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/sed-opal.h>
  15 #include <linux/fault-inject.h>
  16 #include <linux/rcupdate.h>
  17 #include <linux/wait.h>
  18 #include <linux/t10-pi.h>
  19
  20 #include <trace/events/block.h>
  21
  22 extern unsigned int nvme_io_timeout;
  23 #define NVME_IO_TIMEOUT (nvme_io_timeout * HZ)
  24
  25 extern unsigned int admin_timeout;
  26 #define NVME_ADMIN_TIMEOUT      (admin_timeout * HZ)
  27
  28 #define NVME_DEFAULT_KATO       5
  29
  30 #ifdef CONFIG_ARCH_NO_SG_CHAIN
  31 #define  NVME_INLINE_SG_CNT  0
  32 #define  NVME_INLINE_METADATA_SG_CNT  0
  33 #else
  34 #define  NVME_INLINE_SG_CNT  2
  35 #define  NVME_INLINE_METADATA_SG_CNT  1
  36 #endif
  37
  38 /*
  39  * Default to a 4K page size, with the intention to update this
  40  * path in the future to accommodate architectures with differing
  41  * kernel and IO page sizes.
  42  */
  43 #define NVME_CTRL_PAGE_SHIFT    12
  44 #define NVME_CTRL_PAGE_SIZE     (1 << NVME_CTRL_PAGE_SHIFT)
  45
  46 extern struct workqueue_struct *nvme_wq;
  47 extern struct workqueue_struct *nvme_reset_wq;
  48 extern struct workqueue_struct *nvme_delete_wq;
  49
  50 /*
  51  * List of workarounds for devices that required behavior not specified in
  52  * the standard.
  53  */
  54 enum nvme_quirks {
  55         /*
  56          * Prefers I/O aligned to a stripe size specified in a vendor
  57          * specific Identify field.
  58          */
  59         NVME_QUIRK_STRIPE_SIZE                  = (1 << 0),
  60
  61         /*
  62          * The controller doesn't handle Identify value others than 0 or 1
  63          * correctly.
  64          */
  65         NVME_QUIRK_IDENTIFY_CNS                 = (1 << 1),
  66
  67         /*
  68          * The controller deterministically returns O's on reads to
  69          * logical blocks that deallocate was called on.
  70          */
  71         NVME_QUIRK_DEALLOCATE_ZEROES            = (1 << 2),
  72
  73         /*
  74          * The controller needs a delay before starts checking the device
  75          * readiness, which is done by reading the NVME_CSTS_RDY bit.
  76          */
  77         NVME_QUIRK_DELAY_BEFORE_CHK_RDY         = (1 << 3),
  78
  79         /*
  80          * APST should not be used.
  81          */
  82         NVME_QUIRK_NO_APST                      = (1 << 4),
  83
  84         /*
  85          * The deepest sleep state should not be used.
  86          */
  87         NVME_QUIRK_NO_DEEPEST_PS                = (1 << 5),
  88
  89         /*
  90          * Set MEDIUM priority on SQ creation
  91          */
  92         NVME_QUIRK_MEDIUM_PRIO_SQ               = (1 << 7),
  93
  94         /*
  95          * Ignore device provided subnqn.
  96          */
  97         NVME_QUIRK_IGNORE_DEV_SUBNQN            = (1 << 8),
  98
  99         /*
 100          * Broken Write Zeroes.
 101          */
 102         NVME_QUIRK_DISABLE_WRITE_ZEROES         = (1 << 9),
 103
 104         /*
 105          * Force simple suspend/resume path.
 106          */
 107         NVME_QUIRK_SIMPLE_SUSPEND               = (1 << 10),
 108
 109         /*
 110          * Use only one interrupt vector for all queues
 111          */
 112         NVME_QUIRK_SINGLE_VECTOR                = (1 << 11),
 113
 114         /*
 115          * Use non-standard 128 bytes SQEs.
 116          */
 117         NVME_QUIRK_128_BYTES_SQES               = (1 << 12),
 118
 119         /*
 120          * Prevent tag overlap between queues
 121          */
 122         NVME_QUIRK_SHARED_TAGS                  = (1 << 13),
 123
 124         /*
 125          * Don't change the value of the temperature threshold feature
 126          */
 127         NVME_QUIRK_NO_TEMP_THRESH_CHANGE        = (1 << 14),
 128
 129         /*
 130          * The controller doesn't handle the Identify Namespace
 131          * Identification Descriptor list subcommand despite claiming
 132          * NVMe 1.3 compliance.
 133          */
 134         NVME_QUIRK_NO_NS_DESC_LIST              = (1 << 15),
 135
 136         /*
 137          * The controller does not properly handle DMA addresses over
 138          * 48 bits.
 139          */
 140         NVME_QUIRK_DMA_ADDRESS_BITS_48          = (1 << 16),
 141
 142         /*
 143          * The controller requires the command_id value be limited, so skip
 144          * encoding the generation sequence number.
 145          */
 146         NVME_QUIRK_SKIP_CID_GEN                 = (1 << 17),
 147
 148         /*
 149          * Reports garbage in the namespace identifiers (eui64, nguid, uuid).
 150          */
 151         NVME_QUIRK_BOGUS_NID                    = (1 << 18),
 152 };
 153
 154 /*
 155  * Common request structure for NVMe passthrough.  All drivers must have
 156  * this structure as the first member of their request-private data.
 157  */
 158 struct nvme_request {
 159         struct nvme_command     *cmd;
 160         union nvme_result       result;
 161         u8                      genctr;
 162         u8                      retries;
 163         u8                      flags;
 164         u16                     status;
 165         struct nvme_ctrl        *ctrl;
 166 };
 167
 168 /*
 169  * Mark a bio as coming in through the mpath node.
