hammer2 - hardlink stabilization (3), data and inode count propagation.
[dragonfly.git] / sys / vfs / hammer2 / hammer2.h
1 /*
2  * Copyright (c) 2011-2012 The DragonFly Project.  All rights reserved.
3  *
4  * This code is derived from software contributed to The DragonFly Project
5  * by Matthew Dillon <dillon@dragonflybsd.org>
6  * by Venkatesh Srinivas <vsrinivas@dragonflybsd.org>
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  *
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in
16  *    the documentation and/or other materials provided with the
17  *    distribution.
18  * 3. Neither the name of The DragonFly Project nor the names of its
19  *    contributors may be used to endorse or promote products derived
20  *    from this software without specific, prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
23  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
24  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
25  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
26  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
27  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
28  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
29  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
30  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
31  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
32  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33  * SUCH DAMAGE.
34  */
35
36 /*
37  * This header file contains structures used internally by the HAMMER2
38  * implementation.  See hammer2_disk.h for on-disk structures.
39  */
40
41 #ifndef _VFS_HAMMER2_HAMMER2_H_
42 #define _VFS_HAMMER2_HAMMER2_H_
43
44 #include <sys/param.h>
45 #include <sys/types.h>
46 #include <sys/kernel.h>
47 #include <sys/conf.h>
48 #include <sys/systm.h>
49 #include <sys/tree.h>
50 #include <sys/malloc.h>
51 #include <sys/mount.h>
52 #include <sys/vnode.h>
53 #include <sys/proc.h>
54 #include <sys/mountctl.h>
55 #include <sys/priv.h>
56 #include <sys/stat.h>
57 #include <sys/globaldata.h>
58 #include <sys/lockf.h>
59 #include <sys/buf.h>
60 #include <sys/queue.h>
61 #include <sys/limits.h>
62 #include <sys/buf2.h>
63 #include <sys/signal2.h>
64 #include <sys/tree.h>
65
66 #include "hammer2_disk.h"
67 #include "hammer2_mount.h"
68 #include "hammer2_ioctl.h"
69
70 struct hammer2_chain;
71 struct hammer2_inode;
72 struct hammer2_mount;
73 struct hammer2_pfsmount;
74
75 /*
76  * The chain structure tracks blockref recursions all the way to
77  * the root volume.  These consist of indirect blocks, inodes,
78  * and eventually the volume header.
79  *
80  * The chain structure is embedded in the hammer2_mount, hammer2_inode,
81  * and other system memory structures.  The chain structure typically
82  * implements the reference count and busy flag for the larger structure.
83  *
84  * It is always possible to track a chain element all the way back to the
85  * root by following the (parent) links.  (index) is a type-dependent index
86  * in the parent indicating where in the parent the chain element resides.
87  *
88  * When a blockref is added or deleted the related chain element is marked
89  * modified and all of its parents are marked SUBMODIFIED (the parent
90  * recursion can stop once we hit a node that is already marked SUBMODIFIED).
91  * A deleted chain element must remain intact until synchronized against
92  * its parent.
93  *
94  * The blockref at (parent, index) is not adjusted until the modified chain
95  * element is flushed and unmarked.  Until then the child's blockref may
96  * not match the blockref at (parent, index).
97  */
98 SPLAY_HEAD(hammer2_chain_splay, hammer2_chain);
99
100 struct hammer2_chain {
101         struct hammer2_blockref bref;
102         struct hammer2_chain *parent;           /* return chain to root */
103         struct hammer2_chain_splay shead;
104         SPLAY_ENTRY(hammer2_chain) snode;
105         TAILQ_ENTRY(hammer2_chain) flush_node;  /* flush deferral list */
106         union {
107                 struct hammer2_inode *ip;
108                 struct hammer2_indblock *np;
109                 struct hammer2_data *dp;
110                 void *mem;
111         } u;
112
113         struct buf      *bp;            /* buffer cache (ro) */
114         hammer2_media_data_t *data;     /* modified copy of data (rw) */
115         u_int           bytes;          /* physical size of data */
116         struct lock     lk;             /* lockmgr lock */
117         int             index;          /* index in parent */
118         u_int           refs;
119         u_int           busy;           /* soft-busy */
120         u_int           flags;
121 };
122
123 typedef struct hammer2_chain hammer2_chain_t;
124
125 int hammer2_chain_cmp(hammer2_chain_t *chain1, hammer2_chain_t *chain2);
126 SPLAY_PROTOTYPE(hammer2_chain_splay, hammer2_chain, snode, hammer2_chain_cmp);
127
128 /*
129  * MOVED - This bit is set during the flush when the MODIFIED bit is cleared,
130  *         indicating that the parent's blocktable must inherit a change to
131  *         the bref (typically a block reallocation)
132  *
133  *         It must also be set in situations where a chain is not MODIFIED
134  *         but whos bref has changed (typically due to fields other than
135  *         a block reallocation).
