drivers/nvme/host/nvme.h

   1 /* SPDX-License-Identifier: GPL-2.0 */
   2 /*
   3  * Copyright (c) 2011-2014, Intel Corporation.
   4  */
   5
   6 #ifndef _NVME_H
   7 #define _NVME_H
   8
   9 #include <linux/nvme.h>
  10 #include <linux/cdev.h>
  11 #include <linux/pci.h>
  12 #include <linux/kref.h>
  13 #include <linux/blk-mq.h>
  14 #include <linux/lightnvm.h>
  15 #include <linux/sed-opal.h>
  16 #include <linux/fault-inject.h>
  17 #include <linux/rcupdate.h>
  18
  19 extern unsigned int nvme_io_timeout;
  20 #define NVME_IO_TIMEOUT (nvme_io_timeout * HZ)
  21
  22 extern unsigned int admin_timeout;
  23 #define ADMIN_TIMEOUT   (admin_timeout * HZ)
  24
  25 #define NVME_DEFAULT_KATO       5
  26 #define NVME_KATO_GRACE         10
  27
  28 extern struct workqueue_struct *nvme_wq;
  29 extern struct workqueue_struct *nvme_reset_wq;
  30 extern struct workqueue_struct *nvme_delete_wq;
  31
  32 enum {
  33         NVME_NS_LBA             = 0,
  34         NVME_NS_LIGHTNVM        = 1,
  35 };
  36
  37 /*
  38  * List of workarounds for devices that required behavior not specified in
  39  * the standard.
  40  */
  41 enum nvme_quirks {
  42         /*
  43          * Prefers I/O aligned to a stripe size specified in a vendor
  44          * specific Identify field.
  45          */
  46         NVME_QUIRK_STRIPE_SIZE                  = (1 << 0),
  47
  48         /*
  49          * The controller doesn't handle Identify value others than 0 or 1
  50          * correctly.
  51          */
  52         NVME_QUIRK_IDENTIFY_CNS                 = (1 << 1),
  53
  54         /*
  55          * The controller deterministically returns O's on reads to
  56          * logical blocks that deallocate was called on.
  57          */
  58         NVME_QUIRK_DEALLOCATE_ZEROES            = (1 << 2),
  59
  60         /*
  61          * The controller needs a delay before starts checking the device
  62          * readiness, which is done by reading the NVME_CSTS_RDY bit.
  63          */
  64         NVME_QUIRK_DELAY_BEFORE_CHK_RDY         = (1 << 3),
  65
  66         /*
  67          * APST should not be used.
  68          */
  69         NVME_QUIRK_NO_APST                      = (1 << 4),
  70
  71         /*
  72          * The deepest sleep state should not be used.
  73          */
  74         NVME_QUIRK_NO_DEEPEST_PS                = (1 << 5),
  75
  76         /*
  77          * Supports the LighNVM command set if indicated in vs[1].
  78          */
  79         NVME_QUIRK_LIGHTNVM                     = (1 << 6),
  80
  81         /*
  82          * Set MEDIUM priority on SQ creation
  83          */
  84         NVME_QUIRK_MEDIUM_PRIO_SQ               = (1 << 7),
  85
  86         /*
  87          * Ignore device provided subnqn.
  88          */
  89         NVME_QUIRK_IGNORE_DEV_SUBNQN            = (1 << 8),
  90
  91         /*
  92          * Broken Write Zeroes.
  93          */
  94         NVME_QUIRK_DISABLE_WRITE_ZEROES         = (1 << 9),
  95 };
  96
  97 /*
  98  * Common request structure for NVMe passthrough.  All drivers must have
  99  * this structure as the first member of their request-private data.
 100  */
 101 struct nvme_request {
 102         struct nvme_command     *cmd;
 103         union nvme_result       result;
 104         u8                      retries;
 105         u8                      flags;
 106         u16                     status;
 107         struct nvme_ctrl        *ctrl;
 108 };
 109
 110 /*
 111  * Mark a bio as coming in through the mpath node.
