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 __u32 lsm_md_layout_version;
49 __u32 lsm_md_migrate_offset;
50 __u32 lsm_md_migrate_hash;
51 __u32 lsm_md_default_count;
52 __u32 lsm_md_default_index;
53 char lsm_md_pool_name[LOV_MAXPOOLNAME + 1];
54 struct lmv_oinfo lsm_md_oinfo[0];
57 static inline bool lmv_dir_striped(const struct lmv_stripe_md *lsm)
59 return lsm && lsm->lsm_md_magic == LMV_MAGIC;
62 static inline bool lmv_dir_foreign(const struct lmv_stripe_md *lsm)
64 return lsm && lsm->lsm_md_magic == LMV_MAGIC_FOREIGN;
67 static inline bool lmv_dir_layout_changing(const struct lmv_stripe_md *lsm)
69 return lmv_dir_striped(lsm) &&
70 lmv_hash_is_layout_changing(lsm->lsm_md_hash_type);
73 static inline bool lmv_dir_bad_hash(const struct lmv_stripe_md *lsm)
75 if (!lmv_dir_striped(lsm))
78 if (lsm->lsm_md_hash_type & LMV_HASH_FLAG_BAD_TYPE)
81 return !lmv_is_known_hash_type(lsm->lsm_md_hash_type);
85 lsm_md_eq(const struct lmv_stripe_md *lsm1, const struct lmv_stripe_md *lsm2)
89 if (lsm1->lsm_md_magic != lsm2->lsm_md_magic ||
90 lsm1->lsm_md_stripe_count != lsm2->lsm_md_stripe_count ||
91 lsm1->lsm_md_master_mdt_index !=
92 lsm2->lsm_md_master_mdt_index ||
93 lsm1->lsm_md_hash_type != lsm2->lsm_md_hash_type ||
94 lsm1->lsm_md_layout_version !=
95 lsm2->lsm_md_layout_version ||
96 lsm1->lsm_md_migrate_offset !=
97 lsm2->lsm_md_migrate_offset ||
98 lsm1->lsm_md_migrate_hash !=
99 lsm2->lsm_md_migrate_hash ||
100 strncmp(lsm1->lsm_md_pool_name, lsm2->lsm_md_pool_name,
101 sizeof(lsm1->lsm_md_pool_name)) != 0)
104 if (lmv_dir_striped(lsm1)) {
105 for (idx = 0; idx < lsm1->lsm_md_stripe_count; idx++) {
106 if (!lu_fid_eq(&lsm1->lsm_md_oinfo[idx].lmo_fid,
107 &lsm2->lsm_md_oinfo[idx].lmo_fid))
115 static inline void lsm_md_dump(int mask, const struct lmv_stripe_md *lsm)
119 /* If lsm_md_magic == LMV_MAGIC_FOREIGN pool_name may not be a null
120 * terminated string so only print LOV_MAXPOOLNAME bytes.
123 "magic %#x stripe count %d master mdt %d hash type %#x version %d migrate offset %d migrate hash %#x pool %.*s\n",
124 lsm->lsm_md_magic, lsm->lsm_md_stripe_count,
125 lsm->lsm_md_master_mdt_index, lsm->lsm_md_hash_type,
126 lsm->lsm_md_layout_version, lsm->lsm_md_migrate_offset,
127 lsm->lsm_md_migrate_hash,
128 LOV_MAXPOOLNAME, lsm->lsm_md_pool_name);
130 if (!lmv_dir_striped(lsm))
133 for (i = 0; i < lsm->lsm_md_stripe_count; i++)
134 CDEBUG(mask, "stripe[%d] "DFID"\n",
135 i, PFID(&lsm->lsm_md_oinfo[i].lmo_fid));
140 void lmv_free_memmd(struct lmv_stripe_md *lsm);
142 static inline void lmv1_le_to_cpu(struct lmv_mds_md_v1 *lmv_dst,
143 const struct lmv_mds_md_v1 *lmv_src)
147 lmv_dst->lmv_magic = le32_to_cpu(lmv_src->lmv_magic);
148 lmv_dst->lmv_stripe_count = le32_to_cpu(lmv_src->lmv_stripe_count);
149 lmv_dst->lmv_master_mdt_index =
150 le32_to_cpu(lmv_src->lmv_master_mdt_index);
151 lmv_dst->lmv_hash_type = le32_to_cpu(lmv_src->lmv_hash_type);
152 lmv_dst->lmv_layout_version = le32_to_cpu(lmv_src->lmv_layout_version);
153 for (i = 0; i < lmv_src->lmv_stripe_count; i++)
154 fid_le_to_cpu(&lmv_dst->lmv_stripe_fids[i],
155 &lmv_src->lmv_stripe_fids[i]);
158 static inline void lmv_le_to_cpu(union lmv_mds_md *lmv_dst,
159 const union lmv_mds_md *lmv_src)
161 switch (le32_to_cpu(lmv_src->lmv_magic)) {
163 lmv1_le_to_cpu(&lmv_dst->lmv_md_v1, &lmv_src->lmv_md_v1);
170 /* This hash is only for testing purpose */
171 static inline unsigned int
172 lmv_hash_all_chars(unsigned int count, const char *name, int namelen)
175 const unsigned char *p = (const unsigned char *)name;
177 while (--namelen >= 0)
185 static inline unsigned int
186 lmv_hash_fnv1a(unsigned int count, const char *name, int namelen)
190 hash = lustre_hash_fnv_1a_64(name, namelen);
192 return do_div(hash, count);
196 * Robert Jenkins' function for mixing 32-bit values
197 * http://burtleburtle.net/bob/hash/evahash.html
198 * a, b = random bits, c = input and output
200 * Mixing inputs to generate an evenly distributed hash.
