]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * mdadm - manage Linux "md" devices aka RAID arrays. | |
3 | * | |
4 | * Copyright (C) 2001-2009 Neil Brown <neilb@suse.de> | |
5 | * | |
6 | * | |
7 | * This program is free software; you can redistribute it and/or modify | |
8 | * it under the terms of the GNU General Public License as published by | |
9 | * the Free Software Foundation; either version 2 of the License, or | |
10 | * (at your option) any later version. | |
11 | * | |
12 | * This program is distributed in the hope that it will be useful, | |
13 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
14 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
15 | * GNU General Public License for more details. | |
16 | * | |
17 | * You should have received a copy of the GNU General Public License | |
18 | * along with this program; if not, write to the Free Software | |
19 | * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | |
20 | * | |
21 | * Author: Neil Brown | |
22 | * Email: <neilb@suse.de> | |
23 | */ | |
24 | #include "mdadm.h" | |
25 | #include "dlink.h" | |
26 | #include <sys/mman.h> | |
27 | ||
28 | #if ! defined(__BIG_ENDIAN) && ! defined(__LITTLE_ENDIAN) | |
29 | #error no endian defined | |
30 | #endif | |
31 | #include "md_u.h" | |
32 | #include "md_p.h" | |
33 | ||
34 | #ifndef offsetof | |
35 | #define offsetof(t,f) ((size_t)&(((t*)0)->f)) | |
36 | #endif | |
37 | ||
38 | int Grow_Add_device(char *devname, int fd, char *newdev) | |
39 | { | |
40 | /* Add a device to an active array. | |
41 | * Currently, just extend a linear array. | |
42 | * This requires writing a new superblock on the | |
43 | * new device, calling the kernel to add the device, | |
44 | * and if that succeeds, update the superblock on | |
45 | * all other devices. | |
46 | * This means that we need to *find* all other devices. | |
47 | */ | |
48 | struct mdinfo info; | |
49 | ||
50 | struct stat stb; | |
51 | int nfd, fd2; | |
52 | int d, nd; | |
53 | struct supertype *st = NULL; | |
54 | char *subarray = NULL; | |
55 | ||
56 | if (ioctl(fd, GET_ARRAY_INFO, &info.array) < 0) { | |
57 | fprintf(stderr, Name ": cannot get array info for %s\n", devname); | |
58 | return 1; | |
59 | } | |
60 | ||
61 | if (info.array.level != -1) { | |
62 | fprintf(stderr, Name ": can only add devices to linear arrays\n"); | |
63 | return 1; | |
64 | } | |
65 | ||
66 | st = super_by_fd(fd, &subarray); | |
67 | if (!st) { | |
68 | fprintf(stderr, Name ": cannot handle arrays with superblock version %d\n", info.array.major_version); | |
69 | return 1; | |
70 | } | |
71 | ||
72 | if (subarray) { | |
73 | fprintf(stderr, Name ": Cannot grow linear sub-arrays yet\n"); | |
74 | free(subarray); | |
75 | free(st); | |
76 | } | |
77 | ||
78 | nfd = open(newdev, O_RDWR|O_EXCL|O_DIRECT); | |
79 | if (nfd < 0) { | |
80 | fprintf(stderr, Name ": cannot open %s\n", newdev); | |
81 | free(st); | |
82 | return 1; | |
83 | } | |
84 | fstat(nfd, &stb); | |
85 | if ((stb.st_mode & S_IFMT) != S_IFBLK) { | |
86 | fprintf(stderr, Name ": %s is not a block device!\n", newdev); | |
87 | close(nfd); | |
88 | free(st); | |
89 | return 1; | |
90 | } | |
91 | /* now check out all the devices and make sure we can read the superblock */ | |
92 | for (d=0 ; d < info.array.raid_disks ; d++) { | |
93 | mdu_disk_info_t disk; | |
94 | char *dv; | |
95 | ||
96 | st->ss->free_super(st); | |
97 | ||
98 | disk.number = d; | |
99 | if (ioctl(fd, GET_DISK_INFO, &disk) < 0) { | |
100 | fprintf(stderr, Name ": cannot get device detail for device %d\n", | |
101 | d); | |
102 | close(nfd); | |
103 | free(st); | |
104 | return 1; | |
105 | } | |
106 | dv = map_dev(disk.major, disk.minor, 1); | |
107 | if (!dv) { | |
108 | fprintf(stderr, Name ": cannot find device file for device %d\n", | |
109 | d); | |
110 | close(nfd); | |
111 | free(st); | |
112 | return 1; | |
113 | } | |
114 | fd2 = dev_open(dv, O_RDWR); | |
115 | if (!fd2) { | |
116 | fprintf(stderr, Name ": cannot open device file %s\n", dv); | |
117 | close(nfd); | |
118 | free(st); | |
119 | return 1; | |
120 | } | |
121 | ||
122 | if (st->ss->load_super(st, fd2, NULL)) { | |
123 | fprintf(stderr, Name ": cannot find super block on %s\n", dv); | |
124 | close(nfd); | |
125 | close(fd2); | |
126 | free(st); | |
127 | return 1; | |
128 | } | |
129 | close(fd2); | |
130 | } | |
131 | /* Ok, looks good. Lets update the superblock and write it out to | |
132 | * newdev. | |
133 | */ | |
134 | ||
135 | info.disk.number = d; | |
136 | info.disk.major = major(stb.st_rdev); | |
137 | info.disk.minor = minor(stb.st_rdev); | |
138 | info.disk.raid_disk = d; | |
139 | info.disk.state = (1 << MD_DISK_SYNC) | (1 << MD_DISK_ACTIVE); | |
140 | st->ss->update_super(st, &info, "linear-grow-new", newdev, | |
141 | 0, 0, NULL); | |
142 | ||
143 | if (st->ss->store_super(st, nfd)) { | |
144 | fprintf(stderr, Name ": Cannot store new superblock on %s\n", | |
145 | newdev); | |
146 | close(nfd); | |
147 | return 1; | |
148 | } | |
149 | close(nfd); | |
150 | ||
151 | if (ioctl(fd, ADD_NEW_DISK, &info.disk) != 0) { | |
152 | fprintf(stderr, Name ": Cannot add new disk to this array\n"); | |
153 | return 1; | |
154 | } | |
155 | /* Well, that seems to have worked. | |
156 | * Now go through and update all superblocks | |
157 | */ | |
158 | ||
159 | if (ioctl(fd, GET_ARRAY_INFO, &info.array) < 0) { | |
160 | fprintf(stderr, Name ": cannot get array info for %s\n", devname); | |
161 | return 1; | |
162 | } | |
163 | ||
164 | nd = d; | |
165 | for (d=0 ; d < info.array.raid_disks ; d++) { | |
166 | mdu_disk_info_t disk; | |
167 | char *dv; | |
168 | ||
169 | disk.number = d; | |
170 | if (ioctl(fd, GET_DISK_INFO, &disk) < 0) { | |
171 | fprintf(stderr, Name ": cannot get device detail for device %d\n", | |
172 | d); | |
173 | return 1; | |
174 | } | |
175 | dv = map_dev(disk.major, disk.minor, 1); | |
176 | if (!dv) { | |
177 | fprintf(stderr, Name ": cannot find device file for device %d\n", | |
178 | d); | |
179 | return 1; | |
180 | } | |
181 | fd2 = dev_open(dv, O_RDWR); | |
182 | if (fd2 < 0) { | |
183 | fprintf(stderr, Name ": cannot open device file %s\n", dv); | |
184 | return 1; | |
185 | } | |
186 | if (st->ss->load_super(st, fd2, NULL)) { | |
187 | fprintf(stderr, Name ": cannot find super block on %s\n", dv); | |
188 | close(fd); | |
189 | return 1; | |
190 | } | |
191 | info.array.raid_disks = nd+1; | |
192 | info.array.nr_disks = nd+1; | |
193 | info.array.active_disks = nd+1; | |
194 | info.array.working_disks = nd+1; | |
195 | ||
196 | st->ss->update_super(st, &info, "linear-grow-update", dv, | |
197 | 0, 0, NULL); | |
198 | ||
199 | if (st->ss->store_super(st, fd2)) { | |
200 | fprintf(stderr, Name ": Cannot store new superblock on %s\n", dv); | |
201 | close(fd2); | |
202 | return 1; | |
203 | } | |
204 | close(fd2); | |
205 | } | |
206 | ||
207 | return 0; | |
208 | } | |
209 | ||
210 | int Grow_addbitmap(char *devname, int fd, char *file, int chunk, int delay, int write_behind, int force) | |
211 | { | |
212 | /* | |
213 | * First check that array doesn't have a bitmap | |
214 | * Then create the bitmap | |
215 | * Then add it | |
216 | * | |
217 | * For internal bitmaps, we need to check the version, | |
218 | * find all the active devices, and write the bitmap block | |
219 | * to all devices | |
220 | */ | |
221 | mdu_bitmap_file_t bmf; | |
222 | mdu_array_info_t array; | |
223 | struct supertype *st; | |
224 | char *subarray = NULL; | |
225 | int major = BITMAP_MAJOR_HI; | |
226 | int vers = md_get_version(fd); | |
227 | unsigned long long bitmapsize, array_size; | |
228 | ||
229 | if (vers < 9003) { | |
230 | major = BITMAP_MAJOR_HOSTENDIAN; | |
231 | fprintf(stderr, Name ": Warning - bitmaps created on this kernel" | |
232 | " are not portable\n" | |
233 | " between different architectures. Consider upgrading" | |
234 | " the Linux kernel.\n"); | |
235 | } | |
236 | ||
237 | if (ioctl(fd, GET_BITMAP_FILE, &bmf) != 0) { | |
238 | if (errno == ENOMEM) | |
239 | fprintf(stderr, Name ": Memory allocation failure.\n"); | |
240 | else | |
241 | fprintf(stderr, Name ": bitmaps not supported by this kernel.\n"); | |
242 | return 1; | |
243 | } | |
244 | if (bmf.pathname[0]) { | |
245 | if (strcmp(file,"none")==0) { | |
246 | if (ioctl(fd, SET_BITMAP_FILE, -1)!= 0) { | |
247 | fprintf(stderr, Name ": failed to remove bitmap %s\n", | |
248 | bmf.pathname); | |
249 | return 1; | |
250 | } | |
251 | return 0; | |
252 | } | |
253 | fprintf(stderr, Name ": %s already has a bitmap (%s)\n", | |
254 | devname, bmf.pathname); | |
255 | return 1; | |
256 | } | |
257 | if (ioctl(fd, GET_ARRAY_INFO, &array) != 0) { | |
258 | fprintf(stderr, Name ": cannot get array status for %s\n", devname); | |
259 | return 1; | |
260 | } | |
261 | if (array.state & (1<<MD_SB_BITMAP_PRESENT)) { | |
262 | if (strcmp(file, "none")==0) { | |
263 | array.state &= ~(1<<MD_SB_BITMAP_PRESENT); | |
264 | if (ioctl(fd, SET_ARRAY_INFO, &array)!= 0) { | |
265 | fprintf(stderr, Name ": failed to remove internal bitmap.\n"); | |
266 | return 1; | |
267 | } | |
268 | return 0; | |
269 | } | |
270 | fprintf(stderr, Name ": Internal bitmap already present on %s\n", | |
271 | devname); | |
272 | return 1; | |
273 | } | |
274 | ||
275 | if (strcmp(file, "none") == 0) { | |
276 | fprintf(stderr, Name ": no bitmap found on %s\n", devname); | |
277 | return 1; | |
278 | } | |
279 | if (array.level <= 0) { | |
280 | fprintf(stderr, Name ": Bitmaps not meaningful with level %s\n", | |
281 | map_num(pers, array.level)?:"of this array"); | |
282 | return 1; | |
283 | } | |
284 | bitmapsize = array.size; | |
285 | bitmapsize <<= 1; | |
286 | if (get_dev_size(fd, NULL, &array_size) && | |
287 | array_size > (0x7fffffffULL<<9)) { | |
288 | /* Array is big enough that we cannot trust array.size | |
289 | * try other approaches | |
290 | */ | |
291 | bitmapsize = get_component_size(fd); | |
292 | } | |
293 | if (bitmapsize == 0) { | |
294 | fprintf(stderr, Name ": Cannot reliably determine size of array to create bitmap - sorry.\n"); | |
295 | return 1; | |
296 | } | |
297 | ||
298 | if (array.level == 10) { | |
299 | int ncopies = (array.layout&255)*((array.layout>>8)&255); | |
300 | bitmapsize = bitmapsize * array.raid_disks / ncopies; | |
301 | } | |
302 | ||
303 | st = super_by_fd(fd, &subarray); | |
304 | if (!st) { | |
305 | fprintf(stderr, Name ": Cannot understand version %d.%d\n", | |
306 | array.major_version, array.minor_version); | |
307 | return 1; | |
308 | } | |
309 | if (subarray) { | |
310 | fprintf(stderr, Name ": Cannot add bitmaps to sub-arrays yet\n"); | |
311 | free(subarray); | |
312 | free(st); | |
313 | return 1; | |
314 | } | |
315 | if (strcmp(file, "internal") == 0) { | |
316 | int d; | |
317 | if (st->ss->add_internal_bitmap == NULL) { | |
318 | fprintf(stderr, Name ": Internal bitmaps not supported " | |
319 | "with %s metadata\n", st->ss->name); | |
320 | return 1; | |
321 | } | |
322 | for (d=0; d< st->max_devs; d++) { | |
323 | mdu_disk_info_t disk; | |
324 | char *dv; | |
325 | disk.number = d; | |
326 | if (ioctl(fd, GET_DISK_INFO, &disk) < 0) | |
327 | continue; | |
328 | if (disk.major == 0 && | |
329 | disk.minor == 0) | |
330 | continue; | |
331 | if ((disk.state & (1<<MD_DISK_SYNC))==0) | |
332 | continue; | |
333 | dv = map_dev(disk.major, disk.minor, 1); | |
334 | if (dv) { | |
335 | int fd2 = dev_open(dv, O_RDWR); | |
336 | if (fd2 < 0) | |
337 | continue; | |
338 | if (st->ss->load_super(st, fd2, NULL)==0) { | |
339 | if (st->ss->add_internal_bitmap( | |
340 | st, | |
341 | &chunk, delay, write_behind, | |
342 | bitmapsize, 0, major) | |
343 | ) | |
344 | st->ss->write_bitmap(st, fd2); | |
345 | else { | |
346 | fprintf(stderr, Name ": failed to create internal bitmap - chunksize problem.\n"); | |
347 | close(fd2); | |
348 | return 1; | |
349 | } | |
350 | } | |
351 | close(fd2); | |
352 | } | |
353 | } | |
354 | array.state |= (1<<MD_SB_BITMAP_PRESENT); | |
355 | if (ioctl(fd, SET_ARRAY_INFO, &array)!= 0) { | |
356 | if (errno == EBUSY) | |
357 | fprintf(stderr, Name | |
358 | ": Cannot add bitmap while array is" | |
359 | " resyncing or reshaping etc.\n"); | |
360 | fprintf(stderr, Name ": failed to set internal bitmap.\n"); | |
361 | return 1; | |
362 | } | |
363 | } else { | |
364 | int uuid[4]; | |
365 | int bitmap_fd; | |
366 | int d; | |
367 | int max_devs = st->max_devs; | |
368 | ||
369 | /* try to load a superblock */ | |
370 | for (d=0; d<max_devs; d++) { | |
371 | mdu_disk_info_t disk; | |
372 | char *dv; | |
373 | int fd2; | |
374 | disk.number = d; | |
375 | if (ioctl(fd, GET_DISK_INFO, &disk) < 0) | |
376 | continue; | |
377 | if ((disk.major==0 && disk.minor==0) || | |
378 | (disk.state & (1<<MD_DISK_REMOVED))) | |
379 | continue; | |
380 | dv = map_dev(disk.major, disk.minor, 1); | |
381 | if (!dv) continue; | |
382 | fd2 = dev_open(dv, O_RDONLY); | |
383 | if (fd2 >= 0 && | |
384 | st->ss->load_super(st, fd2, NULL) == 0) { | |
385 | close(fd2); | |
386 | st->ss->uuid_from_super(st, uuid); | |
387 | break; | |
388 | } | |
389 | close(fd2); | |
390 | } | |
391 | if (d == max_devs) { | |
392 | fprintf(stderr, Name ": cannot find UUID for array!\n"); | |
393 | return 1; | |
394 | } | |
395 | if (CreateBitmap(file, force, (char*)uuid, chunk, | |
396 | delay, write_behind, bitmapsize, major)) { | |
397 | return 1; | |
398 | } | |
399 | bitmap_fd = open(file, O_RDWR); | |
400 | if (bitmap_fd < 0) { | |
401 | fprintf(stderr, Name ": weird: %s cannot be opened\n", | |
402 | file); | |
403 | return 1; | |
404 | } | |
405 | if (ioctl(fd, SET_BITMAP_FILE, bitmap_fd) < 0) { | |
406 | int err = errno; | |
407 | if (errno == EBUSY) | |
408 | fprintf(stderr, Name | |
409 | ": Cannot add bitmap while array is" | |
410 | " resyncing or reshaping etc.\n"); | |
411 | fprintf(stderr, Name ": Cannot set bitmap file for %s: %s\n", | |
412 | devname, strerror(err)); | |
413 | return 1; | |
414 | } | |
415 | } | |
416 | ||
417 | return 0; | |
418 | } | |
419 | ||
420 | ||
421 | /* | |
422 | * When reshaping an array we might need to backup some data. | |
423 | * This is written to all spares with a 'super_block' describing it. | |
424 | * The superblock goes 4K from the end of the used space on the | |
425 | * device. | |
426 | * It if written after the backup is complete. | |
427 | * It has the following structure. | |
428 | */ | |
429 | ||
430 | static struct mdp_backup_super { | |
431 | char magic[16]; /* md_backup_data-1 or -2 */ | |
432 | __u8 set_uuid[16]; | |
433 | __u64 mtime; | |
434 | /* start/sizes in 512byte sectors */ | |
435 | __u64 devstart; /* address on backup device/file of data */ | |
436 | __u64 arraystart; | |
437 | __u64 length; | |
438 | __u32 sb_csum; /* csum of preceeding bytes. */ | |
439 | __u32 pad1; | |
440 | __u64 devstart2; /* offset in to data of second section */ | |
441 | __u64 arraystart2; | |
442 | __u64 length2; | |
443 | __u32 sb_csum2; /* csum of preceeding bytes. */ | |
444 | __u8 pad[512-68-32]; | |
445 | } __attribute__((aligned(512))) bsb, bsb2; | |
446 | ||
447 | static __u32 bsb_csum(char *buf, int len) | |
448 | { | |
