]>
Commit | Line | Data |
---|---|---|
7c673cae FG |
1 | // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*- |
2 | // vim: ts=8 sw=2 smarttab | |
3 | /* | |
4 | * Ceph ObjectStore engine | |
5 | * | |
6 | * IO engine using Ceph's ObjectStore class to test low-level performance of | |
7 | * Ceph OSDs. | |
8 | * | |
9 | */ | |
10 | ||
11 | #include <memory> | |
12 | #include <system_error> | |
13 | #include <vector> | |
14 | ||
15 | #include "os/ObjectStore.h" | |
16 | #include "global/global_init.h" | |
17 | #include "common/errno.h" | |
18 | #include "include/intarith.h" | |
19 | #include "include/stringify.h" | |
20 | ||
21 | #include <fio.h> | |
22 | #include <optgroup.h> | |
23 | ||
24 | #include "include/assert.h" // fio.h clobbers our assert.h | |
25 | ||
26 | #define dout_context g_ceph_context | |
27 | #define dout_subsys ceph_subsys_ | |
28 | ||
29 | namespace { | |
30 | ||
31 | /// fio configuration options read from the job file | |
32 | struct Options { | |
33 | thread_data* td; | |
34 | char* conf; | |
35 | }; | |
36 | ||
37 | template <class Func> // void Func(fio_option&) | |
38 | fio_option make_option(Func&& func) | |
39 | { | |
40 | // zero-initialize and set common defaults | |
41 | auto o = fio_option{}; | |
42 | o.category = FIO_OPT_C_ENGINE; | |
43 | o.group = FIO_OPT_G_RBD; | |
44 | func(std::ref(o)); | |
45 | return o; | |
46 | } | |
47 | ||
48 | static std::vector<fio_option> ceph_options{ | |
49 | make_option([] (fio_option& o) { | |
50 | o.name = "conf"; | |
51 | o.lname = "ceph configuration file"; | |
52 | o.type = FIO_OPT_STR_STORE; | |
53 | o.help = "Path to a ceph configuration file"; | |
54 | o.off1 = offsetof(Options, conf); | |
55 | }), | |
56 | {} // fio expects a 'null'-terminated list | |
57 | }; | |
58 | ||
59 | ||
60 | /// global engine state shared between all jobs within the process. this | |
61 | /// includes g_ceph_context and the ObjectStore instance | |
62 | struct Engine { | |
63 | /// the initial g_ceph_context reference to be dropped on destruction | |
64 | boost::intrusive_ptr<CephContext> cct; | |
65 | std::unique_ptr<ObjectStore> os; | |
66 | ||
67 | std::mutex lock; | |
68 | int ref_count; | |
69 | ||
70 | Engine(const thread_data* td); | |
71 | ~Engine(); | |
72 | ||
73 | static Engine* get_instance(thread_data* td) { | |
74 | // note: creates an Engine with the options associated with the first job | |
75 | static Engine engine(td); | |
76 | return &engine; | |
77 | } | |
78 | ||
79 | void ref() { | |
80 | std::lock_guard<std::mutex> l(lock); | |
81 | ++ref_count; | |
82 | } | |
83 | void deref() { | |
84 | std::lock_guard<std::mutex> l(lock); | |
85 | --ref_count; | |
86 | if (!ref_count) { | |
87 | ostringstream ostr; | |
88 | Formatter* f = Formatter::create("json-pretty", "json-pretty", "json-pretty"); | |
89 | os->dump_perf_counters(f); | |
31f18b77 | 90 | ostr << "FIO plugin "; |
7c673cae | 91 | f->flush(ostr); |
31f18b77 FG |
92 | if (g_conf->rocksdb_perf) { |
93 | os->get_db_statistics(f); | |
94 | ostr << "FIO get_db_statistics "; | |
95 | f->flush(ostr); | |
96 | } | |
7c673cae FG |
97 | delete f; |
98 | os->umount(); | |
31f18b77 | 99 | dout(0) << ostr.str() << dendl; |
7c673cae FG |
100 | } |
101 | } | |
102 | }; | |
103 | ||
104 | Engine::Engine(const thread_data* td) : ref_count(0) | |
105 | { | |
106 | // add the ceph command line arguments | |
107 | auto o = static_cast<const Options*>(td->eo); | |
108 | if (!o->conf) { | |
109 | throw std::runtime_error("missing conf option for ceph configuration file"); | |
110 | } | |
111 | std::vector<const char*> args{ | |
112 | "-i", "0", // identify as osd.0 for osd_data and osd_journal | |
113 | "--conf", o->conf, // use the requested conf file | |
114 | }; | |
115 | if (td->o.directory) { // allow conf files to use ${fio_dir} for data | |
116 | args.emplace_back("--fio_dir"); | |
117 | args.emplace_back(td->o.directory); | |
