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>
19 #include <linux/t10-pi.h>
21 #include <trace/events/block.h>
23 extern unsigned int nvme_io_timeout;
24 #define NVME_IO_TIMEOUT (nvme_io_timeout * HZ)
26 extern unsigned int admin_timeout;
27 #define ADMIN_TIMEOUT (admin_timeout * HZ)
29 #define NVME_DEFAULT_KATO 5
30 #define NVME_KATO_GRACE 10
32 #ifdef CONFIG_ARCH_NO_SG_CHAIN
33 #define NVME_INLINE_SG_CNT 0
34 #define NVME_INLINE_METADATA_SG_CNT 0
36 #define NVME_INLINE_SG_CNT 2
37 #define NVME_INLINE_METADATA_SG_CNT 1
41 * Default to a 4K page size, with the intention to update this
42 * path in the future to accommodate architectures with differing
43 * kernel and IO page sizes.
45 #define NVME_CTRL_PAGE_SHIFT 12
46 #define NVME_CTRL_PAGE_SIZE (1 << NVME_CTRL_PAGE_SHIFT)
48 extern struct workqueue_struct *nvme_wq;
49 extern struct workqueue_struct *nvme_reset_wq;
50 extern struct workqueue_struct *nvme_delete_wq;
58 * List of workarounds for devices that required behavior not specified in
63 * Prefers I/O aligned to a stripe size specified in a vendor
64 * specific Identify field.
66 NVME_QUIRK_STRIPE_SIZE = (1 << 0),
69 * The controller doesn't handle Identify value others than 0 or 1
72 NVME_QUIRK_IDENTIFY_CNS = (1 << 1),
75 * The controller deterministically returns O's on reads to
76 * logical blocks that deallocate was called on.
78 NVME_QUIRK_DEALLOCATE_ZEROES = (1 << 2),
81 * The controller needs a delay before starts checking the device
82 * readiness, which is done by reading the NVME_CSTS_RDY bit.
84 NVME_QUIRK_DELAY_BEFORE_CHK_RDY = (1 << 3),
87 * APST should not be used.
89 NVME_QUIRK_NO_APST = (1 << 4),
92 * The deepest sleep state should not be used.
94 NVME_QUIRK_NO_DEEPEST_PS = (1 << 5),
97 * Supports the LighNVM command set if indicated in vs[1].
99 NVME_QUIRK_LIGHTNVM = (1 << 6),
102 * Set MEDIUM priority on SQ creation
104 NVME_QUIRK_MEDIUM_PRIO_SQ = (1 << 7),
107 * Ignore device provided subnqn.
109 NVME_QUIRK_IGNORE_DEV_SUBNQN = (1 << 8),
112 * Broken Write Zeroes.
114 NVME_QUIRK_DISABLE_WRITE_ZEROES = (1 << 9),
117 * Force simple suspend/resume path.
119 NVME_QUIRK_SIMPLE_SUSPEND = (1 << 10),
122 * Use only one interrupt vector for all queues
124 NVME_QUIRK_SINGLE_VECTOR = (1 << 11),
127 * Use non-standard 128 bytes SQEs.
129 NVME_QUIRK_128_BYTES_SQES = (1 << 12),
132 * Prevent tag overlap between queues
134 NVME_QUIRK_SHARED_TAGS = (1 << 13),
137 * Don't change the value of the temperature threshold feature
139 NVME_QUIRK_NO_TEMP_THRESH_CHANGE = (1 << 14),
143 * Common request structure for NVMe passthrough. All drivers must have
144 * this structure as the first member of their request-private data.
146 struct nvme_request {
147 struct nvme_command *cmd;
148 union nvme_result result;
152 struct nvme_ctrl *ctrl;
156 * Mark a bio as coming in through the mpath node.
