev_pm_qos_hide_latency_tolerance(ctrl->device);
4624         cdev_device_del(&ctrl->cdev, ctrl->device);
4625         nvme_put_ctrl(ctrl);
4626 }
4627 EXPORT_SYMBOL_GPL(nvme_uninit_ctrl);
4628
4629 static void nvme_free_cels(struct nvme_ctrl *ctrl)
4630 {
4631         struct nvme_effects_log *cel;
4632         unsigned long i;
4633
4634         xa_for_each(&ctrl->cels, i, cel) {
4635                 xa_erase(&ctrl->cels, i);
4636                 kfree(cel);
4637         }
4638
4639         xa_destroy(&ctrl->cels);
4640 }
4641
4642 static void nvme_free_ctrl(struct device *dev)
4643 {
4644         struct nvme_ctrl *ctrl =
4645                 container_of(dev, struct nvme_ctrl, ctrl_device);
4646         struct nvme_subsystem *subsys = ctrl->subsys;
4647
4648         if (!subsys || ctrl->instance != subsys->instance)
4649                 ida_free(&nvme_instance_ida, ctrl->instance);
4650
4651         nvme_free_cels(ctrl);
4652         nvme_mpath_uninit(ctrl);
4653         __free_page(ctrl->discard_page);
4654
4655         if (subsys) {
4656                 mutex_lock(&nvme_subsystems_lock);
4657                 list_del(&ctrl->subsys_entry);
4658                 sysfs_remove_link(&subsys->dev.kobj, dev_name(ctrl->device));
4659                 mutex_unlock(&nvme_subsystems_lock);
4660         }
4661
4662         ctrl->ops->free_ctrl(ctrl);
4663
4664         if (subsys)
4665                 nvme_put_subsystem(subsys);
4666 }
4667
4668 /*
4669  * Initialize a NVMe controller structures.  This needs to be called during
4670  * earliest initialization so that we have the initialized structured around
4671  * during probing.
4672  */
4673 int nvme_init_ctrl(struct nvme_ctrl *ctrl, struct device *dev,
4674                 const struct nvme_ctrl_ops *ops, unsigned long quirks)
4675 {
4676         int ret;
4677
4678         ctrl->state = NVME_CTRL_NEW;
4679         clear_bit(NVME_CTRL_FAILFAST_EXPIRED, &ctrl->flags);
4680         spin_lock_init(&ctrl->lock);
4681         mutex_init(&ctrl->scan_lock);
4682         INIT_LIST_HEAD(&ctrl->namespaces);
4683         xa_init(&ctrl->cels);
4684         init_rwsem(&ctrl->namespaces_rwsem);
4685         ctrl->dev = dev;
4686         ctrl->ops = ops;
4687         ctrl->quirks = quirks;
4688         ctrl->numa_node = NUMA_NO_NODE;
4689         INIT_WORK(&ctrl->scan_work, nvme_scan_work);
4690         INIT_WORK(&ctrl->async_event_work, nvme_async_event_work);
4691         INIT_WORK(&ctrl->fw_act_work, nvme_fw_act_work);
4692         INIT_WORK(&ctrl->delete_work, nvme_delete_ctrl_work);
4693         init_waitqueue_head(&ctrl->state_wq);
4694
4695         INIT_DELAYED_WORK(&ctrl->ka_work, nvme_keep_alive_work);
4696         INIT_DELAYED_WORK(&ctrl->failfast_work, nvme_failfast_work);
4697         memset(&ctrl->ka_cmd, 0, sizeof(ctrl->ka_cmd));
4698         ctrl->ka_cmd.common.opcode = nvme_admin_keep_alive;
4699
4700         BUILD_BUG_ON(NVME_DSM_MAX_RANGES * sizeof(struct nvme_dsm_range) >
4701                         PAGE_SIZE);
4702         ctrl->discard_page = alloc_page(GFP_KERNEL);
4703         if (!ctrl->discard_page) {
4704                 ret = -ENOMEM;
4705                 goto out;
4706         }
4707
4708         ret = ida_alloc(&nvme_instance_ida, GFP_KERNEL);
4709         if (ret < 0)
