]>
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" | |
11fdf7f2 TL |
15 | #include "common/perf_counters.h" |
16 | #include "osdc/Striper.h" | |
7c673cae FG |
17 | #include "include/stringify.h" |
18 | ||
11fdf7f2 | 19 | #include "cls/lock/cls_lock_client.h" |
7c673cae FG |
20 | #include "cls/rbd/cls_rbd.h" |
21 | #include "cls/rbd/cls_rbd_types.h" | |
22 | #include "cls/rbd/cls_rbd_client.h" | |
23 | #include "cls/journal/cls_journal_types.h" | |
24 | #include "cls/journal/cls_journal_client.h" | |
25 | ||
f67539c2 | 26 | #include "librbd/AsioEngine.h" |
7c673cae FG |
27 | #include "librbd/ExclusiveLock.h" |
28 | #include "librbd/ImageCtx.h" | |
29 | #include "librbd/ImageState.h" | |
30 | #include "librbd/internal.h" | |
31 | #include "librbd/Journal.h" | |
32 | #include "librbd/ObjectMap.h" | |
33 | #include "librbd/Operations.h" | |
f67539c2 | 34 | #include "librbd/PluginRegistry.h" |
7c673cae FG |
35 | #include "librbd/Types.h" |
36 | #include "librbd/Utils.h" | |
11fdf7f2 | 37 | #include "librbd/api/Config.h" |
7c673cae | 38 | #include "librbd/api/Image.h" |
f67539c2 TL |
39 | #include "librbd/api/Io.h" |
40 | #include "librbd/cache/Utils.h" | |
7c673cae FG |
41 | #include "librbd/exclusive_lock/AutomaticPolicy.h" |
42 | #include "librbd/exclusive_lock/StandardPolicy.h" | |
9f95a23c | 43 | #include "librbd/deep_copy/MetadataCopyRequest.h" |
7c673cae FG |
44 | #include "librbd/image/CloneRequest.h" |
45 | #include "librbd/image/CreateRequest.h" | |
9f95a23c | 46 | #include "librbd/image/GetMetadataRequest.h" |
1911f103 | 47 | #include "librbd/image/Types.h" |
7c673cae | 48 | #include "librbd/io/AioCompletion.h" |
f67539c2 TL |
49 | #include "librbd/io/ImageDispatchSpec.h" |
50 | #include "librbd/io/ImageDispatcherInterface.h" | |
51 | #include "librbd/io/ObjectDispatcherInterface.h" | |
7c673cae FG |
52 | #include "librbd/io/ObjectRequest.h" |
53 | #include "librbd/io/ReadResult.h" | |
54 | #include "librbd/journal/Types.h" | |
55 | #include "librbd/managed_lock/Types.h" | |
56 | #include "librbd/mirror/EnableRequest.h" | |
57 | #include "librbd/operation/TrimRequest.h" | |
58 | ||
59 | #include "journal/Journaler.h" | |
60 | ||
61 | #include <boost/scope_exit.hpp> | |
62 | #include <boost/variant.hpp> | |
11fdf7f2 | 63 | #include "include/ceph_assert.h" |
7c673cae FG |
64 | |
65 | #define dout_subsys ceph_subsys_rbd | |
66 | #undef dout_prefix | |
67 | #define dout_prefix *_dout << "librbd: " | |
68 | ||
69 | #define rbd_howmany(x, y) (((x) + (y) - 1) / (y)) | |
70 | ||
20effc67 | 71 | using std::istringstream; |
7c673cae FG |
72 | using std::map; |
73 | using std::pair; | |
74 | using std::set; | |
75 | using std::string; | |
76 | using std::vector; | |
77 | // list binds to list() here, so std::list is explicitly used below | |
78 | ||
79 | using ceph::bufferlist; | |
80 | using librados::snap_t; | |
81 | using librados::IoCtx; | |
82 | using librados::Rados; | |
83 | ||
84 | namespace librbd { | |
85 | ||
86 | namespace { | |
87 | ||
88 | int validate_pool(IoCtx &io_ctx, CephContext *cct) { | |
11fdf7f2 | 89 | if (!cct->_conf.get_val<bool>("rbd_validate_pool")) { |
7c673cae FG |
90 | return 0; |
91 | } | |
92 | ||
93 | int r = io_ctx.stat(RBD_DIRECTORY, NULL, NULL); | |
94 | if (r == 0) { | |
95 | return 0; | |
96 | } else if (r < 0 && r != -ENOENT) { | |
97 | lderr(cct) << "failed to stat RBD directory: " << cpp_strerror(r) << dendl; | |
98 | return r; | |
99 | } | |
100 | ||
101 | // allocate a self-managed snapshot id if this a new pool to force | |
102 | // self-managed snapshot mode | |
103 | uint64_t snap_id; | |
104 | r = io_ctx.selfmanaged_snap_create(&snap_id); | |
105 | if (r == -EINVAL) { | |
106 | lderr(cct) << "pool not configured for self-managed RBD snapshot support" | |
107 | << dendl; | |
108 | return r; | |
109 | } else if (r < 0) { | |
110 | lderr(cct) << "failed to allocate self-managed snapshot: " | |
111 | << cpp_strerror(r) << dendl; | |
112 | return r; | |
113 | } | |
114 | ||
115 | r = io_ctx.selfmanaged_snap_remove(snap_id); | |
116 | if (r < 0) { | |
117 | lderr(cct) << "failed to release self-managed snapshot " << snap_id | |
118 | << ": " << cpp_strerror(r) << dendl; | |
119 | } | |
120 | return 0; | |
121 | } | |
122 | ||
7c673cae FG |
123 | } // anonymous namespace |
124 | ||
125 | int detect_format(IoCtx &io_ctx, const string &name, | |
126 | bool *old_format, uint64_t *size) | |
127 | { | |
128 | CephContext *cct = (CephContext *)io_ctx.cct(); | |
129 | if (old_format) | |
130 | *old_format = true; | |
131 | int r = io_ctx.stat(util::old_header_name(name), size, NULL); | |
132 | if (r == -ENOENT) { | |
133 | if (old_format) | |
134 | *old_format = false; | |
135 | r = io_ctx.stat(util::id_obj_name(name), size, NULL); | |
136 | if (r < 0) | |
137 | return r; | |
138 | } else if (r < 0) { | |
139 | return r; | |
140 | } | |
141 | ||
142 | ldout(cct, 20) << "detect format of " << name << " : " | |
143 | << (old_format ? (*old_format ? "old" : "new") : | |
144 | "don't care") << dendl; | |
145 | return 0; | |
146 | } | |
147 | ||
148 | bool has_parent(int64_t parent_pool_id, uint64_t off, uint64_t overlap) | |
149 | { | |
150 | return (parent_pool_id != -1 && off <= overlap); | |
151 | } | |
152 | ||
153 | void init_rbd_header(struct rbd_obj_header_ondisk& ondisk, | |
154 | uint64_t size, int order, uint64_t bid) | |
155 | { | |
156 | uint32_t hi = bid >> 32; | |
157 | uint32_t lo = bid & 0xFFFFFFFF; | |
158 | uint32_t extra = rand() % 0xFFFFFFFF; | |
92f5a8d4 | 159 | // FIPS zeroization audit 20191117: this memset is not security related. |
7c673cae FG |
160 | memset(&ondisk, 0, sizeof(ondisk)); |
161 | ||
162 | memcpy(&ondisk.text, RBD_HEADER_TEXT, sizeof(RBD_HEADER_TEXT)); | |
163 | memcpy(&ondisk.signature, RBD_HEADER_SIGNATURE, | |
164 | sizeof(RBD_HEADER_SIGNATURE)); | |
165 | memcpy(&ondisk.version, RBD_HEADER_VERSION, sizeof(RBD_HEADER_VERSION)); | |
166 | ||
167 | snprintf(ondisk.block_name, sizeof(ondisk.block_name), "rb.%x.%x.%x", | |
168 | hi, lo, extra); | |
169 | ||
170 | ondisk.image_size = size; | |
171 | ondisk.options.order = order; | |
172 | ondisk.options.crypt_type = RBD_CRYPT_NONE; | |
173 | ondisk.options.comp_type = RBD_COMP_NONE; | |
174 | ondisk.snap_seq = 0; | |
175 | ondisk.snap_count = 0; | |
176 | ondisk.reserved = 0; | |
177 | ondisk.snap_names_len = 0; | |
178 | } | |
179 | ||
180 | void image_info(ImageCtx *ictx, image_info_t& info, size_t infosize) | |
181 | { | |
182 | int obj_order = ictx->order; | |
9f95a23c TL |
183 | { |
184 | std::shared_lock locker{ictx->image_lock}; | |
f67539c2 | 185 | info.size = ictx->get_effective_image_size(ictx->snap_id); |
9f95a23c | 186 | } |
7c673cae FG |
187 | info.obj_size = 1ULL << obj_order; |
188 | info.num_objs = Striper::get_num_objects(ictx->layout, info.size); | |
189 | info.order = obj_order; | |
190 | strncpy(info.block_name_prefix, ictx->object_prefix.c_str(), | |
191 | RBD_MAX_BLOCK_NAME_SIZE); | |
192 | info.block_name_prefix[RBD_MAX_BLOCK_NAME_SIZE - 1] = '\0'; | |
193 | ||
194 | // clear deprecated fields | |
195 | info.parent_pool = -1L; | |
196 | info.parent_name[0] = '\0'; | |
197 | } | |
198 | ||
199 | uint64_t oid_to_object_no(const string& oid, const string& object_prefix) | |
200 | { | |
201 | istringstream iss(oid); | |
202 | // skip object prefix and separator | |
203 | iss.ignore(object_prefix.length() + 1); | |
204 | uint64_t num; | |
205 | iss >> std::hex >> num; | |
206 | return num; | |
207 | } | |
208 | ||
209 | void trim_image(ImageCtx *ictx, uint64_t newsize, ProgressContext& prog_ctx) | |
210 | { | |
9f95a23c | 211 | ceph_assert(ceph_mutex_is_locked(ictx->owner_lock)); |
11fdf7f2 TL |
212 | ceph_assert(ictx->exclusive_lock == nullptr || |
213 | ictx->exclusive_lock->is_lock_owner()); | |
7c673cae FG |
214 | |
215 | C_SaferCond ctx; | |
9f95a23c | 216 | ictx->image_lock.lock_shared(); |
7c673cae FG |