449 | int i; | |
450 | int csum = 0; | |
451 | for (i=0; i<len; i++) | |
452 | csum = (csum<<3) + buf[0]; | |
453 | return __cpu_to_le32(csum); | |
454 | } | |
455 | ||
456 | static int check_idle(struct supertype *st) | |
457 | { | |
458 | /* Check that all member arrays for this container, or the | |
459 | * container of this array, are idle | |
460 | */ | |
461 | int container_dev = (st->container_dev != NoMdDev | |
462 | ? st->container_dev : st->devnum); | |
463 | char container[40]; | |
464 | struct mdstat_ent *ent, *e; | |
465 | int is_idle = 1; | |
466 | ||
467 | fmt_devname(container, container_dev); | |
468 | ent = mdstat_read(0, 0); | |
469 | for (e = ent ; e; e = e->next) { | |
470 | if (!is_container_member(e, container)) | |
471 | continue; | |
472 | if (e->percent >= 0) { | |
473 | is_idle = 0; | |
474 | break; | |
475 | } | |
476 | } | |
477 | free_mdstat(ent); | |
478 | return is_idle; | |
479 | } | |
480 | ||
481 | static int freeze_container(struct supertype *st) | |
482 | { | |
483 | int container_dev = (st->container_dev != NoMdDev | |
484 | ? st->container_dev : st->devnum); | |
485 | char container[40]; | |
486 | ||
487 | if (!check_idle(st)) | |
488 | return -1; | |
489 | ||
490 | fmt_devname(container, container_dev); | |
491 | ||
492 | if (block_monitor(container, 1)) { | |
493 | fprintf(stderr, Name ": failed to freeze container\n"); | |
494 | return -2; | |
495 | } | |
496 | ||
497 | return 1; | |
498 | } | |
499 | ||
500 | static void unfreeze_container(struct supertype *st) | |
501 | { | |
502 | int container_dev = (st->container_dev != NoMdDev | |
503 | ? st->container_dev : st->devnum); | |
504 | char container[40]; | |
505 | ||
506 | fmt_devname(container, container_dev); | |
507 | ||
508 | unblock_monitor(container, 1); | |
509 | } | |
510 | ||
511 | static int freeze(struct supertype *st) | |
512 | { | |
513 | /* Try to freeze resync/rebuild on this array/container. | |
514 | * Return -1 if the array is busy, | |
515 | * return -2 container cannot be frozen, | |
516 | * return 0 if this kernel doesn't support 'frozen' | |
517 | * return 1 if it worked. | |
518 | */ | |
519 | if (st->ss->external) | |
520 | return freeze_container(st); | |
521 | else { | |
522 | struct mdinfo *sra = sysfs_read(-1, st->devnum, GET_VERSION); | |
523 | int err; | |
524 | ||
525 | if (!sra) | |
526 | return -1; | |
527 | err = sysfs_freeze_array(sra); | |
528 | sysfs_free(sra); | |
529 | return err; | |
530 | } | |
531 | } | |
532 | ||
533 | static void unfreeze(struct supertype *st) | |
534 | { | |
535 | if (st->ss->external) | |
536 | return unfreeze_container(st); | |
537 | else { | |
538 | struct mdinfo *sra = sysfs_read(-1, st->devnum, GET_VERSION); | |
539 | ||
540 | if (sra) | |
541 | sysfs_set_str(sra, NULL, "sync_action", "idle"); | |
542 | else | |
543 | fprintf(stderr, Name ": failed to unfreeze array\n"); | |
544 | sysfs_free(sra); | |
545 | } | |
546 | } | |
547 | ||
548 | static void wait_reshape(struct mdinfo *sra) | |
549 | { | |
550 | int fd = sysfs_get_fd(sra, NULL, "sync_action"); | |
551 | char action[20]; | |
552 | ||
553 | if (fd < 0) | |
554 | return; | |
555 | ||
556 | while (sysfs_fd_get_str(fd, action, 20) > 0 && | |
557 | strncmp(action, "reshape", 7) == 0) { | |
558 | fd_set rfds; | |
559 | FD_ZERO(&rfds); | |
560 | FD_SET(fd, &rfds); | |
561 | select(fd+1, NULL, NULL, &rfds, NULL); | |
562 | } | |
563 | close(fd); | |
564 | } | |
565 | ||
566 | static int reshape_super(struct supertype *st, long long size, int level, | |
567 | int layout, int chunksize, int raid_disks, | |
568 | char *backup_file, char *dev, int verbose) | |
569 | { | |
570 | /* nothing extra to check in the native case */ | |
571 | if (!st->ss->external) | |
572 | return 0; | |
573 | if (!st->ss->reshape_super || | |
574 | !st->ss->manage_reshape) { | |
575 | fprintf(stderr, Name ": %s metadata does not support reshape\n", | |
576 | st->ss->name); | |
577 | return 1; | |
578 | } | |
579 | ||
580 | return st->ss->reshape_super(st, size, level, layout, chunksize, | |
581 | raid_disks, backup_file, dev, verbose); | |
582 | } | |
583 | ||
584 | static void sync_metadata(struct supertype *st) | |
585 | { | |
586 | if (st->ss->external) { | |
587 | if (st->update_tail) { | |
588 | flush_metadata_updates(st); | |
589 | st->update_tail = &st->updates; | |
590 | } else | |
591 | st->ss->sync_metadata(st); | |
592 | } | |
593 | } | |
594 | ||
595 | static int subarray_set_num(char *container, struct mdinfo *sra, char *name, int n) | |
596 | { | |
597 | /* when dealing with external metadata subarrays we need to be | |
598 | * prepared to handle EAGAIN. The kernel may need to wait for | |
599 | * mdmon to mark the array active so the kernel can handle | |
600 | * allocations/writeback when preparing the reshape action | |
601 | * (md_allow_write()). We temporarily disable safe_mode_delay | |
602 | * to close a race with the array_state going clean before the | |
603 | * next write to raid_disks / stripe_cache_size | |
604 | */ | |
605 | char safe[50]; | |
606 | int rc; | |
607 | ||
608 | /* only 'raid_disks' and 'stripe_cache_size' trigger md_allow_write */ | |
609 | if (!container || | |
610 | (strcmp(name, "raid_disks") != 0 && | |
611 | strcmp(name, "stripe_cache_size") != 0)) | |
612 | return sysfs_set_num(sra, NULL, name, n); | |
613 | ||
614 | rc = sysfs_get_str(sra, NULL, "safe_mode_delay", safe, sizeof(safe)); | |
615 | if (rc <= 0) | |
616 | return -1; | |
617 | sysfs_set_num(sra, NULL, "safe_mode_delay", 0); | |
618 | rc = sysfs_set_num(sra, NULL, name, n); | |
619 | if (rc < 0 && errno == EAGAIN) { | |
620 | ping_monitor(container); | |
621 | /* if we get EAGAIN here then the monitor is not active | |
622 | * so stop trying | |
623 | */ | |
624 | rc = sysfs_set_num(sra, NULL, name, n); | |
625 | } | |
626 | sysfs_set_str(sra, NULL, "safe_mode_delay", safe); | |
627 | return rc; | |
628 | } | |
629 | ||
630 | int start_reshape(struct mdinfo *sra, int already_running) | |
631 | { | |
632 | int err; | |
633 | sysfs_set_num(sra, NULL, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL); | |
634 | err = sysfs_set_num(sra, NULL, "suspend_hi", 0); | |
635 | err = err ?: sysfs_set_num(sra, NULL, "suspend_lo", 0); | |
636 | if (!already_running) | |
637 | sysfs_set_num(sra, NULL, "sync_min", 0); | |
638 | err = err ?: sysfs_set_num(sra, NULL, "sync_max", 0); | |
639 | if (!already_running) | |
640 | err = err ?: sysfs_set_str(sra, NULL, "sync_action", "reshape"); | |
641 | ||
642 | return err; | |
643 | } | |
644 | ||
645 | void abort_reshape(struct mdinfo *sra) | |
646 | { | |
647 | sysfs_set_str(sra, NULL, "sync_action", "idle"); | |
648 | sysfs_set_num(sra, NULL, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL); | |
649 | sysfs_set_num(sra, NULL, "suspend_hi", 0); | |
650 | sysfs_set_num(sra, NULL, "suspend_lo", 0); | |
651 | sysfs_set_num(sra, NULL, "sync_min", 0); | |
652 | sysfs_set_str(sra, NULL, "sync_max", "max"); | |
653 | } | |
654 | ||
655 | int remove_disks_for_takeover(struct supertype *st, | |
656 | struct mdinfo *sra, | |
657 | int layout) | |
658 | { | |
659 | int nr_of_copies; | |
660 | struct mdinfo *remaining; | |
661 | int slot; | |
662 | ||
663 | if (sra->array.level == 10) | |
664 | nr_of_copies = layout & 0xff; | |
665 | else if (sra->array.level == 1) | |
666 | nr_of_copies = sra->array.raid_disks; | |
667 | else | |
668 | return 1; | |
669 | ||
670 | remaining = sra->devs; | |
671 | sra->devs = NULL; | |
672 | /* for each 'copy', select one device and remove from the list. */ | |
673 | for (slot = 0; slot < sra->array.raid_disks; slot += nr_of_copies) { | |
674 | struct mdinfo **diskp; | |
675 | int found = 0; | |
676 | ||
677 | /* Find a working device to keep */ | |
678 | for (diskp = &remaining; *diskp ; diskp = &(*diskp)->next) { | |
679 | struct mdinfo *disk = *diskp; | |
680 | ||
681 | if (disk->disk.raid_disk < slot) | |
682 | continue; | |
683 | if (disk->disk.raid_disk >= slot + nr_of_copies) | |
684 | continue; | |
685 | if (disk->disk.state & (1<<MD_DISK_REMOVED)) | |
686 | continue; | |
687 | if (disk->disk.state & (1<<MD_DISK_FAULTY)) | |
688 | continue; | |
689 | if (!(disk->disk.state & (1<<MD_DISK_SYNC))) | |
690 | continue; | |
691 | ||
692 | /* We have found a good disk to use! */ | |
693 | *diskp = disk->next; | |
694 | disk->next = sra->devs; | |
695 | sra->devs = disk; | |
696 | found = 1; | |
697 | break; | |
698 | } | |
699 | if (!found) | |
700 | break; | |
701 | } | |
702 | ||
703 | if (slot < sra->array.raid_disks) { | |
704 | /* didn't find all slots */ | |
705 | struct mdinfo **e; | |
706 | e = &remaining; | |
707 | while (*e) | |
708 | e = &(*e)->next; | |
709 | *e = sra->devs; | |
710 | sra->devs = remaining; | |
711 | return 1; | |
712 | } | |
713 | ||
714 | /* Remove all 'remaining' devices from the array */ | |
715 | while (remaining) { | |
716 | struct mdinfo *sd = remaining; | |
717 | remaining = sd->next; | |
718 | ||
719 | sysfs_set_str(sra, sd, "state", "faulty"); | |
720 | sysfs_set_str(sra, sd, "slot", "none"); | |
721 | /* for external metadata disks should be removed in mdmon */ | |
722 | if (!st->ss->external) | |
723 | sysfs_set_str(sra, sd, "state", "remove"); | |
724 | sd->disk.state |= (1<<MD_DISK_REMOVED); | |
725 | sd->disk.state &= ~(1<<MD_DISK_SYNC); | |
726 | sd->next = sra->devs; | |
727 | sra->devs = sd; | |
728 | } | |
729 | return 0; | |
730 | } | |
731 | ||
732 | void reshape_free_fdlist(int *fdlist, | |
733 | unsigned long long *offsets, | |
734 | int size) | |
735 | { | |
736 | int i; | |
737 | ||
738 | for (i = 0; i < size; i++) | |
739 | if (fdlist[i] >= 0) | |
740 | close(fdlist[i]); | |
741 | ||
742 | free(fdlist); | |
743 | free(offsets); | |
744 | } | |
745 | ||
746 | int reshape_prepare_fdlist(char *devname, | |
747 | struct mdinfo *sra, | |
748 | int raid_disks, | |
749 | int nrdisks, | |
750 | unsigned long blocks, | |
751 | char *backup_file, | |
752 | int *fdlist, | |
753 | unsigned long long *offsets) | |
754 | { | |
755 | int d = 0; | |
756 | struct mdinfo *sd; | |
757 | ||
758 | for (d = 0; d <= nrdisks; d++) | |
759 | fdlist[d] = -1; | |
760 | d = raid_disks; | |
761 | for (sd = sra->devs; sd; sd = sd->next) { | |
762 | if (sd->disk.state & (1<<MD_DISK_FAULTY)) | |
763 | continue; | |
764 | if (sd->disk.state & (1<<MD_DISK_SYNC)) { | |
765 | char *dn = map_dev(sd->disk.major, | |
766 | sd->disk.minor, 1); | |
767 | fdlist[sd->disk.raid_disk] | |
768 | = dev_open(dn, O_RDONLY); | |
769 | offsets[sd->disk.raid_disk] = sd->data_offset*512; | |
770 | if (fdlist[sd->disk.raid_disk] < 0) { | |
771 | fprintf(stderr, | |
772 | Name ": %s: cannot open component %s\n", | |
773 | devname, dn ? dn : "-unknown-"); | |
774 | d = -1; | |
775 | goto release; | |
776 | } | |
777 | } else if (backup_file == NULL) { | |
778 | /* spare */ | |
779 | char *dn = map_dev(sd->disk.major, | |
780 | sd->disk.minor, 1); | |
781 | fdlist[d] = dev_open(dn, O_RDWR); | |
782 | offsets[d] = (sd->data_offset + sra->component_size - blocks - 8)*512; | |
783 | if (fdlist[d] < 0) { | |
784 | fprintf(stderr, Name ": %s: cannot open component %s\n", | |
785 | devname, dn ? dn : "-unknown-"); | |
786 | d = -1; | |
787 | goto release; | |
788 | } | |
789 | d++; | |
790 | } | |
791 | } | |
792 | release: | |
793 | return d; | |
794 | } | |
795 | ||
796 | int reshape_open_backup_file(char *backup_file, | |
797 | int fd, | |
798 | char *devname, | |
799 | long blocks, | |
800 | int *fdlist, | |
801 | unsigned long long *offsets, | |
802 | int restart) | |
803 | { | |
804 | /* Return 1 on success, 0 on any form of failure */ | |
805 | /* need to check backup file is large enough */ | |
806 | char buf[512]; | |
807 | struct stat stb; | |
808 | unsigned int dev; | |
809 | int i; | |
810 | ||
811 | *fdlist = open(backup_file, O_RDWR|O_CREAT|(restart ? O_TRUNC : O_EXCL), | |
812 | S_IRUSR | S_IWUSR); | |
813 | *offsets = 8 * 512; | |
814 | if (*fdlist < 0) { | |
815 | fprintf(stderr, Name ": %s: cannot create backup file %s: %s\n", | |
816 | devname, backup_file, strerror(errno)); | |
817 | return 0; | |
818 | } | |
819 | /* Guard against backup file being on array device. | |
820 | * If array is partitioned or if LVM etc is in the | |
821 | * way this will not notice, but it is better than | |
822 | * nothing. | |
823 | */ | |
824 | fstat(*fdlist, &stb); | |
825 | dev = stb.st_dev; | |
826 | fstat(fd, &stb); | |
827 | if (stb.st_rdev == dev) { | |
828 | fprintf(stderr, Name ": backup file must NOT be" | |
829 | " on the array being reshaped.\n"); | |
830 | close(*fdlist); | |
831 | return 0; | |
832 | } | |
833 | ||
834 | memset(buf, 0, 512); | |
835 | for (i=0; i < blocks + 8 ; i++) { | |
836 | if (write(*fdlist, buf, 512) != 512) { | |
837 | fprintf(stderr, Name ": %s: cannot create" | |
838 | " backup file %s: %s\n", | |
839 | devname, backup_file, strerror(errno)); | |
840 | return 0; | |
841 | } | |
842 | } | |
843 | if (fsync(*fdlist) != 0) { | |
844 | fprintf(stderr, Name ": %s: cannot create backup file %s: %s\n", | |
845 | devname, backup_file, strerror(errno)); | |
846 | return 0; | |
847 | } | |
848 | ||
849 | return 1; | |
850 | } | |
851 | ||
852 | unsigned long compute_backup_blocks(int nchunk, int ochunk, | |
853 | unsigned int ndata, unsigned int odata) | |
854 | { | |
855 | unsigned long a, b, blocks; | |
856 | /* So how much do we need to backup. | |
857 | * We need an amount of data which is both a whole number of | |
858 | * old stripes and a whole number of new stripes. | |
859 | * So LCM for (chunksize*datadisks). | |
860 | */ | |
861 | a = (ochunk/512) * odata; | |
862 | b = (nchunk/512) * ndata; | |
863 | /* Find GCD */ | |
864 | while (a != b) { | |
865 | if (a < b) | |
866 | b -= a; | |
867 | if (b < a) | |
868 | a -= b; | |
869 | } | |
870 | /* LCM == product / GCD */ | |
871 | blocks = (ochunk/512) * (nchunk/512) * odata * ndata / a; | |
872 | ||
873 | return blocks; | |
874 | } | |
875 | ||
876 | char *analyse_change(struct mdinfo *info, struct reshape *re) | |
877 | { | |
878 | /* Based on the current array state in info->array and | |
879 | * the changes in info->new_* etc, determine: | |
880 | * - whether the change is possible | |
881 | * - Intermediate level/raid_disks/layout | |
882 | * - whether a restriping reshape is needed | |
883 | * - number of sectors in minimum change unit. This | |
884 | * will cover a whole number of stripes in 'before' and | |
885 | * 'after'. | |
886 | * | |
887 | * Return message if the change should be rejected | |
888 | * NULL if the change can be achieved | |
889 | * | |
890 | * This can be called as part of starting a reshape, or | |
891 | * when assembling an array that is undergoing reshape. | |
892 | */ | |
893 | int new_disks; | |
894 | ||
895 | /* If a new level not explicitly given, we assume no-change */ | |
896 | if (info->new_level == UnSet) | |
897 | info->new_level = info->array.level; | |
898 | ||
899 | if (info->new_chunk) | |
900 | switch (info->new_level) { | |
901 | case 0: | |
902 | case 4: | |
903 | case 5: | |
904 | case 6: | |
905 | case 10: | |
906 | /* chunk size is meaningful, must divide component_size | |
