]>
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 | /* | |
5 | * rbd-nbd - RBD in userspace | |
6 | * | |
7 | * Copyright (C) 2015 - 2016 Kylin Corporation | |
8 | * | |
9 | * Author: Yunchuan Wen <yunchuan.wen@kylin-cloud.com> | |
10 | * Li Wang <li.wang@kylin-cloud.com> | |
11 | * | |
12 | * This is free software; you can redistribute it and/or | |
13 | * modify it under the terms of the GNU Lesser General Public | |
14 | * License version 2.1, as published by the Free Software | |
15 | * Foundation. See file COPYING. | |
16 | * | |
17 | */ | |
18 | ||
19 | #include "include/int_types.h" | |
20 | ||
21 | #include <stdio.h> | |
22 | #include <stdlib.h> | |
23 | #include <stddef.h> | |
24 | #include <errno.h> | |
25 | #include <fcntl.h> | |
26 | #include <string.h> | |
27 | #include <sys/types.h> | |
28 | #include <unistd.h> | |
29 | ||
30 | #include <linux/nbd.h> | |
31 | #include <linux/fs.h> | |
32 | #include <sys/ioctl.h> | |
33 | #include <sys/socket.h> | |
34 | ||
eafe8130 TL |
35 | #include "nbd-netlink.h" |
36 | #include <libnl3/netlink/genl/genl.h> | |
37 | #include <libnl3/netlink/genl/ctrl.h> | |
38 | #include <libnl3/netlink/genl/mngt.h> | |
39 | ||
7c673cae | 40 | #include <fstream> |
11fdf7f2 TL |
41 | #include <iostream> |
42 | #include <memory> | |
43 | #include <regex> | |
44 | #include <boost/algorithm/string/predicate.hpp> | |
7c673cae | 45 | |
11fdf7f2 TL |
46 | #include "common/Formatter.h" |
47 | #include "common/Preforker.h" | |
48 | #include "common/TextTable.h" | |
49 | #include "common/ceph_argparse.h" | |
7c673cae FG |
50 | #include "common/config.h" |
51 | #include "common/dout.h" | |
7c673cae FG |
52 | #include "common/errno.h" |
53 | #include "common/module.h" | |
54 | #include "common/safe_io.h" | |
224ce89b | 55 | #include "common/version.h" |
11fdf7f2 | 56 | |
7c673cae FG |
57 | #include "global/global_init.h" |
58 | #include "global/signal_handler.h" | |
59 | ||
60 | #include "include/rados/librados.hpp" | |
61 | #include "include/rbd/librbd.hpp" | |
62 | #include "include/stringify.h" | |
63 | #include "include/xlist.h" | |
64 | ||
11fdf7f2 TL |
65 | #include "mon/MonClient.h" |
66 | ||
7c673cae FG |
67 | #define dout_context g_ceph_context |
68 | #define dout_subsys ceph_subsys_rbd | |
69 | #undef dout_prefix | |
70 | #define dout_prefix *_dout << "rbd-nbd: " | |
71 | ||
31f18b77 FG |
72 | struct Config { |
73 | int nbds_max = 0; | |
74 | int max_part = 255; | |
11fdf7f2 | 75 | int timeout = -1; |
31f18b77 FG |
76 | |
77 | bool exclusive = false; | |
78 | bool readonly = false; | |
79 | bool set_max_part = false; | |
eafe8130 | 80 | bool try_netlink = false; |
31f18b77 FG |
81 | |
82 | std::string poolname; | |
11fdf7f2 | 83 | std::string nsname; |
31f18b77 FG |
84 | std::string imgname; |
85 | std::string snapname; | |
86 | std::string devpath; | |
11fdf7f2 TL |
87 | |
88 | std::string format; | |
89 | bool pretty_format = false; | |
31f18b77 FG |
90 | }; |
91 | ||
7c673cae FG |
92 | static void usage() |
93 | { | |
94 | std::cout << "Usage: rbd-nbd [options] map <image-or-snap-spec> Map an image to nbd device\n" | |
11fdf7f2 TL |
95 | << " unmap <device|image-or-snap-spec> Unmap nbd device\n" |
96 | << " [options] list-mapped List mapped nbd devices\n" | |
97 | << "Map options:\n" | |
494da23a | 98 | << " --device <device path> Specify nbd device path (/dev/nbd{num})\n" |
7c673cae FG |
99 | << " --read-only Map read-only\n" |
100 | << " --nbds_max <limit> Override for module param nbds_max\n" | |
101 | << " --max_part <limit> Override for module param max_part\n" | |
102 | << " --exclusive Forbid writes by other clients\n" | |
11fdf7f2 | 103 | << " --timeout <seconds> Set nbd request timeout\n" |
eafe8130 | 104 | << " --try-netlink Use the nbd netlink interface\n" |
11fdf7f2 TL |
105 | << "\n" |
106 | << "List options:\n" | |
107 | << " --format plain|json|xml Output format (default: plain)\n" | |
108 | << " --pretty-format Pretty formatting (json and xml)\n" | |
7c673cae FG |
109 | << std::endl; |
110 | generic_server_usage(); | |
111 | } | |
112 | ||
7c673cae | 113 | static int nbd = -1; |
eafe8130 | 114 | static int nbd_index = -1; |
7c673cae | 115 | |
11fdf7f2 | 116 | enum Command { |
31f18b77 FG |
117 | None, |
118 | Connect, | |
119 | Disconnect, | |
120 | List | |
11fdf7f2 TL |
121 | }; |
122 | ||
123 | static Command cmd = None; | |
31f18b77 | 124 | |
7c673cae FG |
125 | #define RBD_NBD_BLKSIZE 512UL |
126 | ||
224ce89b WB |
127 | #define HELP_INFO 1 |
128 | #define VERSION_INFO 2 | |
129 | ||
7c673cae FG |
130 | #ifdef CEPH_BIG_ENDIAN |
131 | #define ntohll(a) (a) | |
132 | #elif defined(CEPH_LITTLE_ENDIAN) | |
133 | #define ntohll(a) swab(a) | |
134 | #else | |
135 | #error "Could not determine endianess" | |
136 | #endif | |
137 | #define htonll(a) ntohll(a) | |
138 | ||
11fdf7f2 TL |
139 | static int parse_args(vector<const char*>& args, std::ostream *err_msg, |
140 | Command *command, Config *cfg); | |
eafe8130 | 141 | static int netlink_resize(int nbd_index, uint64_t size); |
7c673cae FG |
142 | |
143 | class NBDServer | |
144 | { | |
145 | private: | |
146 | int fd; | |
147 | librbd::Image ℑ | |
148 | ||
149 | public: | |
150 | NBDServer(int _fd, librbd::Image& _image) | |
151 | : fd(_fd) | |
152 | , image(_image) | |
eafe8130 | 153 | , disconnect_lock("NBDServer::DisconnectLocker") |
7c673cae FG |
154 | , lock("NBDServer::Locker") |
155 | , reader_thread(*this, &NBDServer::reader_entry) | |
156 | , writer_thread(*this, &NBDServer::writer_entry) | |
157 | , started(false) | |
158 | {} | |
159 | ||
160 | private: | |
eafe8130 TL |
161 | Mutex disconnect_lock; |
162 | Cond disconnect_cond; | |
7c673cae FG |
163 | std::atomic<bool> terminated = { false }; |
164 | ||
165 | void shutdown() | |
166 | { | |
167 | bool expected = false; | |
168 | if (terminated.compare_exchange_strong(expected, true)) { | |
169 | ::shutdown(fd, SHUT_RDWR); | |
170 | ||
171 | Mutex::Locker l(lock); | |
172 | cond.Signal(); | |
173 | } | |
174 | } | |
175 | ||
176 | struct IOContext | |
177 | { | |
178 | xlist<IOContext*>::item item; | |
11fdf7f2 | 179 | NBDServer *server = nullptr; |
7c673cae FG |
180 | struct nbd_request request; |
181 | struct nbd_reply reply; | |
182 | bufferlist data; | |
11fdf7f2 | 183 | int command = 0; |
7c673cae FG |
184 | |
185 | IOContext() | |
186 | : item(this) | |
187 | {} | |
188 | }; | |
189 | ||
190 | friend std::ostream &operator<<(std::ostream &os, const IOContext &ctx); | |
191 | ||
192 | Mutex lock; | |
193 | Cond cond; | |
194 | xlist<IOContext*> io_pending; | |
195 | xlist<IOContext*> io_finished; | |
196 | ||
197 | void io_start(IOContext *ctx) | |
