]>
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 | #include "include/int_types.h" | |
4 | ||
5 | #include <errno.h> | |
6 | #include <limits.h> | |
7 | ||
8 | #include "include/types.h" | |
9 | #include "include/uuid.h" | |
10 | #include "common/ceph_context.h" | |
11 | #include "common/dout.h" | |
12 | #include "common/errno.h" | |
13 | #include "common/Throttle.h" | |
14 | #include "common/event_socket.h" | |
15 | #include "cls/lock/cls_lock_client.h" | |
16 | #include "include/stringify.h" | |
17 | ||
18 | #include "cls/rbd/cls_rbd.h" | |
19 | #include "cls/rbd/cls_rbd_types.h" | |
20 | #include "cls/rbd/cls_rbd_client.h" | |
21 | #include "cls/journal/cls_journal_types.h" | |
22 | #include "cls/journal/cls_journal_client.h" | |
23 | ||
24 | #include "librbd/ExclusiveLock.h" | |
25 | #include "librbd/ImageCtx.h" | |
26 | #include "librbd/ImageState.h" | |
27 | #include "librbd/internal.h" | |
28 | #include "librbd/Journal.h" | |
29 | #include "librbd/ObjectMap.h" | |
30 | #include "librbd/Operations.h" | |
31 | #include "librbd/Types.h" | |
32 | #include "librbd/Utils.h" | |
33 | #include "librbd/api/Image.h" | |
34 | #include "librbd/exclusive_lock/AutomaticPolicy.h" | |
35 | #include "librbd/exclusive_lock/StandardPolicy.h" | |
36 | #include "librbd/image/CloneRequest.h" | |
37 | #include "librbd/image/CreateRequest.h" | |
38 | #include "librbd/image/RemoveRequest.h" | |
39 | #include "librbd/io/AioCompletion.h" | |
40 | #include "librbd/io/ImageRequest.h" | |
41 | #include "librbd/io/ImageRequestWQ.h" | |
42 | #include "librbd/io/ObjectRequest.h" | |
43 | #include "librbd/io/ReadResult.h" | |
44 | #include "librbd/journal/Types.h" | |
45 | #include "librbd/managed_lock/Types.h" | |
46 | #include "librbd/mirror/EnableRequest.h" | |
47 | #include "librbd/operation/TrimRequest.h" | |
48 | ||
49 | #include "journal/Journaler.h" | |
50 | ||
51 | #include <boost/scope_exit.hpp> | |
52 | #include <boost/variant.hpp> | |
53 | #include "include/assert.h" | |
54 | ||
55 | #define dout_subsys ceph_subsys_rbd | |
56 | #undef dout_prefix | |
57 | #define dout_prefix *_dout << "librbd: " | |
58 | ||
59 | #define rbd_howmany(x, y) (((x) + (y) - 1) / (y)) | |
60 | ||
61 | using std::map; | |
62 | using std::pair; | |
63 | using std::set; | |
64 | using std::string; | |
65 | using std::vector; | |
66 | // list binds to list() here, so std::list is explicitly used below | |
67 | ||
68 | using ceph::bufferlist; | |
69 | using librados::snap_t; | |
70 | using librados::IoCtx; | |
71 | using librados::Rados; | |
72 | ||
73 | namespace librbd { | |
74 | ||
75 | namespace { | |
76 | ||
77 | int validate_pool(IoCtx &io_ctx, CephContext *cct) { | |
181888fb | 78 | if (!cct->_conf->get_val<bool>("rbd_validate_pool")) { |
7c673cae FG |
79 | return 0; |
80 | } | |
81 | ||
82 | int r = io_ctx.stat(RBD_DIRECTORY, NULL, NULL); | |
83 | if (r == 0) { | |
84 | return 0; | |
85 | } else if (r < 0 && r != -ENOENT) { | |
86 | lderr(cct) << "failed to stat RBD directory: " << cpp_strerror(r) << dendl; | |
87 | return r; | |
88 | } | |
89 | ||
90 | // allocate a self-managed snapshot id if this a new pool to force | |
91 | // self-managed snapshot mode | |
92 | uint64_t snap_id; | |
93 | r = io_ctx.selfmanaged_snap_create(&snap_id); | |
94 | if (r == -EINVAL) { | |
95 | lderr(cct) << "pool not configured for self-managed RBD snapshot support" | |
96 | << dendl; | |
97 | return r; | |
98 | } else if (r < 0) { | |
99 | lderr(cct) << "failed to allocate self-managed snapshot: " | |
100 | << cpp_strerror(r) << dendl; | |
101 | return r; | |
102 | } | |
103 | ||
104 | r = io_ctx.selfmanaged_snap_remove(snap_id); | |
105 | if (r < 0) { | |
106 | lderr(cct) << "failed to release self-managed snapshot " << snap_id | |
107 | << ": " << cpp_strerror(r) << dendl; | |
108 | } | |
109 | return 0; | |
110 | } | |
111 | ||
112 | ||
113 | } // anonymous namespace | |
114 | ||
115 | int detect_format(IoCtx &io_ctx, const string &name, | |
116 | bool *old_format, uint64_t *size) | |
117 | { | |
118 | CephContext *cct = (CephContext *)io_ctx.cct(); | |
119 | if (old_format) | |
120 | *old_format = true; | |
121 | int r = io_ctx.stat(util::old_header_name(name), size, NULL); | |
122 | if (r == -ENOENT) { | |
123 | if (old_format) | |
124 | *old_format = false; | |
125 | r = io_ctx.stat(util::id_obj_name(name), size, NULL); | |
126 | if (r < 0) | |
127 | return r; | |
128 | } else if (r < 0) { | |
129 | return r; | |
130 | } | |
131 | ||
132 | ldout(cct, 20) << "detect format of " << name << " : " | |
133 | << (old_format ? (*old_format ? "old" : "new") : | |
134 | "don't care") << dendl; | |
135 | return 0; | |
136 | } | |
137 | ||
138 | bool has_parent(int64_t parent_pool_id, uint64_t off, uint64_t overlap) | |
139 | { | |
140 | return (parent_pool_id != -1 && off <= overlap); | |
141 | } | |
142 | ||
143 | void init_rbd_header(struct rbd_obj_header_ondisk& ondisk, | |
144 | uint64_t size, int order, uint64_t bid) | |
145 | { | |
146 | uint32_t hi = bid >> 32; | |
147 | uint32_t lo = bid & 0xFFFFFFFF; | |
148 | uint32_t extra = rand() % 0xFFFFFFFF; | |
149 | memset(&ondisk, 0, sizeof(ondisk)); | |
150 | ||
151 | memcpy(&ondisk.text, RBD_HEADER_TEXT, sizeof(RBD_HEADER_TEXT)); | |
152 | memcpy(&ondisk.signature, RBD_HEADER_SIGNATURE, | |
153 | sizeof(RBD_HEADER_SIGNATURE)); | |
154 | memcpy(&ondisk.version, RBD_HEADER_VERSION, sizeof(RBD_HEADER_VERSION)); | |
155 | ||
156 | snprintf(ondisk.block_name, sizeof(ondisk.block_name), "rb.%x.%x.%x", | |
157 | hi, lo, extra); | |
158 | ||
159 | ondisk.image_size = size; | |
160 | ondisk.options.order = order; | |
161 | ondisk.options.crypt_type = RBD_CRYPT_NONE; | |
162 | ondisk.options.comp_type = RBD_COMP_NONE; | |
163 | ondisk.snap_seq = 0; | |
164 | ondisk.snap_count = 0; | |
165 | ondisk.reserved = 0; | |
166 | ondisk.snap_names_len = 0; | |
167 | } | |
168 | ||
169 | void image_info(ImageCtx *ictx, image_info_t& info, size_t infosize) | |
170 | { | |
171 | int obj_order = ictx->order; | |
172 | ictx->snap_lock.get_read(); | |
173 | info.size = ictx->get_image_size(ictx->snap_id); | |
174 | ictx->snap_lock.put_read(); | |
175 | info.obj_size = 1ULL << obj_order; | |
176 | info.num_objs = Striper::get_num_objects(ictx->layout, info.size); | |
177 | info.order = obj_order; | |
178 | strncpy(info.block_name_prefix, ictx->object_prefix.c_str(), | |
179 | RBD_MAX_BLOCK_NAME_SIZE); | |
180 | info.block_name_prefix[RBD_MAX_BLOCK_NAME_SIZE - 1] = '\0'; | |
181 | ||
182 | // clear deprecated fields | |
183 | info.parent_pool = -1L; | |
184 | info.parent_name[0] = '\0'; | |
185 | } | |
186 | ||
187 | uint64_t oid_to_object_no(const string& oid, const string& object_prefix) | |
188 | { | |
189 | istringstream iss(oid); | |
190 | // skip object prefix and separator | |
191 | iss.ignore(object_prefix.length() + 1); | |
192 | uint64_t num; | |
193 | iss >> std::hex >> num; | |
194 | return num; | |
195 | } | |
196 | ||
197 | void trim_image(ImageCtx *ictx, uint64_t newsize, ProgressContext& prog_ctx) | |
198 | { | |
199 | assert(ictx->owner_lock.is_locked()); | |
200 | assert(ictx->exclusive_lock == nullptr || | |
201 | ictx->exclusive_lock->is_lock_owner()); | |
202 | ||
203 | C_SaferCond ctx; | |
204 | ictx->snap_lock.get_read(); | |
205 | operation::TrimRequest<> *req = operation::TrimRequest<>::create( | |
206 | *ictx, &ctx, ictx->size, newsize, prog_ctx); | |
207 | ictx->snap_lock.put_read(); | |
208 | req->send(); | |
209 | ||
210 | int r = ctx.wait(); | |
211 | if (r < 0) { | |
212 | lderr(ictx->cct) << "warning: failed to remove some object(s): " | |
213 | << cpp_strerror(r) << dendl; | |
214 | } | |
215 | } | |
216 | ||
217 | int read_header_bl(IoCtx& io_ctx, const string& header_oid, | |
218 | bufferlist& header, uint64_t *ver) | |
219 | { | |
220 | int r; | |
221 | uint64_t off = 0; | |
222 | #define READ_SIZE 4096 | |
223 | do { | |
224 | bufferlist bl; | |
225 | r = io_ctx.read(header_oid, bl, READ_SIZE, off); | |
226 | if (r < 0) | |
227 | return r; | |
228 | header.claim_append(bl); | |
229 | off += r; | |
230 | } while (r == READ_SIZE); | |
231 | ||
232 | if (header.length() < sizeof(RBD_HEADER_TEXT) || | |
233 | memcmp(RBD_HEADER_TEXT, header.c_str(), sizeof(RBD_HEADER_TEXT))) { | |
234 | CephContext *cct = (CephContext *)io_ctx.cct(); | |
235 | lderr(cct) << "unrecognized header format" << dendl; | |
236 | return -ENXIO; | |
237 | } | |
238 | ||
239 | if (ver) | |
240 | *ver = io_ctx.get_last_version(); | |
241 | ||
242 | return 0; | |
243 | } | |
244 | ||
245 | int read_header(IoCtx& io_ctx, const string& header_oid, | |
246 | struct rbd_obj_header_ondisk *header, uint64_t *ver) | |
247 | { | |
248 | bufferlist header_bl; | |
249 | int r = read_header_bl(io_ctx, header_oid, header_bl, ver); | |
250 | if (r < 0) | |
251 | return r; | |
252 | if (header_bl.length() < (int)sizeof(*header)) | |
253 | return -EIO; | |
254 | memcpy(header, header_bl.c_str(), sizeof(*header)); | |
255 | ||
256 | return 0; | |
257 | } | |
258 | ||
259 | int tmap_set(IoCtx& io_ctx, const string& imgname) | |
260 | { | |
261 | bufferlist cmdbl, emptybl; | |
262 | __u8 c = CEPH_OSD_TMAP_SET; | |
263 | ::encode(c, cmdbl); | |
264 | ::encode(imgname, cmdbl); | |
265 | ::encode(emptybl, cmdbl); | |
266 | return io_ctx.tmap_update(RBD_DIRECTORY, cmdbl); | |
267 | } | |
268 | ||
269 | int tmap_rm(IoCtx& io_ctx, const string& imgname) | |
270 | { | |
271 | bufferlist cmdbl; | |
272 | __u8 c = CEPH_OSD_TMAP_RM; | |
273 | ::encode(c, cmdbl); | |
274 | ::encode(imgname, cmdbl); | |
275 | return io_ctx.tmap_update(RBD_DIRECTORY, cmdbl); | |
276 | } | |
277 | ||
278 | typedef boost::variant<std::string,uint64_t> image_option_value_t; | |
279 | typedef std::map<int,image_option_value_t> image_options_t; | |
280 | typedef std::shared_ptr<image_options_t> image_options_ref; | |
281 | ||
282 | enum image_option_type_t { | |
283 | STR, | |
284 | UINT64, | |
285 | }; | |
286 | ||
287 | const std::map<int, image_option_type_t> IMAGE_OPTIONS_TYPE_MAPPING = { | |
288 | {RBD_IMAGE_OPTION_FORMAT, UINT64}, | |
289 | {RBD_IMAGE_OPTION_FEATURES, UINT64}, | |
290 | {RBD_IMAGE_OPTION_ORDER, UINT64}, | |
291 | {RBD_IMAGE_OPTION_STRIPE_UNIT, UINT64}, | |
292 | {RBD_IMAGE_OPTION_STRIPE_COUNT, UINT64}, | |
293 | {RBD_IMAGE_OPTION_JOURNAL_ORDER, UINT64}, | |
294 | {RBD_IMAGE_OPTION_JOURNAL_SPLAY_WIDTH, UINT64}, | |
295 | {RBD_IMAGE_OPTION_JOURNAL_POOL, STR}, | |
296 | {RBD_IMAGE_OPTION_FEATURES_SET, UINT64}, | |
