]> git.proxmox.com Git - ceph.git/blob - ceph/src/os/filestore/BtrfsFileStoreBackend.cc
df1d2452a1fc7d69d0e3155de10a32395c5c1a3f
[ceph.git] / ceph / src / os / filestore / BtrfsFileStoreBackend.cc
1 // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
2 // vim: ts=8 sw=2 smarttab
3 /*
4 * Ceph - scalable distributed file system
5 *
6 * Copyright (C) 2004-2006 Sage Weil <sage@newdream.net>
7 *
8 * This is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License version 2.1, as published by the Free Software
11 * Foundation. See file COPYING.
12 *
13 */
14
15 #include "include/int_types.h"
16 #include "include/types.h"
17
18 #include <unistd.h>
19 #include <fcntl.h>
20 #include <errno.h>
21 #include <stdlib.h>
22 #include <sys/types.h>
23 #include <sys/stat.h>
24 #include <sys/ioctl.h>
25 #include "include/compat.h"
26 #include "include/linux_fiemap.h"
27 #include "include/color.h"
28 #include "include/buffer.h"
29 #include "include/ceph_assert.h"
30
31 #ifndef __CYGWIN__
32 #include "os/fs/btrfs_ioctl.h"
33 #endif
34
35 #include <iostream>
36 #include <fstream>
37 #include <sstream>
38
39 #include "BtrfsFileStoreBackend.h"
40
41 #include "common/errno.h"
42 #include "common/config.h"
43
44 #if defined(__linux__)
45
46 #define dout_context cct()
47 #define dout_subsys ceph_subsys_filestore
48 #undef dout_prefix
49 #define dout_prefix *_dout << "btrfsfilestorebackend(" << get_basedir_path() << ") "
50
51 using std::cerr;
52 using std::list;
53 using std::string;
54
55 #define ALIGN_DOWN(x, by) ((x) - ((x) % (by)))
56 #define ALIGNED(x, by) (!((x) % (by)))
57 #define ALIGN_UP(x, by) (ALIGNED((x), (by)) ? (x) : (ALIGN_DOWN((x), (by)) + (by)))
58
59 BtrfsFileStoreBackend::BtrfsFileStoreBackend(FileStore *fs):
60 GenericFileStoreBackend(fs), has_clone_range(false),
61 has_snap_create(false), has_snap_destroy(false),
62 has_snap_create_v2(false), has_wait_sync(false), stable_commits(false),
63 m_filestore_btrfs_clone_range(cct()->_conf->filestore_btrfs_clone_range),
64 m_filestore_btrfs_snap (cct()->_conf->filestore_btrfs_snap) { }
65
66 int BtrfsFileStoreBackend::detect_features()
67 {
68 int r;
69
70 r = GenericFileStoreBackend::detect_features();
71 if (r < 0)
72 return r;
73
74 // clone_range?
75 if (m_filestore_btrfs_clone_range) {
76 int fd = ::openat(get_basedir_fd(), "clone_range_test", O_CREAT|O_WRONLY|O_CLOEXEC, 0600);
77 if (fd >= 0) {
78 if (::unlinkat(get_basedir_fd(), "clone_range_test", 0) < 0) {
79 r = -errno;
80 dout(0) << "detect_feature: failed to unlink test file for CLONE_RANGE ioctl: "
81 << cpp_strerror(r) << dendl;
82 }
83 btrfs_ioctl_clone_range_args clone_args;
84 memset(&clone_args, 0, sizeof(clone_args));
85 clone_args.src_fd = -1;
86 r = ::ioctl(fd, BTRFS_IOC_CLONE_RANGE, &clone_args);
87 if (r < 0 && errno == EBADF) {
88 dout(0) << "detect_feature: CLONE_RANGE ioctl is supported" << dendl;
89 has_clone_range = true;
90 } else {
91 r = -errno;
92 dout(0) << "detect_feature: CLONE_RANGE ioctl is NOT supported: " << cpp_strerror(r) << dendl;
93 }
94 TEMP_FAILURE_RETRY(::close(fd));
95 } else {
96 r = -errno;
97 dout(0) << "detect_feature: failed to create test file for CLONE_RANGE ioctl: "
98 << cpp_strerror(r) << dendl;
99 }
100 } else {
101 dout(0) << "detect_feature: CLONE_RANGE ioctl is DISABLED via 'filestore btrfs clone range' option" << dendl;
102 }
103
104 struct btrfs_ioctl_vol_args vol_args;
105 memset(&vol_args, 0, sizeof(vol_args));
106
107 // create test source volume
108 vol_args.fd = 0;
109 strcpy(vol_args.name, "test_subvol");
110 r = ::ioctl(get_basedir_fd(), BTRFS_IOC_SUBVOL_CREATE, &vol_args);
111 if (r != 0) {
112 r = -errno;
113 dout(0) << "detect_feature: failed to create simple subvolume " << vol_args.name << ": " << cpp_strerror(r) << dendl;
114 }
115 int srcfd = ::openat(get_basedir_fd(), vol_args.name, O_RDONLY|O_CLOEXEC);
116 if (srcfd < 0) {
117 r = -errno;
118 dout(0) << "detect_feature: failed to open " << vol_args.name << ": " << cpp_strerror(r) << dendl;
119 }
120
121 // snap_create and snap_destroy?