118 | } | |
119 | ||
120 | // claim the g_ceph_context reference and release it on destruction | |
121 | cct = global_init(nullptr, args, CEPH_ENTITY_TYPE_OSD, | |
122 | CODE_ENVIRONMENT_UTILITY, 0); | |
123 | common_init_finish(g_ceph_context); | |
124 | ||
125 | // create the ObjectStore | |
126 | os.reset(ObjectStore::create(g_ceph_context, | |
127 | g_conf->osd_objectstore, | |
128 | g_conf->osd_data, | |
129 | g_conf->osd_journal)); | |
130 | if (!os) | |
131 | throw std::runtime_error("bad objectstore type " + g_conf->osd_objectstore); | |
132 | ||
31f18b77 FG |
133 | unsigned num_shards; |
134 | if(g_conf->osd_op_num_shards) | |
135 | num_shards = g_conf->osd_op_num_shards; | |
136 | else if(os->is_rotational()) | |
137 | num_shards = g_conf->osd_op_num_shards_hdd; | |
138 | else | |
139 | num_shards = g_conf->osd_op_num_shards_ssd; | |
140 | os->set_cache_shards(num_shards); | |
7c673cae FG |
141 | |
142 | int r = os->mkfs(); | |
143 | if (r < 0) | |
144 | throw std::system_error(-r, std::system_category(), "mkfs failed"); | |
145 | ||
146 | r = os->mount(); | |
147 | if (r < 0) | |
148 | throw std::system_error(-r, std::system_category(), "mount failed"); | |
149 | } | |
150 | ||
151 | Engine::~Engine() | |
152 | { | |
153 | assert(!ref_count); | |
154 | } | |
155 | ||
156 | ||
157 | struct Collection { | |
158 | spg_t pg; | |
159 | coll_t cid; | |
160 | ObjectStore::Sequencer sequencer; | |
161 | ||
162 | // use big pool ids to avoid clashing with existing collections | |
163 | static constexpr int64_t MIN_POOL_ID = 0x0000ffffffffffff; | |
164 | ||
165 | Collection(const spg_t& pg) | |
31f18b77 FG |
166 | : pg(pg), cid(pg), sequencer(stringify(pg)) { |
167 | sequencer.shard_hint = pg; | |
168 | } | |
7c673cae FG |
169 | }; |
170 | ||
171 | struct Object { | |
172 | ghobject_t oid; | |
173 | Collection& coll; | |
174 | ||
175 | Object(const char* name, Collection& coll) | |
176 | : oid(hobject_t(name, "", CEPH_NOSNAP, coll.pg.ps(), coll.pg.pool(), "")), | |
177 | coll(coll) {} | |
178 | }; | |
179 | ||
180 | /// treat each fio job like a separate pool with its own collections and objects | |
181 | struct Job { | |
182 | Engine* engine; //< shared ptr to the global Engine | |
183 | std::vector<Collection> collections; //< spread objects over collections | |
184 | std::vector<Object> objects; //< associate an object with each fio_file | |
185 | std::vector<io_u*> events; //< completions for fio_ceph_os_event() | |
186 | const bool unlink; //< unlink objects on destruction | |
187 | ||
188 | Job(Engine* engine, const thread_data* td); | |
189 | ~Job(); | |
190 | }; | |
191 | ||
192 | Job::Job(Engine* engine, const thread_data* td) | |
193 | : engine(engine), | |
194 | events(td->o.iodepth), | |
195 | unlink(td->o.unlink) | |
196 | { | |
197 | engine->ref(); | |
198 | // use the fio thread_number for our unique pool id | |
199 | const uint64_t pool = Collection::MIN_POOL_ID + td->thread_number; | |
200 | ||
201 | // create a collection for each object, up to osd_pool_default_pg_num | |
202 | uint32_t count = g_conf->osd_pool_default_pg_num; | |
203 | if (count > td->o.nr_files) | |
204 | count = td->o.nr_files; | |
205 | ||
206 | assert(count > 0); | |
207 | collections.reserve(count); | |
208 | ||
209 | const int split_bits = cbits(count - 1); | |
210 | ||
211 | ObjectStore::Transaction t; | |
212 | for (uint32_t i = 0; i < count; i++) { | |
213 | auto pg = spg_t{pg_t{i, pool}}; | |
214 | collections.emplace_back(pg); | |
215 | ||
216 | auto& cid = collections.back().cid; | |
217 | if (!engine->os->collection_exists(cid)) | |
218 | t.create_collection(cid, split_bits); | |
219 | } | |
220 | ||
221 | const uint64_t file_size = td->o.size / max(1u, td->o.nr_files); | |
222 | ||
223 | // create an object for each file in the job | |
224 | for (uint32_t i = 0; i < td->o.nr_files; i++) { | |
225 | auto f = td->files[i]; | |
226 | f->real_file_size = file_size; | |
