sys/cddl/contrib/opensolaris/uts/common/fs/zfs/sys/spa_impl.h

   1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  23  * Copyright (c) 2011, 2015 by Delphix. All rights reserved.
  24  * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
  25  * Copyright 2013 Martin Matuska <mm@FreeBSD.org>. All rights reserved.
  26  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
  27  * Copyright 2013 Saso Kiselkov. All rights reserved.
  28  * Copyright (c) 2017 Datto Inc.
  29  */
  30
  31 #ifndef _SYS_SPA_IMPL_H
  32 #define _SYS_SPA_IMPL_H
  33
  34 #include <sys/spa.h>
  35 #include <sys/vdev.h>
  36 #include <sys/metaslab.h>
  37 #include <sys/dmu.h>
  38 #include <sys/dsl_pool.h>
  39 #include <sys/uberblock_impl.h>
  40 #include <sys/zfs_context.h>
  41 #include <sys/avl.h>
  42 #include <sys/refcount.h>
  43 #include <sys/bplist.h>
  44 #include <sys/bpobj.h>
  45 #include <sys/zfeature.h>
  46 #include <zfeature_common.h>
  47
  48 #ifdef  __cplusplus
  49 extern "C" {
  50 #endif
  51
  52 typedef struct spa_error_entry {
  53         zbookmark_phys_t        se_bookmark;
  54         char                    *se_name;
  55         avl_node_t              se_avl;
  56 } spa_error_entry_t;
  57
  58 typedef struct spa_history_phys {
  59         uint64_t sh_pool_create_len;    /* ending offset of zpool create */
  60         uint64_t sh_phys_max_off;       /* physical EOF */
  61         uint64_t sh_bof;                /* logical BOF */
  62         uint64_t sh_eof;                /* logical EOF */
  63         uint64_t sh_records_lost;       /* num of records overwritten */
  64 } spa_history_phys_t;
  65
  66 struct spa_aux_vdev {
  67         uint64_t        sav_object;             /* MOS object for device list */
  68         nvlist_t        *sav_config;            /* cached device config */
  69         vdev_t          **sav_vdevs;            /* devices */
  70         int             sav_count;              /* number devices */
  71         boolean_t       sav_sync;               /* sync the device list */
  72         nvlist_t        **sav_pending;          /* pending device additions */
  73         uint_t          sav_npending;           /* # pending devices */
  74 };
  75
  76 typedef struct spa_config_lock {
  77         kmutex_t        scl_lock;
  78         kthread_t       *scl_writer;
  79         int             scl_write_wanted;
  80         kcondvar_t      scl_cv;
  81         refcount_t      scl_count;
  82 } spa_config_lock_t;
  83
  84 typedef struct spa_config_dirent {
  85         list_node_t     scd_link;
  86         char            *scd_path;
  87 } spa_config_dirent_t;
  88
  89 typedef enum zio_taskq_type {
  90         ZIO_TASKQ_ISSUE = 0,
  91         ZIO_TASKQ_ISSUE_HIGH,
  92         ZIO_TASKQ_INTERRUPT,
  93         ZIO_TASKQ_INTERRUPT_HIGH,
  94         ZIO_TASKQ_TYPES
  95 } zio_taskq_type_t;
  96
  97 /*
  98  * State machine for the zpool-poolname process.  The states transitions
  99  * are done as follows:
 100  *
 101  *      From               To                   Routine
 102  *      PROC_NONE       -> PROC_CREATED         spa_activate()
 103  *      PROC_CREATED    -> PROC_ACTIVE          spa_thread()
 104  *      PROC_ACTIVE     -> PROC_DEACTIVATE      spa_deactivate()
 105  *      PROC_DEACTIVATE -> PROC_GONE            spa_thread()
 106  *      PROC_GONE       -> PROC_NONE            spa_deactivate()
 107  */
 108 typedef enum spa_proc_state {
 109         SPA_PROC_NONE,          /* spa_proc = &p0, no process created */
 110         SPA_PROC_CREATED,       /* spa_activate() has proc, is waiting */
 111         SPA_PROC_ACTIVE,        /* taskqs created, spa_proc set */
 112         SPA_PROC_DEACTIVATE,    /* spa_deactivate() requests process exit */
 113         SPA_PROC_GONE           /* spa_thread() is exiting, spa_proc = &p0 */
 114 } spa_proc_state_t;
 115
 116 typedef struct spa_taskqs {
 117         uint_t stqs_count;
 118         taskq_t **stqs_taskq;
 119 } spa_taskqs_t;
 120
 121 typedef enum spa_all_vdev_zap_action {
 122         AVZ_ACTION_NONE = 0,
 123         AVZ_ACTION_DESTROY,     /* Destroy all per-vdev ZAPs and the AVZ. */
 124         AVZ_ACTION_REBUILD,     /* Populate the new AVZ, see spa_avz_rebuild */
 125         AVZ_ACTION_INITIALIZE
 126 } spa_avz_action_t;
 127
 128 struct spa {
 129         /*
 130          * Fields protected by spa_namespace_lock.
