]> git.ipfire.org Git - people/ms/linux.git/blame - fs/ceph/mdsmap.c
Merge tag 'trace-v6.0-rc1-2' of git://git.kernel.org/pub/scm/linux/kernel/git/rostedt...
[people/ms/linux.git] / fs / ceph / mdsmap.c
CommitLineData
b2441318 1// SPDX-License-Identifier: GPL-2.0
3d14c5d2 2#include <linux/ceph/ceph_debug.h>
2f2dc053
SW
3
4#include <linux/bug.h>
5#include <linux/err.h>
6#include <linux/random.h>
7#include <linux/slab.h>
8#include <linux/types.h>
9
3d14c5d2
YS
10#include <linux/ceph/mdsmap.h>
11#include <linux/ceph/messenger.h>
12#include <linux/ceph/decode.h>
2f2dc053
SW
13
14#include "super.h"
15
5d47648f 16#define CEPH_MDS_IS_READY(i, ignore_laggy) \
b38c9eb4 17 (m->m_info[i].state > 0 && ignore_laggy ? true : !m->m_info[i].laggy)
2f2dc053 18
5d47648f 19static int __mdsmap_get_random_mds(struct ceph_mdsmap *m, bool ignore_laggy)
2f2dc053
SW
20{
21 int n = 0;
74d6f030 22 int i, j;
a84cd293 23
2f2dc053 24 /* count */
b38c9eb4 25 for (i = 0; i < m->possible_max_rank; i++)
5d47648f 26 if (CEPH_MDS_IS_READY(i, ignore_laggy))
2f2dc053
SW
27 n++;
28 if (n == 0)
29 return -1;
30
31 /* pick */
a84cd293 32 n = prandom_u32() % n;
b38c9eb4 33 for (j = 0, i = 0; i < m->possible_max_rank; i++) {
5d47648f 34 if (CEPH_MDS_IS_READY(i, ignore_laggy))
74d6f030
XL
35 j++;
36 if (j > n)
37 break;
38 }
2f2dc053
SW
39
40 return i;
41}
42
5d47648f
XL
43/*
44 * choose a random mds that is "up" (i.e. has a state > 0), or -1.
45 */
46int ceph_mdsmap_get_random_mds(struct ceph_mdsmap *m)
47{
48 int mds;
49
50 mds = __mdsmap_get_random_mds(m, false);
b38c9eb4 51 if (mds == m->possible_max_rank || mds == -1)
5d47648f
XL
52 mds = __mdsmap_get_random_mds(m, true);
53
b38c9eb4 54 return mds == m->possible_max_rank ? -1 : mds;
5d47648f
XL
55}
56
e9e427f0
YZ
57#define __decode_and_drop_type(p, end, type, bad) \
58 do { \
59 if (*p + sizeof(type) > end) \
60 goto bad; \
61 *p += sizeof(type); \
62 } while (0)
63
64#define __decode_and_drop_set(p, end, type, bad) \
65 do { \
66 u32 n; \
67 size_t need; \
68 ceph_decode_32_safe(p, end, n, bad); \
69 need = sizeof(type) * n; \
70 ceph_decode_need(p, end, need, bad); \
71 *p += need; \
72 } while (0)
73
74#define __decode_and_drop_map(p, end, ktype, vtype, bad) \
75 do { \
76 u32 n; \
77 size_t need; \
78 ceph_decode_32_safe(p, end, n, bad); \
79 need = (sizeof(ktype) + sizeof(vtype)) * n; \
80 ceph_decode_need(p, end, need, bad); \
81 *p += need; \
82 } while (0)
83
84
85static int __decode_and_drop_compat_set(void **p, void* end)
86{
87 int i;
88 /* compat, ro_compat, incompat*/
89 for (i = 0; i < 3; i++) {
90 u32 n;
91 ceph_decode_need(p, end, sizeof(u64) + sizeof(u32), bad);
92 /* mask */
93 *p += sizeof(u64);
94 /* names (map<u64, string>) */
95 n = ceph_decode_32(p);
96 while (n-- > 0) {
97 u32 len;
98 ceph_decode_need(p, end, sizeof(u64) + sizeof(u32),
99 bad);
100 *p += sizeof(u64);
101 len = ceph_decode_32(p);
102 ceph_decode_need(p, end, len, bad);
103 *p += len;
104 }
105 }
106 return 0;
107bad:
108 return -1;
109}
110
2f2dc053
SW
111/*
112 * Decode an MDS map
113 *
114 * Ignore any fields we don't care about (there are quite a few of
115 * them).
