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
  25 extern unsigned int nvme_io_timeout;
  26 #define NVME_IO_TIMEOUT (nvme_io_timeout * HZ)
  27
  28 extern unsigned int admin_timeout;
  29 #define ADMIN_TIMEOUT   (admin_timeout * HZ)
  30
  31 #define NVME_DEFAULT_KATO       5
  32 #define NVME_KATO_GRACE         10
  33
  34 extern struct workqueue_struct *nvme_wq;
  35 extern struct workqueue_struct *nvme_reset_wq;
  36 extern struct workqueue_struct *nvme_delete_wq;
  37
  38 enum {
  39         NVME_NS_LBA             = 0,
  40         NVME_NS_LIGHTNVM        = 1,
  41 };
  42
  43 /*
  44  * List of workarounds for devices that required behavior not specified in
  45  * the standard.
  46  */
  47 enum nvme_quirks {
  48         /*
  49          * Prefers I/O aligned to a stripe size specified in a vendor
  50          * specific Identify field.
  51          */
  52         NVME_QUIRK_STRIPE_SIZE                  = (1 << 0),
  53
  54         /*
  55          * The controller doesn't handle Identify value others than 0 or 1
  56          * correctly.
  57          */
  58         NVME_QUIRK_IDENTIFY_CNS                 = (1 << 1),
  59
  60         /*
  61          * The controller deterministically returns O's on reads to
  62          * logical blocks that deallocate was called on.
  63          */
  64         NVME_QUIRK_DEALLOCATE_ZEROES            = (1 << 2),
  65
  66         /*
  67          * The controller needs a delay before starts checking the device
  68          * readiness, which is done by reading the NVME_CSTS_RDY bit.
  69          */
  70         NVME_QUIRK_DELAY_BEFORE_CHK_RDY         = (1 << 3),
  71
  72         /*
  73          * APST should not be used.
  74          */
  75         NVME_QUIRK_NO_APST                      = (1 << 4),
  76
  77         /*
  78          * The deepest sleep state should not be used.
  79          */
  80         NVME_QUIRK_NO_DEEPEST_PS                = (1 << 5),
  81
  82         /*
  83          * Supports the LighNVM command set if indicated in vs[1].
  84          */
  85         NVME_QUIRK_LIGHTNVM                     = (1 << 6),
  86 };
  87
  88 /*
  89  * Common request structure for NVMe passthrough.  All drivers must have
  90  * this structure as the first member of their request-private data.
  91  */
  92 struct nvme_request {
  93         struct nvme_command     *cmd;
  94         union nvme_result       result;
  95         u8                      retries;
  96         u8                      flags;
  97         u16                     status;
  98 };
  99
 100 /*
 101  * Mark a bio as coming in through the mpath node.
 102  */
 103 #define REQ_NVME_MPATH          REQ_DRV
 104
 105 enum {
 106         NVME_REQ_CANCELLED              = (1 << 0),
 107 };
 108
 109 static inline struct nvme_request *nvme_req(struct request *req)
 110 {
 111         return blk_mq_rq_to_pdu(req);
 112 }
 113
 114 /* The below value is the specific amount of delay needed before checking
 115  * readiness in case of the PCI_DEVICE(0x1c58, 0x0003), which needs the
 116  * NVME_QUIRK_DELAY_BEFORE_CHK_RDY quirk enabled. The value (in ms) was
 117  * found empirically.
