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/delay.h>
  10 #include <linux/errno.h>
  11 #include <linux/hdreg.h>
  12 #include <linux/kernel.h>
  13 #include <linux/module.h>
  14 #include <linux/backing-dev.h>
  15 #include <linux/list_sort.h>
  16 #include <linux/slab.h>
  17 #include <linux/types.h>
  18 #include <linux/pr.h>
  19 #include <linux/ptrace.h>
  20 #include <linux/nvme_ioctl.h>
  21 #include <linux/t10-pi.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 bool streams;
  61 module_param(streams, bool, 0644);
  62 MODULE_PARM_DESC(streams, "turn on support for Streams write directives");
  63
  64 /*
  65  * nvme_wq - hosts nvme related works that are not reset or delete
  66  * nvme_reset_wq - hosts nvme reset works
  67  * nvme_delete_wq - hosts nvme delete works
  68  *
  69  * nvme_wq will host works such are scan, aen handling, fw activation,
  70  * keep-alive error recovery, periodic reconnects etc. nvme_reset_wq
  71  * runs reset works which also flush works hosted on nvme_wq for
  72  * serialization purposes. nvme_delete_wq host controller deletion
  73  * works which flush reset works for serialization.
  74  */
  75 struct workqueue_struct *nvme_wq;
  76 EXPORT_SYMBOL_GPL(nvme_wq);
  77
  78 struct workqueue_struct *nvme_reset_wq;
  79 EXPORT_SYMBOL_GPL(nvme_reset_wq);
  80
  81 struct workqueue_struct *nvme_delete_wq;
  82 EXPORT_SYMBOL_GPL(nvme_delete_wq);
  83
  84 static LIST_HEAD(nvme_subsystems);
  85 static DEFINE_MUTEX(nvme_subsystems_lock);
  86
  87 static DEFINE_IDA(nvme_instance_ida);
  88 static dev_t nvme_chr_devt;
  89 static struct class *nvme_class;
  90 static struct class *nvme_subsys_class;
  91
  92 static int nvme_revalidate_disk(struct gendisk *disk);
  93 static void nvme_put_subsystem(struct nvme_subsystem *subsys);
  94 static void nvme_remove_invalid_namespaces(struct nvme_ctrl *ctrl,
  95                                            unsigned nsid);
  96
  97 static void nvme_set_queue_dying(struct nvme_ns *ns)
  98 {
  99         /*
 100          * Revalidating a dead namespace sets capacity to 0. This will end
 101          * buffered writers dirtying pages that can't be synced.
 102          */
 103         if (!ns->disk || test_and_set_bit(NVME_NS_DEAD, &ns->flags))
 104                 return;
 105         blk_set_queue_dying(ns->queue);
 106         /* Forcibly unquiesce queues to avoid blocking dispatch */
 107         blk_mq_unquiesce_queue(ns->queue);
 108         /*
 109          * Revalidate after unblocking dispatchers that may be holding bd_butex
 110          */
 111         revalidate_disk(ns->disk);
 112 }
 113
 114 static void nvme_queue_scan(struct nvme_ctrl *ctrl)
 115 {
 116         /*
 117          * Only new queue scan work when admin and IO queues are both alive
 118          */
 119         if (ctrl->state == NVME_CTRL_LIVE && ctrl->tagset)
 120                 queue_work(nvme_wq, &ctrl->scan_work);
 121 }
 122
 123 /*
 124  * Use this function to proceed with scheduling reset_work for a controller
 125  * that had previously been set to the resetting state. This is intended for
 126  * code paths that can't be interrupted by other reset attempts. A hot removal
 127  * may prevent this from succeeding.
 128  */
 129 int nvme_try_sched_reset(struct nvme_ctrl *ctrl)
 130 {
 131         if (ctrl->state != NVME_CTRL_RESETTING)
 132                 return -EBUSY;
 133         if (!queue_work(nvme_reset_wq, &ctrl->reset_work))
 134                 return -EBUSY;
 135         return 0;
 136 }
 137 EXPORT_SYMBOL_GPL(nvme_try_sched_reset);
 138
 139 int nvme_reset_ctrl(struct nvme_ctrl *ctrl)
 140 {
 141         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING))
 142                 return -EBUSY;
 143         if (!queue_work(nvme_reset_wq, &ctrl->reset_work))
 144                 return -EBUSY;
 145         return 0;
 146 }
 147 EXPORT_SYMBOL_GPL(nvme_reset_ctrl);
 148
 149 int nvme_reset_ctrl_sync(struct nvme_ctrl *ctrl)
 150 {
 151         int ret;
 152
 153         ret = nvme_reset_ctrl(ctrl);
 154         if (!ret) {
 155                 flush_work(&ctrl->reset_work);
 156                 if (ctrl->state != NVME_CTRL_LIVE)
 157                         ret = -ENETRESET;
 158         }
 159
 160         return ret;
 161 }
 162 EXPORT_SYMBOL_GPL(nvme_reset_ctrl_sync);
 163
 164 static void nvme_do_delete_ctrl(struct nvme_ctrl *ctrl)
 165 {
 166         dev_info(ctrl->device,
 167                  "Removing ctrl: NQN \"%s\"\n", ctrl->opts->subsysnqn);
 168
 169         flush_work(&ctrl->reset_work);
 170         nvme_stop_ctrl(ctrl);
 171         nvme_remove_namespaces(ctrl);
 172         ctrl->ops->delete_ctrl(ctrl);
 173         nvme_uninit_ctrl(ctrl);
 174         nvme_put_ctrl(ctrl);
 175 }
 176
 177 static void nvme_delete_ctrl_work(struct work_struct *work)
 178 {
 179         struct nvme_ctrl *ctrl =
 180                 container_of(work, struct nvme_ctrl, delete_work);
 181
 182         nvme_do_delete_ctrl(ctrl);
 183 }
 184
 185 int nvme_delete_ctrl(struct nvme_ctrl *ctrl)
 186 {
 187         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_DELETING))
 188                 return -EBUSY;
 189         if (!queue_work(nvme_delete_wq, &ctrl->delete_work))
 190                 return -EBUSY;
 191         return 0;
 192 }
 193 EXPORT_SYMBOL_GPL(nvme_delete_ctrl);
 194
 195 static int nvme_delete_ctrl_sync(struct nvme_ctrl *ctrl)
 196 {
 197         int ret = 0;
 198
 199         /*
 200          * Keep a reference until nvme_do_delete_ctrl() complete,
 201          * since ->delete_ctrl can free the controller.
 202          */
 203         nvme_get_ctrl(ctrl);
 204         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_DELETING))
 205                 ret = -EBUSY;
 206         if (!ret)
 207                 nvme_do_delete_ctrl(ctrl);
 208         nvme_put_ctrl(ctrl);
 209         return ret;
 210 }
 211
 212 static inline bool nvme_ns_has_pi(struct nvme_ns *ns)
 213 {
 214         return ns->pi_type && ns->ms == sizeof(struct t10_pi_tuple);
 215 }
 216
 217 static blk_status_t nvme_error_status(u16 status)
 218 {
 219         switch (status & 0x7ff) {
 220         case NVME_SC_SUCCESS:
 221                 return BLK_STS_OK;
 222         case NVME_SC_CAP_EXCEEDED:
 223                 return BLK_STS_NOSPC;
 224         case NVME_SC_LBA_RANGE:
 225                 return BLK_STS_TARGET;
 226         case NVME_SC_BAD_ATTRIBUTES:
 227         case NVME_SC_ONCS_NOT_SUPPORTED:
 228         case NVME_SC_INVALID_OPCODE:
 229         case NVME_SC_INVALID_FIELD:
 230         case NVME_SC_INVALID_NS:
 231                 return BLK_STS_NOTSUPP;
 232         case NVME_SC_WRITE_FAULT:
 233         case NVME_SC_READ_ERROR:
 234         case NVME_SC_UNWRITTEN_BLOCK:
 235         case NVME_SC_ACCESS_DENIED:
 236         case NVME_SC_READ_ONLY:
 237         case NVME_SC_COMPARE_FAILED:
 238                 return BLK_STS_MEDIUM;
 239         case NVME_SC_GUARD_CHECK:
 240         case NVME_SC_APPTAG_CHECK:
 241         case NVME_SC_REFTAG_CHECK:
 242         case NVME_SC_INVALID_PI:
 243                 return BLK_STS_PROTECTION;
 244         case NVME_SC_RESERVATION_CONFLICT:
 245                 return BLK_STS_NEXUS;
 246         case NVME_SC_HOST_PATH_ERROR:
 247                 return BLK_STS_TRANSPORT;
 248         default:
 249                 return BLK_STS_IOERR;
 250         }
 251 }
 252
 253 static inline bool nvme_req_needs_retry(struct request *req)
 254 {
 255         if (blk_noretry_request(req))
 256                 return false;
 257         if (nvme_req(req)->status & NVME_SC_DNR)
 258                 return false;
 259         if (nvme_req(req)->retries >= nvme_max_retries)
 260                 return false;
 261         return true;
 262 }
 263
 264 static void nvme_retry_req(struct request *req)
 265 {
 266         struct nvme_ns *ns = req->q->queuedata;
 267         unsigned long delay = 0;
 268         u16 crd;
 269
 270         /* The mask and shift result must be <= 3 */
 271         crd = (nvme_req(req)->status & NVME_SC_CRD) >> 11;
 272         if (ns && crd)
 273                 delay = ns->ctrl->crdt[crd - 1] * 100;
 274
 275         nvme_req(req)->retries++;
 276         blk_mq_requeue_request(req, false);
 277         blk_mq_delay_kick_requeue_list(req->q, delay);
 278 }
 279
 280 void nvme_complete_rq(struct request *req)
 281 {
 282         blk_status_t status = nvme_error_status(nvme_req(req)->status);
 283
 284         trace_nvme_complete_rq(req);
 285
 286         nvme_cleanup_cmd(req);
 287
 288         if (nvme_req(req)->ctrl->kas)
 289                 nvme_req(req)->ctrl->comp_seen = true;
 290
 291         if (unlikely(status != BLK_STS_OK && nvme_req_needs_retry(req))) {
 292                 if ((req->cmd_flags & REQ_NVME_MPATH) &&
 293                     blk_path_error(status)) {
 294                         nvme_failover_req(req);
 295                         return;
 296                 }
 297
 298                 if (!blk_queue_dying(req->q)) {
 299                         nvme_retry_req(req);
 300                         return;
 301                 }
 302         }
 303
 304         nvme_trace_bio_complete(req, status);
 305         blk_mq_end_request(req, status);
 306 }
 307 EXPORT_SYMBOL_GPL(nvme_complete_rq);
 308
 309 bool nvme_cancel_request(struct request *req, void *data, bool reserved)
 310 {
 311         dev_dbg_ratelimited(((struct nvme_ctrl *) data)->device,
 312                                 "Cancelling I/O %d", req->tag);
 313
 314         /* don't abort one completed request */
 315         if (blk_mq_request_completed(req))
 316                 return true;
 317
 318         nvme_req(req)->status = NVME_SC_HOST_ABORTED_CMD;
 319         blk_mq_complete_request(req);
 320         return true;
 321 }
 322 EXPORT_SYMBOL_GPL(nvme_cancel_request);
 323
 324 bool nvme_change_ctrl_state(struct nvme_ctrl *ctrl,
 325                 enum nvme_ctrl_state new_state)
 326 {
 327         enum nvme_ctrl_state old_state;
 328         unsigned long flags;
 329         bool changed = false;
 330
 331         spin_lock_irqsave(&ctrl->lock, flags);
 332
 333         old_state = ctrl->state;
 334         switch (new_state) {
 335         case NVME_CTRL_LIVE:
 336                 switch (old_state) {
 337                 case NVME_CTRL_NEW:
 338                 case NVME_CTRL_RESETTING:
 339                 case NVME_CTRL_CONNECTING:
 340                         changed = true;
 341                         /* FALLTHRU */
 342                 default:
 343                         break;
 344                 }
 345                 break;
 346         case NVME_CTRL_RESETTING:
 347                 switch (old_state) {
 348                 case NVME_CTRL_NEW:
 349                 case NVME_CTRL_LIVE:
 350                         changed = true;
 351                         /* FALLTHRU */
 352                 default:
 353                         break;
 354                 }
 355                 break;
 356         case NVME_CTRL_CONNECTING:
 357                 switch (old_state) {
 358                 case NVME_CTRL_NEW:
 359                 case NVME_CTRL_RESETTING:
 360                         changed = true;
 361                         /* FALLTHRU */
 362                 default:
 363                         break;
 364                 }
 365                 break;
 366         case NVME_CTRL_DELETING:
 367                 switch (old_state) {
 368                 case NVME_CTRL_LIVE:
 369                 case NVME_CTRL_RESETTING:
 370                 case NVME_CTRL_CONNECTING:
 371                         changed = true;
 372                         /* FALLTHRU */
 373                 default:
 374                         break;
 375                 }
 376                 break;
 377         case NVME_CTRL_DEAD:
 378                 switch (old_state) {
 379                 case NVME_CTRL_DELETING:
 380                         changed = true;
 381                         /* FALLTHRU */
 382                 default:
 383                         break;
 384                 }
 385                 break;
 386         default:
 387                 break;
 388         }
 389
 390         if (changed) {
 391                 ctrl->state = new_state;
 392                 wake_up_all(&ctrl->state_wq);
 393         }
 394
 395         spin_unlock_irqrestore(&ctrl->lock, flags);
 396         if (changed && ctrl->state == NVME_CTRL_LIVE)
 397                 nvme_kick_requeue_lists(ctrl);
 398         return changed;
 399 }
 400 EXPORT_SYMBOL_GPL(nvme_change_ctrl_state);
 401
 402 /*
 403  * Returns true for sink states that can't ever transition back to live.
 404  */
 405 static bool nvme_state_terminal(struct nvme_ctrl *ctrl)
 406 {
 407         switch (ctrl->state) {
 408         case NVME_CTRL_NEW:
 409         case NVME_CTRL_LIVE:
 410         case NVME_CTRL_RESETTING:
 411         case NVME_CTRL_CONNECTING:
 412                 return false;
 413         case NVME_CTRL_DELETING:
 414         case NVME_CTRL_DEAD:
 415                 return true;
 416         default:
 417                 WARN_ONCE(1, "Unhandled ctrl state:%d", ctrl->state);
 418                 return true;
 419         }
 420 }
 421
 422 /*
 423  * Waits for the controller state to be resetting, or returns false if it is
 424  * not possible to ever transition to that state.
 425  */
 426 bool nvme_wait_reset(struct nvme_ctrl *ctrl)
 427 {
 428         wait_event(ctrl->state_wq,
 429                    nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING) ||
 430                    nvme_state_terminal(ctrl));
 431         return ctrl->state == NVME_CTRL_RESETTING;
 432 }
 433 EXPORT_SYMBOL_GPL(nvme_wait_reset);
 434
 435 static void nvme_free_ns_head(struct kref *ref)
 436 {
 437         struct nvme_ns_head *head =
 438                 container_of(ref, struct nvme_ns_head, ref);
 439
 440         nvme_mpath_remove_disk(head);
 441         ida_simple_remove(&head->subsys->ns_ida, head->instance);
 442         list_del_init(&head->entry);
 443         cleanup_srcu_struct(&head->srcu);
 444         nvme_put_subsystem(head->subsys);
 445         kfree(head);
 446 }
 447
 448 static void nvme_put_ns_head(struct nvme_ns_head *head)
 449 {
 450         kref_put(&head->ref, nvme_free_ns_head);
 451 }
 452
 453 static void nvme_free_ns(struct kref *kref)
 454 {
 455         struct nvme_ns *ns = container_of(kref, struct nvme_ns, kref);
 456
 457         if (ns->ndev)
 458                 nvme_nvm_unregister(ns);
 459
 460         put_disk(ns->disk);
 461         nvme_put_ns_head(ns->head);
 462         nvme_put_ctrl(ns->ctrl);
 463         kfree(ns);
 464 }
 465
 466 static void nvme_put_ns(struct nvme_ns *ns)
 467 {
 468         kref_put(&ns->kref, nvme_free_ns);
 469 }
 470
 471 static inline void nvme_clear_nvme_request(struct request *req)
 472 {
 473         if (!(req->rq_flags & RQF_DONTPREP)) {
 474                 nvme_req(req)->retries = 0;
 475                 nvme_req(req)->flags = 0;
 476                 req->rq_flags |= RQF_DONTPREP;
 477         }
 478 }
 479
 480 struct request *nvme_alloc_request(struct request_queue *q,
 481                 struct nvme_command *cmd, blk_mq_req_flags_t flags, int qid)
 482 {
 483         unsigned op = nvme_is_write(cmd) ? REQ_OP_DRV_OUT : REQ_OP_DRV_IN;
 484         struct request *req;
 485
 486         if (qid == NVME_QID_ANY) {
 487                 req = blk_mq_alloc_request(q, op, flags);
 488         } else {
 489                 req = blk_mq_alloc_request_hctx(q, op, flags,
 490                                 qid ? qid - 1 : 0);
 491         }
 492         if (IS_ERR(req))
 493                 return req;
 494
 495         req->cmd_flags |= REQ_FAILFAST_DRIVER;
 496         nvme_clear_nvme_request(req);
 497         nvme_req(req)->cmd = cmd;
 498
 499         return req;
 500 }
 501 EXPORT_SYMBOL_GPL(nvme_alloc_request);
 502
 503 static int nvme_toggle_streams(struct nvme_ctrl *ctrl, bool enable)
 504 {
 505         struct nvme_command c;
 506
 507         memset(&c, 0, sizeof(c));
 508
 509         c.directive.opcode = nvme_admin_directive_send;
 510         c.directive.nsid = cpu_to_le32(NVME_NSID_ALL);
 511         c.directive.doper = NVME_DIR_SND_ID_OP_ENABLE;
 512         c.directive.dtype = NVME_DIR_IDENTIFY;
 513         c.directive.tdtype = NVME_DIR_STREAMS;
 514         c.directive.endir = enable ? NVME_DIR_ENDIR : 0;
 515
 516         return nvme_submit_sync_cmd(ctrl->admin_q, &c, NULL, 0);
 517 }
 518
 519 static int nvme_disable_streams(struct nvme_ctrl *ctrl)
 520 {
 521         return nvme_toggle_streams(ctrl, false);
 522 }
 523
 524 static int nvme_enable_streams(struct nvme_ctrl *ctrl)
 525 {
 526         return nvme_toggle_streams(ctrl, true);
 527 }
 528
 529 static int nvme_get_stream_params(struct nvme_ctrl *ctrl,
 530                                   struct streams_directive_params *s, u32 nsid)
 531 {
 532         struct nvme_command c;
 533
 534         memset(&c, 0, sizeof(c));
 535         memset(s, 0, sizeof(*s));
 536
 537         c.directive.opcode = nvme_admin_directive_recv;
 538         c.directive.nsid = cpu_to_le32(nsid);
 539         c.directive.numd = cpu_to_le32((sizeof(*s) >> 2) - 1);
 540         c.directive.doper = NVME_DIR_RCV_ST_OP_PARAM;
 541         c.directive.dtype = NVME_DIR_STREAMS;
 542
 543         return nvme_submit_sync_cmd(ctrl->admin_q, &c, s, sizeof(*s));
 544 }
 545
 546 static int nvme_configure_directives(struct nvme_ctrl *ctrl)
 547 {
 548         struct streams_directive_params s;
 549         int ret;
 550
 551         if (!(ctrl->oacs & NVME_CTRL_OACS_DIRECTIVES))
 552                 return 0;
 553         if (!streams)
 554                 return 0;
 555
 556         ret = nvme_enable_streams(ctrl);
 557         if (ret)
 558                 return ret;
 559
 560         ret = nvme_get_stream_params(ctrl, &s, NVME_NSID_ALL);
 561         if (ret)
 562                 return ret;
 563
 564         ctrl->nssa = le16_to_cpu(s.nssa);
 565         if (ctrl->nssa < BLK_MAX_WRITE_HINTS - 1) {
 566                 dev_info(ctrl->device, "too few streams (%u) available\n",
 567                                         ctrl->nssa);
 568                 nvme_disable_streams(ctrl);
 569                 return 0;
 570         }
 571
 572         ctrl->nr_streams = min_t(unsigned, ctrl->nssa, BLK_MAX_WRITE_HINTS - 1);
 573         dev_info(ctrl->device, "Using %u streams\n", ctrl->nr_streams);
 574         return 0;
 575 }
 576
 577 /*
 578  * Check if 'req' has a write hint associated with it. If it does, assign
 579  * a valid namespace stream to the write.
 580  */
 581 static void nvme_assign_write_stream(struct nvme_ctrl *ctrl,
 582                                      struct request *req, u16 *control,
 583                                      u32 *dsmgmt)
 584 {
 585         enum rw_hint streamid = req->write_hint;
 586
 587         if (streamid == WRITE_LIFE_NOT_SET || streamid == WRITE_LIFE_NONE)
 588                 streamid = 0;
 589         else {
 590                 streamid--;
 591                 if (WARN_ON_ONCE(streamid > ctrl->nr_streams))
 592                         return;
 593
 594                 *control |= NVME_RW_DTYPE_STREAMS;
 595                 *dsmgmt |= streamid << 16;
 596         }
 597
 598         if (streamid < ARRAY_SIZE(req->q->write_hints))
 599                 req->q->write_hints[streamid] += blk_rq_bytes(req) >> 9;
 600 }
 601
 602 static inline void nvme_setup_flush(struct nvme_ns *ns,
 603                 struct nvme_command *cmnd)
 604 {
 605         cmnd->common.opcode = nvme_cmd_flush;
 606         cmnd->common.nsid = cpu_to_le32(ns->head->ns_id);
 607 }
 608
 609 static blk_status_t nvme_setup_discard(struct nvme_ns *ns, struct request *req,
 610                 struct nvme_command *cmnd)
 611 {
 612         unsigned short segments = blk_rq_nr_discard_segments(req), n = 0;
 613         struct nvme_dsm_range *range;
 614         struct bio *bio;
 615
 616         /*
 617          * Some devices do not consider the DSM 'Number of Ranges' field when
 618          * determining how much data to DMA. Always allocate memory for maximum
 619          * number of segments to prevent device reading beyond end of buffer.
