]> git.ipfire.org Git - thirdparty/mdadm.git/blob - Grow.c
projects / thirdparty / mdadm.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
mdadm: add map_num_s()
[thirdparty/mdadm.git] / Grow.c
1 /*
2 * mdadm - manage Linux "md" devices aka RAID arrays.
3 *
4 * Copyright (C) 2001-2013 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 #include <stddef.h>
28 #include <stdint.h>
29 #include <sys/wait.h>
30
31 #if ! defined(__BIG_ENDIAN) && ! defined(__LITTLE_ENDIAN)
32 #error no endian defined
33 #endif
34 #include "md_u.h"
35 #include "md_p.h"
36
37 int restore_backup(struct supertype *st,
38 struct mdinfo *content,
39 int working_disks,
40 int next_spare,
41 char **backup_filep,
42 int verbose)
43 {
44 int i;
45 int *fdlist;
46 struct mdinfo *dev;
47 int err;
48 int disk_count = next_spare + working_disks;
49 char *backup_file = *backup_filep;
50
51 dprintf("Called restore_backup()\n");
52 fdlist = xmalloc(sizeof(int) * disk_count);
53
54 enable_fds(next_spare);
55 for (i = 0; i < next_spare; i++)
56 fdlist[i] = -1;
57 for (dev = content->devs; dev; dev = dev->next) {
58 char buf[22];
59 int fd;
60
61 sprintf(buf, "%d:%d", dev->disk.major, dev->disk.minor);
62 fd = dev_open(buf, O_RDWR);
63
64 if (dev->disk.raid_disk >= 0)
65 fdlist[dev->disk.raid_disk] = fd;
66 else
67 fdlist[next_spare++] = fd;
68 }
69
70 if (!backup_file) {
71 backup_file = locate_backup(content->sys_name);
72 *backup_filep = backup_file;
73 }
74
75 if (st->ss->external && st->ss->recover_backup)
76 err = st->ss->recover_backup(st, content);
77 else
78 err = Grow_restart(st, content, fdlist, next_spare,
79 backup_file, verbose > 0);
80
81 while (next_spare > 0) {
82 next_spare--;
83 if (fdlist[next_spare] >= 0)
84 close(fdlist[next_spare]);
85 }
86 free(fdlist);
87 if (err) {
88 pr_err("Failed to restore critical section for reshape - sorry.\n");
89 if (!backup_file)
90 pr_err("Possibly you need to specify a --backup-file\n");
91 return 1;
92 }
93
94 dprintf("restore_backup() returns status OK.\n");
95 return 0;
96 }
97
98 int Grow_Add_device(char *devname, int fd, char *newdev)
99 {
100 /* Add a device to an active array.
101 * Currently, just extend a linear array.
102 * This requires writing a new superblock on the
103 * new device, calling the kernel to add the device,
104 * and if that succeeds, update the superblock on
105 * all other devices.
106 * This means that we need to *find* all other devices.
107 */
108 struct mdinfo info;
109
110 dev_t rdev;
111 int nfd, fd2;
112 int d, nd;
113 struct supertype *st = NULL;
114 char *subarray = NULL;
115
116 if (md_get_array_info(fd, &info.array) < 0) {
117 pr_err("cannot get array info for %s\n", devname);
118 return 1;
119 }
120
121 if (info.array.level != -1) {
122 pr_err("can only add devices to linear arrays\n");
123 return 1;
124 }
125
126 st = super_by_fd(fd, &subarray);
127 if (!st) {
128 pr_err("cannot handle arrays with superblock version %d\n",
129 info.array.major_version);
130 return 1;
131 }
132
133 if (subarray) {
134 pr_err("Cannot grow linear sub-arrays yet\n");
135 free(subarray);
136 free(st);
137 return 1;
138 }
139
140 nfd = open(newdev, O_RDWR|O_EXCL|O_DIRECT);
141 if (nfd < 0) {
142 pr_err("cannot open %s\n", newdev);
143 free(st);
144 return 1;
145 }
146 if (!fstat_is_blkdev(nfd, newdev, &rdev)) {
147 close(nfd);
148 free(st);
149 return 1;
150 }
151 /* now check out all the devices and make sure we can read the
152 * superblock */
153 for (d=0 ; d < info.array.raid_disks ; d++) {
154 mdu_disk_info_t disk;
155 char *dv;
156
157 st->ss->free_super(st);
158
159 disk.number = d;
160 if (md_get_disk_info(fd, &disk) < 0) {
161 pr_err("cannot get device detail for device %d\n", d);
162 close(nfd);
163 free(st);
164 return 1;
165 }
166 dv = map_dev(disk.major, disk.minor, 1);
167 if (!dv) {
168 pr_err("cannot find device file for device %d\n", d);
169 close(nfd);
170 free(st);
171 return 1;
172 }
173 fd2 = dev_open(dv, O_RDWR);
174 if (fd2 < 0) {
175 pr_err("cannot open device file %s\n", dv);
176 close(nfd);
177 free(st);
178 return 1;
179 }
180
181 if (st->ss->load_super(st, fd2, NULL)) {
182 pr_err("cannot find super block on %s\n", dv);
183 close(nfd);
184 close(fd2);
185 free(st);
186 return 1;
187 }
188 close(fd2);
189 }
190 /* Ok, looks good. Lets update the superblock and write it out to
191 * newdev.
192 */
193
194 info.disk.number = d;
195 info.disk.major = major(rdev);
196 info.disk.minor = minor(rdev);
197 info.disk.raid_disk = d;
198 info.disk.state = (1 << MD_DISK_SYNC) | (1 << MD_DISK_ACTIVE);
199 if (st->ss->update_super(st, &info, "linear-grow-new", newdev,
200 0, 0, NULL) != 0) {
201 pr_err("Preparing new metadata failed on %s\n", newdev);
202 close(nfd);
203 return 1;
204 }
205
206 if (st->ss->store_super(st, nfd)) {
207 pr_err("Cannot store new superblock on %s\n", newdev);
208 close(nfd);
209 return 1;
210 }
211 close(nfd);
212
213 if (ioctl(fd, ADD_NEW_DISK, &info.disk) != 0) {
214 pr_err("Cannot add new disk to this array\n");
215 return 1;
216 }
217 /* Well, that seems to have worked.
