]>
Commit | Line | Data |
---|---|---|
7c673cae FG |
1 | // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*- |
2 | // vim: ts=8 sw=2 smarttab | |
3 | /* | |
4 | * Ceph distributed storage system | |
5 | * | |
6 | * Copyright (C) 2014 Cloudwatt <libre.licensing@cloudwatt.com> | |
7 | * Copyright (C) 2014 Red Hat <contact@redhat.com> | |
8 | * | |
9 | * Author: Loic Dachary <loic@dachary.org> | |
10 | * | |
11 | * This library is free software; you can redistribute it and/or | |
12 | * modify it under the terms of the GNU Lesser General Public | |
13 | * License as published by the Free Software Foundation; either | |
14 | * version 2.1 of the License, or (at your option) any later version. | |
15 | * | |
16 | */ | |
17 | ||
7c673cae | 18 | #include <algorithm> |
9f95a23c | 19 | #include <cerrno> |
7c673cae | 20 | |
7c673cae | 21 | #include "ErasureCode.h" |
31f18b77 FG |
22 | |
23 | #include "common/strtol.h" | |
7c673cae | 24 | #include "include/buffer.h" |
224ce89b WB |
25 | #include "crush/CrushWrapper.h" |
26 | #include "osd/osd_types.h" | |
7c673cae | 27 | |
224ce89b WB |
28 | #define DEFAULT_RULE_ROOT "default" |
29 | #define DEFAULT_RULE_FAILURE_DOMAIN "host" | |
30 | ||
9f95a23c TL |
31 | using std::make_pair; |
32 | using std::map; | |
33 | using std::ostream; | |
34 | using std::pair; | |
35 | using std::set; | |
36 | using std::string; | |
37 | using std::vector; | |
38 | ||
39 | using ceph::bufferlist; | |
40 | ||
41 | namespace ceph { | |
11fdf7f2 TL |
42 | const unsigned ErasureCode::SIMD_ALIGN = 32; |
43 | ||
224ce89b WB |
44 | int ErasureCode::init( |
45 | ErasureCodeProfile &profile, | |
46 | std::ostream *ss) | |
47 | { | |
48 | int err = 0; | |
49 | err |= to_string("crush-root", profile, | |
50 | &rule_root, | |
51 | DEFAULT_RULE_ROOT, ss); | |
52 | err |= to_string("crush-failure-domain", profile, | |
53 | &rule_failure_domain, | |
54 | DEFAULT_RULE_FAILURE_DOMAIN, ss); | |
55 | err |= to_string("crush-device-class", profile, | |
56 | &rule_device_class, | |
57 | "", ss); | |
58 | if (err) | |
59 | return err; | |
60 | _profile = profile; | |
61 | return 0; | |
62 | } | |
63 | ||
64 | int ErasureCode::create_rule( | |
65 | const std::string &name, | |
66 | CrushWrapper &crush, | |
67 | std::ostream *ss) const | |
68 | { | |
69 | int ruleid = crush.add_simple_rule( | |
70 | name, | |
71 | rule_root, | |
72 | rule_failure_domain, | |
73 | rule_device_class, | |
74 | "indep", | |
75 | pg_pool_t::TYPE_ERASURE, | |
76 | ss); | |
77 | ||
78 | if (ruleid < 0) | |
79 | return ruleid; | |
80 | ||
81 | crush.set_rule_mask_max_size(ruleid, get_chunk_count()); | |
82 | return ruleid; | |
83 | } | |
84 | ||
11fdf7f2 | 85 | int ErasureCode::sanity_check_k_m(int k, int m, ostream *ss) |
7c673cae FG |
86 | { |
87 | if (k < 2) { | |
88 | *ss << "k=" << k << " must be >= 2" << std::endl; | |
89 | return -EINVAL; | |
7c673cae | 90 | } |
11fdf7f2 TL |
91 | if (m < 1) { |
92 | *ss << "m=" << m << " must be >= 1" << std::endl; | |
93 | return -EINVAL; | |
94 | } | |
95 | return 0; | |
7c673cae FG |
96 | } |
97 | ||
98 | int ErasureCode::chunk_index(unsigned int i) const | |
99 | { | |
100 | return chunk_mapping.size() > i ? chunk_mapping[i] : i; | |
101 | } | |
102 | ||