907 | * evenly | |
908 | */ | |
909 | if (info->component_size % (info->new_chunk/512)) | |
910 | return "New chunk size does not" | |
911 | " divide component size"; | |
912 | break; | |
913 | default: | |
914 | return "chunk size not meaningful for this level"; | |
915 | } | |
916 | else | |
917 | info->new_chunk = info->array.chunk_size; | |
918 | ||
919 | switch (info->array.level) { | |
920 | case 1: | |
921 | /* RAID1 can convert to RAID1 with different disks, or | |
922 | * raid5 with 2 disks, or | |
923 | * raid0 with 1 disk | |
924 | */ | |
925 | if (info->new_level == 0) { | |
926 | re->level = 0; | |
927 | re->before.data_disks = 1; | |
928 | re->after.data_disks = 1; | |
929 | re->before.layout = 0; | |
930 | re->backup_blocks = 0; | |
931 | re->parity = 0; | |
932 | return NULL; | |
933 | } | |
934 | if (info->new_level == 1) { | |
935 | if (info->delta_disks == UnSet) | |
936 | /* Don't know what to do */ | |
937 | return "no change requested for Growing RAID1"; | |
938 | re->level = 1; | |
939 | re->backup_blocks = 0; | |
940 | re->parity = 0; | |
941 | return NULL; | |
942 | } | |
943 | if (info->array.raid_disks == 2 && | |
944 | info->new_level == 5) { | |
945 | re->level = 5; | |
946 | re->before.data_disks = 1; | |
947 | re->after.data_disks = 1; | |
948 | re->before.layout = ALGORITHM_LEFT_SYMMETRIC; | |
949 | info->array.chunk_size = 65536; | |
950 | break; | |
951 | } | |
952 | /* Could do some multi-stage conversions, but leave that to | |
953 | * later. | |
954 | */ | |
955 | return "Impossibly level change request for RAID1"; | |
956 | ||
957 | case 10: | |
958 | /* RAID10 can only be converted from near mode to | |
959 | * RAID0 by removing some devices | |
960 | */ | |
961 | if ((info->array.layout & ~0xff) != 0x100) | |
962 | return "Cannot Grow RAID10 with far/offset layout"; | |
963 | /* number of devices must be multiple of number of copies */ | |
964 | if (info->array.raid_disks % (info->array.layout & 0xff)) | |
965 | return "RAID10 layout too complex for Grow operation"; | |
966 | ||
967 | if (info->new_level != 0) | |
968 | return "RAID10 can only be changed to RAID0"; | |
969 | new_disks = (info->array.raid_disks | |
970 | / (info->array.layout & 0xff)); | |
971 | if (info->delta_disks == UnSet) { | |
972 | info->delta_disks = (new_disks | |
973 | - info->array.raid_disks); | |
974 | } | |
975 | if (info->delta_disks != new_disks - info->array.raid_disks) | |
976 | return "New number of raid-devices impossible for RAID10"; | |
977 | if (info->new_chunk && | |
978 | info->new_chunk != info->array.chunk_size) | |
979 | return "Cannot change chunk-size with RAID10 Grow"; | |
980 | ||
981 | /* looks good */ | |
982 | re->level = 0; | |
983 | re->parity = 0; | |
984 | re->before.data_disks = new_disks; | |
985 | re->after.data_disks = re->before.data_disks; | |
986 | re->before.layout = 0; | |
987 | re->backup_blocks = 0; | |
988 | return NULL; | |
989 | ||
990 | case 0: | |
991 | /* RAID0 can be converted to RAID10, or to RAID456 */ | |
992 | if (info->new_level == 10) { | |
993 | if (info->new_layout == UnSet && info->delta_disks == UnSet) { | |
994 | /* Assume near=2 layout */ | |
995 | info->new_layout = 0x102; | |
996 | info->delta_disks = info->array.raid_disks; | |
997 | } | |
998 | if (info->new_layout == UnSet) { | |
999 | int copies = 1 + (info->delta_disks | |
1000 | / info->array.raid_disks); | |
1001 | if (info->array.raid_disks * (copies-1) | |
1002 | != info->delta_disks) | |
1003 | return "Impossible number of devices" | |
1004 | " for RAID0->RAID10"; | |
1005 | info->new_layout = 0x100 + copies; | |
1006 | } | |
1007 | if (info->delta_disks == UnSet) { | |
1008 | int copies = info->new_layout & 0xff; | |
1009 | if (info->new_layout != 0x100 + copies) | |
1010 | return "New layout impossible" | |
1011 | " for RAID0->RAID10";; | |
1012 | info->delta_disks = (copies - 1) * | |
1013 | info->array.raid_disks; | |
1014 | } | |
1015 | if (info->new_chunk && | |
1016 | info->new_chunk != info->array.chunk_size) | |
1017 | return "Cannot change chunk-size with RAID0->RAID10"; | |
1018 | /* looks good */ | |
1019 | re->level = 10; | |
1020 | re->parity = 0; | |
1021 | re->before.data_disks = (info->array.raid_disks + | |
1022 | info->delta_disks); | |
1023 | re->after.data_disks = re->before.data_disks; | |
1024 | re->before.layout = info->new_layout; | |
1025 | re->backup_blocks = 0; | |
1026 | return NULL; | |
1027 | } | |
1028 | ||
1029 | /* RAID0 can also covert to RAID0/4/5/6 by first converting to | |
1030 | * a raid4 style layout of the final level. | |
1031 | */ | |
1032 | switch (info->new_level) { | |
1033 | case 0: | |
1034 | case 4: | |
1035 | re->level = 4; | |
1036 | re->before.layout = 0; | |
1037 | break; | |
1038 | case 5: | |
1039 | re->level = 5; | |
1040 | re->before.layout = ALGORITHM_PARITY_N; | |
1041 | break; | |
1042 | case 6: | |
1043 | re->level = 6; | |
1044 | re->before.layout = ALGORITHM_PARITY_N; | |
1045 | break; | |
1046 | default: | |
1047 | return "Impossible level change requested"; | |
1048 | } | |
1049 | re->before.data_disks = info->array.raid_disks; | |
1050 | /* determining 'after' layout happens outside this 'switch' */ | |
1051 | break; | |
1052 | ||
1053 | case 4: | |
1054 | info->array.layout = ALGORITHM_PARITY_N; | |
1055 | case 5: | |
1056 | switch (info->new_level) { | |
1057 | case 4: | |
1058 | re->level = info->array.level; | |
1059 | re->before.data_disks = info->array.raid_disks - 1; | |
1060 | re->before.layout = info->array.layout; | |
1061 | break; | |
1062 | case 5: | |
1063 | re->level = 5; | |
1064 | re->before.data_disks = info->array.raid_disks - 1; | |
1065 | re->before.layout = info->array.layout; | |
1066 | break; | |
1067 | case 6: | |
1068 | re->level = 6; | |
1069 | re->before.data_disks = info->array.raid_disks - 1; | |
1070 | switch (info->array.layout) { | |
1071 | case ALGORITHM_LEFT_ASYMMETRIC: | |
1072 | re->before.layout = ALGORITHM_LEFT_ASYMMETRIC_6; | |
1073 | break; | |
1074 | case ALGORITHM_RIGHT_ASYMMETRIC: | |
1075 | re->before.layout = ALGORITHM_RIGHT_ASYMMETRIC_6; | |
1076 | break; | |
1077 | case ALGORITHM_LEFT_SYMMETRIC: | |
1078 | re->before.layout = ALGORITHM_LEFT_SYMMETRIC_6; | |
1079 | break; | |
1080 | case ALGORITHM_RIGHT_SYMMETRIC: | |
1081 | re->before.layout = ALGORITHM_RIGHT_SYMMETRIC_6; | |
1082 | break; | |
1083 | case ALGORITHM_PARITY_0: | |
1084 | re->before.layout = ALGORITHM_PARITY_0_6; | |
1085 | break; | |
1086 | case ALGORITHM_PARITY_N: | |
1087 | re->before.layout = ALGORITHM_PARITY_N_6; | |
1088 | break; | |
1089 | default: | |
1090 | return "Cannot convert an array with this layout"; | |
1091 | } | |
1092 | break; | |
1093 | case 1: | |
1094 | if (info->array.raid_disks != 2) | |
1095 | return "Can only convert a 2-device array to RAID1"; | |
1096 | re->level = 1; | |
1097 | break; | |
1098 | default: | |
1099 | return "Impossible level change requested"; | |
1100 | } | |
1101 | break; | |
1102 | case 6: | |
1103 | switch (info->new_level) { | |
1104 | case 4: | |
1105 | case 5: | |
1106 | case 6: | |
1107 | re->level = 6; | |
1108 | re->before.data_disks = info->array.raid_disks - 2; | |
1109 | re->before.layout = info->array.layout; | |
1110 | break; | |
1111 | default: | |
1112 | return "Impossible level change requested"; | |
1113 | } | |
1114 | break; | |
1115 | } | |
1116 | ||
1117 | /* If we reached here then it looks like a re-stripe is | |
1118 | * happening. We have determined the intermediate level | |
1119 | * and initial raid_disks/layout and stored these in 're'. | |
1120 | * | |
1121 | * We need to deduce the final layout that can be atomically | |
1122 | * converted to the end state. | |
1123 | */ | |
1124 | switch (info->new_level) { | |
1125 | case 0: | |
1126 | /* We can only get to RAID0 from RAID4 or RAID5 | |
1127 | * with appropriate layout and one extra device | |
1128 | */ | |
1129 | if (re->level != 4 && re->level != 5) | |
1130 | return "Cannot covert to RAID0 from this level"; | |
1131 | if (info->delta_disks == UnSet) | |
1132 | re->after.data_disks = re->before.data_disks; | |
1133 | else | |
1134 | re->after.data_disks = | |
1135 | info->array.raid_disks + info->delta_disks; | |
1136 | switch (re->level) { | |
1137 | case 4: | |
1138 | re->after.layout = 0 ; break; | |
1139 | case 5: | |
1140 | re->after.layout = ALGORITHM_PARITY_N; break; | |
1141 | } | |
1142 | break; | |
1143 | ||
1144 | case 4: | |
1145 | /* We can only get to RAID4 from RAID5 */ | |
1146 | if (re->level != 4 && re->level != 5) | |
1147 | return "Cannot convert to RAID4 from this level"; | |
1148 | if (info->delta_disks == UnSet) | |
1149 | re->after.data_disks = re->before.data_disks; | |
1150 | else | |
1151 | re->after.data_disks = | |
1152 | re->before.data_disks + info->delta_disks; | |
1153 | switch (re->level) { | |
1154 | case 4: | |
1155 | re->after.layout = 0 ; break; | |
1156 | case 5: | |
1157 | re->after.layout = ALGORITHM_PARITY_N; break; | |
1158 | } | |
1159 | break; | |
1160 | ||
1161 | case 5: | |
1162 | /* We get to RAID5 for RAID5 or RAID6 */ | |
1163 | if (re->level != 5 && re->level != 6) | |
1164 | return "Cannot convert to RAID5 from this level"; | |
1165 | if (info->delta_disks == UnSet) | |
1166 | re->after.data_disks = re->before.data_disks; | |
1167 | else if (re->level == 5) | |
1168 | re->after.data_disks = | |
1169 | re->before.data_disks + info->delta_disks; | |
1170 | else | |
1171 | re->after.data_disks = | |
1172 | info->array.raid_disks + info->delta_disks - 1; | |
1173 | switch (re->level) { | |
1174 | case 5: | |
1175 | if (info->new_layout == UnSet) | |
1176 | re->after.layout = re->before.layout; | |
1177 | else | |
1178 | re->after.layout = info->new_layout; | |
1179 | break; | |
1180 | case 6: | |
1181 | if (info->new_layout == UnSet) | |
1182 | info->new_layout = re->before.layout; | |
1183 | ||
1184 | /* after.layout needs to be raid6 version of new_layout */ | |
1185 | if (info->new_layout == ALGORITHM_PARITY_N) | |
1186 | re->after.layout = ALGORITHM_PARITY_N; | |
1187 | else { | |
1188 | char layout[40]; | |
1189 | char *ls = map_num(r5layout, info->new_layout); | |
1190 | int l; | |
1191 | strcat(strcpy(layout, ls), "-6"); | |
1192 | l = map_name(r6layout, layout); | |
1193 | if (l == UnSet) | |
1194 | return "Cannot find RAID6 layout" | |
1195 | " to convert to"; | |
1196 | re->after.layout = l; | |
1197 | } | |
1198 | } | |
1199 | break; | |
1200 | ||
1201 | case 6: | |
1202 | /* We must already be at level 6 */ | |
1203 | if (re->level != 6) | |
1204 | return "Impossible level change"; | |
1205 | if (info->delta_disks == UnSet) | |
1206 | re->after.data_disks = re->before.data_disks; | |
1207 | else | |
1208 | re->after.data_disks = (info->array.raid_disks + | |
1209 | info->delta_disks) - 2; | |
1210 | if (info->new_layout == UnSet) | |
1211 | re->after.layout = info->array.layout; | |
1212 | else | |
1213 | re->after.layout = info->new_layout; | |
1214 | break; | |
1215 | default: | |
1216 | return "Impossible level change requested"; | |
1217 | } | |
1218 | switch (re->level) { | |
1219 | case 6: re->parity = 2; break; | |
1220 | case 4: | |
1221 | case 5: re->parity = 1; break; | |
1222 | default: re->parity = 0; break; | |
1223 | } | |
1224 | /* So we have a restripe operation, we need to calculate the number | |
1225 | * of blocks per reshape operation. | |
1226 | */ | |
1227 | if (info->new_chunk == 0) | |
1228 | info->new_chunk = info->array.chunk_size; | |
1229 | if (re->after.data_disks == re->before.data_disks && | |
1230 | re->after.layout == re->before.layout && | |
1231 | info->new_chunk == info->array.chunk_size) { | |
1232 | /* Nothing to change */ | |
1233 | re->backup_blocks = 0; | |
1234 | return NULL; | |
1235 | } | |
1236 | if (re->after.data_disks == 1 && re->before.data_disks == 1) { | |
1237 | /* chunk and layout changes make no difference */ | |
1238 | re->backup_blocks = 0; | |
1239 | return NULL; | |
1240 | } | |
1241 | ||
1242 | if (re->after.data_disks == re->before.data_disks && | |
1243 | get_linux_version() < 2006032) | |
1244 | return "in-place reshape is not safe before 2.6.32 - sorry."; | |
1245 | ||
1246 | if (re->after.data_disks < re->before.data_disks && | |
1247 | get_linux_version() < 2006030) | |
1248 | return "reshape to fewer devices is not supported before 2.6.32 - sorry."; | |
1249 | ||
1250 | re->backup_blocks = compute_backup_blocks( | |
1251 | info->new_chunk, info->array.chunk_size, | |
1252 | re->after.data_disks, | |
1253 | re->before.data_disks); | |
1254 | ||
1255 | re->new_size = info->component_size * re->after.data_disks; | |
1256 | return NULL; | |
1257 | } | |
1258 | ||
1259 | static int reshape_array(char *container, int fd, char *devname, | |
1260 | struct supertype *st, struct mdinfo *info, | |
1261 | int force, char *backup_file, int quiet, int forked, | |
1262 | int restart); | |
1263 | static int reshape_container(char *container, int cfd, char *devname, | |
1264 | struct supertype *st, | |
1265 | struct mdinfo *info, | |
1266 | int force, | |
1267 | char *backup_file, | |
1268 | int quiet); | |
1269 | ||
1270 | int Grow_reshape(char *devname, int fd, int quiet, char *backup_file, | |
1271 | long long size, | |
1272 | int level, char *layout_str, int chunksize, int raid_disks, | |
1273 | int force) | |
1274 | { | |
1275 | /* Make some changes in the shape of an array. | |
1276 | * The kernel must support the change. | |
1277 | * | |
1278 | * There are three different changes. Each can trigger | |
1279 | * a resync or recovery so we freeze that until we have | |
1280 | * requested everything (if kernel supports freezing - 2.6.30). | |
1281 | * The steps are: | |
1282 | * - change size (i.e. component_size) | |
1283 | * - change level | |
1284 | * - change layout/chunksize/ndisks | |
1285 | * | |
1286 | * The last can require a reshape. It is different on different | |
1287 | * levels so we need to check the level before actioning it. | |
1288 | * Some times the level change needs to be requested after the | |
1289 | * reshape (e.g. raid6->raid5, raid5->raid0) | |
1290 | * | |
1291 | */ | |
1292 | struct mdu_array_info_s array; | |
1293 | int rv = 0; | |
1294 | struct supertype *st; | |
1295 | char *subarray = NULL; | |
1296 | ||
1297 | int frozen; | |
1298 | int changed = 0; | |
1299 | char *container = NULL; | |
1300 | char container_buf[20]; | |
1301 | int cfd = -1; | |
1302 | ||
1303 | struct mdinfo info; | |
1304 | struct mdinfo *sra; | |
1305 | ||
1306 | if (ioctl(fd, GET_ARRAY_INFO, &array) < 0) { | |
1307 | fprintf(stderr, Name ": %s is not an active md array - aborting\n", | |
1308 | devname); | |
1309 | return 1; | |
1310 | } | |
1311 | ||
1312 | if (size >= 0 && | |
1313 | (chunksize || level!= UnSet || layout_str || raid_disks)) { | |
1314 | fprintf(stderr, Name ": cannot change component size at the same time " | |
1315 | "as other changes.\n" | |
1316 | " Change size first, then check data is intact before " | |
1317 | "making other changes.\n"); | |
1318 | return 1; | |
1319 | } | |
1320 | ||
1321 | if (raid_disks && raid_disks < array.raid_disks && array.level > 1 && | |