 170  */
 171 #define REQ_NVME_MPATH          REQ_DRV
 172
 173 enum {
 174         NVME_REQ_CANCELLED              = (1 << 0),
 175         NVME_REQ_USERCMD                = (1 << 1),
 176 };
 177
 178 static inline struct nvme_request *nvme_req(struct request *req)
 179 {
 180         return blk_mq_rq_to_pdu(req);
 181 }
 182
 183 static inline u16 nvme_req_qid(struct request *req)
 184 {
 185         if (!req->q->queuedata)
 186                 return 0;
 187
 188         return req->mq_hctx->queue_num + 1;
 189 }
 190
 191 /* The below value is the specific amount of delay needed before checking
 192  * readiness in case of the PCI_DEVICE(0x1c58, 0x0003), which needs the
 193  * NVME_QUIRK_DELAY_BEFORE_CHK_RDY quirk enabled. The value (in ms) was
 194  * found empirically.
 195  */
 196 #define NVME_QUIRK_DELAY_AMOUNT         2300
 197
 198 /*
 199  * enum nvme_ctrl_state: Controller state
 200  *
 201  * @NVME_CTRL_NEW:              New controller just allocated, initial state
 202  * @NVME_CTRL_LIVE:             Controller is connected and I/O capable
 203  * @NVME_CTRL_RESETTING:        Controller is resetting (or scheduled reset)
 204  * @NVME_CTRL_CONNECTING:       Controller is disconnected, now connecting the
 205  *                              transport
 206  * @NVME_CTRL_DELETING:         Controller is deleting (or scheduled deletion)
 207  * @NVME_CTRL_DELETING_NOIO:    Controller is deleting and I/O is not
 208  *                              disabled/failed immediately. This state comes
 209  *                              after all async event processing took place and
 210  *                              before ns removal and the controller deletion
 211  *                              progress
 212  * @NVME_CTRL_DEAD:             Controller is non-present/unresponsive during
 213  *                              shutdown or removal. In this case we forcibly
 214  *                              kill all inflight I/O as they have no chance to
 215  *                              complete
 216  */
 217 enum nvme_ctrl_state {
 218         NVME_CTRL_NEW,
 219         NVME_CTRL_LIVE,
 220         NVME_CTRL_RESETTING,
 221         NVME_CTRL_CONNECTING,
 222         NVME_CTRL_DELETING,
 223         NVME_CTRL_DELETING_NOIO,
 224         NVME_CTRL_DEAD,
 225 };
 226
 227 struct nvme_fault_inject {
 228 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
 229         struct fault_attr attr;
 230         struct dentry *parent;
 231         bool dont_retry;        /* DNR, do not retry */
 232         u16 status;             /* status code */
 233 #endif
 234 };
 235
 236 enum nvme_ctrl_flags {
 237         NVME_CTRL_FAILFAST_EXPIRED      = 0,
 238         NVME_CTRL_ADMIN_Q_STOPPED       = 1,
 239         NVME_CTRL_STARTED_ONCE          = 2,
 240         NVME_CTRL_STOPPED               = 3,
 241 };
 242
 243 struct nvme_ctrl {
 244         bool comp_seen;
 245         enum nvme_ctrl_state state;
 246         bool identified;
 247         spinlock_t lock;
 248         struct mutex scan_lock;
 249         const struct nvme_ctrl_ops *ops;
 250         struct request_queue *admin_q;
 251         struct request_queue *connect_q;
 252         struct request_queue *fabrics_q;
 253         struct device *dev;
 254         int instance;
 255         int numa_node;
 256         struct blk_mq_tag_set *tagset;
 257         struct blk_mq_tag_set *admin_tagset;
 258         struct list_head namespaces;
 259         struct rw_semaphore namespaces_rwsem;
 260         struct device ctrl_device;
 261         struct device *device;  /* char device */
 262 #ifdef CONFIG_NVME_HWMON
 263         struct device *hwmon_device;
 264 #endif
 265         struct cdev cdev;
 266         struct work_struct reset_work;
 267         struct work_struct delete_work;
 268         wait_queue_head_t state_wq;
 269
 270         struct nvme_subsystem *subsys;
 271         struct list_head subsys_entry;
 272
 273         struct opal_dev *opal_dev;
 274
 275         char name[12];
 276         u16 cntlid;
 277
 278         u32 ctrl_config;
 279         u16 mtfa;
 280         u32 queue_count;
 281
 282         u64 cap;
 283         u32 max_hw_sectors;
 284         u32 max_segments;
 285         u32 max_integrity_segments;
 286         u32 max_discard_sectors;
 287         u32 max_discard_segments;
 288         u32 max_zeroes_sectors;
 289 #ifdef CONFIG_BLK_DEV_ZONED
 290         u32 max_zone_append;
 291 #endif
 292         u16 crdt[3];
 293         u16 oncs;
 294         u32 dmrsl;
 295         u16 oacs;
 296         u16 sqsize;
 297         u32 max_namespaces;
 298         atomic_t abort_limit;
 299         u8 vwc;
 300         u32 vs;
 301         u32 sgls;
 302         u16 kas;
 303         u8 npss;
 304         u8 apsta;
 305         u16 wctemp;
 306         u16 cctemp;
 307         u32 oaes;
 308         u32 aen_result;
 309         u32 ctratt;
 310         unsigned int shutdown_timeout;
 311         unsigned int kato;
 312         bool subsystem;
 313         unsigned long quirks;
 314         struct nvme_id_power_state psd[32];
 315         struct nvme_effects_log *effects;
 316         struct xarray cels;
 317         struct work_struct scan_work;
 318         struct work_struct async_event_work;
 319         struct delayed_work ka_work;
 320         struct delayed_work failfast_work;
 321         struct nvme_command ka_cmd;
 322         struct work_struct fw_act_work;
