include/sys/vdev_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  */
  25
  26 #ifndef _SYS_VDEV_IMPL_H
  27 #define _SYS_VDEV_IMPL_H
  28
  29 #include <sys/avl.h>
  30 #include <sys/dmu.h>
  31 #include <sys/metaslab.h>
  32 #include <sys/nvpair.h>
  33 #include <sys/space_map.h>
  34 #include <sys/vdev.h>
  35 #include <sys/dkio.h>
  36 #include <sys/uberblock_impl.h>
  37 #include <sys/zfs_ratelimit.h>
  38
  39 #ifdef  __cplusplus
  40 extern "C" {
  41 #endif
  42
  43 /*
  44  * Virtual device descriptors.
  45  *
  46  * All storage pool operations go through the virtual device framework,
  47  * which provides data replication and I/O scheduling.
  48  */
  49
  50 /*
  51  * Forward declarations that lots of things need.
  52  */
  53 typedef struct vdev_queue vdev_queue_t;
  54 typedef struct vdev_cache vdev_cache_t;
  55 typedef struct vdev_cache_entry vdev_cache_entry_t;
  56 struct abd;
  57
  58 extern int zfs_vdev_queue_depth_pct;
  59 extern uint32_t zfs_vdev_async_write_max_active;
  60
  61 /*
  62  * Virtual device operations
  63  */
  64 typedef int     vdev_open_func_t(vdev_t *vd, uint64_t *size, uint64_t *max_size,
  65     uint64_t *ashift);
  66 typedef void    vdev_close_func_t(vdev_t *vd);
  67 typedef uint64_t vdev_asize_func_t(vdev_t *vd, uint64_t psize);
  68 typedef void    vdev_io_start_func_t(zio_t *zio);
  69 typedef void    vdev_io_done_func_t(zio_t *zio);
  70 typedef void    vdev_state_change_func_t(vdev_t *vd, int, int);
  71 typedef boolean_t vdev_need_resilver_func_t(vdev_t *vd, uint64_t, size_t);
  72 typedef void    vdev_hold_func_t(vdev_t *vd);
  73 typedef void    vdev_rele_func_t(vdev_t *vd);
  74
  75 typedef const struct vdev_ops {
  76         vdev_open_func_t                *vdev_op_open;
  77         vdev_close_func_t               *vdev_op_close;
  78         vdev_asize_func_t               *vdev_op_asize;
  79         vdev_io_start_func_t            *vdev_op_io_start;
  80         vdev_io_done_func_t             *vdev_op_io_done;
  81         vdev_state_change_func_t        *vdev_op_state_change;
  82         vdev_need_resilver_func_t       *vdev_op_need_resilver;
  83         vdev_hold_func_t                *vdev_op_hold;
  84         vdev_rele_func_t                *vdev_op_rele;
  85         char                            vdev_op_type[16];
  86         boolean_t                       vdev_op_leaf;
  87 } vdev_ops_t;
  88
  89 /*
  90  * Virtual device properties
  91  */
  92 struct vdev_cache_entry {
  93         struct abd      *ve_abd;
  94         uint64_t        ve_offset;
  95         clock_t         ve_lastused;
  96         avl_node_t      ve_offset_node;
  97         avl_node_t      ve_lastused_node;
  98         uint32_t        ve_hits;
  99         uint16_t        ve_missed_update;
 100         zio_t           *ve_fill_io;
 101 };
 102
 103 struct vdev_cache {
 104         avl_tree_t      vc_offset_tree;
 105         avl_tree_t      vc_lastused_tree;
 106         kmutex_t        vc_lock;
 107 };
 108
 109 typedef struct vdev_queue_class {
 110         uint32_t        vqc_active;
 111
 112         /*
 113          * Sorted by offset or timestamp, depending on if the queue is
 114          * LBA-ordered vs FIFO.