 118  */
 119 #define NVME_QUIRK_DELAY_AMOUNT         2300
 120
 121 enum nvme_ctrl_state {
 122         NVME_CTRL_NEW,
 123         NVME_CTRL_LIVE,
 124         NVME_CTRL_ADMIN_ONLY,    /* Only admin queue live */
 125         NVME_CTRL_RESETTING,
 126         NVME_CTRL_RECONNECTING,
 127         NVME_CTRL_DELETING,
 128         NVME_CTRL_DEAD,
 129 };
 130
 131 struct nvme_ctrl {
 132         enum nvme_ctrl_state state;
 133         bool identified;
 134         spinlock_t lock;
 135         const struct nvme_ctrl_ops *ops;
 136         struct request_queue *admin_q;
 137         struct request_queue *connect_q;
 138         struct device *dev;
 139         int instance;
 140         struct blk_mq_tag_set *tagset;
 141         struct blk_mq_tag_set *admin_tagset;
 142         struct list_head namespaces;
 143         struct mutex namespaces_mutex;
 144         struct device ctrl_device;
 145         struct device *device;  /* char device */
 146         struct cdev cdev;
 147         struct work_struct reset_work;
 148         struct work_struct delete_work;
 149
 150         struct nvme_subsystem *subsys;
 151         struct list_head subsys_entry;
 152
 153         struct opal_dev *opal_dev;
 154
 155         char name[12];
 156         u16 cntlid;
 157
 158         u32 ctrl_config;
 159         u16 mtfa;
 160         u32 queue_count;
 161
 162         u64 cap;
 163         u32 page_size;
 164         u32 max_hw_sectors;
 165         u16 oncs;
 166         u16 oacs;
 167         u16 nssa;
 168         u16 nr_streams;
 169         atomic_t abort_limit;
 170         u8 vwc;
 171         u32 vs;
 172         u32 sgls;
 173         u16 kas;
 174         u8 npss;
 175         u8 apsta;
 176         u32 aen_result;
 177         unsigned int shutdown_timeout;
 178         unsigned int kato;
 179         bool subsystem;
 180         unsigned long quirks;
 181         struct nvme_id_power_state psd[32];
 182         struct nvme_effects_log *effects;
 183         struct work_struct scan_work;
 184         struct work_struct async_event_work;
 185         struct delayed_work ka_work;
 186         struct work_struct fw_act_work;
 187
 188         /* Power saving configuration */
 189         u64 ps_max_latency_us;
 190         bool apst_enabled;
 191
 192         /* PCIe only: */
 193         u32 hmpre;
 194         u32 hmmin;
 195         u32 hmminds;
 196         u16 hmmaxd;
 197
 198         /* Fabrics only */
 199         u16 sqsize;
 200         u32 ioccsz;
 201         u32 iorcsz;
 202         u16 icdoff;
 203         u16 maxcmd;
 204         int nr_reconnects;
 205         struct nvmf_ctrl_options *opts;
 206 };
 207
 208 struct nvme_subsystem {
 209         int                     instance;
 210         struct device           dev;
 211         /*
 212          * Because we unregister the device on the last put we need
 213          * a separate refcount.
 214          */
 215         struct kref             ref;
 216         struct list_head        entry;
 217         struct mutex            lock;
 218         struct list_head        ctrls;
 219         struct list_head        nsheads;
 220         char                    subnqn[NVMF_NQN_SIZE];
 221         char                    serial[20];
 222         char                    model[40];
 223         char                    firmware_rev[8];
 224         u8                      cmic;
 225         u16                     vendor_id;
 226         struct ida              ns_ida;
 227 };
 228
 229 /*
 230  * Container structure for uniqueue namespace identifiers.
 231  */
 232 struct nvme_ns_ids {
 233         u8      eui64[8];
 234         u8      nguid[16];
 235         uuid_t  uuid;
 236 };
 237
 238 /*
 239  * Anchor structure for namespaces.  There is one for each namespace in a
 240  * NVMe subsystem that any of our controllers can see, and the namespace
 241  * structure for each controller is chained of it.  For private namespaces
 242  * there is a 1:1 relation to our namespace structures, that is ->list
 243  * only ever has a single entry for private namespaces.
