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, 2018 by Delphix. All rights reserved.
24  * Copyright (c) 2011, 2019 by Delphix. All rights reserved.
25  * Copyright (c) 2017, Intel Corporation.
26  * Copyright 2019 Joyent, Inc.
27  */
28 
29 #ifndef _SYS_VDEV_IMPL_H
30 #define	_SYS_VDEV_IMPL_H
31 
32 #include <sys/avl.h>
33 #include <sys/bpobj.h>
34 #include <sys/dmu.h>
35 #include <sys/metaslab.h>
36 #include <sys/nvpair.h>
37 #include <sys/space_map.h>
38 #include <sys/vdev.h>
39 #include <sys/dkio.h>
40 #include <sys/uberblock_impl.h>
41 #include <sys/vdev_indirect_mapping.h>
42 #include <sys/vdev_indirect_births.h>
43 #include <sys/vdev_removal.h>
44 
45 #ifdef	__cplusplus
46 extern "C" {
47 #endif
48 
49 /*
50  * Virtual device descriptors.
51  *
52  * All storage pool operations go through the virtual device framework,
53  * which provides data replication and I/O scheduling.
54  */
55 
56 /*
57  * Forward declarations that lots of things need.
58  */
59 typedef struct vdev_queue vdev_queue_t;
60 typedef struct vdev_cache vdev_cache_t;
61 typedef struct vdev_cache_entry vdev_cache_entry_t;
62 struct abd;
63 
64 extern int zfs_vdev_queue_depth_pct;
65 extern int zfs_vdev_def_queue_depth;
66 extern uint32_t zfs_vdev_async_write_max_active;
67 
68 /*
69  * Virtual device operations
70  */
71 typedef int	vdev_open_func_t(vdev_t *vd, uint64_t *size, uint64_t *max_size,
72     uint64_t *ashift);
73 typedef void	vdev_close_func_t(vdev_t *vd);
74 typedef uint64_t vdev_asize_func_t(vdev_t *vd, uint64_t psize);
75 typedef void	vdev_io_start_func_t(zio_t *zio);
76 typedef void	vdev_io_done_func_t(zio_t *zio);
77 typedef void	vdev_state_change_func_t(vdev_t *vd, int, int);
78 typedef boolean_t vdev_need_resilver_func_t(vdev_t *vd, uint64_t, size_t);
79 typedef void	vdev_hold_func_t(vdev_t *vd);
80 typedef void	vdev_rele_func_t(vdev_t *vd);
81 
82 typedef void	vdev_remap_cb_t(uint64_t inner_offset, vdev_t *vd,
83     uint64_t offset, uint64_t size, void *arg);
84 typedef void	vdev_remap_func_t(vdev_t *vd, uint64_t offset, uint64_t size,
85     vdev_remap_cb_t callback, void *arg);
86 typedef int	vdev_dumpio_func_t(vdev_t *vd, caddr_t data, size_t size,
87     uint64_t offset, uint64_t origoffset, boolean_t doread, boolean_t isdump);
88 /*
89  * Given a target vdev, translates the logical range "in" to the physical
90  * range "res"
91  */
92 typedef void vdev_xlation_func_t(vdev_t *cvd, const range_seg64_t *in,
93     range_seg64_t *res);
94 
95 typedef struct vdev_ops {
96 	vdev_open_func_t		*vdev_op_open;
97 	vdev_close_func_t		*vdev_op_close;
98 	vdev_asize_func_t		*vdev_op_asize;
99 	vdev_io_start_func_t		*vdev_op_io_start;
100 	vdev_io_done_func_t		*vdev_op_io_done;
101 	vdev_state_change_func_t	*vdev_op_state_change;
102 	vdev_need_resilver_func_t	*vdev_op_need_resilver;
103 	vdev_hold_func_t		*vdev_op_hold;
104 	vdev_rele_func_t		*vdev_op_rele;
105 	vdev_remap_func_t		*vdev_op_remap;
106 	/*
107 	 * For translating ranges from non-leaf vdevs (e.g. raidz) to leaves.