4710                 goto out;
4711         ctrl->instance = ret;
4712
4713         device_initialize(&ctrl->ctrl_device);
4714         ctrl->device = &ctrl->ctrl_device;
4715         ctrl->device->devt = MKDEV(MAJOR(nvme_ctrl_base_chr_devt),
4716                         ctrl->instance);
4717         ctrl->device->class = nvme_class;
4718         ctrl->device->parent = ctrl->dev;
4719         ctrl->device->groups = nvme_dev_attr_groups;
4720         ctrl->device->release = nvme_free_ctrl;
4721         dev_set_drvdata(ctrl->device, ctrl);
4722         ret = dev_set_name(ctrl->device, "nvme%d", ctrl->instance);
4723         if (ret)
4724                 goto out_release_instance;
4725
4726         nvme_get_ctrl(ctrl);
4727         cdev_init(&ctrl->cdev, &nvme_dev_fops);
4728         ctrl->cdev.owner = ops->module;
4729         ret = cdev_device_add(&ctrl->cdev, ctrl->device);
4730         if (ret)
4731                 goto out_free_name;
4732
4733         /*
4734          * Initialize latency tolerance controls.  The sysfs files won't
4735          * be visible to userspace unless the device actually supports APST.
4736          */
4737         ctrl->device->power.set_latency_tolerance = nvme_set_latency_tolerance;
4738         dev_pm_qos_update_user_latency_tolerance(ctrl->device,
4739                 min(default_ps_max_latency_us, (unsigned long)S32_MAX));
4740
4741         nvme_fault_inject_init(&ctrl->fault_inject, dev_name(ctrl->device));
4742         nvme_mpath_init_ctrl(ctrl);
4743
4744         return 0;
4745 out_free_name:
4746         nvme_put_ctrl(ctrl);
4747         kfree_const(ctrl->device->kobj.name);
4748 out_release_instance:
4749         ida_free(&nvme_instance_ida, ctrl->instance);
4750 out:
4751         if (ctrl->discard_page)
4752                 __free_page(ctrl->discard_page);
4753         return ret;
4754 }
4755 EXPORT_SYMBOL_GPL(nvme_init_ctrl);
4756
4757 static void nvme_start_ns_queue(struct nvme_ns *ns)
4758 {
4759         if (test_and_clear_bit(NVME_NS_STOPPED, &ns->flags))
4760                 blk_mq_unquiesce_queue(ns->queue);
4761 }
4762
4763 static void nvme_stop_ns_queue(struct nvme_ns *ns)
4764 {
4765         if (!test_and_set_bit(NVME_NS_STOPPED, &ns->flags))
4766                 blk_mq_quiesce_queue(ns->queue);
4767         else
4768                 blk_mq_wait_quiesce_done(ns->queue);
4769 }
4770
4771 /*
4772  * Prepare a queue for teardown.
4773  *
4774  * This must forcibly unquiesce queues to avoid blocking dispatch, and only set
4775  * the capacity to 0 after that to avoid blocking dispatchers that may be
4776  * holding bd_butex.  This will end buffered writers dirtying pages that can't
4777  * be synced.
4778  */
4779 static void nvme_set_queue_dying(struct nvme_ns *ns)
4780 {
4781         if (test_and_set_bit(NVME_NS_DEAD, &ns->flags))
4782                 return;
4783
4784         blk_mark_disk_dead(ns->disk);
4785         nvme_start_ns_queue(ns);
4786
4787         set_capacity_and_notify(ns->disk, 0);
4788 }
4789
4790 /**
4791  * nvme_kill_queues(): Ends all namespace queues
4792  * @ctrl: the dead controller that needs to end
4793  *
4794  * Call this function when the driver determines it is unable to get the
4795  * controller in a state capable of servicing IO.