 244  */
 245 struct nvme_ns_head {
 246 #ifdef CONFIG_NVME_MULTIPATH
 247         struct gendisk          *disk;
 248         struct nvme_ns __rcu    *current_path;
 249         struct bio_list         requeue_list;
 250         spinlock_t              requeue_lock;
 251         struct work_struct      requeue_work;
 252 #endif
 253         struct list_head        list;
 254         struct srcu_struct      srcu;
 255         struct nvme_subsystem   *subsys;
 256         unsigned                ns_id;
 257         struct nvme_ns_ids      ids;
 258         struct list_head        entry;
 259         struct kref             ref;
 260         int                     instance;
 261 };
 262
 263 struct nvme_ns {
 264         struct list_head list;
 265
 266         struct nvme_ctrl *ctrl;
 267         struct request_queue *queue;
 268         struct gendisk *disk;
 269         struct list_head siblings;
 270         struct nvm_dev *ndev;
 271         struct kref kref;
 272         struct nvme_ns_head *head;
 273
 274         int lba_shift;
 275         u16 ms;
 276         u16 sgs;
 277         u32 sws;
 278         bool ext;
 279         u8 pi_type;
 280         unsigned long flags;
 281 #define NVME_NS_REMOVING 0
 282 #define NVME_NS_DEAD     1
 283         u16 noiob;
 284 };
 285
 286 struct nvme_ctrl_ops {
 287         const char *name;
 288         struct module *module;
 289         unsigned int flags;
 290 #define NVME_F_FABRICS                  (1 << 0)
 291 #define NVME_F_METADATA_SUPPORTED       (1 << 1)
 292         int (*reg_read32)(struct nvme_ctrl *ctrl, u32 off, u32 *val);
 293         int (*reg_write32)(struct nvme_ctrl *ctrl, u32 off, u32 val);
 294         int (*reg_read64)(struct nvme_ctrl *ctrl, u32 off, u64 *val);
 295         void (*free_ctrl)(struct nvme_ctrl *ctrl);
 296         void (*submit_async_event)(struct nvme_ctrl *ctrl);
 297         void (*delete_ctrl)(struct nvme_ctrl *ctrl);
 298         int (*get_address)(struct nvme_ctrl *ctrl, char *buf, int size);
 299         int (*reinit_request)(void *data, struct request *rq);
 300 };
 301
 302 static inline bool nvme_ctrl_ready(struct nvme_ctrl *ctrl)
 303 {
 304         u32 val = 0;
 305
 306         if (ctrl->ops->reg_read32(ctrl, NVME_REG_CSTS, &val))
 307                 return false;
 308         return val & NVME_CSTS_RDY;
 309 }
 310
 311 static inline int nvme_reset_subsystem(struct nvme_ctrl *ctrl)
 312 {
 313         if (!ctrl->subsystem)
 314                 return -ENOTTY;
 315         return ctrl->ops->reg_write32(ctrl, NVME_REG_NSSR, 0x4E564D65);
 316 }
 317
 318 static inline u64 nvme_block_nr(struct nvme_ns *ns, sector_t sector)
 319 {
 320         return (sector >> (ns->lba_shift - 9));
 321 }
 322
 323 static inline void nvme_cleanup_cmd(struct request *req)
 324 {
 325         if (req->rq_flags & RQF_SPECIAL_PAYLOAD) {
 326                 kfree(page_address(req->special_vec.bv_page) +
 327                       req->special_vec.bv_offset);
 328         }
 329 }
 330
 331 static inline void nvme_end_request(struct request *req, __le16 status,
 332                 union nvme_result result)
 333 {
 334         struct nvme_request *rq = nvme_req(req);
 335
 336         rq->status = le16_to_cpu(status) >> 1;
 337         rq->result = result;
 338         blk_mq_complete_request(req);