 323         unsigned long events;
 324
 325 #ifdef CONFIG_NVME_MULTIPATH
 326         /* asymmetric namespace access: */
 327         u8 anacap;
 328         u8 anatt;
 329         u32 anagrpmax;
 330         u32 nanagrpid;
 331         struct mutex ana_lock;
 332         struct nvme_ana_rsp_hdr *ana_log_buf;
 333         size_t ana_log_size;
 334         struct timer_list anatt_timer;
 335         struct work_struct ana_work;
 336 #endif
 337
 338 #ifdef CONFIG_NVME_AUTH
 339         struct work_struct dhchap_auth_work;
 340         struct mutex dhchap_auth_mutex;
 341         struct nvme_dhchap_queue_context *dhchap_ctxs;
 342         struct nvme_dhchap_key *host_key;
 343         struct nvme_dhchap_key *ctrl_key;
 344         u16 transaction;
 345 #endif
 346
 347         /* Power saving configuration */
 348         u64 ps_max_latency_us;
 349         bool apst_enabled;
 350
 351         /* PCIe only: */
 352         u32 hmpre;
 353         u32 hmmin;
 354         u32 hmminds;
 355         u16 hmmaxd;
 356
 357         /* Fabrics only */
 358         u32 ioccsz;
 359         u32 iorcsz;
 360         u16 icdoff;
 361         u16 maxcmd;
 362         int nr_reconnects;
 363         unsigned long flags;
 364         struct nvmf_ctrl_options *opts;
 365
 366         struct page *discard_page;
 367         unsigned long discard_page_busy;
 368
 369         struct nvme_fault_inject fault_inject;
 370
 371         enum nvme_ctrl_type cntrltype;
 372         enum nvme_dctype dctype;
 373 };
 374
 375 enum nvme_iopolicy {
 376         NVME_IOPOLICY_NUMA,
 377         NVME_IOPOLICY_RR,
 378 };
 379
 380 struct nvme_subsystem {
 381         int                     instance;
 382         struct device           dev;
 383         /*
 384          * Because we unregister the device on the last put we need
 385          * a separate refcount.
 386          */
 387         struct kref             ref;
 388         struct list_head        entry;
 389         struct mutex            lock;
 390         struct list_head        ctrls;
 391         struct list_head        nsheads;
 392         char                    subnqn[NVMF_NQN_SIZE];
 393         char                    serial[20];
 394         char                    model[40];
 395         char                    firmware_rev[8];
 396         u8                      cmic;
 397         enum nvme_subsys_type   subtype;
 398         u16                     vendor_id;
 399         u16                     awupf;  /* 0's based awupf value. */
 400         struct ida              ns_ida;
 401 #ifdef CONFIG_NVME_MULTIPATH
 402         enum nvme_iopolicy      iopolicy;
 403 #endif
 404 };
 405
 406 /*
 407  * Container structure for uniqueue namespace identifiers.
 408  */
 409 struct nvme_ns_ids {
 410         u8      eui64[8];
 411         u8      nguid[16];
 412         uuid_t  uuid;
 413         u8      csi;
 414 };
 415
 416 /*
 417  * Anchor structure for namespaces.  There is one for each namespace in a
 418  * NVMe subsystem that any of our controllers can see, and the namespace
 419  * structure for each controller is chained of it.  For private namespaces
 420  * there is a 1:1 relation to our namespace structures, that is ->list
 421  * only ever has a single entry for private namespaces.
 422  */
 423 struct nvme_ns_head {
 424         struct list_head        list;
 425         struct srcu_struct      srcu;
 426         struct nvme_subsystem   *subsys;
 427         unsigned                ns_id;
 428         struct nvme_ns_ids      ids;
 429         struct list_head        entry;
 430         struct kref             ref;
 431         bool                    shared;
 432         int                     instance;
 433         struct nvme_effects_log *effects;
 434
 435         struct cdev             cdev;
 436         struct device           cdev_device;
 437
 438         struct gendisk          *disk;
 439 #ifdef CONFIG_NVME_MULTIPATH
 440         struct bio_list         requeue_list;
 441         spinlock_t              requeue_lock;
 442         struct work_struct      requeue_work;
 443         struct mutex            lock;
 444         unsigned long           flags;
 445 #define NVME_NSHEAD_DISK_LIVE   0
 446         struct nvme_ns __rcu    *current_path[];
 447 #endif
 448 };
 449
 450 static inline bool nvme_ns_head_multipath(struct nvme_ns_head *head)
 451 {
 452         return IS_ENABLED(CONFIG_NVME_MULTIPATH) && head->disk;
 453 }
 454
 455 enum nvme_ns_features {
 456         NVME_NS_EXT_LBAS = 1 << 0, /* support extended LBA format */
 457         NVME_NS_METADATA_SUPPORTED = 1 << 1, /* support getting generated md */
 458         NVME_NS_DEAC,           /* DEAC bit in Write Zeores supported */
 459 };
 460
 461 struct nvme_ns {
 462         struct list_head list;
 463
 464         struct nvme_ctrl *ctrl;
 465         struct request_queue *queue;
 466         struct gendisk *disk;
 467 #ifdef CONFIG_NVME_MULTIPATH
 468         enum nvme_ana_state ana_state;
 469         u32 ana_grpid;
 470 #endif
 471         struct list_head siblings;
 472         struct kref kref;
 473         struct nvme_ns_head *head;
 474
 475         int lba_shift;
 476         u16 ms;
 477         u16 pi_size;
 478         u16 sgs;
 479         u32 sws;
 480         u8 pi_type;
 481         u8 guard_type;
 482 #ifdef CONFIG_BLK_DEV_ZONED
 483         u64 zsze;
 484 #endif
 485         unsigned long features;
 486         unsigned long flags;
