4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License version 2 for more details. A copy is
14 * included in the COPYING file that accompanied this code.
16 * You should have received a copy of the GNU General Public License
17 * along with this program; if not, write to the Free Software
18 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
23 * Copyright (c) 2014, 2016, Intel Corporation.
26 * lustre/include/lustre_lmv.h
28 * Lustre LMV structures and functions.
30 * Author: Di Wang <di.wang@intel.com>
35 #include <uapi/linux/lustre/lustre_idl.h>
38 struct lu_fid lmo_fid;
40 struct inode *lmo_root;
43 struct lmv_stripe_md {
45 __u32 lsm_md_stripe_count;
46 __u32 lsm_md_master_mdt_index;
47 __u32 lsm_md_hash_type;
48 __u8 lsm_md_max_inherit;
49 __u8 lsm_md_max_inherit_rr;
50 __u32 lsm_md_layout_version;
51 __u32 lsm_md_migrate_offset;
52 __u32 lsm_md_migrate_hash;
53 __u32 lsm_md_default_count;
54 __u32 lsm_md_default_index;
55 char lsm_md_pool_name[LOV_MAXPOOLNAME + 1];
56 struct lmv_oinfo lsm_md_oinfo[0];
59 static inline bool lmv_dir_striped(const struct lmv_stripe_md *lsm)
61 return lsm && lsm->lsm_md_magic == LMV_MAGIC;
64 static inline bool lmv_dir_foreign(const struct lmv_stripe_md *lsm)
66 return lsm && lsm->lsm_md_magic == LMV_MAGIC_FOREIGN;
69 static inline bool lmv_dir_layout_changing(const struct lmv_stripe_md *lsm)
71 return lmv_dir_striped(lsm) &&
72 lmv_hash_is_layout_changing(lsm->lsm_md_hash_type);
75 static inline bool lmv_dir_bad_hash(const struct lmv_stripe_md *lsm)
77 if (!lmv_dir_striped(lsm))
80 if (lsm->lsm_md_hash_type & LMV_HASH_FLAG_BAD_TYPE)
83 return !lmv_is_known_hash_type(lsm->lsm_md_hash_type);
87 lsm_md_eq(const struct lmv_stripe_md *lsm1, const struct lmv_stripe_md *lsm2)
91 if (lsm1->lsm_md_magic != lsm2->lsm_md_magic ||
92 lsm1->lsm_md_stripe_count != lsm2->lsm_md_stripe_count ||
93 lsm1->lsm_md_master_mdt_index !=
94 lsm2->lsm_md_master_mdt_index ||
95 lsm1->lsm_md_hash_type != lsm2->lsm_md_hash_type ||
96 lsm1->lsm_md_layout_version !=
97 lsm2->lsm_md_layout_version ||
98 lsm1->lsm_md_migrate_offset !=
99 lsm2->lsm_md_migrate_offset ||
100 lsm1->lsm_md_migrate_hash !=
101 lsm2->lsm_md_migrate_hash ||
102 strncmp(lsm1->lsm_md_pool_name, lsm2->lsm_md_pool_name,
103 sizeof(lsm1->lsm_md_pool_name)) != 0)
106 if (lmv_dir_striped(lsm1)) {
107 for (idx = 0; idx < lsm1->lsm_md_stripe_count; idx++) {
108 if (!lu_fid_eq(&lsm1->lsm_md_oinfo[idx].lmo_fid,
109 &lsm2->lsm_md_oinfo[idx].lmo_fid))
117 static inline void lsm_md_dump(int mask, const struct lmv_stripe_md *lsm)
121 /* If lsm_md_magic == LMV_MAGIC_FOREIGN pool_name may not be a null
122 * terminated string so only print LOV_MAXPOOLNAME bytes.