 131          */
 132         char            spa_name[ZFS_MAX_DATASET_NAME_LEN];     /* pool name */
 133         char            *spa_comment;           /* comment */
 134         avl_node_t      spa_avl;                /* node in spa_namespace_avl */
 135         nvlist_t        *spa_config;            /* last synced config */
 136         nvlist_t        *spa_config_syncing;    /* currently syncing config */
 137         nvlist_t        *spa_config_splitting;  /* config for splitting */
 138         nvlist_t        *spa_load_info;         /* info and errors from load */
 139         uint64_t        spa_config_txg;         /* txg of last config change */
 140         int             spa_sync_pass;          /* iterate-to-convergence */
 141         pool_state_t    spa_state;              /* pool state */
 142         int             spa_inject_ref;         /* injection references */
 143         uint8_t         spa_sync_on;            /* sync threads are running */
 144         spa_load_state_t spa_load_state;        /* current load operation */
 145         uint64_t        spa_import_flags;       /* import specific flags */
 146         spa_taskqs_t    spa_zio_taskq[ZIO_TYPES][ZIO_TASKQ_TYPES];
 147         dsl_pool_t      *spa_dsl_pool;
 148         boolean_t       spa_is_initializing;    /* true while opening pool */
 149         metaslab_class_t *spa_normal_class;     /* normal data class */
 150         metaslab_class_t *spa_log_class;        /* intent log data class */
 151         uint64_t        spa_first_txg;          /* first txg after spa_open() */
 152         uint64_t        spa_final_txg;          /* txg of export/destroy */
 153         uint64_t        spa_freeze_txg;         /* freeze pool at this txg */
 154         uint64_t        spa_load_max_txg;       /* best initial ub_txg */
 155         uint64_t        spa_claim_max_txg;      /* highest claimed birth txg */
 156         timespec_t      spa_loaded_ts;          /* 1st successful open time */
 157         objset_t        *spa_meta_objset;       /* copy of dp->dp_meta_objset */
 158         kmutex_t        spa_evicting_os_lock;   /* Evicting objset list lock */
 159         list_t          spa_evicting_os_list;   /* Objsets being evicted. */
 160         kcondvar_t      spa_evicting_os_cv;     /* Objset Eviction Completion */
 161         txg_list_t      spa_vdev_txg_list;      /* per-txg dirty vdev list */
 162         vdev_t          *spa_root_vdev;         /* top-level vdev container */
 163         int             spa_min_ashift;         /* of vdevs in normal class */
 164         int             spa_max_ashift;         /* of vdevs in normal class */
 165         uint64_t        spa_config_guid;        /* config pool guid */
 166         uint64_t        spa_load_guid;          /* spa_load initialized guid */
 167         uint64_t        spa_last_synced_guid;   /* last synced guid */
 168         list_t          spa_config_dirty_list;  /* vdevs with dirty config */
 169         list_t          spa_state_dirty_list;   /* vdevs with dirty state */
 170         kmutex_t        spa_alloc_lock;
 171         avl_tree_t      spa_alloc_tree;
 172         spa_aux_vdev_t  spa_spares;             /* hot spares */
 173         spa_aux_vdev_t  spa_l2cache;            /* L2ARC cache devices */
 174         nvlist_t        *spa_label_features;    /* Features for reading MOS */
 175         uint64_t        spa_config_object;      /* MOS object for pool config */
 176         uint64_t        spa_config_generation;  /* config generation number */
 177         uint64_t        spa_syncing_txg;        /* txg currently syncing */
 178         bpobj_t         spa_deferred_bpobj;     /* deferred-free bplist */
 179         bplist_t        spa_free_bplist[TXG_SIZE]; /* bplist of stuff to free */
 180         zio_cksum_salt_t spa_cksum_salt;        /* secret salt for cksum */
 181         /* checksum context templates */
 182         kmutex_t        spa_cksum_tmpls_lock;
 183         void            *spa_cksum_tmpls[ZIO_CHECKSUM_FUNCTIONS];
 184         uberblock_t     spa_ubsync;             /* last synced uberblock */
 185         uberblock_t     spa_uberblock;          /* current uberblock */
 186         boolean_t       spa_extreme_rewind;     /* rewind past deferred frees */