 112  */
 113 #define REQ_NVME_MPATH          REQ_DRV
 114
 115 enum {
 116         NVME_REQ_CANCELLED              = (1 << 0),
 117         NVME_REQ_USERCMD                = (1 << 1),
 118 };
 119
 120 static inline struct nvme_request *nvme_req(struct request *req)
 121 {
 122         return blk_mq_rq_to_pdu(req);
 123 }
 124
 125 static inline u16 nvme_req_qid(struct request *req)
 126 {
 127         if (!req->rq_disk)
 128                 return 0;
 129         return blk_mq_unique_tag_to_hwq(blk_mq_unique_tag(req)) + 1;
 130 }
 131
 132 /* The below value is the specific amount of delay needed before checking
 133  * readiness in case of the PCI_DEVICE(0x1c58, 0x0003), which needs the
 134  * NVME_QUIRK_DELAY_BEFORE_CHK_RDY quirk enabled. The value (in ms) was
 135  * found empirically.
 136  */
 137 #define NVME_QUIRK_DELAY_AMOUNT         2300
 138
 139 enum nvme_ctrl_state {
 140         NVME_CTRL_NEW,
 141         NVME_CTRL_LIVE,
 142         NVME_CTRL_ADMIN_ONLY,    /* Only admin queue live */
 143         NVME_CTRL_RESETTING,
 144         NVME_CTRL_CONNECTING,
 145         NVME_CTRL_DELETING,
 146         NVME_CTRL_DEAD,
 147 };
 148
 149 struct nvme_fault_inject {
 150 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
 151         struct fault_attr attr;
 152         struct dentry *parent;
 153         bool dont_retry;        /* DNR, do not retry */
 154         u16 status;             /* status code */
 155 #endif
 156 };
 157
 158 struct nvme_ctrl {
 159         bool comp_seen;
 160         enum nvme_ctrl_state state;
 161         bool identified;
 162         spinlock_t lock;
 163         struct mutex scan_lock;
 164         const struct nvme_ctrl_ops *ops;
 165         struct request_queue *admin_q;
 166         struct request_queue *connect_q;
 167         struct device *dev;
 168         int instance;
 169         int numa_node;
 170         struct blk_mq_tag_set *tagset;
 171         struct blk_mq_tag_set *admin_tagset;
 172         struct list_head namespaces;
 173         struct rw_semaphore namespaces_rwsem;
 174         struct device ctrl_device;
 175         struct device *device;  /* char device */
 176         struct cdev cdev;
 177         struct work_struct reset_work;
 178         struct work_struct delete_work;
 179
 180         struct nvme_subsystem *subsys;
 181         struct list_head subsys_entry;
 182
 183         struct opal_dev *opal_dev;
 184
 185         char name[12];
 186         u16 cntlid;
 187
 188         u32 ctrl_config;
 189         u16 mtfa;
 190         u32 queue_count;
 191
 192         u64 cap;
 193         u32 page_size;
 194         u32 max_hw_sectors;
 195         u32 max_segments;
 196         u16 crdt[3];
 197         u16 oncs;
 198         u16 oacs;
 199         u16 nssa;
 200         u16 nr_streams;
 201         u32 max_namespaces;
 202         atomic_t abort_limit;
 203         u8 vwc;
 204         u32 vs;
 205         u32 sgls;
 206         u16 kas;
 207         u8 npss;
 208         u8 apsta;
 209         u32 oaes;
 210         u32 aen_result;
 211         u32 ctratt;
 212         unsigned int shutdown_timeout;
 213         unsigned int kato;
 214         bool subsystem;
 215         unsigned long quirks;
 216         struct nvme_id_power_state psd[32];
 217         struct nvme_effects_log *effects;
 218         struct work_struct scan_work;
 219         struct work_struct async_event_work;
 220         struct delayed_work ka_work;
 221         struct nvme_command ka_cmd;
 222         struct work_struct fw_act_work;
 223         unsigned long events;
 224
 225 #ifdef CONFIG_NVME_MULTIPATH
 226         /* asymmetric namespace access: */
 227         u8 anacap;
 228         u8 anatt;
 229         u32 anagrpmax;
 230         u32 nanagrpid;
 231         struct mutex ana_lock;
