Merge tag 'nios2-v5.4-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/lftan...
[linux-2.6-block.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
16
17/*
18 * choose a random mds that is "up" (i.e. has a state > 0), or -1.
19 */
20int ceph_mdsmap_get_random_mds(struct ceph_mdsmap *m)
21{
22 int n = 0;
23 int i;
a84cd293
SL
24
25 /* special case for one mds */
76201b63 26 if (1 == m->m_num_mds && m->m_info[0].state > 0)
a84cd293 27 return 0;
2f2dc053
SW
28
29 /* count */
76201b63 30 for (i = 0; i < m->m_num_mds; i++)
2f2dc053
SW
31 if (m->m_info[i].state > 0)
32 n++;
33 if (n == 0)
34 return -1;
35
36 /* pick */
a84cd293 37 n = prandom_u32() % n;
2f2dc053
SW
38 for (i = 0; n > 0; i++, n--)
39 while (m->m_info[i].state <= 0)
40 i++;
41
42 return i;
43}
44
e9e427f0
YZ
45#define __decode_and_drop_type(p, end, type, bad) \
46 do { \
47 if (*p + sizeof(type) > end) \
48 goto bad; \
49 *p += sizeof(type); \
50 } while (0)
51
52#define __decode_and_drop_set(p, end, type, bad) \
53 do { \
54 u32 n; \
55 size_t need; \
56 ceph_decode_32_safe(p, end, n, bad); \
57 need = sizeof(type) * n; \
58 ceph_decode_need(p, end, need, bad); \
59 *p += need; \
60 } while (0)
61
62#define __decode_and_drop_map(p, end, ktype, vtype, bad) \
63 do { \
64 u32 n; \
65 size_t need; \
66 ceph_decode_32_safe(p, end, n, bad); \
67 need = (sizeof(ktype) + sizeof(vtype)) * n; \
68 ceph_decode_need(p, end, need, bad); \
69 *p += need; \
70 } while (0)
71
72
73static int __decode_and_drop_compat_set(void **p, void* end)
74{
75 int i;
76 /* compat, ro_compat, incompat*/
77 for (i = 0; i < 3; i++) {
78 u32 n;
79 ceph_decode_need(p, end, sizeof(u64) + sizeof(u32), bad);
80 /* mask */
81 *p += sizeof(u64);
82 /* names (map<u64, string>) */
83 n = ceph_decode_32(p);
84 while (n-- > 0) {
85 u32 len;
86 ceph_decode_need(p, end, sizeof(u64) + sizeof(u32),
87 bad);
88 *p += sizeof(u64);
89 len = ceph_decode_32(p);
90 ceph_decode_need(p, end, len, bad);
91 *p += len;
92 }
93 }
94 return 0;
95bad:
96 return -1;
97}
98
2f2dc053
SW
99/*
100 * Decode an MDS map
101 *
102 * Ignore any fields we don't care about (there are quite a few of
103 * them).