108 	 * Used when initializing vdevs. Isn't used by leaf ops.
109 	 */
110 	vdev_xlation_func_t		*vdev_op_xlate;
111 	vdev_dumpio_func_t		*vdev_op_dumpio;
112 	char				vdev_op_type[16];
113 	boolean_t			vdev_op_leaf;
114 } vdev_ops_t;
115 
116 /*
117  * Virtual device properties
118  */
119 struct vdev_cache_entry {
120 	struct abd	*ve_abd;
121 	uint64_t	ve_offset;
122 	uint64_t	ve_lastused;
123 	avl_node_t	ve_offset_node;
124 	avl_node_t	ve_lastused_node;
125 	uint32_t	ve_hits;
126 	uint16_t	ve_missed_update;
127 	zio_t		*ve_fill_io;
128 };
129 
130 struct vdev_cache {
131 	avl_tree_t	vc_offset_tree;
132 	avl_tree_t	vc_lastused_tree;
133 	kmutex_t	vc_lock;
134 };
135 
136 typedef struct vdev_queue_class {
137 	uint32_t	vqc_active;
138 
139 	/*
140 	 * Sorted by offset or timestamp, depending on if the queue is
141 	 * LBA-ordered vs FIFO.
142 	 */
143 	avl_tree_t	vqc_queued_tree;
144 } vdev_queue_class_t;
145 
146 struct vdev_queue {
147 	vdev_t		*vq_vdev;
148 	vdev_queue_class_t vq_class[ZIO_PRIORITY_NUM_QUEUEABLE];
149 	avl_tree_t	vq_active_tree;
150 	avl_tree_t	vq_read_offset_tree;
151 	avl_tree_t	vq_write_offset_tree;
152 	avl_tree_t	vq_trim_offset_tree;
153 	uint64_t	vq_last_offset;
154 	hrtime_t	vq_io_complete_ts; /* time last i/o completed */
155 	kmutex_t	vq_lock;
156 };
157 
158 typedef enum vdev_alloc_bias {
159 	VDEV_BIAS_NONE,
160 	VDEV_BIAS_LOG,		/* dedicated to ZIL data (SLOG) */
161 	VDEV_BIAS_SPECIAL,	/* dedicated to ddt, metadata, and small blks */
162 	VDEV_BIAS_DEDUP		/* dedicated to dedup metadata */
163 } vdev_alloc_bias_t;
164 
165 
166 /*
167  * On-disk indirect vdev state.
168  *
169  * An indirect vdev is described exclusively in the MOS config of a pool.
170  * The config for an indirect vdev includes several fields, which are
171  * accessed in memory by a vdev_indirect_config_t.
172  */
173 typedef struct vdev_indirect_config {
174 	/*
175 	 * Object (in MOS) which contains the indirect mapping. This object
176 	 * contains an array of vdev_indirect_mapping_entry_phys_t ordered by
177 	 * vimep_src. The bonus buffer for this object is a
178 	 * vdev_indirect_mapping_phys_t. This object is allocated when a vdev
179 	 * removal is initiated.
180 	 *
181 	 * Note that this object can be empty if none of the data on the vdev
182 	 * has been copied yet.
183 	 */
184 	uint64_t	vic_mapping_object;
185 
186 	/*
187 	 * Object (in MOS) which contains the birth times for the mapping
188 	 * entries. This object contains an array of
189 	 * vdev_indirect_birth_entry_phys_t sorted by vibe_offset. The bonus
190 	 * buffer for this object is a vdev_indirect_birth_phys_t. This object
191 	 * is allocated when a vdev removal is initiated.
192 	 *
193 	 * Note that this object can be empty if none of the vdev has yet been
194 	 * copied.
195 	 */
196 	uint64_t	vic_births_object;
197 
198 	/*
199 	 * This is the vdev ID which was removed previous to this vdev, or
200 	 * UINT64_MAX if there are no previously removed vdevs.
201 	 */
202 	uint64_t	vic_prev_indirect_vdev;
203 } vdev_indirect_config_t;
204 
205 /*
206  * Virtual device descriptor
207  */
208 struct vdev {
209 	/*
210 	 * Common to all vdev types.
