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 #include <trace/events/block.h> 20 21 extern unsigned int nvme_io_timeout; 22 #define NVME_IO_TIMEOUT (nvme_io_timeout * HZ) 23 24 extern unsigned int admin_timeout; 25 #define ADMIN_TIMEOUT (admin_timeout * HZ) 26 27 #define NVME_DEFAULT_KATO 5 28 #define NVME_KATO_GRACE 10 29 30 extern struct workqueue_struct *nvme_wq; 31 extern struct workqueue_struct *nvme_reset_wq; 32 extern struct workqueue_struct *nvme_delete_wq; 33 34 enum { 35 NVME_NS_LBA = 0, 36 NVME_NS_LIGHTNVM = 1, 37 }; 38 39 /* 40 * List of workarounds for devices that required behavior not specified in 41 * the standard. 42 */ 43 enum nvme_quirks { 44 /* 45 * Prefers I/O aligned to a stripe size specified in a vendor 46 * specific Identify field. 47 */ 48 NVME_QUIRK_STRIPE_SIZE = (1 << 0), 49 50 /* 51 * The controller doesn't handle Identify value others than 0 or 1 52 * correctly. 53 */ 54 NVME_QUIRK_IDENTIFY_CNS = (1 << 1), 55 56 /* 57 * The controller deterministically returns O's on reads to 58 * logical blocks that deallocate was called on. 59 */ 60 NVME_QUIRK_DEALLOCATE_ZEROES = (1 << 2), 61 62 /* 63 * The controller needs a delay before starts checking the device 64 * readiness, which is done by reading the NVME_CSTS_RDY bit. 65 */ 66 NVME_QUIRK_DELAY_BEFORE_CHK_RDY = (1 << 3), 67 68 /* 69 * APST should not be used. 70 */ 71 NVME_QUIRK_NO_APST = (1 << 4), 72 73 /* 74 * The deepest sleep state should not be used. 75 */ 76 NVME_QUIRK_NO_DEEPEST_PS = (1 << 5), 77 78 /* 79 * Supports the LighNVM command set if indicated in vs[1]. 80 */ 81 NVME_QUIRK_LIGHTNVM = (1 << 6), 82 83 /* 84 * Set MEDIUM priority on SQ creation 85 */ 86 NVME_QUIRK_MEDIUM_PRIO_SQ = (1 << 7), 87 88 /* 89 * Ignore device provided subnqn. 90 */ 91 NVME_QUIRK_IGNORE_DEV_SUBNQN = (1 << 8), 92 93 /* 94 * Broken Write Zeroes. 95 */ 96 NVME_QUIRK_DISABLE_WRITE_ZEROES = (1 << 9), 97 98 /* 99 * Force simple suspend/resume path. 100 */ 101 NVME_QUIRK_SIMPLE_SUSPEND = (1 << 10), 102 103 /* 104 * Use only one interrupt vector for all queues 105 */ 106 NVME_QUIRK_SINGLE_VECTOR = (1 << 11), 107 108 /* 109 * Use non-standard 128 bytes SQEs. 110 */ 111 NVME_QUIRK_128_BYTES_SQES = (1 << 12), 112 113 /* 114 * Prevent tag overlap between queues 115 */ 116 NVME_QUIRK_SHARED_TAGS = (1 << 13), 117 }; 118 119 /* 120 * Common request structure for NVMe passthrough. All drivers must have 121 * this structure as the first member of their request-private data. 122 */ 123 struct nvme_request { 124 struct nvme_command *cmd; 125 union nvme_result result; 126 u8 retries; 127 u8 flags; 128 u16 status; 129 struct nvme_ctrl *ctrl; 130 }; 131 132 /* 133 * Mark a bio as coming in through the mpath node. 134 */ 135 #define REQ_NVME_MPATH REQ_DRV 136 137 enum { 138 NVME_REQ_CANCELLED = (1 << 0), 139 NVME_REQ_USERCMD = (1 << 1), 140 }; 141 142 static inline struct nvme_request *nvme_req(struct request *req) 143 { 144 return blk_mq_rq_to_pdu(req); 145 } 146 147 static inline u16 nvme_req_qid(struct request *req) 148 { 149 if (!req->rq_disk) 150 return 0; 151 return blk_mq_unique_tag_to_hwq(blk_mq_unique_tag(req)) + 1; 152 } 153 154 /* The below value is the specific amount of delay needed before checking 155 * readiness in case of the PCI_DEVICE(0x1c58, 0x0003), which needs the 156 * NVME_QUIRK_DELAY_BEFORE_CHK_RDY quirk enabled. The value (in ms) was 157 * found empirically. 