125 "magic %#x stripe count %d master mdt %d hash type %#x max inherit %hhu version %d migrate offset %d migrate hash %#x pool %.*s\n",
126 lsm->lsm_md_magic, lsm->lsm_md_stripe_count,
127 lsm->lsm_md_master_mdt_index, lsm->lsm_md_hash_type,
128 lsm->lsm_md_max_inherit, lsm->lsm_md_layout_version,
129 lsm->lsm_md_migrate_offset, lsm->lsm_md_migrate_hash,
130 LOV_MAXPOOLNAME, lsm->lsm_md_pool_name);
132 if (!lmv_dir_striped(lsm))
135 for (i = 0; i < lsm->lsm_md_stripe_count; i++)
136 CDEBUG(mask, "stripe[%d] "DFID"\n",
137 i, PFID(&lsm->lsm_md_oinfo[i].lmo_fid));
142 void lmv_free_memmd(struct lmv_stripe_md *lsm);
144 static inline void lmv1_le_to_cpu(struct lmv_mds_md_v1 *lmv_dst,
145 const struct lmv_mds_md_v1 *lmv_src)
149 lmv_dst->lmv_magic = le32_to_cpu(lmv_src->lmv_magic);
150 lmv_dst->lmv_stripe_count = le32_to_cpu(lmv_src->lmv_stripe_count);
151 lmv_dst->lmv_master_mdt_index =
152 le32_to_cpu(lmv_src->lmv_master_mdt_index);
153 lmv_dst->lmv_hash_type = le32_to_cpu(lmv_src->lmv_hash_type);
154 lmv_dst->lmv_layout_version = le32_to_cpu(lmv_src->lmv_layout_version);
155 if (lmv_src->lmv_stripe_count > LMV_MAX_STRIPE_COUNT)
157 for (i = 0; i < lmv_src->lmv_stripe_count; i++)
158 fid_le_to_cpu(&lmv_dst->lmv_stripe_fids[i],
159 &lmv_src->lmv_stripe_fids[i]);
162 static inline void lmv_le_to_cpu(union lmv_mds_md *lmv_dst,
163 const union lmv_mds_md *lmv_src)
165 switch (le32_to_cpu(lmv_src->lmv_magic)) {
167 lmv1_le_to_cpu(&lmv_dst->lmv_md_v1, &lmv_src->lmv_md_v1);
174 /* This hash is only for testing purpose */
175 static inline unsigned int
176 lmv_hash_all_chars(unsigned int count, const char *name, int namelen)
179 const unsigned char *p = (const unsigned char *)name;
181 while (--namelen >= 0)
189 static inline unsigned int
190 lmv_hash_fnv1a(unsigned int count, const char *name, int namelen)
194 hash = lustre_hash_fnv_1a_64(name, namelen);
196 return do_div(hash, count);
200 * Robert Jenkins' function for mixing 32-bit values
201 * http://burtleburtle.net/bob/hash/evahash.html
202 * a, b = random bits, c = input and output
204 * Mixing inputs to generate an evenly distributed hash.
206 #define crush_hashmix(a, b, c) \
208 a = a - b; a = a - c; a = a ^ (c >> 13); \
209 b = b - c; b = b - a; b = b ^ (a << 8); \
210 c = c - a; c = c - b; c = c ^ (b >> 13); \
211 a = a - b; a = a - c; a = a ^ (c >> 12); \
212 b = b - c; b = b - a; b = b ^ (a << 16); \
213 c = c - a; c = c - b; c = c ^ (b >> 5); \
214 a = a - b; a = a - c; a = a ^ (c >> 3); \
215 b = b - c; b = b - a; b = b ^ (a << 10); \
216 c = c - a; c = c - b; c = c ^ (b >> 15); \
219 #define crush_hash_seed 1315423911
221 static inline __u32 crush_hash(__u32 a, __u32 b)
223 __u32 hash = crush_hash_seed ^ a ^ b;
227 crush_hashmix(a, b, hash);
228 crush_hashmix(x, a, hash);
229 crush_hashmix(b, y, hash);
234 /* refer to https://github.com/ceph/ceph/blob/master/src/crush/hash.c and
235 * https://www.ssrc.ucsc.edu/Papers/weil-sc06.pdf for details of CRUSH
238 static inline unsigned int
239 lmv_hash_crush(unsigned int count, const char *name, int namelen)
241 unsigned long long straw;
242 unsigned long long highest_straw = 0;
244 unsigned int idx = 0;
247 /* put temp and backup file on the same MDT where target is located.
