drivers/nvme/host/nvme.h

   1 /*
   2  * Copyright (c) 2011-2014, Intel Corporation.
   3  *
   4  * This program is free software; you can redistribute it and/or modify it
   5  * under the terms and conditions of the GNU General Public License,
   6  * version 2, as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope it will be useful, but WITHOUT
   9  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  10  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  11  * more details.
  12  */
  13
  14 #ifndef _NVME_H
  15 #define _NVME_H
  16
  17 #include <linux/nvme.h>
  18 #include <linux/cdev.h>
  19 #include <linux/pci.h>
  20 #include <linux/kref.h>
  21 #include <linux/blk-mq.h>
  22 #include <linux/lightnvm.h>
  23 #include <linux/sed-opal.h>
  24 #include <linux/fault-inject.h>
  25 #include <linux/rcupdate.h>
  26
  27 extern unsigned int nvme_io_timeout;
  28 #define NVME_IO_TIMEOUT (nvme_io_timeout * HZ)
  29
  30 extern unsigned int admin_timeout;
  31 #define ADMIN_TIMEOUT   (admin_timeout * HZ)
  32
  33 #define NVME_DEFAULT_KATO       5
  34 #define NVME_KATO_GRACE         10
  35
  36 extern struct workqueue_struct *nvme_wq;
  37 extern struct workqueue_struct *nvme_reset_wq;
  38 extern struct workqueue_struct *nvme_delete_wq;
  39
  40 enum {
  41         NVME_NS_LBA             = 0,
  42         NVME_NS_LIGHTNVM        = 1,
  43 };
  44
  45 /*
  46  * List of workarounds for devices that required behavior not specified in
  47  * the standard.
  48  */
  49 enum nvme_quirks {
  50         /*
  51          * Prefers I/O aligned to a stripe size specified in a vendor
  52          * specific Identify field.
  53          */
  54         NVME_QUIRK_STRIPE_SIZE                  = (1 << 0),
  55
  56         /*
  57          * The controller doesn't handle Identify value others than 0 or 1
  58          * correctly.
  59          */
  60         NVME_QUIRK_IDENTIFY_CNS                 = (1 << 1),
  61
  62         /*
  63          * The controller deterministically returns O's on reads to
  64          * logical blocks that deallocate was called on.
  65          */
  66         NVME_QUIRK_DEALLOCATE_ZEROES            = (1 << 2),
  67
  68         /*
  69          * The controller needs a delay before starts checking the device
  70          * readiness, which is done by reading the NVME_CSTS_RDY bit.
  71          */
  72         NVME_QUIRK_DELAY_BEFORE_CHK_RDY         = (1 << 3),
  73
  74         /*
  75          * APST should not be used.
  76          */
  77         NVME_QUIRK_NO_APST                      = (1 << 4),
  78
  79         /*
  80          * The deepest sleep state should not be used.
  81          */
  82         NVME_QUIRK_NO_DEEPEST_PS                = (1 << 5),
  83
  84         /*
  85          * Supports the LighNVM command set if indicated in vs[1].
  86          */
  87         NVME_QUIRK_LIGHTNVM                     = (1 << 6),
  88
  89         /*
  90          * Set MEDIUM priority on SQ creation
  91          */
  92         NVME_QUIRK_MEDIUM_PRIO_SQ               = (1 << 7),
  93 };
  94
  95 /*
  96  * Common request structure for NVMe passthrough.  All drivers must have
  97  * this structure as the first member of their request-private data.
  98  */
  99 struct nvme_request {
 100         struct nvme_command     *cmd;
 101         union nvme_result       result;
 102         u8                      retries;
 103         u8                      flags;
 104         u16                     status;
 105         struct nvme_ctrl        *ctrl;
 106 };
 107
 108 /*
 109  * Mark a bio as coming in through the mpath node.