211 	 */
212 	uint64_t	vdev_id;	/* child number in vdev parent	*/
213 	uint64_t	vdev_guid;	/* unique ID for this vdev	*/
214 	uint64_t	vdev_guid_sum;	/* self guid + all child guids	*/
215 	uint64_t	vdev_orig_guid;	/* orig. guid prior to remove	*/
216 	uint64_t	vdev_asize;	/* allocatable device capacity	*/
217 	uint64_t	vdev_min_asize;	/* min acceptable asize		*/
218 	uint64_t	vdev_max_asize;	/* max acceptable asize		*/
219 	uint64_t	vdev_ashift;	/* block alignment shift	*/
220 	uint64_t	vdev_state;	/* see VDEV_STATE_* #defines	*/
221 	uint64_t	vdev_prevstate;	/* used when reopening a vdev	*/
222 	vdev_ops_t	*vdev_ops;	/* vdev operations		*/
223 	spa_t		*vdev_spa;	/* spa for this vdev		*/
224 	void		*vdev_tsd;	/* type-specific data		*/
225 	vnode_t		*vdev_name_vp;	/* vnode for pathname		*/
226 	vnode_t		*vdev_devid_vp;	/* vnode for devid		*/
227 	vdev_t		*vdev_top;	/* top-level vdev		*/
228 	vdev_t		*vdev_parent;	/* parent vdev			*/
229 	vdev_t		**vdev_child;	/* array of children		*/
230 	uint64_t	vdev_children;	/* number of children		*/
231 	vdev_stat_t	vdev_stat;	/* virtual device statistics	*/
232 	boolean_t	vdev_expanding;	/* expand the vdev?		*/
233 	boolean_t	vdev_reopening;	/* reopen in progress?		*/
234 	boolean_t	vdev_nonrot;	/* true if solid state		*/
235 	int		vdev_open_error; /* error on last open		*/
236 	kthread_t	*vdev_open_thread; /* thread opening children	*/
237 	uint64_t	vdev_crtxg;	/* txg when top-level was added */
238 
239 	/*
240 	 * Top-level vdev state.
241 	 */
242 	uint64_t	vdev_ms_array;	/* metaslab array object	*/
243 	uint64_t	vdev_ms_shift;	/* metaslab size shift		*/
244 	uint64_t	vdev_ms_count;	/* number of metaslabs		*/
245 	metaslab_group_t *vdev_mg;	/* metaslab group		*/
246 	metaslab_t	**vdev_ms;	/* metaslab array		*/
247 	txg_list_t	vdev_ms_list;	/* per-txg dirty metaslab lists	*/
248 	txg_list_t	vdev_dtl_list;	/* per-txg dirty DTL lists	*/
249 	txg_node_t	vdev_txg_node;	/* per-txg dirty vdev linkage	*/
250 	boolean_t	vdev_remove_wanted; /* async remove wanted?	*/
251 	boolean_t	vdev_probe_wanted; /* async probe wanted?	*/
252 	list_node_t	vdev_config_dirty_node; /* config dirty list	*/
253 	list_node_t	vdev_state_dirty_node; /* state dirty list	*/
254 	uint64_t	vdev_deflate_ratio; /* deflation ratio (x512)	*/
255 	uint64_t	vdev_islog;	/* is an intent log device	*/
256 	uint64_t	vdev_removing;	/* device is being removed?	*/
257 	boolean_t	vdev_ishole;	/* is a hole in the namespace	*/
258 	uint64_t	vdev_top_zap;
259 	vdev_alloc_bias_t vdev_alloc_bias; /* metaslab allocation bias	*/
260 
261 	/* pool checkpoint related */
262 	space_map_t	*vdev_checkpoint_sm;	/* contains reserved blocks */
263 
264 	/* Initialize related */
265 	boolean_t	vdev_initialize_exit_wanted;
266 	vdev_initializing_state_t	vdev_initialize_state;
267 	list_node_t	vdev_initialize_node;
268 	kthread_t	*vdev_initialize_thread;
269 	/* Protects vdev_initialize_thread and vdev_initialize_state. */