158 #define REQ_NVME_MPATH REQ_DRV
161 NVME_REQ_CANCELLED = (1 << 0),
162 NVME_REQ_USERCMD = (1 << 1),
165 static inline struct nvme_request *nvme_req(struct request *req)
167 return blk_mq_rq_to_pdu(req);
170 static inline u16 nvme_req_qid(struct request *req)
174 return blk_mq_unique_tag_to_hwq(blk_mq_unique_tag(req)) + 1;
177 /* The below value is the specific amount of delay needed before checking
178 * readiness in case of the PCI_DEVICE(0x1c58, 0x0003), which needs the
179 * NVME_QUIRK_DELAY_BEFORE_CHK_RDY quirk enabled. The value (in ms) was
182 #define NVME_QUIRK_DELAY_AMOUNT 2300
185 * enum nvme_ctrl_state: Controller state
187 * @NVME_CTRL_NEW: New controller just allocated, initial state
188 * @NVME_CTRL_LIVE: Controller is connected and I/O capable
189 * @NVME_CTRL_RESETTING: Controller is resetting (or scheduled reset)
190 * @NVME_CTRL_CONNECTING: Controller is disconnected, now connecting the
192 * @NVME_CTRL_DELETING: Controller is deleting (or scheduled deletion)
193 * @NVME_CTRL_DELETING_NOIO: Controller is deleting and I/O is not
194 * disabled/failed immediately. This state comes
195 * after all async event processing took place and
196 * before ns removal and the controller deletion
198 * @NVME_CTRL_DEAD: Controller is non-present/unresponsive during
199 * shutdown or removal. In this case we forcibly
200 * kill all inflight I/O as they have no chance to
203 enum nvme_ctrl_state {
207 NVME_CTRL_CONNECTING,
209 NVME_CTRL_DELETING_NOIO,
213 struct nvme_fault_inject {
214 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
215 struct fault_attr attr;
216 struct dentry *parent;
217 bool dont_retry; /* DNR, do not retry */
218 u16 status; /* status code */
223 struct list_head entry;
224 struct nvme_effects_log log;
230 enum nvme_ctrl_state state;
233 struct mutex scan_lock;
234 const struct nvme_ctrl_ops *ops;
235 struct request_queue *admin_q;
236 struct request_queue *connect_q;
237 struct request_queue *fabrics_q;
241 struct blk_mq_tag_set *tagset;
242 struct blk_mq_tag_set *admin_tagset;
243 struct list_head namespaces;
244 struct rw_semaphore namespaces_rwsem;
245 struct device ctrl_device;
246 struct device *device; /* char device */
248 struct work_struct reset_work;
249 struct work_struct delete_work;
250 wait_queue_head_t state_wq;
252 struct nvme_subsystem *subsys;
253 struct list_head subsys_entry;
255 struct opal_dev *opal_dev;
267 u32 max_integrity_segments;
268 #ifdef CONFIG_BLK_DEV_ZONED
278 atomic_t abort_limit;
290 unsigned int shutdown_timeout;
293 unsigned long quirks;
294 struct nvme_id_power_state psd[32];
295 struct nvme_effects_log *effects;
296 struct list_head cels;
297 struct work_struct scan_work;
298 struct work_struct async_event_work;
299 struct delayed_work ka_work;
300 struct nvme_command ka_cmd;
301 struct work_struct fw_act_work;
302 unsigned long events;
305 #ifdef CONFIG_NVME_MULTIPATH
306 /* asymmetric namespace access: */
311 struct mutex ana_lock;
312 struct nvme_ana_rsp_hdr *ana_log_buf;
314 struct timer_list anatt_timer;
315 struct work_struct ana_work;
318 /* Power saving configuration */
319 u64 ps_max_latency_us;
334 struct nvmf_ctrl_options *opts;
336 struct page *discard_page;
337 unsigned long discard_page_busy;
339 struct nvme_fault_inject fault_inject;
347 struct nvme_subsystem {
351 * Because we unregister the device on the last put we need
352 * a separate refcount.