217 | operation::TrimRequest<> *req = operation::TrimRequest<>::create( |
218 | *ictx, &ctx, ictx->size, newsize, prog_ctx); | |
9f95a23c | 219 | ictx->image_lock.unlock_shared(); |
7c673cae FG |
220 | req->send(); |
221 | ||
222 | int r = ctx.wait(); | |
223 | if (r < 0) { | |
224 | lderr(ictx->cct) << "warning: failed to remove some object(s): " | |
225 | << cpp_strerror(r) << dendl; | |
226 | } | |
227 | } | |
228 | ||
229 | int read_header_bl(IoCtx& io_ctx, const string& header_oid, | |
230 | bufferlist& header, uint64_t *ver) | |
231 | { | |
232 | int r; | |
233 | uint64_t off = 0; | |
234 | #define READ_SIZE 4096 | |
235 | do { | |
236 | bufferlist bl; | |
237 | r = io_ctx.read(header_oid, bl, READ_SIZE, off); | |
238 | if (r < 0) | |
239 | return r; | |
240 | header.claim_append(bl); | |
241 | off += r; | |
242 | } while (r == READ_SIZE); | |
243 | ||
11fdf7f2 TL |
244 | static_assert(sizeof(RBD_HEADER_TEXT) == sizeof(RBD_MIGRATE_HEADER_TEXT), |
245 | "length of rbd headers must be the same"); | |
246 | ||
7c673cae | 247 | if (header.length() < sizeof(RBD_HEADER_TEXT) || |
11fdf7f2 TL |
248 | (memcmp(RBD_HEADER_TEXT, header.c_str(), |
249 | sizeof(RBD_HEADER_TEXT)) != 0 && | |
250 | memcmp(RBD_MIGRATE_HEADER_TEXT, header.c_str(), | |
251 | sizeof(RBD_MIGRATE_HEADER_TEXT)) != 0)) { | |
7c673cae FG |
252 | CephContext *cct = (CephContext *)io_ctx.cct(); |
253 | lderr(cct) << "unrecognized header format" << dendl; | |
254 | return -ENXIO; | |
255 | } | |
256 | ||
257 | if (ver) | |
258 | *ver = io_ctx.get_last_version(); | |
259 | ||
260 | return 0; | |
261 | } | |
262 | ||
263 | int read_header(IoCtx& io_ctx, const string& header_oid, | |
264 | struct rbd_obj_header_ondisk *header, uint64_t *ver) | |
265 | { | |
266 | bufferlist header_bl; | |
267 | int r = read_header_bl(io_ctx, header_oid, header_bl, ver); | |
268 | if (r < 0) | |
269 | return r; | |
270 | if (header_bl.length() < (int)sizeof(*header)) | |
271 | return -EIO; | |
272 | memcpy(header, header_bl.c_str(), sizeof(*header)); | |
273 | ||
274 | return 0; | |
275 | } | |
276 | ||
277 | int tmap_set(IoCtx& io_ctx, const string& imgname) | |
278 | { | |
279 | bufferlist cmdbl, emptybl; | |
280 | __u8 c = CEPH_OSD_TMAP_SET; | |
11fdf7f2 TL |
281 | encode(c, cmdbl); |
282 | encode(imgname, cmdbl); | |
283 | encode(emptybl, cmdbl); | |
7c673cae FG |
284 | return io_ctx.tmap_update(RBD_DIRECTORY, cmdbl); |
285 | } | |
286 | ||
287 | int tmap_rm(IoCtx& io_ctx, const string& imgname) | |
288 | { | |
289 | bufferlist cmdbl; | |
290 | __u8 c = CEPH_OSD_TMAP_RM; | |
11fdf7f2 TL |
291 | encode(c, cmdbl); |
292 | encode(imgname, cmdbl); | |
7c673cae FG |
293 | return io_ctx.tmap_update(RBD_DIRECTORY, cmdbl); |
294 | } | |
295 | ||
296 | typedef boost::variant<std::string,uint64_t> image_option_value_t; | |
297 | typedef std::map<int,image_option_value_t> image_options_t; | |
298 | typedef std::shared_ptr<image_options_t> image_options_ref; | |
299 | ||
300 | enum image_option_type_t { | |
301 | STR, | |
302 | UINT64, | |
303 | }; | |
304 | ||
305 | const std::map<int, image_option_type_t> IMAGE_OPTIONS_TYPE_MAPPING = { | |
306 | {RBD_IMAGE_OPTION_FORMAT, UINT64}, | |
307 | {RBD_IMAGE_OPTION_FEATURES, UINT64}, | |
308 | {RBD_IMAGE_OPTION_ORDER, UINT64}, | |
309 | {RBD_IMAGE_OPTION_STRIPE_UNIT, UINT64}, | |
310 | {RBD_IMAGE_OPTION_STRIPE_COUNT, UINT64}, | |
311 | {RBD_IMAGE_OPTION_JOURNAL_ORDER, UINT64}, | |
312 | {RBD_IMAGE_OPTION_JOURNAL_SPLAY_WIDTH, UINT64}, | |
313 | {RBD_IMAGE_OPTION_JOURNAL_POOL, STR}, | |
314 | {RBD_IMAGE_OPTION_FEATURES_SET, UINT64}, | |
315 | {RBD_IMAGE_OPTION_FEATURES_CLEAR, UINT64}, | |
316 | {RBD_IMAGE_OPTION_DATA_POOL, STR}, | |
11fdf7f2 | 317 | {RBD_IMAGE_OPTION_FLATTEN, UINT64}, |
92f5a8d4 | 318 | {RBD_IMAGE_OPTION_CLONE_FORMAT, UINT64}, |
1911f103 | 319 | {RBD_IMAGE_OPTION_MIRROR_IMAGE_MODE, UINT64}, |
7c673cae FG |
320 | }; |
321 | ||
322 | std::string image_option_name(int optname) { | |
323 | switch (optname) { | |
324 | case RBD_IMAGE_OPTION_FORMAT: | |
325 | return "format"; | |
326 | case RBD_IMAGE_OPTION_FEATURES: | |
327 | return "features"; | |
328 | case RBD_IMAGE_OPTION_ORDER: | |
329 | return "order"; | |
330 | case RBD_IMAGE_OPTION_STRIPE_UNIT: | |
331 | return "stripe_unit"; | |
332 | case RBD_IMAGE_OPTION_STRIPE_COUNT: | |
333 | return "stripe_count"; | |
334 | case RBD_IMAGE_OPTION_JOURNAL_ORDER: | |
335 | return "journal_order"; | |
336 | case RBD_IMAGE_OPTION_JOURNAL_SPLAY_WIDTH: | |
337 | return "journal_splay_width"; | |
338 | case RBD_IMAGE_OPTION_JOURNAL_POOL: | |
339 | return "journal_pool"; | |
340 | case RBD_IMAGE_OPTION_FEATURES_SET: | |
341 | return "features_set"; | |
342 | case RBD_IMAGE_OPTION_FEATURES_CLEAR: | |
343 | return "features_clear"; | |
344 | case RBD_IMAGE_OPTION_DATA_POOL: | |
345 | return "data_pool"; | |
11fdf7f2 TL |
346 | case RBD_IMAGE_OPTION_FLATTEN: |
347 | return "flatten"; | |
92f5a8d4 TL |
348 | case RBD_IMAGE_OPTION_CLONE_FORMAT: |
349 | return "clone_format"; | |
1911f103 TL |
350 | case RBD_IMAGE_OPTION_MIRROR_IMAGE_MODE: |
351 | return "mirror_image_mode"; | |
7c673cae FG |
352 | default: |
353 | return "unknown (" + stringify(optname) + ")"; | |
354 | } | |
355 | } | |
356 | ||
7c673cae FG |
357 | void image_options_create(rbd_image_options_t* opts) |
358 | { | |
359 | image_options_ref* opts_ = new image_options_ref(new image_options_t()); | |
360 | ||
361 | *opts = static_cast<rbd_image_options_t>(opts_); | |
362 | } | |
363 | ||
364 | void image_options_create_ref(rbd_image_options_t* opts, | |
365 | rbd_image_options_t orig) | |
366 | { | |
367 | image_options_ref* orig_ = static_cast<image_options_ref*>(orig); | |
368 | image_options_ref* opts_ = new image_options_ref(*orig_); | |
369 | ||
370 | *opts = static_cast<rbd_image_options_t>(opts_); | |
371 | } | |
372 | ||
373 | void image_options_copy(rbd_image_options_t* opts, | |
374 | const ImageOptions &orig) | |
375 | { | |
376 | image_options_ref* opts_ = new image_options_ref(new image_options_t()); | |
377 | ||
378 | *opts = static_cast<rbd_image_options_t>(opts_); | |
379 | ||
380 | std::string str_val; | |
381 | uint64_t uint64_val; | |
382 | for (auto &i : IMAGE_OPTIONS_TYPE_MAPPING) { | |
383 | switch (i.second) { | |
384 | case STR: | |
385 | if (orig.get(i.first, &str_val) == 0) { | |
386 | image_options_set(*opts, i.first, str_val); | |
387 | } | |
388 | continue; | |
389 | case UINT64: | |
390 | if (orig.get(i.first, &uint64_val) == 0) { | |
391 | image_options_set(*opts, i.first, uint64_val); | |
392 | } | |
393 | continue; | |
394 | } | |
395 | } | |
396 | } | |
397 | ||
398 | void image_options_destroy(rbd_image_options_t opts) | |
399 | { | |
400 | image_options_ref* opts_ = static_cast<image_options_ref*>(opts); | |
401 | ||
402 | delete opts_; | |
403 | } | |
404 | ||
405 | int image_options_set(rbd_image_options_t opts, int optname, | |
406 | const std::string& optval) | |
407 | { | |
408 | image_options_ref* opts_ = static_cast<image_options_ref*>(opts); | |
409 | ||
410 | std::map<int, image_option_type_t>::const_iterator i = | |
411 | IMAGE_OPTIONS_TYPE_MAPPING.find(optname); | |
412 | ||
413 | if (i == IMAGE_OPTIONS_TYPE_MAPPING.end() || i->second != STR) { | |
414 | return -EINVAL; | |
415 | } | |
416 | ||
417 | (*opts_->get())[optname] = optval; | |
418 | return 0; | |
419 | } | |
420 | ||
421 | int image_options_set(rbd_image_options_t opts, int optname, uint64_t optval) | |
422 | { | |
423 | image_options_ref* opts_ = static_cast<image_options_ref*>(opts); | |
424 | ||
425 | std::map<int, image_option_type_t>::const_iterator i = | |
426 | IMAGE_OPTIONS_TYPE_MAPPING.find(optname); | |
427 | ||
428 | if (i == IMAGE_OPTIONS_TYPE_MAPPING.end() || i->second != UINT64) { | |
429 | return -EINVAL; | |
430 | } | |
431 | ||
432 | (*opts_->get())[optname] = optval; | |
433 | return 0; | |
434 | } | |
435 | ||
436 | int image_options_get(rbd_image_options_t opts, int optname, | |