136  */
137 #define HAMMER2_CHAIN_MODIFIED          0x00000001      /* active mods */
138 #define HAMMER2_CHAIN_DIRTYEMBED        0x00000002      /* inode embedded */
139 #define HAMMER2_CHAIN_DIRTYBP           0x00000004      /* dirty on unlock */
140 #define HAMMER2_CHAIN_SUBMODIFIED       0x00000008      /* 1+ subs modified */
141 #define HAMMER2_CHAIN_DELETED           0x00000010
142 #define HAMMER2_CHAIN_INITIAL           0x00000020      /* initial create */
143 #define HAMMER2_CHAIN_FLUSHED           0x00000040      /* flush on unlock */
144 #define HAMMER2_CHAIN_MOVED             0x00000080      /* bref changed */
145 #define HAMMER2_CHAIN_IOFLUSH           0x00000100      /* bawrite on put */
146 #define HAMMER2_CHAIN_DEFERRED          0x00000200      /* on a deferral list*/
147 #define HAMMER2_CHAIN_DESTROYED         0x00000400      /* destroying */
148 #define HAMMER2_CHAIN_MODIFIED_AUX      0x00000800      /* hmp->vchain only */
149 #define HAMMER2_CHAIN_MODIFY_TID        0x00001000      /* mod updates field */
150 #define HAMMER2_CHAIN_MOUNTED           0x00002000      /* PFS is mounted */
151
152 /*
153  * Flags passed to hammer2_chain_lookup() and hammer2_chain_next()
154  */
155 #define HAMMER2_LOOKUP_NOLOCK           0x00000001      /* ref only */
156 #define HAMMER2_LOOKUP_NODATA           0x00000002      /* data left NULL */
157
158 /*
159  * Flags passed to hammer2_chain_modify() and hammer2_chain_resize()
160  *
161  * NOTE: OPTDATA allows us to avoid instantiating buffers for INDIRECT
162  *       blocks in the INITIAL-create state.
163  *
164  * NOTE: NO_MODIFY_TID tells the function to not set HAMMER2_CHAIN_MODIFY_TID
165  *       when marking the chain modified (used when a sub-chain modification
166  *       propagates upward).
167  */
168 #define HAMMER2_MODIFY_NOSUB            0x00000001      /* do not set SUBMOD */
169 #define HAMMER2_MODIFY_OPTDATA          0x00000002      /* data can be NULL */
170 #define HAMMER2_MODIFY_NO_MODIFY_TID    0x00000004
171
172 /*
173  * Flags passed to hammer2_chain_lock()
174  */
175 #define HAMMER2_RESOLVE_NEVER           1
176 #define HAMMER2_RESOLVE_MAYBE           2
177 #define HAMMER2_RESOLVE_ALWAYS          3
178
179 /*
180  * Cluster different types of storage together for allocations
181  */
182 #define HAMMER2_FREECACHE_INODE         0
183 #define HAMMER2_FREECACHE_INDIR         1
184 #define HAMMER2_FREECACHE_DATA          2
185 #define HAMMER2_FREECACHE_UNUSED3       3
186 #define HAMMER2_FREECACHE_TYPES         4
187
188 /*
189  * BMAP read-ahead maximum parameters
190  */
191 #define HAMMER2_BMAP_COUNT              16      /* max bmap read-ahead */
192 #define HAMMER2_BMAP_BYTES              (HAMMER2_PBUFSIZE * HAMMER2_BMAP_COUNT)
193
194 /*
195  * Misc
196  */
197 #define HAMMER2_FLUSH_DEPTH_LIMIT       40      /* stack recursion limit */
198
199 /*
200  * HAMMER2 IN-MEMORY CACHE OF MEDIA STRUCTURES
201  *
202  * There is an in-memory representation of all on-media data structure.