 232         struct nvme_ana_rsp_hdr *ana_log_buf;
 233         size_t ana_log_size;
 234         struct timer_list anatt_timer;
 235         struct work_struct ana_work;
 236 #endif
 237
 238         /* Power saving configuration */
 239         u64 ps_max_latency_us;
 240         bool apst_enabled;
 241
 242         /* PCIe only: */
 243         u32 hmpre;
 244         u32 hmmin;
 245         u32 hmminds;
 246         u16 hmmaxd;
 247
 248         /* Fabrics only */
 249         u16 sqsize;
 250         u32 ioccsz;
 251         u32 iorcsz;
 252         u16 icdoff;
 253         u16 maxcmd;
 254         int nr_reconnects;
 255         struct nvmf_ctrl_options *opts;
 256
 257         struct page *discard_page;
 258         unsigned long discard_page_busy;
 259
 260         struct nvme_fault_inject fault_inject;
 261 };
 262
 263 enum nvme_iopolicy {
 264         NVME_IOPOLICY_NUMA,
 265         NVME_IOPOLICY_RR,
 266 };
 267
 268 struct nvme_subsystem {
 269         int                     instance;
 270         struct device           dev;
 271         /*
 272          * Because we unregister the device on the last put we need
 273          * a separate refcount.
 274          */
 275         struct kref             ref;
 276         struct list_head        entry;
 277         struct mutex            lock;
 278         struct list_head        ctrls;
 279         struct list_head        nsheads;
 280         char                    subnqn[NVMF_NQN_SIZE];
 281         char                    serial[20];
 282         char                    model[40];
 283         char                    firmware_rev[8];
 284         u8                      cmic;
 285         u16                     vendor_id;
 286         u16                     awupf;  /* 0's based awupf value. */
 287         struct ida              ns_ida;
 288 #ifdef CONFIG_NVME_MULTIPATH
 289         enum nvme_iopolicy      iopolicy;
 290 #endif
 291 };
 292
 293 /*
 294  * Container structure for uniqueue namespace identifiers.
 295  */
 296 struct nvme_ns_ids {
 297         u8      eui64[8];
 298         u8      nguid[16];
 299         uuid_t  uuid;
 300 };
 301
 302 /*
 303  * Anchor structure for namespaces.  There is one for each namespace in a
 304  * NVMe subsystem that any of our controllers can see, and the namespace
 305  * structure for each controller is chained of it.  For private namespaces
 306  * there is a 1:1 relation to our namespace structures, that is ->list
 307  * only ever has a single entry for private namespaces.
 308  */
 309 struct nvme_ns_head {
 310         struct list_head        list;
 311         struct srcu_struct      srcu;
 312         struct nvme_subsystem   *subsys;
 313         unsigned                ns_id;
 314         struct nvme_ns_ids      ids;
 315         struct list_head        entry;
 316         struct kref             ref;
 317         int                     instance;
 318 #ifdef CONFIG_NVME_MULTIPATH
 319         struct gendisk          *disk;
 320         struct bio_list         requeue_list;
 321         spinlock_t              requeue_lock;
 322         struct work_struct      requeue_work;
 323         struct mutex            lock;
 324         struct nvme_ns __rcu    *current_path[];
 325 #endif
 326 };
 327
 328 struct nvme_ns {
 329         struct list_head list;
 330
 331         struct nvme_ctrl *ctrl;
 332         struct request_queue *queue;
 333         struct gendisk *disk;
 334 #ifdef CONFIG_NVME_MULTIPATH
 335         enum nvme_ana_state ana_state;
 336         u32 ana_grpid;
 337 #endif
 338         struct list_head siblings;
 339         struct nvm_dev *ndev;
 340         struct kref kref;
 341         struct nvme_ns_head *head;
 342
 343         int lba_shift;
 344         u16 ms;
 345         u16 sgs;
 346         u32 sws;
 347         bool ext;