 620          */
 621         static const size_t alloc_size = sizeof(*range) * NVME_DSM_MAX_RANGES;
 622
 623         range = kzalloc(alloc_size, GFP_ATOMIC | __GFP_NOWARN);
 624         if (!range) {
 625                 /*
 626                  * If we fail allocation our range, fallback to the controller
 627                  * discard page. If that's also busy, it's safe to return
 628                  * busy, as we know we can make progress once that's freed.
 629                  */
 630                 if (test_and_set_bit_lock(0, &ns->ctrl->discard_page_busy))
 631                         return BLK_STS_RESOURCE;
 632
 633                 range = page_address(ns->ctrl->discard_page);
 634         }
 635
 636         __rq_for_each_bio(bio, req) {
 637                 u64 slba = nvme_sect_to_lba(ns, bio->bi_iter.bi_sector);
 638                 u32 nlb = bio->bi_iter.bi_size >> ns->lba_shift;
 639
 640                 if (n < segments) {
 641                         range[n].cattr = cpu_to_le32(0);
 642                         range[n].nlb = cpu_to_le32(nlb);
 643                         range[n].slba = cpu_to_le64(slba);
 644                 }
 645                 n++;
 646         }
 647
 648         if (WARN_ON_ONCE(n != segments)) {
 649                 if (virt_to_page(range) == ns->ctrl->discard_page)
 650                         clear_bit_unlock(0, &ns->ctrl->discard_page_busy);
 651                 else
 652                         kfree(range);
 653                 return BLK_STS_IOERR;
 654         }
 655
 656         cmnd->dsm.opcode = nvme_cmd_dsm;
 657         cmnd->dsm.nsid = cpu_to_le32(ns->head->ns_id);
 658         cmnd->dsm.nr = cpu_to_le32(segments - 1);
 659         cmnd->dsm.attributes = cpu_to_le32(NVME_DSMGMT_AD);
 660
 661         req->special_vec.bv_page = virt_to_page(range);
 662         req->special_vec.bv_offset = offset_in_page(range);
 663         req->special_vec.bv_len = alloc_size;
 664         req->rq_flags |= RQF_SPECIAL_PAYLOAD;
 665
 666         return BLK_STS_OK;
 667 }
 668
 669 static inline blk_status_t nvme_setup_write_zeroes(struct nvme_ns *ns,
 670                 struct request *req, struct nvme_command *cmnd)
 671 {
 672         if (ns->ctrl->quirks & NVME_QUIRK_DEALLOCATE_ZEROES)
 673                 return nvme_setup_discard(ns, req, cmnd);
 674
 675         cmnd->write_zeroes.opcode = nvme_cmd_write_zeroes;
 676         cmnd->write_zeroes.nsid = cpu_to_le32(ns->head->ns_id);
 677         cmnd->write_zeroes.slba =
 678                 cpu_to_le64(nvme_sect_to_lba(ns, blk_rq_pos(req)));
 679         cmnd->write_zeroes.length =
 680                 cpu_to_le16((blk_rq_bytes(req) >> ns->lba_shift) - 1);
 681         cmnd->write_zeroes.control = 0;
 682         return BLK_STS_OK;
 683 }
 684
 685 static inline blk_status_t nvme_setup_rw(struct nvme_ns *ns,
 686                 struct request *req, struct nvme_command *cmnd)
 687 {
 688         struct nvme_ctrl *ctrl = ns->ctrl;
 689         u16 control = 0;
 690         u32 dsmgmt = 0;
 691
 692         if (req->cmd_flags & REQ_FUA)
 693                 control |= NVME_RW_FUA;
 694         if (req->cmd_flags & (REQ_FAILFAST_DEV | REQ_RAHEAD))
 695                 control |= NVME_RW_LR;
 696
 697         if (req->cmd_flags & REQ_RAHEAD)
 698                 dsmgmt |= NVME_RW_DSM_FREQ_PREFETCH;
 699
 700         cmnd->rw.opcode = (rq_data_dir(req) ? nvme_cmd_write : nvme_cmd_read);
 701         cmnd->rw.nsid = cpu_to_le32(ns->head->ns_id);
 702         cmnd->rw.slba = cpu_to_le64(nvme_sect_to_lba(ns, blk_rq_pos(req)));
 703         cmnd->rw.length = cpu_to_le16((blk_rq_bytes(req) >> ns->lba_shift) - 1);
 704
 705         if (req_op(req) == REQ_OP_WRITE && ctrl->nr_streams)
 706                 nvme_assign_write_stream(ctrl, req, &control, &dsmgmt);
 707
 708         if (ns->ms) {
 709                 /*
 710                  * If formated with metadata, the block layer always provides a
 711                  * metadata buffer if CONFIG_BLK_DEV_INTEGRITY is enabled.  Else
 712                  * we enable the PRACT bit for protection information or set the
 713                  * namespace capacity to zero to prevent any I/O.
 714                  */
 715                 if (!blk_integrity_rq(req)) {
 716                         if (WARN_ON_ONCE(!nvme_ns_has_pi(ns)))
 717                                 return BLK_STS_NOTSUPP;
 718                         control |= NVME_RW_PRINFO_PRACT;
 719                 }
 720
 721                 switch (ns->pi_type) {
 722                 case NVME_NS_DPS_PI_TYPE3:
 723                         control |= NVME_RW_PRINFO_PRCHK_GUARD;
 724                         break;
 725                 case NVME_NS_DPS_PI_TYPE1:
 726                 case NVME_NS_DPS_PI_TYPE2:
 727                         control |= NVME_RW_PRINFO_PRCHK_GUARD |
 728                                         NVME_RW_PRINFO_PRCHK_REF;
 729                         cmnd->rw.reftag = cpu_to_le32(t10_pi_ref_tag(req));
 730                         break;
 731                 }
 732         }
 733
 734         cmnd->rw.control = cpu_to_le16(control);
 735         cmnd->rw.dsmgmt = cpu_to_le32(dsmgmt);
 736         return 0;
 737 }
 738
 739 void nvme_cleanup_cmd(struct request *req)
 740 {
 741         if (req->rq_flags & RQF_SPECIAL_PAYLOAD) {
 742                 struct nvme_ns *ns = req->rq_disk->private_data;
 743                 struct page *page = req->special_vec.bv_page;
 744
 745                 if (page == ns->ctrl->discard_page)
 746                         clear_bit_unlock(0, &ns->ctrl->discard_page_busy);
 747                 else
 748                         kfree(page_address(page) + req->special_vec.bv_offset);
 749         }
 750 }
 751 EXPORT_SYMBOL_GPL(nvme_cleanup_cmd);
 752
 753 blk_status_t nvme_setup_cmd(struct nvme_ns *ns, struct request *req,
 754                 struct nvme_command *cmd)
 755 {
 756         blk_status_t ret = BLK_STS_OK;
 757
 758         nvme_clear_nvme_request(req);
 759
 760         memset(cmd, 0, sizeof(*cmd));
 761         switch (req_op(req)) {
 762         case REQ_OP_DRV_IN:
 763         case REQ_OP_DRV_OUT:
 764                 memcpy(cmd, nvme_req(req)->cmd, sizeof(*cmd));
 765                 break;
 766         case REQ_OP_FLUSH:
 767                 nvme_setup_flush(ns, cmd);
 768                 break;
 769         case REQ_OP_WRITE_ZEROES:
 770                 ret = nvme_setup_write_zeroes(ns, req, cmd);
 771                 break;
 772         case REQ_OP_DISCARD:
 773                 ret = nvme_setup_discard(ns, req, cmd);
 774                 break;
 775         case REQ_OP_READ:
 776         case REQ_OP_WRITE:
 777                 ret = nvme_setup_rw(ns, req, cmd);
 778                 break;
 779         default:
 780                 WARN_ON_ONCE(1);
 781                 return BLK_STS_IOERR;
 782         }
 783
 784         cmd->common.command_id = req->tag;
 785         trace_nvme_setup_cmd(req, cmd);
 786         return ret;
 787 }
 788 EXPORT_SYMBOL_GPL(nvme_setup_cmd);
 789
 790 static void nvme_end_sync_rq(struct request *rq, blk_status_t error)
 791 {
 792         struct completion *waiting = rq->end_io_data;
 793
 794         rq->end_io_data = NULL;
 795         complete(waiting);
 796 }
 797
 798 static void nvme_execute_rq_polled(struct request_queue *q,
 799                 struct gendisk *bd_disk, struct request *rq, int at_head)
 800 {
 801         DECLARE_COMPLETION_ONSTACK(wait);
 802
 803         WARN_ON_ONCE(!test_bit(QUEUE_FLAG_POLL, &q->queue_flags));
 804
 805         rq->cmd_flags |= REQ_HIPRI;
 806         rq->end_io_data = &wait;
 807         blk_execute_rq_nowait(q, bd_disk, rq, at_head, nvme_end_sync_rq);
 808
 809         while (!completion_done(&wait)) {
 810                 blk_poll(q, request_to_qc_t(rq->mq_hctx, rq), true);
 811                 cond_resched();
 812         }
 813 }
 814
 815 /*
 816  * Returns 0 on success.  If the result is negative, it's a Linux error code;
 817  * if the result is positive, it's an NVM Express status code
 818  */
 819 int __nvme_submit_sync_cmd(struct request_queue *q, struct nvme_command *cmd,
 820                 union nvme_result *result, void *buffer, unsigned bufflen,
 821                 unsigned timeout, int qid, int at_head,
 822                 blk_mq_req_flags_t flags, bool poll)
 823 {
 824         struct request *req;
 825         int ret;
 826
 827         req = nvme_alloc_request(q, cmd, flags, qid);
 828         if (IS_ERR(req))
 829                 return PTR_ERR(req);
 830
 831         req->timeout = timeout ? timeout : ADMIN_TIMEOUT;
 832
 833         if (buffer && bufflen) {
 834                 ret = blk_rq_map_kern(q, req, buffer, bufflen, GFP_KERNEL);
 835                 if (ret)
 836                         goto out;
 837         }
 838
 839         if (poll)
 840                 nvme_execute_rq_polled(req->q, NULL, req, at_head);
 841         else
 842                 blk_execute_rq(req->q, NULL, req, at_head);
 843         if (result)
 844                 *result = nvme_req(req)->result;
 845         if (nvme_req(req)->flags & NVME_REQ_CANCELLED)
 846                 ret = -EINTR;
 847         else
 848                 ret = nvme_req(req)->status;
 849  out:
 850         blk_mq_free_request(req);
 851         return ret;
 852 }
 853 EXPORT_SYMBOL_GPL(__nvme_submit_sync_cmd);
 854
 855 int nvme_submit_sync_cmd(struct request_queue *q, struct nvme_command *cmd,
 856                 void *buffer, unsigned bufflen)
 857 {
 858         return __nvme_submit_sync_cmd(q, cmd, NULL, buffer, bufflen, 0,
 859                         NVME_QID_ANY, 0, 0, false);
 860 }
 861 EXPORT_SYMBOL_GPL(nvme_submit_sync_cmd);
 862
 863 static void *nvme_add_user_metadata(struct bio *bio, void __user *ubuf,
 864                 unsigned len, u32 seed, bool write)
 865 {
 866         struct bio_integrity_payload *bip;
 867         int ret = -ENOMEM;
 868         void *buf;
 869
 870         buf = kmalloc(len, GFP_KERNEL);
 871         if (!buf)
 872                 goto out;
 873
 874         ret = -EFAULT;
 875         if (write && copy_from_user(buf, ubuf, len))
 876                 goto out_free_meta;
 877
 878         bip = bio_integrity_alloc(bio, GFP_KERNEL, 1);
 879         if (IS_ERR(bip)) {
 880                 ret = PTR_ERR(bip);
 881                 goto out_free_meta;
 882         }
 883
 884         bip->bip_iter.bi_size = len;
 885         bip->bip_iter.bi_sector = seed;
 886         ret = bio_integrity_add_page(bio, virt_to_page(buf), len,
 887                         offset_in_page(buf));
 888         if (ret == len)
 889                 return buf;
 890         ret = -ENOMEM;
 891 out_free_meta:
 892         kfree(buf);
 893 out:
 894         return ERR_PTR(ret);
 895 }
 896
 897 static int nvme_submit_user_cmd(struct request_queue *q,
 898                 struct nvme_command *cmd, void __user *ubuffer,
 899                 unsigned bufflen, void __user *meta_buffer, unsigned meta_len,
 900                 u32 meta_seed, u64 *result, unsigned timeout)
 901 {
 902         bool write = nvme_is_write(cmd);
 903         struct nvme_ns *ns = q->queuedata;
 904         struct gendisk *disk = ns ? ns->disk : NULL;
 905         struct request *req;
 906         struct bio *bio = NULL;
 907         void *meta = NULL;
 908         int ret;
 909
 910         req = nvme_alloc_request(q, cmd, 0, NVME_QID_ANY);
 911         if (IS_ERR(req))
 912                 return PTR_ERR(req);
 913
 914         req->timeout = timeout ? timeout : ADMIN_TIMEOUT;
 915         nvme_req(req)->flags |= NVME_REQ_USERCMD;
 916
 917         if (ubuffer && bufflen) {
 918                 ret = blk_rq_map_user(q, req, NULL, ubuffer, bufflen,
 919                                 GFP_KERNEL);
 920                 if (ret)
 921                         goto out;
 922                 bio = req->bio;
 923                 bio->bi_disk = disk;
 924                 if (disk && meta_buffer && meta_len) {
 925                         meta = nvme_add_user_metadata(bio, meta_buffer, meta_len,
 926                                         meta_seed, write);
 927                         if (IS_ERR(meta)) {
 928                                 ret = PTR_ERR(meta);
 929                                 goto out_unmap;
 930                         }
 931                         req->cmd_flags |= REQ_INTEGRITY;
 932                 }
 933         }
 934
 935         blk_execute_rq(req->q, disk, req, 0);
 936         if (nvme_req(req)->flags & NVME_REQ_CANCELLED)
 937                 ret = -EINTR;
 938         else
 939                 ret = nvme_req(req)->status;
 940         if (result)
 941                 *result = le64_to_cpu(nvme_req(req)->result.u64);
 942         if (meta && !ret && !write) {
 943                 if (copy_to_user(meta_buffer, meta, meta_len))
 944                         ret = -EFAULT;
 945         }
 946         kfree(meta);
 947  out_unmap:
 948         if (bio)
 949                 blk_rq_unmap_user(bio);
 950  out:
 951         blk_mq_free_request(req);
 952         return ret;
 953 }
 954
 955 static void nvme_keep_alive_end_io(struct request *rq, blk_status_t status)
 956 {
 957         struct nvme_ctrl *ctrl = rq->end_io_data;
 958         unsigned long flags;
 959         bool startka = false;
 960
 961         blk_mq_free_request(rq);
 962
 963         if (status) {
 964                 dev_err(ctrl->device,
 965                         "failed nvme_keep_alive_end_io error=%d\n",
 966                                 status);
 967                 return;
 968         }
 969
 970         ctrl->comp_seen = false;
 971         spin_lock_irqsave(&ctrl->lock, flags);
 972         if (ctrl->state == NVME_CTRL_LIVE ||
 973             ctrl->state == NVME_CTRL_CONNECTING)
 974                 startka = true;
 975         spin_unlock_irqrestore(&ctrl->lock, flags);
 976         if (startka)
 977                 schedule_delayed_work(&ctrl->ka_work, ctrl->kato * HZ);
 978 }
 979
 980 static int nvme_keep_alive(struct nvme_ctrl *ctrl)
 981 {
 982         struct request *rq;
 983
 984         rq = nvme_alloc_request(ctrl->admin_q, &ctrl->ka_cmd, BLK_MQ_REQ_RESERVED,
 985                         NVME_QID_ANY);
 986         if (IS_ERR(rq))
 987                 return PTR_ERR(rq);
 988
 989         rq->timeout = ctrl->kato * HZ;
 990         rq->end_io_data = ctrl;
 991
 992         blk_execute_rq_nowait(rq->q, NULL, rq, 0, nvme_keep_alive_end_io);
 993
 994         return 0;
 995 }
 996
 997 static void nvme_keep_alive_work(struct work_struct *work)
 998 {
 999         struct nvme_ctrl *ctrl = container_of(to_delayed_work(work),
1000                         struct nvme_ctrl, ka_work);
1001         bool comp_seen = ctrl->comp_seen;
1002
1003         if ((ctrl->ctratt & NVME_CTRL_ATTR_TBKAS) && comp_seen) {
1004                 dev_dbg(ctrl->device,
1005                         "reschedule traffic based keep-alive timer\n");
1006                 ctrl->comp_seen = false;
1007                 schedule_delayed_work(&ctrl->ka_work, ctrl->kato * HZ);
1008                 return;
1009         }
1010
1011         if (nvme_keep_alive(ctrl)) {
1012                 /* allocation failure, reset the controller */
1013                 dev_err(ctrl->device, "keep-alive failed\n");
1014                 nvme_reset_ctrl(ctrl);
1015                 return;
1016         }
1017 }
1018
1019 static void nvme_start_keep_alive(struct nvme_ctrl *ctrl)
1020 {
1021         if (unlikely(ctrl->kato == 0))
1022                 return;
1023
1024         schedule_delayed_work(&ctrl->ka_work, ctrl->kato * HZ);
1025 }
1026
1027 void nvme_stop_keep_alive(struct nvme_ctrl *ctrl)
1028 {
1029         if (unlikely(ctrl->kato == 0))
1030                 return;
1031
1032         cancel_delayed_work_sync(&ctrl->ka_work);
1033 }
1034 EXPORT_SYMBOL_GPL(nvme_stop_keep_alive);
1035
1036 static int nvme_identify_ctrl(struct nvme_ctrl *dev, struct nvme_id_ctrl **id)
1037 {
1038         struct nvme_command c = { };
1039         int error;
1040
1041         /* gcc-4.4.4 (at least) has issues with initializers and anon unions */
1042         c.identify.opcode = nvme_admin_identify;
1043         c.identify.cns = NVME_ID_CNS_CTRL;
1044
1045         *id = kmalloc(sizeof(struct nvme_id_ctrl), GFP_KERNEL);
1046         if (!*id)
1047                 return -ENOMEM;
1048
1049         error = nvme_submit_sync_cmd(dev->admin_q, &c, *id,
1050                         sizeof(struct nvme_id_ctrl));
1051         if (error)
1052                 kfree(*id);
1053         return error;
1054 }
1055
1056 static int nvme_identify_ns_descs(struct nvme_ctrl *ctrl, unsigned nsid,
1057                 struct nvme_ns_ids *ids)
1058 {
1059         struct nvme_command c = { };
1060         int status;
1061         void *data;
1062         int pos;
1063         int len;
1064
1065         c.identify.opcode = nvme_admin_identify;
1066         c.identify.nsid = cpu_to_le32(nsid);
1067         c.identify.cns = NVME_ID_CNS_NS_DESC_LIST;
1068
1069         data = kzalloc(NVME_IDENTIFY_DATA_SIZE, GFP_KERNEL);
1070         if (!data)
1071                 return -ENOMEM;
1072
1073         status = nvme_submit_sync_cmd(ctrl->admin_q, &c, data,
1074                                       NVME_IDENTIFY_DATA_SIZE);
1075         if (status)
1076                 goto free_data;
1077
1078         for (pos = 0; pos < NVME_IDENTIFY_DATA_SIZE; pos += len) {
1079                 struct nvme_ns_id_desc *cur = data + pos;
1080
1081                 if (cur->nidl == 0)
1082                         break;
1083
1084                 switch (cur->nidt) {
1085                 case NVME_NIDT_EUI64:
1086                         if (cur->nidl != NVME_NIDT_EUI64_LEN) {
1087                                 dev_warn(ctrl->device,
1088                                          "ctrl returned bogus length: %d for NVME_NIDT_EUI64\n",
1089                                          cur->nidl);
1090                                 goto free_data;
1091                         }
1092                         len = NVME_NIDT_EUI64_LEN;
1093                         memcpy(ids->eui64, data + pos + sizeof(*cur), len);
1094                         break;
1095                 case NVME_NIDT_NGUID:
1096                         if (cur->nidl != NVME_NIDT_NGUID_LEN) {
1097                                 dev_warn(ctrl->device,
1098                                          "ctrl returned bogus length: %d for NVME_NIDT_NGUID\n",
1099                                          cur->nidl);
1100                                 goto free_data;
1101                         }
1102                         len = NVME_NIDT_NGUID_LEN;
1103                         memcpy(ids->nguid, data + pos + sizeof(*cur), len);
1104                         break;
1105                 case NVME_NIDT_UUID:
1106                         if (cur->nidl != NVME_NIDT_UUID_LEN) {
1107                                 dev_warn(ctrl->device,
1108                                          "ctrl returned bogus length: %d for NVME_NIDT_UUID\n",
1109                                          cur->nidl);
1110                                 goto free_data;
1111                         }
1112                         len = NVME_NIDT_UUID_LEN;
1113                         uuid_copy(&ids->uuid, data + pos + sizeof(*cur));
1114                         break;
1115                 default:
1116                         /* Skip unknown types */
1117                         len = cur->nidl;
1118                         break;
1119                 }
1120
1121                 len += sizeof(*cur);
1122         }
1123 free_data:
1124         kfree(data);
1125         return status;
1126 }
1127
1128 static int nvme_identify_ns_list(struct nvme_ctrl *dev, unsigned nsid, __le32 *ns_list)
1129 {
1130         struct nvme_command c = { };
1131
1132         c.identify.opcode = nvme_admin_identify;
1133         c.identify.cns = NVME_ID_CNS_NS_ACTIVE_LIST;
1134         c.identify.nsid = cpu_to_le32(nsid);
1135         return nvme_submit_sync_cmd(dev->admin_q, &c, ns_list,
1136                                     NVME_IDENTIFY_DATA_SIZE);
1137 }
1138
1139 static int nvme_identify_ns(struct nvme_ctrl *ctrl,
1140                 unsigned nsid, struct nvme_id_ns **id)
1141 {
1142         struct nvme_command c = { };
1143         int error;
1144
1145         /* gcc-4.4.4 (at least) has issues with initializers and anon unions */
1146         c.identify.opcode = nvme_admin_identify;
1147         c.identify.nsid = cpu_to_le32(nsid);
1148         c.identify.cns = NVME_ID_CNS_NS;
1149
1150         *id = kmalloc(sizeof(**id), GFP_KERNEL);
1151         if (!*id)
1152                 return -ENOMEM;
1153
1154         error = nvme_submit_sync_cmd(ctrl->admin_q, &c, *id, sizeof(**id));
1155         if (error) {
1156                 dev_warn(ctrl->device, "Identify namespace failed (%d)\n", error);
1157                 kfree(*id);
1158         }
1159
1160         return error;
1161 }
1162
1163 static int nvme_features(struct nvme_ctrl *dev, u8 op, unsigned int fid,
1164                 unsigned int dword11, void *buffer, size_t buflen, u32 *result)
1165 {
1166         struct nvme_command c;
1167         union nvme_result res;
1168         int ret;
1169
1170         memset(&c, 0, sizeof(c));
1171         c.features.opcode = op;
1172         c.features.fid = cpu_to_le32(fid);
1173         c.features.dword11 = cpu_to_le32(dword11);
1174
1175         ret = __nvme_submit_sync_cmd(dev->admin_q, &c, &res,
1176                         buffer, buflen, 0, NVME_QID_ANY, 0, 0, false);
1177         if (ret >= 0 && result)
1178                 *result = le32_to_cpu(res.u32);
1179         return ret;
1180 }
1181
1182 int nvme_set_features(struct nvme_ctrl *dev, unsigned int fid,
1183                       unsigned int dword11, void *buffer, size_t buflen,
1184                       u32 *result)
1185 {
1186         return nvme_features(dev, nvme_admin_set_features, fid, dword11, buffer,
1187                              buflen, result);
1188 }
1189 EXPORT_SYMBOL_GPL(nvme_set_features);
1190
1191 int nvme_get_features(struct nvme_ctrl *dev, unsigned int fid,
1192                       unsigned int dword11, void *buffer, size_t buflen,
1193                       u32 *result)
1194 {
1195         return nvme_features(dev, nvme_admin_get_features, fid, dword11, buffer,
1196                              buflen, result);
1197 }
1198 EXPORT_SYMBOL_GPL(nvme_get_features);
1199
1200 int nvme_set_queue_count(struct nvme_ctrl *ctrl, int *count)
1201 {
1202         u32 q_count = (*count - 1) | ((*count - 1) << 16);
1203         u32 result;
1204         int status, nr_io_queues;
1205
1206         status = nvme_set_features(ctrl, NVME_FEAT_NUM_QUEUES, q_count, NULL, 0,
1207                         &result);
1208         if (status < 0)
1209                 return status;
1210
1211         /*
1212          * Degraded controllers might return an error when setting the queue
1213          * count.  We still want to be able to bring them online and offer
1214          * access to the admin queue, as that might be only way to fix them up.