203  *
204  * When accessed read-only the data will be mapped to the related buffer
205  * cache buffer.
206  *
207  * When accessed read-write (marked modified) a kmalloc()'d copy of the
208  * is created which can then be modified.  The copy is destroyed when a
209  * filesystem block is allocated to replace it.
210  *
211  * Active inodes (those with vnodes attached) will maintain the kmalloc()'d
212  * copy for both the read-only and the read-write case.  The combination of
213  * (bp) and (data) determines whether (data) was allocated or not.
214  *
215  * The in-memory representation may remain cached (for example in order to
216  * placemark clustering locks) even after the related data has been
217  * detached.
218  */
219
220 /*
221  * A hammer2 inode.
222  */
223 struct hammer2_inode {
224         struct hammer2_mount    *hmp;           /* Global mount */
225         struct hammer2_pfsmount *pmp;           /* PFS mount */
226         struct hammer2_inode    *pip;           /* parent inode */
227         struct vnode            *vp;
228         hammer2_chain_t         chain;
229         struct hammer2_inode_data ip_data;
230         struct lockf            advlock;
231         u_int                   depth;          /* directory depth */
232         hammer2_off_t           delta_dcount;   /* adjust data_count */
233         hammer2_off_t           delta_icount;   /* adjust inode_count */
234 };
235
236 typedef struct hammer2_inode hammer2_inode_t;
237
238 /*
239  * A hammer2 indirect block
240  */
241 struct hammer2_indblock {
242         hammer2_chain_t         chain;
243 };
244
245 typedef struct hammer2_indblock hammer2_indblock_t;
246
247 /*
248  * A hammer2 data block
249  */
250 struct hammer2_data {
251         hammer2_chain_t         chain;
252 };
253
254 typedef struct hammer2_data hammer2_data_t;
255
256 /*
257  * Global (per device) mount structure for device (aka vp->v_mount->hmp)
258  */
259 struct hammer2_mount {
260         struct vnode    *devvp;         /* device vnode */
261         int             ronly;          /* read-only mount */
262         int             pmp_count;      /* PFS mounts backed by us */
263         TAILQ_ENTRY(hammer2_mount) mntentry; /* hammer2_mntlist */
264
265         struct malloc_type *minode;
266         int             ninodes;
267         int             maxinodes;
268
269         struct malloc_type *mchain;
270         int             nipstacks;
271         int             maxipstacks;
272         hammer2_chain_t vchain;         /* anchor chain */
273         hammer2_chain_t *schain;        /* super-root */
274         struct lock     alloclk;        /* lockmgr lock */
275         struct lock     voldatalk;      /* lockmgr lock */
276
277         hammer2_volume_data_t voldata;
278         hammer2_off_t   freecache[HAMMER2_FREECACHE_TYPES][HAMMER2_MAX_RADIX+1];
279 };
280
281 typedef struct hammer2_mount hammer2_mount_t;
282
283 /*
284  * Per-PFS mount structure for device (aka vp->v_mount)
285  */
286 struct hammer2_pfsmount {
287         struct mount            *mp;            /* kernel mount */
288         struct hammer2_mount    *hmp;           /* device global mount */
289         hammer2_chain_t         *rchain;        /* PFS root chain */
290         hammer2_inode_t         *iroot;         /* PFS root inode */
291         struct netexport        export;         /* nfs export */
292         int                     ronly;          /* read-only mount */
293 };
294
295 typedef struct hammer2_pfsmount hammer2_pfsmount_t;
296
