]>
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 | #include "rgw_gc.h" | |
5 | #include "include/rados/librados.hpp" | |
6 | #include "cls/rgw/cls_rgw_client.h" | |
7 | #include "cls/refcount/cls_refcount_client.h" | |
8 | #include "cls/lock/cls_lock_client.h" | |
9 | #include "auth/Crypto.h" | |
10 | ||
11 | #include <list> | |
12 | ||
13 | #define dout_context g_ceph_context | |
14 | #define dout_subsys ceph_subsys_rgw | |
15 | ||
16 | using namespace std; | |
17 | using namespace librados; | |
18 | ||
19 | static string gc_oid_prefix = "gc"; | |
20 | static string gc_index_lock_name = "gc_process"; | |
21 | ||
22 | ||
23 | #define HASH_PRIME 7877 | |
24 | ||
25 | void RGWGC::initialize(CephContext *_cct, RGWRados *_store) { | |
26 | cct = _cct; | |
27 | store = _store; | |
28 | ||
29 | max_objs = cct->_conf->rgw_gc_max_objs; | |
30 | if (max_objs > HASH_PRIME) | |
31 | max_objs = HASH_PRIME; | |
32 | ||
33 | obj_names = new string[max_objs]; | |
34 | ||
35 | for (int i = 0; i < max_objs; i++) { | |
36 | obj_names[i] = gc_oid_prefix; | |
37 | char buf[32]; | |
38 | snprintf(buf, 32, ".%d", i); | |
39 | obj_names[i].append(buf); | |
40 | } | |
41 | } | |
42 | ||
43 | void RGWGC::finalize() | |
44 | { | |
45 | delete[] obj_names; | |
46 | } | |
47 | ||
48 | int RGWGC::tag_index(const string& tag) | |
49 | { | |
50 | return ceph_str_hash_linux(tag.c_str(), tag.size()) % HASH_PRIME % max_objs; | |
51 | } | |
52 | ||
53 | void RGWGC::add_chain(ObjectWriteOperation& op, cls_rgw_obj_chain& chain, const string& tag) | |
54 | { | |
55 | cls_rgw_gc_obj_info info; | |
56 | info.chain = chain; | |
57 | info.tag = tag; | |
58 | ||
59 | cls_rgw_gc_set_entry(op, cct->_conf->rgw_gc_obj_min_wait, info); | |
60 | } | |
61 | ||
62 | int RGWGC::send_chain(cls_rgw_obj_chain& chain, const string& tag, bool sync) | |
63 | { | |
64 | ObjectWriteOperation op; | |
65 | add_chain(op, chain, tag); | |
66 | ||
67 | int i = tag_index(tag); | |
68 | ||
69 | if (sync) | |
70 | return store->gc_operate(obj_names[i], &op); | |
71 | ||
72 | return store->gc_aio_operate(obj_names[i], &op); | |
73 | } | |
74 | ||
75 | int RGWGC::defer_chain(const string& tag, bool sync) | |
76 | { | |
77 | ObjectWriteOperation op; | |
78 | cls_rgw_gc_defer_entry(op, cct->_conf->rgw_gc_obj_min_wait, tag); | |
79 | ||
80 | int i = tag_index(tag); | |
81 | ||
82 | if (sync) | |
83 | return store->gc_operate(obj_names[i], &op); | |
84 | ||
85 | return store->gc_aio_operate(obj_names[i], &op); | |
86 | } | |
87 | ||
88 | int RGWGC::remove(int index, const std::list<string>& tags) | |
89 | { | |
90 | ObjectWriteOperation op; | |
91 | cls_rgw_gc_remove(op, tags); | |
92 | return store->gc_operate(obj_names[index], &op); | |
93 | } | |
94 | ||
95 | int RGWGC::list(int *index, string& marker, uint32_t max, bool expired_only, std::list<cls_rgw_gc_obj_info>& result, bool *truncated) | |
96 | { | |
97 | result.clear(); | |
98 | ||
99 | for (; *index < max_objs && result.size() < max; (*index)++, marker.clear()) { | |
100 | std::list<cls_rgw_gc_obj_info> entries; | |
101 | int ret = cls_rgw_gc_list(store->gc_pool_ctx, obj_names[*index], marker, max - result.size(), expired_only, entries, truncated); | |
