1 /* SPDX-License-Identifier: GPL-2.0 */
3 * Copyright (c) 2011-2014, Intel Corporation.
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 #include <linux/wait.h>
20 #include <trace/events/block.h>
22 extern unsigned int nvme_io_timeout;
23 #define NVME_IO_TIMEOUT (nvme_io_timeout * HZ)
25 extern unsigned int admin_timeout;
26 #define ADMIN_TIMEOUT (admin_timeout * HZ)
28 #define NVME_DEFAULT_KATO 5
29 #define NVME_KATO_GRACE 10
31 #ifdef CONFIG_ARCH_NO_SG_CHAIN
32 #define NVME_INLINE_SG_CNT 0
34 #define NVME_INLINE_SG_CNT 2
37 extern struct workqueue_struct *nvme_wq;
38 extern struct workqueue_struct *nvme_reset_wq;
39 extern struct workqueue_struct *nvme_delete_wq;
47 * List of workarounds for devices that required behavior not specified in
52 * Prefers I/O aligned to a stripe size specified in a vendor
53 * specific Identify field.
55 NVME_QUIRK_STRIPE_SIZE = (1 << 0),
58 * The controller doesn't handle Identify value others than 0 or 1
61 NVME_QUIRK_IDENTIFY_CNS = (1 << 1),
64 * The controller deterministically returns O's on reads to
65 * logical blocks that deallocate was called on.
67 NVME_QUIRK_DEALLOCATE_ZEROES = (1 << 2),
70 * The controller needs a delay before starts checking the device
71 * readiness, which is done by reading the NVME_CSTS_RDY bit.
73 NVME_QUIRK_DELAY_BEFORE_CHK_RDY = (1 << 3),
76 * APST should not be used.
78 NVME_QUIRK_NO_APST = (1 << 4),
81 * The deepest sleep state should not be used.
83 NVME_QUIRK_NO_DEEPEST_PS = (1 << 5),
86 * Supports the LighNVM command set if indicated in vs[1].
88 NVME_QUIRK_LIGHTNVM = (1 << 6),
91 * Set MEDIUM priority on SQ creation
93 NVME_QUIRK_MEDIUM_PRIO_SQ = (1 << 7),
96 * Ignore device provided subnqn.
98 NVME_QUIRK_IGNORE_DEV_SUBNQN = (1 << 8),
101 * Broken Write Zeroes.
103 NVME_QUIRK_DISABLE_WRITE_ZEROES = (1 << 9),
106 * Force simple suspend/resume path.
108 NVME_QUIRK_SIMPLE_SUSPEND = (1 << 10),
111 * Use only one interrupt vector for all queues
113 NVME_QUIRK_SINGLE_VECTOR = (1 << 11),
116 * Use non-standard 128 bytes SQEs.
118 NVME_QUIRK_128_BYTES_SQES = (1 << 12),
121 * Prevent tag overlap between queues
123 NVME_QUIRK_SHARED_TAGS = (1 << 13),
126 * Don't change the value of the temperature threshold feature
128 NVME_QUIRK_NO_TEMP_THRESH_CHANGE = (1 << 14),
132 * Common request structure for NVMe passthrough. All drivers must have
133 * this structure as the first member of their request-private data.
135 struct nvme_request {
136 struct nvme_command *cmd;
137 union nvme_result result;
141 struct nvme_ctrl *ctrl;
145 * Mark a bio as coming in through the mpath node.