 487 #define NVME_NS_REMOVING        0
 488 #define NVME_NS_ANA_PENDING     2
 489 #define NVME_NS_FORCE_RO        3
 490 #define NVME_NS_READY           4
 491
 492         struct cdev             cdev;
 493         struct device           cdev_device;
 494
 495         struct nvme_fault_inject fault_inject;
 496
 497 };
 498
 499 /* NVMe ns supports metadata actions by the controller (generate/strip) */
 500 static inline bool nvme_ns_has_pi(struct nvme_ns *ns)
 501 {
 502         return ns->pi_type && ns->ms == ns->pi_size;
 503 }
 504
 505 struct nvme_ctrl_ops {
 506         const char *name;
 507         struct module *module;
 508         unsigned int flags;
 509 #define NVME_F_FABRICS                  (1 << 0)
 510 #define NVME_F_METADATA_SUPPORTED       (1 << 1)
 511         const struct attribute_group **dev_attr_groups;
 512         int (*reg_read32)(struct nvme_ctrl *ctrl, u32 off, u32 *val);
 513         int (*reg_write32)(struct nvme_ctrl *ctrl, u32 off, u32 val);
 514         int (*reg_read64)(struct nvme_ctrl *ctrl, u32 off, u64 *val);
 515         void (*free_ctrl)(struct nvme_ctrl *ctrl);
 516         void (*submit_async_event)(struct nvme_ctrl *ctrl);
 517         void (*delete_ctrl)(struct nvme_ctrl *ctrl);
 518         void (*stop_ctrl)(struct nvme_ctrl *ctrl);
 519         int (*get_address)(struct nvme_ctrl *ctrl, char *buf, int size);
 520         void (*print_device_info)(struct nvme_ctrl *ctrl);
 521         bool (*supports_pci_p2pdma)(struct nvme_ctrl *ctrl);
 522 };
 523
 524 /*
 525  * nvme command_id is constructed as such:
 526  * | xxxx | xxxxxxxxxxxx |
 527  *   gen    request tag
 528  */
 529 #define nvme_genctr_mask(gen)                   (gen & 0xf)
 530 #define nvme_cid_install_genctr(gen)            (nvme_genctr_mask(gen) << 12)
 531 #define nvme_genctr_from_cid(cid)               ((cid & 0xf000) >> 12)
 532 #define nvme_tag_from_cid(cid)                  (cid & 0xfff)
 533
 534 static inline u16 nvme_cid(struct request *rq)
 535 {
 536         return nvme_cid_install_genctr(nvme_req(rq)->genctr) | rq->tag;
 537 }
 538
 539 static inline struct request *nvme_find_rq(struct blk_mq_tags *tags,
 540                 u16 command_id)
 541 {
 542         u8 genctr = nvme_genctr_from_cid(command_id);
 543         u16 tag = nvme_tag_from_cid(command_id);
 544         struct request *rq;
 545
 546         rq = blk_mq_tag_to_rq(tags, tag);
 547         if (unlikely(!rq)) {
 548                 pr_err("could not locate request for tag %#x\n",
 549                         tag);
 550                 return NULL;
 551         }
 552         if (unlikely(nvme_genctr_mask(nvme_req(rq)->genctr) != genctr)) {
 553                 dev_err(nvme_req(rq)->ctrl->device,
 554                         "request %#x genctr mismatch (got %#x expected %#x)\n",
 555                         tag, genctr, nvme_genctr_mask(nvme_req(rq)->genctr));
 556                 return NULL;
 557         }
 558         return rq;
 559 }
 560
 561 static inline struct request *nvme_cid_to_rq(struct blk_mq_tags *tags,
 562                 u16 command_id)
 563 {
 564         return blk_mq_tag_to_rq(tags, nvme_tag_from_cid(command_id));
 565 }
 566
 567 /*
 568  * Return the length of the string without the space padding
 569  */
 570 static inline int nvme_strlen(char *s, int len)
 571 {
 572         while (s[len - 1] == ' ')
 573                 len--;
 574         return len;
 575 }
 576
 577 static inline void nvme_print_device_info(struct nvme_ctrl *ctrl)
 578 {
 579         struct nvme_subsystem *subsys = ctrl->subsys;
 580
 581         if (ctrl->ops->print_device_info) {
 582                 ctrl->ops->print_device_info(ctrl);
 583                 return;
 584         }
 585
 586         dev_err(ctrl->device,
 587                 "VID:%04x model:%.*s firmware:%.*s\n", subsys->vendor_id,
 588                 nvme_strlen(subsys->model, sizeof(subsys->model)),
 589                 subsys->model, nvme_strlen(subsys->firmware_rev,
 590                                            sizeof(subsys->firmware_rev)),
 591                 subsys->firmware_rev);
 592 }
 593
 594 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
 595 void nvme_fault_inject_init(struct nvme_fault_inject *fault_inj,
 596                             const char *dev_name);
 597 void nvme_fault_inject_fini(struct nvme_fault_inject *fault_inject);
 598 void nvme_should_fail(struct request *req);
 599 #else
 600 static inline void nvme_fault_inject_init(struct nvme_fault_inject *fault_inj,
 601                                           const char *dev_name)
 602 {
 603 }
 604 static inline void nvme_fault_inject_fini(struct nvme_fault_inject *fault_inj)
 605 {
 606 }
 607 static inline void nvme_should_fail(struct request *req) {}
 608 #endif
 609
 610 bool nvme_wait_reset(struct nvme_ctrl *ctrl);
 611 int nvme_try_sched_reset(struct nvme_ctrl *ctrl);
 612
 613 static inline int nvme_reset_subsystem(struct nvme_ctrl *ctrl)
 614 {
 615         int ret;
 616
 617         if (!ctrl->subsystem)
 618                 return -ENOTTY;
 619         if (!nvme_wait_reset(ctrl))
 620                 return -EBUSY;
 621
 622         ret = ctrl->ops->reg_write32(ctrl, NVME_REG_NSSR, 0x4E564D65);
 623         if (ret)
 624                 return ret;
 625
 626         return nvme_try_sched_reset(ctrl);
 627 }
 628
 629 /*
 630  * Convert a 512B sector number to a device logical block number.