1322 | get_linux_version() < 2006032 && | |
1323 | !check_env("MDADM_FORCE_FEWER")) { | |
1324 | fprintf(stderr, Name ": reducing the number of devices is not safe before Linux 2.6.32\n" | |
1325 | " Please use a newer kernel\n"); | |
1326 | return 1; | |
1327 | } | |
1328 | ||
1329 | st = super_by_fd(fd, &subarray); | |
1330 | if (!st) { | |
1331 | fprintf(stderr, Name ": Unable to determine metadata format for %s\n", devname); | |
1332 | return 1; | |
1333 | } | |
1334 | if (raid_disks > st->max_devs) { | |
1335 | fprintf(stderr, Name ": Cannot increase raid-disks on this array" | |
1336 | " beyond %d\n", st->max_devs); | |
1337 | return 1; | |
1338 | } | |
1339 | ||
1340 | /* in the external case we need to check that the requested reshape is | |
1341 | * supported, and perform an initial check that the container holds the | |
1342 | * pre-requisite spare devices (mdmon owns final validation) | |
1343 | */ | |
1344 | if (st->ss->external) { | |
1345 | int container_dev; | |
1346 | int rv; | |
1347 | ||
1348 | if (subarray) { | |
1349 | container_dev = st->container_dev; | |
1350 | cfd = open_dev_excl(st->container_dev); | |
1351 | } else { | |
1352 | container_dev = st->devnum; | |
1353 | close(fd); | |
1354 | cfd = open_dev_excl(st->devnum); | |
1355 | fd = cfd; | |
1356 | } | |
1357 | if (cfd < 0) { | |
1358 | fprintf(stderr, Name ": Unable to open container for %s\n", | |
1359 | devname); | |
1360 | free(subarray); | |
1361 | return 1; | |
1362 | } | |
1363 | ||
1364 | fmt_devname(container_buf, container_dev); | |
1365 | container = container_buf; | |
1366 | ||
1367 | rv = st->ss->load_container(st, cfd, NULL); | |
1368 | ||
1369 | if (rv) { | |
1370 | fprintf(stderr, Name ": Cannot read superblock for %s\n", | |
1371 | devname); | |
1372 | free(subarray); | |
1373 | return 1; | |
1374 | } | |
1375 | ||
1376 | if (mdmon_running(container_dev)) | |
1377 | st->update_tail = &st->updates; | |
1378 | } | |
1379 | ||
1380 | if (raid_disks > array.raid_disks && | |
1381 | array.spare_disks < (raid_disks - array.raid_disks) && | |
1382 | !force) { | |
1383 | fprintf(stderr, | |
1384 | Name ": Need %d spare%s to avoid degraded array," | |
1385 | " and only have %d.\n" | |
1386 | " Use --force to over-ride this check.\n", | |
1387 | raid_disks - array.raid_disks, | |
1388 | raid_disks - array.raid_disks == 1 ? "" : "s", | |
1389 | array.spare_disks); | |
1390 | return 1; | |
1391 | } | |
1392 | ||
1393 | sra = sysfs_read(fd, 0, GET_LEVEL | GET_DISKS | GET_DEVS | |
1394 | | GET_STATE | GET_VERSION); | |
1395 | if (sra) { | |
1396 | if (st->ss->external && subarray == NULL) { | |
1397 | array.level = LEVEL_CONTAINER; | |
1398 | sra->array.level = LEVEL_CONTAINER; | |
1399 | } | |
1400 | } else { | |
1401 | fprintf(stderr, Name ": failed to read sysfs parameters for %s\n", | |
1402 | devname); | |
1403 | return 1; | |
1404 | } | |
1405 | frozen = freeze(st); | |
1406 | if (frozen < -1) { | |
1407 | /* freeze() already spewed the reason */ | |
1408 | return 1; | |
1409 | } else if (frozen < 0) { | |
1410 | fprintf(stderr, Name ": %s is performing resync/recovery and cannot" | |
1411 | " be reshaped\n", devname); | |
1412 | return 1; | |
1413 | } | |
1414 | ||
1415 | /* ========= set size =============== */ | |
1416 | if (size >= 0 && (size == 0 || size != array.size)) { | |
1417 | long long orig_size = array.size; | |
1418 | ||
1419 | if (reshape_super(st, size, UnSet, UnSet, 0, 0, NULL, devname, !quiet)) { | |
1420 | rv = 1; | |
1421 | goto release; | |
1422 | } | |
1423 | sync_metadata(st); | |
1424 | array.size = size; | |
1425 | if (array.size != size) { | |
1426 | /* got truncated to 32bit, write to | |
1427 | * component_size instead | |
1428 | */ | |
1429 | if (sra) | |
1430 | rv = sysfs_set_num(sra, NULL, | |
1431 | "component_size", size); | |
1432 | else | |
1433 | rv = -1; | |
1434 | } else | |
1435 | rv = ioctl(fd, SET_ARRAY_INFO, &array); | |
1436 | if (rv != 0) { | |
1437 | int err = errno; | |
1438 | ||
1439 | /* restore metadata */ | |
1440 | if (reshape_super(st, orig_size, UnSet, UnSet, 0, 0, | |
1441 | NULL, devname, !quiet) == 0) | |
1442 | sync_metadata(st); | |
1443 | fprintf(stderr, Name ": Cannot set device size for %s: %s\n", | |
1444 | devname, strerror(err)); | |
1445 | if (err == EBUSY && | |
1446 | (array.state & (1<<MD_SB_BITMAP_PRESENT))) | |
1447 | fprintf(stderr, " Bitmap must be removed before size can be changed\n"); | |
1448 | rv = 1; | |
1449 | goto release; | |
1450 | } | |
1451 | ioctl(fd, GET_ARRAY_INFO, &array); | |
1452 | size = get_component_size(fd)/2; | |
1453 | if (size == 0) | |
1454 | size = array.size; | |
1455 | if (!quiet) | |
1456 | fprintf(stderr, Name ": component size of %s has been set to %lluK\n", | |
1457 | devname, size); | |
1458 | changed = 1; | |
1459 | } else if (array.level != LEVEL_CONTAINER) { | |
1460 | size = get_component_size(fd)/2; | |
1461 | if (size == 0) | |
1462 | size = array.size; | |
1463 | } | |
1464 | ||
1465 | /* ========= check for Raid10/Raid1 -> Raid0 conversion =============== | |
1466 | * current implementation assumes that following conditions must be met: | |
1467 | * - RAID10: | |
1468 | * - far_copies == 1 | |
1469 | * - near_copies == 2 | |
1470 | */ | |
1471 | if ((level == 0 && array.level == 10 && sra && | |
1472 | array.layout == ((1 << 8) + 2) && !(array.raid_disks & 1)) || | |
1473 | (level == 0 && array.level == 1 && sra)) { | |
1474 | int err; | |
1475 | err = remove_disks_for_takeover(st, sra, array.layout); | |
1476 | if (err) { | |
1477 | dprintf(Name": Array cannot be reshaped\n"); | |
1478 | if (cfd > -1) | |
1479 | close(cfd); | |
1480 | rv = 1; | |
1481 | goto release; | |
1482 | } | |
1483 | /* FIXME this is added with no justification - why is it here */ | |
1484 | ping_monitor(container); | |
1485 | } | |
1486 | ||
1487 | info.array = array; | |
1488 | sysfs_init(&info, fd, NoMdDev); | |
1489 | strcpy(info.text_version, sra->text_version); | |
1490 | info.component_size = size*2; | |
1491 | info.new_level = level; | |
1492 | info.new_chunk = chunksize * 1024; | |
1493 | if (raid_disks) | |
1494 | info.delta_disks = raid_disks - info.array.raid_disks; | |
1495 | else | |
1496 | info.delta_disks = UnSet; | |
1497 | if (layout_str == NULL) { | |
1498 | info.new_layout = UnSet; | |
1499 | if (info.array.level == 6 && | |
1500 | (info.new_level == 6 || info.new_level == UnSet) && | |
1501 | info.array.layout >= 16) { | |
1502 | fprintf(stderr, Name | |
1503 | ": %s has a non-standard layout. If you" | |
1504 | " wish to preserve this\n" | |
1505 | " during the reshape, please specify" | |
1506 | " --layout=preserve\n" | |
1507 | " If you want to change it, specify a" | |
1508 | " layout or use --layout=normalise\n", | |
1509 | devname); | |
1510 | rv = 1; | |
1511 | goto release; | |
1512 | } | |
1513 | } else if (strcmp(layout_str, "normalise") == 0 || | |
1514 | strcmp(layout_str, "normalize") == 0) { | |
1515 | /* If we have a -6 RAID6 layout, remove the '-6'. */ | |
1516 | info.new_layout = UnSet; | |
1517 | if (info.array.level == 6 && info.new_level == UnSet) { | |
1518 | char l[40], *h; | |
1519 | strcpy(l, map_num(r6layout, info.array.layout)); | |
1520 | h = strrchr(l, '-'); | |
1521 | if (h && strcmp(h, "-6") == 0) { | |
1522 | *h = 0; | |
1523 | info.new_layout = map_name(r6layout, l); | |
1524 | } | |
1525 | } | |
1526 | } else if (strcmp(layout_str, "preserve") == 0) { | |
1527 | info.new_layout = UnSet; | |
1528 | } else { | |
1529 | int l = info.new_level; | |
1530 | if (l == UnSet) | |
1531 | l = info.array.level; | |
1532 | switch (l) { | |
1533 | case 5: | |
1534 | info.new_layout = map_name(r5layout, layout_str); | |
1535 | break; | |
1536 | case 6: | |
1537 | info.new_layout = map_name(r6layout, layout_str); | |
1538 | break; | |
1539 | case 10: | |
1540 | info.new_layout = parse_layout_10(layout_str); | |
1541 | break; | |
1542 | case LEVEL_FAULTY: | |
1543 | info.new_layout = parse_layout_faulty(layout_str); | |
1544 | break; | |
1545 | default: | |
1546 | fprintf(stderr, Name ": layout not meaningful" | |
1547 | " with this level\n"); | |
1548 | rv = 1; | |
1549 | goto release; | |
1550 | } | |
1551 | if (info.new_layout == UnSet) { | |
1552 | fprintf(stderr, Name ": layout %s not understood" | |
1553 | " for this level\n", | |
1554 | layout_str); | |
1555 | rv = 1; | |
1556 | goto release; | |
1557 | } | |
1558 | } | |
1559 | ||
1560 | if (array.level == LEVEL_CONTAINER) { | |
1561 | /* This change is to be applied to every array in the | |
1562 | * container. This is only needed when the metadata imposes | |
1563 | * restraints of the various arrays in the container. | |
1564 | * Currently we only know that IMSM requires all arrays | |
1565 | * to have the same number of devices so changing the | |
1566 | * number of devices (On-Line Capacity Expansion) must be | |
1567 | * performed at the level of the container | |
1568 | */ | |
1569 | rv = reshape_container(container, fd, devname, st, &info, | |
1570 | force, backup_file, quiet); | |
1571 | frozen = 0; | |
1572 | } else { | |
1573 | /* Impose these changes on a single array. First | |
1574 | * check that the metadata is OK with the change. */ | |
1575 | ||
1576 | if (reshape_super(st, info.component_size, info.new_level, | |
1577 | info.new_layout, info.new_chunk, | |
1578 | info.array.raid_disks + info.delta_disks, | |
1579 | backup_file, devname, quiet)) { | |
1580 | rv = 1; | |
1581 | goto release; | |
1582 | } | |
1583 | sync_metadata(st); | |
1584 | rv = reshape_array(container, fd, devname, st, &info, force, | |
1585 | backup_file, quiet, 0, 0); | |
1586 | frozen = 0; | |
1587 | } | |
1588 | release: | |
1589 | if (frozen > 0) | |
1590 | unfreeze(st); | |
1591 | return rv; | |
1592 | } | |
1593 | ||
1594 | static int reshape_array(char *container, int fd, char *devname, | |
1595 | struct supertype *st, struct mdinfo *info, | |
1596 | int force, | |
1597 | char *backup_file, int quiet, int forked, | |
1598 | int restart) | |
1599 | { | |
1600 | struct reshape reshape; | |
1601 | int spares_needed; | |
1602 | char *msg; | |
1603 | int orig_level = UnSet; | |
1604 | int disks, odisks; | |
1605 | ||
1606 | struct mdu_array_info_s array; | |
1607 | char *c; | |
1608 | ||
1609 | int *fdlist; | |
1610 | unsigned long long *offsets; | |
1611 | int d; | |
1612 | int nrdisks; | |
1613 | int err; | |
1614 | unsigned long blocks; | |
1615 | unsigned long cache; | |
1616 | unsigned long long array_size; | |
1617 | int done; | |
1618 | struct mdinfo *sra = NULL; | |
1619 | ||
1620 | msg = analyse_change(info, &reshape); | |
1621 | if (msg) { | |
1622 | fprintf(stderr, Name ": %s\n", msg); | |
1623 | goto release; | |
1624 | } | |
1625 | if (ioctl(fd, GET_ARRAY_INFO, &array) != 0) { | |
1626 | dprintf("Canot get array information.\n"); | |
1627 | goto release; | |
1628 | } | |
1629 | ||
1630 | if (restart) { | |
1631 | /* reshape already started. just skip to monitoring the reshape */ | |
1632 | if (reshape.backup_blocks == 0) | |
1633 | return 0; | |
1634 | goto started; | |
1635 | } | |
1636 | /* The container is frozen but the array may not be. | |
1637 | * So freeze the array so spares don't get put to the wrong use | |
1638 | * FIXME there should probably be a cleaner separation between | |
1639 | * freeze_array and freeze_container. | |
1640 | */ | |
1641 | sysfs_freeze_array(info); | |
1642 | spares_needed = max(reshape.before.data_disks, | |
1643 | reshape.after.data_disks) | |
1644 | + reshape.parity - array.raid_disks; | |
1645 | ||
1646 | if (!force && | |
1647 | info->new_level > 1 && | |
1648 | spares_needed > info->array.spare_disks) { | |
1649 | fprintf(stderr, | |
1650 | Name ": Need %d spare%s to avoid degraded array," | |
1651 | " and only have %d.\n" | |
1652 | " Use --force to over-ride this check.\n", | |
1653 | spares_needed, | |
1654 | spares_needed == 1 ? "" : "s", | |
1655 | info->array.spare_disks); | |
1656 | goto release; | |
1657 | } | |
1658 | ||
1659 | if (reshape.level != info->array.level) { | |
1660 | char *c = map_num(pers, reshape.level); | |
1661 | int err; | |
1662 | if (c == NULL) | |
1663 | goto release; | |
1664 | ||
1665 | err = sysfs_set_str(info, NULL, "level", c); | |
1666 | if (err) { | |
1667 | err = errno; | |
1668 | fprintf(stderr, Name ": %s: could not set level to %s\n", | |
1669 | devname, c); | |
1670 | if (err == EBUSY && | |
1671 | (info->array.state & (1<<MD_SB_BITMAP_PRESENT))) | |
1672 | fprintf(stderr, " Bitmap must be removed" | |
1673 | " before level can be changed\n"); | |
1674 | goto release; | |
1675 | } | |
1676 | if (!quiet) | |
1677 | fprintf(stderr, Name ": level of %s changed to %s\n", | |
1678 | devname, c); | |
1679 | orig_level = info->array.level; | |
1680 | ||
1681 | if (reshape.level > 0 && st->ss->external) { | |
1682 | /* make sure mdmon is aware of the new level */ | |
1683 | if (!mdmon_running(st->container_dev)) | |
1684 | start_mdmon(st->container_dev); | |
1685 | ping_monitor(container); | |
1686 | } | |
1687 | } | |
1688 | /* ->reshape_super might have chosen some spares from the | |
1689 | * container that it wants to be part of the new array. | |
1690 | * We can collect them with ->container_content and give | |
1691 | * them to the kernel. | |
1692 | */ | |
1693 | if (st->ss->reshape_super && st->ss->container_content) { | |
1694 | char *subarray = strchr(info->text_version+1, '/')+1; | |
1695 | struct mdinfo *info2 = | |
1696 | st->ss->container_content(st, subarray); | |
1697 | struct mdinfo *d; | |
1698 | ||
1699 | if (info2) { | |
1700 | sysfs_init(info2, fd, st->devnum); | |
1701 | for (d = info2->devs; d; d = d->next) { | |
1702 | if (d->disk.state == 0 && | |
1703 | d->disk.raid_disk >= 0) { | |
1704 | /* This is a spare that wants to | |
1705 | * be part of the array. | |
1706 | */ | |
1707 | add_disk(fd, st, info2, d); | |
1708 | } | |
1709 | } | |
1710 | sysfs_free(info2); | |
1711 | } | |
1712 | } | |
1713 | ||
1714 | if (reshape.backup_blocks == 0) { | |
1715 | /* No restriping needed, but we might need to impose | |
1716 | * some more changes: layout, raid_disks, chunk_size | |
1717 | */ | |
1718 | /* read current array info */ | |
1719 | if (ioctl(fd, GET_ARRAY_INFO, &array) != 0) { | |
1720 | dprintf("Canot get array information.\n"); | |
1721 | goto release; | |
1722 | } | |
1723 | /* compare current array info with new values and if | |
1724 | * it is different update them to new */ | |
1725 | if (info->new_layout != UnSet && | |
1726 | info->new_layout != array.layout) { | |
1727 | array.layout = info->new_layout; | |
1728 | if (ioctl(fd, SET_ARRAY_INFO, &array) != 0) { | |
1729 | fprintf(stderr, Name ": failed to set new layout\n"); | |
1730 | goto release; | |
1731 | } else if (!quiet) | |
1732 | printf("layout for %s set to %d\n", | |
1733 | devname, array.layout); | |
1734 | } | |
1735 | if (info->delta_disks != UnSet && | |
1736 | info->delta_disks != 0 && | |
1737 | array.raid_disks != (info->array.raid_disks + info->delta_disks)) { | |
1738 | array.raid_disks += info->delta_disks; | |
1739 | if (ioctl(fd, SET_ARRAY_INFO, &array) != 0) { | |