122 vol_args.fd = srcfd;
123 strcpy(vol_args.name, "sync_snap_test");
124 r = ::ioctl(get_basedir_fd(), BTRFS_IOC_SNAP_CREATE, &vol_args);
125 int err = errno;
126 if (r == 0 || errno == EEXIST) {
127 dout(0) << "detect_feature: SNAP_CREATE is supported" << dendl;
128 has_snap_create = true;
129
130 r = ::ioctl(get_basedir_fd(), BTRFS_IOC_SNAP_DESTROY, &vol_args);
131 if (r == 0) {
132 dout(0) << "detect_feature: SNAP_DESTROY is supported" << dendl;
133 has_snap_destroy = true;
134 } else {
135 err = -errno;
136 dout(0) << "detect_feature: SNAP_DESTROY failed: " << cpp_strerror(err) << dendl;
137
138 if (err == -EPERM && getuid() != 0) {
139 dout(0) << "detect_feature: failed with EPERM as non-root; remount with -o user_subvol_rm_allowed" << dendl;
140 cerr << TEXT_YELLOW
141 << "btrfs SNAP_DESTROY failed as non-root; remount with -o user_subvol_rm_allowed"
142 << TEXT_NORMAL << std::endl;
143 } else if (err == -EOPNOTSUPP) {
144 derr << "btrfs SNAP_DESTROY ioctl not supported; you need a kernel newer than 2.6.32" << dendl;
145 }
146 }
147 } else {
148 dout(0) << "detect_feature: SNAP_CREATE failed: " << cpp_strerror(err) << dendl;
149 }
150
151 if (m_filestore_btrfs_snap) {
152 if (has_snap_destroy)
153 stable_commits = true;
154 else
155 dout(0) << "detect_feature: snaps enabled, but no SNAP_DESTROY ioctl; DISABLING" << dendl;
156 }
157
158 // start_sync?
159 __u64 transid = 0;
160 r = ::ioctl(get_basedir_fd(), BTRFS_IOC_START_SYNC, &transid);
161 if (r < 0) {
162 int err = errno;
163 dout(0) << "detect_feature: START_SYNC got " << cpp_strerror(err) << dendl;
164 }
165 if (r == 0 && transid > 0) {
166 dout(0) << "detect_feature: START_SYNC is supported (transid " << transid << ")" << dendl;
167
168 // do we have wait_sync too?
169 r = ::ioctl(get_basedir_fd(), BTRFS_IOC_WAIT_SYNC, &transid);
170 if (r == 0 || errno == ERANGE) {
171 dout(0) << "detect_feature: WAIT_SYNC is supported" << dendl;
172 has_wait_sync = true;
173 } else {
174 int err = errno;
175 dout(0) << "detect_feature: WAIT_SYNC is NOT supported: " << cpp_strerror(err) << dendl;
176 }
177 } else {
178 int err = errno;
179 dout(0) << "detect_feature: START_SYNC is NOT supported: " << cpp_strerror(err) << dendl;
180 }
181
182 if (has_wait_sync) {
183 // async snap creation?