11fdf7f2 | 103 | int ErasureCode::_minimum_to_decode(const set<int> &want_to_read, |
7c673cae FG |
104 | const set<int> &available_chunks, |
105 | set<int> *minimum) | |
106 | { | |
107 | if (includes(available_chunks.begin(), available_chunks.end(), | |
108 | want_to_read.begin(), want_to_read.end())) { | |
109 | *minimum = want_to_read; | |
110 | } else { | |
111 | unsigned int k = get_data_chunk_count(); | |
112 | if (available_chunks.size() < (unsigned)k) | |
113 | return -EIO; | |
114 | set<int>::iterator i; | |
115 | unsigned j; | |
116 | for (i = available_chunks.begin(), j = 0; j < (unsigned)k; ++i, j++) | |
117 | minimum->insert(*i); | |
118 | } | |
119 | return 0; | |
120 | } | |
121 | ||
11fdf7f2 TL |
122 | int ErasureCode::minimum_to_decode(const set<int> &want_to_read, |
123 | const set<int> &available_chunks, | |
124 | map<int, vector<pair<int, int>>> *minimum) | |
125 | { | |
126 | set<int> minimum_shard_ids; | |
127 | int r = _minimum_to_decode(want_to_read, available_chunks, &minimum_shard_ids); | |
128 | if (r != 0) { | |
129 | return r; | |
130 | } | |
131 | vector<pair<int, int>> default_subchunks; | |
132 | default_subchunks.push_back(make_pair(0, get_sub_chunk_count())); | |
133 | for (auto &&id : minimum_shard_ids) { | |
134 | minimum->insert(make_pair(id, default_subchunks)); | |
135 | } | |
136 | return 0; | |
137 | } | |
138 | ||
7c673cae FG |
139 | int ErasureCode::minimum_to_decode_with_cost(const set<int> &want_to_read, |
140 | const map<int, int> &available, | |
141 | set<int> *minimum) | |
142 | { | |
143 | set <int> available_chunks; | |
144 | for (map<int, int>::const_iterator i = available.begin(); | |
145 | i != available.end(); | |
146 | ++i) | |
147 | available_chunks.insert(i->first); | |
11fdf7f2 | 148 | return _minimum_to_decode(want_to_read, available_chunks, minimum); |
7c673cae FG |
149 | } |
150 | ||
151 | int ErasureCode::encode_prepare(const bufferlist &raw, | |
152 | map<int, bufferlist> &encoded) const | |
153 | { | |
154 | unsigned int k = get_data_chunk_count(); | |
155 | unsigned int m = get_chunk_count() - k; | |
156 | unsigned blocksize = get_chunk_size(raw.length()); | |
157 | unsigned padded_chunks = k - raw.length() / blocksize; | |
158 | bufferlist prepared = raw; | |
159 | ||
160 | for (unsigned int i = 0; i < k - padded_chunks; i++) { | |
161 | bufferlist &chunk = encoded[chunk_index(i)]; | |
162 | chunk.substr_of(prepared, i * blocksize, blocksize); | |
163 | chunk.rebuild_aligned_size_and_memory(blocksize, SIMD_ALIGN); | |
11fdf7f2 | 164 | ceph_assert(chunk.is_contiguous()); |
7c673cae FG |
165 | } |
166 | if (padded_chunks) { | |
167 | unsigned remainder = raw.length() - (k - padded_chunks) * blocksize; | |
168 | bufferptr buf(buffer::create_aligned(blocksize, SIMD_ALIGN)); | |
169 | ||
9f95a23c | 170 | raw.begin((k - padded_chunks) * blocksize).copy(remainder, buf.c_str()); |
7c673cae FG |
171 | buf.zero(remainder, blocksize - remainder); |
172 | encoded[chunk_index(k-padded_chunks)].push_back(std::move(buf)); | |
173 | ||
174 | for (unsigned int i = k - padded_chunks + 1; i < k; i++) { | |
175 | bufferptr buf(buffer::create_aligned(blocksize, SIMD_ALIGN)); | |
176 | buf.zero(); | |
177 | encoded[chunk_index(i)].push_back(std::move(buf)); | |
178 | } | |
179 | } | |
180 | for (unsigned int i = k; i < k + m; i++) { | |