437 | std::string* optval) | |
438 | { | |
439 | image_options_ref* opts_ = static_cast<image_options_ref*>(opts); | |
440 | ||
441 | std::map<int, image_option_type_t>::const_iterator i = | |
442 | IMAGE_OPTIONS_TYPE_MAPPING.find(optname); | |
443 | ||
444 | if (i == IMAGE_OPTIONS_TYPE_MAPPING.end() || i->second != STR) { | |
445 | return -EINVAL; | |
446 | } | |
447 | ||
448 | image_options_t::const_iterator j = (*opts_)->find(optname); | |
449 | ||
450 | if (j == (*opts_)->end()) { | |
451 | return -ENOENT; | |
452 | } | |
453 | ||
454 | *optval = boost::get<std::string>(j->second); | |
455 | return 0; | |
456 | } | |
457 | ||
458 | int image_options_get(rbd_image_options_t opts, int optname, uint64_t* optval) | |
459 | { | |
460 | image_options_ref* opts_ = static_cast<image_options_ref*>(opts); | |
461 | ||
462 | std::map<int, image_option_type_t>::const_iterator i = | |
463 | IMAGE_OPTIONS_TYPE_MAPPING.find(optname); | |
464 | ||
465 | if (i == IMAGE_OPTIONS_TYPE_MAPPING.end() || i->second != UINT64) { | |
466 | return -EINVAL; | |
467 | } | |
468 | ||
469 | image_options_t::const_iterator j = (*opts_)->find(optname); | |
470 | ||
471 | if (j == (*opts_)->end()) { | |
472 | return -ENOENT; | |
473 | } | |
474 | ||
475 | *optval = boost::get<uint64_t>(j->second); | |
476 | return 0; | |
477 | } | |
478 | ||
479 | int image_options_is_set(rbd_image_options_t opts, int optname, | |
480 | bool* is_set) | |
481 | { | |
482 | if (IMAGE_OPTIONS_TYPE_MAPPING.find(optname) == | |
483 | IMAGE_OPTIONS_TYPE_MAPPING.end()) { | |
484 | return -EINVAL; | |
485 | } | |
486 | ||
487 | image_options_ref* opts_ = static_cast<image_options_ref*>(opts); | |
488 | *is_set = ((*opts_)->find(optname) != (*opts_)->end()); | |
489 | return 0; | |
490 | } | |
491 | ||
492 | int image_options_unset(rbd_image_options_t opts, int optname) | |
493 | { | |
494 | image_options_ref* opts_ = static_cast<image_options_ref*>(opts); | |
495 | ||
496 | std::map<int, image_option_type_t>::const_iterator i = | |
497 | IMAGE_OPTIONS_TYPE_MAPPING.find(optname); | |
498 | ||
499 | if (i == IMAGE_OPTIONS_TYPE_MAPPING.end()) { | |
11fdf7f2 | 500 | ceph_assert((*opts_)->find(optname) == (*opts_)->end()); |
7c673cae FG |
501 | return -EINVAL; |
502 | } | |
503 | ||
504 | image_options_t::const_iterator j = (*opts_)->find(optname); | |
505 | ||
506 | if (j == (*opts_)->end()) { | |
507 | return -ENOENT; | |
508 | } | |
509 | ||
510 | (*opts_)->erase(j); | |
511 | return 0; | |
512 | } | |
513 | ||
514 | void image_options_clear(rbd_image_options_t opts) | |
515 | { | |
516 | image_options_ref* opts_ = static_cast<image_options_ref*>(opts); | |
517 | ||
518 | (*opts_)->clear(); | |
519 | } | |
520 | ||
521 | bool image_options_is_empty(rbd_image_options_t opts) | |
522 | { | |
523 | image_options_ref* opts_ = static_cast<image_options_ref*>(opts); | |
524 | ||
525 | return (*opts_)->empty(); | |
526 | } | |
527 | ||
7c673cae FG |
528 | int create_v1(IoCtx& io_ctx, const char *imgname, uint64_t size, int order) |
529 | { | |
530 | CephContext *cct = (CephContext *)io_ctx.cct(); | |
531 | ||
532 | ldout(cct, 20) << __func__ << " " << &io_ctx << " name = " << imgname | |
533 | << " size = " << size << " order = " << order << dendl; | |
534 | int r = validate_pool(io_ctx, cct); | |
535 | if (r < 0) { | |
536 | return r; | |
537 | } | |
538 | ||
11fdf7f2 TL |
539 | if (!io_ctx.get_namespace().empty()) { |
540 | lderr(cct) << "attempting to add v1 image to namespace" << dendl; | |
541 | return -EINVAL; | |
542 | } | |
543 | ||
7c673cae FG |
544 | ldout(cct, 2) << "adding rbd image to directory..." << dendl; |
545 | r = tmap_set(io_ctx, imgname); | |
546 | if (r < 0) { | |
547 | lderr(cct) << "error adding image to directory: " << cpp_strerror(r) | |
548 | << dendl; | |
549 | return r; | |
550 | } | |
551 | ||
552 | Rados rados(io_ctx); | |
553 | uint64_t bid = rados.get_instance_id(); | |
554 | ||
555 | ldout(cct, 2) << "creating rbd image..." << dendl; | |
556 | struct rbd_obj_header_ondisk header; | |
557 | init_rbd_header(header, size, order, bid); | |
558 | ||
559 | bufferlist bl; | |
560 | bl.append((const char *)&header, sizeof(header)); | |
561 | ||
562 | string header_oid = util::old_header_name(imgname); | |
563 | r = io_ctx.write(header_oid, bl, bl.length(), 0); | |
564 | if (r < 0) { | |
565 | lderr(cct) << "Error writing image header: " << cpp_strerror(r) | |
566 | << dendl; | |
567 | int remove_r = tmap_rm(io_ctx, imgname); | |
568 | if (remove_r < 0) { | |
569 | lderr(cct) << "Could not remove image from directory after " | |
570 | << "header creation failed: " | |
571 | << cpp_strerror(remove_r) << dendl; | |
572 | } | |
573 | return r; | |
574 | } | |
575 | ||
576 | ldout(cct, 2) << "done." << dendl; | |
577 | return 0; | |
578 | } | |
579 | ||
580 | int create(librados::IoCtx& io_ctx, const char *imgname, uint64_t size, | |
581 | int *order) | |
582 | { | |
583 | uint64_t order_ = *order; | |
584 | ImageOptions opts; | |
585 | ||
586 | int r = opts.set(RBD_IMAGE_OPTION_ORDER, order_); | |
11fdf7f2 | 587 | ceph_assert(r == 0); |
7c673cae FG |
588 | |
589 | r = create(io_ctx, imgname, "", size, opts, "", "", false); | |
590 | ||
591 | int r1 = opts.get(RBD_IMAGE_OPTION_ORDER, &order_); | |
11fdf7f2 | 592 | ceph_assert(r1 == 0); |
7c673cae FG |
593 | *order = order_; |
594 | ||
595 | return r; | |
596 | } | |
597 | ||
598 | int create(IoCtx& io_ctx, const char *imgname, uint64_t size, | |
599 | bool old_format, uint64_t features, int *order, | |
600 | uint64_t stripe_unit, uint64_t stripe_count) | |
601 | { | |
602 | if (!order) | |
603 | return -EINVAL; | |
604 | ||
605 | uint64_t order_ = *order; | |
606 | uint64_t format = old_format ? 1 : 2; | |
607 | ImageOptions opts; | |
608 | int r; | |
609 | ||
610 | r = opts.set(RBD_IMAGE_OPTION_FORMAT, format); | |
11fdf7f2 | 611 | ceph_assert(r == 0); |
7c673cae | 612 | r = opts.set(RBD_IMAGE_OPTION_FEATURES, features); |
11fdf7f2 | 613 | ceph_assert(r == 0); |
7c673cae | 614 | r = opts.set(RBD_IMAGE_OPTION_ORDER, order_); |
11fdf7f2 | 615 | ceph_assert(r == 0); |
7c673cae | 616 | r = opts.set(RBD_IMAGE_OPTION_STRIPE_UNIT, stripe_unit); |
11fdf7f2 | 617 | ceph_assert(r == 0); |
7c673cae | 618 | r = opts.set(RBD_IMAGE_OPTION_STRIPE_COUNT, stripe_count); |
11fdf7f2 | 619 | ceph_assert(r == 0); |
7c673cae FG |
620 | |
621 | r = create(io_ctx, imgname, "", size, opts, "", "", false); | |
622 | ||
623 | int r1 = opts.get(RBD_IMAGE_OPTION_ORDER, &order_); | |
11fdf7f2 | 624 | ceph_assert(r1 == 0); |
7c673cae FG |
625 | *order = order_; |
626 | ||
627 | return r; | |
628 | } | |
629 | ||
630 | int create(IoCtx& io_ctx, const std::string &image_name, | |
631 | const std::string &image_id, uint64_t size, | |
632 | ImageOptions& opts, | |
633 | const std::string &non_primary_global_image_id, | |
634 | const std::string &primary_mirror_uuid, | |
635 | bool skip_mirror_enable) | |
636 | { | |
637 | std::string id(image_id); | |
638 | if (id.empty()) { | |
639 | id = util::generate_image_id(io_ctx); | |
640 | } | |
641 | ||
642 | CephContext *cct = (CephContext *)io_ctx.cct(); | |
92f5a8d4 TL |
643 | uint64_t option; |
644 | if (opts.get(RBD_IMAGE_OPTION_FLATTEN, &option) == 0) { | |
11fdf7f2 TL |
645 | lderr(cct) << "create does not support 'flatten' image option" << dendl; |
646 | return -EINVAL; | |
647 | } | |
92f5a8d4 TL |
648 | if (opts.get(RBD_IMAGE_OPTION_CLONE_FORMAT, &option) == 0) { |
649 | lderr(cct) << "create does not support 'clone_format' image option" | |
650 | << dendl; | |
651 | return -EINVAL; | |
652 | } | |
11fdf7f2 | 653 | |
7c673cae FG |
654 | ldout(cct, 10) << __func__ << " name=" << image_name << ", " |
655 | << "id= " << id << ", " | |
656 | << "size=" << size << ", opts=" << opts << dendl; | |
657 | ||
658 | uint64_t format; | |
659 | if (opts.get(RBD_IMAGE_OPTION_FORMAT, &format) != 0) | |
11fdf7f2 | 660 | format = cct->_conf.get_val<uint64_t>("rbd_default_format"); |