297 | {RBD_IMAGE_OPTION_FEATURES_CLEAR, UINT64}, | |
298 | {RBD_IMAGE_OPTION_DATA_POOL, STR}, | |
299 | }; | |
300 | ||
301 | std::string image_option_name(int optname) { | |
302 | switch (optname) { | |
303 | case RBD_IMAGE_OPTION_FORMAT: | |
304 | return "format"; | |
305 | case RBD_IMAGE_OPTION_FEATURES: | |
306 | return "features"; | |
307 | case RBD_IMAGE_OPTION_ORDER: | |
308 | return "order"; | |
309 | case RBD_IMAGE_OPTION_STRIPE_UNIT: | |
310 | return "stripe_unit"; | |
311 | case RBD_IMAGE_OPTION_STRIPE_COUNT: | |
312 | return "stripe_count"; | |
313 | case RBD_IMAGE_OPTION_JOURNAL_ORDER: | |
314 | return "journal_order"; | |
315 | case RBD_IMAGE_OPTION_JOURNAL_SPLAY_WIDTH: | |
316 | return "journal_splay_width"; | |
317 | case RBD_IMAGE_OPTION_JOURNAL_POOL: | |
318 | return "journal_pool"; | |
319 | case RBD_IMAGE_OPTION_FEATURES_SET: | |
320 | return "features_set"; | |
321 | case RBD_IMAGE_OPTION_FEATURES_CLEAR: | |
322 | return "features_clear"; | |
323 | case RBD_IMAGE_OPTION_DATA_POOL: | |
324 | return "data_pool"; | |
325 | default: | |
326 | return "unknown (" + stringify(optname) + ")"; | |
327 | } | |
328 | } | |
329 | ||
31f18b77 | 330 | std::ostream &operator<<(std::ostream &os, const ImageOptions &opts) { |
7c673cae FG |
331 | os << "["; |
332 | ||
333 | const char *delimiter = ""; | |
334 | for (auto &i : IMAGE_OPTIONS_TYPE_MAPPING) { | |
335 | if (i.second == STR) { | |
336 | std::string val; | |
337 | if (opts.get(i.first, &val) == 0) { | |
338 | os << delimiter << image_option_name(i.first) << "=" << val; | |
339 | delimiter = ", "; | |
340 | } | |
341 | } else if (i.second == UINT64) { | |
342 | uint64_t val; | |
343 | if (opts.get(i.first, &val) == 0) { | |
344 | os << delimiter << image_option_name(i.first) << "=" << val; | |
345 | delimiter = ", "; | |
346 | } | |
347 | } | |
348 | } | |
349 | ||
350 | os << "]"; | |
351 | ||
352 | return os; | |
353 | } | |
354 | ||
355 | void image_options_create(rbd_image_options_t* opts) | |
356 | { | |
357 | image_options_ref* opts_ = new image_options_ref(new image_options_t()); | |
358 | ||
359 | *opts = static_cast<rbd_image_options_t>(opts_); | |
360 | } | |
361 | ||
362 | void image_options_create_ref(rbd_image_options_t* opts, | |
363 | rbd_image_options_t orig) | |
364 | { | |
365 | image_options_ref* orig_ = static_cast<image_options_ref*>(orig); | |
366 | image_options_ref* opts_ = new image_options_ref(*orig_); | |
367 | ||
368 | *opts = static_cast<rbd_image_options_t>(opts_); | |
369 | } | |
370 | ||
371 | void image_options_copy(rbd_image_options_t* opts, | |
372 | const ImageOptions &orig) | |
373 | { | |
374 | image_options_ref* opts_ = new image_options_ref(new image_options_t()); | |
375 | ||
376 | *opts = static_cast<rbd_image_options_t>(opts_); | |
377 | ||
378 | std::string str_val; | |
379 | uint64_t uint64_val; | |
380 | for (auto &i : IMAGE_OPTIONS_TYPE_MAPPING) { | |
381 | switch (i.second) { | |
382 | case STR: | |
383 | if (orig.get(i.first, &str_val) == 0) { | |
384 | image_options_set(*opts, i.first, str_val); | |
385 | } | |
386 | continue; | |
387 | case UINT64: | |
388 | if (orig.get(i.first, &uint64_val) == 0) { | |
389 | image_options_set(*opts, i.first, uint64_val); | |
390 | } | |
391 | continue; | |
392 | } | |
393 | } | |
394 | } | |
395 | ||
396 | void image_options_destroy(rbd_image_options_t opts) | |
397 | { | |
398 | image_options_ref* opts_ = static_cast<image_options_ref*>(opts); | |
399 | ||
400 | delete opts_; | |
401 | } | |
402 | ||
403 | int image_options_set(rbd_image_options_t opts, int optname, | |
404 | const std::string& optval) | |
405 | { | |
406 | image_options_ref* opts_ = static_cast<image_options_ref*>(opts); | |
407 | ||
408 | std::map<int, image_option_type_t>::const_iterator i = | |
409 | IMAGE_OPTIONS_TYPE_MAPPING.find(optname); | |
410 | ||
411 | if (i == IMAGE_OPTIONS_TYPE_MAPPING.end() || i->second != STR) { | |
412 | return -EINVAL; | |
413 | } | |
414 | ||
415 | (*opts_->get())[optname] = optval; | |
416 | return 0; | |
417 | } | |
418 | ||
419 | int image_options_set(rbd_image_options_t opts, int optname, uint64_t optval) | |
420 | { | |
421 | image_options_ref* opts_ = static_cast<image_options_ref*>(opts); | |
422 | ||
423 | std::map<int, image_option_type_t>::const_iterator i = | |
424 | IMAGE_OPTIONS_TYPE_MAPPING.find(optname); | |
425 | ||
426 | if (i == IMAGE_OPTIONS_TYPE_MAPPING.end() || i->second != UINT64) { | |
427 | return -EINVAL; | |
428 | } | |
429 | ||
430 | (*opts_->get())[optname] = optval; | |
431 | return 0; | |
432 | } | |
433 | ||
434 | int image_options_get(rbd_image_options_t opts, int optname, | |
435 | std::string* optval) | |
436 | { | |
437 | image_options_ref* opts_ = static_cast<image_options_ref*>(opts); | |
438 | ||
439 | std::map<int, image_option_type_t>::const_iterator i = | |
440 | IMAGE_OPTIONS_TYPE_MAPPING.find(optname); | |
441 | ||
442 | if (i == IMAGE_OPTIONS_TYPE_MAPPING.end() || i->second != STR) { | |
443 | return -EINVAL; | |
444 | } | |
445 | ||
446 | image_options_t::const_iterator j = (*opts_)->find(optname); | |
447 | ||
448 | if (j == (*opts_)->end()) { | |
449 | return -ENOENT; | |
450 | } | |
451 | ||
452 | *optval = boost::get<std::string>(j->second); | |
453 | return 0; | |
454 | } | |
455 | ||
456 | int image_options_get(rbd_image_options_t opts, int optname, uint64_t* optval) | |
457 | { | |
458 | image_options_ref* opts_ = static_cast<image_options_ref*>(opts); | |
459 | ||
460 | std::map<int, image_option_type_t>::const_iterator i = | |
461 | IMAGE_OPTIONS_TYPE_MAPPING.find(optname); | |
462 | ||
463 | if (i == IMAGE_OPTIONS_TYPE_MAPPING.end() || i->second != UINT64) { | |
464 | return -EINVAL; | |
465 | } | |
466 | ||
467 | image_options_t::const_iterator j = (*opts_)->find(optname); | |
468 | ||
469 | if (j == (*opts_)->end()) { | |
470 | return -ENOENT; | |
471 | } | |
472 | ||
473 | *optval = boost::get<uint64_t>(j->second); | |
474 | return 0; | |
475 | } | |
476 | ||
477 | int image_options_is_set(rbd_image_options_t opts, int optname, | |
478 | bool* is_set) | |
479 | { | |
480 | if (IMAGE_OPTIONS_TYPE_MAPPING.find(optname) == | |
481 | IMAGE_OPTIONS_TYPE_MAPPING.end()) { | |
482 | return -EINVAL; | |
483 | } | |
484 | ||
485 | image_options_ref* opts_ = static_cast<image_options_ref*>(opts); | |
486 | *is_set = ((*opts_)->find(optname) != (*opts_)->end()); | |
487 | return 0; | |
488 | } | |
489 | ||
490 | int image_options_unset(rbd_image_options_t opts, int optname) | |
491 | { | |
492 | image_options_ref* opts_ = static_cast<image_options_ref*>(opts); | |
493 | ||
494 | std::map<int, image_option_type_t>::const_iterator i = | |
495 | IMAGE_OPTIONS_TYPE_MAPPING.find(optname); | |
496 | ||
497 | if (i == IMAGE_OPTIONS_TYPE_MAPPING.end()) { | |
498 | assert((*opts_)->find(optname) == (*opts_)->end()); | |
499 | return -EINVAL; | |
500 | } | |
501 | ||
502 | image_options_t::const_iterator j = (*opts_)->find(optname); | |
503 | ||
504 | if (j == (*opts_)->end()) { | |
505 | return -ENOENT; | |
506 | } | |
507 | ||
508 | (*opts_)->erase(j); | |
509 | return 0; | |
510 | } | |
511 | ||
512 | void image_options_clear(rbd_image_options_t opts) | |
513 | { | |
514 | image_options_ref* opts_ = static_cast<image_options_ref*>(opts); | |
515 | ||
516 | (*opts_)->clear(); | |
517 | } | |
518 | ||
519 | bool image_options_is_empty(rbd_image_options_t opts) | |
520 | { | |
521 | image_options_ref* opts_ = static_cast<image_options_ref*>(opts); | |
522 | ||
523 | return (*opts_)->empty(); | |
524 | } | |
525 | ||
526 | int list(IoCtx& io_ctx, vector<string>& names) | |
527 | { | |
528 | CephContext *cct = (CephContext *)io_ctx.cct(); | |
529 | ldout(cct, 20) << "list " << &io_ctx << dendl; | |
530 | ||
531 | bufferlist bl; | |
532 | int r = io_ctx.read(RBD_DIRECTORY, bl, 0, 0); | |
31f18b77 FG |
533 | if (r < 0) { |
534 | if (r == -ENOENT) { | |
535 | r = 0; | |
536 | } | |
7c673cae | 537 | return r; |
31f18b77 | 538 | } |
7c673cae FG |
539 | |
540 | // old format images are in a tmap | |
541 | if (bl.length()) { | |
542 | bufferlist::iterator p = bl.begin(); | |
543 | bufferlist header; | |
544 | map<string,bufferlist> m; | |
545 | ::decode(header, p); | |
546 | ::decode(m, p); | |
547 | for (map<string,bufferlist>::iterator q = m.begin(); q != m.end(); ++q) { | |
548 | names.push_back(q->first); | |
549 | } | |
550 | } | |
551 | ||
552 | map<string, string> images; | |
553 | r = api::Image<>::list_images(io_ctx, &images); | |
554 | if (r < 0) { | |
555 | lderr(cct) << "error listing v2 images: " << cpp_strerror(r) << dendl; | |
556 | return r; | |
557 | } | |
558 | for (const auto& img_pair : images) { | |
559 | names.push_back(img_pair.first); | |
560 | } | |
561 | ||
562 | return 0; | |
563 | } | |
564 | ||
565 | int flatten_children(ImageCtx *ictx, const char* snap_name, | |
566 | ProgressContext& pctx) | |
567 | { | |
568 | CephContext *cct = ictx->cct; | |
569 | ldout(cct, 20) << "children flatten " << ictx->name << dendl; | |
570 | ||
b32b8144 FG |
571 | int r = ictx->state->refresh_if_required(); |
572 | if (r < 0) { | |
573 | return r; | |
574 | } | |
575 | ||
7c673cae FG |
576 | RWLock::RLocker l(ictx->snap_lock); |
577 | snap_t snap_id = ictx->get_snap_id(cls::rbd::UserSnapshotNamespace(), snap_name); | |
578 | ParentSpec parent_spec(ictx->md_ctx.get_id(), ictx->id, snap_id); | |
579 | map< pair<int64_t, string>, set<string> > image_info; | |
580 | ||
b32b8144 | 581 | r = api::Image<>::list_children(ictx, parent_spec, &image_info); |
7c673cae FG |
582 | if (r < 0) { |
583 | return r; | |
584 | } | |
585 | ||
586 | size_t size = image_info.size(); | |
587 | if (size == 0) | |
588 | return 0; | |
589 | ||
590 | size_t i = 0; | |
591 | Rados rados(ictx->md_ctx); | |
592 | for ( auto &info : image_info){ | |
593 | string pool = info.first.second; | |
594 | IoCtx ioctx; | |
595 | r = rados.ioctx_create2(info.first.first, ioctx); | |
596 | if (r < 0) { | |
597 | lderr(cct) << "Error accessing child image pool " << pool | |
598 | << dendl; | |
599 | return r; | |
600 | } | |
601 | ||
602 | for (auto &id_it : info.second) { | |
603 | ImageCtx *imctx = new ImageCtx("", id_it, NULL, ioctx, false); | |
604 | int r = imctx->state->open(false); | |
605 | if (r < 0) { | |
606 | lderr(cct) << "error opening image: " | |
607 | << cpp_strerror(r) << dendl; | |
608 | return r; | |
609 | } | |
610 | ||
b32b8144 FG |
611 | if ((imctx->features & RBD_FEATURE_DEEP_FLATTEN) == 0 && |