181 | bufferlist &chunk = encoded[chunk_index(i)]; | |
182 | chunk.push_back(buffer::create_aligned(blocksize, SIMD_ALIGN)); | |
183 | } | |
184 | ||
185 | return 0; | |
186 | } | |
187 | ||
188 | int ErasureCode::encode(const set<int> &want_to_encode, | |
189 | const bufferlist &in, | |
190 | map<int, bufferlist> *encoded) | |
191 | { | |
192 | unsigned int k = get_data_chunk_count(); | |
193 | unsigned int m = get_chunk_count() - k; | |
194 | bufferlist out; | |
195 | int err = encode_prepare(in, *encoded); | |
196 | if (err) | |
197 | return err; | |
198 | encode_chunks(want_to_encode, encoded); | |
199 | for (unsigned int i = 0; i < k + m; i++) { | |
200 | if (want_to_encode.count(i) == 0) | |
201 | encoded->erase(i); | |
202 | } | |
203 | return 0; | |
204 | } | |
205 | ||
206 | int ErasureCode::encode_chunks(const set<int> &want_to_encode, | |
207 | map<int, bufferlist> *encoded) | |
208 | { | |
11fdf7f2 | 209 | ceph_abort_msg("ErasureCode::encode_chunks not implemented"); |
7c673cae FG |
210 | } |
211 | ||
11fdf7f2 TL |
212 | int ErasureCode::_decode(const set<int> &want_to_read, |
213 | const map<int, bufferlist> &chunks, | |
214 | map<int, bufferlist> *decoded) | |
7c673cae FG |
215 | { |
216 | vector<int> have; | |
217 | have.reserve(chunks.size()); | |
218 | for (map<int, bufferlist>::const_iterator i = chunks.begin(); | |
219 | i != chunks.end(); | |
220 | ++i) { | |
221 | have.push_back(i->first); | |
222 | } | |
223 | if (includes( | |
224 | have.begin(), have.end(), want_to_read.begin(), want_to_read.end())) { | |
225 | for (set<int>::iterator i = want_to_read.begin(); | |
226 | i != want_to_read.end(); | |
227 | ++i) { | |
228 | (*decoded)[*i] = chunks.find(*i)->second; | |
229 | } | |
230 | return 0; | |
231 | } | |
232 | unsigned int k = get_data_chunk_count(); | |
233 | unsigned int m = get_chunk_count() - k; | |
234 | unsigned blocksize = (*chunks.begin()).second.length(); | |
235 | for (unsigned int i = 0; i < k + m; i++) { | |
236 | if (chunks.find(i) == chunks.end()) { | |
11fdf7f2 | 237 | bufferlist tmp; |
7c673cae | 238 | bufferptr ptr(buffer::create_aligned(blocksize, SIMD_ALIGN)); |
11fdf7f2 TL |
239 | tmp.push_back(ptr); |
240 | tmp.claim_append((*decoded)[i]); | |
241 | (*decoded)[i].swap(tmp); | |
7c673cae FG |
242 | } else { |
243 | (*decoded)[i] = chunks.find(i)->second; | |
244 | (*decoded)[i].rebuild_aligned(SIMD_ALIGN); | |
245 | } | |
246 | } | |
247 | return decode_chunks(want_to_read, chunks, decoded); | |
248 | } | |
249 | ||
11fdf7f2 TL |
250 | int ErasureCode::decode(const set<int> &want_to_read, |
251 | const map<int, bufferlist> &chunks, | |
252 | map<int, bufferlist> *decoded, int chunk_size) | |
253 | { | |
254 | return _decode(want_to_read, chunks, decoded); | |
255 | } | |
256 | ||
7c673cae FG |
257 | int ErasureCode::decode_chunks(const set<int> &want_to_read, |
258 | const map<int, bufferlist> &chunks, | |
259 | map<int, bufferlist> *decoded) | |
260 | { | |
11fdf7f2 | 261 | ceph_abort_msg("ErasureCode::decode_chunks not implemented"); |
7c673cae FG |
262 | } |
263 | ||
264 | int ErasureCode::parse(const ErasureCodeProfile &profile, | |
265 | ostream *ss) | |
266 | { | |
267 | return to_mapping(profile, ss); | |
268 | } | |
269 | ||