 631  */
 632 static inline u64 nvme_sect_to_lba(struct nvme_ns *ns, sector_t sector)
 633 {
 634         return sector >> (ns->lba_shift - SECTOR_SHIFT);
 635 }
 636
 637 /*
 638  * Convert a device logical block number to a 512B sector number.
 639  */
 640 static inline sector_t nvme_lba_to_sect(struct nvme_ns *ns, u64 lba)
 641 {
 642         return lba << (ns->lba_shift - SECTOR_SHIFT);
 643 }
 644
 645 /*
 646  * Convert byte length to nvme's 0-based num dwords
 647  */
 648 static inline u32 nvme_bytes_to_numd(size_t len)
 649 {
 650         return (len >> 2) - 1;
 651 }
 652
 653 static inline bool nvme_is_ana_error(u16 status)
 654 {
 655         switch (status & 0x7ff) {
 656         case NVME_SC_ANA_TRANSITION:
 657         case NVME_SC_ANA_INACCESSIBLE:
 658         case NVME_SC_ANA_PERSISTENT_LOSS:
 659                 return true;
 660         default:
 661                 return false;
 662         }
 663 }
 664
 665 static inline bool nvme_is_path_error(u16 status)
 666 {
 667         /* check for a status code type of 'path related status' */
 668         return (status & 0x700) == 0x300;
 669 }
 670
 671 /*
 672  * Fill in the status and result information from the CQE, and then figure out
 673  * if blk-mq will need to use IPI magic to complete the request, and if yes do
 674  * so.  If not let the caller complete the request without an indirect function
 675  * call.
 676  */
 677 static inline bool nvme_try_complete_req(struct request *req, __le16 status,
 678                 union nvme_result result)
 679 {
 680         struct nvme_request *rq = nvme_req(req);
 681         struct nvme_ctrl *ctrl = rq->ctrl;
 682
 683         if (!(ctrl->quirks & NVME_QUIRK_SKIP_CID_GEN))
 684                 rq->genctr++;
 685
 686         rq->status = le16_to_cpu(status) >> 1;
 687         rq->result = result;
 688         /* inject error when permitted by fault injection framework */
 689         nvme_should_fail(req);
 690         if (unlikely(blk_should_fake_timeout(req->q)))
 691                 return true;
 692         return blk_mq_complete_request_remote(req);
 693 }
 694
 695 static inline void nvme_get_ctrl(struct nvme_ctrl *ctrl)
 696 {
 697         get_device(ctrl->device);
 698 }
 699
 700 static inline void nvme_put_ctrl(struct nvme_ctrl *ctrl)
 701 {
 702         put_device(ctrl->device);
 703 }
 704
 705 static inline bool nvme_is_aen_req(u16 qid, __u16 command_id)
 706 {
 707         return !qid &&
 708                 nvme_tag_from_cid(command_id) >= NVME_AQ_BLK_MQ_DEPTH;
 709 }
 710
 711 void nvme_complete_rq(struct request *req);
 712 void nvme_complete_batch_req(struct request *req);
 713
 714 static __always_inline void nvme_complete_batch(struct io_comp_batch *iob,
 715                                                 void (*fn)(struct request *rq))
 716 {
 717         struct request *req;
 718
 719         rq_list_for_each(&iob->req_list, req) {
 720                 fn(req);
 721                 nvme_complete_batch_req(req);
 722         }
 723         blk_mq_end_request_batch(iob);
 724 }
 725
 726 blk_status_t nvme_host_path_error(struct request *req);
 727 bool nvme_cancel_request(struct request *req, void *data);
 728 void nvme_cancel_tagset(struct nvme_ctrl *ctrl);
 729 void nvme_cancel_admin_tagset(struct nvme_ctrl *ctrl);
 730 bool nvme_change_ctrl_state(struct nvme_ctrl *ctrl,
 731                 enum nvme_ctrl_state new_state);
 732 int nvme_disable_ctrl(struct nvme_ctrl *ctrl);
 733 int nvme_enable_ctrl(struct nvme_ctrl *ctrl);
 734 int nvme_shutdown_ctrl(struct nvme_ctrl *ctrl);
 735 int nvme_init_ctrl(struct nvme_ctrl *ctrl, struct device *dev,
 736                 const struct nvme_ctrl_ops *ops, unsigned long quirks);
 737 void nvme_uninit_ctrl(struct nvme_ctrl *ctrl);
 738 void nvme_start_ctrl(struct nvme_ctrl *ctrl);
 739 void nvme_stop_ctrl(struct nvme_ctrl *ctrl);
 740 int nvme_init_ctrl_finish(struct nvme_ctrl *ctrl, bool was_suspended);
 741 int nvme_alloc_admin_tag_set(struct nvme_ctrl *ctrl, struct blk_mq_tag_set *set,
 742                 const struct blk_mq_ops *ops, unsigned int flags,
 743                 unsigned int cmd_size);
 744 void nvme_remove_admin_tag_set(struct nvme_ctrl *ctrl);
 745 int nvme_alloc_io_tag_set(struct nvme_ctrl *ctrl, struct blk_mq_tag_set *set,
 746                 const struct blk_mq_ops *ops, unsigned int flags,
 747                 unsigned int cmd_size);
 748 void nvme_remove_io_tag_set(struct nvme_ctrl *ctrl);
 749
 750 void nvme_remove_namespaces(struct nvme_ctrl *ctrl);
 751
 752 void nvme_complete_async_event(struct nvme_ctrl *ctrl, __le16 status,