4796  */
4797 void nvme_kill_queues(struct nvme_ctrl *ctrl)
4798 {
4799         struct nvme_ns *ns;
4800
4801         down_read(&ctrl->namespaces_rwsem);
4802
4803         /* Forcibly unquiesce queues to avoid blocking dispatch */
4804         if (ctrl->admin_q && !blk_queue_dying(ctrl->admin_q))
4805                 nvme_start_admin_queue(ctrl);
4806
4807         list_for_each_entry(ns, &ctrl->namespaces, list)
4808                 nvme_set_queue_dying(ns);
4809
4810         up_read(&ctrl->namespaces_rwsem);
4811 }
4812 EXPORT_SYMBOL_GPL(nvme_kill_queues);
4813
4814 void nvme_unfreeze(struct nvme_ctrl *ctrl)
4815 {
4816         struct nvme_ns *ns;
4817
4818         down_read(&ctrl->namespaces_rwsem);
4819         list_for_each_entry(ns, &ctrl->namespaces, list)
4820                 blk_mq_unfreeze_queue(ns->queue);
4821         up_read(&ctrl->namespaces_rwsem);
4822 }
4823 EXPORT_SYMBOL_GPL(nvme_unfreeze);
4824
4825 int nvme_wait_freeze_timeout(struct nvme_ctrl *ctrl, long timeout)
4826 {
4827         struct nvme_ns *ns;
4828
4829         down_read(&ctrl->namespaces_rwsem);
4830         list_for_each_entry(ns, &ctrl->namespaces, list) {
4831                 timeout = blk_mq_freeze_queue_wait_timeout(ns->queue, timeout);
4832                 if (timeout <= 0)
4833                         break;
4834         }
4835         up_read(&ctrl->namespaces_rwsem);
4836         return timeout;
4837 }
4838 EXPORT_SYMBOL_GPL(nvme_wait_freeze_timeout);
4839
4840 void nvme_wait_freeze(struct nvme_ctrl *ctrl)
4841 {
4842         struct nvme_ns *ns;
4843
4844         down_read(&ctrl->namespaces_rwsem);
4845         list_for_each_entry(ns, &ctrl->namespaces, list)
4846                 blk_mq_freeze_queue_wait(ns->queue);
4847         up_read(&ctrl->namespaces_rwsem);
4848 }
4849 EXPORT_SYMBOL_GPL(nvme_wait_freeze);
4850
4851 void nvme_start_freeze(struct nvme_ctrl *ctrl)
4852 {
4853         struct nvme_ns *ns;
4854
4855         down_read(&ctrl->namespaces_rwsem);
4856         list_for_each_entry(ns, &ctrl->namespaces, list)
4857                 blk_freeze_queue_start(ns->queue);
4858         up_read(&ctrl->namespaces_rwsem);
4859 }
4860 EXPORT_SYMBOL_GPL(nvme_start_freeze);
4861
4862 void nvme_stop_queues(struct nvme_ctrl *ctrl)
4863 {
4864         struct nvme_ns *ns;
4865
4866         down_read(&ctrl->namespaces_rwsem);
4867         list_for_each_entry(ns, &ctrl->namespaces, list)
4868                 nvme_stop_ns_queue(ns);
4869         up_read(&ctrl->namespaces_rwsem);
4870 }
4871 EXPORT_SYMBOL_GPL(nvme_stop_queues);
4872
4873 void nvme_start_queues(struct nvme_ctrl *ctrl)
4874 {
4875         struct nvme_ns *ns;
4876
4877         down_read(&ctrl->namespaces_rwsem);
4878         list_for_each_entry(ns, &ctrl->namespaces, list)
4879                 nvme_start_ns_queue(ns);
4880         up_read(&ctrl->namespaces_rwsem);
4881 }
4882 EXPORT_SYMBOL_GPL(nvme_start_queues);
4883
4884 void nvme_stop_admin_queue(struct nvme_ctrl *ctrl)
4885 {
4886         if (!test_and_set_bit(NVME_CTRL_ADMIN_Q_STOPPED, &ctrl->flags))
4887                 blk_mq_quiesce_queue(ctrl->admin_q);
4888         else
4889                 blk_mq_wait_quiesce_done(ctrl->admin_q);