1215          */
1216         if (status > 0) {
1217                 dev_err(ctrl->device, "Could not set queue count (%d)\n", status);
1218                 *count = 0;
1219         } else {
1220                 nr_io_queues = min(result & 0xffff, result >> 16) + 1;
1221                 *count = min(*count, nr_io_queues);
1222         }
1223
1224         return 0;
1225 }
1226 EXPORT_SYMBOL_GPL(nvme_set_queue_count);
1227
1228 #define NVME_AEN_SUPPORTED \
1229         (NVME_AEN_CFG_NS_ATTR | NVME_AEN_CFG_FW_ACT | \
1230          NVME_AEN_CFG_ANA_CHANGE | NVME_AEN_CFG_DISC_CHANGE)
1231
1232 static void nvme_enable_aen(struct nvme_ctrl *ctrl)
1233 {
1234         u32 result, supported_aens = ctrl->oaes & NVME_AEN_SUPPORTED;
1235         int status;
1236
1237         if (!supported_aens)
1238                 return;
1239
1240         status = nvme_set_features(ctrl, NVME_FEAT_ASYNC_EVENT, supported_aens,
1241                         NULL, 0, &result);
1242         if (status)
1243                 dev_warn(ctrl->device, "Failed to configure AEN (cfg %x)\n",
1244                          supported_aens);
1245
1246         queue_work(nvme_wq, &ctrl->async_event_work);
1247 }
1248
1249 static int nvme_submit_io(struct nvme_ns *ns, struct nvme_user_io __user *uio)
1250 {
1251         struct nvme_user_io io;
1252         struct nvme_command c;
1253         unsigned length, meta_len;
1254         void __user *metadata;
1255
1256         if (copy_from_user(&io, uio, sizeof(io)))
1257                 return -EFAULT;
1258         if (io.flags)
1259                 return -EINVAL;
1260
1261         switch (io.opcode) {
1262         case nvme_cmd_write:
1263         case nvme_cmd_read:
1264         case nvme_cmd_compare:
1265                 break;
1266         default:
1267                 return -EINVAL;
1268         }
1269
1270         length = (io.nblocks + 1) << ns->lba_shift;
1271         meta_len = (io.nblocks + 1) * ns->ms;
1272         metadata = (void __user *)(uintptr_t)io.metadata;
1273
1274         if (ns->ext) {
1275                 length += meta_len;
1276                 meta_len = 0;
1277         } else if (meta_len) {
1278                 if ((io.metadata & 3) || !io.metadata)
1279                         return -EINVAL;
1280         }
1281
1282         memset(&c, 0, sizeof(c));
1283         c.rw.opcode = io.opcode;
1284         c.rw.flags = io.flags;
1285         c.rw.nsid = cpu_to_le32(ns->head->ns_id);
1286         c.rw.slba = cpu_to_le64(io.slba);
1287         c.rw.length = cpu_to_le16(io.nblocks);
1288         c.rw.control = cpu_to_le16(io.control);
1289         c.rw.dsmgmt = cpu_to_le32(io.dsmgmt);
1290         c.rw.reftag = cpu_to_le32(io.reftag);
1291         c.rw.apptag = cpu_to_le16(io.apptag);
1292         c.rw.appmask = cpu_to_le16(io.appmask);
1293
1294         return nvme_submit_user_cmd(ns->queue, &c,
1295                         (void __user *)(uintptr_t)io.addr, length,
1296                         metadata, meta_len, lower_32_bits(io.slba), NULL, 0);
1297 }
1298
1299 static u32 nvme_known_admin_effects(u8 opcode)
1300 {
1301         switch (opcode) {
1302         case nvme_admin_format_nvm:
1303                 return NVME_CMD_EFFECTS_CSUPP | NVME_CMD_EFFECTS_LBCC |
1304                                         NVME_CMD_EFFECTS_CSE_MASK;
1305         case nvme_admin_sanitize_nvm:
1306                 return NVME_CMD_EFFECTS_CSE_MASK;
1307         default:
1308                 break;
1309         }
1310         return 0;
1311 }
1312
1313 static u32 nvme_passthru_start(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
1314                                                                 u8 opcode)
1315 {
1316         u32 effects = 0;
1317
1318         if (ns) {
1319                 if (ctrl->effects)
1320                         effects = le32_to_cpu(ctrl->effects->iocs[opcode]);
1321                 if (effects & ~(NVME_CMD_EFFECTS_CSUPP | NVME_CMD_EFFECTS_LBCC))
1322                         dev_warn(ctrl->device,
1323                                  "IO command:%02x has unhandled effects:%08x\n",
1324                                  opcode, effects);
1325                 return 0;
1326         }
1327
1328         if (ctrl->effects)
1329                 effects = le32_to_cpu(ctrl->effects->acs[opcode]);
1330         effects |= nvme_known_admin_effects(opcode);
1331
1332         /*
1333          * For simplicity, IO to all namespaces is quiesced even if the command
1334          * effects say only one namespace is affected.
1335          */
1336         if (effects & (NVME_CMD_EFFECTS_LBCC | NVME_CMD_EFFECTS_CSE_MASK)) {
1337                 mutex_lock(&ctrl->scan_lock);
1338                 mutex_lock(&ctrl->subsys->lock);
1339                 nvme_mpath_start_freeze(ctrl->subsys);
1340                 nvme_mpath_wait_freeze(ctrl->subsys);
1341                 nvme_start_freeze(ctrl);
1342                 nvme_wait_freeze(ctrl);
1343         }
1344         return effects;
1345 }
1346
1347 static void nvme_update_formats(struct nvme_ctrl *ctrl)
1348 {
1349         struct nvme_ns *ns;
1350
1351         down_read(&ctrl->namespaces_rwsem);
1352         list_for_each_entry(ns, &ctrl->namespaces, list)
1353                 if (ns->disk && nvme_revalidate_disk(ns->disk))
1354                         nvme_set_queue_dying(ns);
1355         up_read(&ctrl->namespaces_rwsem);
1356 }
1357
1358 static void nvme_passthru_end(struct nvme_ctrl *ctrl, u32 effects)
1359 {
1360         /*
1361          * Revalidate LBA changes prior to unfreezing. This is necessary to
1362          * prevent memory corruption if a logical block size was changed by
1363          * this command.
1364          */
1365         if (effects & NVME_CMD_EFFECTS_LBCC)
1366                 nvme_update_formats(ctrl);
1367         if (effects & (NVME_CMD_EFFECTS_LBCC | NVME_CMD_EFFECTS_CSE_MASK)) {
1368                 nvme_unfreeze(ctrl);
1369                 nvme_mpath_unfreeze(ctrl->subsys);
1370                 mutex_unlock(&ctrl->subsys->lock);
1371                 nvme_remove_invalid_namespaces(ctrl, NVME_NSID_ALL);
1372                 mutex_unlock(&ctrl->scan_lock);
1373         }
1374         if (effects & NVME_CMD_EFFECTS_CCC)
1375                 nvme_init_identify(ctrl);
1376         if (effects & (NVME_CMD_EFFECTS_NIC | NVME_CMD_EFFECTS_NCC))
1377                 nvme_queue_scan(ctrl);
1378 }
1379
1380 static int nvme_user_cmd(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
1381                         struct nvme_passthru_cmd __user *ucmd)
1382 {
1383         struct nvme_passthru_cmd cmd;
1384         struct nvme_command c;
1385         unsigned timeout = 0;
1386         u32 effects;
1387         u64 result;
1388         int status;
1389
1390         if (!capable(CAP_SYS_ADMIN))
1391                 return -EACCES;
1392         if (copy_from_user(&cmd, ucmd, sizeof(cmd)))
1393                 return -EFAULT;
1394         if (cmd.flags)
1395                 return -EINVAL;
1396
1397         memset(&c, 0, sizeof(c));
1398         c.common.opcode = cmd.opcode;
1399         c.common.flags = cmd.flags;
1400         c.common.nsid = cpu_to_le32(cmd.nsid);
1401         c.common.cdw2[0] = cpu_to_le32(cmd.cdw2);
1402         c.common.cdw2[1] = cpu_to_le32(cmd.cdw3);
1403         c.common.cdw10 = cpu_to_le32(cmd.cdw10);
1404         c.common.cdw11 = cpu_to_le32(cmd.cdw11);
1405         c.common.cdw12 = cpu_to_le32(cmd.cdw12);
1406         c.common.cdw13 = cpu_to_le32(cmd.cdw13);
1407         c.common.cdw14 = cpu_to_le32(cmd.cdw14);
1408         c.common.cdw15 = cpu_to_le32(cmd.cdw15);
1409
1410         if (cmd.timeout_ms)
1411                 timeout = msecs_to_jiffies(cmd.timeout_ms);
1412
1413         effects = nvme_passthru_start(ctrl, ns, cmd.opcode);
1414         status = nvme_submit_user_cmd(ns ? ns->queue : ctrl->admin_q, &c,
1415                         (void __user *)(uintptr_t)cmd.addr, cmd.data_len,
1416                         (void __user *)(uintptr_t)cmd.metadata,
1417                         cmd.metadata_len, 0, &result, timeout);
1418         nvme_passthru_end(ctrl, effects);
1419
1420         if (status >= 0) {
1421                 if (put_user(result, &ucmd->result))
1422                         return -EFAULT;
1423         }
1424
1425         return status;
1426 }
1427
1428 static int nvme_user_cmd64(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
1429                         struct nvme_passthru_cmd64 __user *ucmd)
1430 {
1431         struct nvme_passthru_cmd64 cmd;
1432         struct nvme_command c;
1433         unsigned timeout = 0;
1434         u32 effects;
1435         int status;
1436
1437         if (!capable(CAP_SYS_ADMIN))
1438                 return -EACCES;
1439         if (copy_from_user(&cmd, ucmd, sizeof(cmd)))
1440                 return -EFAULT;
1441         if (cmd.flags)
1442                 return -EINVAL;
1443
1444         memset(&c, 0, sizeof(c));
1445         c.common.opcode = cmd.opcode;
1446         c.common.flags = cmd.flags;
1447         c.common.nsid = cpu_to_le32(cmd.nsid);
1448         c.common.cdw2[0] = cpu_to_le32(cmd.cdw2);
1449         c.common.cdw2[1] = cpu_to_le32(cmd.cdw3);
1450         c.common.cdw10 = cpu_to_le32(cmd.cdw10);
1451         c.common.cdw11 = cpu_to_le32(cmd.cdw11);
1452         c.common.cdw12 = cpu_to_le32(cmd.cdw12);
1453         c.common.cdw13 = cpu_to_le32(cmd.cdw13);
1454         c.common.cdw14 = cpu_to_le32(cmd.cdw14);
1455         c.common.cdw15 = cpu_to_le32(cmd.cdw15);
1456
1457         if (cmd.timeout_ms)
1458                 timeout = msecs_to_jiffies(cmd.timeout_ms);
1459
1460         effects = nvme_passthru_start(ctrl, ns, cmd.opcode);
1461         status = nvme_submit_user_cmd(ns ? ns->queue : ctrl->admin_q, &c,
1462                         (void __user *)(uintptr_t)cmd.addr, cmd.data_len,
1463                         (void __user *)(uintptr_t)cmd.metadata, cmd.metadata_len,
1464                         0, &cmd.result, timeout);
1465         nvme_passthru_end(ctrl, effects);
1466
1467         if (status >= 0) {
1468                 if (put_user(cmd.result, &ucmd->result))
1469                         return -EFAULT;
1470         }
1471
1472         return status;
1473 }
1474
1475 /*
1476  * Issue ioctl requests on the first available path.  Note that unlike normal
1477  * block layer requests we will not retry failed request on another controller.
1478  */
1479 static struct nvme_ns *nvme_get_ns_from_disk(struct gendisk *disk,
1480                 struct nvme_ns_head **head, int *srcu_idx)
1481 {
1482 #ifdef CONFIG_NVME_MULTIPATH
1483         if (disk->fops == &nvme_ns_head_ops) {
1484                 struct nvme_ns *ns;
1485
1486                 *head = disk->private_data;
1487                 *srcu_idx = srcu_read_lock(&(*head)->srcu);
1488                 ns = nvme_find_path(*head);
1489                 if (!ns)
1490                         srcu_read_unlock(&(*head)->srcu, *srcu_idx);
1491                 return ns;
1492         }
1493 #endif
1494         *head = NULL;
1495         *srcu_idx = -1;
1496         return disk->private_data;
1497 }
1498
1499 static void nvme_put_ns_from_disk(struct nvme_ns_head *head, int idx)
1500 {
1501         if (head)
1502                 srcu_read_unlock(&head->srcu, idx);
1503 }
1504
1505 static bool is_ctrl_ioctl(unsigned int cmd)
1506 {
1507         if (cmd == NVME_IOCTL_ADMIN_CMD || cmd == NVME_IOCTL_ADMIN64_CMD)
1508                 return true;
1509         if (is_sed_ioctl(cmd))
1510                 return true;
1511         return false;
1512 }
1513
1514 static int nvme_handle_ctrl_ioctl(struct nvme_ns *ns, unsigned int cmd,
1515                                   void __user *argp,
1516                                   struct nvme_ns_head *head,
1517                                   int srcu_idx)
1518 {
1519         struct nvme_ctrl *ctrl = ns->ctrl;
1520         int ret;
1521
1522         nvme_get_ctrl(ns->ctrl);
1523         nvme_put_ns_from_disk(head, srcu_idx);
1524
1525         switch (cmd) {
1526         case NVME_IOCTL_ADMIN_CMD:
1527                 ret = nvme_user_cmd(ctrl, NULL, argp);
1528                 break;
1529         case NVME_IOCTL_ADMIN64_CMD:
1530                 ret = nvme_user_cmd64(ctrl, NULL, argp);
1531                 break;
1532         default:
1533                 ret = sed_ioctl(ctrl->opal_dev, cmd, argp);
1534                 break;
1535         }
1536         nvme_put_ctrl(ctrl);
1537         return ret;
1538 }
1539
1540 static int nvme_ioctl(struct block_device *bdev, fmode_t mode,
1541                 unsigned int cmd, unsigned long arg)
1542 {
1543         struct nvme_ns_head *head = NULL;
1544         void __user *argp = (void __user *)arg;
1545         struct nvme_ns *ns;
1546         int srcu_idx, ret;
1547
1548         ns = nvme_get_ns_from_disk(bdev->bd_disk, &head, &srcu_idx);
1549         if (unlikely(!ns))
1550                 return -EWOULDBLOCK;
1551
1552         /*
1553          * Handle ioctls that apply to the controller instead of the namespace
1554          * seperately and drop the ns SRCU reference early.  This avoids a
1555          * deadlock when deleting namespaces using the passthrough interface.
1556          */
1557         if (is_ctrl_ioctl(cmd))
1558                 return nvme_handle_ctrl_ioctl(ns, cmd, argp, head, srcu_idx);
1559
1560         switch (cmd) {
1561         case NVME_IOCTL_ID:
1562                 force_successful_syscall_return();
1563                 ret = ns->head->ns_id;
1564                 break;
1565         case NVME_IOCTL_IO_CMD:
1566                 ret = nvme_user_cmd(ns->ctrl, ns, argp);
1567                 break;
1568         case NVME_IOCTL_SUBMIT_IO:
1569                 ret = nvme_submit_io(ns, argp);
1570                 break;
1571         case NVME_IOCTL_IO64_CMD:
1572                 ret = nvme_user_cmd64(ns->ctrl, ns, argp);
1573                 break;
1574         default:
1575                 if (ns->ndev)
1576                         ret = nvme_nvm_ioctl(ns, cmd, arg);
1577                 else
1578                         ret = -ENOTTY;
1579         }
1580
1581         nvme_put_ns_from_disk(head, srcu_idx);
1582         return ret;
1583 }
1584
1585 static int nvme_open(struct block_device *bdev, fmode_t mode)
1586 {
1587         struct nvme_ns *ns = bdev->bd_disk->private_data;
1588
1589 #ifdef CONFIG_NVME_MULTIPATH
1590         /* should never be called due to GENHD_FL_HIDDEN */
1591         if (WARN_ON_ONCE(ns->head->disk))
1592                 goto fail;
1593 #endif
1594         if (!kref_get_unless_zero(&ns->kref))
1595                 goto fail;
1596         if (!try_module_get(ns->ctrl->ops->module))
1597                 goto fail_put_ns;
1598
1599         return 0;
1600
1601 fail_put_ns:
1602         nvme_put_ns(ns);
1603 fail:
1604         return -ENXIO;
1605 }
1606
1607 static void nvme_release(struct gendisk *disk, fmode_t mode)
1608 {
1609         struct nvme_ns *ns = disk->private_data;
1610
1611         module_put(ns->ctrl->ops->module);
1612         nvme_put_ns(ns);
1613 }
1614
1615 static int nvme_getgeo(struct block_device *bdev, struct hd_geometry *geo)
1616 {
1617         /* some standard values */
1618         geo->heads = 1 << 6;
1619         geo->sectors = 1 << 5;
1620         geo->cylinders = get_capacity(bdev->bd_disk) >> 11;
1621         return 0;
1622 }
1623
1624 #ifdef CONFIG_BLK_DEV_INTEGRITY
1625 static void nvme_init_integrity(struct gendisk *disk, u16 ms, u8 pi_type)
1626 {
1627         struct blk_integrity integrity;
1628
1629         memset(&integrity, 0, sizeof(integrity));
1630         switch (pi_type) {
1631         case NVME_NS_DPS_PI_TYPE3:
1632                 integrity.profile = &t10_pi_type3_crc;
1633                 integrity.tag_size = sizeof(u16) + sizeof(u32);
1634                 integrity.flags |= BLK_INTEGRITY_DEVICE_CAPABLE;
1635                 break;
1636         case NVME_NS_DPS_PI_TYPE1:
1637         case NVME_NS_DPS_PI_TYPE2:
1638                 integrity.profile = &t10_pi_type1_crc;
1639                 integrity.tag_size = sizeof(u16);
1640                 integrity.flags |= BLK_INTEGRITY_DEVICE_CAPABLE;
1641                 break;
1642         default:
1643                 integrity.profile = NULL;
1644                 break;
1645         }
1646         integrity.tuple_size = ms;
1647         blk_integrity_register(disk, &integrity);
1648         blk_queue_max_integrity_segments(disk->queue, 1);
1649 }
1650 #else
1651 static void nvme_init_integrity(struct gendisk *disk, u16 ms, u8 pi_type)
1652 {
1653 }
1654 #endif /* CONFIG_BLK_DEV_INTEGRITY */
1655
1656 static void nvme_set_chunk_size(struct nvme_ns *ns)
1657 {
1658         u32 chunk_size = nvme_lba_to_sect(ns, ns->noiob);
1659         blk_queue_chunk_sectors(ns->queue, rounddown_pow_of_two(chunk_size));
1660 }
1661
1662 static void nvme_config_discard(struct gendisk *disk, struct nvme_ns *ns)
1663 {
1664         struct nvme_ctrl *ctrl = ns->ctrl;
1665         struct request_queue *queue = disk->queue;
1666         u32 size = queue_logical_block_size(queue);
1667
1668         if (!(ctrl->oncs & NVME_CTRL_ONCS_DSM)) {
1669                 blk_queue_flag_clear(QUEUE_FLAG_DISCARD, queue);
1670                 return;
1671         }
1672
1673         if (ctrl->nr_streams && ns->sws && ns->sgs)
1674                 size *= ns->sws * ns->sgs;
1675
1676         BUILD_BUG_ON(PAGE_SIZE / sizeof(struct nvme_dsm_range) <
1677                         NVME_DSM_MAX_RANGES);
1678
1679         queue->limits.discard_alignment = 0;
1680         queue->limits.discard_granularity = size;
1681
1682         /* If discard is already enabled, don't reset queue limits */
1683         if (blk_queue_flag_test_and_set(QUEUE_FLAG_DISCARD, queue))
1684                 return;
1685
1686         blk_queue_max_discard_sectors(queue, UINT_MAX);
1687         blk_queue_max_discard_segments(queue, NVME_DSM_MAX_RANGES);
1688
1689         if (ctrl->quirks & NVME_QUIRK_DEALLOCATE_ZEROES)
1690                 blk_queue_max_write_zeroes_sectors(queue, UINT_MAX);
1691 }
1692
1693 static void nvme_config_write_zeroes(struct gendisk *disk, struct nvme_ns *ns)
1694 {
1695         u64 max_blocks;
1696
1697         if (!(ns->ctrl->oncs & NVME_CTRL_ONCS_WRITE_ZEROES) ||
1698             (ns->ctrl->quirks & NVME_QUIRK_DISABLE_WRITE_ZEROES))
1699                 return;
1700         /*
1701          * Even though NVMe spec explicitly states that MDTS is not
1702          * applicable to the write-zeroes:- "The restriction does not apply to
1703          * commands that do not transfer data between the host and the
1704          * controller (e.g., Write Uncorrectable ro Write Zeroes command).".