198 | { | |
199 | Mutex::Locker l(lock); | |
200 | io_pending.push_back(&ctx->item); | |
201 | } | |
202 | ||
203 | void io_finish(IOContext *ctx) | |
204 | { | |
205 | Mutex::Locker l(lock); | |
11fdf7f2 | 206 | ceph_assert(ctx->item.is_on_list()); |
7c673cae FG |
207 | ctx->item.remove_myself(); |
208 | io_finished.push_back(&ctx->item); | |
209 | cond.Signal(); | |
210 | } | |
211 | ||
212 | IOContext *wait_io_finish() | |
213 | { | |
214 | Mutex::Locker l(lock); | |
215 | while(io_finished.empty() && !terminated) | |
216 | cond.Wait(lock); | |
217 | ||
218 | if (io_finished.empty()) | |
219 | return NULL; | |
220 | ||
221 | IOContext *ret = io_finished.front(); | |
222 | io_finished.pop_front(); | |
223 | ||
224 | return ret; | |
225 | } | |
226 | ||
227 | void wait_clean() | |
228 | { | |
11fdf7f2 | 229 | ceph_assert(!reader_thread.is_started()); |
7c673cae FG |
230 | Mutex::Locker l(lock); |
231 | while(!io_pending.empty()) | |
232 | cond.Wait(lock); | |
233 | ||
234 | while(!io_finished.empty()) { | |
11fdf7f2 | 235 | std::unique_ptr<IOContext> free_ctx(io_finished.front()); |
7c673cae FG |
236 | io_finished.pop_front(); |
237 | } | |
238 | } | |
239 | ||
240 | static void aio_callback(librbd::completion_t cb, void *arg) | |
241 | { | |
242 | librbd::RBD::AioCompletion *aio_completion = | |
243 | reinterpret_cast<librbd::RBD::AioCompletion*>(cb); | |
244 | ||
245 | IOContext *ctx = reinterpret_cast<IOContext *>(arg); | |
246 | int ret = aio_completion->get_return_value(); | |
247 | ||
248 | dout(20) << __func__ << ": " << *ctx << dendl; | |
249 | ||
250 | if (ret == -EINVAL) { | |
251 | // if shrinking an image, a pagecache writeback might reference | |
252 | // extents outside of the range of the new image extents | |
181888fb | 253 | dout(0) << __func__ << ": masking IO out-of-bounds error" << dendl; |
7c673cae FG |
254 | ctx->data.clear(); |
255 | ret = 0; | |
256 | } | |
257 | ||
258 | if (ret < 0) { | |
259 | ctx->reply.error = htonl(-ret); | |
260 | } else if ((ctx->command == NBD_CMD_READ) && | |
261 | ret < static_cast<int>(ctx->request.len)) { | |
262 | int pad_byte_count = static_cast<int> (ctx->request.len) - ret; | |
263 | ctx->data.append_zero(pad_byte_count); | |
264 | dout(20) << __func__ << ": " << *ctx << ": Pad byte count: " | |
265 | << pad_byte_count << dendl; | |
11fdf7f2 | 266 | ctx->reply.error = htonl(0); |
7c673cae FG |
267 | } else { |
268 | ctx->reply.error = htonl(0); | |
269 | } | |
270 | ctx->server->io_finish(ctx); | |
271 | ||
272 | aio_completion->release(); | |
273 | } | |
274 | ||
275 | void reader_entry() | |
276 | { | |
277 | while (!terminated) { | |
11fdf7f2 | 278 | std::unique_ptr<IOContext> ctx(new IOContext()); |
7c673cae FG |
279 | ctx->server = this; |
280 | ||
281 | dout(20) << __func__ << ": waiting for nbd request" << dendl; | |
282 | ||
283 | int r = safe_read_exact(fd, &ctx->request, sizeof(struct nbd_request)); | |
284 | if (r < 0) { | |
285 | derr << "failed to read nbd request header: " << cpp_strerror(r) | |
286 | << dendl; | |
eafe8130 | 287 | goto signal; |
7c673cae FG |
288 | } |
289 | ||
290 | if (ctx->request.magic != htonl(NBD_REQUEST_MAGIC)) { | |
291 | derr << "invalid nbd request header" << dendl; | |
eafe8130 | 292 | goto signal; |
7c673cae FG |
293 | } |
294 | ||
295 | ctx->request.from = ntohll(ctx->request.from); | |
296 | ctx->request.type = ntohl(ctx->request.type); | |
297 | ctx->request.len = ntohl(ctx->request.len); | |
298 | ||
299 | ctx->reply.magic = htonl(NBD_REPLY_MAGIC); | |
300 | memcpy(ctx->reply.handle, ctx->request.handle, sizeof(ctx->reply.handle)); | |
301 | ||
302 | ctx->command = ctx->request.type & 0x0000ffff; | |
303 | ||
304 | dout(20) << *ctx << ": start" << dendl; | |
305 | ||
306 | switch (ctx->command) | |
307 | { | |
308 | case NBD_CMD_DISC: | |
309 | // NBD_DO_IT will return when pipe is closed | |
310 | dout(0) << "disconnect request received" << dendl; | |
eafe8130 | 311 | goto signal; |
7c673cae FG |
312 | case NBD_CMD_WRITE: |
313 | bufferptr ptr(ctx->request.len); | |
314 | r = safe_read_exact(fd, ptr.c_str(), ctx->request.len); | |
315 | if (r < 0) { | |
316 | derr << *ctx << ": failed to read nbd request data: " | |
317 | << cpp_strerror(r) << dendl; | |
eafe8130 | 318 | goto signal; |
7c673cae FG |
319 | } |
320 | ctx->data.push_back(ptr); | |
321 | break; | |
322 | } | |
323 | ||
324 | IOContext *pctx = ctx.release(); | |
325 | io_start(pctx); | |
326 | librbd::RBD::AioCompletion *c = new librbd::RBD::AioCompletion(pctx, aio_callback); | |
327 | switch (pctx->command) | |
328 | { | |
329 | case NBD_CMD_WRITE: | |
330 | image.aio_write(pctx->request.from, pctx->request.len, pctx->data, c); | |
331 | break; | |
332 | case NBD_CMD_READ: | |
333 | image.aio_read(pctx->request.from, pctx->request.len, pctx->data, c); | |
334 | break; | |
335 | case NBD_CMD_FLUSH: | |
336 | image.aio_flush(c); | |
337 | break; | |
338 | case NBD_CMD_TRIM: | |
339 | image.aio_discard(pctx->request.from, pctx->request.len, c); | |
340 | break; | |
341 | default: | |
342 | derr << *pctx << ": invalid request command" << dendl; | |
343 | c->release(); | |
eafe8130 | 344 | goto signal; |
7c673cae FG |
345 | } |
346 | } | |
347 | dout(20) << __func__ << ": terminated" << dendl; | |
eafe8130 TL |
348 | |
349 | signal: | |
350 | Mutex::Locker l(disconnect_lock); | |
351 | disconnect_cond.Signal(); | |
7c673cae FG |
352 | } |
353 | ||
354 | void writer_entry() | |
355 | { | |
356 | while (!terminated) { | |
357 | dout(20) << __func__ << ": waiting for io request" << dendl; | |
11fdf7f2 | 358 | std::unique_ptr<IOContext> ctx(wait_io_finish()); |
7c673cae FG |
359 | if (!ctx) { |
360 | dout(20) << __func__ << ": no io requests, terminating" << dendl; | |
361 | return; | |
362 | } | |
363 | ||
364 | dout(20) << __func__ << ": got: " << *ctx << dendl; | |
365 | ||
366 | int r = safe_write(fd, &ctx->reply, sizeof(struct nbd_reply)); | |
367 | if (r < 0) { | |
368 | derr << *ctx << ": failed to write reply header: " << cpp_strerror(r) | |
369 | << dendl; | |
370 | return; | |
371 | } | |
372 | if (ctx->command == NBD_CMD_READ && ctx->reply.error == htonl(0)) { | |
373 | r = ctx->data.write_fd(fd); | |
374 | if (r < 0) { | |
375 | derr << *ctx << ": failed to write replay data: " << cpp_strerror(r) | |
376 | << dendl; | |
377 | return; | |
378 | } | |
379 | } | |
380 | dout(20) << *ctx << ": finish" << dendl; | |
381 | } | |
382 | dout(20) << __func__ << ": terminated" << dendl; | |
383 | } | |
384 | ||
385 | class ThreadHelper : public Thread | |
386 | { | |
387 | public: | |
388 | typedef void (NBDServer::*entry_func)(); | |