355 struct list_head entry;
357 struct list_head ctrls;
358 struct list_head nsheads;
359 char subnqn[NVMF_NQN_SIZE];
362 char firmware_rev[8];
365 u16 awupf; /* 0's based awupf value. */
367 #ifdef CONFIG_NVME_MULTIPATH
368 enum nvme_iopolicy iopolicy;
373 * Container structure for uniqueue namespace identifiers.
383 * Anchor structure for namespaces. There is one for each namespace in a
384 * NVMe subsystem that any of our controllers can see, and the namespace
385 * structure for each controller is chained of it. For private namespaces
386 * there is a 1:1 relation to our namespace structures, that is ->list
387 * only ever has a single entry for private namespaces.
389 struct nvme_ns_head {
390 struct list_head list;
391 struct srcu_struct srcu;
392 struct nvme_subsystem *subsys;
394 struct nvme_ns_ids ids;
395 struct list_head entry;
399 struct nvme_effects_log *effects;
400 #ifdef CONFIG_NVME_MULTIPATH
401 struct gendisk *disk;
402 struct bio_list requeue_list;
403 spinlock_t requeue_lock;
404 struct work_struct requeue_work;
407 #define NVME_NSHEAD_DISK_LIVE 0
408 struct nvme_ns __rcu *current_path[];
412 enum nvme_ns_features {
413 NVME_NS_EXT_LBAS = 1 << 0, /* support extended LBA format */
414 NVME_NS_METADATA_SUPPORTED = 1 << 1, /* support getting generated md */
418 struct list_head list;
420 struct nvme_ctrl *ctrl;
421 struct request_queue *queue;
422 struct gendisk *disk;
423 #ifdef CONFIG_NVME_MULTIPATH
424 enum nvme_ana_state ana_state;
427 struct list_head siblings;
428 struct nvm_dev *ndev;
430 struct nvme_ns_head *head;
437 #ifdef CONFIG_BLK_DEV_ZONED
440 unsigned long features;
442 #define NVME_NS_REMOVING 0
443 #define NVME_NS_DEAD 1
444 #define NVME_NS_ANA_PENDING 2
446 struct nvme_fault_inject fault_inject;
450 /* NVMe ns supports metadata actions by the controller (generate/strip) */
451 static inline bool nvme_ns_has_pi(struct nvme_ns *ns)
453 return ns->pi_type && ns->ms == sizeof(struct t10_pi_tuple);
456 struct nvme_ctrl_ops {
458 struct module *module;
460 #define NVME_F_FABRICS (1 << 0)
461 #define NVME_F_METADATA_SUPPORTED (1 << 1)
462 #define NVME_F_PCI_P2PDMA (1 << 2)
463 int (*reg_read32)(struct nvme_ctrl *ctrl, u32 off, u32 *val);
464 int (*reg_write32)(struct nvme_ctrl *ctrl, u32 off, u32 val);
465 int (*reg_read64)(struct nvme_ctrl *ctrl, u32 off, u64 *val);
466 void (*free_ctrl)(struct nvme_ctrl *ctrl);
467 void (*submit_async_event)(struct nvme_ctrl *ctrl);
468 void (*delete_ctrl)(struct nvme_ctrl *ctrl);
469 int (*get_address)(struct nvme_ctrl *ctrl, char *buf, int size);
472 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
473 void nvme_fault_inject_init(struct nvme_fault_inject *fault_inj,
474 const char *dev_name);
475 void nvme_fault_inject_fini(struct nvme_fault_inject *fault_inject);
476 void nvme_should_fail(struct request *req);
478 static inline void nvme_fault_inject_init(struct nvme_fault_inject *fault_inj,
479 const char *dev_name)
482 static inline void nvme_fault_inject_fini(struct nvme_fault_inject *fault_inj)
485 static inline void nvme_should_fail(struct request *req) {}
488 static inline int nvme_reset_subsystem(struct nvme_ctrl *ctrl)
490 if (!ctrl->subsystem)
492 return ctrl->ops->reg_write32(ctrl, NVME_REG_NSSR, 0x4E564D65);
496 * Convert a 512B sector number to a device logical block number.