4890 }
4891 EXPORT_SYMBOL_GPL(nvme_stop_admin_queue);
4892
4893 void nvme_start_admin_queue(struct nvme_ctrl *ctrl)
4894 {
4895         if (test_and_clear_bit(NVME_CTRL_ADMIN_Q_STOPPED, &ctrl->flags))
4896                 blk_mq_unquiesce_queue(ctrl->admin_q);
4897 }
4898 EXPORT_SYMBOL_GPL(nvme_start_admin_queue);
4899
4900 void nvme_sync_io_queues(struct nvme_ctrl *ctrl)
4901 {
4902         struct nvme_ns *ns;
4903
4904         down_read(&ctrl->namespaces_rwsem);
4905         list_for_each_entry(ns, &ctrl->namespaces, list)
4906                 blk_sync_queue(ns->queue);
4907         up_read(&ctrl->namespaces_rwsem);
4908 }
4909 EXPORT_SYMBOL_GPL(nvme_sync_io_queues);
4910
4911 void nvme_sync_queues(struct nvme_ctrl *ctrl)
4912 {
4913         nvme_sync_io_queues(ctrl);
4914         if (ctrl->admin_q)
4915                 blk_sync_queue(ctrl->admin_q);
4916 }
4917 EXPORT_SYMBOL_GPL(nvme_sync_queues);
4918
4919 struct nvme_ctrl *nvme_ctrl_from_file(struct file *file)
4920 {
4921         if (file->f_op != &nvme_dev_fops)
4922                 return NULL;
4923         return file->private_data;
4924 }
4925 EXPORT_SYMBOL_NS_GPL(nvme_ctrl_from_file, NVME_TARGET_PASSTHRU);
4926
4927 /*
4928  * Check we didn't inadvertently grow the command structure sizes:
4929  */
4930 static inline void _nvme_check_size(void)
4931 {
4932         BUILD_BUG_ON(sizeof(struct nvme_common_command) != 64);
4933         BUILD_BUG_ON(sizeof(struct nvme_rw_command) != 64);
4934         BUILD_BUG_ON(sizeof(struct nvme_identify) != 64);
4935         BUILD_BUG_ON(sizeof(struct nvme_features) != 64);
4936         BUILD_BUG_ON(sizeof(struct nvme_download_firmware) != 64);
4937         BUILD_BUG_ON(sizeof(struct nvme_format_cmd) != 64);
4938         BUILD_BUG_ON(sizeof(struct nvme_dsm_cmd) != 64);
4939         BUILD_BUG_ON(sizeof(struct nvme_write_zeroes_cmd) != 64);
4940         BUILD_BUG_ON(sizeof(struct nvme_abort_cmd) != 64);
4941         BUILD_BUG_ON(sizeof(struct nvme_get_log_page_command) != 64);
4942         BUILD_BUG_ON(sizeof(struct nvme_command) != 64);
4943         BUILD_BUG_ON(sizeof(struct nvme_id_ctrl) != NVME_IDENTIFY_DATA_SIZE);
4944         BUILD_BUG_ON(sizeof(struct nvme_id_ns) != NVME_IDENTIFY_DATA_SIZE);
4945         BUILD_BUG_ON(sizeof(struct nvme_id_ns_cs_indep) !=
4946                         NVME_IDENTIFY_DATA_SIZE);
4947         BUILD_BUG_ON(sizeof(struct nvme_id_ns_zns) != NVME_IDENTIFY_DATA_SIZE);
4948         BUILD_BUG_ON(sizeof(struct nvme_id_ns_nvm) != NVME_IDENTIFY_DATA_SIZE);
4949         BUILD_BUG_ON(sizeof(struct nvme_id_ctrl_zns) != NVME_IDENTIFY_DATA_SIZE);
4950         BUILD_BUG_ON(sizeof(struct nvme_id_ctrl_nvm) != NVME_IDENTIFY_DATA_SIZE);
4951         BUILD_BUG_ON(sizeof(struct nvme_lba_range_type) != 64);
4952         BUILD_BUG_ON(sizeof(struct nvme_smart_log) != 512);
4953         BUILD_BUG_ON(sizeof(struct nvme_dbbuf) != 64);
4954         BUILD_BUG_ON(sizeof(struct nvme_directive_cmd) != 64);
4955         BUILD_BUG_ON(sizeof(struct nvme_feat_host_behavior) != 512);
4956 }
4957
4958
4959 static int __init nvme_core_init(void)
4960 {
4961         int result = -ENOMEM;