389 | private: | |
390 | NBDServer &server; | |
391 | entry_func func; | |
392 | public: | |
393 | ThreadHelper(NBDServer &_server, entry_func _func) | |
394 | :server(_server) | |
395 | ,func(_func) | |
396 | {} | |
397 | protected: | |
398 | void* entry() override | |
399 | { | |
400 | (server.*func)(); | |
401 | server.shutdown(); | |
402 | return NULL; | |
403 | } | |
404 | } reader_thread, writer_thread; | |
405 | ||
406 | bool started; | |
407 | public: | |
408 | void start() | |
409 | { | |
410 | if (!started) { | |
411 | dout(10) << __func__ << ": starting" << dendl; | |
412 | ||
413 | started = true; | |
414 | ||
415 | reader_thread.create("rbd_reader"); | |
416 | writer_thread.create("rbd_writer"); | |
417 | } | |
418 | } | |
419 | ||
eafe8130 TL |
420 | void wait_for_disconnect() |
421 | { | |
422 | if (!started) | |
423 | return; | |
424 | ||
425 | Mutex::Locker l(disconnect_lock); | |
426 | disconnect_cond.Wait(disconnect_lock); | |
427 | } | |
428 | ||
11fdf7f2 | 429 | ~NBDServer() |
7c673cae FG |
430 | { |
431 | if (started) { | |
432 | dout(10) << __func__ << ": terminating" << dendl; | |
433 | ||
434 | shutdown(); | |
435 | ||
436 | reader_thread.join(); | |
437 | writer_thread.join(); | |
438 | ||
439 | wait_clean(); | |
440 | ||
441 | started = false; | |
442 | } | |
443 | } | |
7c673cae FG |
444 | }; |
445 | ||
446 | std::ostream &operator<<(std::ostream &os, const NBDServer::IOContext &ctx) { | |
447 | ||
448 | os << "[" << std::hex << ntohll(*((uint64_t *)ctx.request.handle)); | |
449 | ||
450 | switch (ctx.command) | |
451 | { | |
452 | case NBD_CMD_WRITE: | |
453 | os << " WRITE "; | |
454 | break; | |
455 | case NBD_CMD_READ: | |
456 | os << " READ "; | |
457 | break; | |
458 | case NBD_CMD_FLUSH: | |
459 | os << " FLUSH "; | |
460 | break; | |
461 | case NBD_CMD_TRIM: | |
462 | os << " TRIM "; | |
463 | break; | |
464 | default: | |
11fdf7f2 | 465 | os << " UNKNOWN(" << ctx.command << ") "; |
7c673cae FG |
466 | break; |
467 | } | |
468 | ||
469 | os << ctx.request.from << "~" << ctx.request.len << " " | |
11fdf7f2 | 470 | << std::dec << ntohl(ctx.reply.error) << "]"; |
7c673cae FG |
471 | |
472 | return os; | |
473 | } | |
474 | ||
475 | class NBDWatchCtx : public librbd::UpdateWatchCtx | |
476 | { | |
477 | private: | |
478 | int fd; | |
eafe8130 TL |
479 | int nbd_index; |
480 | bool use_netlink; | |
7c673cae FG |
481 | librados::IoCtx &io_ctx; |
482 | librbd::Image ℑ | |
483 | unsigned long size; | |
484 | public: | |
485 | NBDWatchCtx(int _fd, | |
eafe8130 TL |
486 | int _nbd_index, |
487 | bool _use_netlink, | |
7c673cae FG |
488 | librados::IoCtx &_io_ctx, |
489 | librbd::Image &_image, | |
490 | unsigned long _size) | |
491 | : fd(_fd) | |
eafe8130 TL |
492 | , nbd_index(_nbd_index) |
493 | , use_netlink(_use_netlink) | |
7c673cae FG |
494 | , io_ctx(_io_ctx) |
495 | , image(_image) | |
496 | , size(_size) | |
497 | { } | |
498 | ||
499 | ~NBDWatchCtx() override {} | |
500 | ||
501 | void handle_notify() override | |
502 | { | |
503 | librbd::image_info_t info; | |
504 | if (image.stat(info, sizeof(info)) == 0) { | |
505 | unsigned long new_size = info.size; | |
eafe8130 | 506 | int ret; |
7c673cae FG |
507 | |
508 | if (new_size != size) { | |
b32b8144 | 509 | dout(5) << "resize detected" << dendl; |
7c673cae | 510 | if (ioctl(fd, BLKFLSBUF, NULL) < 0) |
eafe8130 TL |
511 | derr << "invalidate page cache failed: " << cpp_strerror(errno) |
512 | << dendl; | |
513 | if (use_netlink) { | |
514 | ret = netlink_resize(nbd_index, new_size); | |
515 | } else { | |
516 | ret = ioctl(fd, NBD_SET_SIZE, new_size); | |
517 | if (ret < 0) | |
7c673cae | 518 | derr << "resize failed: " << cpp_strerror(errno) << dendl; |
eafe8130 TL |
519 | } |
520 | ||
521 | if (!ret) | |
7c673cae | 522 | size = new_size; |
eafe8130 | 523 | |
b32b8144 FG |
524 | if (ioctl(fd, BLKRRPART, NULL) < 0) { |
525 | derr << "rescan of partition table failed: " << cpp_strerror(errno) | |
526 | << dendl; | |
527 | } | |
7c673cae | 528 | if (image.invalidate_cache() < 0) |
eafe8130 | 529 | derr << "invalidate rbd cache failed" << dendl; |
7c673cae FG |
530 | } |
531 | } | |
532 | } | |
533 | }; | |
534 | ||
11fdf7f2 TL |
535 | class NBDListIterator { |
536 | public: | |
537 | bool get(int *pid, Config *cfg) { | |
538 | while (true) { | |
539 | std::string nbd_path = "/sys/block/nbd" + stringify(m_index); | |
540 | if(access(nbd_path.c_str(), F_OK) != 0) { | |
541 | return false; | |
542 | } | |
543 | ||
544 | *cfg = Config(); | |
545 | cfg->devpath = "/dev/nbd" + stringify(m_index++); | |
546 | ||
547 | std::ifstream ifs; | |
548 | ifs.open(nbd_path + "/pid", std::ifstream::in); | |
549 | if (!ifs.is_open()) { | |
550 | continue; | |
551 | } | |
552 | ifs >> *pid; | |
553 | ||
554 | int r = get_mapped_info(*pid, cfg); | |
555 | if (r < 0) { | |
556 | continue; | |
557 | } | |
558 | ||
559 | return true; | |
560 | } | |
561 | } | |
562 | ||
563 | private: | |
564 | int m_index = 0; | |
565 | ||
566 | int get_mapped_info(int pid, Config *cfg) { | |
567 | int r; | |
568 | std::string path = "/proc/" + stringify(pid) + "/cmdline"; | |
569 | std::ifstream ifs; | |
570 | std::string cmdline; | |
571 | std::vector<const char*> args; | |
572 | ||
573 | ifs.open(path.c_str(), std::ifstream::in); | |
574 | if (!ifs.is_open()) | |
575 | return -1; | |
576 | ifs >> cmdline; | |
577 | ||
578 | for (unsigned i = 0; i < cmdline.size(); i++) { | |
579 | const char *arg = &cmdline[i]; | |
580 | if (i == 0) { | |
581 | if (strcmp(basename(arg) , "rbd-nbd") != 0) { | |
582 | return -EINVAL; | |
583 | } | |
584 | } else { | |
585 | args.push_back(arg); | |
586 | } | |
587 | ||
588 | while (cmdline[i] != '\0') { | |
589 | i++; | |
590 | } | |
591 | } | |
592 | ||
593 | std::ostringstream err_msg; | |
594 | Command command; | |
595 | r = parse_args(args, &err_msg, &command, cfg); | |
596 | if (r < 0) { | |
597 | return r; | |
598 | } | |
599 | ||
600 | if (command != Connect) { | |
601 | return -ENOENT; | |
602 | } | |
603 | ||
604 | return 0; | |
605 | } | |
606 | }; | |
607 | ||
eafe8130 | 608 | static int load_module(Config *cfg) |
7c673cae | 609 | { |
eafe8130 TL |
610 | ostringstream param; |
611 | int ret; | |
11fdf7f2 | 612 | |
eafe8130 TL |
613 | if (cfg->nbds_max) |
614 | param << "nbds_max=" << cfg->nbds_max; | |
615 | ||
616 | if (cfg->max_part) | |
617 | param << " max_part=" << cfg->max_part; | |
618 | ||
619 | if (!access("/sys/module/nbd", F_OK)) { | |
620 | if (cfg->nbds_max || cfg->set_max_part) | |
621 | cerr << "rbd-nbd: ignoring kernel module parameter options: nbd module already loaded" | |