270 	kmutex_t	vdev_initialize_lock;
271 	kcondvar_t	vdev_initialize_cv;
272 	uint64_t	vdev_initialize_offset[TXG_SIZE];
273 	uint64_t	vdev_initialize_last_offset;
274 	range_tree_t	*vdev_initialize_tree;	/* valid while initializing */
275 	uint64_t	vdev_initialize_bytes_est;
276 	uint64_t	vdev_initialize_bytes_done;
277 	time_t		vdev_initialize_action_time;	/* start and end time */
278 
279 	/* TRIM related */
280 	boolean_t	vdev_trim_exit_wanted;
281 	boolean_t	vdev_autotrim_exit_wanted;
282 	vdev_trim_state_t	vdev_trim_state;
283 	list_node_t	vdev_trim_node;
284 	kmutex_t	vdev_autotrim_lock;
285 	kcondvar_t	vdev_autotrim_cv;
286 	kthread_t	*vdev_autotrim_thread;
287 	/* Protects vdev_trim_thread and vdev_trim_state. */
288 	kmutex_t	vdev_trim_lock;
289 	kcondvar_t	vdev_trim_cv;
290 	kthread_t	*vdev_trim_thread;
291 	uint64_t	vdev_trim_offset[TXG_SIZE];
292 	uint64_t	vdev_trim_last_offset;
293 	uint64_t	vdev_trim_bytes_est;
294 	uint64_t	vdev_trim_bytes_done;
295 	uint64_t	vdev_trim_rate;		/* requested rate (bytes/sec) */
296 	uint64_t	vdev_trim_partial;	/* requested partial TRIM */
297 	uint64_t	vdev_trim_secure;	/* requested secure TRIM */
298 	time_t		vdev_trim_action_time;	/* start and end time */
299 
300 	/* The following is not in ZoL, but used for auto-trim test progress */
301 	uint64_t	vdev_autotrim_bytes_done;
302 
303 	/* for limiting outstanding I/Os (initialize and TRIM) */
304 	kmutex_t	vdev_initialize_io_lock;
305 	kcondvar_t	vdev_initialize_io_cv;
306 	uint64_t	vdev_initialize_inflight;
307 	kmutex_t	vdev_trim_io_lock;
308 	kcondvar_t	vdev_trim_io_cv;
309 	uint64_t	vdev_trim_inflight[2];
310 
311 	/*
312 	 * Values stored in the config for an indirect or removing vdev.
313 	 */
314 	vdev_indirect_config_t	vdev_indirect_config;
315 
316 	/*
317 	 * The vdev_indirect_rwlock protects the vdev_indirect_mapping
318 	 * pointer from changing on indirect vdevs (when it is condensed).
319 	 * Note that removing (not yet indirect) vdevs have different
320 	 * access patterns (the mapping is not accessed from open context,
321 	 * e.g. from zio_read) and locking strategy (e.g. svr_lock).
322 	 */
323 	krwlock_t vdev_indirect_rwlock;
324 	vdev_indirect_mapping_t *vdev_indirect_mapping;
325 	vdev_indirect_births_t *vdev_indirect_births;
326 
327 	/*
328 	 * In memory data structures used to manage the obsolete sm, for
329 	 * indirect or removing vdevs.
330 	 *
331 	 * The vdev_obsolete_segments is the in-core record of the segments
332 	 * that are no longer referenced anywhere in the pool (due to
333 	 * being freed or remapped and not referenced by any snapshots).
334 	 * During a sync, segments are added to vdev_obsolete_segments
335 	 * via vdev_indirect_mark_obsolete(); at the end of each sync
336 	 * pass, this is appended to vdev_obsolete_sm via
337 	 * vdev_indirect_sync_obsolete().  The vdev_obsolete_lock
338 	 * protects against concurrent modifications of vdev_obsolete_segments
339 	 * from multiple zio threads.