7c673cae FG |
661 | bool old_format = format == 1; |
662 | ||
663 | // make sure it doesn't already exist, in either format | |
664 | int r = detect_format(io_ctx, image_name, NULL, NULL); | |
665 | if (r != -ENOENT) { | |
666 | if (r) { | |
667 | lderr(cct) << "Could not tell if " << image_name << " already exists" | |
668 | << dendl; | |
669 | return r; | |
670 | } | |
671 | lderr(cct) << "rbd image " << image_name << " already exists" << dendl; | |
672 | return -EEXIST; | |
673 | } | |
674 | ||
675 | uint64_t order = 0; | |
676 | if (opts.get(RBD_IMAGE_OPTION_ORDER, &order) != 0 || order == 0) { | |
11fdf7f2 | 677 | order = cct->_conf.get_val<uint64_t>("rbd_default_order"); |
7c673cae FG |
678 | } |
679 | r = image::CreateRequest<>::validate_order(cct, order); | |
680 | if (r < 0) { | |
681 | return r; | |
682 | } | |
683 | ||
684 | if (old_format) { | |
11fdf7f2 TL |
685 | if ( !getenv("RBD_FORCE_ALLOW_V1") ) { |
686 | lderr(cct) << "Format 1 image creation unsupported. " << dendl; | |
687 | return -EINVAL; | |
688 | } | |
689 | lderr(cct) << "Forced V1 image creation. " << dendl; | |
7c673cae FG |
690 | r = create_v1(io_ctx, image_name.c_str(), size, order); |
691 | } else { | |
f67539c2 | 692 | AsioEngine asio_engine(io_ctx); |
7c673cae | 693 | |
11fdf7f2 TL |
694 | ConfigProxy config{cct->_conf}; |
695 | api::Config<>::apply_pool_overrides(io_ctx, &config); | |
696 | ||
1911f103 TL |
697 | uint32_t create_flags = 0U; |
698 | uint64_t mirror_image_mode = RBD_MIRROR_IMAGE_MODE_JOURNAL; | |
699 | if (skip_mirror_enable) { | |
700 | create_flags = image::CREATE_FLAG_SKIP_MIRROR_ENABLE; | |
701 | } else if (opts.get(RBD_IMAGE_OPTION_MIRROR_IMAGE_MODE, | |
702 | &mirror_image_mode) == 0) { | |
703 | create_flags = image::CREATE_FLAG_FORCE_MIRROR_ENABLE; | |
704 | } | |
705 | ||
7c673cae FG |
706 | C_SaferCond cond; |
707 | image::CreateRequest<> *req = image::CreateRequest<>::create( | |
1911f103 TL |
708 | config, io_ctx, image_name, id, size, opts, create_flags, |
709 | static_cast<cls::rbd::MirrorImageMode>(mirror_image_mode), | |
f67539c2 TL |
710 | non_primary_global_image_id, primary_mirror_uuid, |
711 | asio_engine.get_work_queue(), &cond); | |
7c673cae FG |
712 | req->send(); |
713 | ||
714 | r = cond.wait(); | |
715 | } | |
716 | ||
717 | int r1 = opts.set(RBD_IMAGE_OPTION_ORDER, order); | |
11fdf7f2 | 718 | ceph_assert(r1 == 0); |
7c673cae FG |
719 | |
720 | return r; | |
721 | } | |
722 | ||
723 | /* | |
724 | * Parent may be in different pool, hence different IoCtx | |
725 | */ | |
726 | int clone(IoCtx& p_ioctx, const char *p_name, const char *p_snap_name, | |
727 | IoCtx& c_ioctx, const char *c_name, | |
728 | uint64_t features, int *c_order, | |
729 | uint64_t stripe_unit, int stripe_count) | |
730 | { | |
731 | uint64_t order = *c_order; | |
732 | ||
733 | ImageOptions opts; | |
734 | opts.set(RBD_IMAGE_OPTION_FEATURES, features); | |
735 | opts.set(RBD_IMAGE_OPTION_ORDER, order); | |
736 | opts.set(RBD_IMAGE_OPTION_STRIPE_UNIT, stripe_unit); | |
737 | opts.set(RBD_IMAGE_OPTION_STRIPE_COUNT, stripe_count); | |
738 | ||
11fdf7f2 TL |
739 | int r = clone(p_ioctx, nullptr, p_name, p_snap_name, c_ioctx, nullptr, |
740 | c_name, opts, "", ""); | |
7c673cae FG |
741 | opts.get(RBD_IMAGE_OPTION_ORDER, &order); |
742 | *c_order = order; | |
743 | return r; | |
744 | } | |
745 | ||
11fdf7f2 TL |
746 | int clone(IoCtx& p_ioctx, const char *p_id, const char *p_name, |
747 | const char *p_snap_name, IoCtx& c_ioctx, const char *c_id, | |
748 | const char *c_name, ImageOptions& c_opts, | |
749 | const std::string &non_primary_global_image_id, | |
750 | const std::string &primary_mirror_uuid) | |
7c673cae | 751 | { |
11fdf7f2 TL |
752 | ceph_assert((p_id == nullptr) ^ (p_name == nullptr)); |
753 | ||
7c673cae | 754 | CephContext *cct = (CephContext *)p_ioctx.cct(); |
11fdf7f2 | 755 | if (p_snap_name == nullptr) { |
7c673cae FG |
756 | lderr(cct) << "image to be cloned must be a snapshot" << dendl; |
757 | return -EINVAL; | |
758 | } | |
759 | ||
11fdf7f2 TL |
760 | uint64_t flatten; |
761 | if (c_opts.get(RBD_IMAGE_OPTION_FLATTEN, &flatten) == 0) { | |
762 | lderr(cct) << "clone does not support 'flatten' image option" << dendl; | |
763 | return -EINVAL; | |
7c673cae FG |
764 | } |
765 | ||
11fdf7f2 TL |
766 | int r; |
767 | std::string parent_id; | |
768 | if (p_id == nullptr) { | |
769 | r = cls_client::dir_get_id(&p_ioctx, RBD_DIRECTORY, p_name, | |
770 | &parent_id); | |
771 | if (r < 0) { | |
772 | if (r != -ENOENT) { | |
773 | lderr(cct) << "failed to retrieve parent image id: " | |
774 | << cpp_strerror(r) << dendl; | |
775 | } | |
776 | return r; | |
777 | } | |
778 | } else { | |
779 | parent_id = p_id; | |
7c673cae | 780 | } |
7c673cae | 781 | |
11fdf7f2 TL |
782 | std::string clone_id; |
783 | if (c_id == nullptr) { | |
784 | clone_id = util::generate_image_id(c_ioctx); | |
785 | } else { | |
786 | clone_id = c_id; | |
7c673cae FG |
787 | } |
788 | ||
7c673cae FG |
789 | ldout(cct, 10) << __func__ << " " |
790 | << "c_name=" << c_name << ", " | |
11fdf7f2 | 791 | << "c_id= " << clone_id << ", " |
7c673cae FG |
792 | << "c_opts=" << c_opts << dendl; |
793 | ||
11fdf7f2 TL |
794 | ConfigProxy config{reinterpret_cast<CephContext *>(c_ioctx.cct())->_conf}; |
795 | api::Config<>::apply_pool_overrides(c_ioctx, &config); | |
796 | ||
f67539c2 | 797 | AsioEngine asio_engine(p_ioctx); |
7c673cae FG |
798 | |
799 | C_SaferCond cond; | |
800 | auto *req = image::CloneRequest<>::create( | |
9f95a23c TL |
801 | config, p_ioctx, parent_id, p_snap_name, |
802 | {cls::rbd::UserSnapshotNamespace{}}, CEPH_NOSNAP, c_ioctx, c_name, | |
803 | clone_id, c_opts, cls::rbd::MIRROR_IMAGE_MODE_JOURNAL, | |
f67539c2 TL |
804 | non_primary_global_image_id, primary_mirror_uuid, |
805 | asio_engine.get_work_queue(), &cond); | |
7c673cae FG |
806 | req->send(); |
807 | ||
11fdf7f2 TL |
808 | r = cond.wait(); |
809 | if (r < 0) { | |
810 | return r; | |
811 | } | |
812 | ||
813 | return 0; | |
7c673cae FG |
814 | } |
815 | ||
816 | int rename(IoCtx& io_ctx, const char *srcname, const char *dstname) | |
817 | { | |
818 | CephContext *cct = (CephContext *)io_ctx.cct(); | |
819 | ldout(cct, 20) << "rename " << &io_ctx << " " << srcname << " -> " | |
820 | << dstname << dendl; | |
821 | ||
822 | ImageCtx *ictx = new ImageCtx(srcname, "", "", io_ctx, false); | |
11fdf7f2 | 823 | int r = ictx->state->open(0); |
7c673cae | 824 | if (r < 0) { |
181888fb | 825 | lderr(cct) << "error opening source image: " << cpp_strerror(r) << dendl; |
7c673cae FG |
826 | return r; |
827 | } | |
828 | BOOST_SCOPE_EXIT((ictx)) { | |
829 | ictx->state->close(); | |
830 | } BOOST_SCOPE_EXIT_END | |
831 | ||
832 | return ictx->operations->rename(dstname); | |
833 | } | |
834 | ||
835 | int info(ImageCtx *ictx, image_info_t& info, size_t infosize) | |
836 | { | |
837 | ldout(ictx->cct, 20) << "info " << ictx << dendl; | |
838 | ||
839 | int r = ictx->state->refresh_if_required(); | |
840 | if (r < 0) | |
841 | return r; | |
842 | ||
843 | image_info(ictx, info, infosize); | |
844 | return 0; | |
845 | } | |
846 | ||
847 | int get_old_format(ImageCtx *ictx, uint8_t *old) | |
848 | { | |
849 | int r = ictx->state->refresh_if_required(); | |
850 | if (r < 0) | |
851 | return r; | |
852 | *old = ictx->old_format; | |
853 | return 0; | |
854 | } | |
855 | ||
856 | int get_size(ImageCtx *ictx, uint64_t *size) | |
857 | { | |
858 | int r = ictx->state->refresh_if_required(); | |
859 | if (r < 0) | |
860 | return r; | |
9f95a23c | 861 | std::shared_lock l2{ictx->image_lock}; |
f67539c2 | 862 | *size = ictx->get_effective_image_size(ictx->snap_id); |
7c673cae FG |
863 | return 0; |
864 | } | |
865 | ||
866 | int get_features(ImageCtx *ictx, uint64_t *features) | |
867 | { | |
868 | int r = ictx->state->refresh_if_required(); | |
869 | if (r < 0) | |
870 | return r; | |
9f95a23c | 871 | std::shared_lock l{ictx->image_lock}; |