1740 | fprintf(stderr, Name ": failed to set raid disks\n"); | |
1741 | goto release; | |
1742 | } else if (!quiet) { | |
1743 | printf("raid_disks for %s set to %d\n", | |
1744 | devname, array.raid_disks); | |
1745 | } | |
1746 | } | |
1747 | if (info->new_chunk != 0 && | |
1748 | info->new_chunk != array.chunk_size) { | |
1749 | if (sysfs_set_num(info, NULL, | |
1750 | "chunk_size", info->new_chunk) != 0) { | |
1751 | fprintf(stderr, Name ": failed to set chunk size\n"); | |
1752 | goto release; | |
1753 | } else if (!quiet) | |
1754 | printf("chunk size for %s set to %d\n", | |
1755 | devname, array.chunk_size); | |
1756 | } | |
1757 | unfreeze(st); | |
1758 | return 0; | |
1759 | } | |
1760 | ||
1761 | /* | |
1762 | * There are three possibilities. | |
1763 | * 1/ The array will shrink. | |
1764 | * We need to ensure the reshape will pause before reaching | |
1765 | * the 'critical section'. We also need to fork and wait for | |
1766 | * that to happen. When it does we | |
1767 | * suspend/backup/complete/unfreeze | |
1768 | * | |
1769 | * 2/ The array will not change size. | |
1770 | * This requires that we keep a backup of a sliding window | |
1771 | * so that we can restore data after a crash. So we need | |
1772 | * to fork and monitor progress. | |
1773 | * In future we will allow the data_offset to change, so | |
1774 | * a sliding backup becomes unnecessary. | |
1775 | * | |
1776 | * 3/ The array will grow. This is relatively easy. | |
1777 | * However the kernel's restripe routines will cheerfully | |
1778 | * overwrite some early data before it is safe. So we | |
1779 | * need to make a backup of the early parts of the array | |
1780 | * and be ready to restore it if rebuild aborts very early. | |
1781 | * For externally managed metadata, we still need a forked | |
1782 | * child to monitor the reshape and suspend IO over the region | |
1783 | * that is being reshaped. | |
1784 | * | |
1785 | * We backup data by writing it to one spare, or to a | |
1786 | * file which was given on command line. | |
1787 | * | |
1788 | * In each case, we first make sure that storage is available | |
1789 | * for the required backup. | |
1790 | * Then we: | |
1791 | * - request the shape change. | |
1792 | * - fork to handle backup etc. | |
1793 | */ | |
1794 | started: | |
1795 | /* Check that we can hold all the data */ | |
1796 | get_dev_size(fd, NULL, &array_size); | |
1797 | if (reshape.new_size < (array_size/512)) { | |
1798 | fprintf(stderr, | |
1799 | Name ": this change will reduce the size of the array.\n" | |
1800 | " use --grow --array-size first to truncate array.\n" | |
1801 | " e.g. mdadm --grow %s --array-size %llu\n", | |
1802 | devname, reshape.new_size/2); | |
1803 | goto release; | |
1804 | } | |
1805 | ||
1806 | sra = sysfs_read(fd, 0, | |
1807 | GET_COMPONENT|GET_DEVS|GET_OFFSET|GET_STATE|GET_CHUNK| | |
1808 | GET_CACHE); | |
1809 | if (!sra) { | |
1810 | fprintf(stderr, Name ": %s: Cannot get array details from sysfs\n", | |
1811 | devname); | |
1812 | goto release; | |
1813 | } | |
1814 | ||
1815 | /* Decide how many blocks (sectors) for a reshape | |
1816 | * unit. The number we have so far is just a minimum | |
1817 | */ | |
1818 | blocks = reshape.backup_blocks; | |
1819 | if (reshape.before.data_disks == | |
1820 | reshape.after.data_disks) { | |
1821 | /* Make 'blocks' bigger for better throughput, but | |
1822 | * not so big that we reject it below. | |
1823 | * Try for 16 megabytes | |
1824 | */ | |
1825 | while (blocks * 32 < sra->component_size && | |
1826 | blocks < 16*1024*2) | |
1827 | blocks *= 2; | |
1828 | } else | |
1829 | fprintf(stderr, Name ": Need to backup %luK of critical " | |
1830 | "section..\n", blocks/2); | |
1831 | ||
1832 | if (blocks >= sra->component_size/2) { | |
1833 | fprintf(stderr, Name ": %s: Something wrong" | |
1834 | " - reshape aborted\n", | |
1835 | devname); | |
1836 | goto release; | |
1837 | } | |
1838 | ||
1839 | /* Now we need to open all these devices so we can read/write. | |
1840 | */ | |
1841 | nrdisks = max(reshape.before.data_disks, | |
1842 | reshape.after.data_disks) + reshape.parity | |
1843 | + sra->array.spare_disks; | |
1844 | fdlist = malloc((1+nrdisks) * sizeof(int)); | |
1845 | offsets = malloc((1+nrdisks) * sizeof(offsets[0])); | |
1846 | if (!fdlist || !offsets) { | |
1847 | fprintf(stderr, Name ": malloc failed: grow aborted\n"); | |
1848 | goto release; | |
1849 | } | |
1850 | ||
1851 | odisks = reshape.before.data_disks + reshape.parity; | |
1852 | d = reshape_prepare_fdlist(devname, sra, odisks, | |
1853 | nrdisks, blocks, backup_file, | |
1854 | fdlist, offsets); | |
1855 | if (d < 0) { | |
1856 | goto release; | |
1857 | } | |
1858 | if (backup_file == NULL) { | |
1859 | if (reshape.after.data_disks <= reshape.before.data_disks) { | |
1860 | fprintf(stderr, | |
1861 | Name ": %s: Cannot grow - need backup-file\n", | |
1862 | devname); | |
1863 | goto release; | |
1864 | } else if (sra->array.spare_disks == 0) { | |
1865 | fprintf(stderr, Name ": %s: Cannot grow - need a spare or " | |
1866 | "backup-file to backup critical section\n", | |
1867 | devname); | |
1868 | goto release; | |
1869 | } | |
1870 | } else { | |
1871 | if (!reshape_open_backup_file(backup_file, fd, devname, | |
1872 | (signed)blocks, | |
1873 | fdlist+d, offsets+d, restart)) { | |
1874 | goto release; | |
1875 | } | |
1876 | d++; | |
1877 | } | |
1878 | ||
1879 | /* lastly, check that the internal stripe cache is | |
1880 | * large enough, or it won't work. | |
1881 | * It must hold at least 4 stripes of the larger | |
1882 | * chunk size | |
1883 | */ | |
1884 | cache = max(info->array.chunk_size, info->new_chunk); | |
1885 | cache *= 4; /* 4 stripes minimum */ | |
1886 | cache /= 512; /* convert to sectors */ | |
1887 | disks = min(reshape.before.data_disks, reshape.after.data_disks); | |
1888 | /* make sure there is room for 'blocks' with a bit to spare */ | |
1889 | if (cache < 16 + blocks / disks) | |
1890 | cache = 16 + blocks / disks; | |
1891 | cache /= (4096/512); /* Covert from sectors to pages */ | |
1892 | ||
1893 | if (sra->cache_size < cache) | |
1894 | subarray_set_num(container, sra, "stripe_cache_size", | |
1895 | cache+1); | |
1896 | ||
1897 | /* Right, everything seems fine. Let's kick things off. | |
1898 | * If only changing raid_disks, use ioctl, else use | |
1899 | * sysfs. | |
1900 | */ | |
1901 | sync_metadata(st); | |
1902 | ||
1903 | sra->new_chunk = info->new_chunk; | |
1904 | ||
1905 | if (info->reshape_active) | |
1906 | sra->reshape_progress = info->reshape_progress; | |
1907 | else { | |
1908 | sra->reshape_progress = 0; | |
1909 | if (reshape.after.data_disks < reshape.before.data_disks) | |
1910 | /* start from the end of the new array */ | |
1911 | sra->reshape_progress = (sra->component_size | |
1912 | * reshape.after.data_disks); | |
1913 | } | |
1914 | ||
1915 | if (info->array.chunk_size == info->new_chunk && | |
1916 | reshape.before.layout == reshape.after.layout && | |
1917 | st->ss->external == 0) { | |
1918 | /* use SET_ARRAY_INFO but only if reshape hasn't started */ | |
1919 | ioctl(fd, GET_ARRAY_INFO, &array); | |
1920 | array.raid_disks = reshape.after.data_disks + reshape.parity; | |
1921 | if (!info->reshape_active && | |
1922 | ioctl(fd, SET_ARRAY_INFO, &array) != 0) { | |
1923 | int err = errno; | |
1924 | ||
1925 | fprintf(stderr, | |
1926 | Name ": Cannot set device shape for %s: %s\n", | |
1927 | devname, strerror(errno)); | |
1928 | ||
1929 | if (err == EBUSY && | |
1930 | (array.state & (1<<MD_SB_BITMAP_PRESENT))) | |
1931 | fprintf(stderr, | |
1932 | " Bitmap must be removed before" | |
1933 | " shape can be changed\n"); | |
1934 | ||
1935 | goto release; | |
1936 | } | |
1937 | } else if (info->reshape_active && !st->ss->external) { | |
1938 | /* We don't need to set anything here for internal | |
1939 | * metadata, and for kernels before 2.6.38 we can | |
1940 | * fail if we try. | |
1941 | */ | |
1942 | } else { | |
1943 | /* set them all just in case some old 'new_*' value | |
1944 | * persists from some earlier problem. | |
1945 | * We even set them when restarting in the middle. They will | |
1946 | * already be set in that case so this will be a no-op, | |
1947 | * but it is hard to tell the difference. | |
1948 | */ | |
1949 | int err = 0; | |
1950 | if (sysfs_set_num(sra, NULL, "chunk_size", info->new_chunk) < 0) | |
1951 | err = errno; | |
1952 | if (!err && sysfs_set_num(sra, NULL, "layout", | |
1953 | reshape.after.layout) < 0) | |
1954 | err = errno; | |
1955 | if (!err && subarray_set_num(container, sra, "raid_disks", | |
1956 | reshape.after.data_disks + | |
1957 | reshape.parity) < 0) | |
1958 | err = errno; | |
1959 | if (err) { | |
1960 | fprintf(stderr, Name ": Cannot set device shape for %s\n", | |
1961 | devname); | |
1962 | ||
1963 | if (err == EBUSY && | |
1964 | (array.state & (1<<MD_SB_BITMAP_PRESENT))) | |
1965 | fprintf(stderr, | |
1966 | " Bitmap must be removed before" | |
1967 | " shape can be changed\n"); | |
1968 | goto release; | |
1969 | } | |
1970 | } | |
1971 | ||
1972 | err = start_reshape(sra, (info->reshape_active && !st->ss->external)); | |
1973 | if (err) { | |
1974 | fprintf(stderr, Name ": Cannot start reshape for %s\n", | |
1975 | devname); | |
1976 | goto release; | |
1977 | } | |
1978 | if (restart) | |
1979 | sysfs_set_str(sra, NULL, "array_state", "active"); | |
1980 | ||
1981 | /* Now we just need to kick off the reshape and watch, while | |
1982 | * handling backups of the data... | |
1983 | * This is all done by a forked background process. | |
1984 | */ | |
1985 | switch(forked ? 0 : fork()) { | |
1986 | case -1: | |
1987 | fprintf(stderr, Name ": Cannot run child to monitor reshape: %s\n", | |
1988 | strerror(errno)); | |
1989 | abort_reshape(sra); | |
1990 | goto release; | |
1991 | default: | |
1992 | return 0; | |
1993 | case 0: | |
1994 | break; | |
1995 | } | |
1996 | ||
1997 | close(fd); | |
1998 | if (check_env("MDADM_GROW_VERIFY")) | |
1999 | fd = open(devname, O_RDONLY | O_DIRECT); | |
2000 | else | |
2001 | fd = -1; | |
2002 | mlockall(MCL_FUTURE); | |
2003 | ||
2004 | if (st->ss->external) { | |
2005 | /* metadata handler takes it from here */ | |
2006 | done = st->ss->manage_reshape( | |
2007 | fd, sra, &reshape, st, blocks, | |
2008 | fdlist, offsets, | |
2009 | d - odisks, fdlist+odisks, | |
2010 | offsets+odisks); | |
2011 | } else | |
2012 | done = child_monitor( | |
2013 | fd, sra, &reshape, st, blocks, | |
2014 | fdlist, offsets, | |
2015 | d - odisks, fdlist+odisks, | |
2016 | offsets+odisks); | |
2017 | ||
2018 | if (backup_file && done) | |
2019 | unlink(backup_file); | |
2020 | if (!done) { | |
2021 | abort_reshape(sra); | |
2022 | goto out; | |
2023 | } | |
2024 | ||
2025 | if (!st->ss->external && | |
2026 | !(reshape.before.data_disks != reshape.after.data_disks | |
2027 | && info->custom_array_size) && | |
2028 | info->new_level == reshape.level && | |
2029 | !forked) { | |
2030 | /* no need to wait for the reshape to finish as | |
2031 | * there is nothing more to do. | |
2032 | */ | |
2033 | exit(0); | |
2034 | } | |
2035 | wait_reshape(sra); | |
2036 | ||
2037 | if (st->ss->external) { | |
2038 | /* Re-load the metadata as much could have changed */ | |
2039 | int cfd = open_dev(st->container_dev); | |
2040 | if (cfd >= 0) { | |
2041 | ping_monitor(container); | |
2042 | st->ss->free_super(st); | |
2043 | st->ss->load_container(st, cfd, container); | |
2044 | close(cfd); | |
2045 | } | |
2046 | } | |
2047 | ||
2048 | /* set new array size if required customer_array_size is used | |
2049 | * by this metadata. | |
2050 | */ | |
2051 | if (reshape.before.data_disks != | |
2052 | reshape.after.data_disks && | |
2053 | info->custom_array_size) { | |
2054 | struct mdinfo *info2; | |
2055 | char *subarray = strchr(info->text_version+1, '/')+1; | |
2056 | ||
2057 | info2 = st->ss->container_content(st, subarray); | |
2058 | if (info2) { | |
2059 | unsigned long long current_size = 0; | |
2060 | unsigned long long new_size = | |
2061 | info2->custom_array_size/2; | |
2062 | ||
2063 | if (sysfs_get_ll(sra, | |
2064 | NULL, | |
2065 | "array_size", | |
2066 | ¤t_size) == 0 && | |
2067 | new_size > current_size) { | |
2068 | if (sysfs_set_num(sra, NULL, | |
2069 | "array_size", new_size) | |
2070 | < 0) | |
2071 | dprintf("Error: Cannot" | |
2072 | " set array size"); | |
2073 | else | |
2074 | dprintf("Array size " | |
2075 | "changed"); | |
2076 | dprintf(" from %llu to %llu.\n", | |
2077 | current_size, new_size); | |
2078 | } | |
2079 | sysfs_free(info2); | |
2080 | } | |
2081 | } | |
2082 | ||
2083 | if (info->new_level != reshape.level) { | |
2084 | ||
2085 | c = map_num(pers, info->new_level); | |
2086 | if (c) { | |
2087 | err = sysfs_set_str(sra, NULL, "level", c); | |
2088 | if (err) | |
2089 | fprintf(stderr, Name\ | |
2090 | ": %s: could not set level " | |
2091 | "to %s\n", devname, c); | |
2092 | } | |
2093 | } | |
2094 | out: | |
2095 | if (forked) | |
2096 | return 0; | |
2097 | exit(0); | |
2098 | ||
2099 | release: | |
2100 | if (orig_level != UnSet && sra) { | |
2101 | c = map_num(pers, orig_level); | |
2102 | if (c && sysfs_set_str(sra, NULL, "level", c) == 0) | |
2103 | fprintf(stderr, Name ": aborting level change\n"); | |
2104 | } | |
2105 | if (!forked) | |
2106 | unfreeze(st); | |
2107 | return 1; | |
2108 | } | |
2109 | ||
2110 | int reshape_container(char *container, int cfd, char *devname, | |
2111 | struct supertype *st, | |
2112 | struct mdinfo *info, | |
2113 | int force, | |
2114 | char *backup_file, | |
2115 | int quiet) | |
2116 | { | |
2117 | struct mdinfo *cc = NULL; | |
2118 | ||
2119 | /* component_size is not meaningful for a container, | |
2120 | * so pass '-1' meaning 'no change' | |
2121 | */ | |
2122 | if (reshape_super(st, -1, info->new_level, | |
2123 | info->new_layout, info->new_chunk, | |
2124 | info->array.raid_disks + info->delta_disks, | |
2125 | backup_file, devname, quiet)) | |
2126 | return 1; | |
2127 | ||
2128 | sync_metadata(st); | |
2129 | ||
2130 | /* ping monitor to be sure that update is on disk | |
2131 | */ | |
2132 | ping_monitor(container); | |
2133 | ||
2134 | switch (fork()) { | |
2135 | case -1: /* error */ | |
2136 | perror("Cannot fork to complete reshape\n"); | |
2137 | return 1; | |
2138 | default: /* parent */ | |
2139 | printf(Name ": multi-array reshape continues in background\n"); | |
2140 | return 0; | |
2141 | case 0: /* child */ | |
2142 | break; | |
2143 | } | |
2144 | ||
2145 | while(1) { | |
2146 | /* For each member array with reshape_active, | |
2147 | * we need to perform the reshape. | |
2148 | * We pick the first array that needs reshaping and | |
2149 | * reshape it. reshape_array() will re-read the metadata | |
2150 | * so the next time through a different array should be | |
2151 | * ready for reshape. | |
2152 | */ | |
2153 | struct mdinfo *content; | |
2154 | int rv; | |
2155 | int fd; | |
2156 | struct mdstat_ent *mdstat; | |
2157 | char *adev; | |
2158 | ||
2159 | sysfs_free(cc); | |
2160 | ||
2161 | cc = st->ss->container_content(st, NULL); | |
2162 | ||