340 	 */
341 	kmutex_t	vdev_obsolete_lock;
342 	range_tree_t	*vdev_obsolete_segments;
343 	space_map_t	*vdev_obsolete_sm;
344 
345 	/*
346 	 * Protects the vdev_scan_io_queue field itself as well as the
347 	 * structure's contents (when present).
348 	 */
349 	kmutex_t	vdev_scan_io_queue_lock;
350 	struct dsl_scan_io_queue	*vdev_scan_io_queue;
351 
352 	/*
353 	 * Leaf vdev state.
354 	 */
355 	range_tree_t	*vdev_dtl[DTL_TYPES]; /* dirty time logs	*/
356 	space_map_t	*vdev_dtl_sm;	/* dirty time log space map	*/
357 	txg_node_t	vdev_dtl_node;	/* per-txg dirty DTL linkage	*/
358 	uint64_t	vdev_dtl_object; /* DTL object			*/
359 	uint64_t	vdev_psize;	/* physical device capacity	*/
360 	uint64_t	vdev_wholedisk;	/* true if this is a whole disk */
361 	uint64_t	vdev_offline;	/* persistent offline state	*/
362 	uint64_t	vdev_faulted;	/* persistent faulted state	*/
363 	uint64_t	vdev_degraded;	/* persistent degraded state	*/
364 	uint64_t	vdev_removed;	/* persistent removed state	*/
365 	uint64_t	vdev_resilver_txg; /* persistent resilvering state */
366 	uint64_t	vdev_nparity;	/* number of parity devices for raidz */
367 	char		*vdev_path;	/* vdev path (if any)		*/
368 	char		*vdev_devid;	/* vdev devid (if any)		*/
369 	char		*vdev_physpath;	/* vdev device path (if any)	*/
370 	char		*vdev_fru;	/* physical FRU location	*/
371 	uint64_t	vdev_not_present; /* not present during import	*/
372 	uint64_t	vdev_unspare;	/* unspare when resilvering done */
373 	boolean_t	vdev_nowritecache; /* true if flushwritecache failed */
374 	boolean_t	vdev_has_trim;	/* TRIM is supported		*/
375 	boolean_t	vdev_has_securetrim; /* secure TRIM is supported */
376 	boolean_t	vdev_checkremove; /* temporary online test	*/
377 	boolean_t	vdev_forcefault; /* force online fault		*/
378 	boolean_t	vdev_splitting;	/* split or repair in progress  */
379 	boolean_t	vdev_delayed_close; /* delayed device close?	*/
380 	boolean_t	vdev_tmpoffline; /* device taken offline temporarily? */
381 	boolean_t	vdev_detached;	/* device detached?		*/
382 	boolean_t	vdev_cant_read;	/* vdev is failing all reads	*/
383 	boolean_t	vdev_cant_write; /* vdev is failing all writes	*/
384 	boolean_t	vdev_isspare;	/* was a hot spare		*/
385 	boolean_t	vdev_isl2cache;	/* was a l2cache device		*/
386 	boolean_t	vdev_resilver_deferred;  /* resilver deferred */
387 	vdev_queue_t	vdev_queue;	/* I/O deadline schedule queue	*/
388 	vdev_cache_t	vdev_cache;	/* physical block cache		*/
389 	spa_aux_vdev_t	*vdev_aux;	/* for l2cache and spares vdevs	*/
390 	zio_t		*vdev_probe_zio; /* root of current probe	*/
391 	vdev_aux_t	vdev_label_aux;	/* on-disk aux state		*/
392 	uint64_t	vdev_leaf_zap;
393 	hrtime_t	vdev_mmp_pending; /* 0 if write finished	*/
394 	uint64_t	vdev_mmp_kstat_id;	/* to find kstat entry */
395 	list_node_t	vdev_leaf_node;		/* leaf vdev list */
396 
397 	/*
398 	 * For DTrace to work in userland (libzpool) context, these fields must
399 	 * remain at the end of the structure.  DTrace will use the kernel's
400 	 * CTF definition for 'struct vdev', and since the size of a kmutex_t is
401 	 * larger in userland, the offsets for the rest of the fields would be
402 	 * incorrect.