7c673cae FG |
872 | *features = ictx->features; |
873 | return 0; | |
874 | } | |
875 | ||
876 | int get_overlap(ImageCtx *ictx, uint64_t *overlap) | |
877 | { | |
878 | int r = ictx->state->refresh_if_required(); | |
879 | if (r < 0) | |
880 | return r; | |
9f95a23c | 881 | std::shared_lock image_locker{ictx->image_lock}; |
7c673cae FG |
882 | return ictx->get_parent_overlap(ictx->snap_id, overlap); |
883 | } | |
884 | ||
7c673cae FG |
885 | int get_flags(ImageCtx *ictx, uint64_t *flags) |
886 | { | |
887 | int r = ictx->state->refresh_if_required(); | |
888 | if (r < 0) { | |
889 | return r; | |
890 | } | |
891 | ||
9f95a23c | 892 | std::shared_lock l2{ictx->image_lock}; |
7c673cae FG |
893 | return ictx->get_flags(ictx->snap_id, flags); |
894 | } | |
895 | ||
896 | int set_image_notification(ImageCtx *ictx, int fd, int type) | |
897 | { | |
898 | CephContext *cct = ictx->cct; | |
899 | ldout(cct, 20) << __func__ << " " << ictx << " fd " << fd << " type" << type << dendl; | |
900 | ||
901 | int r = ictx->state->refresh_if_required(); | |
902 | if (r < 0) { | |
903 | return r; | |
904 | } | |
905 | ||
906 | if (ictx->event_socket.is_valid()) | |
907 | return -EINVAL; | |
908 | return ictx->event_socket.init(fd, type); | |
909 | } | |
910 | ||
911 | int is_exclusive_lock_owner(ImageCtx *ictx, bool *is_owner) | |
912 | { | |
91327a77 AA |
913 | CephContext *cct = ictx->cct; |
914 | ldout(cct, 20) << __func__ << ": ictx=" << ictx << dendl; | |
7c673cae FG |
915 | *is_owner = false; |
916 | ||
9f95a23c | 917 | std::shared_lock owner_locker{ictx->owner_lock}; |
91327a77 | 918 | if (ictx->exclusive_lock == nullptr) { |
7c673cae FG |
919 | return 0; |
920 | } | |
921 | ||
f67539c2 | 922 | // might have been blocklisted by peer -- ensure we still own |
7c673cae FG |
923 | // the lock by pinging the OSD |
924 | int r = ictx->exclusive_lock->assert_header_locked(); | |
31f18b77 FG |
925 | if (r == -EBUSY || r == -ENOENT) { |
926 | return 0; | |
927 | } else if (r < 0) { | |
7c673cae FG |
928 | return r; |
929 | } | |
930 | ||
931 | *is_owner = true; | |
932 | return 0; | |
933 | } | |
934 | ||
935 | int lock_acquire(ImageCtx *ictx, rbd_lock_mode_t lock_mode) | |
936 | { | |
937 | CephContext *cct = ictx->cct; | |
938 | ldout(cct, 20) << __func__ << ": ictx=" << ictx << ", " | |
939 | << "lock_mode=" << lock_mode << dendl; | |
940 | ||
941 | if (lock_mode != RBD_LOCK_MODE_EXCLUSIVE) { | |
942 | return -EOPNOTSUPP; | |
943 | } | |
944 | ||
945 | C_SaferCond lock_ctx; | |
946 | { | |
9f95a23c | 947 | std::unique_lock l{ictx->owner_lock}; |
7c673cae FG |
948 | |
949 | if (ictx->exclusive_lock == nullptr) { | |
950 | lderr(cct) << "exclusive-lock feature is not enabled" << dendl; | |
951 | return -EINVAL; | |
952 | } | |
953 | ||
954 | if (ictx->get_exclusive_lock_policy()->may_auto_request_lock()) { | |
955 | ictx->set_exclusive_lock_policy( | |
956 | new exclusive_lock::StandardPolicy(ictx)); | |
957 | } | |
958 | ||
959 | if (ictx->exclusive_lock->is_lock_owner()) { | |
960 | return 0; | |
961 | } | |
962 | ||
963 | ictx->exclusive_lock->acquire_lock(&lock_ctx); | |
964 | } | |
965 | ||
966 | int r = lock_ctx.wait(); | |
967 | if (r < 0) { | |
968 | lderr(cct) << "failed to request exclusive lock: " << cpp_strerror(r) | |
969 | << dendl; | |
970 | return r; | |
971 | } | |
972 | ||
9f95a23c | 973 | std::shared_lock l{ictx->owner_lock}; |
91327a77 AA |
974 | if (ictx->exclusive_lock == nullptr) { |
975 | return -EINVAL; | |
976 | } else if (!ictx->exclusive_lock->is_lock_owner()) { | |
7c673cae | 977 | lderr(cct) << "failed to acquire exclusive lock" << dendl; |
91327a77 | 978 | return ictx->exclusive_lock->get_unlocked_op_error(); |
7c673cae FG |
979 | } |
980 | ||
981 | return 0; | |
982 | } | |
983 | ||
984 | int lock_release(ImageCtx *ictx) | |
985 | { | |
986 | CephContext *cct = ictx->cct; | |
987 | ldout(cct, 20) << __func__ << ": ictx=" << ictx << dendl; | |
988 | ||
989 | C_SaferCond lock_ctx; | |
990 | { | |
9f95a23c | 991 | std::unique_lock l{ictx->owner_lock}; |
7c673cae FG |
992 | |
993 | if (ictx->exclusive_lock == nullptr || | |
994 | !ictx->exclusive_lock->is_lock_owner()) { | |
995 | lderr(cct) << "not exclusive lock owner" << dendl; | |
996 | return -EINVAL; | |
997 | } | |
998 | ||
999 | ictx->exclusive_lock->release_lock(&lock_ctx); | |
1000 | } | |
1001 | ||
1002 | int r = lock_ctx.wait(); | |
1003 | if (r < 0) { | |
1004 | lderr(cct) << "failed to release exclusive lock: " << cpp_strerror(r) | |
1005 | << dendl; | |
1006 | return r; | |
1007 | } | |
1008 | return 0; | |
1009 | } | |
1010 | ||
1011 | int lock_get_owners(ImageCtx *ictx, rbd_lock_mode_t *lock_mode, | |
1012 | std::list<std::string> *lock_owners) | |
1013 | { | |
1014 | CephContext *cct = ictx->cct; | |
1015 | ldout(cct, 20) << __func__ << ": ictx=" << ictx << dendl; | |
1016 | ||
7c673cae FG |
1017 | managed_lock::Locker locker; |
1018 | C_SaferCond get_owner_ctx; | |
9f95a23c TL |
1019 | { |
1020 | std::shared_lock owner_locker{ictx->owner_lock}; | |
1021 | ||
1022 | if (ictx->exclusive_lock == nullptr) { | |
1023 | lderr(cct) << "exclusive-lock feature is not enabled" << dendl; | |
1024 | return -EINVAL; | |
1025 | } | |
1026 | ||
1027 | ictx->exclusive_lock->get_locker(&locker, &get_owner_ctx); | |
1028 | } | |
1029 | ||
7c673cae FG |
1030 | int r = get_owner_ctx.wait(); |
1031 | if (r == -ENOENT) { | |
1032 | return r; | |
1033 | } else if (r < 0) { | |
1034 | lderr(cct) << "failed to determine current lock owner: " | |
1035 | << cpp_strerror(r) << dendl; | |
1036 | return r; | |
1037 | } | |
1038 | ||
1039 | *lock_mode = RBD_LOCK_MODE_EXCLUSIVE; | |
1040 | lock_owners->clear(); | |
1041 | lock_owners->emplace_back(locker.address); | |
1042 | return 0; | |
1043 | } | |
1044 | ||
1045 | int lock_break(ImageCtx *ictx, rbd_lock_mode_t lock_mode, | |
11fdf7f2 | 1046 | const std::string &lock_owner) { |
7c673cae FG |
1047 | CephContext *cct = ictx->cct; |
1048 | ldout(cct, 20) << __func__ << ": ictx=" << ictx << ", " | |
1049 | << "lock_mode=" << lock_mode << ", " | |
1050 | << "lock_owner=" << lock_owner << dendl; | |
1051 | ||
1052 | if (lock_mode != RBD_LOCK_MODE_EXCLUSIVE) { | |
1053 | return -EOPNOTSUPP; | |
1054 | } | |
1055 | ||
1056 | if (ictx->read_only) { | |
1057 | return -EROFS; | |
1058 | } | |
1059 | ||
1060 | managed_lock::Locker locker; | |
1061 | C_SaferCond get_owner_ctx; | |
1062 | { | |
9f95a23c | 1063 | std::shared_lock l{ictx->owner_lock}; |
7c673cae FG |
1064 | |
1065 | if (ictx->exclusive_lock == nullptr) { | |
1066 | lderr(cct) << "exclusive-lock feature is not enabled" << dendl; | |
1067 | return -EINVAL; | |
1068 | } | |
1069 | ||
1070 | ictx->exclusive_lock->get_locker(&locker, &get_owner_ctx); | |
1071 | } | |
1072 | int r = get_owner_ctx.wait(); | |
1073 | if (r == -ENOENT) { | |
1074 | return r; | |
1075 | } else if (r < 0) { | |
1076 | lderr(cct) << "failed to determine current lock owner: " | |
1077 | << cpp_strerror(r) << dendl; | |
1078 | return r; | |
1079 | } | |
1080 | ||
1081 | if (locker.address != lock_owner) { | |
1082 | return -EBUSY; | |
1083 | } | |
1084 | ||
1085 | C_SaferCond break_ctx; | |
1086 | { | |
9f95a23c | 1087 | std::shared_lock l{ictx->owner_lock}; |
7c673cae FG |
1088 | |
1089 | if (ictx->exclusive_lock == nullptr) { | |
1090 | lderr(cct) << "exclusive-lock feature is not enabled" << dendl; | |
1091 | return -EINVAL; | |
1092 | } | |
1093 | ||
1094 | ictx->exclusive_lock->break_lock(locker, true, &break_ctx); | |
1095 | } | |
1096 | r = break_ctx.wait(); | |
1097 | if (r == -ENOENT) { | |
1098 | return r; | |
1099 | } else if (r < 0) { | |
1100 | lderr(cct) << "failed to break lock: " << cpp_strerror(r) << dendl; | |
1101 | return r; | |
1102 | } | |
1103 | return 0; | |
1104 | } | |
1105 | ||
7c673cae FG |
1106 | int copy(ImageCtx *src, IoCtx& dest_md_ctx, const char *destname, |
1107 | ImageOptions& opts, ProgressContext &prog_ctx, size_t sparse_size) | |