116 */
a5cbd5fc 117struct ceph_mdsmap *ceph_mdsmap_decode(void **p, void *end, bool msgr2)
2f2dc053
SW
118{
119 struct ceph_mdsmap *m;
9ec7cab1 120 const void *start = *p;
2f2dc053 121 int i, j, n;
f3848af1 122 int err;
8e298deb 123 u8 mdsmap_v;
e9e427f0 124 u16 mdsmap_ev;
d517b398 125 u32 target;
2f2dc053
SW
126
127 m = kzalloc(sizeof(*m), GFP_NOFS);
d37b1d99 128 if (!m)
2f2dc053
SW
129 return ERR_PTR(-ENOMEM);
130
d463a43d
YZ
131 ceph_decode_need(p, end, 1 + 1, bad);
132 mdsmap_v = ceph_decode_8(p);
8e298deb 133 *p += sizeof(u8); /* mdsmap_cv */
d463a43d
YZ
134 if (mdsmap_v >= 4) {
135 u32 mdsmap_len;
136 ceph_decode_32_safe(p, end, mdsmap_len, bad);
137 if (end < *p + mdsmap_len)
138 goto bad;
139 end = *p + mdsmap_len;
4f6a7e5e 140 }
2f2dc053
SW
141
142 ceph_decode_need(p, end, 8*sizeof(u32) + sizeof(u64), bad);
c89136ea
SW
143 m->m_epoch = ceph_decode_32(p);
144 m->m_client_epoch = ceph_decode_32(p);
145 m->m_last_failure = ceph_decode_32(p);
146 m->m_root = ceph_decode_32(p);
147 m->m_session_timeout = ceph_decode_32(p);
148 m->m_session_autoclose = ceph_decode_32(p);
149 m->m_max_file_size = ceph_decode_64(p);
150 m->m_max_mds = ceph_decode_32(p);
4d7ace02
XL
151
152 /*
b38c9eb4
XL
153 * pick out the active nodes as the m_num_active_mds, the
154 * m_num_active_mds maybe larger than m_max_mds when decreasing
155 * the max_mds in cluster side, in other case it should less
156 * than or equal to m_max_mds.
4d7ace02 157 */
b38c9eb4 158 m->m_num_active_mds = n = ceph_decode_32(p);
4d7ace02
XL
159
160 /*
b38c9eb4 161 * the possible max rank, it maybe larger than the m_num_active_mds,
4d7ace02
XL
162 * for example if the mds_max == 2 in the cluster, when the MDS(0)
163 * was laggy and being replaced by a new MDS, we will temporarily
164 * receive a new mds map with n_num_mds == 1 and the active MDS(1),
b38c9eb4 165 * and the mds rank >= m_num_active_mds.