158 */ 159 #define NVME_QUIRK_DELAY_AMOUNT 2300 160 161 enum nvme_ctrl_state { 162 NVME_CTRL_NEW, 163 NVME_CTRL_LIVE, 164 NVME_CTRL_ADMIN_ONLY, /* Only admin queue live */ 165 NVME_CTRL_RESETTING, 166 NVME_CTRL_CONNECTING, 167 NVME_CTRL_DELETING, 168 NVME_CTRL_DEAD, 169 }; 170 171 struct nvme_fault_inject { 172 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 173 struct fault_attr attr; 174 struct dentry *parent; 175 bool dont_retry; /* DNR, do not retry */ 176 u16 status; /* status code */ 177 #endif 178 }; 179 180 struct nvme_ctrl { 181 bool comp_seen; 182 enum nvme_ctrl_state state; 183 bool identified; 184 spinlock_t lock; 185 struct mutex scan_lock; 186 const struct nvme_ctrl_ops *ops; 187 struct request_queue *admin_q; 188 struct request_queue *connect_q; 189 struct request_queue *fabrics_q; 190 struct device *dev; 191 int instance; 192 int numa_node; 193 struct blk_mq_tag_set *tagset; 194 struct blk_mq_tag_set *admin_tagset; 195 struct list_head namespaces; 196 struct rw_semaphore namespaces_rwsem; 197 struct device ctrl_device; 198 struct device *device; /* char device */ 199 struct cdev cdev; 200 struct work_struct reset_work; 201 struct work_struct delete_work; 202 203 struct nvme_subsystem *subsys; 204 struct list_head subsys_entry; 205 206 struct opal_dev *opal_dev; 207 208 char name[12]; 209 u16 cntlid; 210 211 u32 ctrl_config; 212 u16 mtfa; 213 u32 queue_count; 214 215 u64 cap; 216 u32 page_size; 217 u32 max_hw_sectors; 218 u32 max_segments; 219 u16 crdt[3]; 220 u16 oncs; 221 u16 oacs; 222 u16 nssa; 223 u16 nr_streams; 224 u32 max_namespaces; 225 atomic_t abort_limit; 226 u8 vwc; 227 u32 vs; 228 u32 sgls; 229 u16 kas; 230 u8 npss; 231 u8 apsta; 232 u32 oaes; 233 u32 aen_result; 234 u32 ctratt; 235 unsigned int shutdown_timeout; 236 unsigned int kato; 237 bool subsystem; 238 unsigned long quirks; 239 struct nvme_id_power_state psd[32]; 240 struct nvme_effects_log *effects; 241 struct work_struct scan_work; 242 struct work_struct async_event_work; 243 struct delayed_work ka_work; 244 struct nvme_command ka_cmd; 245 struct work_struct fw_act_work; 246 unsigned long events; 247 248 #ifdef CONFIG_NVME_MULTIPATH 249 /* asymmetric namespace access: */ 250 u8 anacap; 251 u8 anatt; 252 u32 anagrpmax; 253 u32 nanagrpid; 254 struct mutex ana_lock; 255 struct nvme_ana_rsp_hdr *ana_log_buf; 256 size_t ana_log_size; 257 struct timer_list anatt_timer; 258 struct work_struct ana_work; 259 #endif 260 261 /* Power saving configuration */ 262 u64 ps_max_latency_us; 263 bool apst_enabled; 264 265 /* PCIe only: */ 266 u32 hmpre; 267 u32 hmmin; 268 u32 hmminds; 269 u16 hmmaxd; 270 271 /* Fabrics only */ 272 u16 sqsize; 273 u32 ioccsz; 274 u32 iorcsz; 275 u16 icdoff; 276 u16 maxcmd; 277 int nr_reconnects; 278 struct nvmf_ctrl_options *opts; 279 280 struct page *discard_page; 281 unsigned long discard_page_busy; 282 283 struct nvme_fault_inject fault_inject; 284 }; 285 286 enum nvme_iopolicy { 287 NVME_IOPOLICY_NUMA, 288 NVME_IOPOLICY_RR, 289 }; 290 291 struct nvme_subsystem { 292 int instance; 293 struct device dev; 294 /* 295 * Because we unregister the device on the last put we need 296 * a separate refcount. 