270 | const vector<int> &ErasureCode::get_chunk_mapping() const { | |
271 | return chunk_mapping; | |
272 | } | |
273 | ||
274 | int ErasureCode::to_mapping(const ErasureCodeProfile &profile, | |
275 | ostream *ss) | |
276 | { | |
277 | if (profile.find("mapping") != profile.end()) { | |
278 | std::string mapping = profile.find("mapping")->second; | |
279 | int position = 0; | |
280 | vector<int> coding_chunk_mapping; | |
281 | for(std::string::iterator it = mapping.begin(); it != mapping.end(); ++it) { | |
282 | if (*it == 'D') | |
283 | chunk_mapping.push_back(position); | |
284 | else | |
285 | coding_chunk_mapping.push_back(position); | |
286 | position++; | |
287 | } | |
288 | chunk_mapping.insert(chunk_mapping.end(), | |
289 | coding_chunk_mapping.begin(), | |
290 | coding_chunk_mapping.end()); | |
291 | } | |
292 | return 0; | |
293 | } | |
294 | ||
295 | int ErasureCode::to_int(const std::string &name, | |
296 | ErasureCodeProfile &profile, | |
297 | int *value, | |
298 | const std::string &default_value, | |
299 | ostream *ss) | |
300 | { | |
301 | if (profile.find(name) == profile.end() || | |
302 | profile.find(name)->second.size() == 0) | |
303 | profile[name] = default_value; | |
304 | std::string p = profile.find(name)->second; | |
305 | std::string err; | |
306 | int r = strict_strtol(p.c_str(), 10, &err); | |
307 | if (!err.empty()) { | |
308 | *ss << "could not convert " << name << "=" << p | |
309 | << " to int because " << err | |
310 | << ", set to default " << default_value << std::endl; | |
311 | *value = strict_strtol(default_value.c_str(), 10, &err); | |
312 | return -EINVAL; | |
313 | } | |
314 | *value = r; | |
315 | return 0; | |
316 | } | |
317 | ||
318 | int ErasureCode::to_bool(const std::string &name, | |
319 | ErasureCodeProfile &profile, | |
320 | bool *value, | |
321 | const std::string &default_value, | |
322 | ostream *ss) | |
323 | { | |
324 | if (profile.find(name) == profile.end() || | |
325 | profile.find(name)->second.size() == 0) | |
326 | profile[name] = default_value; | |
327 | const std::string p = profile.find(name)->second; | |
328 | *value = (p == "yes") || (p == "true"); | |
329 | return 0; | |
330 | } | |
331 | ||
332 | int ErasureCode::to_string(const std::string &name, | |
333 | ErasureCodeProfile &profile, | |
334 | std::string *value, | |
335 | const std::string &default_value, | |
336 | ostream *ss) | |
337 | { | |
338 | if (profile.find(name) == profile.end() || | |
339 | profile.find(name)->second.size() == 0) | |
340 | profile[name] = default_value; | |
341 | *value = profile[name]; | |
342 | return 0; | |
343 | } | |
344 | ||
345 | int ErasureCode::decode_concat(const map<int, bufferlist> &chunks, | |
346 | bufferlist *decoded) | |
347 | { | |
348 | set<int> want_to_read; | |
349 | ||
350 | for (unsigned int i = 0; i < get_data_chunk_count(); i++) { | |
351 | want_to_read.insert(chunk_index(i)); | |
352 | } | |
353 | map<int, bufferlist> decoded_map; | |
11fdf7f2 | 354 | int r = _decode(want_to_read, chunks, &decoded_map); |
7c673cae FG |
355 | if (r == 0) { |
356 | for (unsigned int i = 0; i < get_data_chunk_count(); i++) { | |
357 | decoded->claim_append(decoded_map[chunk_index(i)]); | |
358 | } | |
359 | } | |
360 | return r; | |
361 | } | |
9f95a23c | 362 | } |