1108 | { | |
1109 | CephContext *cct = (CephContext *)dest_md_ctx.cct(); | |
92f5a8d4 TL |
1110 | uint64_t option; |
1111 | if (opts.get(RBD_IMAGE_OPTION_FLATTEN, &option) == 0) { | |
11fdf7f2 TL |
1112 | lderr(cct) << "copy does not support 'flatten' image option" << dendl; |
1113 | return -EINVAL; | |
1114 | } | |
92f5a8d4 TL |
1115 | if (opts.get(RBD_IMAGE_OPTION_CLONE_FORMAT, &option) == 0) { |
1116 | lderr(cct) << "copy does not support 'clone_format' image option" | |
1117 | << dendl; | |
1118 | return -EINVAL; | |
1119 | } | |
11fdf7f2 | 1120 | |
7c673cae FG |
1121 | ldout(cct, 20) << "copy " << src->name |
1122 | << (src->snap_name.length() ? "@" + src->snap_name : "") | |
1123 | << " -> " << destname << " opts = " << opts << dendl; | |
1124 | ||
9f95a23c | 1125 | src->image_lock.lock_shared(); |
7c673cae FG |
1126 | uint64_t features = src->features; |
1127 | uint64_t src_size = src->get_image_size(src->snap_id); | |
9f95a23c | 1128 | src->image_lock.unlock_shared(); |
e306af50 | 1129 | uint64_t format = 2; |
7c673cae FG |
1130 | if (opts.get(RBD_IMAGE_OPTION_FORMAT, &format) != 0) { |
1131 | opts.set(RBD_IMAGE_OPTION_FORMAT, format); | |
1132 | } | |
1133 | uint64_t stripe_unit = src->stripe_unit; | |
1134 | if (opts.get(RBD_IMAGE_OPTION_STRIPE_UNIT, &stripe_unit) != 0) { | |
1135 | opts.set(RBD_IMAGE_OPTION_STRIPE_UNIT, stripe_unit); | |
1136 | } | |
1137 | uint64_t stripe_count = src->stripe_count; | |
1138 | if (opts.get(RBD_IMAGE_OPTION_STRIPE_COUNT, &stripe_count) != 0) { | |
1139 | opts.set(RBD_IMAGE_OPTION_STRIPE_COUNT, stripe_count); | |
1140 | } | |
1141 | uint64_t order = src->order; | |
1142 | if (opts.get(RBD_IMAGE_OPTION_ORDER, &order) != 0) { | |
1143 | opts.set(RBD_IMAGE_OPTION_ORDER, order); | |
1144 | } | |
1145 | if (opts.get(RBD_IMAGE_OPTION_FEATURES, &features) != 0) { | |
1146 | opts.set(RBD_IMAGE_OPTION_FEATURES, features); | |
1147 | } | |
1148 | if (features & ~RBD_FEATURES_ALL) { | |
1149 | lderr(cct) << "librbd does not support requested features" << dendl; | |
1150 | return -ENOSYS; | |
1151 | } | |
1152 | ||
1153 | int r = create(dest_md_ctx, destname, "", src_size, opts, "", "", false); | |
1154 | if (r < 0) { | |
1155 | lderr(cct) << "header creation failed" << dendl; | |
1156 | return r; | |
1157 | } | |
1158 | opts.set(RBD_IMAGE_OPTION_ORDER, static_cast<uint64_t>(order)); | |
1159 | ||
11fdf7f2 TL |
1160 | ImageCtx *dest = new librbd::ImageCtx(destname, "", nullptr, dest_md_ctx, |
1161 | false); | |
1162 | r = dest->state->open(0); | |
7c673cae FG |
1163 | if (r < 0) { |
1164 | lderr(cct) << "failed to read newly created header" << dendl; | |
1165 | return r; | |
1166 | } | |
1167 | ||
1168 | r = copy(src, dest, prog_ctx, sparse_size); | |
1169 | ||
1170 | int close_r = dest->state->close(); | |
1171 | if (r == 0 && close_r < 0) { | |
1172 | r = close_r; | |
1173 | } | |
1174 | return r; | |
1175 | } | |
1176 | ||
1177 | class C_CopyWrite : public Context { | |
1178 | public: | |
1179 | C_CopyWrite(bufferlist *bl, Context* ctx) | |
1180 | : m_bl(bl), m_ctx(ctx) {} | |
1181 | void finish(int r) override { | |
1182 | delete m_bl; | |
1183 | m_ctx->complete(r); | |
1184 | } | |
1185 | private: | |
1186 | bufferlist *m_bl; | |
1187 | Context *m_ctx; | |
1188 | }; | |
1189 | ||
1190 | class C_CopyRead : public Context { | |
1191 | public: | |
1192 | C_CopyRead(SimpleThrottle *throttle, ImageCtx *dest, uint64_t offset, | |
1193 | bufferlist *bl, size_t sparse_size) | |
1194 | : m_throttle(throttle), m_dest(dest), m_offset(offset), m_bl(bl), | |
1195 | m_sparse_size(sparse_size) { | |
1196 | m_throttle->start_op(); | |
1197 | } | |
1198 | void finish(int r) override { | |
1199 | if (r < 0) { | |
1200 | lderr(m_dest->cct) << "error reading from source image at offset " | |
1201 | << m_offset << ": " << cpp_strerror(r) << dendl; | |
1202 | delete m_bl; | |
1203 | m_throttle->end_op(r); | |
1204 | return; | |
1205 | } | |
11fdf7f2 | 1206 | ceph_assert(m_bl->length() == (size_t)r); |
7c673cae FG |
1207 | |
1208 | if (m_bl->is_zero()) { | |
1209 | delete m_bl; | |
1210 | m_throttle->end_op(r); | |
1211 | return; | |
1212 | } | |
1213 | ||
1214 | if (!m_sparse_size) { | |
1215 | m_sparse_size = (1 << m_dest->order); | |
1216 | } | |
1217 | ||
1218 | auto *throttle = m_throttle; | |
9f95a23c | 1219 | auto *end_op_ctx = new LambdaContext([throttle](int r) { |
7c673cae FG |
1220 | throttle->end_op(r); |
1221 | }); | |
1222 | auto gather_ctx = new C_Gather(m_dest->cct, end_op_ctx); | |
1223 | ||
11fdf7f2 | 1224 | m_bl->rebuild(buffer::ptr_node::create(m_bl->length())); |
7c673cae FG |
1225 | size_t write_offset = 0; |
1226 | size_t write_length = 0; | |
1227 | size_t offset = 0; | |
1228 | size_t length = m_bl->length(); | |
11fdf7f2 | 1229 | const auto& m_ptr = m_bl->front(); |
7c673cae FG |
1230 | while (offset < length) { |
1231 | if (util::calc_sparse_extent(m_ptr, | |
1232 | m_sparse_size, | |
1233 | length, | |
1234 | &write_offset, | |
1235 | &write_length, | |
1236 | &offset)) { | |
7c673cae | 1237 | bufferlist *write_bl = new bufferlist(); |
11fdf7f2 TL |
1238 | write_bl->push_back( |
1239 | buffer::ptr_node::create(m_ptr, write_offset, write_length)); | |
7c673cae FG |
1240 | Context *ctx = new C_CopyWrite(write_bl, gather_ctx->new_sub()); |
1241 | auto comp = io::AioCompletion::create(ctx); | |
1242 | ||
1243 | // coordinate through AIO WQ to ensure lock is acquired if needed | |
f67539c2 TL |
1244 | api::Io<>::aio_write(*m_dest, comp, m_offset + write_offset, |
1245 | write_length, std::move(*write_bl), | |
1246 | LIBRADOS_OP_FLAG_FADVISE_DONTNEED, | |
1247 | std::move(read_trace)); | |
7c673cae FG |
1248 | write_offset = offset; |
1249 | write_length = 0; | |
1250 | } | |
1251 | } | |
1252 | delete m_bl; | |
11fdf7f2 | 1253 | ceph_assert(gather_ctx->get_sub_created_count() > 0); |
7c673cae FG |
1254 | gather_ctx->activate(); |
1255 | } | |
1256 | ||
31f18b77 FG |
1257 | ZTracer::Trace read_trace; |
1258 | ||
7c673cae FG |
1259 | private: |
1260 | SimpleThrottle *m_throttle; | |
1261 | ImageCtx *m_dest; | |
1262 | uint64_t m_offset; | |
1263 | bufferlist *m_bl; | |
1264 | size_t m_sparse_size; | |
1265 | }; | |
1266 | ||
1267 | int copy(ImageCtx *src, ImageCtx *dest, ProgressContext &prog_ctx, size_t sparse_size) | |
1268 | { | |
9f95a23c | 1269 | src->image_lock.lock_shared(); |
7c673cae | 1270 | uint64_t src_size = src->get_image_size(src->snap_id); |
9f95a23c | 1271 | src->image_lock.unlock_shared(); |
7c673cae | 1272 | |
9f95a23c | 1273 | dest->image_lock.lock_shared(); |
7c673cae | 1274 | uint64_t dest_size = dest->get_image_size(dest->snap_id); |
9f95a23c | 1275 | dest->image_lock.unlock_shared(); |
7c673cae FG |
1276 | |
1277 | CephContext *cct = src->cct; | |
1278 | if (dest_size < src_size) { | |
1279 | lderr(cct) << " src size " << src_size << " > dest size " | |
1280 | << dest_size << dendl; | |
1281 | return -EINVAL; | |
1282 | } | |
b32b8144 | 1283 | |
20effc67 TL |
1284 | // ensure previous writes are visible to dest |
1285 | C_SaferCond flush_ctx; | |
1286 | { | |
1287 | auto aio_comp = io::AioCompletion::create_and_start(&flush_ctx, src, | |
1288 | io::AIO_TYPE_FLUSH); | |
1289 | auto req = io::ImageDispatchSpec::create_flush( | |
1290 | *src, io::IMAGE_DISPATCH_LAYER_INTERNAL_START, | |
1291 | aio_comp, io::FLUSH_SOURCE_INTERNAL, {}); | |
1292 | req->send(); | |
1293 | } | |
1294 | int r = flush_ctx.wait(); | |
1295 | if (r < 0) { | |
1296 | return r; | |
1297 | } | |
1298 | ||
9f95a23c TL |
1299 | C_SaferCond ctx; |
1300 | auto req = deep_copy::MetadataCopyRequest<>::create( | |
1301 | src, dest, &ctx); | |
1302 | req->send(); | |
b32b8144 | 1303 | |
20effc67 | 1304 | r = ctx.wait(); |
9f95a23c TL |
1305 | if (r < 0) { |
1306 | lderr(cct) << "failed to copy metadata: " << cpp_strerror(r) << dendl; | |
1307 | return r; | |
7c673cae FG |
1308 | } |
1309 | ||
31f18b77 | 1310 | ZTracer::Trace trace; |