202 #define crush_hashmix(a, b, c) \
204 a = a - b; a = a - c; a = a ^ (c >> 13); \
205 b = b - c; b = b - a; b = b ^ (a << 8); \
206 c = c - a; c = c - b; c = c ^ (b >> 13); \
207 a = a - b; a = a - c; a = a ^ (c >> 12); \
208 b = b - c; b = b - a; b = b ^ (a << 16); \
209 c = c - a; c = c - b; c = c ^ (b >> 5); \
210 a = a - b; a = a - c; a = a ^ (c >> 3); \
211 b = b - c; b = b - a; b = b ^ (a << 10); \
212 c = c - a; c = c - b; c = c ^ (b >> 15); \
215 #define crush_hash_seed 1315423911
217 static inline __u32 crush_hash(__u32 a, __u32 b)
219 __u32 hash = crush_hash_seed ^ a ^ b;
223 crush_hashmix(a, b, hash);
224 crush_hashmix(x, a, hash);
225 crush_hashmix(b, y, hash);
230 /* refer to https://github.com/ceph/ceph/blob/master/src/crush/hash.c and
231 * https://www.ssrc.ucsc.edu/Papers/weil-sc06.pdf for details of CRUSH
234 static inline unsigned int
235 lmv_hash_crush(unsigned int count, const char *name, int namelen)
237 unsigned long long straw;
238 unsigned long long highest_straw = 0;
240 unsigned int idx = 0;
243 /* put temp and backup file on the same MDT where target is located.
244 * temporary file naming rule:
245 * 1. rsync: .<target>.XXXXXX
246 * 2. dstripe: <target>.XXXXXXXX
248 if (lu_name_is_temp_file(name, namelen, true, 6)) {
251 } else if (lu_name_is_temp_file(name, namelen, false, 8)) {
253 } else if (lu_name_is_backup_file(name, namelen, &i)) {
254 LASSERT(i < namelen);
258 pg_id = lmv_hash_fnv1a(LMV_CRUSH_PG_COUNT, name, namelen);
260 /* distribute PG among all stripes pseudo-randomly, so they are almost
261 * evenly distributed, and when stripe count changes, only (delta /
262 * total) sub files need to be moved, herein 'delta' is added or removed
263 * stripe count, 'total' is total stripe count before change for
264 * removal, or count after change for addition.
266 for (i = 0; i < count; i++) {
267 straw = crush_hash(pg_id, i);
268 if (straw > highest_straw) {
269 highest_straw = straw;
273 LASSERT(idx < count);
278 /* directory layout may change in three ways:
279 * 1. directory migration, in its LMV source stripes are appended after
280 * target stripes, \a migrate_hash is source hash type, \a migrate_offset is
281 * target stripe count,
282 * 2. directory split, \a migrate_hash is hash type before split,
283 * \a migrate_offset is stripe count before split.
284 * 3. directory merge, \a migrate_hash is hash type after merge,
285 * \a migrate_offset is stripe count after merge.