 110  */
 111 #define REQ_NVME_MPATH          REQ_DRV
 112
 113 enum {
 114         NVME_REQ_CANCELLED              = (1 << 0),
 115         NVME_REQ_USERCMD                = (1 << 1),
 116 };
 117
 118 static inline struct nvme_request *nvme_req(struct request *req)
 119 {
 120         return blk_mq_rq_to_pdu(req);
 121 }
 122
 123 static inline u16 nvme_req_qid(struct request *req)
 124 {
 125         if (!req->rq_disk)
 126                 return 0;
 127         return blk_mq_unique_tag_to_hwq(blk_mq_unique_tag(req)) + 1;
 128 }
 129
 130 /* The below value is the specific amount of delay needed before checking
 131  * readiness in case of the PCI_DEVICE(0x1c58, 0x0003), which needs the
 132  * NVME_QUIRK_DELAY_BEFORE_CHK_RDY quirk enabled. The value (in ms) was
 133  * found empirically.
 134  */
 135 #define NVME_QUIRK_DELAY_AMOUNT         2300
 136
 137 enum nvme_ctrl_state {
 138         NVME_CTRL_NEW,
 139         NVME_CTRL_LIVE,
 140         NVME_CTRL_ADMIN_ONLY,    /* Only admin queue live */
 141         NVME_CTRL_RESETTING,
 142         NVME_CTRL_CONNECTING,
 143         NVME_CTRL_DELETING,
 144         NVME_CTRL_DEAD,
 145 };
 146
 147 struct nvme_ctrl {
 148         enum nvme_ctrl_state state;
 149         bool identified;
 150         spinlock_t lock;
 151         struct mutex scan_lock;
 152         const struct nvme_ctrl_ops *ops;
 153         struct request_queue *admin_q;
 154         struct request_queue *connect_q;
 155         struct device *dev;
 156         int instance;
 157         struct blk_mq_tag_set *tagset;
 158         struct blk_mq_tag_set *admin_tagset;
 159         struct list_head namespaces;
 160         struct rw_semaphore namespaces_rwsem;
 161         struct device ctrl_device;
 162         struct device *device;  /* char device */
 163         struct cdev cdev;
 164         struct work_struct reset_work;
 165         struct work_struct delete_work;
 166
 167         struct nvme_subsystem *subsys;
 168         struct list_head subsys_entry;
 169
 170         struct opal_dev *opal_dev;
 171
 172         char name[12];
 173         u16 cntlid;
 174
 175         u32 ctrl_config;
 176         u16 mtfa;
 177         u32 queue_count;
 178
 179         u64 cap;
 180         u32 page_size;
 181         u32 max_hw_sectors;
 182         u32 max_segments;
 183         u16 oncs;
 184         u16 oacs;
 185         u16 nssa;
 186         u16 nr_streams;
 187         u32 max_namespaces;
 188         atomic_t abort_limit;
 189         u8 vwc;
 190         u32 vs;
 191         u32 sgls;
 192         u16 kas;
 193         u8 npss;
 194         u8 apsta;
 195         u32 oaes;
 196         u32 aen_result;
 197         unsigned int shutdown_timeout;
 198         unsigned int kato;
 199         bool subsystem;
 200         unsigned long quirks;
 201         struct nvme_id_power_state psd[32];
 202         struct nvme_effects_log *effects;
 203         struct work_struct scan_work;
 204         struct work_struct async_event_work;
 205         struct delayed_work ka_work;
 206         struct nvme_command ka_cmd;
 207         struct work_struct fw_act_work;
 208         unsigned long events;
 209         bool created;
 210
 211 #ifdef CONFIG_NVME_MULTIPATH
 212         /* asymmetric namespace access: */
 213         u8 anacap;
 214         u8 anatt;
 215         u32 anagrpmax;
 216         u32 nanagrpid;
 217         struct mutex ana_lock;
 218         struct nvme_ana_rsp_hdr *ana_log_buf;
 219         size_t ana_log_size;
 220         struct timer_list anatt_timer;
 221         struct work_struct ana_work;
 222 #endif
 223
 224         /* Power saving configuration */
 225         u64 ps_max_latency_us;
 226         bool apst_enabled;
 227
 228         /* PCIe only: */
 229         u32 hmpre;
 230         u32 hmmin;
 231         u32 hmminds;
 232         u16 hmmaxd;
 233
 234         /* Fabrics only */
 235         u16 sqsize;
 236         u32 ioccsz;
 237         u32 iorcsz;
 238         u16 icdoff;
 239         u16 maxcmd;
 240         int nr_reconnects;
 241         struct nvmf_ctrl_options *opts;
 242
 243         struct page *discard_page;
 244         unsigned long discard_page_busy;
 245 };
 246
 247 struct nvme_subsystem {
 248         int                     instance;
 249         struct device           dev;
 250         /*
 251          * Because we unregister the device on the last put we need
 252          * a separate refcount.