622 | << std::endl; | |
623 | return 0; | |
7c673cae FG |
624 | } |
625 | ||
eafe8130 TL |
626 | ret = module_load("nbd", param.str().c_str()); |
627 | if (ret < 0) | |
628 | cerr << "rbd-nbd: failed to load nbd kernel module: " << cpp_strerror(-ret) | |
7c673cae | 629 | << std::endl; |
11fdf7f2 | 630 | |
eafe8130 | 631 | return ret; |
7c673cae FG |
632 | } |
633 | ||
634 | static int check_device_size(int nbd_index, unsigned long expected_size) | |
635 | { | |
636 | // There are bugs with some older kernel versions that result in an | |
637 | // overflow for large image sizes. This check is to ensure we are | |
638 | // not affected. | |
639 | ||
640 | unsigned long size = 0; | |
641 | std::string path = "/sys/block/nbd" + stringify(nbd_index) + "/size"; | |
642 | std::ifstream ifs; | |
643 | ifs.open(path.c_str(), std::ifstream::in); | |
644 | if (!ifs.is_open()) { | |
645 | cerr << "rbd-nbd: failed to open " << path << std::endl; | |
646 | return -EINVAL; | |
647 | } | |
648 | ifs >> size; | |
649 | size *= RBD_NBD_BLKSIZE; | |
650 | ||
651 | if (size == 0) { | |
652 | // Newer kernel versions will report real size only after nbd | |
653 | // connect. Assume this is the case and return success. | |
654 | return 0; | |
655 | } | |
656 | ||
657 | if (size != expected_size) { | |
658 | cerr << "rbd-nbd: kernel reported invalid device size (" << size | |
659 | << ", expected " << expected_size << ")" << std::endl; | |
660 | return -EINVAL; | |
661 | } | |
662 | ||
663 | return 0; | |
664 | } | |
665 | ||
eafe8130 TL |
666 | static int parse_nbd_index(const std::string& devpath) |
667 | { | |
668 | int index, ret; | |
669 | ||
670 | ret = sscanf(devpath.c_str(), "/dev/nbd%d", &index); | |
671 | if (ret <= 0) { | |
672 | // mean an early matching failure. But some cases need a negative value. | |
673 | if (ret == 0) | |
674 | ret = -EINVAL; | |
675 | cerr << "rbd-nbd: invalid device path: " << devpath | |
676 | << " (expected /dev/nbd{num})" << std::endl; | |
677 | return ret; | |
678 | } | |
679 | ||
680 | return index; | |
681 | } | |
682 | ||
683 | static int try_ioctl_setup(Config *cfg, int fd, uint64_t size, uint64_t flags) | |
684 | { | |
685 | int index = 0, r; | |
686 | ||
687 | if (cfg->devpath.empty()) { | |
688 | char dev[64]; | |
689 | const char *path = "/sys/module/nbd/parameters/nbds_max"; | |
690 | int nbds_max = -1; | |
691 | if (access(path, F_OK) == 0) { | |
692 | std::ifstream ifs; | |
693 | ifs.open(path, std::ifstream::in); | |
694 | if (ifs.is_open()) { | |
695 | ifs >> nbds_max; | |
696 | ifs.close(); | |
697 | } | |
698 | } | |
699 | ||
700 | while (true) { | |
701 | snprintf(dev, sizeof(dev), "/dev/nbd%d", index); | |
702 | ||
703 | nbd = open(dev, O_RDWR); | |
704 | if (nbd < 0) { | |
705 | if (nbd == -EPERM && nbds_max != -1 && index < (nbds_max-1)) { | |
706 | ++index; | |
707 | continue; | |
708 | } | |
709 | r = nbd; | |
710 | cerr << "rbd-nbd: failed to find unused device" << std::endl; | |
711 | goto done; | |
712 | } | |
713 | ||
714 | r = ioctl(nbd, NBD_SET_SOCK, fd); | |
715 | if (r < 0) { | |
716 | close(nbd); | |
717 | ++index; | |
718 | continue; | |
719 | } | |
720 | ||
721 | cfg->devpath = dev; | |
722 | break; | |
723 | } | |
724 | } else { | |
725 | r = parse_nbd_index(cfg->devpath); | |
726 | if (r < 0) | |
727 | goto done; | |
728 | index = r; | |
729 | ||
730 | nbd = open(cfg->devpath.c_str(), O_RDWR); | |
731 | if (nbd < 0) { | |
732 | r = nbd; | |
733 | cerr << "rbd-nbd: failed to open device: " << cfg->devpath << std::endl; | |
734 | goto done; | |
735 | } | |
736 | ||
737 | r = ioctl(nbd, NBD_SET_SOCK, fd); | |
738 | if (r < 0) { | |
739 | r = -errno; | |
740 | cerr << "rbd-nbd: the device " << cfg->devpath << " is busy" << std::endl; | |
741 | close(nbd); | |
742 | goto done; | |
743 | } | |
744 | } | |
745 | ||
746 | r = ioctl(nbd, NBD_SET_BLKSIZE, RBD_NBD_BLKSIZE); | |
747 | if (r < 0) { | |
748 | r = -errno; | |
749 | goto close_nbd; | |
750 | } | |
751 | ||
752 | r = ioctl(nbd, NBD_SET_SIZE, size); | |
753 | if (r < 0) { | |
754 | r = -errno; | |
755 | goto close_nbd; | |
756 | } | |
757 | ||
758 | ioctl(nbd, NBD_SET_FLAGS, flags); | |
759 | ||
760 | if (cfg->timeout >= 0) { | |
761 | r = ioctl(nbd, NBD_SET_TIMEOUT, (unsigned long)cfg->timeout); | |
762 | if (r < 0) { | |
763 | r = -errno; | |
764 | cerr << "rbd-nbd: failed to set timeout: " << cpp_strerror(r) | |
765 | << std::endl; | |
766 | goto close_nbd; | |
767 | } | |
768 | } | |
769 | ||
770 | dout(10) << "ioctl setup complete for " << cfg->devpath << dendl; | |
771 | nbd_index = index; | |
772 | return 0; | |
773 | ||
774 | close_nbd: | |
775 | if (r < 0) { | |
776 | ioctl(nbd, NBD_CLEAR_SOCK); | |
777 | cerr << "rbd-nbd: failed to map, status: " << cpp_strerror(-r) << std::endl; | |
778 | } | |
779 | close(nbd); | |
780 | done: | |
781 | return r; | |
782 | } | |
783 | ||
784 | static void netlink_cleanup(struct nl_sock *sock) | |
785 | { | |
786 | if (!sock) | |
787 | return; | |
788 | ||
789 | nl_close(sock); | |
790 | nl_socket_free(sock); | |
791 | } | |
792 | ||
793 | static struct nl_sock *netlink_init(int *id) | |
794 | { | |
795 | struct nl_sock *sock; | |
796 | int ret; | |
797 | ||
798 | sock = nl_socket_alloc(); | |
799 | if (!sock) { | |
800 | cerr << "rbd-nbd: Could not allocate netlink socket." << std::endl; | |
801 | return NULL; | |
802 | } | |
803 | ||
804 | ret = genl_connect(sock); | |
805 | if (ret < 0) { | |
806 | cerr << "rbd-nbd: Could not connect netlink socket. Error " << ret | |
807 | << std::endl; | |
808 | goto free_sock; | |
809 | } | |
810 | ||
811 | *id = genl_ctrl_resolve(sock, "nbd"); | |
812 | if (*id < 0) | |
813 | // nbd netlink interface not supported. | |
814 | goto close_sock; | |
815 | ||
816 | return sock; | |
817 | ||
818 | close_sock: | |
819 | nl_close(sock); | |
820 | free_sock: | |
821 | nl_socket_free(sock); | |
822 | return NULL; | |
823 | } | |
824 | ||
825 | static int netlink_disconnect(int index) | |
826 | { | |
827 | struct nl_sock *sock; | |
828 | struct nl_msg *msg; | |
829 | int ret, nl_id; | |
830 | ||
831 | sock = netlink_init(&nl_id); | |
832 | if (!sock) | |
833 | // Try ioctl | |
834 | return 1; | |
835 | ||
836 | nl_socket_modify_cb(sock, NL_CB_VALID, NL_CB_CUSTOM, genl_handle_msg, NULL); | |
837 | ||
838 | msg = nlmsg_alloc(); | |
839 | if (!msg) { | |
840 | cerr << "rbd-nbd: Could not allocate netlink message." << std::endl; | |
841 | goto free_sock; | |
842 | } | |
843 | ||
844 | if (!genlmsg_put(msg, NL_AUTO_PORT, NL_AUTO_SEQ, nl_id, 0, 0, | |
845 | NBD_CMD_DISCONNECT, 0)) { | |
846 | cerr << "rbd-nbd: Could not setup message." << std::endl; | |