 753                 volatile union nvme_result *res);
 754
 755 void nvme_stop_queues(struct nvme_ctrl *ctrl);
 756 void nvme_start_queues(struct nvme_ctrl *ctrl);
 757 void nvme_stop_admin_queue(struct nvme_ctrl *ctrl);
 758 void nvme_start_admin_queue(struct nvme_ctrl *ctrl);
 759 void nvme_mark_namespaces_dead(struct nvme_ctrl *ctrl);
 760 void nvme_sync_queues(struct nvme_ctrl *ctrl);
 761 void nvme_sync_io_queues(struct nvme_ctrl *ctrl);
 762 void nvme_unfreeze(struct nvme_ctrl *ctrl);
 763 void nvme_wait_freeze(struct nvme_ctrl *ctrl);
 764 int nvme_wait_freeze_timeout(struct nvme_ctrl *ctrl, long timeout);
 765 void nvme_start_freeze(struct nvme_ctrl *ctrl);
 766
 767 static inline enum req_op nvme_req_op(struct nvme_command *cmd)
 768 {
 769         return nvme_is_write(cmd) ? REQ_OP_DRV_OUT : REQ_OP_DRV_IN;
 770 }
 771
 772 #define NVME_QID_ANY -1
 773 void nvme_init_request(struct request *req, struct nvme_command *cmd);
 774 void nvme_cleanup_cmd(struct request *req);
 775 blk_status_t nvme_setup_cmd(struct nvme_ns *ns, struct request *req);
 776 blk_status_t nvme_fail_nonready_command(struct nvme_ctrl *ctrl,
 777                 struct request *req);
 778 bool __nvme_check_ready(struct nvme_ctrl *ctrl, struct request *rq,
 779                 bool queue_live);
 780
 781 static inline bool nvme_check_ready(struct nvme_ctrl *ctrl, struct request *rq,
 782                 bool queue_live)
 783 {
 784         if (likely(ctrl->state == NVME_CTRL_LIVE))
 785                 return true;
 786         if (ctrl->ops->flags & NVME_F_FABRICS &&
 787             ctrl->state == NVME_CTRL_DELETING)
 788                 return queue_live;
 789         return __nvme_check_ready(ctrl, rq, queue_live);
 790 }
 791
 792 /*
 793  * NSID shall be unique for all shared namespaces, or if at least one of the
 794  * following conditions is met:
 795  *   1. Namespace Management is supported by the controller
 796  *   2. ANA is supported by the controller
 797  *   3. NVM Set are supported by the controller
 798  *
 799  * In other case, private namespace are not required to report a unique NSID.
 800  */
 801 static inline bool nvme_is_unique_nsid(struct nvme_ctrl *ctrl,
 802                 struct nvme_ns_head *head)
 803 {
 804         return head->shared ||
 805                 (ctrl->oacs & NVME_CTRL_OACS_NS_MNGT_SUPP) ||
 806                 (ctrl->subsys->cmic & NVME_CTRL_CMIC_ANA) ||
 807                 (ctrl->ctratt & NVME_CTRL_CTRATT_NVM_SETS);
 808 }
 809
 810 int nvme_submit_sync_cmd(struct request_queue *q, struct nvme_command *cmd,
 811                 void *buf, unsigned bufflen);
 812 int __nvme_submit_sync_cmd(struct request_queue *q, struct nvme_command *cmd,
 813                 union nvme_result *result, void *buffer, unsigned bufflen,
 814                 int qid, int at_head,
 815                 blk_mq_req_flags_t flags);
 816 int nvme_set_features(struct nvme_ctrl *dev, unsigned int fid,
 817                       unsigned int dword11, void *buffer, size_t buflen,
 818                       u32 *result);
 819 int nvme_get_features(struct nvme_ctrl *dev, unsigned int fid,
 820                       unsigned int dword11, void *buffer, size_t buflen,
 821                       u32 *result);
 822 int nvme_set_queue_count(struct nvme_ctrl *ctrl, int *count);
 823 void nvme_stop_keep_alive(struct nvme_ctrl *ctrl);
 824 int nvme_reset_ctrl(struct nvme_ctrl *ctrl);
 825 int nvme_reset_ctrl_sync(struct nvme_ctrl *ctrl);
 826 int nvme_delete_ctrl(struct nvme_ctrl *ctrl);
 827 void nvme_queue_scan(struct nvme_ctrl *ctrl);
 828 int nvme_get_log(struct nvme_ctrl *ctrl, u32 nsid, u8 log_page, u8 lsp, u8 csi,
 829                 void *log, size_t size, u64 offset);
 830 bool nvme_tryget_ns_head(struct nvme_ns_head *head);
 831 void nvme_put_ns_head(struct nvme_ns_head *head);
 832 int nvme_cdev_add(struct cdev *cdev, struct device *cdev_device,
 833                 const struct file_operations *fops, struct module *owner);
 834 void nvme_cdev_del(struct cdev *cdev, struct device *cdev_device);
 835 int nvme_ioctl(struct block_device *bdev, fmode_t mode,
 836                 unsigned int cmd, unsigned long arg);
 837 long nvme_ns_chr_ioctl(struct file *file, unsigned int cmd, unsigned long arg);
 838 int nvme_ns_head_ioctl(struct block_device *bdev, fmode_t mode,
 839                 unsigned int cmd, unsigned long arg);
 840 long nvme_ns_head_chr_ioctl(struct file *file, unsigned int cmd,
 841                 unsigned long arg);
 842 long nvme_dev_ioctl(struct file *file, unsigned int cmd,