218 * Now go through and update all superblocks
219 */
220
221 if (md_get_array_info(fd, &info.array) < 0) {
222 pr_err("cannot get array info for %s\n", devname);
223 return 1;
224 }
225
226 nd = d;
227 for (d=0 ; d < info.array.raid_disks ; d++) {
228 mdu_disk_info_t disk;
229 char *dv;
230
231 disk.number = d;
232 if (md_get_disk_info(fd, &disk) < 0) {
233 pr_err("cannot get device detail for device %d\n", d);
234 return 1;
235 }
236 dv = map_dev(disk.major, disk.minor, 1);
237 if (!dv) {
238 pr_err("cannot find device file for device %d\n", d);
239 return 1;
240 }
241 fd2 = dev_open(dv, O_RDWR);
242 if (fd2 < 0) {
243 pr_err("cannot open device file %s\n", dv);
244 return 1;
245 }
246 if (st->ss->load_super(st, fd2, NULL)) {
247 pr_err("cannot find super block on %s\n", dv);
248 close(fd);
249 close(fd2);
250 return 1;
251 }
252 info.array.raid_disks = nd+1;
253 info.array.nr_disks = nd+1;
254 info.array.active_disks = nd+1;
255 info.array.working_disks = nd+1;
256
257 if (st->ss->update_super(st, &info, "linear-grow-update", dv,
258 0, 0, NULL) != 0) {
259 pr_err("Updating metadata failed on %s\n", dv);
260 close(fd2);
261 return 1;
262 }
263
264 if (st->ss->store_super(st, fd2)) {
265 pr_err("Cannot store new superblock on %s\n", dv);
266 close(fd2);
267 return 1;
268 }
269 close(fd2);
270 }
271
272 return 0;
273 }
274
275 int Grow_addbitmap(char *devname, int fd, struct context *c, struct shape *s)
276 {
277 /*
278 * First check that array doesn't have a bitmap
279 * Then create the bitmap
280 * Then add it
281 *
282 * For internal bitmaps, we need to check the version,
283 * find all the active devices, and write the bitmap block
284 * to all devices
285 */
286 mdu_bitmap_file_t bmf;
287 mdu_array_info_t array;
288 struct supertype *st;
289 char *subarray = NULL;
290 int major = BITMAP_MAJOR_HI;
291 unsigned long long bitmapsize, array_size;
292 struct mdinfo *mdi;
293
294 /*
295 * We only ever get called if s->bitmap_file is != NULL, so this check
296 * is just here to quiet down static code checkers.
297 */
298 if (!s->bitmap_file)
299 return 1;
300
301 if (strcmp(s->bitmap_file, "clustered") == 0)
302 major = BITMAP_MAJOR_CLUSTERED;
303
304 if (ioctl(fd, GET_BITMAP_FILE, &bmf) != 0) {
305 if (errno == ENOMEM)
306 pr_err("Memory allocation failure.\n");
307 else
308 pr_err("bitmaps not supported by this kernel.\n");
309 return 1;
310 }
311 if (bmf.pathname[0]) {
312 if (strcmp(s->bitmap_file,"none") == 0) {
313 if (ioctl(fd, SET_BITMAP_FILE, -1) != 0) {
314 pr_err("failed to remove bitmap %s\n",
315 bmf.pathname);
316 return 1;
317 }
318 return 0;
319 }
320 pr_err("%s already has a bitmap (%s)\n", devname, bmf.pathname);
321 return 1;
322 }
323 if (md_get_array_info(fd, &array) != 0) {
324 pr_err("cannot get array status for %s\n", devname);
325 return 1;
326 }
327 if (array.state & (1 << MD_SB_BITMAP_PRESENT)) {
328 if (strcmp(s->bitmap_file, "none")==0) {
329 array.state &= ~(1 << MD_SB_BITMAP_PRESENT);
330 if (md_set_array_info(fd, &array) != 0) {
331 if (array.state & (1 << MD_SB_CLUSTERED))
332 pr_err("failed to remove clustered bitmap.\n");
333 else
334 pr_err("failed to remove internal bitmap.\n");
335 return 1;
336 }
337 return 0;
338 }
339 pr_err("bitmap already present on %s\n", devname);
340 return 1;
341 }
342
343 if (strcmp(s->bitmap_file, "none") == 0) {
344 pr_err("no bitmap found on %s\n", devname);
345 return 1;
346 }
347 if (array.level <= 0) {
348 pr_err("Bitmaps not meaningful with level %s\n",
349 map_num(pers, array.level)?:"of this array");
350 return 1;
351 }
352 bitmapsize = array.size;
353 bitmapsize <<= 1;
354 if (get_dev_size(fd, NULL, &array_size) &&
355 array_size > (0x7fffffffULL << 9)) {
356 /* Array is big enough that we cannot trust array.size
357 * try other approaches
358 */
359 bitmapsize = get_component_size(fd);
360 }
361 if (bitmapsize == 0) {
362 pr_err("Cannot reliably determine size of array to create bitmap - sorry.\n");
363 return 1;
364 }
365
366 if (array.level == 10) {
367 int ncopies;
368
369 ncopies = (array.layout & 255) * ((array.layout >> 8) & 255);
370 bitmapsize = bitmapsize * array.raid_disks / ncopies;
371
372 if (strcmp(s->bitmap_file, "clustered") == 0 &&
373 !is_near_layout_10(array.layout)) {
374 pr_err("only near layout is supported with clustered raid10\n");
375 return 1;
376 }
377 }
378
379 st = super_by_fd(fd, &subarray);
380 if (!st) {
381 pr_err("Cannot understand version %d.%d\n",
382 array.major_version, array.minor_version);
383 return 1;
384 }
385 if (subarray) {
386 pr_err("Cannot add bitmaps to sub-arrays yet\n");
387 free(subarray);
388 free(st);
389 return 1;
390 }
391
392 mdi = sysfs_read(fd, NULL, GET_CONSISTENCY_POLICY);
393 if (mdi) {
394 if (mdi->consistency_policy == CONSISTENCY_POLICY_PPL) {
395 pr_err("Cannot add bitmap to array with PPL\n");
396 free(mdi);
397 free(st);
398 return 1;
399 }
400 free(mdi);
401 }
402
403 if (strcmp(s->bitmap_file, "internal") == 0 ||
404 strcmp(s->bitmap_file, "clustered") == 0) {
405 int rv;
406 int d;
407 int offset_setable = 0;
408 if (st->ss->add_internal_bitmap == NULL) {
409 pr_err("Internal bitmaps not supported with %s metadata\n", st->ss->name);