297 */ 298 struct kref ref; 299 struct list_head entry; 300 struct mutex lock; 301 struct list_head ctrls; 302 struct list_head nsheads; 303 char subnqn[NVMF_NQN_SIZE]; 304 char serial[20]; 305 char model[40]; 306 char firmware_rev[8]; 307 u8 cmic; 308 u16 vendor_id; 309 u16 awupf; /* 0's based awupf value. */ 310 struct ida ns_ida; 311 #ifdef CONFIG_NVME_MULTIPATH 312 enum nvme_iopolicy iopolicy; 313 #endif 314 }; 315 316 /* 317 * Container structure for uniqueue namespace identifiers. 318 */ 319 struct nvme_ns_ids { 320 u8 eui64[8]; 321 u8 nguid[16]; 322 uuid_t uuid; 323 }; 324 325 /* 326 * Anchor structure for namespaces. There is one for each namespace in a 327 * NVMe subsystem that any of our controllers can see, and the namespace 328 * structure for each controller is chained of it. For private namespaces 329 * there is a 1:1 relation to our namespace structures, that is ->list 330 * only ever has a single entry for private namespaces. 331 */ 332 struct nvme_ns_head { 333 struct list_head list; 334 struct srcu_struct srcu; 335 struct nvme_subsystem *subsys; 336 unsigned ns_id; 337 struct nvme_ns_ids ids; 338 struct list_head entry; 339 struct kref ref; 340 int instance; 341 #ifdef CONFIG_NVME_MULTIPATH 342 struct gendisk *disk; 343 struct bio_list requeue_list; 344 spinlock_t requeue_lock; 345 struct work_struct requeue_work; 346 struct mutex lock; 347 struct nvme_ns __rcu *current_path[]; 348 #endif 349 }; 350 351 struct nvme_ns { 352 struct list_head list; 353 354 struct nvme_ctrl *ctrl; 355 struct request_queue *queue; 356 struct gendisk *disk; 357 #ifdef CONFIG_NVME_MULTIPATH 358 enum nvme_ana_state ana_state; 359 u32 ana_grpid; 360 #endif 361 struct list_head siblings; 362 struct nvm_dev *ndev; 363 struct kref kref; 364 struct nvme_ns_head *head; 365 366 int lba_shift; 367 u16 ms; 368 u16 sgs; 369 u32 sws; 370 bool ext; 371 u8 pi_type; 372 unsigned long flags; 373 #define NVME_NS_REMOVING 0 374 #define NVME_NS_DEAD 1 375 #define NVME_NS_ANA_PENDING 2 376 u16 noiob; 377 378 struct nvme_fault_inject fault_inject; 379 380 }; 381 382 struct nvme_ctrl_ops { 383 const char *name; 384 struct module *module; 385 unsigned int flags; 386 #define NVME_F_FABRICS (1 << 0) 387 #define NVME_F_METADATA_SUPPORTED (1 << 1) 388 #define NVME_F_PCI_P2PDMA (1 << 2) 389 int (*reg_read32)(struct nvme_ctrl *ctrl, u32 off, u32 *val); 390 int (*reg_write32)(struct nvme_ctrl *ctrl, u32 off, u32 val); 391 int (*reg_read64)(struct nvme_ctrl *ctrl, u32 off, u64 *val); 392 void (*free_ctrl)(struct nvme_ctrl *ctrl); 393 void (*submit_async_event)(struct nvme_ctrl *ctrl); 394 void (*delete_ctrl)(struct nvme_ctrl *ctrl); 