 187         uint64_t        spa_last_io;            /* lbolt of last non-scan I/O */
 188         kmutex_t        spa_scrub_lock;         /* resilver/scrub lock */
 189         uint64_t        spa_scrub_inflight;     /* in-flight scrub I/Os */
 190         kcondvar_t      spa_scrub_io_cv;        /* scrub I/O completion */
 191         uint8_t         spa_scrub_active;       /* active or suspended? */
 192         uint8_t         spa_scrub_type;         /* type of scrub we're doing */
 193         uint8_t         spa_scrub_finished;     /* indicator to rotate logs */
 194         uint8_t         spa_scrub_started;      /* started since last boot */
 195         uint8_t         spa_scrub_reopen;       /* scrub doing vdev_reopen */
 196         uint64_t        spa_scan_pass_start;    /* start time per pass/reboot */
 197         uint64_t        spa_scan_pass_scrub_pause; /* scrub pause time */
 198         uint64_t        spa_scan_pass_scrub_spent_paused; /* total paused */
 199         uint64_t        spa_scan_pass_exam;     /* examined bytes per pass */
 200         kmutex_t        spa_async_lock;         /* protect async state */
 201         kthread_t       *spa_async_thread;      /* thread doing async task */
 202         kthread_t       *spa_async_thread_vd;   /* thread doing vd async task */
 203         int             spa_async_suspended;    /* async tasks suspended */
 204         kcondvar_t      spa_async_cv;           /* wait for thread_exit() */
 205         uint16_t        spa_async_tasks;        /* async task mask */
 206         char            *spa_root;              /* alternate root directory */
 207         uint64_t        spa_ena;                /* spa-wide ereport ENA */
 208         int             spa_last_open_failed;   /* error if last open failed */
 209         uint64_t        spa_last_ubsync_txg;    /* "best" uberblock txg */
 210         uint64_t        spa_last_ubsync_txg_ts; /* timestamp from that ub */
 211         uint64_t        spa_load_txg;           /* ub txg that loaded */
 212         uint64_t        spa_load_txg_ts;        /* timestamp from that ub */
 213         uint64_t        spa_load_meta_errors;   /* verify metadata err count */
 214         uint64_t        spa_load_data_errors;   /* verify data err count */
 215         uint64_t        spa_verify_min_txg;     /* start txg of verify scrub */
 216         kmutex_t        spa_errlog_lock;        /* error log lock */
 217         uint64_t        spa_errlog_last;        /* last error log object */
 218         uint64_t        spa_errlog_scrub;       /* scrub error log object */
 219         kmutex_t        spa_errlist_lock;       /* error list/ereport lock */
 220         avl_tree_t      spa_errlist_last;       /* last error list */
 221         avl_tree_t      spa_errlist_scrub;      /* scrub error list */
 222         uint64_t        spa_deflate;            /* should we deflate? */
 223         uint64_t        spa_history;            /* history object */
 224         kmutex_t        spa_history_lock;       /* history lock */
 225         vdev_t          *spa_pending_vdev;      /* pending vdev additions */
 226         kmutex_t        spa_props_lock;         /* property lock */
 227         uint64_t        spa_pool_props_object;  /* object for properties */
 228         uint64_t        spa_bootfs;             /* default boot filesystem */
 229         uint64_t        spa_failmode;           /* failure mode for the pool */
 230         uint64_t        spa_delegation;         /* delegation on/off */
 231         list_t          spa_config_list;        /* previous cache file(s) */
 232         /* per-CPU array of root of async I/O: */
 233         zio_t           **spa_async_zio_root;
 234         zio_t           *spa_suspend_zio_root;  /* root of all suspended I/O */
 235         kmutex_t        spa_suspend_lock;       /* protects suspend_zio_root */
 236         kcondvar_t      spa_suspend_cv;         /* notification of resume */
 237         uint8_t         spa_suspended;          /* pool is suspended */
 238         uint8_t         spa_claiming;           /* pool is doing zil_claim() */
 239         boolean_t       spa_debug;              /* debug enabled? */
 240         boolean_t       spa_is_root;            /* pool is root */