410 return 1;
411 }
412 st->nodes = c->nodes;
413 st->cluster_name = c->homecluster;
414 mdi = sysfs_read(fd, NULL, GET_BITMAP_LOCATION);
415 if (mdi)
416 offset_setable = 1;
417 for (d = 0; d < st->max_devs; d++) {
418 mdu_disk_info_t disk;
419 char *dv;
420 int fd2;
421
422 disk.number = d;
423 if (md_get_disk_info(fd, &disk) < 0)
424 continue;
425 if (disk.major == 0 && disk.minor == 0)
426 continue;
427 if ((disk.state & (1 << MD_DISK_SYNC)) == 0)
428 continue;
429 dv = map_dev(disk.major, disk.minor, 1);
430 if (!dv)
431 continue;
432 if (((disk.state & (1 << MD_DISK_WRITEMOSTLY)) == 0) &&
433 (strcmp(s->bitmap_file, "clustered") == 0)) {
434 pr_err("%s disks marked write-mostly are not supported with clustered bitmap\n",devname);
435 return 1;
436 }
437 fd2 = dev_open(dv, O_RDWR);
438 if (fd2 < 0)
439 continue;
440 rv = st->ss->load_super(st, fd2, NULL);
441 if (!rv) {
442 rv = st->ss->add_internal_bitmap(
443 st, &s->bitmap_chunk, c->delay,
444 s->write_behind, bitmapsize,
445 offset_setable, major);
446 if (!rv) {
447 st->ss->write_bitmap(st, fd2,
448 NodeNumUpdate);
449 } else {
450 pr_err("failed to create internal bitmap - chunksize problem.\n");
451 }
452 } else {
453 pr_err("failed to load super-block.\n");
454 }
455 close(fd2);
456 if (rv)
457 return 1;
458 }
459 if (offset_setable) {
460 st->ss->getinfo_super(st, mdi, NULL);
461 if (sysfs_init(mdi, fd, NULL)) {
462 pr_err("failed to initialize sysfs.\n");
463 free(mdi);
464 }
465 rv = sysfs_set_num_signed(mdi, NULL, "bitmap/location",
466 mdi->bitmap_offset);
467 free(mdi);
468 } else {
469 if (strcmp(s->bitmap_file, "clustered") == 0)
470 array.state |= (1 << MD_SB_CLUSTERED);
471 array.state |= (1 << MD_SB_BITMAP_PRESENT);
472 rv = md_set_array_info(fd, &array);
473 }
474 if (rv < 0) {
475 if (errno == EBUSY)
476 pr_err("Cannot add bitmap while array is resyncing or reshaping etc.\n");
477 pr_err("failed to set internal bitmap.\n");
478 return 1;
479 }
480 } else {
481 int uuid[4];
482 int bitmap_fd;
483 int d;
484 int max_devs = st->max_devs;
485
486 /* try to load a superblock */
487 for (d = 0; d < max_devs; d++) {
488 mdu_disk_info_t disk;
489 char *dv;
490 int fd2;
491 disk.number = d;
492 if (md_get_disk_info(fd, &disk) < 0)
493 continue;
494 if ((disk.major==0 && disk.minor == 0) ||
495 (disk.state & (1 << MD_DISK_REMOVED)))
496 continue;
497 dv = map_dev(disk.major, disk.minor, 1);
498 if (!dv)
499 continue;
500 fd2 = dev_open(dv, O_RDONLY);
501 if (fd2 >= 0) {
502 if (st->ss->load_super(st, fd2, NULL) == 0) {
503 close(fd2);
504 st->ss->uuid_from_super(st, uuid);
505 break;
506 }
507 close(fd2);
508 }
509 }
510 if (d == max_devs) {
511 pr_err("cannot find UUID for array!\n");
512 return 1;
513 }
514 if (CreateBitmap(s->bitmap_file, c->force, (char*)uuid,
515 s->bitmap_chunk, c->delay, s->write_behind,
516 bitmapsize, major)) {
517 return 1;
518 }
519 bitmap_fd = open(s->bitmap_file, O_RDWR);
520 if (bitmap_fd < 0) {
521 pr_err("weird: %s cannot be opened\n", s->bitmap_file);
522 return 1;
523 }
524 if (ioctl(fd, SET_BITMAP_FILE, bitmap_fd) < 0) {
525 int err = errno;
526 if (errno == EBUSY)
527 pr_err("Cannot add bitmap while array is resyncing or reshaping etc.\n");
528 pr_err("Cannot set bitmap file for %s: %s\n",
529 devname, strerror(err));
530 return 1;
531 }
532 }
533
534 return 0;
535 }
536
537 int Grow_consistency_policy(char *devname, int fd, struct context *c, struct shape *s)
538 {
539 struct supertype *st;
540 struct mdinfo *sra;
541 struct mdinfo *sd;
542 char *subarray = NULL;
543 int ret = 0;
544 char container_dev[PATH_MAX];
545 char buf[20];
546
547 if (s->consistency_policy != CONSISTENCY_POLICY_RESYNC &&
548 s->consistency_policy != CONSISTENCY_POLICY_PPL) {
549 pr_err("Operation not supported for consistency policy %s\n",
550 map_num_s(consistency_policies, s->consistency_policy));
551 return 1;
552 }
553
554 st = super_by_fd(fd, &subarray);
555 if (!st)
556 return 1;
557
558 sra = sysfs_read(fd, NULL, GET_CONSISTENCY_POLICY|GET_LEVEL|
559 GET_DEVS|GET_STATE);
560 if (!sra) {
561 ret = 1;
562 goto free_st;
563 }
564
565 if (s->consistency_policy == CONSISTENCY_POLICY_PPL &&
566 !st->ss->write_init_ppl) {
567 pr_err("%s metadata does not support PPL\n", st->ss->name);
568 ret = 1;
569 goto free_info;
570 }
571
572 if (sra->array.level != 5) {
573 pr_err("Operation not supported for array level %d\n",
574 sra->array.level);
575 ret = 1;
576 goto free_info;
577 }
578
579 if (sra->consistency_policy == (unsigned)s->consistency_policy) {
580 pr_err("Consistency policy is already %s\n",
581 map_num_s(consistency_policies, s->consistency_policy));
582 ret = 1;
583 goto free_info;
584 } else if (sra->consistency_policy != CONSISTENCY_POLICY_RESYNC &&
585 sra->consistency_policy != CONSISTENCY_POLICY_PPL) {
586 pr_err("Current consistency policy is %s, cannot change to %s\n",
587 map_num_s(consistency_policies, sra->consistency_policy),
588 map_num_s(consistency_policies, s->consistency_policy));
589 ret = 1;
590 goto free_info;
591 }
592
593 if (s->consistency_policy == CONSISTENCY_POLICY_PPL) {
594 if (sysfs_get_str(sra, NULL, "sync_action", buf, 20) <= 0) {