1705          * In order to be more cautious use controller's max_hw_sectors value
1706          * to configure the maximum sectors for the write-zeroes which is
1707          * configured based on the controller's MDTS field in the
1708          * nvme_init_identify() if available.
1709          */
1710         if (ns->ctrl->max_hw_sectors == UINT_MAX)
1711                 max_blocks = (u64)USHRT_MAX + 1;
1712         else
1713                 max_blocks = ns->ctrl->max_hw_sectors + 1;
1714
1715         blk_queue_max_write_zeroes_sectors(disk->queue,
1716                                            nvme_lba_to_sect(ns, max_blocks));
1717 }
1718
1719 static int nvme_report_ns_ids(struct nvme_ctrl *ctrl, unsigned int nsid,
1720                 struct nvme_id_ns *id, struct nvme_ns_ids *ids)
1721 {
1722         int ret = 0;
1723
1724         memset(ids, 0, sizeof(*ids));
1725
1726         if (ctrl->vs >= NVME_VS(1, 1, 0))
1727                 memcpy(ids->eui64, id->eui64, sizeof(id->eui64));
1728         if (ctrl->vs >= NVME_VS(1, 2, 0))
1729                 memcpy(ids->nguid, id->nguid, sizeof(id->nguid));
1730         if (ctrl->vs >= NVME_VS(1, 3, 0)) {
1731                  /* Don't treat error as fatal we potentially
1732                   * already have a NGUID or EUI-64
1733                   */
1734                 ret = nvme_identify_ns_descs(ctrl, nsid, ids);
1735                 if (ret)
1736                         dev_warn(ctrl->device,
1737                                  "Identify Descriptors failed (%d)\n", ret);
1738         }
1739         return ret;
1740 }
1741
1742 static bool nvme_ns_ids_valid(struct nvme_ns_ids *ids)
1743 {
1744         return !uuid_is_null(&ids->uuid) ||
1745                 memchr_inv(ids->nguid, 0, sizeof(ids->nguid)) ||
1746                 memchr_inv(ids->eui64, 0, sizeof(ids->eui64));
1747 }
1748
1749 static bool nvme_ns_ids_equal(struct nvme_ns_ids *a, struct nvme_ns_ids *b)
1750 {
1751         return uuid_equal(&a->uuid, &b->uuid) &&
1752                 memcmp(&a->nguid, &b->nguid, sizeof(a->nguid)) == 0 &&
1753                 memcmp(&a->eui64, &b->eui64, sizeof(a->eui64)) == 0;
1754 }
1755
1756 static void nvme_update_disk_info(struct gendisk *disk,
1757                 struct nvme_ns *ns, struct nvme_id_ns *id)
1758 {
1759         sector_t capacity = nvme_lba_to_sect(ns, le64_to_cpu(id->nsze));
1760         unsigned short bs = 1 << ns->lba_shift;
1761         u32 atomic_bs, phys_bs, io_opt;
1762
1763         if (ns->lba_shift > PAGE_SHIFT) {
1764                 /* unsupported block size, set capacity to 0 later */
1765                 bs = (1 << 9);
1766         }
1767         blk_mq_freeze_queue(disk->queue);
1768         blk_integrity_unregister(disk);
1769
1770         if (id->nabo == 0) {
1771                 /*
1772                  * Bit 1 indicates whether NAWUPF is defined for this namespace
1773                  * and whether it should be used instead of AWUPF. If NAWUPF ==
1774                  * 0 then AWUPF must be used instead.
1775                  */
1776                 if (id->nsfeat & (1 << 1) && id->nawupf)
1777                         atomic_bs = (1 + le16_to_cpu(id->nawupf)) * bs;
1778                 else
1779                         atomic_bs = (1 + ns->ctrl->subsys->awupf) * bs;
1780         } else {
1781                 atomic_bs = bs;
1782         }
1783         phys_bs = bs;
1784         io_opt = bs;
1785         if (id->nsfeat & (1 << 4)) {
1786                 /* NPWG = Namespace Preferred Write Granularity */
1787                 phys_bs *= 1 + le16_to_cpu(id->npwg);
1788                 /* NOWS = Namespace Optimal Write Size */
1789                 io_opt *= 1 + le16_to_cpu(id->nows);
1790         }
1791
1792         blk_queue_logical_block_size(disk->queue, bs);
1793         /*
1794          * Linux filesystems assume writing a single physical block is
1795          * an atomic operation. Hence limit the physical block size to the
1796          * value of the Atomic Write Unit Power Fail parameter.
1797          */
1798         blk_queue_physical_block_size(disk->queue, min(phys_bs, atomic_bs));
1799         blk_queue_io_min(disk->queue, phys_bs);
1800         blk_queue_io_opt(disk->queue, io_opt);
1801
1802         if (ns->ms && !ns->ext &&
1803             (ns->ctrl->ops->flags & NVME_F_METADATA_SUPPORTED))
1804                 nvme_init_integrity(disk, ns->ms, ns->pi_type);
1805         if ((ns->ms && !nvme_ns_has_pi(ns) && !blk_get_integrity(disk)) ||
1806             ns->lba_shift > PAGE_SHIFT)
1807                 capacity = 0;
1808
1809         set_capacity(disk, capacity);
1810
1811         nvme_config_discard(disk, ns);
1812         nvme_config_write_zeroes(disk, ns);
1813
1814         if (id->nsattr & (1 << 0))
1815                 set_disk_ro(disk, true);
1816         else
1817                 set_disk_ro(disk, false);
1818
1819         blk_mq_unfreeze_queue(disk->queue);
1820 }
1821
1822 static void __nvme_revalidate_disk(struct gendisk *disk, struct nvme_id_ns *id)
1823 {
1824         struct nvme_ns *ns = disk->private_data;
1825
1826         /*
1827          * If identify namespace failed, use default 512 byte block size so
1828          * block layer can use before failing read/write for 0 capacity.
1829          */
1830         ns->lba_shift = id->lbaf[id->flbas & NVME_NS_FLBAS_LBA_MASK].ds;
1831         if (ns->lba_shift == 0)
1832                 ns->lba_shift = 9;
1833         ns->noiob = le16_to_cpu(id->noiob);
1834         ns->ms = le16_to_cpu(id->lbaf[id->flbas & NVME_NS_FLBAS_LBA_MASK].ms);
1835         ns->ext = ns->ms && (id->flbas & NVME_NS_FLBAS_META_EXT);
1836         /* the PI implementation requires metadata equal t10 pi tuple size */
1837         if (ns->ms == sizeof(struct t10_pi_tuple))
1838                 ns->pi_type = id->dps & NVME_NS_DPS_PI_MASK;
1839         else
1840                 ns->pi_type = 0;
1841
1842         if (ns->noiob)
1843                 nvme_set_chunk_size(ns);
1844         nvme_update_disk_info(disk, ns, id);
1845 #ifdef CONFIG_NVME_MULTIPATH
1846         if (ns->head->disk) {
1847                 nvme_update_disk_info(ns->head->disk, ns, id);
1848                 blk_queue_stack_limits(ns->head->disk->queue, ns->queue);
1849                 revalidate_disk(ns->head->disk);
1850         }
1851 #endif
1852 }
1853
1854 static int nvme_revalidate_disk(struct gendisk *disk)
1855 {
1856         struct nvme_ns *ns = disk->private_data;
1857         struct nvme_ctrl *ctrl = ns->ctrl;
1858         struct nvme_id_ns *id;
1859         struct nvme_ns_ids ids;
1860         int ret = 0;
1861
1862         if (test_bit(NVME_NS_DEAD, &ns->flags)) {
1863                 set_capacity(disk, 0);
1864                 return -ENODEV;
1865         }
1866
1867         ret = nvme_identify_ns(ctrl, ns->head->ns_id, &id);
1868         if (ret)
1869                 goto out;
1870
1871         if (id->ncap == 0) {
1872                 ret = -ENODEV;
1873                 goto free_id;
1874         }
1875
1876         __nvme_revalidate_disk(disk, id);
1877         ret = nvme_report_ns_ids(ctrl, ns->head->ns_id, id, &ids);
1878         if (ret)
1879                 goto free_id;
1880
1881         if (!nvme_ns_ids_equal(&ns->head->ids, &ids)) {
1882                 dev_err(ctrl->device,
1883                         "identifiers changed for nsid %d\n", ns->head->ns_id);
1884                 ret = -ENODEV;
1885         }
1886
1887 free_id:
1888         kfree(id);
1889 out:
1890         /*
1891          * Only fail the function if we got a fatal error back from the
1892          * device, otherwise ignore the error and just move on.
1893          */
1894         if (ret == -ENOMEM || (ret > 0 && !(ret & NVME_SC_DNR)))
1895                 ret = 0;
1896         else if (ret > 0)
1897                 ret = blk_status_to_errno(nvme_error_status(ret));
1898         return ret;
1899 }
1900
1901 static char nvme_pr_type(enum pr_type type)
1902 {
1903         switch (type) {
1904         case PR_WRITE_EXCLUSIVE:
1905                 return 1;
1906         case PR_EXCLUSIVE_ACCESS:
1907                 return 2;
1908         case PR_WRITE_EXCLUSIVE_REG_ONLY:
1909                 return 3;
1910         case PR_EXCLUSIVE_ACCESS_REG_ONLY:
1911                 return 4;
1912         case PR_WRITE_EXCLUSIVE_ALL_REGS:
1913                 return 5;
1914         case PR_EXCLUSIVE_ACCESS_ALL_REGS:
1915                 return 6;
1916         default:
1917                 return 0;
1918         }
1919 };
1920
1921 static int nvme_pr_command(struct block_device *bdev, u32 cdw10,
1922                                 u64 key, u64 sa_key, u8 op)
1923 {
1924         struct nvme_ns_head *head = NULL;
1925         struct nvme_ns *ns;
1926         struct nvme_command c;
1927         int srcu_idx, ret;
1928         u8 data[16] = { 0, };
1929
1930         ns = nvme_get_ns_from_disk(bdev->bd_disk, &head, &srcu_idx);
1931         if (unlikely(!ns))
1932                 return -EWOULDBLOCK;
1933
1934         put_unaligned_le64(key, &data[0]);
1935         put_unaligned_le64(sa_key, &data[8]);
1936
1937         memset(&c, 0, sizeof(c));
1938         c.common.opcode = op;
1939         c.common.nsid = cpu_to_le32(ns->head->ns_id);
1940         c.common.cdw10 = cpu_to_le32(cdw10);
1941
1942         ret = nvme_submit_sync_cmd(ns->queue, &c, data, 16);
1943         nvme_put_ns_from_disk(head, srcu_idx);
1944         return ret;
1945 }
1946
1947 static int nvme_pr_register(struct block_device *bdev, u64 old,
1948                 u64 new, unsigned flags)
1949 {
1950         u32 cdw10;
1951
1952         if (flags & ~PR_FL_IGNORE_KEY)
1953                 return -EOPNOTSUPP;
1954
1955         cdw10 = old ? 2 : 0;
1956         cdw10 |= (flags & PR_FL_IGNORE_KEY) ? 1 << 3 : 0;
1957         cdw10 |= (1 << 30) | (1 << 31); /* PTPL=1 */
1958         return nvme_pr_command(bdev, cdw10, old, new, nvme_cmd_resv_register);
1959 }
1960
1961 static int nvme_pr_reserve(struct block_device *bdev, u64 key,
1962                 enum pr_type type, unsigned flags)
1963 {
1964         u32 cdw10;
1965
1966         if (flags & ~PR_FL_IGNORE_KEY)
1967                 return -EOPNOTSUPP;
1968
1969         cdw10 = nvme_pr_type(type) << 8;
1970         cdw10 |= ((flags & PR_FL_IGNORE_KEY) ? 1 << 3 : 0);
1971         return nvme_pr_command(bdev, cdw10, key, 0, nvme_cmd_resv_acquire);
1972 }
1973
1974 static int nvme_pr_preempt(struct block_device *bdev, u64 old, u64 new,
1975                 enum pr_type type, bool abort)
1976 {
1977         u32 cdw10 = nvme_pr_type(type) << 8 | (abort ? 2 : 1);
1978         return nvme_pr_command(bdev, cdw10, old, new, nvme_cmd_resv_acquire);
1979 }
1980
1981 static int nvme_pr_clear(struct block_device *bdev, u64 key)
1982 {
1983         u32 cdw10 = 1 | (key ? 1 << 3 : 0);
1984         return nvme_pr_command(bdev, cdw10, key, 0, nvme_cmd_resv_register);
1985 }
1986
1987 static int nvme_pr_release(struct block_device *bdev, u64 key, enum pr_type type)
1988 {
1989         u32 cdw10 = nvme_pr_type(type) << 8 | (key ? 1 << 3 : 0);
1990         return nvme_pr_command(bdev, cdw10, key, 0, nvme_cmd_resv_release);
1991 }
1992
1993 static const struct pr_ops nvme_pr_ops = {
1994         .pr_register    = nvme_pr_register,
1995         .pr_reserve     = nvme_pr_reserve,
1996         .pr_release     = nvme_pr_release,
1997         .pr_preempt     = nvme_pr_preempt,
1998         .pr_clear       = nvme_pr_clear,
1999 };
2000
2001 #ifdef CONFIG_BLK_SED_OPAL
2002 int nvme_sec_submit(void *data, u16 spsp, u8 secp, void *buffer, size_t len,
2003                 bool send)
2004 {
2005         struct nvme_ctrl *ctrl = data;
2006         struct nvme_command cmd;
2007
2008         memset(&cmd, 0, sizeof(cmd));
2009         if (send)
2010                 cmd.common.opcode = nvme_admin_security_send;
2011         else
2012                 cmd.common.opcode = nvme_admin_security_recv;
2013         cmd.common.nsid = 0;
2014         cmd.common.cdw10 = cpu_to_le32(((u32)secp) << 24 | ((u32)spsp) << 8);
2015         cmd.common.cdw11 = cpu_to_le32(len);
2016
2017         return __nvme_submit_sync_cmd(ctrl->admin_q, &cmd, NULL, buffer, len,
2018                                       ADMIN_TIMEOUT, NVME_QID_ANY, 1, 0, false);
2019 }
2020 EXPORT_SYMBOL_GPL(nvme_sec_submit);
2021 #endif /* CONFIG_BLK_SED_OPAL */
2022
2023 static const struct block_device_operations nvme_fops = {
2024         .owner          = THIS_MODULE,
2025         .ioctl          = nvme_ioctl,
2026         .compat_ioctl   = nvme_ioctl,
2027         .open           = nvme_open,
2028         .release        = nvme_release,
2029         .getgeo         = nvme_getgeo,
2030         .revalidate_disk= nvme_revalidate_disk,
2031         .pr_ops         = &nvme_pr_ops,
2032 };
2033
2034 #ifdef CONFIG_NVME_MULTIPATH
2035 static int nvme_ns_head_open(struct block_device *bdev, fmode_t mode)
2036 {
2037         struct nvme_ns_head *head = bdev->bd_disk->private_data;
2038
2039         if (!kref_get_unless_zero(&head->ref))
2040                 return -ENXIO;
2041         return 0;
2042 }
2043
2044 static void nvme_ns_head_release(struct gendisk *disk, fmode_t mode)
2045 {
2046         nvme_put_ns_head(disk->private_data);
2047 }
2048
2049 const struct block_device_operations nvme_ns_head_ops = {
2050         .owner          = THIS_MODULE,
2051         .open           = nvme_ns_head_open,
2052         .release        = nvme_ns_head_release,
2053         .ioctl          = nvme_ioctl,
2054         .compat_ioctl   = nvme_ioctl,
2055         .getgeo         = nvme_getgeo,
2056         .pr_ops         = &nvme_pr_ops,
2057 };
2058 #endif /* CONFIG_NVME_MULTIPATH */
2059
2060 static int nvme_wait_ready(struct nvme_ctrl *ctrl, u64 cap, bool enabled)
2061 {
2062         unsigned long timeout =
2063                 ((NVME_CAP_TIMEOUT(cap) + 1) * HZ / 2) + jiffies;
2064         u32 csts, bit = enabled ? NVME_CSTS_RDY : 0;
2065         int ret;
2066
2067         while ((ret = ctrl->ops->reg_read32(ctrl, NVME_REG_CSTS, &csts)) == 0) {
2068                 if (csts == ~0)
2069                         return -ENODEV;
2070                 if ((csts & NVME_CSTS_RDY) == bit)
2071                         break;
2072
2073                 msleep(100);
2074                 if (fatal_signal_pending(current))
2075                         return -EINTR;
2076                 if (time_after(jiffies, timeout)) {
2077                         dev_err(ctrl->device,
2078                                 "Device not ready; aborting %s\n", enabled ?
2079                                                 "initialisation" : "reset");
2080                         return -ENODEV;
2081                 }
2082         }
2083
2084         return ret;
2085 }
2086
2087 /*
2088  * If the device has been passed off to us in an enabled state, just clear
2089  * the enabled bit.  The spec says we should set the 'shutdown notification
2090  * bits', but doing so may cause the device to complete commands to the
2091  * admin queue ... and we don't know what memory that might be pointing at!
2092  */
2093 int nvme_disable_ctrl(struct nvme_ctrl *ctrl)
2094 {
2095         int ret;
2096
2097         ctrl->ctrl_config &= ~NVME_CC_SHN_MASK;
2098         ctrl->ctrl_config &= ~NVME_CC_ENABLE;
2099
2100         ret = ctrl->ops->reg_write32(ctrl, NVME_REG_CC, ctrl->ctrl_config);
2101         if (ret)
2102                 return ret;
2103
2104         if (ctrl->quirks & NVME_QUIRK_DELAY_BEFORE_CHK_RDY)
2105                 msleep(NVME_QUIRK_DELAY_AMOUNT);
2106
2107         return nvme_wait_ready(ctrl, ctrl->cap, false);
2108 }
2109 EXPORT_SYMBOL_GPL(nvme_disable_ctrl);
2110
2111 int nvme_enable_ctrl(struct nvme_ctrl *ctrl)
2112 {
2113         /*
2114          * Default to a 4K page size, with the intention to update this
2115          * path in the future to accomodate architectures with differing
2116          * kernel and IO page sizes.