104 */
105struct ceph_mdsmap *ceph_mdsmap_decode(void **p, void *end)
106{
107 struct ceph_mdsmap *m;
9ec7cab1 108 const void *start = *p;
2f2dc053 109 int i, j, n;
f3848af1 110 int err;
d463a43d 111 u8 mdsmap_v, mdsmap_cv;
e9e427f0 112 u16 mdsmap_ev;
2f2dc053
SW
113
114 m = kzalloc(sizeof(*m), GFP_NOFS);
d37b1d99 115 if (!m)
2f2dc053
SW
116 return ERR_PTR(-ENOMEM);
117
d463a43d
YZ
118 ceph_decode_need(p, end, 1 + 1, bad);
119 mdsmap_v = ceph_decode_8(p);
120 mdsmap_cv = ceph_decode_8(p);
121 if (mdsmap_v >= 4) {
122 u32 mdsmap_len;
123 ceph_decode_32_safe(p, end, mdsmap_len, bad);
124 if (end < *p + mdsmap_len)
125 goto bad;
126 end = *p + mdsmap_len;
4f6a7e5e 127 }
2f2dc053
SW
128
129 ceph_decode_need(p, end, 8*sizeof(u32) + sizeof(u64), bad);
c89136ea
SW
130 m->m_epoch = ceph_decode_32(p);
131 m->m_client_epoch = ceph_decode_32(p);
132 m->m_last_failure = ceph_decode_32(p);
133 m->m_root = ceph_decode_32(p);
134 m->m_session_timeout = ceph_decode_32(p);
135 m->m_session_autoclose = ceph_decode_32(p);
136 m->m_max_file_size = ceph_decode_64(p);
137 m->m_max_mds = ceph_decode_32(p);
76201b63 138 m->m_num_mds = m->m_max_mds;
2f2dc053 139
76201b63 140 m->m_info = kcalloc(m->m_num_mds, sizeof(*m->m_info), GFP_NOFS);
d37b1d99 141 if (!m->m_info)
e9e427f0 142 goto nomem;
2f2dc053
SW
143
144 /* pick out active nodes from mds_info (state > 0) */
c89136ea 145 n = ceph_decode_32(p);
2f2dc053 146 for (i = 0; i < n; i++) {
94045e11 147 u64 global_id;
2f2dc053
SW
148 u32 namelen;
149 s32 mds, inc, state;
150 u64 state_seq;
d463a43d
YZ
151 u8 info_v;
152 void *info_end = NULL;
2f2dc053
SW
153 struct ceph_entity_addr addr;
154 u32 num_export_targets;
155 void *pexport_targets = NULL;
0deb01c9 156 struct ceph_timespec laggy_since;
6af86528 157 struct ceph_mds_info *info;
2f2dc053 158
d463a43d 159 ceph_decode_need(p, end, sizeof(u64) + 1, bad);
94045e11 160 global_id = ceph_decode_64(p);
d463a43d
YZ
161 info_v= ceph_decode_8(p);
162 if (info_v >= 4) {
163 u32 info_len;
164 u8 info_cv;
165 ceph_decode_need(p, end, 1 + sizeof(u32), bad);
166 info_cv = ceph_decode_8(p);
167 info_len = ceph_decode_32(p);
168 info_end = *p + info_len;
169 if (info_end > end)
170 goto bad;
171 }
172
173 ceph_decode_need(p, end, sizeof(u64) + sizeof(u32), bad);
94045e11 174 *p += sizeof(u64);
c89136ea 175 namelen = ceph_decode_32(p); /* skip mds name */
2f2dc053
SW
176 *p += namelen;
177
178 ceph_decode_need(p, end,
e251e288 179 4*sizeof(u32) + sizeof(u64) +
2f2dc053
SW
180 sizeof(addr) + sizeof(struct ceph_timespec),
181 bad);
c89136ea
SW
182 mds = ceph_decode_32(p);
183 inc = ceph_decode_32(p);
184 state = ceph_decode_32(p);
185 state_seq = ceph_decode_64(p);
f3848af1
JL
186 err = ceph_decode_entity_addr(p, end, &addr);
187 if (err)
188 goto corrupt;
0deb01c9 189 ceph_decode_copy(p, &laggy_since, sizeof(laggy_since));
2f2dc053
SW
190 *p += sizeof(u32);
191 ceph_decode_32_safe(p, end, namelen, bad);
e251e288 192 *p += namelen;