847 | goto nla_put_failure; | |
848 | } | |
849 | ||
850 | NLA_PUT_U32(msg, NBD_ATTR_INDEX, index); | |
851 | ||
852 | ret = nl_send_sync(sock, msg); | |
853 | netlink_cleanup(sock); | |
854 | if (ret < 0) { | |
855 | cerr << "rbd-nbd: netlink disconnect failed: " << nl_geterror(-ret) | |
856 | << std::endl; | |
857 | return -EIO; | |
858 | } | |
859 | ||
860 | return 0; | |
861 | ||
862 | nla_put_failure: | |
863 | nlmsg_free(msg); | |
864 | free_sock: | |
865 | netlink_cleanup(sock); | |
866 | return -EIO; | |
867 | } | |
868 | ||
869 | static int netlink_disconnect_by_path(const std::string& devpath) | |
870 | { | |
871 | int index; | |
872 | ||
873 | index = parse_nbd_index(devpath); | |
874 | if (index < 0) | |
875 | return index; | |
876 | ||
877 | return netlink_disconnect(index); | |
878 | } | |
879 | ||
880 | static int netlink_resize(int nbd_index, uint64_t size) | |
881 | { | |
882 | struct nl_sock *sock; | |
883 | struct nl_msg *msg; | |
884 | int nl_id, ret; | |
885 | ||
886 | sock = netlink_init(&nl_id); | |
887 | if (!sock) { | |
888 | cerr << "rbd-nbd: Netlink interface not supported." << std::endl; | |
889 | return 1; | |
890 | } | |
891 | ||
892 | nl_socket_modify_cb(sock, NL_CB_VALID, NL_CB_CUSTOM, genl_handle_msg, NULL); | |
893 | ||
894 | msg = nlmsg_alloc(); | |
895 | if (!msg) { | |
896 | cerr << "rbd-nbd: Could not allocate netlink message." << std::endl; | |
897 | goto free_sock; | |
898 | } | |
899 | ||
900 | if (!genlmsg_put(msg, NL_AUTO_PORT, NL_AUTO_SEQ, nl_id, 0, 0, | |
901 | NBD_CMD_RECONFIGURE, 0)) { | |
902 | cerr << "rbd-nbd: Could not setup message." << std::endl; | |
903 | goto free_msg; | |
904 | } | |
905 | ||
906 | NLA_PUT_U32(msg, NBD_ATTR_INDEX, nbd_index); | |
907 | NLA_PUT_U64(msg, NBD_ATTR_SIZE_BYTES, size); | |
908 | ||
909 | ret = nl_send_sync(sock, msg); | |
910 | if (ret < 0) { | |
911 | cerr << "rbd-nbd: netlink resize failed: " << nl_geterror(ret) << std::endl; | |
912 | goto free_sock; | |
913 | } | |
914 | ||
915 | netlink_cleanup(sock); | |
916 | dout(10) << "netlink resize complete for nbd" << nbd_index << dendl; | |
917 | return 0; | |
918 | ||
919 | nla_put_failure: | |
920 | free_msg: | |
921 | nlmsg_free(msg); | |
922 | free_sock: | |
923 | netlink_cleanup(sock); | |
924 | return -EIO; | |
925 | } | |
926 | ||
927 | static int netlink_connect_cb(struct nl_msg *msg, void *arg) | |
928 | { | |
929 | struct genlmsghdr *gnlh = (struct genlmsghdr *)nlmsg_data(nlmsg_hdr(msg)); | |
930 | Config *cfg = (Config *)arg; | |
931 | struct nlattr *msg_attr[NBD_ATTR_MAX + 1]; | |
932 | uint32_t index; | |
933 | int ret; | |
934 | ||
935 | ret = nla_parse(msg_attr, NBD_ATTR_MAX, genlmsg_attrdata(gnlh, 0), | |
936 | genlmsg_attrlen(gnlh, 0), NULL); | |
937 | if (ret) { | |
938 | cerr << "rbd-nbd: Unsupported netlink reply" << std::endl; | |
939 | return -NLE_MSGTYPE_NOSUPPORT; | |
940 | } | |
941 | ||
942 | if (!msg_attr[NBD_ATTR_INDEX]) { | |
943 | cerr << "rbd-nbd: netlink connect reply missing device index." << std::endl; | |
944 | return -NLE_MSGTYPE_NOSUPPORT; | |
945 | } | |
946 | ||
947 | index = nla_get_u32(msg_attr[NBD_ATTR_INDEX]); | |
948 | cfg->devpath = "/dev/nbd" + stringify(index); | |
949 | nbd_index = index; | |
950 | ||
951 | return NL_OK; | |
952 | } | |
953 | ||
954 | static int netlink_connect(Config *cfg, struct nl_sock *sock, int nl_id, int fd, | |
955 | uint64_t size, uint64_t flags) | |
956 | { | |
957 | struct nlattr *sock_attr; | |
958 | struct nlattr *sock_opt; | |
959 | struct nl_msg *msg; | |
960 | int ret; | |
961 | ||
962 | nl_socket_modify_cb(sock, NL_CB_VALID, NL_CB_CUSTOM, netlink_connect_cb, cfg); | |
963 | ||
964 | msg = nlmsg_alloc(); | |
965 | if (!msg) { | |
966 | cerr << "rbd-nbd: Could not allocate netlink message." << std::endl; | |
967 | return -ENOMEM; | |
968 | } | |
969 | ||
970 | if (!genlmsg_put(msg, NL_AUTO_PORT, NL_AUTO_SEQ, nl_id, 0, 0, NBD_CMD_CONNECT, | |
971 | 0)) { | |
972 | cerr << "rbd-nbd: Could not setup message." << std::endl; | |
973 | goto free_msg; | |
974 | } | |
975 | ||
976 | if (!cfg->devpath.empty()) { | |
977 | ret = parse_nbd_index(cfg->devpath); | |
978 | if (ret < 0) | |
979 | goto free_msg; | |
980 | ||
981 | NLA_PUT_U32(msg, NBD_ATTR_INDEX, ret); | |
982 | } | |
983 | ||
984 | if (cfg->timeout >= 0) | |
985 | NLA_PUT_U64(msg, NBD_ATTR_TIMEOUT, cfg->timeout); | |
986 | ||
987 | NLA_PUT_U64(msg, NBD_ATTR_SIZE_BYTES, size); | |
988 | NLA_PUT_U64(msg, NBD_ATTR_BLOCK_SIZE_BYTES, RBD_NBD_BLKSIZE); | |
989 | NLA_PUT_U64(msg, NBD_ATTR_SERVER_FLAGS, flags); | |
990 | ||
991 | sock_attr = nla_nest_start(msg, NBD_ATTR_SOCKETS); | |
992 | if (!sock_attr) { | |
993 | cerr << "rbd-nbd: Could not init sockets in netlink message." << std::endl; | |
994 | goto free_msg; | |
995 | } | |
996 | ||
997 | sock_opt = nla_nest_start(msg, NBD_SOCK_ITEM); | |
998 | if (!sock_opt) { | |
999 | cerr << "rbd-nbd: Could not init sock in netlink message." << std::endl; | |
1000 | goto free_msg; | |
1001 | } | |
1002 | ||
1003 | NLA_PUT_U32(msg, NBD_SOCK_FD, fd); | |
1004 | nla_nest_end(msg, sock_opt); | |
1005 | nla_nest_end(msg, sock_attr); | |
1006 | ||
1007 | ret = nl_send_sync(sock, msg); | |
1008 | if (ret < 0) { | |
1009 | cerr << "rbd-nbd: netlink connect failed: " << nl_geterror(ret) | |
1010 | << std::endl; | |
1011 | return -EIO; | |
1012 | } | |
1013 | ||
1014 | dout(10) << "netlink connect complete for " << cfg->devpath << dendl; | |
1015 | return 0; | |
1016 | ||
1017 | nla_put_failure: | |
1018 | free_msg: | |
1019 | nlmsg_free(msg); | |
1020 | return -EIO; | |
1021 | } | |
1022 | ||
1023 | static int try_netlink_setup(Config *cfg, int fd, uint64_t size, uint64_t flags) | |
1024 | { | |
1025 | struct nl_sock *sock; | |
1026 | int nl_id, ret; | |
1027 | ||
1028 | sock = netlink_init(&nl_id); | |
1029 | if (!sock) { | |
1030 | cerr << "rbd-nbd: Netlink interface not supported. Using ioctl interface." | |
1031 | << std::endl; | |
1032 | return 1; | |
1033 | } | |
1034 | ||
1035 | dout(10) << "netlink interface supported." << dendl; | |
1036 | ||
1037 | ret = netlink_connect(cfg, sock, nl_id, fd, size, flags); | |
1038 | netlink_cleanup(sock); | |
1039 | ||
1040 | if (ret != 0) | |
1041 | return ret; | |
1042 | ||
1043 | nbd = open(cfg->devpath.c_str(), O_RDWR); | |
1044 | if (nbd < 0) { | |
1045 | cerr << "rbd-nbd: failed to open device: " << cfg->devpath << std::endl; | |
1046 | return nbd; | |
1047 | } | |
1048 | ||
1049 | return 0; | |
1050 | } | |
1051 | ||
1052 | static void handle_signal(int signum) | |
1053 | { | |
1054 | int ret; | |
1055 | ||
1056 | ceph_assert(signum == SIGINT || signum == SIGTERM); | |