498 static inline u64 nvme_sect_to_lba(struct nvme_ns *ns, sector_t sector)
500 return sector >> (ns->lba_shift - SECTOR_SHIFT);
504 * Convert a device logical block number to a 512B sector number.
506 static inline sector_t nvme_lba_to_sect(struct nvme_ns *ns, u64 lba)
508 return lba << (ns->lba_shift - SECTOR_SHIFT);
512 * Convert byte length to nvme's 0-based num dwords
514 static inline u32 nvme_bytes_to_numd(size_t len)
516 return (len >> 2) - 1;
519 static inline bool nvme_end_request(struct request *req, __le16 status,
520 union nvme_result result)
522 struct nvme_request *rq = nvme_req(req);
524 rq->status = le16_to_cpu(status) >> 1;
526 /* inject error when permitted by fault injection framework */
527 nvme_should_fail(req);
528 if (unlikely(blk_should_fake_timeout(req->q)))
530 return blk_mq_complete_request_remote(req);
533 static inline void nvme_get_ctrl(struct nvme_ctrl *ctrl)
535 get_device(ctrl->device);
538 static inline void nvme_put_ctrl(struct nvme_ctrl *ctrl)
540 put_device(ctrl->device);
543 static inline bool nvme_is_aen_req(u16 qid, __u16 command_id)
545 return !qid && command_id >= NVME_AQ_BLK_MQ_DEPTH;
548 void nvme_complete_rq(struct request *req);
549 bool nvme_cancel_request(struct request *req, void *data, bool reserved);
550 bool nvme_change_ctrl_state(struct nvme_ctrl *ctrl,
551 enum nvme_ctrl_state new_state);
552 bool nvme_wait_reset(struct nvme_ctrl *ctrl);
553 int nvme_disable_ctrl(struct nvme_ctrl *ctrl);
554 int nvme_enable_ctrl(struct nvme_ctrl *ctrl);
555 int nvme_shutdown_ctrl(struct nvme_ctrl *ctrl);
556 int nvme_init_ctrl(struct nvme_ctrl *ctrl, struct device *dev,
557 const struct nvme_ctrl_ops *ops, unsigned long quirks);
558 void nvme_uninit_ctrl(struct nvme_ctrl *ctrl);
559 void nvme_start_ctrl(struct nvme_ctrl *ctrl);
560 void nvme_stop_ctrl(struct nvme_ctrl *ctrl);
561 int nvme_init_identify(struct nvme_ctrl *ctrl);
563 void nvme_remove_namespaces(struct nvme_ctrl *ctrl);
565 int nvme_sec_submit(void *data, u16 spsp, u8 secp, void *buffer, size_t len,
568 void nvme_complete_async_event(struct nvme_ctrl *ctrl, __le16 status,
569 volatile union nvme_result *res);
571 void nvme_stop_queues(struct nvme_ctrl *ctrl);
572 void nvme_start_queues(struct nvme_ctrl *ctrl);
573 void nvme_kill_queues(struct nvme_ctrl *ctrl);
574 void nvme_sync_queues(struct nvme_ctrl *ctrl);
575 void nvme_unfreeze(struct nvme_ctrl *ctrl);
576 void nvme_wait_freeze(struct nvme_ctrl *ctrl);
577 void nvme_wait_freeze_timeout(struct nvme_ctrl *ctrl, long timeout);
578 void nvme_start_freeze(struct nvme_ctrl *ctrl);
580 #define NVME_QID_ANY -1
581 struct request *nvme_alloc_request(struct request_queue *q,
582 struct nvme_command *cmd, blk_mq_req_flags_t flags, int qid);
583 void nvme_cleanup_cmd(struct request *req);
584 blk_status_t nvme_setup_cmd(struct nvme_ns *ns, struct request *req,
585 struct nvme_command *cmd);
586 int nvme_submit_sync_cmd(struct request_queue *q, struct nvme_command *cmd,
587 void *buf, unsigned bufflen);
588 int __nvme_submit_sync_cmd(struct request_queue *q, struct nvme_command *cmd,
589 union nvme_result *result, void *buffer, unsigned bufflen,