d463a43d 193 if (info_v >= 2) {
2f2dc053
SW
194 ceph_decode_32_safe(p, end, num_export_targets, bad);
195 pexport_targets = *p;
e251e288 196 *p += num_export_targets * sizeof(u32);
2f2dc053
SW
197 } else {
198 num_export_targets = 0;
199 }
200
d463a43d
YZ
201 if (info_end && *p != info_end) {
202 if (*p > info_end)
203 goto bad;
204 *p = info_end;
205 }
206
94045e11 207 dout("mdsmap_decode %d/%d %lld mds%d.%d %s %s\n",
3d14c5d2 208 i+1, n, global_id, mds, inc,
b726ec97 209 ceph_pr_addr(&addr),
2f2dc053 210 ceph_mds_state_name(state));
6af86528 211
76201b63 212 if (mds < 0 || state <= 0)
6af86528
DC
213 continue;
214
76201b63
YZ
215 if (mds >= m->m_num_mds) {
216 int new_num = max(mds + 1, m->m_num_mds * 2);
217 void *new_m_info = krealloc(m->m_info,
218 new_num * sizeof(*m->m_info),
219 GFP_NOFS | __GFP_ZERO);
220 if (!new_m_info)
221 goto nomem;
222 m->m_info = new_m_info;
223 m->m_num_mds = new_num;
224 }
225
6af86528
DC
226 info = &m->m_info[mds];
227 info->global_id = global_id;
228 info->state = state;
229 info->addr = addr;
230 info->laggy = (laggy_since.tv_sec != 0 ||
231 laggy_since.tv_nsec != 0);
232 info->num_export_targets = num_export_targets;
233 if (num_export_targets) {
234 info->export_targets = kcalloc(num_export_targets,
235 sizeof(u32), GFP_NOFS);
d37b1d99 236 if (!info->export_targets)
e9e427f0 237 goto nomem;
6af86528
DC
238 for (j = 0; j < num_export_targets; j++)
239 info->export_targets[j] =
240 ceph_decode_32(&pexport_targets);
241 } else {
242 info->export_targets = NULL;
2f2dc053
SW
243 }
244 }
76201b63
YZ
245 if (m->m_num_mds > m->m_max_mds) {
246 /* find max up mds */
247 for (i = m->m_num_mds; i >= m->m_max_mds; i--) {
248 if (i == 0 || m->m_info[i-1].state > 0)
249 break;
250 }
251 m->m_num_mds = i;
252 }
2f2dc053
SW
253
254 /* pg_pools */
255 ceph_decode_32_safe(p, end, n, bad);
256 m->m_num_data_pg_pools = n;
4f6a7e5e 257 m->m_data_pg_pools = kcalloc(n, sizeof(u64), GFP_NOFS);
2f2dc053 258 if (!m->m_data_pg_pools)
e9e427f0 259 goto nomem;
4f6a7e5e 260 ceph_decode_need(p, end, sizeof(u64)*(n+1), bad);
2f2dc053 261 for (i = 0; i < n; i++)
4f6a7e5e
SW
262 m->m_data_pg_pools[i] = ceph_decode_64(p);
263 m->m_cas_pg_pool = ceph_decode_64(p);
e9e427f0
YZ
264 m->m_enabled = m->m_epoch > 1;
265
266 mdsmap_ev = 1;
267 if (mdsmap_v >= 2) {
268 ceph_decode_16_safe(p, end, mdsmap_ev, bad_ext);
269 }
270 if (mdsmap_ev >= 3) {
271 if (__decode_and_drop_compat_set(p, end) < 0)
272 goto bad_ext;
273 }
274 /* metadata_pool */
275 if (mdsmap_ev < 5) {
276 __decode_and_drop_type(p, end, u32, bad_ext);
277 } else {
278 __decode_and_drop_type(p, end, u64, bad_ext);
279 }
2f2dc053 280
e9e427f0
YZ
281 /* created + modified + tableserver */
282 __decode_and_drop_type(p, end, struct ceph_timespec, bad_ext);
283 __decode_and_drop_type(p, end, struct ceph_timespec, bad_ext);
284 __decode_and_drop_type(p, end, u32, bad_ext);
285
286 /* in */
287 {
288 int num_laggy = 0;
289 ceph_decode_32_safe(p, end, n, bad_ext);