 253          */
 254         struct kref             ref;
 255         struct list_head        entry;
 256         struct mutex            lock;
 257         struct list_head        ctrls;
 258         struct list_head        nsheads;
 259         char                    subnqn[NVMF_NQN_SIZE];
 260         char                    serial[20];
 261         char                    model[40];
 262         char                    firmware_rev[8];
 263         u8                      cmic;
 264         u16                     vendor_id;
 265         struct ida              ns_ida;
 266 };
 267
 268 /*
 269  * Container structure for uniqueue namespace identifiers.
 270  */
 271 struct nvme_ns_ids {
 272         u8      eui64[8];
 273         u8      nguid[16];
 274         uuid_t  uuid;
 275 };
 276
 277 /*
 278  * Anchor structure for namespaces.  There is one for each namespace in a
 279  * NVMe subsystem that any of our controllers can see, and the namespace
 280  * structure for each controller is chained of it.  For private namespaces
 281  * there is a 1:1 relation to our namespace structures, that is ->list
 282  * only ever has a single entry for private namespaces.
 283  */
 284 struct nvme_ns_head {
 285 #ifdef CONFIG_NVME_MULTIPATH
 286         struct gendisk          *disk;
 287         struct nvme_ns __rcu    *current_path;
 288         struct bio_list         requeue_list;
 289         spinlock_t              requeue_lock;
 290         struct work_struct      requeue_work;
 291         struct mutex            lock;
 292 #endif
 293         struct list_head        list;
 294         struct srcu_struct      srcu;
 295         struct nvme_subsystem   *subsys;
 296         unsigned                ns_id;
 297         struct nvme_ns_ids      ids;
 298         struct list_head        entry;
 299         struct kref             ref;
 300         int                     instance;
 301 };
 302
 303 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
 304 struct nvme_fault_inject {
 305         struct fault_attr attr;
 306         struct dentry *parent;
 307         bool dont_retry;        /* DNR, do not retry */
 308         u16 status;             /* status code */
 309 };
 310 #endif
 311
 312 struct nvme_ns {
 313         struct list_head list;
 314
 315         struct nvme_ctrl *ctrl;
 316         struct request_queue *queue;
 317         struct gendisk *disk;
 318 #ifdef CONFIG_NVME_MULTIPATH
 319         enum nvme_ana_state ana_state;
 320         u32 ana_grpid;
 321 #endif
 322         struct list_head siblings;
 323         struct nvm_dev *ndev;
 324         struct kref kref;
 325         struct nvme_ns_head *head;
 326
 327         int lba_shift;
 328         u16 ms;
 329         u16 sgs;
 330         u32 sws;
 331         bool ext;
 332         u8 pi_type;
 333         unsigned long flags;
 334 #define NVME_NS_REMOVING        0
 335 #define NVME_NS_DEAD            1
 336 #define NVME_NS_ANA_PENDING     2
 337         u16 noiob;
 338
 339 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
 340         struct nvme_fault_inject fault_inject;
 341 #endif
 342
 343 };
 344
 345 struct nvme_ctrl_ops {
 346         const char *name;
 347         struct module *module;
 348         unsigned int flags;
 349 #define NVME_F_FABRICS                  (1 << 0)
 350 #define NVME_F_METADATA_SUPPORTED       (1 << 1)
 351         int (*reg_read32)(struct nvme_ctrl *ctrl, u32 off, u32 *val);
 352         int (*reg_write32)(struct nvme_ctrl *ctrl, u32 off, u32 val);
 353         int (*reg_read64)(struct nvme_ctrl *ctrl, u32 off, u64 *val);
 354         void (*free_ctrl)(struct nvme_ctrl *ctrl);