297 #if defined(_KERNEL)
298
299 MALLOC_DECLARE(M_HAMMER2);
300
301 #define VTOI(vp)        ((hammer2_inode_t *)(vp)->v_data)
302 #define ITOV(ip)        ((ip)->vp)
303
304 static __inline
305 hammer2_pfsmount_t *
306 MPTOPMP(struct mount *mp)
307 {
308         return ((hammer2_pfsmount_t *)mp->mnt_data);
309 }
310
311 static __inline
312 hammer2_mount_t *
313 MPTOHMP(struct mount *mp)
314 {
315         return (((hammer2_pfsmount_t *)mp->mnt_data)->hmp);
316 }
317
318 extern struct vop_ops hammer2_vnode_vops;
319 extern struct vop_ops hammer2_spec_vops;
320 extern struct vop_ops hammer2_fifo_vops;
321
322 extern int hammer2_debug;
323 extern int hammer2_cluster_enable;
324 extern int hammer2_hardlink_enable;
325 extern long hammer2_iod_file_read;
326 extern long hammer2_iod_meta_read;
327 extern long hammer2_iod_indr_read;
328 extern long hammer2_iod_file_write;
329 extern long hammer2_iod_meta_write;
330 extern long hammer2_iod_indr_write;
331 extern long hammer2_iod_volu_write;
332 extern long hammer2_ioa_file_read;
333 extern long hammer2_ioa_meta_read;
334 extern long hammer2_ioa_indr_read;
335 extern long hammer2_ioa_file_write;
336 extern long hammer2_ioa_meta_write;
337 extern long hammer2_ioa_indr_write;
338 extern long hammer2_ioa_volu_write;
339
340 /*
341  * hammer2_subr.c
342  */
343 void hammer2_inode_lock_ex(hammer2_inode_t *ip);
344 void hammer2_inode_unlock_ex(hammer2_inode_t *ip);
345 void hammer2_inode_lock_sh(hammer2_inode_t *ip);
346 void hammer2_inode_unlock_sh(hammer2_inode_t *ip);
347 void hammer2_inode_busy(hammer2_inode_t *ip);
348 void hammer2_inode_unbusy(hammer2_inode_t *ip);
349 void hammer2_voldata_lock(hammer2_mount_t *hmp);
350 void hammer2_voldata_unlock(hammer2_mount_t *hmp);
351
352 void hammer2_mount_exlock(hammer2_mount_t *hmp);
353 void hammer2_mount_shlock(hammer2_mount_t *hmp);
354 void hammer2_mount_unlock(hammer2_mount_t *hmp);
355
356 int hammer2_get_dtype(hammer2_inode_t *ip);
357 int hammer2_get_vtype(hammer2_inode_t *ip);
358 u_int8_t hammer2_get_obj_type(enum vtype vtype);
359 void hammer2_time_to_timespec(u_int64_t xtime, struct timespec *ts);
360 u_int64_t hammer2_timespec_to_time(struct timespec *ts);
361 u_int32_t hammer2_to_unix_xid(uuid_t *uuid);
362 void hammer2_guid_to_uuid(uuid_t *uuid, u_int32_t guid);
363
364 hammer2_key_t hammer2_dirhash(const unsigned char *name, size_t len);
365 int hammer2_bytes_to_radix(size_t bytes);
366
367 int hammer2_calc_logical(hammer2_inode_t *ip, hammer2_off_t uoff,
368                          hammer2_key_t *lbasep, hammer2_key_t *leofp);
369 void hammer2_update_time(uint64_t *timep);
370
371 /*
372  * hammer2_inode.c
373  */
374 struct vnode *hammer2_igetv(hammer2_inode_t *ip, int *errorp);
375
376 void hammer2_inode_lock_nlinks(hammer2_inode_t *ip);
377 void hammer2_inode_unlock_nlinks(hammer2_inode_t *ip);
378 hammer2_inode_t *hammer2_inode_alloc(hammer2_pfsmount_t *pmp, void *data);
379 void hammer2_inode_free(hammer2_inode_t *ip);
380 void hammer2_inode_ref(hammer2_inode_t *ip);
381 void hammer2_inode_drop(hammer2_inode_t *ip);
382 int hammer2_inode_calc_alloc(hammer2_key_t filesize);
383
384 int hammer2_inode_create(hammer2_inode_t *dip,
385                         struct vattr *vap, struct ucred *cred,
386                         const uint8_t *name, size_t name_len,
387                         hammer2_inode_t **nipp);
388
389 int hammer2_inode_duplicate(hammer2_inode_t *dip,
390                         hammer2_inode_t *oip, hammer2_inode_t **nipp,
391                         const uint8_t *name, size_t name_len);
