Commit | Line | Data |
---|---|---|
2f2dc053 SW |
1 | #include "ceph_debug.h" |
2 | ||
3 | #include <linux/bug.h> | |
4 | #include <linux/err.h> | |
5 | #include <linux/random.h> | |
6 | #include <linux/slab.h> | |
7 | #include <linux/types.h> | |
8 | ||
9 | #include "mdsmap.h" | |
10 | #include "messenger.h" | |
11 | #include "decode.h" | |
12 | ||
13 | #include "super.h" | |
14 | ||
15 | ||
16 | /* | |
17 | * choose a random mds that is "up" (i.e. has a state > 0), or -1. | |
18 | */ | |
19 | int ceph_mdsmap_get_random_mds(struct ceph_mdsmap *m) | |
20 | { | |
21 | int n = 0; | |
22 | int i; | |
23 | char r; | |
24 | ||
25 | /* count */ | |
26 | for (i = 0; i < m->m_max_mds; i++) | |
27 | if (m->m_info[i].state > 0) | |
28 | n++; | |
29 | if (n == 0) | |
30 | return -1; | |
31 | ||
32 | /* pick */ | |
33 | get_random_bytes(&r, 1); | |
34 | n = r % n; | |
35 | i = 0; | |
36 | for (i = 0; n > 0; i++, n--) | |
37 | while (m->m_info[i].state <= 0) | |
38 | i++; | |
39 | ||
40 | return i; | |
41 | } | |
42 | ||
43 | /* | |
44 | * Decode an MDS map | |
45 | * | |
46 | * Ignore any fields we don't care about (there are quite a few of | |
47 | * them). | |
48 | */ | |
49 | struct ceph_mdsmap *ceph_mdsmap_decode(void **p, void *end) | |
50 | { | |
51 | struct ceph_mdsmap *m; | |
9ec7cab1 | 52 | const void *start = *p; |
2f2dc053 SW |
53 | int i, j, n; |
54 | int err = -EINVAL; | |
55 | u16 version; | |
56 | ||
57 | m = kzalloc(sizeof(*m), GFP_NOFS); | |
58 | if (m == NULL) | |
59 | return ERR_PTR(-ENOMEM); | |
60 | ||
61 | ceph_decode_16_safe(p, end, version, bad); | |
62 | ||
63 | ceph_decode_need(p, end, 8*sizeof(u32) + sizeof(u64), bad); | |
c89136ea SW |
64 | m->m_epoch = ceph_decode_32(p); |
65 | m->m_client_epoch = ceph_decode_32(p); | |
66 | m->m_last_failure = ceph_decode_32(p); | |
67 | m->m_root = ceph_decode_32(p); | |
68 | m->m_session_timeout = ceph_decode_32(p); | |
69 | m->m_session_autoclose = ceph_decode_32(p); | |
70 | m->m_max_file_size = ceph_decode_64(p); | |
71 | m->m_max_mds = ceph_decode_32(p); | |
2f2dc053 SW |
72 | |
73 | m->m_info = kcalloc(m->m_max_mds, sizeof(*m->m_info), GFP_NOFS); | |
74 | if (m->m_info == NULL) | |
75 | goto badmem; | |
76 | ||
77 | /* pick out active nodes from mds_info (state > 0) */ | |
c89136ea | 78 | n = ceph_decode_32(p); |
2f2dc053 | 79 | for (i = 0; i < n; i++) { |
94045e11 | 80 | u64 global_id; |
2f2dc053 SW |
81 | u32 namelen; |
82 | s32 mds, inc, state; | |
83 | u64 state_seq; | |
84 | u8 infoversion; | |
85 | struct ceph_entity_addr addr; | |
86 | u32 num_export_targets; | |
87 | void *pexport_targets = NULL; | |
88 | ||
94045e11 SW |
89 | ceph_decode_need(p, end, sizeof(u64)*2 + 1 + sizeof(u32), bad); |
90 | global_id = ceph_decode_64(p); | |
c89136ea | 91 | infoversion = ceph_decode_8(p); |
94045e11 | 92 | *p += sizeof(u64); |
c89136ea | 93 | namelen = ceph_decode_32(p); /* skip mds name */ |
2f2dc053 SW |
94 | *p += namelen; |
95 | ||
96 | ceph_decode_need(p, end, | |