102 | if (ret == -ENOENT) | |
103 | continue; | |
104 | if (ret < 0) | |
105 | return ret; | |
106 | ||
107 | std::list<cls_rgw_gc_obj_info>::iterator iter; | |
108 | for (iter = entries.begin(); iter != entries.end(); ++iter) { | |
109 | result.push_back(*iter); | |
110 | } | |
111 | ||
112 | if (*index == max_objs - 1) { | |
113 | /* we cut short here, truncated will hold the correct value */ | |
114 | return 0; | |
115 | } | |
116 | ||
117 | if (result.size() == max) { | |
118 | /* close approximation, it might be that the next of the objects don't hold | |
119 | * anything, in this case truncated should have been false, but we can find | |
120 | * that out on the next iteration | |
121 | */ | |
122 | *truncated = true; | |
123 | return 0; | |
124 | } | |
125 | ||
126 | } | |
127 | *truncated = false; | |
128 | ||
129 | return 0; | |
130 | } | |
131 | ||
132 | int RGWGC::process(int index, int max_secs) | |
133 | { | |
134 | rados::cls::lock::Lock l(gc_index_lock_name); | |
135 | utime_t end = ceph_clock_now(); | |
136 | std::list<string> remove_tags; | |
137 | ||
138 | /* max_secs should be greater than zero. We don't want a zero max_secs | |
139 | * to be translated as no timeout, since we'd then need to break the | |
140 | * lock and that would require a manual intervention. In this case | |
141 | * we can just wait it out. */ | |
142 | if (max_secs <= 0) | |
143 | return -EAGAIN; | |
144 | ||
145 | end += max_secs; | |
146 | utime_t time(max_secs, 0); | |
147 | l.set_duration(time); | |
148 | ||
149 | int ret = l.lock_exclusive(&store->gc_pool_ctx, obj_names[index]); | |
150 | if (ret == -EBUSY) { /* already locked by another gc processor */ | |
151 | dout(0) << "RGWGC::process() failed to acquire lock on " << obj_names[index] << dendl; | |
152 | return 0; | |
153 | } | |
154 | if (ret < 0) | |
155 | return ret; | |
156 | ||
157 | string marker; | |
158 | bool truncated; | |
159 | IoCtx *ctx = new IoCtx; | |
160 | do { | |
161 | int max = 100; | |
162 | std::list<cls_rgw_gc_obj_info> entries; | |
163 | ret = cls_rgw_gc_list(store->gc_pool_ctx, obj_names[index], marker, max, true, entries, &truncated); | |
164 | if (ret == -ENOENT) { | |
165 | ret = 0; | |
166 | goto done; | |
167 | } | |
168 | if (ret < 0) | |
169 | goto done; | |
170 | ||
171 | string last_pool; | |
172 | std::list<cls_rgw_gc_obj_info>::iterator iter; | |
173 | for (iter = entries.begin(); iter != entries.end(); ++iter) { | |
174 | bool remove_tag; | |
175 | cls_rgw_gc_obj_info& info = *iter; | |
176 | std::list<cls_rgw_obj>::iterator liter; | |
177 | cls_rgw_obj_chain& chain = info.chain; | |
178 | ||
179 | utime_t now = ceph_clock_now(); | |
180 | if (now >= end) | |
181 | goto done; | |
182 | ||
183 | remove_tag = true; | |
184 | for (liter = chain.objs.begin(); liter != chain.objs.end(); ++liter) { | |
185 | cls_rgw_obj& obj = *liter; | |
186 | ||
187 | if (obj.pool != last_pool) { | |
188 | delete ctx; | |
189 | ctx = new IoCtx; | |
190 | ret = rgw_init_ioctx(store->get_rados_handle(), obj.pool, *ctx); | |
191 | if (ret < 0) { | |
192 | dout(0) << "ERROR: failed to create ioctx pool=" << obj.pool << dendl; | |
193 | continue; | |
194 | } | |
195 | last_pool = obj.pool; | |
196 | } | |
197 | ||