 339 }
 340
 341 static inline void nvme_get_ctrl(struct nvme_ctrl *ctrl)
 342 {
 343         get_device(ctrl->device);
 344 }
 345
 346 static inline void nvme_put_ctrl(struct nvme_ctrl *ctrl)
 347 {
 348         put_device(ctrl->device);
 349 }
 350
 351 void nvme_complete_rq(struct request *req);
 352 void nvme_cancel_request(struct request *req, void *data, bool reserved);
 353 bool nvme_change_ctrl_state(struct nvme_ctrl *ctrl,
 354                 enum nvme_ctrl_state new_state);
 355 int nvme_disable_ctrl(struct nvme_ctrl *ctrl, u64 cap);
 356 int nvme_enable_ctrl(struct nvme_ctrl *ctrl, u64 cap);
 357 int nvme_shutdown_ctrl(struct nvme_ctrl *ctrl);
 358 int nvme_init_ctrl(struct nvme_ctrl *ctrl, struct device *dev,
 359                 const struct nvme_ctrl_ops *ops, unsigned long quirks);
 360 void nvme_uninit_ctrl(struct nvme_ctrl *ctrl);
 361 void nvme_start_ctrl(struct nvme_ctrl *ctrl);
 362 void nvme_stop_ctrl(struct nvme_ctrl *ctrl);
 363 void nvme_put_ctrl(struct nvme_ctrl *ctrl);
 364 int nvme_init_identify(struct nvme_ctrl *ctrl);
 365
 366 void nvme_queue_scan(struct nvme_ctrl *ctrl);
 367 void nvme_remove_namespaces(struct nvme_ctrl *ctrl);
 368
 369 int nvme_sec_submit(void *data, u16 spsp, u8 secp, void *buffer, size_t len,
 370                 bool send);
 371
 372 void nvme_complete_async_event(struct nvme_ctrl *ctrl, __le16 status,
 373                 union nvme_result *res);
 374
 375 void nvme_stop_queues(struct nvme_ctrl *ctrl);
 376 void nvme_start_queues(struct nvme_ctrl *ctrl);
 377 void nvme_kill_queues(struct nvme_ctrl *ctrl);
 378 void nvme_unfreeze(struct nvme_ctrl *ctrl);
 379 void nvme_wait_freeze(struct nvme_ctrl *ctrl);
 380 void nvme_wait_freeze_timeout(struct nvme_ctrl *ctrl, long timeout);
 381 void nvme_start_freeze(struct nvme_ctrl *ctrl);
 382 int nvme_reinit_tagset(struct nvme_ctrl *ctrl, struct blk_mq_tag_set *set);
 383
 384 #define NVME_QID_ANY -1
 385 struct request *nvme_alloc_request(struct request_queue *q,
 386                 struct nvme_command *cmd, blk_mq_req_flags_t flags, int qid);
 387 blk_status_t nvme_setup_cmd(struct nvme_ns *ns, struct request *req,
 388                 struct nvme_command *cmd);
 389 int nvme_submit_sync_cmd(struct request_queue *q, struct nvme_command *cmd,
 390                 void *buf, unsigned bufflen);
 391 int __nvme_submit_sync_cmd(struct request_queue *q, struct nvme_command *cmd,
 392                 union nvme_result *result, void *buffer, unsigned bufflen,
 393                 unsigned timeout, int qid, int at_head,
 394                 blk_mq_req_flags_t flags);
 395 int nvme_set_queue_count(struct nvme_ctrl *ctrl, int *count);
 396 void nvme_start_keep_alive(struct nvme_ctrl *ctrl);
 397 void nvme_stop_keep_alive(struct nvme_ctrl *ctrl);
 398 int nvme_reset_ctrl(struct nvme_ctrl *ctrl);
 399 int nvme_reset_ctrl_sync(struct nvme_ctrl *ctrl);
 400 int nvme_delete_ctrl(struct nvme_ctrl *ctrl);
 401 int nvme_delete_ctrl_sync(struct nvme_ctrl *ctrl);
 402
 403 extern const struct attribute_group nvme_ns_id_attr_group;
 404 extern const struct block_device_operations nvme_ns_head_ops;
 405
 406 #ifdef CONFIG_NVME_MULTIPATH
 407 void nvme_failover_req(struct request *req);
 408 bool nvme_req_needs_failover(struct request *req, blk_status_t error);