612 | !imctx->snaps.empty()) { | |
613 | lderr(cct) << "snapshot in-use by " << pool << "/" << imctx->name | |
614 | << dendl; | |
615 | imctx->state->close(); | |
616 | return -EBUSY; | |
617 | } | |
618 | ||
7c673cae FG |
619 | librbd::NoOpProgressContext prog_ctx; |
620 | r = imctx->operations->flatten(prog_ctx); | |
621 | if (r < 0) { | |
622 | lderr(cct) << "error flattening image: " << pool << "/" << id_it | |
623 | << cpp_strerror(r) << dendl; | |
624 | imctx->state->close(); | |
625 | return r; | |
626 | } | |
627 | ||
7c673cae FG |
628 | r = imctx->state->close(); |
629 | if (r < 0) { | |
630 | lderr(cct) << "failed to close image: " << cpp_strerror(r) << dendl; | |
631 | return r; | |
632 | } | |
633 | } | |
634 | pctx.update_progress(++i, size); | |
635 | assert(i <= size); | |
636 | } | |
637 | ||
638 | return 0; | |
639 | } | |
640 | ||
641 | int list_children(ImageCtx *ictx, set<pair<string, string> >& names) | |
642 | { | |
643 | CephContext *cct = ictx->cct; | |
644 | ldout(cct, 20) << "children list " << ictx->name << dendl; | |
645 | ||
b32b8144 FG |
646 | int r = ictx->state->refresh_if_required(); |
647 | if (r < 0) { | |
648 | return r; | |
649 | } | |
650 | ||
7c673cae FG |
651 | RWLock::RLocker l(ictx->snap_lock); |
652 | ParentSpec parent_spec(ictx->md_ctx.get_id(), ictx->id, ictx->snap_id); | |
653 | map< pair<int64_t, string>, set<string> > image_info; | |
654 | ||
b32b8144 | 655 | r = api::Image<>::list_children(ictx, parent_spec, &image_info); |
7c673cae FG |
656 | if (r < 0) { |
657 | return r; | |
658 | } | |
659 | ||
660 | Rados rados(ictx->md_ctx); | |
661 | for ( auto &info : image_info){ | |
662 | IoCtx ioctx; | |
663 | r = rados.ioctx_create2(info.first.first, ioctx); | |
664 | if (r < 0) { | |
665 | lderr(cct) << "Error accessing child image pool " << info.first.second | |
666 | << dendl; | |
667 | return r; | |
668 | } | |
669 | ||
670 | for (auto &id_it : info.second) { | |
671 | string name; | |
672 | r = cls_client::dir_get_name(&ioctx, RBD_DIRECTORY, id_it, &name); | |
673 | if (r < 0) { | |
674 | lderr(cct) << "Error looking up name for image id " << id_it | |
675 | << " in pool " << info.first.second << dendl; | |
676 | return r; | |
677 | } | |
678 | names.insert(make_pair(info.first.second, name)); | |
679 | } | |
680 | } | |
681 | ||
682 | return 0; | |
683 | } | |
684 | ||
685 | int get_snap_namespace(ImageCtx *ictx, | |
686 | const char *snap_name, | |
687 | cls::rbd::SnapshotNamespace *snap_namespace) { | |
688 | ldout(ictx->cct, 20) << "get_snap_namespace " << ictx << " " << snap_name | |
689 | << dendl; | |
690 | ||
691 | int r = ictx->state->refresh_if_required(); | |
692 | if (r < 0) | |
693 | return r; | |
694 | RWLock::RLocker l(ictx->snap_lock); | |
695 | snap_t snap_id = ictx->get_snap_id(*snap_namespace, snap_name); | |
696 | if (snap_id == CEPH_NOSNAP) | |
697 | return -ENOENT; | |
698 | r = ictx->get_snap_namespace(snap_id, snap_namespace); | |
699 | return r; | |
700 | } | |
701 | ||
702 | int snap_is_protected(ImageCtx *ictx, const char *snap_name, bool *is_protected) | |
703 | { | |
704 | ldout(ictx->cct, 20) << "snap_is_protected " << ictx << " " << snap_name | |
705 | << dendl; | |
706 | ||
707 | int r = ictx->state->refresh_if_required(); | |
708 | if (r < 0) | |
709 | return r; | |
710 | ||
711 | RWLock::RLocker l(ictx->snap_lock); | |
712 | snap_t snap_id = ictx->get_snap_id(cls::rbd::UserSnapshotNamespace(), snap_name); | |
713 | if (snap_id == CEPH_NOSNAP) | |
714 | return -ENOENT; | |
715 | bool is_unprotected; | |
716 | r = ictx->is_snap_unprotected(snap_id, &is_unprotected); | |
717 | // consider both PROTECTED or UNPROTECTING to be 'protected', | |
718 | // since in either state they can't be deleted | |
719 | *is_protected = !is_unprotected; | |
720 | return r; | |
721 | } | |
722 | ||
723 | int create_v1(IoCtx& io_ctx, const char *imgname, uint64_t size, int order) | |
724 | { | |
725 | CephContext *cct = (CephContext *)io_ctx.cct(); | |
726 | ||
727 | ldout(cct, 20) << __func__ << " " << &io_ctx << " name = " << imgname | |
728 | << " size = " << size << " order = " << order << dendl; | |
729 | int r = validate_pool(io_ctx, cct); | |
730 | if (r < 0) { | |
731 | return r; | |
732 | } | |
733 | ||
734 | ldout(cct, 2) << "adding rbd image to directory..." << dendl; | |
735 | r = tmap_set(io_ctx, imgname); | |
736 | if (r < 0) { | |
737 | lderr(cct) << "error adding image to directory: " << cpp_strerror(r) | |
738 | << dendl; | |
739 | return r; | |
740 | } | |
741 | ||
742 | Rados rados(io_ctx); | |
743 | uint64_t bid = rados.get_instance_id(); | |
744 | ||
745 | ldout(cct, 2) << "creating rbd image..." << dendl; | |
746 | struct rbd_obj_header_ondisk header; | |
747 | init_rbd_header(header, size, order, bid); | |
748 | ||
749 | bufferlist bl; | |
750 | bl.append((const char *)&header, sizeof(header)); | |
751 | ||
752 | string header_oid = util::old_header_name(imgname); | |
753 | r = io_ctx.write(header_oid, bl, bl.length(), 0); | |
754 | if (r < 0) { | |
755 | lderr(cct) << "Error writing image header: " << cpp_strerror(r) | |
756 | << dendl; | |
757 | int remove_r = tmap_rm(io_ctx, imgname); | |
758 | if (remove_r < 0) { | |
759 | lderr(cct) << "Could not remove image from directory after " | |
760 | << "header creation failed: " | |
761 | << cpp_strerror(remove_r) << dendl; | |
762 | } | |
763 | return r; | |
764 | } | |
765 | ||
766 | ldout(cct, 2) << "done." << dendl; | |
767 | return 0; | |
768 | } | |
769 | ||
770 | int create(librados::IoCtx& io_ctx, const char *imgname, uint64_t size, | |
771 | int *order) | |
772 | { | |
773 | uint64_t order_ = *order; | |
774 | ImageOptions opts; | |
775 | ||
776 | int r = opts.set(RBD_IMAGE_OPTION_ORDER, order_); | |
777 | assert(r == 0); | |
778 | ||
779 | r = create(io_ctx, imgname, "", size, opts, "", "", false); | |
780 | ||
781 | int r1 = opts.get(RBD_IMAGE_OPTION_ORDER, &order_); | |
782 | assert(r1 == 0); | |
783 | *order = order_; | |
784 | ||
785 | return r; | |
786 | } | |
787 | ||
788 | int create(IoCtx& io_ctx, const char *imgname, uint64_t size, | |
789 | bool old_format, uint64_t features, int *order, | |
790 | uint64_t stripe_unit, uint64_t stripe_count) | |
791 | { | |
792 | if (!order) | |
793 | return -EINVAL; | |
794 | ||
795 | uint64_t order_ = *order; | |
796 | uint64_t format = old_format ? 1 : 2; | |
797 | ImageOptions opts; | |
798 | int r; | |
799 | ||
800 | r = opts.set(RBD_IMAGE_OPTION_FORMAT, format); | |
801 | assert(r == 0); | |
802 | r = opts.set(RBD_IMAGE_OPTION_FEATURES, features); | |
803 | assert(r == 0); | |
804 | r = opts.set(RBD_IMAGE_OPTION_ORDER, order_); | |
805 | assert(r == 0); | |
806 | r = opts.set(RBD_IMAGE_OPTION_STRIPE_UNIT, stripe_unit); | |
807 | assert(r == 0); | |
808 | r = opts.set(RBD_IMAGE_OPTION_STRIPE_COUNT, stripe_count); | |
809 | assert(r == 0); | |
810 | ||
811 | r = create(io_ctx, imgname, "", size, opts, "", "", false); | |
812 | ||
813 | int r1 = opts.get(RBD_IMAGE_OPTION_ORDER, &order_); | |
814 | assert(r1 == 0); | |
815 | *order = order_; | |
816 | ||
817 | return r; | |
818 | } | |
819 | ||
820 | int create(IoCtx& io_ctx, const std::string &image_name, | |
821 | const std::string &image_id, uint64_t size, | |
822 | ImageOptions& opts, | |
823 | const std::string &non_primary_global_image_id, | |
824 | const std::string &primary_mirror_uuid, | |
825 | bool skip_mirror_enable) | |
826 | { | |
827 | std::string id(image_id); | |
828 | if (id.empty()) { | |
829 | id = util::generate_image_id(io_ctx); | |
830 | } | |
831 | ||
832 | CephContext *cct = (CephContext *)io_ctx.cct(); | |
833 | ldout(cct, 10) << __func__ << " name=" << image_name << ", " | |
834 | << "id= " << id << ", " | |
835 | << "size=" << size << ", opts=" << opts << dendl; | |
836 | ||
837 | uint64_t format; | |
838 | if (opts.get(RBD_IMAGE_OPTION_FORMAT, &format) != 0) | |
181888fb | 839 | format = cct->_conf->get_val<int64_t>("rbd_default_format"); |
7c673cae FG |
840 | bool old_format = format == 1; |
841 | ||
842 | // make sure it doesn't already exist, in either format | |
843 | int r = detect_format(io_ctx, image_name, NULL, NULL); | |
844 | if (r != -ENOENT) { | |
845 | if (r) { | |
846 | lderr(cct) << "Could not tell if " << image_name << " already exists" | |
847 | << dendl; | |
848 | return r; | |
849 | } | |
850 | lderr(cct) << "rbd image " << image_name << " already exists" << dendl; | |
851 | return -EEXIST; | |
852 | } | |
853 | ||
854 | uint64_t order = 0; | |
855 | if (opts.get(RBD_IMAGE_OPTION_ORDER, &order) != 0 || order == 0) { | |
181888fb | 856 | order = cct->_conf->get_val<int64_t>("rbd_default_order"); |
7c673cae FG |
857 | } |
858 | r = image::CreateRequest<>::validate_order(cct, order); | |
859 | if (r < 0) { | |
860 | return r; | |
861 | } | |
862 | ||
863 | if (old_format) { | |
864 | r = create_v1(io_ctx, image_name.c_str(), size, order); | |
865 | } else { | |
866 | ThreadPool *thread_pool; | |
867 | ContextWQ *op_work_queue; | |
868 | ImageCtx::get_thread_pool_instance(cct, &thread_pool, &op_work_queue); | |
869 | ||
870 | C_SaferCond cond; | |
871 | image::CreateRequest<> *req = image::CreateRequest<>::create( | |
872 | io_ctx, image_name, id, size, opts, non_primary_global_image_id, | |
873 | primary_mirror_uuid, skip_mirror_enable, op_work_queue, &cond); | |
874 | req->send(); | |
875 | ||
876 | r = cond.wait(); | |
877 | } | |
878 | ||
879 | int r1 = opts.set(RBD_IMAGE_OPTION_ORDER, order); | |
880 | assert(r1 == 0); | |
881 | ||
882 | return r; | |
883 | } | |
884 | ||
885 | /* | |
886 | * Parent may be in different pool, hence different IoCtx | |
887 | */ | |
888 | int clone(IoCtx& p_ioctx, const char *p_name, const char *p_snap_name, | |
889 | IoCtx& c_ioctx, const char *c_name, | |
890 | uint64_t features, int *c_order, | |
891 | uint64_t stripe_unit, int stripe_count) | |
892 | { | |
893 | uint64_t order = *c_order; | |
894 | ||
895 | ImageOptions opts; | |
896 | opts.set(RBD_IMAGE_OPTION_FEATURES, features); | |
897 | opts.set(RBD_IMAGE_OPTION_ORDER, order); | |
898 | opts.set(RBD_IMAGE_OPTION_STRIPE_UNIT, stripe_unit); | |
899 | opts.set(RBD_IMAGE_OPTION_STRIPE_COUNT, stripe_count); | |
900 | ||
901 | int r = clone(p_ioctx, p_name, p_snap_name, c_ioctx, c_name, opts); | |
902 | opts.get(RBD_IMAGE_OPTION_ORDER, &order); | |
903 | *c_order = order; | |
904 | return r; | |
905 | } | |
906 | ||
907 | int clone(IoCtx& p_ioctx, const char *p_name, const char *p_snap_name, | |