4d7ace02 166 */
b38c9eb4 167 m->possible_max_rank = max(m->m_num_active_mds, m->m_max_mds);
2f2dc053 168
b38c9eb4 169 m->m_info = kcalloc(m->possible_max_rank, sizeof(*m->m_info), GFP_NOFS);
d37b1d99 170 if (!m->m_info)
e9e427f0 171 goto nomem;
2f2dc053
SW
172
173 /* pick out active nodes from mds_info (state > 0) */
2f2dc053 174 for (i = 0; i < n; i++) {
94045e11 175 u64 global_id;
2f2dc053
SW
176 u32 namelen;
177 s32 mds, inc, state;
d463a43d
YZ
178 u8 info_v;
179 void *info_end = NULL;
2f2dc053
SW
180 struct ceph_entity_addr addr;
181 u32 num_export_targets;
182 void *pexport_targets = NULL;
0deb01c9 183 struct ceph_timespec laggy_since;
6af86528 184 struct ceph_mds_info *info;
da08e1e1 185 bool laggy;
2f2dc053 186
d463a43d 187 ceph_decode_need(p, end, sizeof(u64) + 1, bad);
94045e11 188 global_id = ceph_decode_64(p);
d463a43d
YZ
189 info_v= ceph_decode_8(p);
190 if (info_v >= 4) {
191 u32 info_len;
d463a43d 192 ceph_decode_need(p, end, 1 + sizeof(u32), bad);
8e298deb 193 *p += sizeof(u8); /* info_cv */
d463a43d
YZ
194 info_len = ceph_decode_32(p);
195 info_end = *p + info_len;
196 if (info_end > end)
197 goto bad;
198 }
199
200 ceph_decode_need(p, end, sizeof(u64) + sizeof(u32), bad);
94045e11 201 *p += sizeof(u64);
c89136ea 202 namelen = ceph_decode_32(p); /* skip mds name */
2f2dc053
SW
203 *p += namelen;
204
a5cbd5fc
ID
205 ceph_decode_32_safe(p, end, mds, bad);
206 ceph_decode_32_safe(p, end, inc, bad);
207 ceph_decode_32_safe(p, end, state, bad);
8e298deb 208 *p += sizeof(u64); /* state_seq */
a5cbd5fc
ID
209 if (info_v >= 8)
210 err = ceph_decode_entity_addrvec(p, end, msgr2, &addr);
211 else
212 err = ceph_decode_entity_addr(p, end, &addr);
f3848af1
JL
213 if (err)
214 goto corrupt;
a5cbd5fc
ID
215
216 ceph_decode_copy_safe(p, end, &laggy_since, sizeof(laggy_since),
217 bad);
da08e1e1 218 laggy = laggy_since.tv_sec != 0 || laggy_since.tv_nsec != 0;
2f2dc053
SW
219 *p += sizeof(u32);
220 ceph_decode_32_safe(p, end, namelen, bad);
e251e288 221 *p += namelen;
d463a43d 222 if (info_v >= 2) {
2f2dc053
SW
223 ceph_decode_32_safe(p, end, num_export_targets, bad);
224 pexport_targets = *p;
e251e288 225 *p += num_export_targets * sizeof(u32);
2f2dc053
SW
226 } else {
227 num_export_targets = 0;
228 }
229
d463a43d
YZ
230 if (info_end && *p != info_end) {
231 if (*p > info_end)
232 goto bad;
233 *p = info_end;
234 }
235
da08e1e1 236 dout("mdsmap_decode %d/%d %lld mds%d.%d %s %s%s\n",
3d14c5d2 237 i+1, n, global_id, mds, inc,
b726ec97 238 ceph_pr_addr(&addr),
da08e1e1
XL
239 ceph_mds_state_name(state),
240 laggy ? "(laggy)" : "");
6af86528 241
b38c9eb4 242 if (mds < 0 || mds >= m->possible_max_rank) {
4d7ace02 243 pr_warn("mdsmap_decode got incorrect mds(%d)\n", mds);
6af86528 244 continue;
4d7ace02 245 }
6af86528 246
4d7ace02 247 if (state <= 0) {
ccd1acdf
LH