1057 | derr << "*** Got signal " << sig_str(signum) << " ***" << dendl; | |
1058 | ||
1059 | if (nbd < 0 || nbd_index < 0) { | |
1060 | dout(20) << __func__ << ": " << "disconnect not needed." << dendl; | |
1061 | return; | |
1062 | } | |
1063 | ||
1064 | dout(20) << __func__ << ": " << "sending NBD_DISCONNECT" << dendl; | |
1065 | ret = netlink_disconnect(nbd_index); | |
1066 | if (ret == 1) | |
1067 | ret = ioctl(nbd, NBD_DISCONNECT); | |
1068 | ||
1069 | if (ret != 0) { | |
1070 | derr << "rbd-nbd: disconnect failed. Error: " << ret << dendl; | |
1071 | } else { | |
1072 | dout(20) << __func__ << ": " << "disconnected" << dendl; | |
1073 | } | |
1074 | } | |
1075 | ||
1076 | static NBDServer *start_server(int fd, librbd::Image& image) | |
1077 | { | |
1078 | NBDServer *server; | |
1079 | ||
1080 | server = new NBDServer(fd, image); | |
1081 | server->start(); | |
1082 | ||
1083 | init_async_signal_handler(); | |
1084 | register_async_signal_handler(SIGHUP, sighup_handler); | |
1085 | register_async_signal_handler_oneshot(SIGINT, handle_signal); | |
1086 | register_async_signal_handler_oneshot(SIGTERM, handle_signal); | |
1087 | ||
1088 | return server; | |
1089 | } | |
1090 | ||
1091 | static void run_server(Preforker& forker, NBDServer *server, bool netlink_used) | |
1092 | { | |
1093 | if (g_conf()->daemonize) { | |
1094 | global_init_postfork_finish(g_ceph_context); | |
1095 | forker.daemonize(); | |
1096 | } | |
1097 | ||
1098 | if (netlink_used) | |
1099 | server->wait_for_disconnect(); | |
1100 | else | |
1101 | ioctl(nbd, NBD_DO_IT); | |
1102 | ||
1103 | unregister_async_signal_handler(SIGHUP, sighup_handler); | |
1104 | unregister_async_signal_handler(SIGINT, handle_signal); | |
1105 | unregister_async_signal_handler(SIGTERM, handle_signal); | |
1106 | shutdown_async_signal_handler(); | |
1107 | } | |
1108 | ||
31f18b77 | 1109 | static int do_map(int argc, const char *argv[], Config *cfg) |
7c673cae FG |
1110 | { |
1111 | int r; | |
1112 | ||
1113 | librados::Rados rados; | |
1114 | librbd::RBD rbd; | |
1115 | librados::IoCtx io_ctx; | |
1116 | librbd::Image image; | |
1117 | ||
1118 | int read_only = 0; | |
1119 | unsigned long flags; | |
1120 | unsigned long size; | |
eafe8130 | 1121 | bool use_netlink; |
7c673cae | 1122 | |
7c673cae FG |
1123 | int fd[2]; |
1124 | ||
1125 | librbd::image_info_t info; | |
1126 | ||
1127 | Preforker forker; | |
eafe8130 | 1128 | NBDServer *server; |
7c673cae FG |
1129 | |
1130 | vector<const char*> args; | |
1131 | argv_to_vec(argc, argv, args); | |
11fdf7f2 TL |
1132 | if (args.empty()) { |
1133 | cerr << argv[0] << ": -h or --help for usage" << std::endl; | |
1134 | exit(1); | |
1135 | } | |
1136 | if (ceph_argparse_need_usage(args)) { | |
1137 | usage(); | |
1138 | exit(0); | |
1139 | } | |
7c673cae FG |
1140 | |
1141 | auto cct = global_init(NULL, args, CEPH_ENTITY_TYPE_CLIENT, | |
1142 | CODE_ENVIRONMENT_DAEMON, | |
1143 | CINIT_FLAG_UNPRIVILEGED_DAEMON_DEFAULTS); | |
11fdf7f2 | 1144 | g_ceph_context->_conf.set_val_or_die("pid_file", ""); |
7c673cae FG |
1145 | |
1146 | if (global_init_prefork(g_ceph_context) >= 0) { | |
1147 | std::string err; | |
1148 | r = forker.prefork(err); | |
1149 | if (r < 0) { | |
1150 | cerr << err << std::endl; | |
1151 | return r; | |
1152 | } | |
7c673cae | 1153 | if (forker.is_parent()) { |
7c673cae FG |
1154 | if (forker.parent_wait(err) != 0) { |
1155 | return -ENXIO; | |
1156 | } | |
1157 | return 0; | |
1158 | } | |
28e407b8 | 1159 | global_init_postfork_start(g_ceph_context); |
7c673cae FG |
1160 | } |
1161 | ||
1162 | common_init_finish(g_ceph_context); | |
1163 | global_init_chdir(g_ceph_context); | |
1164 | ||
1165 | if (socketpair(AF_UNIX, SOCK_STREAM, 0, fd) == -1) { | |
1166 | r = -errno; | |
1167 | goto close_ret; | |
1168 | } | |
1169 | ||
94b18763 FG |
1170 | r = rados.init_with_context(g_ceph_context); |
1171 | if (r < 0) | |
1172 | goto close_fd; | |
1173 | ||
1174 | r = rados.connect(); | |
1175 | if (r < 0) | |
1176 | goto close_fd; | |
1177 | ||
1178 | r = rados.ioctx_create(cfg->poolname.c_str(), io_ctx); | |
1179 | if (r < 0) | |
1180 | goto close_fd; | |
1181 | ||
11fdf7f2 TL |
1182 | io_ctx.set_namespace(cfg->nsname); |
1183 | ||
94b18763 FG |
1184 | r = rbd.open(io_ctx, image, cfg->imgname.c_str()); |
1185 | if (r < 0) | |
1186 | goto close_fd; | |
1187 | ||
1188 | if (cfg->exclusive) { | |
1189 | r = image.lock_acquire(RBD_LOCK_MODE_EXCLUSIVE); | |
1190 | if (r < 0) { | |
1191 | cerr << "rbd-nbd: failed to acquire exclusive lock: " << cpp_strerror(r) | |
1192 | << std::endl; | |
1193 | goto close_fd; | |
1194 | } | |
1195 | } | |
1196 | ||
1197 | if (!cfg->snapname.empty()) { | |
1198 | r = image.snap_set(cfg->snapname.c_str()); | |
1199 | if (r < 0) | |
1200 | goto close_fd; | |
1201 | } | |
1202 | ||
1203 | r = image.stat(info, sizeof(info)); | |
1204 | if (r < 0) | |
1205 | goto close_fd; | |
1206 | ||
7c673cae | 1207 | flags = NBD_FLAG_SEND_FLUSH | NBD_FLAG_SEND_TRIM | NBD_FLAG_HAS_FLAGS; |
31f18b77 | 1208 | if (!cfg->snapname.empty() || cfg->readonly) { |
7c673cae FG |
1209 | flags |= NBD_FLAG_READ_ONLY; |
1210 | read_only = 1; | |
1211 | } | |
1212 | ||
7c673cae FG |
1213 | if (info.size > ULONG_MAX) { |
1214 | r = -EFBIG; | |
1adf2230 AA |
1215 | cerr << "rbd-nbd: image is too large (" << byte_u_t(info.size) |
1216 | << ", max is " << byte_u_t(ULONG_MAX) << ")" << std::endl; | |
eafe8130 | 1217 | goto close_fd; |
7c673cae FG |
1218 | } |
1219 | ||
1220 | size = info.size; | |
1221 | ||
eafe8130 TL |
1222 | r = load_module(cfg); |
1223 | if (r < 0) | |
1224 | goto close_fd; | |
1225 | ||
1226 | server = start_server(fd[1], image); | |
1227 | ||
1228 | use_netlink = cfg->try_netlink; | |
1229 | if (use_netlink) { | |
1230 | r = try_netlink_setup(cfg, fd[0], size, flags); | |
1231 | if (r < 0) { | |
1232 | goto free_server; | |
1233 | } else if (r == 1) { | |
1234 | use_netlink = false; | |
1235 | } | |
7c673cae FG |
1236 | } |
1237 | ||
eafe8130 TL |
1238 | if (!use_netlink) { |
1239 | r = try_ioctl_setup(cfg, fd[0], size, flags); | |
1240 | if (r < 0) | |
1241 | goto free_server; | |
7c673cae FG |
1242 | } |
1243 | ||
eafe8130 TL |
1244 | r = check_device_size(nbd_index, size); |
1245 | if (r < 0) | |
1246 | goto close_nbd; | |
7c673cae FG |
1247 | |
1248 | r = ioctl(nbd, BLKROSET, (unsigned long) &read_only); | |
1249 | if (r < 0) { | |
1250 | r = -errno; | |
1251 | goto close_nbd; | |
1252 | } | |
1253 | ||
1254 | { | |
1255 | uint64_t handle; | |
1256 | ||
eafe8130 TL |
1257 | NBDWatchCtx watch_ctx(nbd, nbd_index, use_netlink, io_ctx, image, |
1258 | info.size); | |