395 int (*get_address)(struct nvme_ctrl *ctrl, char *buf, int size); 396 }; 397 398 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 399 void nvme_fault_inject_init(struct nvme_fault_inject *fault_inj, 400 const char *dev_name); 401 void nvme_fault_inject_fini(struct nvme_fault_inject *fault_inject); 402 void nvme_should_fail(struct request *req); 403 #else 404 static inline void nvme_fault_inject_init(struct nvme_fault_inject *fault_inj, 405 const char *dev_name) 406 { 407 } 408 static inline void nvme_fault_inject_fini(struct nvme_fault_inject *fault_inj) 409 { 410 } 411 static inline void nvme_should_fail(struct request *req) {} 412 #endif 413 414 static inline int nvme_reset_subsystem(struct nvme_ctrl *ctrl) 415 { 416 if (!ctrl->subsystem) 417 return -ENOTTY; 418 return ctrl->ops->reg_write32(ctrl, NVME_REG_NSSR, 0x4E564D65); 419 } 420 421 static inline u64 nvme_block_nr(struct nvme_ns *ns, sector_t sector) 422 { 423 return (sector >> (ns->lba_shift - 9)); 424 } 425 426 static inline void nvme_end_request(struct request *req, __le16 status, 427 union nvme_result result) 428 { 429 struct nvme_request *rq = nvme_req(req); 430 431 rq->status = le16_to_cpu(status) >> 1; 432 rq->result = result; 433 /* inject error when permitted by fault injection framework */ 434 nvme_should_fail(req); 435 blk_mq_complete_request(req); 436 } 437 438 static inline void nvme_get_ctrl(struct nvme_ctrl *ctrl) 439 { 440 get_device(ctrl->device); 441 } 442 443 static inline void nvme_put_ctrl(struct nvme_ctrl *ctrl) 444 { 445 put_device(ctrl->device); 446 } 447 448 void nvme_complete_rq(struct request *req); 449 bool nvme_cancel_request(struct request *req, void *data, bool reserved); 450 bool nvme_change_ctrl_state(struct nvme_ctrl *ctrl, 451 enum nvme_ctrl_state new_state); 452 int nvme_disable_ctrl(struct nvme_ctrl *ctrl); 453 int nvme_enable_ctrl(struct nvme_ctrl *ctrl); 454 int nvme_shutdown_ctrl(struct nvme_ctrl *ctrl); 455 int nvme_init_ctrl(struct nvme_ctrl *ctrl, struct device *dev, 456 const struct nvme_ctrl_ops *ops, unsigned long quirks); 457 void nvme_uninit_ctrl(struct nvme_ctrl *ctrl); 458 void nvme_start_ctrl(struct nvme_ctrl *ctrl); 459 void nvme_stop_ctrl(struct nvme_ctrl *ctrl); 460 void nvme_put_ctrl(struct nvme_ctrl *ctrl); 461 int nvme_init_identify(struct nvme_ctrl *ctrl); 462 463 void nvme_remove_namespaces(struct nvme_ctrl *ctrl); 464 465 int nvme_sec_submit(void *data, u16 spsp, u8 secp, void *buffer, size_t len, 466 bool send); 467 468 void nvme_complete_async_event(struct nvme_ctrl *ctrl, __le16 status, 469 volatile union nvme_result *res); 470 471 void nvme_stop_queues(struct nvme_ctrl *ctrl); 472 void nvme_start_queues(struct nvme_ctrl *ctrl); 473 void nvme_kill_queues(struct nvme_ctrl *ctrl); 474 void nvme_sync_queues(struct nvme_ctrl *ctrl); 475 void nvme_unfreeze(struct nvme_ctrl *ctrl); 476 void nvme_wait_freeze(struct nvme_ctrl *ctrl); 477 void nvme_wait_freeze_timeout(struct nvme_ctrl *ctrl, long timeout); 478 void nvme_start_freeze(struct nvme_ctrl *ctrl); 479 480 #define NVME_QID_ANY -1 481 