288 __lmv_name_to_stripe_index(__u32 hash_type, __u32 stripe_count,
289 __u32 migrate_hash, __u32 migrate_offset,
290 const char *name, int namelen, bool new_layout)
292 __u32 saved_hash = hash_type;
293 __u32 saved_count = stripe_count;
294 int stripe_index = 0;
296 LASSERT(namelen > 0);
297 LASSERT(stripe_count > 0);
299 if (lmv_hash_is_splitting(hash_type)) {
301 hash_type = migrate_hash;
302 stripe_count = migrate_offset;
304 } else if (lmv_hash_is_merging(hash_type)) {
306 hash_type = migrate_hash;
307 stripe_count = migrate_offset;
309 } else if (lmv_hash_is_migrating(hash_type)) {
311 stripe_count = migrate_offset;
313 hash_type = migrate_hash;
314 stripe_count -= migrate_offset;
318 if (stripe_count > 1) {
319 switch (hash_type & LMV_HASH_TYPE_MASK) {
320 case LMV_HASH_TYPE_ALL_CHARS:
321 stripe_index = lmv_hash_all_chars(stripe_count, name,
324 case LMV_HASH_TYPE_FNV_1A_64:
325 stripe_index = lmv_hash_fnv1a(stripe_count, name,
328 case LMV_HASH_TYPE_CRUSH:
329 stripe_index = lmv_hash_crush(stripe_count, name,
337 LASSERT(stripe_index < stripe_count);
339 if (!new_layout && lmv_hash_is_migrating(saved_hash))
340 stripe_index += migrate_offset;
342 LASSERT(stripe_index < saved_count);
344 CDEBUG(D_INFO, "name %.*s hash=%#x/%#x idx=%d/%u/%u under %s layout\n",
345 namelen, name, saved_hash, migrate_hash, stripe_index,
346 saved_count, migrate_offset, new_layout ? "new" : "old");
351 static inline int lmv_name_to_stripe_index(struct lmv_mds_md_v1 *lmv,
352 const char *name, int namelen)
354 if (lmv->lmv_magic == LMV_MAGIC_V1)
355 return __lmv_name_to_stripe_index(lmv->lmv_hash_type,
356 lmv->lmv_stripe_count,
357 lmv->lmv_migrate_hash,
358 lmv->lmv_migrate_offset,
359 name, namelen, true);
361 if (lmv->lmv_magic == cpu_to_le32(LMV_MAGIC_V1))
362 return __lmv_name_to_stripe_index(
363 le32_to_cpu(lmv->lmv_hash_type),
364 le32_to_cpu(lmv->lmv_stripe_count),
365 le32_to_cpu(lmv->lmv_migrate_hash),
366 le32_to_cpu(lmv->lmv_migrate_offset),
367 name, namelen, true);
372 static inline int lmv_name_to_stripe_index_old(struct lmv_mds_md_v1 *lmv,
373 const char *name, int namelen)
375 if (lmv->lmv_magic == LMV_MAGIC_V1 ||
376 lmv->lmv_magic == LMV_MAGIC_STRIPE)
377 return __lmv_name_to_stripe_index(lmv->lmv_hash_type,
378 lmv->lmv_stripe_count,
379 lmv->lmv_migrate_hash,
380 lmv->lmv_migrate_offset,
381 name, namelen, false);
383 if (lmv->lmv_magic == cpu_to_le32(LMV_MAGIC_V1) ||
384 lmv->lmv_magic == cpu_to_le32(LMV_MAGIC_STRIPE))
385 return __lmv_name_to_stripe_index(
386 le32_to_cpu(lmv->lmv_hash_type),
387 le32_to_cpu(lmv->lmv_stripe_count),
388 le32_to_cpu(lmv->lmv_migrate_hash),
389 le32_to_cpu(lmv->lmv_migrate_offset),
390 name, namelen, false);
395 static inline bool lmv_user_magic_supported(__u32 lum_magic)
397 return lum_magic == LMV_USER_MAGIC ||
398 lum_magic == LMV_USER_MAGIC_SPECIFIC ||
399 lum_magic == LMV_MAGIC_FOREIGN;
402 /* master LMV is sane */
403 static inline bool lmv_is_sane(const struct lmv_mds_md_v1 *lmv)
408 if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1)
411 if (le32_to_cpu(lmv->lmv_stripe_count) == 0)
414 if (!lmv_is_known_hash_type(le32_to_cpu(lmv->lmv_hash_type)))
419 LMV_DEBUG(D_ERROR, lmv, "insane");
423 /* LMV can be either master or stripe LMV */
424 static inline bool lmv_is_sane2(const struct lmv_mds_md_v1 *lmv)
429 if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1 &&
430 le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_STRIPE)
433 if (le32_to_cpu(lmv->lmv_stripe_count) == 0)
436 if (!lmv_is_known_hash_type(le32_to_cpu(lmv->lmv_hash_type)))
441 LMV_DEBUG(D_ERROR, lmv, "insane");
445 static inline bool lmv_is_splitting(const struct lmv_mds_md_v1 *lmv)
447 LASSERT(lmv_is_sane2(lmv));
448 return lmv_hash_is_splitting(cpu_to_le32(lmv->lmv_hash_type));
451 static inline bool lmv_is_merging(const struct lmv_mds_md_v1 *lmv)
453 LASSERT(lmv_is_sane2(lmv));
454 return lmv_hash_is_merging(cpu_to_le32(lmv->lmv_hash_type));
457 static inline bool lmv_is_migrating(const struct lmv_mds_md_v1 *lmv)
459 LASSERT(lmv_is_sane(lmv));
460 return lmv_hash_is_migrating(cpu_to_le32(lmv->lmv_hash_type));
463 static inline bool lmv_is_restriping(const struct lmv_mds_md_v1 *lmv)
465 LASSERT(lmv_is_sane2(lmv));
466 return lmv_hash_is_splitting(cpu_to_le32(lmv->lmv_hash_type)) ||
467 lmv_hash_is_merging(cpu_to_le32(lmv->lmv_hash_type));
470 static inline bool lmv_is_layout_changing(const struct lmv_mds_md_v1 *lmv)
472 LASSERT(lmv_is_sane2(lmv));
473 return lmv_hash_is_splitting(cpu_to_le32(lmv->lmv_hash_type)) ||
474 lmv_hash_is_merging(cpu_to_le32(lmv->lmv_hash_type)) ||
475 lmv_hash_is_migrating(cpu_to_le32(lmv->lmv_hash_type));