Merge branch 'vendor/FLEX'
[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/thread.h>
58 #include <sys/globaldata.h>
59 #include <sys/lockf.h>
60 #include <sys/buf.h>
61 #include <sys/queue.h>
62 #include <sys/limits.h>
63 #include <sys/buf2.h>
64 #include <sys/signal2.h>
65 #include <sys/tree.h>
66 #include <sys/dmsg.h>
67
68 #include "hammer2_disk.h"
69 #include "hammer2_mount.h"
70 #include "hammer2_ioctl.h"
71 #include "hammer2_ccms.h"
72
73 struct hammer2_chain;
74 struct hammer2_inode;
75 struct hammer2_mount;
76 struct hammer2_pfsmount;
77 struct hammer2_span;
78 struct hammer2_state;
79 struct hammer2_msg;
80
81 /*
82  * The chain structure tracks blockref recursions all the way to
83  * the root volume.  These consist of indirect blocks, inodes,
84  * and eventually the volume header.
85  *
86  * The chain structure is embedded in the hammer2_mount, hammer2_inode,
87  * and other system memory structures.  The chain structure typically
88  * implements the reference count and busy flag for the larger structure.
89  *
90  * It is always possible to track a chain element all the way back to the
91  * root by following the (parent) links.  (index) is a type-dependent index
92  * in the parent indicating where in the parent the chain element resides.
93  *
94  * When a blockref is added or deleted the related chain element is marked
95  * modified and all of its parents are marked SUBMODIFIED (the parent
96  * recursion can stop once we hit a node that is already marked SUBMODIFIED).
97  * A deleted chain element must remain intact until synchronized against
98  * its parent.
99  *
100  * The blockref at (parent, index) is not adjusted until the modified chain
101  * element is flushed and unmarked.  Until then the child's blockref may
102  * not match the blockref at (parent, index).
103  */
104 RB_HEAD(hammer2_chain_tree, hammer2_chain);
105 TAILQ_HEAD(flush_deferral_list, hammer2_chain);
106
107 struct hammer2_chain {
108         ccms_cst_t      cst;                    /* attr or data cst */
109         struct hammer2_blockref bref;
110         struct hammer2_blockref bref_flush;     /* synchronized w/MOVED bit */
111         struct hammer2_chain    *parent;        /* return chain to root */
112         struct hammer2_chain_tree rbhead;
113         struct hammer2_state    *state;         /* if active cache msg */
114         RB_ENTRY(hammer2_chain) rbnode;
115         TAILQ_ENTRY(hammer2_chain) flush_node;  /* flush deferral list */
116         union {
117                 struct hammer2_inode *ip;
118                 struct hammer2_indblock *np;
119                 struct hammer2_data *dp;
120                 void *mem;
121         } u;
122
123         struct buf      *bp;            /* buffer cache (ro) */
124         hammer2_media_data_t *data;     /* modified copy of data (rw) */
125         u_int           bytes;          /* physical size of data */
126         int             index;          /* index in parent */
127         u_int           flushing;       /* element undergoing flush (count) */
128         u_int           refs;
129         u_int           flags;
130 };
131
132 typedef struct hammer2_chain hammer2_chain_t;
133
134 int hammer2_chain_cmp(hammer2_chain_t *chain1, hammer2_chain_t *chain2);
135 RB_PROTOTYPE(hammer2_chain_tree, hammer2_chain, rbnode, hammer2_chain_cmp);
136
137 /*
138  * MOVED - This bit is set during the flush when the MODIFIED bit is cleared,
139  *         indicating that the parent's blocktable must inherit a change to
140  *         the bref (typically a block reallocation)
141  *
142  *         It must also be set in situations where a chain is not MODIFIED
143  *         but whos bref has changed (typically due to fields other than
144  *         a block reallocation).