595 ret = 1;
596 goto free_info;
597 } else if (strcmp(buf, "reshape\n") == 0) {
598 pr_err("PPL cannot be enabled when reshape is in progress\n");
599 ret = 1;
600 goto free_info;
601 }
602 }
603
604 if (subarray) {
605 char *update;
606
607 if (s->consistency_policy == CONSISTENCY_POLICY_PPL)
608 update = "ppl";
609 else
610 update = "no-ppl";
611
612 sprintf(container_dev, "/dev/%s", st->container_devnm);
613
614 ret = Update_subarray(container_dev, subarray, update, NULL,
615 c->verbose);
616 if (ret)
617 goto free_info;
618 }
619
620 if (s->consistency_policy == CONSISTENCY_POLICY_PPL) {
621 struct mdinfo info;
622
623 if (subarray) {
624 struct mdinfo *mdi;
625 int cfd;
626
627 cfd = open(container_dev, O_RDWR|O_EXCL);
628 if (cfd < 0) {
629 pr_err("Failed to open %s\n", container_dev);
630 ret = 1;
631 goto free_info;
632 }
633
634 ret = st->ss->load_container(st, cfd, st->container_devnm);
635 close(cfd);
636
637 if (ret) {
638 pr_err("Cannot read superblock for %s\n",
639 container_dev);
640 goto free_info;
641 }
642
643 mdi = st->ss->container_content(st, subarray);
644 info = *mdi;
645 free(mdi);
646 }
647
648 for (sd = sra->devs; sd; sd = sd->next) {
649 int dfd;
650 char *devpath;
651
652 devpath = map_dev(sd->disk.major, sd->disk.minor, 0);
653 dfd = dev_open(devpath, O_RDWR);
654 if (dfd < 0) {
655 pr_err("Failed to open %s\n", devpath);
656 ret = 1;
657 goto free_info;
658 }
659
660 if (!subarray) {
661 ret = st->ss->load_super(st, dfd, NULL);
662 if (ret) {
663 pr_err("Failed to load super-block.\n");
664 close(dfd);
665 goto free_info;
666 }
667
668 ret = st->ss->update_super(st, sra, "ppl",
669 devname,
670 c->verbose, 0, NULL);
671 if (ret) {
672 close(dfd);
673 st->ss->free_super(st);
674 goto free_info;
675 }
676 st->ss->getinfo_super(st, &info, NULL);
677 }
678
679 ret |= sysfs_set_num(sra, sd, "ppl_sector",
680 info.ppl_sector);
681 ret |= sysfs_set_num(sra, sd, "ppl_size",
682 info.ppl_size);
683
684 if (ret) {
685 pr_err("Failed to set PPL attributes for %s\n",
686 sd->sys_name);
687 close(dfd);
688 st->ss->free_super(st);
689 goto free_info;
690 }
691
692 ret = st->ss->write_init_ppl(st, &info, dfd);
693 if (ret)
694 pr_err("Failed to write PPL\n");
695
696 close(dfd);
697
698 if (!subarray)
699 st->ss->free_super(st);
700
701 if (ret)
702 goto free_info;
703 }
704 }
705
706 ret = sysfs_set_str(sra, NULL, "consistency_policy",
707 map_num_s(consistency_policies,
708 s->consistency_policy));
709 if (ret)
710 pr_err("Failed to change array consistency policy\n");
711
712 free_info:
713 sysfs_free(sra);
714 free_st:
715 free(st);
716 free(subarray);
717
718 return ret;
719 }
720
721 /*
722 * When reshaping an array we might need to backup some data.
723 * This is written to all spares with a 'super_block' describing it.
724 * The superblock goes 4K from the end of the used space on the
725 * device.
726 * It if written after the backup is complete.
727 * It has the following structure.
728 */
729
730 static struct mdp_backup_super {
731 char magic[16]; /* md_backup_data-1 or -2 */
732 __u8 set_uuid[16];
733 __u64 mtime;
734 /* start/sizes in 512byte sectors */
735 __u64 devstart; /* address on backup device/file of data */
736 __u64 arraystart;
737 __u64 length;
738 __u32 sb_csum; /* csum of preceeding bytes. */
739 __u32 pad1;
740 __u64 devstart2; /* offset in to data of second section */
741 __u64 arraystart2;
742 __u64 length2;
743 __u32 sb_csum2; /* csum of preceeding bytes. */
744 __u8 pad[512-68-32];
745 } __attribute__((aligned(512))) bsb, bsb2;
746
747 static __u32 bsb_csum(char *buf, int len)
748 {
749 int i;
750 int csum = 0;
751 for (i = 0; i < len; i++)
752 csum = (csum<<3) + buf[0];
753 return __cpu_to_le32(csum);
754 }
755
756 static int check_idle(struct supertype *st)
757 {
758 /* Check that all member arrays for this container, or the
759 * container of this array, are idle
760 */
761 char *container = (st->container_devnm[0]
762 ? st->container_devnm : st->devnm);
763 struct mdstat_ent *ent, *e;
764 int is_idle = 1;
765
766 ent = mdstat_read(0, 0);
767 for (e = ent ; e; e = e->next) {
768 if (!is_container_member(e, container))
769 continue;
770 /* frozen array is not idle*/
771 if (e->percent >= 0 || e->metadata_version[9] == '-') {
772 is_idle = 0;
773 break;
774 }
775 }
776 free_mdstat(ent);
777 return is_idle;
778 }
779
780 static int freeze_container(struct supertype *st)
781 {
782 char *container = (st->container_devnm[0]
783 ? st->container_devnm : st->devnm);
784
785 if (!check_idle(st))
786 return -1;
787
788 if (block_monitor(container, 1)) {
789 pr_err("failed to freeze container\n");
790 return -2;
791 }
792
793 return 1;
794 }
795
796 static void unfreeze_container(struct supertype *st)
797 {
798 char *container = (st->container_devnm[0]
799 ? st->container_devnm : st->devnm);
800
801 unblock_monitor(container, 1);
802 }
803
804 static int freeze(struct supertype *st)
805 {
806 /* Try to freeze resync/rebuild on this array/container.
807 * Return -1 if the array is busy,
808 * return -2 container cannot be frozen,
809 * return 0 if this kernel doesn't support 'frozen'
810 * return 1 if it worked.