227 | f->engine_pos = i; | |
228 | ||
229 | // associate each object with a collection in a round-robin fashion | |
230 | auto& coll = collections[i % collections.size()]; | |
231 | ||
232 | objects.emplace_back(f->file_name, coll); | |
233 | auto& oid = objects.back().oid; | |
234 | ||
235 | t.touch(coll.cid, oid); | |
236 | t.truncate(coll.cid, oid, file_size); | |
237 | } | |
238 | ||
239 | // apply the entire transaction synchronously | |
240 | ObjectStore::Sequencer sequencer("job init"); | |
241 | int r = engine->os->apply_transaction(&sequencer, std::move(t)); | |
242 | if (r) { | |
243 | engine->deref(); | |
244 | throw std::system_error(r, std::system_category(), "job init"); | |
245 | } | |
246 | } | |
247 | ||
248 | Job::~Job() | |
249 | { | |
250 | if (unlink) { | |
251 | ObjectStore::Transaction t; | |
252 | // remove our objects | |
253 | for (auto& obj : objects) { | |
254 | t.remove(obj.coll.cid, obj.oid); | |
255 | } | |
256 | // remove our collections | |
257 | for (auto& coll : collections) { | |
258 | t.remove_collection(coll.cid); | |
259 | } | |
260 | ObjectStore::Sequencer sequencer("job cleanup"); | |
261 | int r = engine->os->apply_transaction(&sequencer, std::move(t)); | |
262 | if (r) | |
263 | derr << "job cleanup failed with " << cpp_strerror(-r) << dendl; | |
264 | } | |
265 | engine->deref(); | |
266 | } | |
267 | ||
268 | ||
269 | int fio_ceph_os_setup(thread_data* td) | |
270 | { | |
271 | // if there are multiple jobs, they must run in the same process against a | |
272 | // single instance of the ObjectStore. explicitly disable fio's default | |
273 | // job-per-process configuration | |
274 | td->o.use_thread = 1; | |
275 | ||
276 | try { | |
277 | // get or create the global Engine instance | |
278 | auto engine = Engine::get_instance(td); | |
279 | // create a Job for this thread | |
280 | td->io_ops_data = new Job(engine, td); | |
281 | } catch (std::exception& e) { | |
282 | std::cerr << "setup failed with " << e.what() << std::endl; | |
283 | return -1; | |
284 | } | |
285 | return 0; | |
286 | } | |
287 | ||
288 | void fio_ceph_os_cleanup(thread_data* td) | |
289 | { | |
290 | auto job = static_cast<Job*>(td->io_ops_data); | |
291 | td->io_ops_data = nullptr; | |
292 | delete job; | |
293 | } | |
294 | ||
295 | ||
296 | io_u* fio_ceph_os_event(thread_data* td, int event) | |
297 | { | |
298 | // return the requested event from fio_ceph_os_getevents() | |
299 | auto job = static_cast<Job*>(td->io_ops_data); | |
300 | return job->events[event]; | |
301 | } | |
302 | ||
303 | int fio_ceph_os_getevents(thread_data* td, unsigned int min, | |
304 | unsigned int max, const timespec* t) | |
305 | { | |
306 | auto job = static_cast<Job*>(td->io_ops_data); | |
307 | unsigned int events = 0; | |
308 | io_u* u; | |
309 | unsigned int i; | |
310 | ||
311 | // loop through inflight ios until we find 'min' completions | |
312 | do { | |
313 | io_u_qiter(&td->io_u_all, u, i) { | |
314 | if (!(u->flags & IO_U_F_FLIGHT)) | |
315 | continue; | |
316 | ||
317 | if (u->engine_data) { | |
318 | u->engine_data = nullptr; | |
319 | job->events[events] = u; | |
320 | events++; | |
321 | } | |
322 | } | |
323 | if (events >= min) | |
324 | break; | |
325 | usleep(100); | |
326 | } while (1); | |
327 | ||
328 | return events; | |
329 | } | |
330 | ||
331 | /// completion context for ObjectStore::queue_transaction() | |
332 | class UnitComplete : public Context { | |
333 | io_u* u; | |
334 | public: | |
335 | UnitComplete(io_u* u) : u(u) {} | |
336 | void finish(int r) { | |
337 | // mark the pointer to indicate completion for fio_ceph_os_getevents() | |
338 | u->engine_data = reinterpret_cast<void*>(1ull); | |
339 | } | |
340 | }; | |
341 | ||
342 | int fio_ceph_os_queue(thread_data* td, io_u* u) | |
343 | { | |
344 | fio_ro_check(td, u); | |
345 | ||
346 | auto job = static_cast<Job*>(td->io_ops_data); | |