184 struct btrfs_ioctl_vol_args_v2 async_args;
185 memset(&async_args, 0, sizeof(async_args));
186 async_args.fd = srcfd;
187 async_args.flags = BTRFS_SUBVOL_CREATE_ASYNC;
188 strcpy(async_args.name, "async_snap_test");
189
190 // remove old one, first
191 struct stat st;
192 strcpy(vol_args.name, async_args.name);
193 if (::fstatat(get_basedir_fd(), vol_args.name, &st, 0) == 0) {
194 dout(0) << "detect_feature: removing old async_snap_test" << dendl;
195 r = ::ioctl(get_basedir_fd(), BTRFS_IOC_SNAP_DESTROY, &vol_args);
196 if (r != 0) {
197 int err = errno;
198 dout(0) << "detect_feature: failed to remove old async_snap_test: " << cpp_strerror(err) << dendl;
199 }
200 }
201
202 r = ::ioctl(get_basedir_fd(), BTRFS_IOC_SNAP_CREATE_V2, &async_args);
203 if (r == 0 || errno == EEXIST) {
204 dout(0) << "detect_feature: SNAP_CREATE_V2 is supported" << dendl;
205 has_snap_create_v2 = true;
206
207 // clean up
208 strcpy(vol_args.name, "async_snap_test");
209 r = ::ioctl(get_basedir_fd(), BTRFS_IOC_SNAP_DESTROY, &vol_args);
210 if (r != 0) {
211 int err = errno;
212 dout(0) << "detect_feature: SNAP_DESTROY failed: " << cpp_strerror(err) << dendl;
213 }
214 } else {
215 int err = errno;
216 dout(0) << "detect_feature: SNAP_CREATE_V2 is NOT supported: " << cpp_strerror(err) << dendl;
217 }
218 }
219
220 // clean up test subvol
221 if (srcfd >= 0)
222 TEMP_FAILURE_RETRY(::close(srcfd));
223
224 strcpy(vol_args.name, "test_subvol");
225 r = ::ioctl(get_basedir_fd(), BTRFS_IOC_SNAP_DESTROY, &vol_args);
226 if (r < 0) {
227 r = -errno;
228 dout(0) << "detect_feature: failed to remove " << vol_args.name << ": " << cpp_strerror(r) << dendl;
229 }
230
231 if (m_filestore_btrfs_snap && !has_snap_create_v2) {
232 dout(0) << "mount WARNING: btrfs snaps enabled, but no SNAP_CREATE_V2 ioctl (from kernel 2.6.37+)" << dendl;
233 cerr << TEXT_YELLOW
234 << " ** WARNING: 'filestore btrfs snap' is enabled (for safe transactions,\n"
235 << " rollback), but btrfs does not support the SNAP_CREATE_V2 ioctl\n"
236 << " (added in Linux 2.6.37). Expect slow btrfs sync/commit\n"
237 << " performance.\n"
238 << TEXT_NORMAL;
239 }
240
241 return 0;
242 }
243
244 bool BtrfsFileStoreBackend::can_checkpoint()
245 {
246 return stable_commits;
247 }
248
249 int BtrfsFileStoreBackend::create_current()
250 {
251 struct stat st;
252 int ret = ::stat(get_current_path().c_str(), &st);
253 if (ret == 0) {
254 // current/ exists
255 if (!S_ISDIR(st.st_mode)) {
256 dout(0) << "create_current: current/ exists but is not a directory" << dendl;
257 return -EINVAL;
258 }
259
260 struct stat basest;
261 struct statfs currentfs;
262 ret = ::fstat(get_basedir_fd(), &basest);
263 if (ret < 0) {
264 ret = -errno;
265 dout(0) << "create_current: cannot fstat basedir " << cpp_strerror(ret) << dendl;
266 return ret;
267 }
268 ret = ::statfs(get_current_path().c_str(), &currentfs);
269 if (ret < 0) {
270 ret = -errno;
271 dout(0) << "create_current: cannot statsf basedir " << cpp_strerror(ret) << dendl;
272 return ret;
273 }
274 if (currentfs.f_type == BTRFS_SUPER_MAGIC && basest.st_dev != st.st_dev) {