2117          */
2118         unsigned dev_page_min, page_shift = 12;
2119         int ret;
2120
2121         ret = ctrl->ops->reg_read64(ctrl, NVME_REG_CAP, &ctrl->cap);
2122         if (ret) {
2123                 dev_err(ctrl->device, "Reading CAP failed (%d)\n", ret);
2124                 return ret;
2125         }
2126         dev_page_min = NVME_CAP_MPSMIN(ctrl->cap) + 12;
2127
2128         if (page_shift < dev_page_min) {
2129                 dev_err(ctrl->device,
2130                         "Minimum device page size %u too large for host (%u)\n",
2131                         1 << dev_page_min, 1 << page_shift);
2132                 return -ENODEV;
2133         }
2134
2135         ctrl->page_size = 1 << page_shift;
2136
2137         ctrl->ctrl_config = NVME_CC_CSS_NVM;
2138         ctrl->ctrl_config |= (page_shift - 12) << NVME_CC_MPS_SHIFT;
2139         ctrl->ctrl_config |= NVME_CC_AMS_RR | NVME_CC_SHN_NONE;
2140         ctrl->ctrl_config |= NVME_CC_IOSQES | NVME_CC_IOCQES;
2141         ctrl->ctrl_config |= NVME_CC_ENABLE;
2142
2143         ret = ctrl->ops->reg_write32(ctrl, NVME_REG_CC, ctrl->ctrl_config);
2144         if (ret)
2145                 return ret;
2146         return nvme_wait_ready(ctrl, ctrl->cap, true);
2147 }
2148 EXPORT_SYMBOL_GPL(nvme_enable_ctrl);
2149
2150 int nvme_shutdown_ctrl(struct nvme_ctrl *ctrl)
2151 {
2152         unsigned long timeout = jiffies + (ctrl->shutdown_timeout * HZ);
2153         u32 csts;
2154         int ret;
2155
2156         ctrl->ctrl_config &= ~NVME_CC_SHN_MASK;
2157         ctrl->ctrl_config |= NVME_CC_SHN_NORMAL;
2158
2159         ret = ctrl->ops->reg_write32(ctrl, NVME_REG_CC, ctrl->ctrl_config);
2160         if (ret)
2161                 return ret;
2162
2163         while ((ret = ctrl->ops->reg_read32(ctrl, NVME_REG_CSTS, &csts)) == 0) {
2164                 if ((csts & NVME_CSTS_SHST_MASK) == NVME_CSTS_SHST_CMPLT)
2165                         break;
2166
2167                 msleep(100);
2168                 if (fatal_signal_pending(current))
2169                         return -EINTR;
2170                 if (time_after(jiffies, timeout)) {
2171                         dev_err(ctrl->device,
2172                                 "Device shutdown incomplete; abort shutdown\n");
2173                         return -ENODEV;
2174                 }
2175         }
2176
2177         return ret;
2178 }
2179 EXPORT_SYMBOL_GPL(nvme_shutdown_ctrl);
2180
2181 static void nvme_set_queue_limits(struct nvme_ctrl *ctrl,
2182                 struct request_queue *q)
2183 {
2184         bool vwc = false;
2185
2186         if (ctrl->max_hw_sectors) {
2187                 u32 max_segments =
2188                         (ctrl->max_hw_sectors / (ctrl->page_size >> 9)) + 1;
2189
2190                 max_segments = min_not_zero(max_segments, ctrl->max_segments);
2191                 blk_queue_max_hw_sectors(q, ctrl->max_hw_sectors);
2192                 blk_queue_max_segments(q, min_t(u32, max_segments, USHRT_MAX));
2193         }
2194         if ((ctrl->quirks & NVME_QUIRK_STRIPE_SIZE) &&
2195             is_power_of_2(ctrl->max_hw_sectors))
2196                 blk_queue_chunk_sectors(q, ctrl->max_hw_sectors);
2197         blk_queue_virt_boundary(q, ctrl->page_size - 1);
2198         if (ctrl->vwc & NVME_CTRL_VWC_PRESENT)
2199                 vwc = true;
2200         blk_queue_write_cache(q, vwc, vwc);
2201 }
2202
2203 static int nvme_configure_timestamp(struct nvme_ctrl *ctrl)
2204 {
2205         __le64 ts;
2206         int ret;
2207
2208         if (!(ctrl->oncs & NVME_CTRL_ONCS_TIMESTAMP))
2209                 return 0;
2210
2211         ts = cpu_to_le64(ktime_to_ms(ktime_get_real()));
2212         ret = nvme_set_features(ctrl, NVME_FEAT_TIMESTAMP, 0, &ts, sizeof(ts),
2213                         NULL);
2214         if (ret)
2215                 dev_warn_once(ctrl->device,
2216                         "could not set timestamp (%d)\n", ret);
2217         return ret;
2218 }
2219
2220 static int nvme_configure_acre(struct nvme_ctrl *ctrl)
2221 {
2222         struct nvme_feat_host_behavior *host;
2223         int ret;
2224
2225         /* Don't bother enabling the feature if retry delay is not reported */
2226         if (!ctrl->crdt[0])
2227                 return 0;
2228
2229         host = kzalloc(sizeof(*host), GFP_KERNEL);
2230         if (!host)
2231                 return 0;
2232
2233         host->acre = NVME_ENABLE_ACRE;
2234         ret = nvme_set_features(ctrl, NVME_FEAT_HOST_BEHAVIOR, 0,
2235                                 host, sizeof(*host), NULL);
2236         kfree(host);
2237         return ret;
2238 }
2239
2240 static int nvme_configure_apst(struct nvme_ctrl *ctrl)
2241 {
2242         /*
2243          * APST (Autonomous Power State Transition) lets us program a
2244          * table of power state transitions that the controller will
2245          * perform automatically.  We configure it with a simple
2246          * heuristic: we are willing to spend at most 2% of the time
2247          * transitioning between power states.  Therefore, when running
2248          * in any given state, we will enter the next lower-power
2249          * non-operational state after waiting 50 * (enlat + exlat)
2250          * microseconds, as long as that state's exit latency is under
2251          * the requested maximum latency.
2252          *
2253          * We will not autonomously enter any non-operational state for
2254          * which the total latency exceeds ps_max_latency_us.  Users
2255          * can set ps_max_latency_us to zero to turn off APST.
2256          */
2257
2258         unsigned apste;
2259         struct nvme_feat_auto_pst *table;
2260         u64 max_lat_us = 0;
2261         int max_ps = -1;
2262         int ret;
2263
2264         /*
2265          * If APST isn't supported or if we haven't been initialized yet,
2266          * then don't do anything.
2267          */
2268         if (!ctrl->apsta)
2269                 return 0;
2270
2271         if (ctrl->npss > 31) {
2272                 dev_warn(ctrl->device, "NPSS is invalid; not using APST\n");
2273                 return 0;
2274         }
2275
2276         table = kzalloc(sizeof(*table), GFP_KERNEL);
2277         if (!table)
2278                 return 0;
2279
2280         if (!ctrl->apst_enabled || ctrl->ps_max_latency_us == 0) {
2281                 /* Turn off APST. */
2282                 apste = 0;
2283                 dev_dbg(ctrl->device, "APST disabled\n");
2284         } else {
2285                 __le64 target = cpu_to_le64(0);
2286                 int state;
2287
2288                 /*
2289                  * Walk through all states from lowest- to highest-power.
2290                  * According to the spec, lower-numbered states use more
2291                  * power.  NPSS, despite the name, is the index of the
2292                  * lowest-power state, not the number of states.
2293                  */
2294                 for (state = (int)ctrl->npss; state >= 0; state--) {
2295                         u64 total_latency_us, exit_latency_us, transition_ms;
2296
2297                         if (target)
2298                                 table->entries[state] = target;
2299
2300                         /*
2301                          * Don't allow transitions to the deepest state
2302                          * if it's quirked off.
2303                          */
2304                         if (state == ctrl->npss &&
2305                             (ctrl->quirks & NVME_QUIRK_NO_DEEPEST_PS))
2306                                 continue;
2307
2308                         /*
2309                          * Is this state a useful non-operational state for
2310                          * higher-power states to autonomously transition to?
2311                          */
2312                         if (!(ctrl->psd[state].flags &
2313                               NVME_PS_FLAGS_NON_OP_STATE))
2314                                 continue;
2315
2316                         exit_latency_us =
2317                                 (u64)le32_to_cpu(ctrl->psd[state].exit_lat);
2318                         if (exit_latency_us > ctrl->ps_max_latency_us)
2319                                 continue;
2320
2321                         total_latency_us =
2322                                 exit_latency_us +
2323                                 le32_to_cpu(ctrl->psd[state].entry_lat);
2324
2325                         /*
2326                          * This state is good.  Use it as the APST idle
2327                          * target for higher power states.
2328                          */
2329                         transition_ms = total_latency_us + 19;
2330                         do_div(transition_ms, 20);
2331                         if (transition_ms > (1 << 24) - 1)
2332                                 transition_ms = (1 << 24) - 1;
2333
2334                         target = cpu_to_le64((state << 3) |
2335                                              (transition_ms << 8));
2336
2337                         if (max_ps == -1)
2338                                 max_ps = state;
2339
2340                         if (total_latency_us > max_lat_us)
2341                                 max_lat_us = total_latency_us;
2342                 }
2343
2344                 apste = 1;
2345
2346                 if (max_ps == -1) {
2347                         dev_dbg(ctrl->device, "APST enabled but no non-operational states are available\n");
2348                 } else {
2349                         dev_dbg(ctrl->device, "APST enabled: max PS = %d, max round-trip latency = %lluus, table = %*phN\n",
2350                                 max_ps, max_lat_us, (int)sizeof(*table), table);
2351                 }
2352         }
2353
2354         ret = nvme_set_features(ctrl, NVME_FEAT_AUTO_PST, apste,
2355                                 table, sizeof(*table), NULL);
2356         if (ret)
2357                 dev_err(ctrl->device, "failed to set APST feature (%d)\n", ret);
2358
2359         kfree(table);
2360         return ret;
2361 }
2362
2363 static void nvme_set_latency_tolerance(struct device *dev, s32 val)
2364 {
2365         struct nvme_ctrl *ctrl = dev_get_drvdata(dev);
2366         u64 latency;
2367
2368         switch (val) {
2369         case PM_QOS_LATENCY_TOLERANCE_NO_CONSTRAINT:
2370         case PM_QOS_LATENCY_ANY:
2371                 latency = U64_MAX;
2372                 break;
2373
2374         default:
2375                 latency = val;
2376         }
2377
2378         if (ctrl->ps_max_latency_us != latency) {
2379                 ctrl->ps_max_latency_us = latency;
2380                 nvme_configure_apst(ctrl);
2381         }
2382 }
2383
2384 struct nvme_core_quirk_entry {
2385         /*
2386          * NVMe model and firmware strings are padded with spaces.  For
2387          * simplicity, strings in the quirk table are padded with NULLs
2388          * instead.
2389          */
2390         u16 vid;
2391         const char *mn;
2392         const char *fr;
2393         unsigned long quirks;
2394 };
2395
2396 static const struct nvme_core_quirk_entry core_quirks[] = {
2397         {
2398                 /*
2399                  * This Toshiba device seems to die using any APST states.  See:
2400                  * https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1678184/comments/11
2401                  */
2402                 .vid = 0x1179,
2403                 .mn = "THNSF5256GPUK TOSHIBA",
2404                 .quirks = NVME_QUIRK_NO_APST,
2405         },
2406         {
2407                 /*
2408                  * This LiteON CL1-3D*-Q11 firmware version has a race
2409                  * condition associated with actions related to suspend to idle
2410                  * LiteON has resolved the problem in future firmware
2411                  */
2412                 .vid = 0x14a4,
2413                 .fr = "22301111",
2414                 .quirks = NVME_QUIRK_SIMPLE_SUSPEND,
2415         },
2416         {
2417                 /*
2418                  * This Kingston E8FK11.T firmware version has no interrupt
2419                  * after resume with actions related to suspend to idle
2420                  * https://bugzilla.kernel.org/show_bug.cgi?id=204887
2421                  */
2422                 .vid = 0x2646,
2423                 .fr = "E8FK11.T",
2424                 .quirks = NVME_QUIRK_SIMPLE_SUSPEND,
2425         }
2426 };
2427
2428 /* match is null-terminated but idstr is space-padded. */
2429 static bool string_matches(const char *idstr, const char *match, size_t len)
2430 {
2431         size_t matchlen;
2432
2433         if (!match)
2434                 return true;
2435
2436         matchlen = strlen(match);
2437         WARN_ON_ONCE(matchlen > len);
2438
2439         if (memcmp(idstr, match, matchlen))
2440                 return false;
2441
2442         for (; matchlen < len; matchlen++)
2443                 if (idstr[matchlen] != ' ')
2444                         return false;
2445
2446         return true;
2447 }
2448
2449 static bool quirk_matches(const struct nvme_id_ctrl *id,
2450                           const struct nvme_core_quirk_entry *q)
2451 {
2452         return q->vid == le16_to_cpu(id->vid) &&
2453                 string_matches(id->mn, q->mn, sizeof(id->mn)) &&
2454                 string_matches(id->fr, q->fr, sizeof(id->fr));
2455 }
2456
2457 static void nvme_init_subnqn(struct nvme_subsystem *subsys, struct nvme_ctrl *ctrl,
2458                 struct nvme_id_ctrl *id)
2459 {
2460         size_t nqnlen;
2461         int off;
2462
2463         if(!(ctrl->quirks & NVME_QUIRK_IGNORE_DEV_SUBNQN)) {
2464                 nqnlen = strnlen(id->subnqn, NVMF_NQN_SIZE);
2465                 if (nqnlen > 0 && nqnlen < NVMF_NQN_SIZE) {
2466                         strlcpy(subsys->subnqn, id->subnqn, NVMF_NQN_SIZE);
2467                         return;
2468                 }
2469
2470                 if (ctrl->vs >= NVME_VS(1, 2, 1))
2471                         dev_warn(ctrl->device, "missing or invalid SUBNQN field.\n");
2472         }
2473
2474         /* Generate a "fake" NQN per Figure 254 in NVMe 1.3 + ECN 001 */
2475         off = snprintf(subsys->subnqn, NVMF_NQN_SIZE,
2476                         "nqn.2014.08.org.nvmexpress:%04x%04x",
2477                         le16_to_cpu(id->vid), le16_to_cpu(id->ssvid));
2478         memcpy(subsys->subnqn + off, id->sn, sizeof(id->sn));
2479         off += sizeof(id->sn);
2480         memcpy(subsys->subnqn + off, id->mn, sizeof(id->mn));
2481         off += sizeof(id->mn);
2482         memset(subsys->subnqn + off, 0, sizeof(subsys->subnqn) - off);
2483 }
2484
2485 static void nvme_release_subsystem(struct device *dev)
2486 {
2487         struct nvme_subsystem *subsys =
2488                 container_of(dev, struct nvme_subsystem, dev);
2489
2490         if (subsys->instance >= 0)
2491                 ida_simple_remove(&nvme_instance_ida, subsys->instance);
2492         kfree(subsys);
2493 }
2494
2495 static void nvme_destroy_subsystem(struct kref *ref)
2496 {
2497         struct nvme_subsystem *subsys =
2498                         container_of(ref, struct nvme_subsystem, ref);
2499
2500         mutex_lock(&nvme_subsystems_lock);
2501         list_del(&subsys->entry);
2502         mutex_unlock(&nvme_subsystems_lock);
2503
2504         ida_destroy(&subsys->ns_ida);
2505         device_del(&subsys->dev);
2506         put_device(&subsys->dev);
2507 }
2508
2509 static void nvme_put_subsystem(struct nvme_subsystem *subsys)
2510 {
2511         kref_put(&subsys->ref, nvme_destroy_subsystem);
2512 }
2513
2514 static struct nvme_subsystem *__nvme_find_get_subsystem(const char *subsysnqn)
2515 {
2516         struct nvme_subsystem *subsys;
2517
2518         lockdep_assert_held(&nvme_subsystems_lock);
2519
2520         /*
2521          * Fail matches for discovery subsystems. This results
2522          * in each discovery controller bound to a unique subsystem.
2523          * This avoids issues with validating controller values
2524          * that can only be true when there is a single unique subsystem.
2525          * There may be multiple and completely independent entities
2526          * that provide discovery controllers.
2527          */
2528         if (!strcmp(subsysnqn, NVME_DISC_SUBSYS_NAME))
2529                 return NULL;
2530
2531         list_for_each_entry(subsys, &nvme_subsystems, entry) {
2532                 if (strcmp(subsys->subnqn, subsysnqn))
2533                         continue;
2534                 if (!kref_get_unless_zero(&subsys->ref))
2535                         continue;
2536                 return subsys;
2537         }
2538
2539         return NULL;
2540 }
2541
2542 #define SUBSYS_ATTR_RO(_name, _mode, _show)                     \
2543         struct device_attribute subsys_attr_##_name = \
2544                 __ATTR(_name, _mode, _show, NULL)
2545
2546 static ssize_t nvme_subsys_show_nqn(struct device *dev,
2547                                     struct device_attribute *attr,
2548                                     char *buf)
2549 {
2550         struct nvme_subsystem *subsys =
2551                 container_of(dev, struct nvme_subsystem, dev);
2552
2553         return snprintf(buf, PAGE_SIZE, "%s\n", subsys->subnqn);
2554 }
2555 static SUBSYS_ATTR_RO(subsysnqn, S_IRUGO, nvme_subsys_show_nqn);
2556
2557 #define nvme_subsys_show_str_function(field)                            \
2558 static ssize_t subsys_##field##_show(struct device *dev,                \
2559                             struct device_attribute *attr, char *buf)   \
2560 {                                                                       \
2561         struct nvme_subsystem *subsys =                                 \
2562                 container_of(dev, struct nvme_subsystem, dev);          \
2563         return sprintf(buf, "%.*s\n",                                   \
2564                        (int)sizeof(subsys->field), subsys->field);      \
2565 }                                                                       \
2566 static SUBSYS_ATTR_RO(field, S_IRUGO, subsys_##field##_show);
2567
2568 nvme_subsys_show_str_function(model);
2569 nvme_subsys_show_str_function(serial);
2570 nvme_subsys_show_str_function(firmware_rev);
2571
2572 static struct attribute *nvme_subsys_attrs[] = {
2573         &subsys_attr_model.attr,
2574         &subsys_attr_serial.attr,
2575         &subsys_attr_firmware_rev.attr,
2576         &subsys_attr_subsysnqn.attr,
2577 #ifdef CONFIG_NVME_MULTIPATH
2578         &subsys_attr_iopolicy.attr,
2579 #endif
2580         NULL,
2581 };
2582
2583 static struct attribute_group nvme_subsys_attrs_group = {
2584         .attrs = nvme_subsys_attrs,
2585 };
2586
2587 static const struct attribute_group *nvme_subsys_attrs_groups[] = {
2588         &nvme_subsys_attrs_group,
2589         NULL,
2590 };
2591
2592 static bool nvme_validate_cntlid(struct nvme_subsystem *subsys,
2593                 struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id)
2594 {
2595         struct nvme_ctrl *tmp;
2596
2597         lockdep_assert_held(&nvme_subsystems_lock);
2598
2599         list_for_each_entry(tmp, &subsys->ctrls, subsys_entry) {
2600                 if (tmp->state == NVME_CTRL_DELETING ||
2601                     tmp->state == NVME_CTRL_DEAD)
2602                         continue;
2603
2604                 if (tmp->cntlid == ctrl->cntlid) {
2605                         dev_err(ctrl->device,
2606                                 "Duplicate cntlid %u with %s, rejecting\n",
2607                                 ctrl->cntlid, dev_name(tmp->device));
2608                         return false;
2609                 }
2610
2611                 if ((id->cmic & (1 << 1)) ||
2612                     (ctrl->opts && ctrl->opts->discovery_nqn))
2613                         continue;
2614
2615                 dev_err(ctrl->device,
2616                         "Subsystem does not support multiple controllers\n");
2617                 return false;
2618         }
2619
2620         return true;
2621 }
2622
2623 static int nvme_init_subsystem(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id)
2624 {
2625         struct nvme_subsystem *subsys, *found;
2626         int ret;
2627
2628         subsys = kzalloc(sizeof(*subsys), GFP_KERNEL);
2629         if (!subsys)
2630                 return -ENOMEM;
2631
2632         subsys->instance = -1;
2633         mutex_init(&subsys->lock);
2634         kref_init(&subsys->ref);
2635         INIT_LIST_HEAD(&subsys->ctrls);
2636         INIT_LIST_HEAD(&subsys->nsheads);
2637         nvme_init_subnqn(subsys, ctrl, id);
2638         memcpy(subsys->serial, id->sn, sizeof(subsys->serial));
2639         memcpy(subsys->model, id->mn, sizeof(subsys->model));
2640         memcpy(subsys->firmware_rev, id->fr, sizeof(subsys->firmware_rev));
2641         subsys->vendor_id = le16_to_cpu(id->vid);
2642         subsys->cmic = id->cmic;
2643         subsys->awupf = le16_to_cpu(id->awupf);
2644 #ifdef CONFIG_NVME_MULTIPATH
2645         subsys->iopolicy = NVME_IOPOLICY_NUMA;
2646 #endif
2647
2648         subsys->dev.class = nvme_subsys_class;
2649         subsys->dev.release = nvme_release_subsystem;
2650         subsys->dev.groups = nvme_subsys_attrs_groups;
2651         dev_set_name(&subsys->dev, "nvme-subsys%d", ctrl->instance);
2652         device_initialize(&subsys->dev);
2653
2654         mutex_lock(&nvme_subsystems_lock);
2655         found = __nvme_find_get_subsystem(subsys->subnqn);
2656         if (found) {
2657                 put_device(&subsys->dev);
2658                 subsys = found;
2659
2660                 if (!nvme_validate_cntlid(subsys, ctrl, id)) {
2661                         ret = -EINVAL;
2662                         goto out_put_subsystem;
2663                 }
2664         } else {
2665                 ret = device_add(&subsys->dev);
2666                 if (ret) {
2667                         dev_err(ctrl->device,
2668                                 "failed to register subsystem device.\n");
2669                         put_device(&subsys->dev);
2670                         goto out_unlock;
2671                 }
2672                 ida_init(&subsys->ns_ida);
2673                 list_add_tail(&subsys->entry, &nvme_subsystems);
2674         }
2675
2676         ret = sysfs_create_link(&subsys->dev.kobj, &ctrl->device->kobj,
2677                                 dev_name(ctrl->device));
2678         if (ret) {
2679                 dev_err(ctrl->device,
2680                         "failed to create sysfs link from subsystem.\n");
2681                 goto out_put_subsystem;
2682         }
2683
2684         if (!found)
2685                 subsys->instance = ctrl->instance;
2686         ctrl->subsys = subsys;
2687         list_add_tail(&ctrl->subsys_entry, &subsys->ctrls);
2688         mutex_unlock(&nvme_subsystems_lock);
2689         return 0;
2690
2691 out_put_subsystem:
2692         nvme_put_subsystem(subsys);
2693 out_unlock:
2694         mutex_unlock(&nvme_subsystems_lock);
2695         return ret;
2696 }
2697
2698 int nvme_get_log(struct nvme_ctrl *ctrl, u32 nsid, u8 log_page, u8 lsp,
2699                 void *log, size_t size, u64 offset)
2700 {
2701         struct nvme_command c = { };
2702         unsigned long dwlen = size / 4 - 1;
2703
2704         c.get_log_page.opcode = nvme_admin_get_log_page;
2705         c.get_log_page.nsid = cpu_to_le32(nsid);
2706         c.get_log_page.lid = log_page;
2707         c.get_log_page.lsp = lsp;
2708         c.get_log_page.numdl = cpu_to_le16(dwlen & ((1 << 16) - 1));
2709         c.get_log_page.numdu = cpu_to_le16(dwlen >> 16);
2710         c.get_log_page.lpol = cpu_to_le32(lower_32_bits(offset));
2711         c.get_log_page.lpou = cpu_to_le32(upper_32_bits(offset));
2712
2713         return nvme_submit_sync_cmd(ctrl->admin_q, &c, log, size);
2714 }
2715
2716 static int nvme_get_effects_log(struct nvme_ctrl *ctrl)
2717 {
2718         int ret;
2719
2720         if (!ctrl->effects)
2721                 ctrl->effects = kzalloc(sizeof(*ctrl->effects), GFP_KERNEL);
2722
2723         if (!ctrl->effects)
2724                 return 0;
2725
2726         ret = nvme_get_log(ctrl, NVME_NSID_ALL, NVME_LOG_CMD_EFFECTS, 0,
2727                         ctrl->effects, sizeof(*ctrl->effects), 0);
2728         if (ret) {
2729                 kfree(ctrl->effects);
2730                 ctrl->effects = NULL;
2731         }
2732         return ret;
2733 }
2734
2735 /*
2736  * Initialize the cached copies of the Identify data and various controller
2737  * register in our nvme_ctrl structure.  This should be called as soon as
2738  * the admin queue is fully up and running.