248 * temporary file naming rule:
249 * 1. rsync: .<target>.XXXXXX
250 * 2. dstripe: <target>.XXXXXXXX
252 if (lu_name_is_temp_file(name, namelen, true, 6)) {
255 } else if (lu_name_is_temp_file(name, namelen, false, 8)) {
257 } else if (lu_name_is_backup_file(name, namelen, &i)) {
258 LASSERT(i < namelen);
262 pg_id = lmv_hash_fnv1a(LMV_CRUSH_PG_COUNT, name, namelen);
264 /* distribute PG among all stripes pseudo-randomly, so they are almost
265 * evenly distributed, and when stripe count changes, only (delta /
266 * total) sub files need to be moved, herein 'delta' is added or removed
267 * stripe count, 'total' is total stripe count before change for
268 * removal, or count after change for addition.
270 for (i = 0; i < count; i++) {
271 straw = crush_hash(pg_id, i);
272 if (straw > highest_straw) {
273 highest_straw = straw;
277 LASSERT(idx < count);
282 /* directory layout may change in three ways:
283 * 1. directory migration, in its LMV source stripes are appended after
284 * target stripes, \a migrate_hash is source hash type, \a migrate_offset is
285 * target stripe count,
286 * 2. directory split, \a migrate_hash is hash type before split,
287 * \a migrate_offset is stripe count before split.
288 * 3. directory merge, \a migrate_hash is hash type after merge,
289 * \a migrate_offset is stripe count after merge.
292 __lmv_name_to_stripe_index(__u32 hash_type, __u32 stripe_count,
293 __u32 migrate_hash, __u32 migrate_offset,
294 const char *name, int namelen, bool new_layout)
296 __u32 saved_hash = hash_type;
297 __u32 saved_count = stripe_count;
298 int stripe_index = 0;
300 LASSERT(namelen > 0);
301 LASSERT(stripe_count > 0);
303 if (lmv_hash_is_splitting(hash_type)) {
305 hash_type = migrate_hash;
306 stripe_count = migrate_offset;
308 } else if (lmv_hash_is_merging(hash_type)) {
310 hash_type = migrate_hash;
311 stripe_count = migrate_offset;
313 } else if (lmv_hash_is_migrating(hash_type)) {
315 stripe_count = migrate_offset;
317 hash_type = migrate_hash;
318 stripe_count -= migrate_offset;
322 if (stripe_count > 1) {
323 switch (hash_type & LMV_HASH_TYPE_MASK) {
324 case LMV_HASH_TYPE_ALL_CHARS:
325 stripe_index = lmv_hash_all_chars(stripe_count, name,
328 case LMV_HASH_TYPE_FNV_1A_64:
329 stripe_index = lmv_hash_fnv1a(stripe_count, name,
332 case LMV_HASH_TYPE_CRUSH:
333 stripe_index = lmv_hash_crush(stripe_count, name,
341 LASSERT(stripe_index < stripe_count);
343 if (!new_layout && lmv_hash_is_migrating(saved_hash))
344 stripe_index += migrate_offset;
346 LASSERT(stripe_index < saved_count);
348 CDEBUG(D_INFO, "name %.*s hash=%#x/%#x idx=%d/%u/%u under %s layout\n",
349 namelen, name, saved_hash, migrate_hash, stripe_index,
350 saved_count, migrate_offset, new_layout ? "new" : "old");
355 static inline int lmv_name_to_stripe_index(struct lmv_mds_md_v1 *lmv,
356 const char *name, int namelen)
358 if (lmv->lmv_magic == LMV_MAGIC_V1)
359 return __lmv_name_to_stripe_index(lmv->lmv_hash_type,
360 lmv->lmv_stripe_count,