 409 void nvme_kick_requeue_lists(struct nvme_ctrl *ctrl);
 410 int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl,struct nvme_ns_head *head);
 411 void nvme_mpath_add_disk(struct nvme_ns_head *head);
 412 void nvme_mpath_add_disk_links(struct nvme_ns *ns);
 413 void nvme_mpath_remove_disk(struct nvme_ns_head *head);
 414 void nvme_mpath_remove_disk_links(struct nvme_ns *ns);
 415
 416 static inline void nvme_mpath_clear_current_path(struct nvme_ns *ns)
 417 {
 418         struct nvme_ns_head *head = ns->head;
 419
 420         if (head && ns == srcu_dereference(head->current_path, &head->srcu))
 421                 rcu_assign_pointer(head->current_path, NULL);
 422 }
 423 struct nvme_ns *nvme_find_path(struct nvme_ns_head *head);
 424
 425 static inline void nvme_mpath_check_last_path(struct nvme_ns *ns)
 426 {
 427         struct nvme_ns_head *head = ns->head;
 428
 429         if (head->disk && list_empty(&head->list))
 430                 kblockd_schedule_work(&head->requeue_work);
 431 }
 432
 433 #else
 434 static inline void nvme_failover_req(struct request *req)
 435 {
 436 }
 437 static inline bool nvme_req_needs_failover(struct request *req,
 438                                            blk_status_t error)
 439 {
 440         return false;
 441 }
 442 static inline void nvme_kick_requeue_lists(struct nvme_ctrl *ctrl)
 443 {
 444 }
 445 static inline int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl,
 446                 struct nvme_ns_head *head)
 447 {
 448         return 0;
 449 }
 450 static inline void nvme_mpath_add_disk(struct nvme_ns_head *head)
 451 {
 452 }
 453 static inline void nvme_mpath_remove_disk(struct nvme_ns_head *head)
 454 {
 455 }
 456 static inline void nvme_mpath_add_disk_links(struct nvme_ns *ns)
 457 {
 458 }
 459 static inline void nvme_mpath_remove_disk_links(struct nvme_ns *ns)
 460 {
 461 }
 462 static inline void nvme_mpath_clear_current_path(struct nvme_ns *ns)
 463 {
 464 }
 465 static inline void nvme_mpath_check_last_path(struct nvme_ns *ns)
 466 {
 467 }
 468 #endif /* CONFIG_NVME_MULTIPATH */
 469
 470 #ifdef CONFIG_NVM
 471 int nvme_nvm_register(struct nvme_ns *ns, char *disk_name, int node);
 472 void nvme_nvm_unregister(struct nvme_ns *ns);
 473 int nvme_nvm_register_sysfs(struct nvme_ns *ns);
 474 void nvme_nvm_unregister_sysfs(struct nvme_ns *ns);
 475 int nvme_nvm_ioctl(struct nvme_ns *ns, unsigned int cmd, unsigned long arg);
 476 #else
 477 static inline int nvme_nvm_register(struct nvme_ns *ns, char *disk_name,
 478                                     int node)
 479 {
 480         return 0;
 481 }
 482
 483 static inline void nvme_nvm_unregister(struct nvme_ns *ns) {};
 484 static inline int nvme_nvm_register_sysfs(struct nvme_ns *ns)
 485 {
 486         return 0;
 487 }
 488 static inline void nvme_nvm_unregister_sysfs(struct nvme_ns *ns) {};
 489 static inline int nvme_nvm_ioctl(struct nvme_ns *ns, unsigned int cmd,
 490                                                         unsigned long arg)
 491 {
 492         return -ENOTTY;
 493 }
 494 #endif /* CONFIG_NVM */
 495
 496 static inline struct nvme_ns *nvme_get_ns_from_dev(struct device *dev)
 497 {
 498         return dev_to_disk(dev)->private_data;
 499 }
 500
 501 int __init nvme_core_init(void);
 502 void nvme_core_exit(void);
 503
 504 #endif /* _NVME_H */