403 	 */
404 	kmutex_t	vdev_dtl_lock;	/* vdev_dtl_{map,resilver}	*/
405 	kmutex_t	vdev_stat_lock;	/* vdev_stat			*/
406 	kmutex_t	vdev_probe_lock; /* protects vdev_probe_zio	*/
407 };
408 
409 #define	VDEV_RAIDZ_MAXPARITY	3
410 
411 #define	VDEV_PAD_SIZE		(8 << 10)
412 /* 2 padding areas (vl_pad1 and vl_pad2) to skip */
413 #define	VDEV_SKIP_SIZE		VDEV_PAD_SIZE * 2
414 #define	VDEV_PHYS_SIZE		(112 << 10)
415 #define	VDEV_UBERBLOCK_RING	(128 << 10)
416 
417 /*
418  * MMP blocks occupy the last MMP_BLOCKS_PER_LABEL slots in the uberblock
419  * ring when MMP is enabled.
420  */
421 #define	MMP_BLOCKS_PER_LABEL	1
422 
423 /* The largest uberblock we support is 8k. */
424 #define	MAX_UBERBLOCK_SHIFT (13)
425 #define	VDEV_UBERBLOCK_SHIFT(vd)	\
426 	MIN(MAX((vd)->vdev_top->vdev_ashift, UBERBLOCK_SHIFT), \
427 	    MAX_UBERBLOCK_SHIFT)
428 #define	VDEV_UBERBLOCK_COUNT(vd)	\
429 	(VDEV_UBERBLOCK_RING >> VDEV_UBERBLOCK_SHIFT(vd))
430 #define	VDEV_UBERBLOCK_OFFSET(vd, n)	\
431 	offsetof(vdev_label_t, vl_uberblock[(n) << VDEV_UBERBLOCK_SHIFT(vd)])
432 #define	VDEV_UBERBLOCK_SIZE(vd)		(1ULL << VDEV_UBERBLOCK_SHIFT(vd))
433 
434 typedef struct vdev_phys {
435 	char		vp_nvlist[VDEV_PHYS_SIZE - sizeof (zio_eck_t)];
436 	zio_eck_t	vp_zbt;
437 } vdev_phys_t;
438 
439 typedef struct vdev_label {
440 	char		vl_pad1[VDEV_PAD_SIZE];			/*  8K */
441 	char		vl_pad2[VDEV_PAD_SIZE];			/*  8K */
442 	vdev_phys_t	vl_vdev_phys;				/* 112K	*/
443 	char		vl_uberblock[VDEV_UBERBLOCK_RING];	/* 128K	*/
444 } vdev_label_t;							/* 256K total */
445 
446 /*
447  * vdev_dirty() flags
448  */
449 #define	VDD_METASLAB	0x01
450 #define	VDD_DTL		0x02
451 
452 /* Offset of embedded boot loader region on each label */
453 #define	VDEV_BOOT_OFFSET	(2 * sizeof (vdev_label_t))
454 /*
455  * Size of embedded boot loader region on each label.
456  * The total size of the first two labels plus the boot area is 4MB.
457  */
458 #define	VDEV_BOOT_SIZE		(7ULL << 19)			/* 3.5M */
459 
460 /*
461  * Size of label regions at the start and end of each leaf device.