590 unsigned timeout, int qid, int at_head,
591 blk_mq_req_flags_t flags, bool poll);
592 int nvme_set_features(struct nvme_ctrl *dev, unsigned int fid,
593 unsigned int dword11, void *buffer, size_t buflen,
595 int nvme_get_features(struct nvme_ctrl *dev, unsigned int fid,
596 unsigned int dword11, void *buffer, size_t buflen,
598 int nvme_set_queue_count(struct nvme_ctrl *ctrl, int *count);
599 void nvme_stop_keep_alive(struct nvme_ctrl *ctrl);
600 int nvme_reset_ctrl(struct nvme_ctrl *ctrl);
601 int nvme_reset_ctrl_sync(struct nvme_ctrl *ctrl);
602 int nvme_try_sched_reset(struct nvme_ctrl *ctrl);
603 int nvme_delete_ctrl(struct nvme_ctrl *ctrl);
605 int nvme_get_log(struct nvme_ctrl *ctrl, u32 nsid, u8 log_page, u8 lsp, u8 csi,
606 void *log, size_t size, u64 offset);
607 struct nvme_ns *nvme_get_ns_from_disk(struct gendisk *disk,
608 struct nvme_ns_head **head, int *srcu_idx);
609 void nvme_put_ns_from_disk(struct nvme_ns_head *head, int idx);
611 extern const struct attribute_group *nvme_ns_id_attr_groups[];
612 extern const struct block_device_operations nvme_ns_head_ops;
614 #ifdef CONFIG_NVME_MULTIPATH
615 static inline bool nvme_ctrl_use_ana(struct nvme_ctrl *ctrl)
617 return ctrl->ana_log_buf != NULL;
620 void nvme_mpath_unfreeze(struct nvme_subsystem *subsys);
621 void nvme_mpath_wait_freeze(struct nvme_subsystem *subsys);
622 void nvme_mpath_start_freeze(struct nvme_subsystem *subsys);
623 void nvme_set_disk_name(char *disk_name, struct nvme_ns *ns,
624 struct nvme_ctrl *ctrl, int *flags);
625 bool nvme_failover_req(struct request *req);
626 void nvme_kick_requeue_lists(struct nvme_ctrl *ctrl);
627 int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl,struct nvme_ns_head *head);
628 void nvme_mpath_add_disk(struct nvme_ns *ns, struct nvme_id_ns *id);
629 void nvme_mpath_remove_disk(struct nvme_ns_head *head);
630 int nvme_mpath_init(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id);
631 void nvme_mpath_uninit(struct nvme_ctrl *ctrl);
632 void nvme_mpath_stop(struct nvme_ctrl *ctrl);
633 bool nvme_mpath_clear_current_path(struct nvme_ns *ns);
634 void nvme_mpath_clear_ctrl_paths(struct nvme_ctrl *ctrl);
635 struct nvme_ns *nvme_find_path(struct nvme_ns_head *head);
636 blk_qc_t nvme_ns_head_submit_bio(struct bio *bio);
638 static inline void nvme_mpath_check_last_path(struct nvme_ns *ns)
640 struct nvme_ns_head *head = ns->head;
642 if (head->disk && list_empty(&head->list))
643 kblockd_schedule_work(&head->requeue_work);
646 static inline void nvme_trace_bio_complete(struct request *req,
649 struct nvme_ns *ns = req->q->queuedata;
651 if (req->cmd_flags & REQ_NVME_MPATH)
652 trace_block_bio_complete(ns->head->disk->queue, req->bio);
655 extern struct device_attribute dev_attr_ana_grpid;
656 extern struct device_attribute dev_attr_ana_state;
657 extern struct device_attribute subsys_attr_iopolicy;
660 static inline bool nvme_ctrl_use_ana(struct nvme_ctrl *ctrl)
665 * Without the multipath code enabled, multiple controller per subsystems are
666 * visible as devices and thus we cannot use the subsystem instance.
