module/zfs/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 2009 Sun Microsystems, Inc.  All rights reserved.
  23  * Use is subject to license terms.
  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
  38 #ifdef  __cplusplus
  39 extern "C" {
  40 #endif
  41
  42 /*
  43  * Virtual device descriptors.
  44  *
  45  * All storage pool operations go through the virtual device framework,
  46  * which provides data replication and I/O scheduling.
  47  */
  48
  49 /*
  50  * Forward declarations that lots of things need.
  51  */
  52 typedef struct vdev_queue vdev_queue_t;
  53 typedef struct vdev_cache vdev_cache_t;
  54 typedef struct vdev_cache_entry vdev_cache_entry_t;
  55
  56 /*
  57  * Virtual device operations
  58  */
  59 typedef int     vdev_open_func_t(vdev_t *vd, uint64_t *size, uint64_t *ashift);
  60 typedef void    vdev_close_func_t(vdev_t *vd);
  61 typedef uint64_t vdev_asize_func_t(vdev_t *vd, uint64_t psize);
  62 typedef int     vdev_io_start_func_t(zio_t *zio);
  63 typedef void    vdev_io_done_func_t(zio_t *zio);
  64 typedef void    vdev_state_change_func_t(vdev_t *vd, int, int);
  65
  66 typedef struct vdev_ops {
  67         vdev_open_func_t                *vdev_op_open;
  68         vdev_close_func_t               *vdev_op_close;
  69         vdev_asize_func_t               *vdev_op_asize;
  70         vdev_io_start_func_t            *vdev_op_io_start;
  71         vdev_io_done_func_t             *vdev_op_io_done;
  72         vdev_state_change_func_t        *vdev_op_state_change;
  73         char                            vdev_op_type[16];
  74         boolean_t                       vdev_op_leaf;
  75 } vdev_ops_t;
  76
  77 /*
  78  * Virtual device properties
  79  */
  80 struct vdev_cache_entry {
  81         char            *ve_data;
  82         uint64_t        ve_offset;
  83         uint64_t        ve_lastused;
  84         avl_node_t      ve_offset_node;
  85         avl_node_t      ve_lastused_node;
  86         uint32_t        ve_hits;
  87         uint16_t        ve_missed_update;
  88         zio_t           *ve_fill_io;
  89 };
  90
  91 struct vdev_cache {
  92         avl_tree_t      vc_offset_tree;
  93         avl_tree_t      vc_lastused_tree;
  94         kmutex_t        vc_lock;
  95 };
  96
  97 struct vdev_queue {
  98         avl_tree_t      vq_deadline_tree;
  99         avl_tree_t      vq_read_tree;
 100         avl_tree_t      vq_write_tree;
 101         avl_tree_t      vq_pending_tree;
 102         kmutex_t        vq_lock;
 103 };
 104
 105 /*
 106  * Virtual device descriptor
 107  */
 108 struct vdev {
 109         /*
 110          * Common to all vdev types.
 111          */
 112         uint64_t        vdev_id;        /* child number in vdev parent  */
 113         uint64_t        vdev_guid;      /* unique ID for this vdev      */
 114         uint64_t        vdev_guid_sum;  /* self guid + all child guids  */
 115         uint64_t        vdev_asize;     /* allocatable device capacity  */
 116         uint64_t        vdev_min_asize; /* min acceptable asize         */
 117         uint64_t        vdev_ashift;    /* block alignment shift        */
 118         uint64_t        vdev_state;     /* see VDEV_STATE_* #defines    */
 119         uint64_t        vdev_prevstate; /* used when reopening a vdev   */
 120         vdev_ops_t      *vdev_ops;      /* vdev operations              */
 121         spa_t           *vdev_spa;      /* spa for this vdev            */
 122         void            *vdev_tsd;      /* type-specific data           */
 123         vdev_t          *vdev_top;      /* top-level vdev               */
 124         vdev_t          *vdev_parent;   /* parent vdev                  */
 125         vdev_t          **vdev_child;   /* array of children            */
 126         uint64_t        vdev_children;  /* number of children           */
 127         space_map_t     vdev_dtl[DTL_TYPES]; /* in-core dirty time logs */
 128         vdev_stat_t     vdev_stat;      /* virtual device statistics    */
 129         boolean_t       vdev_expanding; /* expand the vdev?             */
 130         int             vdev_open_error; /* error on last open          */
 131         kthread_t       *vdev_open_thread; /* thread opening children   */
 132
 133         /*
 134          * Top-level vdev state.