2163 | for (content = cc; content ; content = content->next) { | |
2164 | char *subarray; | |
2165 | if (!content->reshape_active) | |
2166 | continue; | |
2167 | ||
2168 | subarray = strchr(content->text_version+1, '/')+1; | |
2169 | mdstat = mdstat_by_subdev(subarray, | |
2170 | devname2devnum(container)); | |
2171 | if (!mdstat) | |
2172 | continue; | |
2173 | break; | |
2174 | } | |
2175 | if (!content) | |
2176 | break; | |
2177 | ||
2178 | fd = open_dev(mdstat->devnum); | |
2179 | if (fd < 0) | |
2180 | break; | |
2181 | adev = map_dev(dev2major(mdstat->devnum), | |
2182 | dev2minor(mdstat->devnum), | |
2183 | 0); | |
2184 | if (!adev) | |
2185 | adev = content->text_version; | |
2186 | ||
2187 | sysfs_init(content, fd, mdstat->devnum); | |
2188 | ||
2189 | rv = reshape_array(container, fd, adev, st, | |
2190 | content, force, | |
2191 | backup_file, quiet, 1, 0); | |
2192 | close(fd); | |
2193 | if (rv) | |
2194 | break; | |
2195 | } | |
2196 | unfreeze(st); | |
2197 | sysfs_free(cc); | |
2198 | exit(0); | |
2199 | } | |
2200 | ||
2201 | /* | |
2202 | * We run a child process in the background which performs the following | |
2203 | * steps: | |
2204 | * - wait for resync to reach a certain point | |
2205 | * - suspend io to the following section | |
2206 | * - backup that section | |
2207 | * - allow resync to proceed further | |
2208 | * - resume io | |
2209 | * - discard the backup. | |
2210 | * | |
2211 | * When are combined in slightly different ways in the three cases. | |
2212 | * Grow: | |
2213 | * - suspend/backup/allow/wait/resume/discard | |
2214 | * Shrink: | |
2215 | * - allow/wait/suspend/backup/allow/wait/resume/discard | |
2216 | * same-size: | |
2217 | * - wait/resume/discard/suspend/backup/allow | |
2218 | * | |
2219 | * suspend/backup/allow always come together | |
2220 | * wait/resume/discard do too. | |
2221 | * For the same-size case we have two backups to improve flow. | |
2222 | * | |
2223 | */ | |
2224 | ||
2225 | int progress_reshape(struct mdinfo *info, struct reshape *reshape, | |
2226 | unsigned long long backup_point, | |
2227 | unsigned long long wait_point, | |
2228 | unsigned long long *suspend_point, | |
2229 | unsigned long long *reshape_completed) | |
2230 | { | |
2231 | /* This function is called repeatedly by the reshape manager. | |
2232 | * It determines how much progress can safely be made and allows | |
2233 | * that progress. | |
2234 | * - 'info' identifies the array and particularly records in | |
2235 | * ->reshape_progress the metadata's knowledge of progress | |
2236 | * This is a sector offset from the start of the array | |
2237 | * of the next array block to be relocated. This number | |
2238 | * may increase from 0 or decrease from array_size, depending | |
2239 | * on the type of reshape that is happening. | |
2240 | * Note that in contrast, 'sync_completed' is a block count of the | |
2241 | * reshape so far. It gives the distance between the start point | |
2242 | * (head or tail of device) and the next place that data will be | |
2243 | * written. It always increases. | |
2244 | * - 'reshape' is the structure created by analyse_change | |
2245 | * - 'backup_point' shows how much the metadata manager has backed-up | |
2246 | * data. For reshapes with increasing progress, it is the next address | |
2247 | * to be backed up, previous addresses have been backed-up. For | |
2248 | * decreasing progress, it is the earliest address that has been | |
2249 | * backed up - later address are also backed up. | |
2250 | * So addresses between reshape_progress and backup_point are | |
2251 | * backed up providing those are in the 'correct' order. | |
2252 | * - 'wait_point' is an array address. When reshape_completed | |
2253 | * passes this point, progress_reshape should return. It might | |
2254 | * return earlier if it determines that ->reshape_progress needs | |
2255 | * to be updated or further backup is needed. | |
2256 | * - suspend_point is maintained by progress_reshape and the caller | |
2257 | * should not touch it except to initialise to zero. | |
2258 | * It is an array address and it only increases in 2.6.37 and earlier. | |
2259 | * This makes it difficult to handle reducing reshapes with | |
2260 | * external metadata. | |
2261 | * However: it is similar to backup_point in that it records the | |
2262 | * other end of a suspended region from reshape_progress. | |
2263 | * it is moved to extend the region that is safe to backup and/or | |
2264 | * reshape | |
2265 | * - reshape_completed is read from sysfs and returned. The caller | |
2266 | * should copy this into ->reshape_progress when it has reason to | |
2267 | * believe that the metadata knows this, and any backup outside this | |
2268 | * has been erased. | |
2269 | * | |
2270 | * Return value is: | |
2271 | * 1 if more data from backup_point - but only as far as suspend_point, | |
2272 | * should be backed up | |
2273 | * 0 if things are progressing smoothly | |
2274 | * -1 if the reshape is finished, either because it is all done, | |
2275 | * or due to an error. | |
2276 | */ | |
2277 | ||
2278 | int advancing = (reshape->after.data_disks | |
2279 | >= reshape->before.data_disks); | |
2280 | unsigned long long need_backup; /* All data between start of array and | |
2281 | * here will at some point need to | |
2282 | * be backed up. | |
2283 | */ | |
2284 | unsigned long long read_offset, write_offset; | |
2285 | unsigned long long write_range; | |
2286 | unsigned long long max_progress, target, completed; | |
2287 | unsigned long long array_size = (info->component_size | |
2288 | * reshape->before.data_disks); | |
2289 | int fd; | |
2290 | char buf[20]; | |
2291 | ||
2292 | /* First, we unsuspend any region that is now known to be safe. | |
2293 | * If suspend_point is on the 'wrong' side of reshape_progress, then | |
2294 | * we don't have or need suspension at the moment. This is true for | |
2295 | * native metadata when we don't need to back-up. | |
2296 | */ | |
2297 | if (advancing) { | |
2298 | if (info->reshape_progress <= *suspend_point) | |
2299 | sysfs_set_num(info, NULL, "suspend_lo", | |
2300 | info->reshape_progress); | |
2301 | } else { | |
2302 | /* Note: this won't work in 2.6.37 and before. | |
2303 | * Something somewhere should make sure we don't need it! | |
2304 | */ | |
2305 | if (info->reshape_progress >= *suspend_point) | |
2306 | sysfs_set_num(info, NULL, "suspend_hi", | |
2307 | info->reshape_progress); | |
2308 | } | |
2309 | ||
2310 | /* Now work out how far it is safe to progress. | |
2311 | * If the read_offset for ->reshape_progress is less than | |
2312 | * 'blocks' beyond the write_offset, we can only progress as far | |
2313 | * as a backup. | |
2314 | * Otherwise we can progress until the write_offset for the new location | |
2315 | * reaches (within 'blocks' of) the read_offset at the current location. | |
2316 | * However that region must be suspended unless we are using native | |
2317 | * metadata. | |
2318 | * If we need to suspend more, we limit it to 128M per device, which is | |
2319 | * rather arbitrary and should be some time-based calculation. | |
2320 | */ | |
2321 | read_offset = info->reshape_progress / reshape->before.data_disks; | |
2322 | write_offset = info->reshape_progress / reshape->after.data_disks; | |
2323 | write_range = info->new_chunk/512; | |
2324 | if (reshape->before.data_disks == reshape->after.data_disks) | |
2325 | need_backup = array_size; | |
2326 | else | |
2327 | need_backup = reshape->backup_blocks; | |
2328 | if (advancing) { | |
2329 | if (read_offset < write_offset + write_range) | |
2330 | max_progress = backup_point; | |
2331 | else | |
2332 | max_progress = | |
2333 | read_offset * | |
2334 | reshape->after.data_disks; | |
2335 | } else { | |
2336 | if (read_offset > write_offset - write_range) | |
2337 | /* Can only progress as far as has been backed up, | |
2338 | * which must be suspended */ | |
2339 | max_progress = backup_point; | |
2340 | else if (info->reshape_progress <= need_backup) | |
2341 | max_progress = backup_point; | |
2342 | else { | |
2343 | if (info->array.major_version >= 0) | |
2344 | /* Can progress until backup is needed */ | |
2345 | max_progress = need_backup; | |
2346 | else { | |
2347 | /* Can progress until metadata update is required */ | |
2348 | max_progress = | |
2349 | read_offset * | |
2350 | reshape->after.data_disks; | |
2351 | /* but data must be suspended */ | |
2352 | if (max_progress < *suspend_point) | |
2353 | max_progress = *suspend_point; | |
2354 | } | |
2355 | } | |
2356 | } | |
2357 | ||
2358 | /* We know it is safe to progress to 'max_progress' providing | |
2359 | * it is suspended or we are using native metadata. | |
2360 | * Consider extending suspend_point 128M per device if it | |
2361 | * is less than 64M per device beyond reshape_progress. | |
2362 | * But always do a multiple of 'blocks' | |
2363 | * FIXME this is too big - it takes to long to complete | |
2364 | * this much. | |
2365 | */ | |
2366 | target = 64*1024*2 * min(reshape->before.data_disks, | |
2367 | reshape->after.data_disks); | |
2368 | target /= reshape->backup_blocks; | |
2369 | if (target < 2) | |
2370 | target = 2; | |
2371 | target *= reshape->backup_blocks; | |
2372 | ||
2373 | /* For externally managed metadata we always need to suspend IO to | |
2374 | * the area being reshaped so we regularly push suspend_point forward. | |
2375 | * For native metadata we only need the suspend if we are going to do | |
2376 | * a backup. | |
2377 | */ | |
2378 | if (advancing) { | |
2379 | if ((need_backup > info->reshape_progress | |
2380 | || info->array.major_version < 0) && | |
2381 | *suspend_point < info->reshape_progress + target) { | |
2382 | if (need_backup < *suspend_point + 2 * target) | |
2383 | *suspend_point = need_backup; | |
2384 | else if (*suspend_point + 2 * target < array_size) | |
2385 | *suspend_point += 2 * target; | |
2386 | else | |
2387 | *suspend_point = array_size; | |
2388 | sysfs_set_num(info, NULL, "suspend_hi", *suspend_point); | |
2389 | if (max_progress > *suspend_point) | |
2390 | max_progress = *suspend_point; | |
2391 | } | |
2392 | } else { | |
2393 | if (info->array.major_version >= 0) { | |
2394 | /* Only need to suspend when about to backup */ | |
2395 | if (info->reshape_progress < need_backup * 2 && | |
2396 | *suspend_point > 0) { | |
2397 | *suspend_point = 0; | |
2398 | sysfs_set_num(info, NULL, "suspend_lo", 0); | |
2399 | sysfs_set_num(info, NULL, "suspend_hi", need_backup); | |
2400 | } | |
2401 | } else { | |
2402 | /* Need to suspend continually */ | |
2403 | if (info->reshape_progress < *suspend_point) | |
2404 | *suspend_point = info->reshape_progress; | |
2405 | if (*suspend_point + target < info->reshape_progress) | |
2406 | /* No need to move suspend region yet */; | |
2407 | else { | |
2408 | if (*suspend_point >= 2 * target) | |
2409 | *suspend_point -= 2 * target; | |
2410 | else | |
2411 | *suspend_point = 0; | |
2412 | sysfs_set_num(info, NULL, "suspend_lo", | |
2413 | *suspend_point); | |
2414 | } | |
2415 | if (max_progress < *suspend_point) | |
2416 | max_progress = *suspend_point; | |
2417 | } | |
2418 | } | |
2419 | ||
2420 | /* now set sync_max to allow that progress. sync_max, like | |
2421 | * sync_completed is a count of sectors written per device, so | |
2422 | * we find the difference between max_progress and the start point, | |
2423 | * and divide that by after.data_disks to get a sync_max | |
2424 | * number. | |
2425 | * At the same time we convert wait_point to a similar number | |
2426 | * for comparing against sync_completed. | |
2427 | */ | |
2428 | /* scale down max_progress to per_disk */ | |
2429 | max_progress /= reshape->after.data_disks; | |
2430 | /* Round to chunk size as some kernels give an erroneously high number */ | |
2431 | max_progress /= info->new_chunk/512; | |
2432 | max_progress *= info->new_chunk/512; | |
2433 | /* And round to old chunk size as the kernel wants that */ | |
2434 | max_progress /= info->array.chunk_size/512; | |
2435 | max_progress *= info->array.chunk_size/512; | |
2436 | /* Limit progress to the whole device */ | |
2437 | if (max_progress > info->component_size) | |
2438 | max_progress = info->component_size; | |
2439 | wait_point /= reshape->after.data_disks; | |
2440 | if (!advancing) { | |
2441 | /* switch from 'device offset' to 'processed block count' */ | |
2442 | max_progress = info->component_size - max_progress; | |
2443 | wait_point = info->component_size - wait_point; | |
2444 | } | |
2445 | ||
2446 | sysfs_set_num(info, NULL, "sync_max", max_progress); | |
2447 | ||
2448 | /* Now wait. If we have already reached the point that we were | |
2449 | * asked to wait to, don't wait at all, else wait for any change. | |
2450 | * We need to select on 'sync_completed' as that is the place that | |
2451 | * notifications happen, but we are really interested in | |
2452 | * 'reshape_position' | |
2453 | */ | |
2454 | fd = sysfs_get_fd(info, NULL, "sync_completed"); | |
2455 | if (fd < 0) | |
2456 | goto check_progress; | |
2457 | ||
2458 | if (sysfs_fd_get_ll(fd, &completed) < 0) { | |
2459 | close(fd); | |
2460 | goto check_progress; | |
2461 | } | |
2462 | while (completed < max_progress && completed < wait_point) { | |
2463 | /* Check that sync_action is still 'reshape' to avoid | |
2464 | * waiting forever on a dead array | |
2465 | */ | |
2466 | char action[20]; | |
2467 | fd_set rfds; | |
2468 | if (sysfs_get_str(info, NULL, "sync_action", | |
2469 | action, 20) <= 0 || | |
2470 | strncmp(action, "reshape", 7) != 0) | |
2471 | break; | |
2472 | /* Some kernels reset 'sync_completed' to zero | |
2473 | * before setting 'sync_action' to 'idle'. | |
2474 | * So we need these extra tests. | |
2475 | */ | |
2476 | if (completed == 0 && advancing | |
2477 | && info->reshape_progress > 0) | |
2478 | break; | |
2479 | if (completed == 0 && !advancing | |
2480 | && info->reshape_progress < (info->component_size | |
2481 | * reshape->after.data_disks)) | |
2482 | break; | |
2483 | FD_ZERO(&rfds); | |
2484 | FD_SET(fd, &rfds); | |
2485 | select(fd+1, NULL, NULL, &rfds, NULL); | |
2486 | if (sysfs_fd_get_ll(fd, &completed) < 0) { | |
2487 | close(fd); | |
2488 | goto check_progress; | |
2489 | } | |
2490 | } | |
2491 | /* Some kernels reset 'sync_completed' to zero, | |
2492 | * we need to have real point we are in md | |
2493 | */ | |
2494 | if (completed == 0) | |
2495 | completed = max_progress; | |
2496 | ||
2497 | /* some kernels can give an incorrectly high 'completed' number */ | |
2498 | completed /= (info->new_chunk/512); | |
2499 | completed *= (info->new_chunk/512); | |
2500 | /* Convert 'completed' back in to a 'progress' number */ | |
2501 | completed *= reshape->after.data_disks; | |
2502 | if (!advancing) { | |
2503 | completed = info->component_size * reshape->after.data_disks | |
2504 | - completed; | |
2505 | } | |
2506 | *reshape_completed = completed; | |
2507 | ||
2508 | close(fd); | |
2509 | ||
2510 | /* We return the need_backup flag. Caller will decide | |
2511 | * how much - a multiple of ->backup_blocks up to *suspend_point | |