 843                 unsigned long arg);
 844 int nvme_ns_chr_uring_cmd_iopoll(struct io_uring_cmd *ioucmd,
 845                 struct io_comp_batch *iob, unsigned int poll_flags);
 846 int nvme_ns_head_chr_uring_cmd_iopoll(struct io_uring_cmd *ioucmd,
 847                 struct io_comp_batch *iob, unsigned int poll_flags);
 848 int nvme_ns_chr_uring_cmd(struct io_uring_cmd *ioucmd,
 849                 unsigned int issue_flags);
 850 int nvme_ns_head_chr_uring_cmd(struct io_uring_cmd *ioucmd,
 851                 unsigned int issue_flags);
 852 int nvme_getgeo(struct block_device *bdev, struct hd_geometry *geo);
 853 int nvme_dev_uring_cmd(struct io_uring_cmd *ioucmd, unsigned int issue_flags);
 854
 855 extern const struct attribute_group *nvme_ns_id_attr_groups[];
 856 extern const struct pr_ops nvme_pr_ops;
 857 extern const struct block_device_operations nvme_ns_head_ops;
 858 extern const struct attribute_group nvme_dev_attrs_group;
 859
 860 struct nvme_ns *nvme_find_path(struct nvme_ns_head *head);
 861 #ifdef CONFIG_NVME_MULTIPATH
 862 static inline bool nvme_ctrl_use_ana(struct nvme_ctrl *ctrl)
 863 {
 864         return ctrl->ana_log_buf != NULL;
 865 }
 866
 867 void nvme_mpath_unfreeze(struct nvme_subsystem *subsys);
 868 void nvme_mpath_wait_freeze(struct nvme_subsystem *subsys);
 869 void nvme_mpath_start_freeze(struct nvme_subsystem *subsys);
 870 void nvme_mpath_default_iopolicy(struct nvme_subsystem *subsys);
 871 void nvme_failover_req(struct request *req);
 872 void nvme_kick_requeue_lists(struct nvme_ctrl *ctrl);
 873 int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl,struct nvme_ns_head *head);
 874 void nvme_mpath_add_disk(struct nvme_ns *ns, __le32 anagrpid);
 875 void nvme_mpath_remove_disk(struct nvme_ns_head *head);
 876 int nvme_mpath_init_identify(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id);
 877 void nvme_mpath_init_ctrl(struct nvme_ctrl *ctrl);
 878 void nvme_mpath_update(struct nvme_ctrl *ctrl);
 879 void nvme_mpath_uninit(struct nvme_ctrl *ctrl);
 880 void nvme_mpath_stop(struct nvme_ctrl *ctrl);
 881 bool nvme_mpath_clear_current_path(struct nvme_ns *ns);
 882 void nvme_mpath_revalidate_paths(struct nvme_ns *ns);
 883 void nvme_mpath_clear_ctrl_paths(struct nvme_ctrl *ctrl);
 884 void nvme_mpath_shutdown_disk(struct nvme_ns_head *head);
 885
 886 static inline void nvme_trace_bio_complete(struct request *req)
 887 {
 888         struct nvme_ns *ns = req->q->queuedata;
 889
 890         if (req->cmd_flags & REQ_NVME_MPATH)
 891                 trace_block_bio_complete(ns->head->disk->queue, req->bio);
 892 }
 893
 894 extern bool multipath;
 895 extern struct device_attribute dev_attr_ana_grpid;
 896 extern struct device_attribute dev_attr_ana_state;
 897 extern struct device_attribute subsys_attr_iopolicy;
 898
 899 #else
 900 #define multipath false
 901 static inline bool nvme_ctrl_use_ana(struct nvme_ctrl *ctrl)
 902 {
 903         return false;
 904 }
 905 static inline void nvme_failover_req(struct request *req)
 906 {
 907 }
 908 static inline void nvme_kick_requeue_lists(struct nvme_ctrl *ctrl)
 909 {
 910 }
 911 static inline int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl,
 912                 struct nvme_ns_head *head)
 913 {
 914         return 0;
 915 }
 916 static inline void nvme_mpath_add_disk(struct nvme_ns *ns, __le32 anagrpid)
 917 {
 918 }
 919 static inline void nvme_mpath_remove_disk(struct nvme_ns_head *head)
 920 {
 921 }
 922 static inline bool nvme_mpath_clear_current_path(struct nvme_ns *ns)
 923 {
 924         return false;
 925 }
 926 static inline void nvme_mpath_revalidate_paths(struct nvme_ns *ns)
 927 {
 928 }
 929 static inline void nvme_mpath_clear_ctrl_paths(struct nvme_ctrl *ctrl)
 930 {
 931 }
 932 static inline void nvme_mpath_shutdown_disk(struct nvme_ns_head *head)
 933 {
 934 }
 935 static inline void nvme_trace_bio_complete(struct request *req)
 936 {
 937 }
 938 static inline void nvme_mpath_init_ctrl(struct nvme_ctrl *ctrl)
 939 {
 940 }
 941 static inline int nvme_mpath_init_identify(struct nvme_ctrl *ctrl,
 942                 struct nvme_id_ctrl *id)
 943 {
 944         if (ctrl->subsys->cmic & NVME_CTRL_CMIC_ANA)
 945                 dev_warn(ctrl->device,
 946 "Please enable CONFIG_NVME_MULTIPATH for full support of multi-port devices.\n");
 947         return 0;
 948 }
 949 static inline void nvme_mpath_update(struct nvme_ctrl *ctrl)
 950 {
 951 }
 952 static inline void nvme_mpath_uninit(struct nvme_ctrl *ctrl)
 953 {
 954 }
 955 static inline void nvme_mpath_stop(struct nvme_ctrl *ctrl)