2739  */
2740 int nvme_init_identify(struct nvme_ctrl *ctrl)
2741 {
2742         struct nvme_id_ctrl *id;
2743         int ret, page_shift;
2744         u32 max_hw_sectors;
2745         bool prev_apst_enabled;
2746
2747         ret = ctrl->ops->reg_read32(ctrl, NVME_REG_VS, &ctrl->vs);
2748         if (ret) {
2749                 dev_err(ctrl->device, "Reading VS failed (%d)\n", ret);
2750                 return ret;
2751         }
2752         page_shift = NVME_CAP_MPSMIN(ctrl->cap) + 12;
2753         ctrl->sqsize = min_t(int, NVME_CAP_MQES(ctrl->cap), ctrl->sqsize);
2754
2755         if (ctrl->vs >= NVME_VS(1, 1, 0))
2756                 ctrl->subsystem = NVME_CAP_NSSRC(ctrl->cap);
2757
2758         ret = nvme_identify_ctrl(ctrl, &id);
2759         if (ret) {
2760                 dev_err(ctrl->device, "Identify Controller failed (%d)\n", ret);
2761                 return -EIO;
2762         }
2763
2764         if (id->lpa & NVME_CTRL_LPA_CMD_EFFECTS_LOG) {
2765                 ret = nvme_get_effects_log(ctrl);
2766                 if (ret < 0)
2767                         goto out_free;
2768         }
2769
2770         if (!(ctrl->ops->flags & NVME_F_FABRICS))
2771                 ctrl->cntlid = le16_to_cpu(id->cntlid);
2772
2773         if (!ctrl->identified) {
2774                 int i;
2775
2776                 ret = nvme_init_subsystem(ctrl, id);
2777                 if (ret)
2778                         goto out_free;
2779
2780                 /*
2781                  * Check for quirks.  Quirk can depend on firmware version,
2782                  * so, in principle, the set of quirks present can change
2783                  * across a reset.  As a possible future enhancement, we
2784                  * could re-scan for quirks every time we reinitialize
2785                  * the device, but we'd have to make sure that the driver
2786                  * behaves intelligently if the quirks change.
2787                  */
2788                 for (i = 0; i < ARRAY_SIZE(core_quirks); i++) {
2789                         if (quirk_matches(id, &core_quirks[i]))
2790                                 ctrl->quirks |= core_quirks[i].quirks;
2791                 }
2792         }
2793
2794         if (force_apst && (ctrl->quirks & NVME_QUIRK_NO_DEEPEST_PS)) {
2795                 dev_warn(ctrl->device, "forcibly allowing all power states due to nvme_core.force_apst -- use at your own risk\n");
2796                 ctrl->quirks &= ~NVME_QUIRK_NO_DEEPEST_PS;
2797         }
2798
2799         ctrl->crdt[0] = le16_to_cpu(id->crdt1);
2800         ctrl->crdt[1] = le16_to_cpu(id->crdt2);
2801         ctrl->crdt[2] = le16_to_cpu(id->crdt3);
2802
2803         ctrl->oacs = le16_to_cpu(id->oacs);
2804         ctrl->oncs = le16_to_cpu(id->oncs);
2805         ctrl->mtfa = le16_to_cpu(id->mtfa);
2806         ctrl->oaes = le32_to_cpu(id->oaes);
2807         ctrl->wctemp = le16_to_cpu(id->wctemp);
2808         ctrl->cctemp = le16_to_cpu(id->cctemp);
2809
2810         atomic_set(&ctrl->abort_limit, id->acl + 1);
2811         ctrl->vwc = id->vwc;
2812         if (id->mdts)
2813                 max_hw_sectors = 1 << (id->mdts + page_shift - 9);
2814         else
2815                 max_hw_sectors = UINT_MAX;
2816         ctrl->max_hw_sectors =
2817                 min_not_zero(ctrl->max_hw_sectors, max_hw_sectors);
2818
2819         nvme_set_queue_limits(ctrl, ctrl->admin_q);
2820         ctrl->sgls = le32_to_cpu(id->sgls);
2821         ctrl->kas = le16_to_cpu(id->kas);
2822         ctrl->max_namespaces = le32_to_cpu(id->mnan);
2823         ctrl->ctratt = le32_to_cpu(id->ctratt);
2824
2825         if (id->rtd3e) {
2826                 /* us -> s */
2827                 u32 transition_time = le32_to_cpu(id->rtd3e) / 1000000;
2828
2829                 ctrl->shutdown_timeout = clamp_t(unsigned int, transition_time,
2830                                                  shutdown_timeout, 60);
2831
2832                 if (ctrl->shutdown_timeout != shutdown_timeout)
2833                         dev_info(ctrl->device,
2834                                  "Shutdown timeout set to %u seconds\n",
2835                                  ctrl->shutdown_timeout);
2836         } else
2837                 ctrl->shutdown_timeout = shutdown_timeout;
2838
2839         ctrl->npss = id->npss;
2840         ctrl->apsta = id->apsta;
2841         prev_apst_enabled = ctrl->apst_enabled;
2842         if (ctrl->quirks & NVME_QUIRK_NO_APST) {
2843                 if (force_apst && id->apsta) {
2844                         dev_warn(ctrl->device, "forcibly allowing APST due to nvme_core.force_apst -- use at your own risk\n");
2845                         ctrl->apst_enabled = true;
2846                 } else {
2847                         ctrl->apst_enabled = false;
2848                 }
2849         } else {
2850                 ctrl->apst_enabled = id->apsta;
2851         }
2852         memcpy(ctrl->psd, id->psd, sizeof(ctrl->psd));
2853
2854         if (ctrl->ops->flags & NVME_F_FABRICS) {
2855                 ctrl->icdoff = le16_to_cpu(id->icdoff);
2856                 ctrl->ioccsz = le32_to_cpu(id->ioccsz);
2857                 ctrl->iorcsz = le32_to_cpu(id->iorcsz);
2858                 ctrl->maxcmd = le16_to_cpu(id->maxcmd);
2859
2860                 /*
2861                  * In fabrics we need to verify the cntlid matches the
2862                  * admin connect
2863                  */
2864                 if (ctrl->cntlid != le16_to_cpu(id->cntlid)) {
2865                         ret = -EINVAL;
2866                         goto out_free;
2867                 }
2868
2869                 if (!ctrl->opts->discovery_nqn && !ctrl->kas) {
2870                         dev_err(ctrl->device,
2871                                 "keep-alive support is mandatory for fabrics\n");
2872                         ret = -EINVAL;
2873                         goto out_free;
2874                 }
2875         } else {
2876                 ctrl->hmpre = le32_to_cpu(id->hmpre);
2877                 ctrl->hmmin = le32_to_cpu(id->hmmin);
2878                 ctrl->hmminds = le32_to_cpu(id->hmminds);
2879                 ctrl->hmmaxd = le16_to_cpu(id->hmmaxd);
2880         }
2881
2882         ret = nvme_mpath_init(ctrl, id);
2883         kfree(id);
2884
2885         if (ret < 0)
2886                 return ret;
2887
2888         if (ctrl->apst_enabled && !prev_apst_enabled)
2889                 dev_pm_qos_expose_latency_tolerance(ctrl->device);
2890         else if (!ctrl->apst_enabled && prev_apst_enabled)
2891                 dev_pm_qos_hide_latency_tolerance(ctrl->device);
2892
2893         ret = nvme_configure_apst(ctrl);
2894         if (ret < 0)
2895                 return ret;
2896
2897         ret = nvme_configure_timestamp(ctrl);
2898         if (ret < 0)
2899                 return ret;
2900
2901         ret = nvme_configure_directives(ctrl);
2902         if (ret < 0)
2903                 return ret;
2904
2905         ret = nvme_configure_acre(ctrl);
2906         if (ret < 0)
2907                 return ret;
2908
2909         if (!ctrl->identified)
2910                 nvme_hwmon_init(ctrl);
2911
2912         ctrl->identified = true;
2913
2914         return 0;
2915
2916 out_free:
2917         kfree(id);
2918         return ret;
2919 }
2920 EXPORT_SYMBOL_GPL(nvme_init_identify);
2921
2922 static int nvme_dev_open(struct inode *inode, struct file *file)
2923 {
2924         struct nvme_ctrl *ctrl =
2925                 container_of(inode->i_cdev, struct nvme_ctrl, cdev);
2926
2927         switch (ctrl->state) {
2928         case NVME_CTRL_LIVE:
2929                 break;
2930         default:
2931                 return -EWOULDBLOCK;
2932         }
2933
2934         file->private_data = ctrl;
2935         return 0;
2936 }
2937
2938 static int nvme_dev_user_cmd(struct nvme_ctrl *ctrl, void __user *argp)
2939 {
2940         struct nvme_ns *ns;
2941         int ret;
2942
2943         down_read(&ctrl->namespaces_rwsem);
2944         if (list_empty(&ctrl->namespaces)) {
2945                 ret = -ENOTTY;
2946                 goto out_unlock;
2947         }
2948
2949         ns = list_first_entry(&ctrl->namespaces, struct nvme_ns, list);
2950         if (ns != list_last_entry(&ctrl->namespaces, struct nvme_ns, list)) {
2951                 dev_warn(ctrl->device,
2952                         "NVME_IOCTL_IO_CMD not supported when multiple namespaces present!\n");
2953                 ret = -EINVAL;
2954                 goto out_unlock;
2955         }
2956
2957         dev_warn(ctrl->device,
2958                 "using deprecated NVME_IOCTL_IO_CMD ioctl on the char device!\n");
2959         kref_get(&ns->kref);
2960         up_read(&ctrl->namespaces_rwsem);
2961
2962         ret = nvme_user_cmd(ctrl, ns, argp);
2963         nvme_put_ns(ns);
2964         return ret;
2965
2966 out_unlock:
2967         up_read(&ctrl->namespaces_rwsem);
2968         return ret;
2969 }
2970
2971 static long nvme_dev_ioctl(struct file *file, unsigned int cmd,
2972                 unsigned long arg)
2973 {
2974         struct nvme_ctrl *ctrl = file->private_data;
2975         void __user *argp = (void __user *)arg;
2976
2977         switch (cmd) {
2978         case NVME_IOCTL_ADMIN_CMD:
2979                 return nvme_user_cmd(ctrl, NULL, argp);
2980         case NVME_IOCTL_ADMIN64_CMD:
2981                 return nvme_user_cmd64(ctrl, NULL, argp);
2982         case NVME_IOCTL_IO_CMD:
2983                 return nvme_dev_user_cmd(ctrl, argp);
2984         case NVME_IOCTL_RESET:
2985                 dev_warn(ctrl->device, "resetting controller\n");
2986                 return nvme_reset_ctrl_sync(ctrl);
2987         case NVME_IOCTL_SUBSYS_RESET:
2988                 return nvme_reset_subsystem(ctrl);
2989         case NVME_IOCTL_RESCAN:
2990                 nvme_queue_scan(ctrl);
2991                 return 0;
2992         default:
2993                 return -ENOTTY;
2994         }
2995 }
2996
2997 static const struct file_operations nvme_dev_fops = {
2998         .owner          = THIS_MODULE,
2999         .open           = nvme_dev_open,
3000         .unlocked_ioctl = nvme_dev_ioctl,
3001         .compat_ioctl   = compat_ptr_ioctl,
3002 };
3003
3004 static ssize_t nvme_sysfs_reset(struct device *dev,
3005                                 struct device_attribute *attr, const char *buf,
3006                                 size_t count)
3007 {
3008         struct nvme_ctrl *ctrl = dev_get_drvdata(dev);
3009         int ret;
3010
3011         ret = nvme_reset_ctrl_sync(ctrl);
3012         if (ret < 0)
3013                 return ret;
3014         return count;
3015 }
3016 static DEVICE_ATTR(reset_controller, S_IWUSR, NULL, nvme_sysfs_reset);
3017
3018 static ssize_t nvme_sysfs_rescan(struct device *dev,
3019                                 struct device_attribute *attr, const char *buf,
3020                                 size_t count)
3021 {
3022         struct nvme_ctrl *ctrl = dev_get_drvdata(dev);
3023
3024         nvme_queue_scan(ctrl);
3025         return count;
3026 }
3027 static DEVICE_ATTR(rescan_controller, S_IWUSR, NULL, nvme_sysfs_rescan);
3028
3029 static inline struct nvme_ns_head *dev_to_ns_head(struct device *dev)
3030 {
3031         struct gendisk *disk = dev_to_disk(dev);
3032
3033         if (disk->fops == &nvme_fops)
3034                 return nvme_get_ns_from_dev(dev)->head;
3035         else
3036                 return disk->private_data;
3037 }
3038
3039 static ssize_t wwid_show(struct device *dev, struct device_attribute *attr,
3040                 char *buf)
3041 {
3042         struct nvme_ns_head *head = dev_to_ns_head(dev);
3043         struct nvme_ns_ids *ids = &head->ids;
3044         struct nvme_subsystem *subsys = head->subsys;
3045         int serial_len = sizeof(subsys->serial);
3046         int model_len = sizeof(subsys->model);
3047
3048         if (!uuid_is_null(&ids->uuid))
3049                 return sprintf(buf, "uuid.%pU\n", &ids->uuid);
3050
3051         if (memchr_inv(ids->nguid, 0, sizeof(ids->nguid)))
3052                 return sprintf(buf, "eui.%16phN\n", ids->nguid);
3053
3054         if (memchr_inv(ids->eui64, 0, sizeof(ids->eui64)))
3055                 return sprintf(buf, "eui.%8phN\n", ids->eui64);
3056
3057         while (serial_len > 0 && (subsys->serial[serial_len - 1] == ' ' ||
3058                                   subsys->serial[serial_len - 1] == '\0'))
3059                 serial_len--;
3060         while (model_len > 0 && (subsys->model[model_len - 1] == ' ' ||
3061                                  subsys->model[model_len - 1] == '\0'))
3062                 model_len--;
3063
3064         return sprintf(buf, "nvme.%04x-%*phN-%*phN-%08x\n", subsys->vendor_id,
3065                 serial_len, subsys->serial, model_len, subsys->model,
3066                 head->ns_id);
3067 }
3068 static DEVICE_ATTR_RO(wwid);
3069
3070 static ssize_t nguid_show(struct device *dev, struct device_attribute *attr,
3071                 char *buf)
3072 {
3073         return sprintf(buf, "%pU\n", dev_to_ns_head(dev)->ids.nguid);
3074 }
3075 static DEVICE_ATTR_RO(nguid);
3076
3077 static ssize_t uuid_show(struct device *dev, struct device_attribute *attr,
3078                 char *buf)
3079 {
3080         struct nvme_ns_ids *ids = &dev_to_ns_head(dev)->ids;
3081
3082         /* For backward compatibility expose the NGUID to userspace if
3083          * we have no UUID set
3084          */
3085         if (uuid_is_null(&ids->uuid)) {
3086                 printk_ratelimited(KERN_WARNING
3087                                    "No UUID available providing old NGUID\n");
3088                 return sprintf(buf, "%pU\n", ids->nguid);
3089         }
3090         return sprintf(buf, "%pU\n", &ids->uuid);
3091 }
3092 static DEVICE_ATTR_RO(uuid);
3093
3094 static ssize_t eui_show(struct device *dev, struct device_attribute *attr,
3095                 char *buf)
3096 {
3097         return sprintf(buf, "%8ph\n", dev_to_ns_head(dev)->ids.eui64);
3098 }
3099 static DEVICE_ATTR_RO(eui);
3100
3101 static ssize_t nsid_show(struct device *dev, struct device_attribute *attr,
3102                 char *buf)
3103 {
3104         return sprintf(buf, "%d\n", dev_to_ns_head(dev)->ns_id);
3105 }
3106 static DEVICE_ATTR_RO(nsid);
3107
3108 static struct attribute *nvme_ns_id_attrs[] = {
3109         &dev_attr_wwid.attr,
3110         &dev_attr_uuid.attr,
3111         &dev_attr_nguid.attr,
3112         &dev_attr_eui.attr,
3113         &dev_attr_nsid.attr,
3114 #ifdef CONFIG_NVME_MULTIPATH
3115         &dev_attr_ana_grpid.attr,
3116         &dev_attr_ana_state.attr,
3117 #endif
3118         NULL,
3119 };
3120
3121 static umode_t nvme_ns_id_attrs_are_visible(struct kobject *kobj,
3122                 struct attribute *a, int n)
3123 {
3124         struct device *dev = container_of(kobj, struct device, kobj);
3125         struct nvme_ns_ids *ids = &dev_to_ns_head(dev)->ids;
3126
3127         if (a == &dev_attr_uuid.attr) {
3128                 if (uuid_is_null(&ids->uuid) &&
3129                     !memchr_inv(ids->nguid, 0, sizeof(ids->nguid)))
3130                         return 0;
3131         }
3132         if (a == &dev_attr_nguid.attr) {
3133                 if (!memchr_inv(ids->nguid, 0, sizeof(ids->nguid)))
3134                         return 0;
3135         }
3136         if (a == &dev_attr_eui.attr) {
3137                 if (!memchr_inv(ids->eui64, 0, sizeof(ids->eui64)))
3138                         return 0;
3139         }
3140 #ifdef CONFIG_NVME_MULTIPATH
3141         if (a == &dev_attr_ana_grpid.attr || a == &dev_attr_ana_state.attr) {
3142                 if (dev_to_disk(dev)->fops != &nvme_fops) /* per-path attr */
3143                         return 0;
3144                 if (!nvme_ctrl_use_ana(nvme_get_ns_from_dev(dev)->ctrl))
3145                         return 0;
3146         }
3147 #endif
3148         return a->mode;
3149 }
3150
3151 static const struct attribute_group nvme_ns_id_attr_group = {
3152         .attrs          = nvme_ns_id_attrs,
3153         .is_visible     = nvme_ns_id_attrs_are_visible,
3154 };
3155
3156 const struct attribute_group *nvme_ns_id_attr_groups[] = {
3157         &nvme_ns_id_attr_group,
3158 #ifdef CONFIG_NVM
3159         &nvme_nvm_attr_group,
3160 #endif
3161         NULL,
3162 };
3163
3164 #define nvme_show_str_function(field)                                           \
3165 static ssize_t  field##_show(struct device *dev,                                \
3166                             struct device_attribute *attr, char *buf)           \
3167 {                                                                               \
3168         struct nvme_ctrl *ctrl = dev_get_drvdata(dev);                          \
3169         return sprintf(buf, "%.*s\n",                                           \
3170                 (int)sizeof(ctrl->subsys->field), ctrl->subsys->field);         \
3171 }                                                                               \
3172 static DEVICE_ATTR(field, S_IRUGO, field##_show, NULL);
3173
3174 nvme_show_str_function(model);
3175 nvme_show_str_function(serial);
3176 nvme_show_str_function(firmware_rev);
3177
3178 #define nvme_show_int_function(field)                                           \
3179 static ssize_t  field##_show(struct device *dev,                                \
3180                             struct device_attribute *attr, char *buf)           \
3181 {                                                                               \
3182         struct nvme_ctrl *ctrl = dev_get_drvdata(dev);                          \
3183         return sprintf(buf, "%d\n", ctrl->field);       \
3184 }                                                                               \
3185 static DEVICE_ATTR(field, S_IRUGO, field##_show, NULL);
3186
3187 nvme_show_int_function(cntlid);
3188 nvme_show_int_function(numa_node);
3189 nvme_show_int_function(queue_count);
3190 nvme_show_int_function(sqsize);
3191
3192 static ssize_t nvme_sysfs_delete(struct device *dev,
3193                                 struct device_attribute *attr, const char *buf,
3194                                 size_t count)
3195 {
3196         struct nvme_ctrl *ctrl = dev_get_drvdata(dev);
3197
3198         if (device_remove_file_self(dev, attr))
3199                 nvme_delete_ctrl_sync(ctrl);
3200         return count;
3201 }
3202 static DEVICE_ATTR(delete_controller, S_IWUSR, NULL, nvme_sysfs_delete);
3203
3204 static ssize_t nvme_sysfs_show_transport(struct device *dev,
3205                                          struct device_attribute *attr,
3206                                          char *buf)
3207 {
3208         struct nvme_ctrl *ctrl = dev_get_drvdata(dev);
3209
3210         return snprintf(buf, PAGE_SIZE, "%s\n", ctrl->ops->name);
3211 }
3212 static DEVICE_ATTR(transport, S_IRUGO, nvme_sysfs_show_transport, NULL);
3213
3214 static ssize_t nvme_sysfs_show_state(struct device *dev,
3215                                      struct device_attribute *attr,
3216                                      char *buf)
3217 {
3218         struct nvme_ctrl *ctrl = dev_get_drvdata(dev);
3219         static const char *const state_name[] = {
3220                 [NVME_CTRL_NEW]         = "new",
3221                 [NVME_CTRL_LIVE]        = "live",
3222                 [NVME_CTRL_RESETTING]   = "resetting",
3223                 [NVME_CTRL_CONNECTING]  = "connecting",
3224                 [NVME_CTRL_DELETING]    = "deleting",
3225                 [NVME_CTRL_DEAD]        = "dead",
3226         };
3227
3228         if ((unsigned)ctrl->state < ARRAY_SIZE(state_name) &&
3229             state_name[ctrl->state])
3230                 return sprintf(buf, "%s\n", state_name[ctrl->state]);