 348         u8 pi_type;
 349         unsigned long flags;
 350 #define NVME_NS_REMOVING        0
 351 #define NVME_NS_DEAD            1
 352 #define NVME_NS_ANA_PENDING     2
 353         u16 noiob;
 354
 355         struct nvme_fault_inject fault_inject;
 356
 357 };
 358
 359 struct nvme_ctrl_ops {
 360         const char *name;
 361         struct module *module;
 362         unsigned int flags;
 363 #define NVME_F_FABRICS                  (1 << 0)
 364 #define NVME_F_METADATA_SUPPORTED       (1 << 1)
 365 #define NVME_F_PCI_P2PDMA               (1 << 2)
 366         int (*reg_read32)(struct nvme_ctrl *ctrl, u32 off, u32 *val);
 367         int (*reg_write32)(struct nvme_ctrl *ctrl, u32 off, u32 val);
 368         int (*reg_read64)(struct nvme_ctrl *ctrl, u32 off, u64 *val);
 369         void (*free_ctrl)(struct nvme_ctrl *ctrl);
 370         void (*submit_async_event)(struct nvme_ctrl *ctrl);
 371         void (*delete_ctrl)(struct nvme_ctrl *ctrl);
 372         int (*get_address)(struct nvme_ctrl *ctrl, char *buf, int size);
 373 };
 374
 375 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
 376 void nvme_fault_inject_init(struct nvme_fault_inject *fault_inj,
 377                             const char *dev_name);
 378 void nvme_fault_inject_fini(struct nvme_fault_inject *fault_inject);
 379 void nvme_should_fail(struct request *req);
 380 #else
 381 static inline void nvme_fault_inject_init(struct nvme_fault_inject *fault_inj,
 382                                           const char *dev_name)
 383 {
 384 }
 385 static inline void nvme_fault_inject_fini(struct nvme_fault_inject *fault_inj)
 386 {
 387 }
 388 static inline void nvme_should_fail(struct request *req) {}
 389 #endif
 390
 391 static inline int nvme_reset_subsystem(struct nvme_ctrl *ctrl)
 392 {
 393         if (!ctrl->subsystem)
 394                 return -ENOTTY;
 395         return ctrl->ops->reg_write32(ctrl, NVME_REG_NSSR, 0x4E564D65);
 396 }
 397
 398 static inline u64 nvme_block_nr(struct nvme_ns *ns, sector_t sector)
 399 {
 400         return (sector >> (ns->lba_shift - 9));
 401 }
 402
 403 static inline void nvme_end_request(struct request *req, __le16 status,
 404                 union nvme_result result)
 405 {
 406         struct nvme_request *rq = nvme_req(req);
 407
 408         rq->status = le16_to_cpu(status) >> 1;
 409         rq->result = result;
 410         /* inject error when permitted by fault injection framework */
 411         nvme_should_fail(req);
 412         blk_mq_complete_request(req);
 413 }
 414
 415 static inline void nvme_get_ctrl(struct nvme_ctrl *ctrl)
 416 {
 417         get_device(ctrl->device);
 418 }
 419
 420 static inline void nvme_put_ctrl(struct nvme_ctrl *ctrl)
 421 {
 422         put_device(ctrl->device);
 423 }
 424
 425 void nvme_complete_rq(struct request *req);
 426 bool nvme_cancel_request(struct request *req, void *data, bool reserved);
 427 bool nvme_change_ctrl_state(struct nvme_ctrl *ctrl,
 428                 enum nvme_ctrl_state new_state);
 429 int nvme_disable_ctrl(struct nvme_ctrl *ctrl, u64 cap);
 430 int nvme_enable_ctrl(struct nvme_ctrl *ctrl, u64 cap);
 431 int nvme_shutdown_ctrl(struct nvme_ctrl *ctrl);
 432 int nvme_init_ctrl(struct nvme_ctrl *ctrl, struct device *dev,
 433                 const struct nvme_ctrl_ops *ops, unsigned long quirks);
 434 void nvme_uninit_ctrl(struct nvme_ctrl *ctrl);
 435 void nvme_start_ctrl(struct nvme_ctrl *ctrl);
 436 void nvme_stop_ctrl(struct nvme_ctrl *ctrl);
 437 void nvme_put_ctrl(struct nvme_ctrl *ctrl);
 438 int nvme_init_identify(struct nvme_ctrl *ctrl);
 439
 440 void nvme_remove_namespaces(struct nvme_ctrl *ctrl);
 441
 442 int nvme_sec_submit(void *data, u16 spsp, u8 secp, void *buffer, size_t len,
 443                 bool send);
 444
 445 void nvme_complete_async_event(struct nvme_ctrl *ctrl, __le16 status,