145  */
146 #define HAMMER2_CHAIN_MODIFIED          0x00000001      /* active mods */
147 #define HAMMER2_CHAIN_DIRTYEMBED        0x00000002      /* inode embedded */
148 #define HAMMER2_CHAIN_DIRTYBP           0x00000004      /* dirty on unlock */
149 #define HAMMER2_CHAIN_SUBMODIFIED       0x00000008      /* 1+ subs modified */
150 #define HAMMER2_CHAIN_DELETED           0x00000010      /* deleted chain */
151 #define HAMMER2_CHAIN_INITIAL           0x00000020      /* initial create */
152 #define HAMMER2_CHAIN_FLUSHED           0x00000040      /* flush on unlock */
153 #define HAMMER2_CHAIN_MOVED             0x00000080      /* bref changed */
154 #define HAMMER2_CHAIN_IOFLUSH           0x00000100      /* bawrite on put */
155 #define HAMMER2_CHAIN_DEFERRED          0x00000200      /* on a deferral list*/
156 #define HAMMER2_CHAIN_DESTROYED         0x00000400      /* destroying inode */
157 #define HAMMER2_CHAIN_MODIFIED_AUX      0x00000800      /* hmp->vchain only */
158 #define HAMMER2_CHAIN_MODIFY_TID        0x00001000      /* mod updates field */
159 #define HAMMER2_CHAIN_MOUNTED           0x00002000      /* PFS is mounted */
160 #define HAMMER2_CHAIN_ONRBTREE          0x00004000      /* on parent RB tree */
161
162 /*
163  * Flags passed to hammer2_chain_lookup() and hammer2_chain_next()
164  */
165 #define HAMMER2_LOOKUP_NOLOCK           0x00000001      /* ref only */
166 #define HAMMER2_LOOKUP_NODATA           0x00000002      /* data left NULL */
167 #define HAMMER2_LOOKUP_SHARED           0x00000100
168
169 /*
170  * Flags passed to hammer2_chain_modify() and hammer2_chain_resize()
171  *
172  * NOTE: OPTDATA allows us to avoid instantiating buffers for INDIRECT
173  *       blocks in the INITIAL-create state.
174  *
175  * NOTE: NO_MODIFY_TID tells the function to not set HAMMER2_CHAIN_MODIFY_TID
176  *       when marking the chain modified (used when a sub-chain modification
177  *       propagates upward).
178  */
179 #define HAMMER2_MODIFY_NOSUB            0x00000001      /* do not set SUBMOD */
180 #define HAMMER2_MODIFY_OPTDATA          0x00000002      /* data can be NULL */
181 #define HAMMER2_MODIFY_NO_MODIFY_TID    0x00000004
182
183 /*
184  * Flags passed to hammer2_chain_lock()
185  */
186 #define HAMMER2_RESOLVE_NEVER           1
187 #define HAMMER2_RESOLVE_MAYBE           2
188 #define HAMMER2_RESOLVE_ALWAYS          3
189 #define HAMMER2_RESOLVE_MASK            0x0F
190
191 #define HAMMER2_RESOLVE_SHARED          0x10
192
193 /*
194  * Cluster different types of storage together for allocations
195  */
196 #define HAMMER2_FREECACHE_INODE         0
197 #define HAMMER2_FREECACHE_INDIR         1
198 #define HAMMER2_FREECACHE_DATA          2
199 #define HAMMER2_FREECACHE_UNUSED3       3
200 #define HAMMER2_FREECACHE_TYPES         4
201
202 /*
203  * BMAP read-ahead maximum parameters
204  */
205 #define HAMMER2_BMAP_COUNT              16      /* max bmap read-ahead */
206 #define HAMMER2_BMAP_BYTES              (HAMMER2_PBUFSIZE * HAMMER2_BMAP_COUNT)
207
208 /*
209  * Misc
210  */
211 #define HAMMER2_FLUSH_DEPTH_LIMIT       40      /* stack recursion limit */
212
213 /*
214  * HAMMER2 IN-MEMORY CACHE OF MEDIA STRUCTURES
215  *
216  * There is an in-memory representation of all on-media data structure.
217  *
218  * When accessed read-only the data will be mapped to the related buffer
219  * cache buffer.
220  *
221  * When accessed read-write (marked modified) a kmalloc()'d copy of the
222  * is created which can then be modified.  The copy is destroyed when a
223  * filesystem block is allocated to replace it.
224  *
225  * Active inodes (those with vnodes attached) will maintain the kmalloc()'d
226  * copy for both the read-only and the read-write case.  The combination of
227  * (bp) and (data) determines whether (data) was allocated or not.
228  *
229  * The in-memory representation may remain cached (for example in order to
230  * placemark clustering locks) even after the related data has been
231  * detached.
232  */
233
234 /*
235  * A hammer2 inode.
236  *
237  * NOTE: The inode's attribute CST which is also used to lock the inode
238  *       is embedded in the chain (chain.cst) and aliased w/ attr_cst.