392 int hammer2_inode_connect(hammer2_inode_t *dip, hammer2_inode_t *oip,
393                         const uint8_t *name, size_t name_len);
394
395 int hammer2_unlink_file(hammer2_inode_t *dip,
396                         const uint8_t *name, size_t name_len, int isdir);
397 int hammer2_hardlink_consolidate(hammer2_inode_t **ipp, hammer2_inode_t *tdip);
398 int hammer2_hardlink_deconsolidate(hammer2_inode_t *dip,
399                         hammer2_chain_t **chainp, hammer2_inode_t **ipp);
400 int hammer2_hardlink_find(hammer2_inode_t *dip, hammer2_chain_t **chainp,
401                         hammer2_inode_t **ipp);
402
403 /*
404  * hammer2_chain.c
405  */
406 void hammer2_modify_volume(hammer2_mount_t *hmp);
407 hammer2_chain_t *hammer2_chain_alloc(hammer2_mount_t *hmp,
408                                 hammer2_blockref_t *bref);
409 void hammer2_chain_free(hammer2_mount_t *hmp, hammer2_chain_t *chain);
410 void hammer2_chain_ref(hammer2_mount_t *hmp, hammer2_chain_t *chain);
411 void hammer2_chain_drop(hammer2_mount_t *hmp, hammer2_chain_t *chain);
412 int hammer2_chain_lock(hammer2_mount_t *hmp, hammer2_chain_t *chain, int how);
413 void hammer2_chain_moved(hammer2_mount_t *hmp, hammer2_chain_t *chain);
414 void hammer2_chain_modify(hammer2_mount_t *hmp, hammer2_chain_t *chain,
415                                 int flags);
416 void hammer2_chain_resize(hammer2_inode_t *ip, hammer2_chain_t *chain,
417                                 int nradix, int flags);
418 void hammer2_chain_unlock(hammer2_mount_t *hmp, hammer2_chain_t *chain);
419 hammer2_chain_t *hammer2_chain_find(hammer2_mount_t *hmp,
420                                 hammer2_chain_t *parent, int index);
421 hammer2_chain_t *hammer2_chain_get(hammer2_mount_t *hmp,
422                                 hammer2_chain_t *parent,
423                                 int index, int flags);
424 hammer2_chain_t *hammer2_chain_lookup(hammer2_mount_t *hmp,
425                                 hammer2_chain_t **parentp,
426                                 hammer2_key_t key_beg, hammer2_key_t key_end,
427                                 int flags);
428 hammer2_chain_t *hammer2_chain_next(hammer2_mount_t *hmp,
429                                 hammer2_chain_t **parentp,
430                                 hammer2_chain_t *chain,
431                                 hammer2_key_t key_beg, hammer2_key_t key_end,
432                                 int flags);
433 hammer2_chain_t *hammer2_chain_create(hammer2_mount_t *hmp,
434                                 hammer2_chain_t *parent,
435                                 hammer2_chain_t *chain,
436                                 hammer2_key_t key, int keybits,
437                                 int type, size_t bytes);
438 void hammer2_chain_delete(hammer2_mount_t *hmp, hammer2_chain_t *parent,
439                                 hammer2_chain_t *chain);
440 void hammer2_chain_flush(hammer2_mount_t *hmp, hammer2_chain_t *chain,
441                                 hammer2_tid_t modify_tid);
442 void hammer2_chain_commit(hammer2_mount_t *hmp, hammer2_chain_t *chain);
443
444 /*
445  * hammer2_ioctl.c
446  */
447 int hammer2_ioctl(hammer2_inode_t *ip, u_long com, void *data,
448                                 int fflag, struct ucred *cred);
449
450 /*
451  * hammer2_freemap.c
452  */
453 hammer2_off_t hammer2_freemap_alloc(hammer2_mount_t *hmp,
454                                 int type, size_t bytes);
455
456 #endif /* !_KERNEL */
457 #endif /* !_VFS_HAMMER2_HAMMER2_H_ */