908 | IoCtx& c_ioctx, const char *c_name, ImageOptions& c_opts) | |
909 | { | |
910 | CephContext *cct = (CephContext *)p_ioctx.cct(); | |
911 | if (p_snap_name == NULL) { | |
912 | lderr(cct) << "image to be cloned must be a snapshot" << dendl; | |
913 | return -EINVAL; | |
914 | } | |
915 | ||
916 | // make sure parent snapshot exists | |
917 | ImageCtx *p_imctx = new ImageCtx(p_name, "", p_snap_name, p_ioctx, true); | |
918 | int r = p_imctx->state->open(false); | |
919 | if (r < 0) { | |
920 | lderr(cct) << "error opening parent image: " | |
921 | << cpp_strerror(r) << dendl; | |
922 | return r; | |
923 | } | |
924 | ||
925 | r = clone(p_imctx, c_ioctx, c_name, "", c_opts, "", ""); | |
926 | ||
927 | int close_r = p_imctx->state->close(); | |
928 | if (r == 0 && close_r < 0) { | |
929 | r = close_r; | |
930 | } | |
931 | ||
932 | if (r < 0) { | |
933 | return r; | |
934 | } | |
935 | return 0; | |
936 | } | |
937 | ||
938 | int clone(ImageCtx *p_imctx, IoCtx& c_ioctx, const std::string &c_name, | |
939 | const std::string &c_id, ImageOptions& c_opts, | |
940 | const std::string &non_primary_global_image_id, | |
941 | const std::string &primary_mirror_uuid) | |
942 | { | |
943 | std::string id(c_id); | |
944 | if (id.empty()) { | |
945 | id = util::generate_image_id(c_ioctx); | |
946 | } | |
947 | ||
948 | CephContext *cct = (CephContext *)c_ioctx.cct(); | |
949 | ldout(cct, 10) << __func__ << " " | |
950 | << "c_name=" << c_name << ", " | |
951 | << "c_id= " << c_id << ", " | |
952 | << "c_opts=" << c_opts << dendl; | |
953 | ||
954 | ThreadPool *thread_pool; | |
955 | ContextWQ *op_work_queue; | |
956 | ImageCtx::get_thread_pool_instance(cct, &thread_pool, &op_work_queue); | |
957 | ||
958 | C_SaferCond cond; | |
959 | auto *req = image::CloneRequest<>::create( | |
960 | p_imctx, c_ioctx, c_name, id, c_opts, | |
961 | non_primary_global_image_id, primary_mirror_uuid, op_work_queue, &cond); | |
962 | req->send(); | |
963 | ||
964 | return cond.wait(); | |
965 | } | |
966 | ||
967 | int rename(IoCtx& io_ctx, const char *srcname, const char *dstname) | |
968 | { | |
969 | CephContext *cct = (CephContext *)io_ctx.cct(); | |
970 | ldout(cct, 20) << "rename " << &io_ctx << " " << srcname << " -> " | |
971 | << dstname << dendl; | |
972 | ||
973 | ImageCtx *ictx = new ImageCtx(srcname, "", "", io_ctx, false); | |
974 | int r = ictx->state->open(false); | |
975 | if (r < 0) { | |
181888fb | 976 | lderr(cct) << "error opening source image: " << cpp_strerror(r) << dendl; |
7c673cae FG |
977 | return r; |
978 | } | |
979 | BOOST_SCOPE_EXIT((ictx)) { | |
980 | ictx->state->close(); | |
981 | } BOOST_SCOPE_EXIT_END | |
982 | ||
983 | return ictx->operations->rename(dstname); | |
984 | } | |
985 | ||
986 | int info(ImageCtx *ictx, image_info_t& info, size_t infosize) | |
987 | { | |
988 | ldout(ictx->cct, 20) << "info " << ictx << dendl; | |
989 | ||
990 | int r = ictx->state->refresh_if_required(); | |
991 | if (r < 0) | |
992 | return r; | |
993 | ||
994 | image_info(ictx, info, infosize); | |
995 | return 0; | |
996 | } | |
997 | ||
998 | int get_old_format(ImageCtx *ictx, uint8_t *old) | |
999 | { | |
1000 | int r = ictx->state->refresh_if_required(); | |
1001 | if (r < 0) | |
1002 | return r; | |
1003 | *old = ictx->old_format; | |
1004 | return 0; | |
1005 | } | |
1006 | ||
1007 | int get_size(ImageCtx *ictx, uint64_t *size) | |
1008 | { | |
1009 | int r = ictx->state->refresh_if_required(); | |
1010 | if (r < 0) | |
1011 | return r; | |
1012 | RWLock::RLocker l2(ictx->snap_lock); | |
1013 | *size = ictx->get_image_size(ictx->snap_id); | |
1014 | return 0; | |
1015 | } | |
1016 | ||
1017 | int get_features(ImageCtx *ictx, uint64_t *features) | |
1018 | { | |
1019 | int r = ictx->state->refresh_if_required(); | |
1020 | if (r < 0) | |
1021 | return r; | |
1022 | RWLock::RLocker l(ictx->snap_lock); | |
1023 | *features = ictx->features; | |
1024 | return 0; | |
1025 | } | |
1026 | ||
1027 | int get_overlap(ImageCtx *ictx, uint64_t *overlap) | |
1028 | { | |
1029 | int r = ictx->state->refresh_if_required(); | |
1030 | if (r < 0) | |
1031 | return r; | |
1032 | RWLock::RLocker l(ictx->snap_lock); | |
1033 | RWLock::RLocker l2(ictx->parent_lock); | |
1034 | return ictx->get_parent_overlap(ictx->snap_id, overlap); | |
1035 | } | |
1036 | ||
1037 | int get_parent_info(ImageCtx *ictx, string *parent_pool_name, | |
1038 | string *parent_name, string *parent_id, | |
1039 | string *parent_snap_name) | |
1040 | { | |
1041 | int r = ictx->state->refresh_if_required(); | |
1042 | if (r < 0) | |
1043 | return r; | |
1044 | ||
1045 | RWLock::RLocker l(ictx->snap_lock); | |
1046 | RWLock::RLocker l2(ictx->parent_lock); | |
1047 | if (ictx->parent == NULL) { | |
1048 | return -ENOENT; | |
1049 | } | |
1050 | ||
1051 | ParentSpec parent_spec; | |
1052 | ||
1053 | if (ictx->snap_id == CEPH_NOSNAP) { | |
1054 | parent_spec = ictx->parent_md.spec; | |
1055 | } else { | |
1056 | r = ictx->get_parent_spec(ictx->snap_id, &parent_spec); | |
1057 | if (r < 0) { | |
1058 | lderr(ictx->cct) << "Can't find snapshot id = " << ictx->snap_id | |
1059 | << dendl; | |
1060 | return r; | |
1061 | } | |
1062 | if (parent_spec.pool_id == -1) | |
1063 | return -ENOENT; | |
1064 | } | |
1065 | if (parent_pool_name) { | |
1066 | Rados rados(ictx->md_ctx); | |
1067 | r = rados.pool_reverse_lookup(parent_spec.pool_id, | |
1068 | parent_pool_name); | |
1069 | if (r < 0) { | |
1070 | lderr(ictx->cct) << "error looking up pool name: " << cpp_strerror(r) | |
1071 | << dendl; | |
1072 | return r; | |
1073 | } | |
1074 | } | |
1075 | ||
1076 | if (parent_snap_name) { | |
1077 | RWLock::RLocker l(ictx->parent->snap_lock); | |
1078 | r = ictx->parent->get_snap_name(parent_spec.snap_id, | |
1079 | parent_snap_name); | |
1080 | if (r < 0) { | |
1081 | lderr(ictx->cct) << "error finding parent snap name: " | |
1082 | << cpp_strerror(r) << dendl; | |
1083 | return r; | |
1084 | } | |
1085 | } | |
1086 | ||
1087 | if (parent_name) { | |
1088 | RWLock::RLocker snap_locker(ictx->parent->snap_lock); | |
1089 | *parent_name = ictx->parent->name; | |
1090 | } | |
1091 | if (parent_id) { | |
1092 | *parent_id = ictx->parent->id; | |
1093 | } | |
1094 | ||
1095 | return 0; | |
1096 | } | |
1097 | ||
1098 | int get_flags(ImageCtx *ictx, uint64_t *flags) | |
1099 | { | |
1100 | int r = ictx->state->refresh_if_required(); | |
1101 | if (r < 0) { | |
1102 | return r; | |
1103 | } | |
1104 | ||
1105 | RWLock::RLocker l2(ictx->snap_lock); | |
1106 | return ictx->get_flags(ictx->snap_id, flags); | |
1107 | } | |
1108 | ||
1109 | int set_image_notification(ImageCtx *ictx, int fd, int type) | |
1110 | { | |
1111 | CephContext *cct = ictx->cct; | |
1112 | ldout(cct, 20) << __func__ << " " << ictx << " fd " << fd << " type" << type << dendl; | |
1113 | ||
1114 | int r = ictx->state->refresh_if_required(); | |
1115 | if (r < 0) { | |
1116 | return r; | |
1117 | } | |
1118 | ||
1119 | if (ictx->event_socket.is_valid()) | |
1120 | return -EINVAL; | |
1121 | return ictx->event_socket.init(fd, type); | |
1122 | } | |
1123 | ||
1124 | int is_exclusive_lock_owner(ImageCtx *ictx, bool *is_owner) | |
1125 | { | |
1126 | *is_owner = false; | |
1127 | ||
1128 | RWLock::RLocker owner_locker(ictx->owner_lock); | |
1129 | if (ictx->exclusive_lock == nullptr || | |
1130 | !ictx->exclusive_lock->is_lock_owner()) { | |
1131 | return 0; | |
1132 | } | |
1133 | ||
1134 | // might have been blacklisted by peer -- ensure we still own | |
1135 | // the lock by pinging the OSD | |
1136 | int r = ictx->exclusive_lock->assert_header_locked(); | |
31f18b77 FG |
1137 | if (r == -EBUSY || r == -ENOENT) { |
1138 | return 0; | |
1139 | } else if (r < 0) { | |
7c673cae FG |
1140 | return r; |
1141 | } | |
1142 | ||
1143 | *is_owner = true; | |
1144 | return 0; | |
1145 | } | |
1146 | ||
1147 | int lock_acquire(ImageCtx *ictx, rbd_lock_mode_t lock_mode) | |
1148 | { | |
1149 | CephContext *cct = ictx->cct; | |
1150 | ldout(cct, 20) << __func__ << ": ictx=" << ictx << ", " | |
1151 | << "lock_mode=" << lock_mode << dendl; | |
1152 | ||
1153 | if (lock_mode != RBD_LOCK_MODE_EXCLUSIVE) { | |
1154 | return -EOPNOTSUPP; | |
1155 | } | |
1156 | ||
1157 | C_SaferCond lock_ctx; | |
1158 | { | |
1159 | RWLock::WLocker l(ictx->owner_lock); | |
1160 | ||
1161 | if (ictx->exclusive_lock == nullptr) { | |
1162 | lderr(cct) << "exclusive-lock feature is not enabled" << dendl; | |
1163 | return -EINVAL; | |
1164 | } | |
1165 | ||
1166 | if (ictx->get_exclusive_lock_policy()->may_auto_request_lock()) { | |
1167 | ictx->set_exclusive_lock_policy( | |
1168 | new exclusive_lock::StandardPolicy(ictx)); | |
1169 | } | |
1170 | ||
1171 | if (ictx->exclusive_lock->is_lock_owner()) { | |
1172 | return 0; | |
1173 | } | |
1174 | ||
1175 | ictx->exclusive_lock->acquire_lock(&lock_ctx); | |
1176 | } | |
1177 | ||
1178 | int r = lock_ctx.wait(); | |
1179 | if (r < 0) { | |
1180 | lderr(cct) << "failed to request exclusive lock: " << cpp_strerror(r) | |
1181 | << dendl; | |
1182 | return r; | |
1183 | } | |
1184 | ||
1185 | RWLock::RLocker l(ictx->owner_lock); | |
1186 | ||
1187 | if (ictx->exclusive_lock == nullptr || | |
1188 | !ictx->exclusive_lock->is_lock_owner()) { | |
1189 | lderr(cct) << "failed to acquire exclusive lock" << dendl; | |
1190 | return -EROFS; | |
1191 | } | |
1192 | ||
1193 | return 0; | |
1194 | } | |
1195 | ||
1196 | int lock_release(ImageCtx *ictx) | |
1197 | { | |
1198 | CephContext *cct = ictx->cct; | |
1199 | ldout(cct, 20) << __func__ << ": ictx=" << ictx << dendl; | |
1200 | ||
1201 | C_SaferCond lock_ctx; | |
1202 | { | |
1203 | RWLock::WLocker l(ictx->owner_lock); | |
1204 | ||
1205 | if (ictx->exclusive_lock == nullptr || | |
1206 | !ictx->exclusive_lock->is_lock_owner()) { | |
1207 | lderr(cct) << "not exclusive lock owner" << dendl; | |
1208 | return -EINVAL; | |
1209 | } | |
1210 | ||
1211 | ictx->exclusive_lock->release_lock(&lock_ctx); | |
1212 | } | |
1213 | ||
1214 | int r = lock_ctx.wait(); | |
1215 | if (r < 0) { | |
1216 | lderr(cct) << "failed to release exclusive lock: " << cpp_strerror(r) | |
1217 | << dendl; | |
1218 | return r; | |
1219 | } | |
1220 | return 0; | |
1221 | } | |
1222 | ||
1223 | int lock_get_owners(ImageCtx *ictx, rbd_lock_mode_t *lock_mode, | |
1224 | std::list<std::string> *lock_owners) | |
1225 | { | |
1226 | CephContext *cct = ictx->cct; | |
1227 | ldout(cct, 20) << __func__ << ": ictx=" << ictx << dendl; | |