 956 {
 957 }
 958 static inline void nvme_mpath_unfreeze(struct nvme_subsystem *subsys)
 959 {
 960 }
 961 static inline void nvme_mpath_wait_freeze(struct nvme_subsystem *subsys)
 962 {
 963 }
 964 static inline void nvme_mpath_start_freeze(struct nvme_subsystem *subsys)
 965 {
 966 }
 967 static inline void nvme_mpath_default_iopolicy(struct nvme_subsystem *subsys)
 968 {
 969 }
 970 #endif /* CONFIG_NVME_MULTIPATH */
 971
 972 int nvme_revalidate_zones(struct nvme_ns *ns);
 973 int nvme_ns_report_zones(struct nvme_ns *ns, sector_t sector,
 974                 unsigned int nr_zones, report_zones_cb cb, void *data);
 975 #ifdef CONFIG_BLK_DEV_ZONED
 976 int nvme_update_zone_info(struct nvme_ns *ns, unsigned lbaf);
 977 blk_status_t nvme_setup_zone_mgmt_send(struct nvme_ns *ns, struct request *req,
 978                                        struct nvme_command *cmnd,
 979                                        enum nvme_zone_mgmt_action action);
 980 #else
 981 static inline blk_status_t nvme_setup_zone_mgmt_send(struct nvme_ns *ns,
 982                 struct request *req, struct nvme_command *cmnd,
 983                 enum nvme_zone_mgmt_action action)
 984 {
 985         return BLK_STS_NOTSUPP;
 986 }
 987
 988 static inline int nvme_update_zone_info(struct nvme_ns *ns, unsigned lbaf)
 989 {
 990         dev_warn(ns->ctrl->device,
 991                  "Please enable CONFIG_BLK_DEV_ZONED to support ZNS devices\n");
 992         return -EPROTONOSUPPORT;
 993 }
 994 #endif
 995
 996 static inline struct nvme_ns *nvme_get_ns_from_dev(struct device *dev)
 997 {
 998         return dev_to_disk(dev)->private_data;
 999 }
1000
1001 #ifdef CONFIG_NVME_HWMON
1002 int nvme_hwmon_init(struct nvme_ctrl *ctrl);
1003 void nvme_hwmon_exit(struct nvme_ctrl *ctrl);
1004 #else
1005 static inline int nvme_hwmon_init(struct nvme_ctrl *ctrl)
1006 {
1007         return 0;
1008 }
1009
1010 static inline void nvme_hwmon_exit(struct nvme_ctrl *ctrl)
1011 {
1012 }
1013 #endif
1014
1015 static inline bool nvme_ctrl_sgl_supported(struct nvme_ctrl *ctrl)
1016 {
1017         return ctrl->sgls & ((1 << 0) | (1 << 1));
1018 }
1019
1020 #ifdef CONFIG_NVME_AUTH
1021 int __init nvme_init_auth(void);
1022 void __exit nvme_exit_auth(void);
1023 int nvme_auth_init_ctrl(struct nvme_ctrl *ctrl);
1024 void nvme_auth_stop(struct nvme_ctrl *ctrl);
1025 int nvme_auth_negotiate(struct nvme_ctrl *ctrl, int qid);
1026 int nvme_auth_wait(struct nvme_ctrl *ctrl, int qid);
1027 void nvme_auth_free(struct nvme_ctrl *ctrl);
1028 #else
1029 static inline int nvme_auth_init_ctrl(struct nvme_ctrl *ctrl)
1030 {
1031         return 0;
1032 }
1033 static inline int __init nvme_init_auth(void)
1034 {
1035         return 0;
1036 }
1037 static inline void __exit nvme_exit_auth(void)
1038 {
1039 }
1040 static inline void nvme_auth_stop(struct nvme_ctrl *ctrl) {};
1041 static inline int nvme_auth_negotiate(struct nvme_ctrl *ctrl, int qid)
1042 {
1043         return -EPROTONOSUPPORT;
1044 }
1045 static inline int nvme_auth_wait(struct nvme_ctrl *ctrl, int qid)
1046 {
1047         return NVME_SC_AUTH_REQUIRED;
1048 }
1049 static inline void nvme_auth_free(struct nvme_ctrl *ctrl) {};
1050 #endif
1051
1052 u32 nvme_command_effects(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
1053                          u8 opcode);
1054 int nvme_execute_passthru_rq(struct request *rq, u32 *effects);
1055 void nvme_passthru_end(struct nvme_ctrl *ctrl, u32 effects,
1056                        struct nvme_command *cmd, int status);
1057 struct nvme_ctrl *nvme_ctrl_from_file(struct file *file);
1058 struct nvme_ns *nvme_find_get_ns(struct nvme_ctrl *ctrl, unsigned nsid);
1059 void nvme_put_ns(struct nvme_ns *ns);
1060
1061 static inline bool nvme_multi_css(struct nvme_ctrl *ctrl)
1062 {
1063         return (ctrl->ctrl_config & NVME_CC_CSS_MASK) == NVME_CC_CSS_CSI;
1064 }
1065
1066 #ifdef CONFIG_NVME_VERBOSE_ERRORS
1067 const unsigned char *nvme_get_error_status_str(u16 status);
1068 const unsigned char *nvme_get_opcode_str(u8 opcode);
1069 const unsigned char *nvme_get_admin_opcode_str(u8 opcode);
1070 #else /* CONFIG_NVME_VERBOSE_ERRORS */
1071 static inline const unsigned char *nvme_get_error_status_str(u16 status)
1072 {
1073         return "I/O Error";
1074 }
1075 static inline const unsigned char *nvme_get_opcode_str(u8 opcode)
1076 {
1077         return "I/O Cmd";
1078 }
1079 static inline const unsigned char *nvme_get_admin_opcode_str(u8 opcode)
1080 {
1081         return "Admin Cmd";
1082 }
1083 #endif /* CONFIG_NVME_VERBOSE_ERRORS */
1084
1085 #endif /* _NVME_H */