 115          */
 116         avl_tree_t      vqc_queued_tree;
 117 } vdev_queue_class_t;
 118
 119 struct vdev_queue {
 120         vdev_t          *vq_vdev;
 121         vdev_queue_class_t vq_class[ZIO_PRIORITY_NUM_QUEUEABLE];
 122         avl_tree_t      vq_active_tree;
 123         avl_tree_t      vq_read_offset_tree;
 124         avl_tree_t      vq_write_offset_tree;
 125         uint64_t        vq_last_offset;
 126         hrtime_t        vq_io_complete_ts; /* time last i/o completed */
 127         hrtime_t        vq_io_delta_ts;
 128         zio_t           vq_io_search; /* used as local for stack reduction */
 129         kmutex_t        vq_lock;
 130         uint64_t        vq_lastoffset;
 131 };
 132
 133 /*
 134  * Virtual device descriptor
 135  */
 136 struct vdev {
 137         /*
 138          * Common to all vdev types.
 139          */
 140         uint64_t        vdev_id;        /* child number in vdev parent  */
 141         uint64_t        vdev_guid;      /* unique ID for this vdev      */
 142         uint64_t        vdev_guid_sum;  /* self guid + all child guids  */
 143         uint64_t        vdev_orig_guid; /* orig. guid prior to remove   */
 144         uint64_t        vdev_asize;     /* allocatable device capacity  */
 145         uint64_t        vdev_min_asize; /* min acceptable asize         */
 146         uint64_t        vdev_max_asize; /* max acceptable asize         */
 147         uint64_t        vdev_ashift;    /* block alignment shift        */
 148         uint64_t        vdev_state;     /* see VDEV_STATE_* #defines    */
 149         uint64_t        vdev_prevstate; /* used when reopening a vdev   */
 150         vdev_ops_t      *vdev_ops;      /* vdev operations              */
 151         spa_t           *vdev_spa;      /* spa for this vdev            */
 152         void            *vdev_tsd;      /* type-specific data           */
 153         vnode_t         *vdev_name_vp;  /* vnode for pathname           */
 154         vnode_t         *vdev_devid_vp; /* vnode for devid              */
 155         vdev_t          *vdev_top;      /* top-level vdev               */
 156         vdev_t          *vdev_parent;   /* parent vdev                  */
 157         vdev_t          **vdev_child;   /* array of children            */
 158         uint64_t        vdev_children;  /* number of children           */
 159         vdev_stat_t     vdev_stat;      /* virtual device statistics    */
 160         vdev_stat_ex_t  vdev_stat_ex;   /* extended statistics          */
 161         boolean_t       vdev_expanding; /* expand the vdev?             */
 162         boolean_t       vdev_reopening; /* reopen in progress?          */
 163         boolean_t       vdev_nonrot;    /* true if solid state          */
 164         int             vdev_open_error; /* error on last open          */
 165         kthread_t       *vdev_open_thread; /* thread opening children   */
 166         uint64_t        vdev_crtxg;     /* txg when top-level was added */
 167
 168         /*
 169          * Top-level vdev state.
 170          */
 171         uint64_t        vdev_ms_array;  /* metaslab array object        */
 172         uint64_t        vdev_ms_shift;  /* metaslab size shift          */
 173         uint64_t        vdev_ms_count;  /* number of metaslabs          */
 174         metaslab_group_t *vdev_mg;      /* metaslab group               */
 175         metaslab_t      **vdev_ms;      /* metaslab array               */
 176         uint64_t        vdev_pending_fastwrite; /* allocated fastwrites */
 177         txg_list_t      vdev_ms_list;   /* per-txg dirty metaslab lists */
 178         txg_list_t      vdev_dtl_list;  /* per-txg dirty DTL lists      */
 179         txg_node_t      vdev_txg_node;  /* per-txg dirty vdev linkage   */
 180         boolean_t       vdev_remove_wanted; /* async remove wanted?     */
 181         boolean_t       vdev_probe_wanted; /* async probe wanted?       */
 182         list_node_t     vdev_config_dirty_node; /* config dirty list    */
 183         list_node_t     vdev_state_dirty_node; /* state dirty list      */
 184         uint64_t        vdev_deflate_ratio; /* deflation ratio (x512)   */
 185         uint64_t        vdev_islog;     /* is an intent log device      */
 186         uint64_t        vdev_removing;  /* device is being removed?     */
 187         boolean_t       vdev_ishole;    /* is a hole in the namespace   */
 188         kmutex_t        vdev_queue_lock; /* protects vdev_queue_depth   */
 189         uint64_t        vdev_top_zap;
 190
 191         /*
 192          * The queue depth parameters determine how many async writes are
 193          * still pending (i.e. allocated by net yet issued to disk) per
 194          * top-level (vdev_async_write_queue_depth) and the maximum allowed
 195          * (vdev_max_async_write_queue_depth). These values only apply to
 196          * top-level vdevs.