147 #define REQ_NVME_MPATH REQ_DRV
150 NVME_REQ_CANCELLED = (1 << 0),
151 NVME_REQ_USERCMD = (1 << 1),
154 static inline struct nvme_request *nvme_req(struct request *req)
156 return blk_mq_rq_to_pdu(req);
159 static inline u16 nvme_req_qid(struct request *req)
163 return blk_mq_unique_tag_to_hwq(blk_mq_unique_tag(req)) + 1;
166 /* The below value is the specific amount of delay needed before checking
167 * readiness in case of the PCI_DEVICE(0x1c58, 0x0003), which needs the
168 * NVME_QUIRK_DELAY_BEFORE_CHK_RDY quirk enabled. The value (in ms) was
171 #define NVME_QUIRK_DELAY_AMOUNT 2300
173 enum nvme_ctrl_state {
177 NVME_CTRL_CONNECTING,
182 struct nvme_fault_inject {
183 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
184 struct fault_attr attr;
185 struct dentry *parent;
186 bool dont_retry; /* DNR, do not retry */
187 u16 status; /* status code */
193 enum nvme_ctrl_state state;
196 struct mutex scan_lock;
197 const struct nvme_ctrl_ops *ops;
198 struct request_queue *admin_q;
199 struct request_queue *connect_q;
200 struct request_queue *fabrics_q;
204 struct blk_mq_tag_set *tagset;
205 struct blk_mq_tag_set *admin_tagset;
206 struct list_head namespaces;
207 struct rw_semaphore namespaces_rwsem;
208 struct device ctrl_device;
209 struct device *device; /* char device */
211 struct work_struct reset_work;
212 struct work_struct delete_work;
213 wait_queue_head_t state_wq;
215 struct nvme_subsystem *subsys;
216 struct list_head subsys_entry;
218 struct opal_dev *opal_dev;
238 atomic_t abort_limit;
250 unsigned int shutdown_timeout;
253 unsigned long quirks;
254 struct nvme_id_power_state psd[32];
255 struct nvme_effects_log *effects;
256 struct work_struct scan_work;
257 struct work_struct async_event_work;
258 struct delayed_work ka_work;
259 struct nvme_command ka_cmd;
260 struct work_struct fw_act_work;
261 unsigned long events;
264 #ifdef CONFIG_NVME_MULTIPATH
265 /* asymmetric namespace access: */
270 struct mutex ana_lock;
271 struct nvme_ana_rsp_hdr *ana_log_buf;
273 struct timer_list anatt_timer;
274 struct work_struct ana_work;
277 /* Power saving configuration */
278 u64 ps_max_latency_us;
293 struct nvmf_ctrl_options *opts;
295 struct page *discard_page;
296 unsigned long discard_page_busy;
298 struct nvme_fault_inject fault_inject;
306 struct nvme_subsystem {
310 * Because we unregister the device on the last put we need
311 * a separate refcount.
314 struct list_head entry;
316 struct list_head ctrls;
317 struct list_head nsheads;
318 char subnqn[NVMF_NQN_SIZE];
321 char firmware_rev[8];
324 u16 awupf; /* 0's based awupf value. */
326 #ifdef CONFIG_NVME_MULTIPATH
327 enum nvme_iopolicy iopolicy;
332 * Container structure for uniqueue namespace identifiers.
341 * Anchor structure for namespaces. There is one for each namespace in a
342 * NVMe subsystem that any of our controllers can see, and the namespace
343 * structure for each controller is chained of it. For private namespaces
344 * there is a 1:1 relation to our namespace structures, that is ->list
345 * only ever has a single entry for private namespaces.