811 */
812 if (st->ss->external)
813 return freeze_container(st);
814 else {
815 struct mdinfo *sra = sysfs_read(-1, st->devnm, GET_VERSION);
816 int err;
817 char buf[20];
818
819 if (!sra)
820 return -1;
821 /* Need to clear any 'read-auto' status */
822 if (sysfs_get_str(sra, NULL, "array_state", buf, 20) > 0 &&
823 strncmp(buf, "read-auto", 9) == 0)
824 sysfs_set_str(sra, NULL, "array_state", "clean");
825
826 err = sysfs_freeze_array(sra);
827 sysfs_free(sra);
828 return err;
829 }
830 }
831
832 static void unfreeze(struct supertype *st)
833 {
834 if (st->ss->external)
835 return unfreeze_container(st);
836 else {
837 struct mdinfo *sra = sysfs_read(-1, st->devnm, GET_VERSION);
838 char buf[20];
839
840 if (sra &&
841 sysfs_get_str(sra, NULL, "sync_action", buf, 20) > 0 &&
842 strcmp(buf, "frozen\n") == 0)
843 sysfs_set_str(sra, NULL, "sync_action", "idle");
844 sysfs_free(sra);
845 }
846 }
847
848 static void wait_reshape(struct mdinfo *sra)
849 {
850 int fd = sysfs_get_fd(sra, NULL, "sync_action");
851 char action[20];
852
853 if (fd < 0)
854 return;
855
856 while (sysfs_fd_get_str(fd, action, 20) > 0 &&
857 strncmp(action, "reshape", 7) == 0)
858 sysfs_wait(fd, NULL);
859 close(fd);
860 }
861
862 static int reshape_super(struct supertype *st, unsigned long long size,
863 int level, int layout, int chunksize, int raid_disks,
864 int delta_disks, char *backup_file, char *dev,
865 int direction, int verbose)
866 {
867 /* nothing extra to check in the native case */
868 if (!st->ss->external)
869 return 0;
870 if (!st->ss->reshape_super || !st->ss->manage_reshape) {
871 pr_err("%s metadata does not support reshape\n",
872 st->ss->name);
873 return 1;
874 }
875
876 return st->ss->reshape_super(st, size, level, layout, chunksize,
877 raid_disks, delta_disks, backup_file, dev,
878 direction, verbose);
879 }
880
881 static void sync_metadata(struct supertype *st)
882 {
883 if (st->ss->external) {
884 if (st->update_tail) {
885 flush_metadata_updates(st);
886 st->update_tail = &st->updates;
887 } else
888 st->ss->sync_metadata(st);
889 }
890 }
891
892 static int subarray_set_num(char *container, struct mdinfo *sra, char *name, int n)
893 {
894 /* when dealing with external metadata subarrays we need to be
895 * prepared to handle EAGAIN. The kernel may need to wait for
896 * mdmon to mark the array active so the kernel can handle
897 * allocations/writeback when preparing the reshape action
898 * (md_allow_write()). We temporarily disable safe_mode_delay
899 * to close a race with the array_state going clean before the
900 * next write to raid_disks / stripe_cache_size
901 */
902 char safe[50];
903 int rc;
904
905 /* only 'raid_disks' and 'stripe_cache_size' trigger md_allow_write */
906 if (!container ||
907 (strcmp(name, "raid_disks") != 0 &&
908 strcmp(name, "stripe_cache_size") != 0))
909 return sysfs_set_num(sra, NULL, name, n);
910
911 rc = sysfs_get_str(sra, NULL, "safe_mode_delay", safe, sizeof(safe));
912 if (rc <= 0)
913 return -1;
914 sysfs_set_num(sra, NULL, "safe_mode_delay", 0);
915 rc = sysfs_set_num(sra, NULL, name, n);
916 if (rc < 0 && errno == EAGAIN) {
917 ping_monitor(container);
918 /* if we get EAGAIN here then the monitor is not active
919 * so stop trying
920 */
921 rc = sysfs_set_num(sra, NULL, name, n);
922 }
923 sysfs_set_str(sra, NULL, "safe_mode_delay", safe);
924 return rc;
925 }
926
927 int start_reshape(struct mdinfo *sra, int already_running,
928 int before_data_disks, int data_disks, struct supertype *st)
929 {
930 int err;
931 unsigned long long sync_max_to_set;
932
933 sysfs_set_num(sra, NULL, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL);
934 err = sysfs_set_num(sra, NULL, "suspend_hi", sra->reshape_progress);
935 err = err ?: sysfs_set_num(sra, NULL, "suspend_lo",
936 sra->reshape_progress);
937 if (before_data_disks <= data_disks)
938 sync_max_to_set = sra->reshape_progress / data_disks;
939 else
940 sync_max_to_set = (sra->component_size * data_disks
941 - sra->reshape_progress) / data_disks;
942
943 if (!already_running)
944 sysfs_set_num(sra, NULL, "sync_min", sync_max_to_set);
945
946 if (st->ss->external)
947 err = err ?: sysfs_set_num(sra, NULL, "sync_max", sync_max_to_set);
948 else
949 err = err ?: sysfs_set_str(sra, NULL, "sync_max", "max");
950
951 if (!already_running && err == 0) {
952 int cnt = 5;
953 do {
954 err = sysfs_set_str(sra, NULL, "sync_action",
955 "reshape");
956 if (err)
957 sleep(1);
958 } while (err && errno == EBUSY && cnt-- > 0);
959 }
960 return err;
961 }
962
963 void abort_reshape(struct mdinfo *sra)
964 {
965 sysfs_set_str(sra, NULL, "sync_action", "idle");
966 /*
967 * Prior to kernel commit: 23ddff3792f6 ("md: allow suspend_lo and
968 * suspend_hi to decrease as well as increase.")
969 * you could only increase suspend_{lo,hi} unless the region they
970 * covered was empty. So to reset to 0, you need to push suspend_lo
971 * up past suspend_hi first. So to maximize the chance of mdadm
972 * working on all kernels, we want to keep doing that.
973 */
974 sysfs_set_num(sra, NULL, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL);
975 sysfs_set_num(sra, NULL, "suspend_hi", 0);
976 sysfs_set_num(sra, NULL, "suspend_lo", 0);
977 sysfs_set_num(sra, NULL, "sync_min", 0);
978 // It isn't safe to reset sync_max as we aren't monitoring.
979 // Array really should be stopped at this point.