4962
4963         _nvme_check_size();
4964
4965         nvme_wq = alloc_workqueue("nvme-wq",
4966                         WQ_UNBOUND | WQ_MEM_RECLAIM | WQ_SYSFS, 0);
4967         if (!nvme_wq)
4968                 goto out;
4969
4970         nvme_reset_wq = alloc_workqueue("nvme-reset-wq",
4971                         WQ_UNBOUND | WQ_MEM_RECLAIM | WQ_SYSFS, 0);
4972         if (!nvme_reset_wq)
4973                 goto destroy_wq;
4974
4975         nvme_delete_wq = alloc_workqueue("nvme-delete-wq",
4976                         WQ_UNBOUND | WQ_MEM_RECLAIM | WQ_SYSFS, 0);
4977         if (!nvme_delete_wq)
4978                 goto destroy_reset_wq;
4979
4980         result = alloc_chrdev_region(&nvme_ctrl_base_chr_devt, 0,
4981                         NVME_MINORS, "nvme");
4982         if (result < 0)
4983                 goto destroy_delete_wq;
4984
4985         nvme_class = class_create(THIS_MODULE, "nvme");
4986         if (IS_ERR(nvme_class)) {
4987                 result = PTR_ERR(nvme_class);
4988                 goto unregister_chrdev;
4989         }
4990         nvme_class->dev_uevent = nvme_class_uevent;
4991
4992         nvme_subsys_class = class_create(THIS_MODULE, "nvme-subsystem");
4993         if (IS_ERR(nvme_subsys_class)) {
4994                 result = PTR_ERR(nvme_subsys_class);
4995                 goto destroy_class;
4996         }
4997
4998         result = alloc_chrdev_region(&nvme_ns_chr_devt, 0, NVME_MINORS,
4999                                      "nvme-generic");
5000         if (result < 0)
5001                 goto destroy_subsys_class;
5002
5003         nvme_ns_chr_class = class_create(THIS_MODULE, "nvme-generic");
5004         if (IS_ERR(nvme_ns_chr_class)) {
5005                 result = PTR_ERR(nvme_ns_chr_class);
5006                 goto unregister_generic_ns;
5007         }
5008
5009         return 0;
5010
5011 unregister_generic_ns:
5012         unregister_chrdev_region(nvme_ns_chr_devt, NVME_MINORS);
5013 destroy_subsys_class:
5014         class_destroy(nvme_subsys_class);
5015 destroy_class:
5016         class_destroy(nvme_class);
5017 unregister_chrdev:
5018         unregister_chrdev_region(nvme_ctrl_base_chr_devt, NVME_MINORS);
5019 destroy_delete_wq:
5020         destroy_workqueue(nvme_delete_wq);
5021 destroy_reset_wq:
5022         destroy_workqueue(nvme_reset_wq);
5023 destroy_wq:
5024         destroy_workqueue(nvme_wq);
5025 out:
5026         return result;
5027 }
5028
5029 static void __exit nvme_core_exit(void)
5030 {
5031         class_destroy(nvme_ns_chr_class);
5032         class_destroy(nvme_subsys_class);
5033         class_destroy(nvme_class);
5034         unregister_chrdev_region(nvme_ns_chr_devt, NVME_MINORS);
5035         unregister_chrdev_region(nvme_ctrl_base_chr_devt, NVME_MINORS);
5036         destroy_workqueue(nvme_delete_wq);
5037         destroy_workqueue(nvme_reset_wq);
5038         destroy_workqueue(nvme_wq);
5039         ida_destroy(&nvme_ns_chr_minor_ida);
5040         ida_destroy(&nvme_instance_ida);
5041 }
5042
5043 MODULE_LICENSE("GPL");
5044 MODULE_VERSION("1.0");
5045 module_init(nvme_core_init);
5046 module_exit(nvme_core_exit);