 197          */
 198         uint64_t        vdev_async_write_queue_depth;
 199         uint64_t        vdev_max_async_write_queue_depth;
 200
 201         /*
 202          * Leaf vdev state.
 203          */
 204         range_tree_t    *vdev_dtl[DTL_TYPES]; /* dirty time logs        */
 205         space_map_t     *vdev_dtl_sm;   /* dirty time log space map     */
 206         txg_node_t      vdev_dtl_node;  /* per-txg dirty DTL linkage    */
 207         uint64_t        vdev_dtl_object; /* DTL object                  */
 208         uint64_t        vdev_psize;     /* physical device capacity     */
 209         uint64_t        vdev_wholedisk; /* true if this is a whole disk */
 210         uint64_t        vdev_offline;   /* persistent offline state     */
 211         uint64_t        vdev_faulted;   /* persistent faulted state     */
 212         uint64_t        vdev_degraded;  /* persistent degraded state    */
 213         uint64_t        vdev_removed;   /* persistent removed state     */
 214         uint64_t        vdev_resilver_txg; /* persistent resilvering state */
 215         uint64_t        vdev_nparity;   /* number of parity devices for raidz */
 216         char            *vdev_path;     /* vdev path (if any)           */
 217         char            *vdev_devid;    /* vdev devid (if any)          */
 218         char            *vdev_physpath; /* vdev device path (if any)    */
 219         char            *vdev_enc_sysfs_path;   /* enclosure sysfs path */
 220         char            *vdev_fru;      /* physical FRU location        */
 221         uint64_t        vdev_not_present; /* not present during import  */
 222         uint64_t        vdev_unspare;   /* unspare when resilvering done */
 223         boolean_t       vdev_nowritecache; /* true if flushwritecache failed */
 224         boolean_t       vdev_checkremove; /* temporary online test      */
 225         boolean_t       vdev_forcefault; /* force online fault          */
 226         boolean_t       vdev_splitting; /* split or repair in progress  */
 227         boolean_t       vdev_delayed_close; /* delayed device close?    */
 228         boolean_t       vdev_tmpoffline; /* device taken offline temporarily? */
 229         boolean_t       vdev_detached;  /* device detached?             */
 230         boolean_t       vdev_cant_read; /* vdev is failing all reads    */
 231         boolean_t       vdev_cant_write; /* vdev is failing all writes  */
 232         boolean_t       vdev_isspare;   /* was a hot spare              */
 233         boolean_t       vdev_isl2cache; /* was a l2cache device         */
 234         boolean_t       vdev_copy_uberblocks;  /* post expand copy uberblocks */
 235         vdev_queue_t    vdev_queue;     /* I/O deadline schedule queue  */
 236         vdev_cache_t    vdev_cache;     /* physical block cache         */
 237         spa_aux_vdev_t  *vdev_aux;      /* for l2cache and spares vdevs */
 238         zio_t           *vdev_probe_zio; /* root of current probe       */
 239         vdev_aux_t      vdev_label_aux; /* on-disk aux state            */
 240         uint64_t        vdev_leaf_zap;
 241         hrtime_t        vdev_mmp_pending; /* 0 if write finished        */
 242
 243         /*
 244          * For DTrace to work in userland (libzpool) context, these fields must
 245          * remain at the end of the structure.  DTrace will use the kernel's
 246          * CTF definition for 'struct vdev', and since the size of a kmutex_t is
 247          * larger in userland, the offsets for the rest of the fields would be
 248          * incorrect.