980 }
981
982 int remove_disks_for_takeover(struct supertype *st,
983 struct mdinfo *sra,
984 int layout)
985 {
986 int nr_of_copies;
987 struct mdinfo *remaining;
988 int slot;
989
990 if (st->ss->external) {
991 int rv = 0;
992 struct mdinfo *arrays = st->ss->container_content(st, NULL);
993 /*
994 * containter_content returns list of arrays in container
995 * If arrays->next is not NULL it means that there are
996 * 2 arrays in container and operation should be blocked
997 */
998 if (arrays) {
999 if (arrays->next)
1000 rv = 1;
1001 sysfs_free(arrays);
1002 if (rv) {
1003 pr_err("Error. Cannot perform operation on %s- for this operation "
1004 "it MUST be single array in container\n", st->devnm);
1005 return rv;
1006 }
1007 }
1008 }
1009
1010 if (sra->array.level == 10)
1011 nr_of_copies = layout & 0xff;
1012 else if (sra->array.level == 1)
1013 nr_of_copies = sra->array.raid_disks;
1014 else
1015 return 1;
1016
1017 remaining = sra->devs;
1018 sra->devs = NULL;
1019 /* for each 'copy', select one device and remove from the list. */
1020 for (slot = 0; slot < sra->array.raid_disks; slot += nr_of_copies) {
1021 struct mdinfo **diskp;
1022 int found = 0;
1023
1024 /* Find a working device to keep */
1025 for (diskp = &remaining; *diskp ; diskp = &(*diskp)->next) {
1026 struct mdinfo *disk = *diskp;
1027
1028 if (disk->disk.raid_disk < slot)
1029 continue;
1030 if (disk->disk.raid_disk >= slot + nr_of_copies)
1031 continue;
1032 if (disk->disk.state & (1<<MD_DISK_REMOVED))
1033 continue;
1034 if (disk->disk.state & (1<<MD_DISK_FAULTY))
1035 continue;
1036 if (!(disk->disk.state & (1<<MD_DISK_SYNC)))
1037 continue;
1038
1039 /* We have found a good disk to use! */
1040 *diskp = disk->next;
1041 disk->next = sra->devs;
1042 sra->devs = disk;
1043 found = 1;
1044 break;
1045 }
1046 if (!found)
1047 break;
1048 }
1049
1050 if (slot < sra->array.raid_disks) {
1051 /* didn't find all slots */
1052 struct mdinfo **e;
1053 e = &remaining;
1054 while (*e)
1055 e = &(*e)->next;
1056 *e = sra->devs;
1057 sra->devs = remaining;
1058 return 1;
1059 }
1060
1061 /* Remove all 'remaining' devices from the array */
1062 while (remaining) {
1063 struct mdinfo *sd = remaining;
1064 remaining = sd->next;
1065
1066 sysfs_set_str(sra, sd, "state", "faulty");
1067 sysfs_set_str(sra, sd, "slot", "none");
1068 /* for external metadata disks should be removed in mdmon */
1069 if (!st->ss->external)
1070 sysfs_set_str(sra, sd, "state", "remove");
1071 sd->disk.state |= (1<<MD_DISK_REMOVED);
1072 sd->disk.state &= ~(1<<MD_DISK_SYNC);
1073 sd->next = sra->devs;
1074 sra->devs = sd;
1075 }
1076 return 0;
1077 }
1078
1079 void reshape_free_fdlist(int *fdlist,
1080 unsigned long long *offsets,
1081 int size)
1082 {
1083 int i;
1084
1085 for (i = 0; i < size; i++)
1086 if (fdlist[i] >= 0)
1087 close(fdlist[i]);
1088
1089 free(fdlist);
1090 free(offsets);
1091 }
1092
1093 int reshape_prepare_fdlist(char *devname,
1094 struct mdinfo *sra,
1095 int raid_disks,
1096 int nrdisks,
1097 unsigned long blocks,
1098 char *backup_file,
1099 int *fdlist,
1100 unsigned long long *offsets)
1101 {
1102 int d = 0;
1103 struct mdinfo *sd;
1104
1105 enable_fds(nrdisks);
1106 for (d = 0; d <= nrdisks; d++)
1107 fdlist[d] = -1;
1108 d = raid_disks;
1109 for (sd = sra->devs; sd; sd = sd->next) {
1110 if (sd->disk.state & (1<<MD_DISK_FAULTY))
1111 continue;
1112 if (sd->disk.state & (1<<MD_DISK_SYNC) &&
1113 sd->disk.raid_disk < raid_disks) {
1114 char *dn = map_dev(sd->disk.major, sd->disk.minor, 1);
1115 fdlist[sd->disk.raid_disk] = dev_open(dn, O_RDONLY);
1116 offsets[sd->disk.raid_disk] = sd->data_offset*512;
1117 if (fdlist[sd->disk.raid_disk] < 0) {
1118 pr_err("%s: cannot open component %s\n",
1119 devname, dn ? dn : "-unknown-");
1120 d = -1;
1121 goto release;
1122 }
1123 } else if (backup_file == NULL) {
1124 /* spare */
1125 char *dn = map_dev(sd->disk.major, sd->disk.minor, 1);
1126 fdlist[d] = dev_open(dn, O_RDWR);
1127 offsets[d] = (sd->data_offset + sra->component_size - blocks - 8)*512;
1128 if (fdlist[d] < 0) {
1129 pr_err("%s: cannot open component %s\n",
1130 devname, dn ? dn : "-unknown-");
1131 d = -1;
1132 goto release;
1133 }
1134 d++;
1135 }
1136 }
1137 release:
1138 return d;
1139 }
1140
1141 int reshape_open_backup_file(char *backup_file,
1142 int fd,
1143 char *devname,
1144 long blocks,
1145 int *fdlist,
1146 unsigned long long *offsets,
1147 char *sys_name,
1148 int restart)
1149 {
1150 /* Return 1 on success, 0 on any form of failure */
1151 /* need to check backup file is large enough */
1152 char buf[512];
1153 struct stat stb;
1154 unsigned int dev;
1155 int i;
1156
1157 *fdlist = open(backup_file, O_RDWR|O_CREAT|(restart ? O_TRUNC : O_EXCL),
1158 S_IRUSR | S_IWUSR);
1159 *offsets = 8 * 512;
1160 if (*fdlist < 0) {
1161 pr_err("%s: cannot create backup file %s: %s\n",
1162 devname, backup_file, strerror(errno));
1163 return 0;
1164 }
1165 /* Guard against backup file being on array device.
1166 * If array is partitioned or if LVM etc is in the
1167 * way this will not notice, but it is better than
1168 * nothing.
1169 */
1170 fstat(*fdlist, &stb);
1171 dev = stb.st_dev;
1172 fstat(fd, &stb);
1173 if (stb.st_rdev == dev) {
1174 pr_err("backup file must NOT be on the array being reshaped.\n");
1175 close(*fdlist);
1176 return 0;
1177 }
1178
1179 memset(buf, 0, 512);
1180 for (i=0; i < blocks + 8 ; i++) {
1181 if (write(*fdlist, buf, 512) != 512) {
1182 pr_err("%s: cannot create backup file %s: %s\n",
1183 devname, backup_file, strerror(errno));
1184 return 0;
1185 }
1186 }
1187 if (fsync(*fdlist) != 0) {
1188 pr_err("%s: cannot create backup file %s: %s\n",
1189 devname, backup_file, strerror(errno));
1190 return 0;
1191 }
1192
1193 if (!restart && strncmp(backup_file, MAP_DIR, strlen(MAP_DIR)) != 0) {
1194 char *bu = make_backup(sys_name);
1195 if (symlink(backup_file, bu))
1196 pr_err("Recording backup file in " MAP_DIR " failed: %s\n",
1197 strerror(errno));
1198 free(bu);
1199 }
1200
1201 return 1;
1202 }
1203
1204 unsigned long compute_backup_blocks(int nchunk, int ochunk,
1205 unsigned int ndata, unsigned int odata)
1206 {
1207 unsigned long a, b, blocks;
1208 /* So how much do we need to backup.
1209 * We need an amount of data which is both a whole number of
1210 * old stripes and a whole number of new stripes.
1211 * So LCM for (chunksize*datadisks).