347 struct nvme_ns_head {
348 struct list_head list;
349 struct srcu_struct srcu;
350 struct nvme_subsystem *subsys;
352 struct nvme_ns_ids ids;
353 struct list_head entry;
357 #ifdef CONFIG_NVME_MULTIPATH
358 struct gendisk *disk;
359 struct bio_list requeue_list;
360 spinlock_t requeue_lock;
361 struct work_struct requeue_work;
363 struct nvme_ns __rcu *current_path[];
368 struct list_head list;
370 struct nvme_ctrl *ctrl;
371 struct request_queue *queue;
372 struct gendisk *disk;
373 #ifdef CONFIG_NVME_MULTIPATH
374 enum nvme_ana_state ana_state;
377 struct list_head siblings;
378 struct nvm_dev *ndev;
380 struct nvme_ns_head *head;
389 #define NVME_NS_REMOVING 0
390 #define NVME_NS_DEAD 1
391 #define NVME_NS_ANA_PENDING 2
393 struct nvme_fault_inject fault_inject;
397 struct nvme_ctrl_ops {
399 struct module *module;
401 #define NVME_F_FABRICS (1 << 0)
402 #define NVME_F_METADATA_SUPPORTED (1 << 1)
403 #define NVME_F_PCI_P2PDMA (1 << 2)
404 int (*reg_read32)(struct nvme_ctrl *ctrl, u32 off, u32 *val);
405 int (*reg_write32)(struct nvme_ctrl *ctrl, u32 off, u32 val);
406 int (*reg_read64)(struct nvme_ctrl *ctrl, u32 off, u64 *val);
407 void (*free_ctrl)(struct nvme_ctrl *ctrl);
408 void (*submit_async_event)(struct nvme_ctrl *ctrl);
409 void (*delete_ctrl)(struct nvme_ctrl *ctrl);
410 int (*get_address)(struct nvme_ctrl *ctrl, char *buf, int size);
413 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
414 void nvme_fault_inject_init(struct nvme_fault_inject *fault_inj,
415 const char *dev_name);
416 void nvme_fault_inject_fini(struct nvme_fault_inject *fault_inject);
417 void nvme_should_fail(struct request *req);
419 static inline void nvme_fault_inject_init(struct nvme_fault_inject *fault_inj,
420 const char *dev_name)
423 static inline void nvme_fault_inject_fini(struct nvme_fault_inject *fault_inj)
426 static inline void nvme_should_fail(struct request *req) {}
429 static inline int nvme_reset_subsystem(struct nvme_ctrl *ctrl)
431 if (!ctrl->subsystem)
433 return ctrl->ops->reg_write32(ctrl, NVME_REG_NSSR, 0x4E564D65);
437 * Convert a 512B sector number to a device logical block number.
439 static inline u64 nvme_sect_to_lba(struct nvme_ns *ns, sector_t sector)
441 return sector >> (ns->lba_shift - SECTOR_SHIFT);
445 * Convert a device logical block number to a 512B sector number.
447 static inline sector_t nvme_lba_to_sect(struct nvme_ns *ns, u64 lba)
449 return lba << (ns->lba_shift - SECTOR_SHIFT);
453 * Convert byte length to nvme's 0-based num dwords
455 static inline u32 nvme_bytes_to_numd(size_t len)
457 return (len >> 2) - 1;
460 static inline void nvme_end_request(struct request *req, __le16 status,
461 union nvme_result result)
463 struct nvme_request *rq = nvme_req(req);
465 rq->status = le16_to_cpu(status) >> 1;
467 /* inject error when permitted by fault injection framework */
468 nvme_should_fail(req);
469 blk_mq_complete_request(req);
472 static inline void nvme_get_ctrl(struct nvme_ctrl *ctrl)
474 get_device(ctrl->device);
477 static inline void nvme_put_ctrl(struct nvme_ctrl *ctrl)
479 put_device(ctrl->device);
482 static inline bool nvme_is_aen_req(u16 qid, __u16 command_id)
484 return !qid && command_id >= NVME_AQ_BLK_MQ_DEPTH;
487 void nvme_complete_rq(struct request *req);
488 bool nvme_cancel_request(struct request *req, void *data, bool reserved);
489 bool nvme_change_ctrl_state(struct nvme_ctrl *ctrl,
490 enum nvme_ctrl_state new_state);
491 bool nvme_wait_reset(struct nvme_ctrl *ctrl);
492 int nvme_disable_ctrl(struct nvme_ctrl *ctrl);
493 int nvme_enable_ctrl(struct nvme_ctrl *ctrl);
494 int nvme_shutdown_ctrl(struct nvme_ctrl *ctrl);
495 int nvme_init_ctrl(struct nvme_ctrl *ctrl, struct device *dev,
496 const struct nvme_ctrl_ops *ops, unsigned long quirks);
497 void nvme_uninit_ctrl(struct nvme_ctrl *ctrl);
498 void nvme_start_ctrl(struct nvme_ctrl *ctrl);
499 void nvme_stop_ctrl(struct nvme_ctrl *ctrl);
500 int nvme_init_identify(struct nvme_ctrl *ctrl);
502 void nvme_remove_namespaces(struct nvme_ctrl *ctrl);