 446                 volatile union nvme_result *res);
 447
 448 void nvme_stop_queues(struct nvme_ctrl *ctrl);
 449 void nvme_start_queues(struct nvme_ctrl *ctrl);
 450 void nvme_kill_queues(struct nvme_ctrl *ctrl);
 451 void nvme_sync_queues(struct nvme_ctrl *ctrl);
 452 void nvme_unfreeze(struct nvme_ctrl *ctrl);
 453 void nvme_wait_freeze(struct nvme_ctrl *ctrl);
 454 void nvme_wait_freeze_timeout(struct nvme_ctrl *ctrl, long timeout);
 455 void nvme_start_freeze(struct nvme_ctrl *ctrl);
 456
 457 #define NVME_QID_ANY -1
 458 struct request *nvme_alloc_request(struct request_queue *q,
 459                 struct nvme_command *cmd, blk_mq_req_flags_t flags, int qid);
 460 void nvme_cleanup_cmd(struct request *req);
 461 blk_status_t nvme_setup_cmd(struct nvme_ns *ns, struct request *req,
 462                 struct nvme_command *cmd);
 463 int nvme_submit_sync_cmd(struct request_queue *q, struct nvme_command *cmd,
 464                 void *buf, unsigned bufflen);
 465 int __nvme_submit_sync_cmd(struct request_queue *q, struct nvme_command *cmd,
 466                 union nvme_result *result, void *buffer, unsigned bufflen,
 467                 unsigned timeout, int qid, int at_head,
 468                 blk_mq_req_flags_t flags, bool poll);
 469 int nvme_set_features(struct nvme_ctrl *dev, unsigned int fid,
 470                       unsigned int dword11, void *buffer, size_t buflen,
 471                       u32 *result);
 472 int nvme_get_features(struct nvme_ctrl *dev, unsigned int fid,
 473                       unsigned int dword11, void *buffer, size_t buflen,
 474                       u32 *result);
 475 int nvme_set_queue_count(struct nvme_ctrl *ctrl, int *count);
 476 void nvme_stop_keep_alive(struct nvme_ctrl *ctrl);
 477 int nvme_reset_ctrl(struct nvme_ctrl *ctrl);
 478 int nvme_reset_ctrl_sync(struct nvme_ctrl *ctrl);
 479 int nvme_delete_ctrl(struct nvme_ctrl *ctrl);
 480
 481 int nvme_get_log(struct nvme_ctrl *ctrl, u32 nsid, u8 log_page, u8 lsp,
 482                 void *log, size_t size, u64 offset);
 483
 484 extern const struct attribute_group *nvme_ns_id_attr_groups[];
 485 extern const struct block_device_operations nvme_ns_head_ops;
 486
 487 #ifdef CONFIG_NVME_MULTIPATH
 488 bool nvme_ctrl_use_ana(struct nvme_ctrl *ctrl);
 489 void nvme_set_disk_name(char *disk_name, struct nvme_ns *ns,
 490                         struct nvme_ctrl *ctrl, int *flags);
 491 void nvme_failover_req(struct request *req);
 492 void nvme_kick_requeue_lists(struct nvme_ctrl *ctrl);
 493 int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl,struct nvme_ns_head *head);
 494 void nvme_mpath_add_disk(struct nvme_ns *ns, struct nvme_id_ns *id);
 495 void nvme_mpath_remove_disk(struct nvme_ns_head *head);
 496 int nvme_mpath_init(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id);
 497 void nvme_mpath_uninit(struct nvme_ctrl *ctrl);
 498 void nvme_mpath_stop(struct nvme_ctrl *ctrl);
 499 void nvme_mpath_clear_current_path(struct nvme_ns *ns);
 500 struct nvme_ns *nvme_find_path(struct nvme_ns_head *head);
 501
 502 static inline void nvme_mpath_check_last_path(struct nvme_ns *ns)
 503 {
 504         struct nvme_ns_head *head = ns->head;
 505
 506         if (head->disk && list_empty(&head->list))
 507                 kblockd_schedule_work(&head->requeue_work);
 508 }
 509
 510 extern struct device_attribute dev_attr_ana_grpid;
 511 extern struct device_attribute dev_attr_ana_state;
 512 extern struct device_attribute subsys_attr_iopolicy;
 513
 514 #else
 515 static inline bool nvme_ctrl_use_ana(struct nvme_ctrl *ctrl)
 516 {
 517         return false;
 518 }
 519 /*
 520  * Without the multipath code enabled, multiple controller per subsystems are
 521  * visible as devices and thus we cannot use the subsystem instance.