1228 | ||
1229 | if (!ictx->test_features(RBD_FEATURE_EXCLUSIVE_LOCK)) { | |
1230 | lderr(cct) << "exclusive-lock feature is not enabled" << dendl; | |
1231 | return -EINVAL; | |
1232 | } | |
1233 | ||
1234 | managed_lock::Locker locker; | |
1235 | C_SaferCond get_owner_ctx; | |
1236 | ExclusiveLock<>(*ictx).get_locker(&locker, &get_owner_ctx); | |
1237 | int r = get_owner_ctx.wait(); | |
1238 | if (r == -ENOENT) { | |
1239 | return r; | |
1240 | } else if (r < 0) { | |
1241 | lderr(cct) << "failed to determine current lock owner: " | |
1242 | << cpp_strerror(r) << dendl; | |
1243 | return r; | |
1244 | } | |
1245 | ||
1246 | *lock_mode = RBD_LOCK_MODE_EXCLUSIVE; | |
1247 | lock_owners->clear(); | |
1248 | lock_owners->emplace_back(locker.address); | |
1249 | return 0; | |
1250 | } | |
1251 | ||
1252 | int lock_break(ImageCtx *ictx, rbd_lock_mode_t lock_mode, | |
1253 | const std::string &lock_owner) | |
1254 | { | |
1255 | CephContext *cct = ictx->cct; | |
1256 | ldout(cct, 20) << __func__ << ": ictx=" << ictx << ", " | |
1257 | << "lock_mode=" << lock_mode << ", " | |
1258 | << "lock_owner=" << lock_owner << dendl; | |
1259 | ||
1260 | if (lock_mode != RBD_LOCK_MODE_EXCLUSIVE) { | |
1261 | return -EOPNOTSUPP; | |
1262 | } | |
1263 | ||
1264 | if (ictx->read_only) { | |
1265 | return -EROFS; | |
1266 | } | |
1267 | ||
1268 | managed_lock::Locker locker; | |
1269 | C_SaferCond get_owner_ctx; | |
1270 | { | |
1271 | RWLock::RLocker l(ictx->owner_lock); | |
1272 | ||
1273 | if (ictx->exclusive_lock == nullptr) { | |
1274 | lderr(cct) << "exclusive-lock feature is not enabled" << dendl; | |
1275 | return -EINVAL; | |
1276 | } | |
1277 | ||
1278 | ictx->exclusive_lock->get_locker(&locker, &get_owner_ctx); | |
1279 | } | |
1280 | int r = get_owner_ctx.wait(); | |
1281 | if (r == -ENOENT) { | |
1282 | return r; | |
1283 | } else if (r < 0) { | |
1284 | lderr(cct) << "failed to determine current lock owner: " | |
1285 | << cpp_strerror(r) << dendl; | |
1286 | return r; | |
1287 | } | |
1288 | ||
1289 | if (locker.address != lock_owner) { | |
1290 | return -EBUSY; | |
1291 | } | |
1292 | ||
1293 | C_SaferCond break_ctx; | |
1294 | { | |
1295 | RWLock::RLocker l(ictx->owner_lock); | |
1296 | ||
1297 | if (ictx->exclusive_lock == nullptr) { | |
1298 | lderr(cct) << "exclusive-lock feature is not enabled" << dendl; | |
1299 | return -EINVAL; | |
1300 | } | |
1301 | ||
1302 | ictx->exclusive_lock->break_lock(locker, true, &break_ctx); | |
1303 | } | |
1304 | r = break_ctx.wait(); | |
1305 | if (r == -ENOENT) { | |
1306 | return r; | |
1307 | } else if (r < 0) { | |
1308 | lderr(cct) << "failed to break lock: " << cpp_strerror(r) << dendl; | |
1309 | return r; | |
1310 | } | |
1311 | return 0; | |
1312 | } | |
1313 | ||
1314 | int remove(IoCtx& io_ctx, const std::string &image_name, | |
1315 | const std::string &image_id, ProgressContext& prog_ctx, | |
1316 | bool force, bool from_trash_remove) | |
1317 | { | |
1318 | CephContext *cct((CephContext *)io_ctx.cct()); | |
1319 | ldout(cct, 20) << "remove " << &io_ctx << " " | |
1320 | << (image_id.empty() ? image_name : image_id) << dendl; | |
1321 | ||
1322 | ThreadPool *thread_pool; | |
1323 | ContextWQ *op_work_queue; | |
1324 | ImageCtx::get_thread_pool_instance(cct, &thread_pool, &op_work_queue); | |
1325 | ||
1326 | C_SaferCond cond; | |
1327 | auto req = librbd::image::RemoveRequest<>::create( | |
1328 | io_ctx, image_name, image_id, force, from_trash_remove, prog_ctx, | |
1329 | op_work_queue, &cond); | |
1330 | req->send(); | |
1331 | ||
1332 | return cond.wait(); | |
1333 | } | |
1334 | ||
1335 | int trash_move(librados::IoCtx &io_ctx, rbd_trash_image_source_t source, | |
1336 | const std::string &image_name, uint64_t delay) { | |
1337 | CephContext *cct((CephContext *)io_ctx.cct()); | |
1338 | ldout(cct, 20) << "trash_move " << &io_ctx << " " << image_name | |
1339 | << dendl; | |
1340 | ||
1341 | std::string image_id; | |
1342 | ImageCtx *ictx = new ImageCtx(image_name, "", nullptr, io_ctx, false); | |
1343 | int r = ictx->state->open(true); | |
1344 | if (r < 0) { | |
1345 | ictx = nullptr; | |
1346 | ||
1347 | if (r != -ENOENT) { | |
1348 | ldout(cct, 2) << "error opening image: " << cpp_strerror(-r) << dendl; | |
1349 | return r; | |
1350 | } | |
1351 | ||
1352 | // try to get image id from the directory | |
1353 | r = cls_client::dir_get_id(&io_ctx, RBD_DIRECTORY, image_name, &image_id); | |
1354 | if (r < 0) { | |
1355 | if (r != -ENOENT) { | |
1356 | ldout(cct, 2) << "error reading image id from dirctory: " | |
1357 | << cpp_strerror(-r) << dendl; | |
1358 | } | |
1359 | return r; | |
1360 | } | |
1361 | } else { | |
1362 | if (ictx->old_format) { | |
1363 | ictx->state->close(); | |
1364 | return -EOPNOTSUPP; | |
1365 | } | |
1366 | ||
1367 | image_id = ictx->id; | |
1368 | ictx->owner_lock.get_read(); | |
1369 | if (ictx->exclusive_lock != nullptr) { | |
b32b8144 FG |
1370 | r = ictx->operations->prepare_image_update(false); |
1371 | if (r < 0) { | |
7c673cae FG |
1372 | lderr(cct) << "cannot obtain exclusive lock - not removing" << dendl; |
1373 | ictx->owner_lock.put_read(); | |
1374 | ictx->state->close(); | |
1375 | return -EBUSY; | |
1376 | } | |
1377 | } | |
1378 | } | |
1379 | ||
1380 | BOOST_SCOPE_EXIT_ALL(ictx, cct) { | |
1381 | if (ictx == nullptr) | |
1382 | return; | |
1383 | ||
1384 | bool is_locked = ictx->exclusive_lock != nullptr && | |
1385 | ictx->exclusive_lock->is_lock_owner(); | |
1386 | if (is_locked) { | |
1387 | C_SaferCond ctx; | |
31f18b77 FG |
1388 | auto exclusive_lock = ictx->exclusive_lock; |
1389 | exclusive_lock->shut_down(&ctx); | |
7c673cae FG |
1390 | ictx->owner_lock.put_read(); |
1391 | int r = ctx.wait(); | |
1392 | if (r < 0) { | |
1393 | lderr(cct) << "error shutting down exclusive lock" << dendl; | |
1394 | } | |
31f18b77 | 1395 | delete exclusive_lock; |
7c673cae FG |
1396 | } else { |
1397 | ictx->owner_lock.put_read(); | |
1398 | } | |
1399 | ictx->state->close(); | |
1400 | }; | |
1401 | ||
1402 | ldout(cct, 2) << "adding image entry to rbd_trash" << dendl; | |
1403 | utime_t ts = ceph_clock_now(); | |
1404 | utime_t deferment_end_time = ts; | |
1405 | deferment_end_time += (double)delay; | |
1406 | cls::rbd::TrashImageSource trash_source = | |
1407 | static_cast<cls::rbd::TrashImageSource>(source); | |
1408 | cls::rbd::TrashImageSpec trash_spec(trash_source, image_name, ts, | |
1409 | deferment_end_time); | |
1410 | r = cls_client::trash_add(&io_ctx, image_id, trash_spec); | |
1411 | if (r < 0 && r != -EEXIST) { | |
1412 | lderr(cct) << "error adding image " << image_name << " to rbd_trash" | |
1413 | << dendl; | |
1414 | return r; | |
1415 | } else if (r == -EEXIST) { | |
1416 | ldout(cct, 10) << "found previous unfinished deferred remove for image:" | |
1417 | << image_id << dendl; | |
1418 | // continue with removing image from directory | |
1419 | } | |
1420 | ||
1421 | ldout(cct, 2) << "removing id object..." << dendl; | |
1422 | r = io_ctx.remove(util::id_obj_name(image_name)); | |
1423 | if (r < 0 && r != -ENOENT) { | |
1424 | lderr(cct) << "error removing id object: " << cpp_strerror(r) | |
1425 | << dendl; | |
1426 | return r; | |
1427 | } | |
1428 | ||
1429 | ldout(cct, 2) << "removing rbd image from v2 directory..." << dendl; | |
1430 | r = cls_client::dir_remove_image(&io_ctx, RBD_DIRECTORY, image_name, | |
1431 | image_id); | |
1432 | if (r < 0) { | |
1433 | if (r != -ENOENT) { | |
1434 | lderr(cct) << "error removing image from v2 directory: " | |
1435 | << cpp_strerror(-r) << dendl; | |
1436 | } | |
1437 | return r; | |
1438 | } | |
1439 | ||
1440 | return 0; | |
1441 | } | |
1442 | ||
1443 | int trash_get(IoCtx &io_ctx, const std::string &id, | |
1444 | trash_image_info_t *info) { | |
1445 | CephContext *cct((CephContext *)io_ctx.cct()); | |
1446 | ldout(cct, 20) << __func__ << " " << &io_ctx << dendl; | |
1447 | ||
1448 | cls::rbd::TrashImageSpec spec; | |
1449 | int r = cls_client::trash_get(&io_ctx, id, &spec); | |
1450 | if (r == -ENOENT) { | |
1451 | return r; | |
1452 | } else if (r < 0) { | |
1453 | lderr(cct) << "error retrieving trash entry: " << cpp_strerror(r) | |
1454 | << dendl; | |
1455 | return r; | |
1456 | } | |
1457 | ||
1458 | rbd_trash_image_source_t source = static_cast<rbd_trash_image_source_t>( | |
1459 | spec.source); | |
1460 | *info = trash_image_info_t{id, spec.name, source, spec.deletion_time.sec(), | |
1461 | spec.deferment_end_time.sec()}; | |
1462 | return 0; | |
1463 | } | |
1464 | ||
1465 | int trash_list(IoCtx &io_ctx, vector<trash_image_info_t> &entries) { | |
1466 | CephContext *cct((CephContext *)io_ctx.cct()); | |
1467 | ldout(cct, 20) << "trash_list " << &io_ctx << dendl; | |
1468 | ||
c07f9fc5 FG |
1469 | bool more_entries; |
1470 | uint32_t max_read = 1024; | |
1471 | std::string last_read = ""; | |
1472 | do { | |
1473 | map<string, cls::rbd::TrashImageSpec> trash_entries; | |
1474 | int r = cls_client::trash_list(&io_ctx, last_read, max_read, | |
1475 | &trash_entries); | |
1476 | if (r < 0 && r != -ENOENT) { | |
1477 | lderr(cct) << "error listing rbd trash entries: " << cpp_strerror(r) | |
7c673cae | 1478 | << dendl; |
c07f9fc5 FG |
1479 | return r; |
1480 | } else if (r == -ENOENT) { | |
1481 | break; | |
7c673cae | 1482 | } |
7c673cae | 1483 | |
c07f9fc5 FG |
1484 | if (trash_entries.empty()) { |
1485 | break; | |
1486 | } | |
1487 | ||
1488 | for (const auto &entry : trash_entries) { | |
1489 | rbd_trash_image_source_t source = | |
1490 | static_cast<rbd_trash_image_source_t>(entry.second.source); | |
1491 | entries.push_back({entry.first, entry.second.name, source, | |
1492 | entry.second.deletion_time.sec(), | |
1493 | entry.second.deferment_end_time.sec()}); | |
1494 | } | |
1495 | last_read = trash_entries.rbegin()->first; | |
1496 | more_entries = (trash_entries.size() >= max_read); | |
1497 | } while (more_entries); | |
1498 | ||
7c673cae FG |
1499 | return 0; |
1500 | } | |
1501 | ||
1502 | int trash_remove(IoCtx &io_ctx, const std::string &image_id, bool force, | |
1503 | ProgressContext& prog_ctx) { | |
1504 | CephContext *cct((CephContext *)io_ctx.cct()); | |
1505 | ldout(cct, 20) << "trash_remove " << &io_ctx << " " << image_id | |
1506 | << " " << force << dendl; | |
1507 | ||
1508 | cls::rbd::TrashImageSpec trash_spec; | |
1509 | int r = cls_client::trash_get(&io_ctx, image_id, &trash_spec); | |
1510 | if (r < 0) { | |
1511 | lderr(cct) << "error getting image id " << image_id | |