 135          */
 136         uint64_t        vdev_ms_array;  /* metaslab array object        */
 137         uint64_t        vdev_ms_shift;  /* metaslab size shift          */
 138         uint64_t        vdev_ms_count;  /* number of metaslabs          */
 139         metaslab_group_t *vdev_mg;      /* metaslab group               */
 140         metaslab_t      **vdev_ms;      /* metaslab array               */
 141         txg_list_t      vdev_ms_list;   /* per-txg dirty metaslab lists */
 142         txg_list_t      vdev_dtl_list;  /* per-txg dirty DTL lists      */
 143         txg_node_t      vdev_txg_node;  /* per-txg dirty vdev linkage   */
 144         boolean_t       vdev_remove_wanted; /* async remove wanted?     */
 145         boolean_t       vdev_probe_wanted; /* async probe wanted?       */
 146         list_node_t     vdev_config_dirty_node; /* config dirty list    */
 147         list_node_t     vdev_state_dirty_node; /* state dirty list      */
 148         uint64_t        vdev_deflate_ratio; /* deflation ratio (x512)   */
 149         uint64_t        vdev_islog;     /* is an intent log device      */
 150
 151         /*
 152          * Leaf vdev state.
 153          */
 154         uint64_t        vdev_psize;     /* physical device capacity     */
 155         space_map_obj_t vdev_dtl_smo;   /* dirty time log space map obj */
 156         txg_node_t      vdev_dtl_node;  /* per-txg dirty DTL linkage    */
 157         uint64_t        vdev_wholedisk; /* true if this is a whole disk */
 158         uint64_t        vdev_offline;   /* persistent offline state     */
 159         uint64_t        vdev_faulted;   /* persistent faulted state     */
 160         uint64_t        vdev_degraded;  /* persistent degraded state    */
 161         uint64_t        vdev_removed;   /* persistent removed state     */
 162         uint64_t        vdev_nparity;   /* number of parity devices for raidz */
 163         char            *vdev_path;     /* vdev path (if any)           */
 164         char            *vdev_devid;    /* vdev devid (if any)          */
 165         char            *vdev_physpath; /* vdev device path (if any)    */
 166         char            *vdev_fru;      /* physical FRU location        */
 167         uint64_t        vdev_not_present; /* not present during import  */
 168         uint64_t        vdev_unspare;   /* unspare when resilvering done */
 169         hrtime_t        vdev_last_try;  /* last reopen time             */
 170         boolean_t       vdev_nowritecache; /* true if flushwritecache failed */
 171         boolean_t       vdev_checkremove; /* temporary online test      */
 172         boolean_t       vdev_forcefault; /* force online fault          */
 173         uint8_t         vdev_tmpoffline; /* device taken offline temporarily? */
 174         uint8_t         vdev_detached;  /* device detached?             */
 175         uint8_t         vdev_cant_read; /* vdev is failing all reads    */
 176         uint8_t         vdev_cant_write; /* vdev is failing all writes  */
 177         uint64_t        vdev_isspare;   /* was a hot spare              */
 178         uint64_t        vdev_isl2cache; /* was a l2cache device         */
 179         vdev_queue_t    vdev_queue;     /* I/O deadline schedule queue  */
 180         vdev_cache_t    vdev_cache;     /* physical block cache         */
 181         spa_aux_vdev_t  *vdev_aux;      /* for l2cache vdevs            */
 182         zio_t           *vdev_probe_zio; /* root of current probe       */
 183
 184         /*
 185          * For DTrace to work in userland (libzpool) context, these fields must
 186          * remain at the end of the structure.  DTrace will use the kernel's
 187          * CTF definition for 'struct vdev', and since the size of a kmutex_t is
 188          * larger in userland, the offsets for the rest fields would be
 189          * incorrect.