 355         void (*submit_async_event)(struct nvme_ctrl *ctrl);
 356         void (*delete_ctrl)(struct nvme_ctrl *ctrl);
 357         int (*get_address)(struct nvme_ctrl *ctrl, char *buf, int size);
 358         void (*stop_ctrl)(struct nvme_ctrl *ctrl);
 359 };
 360
 361 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
 362 void nvme_fault_inject_init(struct nvme_ns *ns);
 363 void nvme_fault_inject_fini(struct nvme_ns *ns);
 364 void nvme_should_fail(struct request *req);
 365 #else
 366 static inline void nvme_fault_inject_init(struct nvme_ns *ns) {}
 367 static inline void nvme_fault_inject_fini(struct nvme_ns *ns) {}
 368 static inline void nvme_should_fail(struct request *req) {}
 369 #endif
 370
 371 static inline bool nvme_ctrl_ready(struct nvme_ctrl *ctrl)
 372 {
 373         u32 val = 0;
 374
 375         if (ctrl->ops->reg_read32(ctrl, NVME_REG_CSTS, &val))
 376                 return false;
 377         return val & NVME_CSTS_RDY;
 378 }
 379
 380 static inline int nvme_reset_subsystem(struct nvme_ctrl *ctrl)
 381 {
 382         if (!ctrl->subsystem)
 383                 return -ENOTTY;
 384         return ctrl->ops->reg_write32(ctrl, NVME_REG_NSSR, 0x4E564D65);
 385 }
 386
 387 static inline u64 nvme_block_nr(struct nvme_ns *ns, sector_t sector)
 388 {
 389         return (sector >> (ns->lba_shift - 9));
 390 }
 391
 392 static inline void nvme_end_request(struct request *req, __le16 status,
 393                 union nvme_result result)
 394 {
 395         struct nvme_request *rq = nvme_req(req);
 396
 397         rq->status = le16_to_cpu(status) >> 1;
 398         rq->result = result;
 399         /* inject error when permitted by fault injection framework */
 400         nvme_should_fail(req);
 401         blk_mq_complete_request(req);
 402 }
 403
 404 static inline void nvme_get_ctrl(struct nvme_ctrl *ctrl)
 405 {
 406         get_device(ctrl->device);
 407 }
 408
 409 static inline void nvme_put_ctrl(struct nvme_ctrl *ctrl)
 410 {
 411         put_device(ctrl->device);
 412 }
 413
 414 void nvme_complete_rq(struct request *req);
 415 void nvme_cancel_request(struct request *req, void *data, bool reserved);
 416 bool nvme_change_ctrl_state(struct nvme_ctrl *ctrl,
 417                 enum nvme_ctrl_state new_state);
 418 int nvme_disable_ctrl(struct nvme_ctrl *ctrl, u64 cap);
 419 int nvme_enable_ctrl(struct nvme_ctrl *ctrl, u64 cap);
 420 int nvme_shutdown_ctrl(struct nvme_ctrl *ctrl);
 421 int nvme_init_ctrl(struct nvme_ctrl *ctrl, struct device *dev,
 422                 const struct nvme_ctrl_ops *ops, unsigned long quirks);
 423 void nvme_uninit_ctrl(struct nvme_ctrl *ctrl);
 424 void nvme_start_ctrl(struct nvme_ctrl *ctrl);
 425 void nvme_stop_ctrl(struct nvme_ctrl *ctrl);
 426 void nvme_put_ctrl(struct nvme_ctrl *ctrl);
 427 int nvme_init_identify(struct nvme_ctrl *ctrl);
 428
 429 void nvme_remove_namespaces(struct nvme_ctrl *ctrl);
 430
 431 int nvme_sec_submit(void *data, u16 spsp, u8 secp, void *buffer, size_t len,
 432                 bool send);
 433
 434 void nvme_complete_async_event(struct nvme_ctrl *ctrl, __le16 status,
 435                 volatile union nvme_result *res);
 436
 437 void nvme_stop_queues(struct nvme_ctrl *ctrl);
 438 void nvme_start_queues(struct nvme_ctrl *ctrl);
 439 void nvme_kill_queues(struct nvme_ctrl *ctrl);
 440 void nvme_unfreeze(struct nvme_ctrl *ctrl);
 441 void nvme_wait_freeze(struct nvme_ctrl *ctrl);
 442 void nvme_wait_freeze_timeout(struct nvme_ctrl *ctrl, long timeout);
 443 void nvme_start_freeze(struct nvme_ctrl *ctrl);
 444
 445 #define NVME_QID_ANY -1
 446 struct request *nvme_alloc_request(struct request_queue *q,