275 dout(2) << "create_current: current appears to be a btrfs subvolume" << dendl;
276 stable_commits = true;
277 }
278 return 0;
279 }
280
281 struct btrfs_ioctl_vol_args volargs;
282 memset(&volargs, 0, sizeof(volargs));
283
284 volargs.fd = 0;
285 strcpy(volargs.name, "current");
286 if (::ioctl(get_basedir_fd(), BTRFS_IOC_SUBVOL_CREATE, (unsigned long int)&volargs) < 0) {
287 ret = -errno;
288 dout(0) << "create_current: BTRFS_IOC_SUBVOL_CREATE failed with error "
289 << cpp_strerror(ret) << dendl;
290 return ret;
291 }
292
293 dout(2) << "create_current: created btrfs subvol " << get_current_path() << dendl;
294 if (::chmod(get_current_path().c_str(), 0755) < 0) {
295 ret = -errno;
296 dout(0) << "create_current: failed to chmod " << get_current_path() << " to 0755: "
297 << cpp_strerror(ret) << dendl;
298 return ret;
299 }
300
301 stable_commits = true;
302 return 0;
303 }
304
305 int BtrfsFileStoreBackend::list_checkpoints(list<string>& ls)
306 {
307 int ret, err = 0;
308
309 struct stat basest;
310 ret = ::fstat(get_basedir_fd(), &basest);
311 if (ret < 0) {
312 ret = -errno;
313 dout(0) << "list_checkpoints: cannot fstat basedir " << cpp_strerror(ret) << dendl;
314 return ret;
315 }
316
317 // get snap list
318 DIR *dir = ::opendir(get_basedir_path().c_str());
319 if (!dir) {
320 ret = -errno;
321 dout(0) << "list_checkpoints: opendir '" << get_basedir_path() << "' failed: "
322 << cpp_strerror(ret) << dendl;
323 return ret;
324 }
325
326 list<string> snaps;
327 char path[PATH_MAX];
328 struct dirent *de;
329 while ((de = ::readdir(dir))) {
330 snprintf(path, sizeof(path), "%s/%s", get_basedir_path().c_str(), de->d_name);
331
332 struct stat st;
333 ret = ::stat(path, &st);
334 if (ret < 0) {
335 err = -errno;
336 dout(0) << "list_checkpoints: stat '" << path << "' failed: "
337 << cpp_strerror(err) << dendl;
338 break;
339 }
340
341 if (!S_ISDIR(st.st_mode))
342 continue;
343
344 struct statfs fs;
345 ret = ::statfs(path, &fs);
346 if (ret < 0) {
347 err = -errno;
348 dout(0) << "list_checkpoints: statfs '" << path << "' failed: "
349 << cpp_strerror(err) << dendl;
350 break;
351 }
352
353 if (fs.f_type == BTRFS_SUPER_MAGIC && basest.st_dev != st.st_dev)
354 snaps.push_back(string(de->d_name));
355 }
356
357 if (::closedir(dir) < 0) {
358 ret = -errno;
359 dout(0) << "list_checkpoints: closedir failed: " << cpp_strerror(ret) << dendl;
360 if (!err)
361 err = ret;
362 }
363
364 if (err)
365 return err;
366
367 ls.swap(snaps);
368 return 0;
369 }
370
371 int BtrfsFileStoreBackend::create_checkpoint(const string& name, uint64_t *transid)
372 {
373 dout(10) << "create_checkpoint: '" << name << "'" << dendl;
374 if (has_snap_create_v2 && transid) {
375 struct btrfs_ioctl_vol_args_v2 async_args;
376 memset(&async_args, 0, sizeof(async_args));
377 async_args.fd = get_current_fd();
378 async_args.flags = BTRFS_SUBVOL_CREATE_ASYNC;
379
380 size_t name_size = sizeof(async_args.name);
381 strncpy(async_args.name, name.c_str(), name_size);
382 async_args.name[name_size-1] = '\0';
383
384 int r = ::ioctl(get_basedir_fd(), BTRFS_IOC_SNAP_CREATE_V2, &async_args);
385 if (r < 0) {
386 r = -errno;