181888fb | 1311 | if (src->blkin_trace_all) { |
31f18b77 FG |
1312 | trace.init("copy", &src->trace_endpoint); |
1313 | } | |
1314 | ||
11fdf7f2 | 1315 | SimpleThrottle throttle(src->config.get_val<uint64_t>("rbd_concurrent_management_ops"), false); |
7c673cae | 1316 | uint64_t period = src->get_stripe_period(); |
31f18b77 FG |
1317 | unsigned fadvise_flags = LIBRADOS_OP_FLAG_FADVISE_SEQUENTIAL | |
1318 | LIBRADOS_OP_FLAG_FADVISE_NOCACHE; | |
11fdf7f2 | 1319 | uint64_t object_id = 0; |
7c673cae FG |
1320 | for (uint64_t offset = 0; offset < src_size; offset += period) { |
1321 | if (throttle.pending_error()) { | |
1322 | return throttle.wait_for_ret(); | |
1323 | } | |
1324 | ||
11fdf7f2 | 1325 | { |
9f95a23c | 1326 | std::shared_lock image_locker{src->image_lock}; |
11fdf7f2 TL |
1327 | if (src->object_map != nullptr) { |
1328 | bool skip = true; | |
1329 | // each period is related to src->stripe_count objects, check them all | |
1330 | for (uint64_t i=0; i < src->stripe_count; i++) { | |
1331 | if (object_id < src->object_map->size() && | |
1332 | src->object_map->object_may_exist(object_id)) { | |
1333 | skip = false; | |
1334 | } | |
1335 | ++object_id; | |
1336 | } | |
1337 | ||
1338 | if (skip) continue; | |
1339 | } else { | |
1340 | object_id += src->stripe_count; | |
1341 | } | |
1342 | } | |
1343 | ||
20effc67 | 1344 | uint64_t len = std::min(period, src_size - offset); |
7c673cae | 1345 | bufferlist *bl = new bufferlist(); |
31f18b77 FG |
1346 | auto ctx = new C_CopyRead(&throttle, dest, offset, bl, sparse_size); |
1347 | auto comp = io::AioCompletion::create_and_start<Context>( | |
1348 | ctx, src, io::AIO_TYPE_READ); | |
f67539c2 TL |
1349 | auto req = io::ImageDispatchSpec::create_read( |
1350 | *src, io::IMAGE_DISPATCH_LAYER_NONE, comp, | |
1351 | {{offset, len}}, io::ReadResult{bl}, | |
1352 | src->get_data_io_context(), fadvise_flags, 0, trace); | |
31f18b77 | 1353 | |
f67539c2 TL |
1354 | ctx->read_trace = trace; |
1355 | req->send(); | |
31f18b77 | 1356 | |
7c673cae FG |
1357 | prog_ctx.update_progress(offset, src_size); |
1358 | } | |
1359 | ||
1360 | r = throttle.wait_for_ret(); | |
1361 | if (r >= 0) | |
1362 | prog_ctx.update_progress(src_size, src_size); | |
1363 | return r; | |
1364 | } | |
1365 | ||
7c673cae FG |
1366 | int list_lockers(ImageCtx *ictx, |
1367 | std::list<locker_t> *lockers, | |
1368 | bool *exclusive, | |
1369 | string *tag) | |
1370 | { | |
1371 | ldout(ictx->cct, 20) << "list_locks on image " << ictx << dendl; | |
1372 | ||
1373 | int r = ictx->state->refresh_if_required(); | |
1374 | if (r < 0) | |
1375 | return r; | |
1376 | ||
9f95a23c | 1377 | std::shared_lock locker{ictx->image_lock}; |
7c673cae FG |
1378 | if (exclusive) |
1379 | *exclusive = ictx->exclusive_locked; | |
1380 | if (tag) | |
1381 | *tag = ictx->lock_tag; | |
1382 | if (lockers) { | |
1383 | lockers->clear(); | |
1384 | map<rados::cls::lock::locker_id_t, | |
1385 | rados::cls::lock::locker_info_t>::const_iterator it; | |
1386 | for (it = ictx->lockers.begin(); it != ictx->lockers.end(); ++it) { | |
1387 | locker_t locker; | |
1388 | locker.client = stringify(it->first.locker); | |
1389 | locker.cookie = it->first.cookie; | |
11fdf7f2 | 1390 | locker.address = it->second.addr.get_legacy_str(); |
7c673cae FG |
1391 | lockers->push_back(locker); |
1392 | } | |
1393 | } | |
1394 | ||
1395 | return 0; | |
1396 | } | |
1397 | ||
1398 | int lock(ImageCtx *ictx, bool exclusive, const string& cookie, | |
1399 | const string& tag) | |
1400 | { | |
1401 | ldout(ictx->cct, 20) << "lock image " << ictx << " exclusive=" << exclusive | |
1402 | << " cookie='" << cookie << "' tag='" << tag << "'" | |
1403 | << dendl; | |
1404 | ||
1405 | int r = ictx->state->refresh_if_required(); | |
1406 | if (r < 0) | |
1407 | return r; | |
1408 | ||
1409 | /** | |
1410 | * If we wanted we could do something more intelligent, like local | |
1411 | * checks that we think we will succeed. But for now, let's not | |
1412 | * duplicate that code. | |
1413 | */ | |
1414 | { | |
9f95a23c | 1415 | std::shared_lock locker{ictx->image_lock}; |
7c673cae | 1416 | r = rados::cls::lock::lock(&ictx->md_ctx, ictx->header_oid, RBD_LOCK_NAME, |
f67539c2 | 1417 | exclusive ? ClsLockType::EXCLUSIVE : ClsLockType::SHARED, |
7c673cae FG |
1418 | cookie, tag, "", utime_t(), 0); |
1419 | if (r < 0) { | |
1420 | return r; | |
1421 | } | |
1422 | } | |
1423 | ||
1424 | ictx->notify_update(); | |
1425 | return 0; | |
1426 | } | |
1427 | ||
1428 | int unlock(ImageCtx *ictx, const string& cookie) | |
1429 | { | |
1430 | ldout(ictx->cct, 20) << "unlock image " << ictx | |
1431 | << " cookie='" << cookie << "'" << dendl; | |
1432 | ||
1433 | int r = ictx->state->refresh_if_required(); | |
1434 | if (r < 0) | |
1435 | return r; | |
1436 | ||
1437 | { | |
9f95a23c | 1438 | std::shared_lock locker{ictx->image_lock}; |
7c673cae FG |
1439 | r = rados::cls::lock::unlock(&ictx->md_ctx, ictx->header_oid, |
1440 | RBD_LOCK_NAME, cookie); | |
1441 | if (r < 0) { | |
1442 | return r; | |
1443 | } | |
1444 | } | |
1445 | ||
1446 | ictx->notify_update(); | |
1447 | return 0; | |
1448 | } | |
1449 | ||
1450 | int break_lock(ImageCtx *ictx, const string& client, | |
1451 | const string& cookie) | |
1452 | { | |
1453 | ldout(ictx->cct, 20) << "break_lock image " << ictx << " client='" << client | |
1454 | << "' cookie='" << cookie << "'" << dendl; | |
1455 | ||
1456 | int r = ictx->state->refresh_if_required(); | |
1457 | if (r < 0) | |
1458 | return r; | |
1459 | ||
1460 | entity_name_t lock_client; | |
1461 | if (!lock_client.parse(client)) { | |
1462 | lderr(ictx->cct) << "Unable to parse client '" << client | |
1463 | << "'" << dendl; | |
1464 | return -EINVAL; | |
1465 | } | |
1466 | ||
f67539c2 | 1467 | if (ictx->config.get_val<bool>("rbd_blocklist_on_break_lock")) { |
7c673cae FG |
1468 | typedef std::map<rados::cls::lock::locker_id_t, |
1469 | rados::cls::lock::locker_info_t> Lockers; | |
1470 | Lockers lockers; | |
1471 | ClsLockType lock_type; | |
1472 | std::string lock_tag; | |
1473 | r = rados::cls::lock::get_lock_info(&ictx->md_ctx, ictx->header_oid, | |
1474 | RBD_LOCK_NAME, &lockers, &lock_type, | |
1475 | &lock_tag); | |
1476 | if (r < 0) { | |
1477 | lderr(ictx->cct) << "unable to retrieve lock info: " << cpp_strerror(r) | |
1478 | << dendl; | |
1479 | return r; | |
1480 | } | |
1481 | ||
1482 | std::string client_address; | |
1483 | for (Lockers::iterator it = lockers.begin(); | |
1484 | it != lockers.end(); ++it) { | |
1485 | if (it->first.locker == lock_client) { | |
11fdf7f2 | 1486 | client_address = it->second.addr.get_legacy_str(); |
7c673cae FG |
1487 | break; |
1488 | } | |
1489 | } | |
1490 | if (client_address.empty()) { | |
1491 | return -ENOENT; | |
1492 | } | |
1493 | ||
7c673cae | 1494 | librados::Rados rados(ictx->md_ctx); |
f67539c2 | 1495 | r = rados.blocklist_add( |
11fdf7f2 | 1496 | client_address, |
f67539c2 | 1497 | ictx->config.get_val<uint64_t>("rbd_blocklist_expire_seconds")); |
7c673cae | 1498 | if (r < 0) { |
f67539c2 | 1499 | lderr(ictx->cct) << "unable to blocklist client: " << cpp_strerror(r) |
7c673cae FG |
1500 | << dendl; |
1501 | return r; | |
1502 | } | |
1503 | } | |
1504 | ||
1505 | r = rados::cls::lock::break_lock(&ictx->md_ctx, ictx->header_oid, | |
1506 | RBD_LOCK_NAME, cookie, lock_client); | |
1507 | if (r < 0) | |
1508 | return r; | |
1509 | ictx->notify_update(); | |
1510 | return 0; | |
1511 | } | |
1512 | ||
1513 | void rbd_ctx_cb(completion_t cb, void *arg) | |
1514 | { | |
1515 | Context *ctx = reinterpret_cast<Context *>(arg); | |
1516 | auto comp = reinterpret_cast<io::AioCompletion *>(cb); | |
1517 | ctx->complete(comp->get_return_value()); | |
1518 | comp->release(); | |
1519 | } | |
1520 | ||
1521 | int64_t read_iterate(ImageCtx *ictx, uint64_t off, uint64_t len, | |
1522 | int (*cb)(uint64_t, size_t, const char *, void *), | |
1523 | void *arg) | |