290 ceph_decode_need(p, end, sizeof(u32) * n, bad_ext);
291
292 for (i = 0; i < n; i++) {
293 s32 mds = ceph_decode_32(p);
76201b63 294 if (mds >= 0 && mds < m->m_num_mds) {
e9e427f0
YZ
295 if (m->m_info[mds].laggy)
296 num_laggy++;
297 }
298 }
299 m->m_num_laggy = num_laggy;
76201b63
YZ
300
301 if (n > m->m_num_mds) {
302 void *new_m_info = krealloc(m->m_info,
303 n * sizeof(*m->m_info),
304 GFP_NOFS | __GFP_ZERO);
305 if (!new_m_info)
306 goto nomem;
307 m->m_info = new_m_info;
308 }
309 m->m_num_mds = n;
e9e427f0
YZ
310 }
311
312 /* inc */
313 __decode_and_drop_map(p, end, u32, u32, bad_ext);
314 /* up */
315 __decode_and_drop_map(p, end, u32, u64, bad_ext);
316 /* failed */
317 __decode_and_drop_set(p, end, u32, bad_ext);
318 /* stopped */
319 __decode_and_drop_set(p, end, u32, bad_ext);
320
321 if (mdsmap_ev >= 4) {
322 /* last_failure_osd_epoch */
323 __decode_and_drop_type(p, end, u32, bad_ext);
324 }
325 if (mdsmap_ev >= 6) {
326 /* ever_allowed_snaps */
327 __decode_and_drop_type(p, end, u8, bad_ext);
328 /* explicitly_allowed_snaps */
329 __decode_and_drop_type(p, end, u8, bad_ext);
330 }
331 if (mdsmap_ev >= 7) {
332 /* inline_data_enabled */
333 __decode_and_drop_type(p, end, u8, bad_ext);
334 }
335 if (mdsmap_ev >= 8) {
336 u32 name_len;
337 /* enabled */
338 ceph_decode_8_safe(p, end, m->m_enabled, bad_ext);
339 ceph_decode_32_safe(p, end, name_len, bad_ext);
340 ceph_decode_need(p, end, name_len, bad_ext);
341 *p += name_len;
342 }
343 /* damaged */
344 if (mdsmap_ev >= 9) {
345 size_t need;
346 ceph_decode_32_safe(p, end, n, bad_ext);
347 need = sizeof(u32) * n;
348 ceph_decode_need(p, end, need, bad_ext);
349 *p += need;
350 m->m_damaged = n > 0;
351 } else {
352 m->m_damaged = false;
353 }
354bad_ext:
d463a43d 355 *p = end;
2f2dc053
SW
356 dout("mdsmap_decode success epoch %u\n", m->m_epoch);
357 return m;
e9e427f0 358nomem:
2f2dc053 359 err = -ENOMEM;
e9e427f0 360 goto out_err;
f3848af1 361corrupt:
2f2dc053 362 pr_err("corrupt mdsmap\n");
9ec7cab1
SW
363 print_hex_dump(KERN_DEBUG, "mdsmap: ",
364 DUMP_PREFIX_OFFSET, 16, 1,
365 start, end - start, true);
e9e427f0 366out_err:
2f2dc053 367 ceph_mdsmap_destroy(m);
c213b50b 368 return ERR_PTR(err);
f3848af1
JL
369bad:
370 err = -EINVAL;
371 goto corrupt;
2f2dc053
SW
372}
373
374void ceph_mdsmap_destroy(struct ceph_mdsmap *m)
375{
376 int i;
377
76201b63 378 for (i = 0; i < m->m_num_mds; i++)
2f2dc053
SW
379 kfree(m->m_info[i].export_targets);
380 kfree(m->m_info);
381 kfree(m->m_data_pg_pools);
382 kfree(m);
383}
e9e427f0
YZ
384
385bool ceph_mdsmap_is_cluster_available(struct ceph_mdsmap *m)
386{
387 int i, nr_active = 0;
388 if (!m->m_enabled)
389 return false;
390 if (m->m_damaged)
391 return false;
392 if (m->m_num_laggy > 0)
393 return false;
76201b63 394 for (i = 0; i < m->m_num_mds; i++) {
e9e427f0
YZ
395 if (m->m_info[i].state == CEPH_MDS_STATE_ACTIVE)
396 nr_active++;
397 }
398 return nr_active > 0;
399}