504 int nvme_sec_submit(void *data, u16 spsp, u8 secp, void *buffer, size_t len,
507 void nvme_complete_async_event(struct nvme_ctrl *ctrl, __le16 status,
508 volatile union nvme_result *res);
510 void nvme_stop_queues(struct nvme_ctrl *ctrl);
511 void nvme_start_queues(struct nvme_ctrl *ctrl);
512 void nvme_kill_queues(struct nvme_ctrl *ctrl);
513 void nvme_sync_queues(struct nvme_ctrl *ctrl);
514 void nvme_unfreeze(struct nvme_ctrl *ctrl);
515 void nvme_wait_freeze(struct nvme_ctrl *ctrl);
516 void nvme_wait_freeze_timeout(struct nvme_ctrl *ctrl, long timeout);
517 void nvme_start_freeze(struct nvme_ctrl *ctrl);
519 #define NVME_QID_ANY -1
520 struct request *nvme_alloc_request(struct request_queue *q,
521 struct nvme_command *cmd, blk_mq_req_flags_t flags, int qid);
522 void nvme_cleanup_cmd(struct request *req);
523 blk_status_t nvme_setup_cmd(struct nvme_ns *ns, struct request *req,
524 struct nvme_command *cmd);
525 int nvme_submit_sync_cmd(struct request_queue *q, struct nvme_command *cmd,
526 void *buf, unsigned bufflen);
527 int __nvme_submit_sync_cmd(struct request_queue *q, struct nvme_command *cmd,
528 union nvme_result *result, void *buffer, unsigned bufflen,
529 unsigned timeout, int qid, int at_head,
530 blk_mq_req_flags_t flags, bool poll);
531 int nvme_set_features(struct nvme_ctrl *dev, unsigned int fid,
532 unsigned int dword11, void *buffer, size_t buflen,
534 int nvme_get_features(struct nvme_ctrl *dev, unsigned int fid,
535 unsigned int dword11, void *buffer, size_t buflen,
537 int nvme_set_queue_count(struct nvme_ctrl *ctrl, int *count);
538 void nvme_stop_keep_alive(struct nvme_ctrl *ctrl);
539 int nvme_reset_ctrl(struct nvme_ctrl *ctrl);
540 int nvme_reset_ctrl_sync(struct nvme_ctrl *ctrl);
541 int nvme_try_sched_reset(struct nvme_ctrl *ctrl);
542 int nvme_delete_ctrl(struct nvme_ctrl *ctrl);
544 int nvme_get_log(struct nvme_ctrl *ctrl, u32 nsid, u8 log_page, u8 lsp,
545 void *log, size_t size, u64 offset);
547 extern const struct attribute_group *nvme_ns_id_attr_groups[];
548 extern const struct block_device_operations nvme_ns_head_ops;
550 #ifdef CONFIG_NVME_MULTIPATH
551 static inline bool nvme_ctrl_use_ana(struct nvme_ctrl *ctrl)
553 return ctrl->ana_log_buf != NULL;
556 void nvme_mpath_unfreeze(struct nvme_subsystem *subsys);
557 void nvme_mpath_wait_freeze(struct nvme_subsystem *subsys);
558 void nvme_mpath_start_freeze(struct nvme_subsystem *subsys);
559 void nvme_set_disk_name(char *disk_name, struct nvme_ns *ns,
560 struct nvme_ctrl *ctrl, int *flags);
561 bool nvme_failover_req(struct request *req);
562 void nvme_kick_requeue_lists(struct nvme_ctrl *ctrl);
563 int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl,struct nvme_ns_head *head);
564 void nvme_mpath_add_disk(struct nvme_ns *ns, struct nvme_id_ns *id);
565 void nvme_mpath_remove_disk(struct nvme_ns_head *head);
566 int nvme_mpath_init(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id);
567 void nvme_mpath_uninit(struct nvme_ctrl *ctrl);
568 void nvme_mpath_stop(struct nvme_ctrl *ctrl);
569 bool nvme_mpath_clear_current_path(struct nvme_ns *ns);
570 void nvme_mpath_clear_ctrl_paths(struct nvme_ctrl *ctrl);
571 struct nvme_ns *nvme_find_path(struct nvme_ns_head *head);
573 static inline void nvme_mpath_check_last_path(struct nvme_ns *ns)
575 struct nvme_ns_head *head = ns->head;
577 if (head->disk && list_empty(&head->list))
578 kblockd_schedule_work(&head->requeue_work);
581 static inline void nvme_trace_bio_complete(struct request *req,
584 struct nvme_ns *ns = req->q->queuedata;
586 if (req->cmd_flags & REQ_NVME_MPATH)
587 trace_block_bio_complete(ns->head->disk->queue,
591 extern struct device_attribute dev_attr_ana_grpid;
592 extern struct device_attribute dev_attr_ana_state;
593 extern struct device_attribute subsys_attr_iopolicy;
596 static inline bool nvme_ctrl_use_ana(struct nvme_ctrl *ctrl)
601 * Without the multipath code enabled, multiple controller per subsystems are
602 * visible as devices and thus we cannot use the subsystem instance.