347 | auto& object = job->objects[u->file->engine_pos]; | |
348 | auto& coll = object.coll; | |
349 | auto& os = job->engine->os; | |
350 | ||
351 | if (u->ddir == DDIR_WRITE) { | |
352 | // provide a hint if we're likely to read this data back | |
353 | const int flags = td_rw(td) ? CEPH_OSD_OP_FLAG_FADVISE_WILLNEED : 0; | |
354 | ||
355 | bufferlist bl; | |
356 | bl.push_back(buffer::copy(reinterpret_cast<char*>(u->xfer_buf), | |
357 | u->xfer_buflen ) ); | |
358 | ||
359 | // enqueue a write transaction on the collection's sequencer | |
360 | ObjectStore::Transaction t; | |
361 | t.write(coll.cid, object.oid, u->offset, u->xfer_buflen, bl, flags); | |
362 | os->queue_transaction(&coll.sequencer, | |
363 | std::move(t), | |
364 | nullptr, | |
365 | new UnitComplete(u)); | |
366 | return FIO_Q_QUEUED; | |
367 | } | |
368 | ||
369 | if (u->ddir == DDIR_READ) { | |
370 | // ObjectStore reads are synchronous, so make the call and return COMPLETED | |
371 | bufferlist bl; | |
372 | int r = os->read(coll.cid, object.oid, u->offset, u->xfer_buflen, bl); | |
373 | if (r < 0) { | |
374 | u->error = r; | |
375 | td_verror(td, u->error, "xfer"); | |
376 | } else { | |
377 | bl.copy(0, bl.length(), static_cast<char*>(u->xfer_buf)); | |
378 | u->resid = u->xfer_buflen - r; | |
379 | } | |
380 | return FIO_Q_COMPLETED; | |
381 | } | |
382 | ||
383 | derr << "WARNING: Only DDIR_READ and DDIR_WRITE are supported!" << dendl; | |
384 | u->error = -EINVAL; | |
385 | td_verror(td, u->error, "xfer"); | |
386 | return FIO_Q_COMPLETED; | |
387 | } | |
388 | ||
389 | int fio_ceph_os_commit(thread_data* td) | |
390 | { | |
391 | // commit() allows the engine to batch up queued requests to be submitted all | |
392 | // at once. it would be natural for queue() to collect transactions in a list, | |
393 | // and use commit() to pass them all to ObjectStore::queue_transactions(). but | |
394 | // because we spread objects over multiple collections, we a) need to use a | |
395 | // different sequencer for each collection, and b) are less likely to see a | |
396 | // benefit from batching requests within a collection | |
397 | return 0; | |
398 | } | |
399 | ||
400 | // open/close are noops. we set the FIO_DISKLESSIO flag in ioengine_ops to | |
401 | // prevent fio from creating the files | |
402 | int fio_ceph_os_open(thread_data* td, fio_file* f) { return 0; } | |
403 | int fio_ceph_os_close(thread_data* td, fio_file* f) { return 0; } | |
404 | ||
405 | int fio_ceph_os_io_u_init(thread_data* td, io_u* u) | |
406 | { | |
407 | // no data is allocated, we just use the pointer as a boolean 'completed' flag | |
408 | u->engine_data = nullptr; | |
409 | return 0; | |
410 | } | |
411 | ||
412 | void fio_ceph_os_io_u_free(thread_data* td, io_u* u) | |
413 | { | |
414 | u->engine_data = nullptr; | |
415 | } | |
416 | ||
417 | ||
418 | // ioengine_ops for get_ioengine() | |
419 | struct ceph_ioengine : public ioengine_ops { | |
420 | ceph_ioengine() : ioengine_ops({}) { | |
421 | name = "ceph-os"; | |
422 | version = FIO_IOOPS_VERSION; | |
423 | flags = FIO_DISKLESSIO; | |
424 | setup = fio_ceph_os_setup; | |
425 | queue = fio_ceph_os_queue; | |
426 | commit = fio_ceph_os_commit; | |
427 | getevents = fio_ceph_os_getevents; | |
428 | event = fio_ceph_os_event; | |
429 | cleanup = fio_ceph_os_cleanup; | |
430 | open_file = fio_ceph_os_open; | |
431 | close_file = fio_ceph_os_close; | |
432 | io_u_init = fio_ceph_os_io_u_init; | |
433 | io_u_free = fio_ceph_os_io_u_free; | |
434 | options = ceph_options.data(); | |
435 | option_struct_size = sizeof(struct Options); | |
436 | } | |
437 | }; | |
438 | ||
439 | } // anonymous namespace | |
440 | ||
441 | extern "C" { | |
442 | // the exported fio engine interface | |
443 | void get_ioengine(struct ioengine_ops** ioengine_ptr) { | |
444 | static ceph_ioengine ioengine; | |
445 | *ioengine_ptr = &ioengine; | |
446 | } | |
447 | } // extern "C" |