3231
3232         return sprintf(buf, "unknown state\n");
3233 }
3234
3235 static DEVICE_ATTR(state, S_IRUGO, nvme_sysfs_show_state, NULL);
3236
3237 static ssize_t nvme_sysfs_show_subsysnqn(struct device *dev,
3238                                          struct device_attribute *attr,
3239                                          char *buf)
3240 {
3241         struct nvme_ctrl *ctrl = dev_get_drvdata(dev);
3242
3243         return snprintf(buf, PAGE_SIZE, "%s\n", ctrl->subsys->subnqn);
3244 }
3245 static DEVICE_ATTR(subsysnqn, S_IRUGO, nvme_sysfs_show_subsysnqn, NULL);
3246
3247 static ssize_t nvme_sysfs_show_address(struct device *dev,
3248                                          struct device_attribute *attr,
3249                                          char *buf)
3250 {
3251         struct nvme_ctrl *ctrl = dev_get_drvdata(dev);
3252
3253         return ctrl->ops->get_address(ctrl, buf, PAGE_SIZE);
3254 }
3255 static DEVICE_ATTR(address, S_IRUGO, nvme_sysfs_show_address, NULL);
3256
3257 static struct attribute *nvme_dev_attrs[] = {
3258         &dev_attr_reset_controller.attr,
3259         &dev_attr_rescan_controller.attr,
3260         &dev_attr_model.attr,
3261         &dev_attr_serial.attr,
3262         &dev_attr_firmware_rev.attr,
3263         &dev_attr_cntlid.attr,
3264         &dev_attr_delete_controller.attr,
3265         &dev_attr_transport.attr,
3266         &dev_attr_subsysnqn.attr,
3267         &dev_attr_address.attr,
3268         &dev_attr_state.attr,
3269         &dev_attr_numa_node.attr,
3270         &dev_attr_queue_count.attr,
3271         &dev_attr_sqsize.attr,
3272         NULL
3273 };
3274
3275 static umode_t nvme_dev_attrs_are_visible(struct kobject *kobj,
3276                 struct attribute *a, int n)
3277 {
3278         struct device *dev = container_of(kobj, struct device, kobj);
3279         struct nvme_ctrl *ctrl = dev_get_drvdata(dev);
3280
3281         if (a == &dev_attr_delete_controller.attr && !ctrl->ops->delete_ctrl)
3282                 return 0;
3283         if (a == &dev_attr_address.attr && !ctrl->ops->get_address)
3284                 return 0;
3285
3286         return a->mode;
3287 }
3288
3289 static struct attribute_group nvme_dev_attrs_group = {
3290         .attrs          = nvme_dev_attrs,
3291         .is_visible     = nvme_dev_attrs_are_visible,
3292 };
3293
3294 static const struct attribute_group *nvme_dev_attr_groups[] = {
3295         &nvme_dev_attrs_group,
3296         NULL,
3297 };
3298
3299 static struct nvme_ns_head *__nvme_find_ns_head(struct nvme_subsystem *subsys,
3300                 unsigned nsid)
3301 {
3302         struct nvme_ns_head *h;
3303
3304         lockdep_assert_held(&subsys->lock);
3305
3306         list_for_each_entry(h, &subsys->nsheads, entry) {
3307                 if (h->ns_id == nsid && kref_get_unless_zero(&h->ref))
3308                         return h;
3309         }
3310
3311         return NULL;
3312 }
3313
3314 static int __nvme_check_ids(struct nvme_subsystem *subsys,
3315                 struct nvme_ns_head *new)
3316 {
3317         struct nvme_ns_head *h;
3318
3319         lockdep_assert_held(&subsys->lock);
3320
3321         list_for_each_entry(h, &subsys->nsheads, entry) {
3322                 if (nvme_ns_ids_valid(&new->ids) &&
3323                     !list_empty(&h->list) &&
3324                     nvme_ns_ids_equal(&new->ids, &h->ids))
3325                         return -EINVAL;
3326         }
3327
3328         return 0;
3329 }
3330
3331 static struct nvme_ns_head *nvme_alloc_ns_head(struct nvme_ctrl *ctrl,
3332                 unsigned nsid, struct nvme_id_ns *id)
3333 {
3334         struct nvme_ns_head *head;
3335         size_t size = sizeof(*head);
3336         int ret = -ENOMEM;
3337
3338 #ifdef CONFIG_NVME_MULTIPATH
3339         size += num_possible_nodes() * sizeof(struct nvme_ns *);
3340 #endif
3341
3342         head = kzalloc(size, GFP_KERNEL);
3343         if (!head)
3344                 goto out;
3345         ret = ida_simple_get(&ctrl->subsys->ns_ida, 1, 0, GFP_KERNEL);
3346         if (ret < 0)
3347                 goto out_free_head;
3348         head->instance = ret;
3349         INIT_LIST_HEAD(&head->list);
3350         ret = init_srcu_struct(&head->srcu);
3351         if (ret)
3352                 goto out_ida_remove;
3353         head->subsys = ctrl->subsys;
3354         head->ns_id = nsid;
3355         kref_init(&head->ref);
3356
3357         ret = nvme_report_ns_ids(ctrl, nsid, id, &head->ids);
3358         if (ret)
3359                 goto out_cleanup_srcu;
3360
3361         ret = __nvme_check_ids(ctrl->subsys, head);
3362         if (ret) {
3363                 dev_err(ctrl->device,
3364                         "duplicate IDs for nsid %d\n", nsid);
3365                 goto out_cleanup_srcu;
3366         }
3367
3368         ret = nvme_mpath_alloc_disk(ctrl, head);
3369         if (ret)
3370                 goto out_cleanup_srcu;
3371
3372         list_add_tail(&head->entry, &ctrl->subsys->nsheads);
3373
3374         kref_get(&ctrl->subsys->ref);
3375
3376         return head;
3377 out_cleanup_srcu:
3378         cleanup_srcu_struct(&head->srcu);
3379 out_ida_remove:
3380         ida_simple_remove(&ctrl->subsys->ns_ida, head->instance);
3381 out_free_head:
3382         kfree(head);
3383 out:
3384         if (ret > 0)
3385                 ret = blk_status_to_errno(nvme_error_status(ret));
3386         return ERR_PTR(ret);
3387 }
3388
3389 static int nvme_init_ns_head(struct nvme_ns *ns, unsigned nsid,
3390                 struct nvme_id_ns *id)
3391 {
3392         struct nvme_ctrl *ctrl = ns->ctrl;
3393         bool is_shared = id->nmic & (1 << 0);
3394         struct nvme_ns_head *head = NULL;
3395         int ret = 0;
3396
3397         mutex_lock(&ctrl->subsys->lock);
3398         if (is_shared)
3399                 head = __nvme_find_ns_head(ctrl->subsys, nsid);
3400         if (!head) {
3401                 head = nvme_alloc_ns_head(ctrl, nsid, id);
3402                 if (IS_ERR(head)) {
3403                         ret = PTR_ERR(head);
3404                         goto out_unlock;
3405                 }
3406         } else {
3407                 struct nvme_ns_ids ids;
3408
3409                 ret = nvme_report_ns_ids(ctrl, nsid, id, &ids);
3410                 if (ret)
3411                         goto out_unlock;
3412
3413                 if (!nvme_ns_ids_equal(&head->ids, &ids)) {
3414                         dev_err(ctrl->device,
3415                                 "IDs don't match for shared namespace %d\n",
3416                                         nsid);
3417                         ret = -EINVAL;
3418                         goto out_unlock;
3419                 }
3420         }
3421
3422         list_add_tail(&ns->siblings, &head->list);
3423         ns->head = head;
3424
3425 out_unlock:
3426         mutex_unlock(&ctrl->subsys->lock);
3427         if (ret > 0)
3428                 ret = blk_status_to_errno(nvme_error_status(ret));
3429         return ret;
3430 }
3431
3432 static int ns_cmp(void *priv, struct list_head *a, struct list_head *b)
3433 {
3434         struct nvme_ns *nsa = container_of(a, struct nvme_ns, list);
3435         struct nvme_ns *nsb = container_of(b, struct nvme_ns, list);
3436
3437         return nsa->head->ns_id - nsb->head->ns_id;
3438 }
3439
3440 static struct nvme_ns *nvme_find_get_ns(struct nvme_ctrl *ctrl, unsigned nsid)
3441 {
3442         struct nvme_ns *ns, *ret = NULL;
3443
3444         down_read(&ctrl->namespaces_rwsem);
3445         list_for_each_entry(ns, &ctrl->namespaces, list) {
3446                 if (ns->head->ns_id == nsid) {
3447                         if (!kref_get_unless_zero(&ns->kref))
3448                                 continue;
3449                         ret = ns;
3450                         break;
3451                 }
3452                 if (ns->head->ns_id > nsid)
3453                         break;
3454         }
3455         up_read(&ctrl->namespaces_rwsem);
3456         return ret;
3457 }
3458
3459 static int nvme_setup_streams_ns(struct nvme_ctrl *ctrl, struct nvme_ns *ns)
3460 {
3461         struct streams_directive_params s;
3462         int ret;
3463
3464         if (!ctrl->nr_streams)
3465                 return 0;
3466
3467         ret = nvme_get_stream_params(ctrl, &s, ns->head->ns_id);
3468         if (ret)
3469                 return ret;
3470
3471         ns->sws = le32_to_cpu(s.sws);
3472         ns->sgs = le16_to_cpu(s.sgs);
3473
3474         if (ns->sws) {
3475                 unsigned int bs = 1 << ns->lba_shift;
3476
3477                 blk_queue_io_min(ns->queue, bs * ns->sws);
3478                 if (ns->sgs)
3479                         blk_queue_io_opt(ns->queue, bs * ns->sws * ns->sgs);
3480         }
3481
3482         return 0;
3483 }
3484
3485 static int nvme_alloc_ns(struct nvme_ctrl *ctrl, unsigned nsid)
3486 {
3487         struct nvme_ns *ns;
3488         struct gendisk *disk;
3489         struct nvme_id_ns *id;
3490         char disk_name[DISK_NAME_LEN];
3491         int node = ctrl->numa_node, flags = GENHD_FL_EXT_DEVT, ret;
3492
3493         ns = kzalloc_node(sizeof(*ns), GFP_KERNEL, node);
3494         if (!ns)
3495                 return -ENOMEM;
3496
3497         ns->queue = blk_mq_init_queue(ctrl->tagset);
3498         if (IS_ERR(ns->queue)) {
3499                 ret = PTR_ERR(ns->queue);
3500                 goto out_free_ns;
3501         }
3502
3503         if (ctrl->opts && ctrl->opts->data_digest)
3504                 ns->queue->backing_dev_info->capabilities
3505                         |= BDI_CAP_STABLE_WRITES;
3506
3507         blk_queue_flag_set(QUEUE_FLAG_NONROT, ns->queue);
3508         if (ctrl->ops->flags & NVME_F_PCI_P2PDMA)
3509                 blk_queue_flag_set(QUEUE_FLAG_PCI_P2PDMA, ns->queue);
3510
3511         ns->queue->queuedata = ns;
3512         ns->ctrl = ctrl;
3513
3514         kref_init(&ns->kref);
3515         ns->lba_shift = 9; /* set to a default value for 512 until disk is validated */
3516
3517         blk_queue_logical_block_size(ns->queue, 1 << ns->lba_shift);
3518         nvme_set_queue_limits(ctrl, ns->queue);
3519
3520         ret = nvme_identify_ns(ctrl, nsid, &id);
3521         if (ret)
3522                 goto out_free_queue;
3523
3524         if (id->ncap == 0) {
3525                 ret = -EINVAL;
3526                 goto out_free_id;
3527         }
3528
3529         ret = nvme_init_ns_head(ns, nsid, id);
3530         if (ret)
3531                 goto out_free_id;
3532         nvme_setup_streams_ns(ctrl, ns);
3533         nvme_set_disk_name(disk_name, ns, ctrl, &flags);
3534
3535         disk = alloc_disk_node(0, node);
3536         if (!disk) {
3537                 ret = -ENOMEM;
3538                 goto out_unlink_ns;
3539         }
3540
3541         disk->fops = &nvme_fops;
3542         disk->private_data = ns;
3543         disk->queue = ns->queue;
3544         disk->flags = flags;
3545         memcpy(disk->disk_name, disk_name, DISK_NAME_LEN);
3546         ns->disk = disk;
3547
3548         __nvme_revalidate_disk(disk, id);
3549
3550         if ((ctrl->quirks & NVME_QUIRK_LIGHTNVM) && id->vs[0] == 0x1) {
3551                 ret = nvme_nvm_register(ns, disk_name, node);
3552                 if (ret) {
3553                         dev_warn(ctrl->device, "LightNVM init failure\n");
3554                         goto out_put_disk;
3555                 }
3556         }
3557
3558         down_write(&ctrl->namespaces_rwsem);
3559         list_add_tail(&ns->list, &ctrl->namespaces);
3560         up_write(&ctrl->namespaces_rwsem);
3561
3562         nvme_get_ctrl(ctrl);
3563
3564         device_add_disk(ctrl->device, ns->disk, nvme_ns_id_attr_groups);
3565
3566         nvme_mpath_add_disk(ns, id);
3567         nvme_fault_inject_init(&ns->fault_inject, ns->disk->disk_name);
3568         kfree(id);
3569
3570         return 0;
3571  out_put_disk:
3572         put_disk(ns->disk);
3573  out_unlink_ns:
3574         mutex_lock(&ctrl->subsys->lock);
3575         list_del_rcu(&ns->siblings);
3576         mutex_unlock(&ctrl->subsys->lock);
3577         nvme_put_ns_head(ns->head);
3578  out_free_id:
3579         kfree(id);
3580  out_free_queue:
3581         blk_cleanup_queue(ns->queue);
3582  out_free_ns:
3583         kfree(ns);
3584         if (ret > 0)
3585                 ret = blk_status_to_errno(nvme_error_status(ret));
3586         return ret;
3587 }
3588
3589 static void nvme_ns_remove(struct nvme_ns *ns)
3590 {
3591         if (test_and_set_bit(NVME_NS_REMOVING, &ns->flags))
3592                 return;
3593
3594         nvme_fault_inject_fini(&ns->fault_inject);
3595
3596         mutex_lock(&ns->ctrl->subsys->lock);
3597         list_del_rcu(&ns->siblings);
3598         mutex_unlock(&ns->ctrl->subsys->lock);
3599         synchronize_rcu(); /* guarantee not available in head->list */
3600         nvme_mpath_clear_current_path(ns);
3601         synchronize_srcu(&ns->head->srcu); /* wait for concurrent submissions */
3602
3603         if (ns->disk && ns->disk->flags & GENHD_FL_UP) {
3604                 del_gendisk(ns->disk);
3605                 blk_cleanup_queue(ns->queue);
3606                 if (blk_get_integrity(ns->disk))
3607                         blk_integrity_unregister(ns->disk);
3608         }
3609
3610         down_write(&ns->ctrl->namespaces_rwsem);
3611         list_del_init(&ns->list);
3612         up_write(&ns->ctrl->namespaces_rwsem);
3613
3614         nvme_mpath_check_last_path(ns);
3615         nvme_put_ns(ns);
3616 }
3617
3618 static void nvme_validate_ns(struct nvme_ctrl *ctrl, unsigned nsid)
3619 {
3620         struct nvme_ns *ns;
3621
3622         ns = nvme_find_get_ns(ctrl, nsid);
3623         if (ns) {
3624                 if (ns->disk && revalidate_disk(ns->disk))
3625                         nvme_ns_remove(ns);
3626                 nvme_put_ns(ns);
3627         } else
3628                 nvme_alloc_ns(ctrl, nsid);
3629 }
3630
3631 static void nvme_remove_invalid_namespaces(struct nvme_ctrl *ctrl,
3632                                         unsigned nsid)
3633 {
3634         struct nvme_ns *ns, *next;
3635         LIST_HEAD(rm_list);
3636
3637         down_write(&ctrl->namespaces_rwsem);
3638         list_for_each_entry_safe(ns, next, &ctrl->namespaces, list) {
3639                 if (ns->head->ns_id > nsid || test_bit(NVME_NS_DEAD, &ns->flags))
3640                         list_move_tail(&ns->list, &rm_list);
3641         }
3642         up_write(&ctrl->namespaces_rwsem);
3643
3644         list_for_each_entry_safe(ns, next, &rm_list, list)
3645                 nvme_ns_remove(ns);
3646
3647 }
3648
3649 static int nvme_scan_ns_list(struct nvme_ctrl *ctrl, unsigned nn)
3650 {
3651         struct nvme_ns *ns;
3652         __le32 *ns_list;
3653         unsigned i, j, nsid, prev = 0;
3654         unsigned num_lists = DIV_ROUND_UP_ULL((u64)nn, 1024);
3655         int ret = 0;
3656
3657         ns_list = kzalloc(NVME_IDENTIFY_DATA_SIZE, GFP_KERNEL);
3658         if (!ns_list)
3659                 return -ENOMEM;
3660
3661         for (i = 0; i < num_lists; i++) {
3662                 ret = nvme_identify_ns_list(ctrl, prev, ns_list);
3663                 if (ret)
3664                         goto free;
3665
3666                 for (j = 0; j < min(nn, 1024U); j++) {
3667                         nsid = le32_to_cpu(ns_list[j]);
3668                         if (!nsid)
3669                                 goto out;
3670
3671                         nvme_validate_ns(ctrl, nsid);
3672
3673                         while (++prev < nsid) {
3674                                 ns = nvme_find_get_ns(ctrl, prev);
3675                                 if (ns) {
3676                                         nvme_ns_remove(ns);
3677                                         nvme_put_ns(ns);
3678                                 }
3679                         }
3680                 }
3681                 nn -= j;
3682         }
3683  out:
3684         nvme_remove_invalid_namespaces(ctrl, prev);
3685  free:
3686         kfree(ns_list);
3687         return ret;
3688 }
3689
3690 static void nvme_scan_ns_sequential(struct nvme_ctrl *ctrl, unsigned nn)
3691 {
3692         unsigned i;
3693
3694         for (i = 1; i <= nn; i++)
3695                 nvme_validate_ns(ctrl, i);
3696
3697         nvme_remove_invalid_namespaces(ctrl, nn);
3698 }
3699
3700 static void nvme_clear_changed_ns_log(struct nvme_ctrl *ctrl)
3701 {
3702         size_t log_size = NVME_MAX_CHANGED_NAMESPACES * sizeof(__le32);
3703         __le32 *log;
3704         int error;
3705
3706         log = kzalloc(log_size, GFP_KERNEL);
3707         if (!log)
3708                 return;
3709
3710         /*
3711          * We need to read the log to clear the AEN, but we don't want to rely
3712          * on it for the changed namespace information as userspace could have
3713          * raced with us in reading the log page, which could cause us to miss
3714          * updates.
3715          */
3716         error = nvme_get_log(ctrl, NVME_NSID_ALL, NVME_LOG_CHANGED_NS, 0, log,
3717                         log_size, 0);
3718         if (error)
3719                 dev_warn(ctrl->device,
3720                         "reading changed ns log failed: %d\n", error);
3721
3722         kfree(log);
3723 }
3724
3725 static void nvme_scan_work(struct work_struct *work)
3726 {
3727         struct nvme_ctrl *ctrl =
3728                 container_of(work, struct nvme_ctrl, scan_work);
3729         struct nvme_id_ctrl *id;
3730         unsigned nn;
3731
3732         /* No tagset on a live ctrl means IO queues could not created */
3733         if (ctrl->state != NVME_CTRL_LIVE || !ctrl->tagset)
3734                 return;
3735
3736         if (test_and_clear_bit(NVME_AER_NOTICE_NS_CHANGED, &ctrl->events)) {
3737                 dev_info(ctrl->device, "rescanning namespaces.\n");
3738                 nvme_clear_changed_ns_log(ctrl);
3739         }
3740
3741         if (nvme_identify_ctrl(ctrl, &id))
3742                 return;
3743
3744         mutex_lock(&ctrl->scan_lock);
3745         nn = le32_to_cpu(id->nn);
3746         if (ctrl->vs >= NVME_VS(1, 1, 0) &&
3747             !(ctrl->quirks & NVME_QUIRK_IDENTIFY_CNS)) {
3748                 if (!nvme_scan_ns_list(ctrl, nn))
3749                         goto out_free_id;
3750         }
3751         nvme_scan_ns_sequential(ctrl, nn);
3752 out_free_id:
3753         mutex_unlock(&ctrl->scan_lock);
3754         kfree(id);
3755         down_write(&ctrl->namespaces_rwsem);
3756         list_sort(NULL, &ctrl->namespaces, ns_cmp);
3757         up_write(&ctrl->namespaces_rwsem);
3758 }
3759
3760 /*
3761  * This function iterates the namespace list unlocked to allow recovery from
3762  * controller failure. It is up to the caller to ensure the namespace list is
3763  * not modified by scan work while this function is executing.
3764  */
3765 void nvme_remove_namespaces(struct nvme_ctrl *ctrl)
3766 {
3767         struct nvme_ns *ns, *next;
3768         LIST_HEAD(ns_list);
3769
3770         /*
3771          * make sure to requeue I/O to all namespaces as these
3772          * might result from the scan itself and must complete
3773          * for the scan_work to make progress
3774          */
3775         nvme_mpath_clear_ctrl_paths(ctrl);
3776
3777         /* prevent racing with ns scanning */
3778         flush_work(&ctrl->scan_work);
3779
3780         /*
3781          * The dead states indicates the controller was not gracefully
3782          * disconnected. In that case, we won't be able to flush any data while
3783          * removing the namespaces' disks; fail all the queues now to avoid
3784          * potentially having to clean up the failed sync later.