604 static inline void nvme_set_disk_name(char *disk_name, struct nvme_ns *ns,
605 struct nvme_ctrl *ctrl, int *flags)
607 sprintf(disk_name, "nvme%dn%d", ctrl->instance, ns->head->instance);
610 static inline bool nvme_failover_req(struct request *req)
614 static inline void nvme_kick_requeue_lists(struct nvme_ctrl *ctrl)
617 static inline int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl,
618 struct nvme_ns_head *head)
622 static inline void nvme_mpath_add_disk(struct nvme_ns *ns,
623 struct nvme_id_ns *id)
626 static inline void nvme_mpath_remove_disk(struct nvme_ns_head *head)
629 static inline bool nvme_mpath_clear_current_path(struct nvme_ns *ns)
633 static inline void nvme_mpath_clear_ctrl_paths(struct nvme_ctrl *ctrl)
636 static inline void nvme_mpath_check_last_path(struct nvme_ns *ns)
639 static inline void nvme_trace_bio_complete(struct request *req,
643 static inline int nvme_mpath_init(struct nvme_ctrl *ctrl,
644 struct nvme_id_ctrl *id)
646 if (ctrl->subsys->cmic & (1 << 3))
647 dev_warn(ctrl->device,
648 "Please enable CONFIG_NVME_MULTIPATH for full support of multi-port devices.\n");
651 static inline void nvme_mpath_uninit(struct nvme_ctrl *ctrl)
654 static inline void nvme_mpath_stop(struct nvme_ctrl *ctrl)
657 static inline void nvme_mpath_unfreeze(struct nvme_subsystem *subsys)
660 static inline void nvme_mpath_wait_freeze(struct nvme_subsystem *subsys)
663 static inline void nvme_mpath_start_freeze(struct nvme_subsystem *subsys)
666 #endif /* CONFIG_NVME_MULTIPATH */
669 int nvme_nvm_register(struct nvme_ns *ns, char *disk_name, int node);
670 void nvme_nvm_unregister(struct nvme_ns *ns);
671 extern const struct attribute_group nvme_nvm_attr_group;
672 int nvme_nvm_ioctl(struct nvme_ns *ns, unsigned int cmd, unsigned long arg);
674 static inline int nvme_nvm_register(struct nvme_ns *ns, char *disk_name,
680 static inline void nvme_nvm_unregister(struct nvme_ns *ns) {};
681 static inline int nvme_nvm_ioctl(struct nvme_ns *ns, unsigned int cmd,
686 #endif /* CONFIG_NVM */
688 static inline struct nvme_ns *nvme_get_ns_from_dev(struct device *dev)
690 return dev_to_disk(dev)->private_data;
693 #ifdef CONFIG_NVME_HWMON
694 void nvme_hwmon_init(struct nvme_ctrl *ctrl);
696 static inline void nvme_hwmon_init(struct nvme_ctrl *ctrl) { }