248 dout("mdsmap_decode got incorrect state(%s)\n",
249 ceph_mds_state_name(state));
4d7ace02 250 continue;
76201b63
YZ
251 }
252
6af86528
DC
253 info = &m->m_info[mds];
254 info->global_id = global_id;
255 info->state = state;
256 info->addr = addr;
da08e1e1 257 info->laggy = laggy;
6af86528
DC
258 info->num_export_targets = num_export_targets;
259 if (num_export_targets) {
260 info->export_targets = kcalloc(num_export_targets,
261 sizeof(u32), GFP_NOFS);
d37b1d99 262 if (!info->export_targets)
e9e427f0 263 goto nomem;
d517b398
XL
264 for (j = 0; j < num_export_targets; j++) {
265 target = ceph_decode_32(&pexport_targets);
d517b398
XL
266 info->export_targets[j] = target;
267 }
6af86528
DC
268 } else {
269 info->export_targets = NULL;
2f2dc053
SW
270 }
271 }
272
273 /* pg_pools */
274 ceph_decode_32_safe(p, end, n, bad);
275 m->m_num_data_pg_pools = n;
4f6a7e5e 276 m->m_data_pg_pools = kcalloc(n, sizeof(u64), GFP_NOFS);
2f2dc053 277 if (!m->m_data_pg_pools)
e9e427f0 278 goto nomem;
4f6a7e5e 279 ceph_decode_need(p, end, sizeof(u64)*(n+1), bad);
2f2dc053 280 for (i = 0; i < n; i++)
4f6a7e5e
SW
281 m->m_data_pg_pools[i] = ceph_decode_64(p);
282 m->m_cas_pg_pool = ceph_decode_64(p);
e9e427f0
YZ
283 m->m_enabled = m->m_epoch > 1;
284
285 mdsmap_ev = 1;
286 if (mdsmap_v >= 2) {
287 ceph_decode_16_safe(p, end, mdsmap_ev, bad_ext);
288 }
289 if (mdsmap_ev >= 3) {
290 if (__decode_and_drop_compat_set(p, end) < 0)
291 goto bad_ext;
292 }
293 /* metadata_pool */
294 if (mdsmap_ev < 5) {
295 __decode_and_drop_type(p, end, u32, bad_ext);
296 } else {
297 __decode_and_drop_type(p, end, u64, bad_ext);
298 }
2f2dc053 299
e9e427f0
YZ
300 /* created + modified + tableserver */
301 __decode_and_drop_type(p, end, struct ceph_timespec, bad_ext);
302 __decode_and_drop_type(p, end, struct ceph_timespec, bad_ext);
303 __decode_and_drop_type(p, end, u32, bad_ext);
304
305 /* in */
306 {
307 int num_laggy = 0;
308 ceph_decode_32_safe(p, end, n, bad_ext);
309 ceph_decode_need(p, end, sizeof(u32) * n, bad_ext);
310
311 for (i = 0; i < n; i++) {
312 s32 mds = ceph_decode_32(p);
b38c9eb4 313 if (mds >= 0 && mds < m->possible_max_rank) {
e9e427f0
YZ
314 if (m->m_info[mds].laggy)
315 num_laggy++;
316 }
317 }
318 m->m_num_laggy = num_laggy;
76201b63 319
b38c9eb4 320 if (n > m->possible_max_rank) {
76201b63
YZ
321 void *new_m_info = krealloc(m->m_info,
322 n * sizeof(*m->m_info),
323 GFP_NOFS | __GFP_ZERO);
324 if (!new_m_info)
325 goto nomem;
326 m->m_info = new_m_info;
327 }
b38c9eb4 328 m->possible_max_rank = n;
e9e427f0
YZ
329 }
330
331 /* inc */
332 __decode_and_drop_map(p, end, u32, u32, bad_ext);
333 /* up */
334 __decode_and_drop_map(p, end, u32, u64, bad_ext);
335 /* failed */
336 __decode_and_drop_set(p, end, u32, bad_ext);
337 /* stopped */
338 __decode_and_drop_set(p, end, u32, bad_ext);
339
340 if (mdsmap_ev >= 4) {