7c673cae FG |
1259 | r = image.update_watch(&watch_ctx, &handle); |
1260 | if (r < 0) | |
1261 | goto close_nbd; | |
1262 | ||
31f18b77 | 1263 | cout << cfg->devpath << std::endl; |
7c673cae | 1264 | |
eafe8130 | 1265 | run_server(forker, server, use_netlink); |
7c673cae FG |
1266 | |
1267 | r = image.update_unwatch(handle); | |
11fdf7f2 | 1268 | ceph_assert(r == 0); |
7c673cae FG |
1269 | } |
1270 | ||
1271 | close_nbd: | |
1272 | if (r < 0) { | |
eafe8130 TL |
1273 | if (use_netlink) { |
1274 | netlink_disconnect(nbd_index); | |
1275 | } else { | |
1276 | ioctl(nbd, NBD_CLEAR_SOCK); | |
1277 | cerr << "rbd-nbd: failed to map, status: " << cpp_strerror(-r) | |
1278 | << std::endl; | |
1279 | } | |
7c673cae FG |
1280 | } |
1281 | close(nbd); | |
eafe8130 TL |
1282 | free_server: |
1283 | delete server; | |
7c673cae FG |
1284 | close_fd: |
1285 | close(fd[0]); | |
1286 | close(fd[1]); | |
1287 | close_ret: | |
1288 | image.close(); | |
1289 | io_ctx.close(); | |
1290 | rados.shutdown(); | |
1291 | ||
1292 | forker.exit(r < 0 ? EXIT_FAILURE : 0); | |
1293 | // Unreachable; | |
1294 | return r; | |
1295 | } | |
1296 | ||
eafe8130 | 1297 | static int do_unmap(Config *cfg) |
7c673cae | 1298 | { |
eafe8130 TL |
1299 | int r, nbd; |
1300 | ||
1301 | /* | |
1302 | * The netlink disconnect call supports devices setup with netlink or ioctl, | |
1303 | * so we always try that first. | |
1304 | */ | |
1305 | r = netlink_disconnect_by_path(cfg->devpath); | |
1306 | if (r != 1) | |
1307 | return r; | |
31f18b77 | 1308 | |
eafe8130 | 1309 | nbd = open(cfg->devpath.c_str(), O_RDWR); |
7c673cae | 1310 | if (nbd < 0) { |
eafe8130 | 1311 | cerr << "rbd-nbd: failed to open device: " << cfg->devpath << std::endl; |
7c673cae FG |
1312 | return nbd; |
1313 | } | |
1314 | ||
31f18b77 FG |
1315 | r = ioctl(nbd, NBD_DISCONNECT); |
1316 | if (r < 0) { | |
eafe8130 | 1317 | cerr << "rbd-nbd: the device is not used" << std::endl; |
7c673cae FG |
1318 | } |
1319 | ||
1320 | close(nbd); | |
31f18b77 | 1321 | return r; |
7c673cae FG |
1322 | } |
1323 | ||
11fdf7f2 TL |
1324 | static int parse_imgpath(const std::string &imgpath, Config *cfg, |
1325 | std::ostream *err_msg) { | |
1326 | std::regex pattern("^(?:([^/]+)/(?:([^/@]+)/)?)?([^@]+)(?:@([^/@]+))?$"); | |
1327 | std::smatch match; | |
1328 | if (!std::regex_match(imgpath, match, pattern)) { | |
7c673cae FG |
1329 | std::cerr << "rbd-nbd: invalid spec '" << imgpath << "'" << std::endl; |
1330 | return -EINVAL; | |
1331 | } | |
1332 | ||
31f18b77 FG |
1333 | if (match[1].matched) { |
1334 | cfg->poolname = match[1]; | |
1335 | } | |
7c673cae | 1336 | |
11fdf7f2 TL |
1337 | if (match[2].matched) { |
1338 | cfg->nsname = match[2]; | |
1339 | } | |
1340 | ||
1341 | cfg->imgname = match[3]; | |
7c673cae | 1342 | |
11fdf7f2 TL |
1343 | if (match[4].matched) |
1344 | cfg->snapname = match[4]; | |
7c673cae FG |
1345 | |
1346 | return 0; | |
1347 | } | |
1348 | ||
11fdf7f2 | 1349 | static int do_list_mapped_devices(const std::string &format, bool pretty_format) |
7c673cae | 1350 | { |
11fdf7f2 TL |
1351 | bool should_print = false; |
1352 | std::unique_ptr<ceph::Formatter> f; | |
1353 | TextTable tbl; | |
7c673cae | 1354 | |
11fdf7f2 TL |
1355 | if (format == "json") { |
1356 | f.reset(new JSONFormatter(pretty_format)); | |
1357 | } else if (format == "xml") { | |
1358 | f.reset(new XMLFormatter(pretty_format)); | |
1359 | } else if (!format.empty() && format != "plain") { | |
1360 | std::cerr << "rbd-nbd: invalid output format: " << format << std::endl; | |
1361 | return -EINVAL; | |
7c673cae FG |
1362 | } |
1363 | ||
11fdf7f2 TL |
1364 | if (f) { |
1365 | f->open_array_section("devices"); | |
1366 | } else { | |
1367 | tbl.define_column("id", TextTable::LEFT, TextTable::LEFT); | |
1368 | tbl.define_column("pool", TextTable::LEFT, TextTable::LEFT); | |
1369 | tbl.define_column("namespace", TextTable::LEFT, TextTable::LEFT); | |
1370 | tbl.define_column("image", TextTable::LEFT, TextTable::LEFT); | |
1371 | tbl.define_column("snap", TextTable::LEFT, TextTable::LEFT); | |
1372 | tbl.define_column("device", TextTable::LEFT, TextTable::LEFT); | |
7c673cae | 1373 | } |
7c673cae | 1374 | |
11fdf7f2 TL |
1375 | int pid; |
1376 | Config cfg; | |
1377 | NBDListIterator it; | |
1378 | while (it.get(&pid, &cfg)) { | |
1379 | if (f) { | |
1380 | f->open_object_section("device"); | |
1381 | f->dump_int("id", pid); | |
1382 | f->dump_string("pool", cfg.poolname); | |
1383 | f->dump_string("namespace", cfg.nsname); | |
1384 | f->dump_string("image", cfg.imgname); | |
1385 | f->dump_string("snap", cfg.snapname); | |
1386 | f->dump_string("device", cfg.devpath); | |
1387 | f->close_section(); | |
1388 | } else { | |
31f18b77 FG |
1389 | should_print = true; |
1390 | if (cfg.snapname.empty()) { | |
1391 | cfg.snapname = "-"; | |
1392 | } | |
11fdf7f2 TL |
1393 | tbl << pid << cfg.poolname << cfg.nsname << cfg.imgname << cfg.snapname |
1394 | << cfg.devpath << TextTable::endrow; | |
31f18b77 | 1395 | } |
31f18b77 FG |
1396 | } |
1397 | ||
11fdf7f2 TL |
1398 | if (f) { |
1399 | f->close_section(); // devices | |
1400 | f->flush(std::cout); | |
1401 | } | |
31f18b77 | 1402 | if (should_print) { |
11fdf7f2 | 1403 | std::cout << tbl; |
31f18b77 FG |
1404 | } |
1405 | return 0; | |
1406 | } | |
1407 | ||
11fdf7f2 TL |
1408 | static bool find_mapped_dev_by_spec(Config *cfg) { |
1409 | int pid; | |
1410 | Config c; | |
1411 | NBDListIterator it; | |
1412 | while (it.get(&pid, &c)) { | |
1413 | if (c.poolname == cfg->poolname && c.imgname == cfg->imgname && | |
1414 | c.snapname == cfg->snapname) { | |
1415 | *cfg = c; | |
1416 | return true; | |
1417 | } | |
1418 | } | |
1419 | return false; | |
1420 | } | |
1421 | ||
1422 | ||
1423 | static int parse_args(vector<const char*>& args, std::ostream *err_msg, | |
1424 | Command *command, Config *cfg) { | |
181888fb FG |
1425 | std::string conf_file_list; |
1426 | std::string cluster; | |
1427 | CephInitParameters iparams = ceph_argparse_early_args( | |
1428 | args, CEPH_ENTITY_TYPE_CLIENT, &cluster, &conf_file_list); | |
7c673cae | 1429 | |
11fdf7f2 TL |
1430 | ConfigProxy config{false}; |
1431 | config->name = iparams.name; | |
1432 | config->cluster = cluster; | |
181888fb FG |
1433 | |
1434 | if (!conf_file_list.empty()) { | |
1435 | config.parse_config_files(conf_file_list.c_str(), nullptr, 0); | |
1436 | } else { | |
1437 | config.parse_config_files(nullptr, nullptr, 0); | |
1438 | } | |
11fdf7f2 | 1439 | config.parse_env(CEPH_ENTITY_TYPE_CLIENT); |
31f18b77 | 1440 | config.parse_argv(args); |
181888fb FG |
1441 | cfg->poolname = config.get_val<std::string>("rbd_default_pool"); |