239  */
240 struct hammer2_inode {
241         struct hammer2_mount    *hmp;           /* Global mount */
242         struct hammer2_pfsmount *pmp;           /* PFS mount */
243         struct hammer2_inode    *pip;           /* parent inode */
244         struct vnode            *vp;
245         ccms_cst_t              topo_cst;       /* directory topology cst */
246         hammer2_chain_t         chain;
247         struct hammer2_inode_data ip_data;
248         struct lockf            advlock;
249         hammer2_off_t           delta_dcount;   /* adjust data_count */
250         hammer2_off_t           delta_icount;   /* adjust inode_count */
251 };
252
253 typedef struct hammer2_inode hammer2_inode_t;
254
255 /*
256  * A hammer2 indirect block
257  */
258 struct hammer2_indblock {
259         hammer2_chain_t         chain;
260 };
261
262 typedef struct hammer2_indblock hammer2_indblock_t;
263
264 /*
265  * A hammer2 data block
266  */
267 struct hammer2_data {
268         hammer2_chain_t         chain;
269 };
270
271 typedef struct hammer2_data hammer2_data_t;
272
273 struct hammer2_freecache {
274         hammer2_off_t   bulk;
275         hammer2_off_t   single;
276 };
277
278 typedef struct hammer2_freecache hammer2_freecache_t;
279
280 /*
281  * Global (per device) mount structure for device (aka vp->v_mount->hmp)
282  */
283 struct hammer2_mount {
284         struct vnode    *devvp;         /* device vnode */
285         int             ronly;          /* read-only mount */
286         int             pmp_count;      /* PFS mounts backed by us */
287         TAILQ_ENTRY(hammer2_mount) mntentry; /* hammer2_mntlist */
288
289         struct malloc_type *minode;
290         int             ninodes;
291         int             maxinodes;
292
293         struct malloc_type *mchain;
294         int             nipstacks;
295         int             maxipstacks;
296         hammer2_chain_t vchain;         /* anchor chain */
297         hammer2_chain_t *schain;        /* super-root */
298         struct lock     alloclk;        /* lockmgr lock */
299         struct lock     voldatalk;      /* lockmgr lock */
300
301         int             volhdrno;       /* last volhdrno written */
302         hammer2_volume_data_t voldata;
303         hammer2_volume_data_t volsync;  /* synchronized voldata */
304         hammer2_freecache_t freecache[HAMMER2_FREECACHE_TYPES]
305                                      [HAMMER2_MAX_RADIX+1];
306 };
307
308 typedef struct hammer2_mount hammer2_mount_t;
309
310 /*
311  * Per-PFS mount structure for device (aka vp->v_mount)
312  */
313 struct hammer2_pfsmount {
314         struct mount            *mp;            /* kernel mount */
315         struct hammer2_mount    *hmp;           /* device global mount */
316         hammer2_chain_t         *rchain;        /* PFS root chain */
317         hammer2_inode_t         *iroot;         /* PFS root inode */
318         ccms_domain_t           ccms_dom;
319         struct netexport        export;         /* nfs export */
320         int                     ronly;          /* read-only mount */
321         struct malloc_type      *mmsg;
322         kdmsg_iocom_t           iocom;
323 };
324
325 typedef struct hammer2_pfsmount hammer2_pfsmount_t;
326
327 #if defined(_KERNEL)
328
329 MALLOC_DECLARE(M_HAMMER2);
330
331 #define VTOI(vp)        ((hammer2_inode_t *)(vp)->v_data)
332 #define ITOV(ip)        ((ip)->vp)
333
334 static __inline
335 hammer2_pfsmount_t *
336 MPTOPMP(struct mount *mp)
337 {
338         return ((hammer2_pfsmount_t *)mp->mnt_data);
339 }
340
341 static __inline
342 hammer2_mount_t *
343 MPTOHMP(struct mount *mp)
344 {
345         return (((hammer2_pfsmount_t *)mp->mnt_data)->hmp);
346 }
347
348 extern struct vop_ops hammer2_vnode_vops;
349 extern struct vop_ops hammer2_spec_vops;