 522  */
 523 static inline void nvme_set_disk_name(char *disk_name, struct nvme_ns *ns,
 524                                       struct nvme_ctrl *ctrl, int *flags)
 525 {
 526         sprintf(disk_name, "nvme%dn%d", ctrl->instance, ns->head->instance);
 527 }
 528
 529 static inline void nvme_failover_req(struct request *req)
 530 {
 531 }
 532 static inline void nvme_kick_requeue_lists(struct nvme_ctrl *ctrl)
 533 {
 534 }
 535 static inline int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl,
 536                 struct nvme_ns_head *head)
 537 {
 538         return 0;
 539 }
 540 static inline void nvme_mpath_add_disk(struct nvme_ns *ns,
 541                 struct nvme_id_ns *id)
 542 {
 543 }
 544 static inline void nvme_mpath_remove_disk(struct nvme_ns_head *head)
 545 {
 546 }
 547 static inline void nvme_mpath_clear_current_path(struct nvme_ns *ns)
 548 {
 549 }
 550 static inline void nvme_mpath_check_last_path(struct nvme_ns *ns)
 551 {
 552 }
 553 static inline int nvme_mpath_init(struct nvme_ctrl *ctrl,
 554                 struct nvme_id_ctrl *id)
 555 {
 556         if (ctrl->subsys->cmic & (1 << 3))
 557                 dev_warn(ctrl->device,
 558 "Please enable CONFIG_NVME_MULTIPATH for full support of multi-port devices.\n");
 559         return 0;
 560 }
 561 static inline void nvme_mpath_uninit(struct nvme_ctrl *ctrl)
 562 {
 563 }
 564 static inline void nvme_mpath_stop(struct nvme_ctrl *ctrl)
 565 {
 566 }
 567 #endif /* CONFIG_NVME_MULTIPATH */
 568
 569 #ifdef CONFIG_NVM
 570 int nvme_nvm_register(struct nvme_ns *ns, char *disk_name, int node);
 571 void nvme_nvm_unregister(struct nvme_ns *ns);
 572 extern const struct attribute_group nvme_nvm_attr_group;
 573 int nvme_nvm_ioctl(struct nvme_ns *ns, unsigned int cmd, unsigned long arg);
 574 #else
 575 static inline int nvme_nvm_register(struct nvme_ns *ns, char *disk_name,
 576                                     int node)
 577 {
 578         return 0;
 579 }
 580
 581 static inline void nvme_nvm_unregister(struct nvme_ns *ns) {};
 582 static inline int nvme_nvm_ioctl(struct nvme_ns *ns, unsigned int cmd,
 583                                                         unsigned long arg)
 584 {
 585         return -ENOTTY;
 586 }
 587 #endif /* CONFIG_NVM */
 588
 589 static inline struct nvme_ns *nvme_get_ns_from_dev(struct device *dev)
 590 {
 591         return dev_to_disk(dev)->private_data;
 592 }
 593
 594 #endif /* _NVME_H */