struct request *nvme_alloc_request(struct request_queue *q, 482 struct nvme_command *cmd, blk_mq_req_flags_t flags, int qid); 483 void nvme_cleanup_cmd(struct request *req); 484 blk_status_t nvme_setup_cmd(struct nvme_ns *ns, struct request *req, 485 struct nvme_command *cmd); 486 int nvme_submit_sync_cmd(struct request_queue *q, struct nvme_command *cmd, 487 void *buf, unsigned bufflen); 488 int __nvme_submit_sync_cmd(struct request_queue *q, struct nvme_command *cmd, 489 union nvme_result *result, void *buffer, unsigned bufflen, 490 unsigned timeout, int qid, int at_head, 491 blk_mq_req_flags_t flags, bool poll); 492 int nvme_set_features(struct nvme_ctrl *dev, unsigned int fid, 493 unsigned int dword11, void *buffer, size_t buflen, 494 u32 *result); 495 int nvme_get_features(struct nvme_ctrl *dev, unsigned int fid, 496 unsigned int dword11, void *buffer, size_t buflen, 497 u32 *result); 498 int nvme_set_queue_count(struct nvme_ctrl *ctrl, int *count); 499 void nvme_stop_keep_alive(struct nvme_ctrl *ctrl); 500 int nvme_reset_ctrl(struct nvme_ctrl *ctrl); 501 int nvme_reset_ctrl_sync(struct nvme_ctrl *ctrl); 502 int nvme_delete_ctrl(struct nvme_ctrl *ctrl); 503 504 int nvme_get_log(struct nvme_ctrl *ctrl, u32 nsid, u8 log_page, u8 lsp, 505 void *log, size_t size, u64 offset); 506 507 extern const struct attribute_group *nvme_ns_id_attr_groups[]; 508 extern const struct block_device_operations nvme_ns_head_ops; 509 510 #ifdef CONFIG_NVME_MULTIPATH 511 static inline bool nvme_ctrl_use_ana(struct nvme_ctrl *ctrl) 512 { 513 return ctrl->ana_log_buf != NULL; 514 } 515 516 void nvme_mpath_unfreeze(struct nvme_subsystem *subsys); 517 void nvme_mpath_wait_freeze(struct nvme_subsystem *subsys); 518 void nvme_mpath_start_freeze(struct nvme_subsystem *subsys); 519 void nvme_set_disk_name(char *disk_name, struct nvme_ns *ns, 520 struct nvme_ctrl *ctrl, int *flags); 521 void nvme_failover_req(struct request *req); 522 void nvme_kick_requeue_lists(struct nvme_ctrl *ctrl); 523 int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl,struct nvme_ns_head *head); 524 void nvme_mpath_add_disk(struct nvme_ns *ns, struct nvme_id_ns *id); 525 void nvme_mpath_remove_disk(struct nvme_ns_head *head); 526 int nvme_mpath_init(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id); 527 void nvme_mpath_uninit(struct nvme_ctrl *ctrl); 528 void nvme_mpath_stop(struct nvme_ctrl *ctrl); 529 bool nvme_mpath_clear_current_path(struct nvme_ns *ns); 530 void nvme_mpath_clear_ctrl_paths(struct nvme_ctrl *ctrl); 531 struct nvme_ns *nvme_find_path(struct nvme_ns_head *head); 532 533 static inline void nvme_mpath_check_last_path(struct nvme_ns *ns) 534 { 535 struct nvme_ns_head *head = ns->head; 536 537 if (head->disk && list_empty(&head->list)) 538 kblockd_schedule_work(&head->requeue_work); 539 } 540 541 static inline void nvme_trace_bio_complete(struct request *req, 542 blk_status_t status) 543 { 544 struct nvme_ns *ns = req->q->queuedata; 545 546 if (req->cmd_flags & REQ_NVME_MPATH) 547 trace_block_bio_complete(ns->head->disk->queue, 548 req->bio, status); 549 } 550 551 extern struct device_attribute dev_attr_ana_grpid; 552 extern struct device_attribute dev_attr_ana_state; 553 extern struct device_attribute subsys_attr_iopolicy; 554 555 #else 556 static inline bool nvme_ctrl_use_ana(struct nvme_ctrl *ctrl) 557 { 558 return false; 559 } 560 /* 561 * Without the multipath code enabled, multiple controller per subsystems are 562 * visible as devices and thus we cannot use the subsystem instance. 