462  */
463 #define	VDEV_LABEL_START_SIZE	(2 * sizeof (vdev_label_t) + VDEV_BOOT_SIZE)
464 #define	VDEV_LABEL_END_SIZE	(2 * sizeof (vdev_label_t))
465 #define	VDEV_LABELS		4
466 #define	VDEV_BEST_LABEL		VDEV_LABELS
467 
468 #define	VDEV_ALLOC_LOAD		0
469 #define	VDEV_ALLOC_ADD		1
470 #define	VDEV_ALLOC_SPARE	2
471 #define	VDEV_ALLOC_L2CACHE	3
472 #define	VDEV_ALLOC_ROOTPOOL	4
473 #define	VDEV_ALLOC_SPLIT	5
474 #define	VDEV_ALLOC_ATTACH	6
475 
476 /*
477  * Allocate or free a vdev
478  */
479 extern vdev_t *vdev_alloc_common(spa_t *spa, uint_t id, uint64_t guid,
480     vdev_ops_t *ops);
481 extern int vdev_alloc(spa_t *spa, vdev_t **vdp, nvlist_t *config,
482     vdev_t *parent, uint_t id, int alloctype);
483 extern void vdev_free(vdev_t *vd);
484 
485 /*
486  * Add or remove children and parents
487  */
488 extern void vdev_add_child(vdev_t *pvd, vdev_t *cvd);
489 extern void vdev_remove_child(vdev_t *pvd, vdev_t *cvd);
490 extern void vdev_compact_children(vdev_t *pvd);
491 extern vdev_t *vdev_add_parent(vdev_t *cvd, vdev_ops_t *ops);
492 extern void vdev_remove_parent(vdev_t *cvd);
493 
494 /*
495  * vdev sync load and sync
496  */
497 extern boolean_t vdev_log_state_valid(vdev_t *vd);
498 extern int vdev_load(vdev_t *vd);
499 extern int vdev_dtl_load(vdev_t *vd);
500 extern void vdev_sync(vdev_t *vd, uint64_t txg);
501 extern void vdev_sync_done(vdev_t *vd, uint64_t txg);
502 extern void vdev_dirty(vdev_t *vd, int flags, void *arg, uint64_t txg);
503 extern void vdev_dirty_leaves(vdev_t *vd, int flags, uint64_t txg);
504 
505 /*
506  * Available vdev types.
507  */
508 extern vdev_ops_t vdev_root_ops;
509 extern vdev_ops_t vdev_mirror_ops;
510 extern vdev_ops_t vdev_replacing_ops;
511 extern vdev_ops_t vdev_raidz_ops;
512 extern vdev_ops_t vdev_disk_ops;
513 extern vdev_ops_t vdev_file_ops;
514 extern vdev_ops_t vdev_missing_ops;
515 extern vdev_ops_t vdev_hole_ops;
516 extern vdev_ops_t vdev_spare_ops;
517 extern vdev_ops_t vdev_indirect_ops;
518 
519 /*
520  * Common size functions
521  */
522 extern void vdev_default_xlate(vdev_t *vd, const range_seg64_t *in,
523     range_seg64_t *out);
524 extern uint64_t vdev_default_asize(vdev_t *vd, uint64_t psize);
525 extern uint64_t vdev_get_min_asize(vdev_t *vd);
526 extern void vdev_set_min_asize(vdev_t *vd);
527 
528 /*
529  * Global variables
530  */
531 extern int zfs_vdev_standard_sm_blksz;
532 /* zdb uses this tunable, so it must be declared here to make lint happy. */
533 extern int zfs_vdev_cache_size;
534 
535 /*
536  * Functions from vdev_indirect.c
537  */
538 extern void vdev_indirect_sync_obsolete(vdev_t *vd, dmu_tx_t *tx);
539 extern boolean_t vdev_indirect_should_condense(vdev_t *vd);
540 extern void spa_condense_indirect_start_sync(vdev_t *vd, dmu_tx_t *tx);
541 extern int vdev_obsolete_sm_object(vdev_t *vd);
542 extern boolean_t vdev_obsolete_counts_are_precise(vdev_t *vd);
543 
544 /*
545  * Other miscellaneous functions
546  */
547 int vdev_checkpoint_sm_object(vdev_t *vd);
548 
549 /*
550  * The vdev_buf_t is used to translate between zio_t and buf_t, and back again.
551  */
552 typedef struct vdev_buf {
553 	buf_t	vb_buf;		/* buffer that describes the io */
554 	zio_t	*vb_io;		/* pointer back to the original zio_t */
555 } vdev_buf_t;
556 
557 #ifdef	__cplusplus
558 }
559 #endif
560 
561 #endif	/* _SYS_VDEV_IMPL_H */
562