2512 | */ | |
2513 | if (advancing) | |
2514 | return need_backup > info->reshape_progress; | |
2515 | else | |
2516 | return need_backup >= info->reshape_progress; | |
2517 | ||
2518 | check_progress: | |
2519 | /* if we couldn't read a number from sync_completed, then | |
2520 | * either the reshape did complete, or it aborted. | |
2521 | * We can tell which by checking for 'none' in reshape_position. | |
2522 | */ | |
2523 | strcpy(buf, "hi"); | |
2524 | if (sysfs_get_str(info, NULL, "reshape_position", buf, sizeof(buf)) < 0 | |
2525 | || strncmp(buf, "none", 4) != 0) | |
2526 | return -2; /* abort */ | |
2527 | else { | |
2528 | /* Maybe racing with array shutdown - check state */ | |
2529 | if (sysfs_get_str(info, NULL, "array_state", buf, sizeof(buf)) < 0 | |
2530 | || strncmp(buf, "inactive", 8) == 0 | |
2531 | || strncmp(buf, "clear",5) == 0) | |
2532 | return -2; /* abort */ | |
2533 | return -1; /* complete */ | |
2534 | } | |
2535 | } | |
2536 | ||
2537 | ||
2538 | /* FIXME return status is never checked */ | |
2539 | static int grow_backup(struct mdinfo *sra, | |
2540 | unsigned long long offset, /* per device */ | |
2541 | unsigned long stripes, /* per device, in old chunks */ | |
2542 | int *sources, unsigned long long *offsets, | |
2543 | int disks, int chunk, int level, int layout, | |
2544 | int dests, int *destfd, unsigned long long *destoffsets, | |
2545 | int part, int *degraded, | |
2546 | char *buf) | |
2547 | { | |
2548 | /* Backup 'blocks' sectors at 'offset' on each device of the array, | |
2549 | * to storage 'destfd' (offset 'destoffsets'), after first | |
2550 | * suspending IO. Then allow resync to continue | |
2551 | * over the suspended section. | |
2552 | * Use part 'part' of the backup-super-block. | |
2553 | */ | |
2554 | int odata = disks; | |
2555 | int rv = 0; | |
2556 | int i; | |
2557 | unsigned long long ll; | |
2558 | int new_degraded; | |
2559 | //printf("offset %llu\n", offset); | |
2560 | if (level >= 4) | |
2561 | odata--; | |
2562 | if (level == 6) | |
2563 | odata--; | |
2564 | ||
2565 | /* Check that array hasn't become degraded, else we might backup the wrong data */ | |
2566 | if (sysfs_get_ll(sra, NULL, "degraded", &ll) < 0) | |
2567 | return -1; /* FIXME this error is ignored */ | |
2568 | new_degraded = (int)ll; | |
2569 | if (new_degraded != *degraded) { | |
2570 | /* check each device to ensure it is still working */ | |
2571 | struct mdinfo *sd; | |
2572 | for (sd = sra->devs ; sd ; sd = sd->next) { | |
2573 | if (sd->disk.state & (1<<MD_DISK_FAULTY)) | |
2574 | continue; | |
2575 | if (sd->disk.state & (1<<MD_DISK_SYNC)) { | |
2576 | char sbuf[20]; | |
2577 | if (sysfs_get_str(sra, sd, "state", sbuf, 20) < 0 || | |
2578 | strstr(sbuf, "faulty") || | |
2579 | strstr(sbuf, "in_sync") == NULL) { | |
2580 | /* this device is dead */ | |
2581 | sd->disk.state = (1<<MD_DISK_FAULTY); | |
2582 | if (sd->disk.raid_disk >= 0 && | |
2583 | sources[sd->disk.raid_disk] >= 0) { | |
2584 | close(sources[sd->disk.raid_disk]); | |
2585 | sources[sd->disk.raid_disk] = -1; | |
2586 | } | |
2587 | } | |
2588 | } | |
2589 | } | |
2590 | *degraded = new_degraded; | |
2591 | } | |
2592 | if (part) { | |
2593 | bsb.arraystart2 = __cpu_to_le64(offset * odata); | |
2594 | bsb.length2 = __cpu_to_le64(stripes * (chunk/512) * odata); | |
2595 | } else { | |
2596 | bsb.arraystart = __cpu_to_le64(offset * odata); | |
2597 | bsb.length = __cpu_to_le64(stripes * (chunk/512) * odata); | |
2598 | } | |
2599 | if (part) | |
2600 | bsb.magic[15] = '2'; | |
2601 | for (i = 0; i < dests; i++) | |
2602 | if (part) | |
2603 | lseek64(destfd[i], destoffsets[i] + __le64_to_cpu(bsb.devstart2)*512, 0); | |
2604 | else | |
2605 | lseek64(destfd[i], destoffsets[i], 0); | |
2606 | ||
2607 | rv = save_stripes(sources, offsets, | |
2608 | disks, chunk, level, layout, | |
2609 | dests, destfd, | |
2610 | offset*512*odata, stripes * chunk * odata, | |
2611 | buf); | |
2612 | ||
2613 | if (rv) | |
2614 | return rv; | |
2615 | bsb.mtime = __cpu_to_le64(time(0)); | |
2616 | for (i = 0; i < dests; i++) { | |
2617 | bsb.devstart = __cpu_to_le64(destoffsets[i]/512); | |
2618 | ||
2619 | bsb.sb_csum = bsb_csum((char*)&bsb, ((char*)&bsb.sb_csum)-((char*)&bsb)); | |
2620 | if (memcmp(bsb.magic, "md_backup_data-2", 16) == 0) | |
2621 | bsb.sb_csum2 = bsb_csum((char*)&bsb, | |
2622 | ((char*)&bsb.sb_csum2)-((char*)&bsb)); | |
2623 | ||
2624 | rv = -1; | |
2625 | if ((unsigned long long)lseek64(destfd[i], destoffsets[i] - 4096, 0) | |
2626 | != destoffsets[i] - 4096) | |
2627 | break; | |
2628 | if (write(destfd[i], &bsb, 512) != 512) | |
2629 | break; | |
2630 | if (destoffsets[i] > 4096) { | |
2631 | if ((unsigned long long)lseek64(destfd[i], destoffsets[i]+stripes*chunk*odata, 0) != | |
2632 | destoffsets[i]+stripes*chunk*odata) | |
2633 | break; | |
2634 | if (write(destfd[i], &bsb, 512) != 512) | |
2635 | break; | |
2636 | } | |
2637 | fsync(destfd[i]); | |
2638 | rv = 0; | |
2639 | } | |
2640 | ||
2641 | return rv; | |
2642 | } | |
2643 | ||
2644 | /* in 2.6.30, the value reported by sync_completed can be | |
2645 | * less that it should be by one stripe. | |
2646 | * This only happens when reshape hits sync_max and pauses. | |
2647 | * So allow wait_backup to either extent sync_max further | |
2648 | * than strictly necessary, or return before the | |
2649 | * sync has got quite as far as we would really like. | |
2650 | * This is what 'blocks2' is for. | |
2651 | * The various caller give appropriate values so that | |
2652 | * every works. | |
2653 | */ | |
2654 | /* FIXME return value is often ignored */ | |
2655 | static int forget_backup( | |
2656 | int dests, int *destfd, unsigned long long *destoffsets, | |
2657 | int part) | |
2658 | { | |
2659 | /* | |
2660 | * Erase backup 'part' (which is 0 or 1) | |
2661 | */ | |
2662 | int i; | |
2663 | int rv; | |
2664 | ||
2665 | if (part) { | |
2666 | bsb.arraystart2 = __cpu_to_le64(0); | |
2667 | bsb.length2 = __cpu_to_le64(0); | |
2668 | } else { | |
2669 | bsb.arraystart = __cpu_to_le64(0); | |
2670 | bsb.length = __cpu_to_le64(0); | |
2671 | } | |
2672 | bsb.mtime = __cpu_to_le64(time(0)); | |
2673 | rv = 0; | |
2674 | for (i = 0; i < dests; i++) { | |
2675 | bsb.devstart = __cpu_to_le64(destoffsets[i]/512); | |
2676 | bsb.sb_csum = bsb_csum((char*)&bsb, ((char*)&bsb.sb_csum)-((char*)&bsb)); | |
2677 | if (memcmp(bsb.magic, "md_backup_data-2", 16) == 0) | |
2678 | bsb.sb_csum2 = bsb_csum((char*)&bsb, | |
2679 | ((char*)&bsb.sb_csum2)-((char*)&bsb)); | |
2680 | if ((unsigned long long)lseek64(destfd[i], destoffsets[i]-4096, 0) != | |
2681 | destoffsets[i]-4096) | |
2682 | rv = -1; | |
2683 | if (rv == 0 && | |
2684 | write(destfd[i], &bsb, 512) != 512) | |
2685 | rv = -1; | |
2686 | fsync(destfd[i]); | |
2687 | } | |
2688 | return rv; | |
2689 | } | |
2690 | ||
2691 | static void fail(char *msg) | |
2692 | { | |
2693 | int rv; | |
2694 | rv = (write(2, msg, strlen(msg)) != (int)strlen(msg)); | |
2695 | rv |= (write(2, "\n", 1) != 1); | |
2696 | exit(rv ? 1 : 2); | |
2697 | } | |
2698 | ||
2699 | static char *abuf, *bbuf; | |
2700 | static unsigned long long abuflen; | |
2701 | static void validate(int afd, int bfd, unsigned long long offset) | |
2702 | { | |
2703 | /* check that the data in the backup against the array. | |
2704 | * This is only used for regression testing and should not | |
2705 | * be used while the array is active | |
2706 | */ | |
2707 | if (afd < 0) | |
2708 | return; | |
2709 | lseek64(bfd, offset - 4096, 0); | |
2710 | if (read(bfd, &bsb2, 512) != 512) | |
2711 | fail("cannot read bsb"); | |
2712 | if (bsb2.sb_csum != bsb_csum((char*)&bsb2, | |
2713 | ((char*)&bsb2.sb_csum)-((char*)&bsb2))) | |
2714 | fail("first csum bad"); | |
2715 | if (memcmp(bsb2.magic, "md_backup_data", 14) != 0) | |
2716 | fail("magic is bad"); | |
2717 | if (memcmp(bsb2.magic, "md_backup_data-2", 16) == 0 && | |
2718 | bsb2.sb_csum2 != bsb_csum((char*)&bsb2, | |
2719 | ((char*)&bsb2.sb_csum2)-((char*)&bsb2))) | |
2720 | fail("second csum bad"); | |
2721 | ||
2722 | if (__le64_to_cpu(bsb2.devstart)*512 != offset) | |
2723 | fail("devstart is wrong"); | |
2724 | ||
2725 | if (bsb2.length) { | |
2726 | unsigned long long len = __le64_to_cpu(bsb2.length)*512; | |
2727 | ||
2728 | if (abuflen < len) { | |
2729 | free(abuf); | |
2730 | free(bbuf); | |
2731 | abuflen = len; | |
2732 | if (posix_memalign((void**)&abuf, 4096, abuflen) || | |
2733 | posix_memalign((void**)&bbuf, 4096, abuflen)) { | |
2734 | abuflen = 0; | |
2735 | /* just stop validating on mem-alloc failure */ | |
2736 | return; | |
2737 | } | |
2738 | } | |
2739 | ||
2740 | lseek64(bfd, offset, 0); | |
2741 | if ((unsigned long long)read(bfd, bbuf, len) != len) { | |
2742 | //printf("len %llu\n", len); | |
2743 | fail("read first backup failed"); | |
2744 | } | |
2745 | lseek64(afd, __le64_to_cpu(bsb2.arraystart)*512, 0); | |
2746 | if ((unsigned long long)read(afd, abuf, len) != len) | |
2747 | fail("read first from array failed"); | |
2748 | if (memcmp(bbuf, abuf, len) != 0) { | |
2749 | #if 0 | |
2750 | int i; | |
2751 | printf("offset=%llu len=%llu\n", | |
2752 | (unsigned long long)__le64_to_cpu(bsb2.arraystart)*512, len); | |
2753 | for (i=0; i<len; i++) | |
2754 | if (bbuf[i] != abuf[i]) { | |
2755 | printf("first diff byte %d\n", i); | |
2756 | break; | |
2757 | } | |
2758 | #endif | |
2759 | fail("data1 compare failed"); | |
2760 | } | |
2761 | } | |
2762 | if (bsb2.length2) { | |
2763 | unsigned long long len = __le64_to_cpu(bsb2.length2)*512; | |
2764 | ||
2765 | if (abuflen < len) { | |
2766 | free(abuf); | |
2767 | free(bbuf); | |
2768 | abuflen = len; | |
2769 | abuf = malloc(abuflen); | |
2770 | bbuf = malloc(abuflen); | |
2771 | } | |
2772 | ||
2773 | lseek64(bfd, offset+__le64_to_cpu(bsb2.devstart2)*512, 0); | |
2774 | if ((unsigned long long)read(bfd, bbuf, len) != len) | |
2775 | fail("read second backup failed"); | |
2776 | lseek64(afd, __le64_to_cpu(bsb2.arraystart2)*512, 0); | |
2777 | if ((unsigned long long)read(afd, abuf, len) != len) | |
2778 | fail("read second from array failed"); | |
2779 | if (memcmp(bbuf, abuf, len) != 0) | |
2780 | fail("data2 compare failed"); | |
2781 | } | |
2782 | } | |
2783 | ||
2784 | int child_monitor(int afd, struct mdinfo *sra, struct reshape *reshape, | |
2785 | struct supertype *st, unsigned long blocks, | |
2786 | int *fds, unsigned long long *offsets, | |
2787 | int dests, int *destfd, unsigned long long *destoffsets) | |
2788 | { | |
2789 | /* Monitor a reshape where backup is being performed using | |
2790 | * 'native' mechanism - either to a backup file, or | |
2791 | * to some space in a spare. | |
2792 | */ | |
2793 | char *buf; | |
2794 | int degraded = -1; | |
2795 | unsigned long long speed; | |
2796 | unsigned long long suspend_point, array_size; | |
2797 | unsigned long long backup_point, wait_point; | |
2798 | unsigned long long reshape_completed; | |
2799 | int done = 0; | |
2800 | int increasing = reshape->after.data_disks >= reshape->before.data_disks; | |
2801 | int part = 0; /* The next part of the backup area to fill. It may already | |
2802 | * be full, so we need to check */ | |
2803 | int level = reshape->level; | |
2804 | int layout = reshape->before.layout; | |
2805 | int data = reshape->before.data_disks; | |
2806 | int disks = reshape->before.data_disks + reshape->parity; | |
2807 | int chunk = sra->array.chunk_size; | |
2808 | struct mdinfo *sd; | |
2809 | unsigned long stripes; | |
2810 | ||
2811 | /* set up the backup-super-block. This requires the | |
2812 | * uuid from the array. | |
2813 | */ | |
2814 | /* Find a superblock */ | |
2815 | for (sd = sra->devs; sd; sd = sd->next) { | |
2816 | char *dn; | |
2817 | int devfd; | |
2818 | int ok; | |
2819 | if (sd->disk.state & (1<<MD_DISK_FAULTY)) | |
2820 | continue; | |
2821 | dn = map_dev(sd->disk.major, sd->disk.minor, 1); | |
2822 | devfd = dev_open(dn, O_RDONLY); | |
2823 | if (devfd < 0) | |
2824 | continue; | |
2825 | ok = st->ss->load_super(st, devfd, NULL); | |
2826 | close(devfd); | |
2827 | if (ok >= 0) | |
2828 | break; | |
2829 | } | |
2830 | if (!sd) { | |
2831 | fprintf(stderr, Name ": Cannot find a superblock\n"); | |
2832 | return 0; | |
2833 | } | |
2834 | ||
2835 | memset(&bsb, 0, 512); | |
2836 | memcpy(bsb.magic, "md_backup_data-1", 16); | |
2837 | st->ss->uuid_from_super(st, (int*)&bsb.set_uuid); | |
2838 | bsb.mtime = __cpu_to_le64(time(0)); | |
2839 | bsb.devstart2 = blocks; | |
2840 | ||
2841 | stripes = blocks / (sra->array.chunk_size/512) / | |
2842 | reshape->before.data_disks; | |
2843 | ||
2844 | if (posix_memalign((void**)&buf, 4096, disks * chunk)) | |
2845 | /* Don't start the 'reshape' */ | |
2846 | return 0; | |
2847 | if (reshape->before.data_disks == reshape->after.data_disks) { | |
2848 | sysfs_get_ll(sra, NULL, "sync_speed_min", &speed); | |
2849 | sysfs_set_num(sra, NULL, "sync_speed_min", 200000); | |
2850 | } | |
2851 | ||
2852 | if (increasing) { | |
2853 | array_size = sra->component_size * reshape->after.data_disks; | |
2854 | backup_point = sra->reshape_progress; | |
2855 | suspend_point = 0; | |
2856 | } else { | |
2857 | array_size = sra->component_size * reshape->before.data_disks; | |
2858 | backup_point = reshape->backup_blocks; | |
2859 | suspend_point = array_size; | |
2860 | } | |
2861 | ||
2862 | while (!done) { | |
2863 | int rv; | |
2864 | ||
2865 | /* Want to return as soon the oldest backup slot can | |
2866 | * be released as that allows us to start backing up | |
2867 | * some more, providing suspend_point has been | |
2868 | * advanced, which it should have. | |
2869 | */ | |
2870 | if (increasing) { | |
2871 | wait_point = array_size; | |
2872 | if (part == 0 && __le64_to_cpu(bsb.length) > 0) | |
2873 | wait_point = (__le64_to_cpu(bsb.arraystart) + | |
2874 | __le64_to_cpu(bsb.length)); | |
2875 | if (part == 1 && __le64_to_cpu(bsb.length2) > 0) | |
2876 | wait_point = (__le64_to_cpu(bsb.arraystart2) + | |
2877 | __le64_to_cpu(bsb.length2)); | |
2878 | } else { | |
2879 | wait_point = 0; | |
2880 | if (part == 0 && __le64_to_cpu(bsb.length) > 0) | |
2881 | wait_point = __le64_to_cpu(bsb.arraystart); | |
2882 | if (part == 1 && __le64_to_cpu(bsb.length2) > 0) | |
2883 | wait_point = __le64_to_cpu(bsb.arraystart2); | |
2884 | } | |
2885 | ||
2886 | rv = progress_reshape(sra, reshape, | |
2887 | backup_point, wait_point, | |
2888 | &suspend_point, &reshape_completed); | |
2889 | /* external metadata would need to ping_monitor here */ | |
2890 | sra->reshape_progress = reshape_completed; | |
2891 | ||
2892 | /* Clear any backup region that is before 'here' */ | |
2893 | if (increasing) { | |
2894 | if (reshape_completed >= (__le64_to_cpu(bsb.arraystart) + | |
2895 | __le64_to_cpu(bsb.length))) | |
2896 | forget_backup(dests, destfd, | |
2897 | destoffsets, 0); | |
2898 | if (reshape_completed >= (__le64_to_cpu(bsb.arraystart2) + | |
2899 | __le64_to_cpu(bsb.length2))) | |
2900 | forget_backup(dests, destfd, | |
2901 | destoffsets, 1); | |
2902 | } else { | |
2903 | if (reshape_completed <= (__le64_to_cpu(bsb.arraystart))) | |