563 */ 564 static inline void nvme_set_disk_name(char *disk_name, struct nvme_ns *ns, 565 struct nvme_ctrl *ctrl, int *flags) 566 { 567 sprintf(disk_name, "nvme%dn%d", ctrl->instance, ns->head->instance); 568 } 569 570 static inline void nvme_failover_req(struct request *req) 571 { 572 } 573 static inline void nvme_kick_requeue_lists(struct nvme_ctrl *ctrl) 574 { 575 } 576 static inline int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl, 577 struct nvme_ns_head *head) 578 { 579 return 0; 580 } 581 static inline void nvme_mpath_add_disk(struct nvme_ns *ns, 582 struct nvme_id_ns *id) 583 { 584 } 585 static inline void nvme_mpath_remove_disk(struct nvme_ns_head *head) 586 { 587 } 588 static inline bool nvme_mpath_clear_current_path(struct nvme_ns *ns) 589 { 590 return false; 591 } 592 static inline void nvme_mpath_clear_ctrl_paths(struct nvme_ctrl *ctrl) 593 { 594 } 595 static inline void nvme_mpath_check_last_path(struct nvme_ns *ns) 596 { 597 } 598 static inline void nvme_trace_bio_complete(struct request *req, 599 blk_status_t status) 600 { 601 } 602 static inline int nvme_mpath_init(struct nvme_ctrl *ctrl, 603 struct nvme_id_ctrl *id) 604 { 605 if (ctrl->subsys->cmic & (1 << 3)) 606 dev_warn(ctrl->device, 607 "Please enable CONFIG_NVME_MULTIPATH for full support of multi-port devices.\n"); 608 return 0; 609 } 610 static inline void nvme_mpath_uninit(struct nvme_ctrl *ctrl) 611 { 612 } 613 static inline void nvme_mpath_stop(struct nvme_ctrl *ctrl) 614 { 615 } 616 static inline void nvme_mpath_unfreeze(struct nvme_subsystem *subsys) 617 { 618 } 619 static inline void nvme_mpath_wait_freeze(struct nvme_subsystem *subsys) 620 { 621 } 622 static inline void nvme_mpath_start_freeze(struct nvme_subsystem *subsys) 623 { 624 } 625 #endif /* CONFIG_NVME_MULTIPATH */ 626 627 #ifdef CONFIG_NVM 628 int nvme_nvm_register(struct nvme_ns *ns, char *disk_name, int node); 629 void nvme_nvm_unregister(struct nvme_ns *ns); 630 extern const struct attribute_group nvme_nvm_attr_group; 631 int nvme_nvm_ioctl(struct nvme_ns *ns, unsigned int cmd, unsigned long arg); 632 #else 633 static inline int nvme_nvm_register(struct nvme_ns *ns, char *disk_name, 634 int node) 635 { 636 return 0; 637 } 638 639 static inline void nvme_nvm_unregister(struct nvme_ns *ns) {}; 640 static inline int nvme_nvm_ioctl(struct nvme_ns *ns, unsigned int cmd, 641 unsigned long arg) 642 { 643 return -ENOTTY; 644 } 645 #endif /* CONFIG_NVM */ 646 647 static inline struct nvme_ns *nvme_get_ns_from_dev(struct device *dev) 648 { 649 return dev_to_disk(dev)->private_data; 650 } 651 652 #endif /* _NVME_H */ 653