350 extern struct vop_ops hammer2_fifo_vops;
351
352 extern int hammer2_debug;
353 extern int hammer2_cluster_enable;
354 extern int hammer2_hardlink_enable;
355 extern long hammer2_iod_file_read;
356 extern long hammer2_iod_meta_read;
357 extern long hammer2_iod_indr_read;
358 extern long hammer2_iod_file_write;
359 extern long hammer2_iod_meta_write;
360 extern long hammer2_iod_indr_write;
361 extern long hammer2_iod_volu_write;
362 extern long hammer2_ioa_file_read;
363 extern long hammer2_ioa_meta_read;
364 extern long hammer2_ioa_indr_read;
365 extern long hammer2_ioa_file_write;
366 extern long hammer2_ioa_meta_write;
367 extern long hammer2_ioa_indr_write;
368 extern long hammer2_ioa_volu_write;
369
370 /*
371  * hammer2_subr.c
372  */
373 #define hammer2_icrc32(buf, size)       iscsi_crc32((buf), (size))
374 #define hammer2_icrc32c(buf, size, crc) iscsi_crc32_ext((buf), (size), (crc))
375
376 void hammer2_inode_lock_ex(hammer2_inode_t *ip);
377 void hammer2_inode_unlock_ex(hammer2_inode_t *ip);
378 void hammer2_inode_lock_sh(hammer2_inode_t *ip);
379 void hammer2_inode_unlock_sh(hammer2_inode_t *ip);
380 void hammer2_inode_busy(hammer2_inode_t *ip);
381 void hammer2_inode_unbusy(hammer2_inode_t *ip);
382 void hammer2_voldata_lock(hammer2_mount_t *hmp);
383 void hammer2_voldata_unlock(hammer2_mount_t *hmp);
384
385 void hammer2_mount_exlock(hammer2_mount_t *hmp);
386 void hammer2_mount_shlock(hammer2_mount_t *hmp);
387 void hammer2_mount_unlock(hammer2_mount_t *hmp);
388
389 int hammer2_get_dtype(hammer2_inode_t *ip);
390 int hammer2_get_vtype(hammer2_inode_t *ip);
391 u_int8_t hammer2_get_obj_type(enum vtype vtype);
392 void hammer2_time_to_timespec(u_int64_t xtime, struct timespec *ts);
393 u_int64_t hammer2_timespec_to_time(struct timespec *ts);
394 u_int32_t hammer2_to_unix_xid(uuid_t *uuid);
395 void hammer2_guid_to_uuid(uuid_t *uuid, u_int32_t guid);
396
397 hammer2_key_t hammer2_dirhash(const unsigned char *name, size_t len);
398 int hammer2_bytes_to_radix(size_t bytes);
399
400 int hammer2_calc_logical(hammer2_inode_t *ip, hammer2_off_t uoff,
401                          hammer2_key_t *lbasep, hammer2_key_t *leofp);
402 void hammer2_update_time(uint64_t *timep);
403
404 /*
405  * hammer2_inode.c
406  */
407 struct vnode *hammer2_igetv(hammer2_inode_t *ip, int *errorp);
408
409 void hammer2_inode_lock_nlinks(hammer2_inode_t *ip);
410 void hammer2_inode_unlock_nlinks(hammer2_inode_t *ip);
411 hammer2_inode_t *hammer2_inode_alloc(hammer2_pfsmount_t *pmp, void *data);
412 void hammer2_inode_free(hammer2_inode_t *ip);
413 void hammer2_inode_ref(hammer2_inode_t *ip);
414 void hammer2_inode_drop(hammer2_inode_t *ip);
415 int hammer2_inode_calc_alloc(hammer2_key_t filesize);
416
417 int hammer2_inode_create(hammer2_inode_t *dip,
418                         struct vattr *vap, struct ucred *cred,
419                         const uint8_t *name, size_t name_len,
420                         hammer2_inode_t **nipp);
421
422 int hammer2_inode_duplicate(hammer2_inode_t *dip,
423                         hammer2_inode_t *oip, hammer2_inode_t **nipp,
424                         const uint8_t *name, size_t name_len);
425 int hammer2_inode_connect(hammer2_inode_t *dip, hammer2_inode_t *oip,
426                         const uint8_t *name, size_t name_len);
427 hammer2_inode_t *hammer2_inode_common_parent(hammer2_mount_t *hmp,
428                         hammer2_inode_t *fdip, hammer2_inode_t *tdip);
429
430 int hammer2_unlink_file(hammer2_inode_t *dip,
431                         const uint8_t *name, size_t name_len,
432                         int isdir, hammer2_inode_t *retain_ip);
433 int hammer2_hardlink_consolidate(hammer2_inode_t **ipp, hammer2_inode_t *tdip);
434 int hammer2_hardlink_deconsolidate(hammer2_inode_t *dip,
435                         hammer2_chain_t **chainp, hammer2_inode_t **ipp);