 190          */
 191         kmutex_t        vdev_dtl_lock;  /* vdev_dtl_{map,resilver}      */
 192         kmutex_t        vdev_stat_lock; /* vdev_stat                    */
 193         kmutex_t        vdev_probe_lock; /* protects vdev_probe_zio     */
 194 };
 195
 196 #define VDEV_PAD_SIZE           (8 << 10)
 197 /* 2 padding areas (vl_pad1 and vl_pad2) to skip */
 198 #define VDEV_SKIP_SIZE          VDEV_PAD_SIZE * 2
 199 #define VDEV_PHYS_SIZE          (112 << 10)
 200 #define VDEV_UBERBLOCK_RING     (128 << 10)
 201
 202 #define VDEV_UBERBLOCK_SHIFT(vd)        \
 203         MAX((vd)->vdev_top->vdev_ashift, UBERBLOCK_SHIFT)
 204 #define VDEV_UBERBLOCK_COUNT(vd)        \
 205         (VDEV_UBERBLOCK_RING >> VDEV_UBERBLOCK_SHIFT(vd))
 206 #define VDEV_UBERBLOCK_OFFSET(vd, n)    \
 207         offsetof(vdev_label_t, vl_uberblock[(n) << VDEV_UBERBLOCK_SHIFT(vd)])
 208 #define VDEV_UBERBLOCK_SIZE(vd)         (1ULL << VDEV_UBERBLOCK_SHIFT(vd))
 209
 210 typedef struct vdev_phys {
 211         char            vp_nvlist[VDEV_PHYS_SIZE - sizeof (zio_block_tail_t)];
 212         zio_block_tail_t vp_zbt;
 213 } vdev_phys_t;
 214
 215 typedef struct vdev_label {
 216         char            vl_pad1[VDEV_PAD_SIZE];                 /*  8K */
 217         char            vl_pad2[VDEV_PAD_SIZE];                 /*  8K */
 218         vdev_phys_t     vl_vdev_phys;                           /* 112K */
 219         char            vl_uberblock[VDEV_UBERBLOCK_RING];      /* 128K */
 220 } vdev_label_t;                                                 /* 256K total */
 221
 222 /*
 223  * vdev_dirty() flags
 224  */
 225 #define VDD_METASLAB    0x01
 226 #define VDD_DTL         0x02
 227
 228 /*
 229  * Size and offset of embedded boot loader region on each label.
 230  * The total size of the first two labels plus the boot area is 4MB.
 231  */
 232 #define VDEV_BOOT_OFFSET        (2 * sizeof (vdev_label_t))
 233 #define VDEV_BOOT_SIZE          (7ULL << 19)                    /* 3.5M */
 234
 235 /*
 236  * Size of label regions at the start and end of each leaf device.
 237  */
 238 #define VDEV_LABEL_START_SIZE   (2 * sizeof (vdev_label_t) + VDEV_BOOT_SIZE)
 239 #define VDEV_LABEL_END_SIZE     (2 * sizeof (vdev_label_t))
 240 #define VDEV_LABELS             4
 241
 242 #define VDEV_ALLOC_LOAD         0
 243 #define VDEV_ALLOC_ADD          1
 244 #define VDEV_ALLOC_SPARE        2
 245 #define VDEV_ALLOC_L2CACHE      3
 246 #define VDEV_ALLOC_ROOTPOOL     4
 247
 248 /*
 249  * Allocate or free a vdev
 250  */
 251 extern int vdev_alloc(spa_t *spa, vdev_t **vdp, nvlist_t *config,
 252     vdev_t *parent, uint_t id, int alloctype);
 253 extern void vdev_free(vdev_t *vd);
 254
 255 /*
 256  * Add or remove children and parents
 257  */
 258 extern void vdev_add_child(vdev_t *pvd, vdev_t *cvd);
 259 extern void vdev_remove_child(vdev_t *pvd, vdev_t *cvd);
 260 extern void vdev_compact_children(vdev_t *pvd);
 261 extern vdev_t *vdev_add_parent(vdev_t *cvd, vdev_ops_t *ops);
 262 extern void vdev_remove_parent(vdev_t *cvd);
 263
 264 /*
 265  * vdev sync load and sync
 266  */
 267 extern void vdev_load_log_state(vdev_t *vd, nvlist_t *nv);
 268 extern void vdev_load(vdev_t *vd);
 269 extern void vdev_sync(vdev_t *vd, uint64_t txg);
 270 extern void vdev_sync_done(vdev_t *vd, uint64_t txg);
 271 extern void vdev_dirty(vdev_t *vd, int flags, void *arg, uint64_t txg);
 272
 273 /*
 274  * Available vdev types.
 275  */
 276 extern vdev_ops_t vdev_root_ops;
 277 extern vdev_ops_t vdev_mirror_ops;
 278 extern vdev_ops_t vdev_replacing_ops;
 279 extern vdev_ops_t vdev_raidz_ops;
 280 extern vdev_ops_t vdev_disk_ops;
 281 extern vdev_ops_t vdev_file_ops;
 282 extern vdev_ops_t vdev_missing_ops;
 283 extern vdev_ops_t vdev_spare_ops;
 284
 285 /*
 286  * Common size functions
 287  */
 288 extern uint64_t vdev_default_asize(vdev_t *vd, uint64_t psize);
 289 extern uint64_t vdev_get_min_asize(vdev_t *vd);
 290 extern void vdev_set_min_asize(vdev_t *vd);
 291
 292 /*
 293  * zdb uses this tunable, so it must be declared here to make lint happy.
 294  */
 295 extern int zfs_vdev_cache_size;
 296
 297 #ifdef  __cplusplus
 298 }
 299 #endif
 300
 301 #endif  /* _SYS_VDEV_IMPL_H */