1442 | ||
1443 | std::vector<const char*>::iterator i; | |
1444 | std::ostringstream err; | |
31f18b77 | 1445 | |
7c673cae FG |
1446 | for (i = args.begin(); i != args.end(); ) { |
1447 | if (ceph_argparse_flag(args, i, "-h", "--help", (char*)NULL)) { | |
224ce89b WB |
1448 | return HELP_INFO; |
1449 | } else if (ceph_argparse_flag(args, i, "-v", "--version", (char*)NULL)) { | |
1450 | return VERSION_INFO; | |
31f18b77 FG |
1451 | } else if (ceph_argparse_witharg(args, i, &cfg->devpath, "--device", (char *)NULL)) { |
1452 | } else if (ceph_argparse_witharg(args, i, &cfg->nbds_max, err, "--nbds_max", (char *)NULL)) { | |
7c673cae | 1453 | if (!err.str().empty()) { |
31f18b77 FG |
1454 | *err_msg << "rbd-nbd: " << err.str(); |
1455 | return -EINVAL; | |
7c673cae | 1456 | } |
31f18b77 FG |
1457 | if (cfg->nbds_max < 0) { |
1458 | *err_msg << "rbd-nbd: Invalid argument for nbds_max!"; | |
1459 | return -EINVAL; | |
7c673cae | 1460 | } |
31f18b77 | 1461 | } else if (ceph_argparse_witharg(args, i, &cfg->max_part, err, "--max_part", (char *)NULL)) { |
7c673cae | 1462 | if (!err.str().empty()) { |
31f18b77 FG |
1463 | *err_msg << "rbd-nbd: " << err.str(); |
1464 | return -EINVAL; | |
7c673cae | 1465 | } |
31f18b77 FG |
1466 | if ((cfg->max_part < 0) || (cfg->max_part > 255)) { |
1467 | *err_msg << "rbd-nbd: Invalid argument for max_part(0~255)!"; | |
1468 | return -EINVAL; | |
7c673cae | 1469 | } |
31f18b77 | 1470 | cfg->set_max_part = true; |
7c673cae | 1471 | } else if (ceph_argparse_flag(args, i, "--read-only", (char *)NULL)) { |
31f18b77 | 1472 | cfg->readonly = true; |
7c673cae | 1473 | } else if (ceph_argparse_flag(args, i, "--exclusive", (char *)NULL)) { |
31f18b77 | 1474 | cfg->exclusive = true; |
11fdf7f2 TL |
1475 | } else if (ceph_argparse_witharg(args, i, &cfg->timeout, err, "--timeout", |
1476 | (char *)NULL)) { | |
1477 | if (!err.str().empty()) { | |
1478 | *err_msg << "rbd-nbd: " << err.str(); | |
1479 | return -EINVAL; | |
1480 | } | |
1481 | if (cfg->timeout < 0) { | |
1482 | *err_msg << "rbd-nbd: Invalid argument for timeout!"; | |
1483 | return -EINVAL; | |
1484 | } | |
1485 | } else if (ceph_argparse_witharg(args, i, &cfg->format, err, "--format", | |
1486 | (char *)NULL)) { | |
1487 | } else if (ceph_argparse_flag(args, i, "--pretty-format", (char *)NULL)) { | |
1488 | cfg->pretty_format = true; | |
eafe8130 TL |
1489 | } else if (ceph_argparse_flag(args, i, "--try-netlink", (char *)NULL)) { |
1490 | cfg->try_netlink = true; | |
7c673cae FG |
1491 | } else { |
1492 | ++i; | |
1493 | } | |
1494 | } | |
1495 | ||
11fdf7f2 | 1496 | Command cmd = None; |
7c673cae FG |
1497 | if (args.begin() != args.end()) { |
1498 | if (strcmp(*args.begin(), "map") == 0) { | |
1499 | cmd = Connect; | |
1500 | } else if (strcmp(*args.begin(), "unmap") == 0) { | |
1501 | cmd = Disconnect; | |
1502 | } else if (strcmp(*args.begin(), "list-mapped") == 0) { | |
1503 | cmd = List; | |
1504 | } else { | |
31f18b77 FG |
1505 | *err_msg << "rbd-nbd: unknown command: " << *args.begin(); |
1506 | return -EINVAL; | |
7c673cae FG |
1507 | } |
1508 | args.erase(args.begin()); | |
1509 | } | |
1510 | ||
1511 | if (cmd == None) { | |
31f18b77 FG |
1512 | *err_msg << "rbd-nbd: must specify command"; |
1513 | return -EINVAL; | |
7c673cae FG |
1514 | } |
1515 | ||
1516 | switch (cmd) { | |
1517 | case Connect: | |
1518 | if (args.begin() == args.end()) { | |
31f18b77 FG |
1519 | *err_msg << "rbd-nbd: must specify image-or-snap-spec"; |
1520 | return -EINVAL; | |
7c673cae | 1521 | } |
11fdf7f2 | 1522 | if (parse_imgpath(*args.begin(), cfg, err_msg) < 0) { |
31f18b77 | 1523 | return -EINVAL; |
11fdf7f2 | 1524 | } |
7c673cae FG |
1525 | args.erase(args.begin()); |
1526 | break; | |
1527 | case Disconnect: | |
1528 | if (args.begin() == args.end()) { | |
11fdf7f2 | 1529 | *err_msg << "rbd-nbd: must specify nbd device or image-or-snap-spec"; |
31f18b77 | 1530 | return -EINVAL; |
7c673cae | 1531 | } |
11fdf7f2 TL |
1532 | if (boost::starts_with(*args.begin(), "/dev/")) { |
1533 | cfg->devpath = *args.begin(); | |
1534 | } else { | |
1535 | if (parse_imgpath(*args.begin(), cfg, err_msg) < 0) { | |
1536 | return -EINVAL; | |
1537 | } | |
1538 | if (!find_mapped_dev_by_spec(cfg)) { | |
1539 | *err_msg << "rbd-nbd: " << *args.begin() << " is not mapped"; | |
1540 | return -ENOENT; | |
1541 | } | |
1542 | } | |
7c673cae FG |
1543 | args.erase(args.begin()); |
1544 | break; | |
1545 | default: | |
1546 | //shut up gcc; | |
1547 | break; | |
1548 | } | |
1549 | ||
1550 | if (args.begin() != args.end()) { | |
31f18b77 FG |
1551 | *err_msg << "rbd-nbd: unknown args: " << *args.begin(); |
1552 | return -EINVAL; | |
1553 | } | |
1554 | ||
11fdf7f2 | 1555 | *command = cmd; |
31f18b77 FG |
1556 | return 0; |
1557 | } | |
1558 | ||
1559 | static int rbd_nbd(int argc, const char *argv[]) | |
1560 | { | |
1561 | int r; | |
1562 | Config cfg; | |
1563 | vector<const char*> args; | |
1564 | argv_to_vec(argc, argv, args); | |
1565 | ||
1566 | std::ostringstream err_msg; | |
11fdf7f2 | 1567 | r = parse_args(args, &err_msg, &cmd, &cfg); |
224ce89b | 1568 | if (r == HELP_INFO) { |
31f18b77 | 1569 | usage(); |
11fdf7f2 | 1570 | return 0; |
224ce89b WB |
1571 | } else if (r == VERSION_INFO) { |
1572 | std::cout << pretty_version_to_str() << std::endl; | |
1573 | return 0; | |
11fdf7f2 | 1574 | } else if (r < 0) { |
31f18b77 FG |
1575 | cerr << err_msg.str() << std::endl; |
1576 | return r; | |
7c673cae FG |
1577 | } |
1578 | ||
1579 | switch (cmd) { | |
1580 | case Connect: | |
31f18b77 | 1581 | if (cfg.imgname.empty()) { |
7c673cae | 1582 | cerr << "rbd-nbd: image name was not specified" << std::endl; |
31f18b77 | 1583 | return -EINVAL; |
7c673cae FG |
1584 | } |
1585 | ||
31f18b77 | 1586 | r = do_map(argc, argv, &cfg); |
7c673cae | 1587 | if (r < 0) |
31f18b77 | 1588 | return -EINVAL; |
7c673cae FG |
1589 | break; |
1590 | case Disconnect: | |
eafe8130 | 1591 | r = do_unmap(&cfg); |
7c673cae | 1592 | if (r < 0) |
31f18b77 | 1593 | return -EINVAL; |
7c673cae FG |
1594 | break; |
1595 | case List: | |
11fdf7f2 | 1596 | r = do_list_mapped_devices(cfg.format, cfg.pretty_format); |
7c673cae | 1597 | if (r < 0) |
31f18b77 | 1598 | return -EINVAL; |
7c673cae FG |
1599 | break; |
1600 | default: | |
1601 | usage(); | |
d2e6a577 | 1602 | break; |
7c673cae FG |
1603 | } |
1604 | ||
1605 | return 0; | |
1606 | } | |
1607 | ||
1608 | int main(int argc, const char *argv[]) | |
1609 | { | |
31f18b77 FG |
1610 | int r = rbd_nbd(argc, argv); |
1611 | if (r < 0) { | |
1612 | return EXIT_FAILURE; | |
1613 | } | |
1614 | return 0; | |
7c673cae | 1615 | } |