436 int hammer2_hardlink_find(hammer2_inode_t *dip, hammer2_chain_t **chainp,
437                         hammer2_inode_t **ipp);
438
439 /*
440  * hammer2_chain.c
441  */
442 void hammer2_modify_volume(hammer2_mount_t *hmp);
443 hammer2_chain_t *hammer2_chain_alloc(hammer2_mount_t *hmp,
444                                 hammer2_blockref_t *bref);
445 void hammer2_chain_free(hammer2_mount_t *hmp, hammer2_chain_t *chain);
446 void hammer2_chain_ref(hammer2_mount_t *hmp, hammer2_chain_t *chain);
447 void hammer2_chain_drop(hammer2_mount_t *hmp, hammer2_chain_t *chain);
448 int hammer2_chain_lock(hammer2_mount_t *hmp, hammer2_chain_t *chain, int how);
449 void hammer2_chain_moved(hammer2_mount_t *hmp, hammer2_chain_t *chain);
450 void hammer2_chain_modify(hammer2_mount_t *hmp, hammer2_chain_t *chain,
451                                 int flags);
452 void hammer2_chain_resize(hammer2_inode_t *ip, hammer2_chain_t *chain,
453                                 int nradix, int flags);
454 void hammer2_chain_unlock(hammer2_mount_t *hmp, hammer2_chain_t *chain);
455 void hammer2_chain_wait(hammer2_mount_t *hmp, hammer2_chain_t *chain);
456 hammer2_chain_t *hammer2_chain_find(hammer2_mount_t *hmp,
457                                 hammer2_chain_t *parent, int index);
458 hammer2_chain_t *hammer2_chain_get(hammer2_mount_t *hmp,
459                                 hammer2_chain_t *parent,
460                                 int index, int flags);
461 hammer2_chain_t *hammer2_chain_lookup(hammer2_mount_t *hmp,
462                                 hammer2_chain_t **parentp,
463                                 hammer2_key_t key_beg, hammer2_key_t key_end,
464                                 int flags);
465 hammer2_chain_t *hammer2_chain_next(hammer2_mount_t *hmp,
466                                 hammer2_chain_t **parentp,
467                                 hammer2_chain_t *chain,
468                                 hammer2_key_t key_beg, hammer2_key_t key_end,
469                                 int flags);
470 hammer2_chain_t *hammer2_chain_create(hammer2_mount_t *hmp,
471                                 hammer2_chain_t *parent,
472                                 hammer2_chain_t *chain,
473                                 hammer2_key_t key, int keybits,
474                                 int type, size_t bytes,
475                                 int *errorp);
476 void hammer2_chain_delete(hammer2_mount_t *hmp, hammer2_chain_t *parent,
477                                 hammer2_chain_t *chain, int retain);
478 void hammer2_chain_flush(hammer2_mount_t *hmp, hammer2_chain_t *chain,
479                                 hammer2_tid_t modify_tid);
480 void hammer2_chain_commit(hammer2_mount_t *hmp, hammer2_chain_t *chain);
481 void hammer2_chain_parent_setsubmod(hammer2_mount_t *hmp,
482                                 hammer2_chain_t *chain);
483
484 /*
485  * hammer2_ioctl.c
486  */
487 int hammer2_ioctl(hammer2_inode_t *ip, u_long com, void *data,
488                                 int fflag, struct ucred *cred);
489
490 /*
491  * hammer2_msgops.c
492  */
493 int hammer2_msg_dbg_rcvmsg(kdmsg_msg_t *msg);
494 int hammer2_msg_adhoc_input(kdmsg_msg_t *msg);
495
496 /*
497  * hammer2_vfsops.c
498  */
499 void hammer2_clusterctl_wakeup(kdmsg_iocom_t *iocom);
500 void hammer2_volconf_update(hammer2_pfsmount_t *pmp, int index);
501 void hammer2_cluster_reconnect(hammer2_pfsmount_t *pmp, struct file *fp);
502
503 /*
504  * hammer2_freemap.c
505  */
506 hammer2_off_t hammer2_freemap_alloc(hammer2_mount_t *hmp,
507                                 int type, size_t bytes);
508 void hammer2_freemap_free(hammer2_mount_t *hmp, hammer2_off_t data_off,
509                                 int type);
510
511 #endif /* !_KERNEL */
512 #endif /* !_VFS_HAMMER2_HAMMER2_H_ */