1212 */
1213 a = (ochunk/512) * odata;
1214 b = (nchunk/512) * ndata;
1215 /* Find GCD */
1216 a = GCD(a, b);
1217 /* LCM == product / GCD */
1218 blocks = (unsigned long)(ochunk/512) * (unsigned long)(nchunk/512) *
1219 odata * ndata / a;
1220
1221 return blocks;
1222 }
1223
1224 char *analyse_change(char *devname, struct mdinfo *info, struct reshape *re)
1225 {
1226 /* Based on the current array state in info->array and
1227 * the changes in info->new_* etc, determine:
1228 * - whether the change is possible
1229 * - Intermediate level/raid_disks/layout
1230 * - whether a restriping reshape is needed
1231 * - number of sectors in minimum change unit. This
1232 * will cover a whole number of stripes in 'before' and
1233 * 'after'.
1234 *
1235 * Return message if the change should be rejected
1236 * NULL if the change can be achieved
1237 *
1238 * This can be called as part of starting a reshape, or
1239 * when assembling an array that is undergoing reshape.
1240 */
1241 int near, far, offset, copies;
1242 int new_disks;
1243 int old_chunk, new_chunk;
1244 /* delta_parity records change in number of devices
1245 * caused by level change
1246 */
1247 int delta_parity = 0;
1248
1249 memset(re, 0, sizeof(*re));
1250
1251 /* If a new level not explicitly given, we assume no-change */
1252 if (info->new_level == UnSet)
1253 info->new_level = info->array.level;
1254
1255 if (info->new_chunk)
1256 switch (info->new_level) {
1257 case 0:
1258 case 4:
1259 case 5:
1260 case 6:
1261 case 10:
1262 /* chunk size is meaningful, must divide component_size
1263 * evenly
1264 */
1265 if (info->component_size % (info->new_chunk/512)) {
1266 unsigned long long shrink = info->component_size;
1267 shrink &= ~(unsigned long long)(info->new_chunk/512-1);
1268 pr_err("New chunk size (%dK) does not evenly divide device size (%lluk)\n",
1269 info->new_chunk/1024, info->component_size/2);
1270 pr_err("After shrinking any filesystem, \"mdadm --grow %s --size %llu\"\n",
1271 devname, shrink/2);
1272 pr_err("will shrink the array so the given chunk size would work.\n");
1273 return "";
1274 }
1275 break;
1276 default:
1277 return "chunk size not meaningful for this level";
1278 }
1279 else
1280 info->new_chunk = info->array.chunk_size;
1281
1282 switch (info->array.level) {
1283 default:
1284 return "No reshape is possibly for this RAID level";
1285 case LEVEL_LINEAR:
1286 if (info->delta_disks != UnSet)
1287 return "Only --add is supported for LINEAR, setting --raid-disks is not needed";
1288 else
1289 return "Only --add is supported for LINEAR, other --grow options are not meaningful";
1290 case 1:
1291 /* RAID1 can convert to RAID1 with different disks, or
1292 * raid5 with 2 disks, or
1293 * raid0 with 1 disk
1294 */
1295 if (info->new_level > 1 && (info->component_size & 7))
1296 return "Cannot convert RAID1 of this size - reduce size to multiple of 4K first.";
1297 if (info->new_level == 0) {
1298 if (info->delta_disks != UnSet &&
1299 info->delta_disks != 0)
1300 return "Cannot change number of disks with RAID1->RAID0 conversion";
1301 re->level = 0;
1302 re->before.data_disks = 1;
1303 re->after.data_disks = 1;
1304 return NULL;
1305 }
1306 if (info->new_level == 1) {
1307 if (info->delta_disks == UnSet)
1308 /* Don't know what to do */
1309 return "no change requested for Growing RAID1";
1310 re->level = 1;
1311 return NULL;
1312 }
1313 if (info->array.raid_disks != 2 && info->new_level == 5)
1314 return "Can only convert a 2-device array to RAID5";
1315 if (info->array.raid_disks == 2 && info->new_level == 5) {
1316 re->level = 5;
1317 re->before.data_disks = 1;
1318 if (info->delta_disks != UnSet &&
1319 info->delta_disks != 0)
1320 re->after.data_disks = 1 + info->delta_disks;
1321 else
1322 re->after.data_disks = 1;
1323 if (re->after.data_disks < 1)
1324 return "Number of disks too small for RAID5";
1325
1326 re->before.layout = ALGORITHM_LEFT_SYMMETRIC;
1327 info->array.chunk_size = 65536;
1328 break;
1329 }
1330 /* Could do some multi-stage conversions, but leave that to
1331 * later.
1332 */
1333 return "Impossibly level change request for RAID1";
1334
1335 case 10:
1336 /* RAID10 can be converted from near mode to
1337 * RAID0 by removing some devices.
1338 * It can also be reshaped if the kernel supports
1339 * new_data_offset.
1340 */
1341 switch (info->new_level) {
1342 case 0:
1343 if ((info->array.layout & ~0xff) != 0x100)
1344 return "Cannot Grow RAID10 with far/offset layout";
1345 /*
1346 * number of devices must be multiple of
1347 * number of copies
1348 */
1349 if (info->array.raid_disks %
1350 (info->array.layout & 0xff))
1351 return "RAID10 layout too complex for Grow operation";
1352
1353 new_disks = (info->array.raid_disks /
1354 (info->array.layout & 0xff));
1355 if (info->delta_disks == UnSet)
1356 info->delta_disks = (new_disks
1357 - info->array.raid_disks);
1358
1359 if (info->delta_disks !=
1360 new_disks - info->array.raid_disks)
1361 return "New number of raid-devices impossible for RAID10";
1362 if (info->new_chunk &&
1363 info->new_chunk != info->array.chunk_size)
1364 return "Cannot change chunk-size with RAID10 Grow";
1365
1366 /* looks good */
1367 re->level = 0;
1368 re->before.data_disks = new_disks;
1369 re->after.data_disks = re->before.data_disks;
1370 return NULL;
1371
1372 case 10:
1373 near = info->array.layout & 0xff;
1374 far = (info->array.layout >> 8) & 0xff;
1375 offset = info->array.layout & 0x10000;
1376 if (far > 1 && !offset)
1377 return "Cannot reshape RAID10 in far-mode";
1378 copies = near * far;
1379
1380 old_chunk = info->array.chunk_size * far;
1381
1382 if (info->new_layout == UnSet)
1383 info->new_layout = info->array.layout;
1384 else {
1385 near = info->new_layout & 0xff;
1386 far = (info->new_layout >> 8) & 0xff;
1387 offset = info->new_layout & 0x10000;
1388 if (far > 1 && !offset)
1389 return "Cannot reshape RAID10 to far-mode";
1390 if (near * far != copies)
1391 return "Cannot change number of copies when reshaping RAID10";
1392 }
1393 if (info->delta_disks == UnSet)
1394 info->delta_disks = 0;
1395 new_disks = (info->array.raid_disks +
1396 info->delta_disks);
1397
1398 new_chunk = info->new_chunk * far;
1399
1400 re->level = 10;
1401 re->before.layout = info->array.layout;
1402 re->before.data_disks = info->array.raid_disks;
1403 re->after.layout = info->new_layout;
1404 re->after.data_disks = new_disks;
1405 /* For RAID10 we don't do backup but do allow reshape,
1406 * so set backup_blocks to INVALID_SECTORS rather than
1407 * zero.