387 dout(0) << "create_checkpoint: async snap create '" << name << "' got " << cpp_strerror(r) << dendl;
388 return r;
389 }
390 dout(20) << "create_checkpoint: async snap create '" << name << "' transid " << async_args.transid << dendl;
391 *transid = async_args.transid;
392 } else {
393 struct btrfs_ioctl_vol_args vol_args;
394 memset(&vol_args, 0, sizeof(vol_args));
395 vol_args.fd = get_current_fd();
396
397 size_t name_size = sizeof(vol_args.name);
398 strncpy(vol_args.name, name.c_str(), name_size);
399 vol_args.name[name_size-1] = '\0';
400
401 int r = ::ioctl(get_basedir_fd(), BTRFS_IOC_SNAP_CREATE, &vol_args);
402 if (r < 0) {
403 r = -errno;
404 dout(0) << "create_checkpoint: snap create '" << name << "' got " << cpp_strerror(r) << dendl;
405 return r;
406 }
407 if (transid)
408 *transid = 0;
409 }
410 return 0;
411 }
412
413 int BtrfsFileStoreBackend::sync_checkpoint(uint64_t transid)
414 {
415 // wait for commit
416 dout(10) << "sync_checkpoint: transid " << transid << " to complete" << dendl;
417 int ret = ::ioctl(get_op_fd(), BTRFS_IOC_WAIT_SYNC, &transid);
418 if (ret < 0) {
419 ret = -errno;
420 dout(0) << "sync_checkpoint: ioctl WAIT_SYNC got " << cpp_strerror(ret) << dendl;
421 return -errno;
422 }
423 dout(20) << "sync_checkpoint: done waiting for transid " << transid << dendl;
424 return 0;
425 }
426
427 int BtrfsFileStoreBackend::rollback_to(const string& name)
428 {
429 dout(10) << "rollback_to: to '" << name << "'" << dendl;
430 char s[PATH_MAX];
431 btrfs_ioctl_vol_args vol_args;
432
433 memset(&vol_args, 0, sizeof(vol_args));
434 vol_args.fd = 0;
435 strcpy(vol_args.name, "current");
436
437 int ret = ::ioctl(get_basedir_fd(), BTRFS_IOC_SNAP_DESTROY, &vol_args);
438 if (ret && errno != ENOENT) {
439 dout(0) << "rollback_to: error removing old current subvol: " << cpp_strerror(ret) << dendl;
440 snprintf(s, sizeof(s), "%s/current.remove.me.%d", get_basedir_path().c_str(), rand());
441 if (::rename(get_current_path().c_str(), s)) {
442 ret = -errno;
443 dout(0) << "rollback_to: error renaming old current subvol: "
444 << cpp_strerror(ret) << dendl;
445 return ret;
446 }
447 }
448
449 snprintf(s, sizeof(s), "%s/%s", get_basedir_path().c_str(), name.c_str());
450
451 // roll back
452 vol_args.fd = ::open(s, O_RDONLY|O_CLOEXEC);
453 if (vol_args.fd < 0) {
454 ret = -errno;
455 dout(0) << "rollback_to: error opening '" << s << "': " << cpp_strerror(ret) << dendl;
456 return ret;
457 }
458 ret = ::ioctl(get_basedir_fd(), BTRFS_IOC_SNAP_CREATE, &vol_args);
459 if (ret < 0 ) {
460 ret = -errno;
461 dout(0) << "rollback_to: ioctl SNAP_CREATE got " << cpp_strerror(ret) << dendl;
462 }
463 TEMP_FAILURE_RETRY(::close(vol_args.fd));
464 return ret;
465 }
466
467 int BtrfsFileStoreBackend::destroy_checkpoint(const string& name)
468 {
469 dout(10) << "destroy_checkpoint: '" << name << "'" << dendl;
470 btrfs_ioctl_vol_args vol_args;
471 memset(&vol_args, 0, sizeof(vol_args));
472 vol_args.fd = 0;
473 strncpy(vol_args.name, name.c_str(), sizeof(vol_args.name) - 1);
474 vol_args.name[sizeof(vol_args.name) - 1] = '\0';
475
476 int ret = ::ioctl(get_basedir_fd(), BTRFS_IOC_SNAP_DESTROY, &vol_args);