e251e288 | 97 | 4*sizeof(u32) + sizeof(u64) + |
2f2dc053 SW |
98 | sizeof(addr) + sizeof(struct ceph_timespec), |
99 | bad); | |
c89136ea SW |
100 | mds = ceph_decode_32(p); |
101 | inc = ceph_decode_32(p); | |
102 | state = ceph_decode_32(p); | |
103 | state_seq = ceph_decode_64(p); | |
94045e11 SW |
104 | ceph_decode_copy(p, &addr, sizeof(addr)); |
105 | ceph_decode_addr(&addr); | |
2f2dc053 SW |
106 | *p += sizeof(struct ceph_timespec); |
107 | *p += sizeof(u32); | |
108 | ceph_decode_32_safe(p, end, namelen, bad); | |
e251e288 | 109 | *p += namelen; |
2f2dc053 SW |
110 | if (infoversion >= 2) { |
111 | ceph_decode_32_safe(p, end, num_export_targets, bad); | |
112 | pexport_targets = *p; | |
e251e288 | 113 | *p += num_export_targets * sizeof(u32); |
2f2dc053 SW |
114 | } else { |
115 | num_export_targets = 0; | |
116 | } | |
117 | ||
94045e11 SW |
118 | dout("mdsmap_decode %d/%d %lld mds%d.%d %s %s\n", |
119 | i+1, n, global_id, mds, inc, pr_addr(&addr.in_addr), | |
2f2dc053 SW |
120 | ceph_mds_state_name(state)); |
121 | if (mds >= 0 && mds < m->m_max_mds && state > 0) { | |
94045e11 | 122 | m->m_info[mds].global_id = global_id; |
2f2dc053 SW |
123 | m->m_info[mds].state = state; |
124 | m->m_info[mds].addr = addr; | |
125 | m->m_info[mds].num_export_targets = num_export_targets; | |
126 | if (num_export_targets) { | |
127 | m->m_info[mds].export_targets = | |
128 | kcalloc(num_export_targets, sizeof(u32), | |
129 | GFP_NOFS); | |
130 | for (j = 0; j < num_export_targets; j++) | |
c89136ea SW |
131 | m->m_info[mds].export_targets[j] = |
132 | ceph_decode_32(&pexport_targets); | |
2f2dc053 SW |
133 | } else { |
134 | m->m_info[mds].export_targets = NULL; | |
135 | } | |
136 | } | |
137 | } | |
138 | ||
139 | /* pg_pools */ | |
140 | ceph_decode_32_safe(p, end, n, bad); | |
141 | m->m_num_data_pg_pools = n; | |
142 | m->m_data_pg_pools = kcalloc(n, sizeof(u32), GFP_NOFS); | |
143 | if (!m->m_data_pg_pools) | |
144 | goto badmem; | |
145 | ceph_decode_need(p, end, sizeof(u32)*(n+1), bad); | |
146 | for (i = 0; i < n; i++) | |
c89136ea SW |
147 | m->m_data_pg_pools[i] = ceph_decode_32(p); |
148 | m->m_cas_pg_pool = ceph_decode_32(p); | |
2f2dc053 SW |
149 | |
150 | /* ok, we don't care about the rest. */ | |
151 | dout("mdsmap_decode success epoch %u\n", m->m_epoch); | |
152 | return m; | |
153 | ||
154 | badmem: | |
155 | err = -ENOMEM; | |
156 | bad: | |
157 | pr_err("corrupt mdsmap\n"); | |
9ec7cab1 SW |
158 | print_hex_dump(KERN_DEBUG, "mdsmap: ", |
159 | DUMP_PREFIX_OFFSET, 16, 1, | |
160 | start, end - start, true); | |
2f2dc053 SW |
161 | ceph_mdsmap_destroy(m); |
162 | return ERR_PTR(-EINVAL); | |
163 | } | |
164 | ||
165 | void ceph_mdsmap_destroy(struct ceph_mdsmap *m) | |
166 | { | |
167 | int i; | |
168 | ||
169 | for (i = 0; i < m->m_max_mds; i++) | |
170 | kfree(m->m_info[i].export_targets); | |
171 | kfree(m->m_info); | |
172 | kfree(m->m_data_pg_pools); | |
173 | kfree(m); | |
174 | } |