3785          */
3786         if (ctrl->state == NVME_CTRL_DEAD)
3787                 nvme_kill_queues(ctrl);
3788
3789         down_write(&ctrl->namespaces_rwsem);
3790         list_splice_init(&ctrl->namespaces, &ns_list);
3791         up_write(&ctrl->namespaces_rwsem);
3792
3793         list_for_each_entry_safe(ns, next, &ns_list, list)
3794                 nvme_ns_remove(ns);
3795 }
3796 EXPORT_SYMBOL_GPL(nvme_remove_namespaces);
3797
3798 static int nvme_class_uevent(struct device *dev, struct kobj_uevent_env *env)
3799 {
3800         struct nvme_ctrl *ctrl =
3801                 container_of(dev, struct nvme_ctrl, ctrl_device);
3802         struct nvmf_ctrl_options *opts = ctrl->opts;
3803         int ret;
3804
3805         ret = add_uevent_var(env, "NVME_TRTYPE=%s", ctrl->ops->name);
3806         if (ret)
3807                 return ret;
3808
3809         if (opts) {
3810                 ret = add_uevent_var(env, "NVME_TRADDR=%s", opts->traddr);
3811                 if (ret)
3812                         return ret;
3813
3814                 ret = add_uevent_var(env, "NVME_TRSVCID=%s",
3815                                 opts->trsvcid ?: "none");
3816                 if (ret)
3817                         return ret;
3818
3819                 ret = add_uevent_var(env, "NVME_HOST_TRADDR=%s",
3820                                 opts->host_traddr ?: "none");
3821         }
3822         return ret;
3823 }
3824
3825 static void nvme_aen_uevent(struct nvme_ctrl *ctrl)
3826 {
3827         char *envp[2] = { NULL, NULL };
3828         u32 aen_result = ctrl->aen_result;
3829
3830         ctrl->aen_result = 0;
3831         if (!aen_result)
3832                 return;
3833
3834         envp[0] = kasprintf(GFP_KERNEL, "NVME_AEN=%#08x", aen_result);
3835         if (!envp[0])
3836                 return;
3837         kobject_uevent_env(&ctrl->device->kobj, KOBJ_CHANGE, envp);
3838         kfree(envp[0]);
3839 }
3840
3841 static void nvme_async_event_work(struct work_struct *work)
3842 {
3843         struct nvme_ctrl *ctrl =
3844                 container_of(work, struct nvme_ctrl, async_event_work);
3845
3846         nvme_aen_uevent(ctrl);
3847         ctrl->ops->submit_async_event(ctrl);
3848 }
3849
3850 static bool nvme_ctrl_pp_status(struct nvme_ctrl *ctrl)
3851 {
3852
3853         u32 csts;
3854
3855         if (ctrl->ops->reg_read32(ctrl, NVME_REG_CSTS, &csts))
3856                 return false;
3857
3858         if (csts == ~0)
3859                 return false;
3860
3861         return ((ctrl->ctrl_config & NVME_CC_ENABLE) && (csts & NVME_CSTS_PP));
3862 }
3863
3864 static void nvme_get_fw_slot_info(struct nvme_ctrl *ctrl)
3865 {
3866         struct nvme_fw_slot_info_log *log;
3867
3868         log = kmalloc(sizeof(*log), GFP_KERNEL);
3869         if (!log)
3870                 return;
3871
3872         if (nvme_get_log(ctrl, NVME_NSID_ALL, 0, NVME_LOG_FW_SLOT, log,
3873                         sizeof(*log), 0))
3874                 dev_warn(ctrl->device, "Get FW SLOT INFO log error\n");
3875         kfree(log);
3876 }
3877
3878 static void nvme_fw_act_work(struct work_struct *work)
3879 {
3880         struct nvme_ctrl *ctrl = container_of(work,
3881                                 struct nvme_ctrl, fw_act_work);
3882         unsigned long fw_act_timeout;
3883
3884         if (ctrl->mtfa)
3885                 fw_act_timeout = jiffies +
3886                                 msecs_to_jiffies(ctrl->mtfa * 100);
3887         else
3888                 fw_act_timeout = jiffies +
3889                                 msecs_to_jiffies(admin_timeout * 1000);
3890
3891         nvme_stop_queues(ctrl);
3892         while (nvme_ctrl_pp_status(ctrl)) {
3893                 if (time_after(jiffies, fw_act_timeout)) {
3894                         dev_warn(ctrl->device,
3895                                 "Fw activation timeout, reset controller\n");
3896                         nvme_try_sched_reset(ctrl);
3897                         return;
3898                 }
3899                 msleep(100);
3900         }
3901
3902         if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_LIVE))
3903                 return;
3904
3905         nvme_start_queues(ctrl);
3906         /* read FW slot information to clear the AER */
3907         nvme_get_fw_slot_info(ctrl);
3908 }
3909
3910 static void nvme_handle_aen_notice(struct nvme_ctrl *ctrl, u32 result)
3911 {
3912         u32 aer_notice_type = (result & 0xff00) >> 8;
3913
3914         trace_nvme_async_event(ctrl, aer_notice_type);
3915
3916         switch (aer_notice_type) {
3917         case NVME_AER_NOTICE_NS_CHANGED:
3918                 set_bit(NVME_AER_NOTICE_NS_CHANGED, &ctrl->events);
3919                 nvme_queue_scan(ctrl);
3920                 break;
3921         case NVME_AER_NOTICE_FW_ACT_STARTING:
3922                 /*
3923                  * We are (ab)using the RESETTING state to prevent subsequent
3924                  * recovery actions from interfering with the controller's
3925                  * firmware activation.
3926                  */
3927                 if (nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING))
3928                         queue_work(nvme_wq, &ctrl->fw_act_work);
3929                 break;
3930 #ifdef CONFIG_NVME_MULTIPATH
3931         case NVME_AER_NOTICE_ANA:
3932                 if (!ctrl->ana_log_buf)
3933                         break;
3934                 queue_work(nvme_wq, &ctrl->ana_work);
3935                 break;
3936 #endif
3937         case NVME_AER_NOTICE_DISC_CHANGED:
3938                 ctrl->aen_result = result;
3939                 break;
3940         default:
3941                 dev_warn(ctrl->device, "async event result %08x\n", result);
3942         }
3943 }
3944
3945 void nvme_complete_async_event(struct nvme_ctrl *ctrl, __le16 status,
3946                 volatile union nvme_result *res)
3947 {
3948         u32 result = le32_to_cpu(res->u32);
3949         u32 aer_type = result & 0x07;
3950
3951         if (le16_to_cpu(status) >> 1 != NVME_SC_SUCCESS)
3952                 return;
3953
3954         switch (aer_type) {
3955         case NVME_AER_NOTICE:
3956                 nvme_handle_aen_notice(ctrl, result);
3957                 break;
3958         case NVME_AER_ERROR:
3959         case NVME_AER_SMART:
3960         case NVME_AER_CSS:
3961         case NVME_AER_VS:
3962                 trace_nvme_async_event(ctrl, aer_type);
3963                 ctrl->aen_result = result;
3964                 break;
3965         default:
3966                 break;
3967         }
3968         queue_work(nvme_wq, &ctrl->async_event_work);
3969 }
3970 EXPORT_SYMBOL_GPL(nvme_complete_async_event);
3971
3972 void nvme_stop_ctrl(struct nvme_ctrl *ctrl)
3973 {
3974         nvme_mpath_stop(ctrl);
3975         nvme_stop_keep_alive(ctrl);
3976         flush_work(&ctrl->async_event_work);
3977         cancel_work_sync(&ctrl->fw_act_work);
3978 }
3979 EXPORT_SYMBOL_GPL(nvme_stop_ctrl);
3980
3981 void nvme_start_ctrl(struct nvme_ctrl *ctrl)
3982 {
3983         if (ctrl->kato)
3984                 nvme_start_keep_alive(ctrl);
3985
3986         nvme_enable_aen(ctrl);
3987
3988         if (ctrl->queue_count > 1) {
3989                 nvme_queue_scan(ctrl);
3990                 nvme_start_queues(ctrl);
3991         }
3992 }
3993 EXPORT_SYMBOL_GPL(nvme_start_ctrl);
3994
3995 void nvme_uninit_ctrl(struct nvme_ctrl *ctrl)
3996 {
3997         nvme_fault_inject_fini(&ctrl->fault_inject);
3998         dev_pm_qos_hide_latency_tolerance(ctrl->device);
3999         cdev_device_del(&ctrl->cdev, ctrl->device);
4000 }
4001 EXPORT_SYMBOL_GPL(nvme_uninit_ctrl);
4002
4003 static void nvme_free_ctrl(struct device *dev)
4004 {
4005         struct nvme_ctrl *ctrl =
4006                 container_of(dev, struct nvme_ctrl, ctrl_device);
4007         struct nvme_subsystem *subsys = ctrl->subsys;
4008
4009         if (subsys && ctrl->instance != subsys->instance)
4010                 ida_simple_remove(&nvme_instance_ida, ctrl->instance);
4011
4012         kfree(ctrl->effects);
4013         nvme_mpath_uninit(ctrl);
4014         __free_page(ctrl->discard_page);
4015
4016         if (subsys) {
4017                 mutex_lock(&nvme_subsystems_lock);
4018                 list_del(&ctrl->subsys_entry);
4019                 sysfs_remove_link(&subsys->dev.kobj, dev_name(ctrl->device));
4020                 mutex_unlock(&nvme_subsystems_lock);
4021         }
4022
4023         ctrl->ops->free_ctrl(ctrl);
4024
4025         if (subsys)
4026                 nvme_put_subsystem(subsys);
4027 }
4028
4029 /*
4030  * Initialize a NVMe controller structures.  This needs to be called during
4031  * earliest initialization so that we have the initialized structured around
4032  * during probing.
4033  */
4034 int nvme_init_ctrl(struct nvme_ctrl *ctrl, struct device *dev,
4035                 const struct nvme_ctrl_ops *ops, unsigned long quirks)
4036 {
4037         int ret;
4038
4039         ctrl->state = NVME_CTRL_NEW;
4040         spin_lock_init(&ctrl->lock);
4041         mutex_init(&ctrl->scan_lock);
4042         INIT_LIST_HEAD(&ctrl->namespaces);
4043         init_rwsem(&ctrl->namespaces_rwsem);
4044         ctrl->dev = dev;
4045         ctrl->ops = ops;
4046         ctrl->quirks = quirks;
4047         INIT_WORK(&ctrl->scan_work, nvme_scan_work);
4048         INIT_WORK(&ctrl->async_event_work, nvme_async_event_work);
4049         INIT_WORK(&ctrl->fw_act_work, nvme_fw_act_work);
4050         INIT_WORK(&ctrl->delete_work, nvme_delete_ctrl_work);
4051         init_waitqueue_head(&ctrl->state_wq);
4052
4053         INIT_DELAYED_WORK(&ctrl->ka_work, nvme_keep_alive_work);
4054         memset(&ctrl->ka_cmd, 0, sizeof(ctrl->ka_cmd));
4055         ctrl->ka_cmd.common.opcode = nvme_admin_keep_alive;
4056
4057         BUILD_BUG_ON(NVME_DSM_MAX_RANGES * sizeof(struct nvme_dsm_range) >
4058                         PAGE_SIZE);
4059         ctrl->discard_page = alloc_page(GFP_KERNEL);
4060         if (!ctrl->discard_page) {
4061                 ret = -ENOMEM;
4062                 goto out;
4063         }
4064
4065         ret = ida_simple_get(&nvme_instance_ida, 0, 0, GFP_KERNEL);
4066         if (ret < 0)
4067                 goto out;
4068         ctrl->instance = ret;
4069
4070         device_initialize(&ctrl->ctrl_device);
4071         ctrl->device = &ctrl->ctrl_device;
4072         ctrl->device->devt = MKDEV(MAJOR(nvme_chr_devt), ctrl->instance);
4073         ctrl->device->class = nvme_class;
4074         ctrl->device->parent = ctrl->dev;
4075         ctrl->device->groups = nvme_dev_attr_groups;
4076         ctrl->device->release = nvme_free_ctrl;
4077         dev_set_drvdata(ctrl->device, ctrl);
4078         ret = dev_set_name(ctrl->device, "nvme%d", ctrl->instance);
4079         if (ret)
4080                 goto out_release_instance;
4081
4082         cdev_init(&ctrl->cdev, &nvme_dev_fops);
4083         ctrl->cdev.owner = ops->module;
4084         ret = cdev_device_add(&ctrl->cdev, ctrl->device);
4085         if (ret)
4086                 goto out_free_name;
4087
4088         /*
4089          * Initialize latency tolerance controls.  The sysfs files won't
4090          * be visible to userspace unless the device actually supports APST.
4091          */
4092         ctrl->device->power.set_latency_tolerance = nvme_set_latency_tolerance;
4093         dev_pm_qos_update_user_latency_tolerance(ctrl->device,
4094                 min(default_ps_max_latency_us, (unsigned long)S32_MAX));
4095
4096         nvme_fault_inject_init(&ctrl->fault_inject, dev_name(ctrl->device));
4097
4098         return 0;
4099 out_free_name:
4100         kfree_const(ctrl->device->kobj.name);
4101 out_release_instance:
4102         ida_simple_remove(&nvme_instance_ida, ctrl->instance);
4103 out:
4104         if (ctrl->discard_page)
4105                 __free_page(ctrl->discard_page);
4106         return ret;
4107 }
4108 EXPORT_SYMBOL_GPL(nvme_init_ctrl);
4109
4110 /**
4111  * nvme_kill_queues(): Ends all namespace queues
4112  * @ctrl: the dead controller that needs to end
4113  *
4114  * Call this function when the driver determines it is unable to get the
4115  * controller in a state capable of servicing IO.
4116  */
4117 void nvme_kill_queues(struct nvme_ctrl *ctrl)
4118 {
4119         struct nvme_ns *ns;
4120
4121         down_read(&ctrl->namespaces_rwsem);
4122
4123         /* Forcibly unquiesce queues to avoid blocking dispatch */
4124         if (ctrl->admin_q && !blk_queue_dying(ctrl->admin_q))
4125                 blk_mq_unquiesce_queue(ctrl->admin_q);
4126
4127         list_for_each_entry(ns, &ctrl->namespaces, list)
4128                 nvme_set_queue_dying(ns);
4129
4130         up_read(&ctrl->namespaces_rwsem);
4131 }
4132 EXPORT_SYMBOL_GPL(nvme_kill_queues);
4133
4134 void nvme_unfreeze(struct nvme_ctrl *ctrl)
4135 {
4136         struct nvme_ns *ns;
4137
4138         down_read(&ctrl->namespaces_rwsem);
4139         list_for_each_entry(ns, &ctrl->namespaces, list)
4140                 blk_mq_unfreeze_queue(ns->queue);
4141         up_read(&ctrl->namespaces_rwsem);
4142 }
4143 EXPORT_SYMBOL_GPL(nvme_unfreeze);
4144
4145 void nvme_wait_freeze_timeout(struct nvme_ctrl *ctrl, long timeout)
4146 {
4147         struct nvme_ns *ns;
4148
4149         down_read(&ctrl->namespaces_rwsem);
4150         list_for_each_entry(ns, &ctrl->namespaces, list) {
4151                 timeout = blk_mq_freeze_queue_wait_timeout(ns->queue, timeout);
4152                 if (timeout <= 0)
4153                         break;
4154         }
4155         up_read(&ctrl->namespaces_rwsem);
4156 }
4157 EXPORT_SYMBOL_GPL(nvme_wait_freeze_timeout);
4158
4159 void nvme_wait_freeze(struct nvme_ctrl *ctrl)
4160 {
4161         struct nvme_ns *ns;
4162
4163         down_read(&ctrl->namespaces_rwsem);
4164         list_for_each_entry(ns, &ctrl->namespaces, list)
4165                 blk_mq_freeze_queue_wait(ns->queue);
4166         up_read(&ctrl->namespaces_rwsem);
4167 }
4168 EXPORT_SYMBOL_GPL(nvme_wait_freeze);
4169
4170 void nvme_start_freeze(struct nvme_ctrl *ctrl)
4171 {
4172         struct nvme_ns *ns;
4173
4174         down_read(&ctrl->namespaces_rwsem);
4175         list_for_each_entry(ns, &ctrl->namespaces, list)
4176                 blk_freeze_queue_start(ns->queue);
4177         up_read(&ctrl->namespaces_rwsem);
4178 }
4179 EXPORT_SYMBOL_GPL(nvme_start_freeze);
4180
4181 void nvme_stop_queues(struct nvme_ctrl *ctrl)
4182 {
4183         struct nvme_ns *ns;
4184
4185         down_read(&ctrl->namespaces_rwsem);
4186         list_for_each_entry(ns, &ctrl->namespaces, list)
4187                 blk_mq_quiesce_queue(ns->queue);
4188         up_read(&ctrl->namespaces_rwsem);
4189 }
4190 EXPORT_SYMBOL_GPL(nvme_stop_queues);
4191
4192 void nvme_start_queues(struct nvme_ctrl *ctrl)
4193 {
4194         struct nvme_ns *ns;
4195
4196         down_read(&ctrl->namespaces_rwsem);
4197         list_for_each_entry(ns, &ctrl->namespaces, list)
4198                 blk_mq_unquiesce_queue(ns->queue);
4199         up_read(&ctrl->namespaces_rwsem);
4200 }
4201 EXPORT_SYMBOL_GPL(nvme_start_queues);
4202
4203
4204 void nvme_sync_queues(struct nvme_ctrl *ctrl)
4205 {
4206         struct nvme_ns *ns;
4207
4208         down_read(&ctrl->namespaces_rwsem);
4209         list_for_each_entry(ns, &ctrl->namespaces, list)
4210                 blk_sync_queue(ns->queue);
4211         up_read(&ctrl->namespaces_rwsem);
4212
4213         if (ctrl->admin_q)
4214                 blk_sync_queue(ctrl->admin_q);
4215 }
4216 EXPORT_SYMBOL_GPL(nvme_sync_queues);
4217
4218 /*
4219  * Check we didn't inadvertently grow the command structure sizes:
4220  */
4221 static inline void _nvme_check_size(void)
4222 {
4223         BUILD_BUG_ON(sizeof(struct nvme_common_command) != 64);
4224         BUILD_BUG_ON(sizeof(struct nvme_rw_command) != 64);
4225         BUILD_BUG_ON(sizeof(struct nvme_identify) != 64);
4226         BUILD_BUG_ON(sizeof(struct nvme_features) != 64);
4227         BUILD_BUG_ON(sizeof(struct nvme_download_firmware) != 64);
4228         BUILD_BUG_ON(sizeof(struct nvme_format_cmd) != 64);
4229         BUILD_BUG_ON(sizeof(struct nvme_dsm_cmd) != 64);
4230         BUILD_BUG_ON(sizeof(struct nvme_write_zeroes_cmd) != 64);
4231         BUILD_BUG_ON(sizeof(struct nvme_abort_cmd) != 64);
4232         BUILD_BUG_ON(sizeof(struct nvme_get_log_page_command) != 64);
4233         BUILD_BUG_ON(sizeof(struct nvme_command) != 64);
4234         BUILD_BUG_ON(sizeof(struct nvme_id_ctrl) != NVME_IDENTIFY_DATA_SIZE);
4235         BUILD_BUG_ON(sizeof(struct nvme_id_ns) != NVME_IDENTIFY_DATA_SIZE);
4236         BUILD_BUG_ON(sizeof(struct nvme_lba_range_type) != 64);
4237         BUILD_BUG_ON(sizeof(struct nvme_smart_log) != 512);
4238         BUILD_BUG_ON(sizeof(struct nvme_dbbuf) != 64);
4239         BUILD_BUG_ON(sizeof(struct nvme_directive_cmd) != 64);
4240 }
4241
4242
4243 static int __init nvme_core_init(void)
4244 {
4245         int result = -ENOMEM;
4246
4247         _nvme_check_size();
4248
4249         nvme_wq = alloc_workqueue("nvme-wq",
4250                         WQ_UNBOUND | WQ_MEM_RECLAIM | WQ_SYSFS, 0);
4251         if (!nvme_wq)
4252                 goto out;
4253
4254         nvme_reset_wq = alloc_workqueue("nvme-reset-wq",
4255                         WQ_UNBOUND | WQ_MEM_RECLAIM | WQ_SYSFS, 0);
4256         if (!nvme_reset_wq)
4257                 goto destroy_wq;
4258
4259         nvme_delete_wq = alloc_workqueue("nvme-delete-wq",
4260                         WQ_UNBOUND | WQ_MEM_RECLAIM | WQ_SYSFS, 0);
4261         if (!nvme_delete_wq)
4262                 goto destroy_reset_wq;
4263
4264         result = alloc_chrdev_region(&nvme_chr_devt, 0, NVME_MINORS, "nvme");
4265         if (result < 0)
4266                 goto destroy_delete_wq;
4267
4268         nvme_class = class_create(THIS_MODULE, "nvme");
4269         if (IS_ERR(nvme_class)) {
4270                 result = PTR_ERR(nvme_class);
4271                 goto unregister_chrdev;
4272         }
4273         nvme_class->dev_uevent = nvme_class_uevent;
4274
4275         nvme_subsys_class = class_create(THIS_MODULE, "nvme-subsystem");
4276         if (IS_ERR(nvme_subsys_class)) {
4277                 result = PTR_ERR(nvme_subsys_class);
4278                 goto destroy_class;
4279         }
4280         return 0;
4281
4282 destroy_class:
4283         class_destroy(nvme_class);
4284 unregister_chrdev:
4285         unregister_chrdev_region(nvme_chr_devt, NVME_MINORS);
4286 destroy_delete_wq:
4287         destroy_workqueue(nvme_delete_wq);
4288 destroy_reset_wq:
4289         destroy_workqueue(nvme_reset_wq);
4290 destroy_wq:
4291         destroy_workqueue(nvme_wq);
4292 out:
4293         return result;
4294 }
4295
4296 static void __exit nvme_core_exit(void)
4297 {
4298         class_destroy(nvme_subsys_class);
4299         class_destroy(nvme_class);
4300         unregister_chrdev_region(nvme_chr_devt, NVME_MINORS);
4301         destroy_workqueue(nvme_delete_wq);
4302         destroy_workqueue(nvme_reset_wq);
4303         destroy_workqueue(nvme_wq);
4304 }
4305
4306 MODULE_LICENSE("GPL");
4307 MODULE_VERSION("1.0");
4308 module_init(nvme_core_init);
4309 module_exit(nvme_core_exit);