1512 | << " info from trash: " << cpp_strerror(r) << dendl; | |
1513 | return r; | |
1514 | } | |
1515 | ||
1516 | utime_t now = ceph_clock_now(); | |
1517 | if (now < trash_spec.deferment_end_time && !force) { | |
1518 | lderr(cct) << "error: deferment time has not expired." << dendl; | |
1519 | return -EPERM; | |
1520 | } | |
1521 | ||
1522 | r = remove(io_ctx, "", image_id, prog_ctx, false, true); | |
1523 | if (r < 0) { | |
1524 | lderr(cct) << "error removing image " << image_id | |
1525 | << ", which is pending deletion" << dendl; | |
1526 | return r; | |
1527 | } | |
1528 | r = cls_client::trash_remove(&io_ctx, image_id); | |
1529 | if (r < 0 && r != -ENOENT) { | |
1530 | lderr(cct) << "error removing image " << image_id | |
1531 | << " from rbd_trash object" << dendl; | |
1532 | return r; | |
1533 | } | |
1534 | return 0; | |
1535 | } | |
1536 | ||
1537 | int trash_restore(librados::IoCtx &io_ctx, const std::string &image_id, | |
1538 | const std::string &image_new_name) { | |
1539 | CephContext *cct((CephContext *)io_ctx.cct()); | |
1540 | ldout(cct, 20) << "trash_restore " << &io_ctx << " " << image_id << " " | |
1541 | << image_new_name << dendl; | |
1542 | ||
1543 | cls::rbd::TrashImageSpec trash_spec; | |
1544 | int r = cls_client::trash_get(&io_ctx, image_id, &trash_spec); | |
1545 | if (r < 0) { | |
1546 | lderr(cct) << "error getting image id " << image_id | |
1547 | << " info from trash: " << cpp_strerror(r) << dendl; | |
1548 | return r; | |
1549 | } | |
1550 | ||
1551 | std::string image_name = image_new_name; | |
1552 | if (image_name.empty()) { | |
1553 | // if user didn't specify a new name, let's try using the old name | |
1554 | image_name = trash_spec.name; | |
1555 | ldout(cct, 20) << "restoring image id " << image_id << " with name " | |
1556 | << image_name << dendl; | |
1557 | } | |
1558 | ||
1559 | // check if no image exists with the same name | |
1560 | bool create_id_obj = true; | |
1561 | std::string existing_id; | |
1562 | r = cls_client::get_id(&io_ctx, util::id_obj_name(image_name), &existing_id); | |
1563 | if (r < 0 && r != -ENOENT) { | |
1564 | lderr(cct) << "error checking if image " << image_name << " exists: " | |
1565 | << cpp_strerror(r) << dendl; | |
1566 | return r; | |
1567 | } else if (r != -ENOENT){ | |
1568 | // checking if we are recovering from an incomplete restore | |
1569 | if (existing_id != image_id) { | |
1570 | ldout(cct, 2) << "an image with the same name already exists" << dendl; | |
1571 | return -EEXIST; | |
1572 | } | |
1573 | create_id_obj = false; | |
1574 | } | |
1575 | ||
1576 | if (create_id_obj) { | |
1577 | ldout(cct, 2) << "adding id object" << dendl; | |
1578 | librados::ObjectWriteOperation op; | |
1579 | op.create(true); | |
1580 | cls_client::set_id(&op, image_id); | |
1581 | r = io_ctx.operate(util::id_obj_name(image_name), &op); | |
1582 | if (r < 0) { | |
1583 | lderr(cct) << "error adding id object for image " << image_name | |
1584 | << ": " << cpp_strerror(r) << dendl; | |
1585 | return r; | |
1586 | } | |
1587 | } | |
1588 | ||
1589 | ldout(cct, 2) << "adding rbd image from v2 directory..." << dendl; | |
1590 | r = cls_client::dir_add_image(&io_ctx, RBD_DIRECTORY, image_name, | |
1591 | image_id); | |
1592 | if (r < 0 && r != -EEXIST) { | |
1593 | lderr(cct) << "error adding image to v2 directory: " | |
1594 | << cpp_strerror(r) << dendl; | |
1595 | return r; | |
1596 | } | |
1597 | ||
1598 | ldout(cct, 2) << "removing image from trash..." << dendl; | |
1599 | r = cls_client::trash_remove(&io_ctx, image_id); | |
1600 | if (r < 0 && r != -ENOENT) { | |
1601 | lderr(cct) << "error removing image id " << image_id << " from trash: " | |
1602 | << cpp_strerror(r) << dendl; | |
1603 | return r; | |
1604 | } | |
1605 | ||
1606 | return 0; | |
1607 | } | |
1608 | ||
1609 | int snap_list(ImageCtx *ictx, vector<snap_info_t>& snaps) | |
1610 | { | |
1611 | ldout(ictx->cct, 20) << "snap_list " << ictx << dendl; | |
1612 | ||
1613 | int r = ictx->state->refresh_if_required(); | |
1614 | if (r < 0) | |
1615 | return r; | |
1616 | ||
1617 | RWLock::RLocker l(ictx->snap_lock); | |
1618 | for (map<snap_t, SnapInfo>::iterator it = ictx->snap_info.begin(); | |
1619 | it != ictx->snap_info.end(); ++it) { | |
1620 | snap_info_t info; | |
1621 | info.name = it->second.name; | |
1622 | info.id = it->first; | |
1623 | info.size = it->second.size; | |
1624 | snaps.push_back(info); | |
1625 | } | |
1626 | ||
1627 | return 0; | |
1628 | } | |
1629 | ||
1630 | int snap_exists(ImageCtx *ictx, const cls::rbd::SnapshotNamespace& snap_namespace, | |
1631 | const char *snap_name, bool *exists) | |
1632 | { | |
1633 | ldout(ictx->cct, 20) << "snap_exists " << ictx << " " << snap_name << dendl; | |
1634 | ||
1635 | int r = ictx->state->refresh_if_required(); | |
1636 | if (r < 0) | |
1637 | return r; | |
1638 | ||
1639 | RWLock::RLocker l(ictx->snap_lock); | |
1640 | *exists = ictx->get_snap_id(snap_namespace, snap_name) != CEPH_NOSNAP; | |
1641 | return 0; | |
1642 | } | |
1643 | ||
1644 | int snap_remove(ImageCtx *ictx, const char *snap_name, uint32_t flags, | |
1645 | ProgressContext& pctx) | |
1646 | { | |
1647 | ldout(ictx->cct, 20) << "snap_remove " << ictx << " " << snap_name << " flags: " << flags << dendl; | |
1648 | ||
1649 | int r = 0; | |
1650 | ||
1651 | r = ictx->state->refresh_if_required(); | |
1652 | if (r < 0) | |
1653 | return r; | |
1654 | ||
1655 | if (flags & RBD_SNAP_REMOVE_FLATTEN) { | |
1656 | r = flatten_children(ictx, snap_name, pctx); | |
1657 | if (r < 0) { | |
1658 | return r; | |
1659 | } | |
1660 | } | |
1661 | ||
1662 | bool is_protected; | |
1663 | r = snap_is_protected(ictx, snap_name, &is_protected); | |
1664 | if (r < 0) { | |
1665 | return r; | |
1666 | } | |
1667 | ||
1668 | if (is_protected && flags & RBD_SNAP_REMOVE_UNPROTECT) { | |
1669 | r = ictx->operations->snap_unprotect(cls::rbd::UserSnapshotNamespace(), snap_name); | |
1670 | if (r < 0) { | |
1671 | lderr(ictx->cct) << "failed to unprotect snapshot: " << snap_name << dendl; | |
1672 | return r; | |
1673 | } | |
1674 | ||
1675 | r = snap_is_protected(ictx, snap_name, &is_protected); | |
1676 | if (r < 0) { | |
1677 | return r; | |
1678 | } | |
1679 | if (is_protected) { | |
1680 | lderr(ictx->cct) << "snapshot is still protected after unprotection" << dendl; | |
1681 | ceph_abort(); | |
1682 | } | |
1683 | } | |
1684 | ||
1685 | C_SaferCond ctx; | |
1686 | ictx->operations->snap_remove(cls::rbd::UserSnapshotNamespace(), snap_name, &ctx); | |
1687 | ||
1688 | r = ctx.wait(); | |
1689 | return r; | |
1690 | } | |
1691 | ||
1692 | int snap_get_timestamp(ImageCtx *ictx, uint64_t snap_id, struct timespec *timestamp) | |
1693 | { | |
1694 | std::map<librados::snap_t, SnapInfo>::iterator snap_it = ictx->snap_info.find(snap_id); | |
1695 | assert(snap_it != ictx->snap_info.end()); | |
1696 | utime_t time = snap_it->second.timestamp; | |
1697 | time.to_timespec(timestamp); | |
1698 | return 0; | |
1699 | } | |
1700 | ||
1701 | int snap_get_limit(ImageCtx *ictx, uint64_t *limit) | |
1702 | { | |
1703 | int r = cls_client::snapshot_get_limit(&ictx->md_ctx, ictx->header_oid, | |
1704 | limit); | |
1705 | if (r == -EOPNOTSUPP) { | |
1706 | *limit = UINT64_MAX; | |
1707 | r = 0; | |
1708 | } | |
1709 | return r; | |
1710 | } | |
1711 | ||
1712 | int snap_set_limit(ImageCtx *ictx, uint64_t limit) | |
1713 | { | |
1714 | return ictx->operations->snap_set_limit(limit); | |
1715 | } | |
1716 | ||
1717 | struct CopyProgressCtx { | |
1718 | explicit CopyProgressCtx(ProgressContext &p) | |
1719 | : destictx(NULL), src_size(0), prog_ctx(p) | |
1720 | { } | |
1721 | ||
1722 | ImageCtx *destictx; | |
1723 | uint64_t src_size; | |
1724 | ProgressContext &prog_ctx; | |
1725 | }; | |
1726 | ||
1727 | int copy(ImageCtx *src, IoCtx& dest_md_ctx, const char *destname, | |
1728 | ImageOptions& opts, ProgressContext &prog_ctx, size_t sparse_size) | |
1729 | { | |
1730 | CephContext *cct = (CephContext *)dest_md_ctx.cct(); | |
1731 | ldout(cct, 20) << "copy " << src->name | |
1732 | << (src->snap_name.length() ? "@" + src->snap_name : "") | |
1733 | << " -> " << destname << " opts = " << opts << dendl; | |
1734 | ||
1735 | src->snap_lock.get_read(); | |
1736 | uint64_t features = src->features; | |
1737 | uint64_t src_size = src->get_image_size(src->snap_id); | |
1738 | src->snap_lock.put_read(); | |
1739 | uint64_t format = src->old_format ? 1 : 2; | |
1740 | if (opts.get(RBD_IMAGE_OPTION_FORMAT, &format) != 0) { | |
1741 | opts.set(RBD_IMAGE_OPTION_FORMAT, format); | |
1742 | } | |
1743 | uint64_t stripe_unit = src->stripe_unit; | |
1744 | if (opts.get(RBD_IMAGE_OPTION_STRIPE_UNIT, &stripe_unit) != 0) { | |
1745 | opts.set(RBD_IMAGE_OPTION_STRIPE_UNIT, stripe_unit); | |
1746 | } | |
1747 | uint64_t stripe_count = src->stripe_count; | |
1748 | if (opts.get(RBD_IMAGE_OPTION_STRIPE_COUNT, &stripe_count) != 0) { | |
1749 | opts.set(RBD_IMAGE_OPTION_STRIPE_COUNT, stripe_count); | |
1750 | } | |
1751 | uint64_t order = src->order; | |
1752 | if (opts.get(RBD_IMAGE_OPTION_ORDER, &order) != 0) { | |
1753 | opts.set(RBD_IMAGE_OPTION_ORDER, order); | |
1754 | } | |
1755 | if (opts.get(RBD_IMAGE_OPTION_FEATURES, &features) != 0) { | |
1756 | opts.set(RBD_IMAGE_OPTION_FEATURES, features); | |
1757 | } | |
1758 | if (features & ~RBD_FEATURES_ALL) { | |
1759 | lderr(cct) << "librbd does not support requested features" << dendl; | |
1760 | return -ENOSYS; | |
1761 | } | |
1762 | ||
1763 | int r = create(dest_md_ctx, destname, "", src_size, opts, "", "", false); | |
1764 | if (r < 0) { | |
1765 | lderr(cct) << "header creation failed" << dendl; | |
1766 | return r; | |
1767 | } | |
1768 | opts.set(RBD_IMAGE_OPTION_ORDER, static_cast<uint64_t>(order)); | |
1769 | ||
1770 | ImageCtx *dest = new librbd::ImageCtx(destname, "", NULL, | |
1771 | dest_md_ctx, false); | |
1772 | r = dest->state->open(false); | |
1773 | if (r < 0) { | |
1774 | lderr(cct) << "failed to read newly created header" << dendl; | |
1775 | return r; | |
1776 | } | |
1777 | ||
1778 | r = copy(src, dest, prog_ctx, sparse_size); | |
1779 | ||
1780 | int close_r = dest->state->close(); | |
1781 | if (r == 0 && close_r < 0) { | |
1782 | r = close_r; | |
1783 | } | |
1784 | return r; | |
1785 | } | |
1786 | ||
1787 | class C_CopyWrite : public Context { | |
1788 | public: | |
1789 | C_CopyWrite(bufferlist *bl, Context* ctx) | |
1790 | : m_bl(bl), m_ctx(ctx) {} | |