198 | ctx->locator_set_key(obj.loc); | |
199 | ||
200 | const string& oid = obj.key.name; /* just stored raw oid there */ | |
201 | ||
202 | dout(0) << "gc::process: removing " << obj.pool << ":" << obj.key.name << dendl; | |
203 | ObjectWriteOperation op; | |
204 | cls_refcount_put(op, info.tag, true); | |
205 | ret = ctx->operate(oid, &op); | |
206 | if (ret == -ENOENT) | |
207 | ret = 0; | |
208 | if (ret < 0) { | |
209 | remove_tag = false; | |
210 | dout(0) << "failed to remove " << obj.pool << ":" << oid << "@" << obj.loc << dendl; | |
211 | } | |
212 | ||
213 | if (going_down()) // leave early, even if tag isn't removed, it's ok | |
214 | goto done; | |
215 | } | |
216 | if (remove_tag) { | |
217 | remove_tags.push_back(info.tag); | |
218 | #define MAX_REMOVE_CHUNK 16 | |
219 | if (remove_tags.size() > MAX_REMOVE_CHUNK) { | |
220 | RGWGC::remove(index, remove_tags); | |
221 | remove_tags.clear(); | |
222 | } | |
223 | } | |
224 | } | |
225 | } while (truncated); | |
226 | ||
227 | done: | |
228 | if (!remove_tags.empty()) | |
229 | RGWGC::remove(index, remove_tags); | |
230 | l.unlock(&store->gc_pool_ctx, obj_names[index]); | |
231 | delete ctx; | |
232 | return 0; | |
233 | } | |
234 | ||
235 | int RGWGC::process() | |
236 | { | |
237 | int max_secs = cct->_conf->rgw_gc_processor_max_time; | |
238 | ||
239 | unsigned start; | |
240 | int ret = get_random_bytes((char *)&start, sizeof(start)); | |
241 | if (ret < 0) | |
242 | return ret; | |
243 | ||
244 | for (int i = 0; i < max_objs; i++) { | |
245 | int index = (i + start) % max_objs; | |
246 | ret = process(index, max_secs); | |
247 | if (ret < 0) | |
248 | return ret; | |
249 | } | |
250 | ||
251 | return 0; | |
252 | } | |
253 | ||
254 | bool RGWGC::going_down() | |
255 | { | |
256 | return down_flag; | |
257 | } | |
258 | ||
259 | void RGWGC::start_processor() | |
260 | { | |
261 | worker = new GCWorker(cct, this); | |
262 | worker->create("rgw_gc"); | |
263 | } | |
264 | ||
265 | void RGWGC::stop_processor() | |
266 | { | |
267 | down_flag = true; | |
268 | if (worker) { | |
269 | worker->stop(); | |
270 | worker->join(); | |
271 | } | |
272 | delete worker; | |
273 | worker = NULL; | |
274 | } | |
275 | ||
276 | void *RGWGC::GCWorker::entry() { | |
277 | do { | |
278 | utime_t start = ceph_clock_now(); | |
279 | dout(2) << "garbage collection: start" << dendl; | |
280 | int r = gc->process(); | |
281 | if (r < 0) { | |
282 | dout(0) << "ERROR: garbage collection process() returned error r=" << r << dendl; | |
283 | } | |
284 | dout(2) << "garbage collection: stop" << dendl; | |
285 | ||
286 | if (gc->going_down()) | |
287 | break; | |
288 | ||
289 | utime_t end = ceph_clock_now(); | |
290 | end -= start; | |
291 | int secs = cct->_conf->rgw_gc_processor_period; | |
292 | ||
293 | if (secs <= end.sec()) | |
294 | continue; // next round | |
295 | ||
296 | secs -= end.sec(); | |
297 | ||
298 | lock.Lock(); | |
299 | cond.WaitInterval(lock, utime_t(secs, 0)); | |
300 | lock.Unlock(); | |
301 | } while (!gc->going_down()); | |
302 | ||
303 | return NULL; | |
304 | } | |
305 | ||
306 | void RGWGC::GCWorker::stop() | |
307 | { | |
308 | Mutex::Locker l(lock); | |
309 | cond.Signal(); | |
310 | } | |
311 |