361 lmv->lmv_migrate_hash,
362 lmv->lmv_migrate_offset,
363 name, namelen, true);
365 if (lmv->lmv_magic == cpu_to_le32(LMV_MAGIC_V1))
366 return __lmv_name_to_stripe_index(
367 le32_to_cpu(lmv->lmv_hash_type),
368 le32_to_cpu(lmv->lmv_stripe_count),
369 le32_to_cpu(lmv->lmv_migrate_hash),
370 le32_to_cpu(lmv->lmv_migrate_offset),
371 name, namelen, true);
376 static inline int lmv_name_to_stripe_index_old(struct lmv_mds_md_v1 *lmv,
377 const char *name, int namelen)
379 if (lmv->lmv_magic == LMV_MAGIC_V1 ||
380 lmv->lmv_magic == LMV_MAGIC_STRIPE)
381 return __lmv_name_to_stripe_index(lmv->lmv_hash_type,
382 lmv->lmv_stripe_count,
383 lmv->lmv_migrate_hash,
384 lmv->lmv_migrate_offset,
385 name, namelen, false);
387 if (lmv->lmv_magic == cpu_to_le32(LMV_MAGIC_V1) ||
388 lmv->lmv_magic == cpu_to_le32(LMV_MAGIC_STRIPE))
389 return __lmv_name_to_stripe_index(
390 le32_to_cpu(lmv->lmv_hash_type),
391 le32_to_cpu(lmv->lmv_stripe_count),
392 le32_to_cpu(lmv->lmv_migrate_hash),
393 le32_to_cpu(lmv->lmv_migrate_offset),
394 name, namelen, false);
399 static inline bool lmv_user_magic_supported(__u32 lum_magic)
401 return lum_magic == LMV_USER_MAGIC ||
402 lum_magic == LMV_USER_MAGIC_SPECIFIC ||
403 lum_magic == LMV_MAGIC_FOREIGN;
406 /* master LMV is sane */
407 static inline bool lmv_is_sane(const struct lmv_mds_md_v1 *lmv)
412 if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1)
415 if (le32_to_cpu(lmv->lmv_stripe_count) == 0)
418 if (!lmv_is_known_hash_type(le32_to_cpu(lmv->lmv_hash_type)))
423 LMV_DEBUG(D_ERROR, lmv, "insane");
427 /* LMV can be either master or stripe LMV */
428 static inline bool lmv_is_sane2(const struct lmv_mds_md_v1 *lmv)
433 if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1 &&
434 le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_STRIPE)
437 if (le32_to_cpu(lmv->lmv_stripe_count) == 0)
440 if (!lmv_is_known_hash_type(le32_to_cpu(lmv->lmv_hash_type)))
445 LMV_DEBUG(D_ERROR, lmv, "insane");
449 static inline bool lmv_is_splitting(const struct lmv_mds_md_v1 *lmv)
451 if (!lmv_is_sane2(lmv))
454 return lmv_hash_is_splitting(cpu_to_le32(lmv->lmv_hash_type));
457 static inline bool lmv_is_merging(const struct lmv_mds_md_v1 *lmv)
459 if (!lmv_is_sane2(lmv))
462 return lmv_hash_is_merging(cpu_to_le32(lmv->lmv_hash_type));
465 static inline bool lmv_is_migrating(const struct lmv_mds_md_v1 *lmv)
467 if (!lmv_is_sane(lmv))
470 return lmv_hash_is_migrating(cpu_to_le32(lmv->lmv_hash_type));
473 static inline bool lmv_is_restriping(const struct lmv_mds_md_v1 *lmv)
475 if (!lmv_is_sane2(lmv))
478 return lmv_hash_is_splitting(cpu_to_le32(lmv->lmv_hash_type)) ||
479 lmv_hash_is_merging(cpu_to_le32(lmv->lmv_hash_type));
482 static inline bool lmv_is_layout_changing(const struct lmv_mds_md_v1 *lmv)
484 if (!lmv_is_sane2(lmv))
487 return lmv_hash_is_splitting(cpu_to_le32(lmv->lmv_hash_type)) ||
488 lmv_hash_is_merging(cpu_to_le32(lmv->lmv_hash_type)) ||
489 lmv_hash_is_migrating(cpu_to_le32(lmv->lmv_hash_type));