1791 | void finish(int r) override { | |
1792 | delete m_bl; | |
1793 | m_ctx->complete(r); | |
1794 | } | |
1795 | private: | |
1796 | bufferlist *m_bl; | |
1797 | Context *m_ctx; | |
1798 | }; | |
1799 | ||
1800 | class C_CopyRead : public Context { | |
1801 | public: | |
1802 | C_CopyRead(SimpleThrottle *throttle, ImageCtx *dest, uint64_t offset, | |
1803 | bufferlist *bl, size_t sparse_size) | |
1804 | : m_throttle(throttle), m_dest(dest), m_offset(offset), m_bl(bl), | |
1805 | m_sparse_size(sparse_size) { | |
1806 | m_throttle->start_op(); | |
1807 | } | |
1808 | void finish(int r) override { | |
1809 | if (r < 0) { | |
1810 | lderr(m_dest->cct) << "error reading from source image at offset " | |
1811 | << m_offset << ": " << cpp_strerror(r) << dendl; | |
1812 | delete m_bl; | |
1813 | m_throttle->end_op(r); | |
1814 | return; | |
1815 | } | |
1816 | assert(m_bl->length() == (size_t)r); | |
1817 | ||
1818 | if (m_bl->is_zero()) { | |
1819 | delete m_bl; | |
1820 | m_throttle->end_op(r); | |
1821 | return; | |
1822 | } | |
1823 | ||
1824 | if (!m_sparse_size) { | |
1825 | m_sparse_size = (1 << m_dest->order); | |
1826 | } | |
1827 | ||
1828 | auto *throttle = m_throttle; | |
1829 | auto *end_op_ctx = new FunctionContext([throttle](int r) { | |
1830 | throttle->end_op(r); | |
1831 | }); | |
1832 | auto gather_ctx = new C_Gather(m_dest->cct, end_op_ctx); | |
1833 | ||
1834 | bufferptr m_ptr(m_bl->length()); | |
1835 | m_bl->rebuild(m_ptr); | |
1836 | size_t write_offset = 0; | |
1837 | size_t write_length = 0; | |
1838 | size_t offset = 0; | |
1839 | size_t length = m_bl->length(); | |
1840 | while (offset < length) { | |
1841 | if (util::calc_sparse_extent(m_ptr, | |
1842 | m_sparse_size, | |
1843 | length, | |
1844 | &write_offset, | |
1845 | &write_length, | |
1846 | &offset)) { | |
1847 | bufferptr write_ptr(m_ptr, write_offset, write_length); | |
1848 | bufferlist *write_bl = new bufferlist(); | |
1849 | write_bl->push_back(write_ptr); | |
1850 | Context *ctx = new C_CopyWrite(write_bl, gather_ctx->new_sub()); | |
1851 | auto comp = io::AioCompletion::create(ctx); | |
1852 | ||
1853 | // coordinate through AIO WQ to ensure lock is acquired if needed | |
1854 | m_dest->io_work_queue->aio_write(comp, m_offset + write_offset, | |
1855 | write_length, | |
1856 | std::move(*write_bl), | |
31f18b77 FG |
1857 | LIBRADOS_OP_FLAG_FADVISE_DONTNEED, |
1858 | std::move(read_trace)); | |
7c673cae FG |
1859 | write_offset = offset; |
1860 | write_length = 0; | |
1861 | } | |
1862 | } | |
1863 | delete m_bl; | |
1864 | assert(gather_ctx->get_sub_created_count() > 0); | |
1865 | gather_ctx->activate(); | |
1866 | } | |
1867 | ||
31f18b77 FG |
1868 | ZTracer::Trace read_trace; |
1869 | ||
7c673cae FG |
1870 | private: |
1871 | SimpleThrottle *m_throttle; | |
1872 | ImageCtx *m_dest; | |
1873 | uint64_t m_offset; | |
1874 | bufferlist *m_bl; | |
1875 | size_t m_sparse_size; | |
1876 | }; | |
1877 | ||
1878 | int copy(ImageCtx *src, ImageCtx *dest, ProgressContext &prog_ctx, size_t sparse_size) | |
1879 | { | |
1880 | src->snap_lock.get_read(); | |
1881 | uint64_t src_size = src->get_image_size(src->snap_id); | |
1882 | src->snap_lock.put_read(); | |
1883 | ||
1884 | dest->snap_lock.get_read(); | |
1885 | uint64_t dest_size = dest->get_image_size(dest->snap_id); | |
1886 | dest->snap_lock.put_read(); | |
1887 | ||
1888 | CephContext *cct = src->cct; | |
1889 | if (dest_size < src_size) { | |
1890 | lderr(cct) << " src size " << src_size << " > dest size " | |
1891 | << dest_size << dendl; | |
1892 | return -EINVAL; | |
1893 | } | |
1894 | int r; | |
b32b8144 | 1895 | const uint32_t MAX_KEYS = 64; |
7c673cae | 1896 | map<string, bufferlist> pairs; |
b32b8144 FG |
1897 | std::string last_key = ""; |
1898 | bool more_results = true; | |
7c673cae | 1899 | |
b32b8144 FG |
1900 | while (more_results) { |
1901 | r = cls_client::metadata_list(&src->md_ctx, src->header_oid, last_key, 0, &pairs); | |
1902 | if (r < 0 && r != -EOPNOTSUPP && r != -EIO) { | |
1903 | lderr(cct) << "couldn't list metadata: " << cpp_strerror(r) << dendl; | |
7c673cae | 1904 | return r; |
b32b8144 FG |
1905 | } else if (r == 0 && !pairs.empty()) { |
1906 | r = cls_client::metadata_set(&dest->md_ctx, dest->header_oid, pairs); | |
1907 | if (r < 0) { | |
1908 | lderr(cct) << "couldn't set metadata: " << cpp_strerror(r) << dendl; | |
1909 | return r; | |
1910 | } | |
1911 | ||
1912 | last_key = pairs.rbegin()->first; | |
7c673cae | 1913 | } |
b32b8144 FG |
1914 | |
1915 | more_results = (pairs.size() == MAX_KEYS); | |
1916 | pairs.clear(); | |
7c673cae FG |
1917 | } |
1918 | ||
31f18b77 | 1919 | ZTracer::Trace trace; |
181888fb | 1920 | if (src->blkin_trace_all) { |
31f18b77 FG |
1921 | trace.init("copy", &src->trace_endpoint); |
1922 | } | |
1923 | ||
7c673cae FG |
1924 | RWLock::RLocker owner_lock(src->owner_lock); |
1925 | SimpleThrottle throttle(src->concurrent_management_ops, false); | |
1926 | uint64_t period = src->get_stripe_period(); | |
31f18b77 FG |
1927 | unsigned fadvise_flags = LIBRADOS_OP_FLAG_FADVISE_SEQUENTIAL | |
1928 | LIBRADOS_OP_FLAG_FADVISE_NOCACHE; | |
7c673cae FG |
1929 | for (uint64_t offset = 0; offset < src_size; offset += period) { |
1930 | if (throttle.pending_error()) { | |
1931 | return throttle.wait_for_ret(); | |
1932 | } | |
1933 | ||
1934 | uint64_t len = min(period, src_size - offset); | |
1935 | bufferlist *bl = new bufferlist(); | |
31f18b77 FG |
1936 | auto ctx = new C_CopyRead(&throttle, dest, offset, bl, sparse_size); |
1937 | auto comp = io::AioCompletion::create_and_start<Context>( | |
1938 | ctx, src, io::AIO_TYPE_READ); | |
1939 | ||
1940 | io::ImageReadRequest<> req(*src, comp, {{offset, len}}, | |
1941 | io::ReadResult{bl}, fadvise_flags, | |
1942 | std::move(trace)); | |
1943 | ctx->read_trace = req.get_trace(); | |
1944 | ||
1945 | req.send(); | |
7c673cae FG |
1946 | prog_ctx.update_progress(offset, src_size); |
1947 | } | |
1948 | ||
1949 | r = throttle.wait_for_ret(); | |
1950 | if (r >= 0) | |
1951 | prog_ctx.update_progress(src_size, src_size); | |
1952 | return r; | |
1953 | } | |
1954 | ||
1955 | int snap_set(ImageCtx *ictx, const cls::rbd::SnapshotNamespace &snap_namespace, | |
1956 | const char *snap_name) | |
1957 | { | |
1958 | ldout(ictx->cct, 20) << "snap_set " << ictx << " snap = " | |
1959 | << (snap_name ? snap_name : "NULL") << dendl; | |
1960 | ||
1961 | // ignore return value, since we may be set to a non-existent | |
1962 | // snapshot and the user is trying to fix that | |
1963 | ictx->state->refresh_if_required(); | |
1964 | ||
1965 | C_SaferCond ctx; | |
1966 | std::string name(snap_name == nullptr ? "" : snap_name); | |
1967 | ictx->state->snap_set(snap_namespace, name, &ctx); | |
1968 | ||
1969 | int r = ctx.wait(); | |
1970 | if (r < 0) { | |
1971 | if (r != -ENOENT) { | |
1972 | lderr(ictx->cct) << "failed to " << (name.empty() ? "un" : "") << "set " | |
1973 | << "snapshot: " << cpp_strerror(r) << dendl; | |
1974 | } | |
1975 | return r; | |
1976 | } | |
1977 | ||
1978 | return 0; | |
1979 | } | |
1980 | ||
1981 | int list_lockers(ImageCtx *ictx, | |
1982 | std::list<locker_t> *lockers, | |
1983 | bool *exclusive, | |
1984 | string *tag) | |
1985 | { | |
1986 | ldout(ictx->cct, 20) << "list_locks on image " << ictx << dendl; | |
1987 | ||
1988 | int r = ictx->state->refresh_if_required(); | |
1989 | if (r < 0) | |
1990 | return r; | |
1991 | ||
1992 | RWLock::RLocker locker(ictx->md_lock); | |
1993 | if (exclusive) | |
1994 | *exclusive = ictx->exclusive_locked; | |
1995 | if (tag) | |
1996 | *tag = ictx->lock_tag; | |
1997 | if (lockers) { | |
1998 | lockers->clear(); | |
1999 | map<rados::cls::lock::locker_id_t, | |
2000 | rados::cls::lock::locker_info_t>::const_iterator it; | |
2001 | for (it = ictx->lockers.begin(); it != ictx->lockers.end(); ++it) { | |
2002 | locker_t locker; | |
2003 | locker.client = stringify(it->first.locker); | |
2004 | locker.cookie = it->first.cookie; | |
2005 | locker.address = stringify(it->second.addr); | |
2006 | lockers->push_back(locker); | |
2007 | } | |
2008 | } | |
2009 | ||
2010 | return 0; | |
2011 | } | |
2012 | ||
2013 | int lock(ImageCtx *ictx, bool exclusive, const string& cookie, | |
2014 | const string& tag) | |
2015 | { | |
2016 | ldout(ictx->cct, 20) << "lock image " << ictx << " exclusive=" << exclusive | |
2017 | << " cookie='" << cookie << "' tag='" << tag << "'" | |
2018 | << dendl; | |
2019 | ||
2020 | int r = ictx->state->refresh_if_required(); | |
2021 | if (r < 0) | |
2022 | return r; | |
2023 | ||
2024 | /** | |
2025 | * If we wanted we could do something more intelligent, like local | |
2026 | * checks that we think we will succeed. But for now, let's not | |
2027 | * duplicate that code. | |
2028 | */ | |
2029 | { | |
2030 | RWLock::RLocker locker(ictx->md_lock); | |
2031 | r = rados::cls::lock::lock(&ictx->md_ctx, ictx->header_oid, RBD_LOCK_NAME, | |
2032 | exclusive ? LOCK_EXCLUSIVE : LOCK_SHARED, | |
2033 | cookie, tag, "", utime_t(), 0); | |
2034 | if (r < 0) { | |
2035 | return r; | |
2036 | } | |
2037 | } | |
2038 | ||
2039 | ictx->notify_update(); | |
2040 | return 0; | |
2041 | } | |
2042 | ||
2043 | int unlock(ImageCtx *ictx, const string& cookie) | |
2044 | { | |
2045 | ldout(ictx->cct, 20) << "unlock image " << ictx | |
2046 | << " cookie='" << cookie << "'" << dendl; | |
2047 | ||
2048 | int r = ictx->state->refresh_if_required(); | |
2049 | if (r < 0) | |
2050 | return r; | |
2051 | ||
2052 | { | |
2053 | RWLock::RLocker locker(ictx->md_lock); | |
2054 | r = rados::cls::lock::unlock(&ictx->md_ctx, ictx->header_oid, | |
2055 | RBD_LOCK_NAME, cookie); | |
2056 | if (r < 0) { | |
2057 | return r; | |
2058 | } | |
2059 | } | |
2060 | ||
2061 | ictx->notify_update(); | |
2062 | return 0; | |
2063 | } | |
2064 | ||
2065 | int break_lock(ImageCtx *ictx, const string& client, | |
2066 | const string& cookie) | |
2067 | { | |
2068 | ldout(ictx->cct, 20) << "break_lock image " << ictx << " client='" << client | |
2069 | << "' cookie='" << cookie << "'" << dendl; | |
2070 | ||
2071 | int r = ictx->state->refresh_if_required(); | |
2072 | if (r < 0) | |
2073 | return r; | |
2074 | ||
2075 | entity_name_t lock_client; | |
2076 | if (!lock_client.parse(client)) { | |
2077 | lderr(ictx->cct) << "Unable to parse client '" << client | |
2078 | << "'" << dendl; | |
2079 | return -EINVAL; | |
2080 | } | |
2081 | ||
2082 | if (ictx->blacklist_on_break_lock) { | |