1524 | { | |
11fdf7f2 TL |
1525 | coarse_mono_time start_time; |
1526 | ceph::timespan elapsed; | |
7c673cae FG |
1527 | |
1528 | ldout(ictx->cct, 20) << "read_iterate " << ictx << " off = " << off | |
1529 | << " len = " << len << dendl; | |
1530 | ||
1531 | int r = ictx->state->refresh_if_required(); | |
1532 | if (r < 0) | |
1533 | return r; | |
1534 | ||
1535 | uint64_t mylen = len; | |
9f95a23c | 1536 | ictx->image_lock.lock_shared(); |
7c673cae | 1537 | r = clip_io(ictx, off, &mylen); |
9f95a23c | 1538 | ictx->image_lock.unlock_shared(); |
7c673cae FG |
1539 | if (r < 0) |
1540 | return r; | |
1541 | ||
1542 | int64_t total_read = 0; | |
1543 | uint64_t period = ictx->get_stripe_period(); | |
1544 | uint64_t left = mylen; | |
1545 | ||
31f18b77 | 1546 | ZTracer::Trace trace; |
181888fb | 1547 | if (ictx->blkin_trace_all) { |
31f18b77 FG |
1548 | trace.init("read_iterate", &ictx->trace_endpoint); |
1549 | } | |
1550 | ||
9f95a23c | 1551 | std::shared_lock owner_locker{ictx->owner_lock}; |
11fdf7f2 | 1552 | start_time = coarse_mono_clock::now(); |
7c673cae FG |
1553 | while (left > 0) { |
1554 | uint64_t period_off = off - (off % period); | |
20effc67 | 1555 | uint64_t read_len = std::min(period_off + period - off, left); |
7c673cae FG |
1556 | |
1557 | bufferlist bl; | |
1558 | ||
1559 | C_SaferCond ctx; | |
1560 | auto c = io::AioCompletion::create_and_start(&ctx, ictx, | |
1561 | io::AIO_TYPE_READ); | |
f67539c2 TL |
1562 | auto req = io::ImageDispatchSpec::create_read( |
1563 | *ictx, io::IMAGE_DISPATCH_LAYER_NONE, c, | |
1564 | {{off, read_len}}, io::ReadResult{&bl}, | |
1565 | ictx->get_data_io_context(), 0, 0, trace); | |
1566 | req->send(); | |
7c673cae FG |
1567 | |
1568 | int ret = ctx.wait(); | |
1569 | if (ret < 0) { | |
1570 | return ret; | |
1571 | } | |
1572 | ||
1573 | r = cb(total_read, ret, bl.c_str(), arg); | |
1574 | if (r < 0) { | |
1575 | return r; | |
1576 | } | |
1577 | ||
1578 | total_read += ret; | |
1579 | left -= ret; | |
1580 | off += ret; | |
1581 | } | |
1582 | ||
11fdf7f2 | 1583 | elapsed = coarse_mono_clock::now() - start_time; |
7c673cae FG |
1584 | ictx->perfcounter->tinc(l_librbd_rd_latency, elapsed); |
1585 | ictx->perfcounter->inc(l_librbd_rd); | |
1586 | ictx->perfcounter->inc(l_librbd_rd_bytes, mylen); | |
1587 | return total_read; | |
1588 | } | |
1589 | ||
1590 | // validate extent against image size; clip to image size if necessary | |
1591 | int clip_io(ImageCtx *ictx, uint64_t off, uint64_t *len) | |
1592 | { | |
9f95a23c | 1593 | ceph_assert(ceph_mutex_is_locked(ictx->image_lock)); |
7c673cae | 1594 | |
f67539c2 TL |
1595 | if (ictx->snap_id != CEPH_NOSNAP && |
1596 | ictx->get_snap_info(ictx->snap_id) == nullptr) { | |
7c673cae | 1597 | return -ENOENT; |
f67539c2 TL |
1598 | } |
1599 | uint64_t image_size = ictx->get_effective_image_size(ictx->snap_id); | |
7c673cae FG |
1600 | |
1601 | // special-case "len == 0" requests: always valid | |
1602 | if (*len == 0) | |
1603 | return 0; | |
1604 | ||
1605 | // can't start past end | |
1606 | if (off >= image_size) | |
1607 | return -EINVAL; | |
1608 | ||
1609 | // clip requests that extend past end to just end | |
1610 | if ((off + *len) > image_size) | |
1611 | *len = (size_t)(image_size - off); | |
1612 | ||
1613 | return 0; | |
1614 | } | |
1615 | ||
11fdf7f2 | 1616 | int invalidate_cache(ImageCtx *ictx) |
7c673cae FG |
1617 | { |
1618 | CephContext *cct = ictx->cct; | |
11fdf7f2 | 1619 | ldout(cct, 20) << "invalidate_cache " << ictx << dendl; |
7c673cae FG |
1620 | |
1621 | int r = ictx->state->refresh_if_required(); | |
1622 | if (r < 0) { | |
1623 | return r; | |
1624 | } | |
1625 | ||
7c673cae FG |
1626 | C_SaferCond ctx; |
1627 | { | |
f67539c2 | 1628 | ictx->io_image_dispatcher->invalidate_cache(&ctx); |
7c673cae FG |
1629 | } |
1630 | r = ctx.wait(); | |
f67539c2 TL |
1631 | |
1632 | if (r < 0) { | |
1633 | ldout(cct, 20) << "failed to invalidate image cache" << dendl; | |
1634 | return r; | |
1635 | } | |
1636 | ||
7c673cae | 1637 | ictx->perfcounter->inc(l_librbd_invalidate_cache); |
f67539c2 TL |
1638 | |
1639 | // Delete writeback cache if it is not initialized | |
1640 | if ((!ictx->exclusive_lock || | |
1641 | !ictx->exclusive_lock->is_lock_owner()) && | |
1642 | ictx->test_features(RBD_FEATURE_DIRTY_CACHE)) { | |
1643 | C_SaferCond ctx3; | |
1644 | ictx->plugin_registry->discard(&ctx3); | |
1645 | r = ctx3.wait(); | |
1646 | } | |
7c673cae FG |
1647 | return r; |
1648 | } | |
1649 | ||
1650 | int poll_io_events(ImageCtx *ictx, io::AioCompletion **comps, int numcomp) | |
1651 | { | |
1652 | if (numcomp <= 0) | |
1653 | return -EINVAL; | |
1654 | CephContext *cct = ictx->cct; | |
1655 | ldout(cct, 20) << __func__ << " " << ictx << " numcomp = " << numcomp | |
1656 | << dendl; | |
1657 | int i = 0; | |
9f95a23c TL |
1658 | while (i < numcomp && ictx->event_socket_completions.pop(comps[i])) { |
1659 | ++i; | |
7c673cae | 1660 | } |
9f95a23c | 1661 | |
7c673cae FG |
1662 | return i; |
1663 | } | |
1664 | ||
1665 | int metadata_get(ImageCtx *ictx, const string &key, string *value) | |
1666 | { | |
1667 | CephContext *cct = ictx->cct; | |
1668 | ldout(cct, 20) << "metadata_get " << ictx << " key=" << key << dendl; | |
1669 | ||
1670 | int r = ictx->state->refresh_if_required(); | |
1671 | if (r < 0) { | |
1672 | return r; | |
1673 | } | |
1674 | ||
1675 | return cls_client::metadata_get(&ictx->md_ctx, ictx->header_oid, key, value); | |
1676 | } | |
1677 | ||
1678 | int metadata_list(ImageCtx *ictx, const string &start, uint64_t max, map<string, bufferlist> *pairs) | |
1679 | { | |
1680 | CephContext *cct = ictx->cct; | |
1681 | ldout(cct, 20) << "metadata_list " << ictx << dendl; | |
1682 | ||
1683 | int r = ictx->state->refresh_if_required(); | |
1684 | if (r < 0) { | |
1685 | return r; | |
1686 | } | |
1687 | ||
9f95a23c TL |
1688 | C_SaferCond ctx; |
1689 | auto req = image::GetMetadataRequest<>::create( | |
1690 | ictx->md_ctx, ictx->header_oid, false, "", start, max, pairs, &ctx); | |
1691 | req->send(); | |
7c673cae | 1692 | |
9f95a23c | 1693 | return ctx.wait(); |
7c673cae FG |
1694 | } |
1695 | ||
11fdf7f2 TL |
1696 | int list_watchers(ImageCtx *ictx, |
1697 | std::list<librbd::image_watcher_t> &watchers) | |
1698 | { | |
1699 | int r; | |
1700 | std::string header_oid; | |
1701 | std::list<obj_watch_t> obj_watchers; | |
1702 | ||
1703 | if (ictx->old_format) { | |
1704 | header_oid = util::old_header_name(ictx->name); | |
1705 | } else { | |
1706 | header_oid = util::header_name(ictx->id); | |
1707 | } | |
1708 | ||
1709 | r = ictx->md_ctx.list_watchers(header_oid, &obj_watchers); | |
1710 | if (r < 0) { | |
1711 | return r; | |
1712 | } | |
1713 | ||
1714 | for (auto i = obj_watchers.begin(); i != obj_watchers.end(); ++i) { | |
1715 | librbd::image_watcher_t watcher; | |
1716 | watcher.addr = i->addr; | |
1717 | watcher.id = i->watcher_id; | |
1718 | watcher.cookie = i->cookie; | |
1719 | ||
1720 | watchers.push_back(watcher); | |
1721 | } | |
1722 | ||
1723 | return 0; | |
1724 | } | |
1725 | ||
1726 | } | |
1727 | ||
1728 | std::ostream &operator<<(std::ostream &os, const librbd::ImageOptions &opts) { | |
1729 | os << "["; | |
1730 | ||
1731 | const char *delimiter = ""; | |
1732 | for (auto &i : librbd::IMAGE_OPTIONS_TYPE_MAPPING) { | |
1733 | if (i.second == librbd::STR) { | |
1734 | std::string val; | |
1735 | if (opts.get(i.first, &val) == 0) { | |
1736 | os << delimiter << librbd::image_option_name(i.first) << "=" << val; | |
1737 | delimiter = ", "; | |
1738 | } | |
1739 | } else if (i.second == librbd::UINT64) { | |
1740 | uint64_t val; | |
1741 | if (opts.get(i.first, &val) == 0) { | |
1742 | os << delimiter << librbd::image_option_name(i.first) << "=" << val; | |
1743 | delimiter = ", "; | |
1744 | } | |
1745 | } | |
1746 | } | |
7c673cae | 1747 | |
11fdf7f2 | 1748 | os << "]"; |
7c673cae | 1749 | |
11fdf7f2 | 1750 | return os; |
7c673cae | 1751 | } |