341 /* last_failure_osd_epoch */
342 __decode_and_drop_type(p, end, u32, bad_ext);
343 }
344 if (mdsmap_ev >= 6) {
345 /* ever_allowed_snaps */
346 __decode_and_drop_type(p, end, u8, bad_ext);
347 /* explicitly_allowed_snaps */
348 __decode_and_drop_type(p, end, u8, bad_ext);
349 }
350 if (mdsmap_ev >= 7) {
351 /* inline_data_enabled */
352 __decode_and_drop_type(p, end, u8, bad_ext);
353 }
354 if (mdsmap_ev >= 8) {
e9e427f0
YZ
355 /* enabled */
356 ceph_decode_8_safe(p, end, m->m_enabled, bad_ext);
d93231a6
LH
357 /* fs_name */
358 ceph_decode_skip_string(p, end, bad_ext);
e9e427f0
YZ
359 }
360 /* damaged */
361 if (mdsmap_ev >= 9) {
362 size_t need;
363 ceph_decode_32_safe(p, end, n, bad_ext);
364 need = sizeof(u32) * n;
365 ceph_decode_need(p, end, need, bad_ext);
366 *p += need;
367 m->m_damaged = n > 0;
368 } else {
369 m->m_damaged = false;
370 }
d93231a6
LH
371 if (mdsmap_ev >= 17) {
372 /* balancer */
373 ceph_decode_skip_string(p, end, bad_ext);
374 /* standby_count_wanted */
375 ceph_decode_skip_32(p, end, bad_ext);
376 /* old_max_mds */
377 ceph_decode_skip_32(p, end, bad_ext);
378 /* min_compat_client */
379 ceph_decode_skip_8(p, end, bad_ext);
380 /* required_client_features */
381 ceph_decode_skip_set(p, end, 64, bad_ext);
382 ceph_decode_64_safe(p, end, m->m_max_xattr_size, bad_ext);
383 } else {
384 /* This forces the usage of the (sync) SETXATTR Op */
385 m->m_max_xattr_size = 0;
386 }
e9e427f0 387bad_ext:
da08e1e1
XL
388 dout("mdsmap_decode m_enabled: %d, m_damaged: %d, m_num_laggy: %d\n",
389 !!m->m_enabled, !!m->m_damaged, m->m_num_laggy);
d463a43d 390 *p = end;
2f2dc053
SW
391 dout("mdsmap_decode success epoch %u\n", m->m_epoch);
392 return m;
e9e427f0 393nomem:
2f2dc053 394 err = -ENOMEM;
e9e427f0 395 goto out_err;
f3848af1 396corrupt:
2f2dc053 397 pr_err("corrupt mdsmap\n");
9ec7cab1
SW
398 print_hex_dump(KERN_DEBUG, "mdsmap: ",
399 DUMP_PREFIX_OFFSET, 16, 1,
400 start, end - start, true);
e9e427f0 401out_err:
2f2dc053 402 ceph_mdsmap_destroy(m);
c213b50b 403 return ERR_PTR(err);
f3848af1
JL
404bad:
405 err = -EINVAL;
406 goto corrupt;
2f2dc053
SW
407}
408
409void ceph_mdsmap_destroy(struct ceph_mdsmap *m)
410{
411 int i;
412
a9e6ffbc
TL
413 if (m->m_info) {
414 for (i = 0; i < m->possible_max_rank; i++)
415 kfree(m->m_info[i].export_targets);
416 kfree(m->m_info);
417 }
2f2dc053
SW
418 kfree(m->m_data_pg_pools);
419 kfree(m);
420}
e9e427f0
YZ
421
422bool ceph_mdsmap_is_cluster_available(struct ceph_mdsmap *m)
423{
424 int i, nr_active = 0;
425 if (!m->m_enabled)
426 return false;
427 if (m->m_damaged)
428 return false;
4d7ace02 429 if (m->m_num_laggy == m->m_num_active_mds)
e9e427f0 430 return false;
b38c9eb4 431 for (i = 0; i < m->possible_max_rank; i++) {
e9e427f0
YZ
432 if (m->m_info[i].state == CEPH_MDS_STATE_ACTIVE)
433 nr_active++;
434 }
435 return nr_active > 0;
436}