 249          */
 250         kmutex_t        vdev_dtl_lock;  /* vdev_dtl_{map,resilver}      */
 251         kmutex_t        vdev_stat_lock; /* vdev_stat                    */
 252         kmutex_t        vdev_probe_lock; /* protects vdev_probe_zio     */
 253
 254         /*
 255          * We rate limit ZIO delay and ZIO checksum events, since they
 256          * can flood ZED with tons of events when a drive is acting up.
 257          */
 258 #define DELAYS_PER_SECOND 5
 259 #define CHECKSUMS_PER_SECOND 5
 260         zfs_ratelimit_t vdev_delay_rl;
 261         zfs_ratelimit_t vdev_checksum_rl;
 262 };
 263
 264 #define VDEV_RAIDZ_MAXPARITY    3
 265
 266 #define VDEV_PAD_SIZE           (8 << 10)
 267 /* 2 padding areas (vl_pad1 and vl_pad2) to skip */
 268 #define VDEV_SKIP_SIZE          VDEV_PAD_SIZE * 2
 269 #define VDEV_PHYS_SIZE          (112 << 10)
 270 #define VDEV_UBERBLOCK_RING     (128 << 10)
 271
 272 /*
 273  * MMP blocks occupy the last MMP_BLOCKS_PER_LABEL slots in the uberblock
 274  * ring when MMP is enabled.
 275  */
 276 #define MMP_BLOCKS_PER_LABEL    1
 277
 278 /* The largest uberblock we support is 8k. */
 279 #define MAX_UBERBLOCK_SHIFT (13)
 280 #define VDEV_UBERBLOCK_SHIFT(vd)        \
 281         MIN(MAX((vd)->vdev_top->vdev_ashift, UBERBLOCK_SHIFT), \
 282             MAX_UBERBLOCK_SHIFT)
 283 #define VDEV_UBERBLOCK_COUNT(vd)        \
 284         (VDEV_UBERBLOCK_RING >> VDEV_UBERBLOCK_SHIFT(vd))
 285 #define VDEV_UBERBLOCK_OFFSET(vd, n)    \
 286         offsetof(vdev_label_t, vl_uberblock[(n) << VDEV_UBERBLOCK_SHIFT(vd)])
 287 #define VDEV_UBERBLOCK_SIZE(vd)         (1ULL << VDEV_UBERBLOCK_SHIFT(vd))
 288
 289 typedef struct vdev_phys {
 290         char            vp_nvlist[VDEV_PHYS_SIZE - sizeof (zio_eck_t)];
 291         zio_eck_t       vp_zbt;
 292 } vdev_phys_t;
 293
 294 typedef struct vdev_label {
 295         char            vl_pad1[VDEV_PAD_SIZE];                 /*  8K */
 296         char            vl_pad2[VDEV_PAD_SIZE];                 /*  8K */
 297         vdev_phys_t     vl_vdev_phys;                           /* 112K */
 298         char            vl_uberblock[VDEV_UBERBLOCK_RING];      /* 128K */
 299 } vdev_label_t;                                                 /* 256K total */
 300
 301 /*
 302  * vdev_dirty() flags
 303  */
 304 #define VDD_METASLAB    0x01
 305 #define VDD_DTL         0x02
 306
 307 /* Offset of embedded boot loader region on each label */
 308 #define VDEV_BOOT_OFFSET        (2 * sizeof (vdev_label_t))
 309 /*
 310  * Size of embedded boot loader region on each label.
 311  * The total size of the first two labels plus the boot area is 4MB.