 447                 struct nvme_command *cmd, blk_mq_req_flags_t flags, int qid);
 448 void nvme_cleanup_cmd(struct request *req);
 449 blk_status_t nvme_setup_cmd(struct nvme_ns *ns, struct request *req,
 450                 struct nvme_command *cmd);
 451 int nvme_submit_sync_cmd(struct request_queue *q, struct nvme_command *cmd,
 452                 void *buf, unsigned bufflen);
 453 int __nvme_submit_sync_cmd(struct request_queue *q, struct nvme_command *cmd,
 454                 union nvme_result *result, void *buffer, unsigned bufflen,
 455                 unsigned timeout, int qid, int at_head,
 456                 blk_mq_req_flags_t flags);
 457 int nvme_set_queue_count(struct nvme_ctrl *ctrl, int *count);
 458 void nvme_stop_keep_alive(struct nvme_ctrl *ctrl);
 459 int nvme_reset_ctrl(struct nvme_ctrl *ctrl);
 460 int nvme_reset_ctrl_sync(struct nvme_ctrl *ctrl);
 461 int nvme_delete_ctrl(struct nvme_ctrl *ctrl);
 462 int nvme_delete_ctrl_sync(struct nvme_ctrl *ctrl);
 463
 464 int nvme_get_log(struct nvme_ctrl *ctrl, u32 nsid, u8 log_page, u8 lsp,
 465                 void *log, size_t size, u64 offset);
 466
 467 extern const struct attribute_group *nvme_ns_id_attr_groups[];
 468 extern const struct block_device_operations nvme_ns_head_ops;
 469
 470 #ifdef CONFIG_NVME_MULTIPATH
 471 static inline bool nvme_ctrl_use_ana(struct nvme_ctrl *ctrl)
 472 {
 473         return ctrl->ana_log_buf != NULL;
 474 }
 475
 476 void nvme_mpath_unfreeze(struct nvme_subsystem *subsys);
 477 void nvme_mpath_wait_freeze(struct nvme_subsystem *subsys);
 478 void nvme_mpath_start_freeze(struct nvme_subsystem *subsys);
 479 void nvme_set_disk_name(char *disk_name, struct nvme_ns *ns,
 480                         struct nvme_ctrl *ctrl, int *flags);
 481 bool nvme_failover_req(struct request *req);
 482 void nvme_kick_requeue_lists(struct nvme_ctrl *ctrl);
 483 int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl,struct nvme_ns_head *head);
 484 void nvme_mpath_add_disk(struct nvme_ns *ns, struct nvme_id_ns *id);
 485 void nvme_mpath_remove_disk(struct nvme_ns_head *head);
 486 int nvme_mpath_init(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id);
 487 void nvme_mpath_uninit(struct nvme_ctrl *ctrl);
 488 void nvme_mpath_stop(struct nvme_ctrl *ctrl);
 489
 490 static inline void nvme_mpath_clear_current_path(struct nvme_ns *ns)
 491 {
 492         struct nvme_ns_head *head = ns->head;
 493
 494         if (head && ns == rcu_access_pointer(head->current_path))
 495                 rcu_assign_pointer(head->current_path, NULL);
 496 }
 497 struct nvme_ns *nvme_find_path(struct nvme_ns_head *head);
 498
 499 static inline void nvme_mpath_check_last_path(struct nvme_ns *ns)
 500 {
 501         struct nvme_ns_head *head = ns->head;
 502
 503         if (head->disk && list_empty(&head->list))
 504                 kblockd_schedule_work(&head->requeue_work);
 505 }
 506
 507 static inline void nvme_mpath_update_disk_size(struct gendisk *disk)
 508 {
 509         struct block_device *bdev = bdget_disk(disk, 0);
 510
 511         if (bdev) {
 512                 bd_set_size(bdev, get_capacity(disk) << SECTOR_SHIFT);
 513                 bdput(bdev);
 514         }
 515 }
 516
 517 extern struct device_attribute dev_attr_ana_grpid;
 518 extern struct device_attribute dev_attr_ana_state;
 519
 520 #else
 521 static inline bool nvme_ctrl_use_ana(struct nvme_ctrl *ctrl)
 522 {
 523         return false;
 524 }
 525 /*
 526  * Without the multipath code enabled, multiple controller per subsystems are
 527  * visible as devices and thus we cannot use the subsystem instance.