477 if (ret) {
478 ret = -errno;
479 dout(0) << "destroy_checkpoint: ioctl SNAP_DESTROY got " << cpp_strerror(ret) << dendl;
480 return ret;
481 }
482 return 0;
483 }
484
485 int BtrfsFileStoreBackend::syncfs()
486 {
487 dout(15) << "syncfs" << dendl;
488 // do a full btrfs commit
489 int ret = ::ioctl(get_op_fd(), BTRFS_IOC_SYNC);
490 if (ret < 0) {
491 ret = -errno;
492 dout(0) << "syncfs: btrfs IOC_SYNC got " << cpp_strerror(ret) << dendl;
493 }
494 return ret;
495 }
496
497 int BtrfsFileStoreBackend::clone_range(int from, int to, uint64_t srcoff, uint64_t len, uint64_t dstoff)
498 {
499 dout(20) << "clone_range: " << srcoff << "~" << len << " to " << dstoff << dendl;
500 size_t blk_size = get_blksize();
501 if (!has_clone_range ||
502 srcoff % blk_size != dstoff % blk_size) {
503 dout(20) << "clone_range: using copy" << dendl;
504 return _copy_range(from, to, srcoff, len, dstoff);
505 }
506
507 int err = 0;
508 int r = 0;
509
510 uint64_t srcoffclone = ALIGN_UP(srcoff, blk_size);
511 uint64_t dstoffclone = ALIGN_UP(dstoff, blk_size);
512 if (srcoffclone >= srcoff + len) {
513 dout(20) << "clone_range: using copy, extent too short to align srcoff" << dendl;
514 return _copy_range(from, to, srcoff, len, dstoff);
515 }
516
517 uint64_t lenclone = len - (srcoffclone - srcoff);
518 if (!ALIGNED(lenclone, blk_size)) {
519 struct stat from_stat, to_stat;
520 err = ::fstat(from, &from_stat);
521 if (err) return -errno;
522 err = ::fstat(to , &to_stat);
523 if (err) return -errno;
524
525 if (srcoff + len != (uint64_t)from_stat.st_size ||
526 dstoff + len < (uint64_t)to_stat.st_size) {
527 // Not to the end of the file, need to align length as well
528 lenclone = ALIGN_DOWN(lenclone, blk_size);
529 }
530 }
531 if (lenclone == 0) {
532 // too short
533 return _copy_range(from, to, srcoff, len, dstoff);
534 }
535
536 dout(20) << "clone_range: cloning " << srcoffclone << "~" << lenclone
537 << " to " << dstoffclone << " = " << r << dendl;
538 btrfs_ioctl_clone_range_args a;
539 a.src_fd = from;
540 a.src_offset = srcoffclone;
541 a.src_length = lenclone;
542 a.dest_offset = dstoffclone;
543 err = ::ioctl(to, BTRFS_IOC_CLONE_RANGE, &a);
544 if (err >= 0) {
545 r += err;
546 } else if (errno == EINVAL) {
547 // Still failed, might be compressed
548 dout(20) << "clone_range: failed CLONE_RANGE call with -EINVAL, using copy" << dendl;
549 return _copy_range(from, to, srcoff, len, dstoff);
550 } else {
551 return -errno;
552 }
553
554 // Take care any trimmed from front
555 if (srcoffclone != srcoff) {
556 err = _copy_range(from, to, srcoff, srcoffclone - srcoff, dstoff);
557 if (err >= 0) {
558 r += err;
559 } else {
560 return err;
561 }
562 }
563
564 // Copy end
565 if (srcoffclone + lenclone != srcoff + len) {
566 err = _copy_range(from, to,
567 srcoffclone + lenclone,
568 (srcoff + len) - (srcoffclone + lenclone),
569 dstoffclone + lenclone);
570 if (err >= 0) {
571 r += err;
572 } else {
573 return err;
574 }
575 }
576 dout(20) << "clone_range: finished " << srcoff << "~" << len
577 << " to " << dstoff << " = " << r << dendl;
578 return r;
579 }
580 #endif