 312  */
 313 #define VDEV_BOOT_SIZE          (7ULL << 19)                    /* 3.5M */
 314
 315 /*
 316  * Size of label regions at the start and end of each leaf device.
 317  */
 318 #define VDEV_LABEL_START_SIZE   (2 * sizeof (vdev_label_t) + VDEV_BOOT_SIZE)
 319 #define VDEV_LABEL_END_SIZE     (2 * sizeof (vdev_label_t))
 320 #define VDEV_LABELS             4
 321 #define VDEV_BEST_LABEL         VDEV_LABELS
 322
 323 #define VDEV_ALLOC_LOAD         0
 324 #define VDEV_ALLOC_ADD          1
 325 #define VDEV_ALLOC_SPARE        2
 326 #define VDEV_ALLOC_L2CACHE      3
 327 #define VDEV_ALLOC_ROOTPOOL     4
 328 #define VDEV_ALLOC_SPLIT        5
 329 #define VDEV_ALLOC_ATTACH       6
 330
 331 /*
 332  * Allocate or free a vdev
 333  */
 334 extern vdev_t *vdev_alloc_common(spa_t *spa, uint_t id, uint64_t guid,
 335     vdev_ops_t *ops);
 336 extern int vdev_alloc(spa_t *spa, vdev_t **vdp, nvlist_t *config,
 337     vdev_t *parent, uint_t id, int alloctype);
 338 extern void vdev_free(vdev_t *vd);
 339
 340 /*
 341  * Add or remove children and parents
 342  */
 343 extern void vdev_add_child(vdev_t *pvd, vdev_t *cvd);
 344 extern void vdev_remove_child(vdev_t *pvd, vdev_t *cvd);
 345 extern void vdev_compact_children(vdev_t *pvd);
 346 extern vdev_t *vdev_add_parent(vdev_t *cvd, vdev_ops_t *ops);
 347 extern void vdev_remove_parent(vdev_t *cvd);
 348
 349 /*
 350  * vdev sync load and sync
 351  */
 352 extern void vdev_load_log_state(vdev_t *nvd, vdev_t *ovd);
 353 extern boolean_t vdev_log_state_valid(vdev_t *vd);
 354 extern void vdev_load(vdev_t *vd);
 355 extern int vdev_dtl_load(vdev_t *vd);
 356 extern void vdev_sync(vdev_t *vd, uint64_t txg);
 357 extern void vdev_sync_done(vdev_t *vd, uint64_t txg);
 358 extern void vdev_dirty(vdev_t *vd, int flags, void *arg, uint64_t txg);
 359 extern void vdev_dirty_leaves(vdev_t *vd, int flags, uint64_t txg);
 360
 361 /*
 362  * Available vdev types.
 363  */
 364 extern vdev_ops_t vdev_root_ops;
 365 extern vdev_ops_t vdev_mirror_ops;
 366 extern vdev_ops_t vdev_replacing_ops;
 367 extern vdev_ops_t vdev_raidz_ops;
 368 extern vdev_ops_t vdev_disk_ops;
 369 extern vdev_ops_t vdev_file_ops;
 370 extern vdev_ops_t vdev_missing_ops;
 371 extern vdev_ops_t vdev_hole_ops;
 372 extern vdev_ops_t vdev_spare_ops;
 373
 374 /*
 375  * Common size functions
 376  */
 377 extern uint64_t vdev_default_asize(vdev_t *vd, uint64_t psize);
 378 extern uint64_t vdev_get_min_asize(vdev_t *vd);
 379 extern void vdev_set_min_asize(vdev_t *vd);
 380
 381 /*
 382  * Global variables
 383  */
 384 /* zdb uses this tunable, so it must be declared here to make lint happy. */
 385 extern int zfs_vdev_cache_size;
 386
 387 #ifdef  __cplusplus
 388 }
 389 #endif
 390
 391 #endif  /* _SYS_VDEV_IMPL_H */