668 static inline void nvme_set_disk_name(char *disk_name, struct nvme_ns *ns,
669 struct nvme_ctrl *ctrl, int *flags)
671 sprintf(disk_name, "nvme%dn%d", ctrl->instance, ns->head->instance);
674 static inline bool nvme_failover_req(struct request *req)
678 static inline void nvme_kick_requeue_lists(struct nvme_ctrl *ctrl)
681 static inline int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl,
682 struct nvme_ns_head *head)
686 static inline void nvme_mpath_add_disk(struct nvme_ns *ns,
687 struct nvme_id_ns *id)
690 static inline void nvme_mpath_remove_disk(struct nvme_ns_head *head)
693 static inline bool nvme_mpath_clear_current_path(struct nvme_ns *ns)
697 static inline void nvme_mpath_clear_ctrl_paths(struct nvme_ctrl *ctrl)
700 static inline void nvme_mpath_check_last_path(struct nvme_ns *ns)
703 static inline void nvme_trace_bio_complete(struct request *req,
707 static inline int nvme_mpath_init(struct nvme_ctrl *ctrl,
708 struct nvme_id_ctrl *id)
710 if (ctrl->subsys->cmic & (1 << 3))
711 dev_warn(ctrl->device,
712 "Please enable CONFIG_NVME_MULTIPATH for full support of multi-port devices.\n");
715 static inline void nvme_mpath_uninit(struct nvme_ctrl *ctrl)
718 static inline void nvme_mpath_stop(struct nvme_ctrl *ctrl)
721 static inline void nvme_mpath_unfreeze(struct nvme_subsystem *subsys)
724 static inline void nvme_mpath_wait_freeze(struct nvme_subsystem *subsys)
727 static inline void nvme_mpath_start_freeze(struct nvme_subsystem *subsys)
730 #endif /* CONFIG_NVME_MULTIPATH */
732 #ifdef CONFIG_BLK_DEV_ZONED
733 int nvme_update_zone_info(struct gendisk *disk, struct nvme_ns *ns,
736 int nvme_report_zones(struct gendisk *disk, sector_t sector,
737 unsigned int nr_zones, report_zones_cb cb, void *data);
739 blk_status_t nvme_setup_zone_mgmt_send(struct nvme_ns *ns, struct request *req,
740 struct nvme_command *cmnd,
741 enum nvme_zone_mgmt_action action);
743 #define nvme_report_zones NULL
745 static inline blk_status_t nvme_setup_zone_mgmt_send(struct nvme_ns *ns,
746 struct request *req, struct nvme_command *cmnd,
747 enum nvme_zone_mgmt_action action)
749 return BLK_STS_NOTSUPP;
752 static inline int nvme_update_zone_info(struct gendisk *disk,
756 dev_warn(ns->ctrl->device,
757 "Please enable CONFIG_BLK_DEV_ZONED to support ZNS devices\n");
758 return -EPROTONOSUPPORT;
763 int nvme_nvm_register(struct nvme_ns *ns, char *disk_name, int node);
764 void nvme_nvm_unregister(struct nvme_ns *ns);
765 extern const struct attribute_group nvme_nvm_attr_group;
766 int nvme_nvm_ioctl(struct nvme_ns *ns, unsigned int cmd, unsigned long arg);
768 static inline int nvme_nvm_register(struct nvme_ns *ns, char *disk_name,
774 static inline void nvme_nvm_unregister(struct nvme_ns *ns) {};
775 static inline int nvme_nvm_ioctl(struct nvme_ns *ns, unsigned int cmd,
780 #endif /* CONFIG_NVM */
782 static inline struct nvme_ns *nvme_get_ns_from_dev(struct device *dev)
784 return dev_to_disk(dev)->private_data;
787 #ifdef CONFIG_NVME_HWMON
788 void nvme_hwmon_init(struct nvme_ctrl *ctrl);
790 static inline void nvme_hwmon_init(struct nvme_ctrl *ctrl) { }