2904 | forget_backup(dests, destfd, | |
2905 | destoffsets, 0); | |
2906 | if (reshape_completed <= (__le64_to_cpu(bsb.arraystart2))) | |
2907 | forget_backup(dests, destfd, | |
2908 | destoffsets, 1); | |
2909 | } | |
2910 | ||
2911 | if (rv < 0) { | |
2912 | if (rv == -1) | |
2913 | done = 1; | |
2914 | break; | |
2915 | } | |
2916 | ||
2917 | while (rv) { | |
2918 | unsigned long long offset; | |
2919 | unsigned long actual_stripes; | |
2920 | /* Need to backup some data. | |
2921 | * If 'part' is not used and the desired | |
2922 | * backup size is suspended, do a backup, | |
2923 | * then consider the next part. | |
2924 | */ | |
2925 | /* Check that 'part' is unused */ | |
2926 | if (part == 0 && __le64_to_cpu(bsb.length) != 0) | |
2927 | break; | |
2928 | if (part == 1 && __le64_to_cpu(bsb.length2) != 0) | |
2929 | break; | |
2930 | ||
2931 | offset = backup_point / data; | |
2932 | actual_stripes = stripes; | |
2933 | if (increasing) { | |
2934 | if (offset + actual_stripes * (chunk/512) > | |
2935 | sra->component_size) | |
2936 | actual_stripes = ((sra->component_size - offset) | |
2937 | / (chunk/512)); | |
2938 | if (offset + actual_stripes * (chunk/512) > | |
2939 | suspend_point/data) | |
2940 | break; | |
2941 | } else { | |
2942 | if (offset < actual_stripes * (chunk/512)) | |
2943 | actual_stripes = offset / (chunk/512); | |
2944 | offset -= actual_stripes * (chunk/512); | |
2945 | if (offset < suspend_point/data) | |
2946 | break; | |
2947 | } | |
2948 | if (actual_stripes == 0) | |
2949 | break; | |
2950 | grow_backup(sra, offset, actual_stripes, | |
2951 | fds, offsets, | |
2952 | disks, chunk, level, layout, | |
2953 | dests, destfd, destoffsets, | |
2954 | part, °raded, buf); | |
2955 | validate(afd, destfd[0], destoffsets[0]); | |
2956 | /* record where 'part' is up to */ | |
2957 | part = !part; | |
2958 | if (increasing) | |
2959 | backup_point += actual_stripes * (chunk/512) * data; | |
2960 | else | |
2961 | backup_point -= actual_stripes * (chunk/512) * data; | |
2962 | } | |
2963 | } | |
2964 | ||
2965 | /* FIXME maybe call progress_reshape one more time instead */ | |
2966 | abort_reshape(sra); /* remove any remaining suspension */ | |
2967 | if (reshape->before.data_disks == reshape->after.data_disks) | |
2968 | sysfs_set_num(sra, NULL, "sync_speed_min", speed); | |
2969 | free(buf); | |
2970 | return done; | |
2971 | } | |
2972 | ||
2973 | /* | |
2974 | * If any spare contains md_back_data-1 which is recent wrt mtime, | |
2975 | * write that data into the array and update the super blocks with | |
2976 | * the new reshape_progress | |
2977 | */ | |
2978 | int Grow_restart(struct supertype *st, struct mdinfo *info, int *fdlist, int cnt, | |
2979 | char *backup_file, int verbose) | |
2980 | { | |
2981 | int i, j; | |
2982 | int old_disks; | |
2983 | unsigned long long *offsets; | |
2984 | unsigned long long nstripe, ostripe; | |
2985 | int ndata, odata; | |
2986 | ||
2987 | if (info->new_level != info->array.level) | |
2988 | return 1; /* Cannot handle level changes (they are instantaneous) */ | |
2989 | ||
2990 | odata = info->array.raid_disks - info->delta_disks - 1; | |
2991 | if (info->array.level == 6) odata--; /* number of data disks */ | |
2992 | ndata = info->array.raid_disks - 1; | |
2993 | if (info->new_level == 6) ndata--; | |
2994 | ||
2995 | old_disks = info->array.raid_disks - info->delta_disks; | |
2996 | ||
2997 | if (info->delta_disks <= 0) | |
2998 | /* Didn't grow, so the backup file must have | |
2999 | * been used | |
3000 | */ | |
3001 | old_disks = cnt; | |
3002 | for (i=old_disks-(backup_file?1:0); i<cnt; i++) { | |
3003 | struct mdinfo dinfo; | |
3004 | int fd; | |
3005 | int bsbsize; | |
3006 | char *devname, namebuf[20]; | |
3007 | unsigned long long lo, hi; | |
3008 | ||
3009 | /* This was a spare and may have some saved data on it. | |
3010 | * Load the superblock, find and load the | |
3011 | * backup_super_block. | |
3012 | * If either fail, go on to next device. | |
3013 | * If the backup contains no new info, just return | |
3014 | * else restore data and update all superblocks | |
3015 | */ | |
3016 | if (i == old_disks-1) { | |
3017 | fd = open(backup_file, O_RDONLY); | |
3018 | if (fd<0) { | |
3019 | fprintf(stderr, Name ": backup file %s inaccessible: %s\n", | |
3020 | backup_file, strerror(errno)); | |
3021 | continue; | |
3022 | } | |
3023 | devname = backup_file; | |
3024 | } else { | |
3025 | fd = fdlist[i]; | |
3026 | if (fd < 0) | |
3027 | continue; | |
3028 | if (st->ss->load_super(st, fd, NULL)) | |
3029 | continue; | |
3030 | ||
3031 | st->ss->getinfo_super(st, &dinfo, NULL); | |
3032 | st->ss->free_super(st); | |
3033 | ||
3034 | if (lseek64(fd, | |
3035 | (dinfo.data_offset + dinfo.component_size - 8) <<9, | |
3036 | 0) < 0) { | |
3037 | fprintf(stderr, Name ": Cannot seek on device %d\n", i); | |
3038 | continue; /* Cannot seek */ | |
3039 | } | |
3040 | sprintf(namebuf, "device-%d", i); | |
3041 | devname = namebuf; | |
3042 | } | |
3043 | if (read(fd, &bsb, sizeof(bsb)) != sizeof(bsb)) { | |
3044 | if (verbose) | |
3045 | fprintf(stderr, Name ": Cannot read from %s\n", devname); | |
3046 | continue; /* Cannot read */ | |
3047 | } | |
3048 | if (memcmp(bsb.magic, "md_backup_data-1", 16) != 0 && | |
3049 | memcmp(bsb.magic, "md_backup_data-2", 16) != 0) { | |
3050 | if (verbose) | |
3051 | fprintf(stderr, Name ": No backup metadata on %s\n", devname); | |
3052 | continue; | |
3053 | } | |
3054 | if (bsb.sb_csum != bsb_csum((char*)&bsb, ((char*)&bsb.sb_csum)-((char*)&bsb))) { | |
3055 | if (verbose) | |
3056 | fprintf(stderr, Name ": Bad backup-metadata checksum on %s\n", devname); | |
3057 | continue; /* bad checksum */ | |
3058 | } | |
3059 | if (memcmp(bsb.magic, "md_backup_data-2", 16) == 0 && | |
3060 | bsb.sb_csum2 != bsb_csum((char*)&bsb, ((char*)&bsb.sb_csum2)-((char*)&bsb))) { | |
3061 | if (verbose) | |
3062 | fprintf(stderr, Name ": Bad backup-metadata checksum2 on %s\n", devname); | |
3063 | continue; /* Bad second checksum */ | |
3064 | } | |
3065 | if (memcmp(bsb.set_uuid,info->uuid, 16) != 0) { | |
3066 | if (verbose) | |
3067 | fprintf(stderr, Name ": Wrong uuid on backup-metadata on %s\n", devname); | |
3068 | continue; /* Wrong uuid */ | |
3069 | } | |
3070 | ||
3071 | /* array utime and backup-mtime should be updated at much the same time, but it seems that | |
3072 | * sometimes they aren't... So allow considerable flexability in matching, and allow | |
3073 | * this test to be overridden by an environment variable. | |
3074 | */ | |
3075 | if (info->array.utime > (int)__le64_to_cpu(bsb.mtime) + 2*60*60 || | |
3076 | info->array.utime < (int)__le64_to_cpu(bsb.mtime) - 10*60) { | |
3077 | if (check_env("MDADM_GROW_ALLOW_OLD")) { | |
3078 | fprintf(stderr, Name ": accepting backup with timestamp %lu " | |
3079 | "for array with timestamp %lu\n", | |
3080 | (unsigned long)__le64_to_cpu(bsb.mtime), | |
3081 | (unsigned long)info->array.utime); | |
3082 | } else { | |
3083 | if (verbose) | |
3084 | fprintf(stderr, Name ": too-old timestamp on " | |
3085 | "backup-metadata on %s\n", devname); | |
3086 | continue; /* time stamp is too bad */ | |
3087 | } | |
3088 | } | |
3089 | ||
3090 | if (bsb.magic[15] == '1') { | |
3091 | if (bsb.length == 0) | |
3092 | continue; | |
3093 | if (info->delta_disks >= 0) { | |
3094 | /* reshape_progress is increasing */ | |
3095 | if (__le64_to_cpu(bsb.arraystart) | |
3096 | + __le64_to_cpu(bsb.length) | |
3097 | < info->reshape_progress) { | |
3098 | nonew: | |
3099 | if (verbose) | |
3100 | fprintf(stderr, Name | |
3101 | ": backup-metadata found on %s but is not needed\n", devname); | |
3102 | continue; /* No new data here */ | |
3103 | } | |
3104 | } else { | |
3105 | /* reshape_progress is decreasing */ | |
3106 | if (__le64_to_cpu(bsb.arraystart) >= | |
3107 | info->reshape_progress) | |
3108 | goto nonew; /* No new data here */ | |
3109 | } | |
3110 | } else { | |
3111 | if (bsb.length == 0 && bsb.length2 == 0) | |
3112 | continue; | |
3113 | if (info->delta_disks >= 0) { | |
3114 | /* reshape_progress is increasing */ | |
3115 | if ((__le64_to_cpu(bsb.arraystart) | |
3116 | + __le64_to_cpu(bsb.length) | |
3117 | < info->reshape_progress) | |
3118 | && | |
3119 | (__le64_to_cpu(bsb.arraystart2) | |
3120 | + __le64_to_cpu(bsb.length2) | |
3121 | < info->reshape_progress)) | |
3122 | goto nonew; /* No new data here */ | |
3123 | } else { | |
3124 | /* reshape_progress is decreasing */ | |
3125 | if (__le64_to_cpu(bsb.arraystart) >= | |
3126 | info->reshape_progress && | |
3127 | __le64_to_cpu(bsb.arraystart2) >= | |
3128 | info->reshape_progress) | |
3129 | goto nonew; /* No new data here */ | |
3130 | } | |
3131 | } | |
3132 | if (lseek64(fd, __le64_to_cpu(bsb.devstart)*512, 0)< 0) { | |
3133 | second_fail: | |
3134 | if (verbose) | |
3135 | fprintf(stderr, Name | |
3136 | ": Failed to verify secondary backup-metadata block on %s\n", | |
3137 | devname); | |
3138 | continue; /* Cannot seek */ | |
3139 | } | |
3140 | /* There should be a duplicate backup superblock 4k before here */ | |
3141 | if (lseek64(fd, -4096, 1) < 0 || | |
3142 | read(fd, &bsb2, sizeof(bsb2)) != sizeof(bsb2)) | |
3143 | goto second_fail; /* Cannot find leading superblock */ | |
3144 | if (bsb.magic[15] == '1') | |
3145 | bsbsize = offsetof(struct mdp_backup_super, pad1); | |
3146 | else | |
3147 | bsbsize = offsetof(struct mdp_backup_super, pad); | |
3148 | if (memcmp(&bsb2, &bsb, bsbsize) != 0) | |
3149 | goto second_fail; /* Cannot find leading superblock */ | |
3150 | ||
3151 | /* Now need the data offsets for all devices. */ | |
3152 | offsets = malloc(sizeof(*offsets)*info->array.raid_disks); | |
3153 | for(j=0; j<info->array.raid_disks; j++) { | |
3154 | if (fdlist[j] < 0) | |
3155 | continue; | |
3156 | if (st->ss->load_super(st, fdlist[j], NULL)) | |
3157 | /* FIXME should be this be an error */ | |
3158 | continue; | |
3159 | st->ss->getinfo_super(st, &dinfo, NULL); | |
3160 | st->ss->free_super(st); | |
3161 | offsets[j] = dinfo.data_offset * 512; | |
3162 | } | |
3163 | printf(Name ": restoring critical section\n"); | |
3164 | ||
3165 | if (restore_stripes(fdlist, offsets, | |
3166 | info->array.raid_disks, | |
3167 | info->new_chunk, | |
3168 | info->new_level, | |
3169 | info->new_layout, | |
3170 | fd, __le64_to_cpu(bsb.devstart)*512, | |
3171 | __le64_to_cpu(bsb.arraystart)*512, | |
3172 | __le64_to_cpu(bsb.length)*512)) { | |
3173 | /* didn't succeed, so giveup */ | |
3174 | if (verbose) | |
3175 | fprintf(stderr, Name ": Error restoring backup from %s\n", | |
3176 | devname); | |
3177 | return 1; | |
3178 | } | |
3179 | ||
3180 | if (bsb.magic[15] == '2' && | |
3181 | restore_stripes(fdlist, offsets, | |
3182 | info->array.raid_disks, | |
3183 | info->new_chunk, | |
3184 | info->new_level, | |
3185 | info->new_layout, | |
3186 | fd, __le64_to_cpu(bsb.devstart)*512 + | |
3187 | __le64_to_cpu(bsb.devstart2)*512, | |
3188 | __le64_to_cpu(bsb.arraystart2)*512, | |
3189 | __le64_to_cpu(bsb.length2)*512)) { | |
3190 | /* didn't succeed, so giveup */ | |
3191 | if (verbose) | |
3192 | fprintf(stderr, Name ": Error restoring second backup from %s\n", | |
3193 | devname); | |
3194 | return 1; | |
3195 | } | |
3196 | ||
3197 | ||
3198 | /* Ok, so the data is restored. Let's update those superblocks. */ | |
3199 | ||
3200 | lo = hi = 0; | |
3201 | if (bsb.length) { | |
3202 | lo = __le64_to_cpu(bsb.arraystart); | |
3203 | hi = lo + __le64_to_cpu(bsb.length); | |
3204 | } | |
3205 | if (bsb.magic[15] == '2' && bsb.length2) { | |
3206 | unsigned long long lo1, hi1; | |
3207 | lo1 = __le64_to_cpu(bsb.arraystart2); | |
3208 | hi1 = lo1 + __le64_to_cpu(bsb.length2); | |
3209 | if (lo == hi) { | |
3210 | lo = lo1; | |
3211 | hi = hi1; | |
3212 | } else if (lo < lo1) | |
3213 | hi = hi1; | |
3214 | else | |
3215 | lo = lo1; | |
3216 | } | |
3217 | if (lo < hi && | |
3218 | (info->reshape_progress < lo || | |
3219 | info->reshape_progress > hi)) | |
3220 | /* backup does not affect reshape_progress*/ ; | |
3221 | else if (info->delta_disks >= 0) { | |
3222 | info->reshape_progress = __le64_to_cpu(bsb.arraystart) + | |
3223 | __le64_to_cpu(bsb.length); | |
3224 | if (bsb.magic[15] == '2') { | |
3225 | unsigned long long p2 = __le64_to_cpu(bsb.arraystart2) + | |
3226 | __le64_to_cpu(bsb.length2); | |
3227 | if (p2 > info->reshape_progress) | |
3228 | info->reshape_progress = p2; | |
3229 | } | |
3230 | } else { | |
3231 | info->reshape_progress = __le64_to_cpu(bsb.arraystart); | |
3232 | if (bsb.magic[15] == '2') { | |
3233 | unsigned long long p2 = __le64_to_cpu(bsb.arraystart2); | |
3234 | if (p2 < info->reshape_progress) | |
3235 | info->reshape_progress = p2; | |
3236 | } | |
3237 | } | |
3238 | for (j=0; j<info->array.raid_disks; j++) { | |
3239 | if (fdlist[j] < 0) continue; | |
3240 | if (st->ss->load_super(st, fdlist[j], NULL)) | |
3241 | continue; | |
3242 | st->ss->getinfo_super(st, &dinfo, NULL); | |
3243 | dinfo.reshape_progress = info->reshape_progress; | |
3244 | st->ss->update_super(st, &dinfo, | |
3245 | "_reshape_progress", | |
3246 | NULL,0, 0, NULL); | |
3247 | st->ss->store_super(st, fdlist[j]); | |
3248 | st->ss->free_super(st); | |
3249 | } | |
3250 | return 0; | |
3251 | } | |
3252 | /* Didn't find any backup data, try to see if any | |
3253 | * was needed. | |
3254 | */ | |
3255 | if (info->delta_disks < 0) { | |
3256 | /* When shrinking, the critical section is at the end. | |
3257 | * So see if we are before the critical section. | |
3258 | */ | |
3259 | unsigned long long first_block; | |
3260 | nstripe = ostripe = 0; | |
3261 | first_block = 0; | |
3262 | while (ostripe >= nstripe) { | |
3263 | ostripe += info->array.chunk_size / 512; | |
3264 | first_block = ostripe * odata; | |
3265 | nstripe = first_block / ndata / (info->new_chunk/512) * | |
3266 | (info->new_chunk/512); | |
3267 | } | |
3268 | ||
3269 | if (info->reshape_progress >= first_block) | |
3270 | return 0; | |
3271 | } | |
3272 | if (info->delta_disks > 0) { | |
3273 | /* See if we are beyond the critical section. */ | |
3274 | unsigned long long last_block; | |
3275 | nstripe = ostripe = 0; | |
3276 | last_block = 0; | |
3277 | while (nstripe >= ostripe) { | |
3278 | nstripe += info->new_chunk / 512; | |
3279 | last_block = nstripe * ndata; | |
3280 | ostripe = last_block / odata / (info->array.chunk_size/512) * | |
3281 | (info->array.chunk_size/512); | |
3282 | } | |
3283 | ||
3284 | if (info->reshape_progress >= last_block) | |
3285 | return 0; | |
3286 | } | |
3287 | /* needed to recover critical section! */ | |
3288 | if (verbose) | |
3289 | fprintf(stderr, Name ": Failed to find backup of critical section\n"); | |
3290 | return 1; | |
3291 | } | |
3292 | ||
3293 | int Grow_continue(int mdfd, struct supertype *st, struct mdinfo *info, | |
3294 | char *backup_file) | |
3295 | { | |
3296 | int err = sysfs_set_str(info, NULL, "array_state", "readonly"); | |
3297 | if (err) | |
3298 | return err; | |
3299 | return reshape_array(NULL, mdfd, "array", st, info, 1, backup_file, 0, 0, 1); | |
3300 | } | |
3301 | ||
3302 |