2083 | typedef std::map<rados::cls::lock::locker_id_t, | |
2084 | rados::cls::lock::locker_info_t> Lockers; | |
2085 | Lockers lockers; | |
2086 | ClsLockType lock_type; | |
2087 | std::string lock_tag; | |
2088 | r = rados::cls::lock::get_lock_info(&ictx->md_ctx, ictx->header_oid, | |
2089 | RBD_LOCK_NAME, &lockers, &lock_type, | |
2090 | &lock_tag); | |
2091 | if (r < 0) { | |
2092 | lderr(ictx->cct) << "unable to retrieve lock info: " << cpp_strerror(r) | |
2093 | << dendl; | |
2094 | return r; | |
2095 | } | |
2096 | ||
2097 | std::string client_address; | |
2098 | for (Lockers::iterator it = lockers.begin(); | |
2099 | it != lockers.end(); ++it) { | |
2100 | if (it->first.locker == lock_client) { | |
2101 | client_address = stringify(it->second.addr); | |
2102 | break; | |
2103 | } | |
2104 | } | |
2105 | if (client_address.empty()) { | |
2106 | return -ENOENT; | |
2107 | } | |
2108 | ||
2109 | RWLock::RLocker locker(ictx->md_lock); | |
2110 | librados::Rados rados(ictx->md_ctx); | |
2111 | r = rados.blacklist_add(client_address, | |
2112 | ictx->blacklist_expire_seconds); | |
2113 | if (r < 0) { | |
2114 | lderr(ictx->cct) << "unable to blacklist client: " << cpp_strerror(r) | |
2115 | << dendl; | |
2116 | return r; | |
2117 | } | |
2118 | } | |
2119 | ||
2120 | r = rados::cls::lock::break_lock(&ictx->md_ctx, ictx->header_oid, | |
2121 | RBD_LOCK_NAME, cookie, lock_client); | |
2122 | if (r < 0) | |
2123 | return r; | |
2124 | ictx->notify_update(); | |
2125 | return 0; | |
2126 | } | |
2127 | ||
2128 | void rbd_ctx_cb(completion_t cb, void *arg) | |
2129 | { | |
2130 | Context *ctx = reinterpret_cast<Context *>(arg); | |
2131 | auto comp = reinterpret_cast<io::AioCompletion *>(cb); | |
2132 | ctx->complete(comp->get_return_value()); | |
2133 | comp->release(); | |
2134 | } | |
2135 | ||
2136 | int64_t read_iterate(ImageCtx *ictx, uint64_t off, uint64_t len, | |
2137 | int (*cb)(uint64_t, size_t, const char *, void *), | |
2138 | void *arg) | |
2139 | { | |
2140 | utime_t start_time, elapsed; | |
2141 | ||
2142 | ldout(ictx->cct, 20) << "read_iterate " << ictx << " off = " << off | |
2143 | << " len = " << len << dendl; | |
2144 | ||
2145 | int r = ictx->state->refresh_if_required(); | |
2146 | if (r < 0) | |
2147 | return r; | |
2148 | ||
2149 | uint64_t mylen = len; | |
2150 | ictx->snap_lock.get_read(); | |
2151 | r = clip_io(ictx, off, &mylen); | |
2152 | ictx->snap_lock.put_read(); | |
2153 | if (r < 0) | |
2154 | return r; | |
2155 | ||
2156 | int64_t total_read = 0; | |
2157 | uint64_t period = ictx->get_stripe_period(); | |
2158 | uint64_t left = mylen; | |
2159 | ||
31f18b77 | 2160 | ZTracer::Trace trace; |
181888fb | 2161 | if (ictx->blkin_trace_all) { |
31f18b77 FG |
2162 | trace.init("read_iterate", &ictx->trace_endpoint); |
2163 | } | |
2164 | ||
7c673cae FG |
2165 | RWLock::RLocker owner_locker(ictx->owner_lock); |
2166 | start_time = ceph_clock_now(); | |
2167 | while (left > 0) { | |
2168 | uint64_t period_off = off - (off % period); | |
2169 | uint64_t read_len = min(period_off + period - off, left); | |
2170 | ||
2171 | bufferlist bl; | |
2172 | ||
2173 | C_SaferCond ctx; | |
2174 | auto c = io::AioCompletion::create_and_start(&ctx, ictx, | |
2175 | io::AIO_TYPE_READ); | |
2176 | io::ImageRequest<>::aio_read(ictx, c, {{off, read_len}}, | |
31f18b77 | 2177 | io::ReadResult{&bl}, 0, std::move(trace)); |
7c673cae FG |
2178 | |
2179 | int ret = ctx.wait(); | |
2180 | if (ret < 0) { | |
2181 | return ret; | |
2182 | } | |
2183 | ||
2184 | r = cb(total_read, ret, bl.c_str(), arg); | |
2185 | if (r < 0) { | |
2186 | return r; | |
2187 | } | |
2188 | ||
2189 | total_read += ret; | |
2190 | left -= ret; | |
2191 | off += ret; | |
2192 | } | |
2193 | ||
2194 | elapsed = ceph_clock_now() - start_time; | |
2195 | ictx->perfcounter->tinc(l_librbd_rd_latency, elapsed); | |
2196 | ictx->perfcounter->inc(l_librbd_rd); | |
2197 | ictx->perfcounter->inc(l_librbd_rd_bytes, mylen); | |
2198 | return total_read; | |
2199 | } | |
2200 | ||
2201 | // validate extent against image size; clip to image size if necessary | |
2202 | int clip_io(ImageCtx *ictx, uint64_t off, uint64_t *len) | |
2203 | { | |
2204 | assert(ictx->snap_lock.is_locked()); | |
2205 | uint64_t image_size = ictx->get_image_size(ictx->snap_id); | |
2206 | bool snap_exists = ictx->snap_exists; | |
2207 | ||
2208 | if (!snap_exists) | |
2209 | return -ENOENT; | |
2210 | ||
2211 | // special-case "len == 0" requests: always valid | |
2212 | if (*len == 0) | |
2213 | return 0; | |
2214 | ||
2215 | // can't start past end | |
2216 | if (off >= image_size) | |
2217 | return -EINVAL; | |
2218 | ||
2219 | // clip requests that extend past end to just end | |
2220 | if ((off + *len) > image_size) | |
2221 | *len = (size_t)(image_size - off); | |
2222 | ||
2223 | return 0; | |
2224 | } | |
2225 | ||
2226 | int flush(ImageCtx *ictx) | |
2227 | { | |
2228 | CephContext *cct = ictx->cct; | |
2229 | ldout(cct, 20) << "flush " << ictx << dendl; | |
2230 | ||
2231 | int r = ictx->state->refresh_if_required(); | |
2232 | if (r < 0) { | |
2233 | return r; | |
2234 | } | |
2235 | ||
2236 | ictx->user_flushed(); | |
2237 | C_SaferCond ctx; | |
2238 | { | |
2239 | RWLock::RLocker owner_locker(ictx->owner_lock); | |
2240 | ictx->flush(&ctx); | |
2241 | } | |
2242 | r = ctx.wait(); | |
2243 | ||
2244 | ictx->perfcounter->inc(l_librbd_flush); | |
2245 | return r; | |
2246 | } | |
2247 | ||
2248 | int invalidate_cache(ImageCtx *ictx) | |
2249 | { | |
2250 | CephContext *cct = ictx->cct; | |
2251 | ldout(cct, 20) << "invalidate_cache " << ictx << dendl; | |
2252 | ||
2253 | int r = ictx->state->refresh_if_required(); | |
2254 | if (r < 0) { | |
2255 | return r; | |
2256 | } | |
2257 | ||
2258 | RWLock::RLocker owner_locker(ictx->owner_lock); | |
7c673cae FG |
2259 | r = ictx->invalidate_cache(false); |
2260 | ictx->perfcounter->inc(l_librbd_invalidate_cache); | |
2261 | return r; | |
2262 | } | |
2263 | ||
2264 | int poll_io_events(ImageCtx *ictx, io::AioCompletion **comps, int numcomp) | |
2265 | { | |
2266 | if (numcomp <= 0) | |
2267 | return -EINVAL; | |
2268 | CephContext *cct = ictx->cct; | |
2269 | ldout(cct, 20) << __func__ << " " << ictx << " numcomp = " << numcomp | |
2270 | << dendl; | |
2271 | int i = 0; | |
2272 | Mutex::Locker l(ictx->completed_reqs_lock); | |
2273 | while (i < numcomp) { | |
2274 | if (ictx->completed_reqs.empty()) | |
2275 | break; | |
2276 | comps[i++] = ictx->completed_reqs.front(); | |
2277 | ictx->completed_reqs.pop_front(); | |
2278 | } | |
2279 | return i; | |
2280 | } | |
2281 | ||
2282 | int metadata_get(ImageCtx *ictx, const string &key, string *value) | |
2283 | { | |
2284 | CephContext *cct = ictx->cct; | |
2285 | ldout(cct, 20) << "metadata_get " << ictx << " key=" << key << dendl; | |
2286 | ||
2287 | int r = ictx->state->refresh_if_required(); | |
2288 | if (r < 0) { | |
2289 | return r; | |
2290 | } | |
2291 | ||
2292 | return cls_client::metadata_get(&ictx->md_ctx, ictx->header_oid, key, value); | |
2293 | } | |
2294 | ||
2295 | int metadata_list(ImageCtx *ictx, const string &start, uint64_t max, map<string, bufferlist> *pairs) | |
2296 | { | |
2297 | CephContext *cct = ictx->cct; | |
2298 | ldout(cct, 20) << "metadata_list " << ictx << dendl; | |
2299 | ||
2300 | int r = ictx->state->refresh_if_required(); | |
2301 | if (r < 0) { | |
2302 | return r; | |
2303 | } | |
2304 | ||
2305 | return cls_client::metadata_list(&ictx->md_ctx, ictx->header_oid, start, max, pairs); | |
2306 | } | |
2307 | ||
2308 | struct C_RBD_Readahead : public Context { | |
2309 | ImageCtx *ictx; | |
2310 | object_t oid; | |
2311 | uint64_t offset; | |
2312 | uint64_t length; | |
2313 | C_RBD_Readahead(ImageCtx *ictx, object_t oid, uint64_t offset, uint64_t length) | |
2314 | : ictx(ictx), oid(oid), offset(offset), length(length) { } | |
2315 | void finish(int r) override { | |
2316 | ldout(ictx->cct, 20) << "C_RBD_Readahead on " << oid << ": " << offset << "+" << length << dendl; | |
2317 | ictx->readahead.dec_pending(); | |
2318 | } | |
2319 | }; | |
2320 | ||
2321 | void readahead(ImageCtx *ictx, | |
2322 | const vector<pair<uint64_t,uint64_t> >& image_extents) | |
2323 | { | |
2324 | uint64_t total_bytes = 0; | |
2325 | for (vector<pair<uint64_t,uint64_t> >::const_iterator p = image_extents.begin(); | |
2326 | p != image_extents.end(); | |
2327 | ++p) { | |
2328 | total_bytes += p->second; | |
2329 | } | |
2330 | ||
2331 | ictx->md_lock.get_write(); | |
2332 | bool abort = ictx->readahead_disable_after_bytes != 0 && | |
2333 | ictx->total_bytes_read > ictx->readahead_disable_after_bytes; | |
2334 | if (abort) { | |
2335 | ictx->md_lock.put_write(); | |
2336 | return; | |
2337 | } | |
2338 | ictx->total_bytes_read += total_bytes; | |
2339 | ictx->snap_lock.get_read(); | |
2340 | uint64_t image_size = ictx->get_image_size(ictx->snap_id); | |
2341 | ictx->snap_lock.put_read(); | |
2342 | ictx->md_lock.put_write(); | |
2343 | ||
2344 | pair<uint64_t, uint64_t> readahead_extent = ictx->readahead.update(image_extents, image_size); | |
2345 | uint64_t readahead_offset = readahead_extent.first; | |
2346 | uint64_t readahead_length = readahead_extent.second; | |
2347 | ||
2348 | if (readahead_length > 0) { | |
2349 | ldout(ictx->cct, 20) << "(readahead logical) " << readahead_offset << "~" << readahead_length << dendl; | |
2350 | map<object_t,vector<ObjectExtent> > readahead_object_extents; | |
2351 | Striper::file_to_extents(ictx->cct, ictx->format_string, &ictx->layout, | |
2352 | readahead_offset, readahead_length, 0, readahead_object_extents); | |
2353 | for (map<object_t,vector<ObjectExtent> >::iterator p = readahead_object_extents.begin(); p != readahead_object_extents.end(); ++p) { | |
2354 | for (vector<ObjectExtent>::iterator q = p->second.begin(); q != p->second.end(); ++q) { | |
2355 | ldout(ictx->cct, 20) << "(readahead) oid " << q->oid << " " << q->offset << "~" << q->length << dendl; | |
2356 | ||
2357 | Context *req_comp = new C_RBD_Readahead(ictx, q->oid, q->offset, q->length); | |
2358 | ictx->readahead.inc_pending(); | |
2359 | ictx->aio_read_from_cache(q->oid, q->objectno, NULL, | |
2360 | q->length, q->offset, | |
31f18b77 | 2361 | req_comp, 0, nullptr); |
7c673cae FG |
2362 | } |
2363 | } | |
2364 | ictx->perfcounter->inc(l_librbd_readahead); | |
2365 | ictx->perfcounter->inc(l_librbd_readahead_bytes, readahead_length); | |
2366 | } | |
2367 | } | |
2368 | ||
2369 | ||
2370 | ||
2371 | } |