 241         int             spa_minref;             /* num refs when first opened */
 242         int             spa_mode;               /* FREAD | FWRITE */
 243         spa_log_state_t spa_log_state;          /* log state */
 244         uint64_t        spa_autoexpand;         /* lun expansion on/off */
 245         ddt_t           *spa_ddt[ZIO_CHECKSUM_FUNCTIONS]; /* in-core DDTs */
 246         uint64_t        spa_ddt_stat_object;    /* DDT statistics */
 247         uint64_t        spa_dedup_ditto;        /* dedup ditto threshold */
 248         uint64_t        spa_dedup_checksum;     /* default dedup checksum */
 249         uint64_t        spa_dspace;             /* dspace in normal class */
 250         kmutex_t        spa_vdev_top_lock;      /* dueling offline/remove */
 251         kmutex_t        spa_proc_lock;          /* protects spa_proc* */
 252         kcondvar_t      spa_proc_cv;            /* spa_proc_state transitions */
 253         spa_proc_state_t spa_proc_state;        /* see definition */
 254         struct proc     *spa_proc;              /* "zpool-poolname" process */
 255         uint64_t        spa_did;                /* if procp != p0, did of t1 */
 256         kthread_t       *spa_trim_thread;       /* thread sending TRIM I/Os */
 257         kmutex_t        spa_trim_lock;          /* protects spa_trim_cv */
 258         kcondvar_t      spa_trim_cv;            /* used to notify TRIM thread */
 259         boolean_t       spa_autoreplace;        /* autoreplace set in open */
 260         int             spa_vdev_locks;         /* locks grabbed */
 261         uint64_t        spa_creation_version;   /* version at pool creation */
 262         uint64_t        spa_prev_software_version; /* See ub_software_version */
 263         uint64_t        spa_feat_for_write_obj; /* required to write to pool */
 264         uint64_t        spa_feat_for_read_obj;  /* required to read from pool */
 265         uint64_t        spa_feat_desc_obj;      /* Feature descriptions */
 266         uint64_t        spa_feat_enabled_txg_obj; /* Feature enabled txg */
 267         /* cache feature refcounts */
 268         uint64_t        spa_feat_refcount_cache[SPA_FEATURES];
 269 #ifdef illumos
 270         cyclic_id_t     spa_deadman_cycid;      /* cyclic id */
 271 #else   /* !illumos */
 272 #ifdef _KERNEL
 273         struct callout  spa_deadman_cycid;      /* callout id */
 274         struct task     spa_deadman_task;
 275 #endif
 276 #endif  /* illumos */
 277         uint64_t        spa_deadman_calls;      /* number of deadman calls */
 278         hrtime_t        spa_sync_starttime;     /* starting time fo spa_sync */
 279         uint64_t        spa_deadman_synctime;   /* deadman expiration timer */
 280         uint64_t        spa_all_vdev_zaps;      /* ZAP of per-vd ZAP obj #s */
 281         spa_avz_action_t        spa_avz_action; /* destroy/rebuild AVZ? */
 282
 283 #ifdef illumos
 284         /*
 285          * spa_iokstat_lock protects spa_iokstat and
 286          * spa_queue_stats[].
 287          */
 288         kmutex_t        spa_iokstat_lock;
 289         struct kstat    *spa_iokstat;           /* kstat of io to this pool */
 290         struct {
 291                 int spa_active;
 292                 int spa_queued;
 293         } spa_queue_stats[ZIO_PRIORITY_NUM_QUEUEABLE];
 294 #endif
 295         hrtime_t        spa_ccw_fail_time;      /* Conf cache write fail time */
 296
 297         /*
 298          * spa_refcount & spa_config_lock must be the last elements
 299          * because refcount_t changes size based on compilation options.
 300          * In order for the MDB module to function correctly, the other
 301          * fields must remain in the same location.
 302          */
 303         spa_config_lock_t spa_config_lock[SCL_LOCKS]; /* config changes */
 304         refcount_t      spa_refcount;           /* number of opens */
 305 #ifndef illumos
 306         boolean_t       spa_splitting_newspa;   /* creating new spa in split */
 307 #endif
 308 };
 309
 310 extern const char *spa_config_path;
 311
 312 extern void spa_taskq_dispatch_ent(spa_t *spa, zio_type_t t, zio_taskq_type_t q,
 313     task_func_t *func, void *arg, uint_t flags, taskq_ent_t *ent);
 314
 315 #ifdef  __cplusplus
 316 }
 317 #endif
 318
 319 #endif  /* _SYS_SPA_IMPL_H */