 528  */
 529 static inline void nvme_set_disk_name(char *disk_name, struct nvme_ns *ns,
 530                                       struct nvme_ctrl *ctrl, int *flags)
 531 {
 532         sprintf(disk_name, "nvme%dn%d", ctrl->instance, ns->head->instance);
 533 }
 534
 535 static inline bool nvme_failover_req(struct request *req)
 536 {
 537         return false;
 538 }
 539 static inline void nvme_kick_requeue_lists(struct nvme_ctrl *ctrl)
 540 {
 541 }
 542 static inline int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl,
 543                 struct nvme_ns_head *head)
 544 {
 545         return 0;
 546 }
 547 static inline void nvme_mpath_add_disk(struct nvme_ns *ns,
 548                 struct nvme_id_ns *id)
 549 {
 550 }
 551 static inline void nvme_mpath_remove_disk(struct nvme_ns_head *head)
 552 {
 553 }
 554 static inline void nvme_mpath_clear_current_path(struct nvme_ns *ns)
 555 {
 556 }
 557 static inline void nvme_mpath_check_last_path(struct nvme_ns *ns)
 558 {
 559 }
 560 static inline int nvme_mpath_init(struct nvme_ctrl *ctrl,
 561                 struct nvme_id_ctrl *id)
 562 {
 563         if (ctrl->subsys->cmic & (1 << 3))
 564                 dev_warn(ctrl->device,
 565 "Please enable CONFIG_NVME_MULTIPATH for full support of multi-port devices.\n");
 566         return 0;
 567 }
 568 static inline void nvme_mpath_uninit(struct nvme_ctrl *ctrl)
 569 {
 570 }
 571 static inline void nvme_mpath_stop(struct nvme_ctrl *ctrl)
 572 {
 573 }
 574 static inline void nvme_mpath_unfreeze(struct nvme_subsystem *subsys)
 575 {
 576 }
 577 static inline void nvme_mpath_wait_freeze(struct nvme_subsystem *subsys)
 578 {
 579 }
 580 static inline void nvme_mpath_start_freeze(struct nvme_subsystem *subsys)
 581 {
 582 }
 583 static inline void nvme_mpath_update_disk_size(struct gendisk *disk)
 584 {
 585 }
 586 #endif /* CONFIG_NVME_MULTIPATH */
 587
 588 #ifdef CONFIG_NVM
 589 void nvme_nvm_update_nvm_info(struct nvme_ns *ns);
 590 int nvme_nvm_register(struct nvme_ns *ns, char *disk_name, int node);
 591 void nvme_nvm_unregister(struct nvme_ns *ns);
 592 extern const struct attribute_group nvme_nvm_attr_group;
 593 int nvme_nvm_ioctl(struct nvme_ns *ns, unsigned int cmd, unsigned long arg);
 594 #else
 595 static inline void nvme_nvm_update_nvm_info(struct nvme_ns *ns) {};
 596 static inline int nvme_nvm_register(struct nvme_ns *ns, char *disk_name,
 597                                     int node)
 598 {
 599         return 0;
 600 }
 601
 602 static inline void nvme_nvm_unregister(struct nvme_ns *ns) {};
 603 static inline int nvme_nvm_ioctl(struct nvme_ns *ns, unsigned int cmd,
 604                                                         unsigned long arg)
 605 {
 606         return -ENOTTY;
 607 }
 608 #endif /* CONFIG_NVM */
 609
 610 static inline struct nvme_ns *nvme_get_ns_from_dev(struct device *dev)
 611 {
 612         return dev_to_disk(dev)->private_data;
 613 }
 614
 615 int __init nvme_core_init(void);
 616 void nvme_core_exit(void);
 617
 618 #endif /* _NVME_H */