1408 * And there is no need to synchronise stripes on both
1409 * 'old' and 'new'. So the important
1410 * number is the minimum data_offset difference
1411 * which is the larger of (offset copies * chunk).
1412 */
1413 re->backup_blocks = INVALID_SECTORS;
1414 re->min_offset_change = max(old_chunk, new_chunk) / 512;
1415 if (new_disks < re->before.data_disks &&
1416 info->space_after < re->min_offset_change)
1417 /* Reduce component size by one chunk */
1418 re->new_size = (info->component_size -
1419 re->min_offset_change);
1420 else
1421 re->new_size = info->component_size;
1422 re->new_size = re->new_size * new_disks / copies;
1423 return NULL;
1424
1425 default:
1426 return "RAID10 can only be changed to RAID0";
1427 }
1428 case 0:
1429 /* RAID0 can be converted to RAID10, or to RAID456 */
1430 if (info->new_level == 10) {
1431 if (info->new_layout == UnSet &&
1432 info->delta_disks == UnSet) {
1433 /* Assume near=2 layout */
1434 info->new_layout = 0x102;
1435 info->delta_disks = info->array.raid_disks;
1436 }
1437 if (info->new_layout == UnSet) {
1438 int copies = 1 + (info->delta_disks
1439 / info->array.raid_disks);
1440 if (info->array.raid_disks * (copies-1) !=
1441 info->delta_disks)
1442 return "Impossible number of devices for RAID0->RAID10";
1443 info->new_layout = 0x100 + copies;
1444 }
1445 if (info->delta_disks == UnSet) {
1446 int copies = info->new_layout & 0xff;
1447 if (info->new_layout != 0x100 + copies)
1448 return "New layout impossible for RAID0->RAID10";;
1449 info->delta_disks = (copies - 1) *
1450 info->array.raid_disks;
1451 }
1452 if (info->new_chunk &&
1453 info->new_chunk != info->array.chunk_size)
1454 return "Cannot change chunk-size with RAID0->RAID10";
1455 /* looks good */
1456 re->level = 10;
1457 re->before.data_disks = (info->array.raid_disks +
1458 info->delta_disks);
1459 re->after.data_disks = re->before.data_disks;
1460 re->before.layout = info->new_layout;
1461 return NULL;
1462 }
1463
1464 /* RAID0 can also covert to RAID0/4/5/6 by first converting to
1465 * a raid4 style layout of the final level.
1466 */
1467 switch (info->new_level) {
1468 case 4:
1469 delta_parity = 1;
1470 case 0:
1471 re->level = 4;
1472 re->before.layout = 0;
1473 break;
1474 case 5:
1475 delta_parity = 1;
1476 re->level = 5;
1477 re->before.layout = ALGORITHM_PARITY_N;
1478 if (info->new_layout == UnSet)
1479 info->new_layout = map_name(r5layout, "default");
1480 break;
1481 case 6:
1482 delta_parity = 2;
1483 re->level = 6;
1484 re->before.layout = ALGORITHM_PARITY_N;
1485 if (info->new_layout == UnSet)
1486 info->new_layout = map_name(r6layout, "default");
1487 break;
1488 default:
1489 return "Impossible level change requested";
1490 }
1491 re->before.data_disks = info->array.raid_disks;
1492 /* determining 'after' layout happens outside this 'switch' */
1493 break;
1494
1495 case 4:
1496 info->array.layout = ALGORITHM_PARITY_N;
1497 case 5:
1498 switch (info->new_level) {
1499 case 0:
1500 delta_parity = -1;
1501 case 4:
1502 re->level = info->array.level;
1503 re->before.data_disks = info->array.raid_disks - 1;
1504 re->before.layout = info->array.layout;
1505 break;
1506 case 5:
1507 re->level = 5;
1508 re->before.data_disks = info->array.raid_disks - 1;
1509 re->before.layout = info->array.layout;
1510 break;
1511 case 6:
1512 delta_parity = 1;
1513 re->level = 6;
1514 re->before.data_disks = info->array.raid_disks - 1;
1515 switch (info->array.layout) {
1516 case ALGORITHM_LEFT_ASYMMETRIC:
1517 re->before.layout = ALGORITHM_LEFT_ASYMMETRIC_6;
1518 break;
1519 case ALGORITHM_RIGHT_ASYMMETRIC:
1520 re->before.layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
1521 break;
1522 case ALGORITHM_LEFT_SYMMETRIC:
1523 re->before.layout = ALGORITHM_LEFT_SYMMETRIC_6;
1524 break;
1525 case ALGORITHM_RIGHT_SYMMETRIC:
1526 re->before.layout = ALGORITHM_RIGHT_SYMMETRIC_6;
1527 break;
1528 case ALGORITHM_PARITY_0:
1529 re->before.layout = ALGORITHM_PARITY_0_6;
1530 break;
1531 case ALGORITHM_PARITY_N:
1532 re->before.layout = ALGORITHM_PARITY_N_6;
1533 break;
1534 default:
1535 return "Cannot convert an array with this layout";
1536 }
1537 break;
1538 case 1:
1539 if (info->array.raid_disks != 2)
1540 return "Can only convert a 2-device array to RAID1";
1541 if (info->delta_disks != UnSet &&
1542 info->delta_disks != 0)
1543 return "Cannot set raid_disk when converting RAID5->RAID1";
1544 re->level = 1;
1545 info->new_chunk = 0;
1546 return NULL;
1547 default:
1548 return "Impossible level change requested";
1549 }
1550 break;
1551 case 6:
1552 switch (info->new_level) {
1553 case 4:
1554 case 5:
1555 delta_parity = -1;
1556 case 6:
1557 re->level = 6;
1558 re->before.data_disks = info->array.raid_disks - 2;
1559 re->before.layout = info->array.layout;
1560 break;
1561 default:
1562 return "Impossible level change requested";
1563 }
1564 break;
1565 }
1566
1567 /* If we reached here then it looks like a re-stripe is
1568 * happening. We have determined the intermediate level
1569 * and initial raid_disks/layout and stored these in 're'.
1570 *
1571 * We need to deduce the final layout that can be atomically
1572 * converted to the end state.
1573 */
1574 switch (info->new_level) {
1575 case 0:
1576 /* We can only get to RAID0 from RAID4 or RAID5
1577 * with appropriate layout and one extra device
1578 */
1579 if (re->level != 4 && re->level != 5)
1580 return "Cannot covert to RAID0 from this level";
1581
1582 switch (re->level) {
1583 case 4:
1584 re->before.layout = 0;
1585 re->after.layout = 0;
1586 break;
1587 case 5:
1588 re->after.layout = ALGORITHM_PARITY_N;
1589 break;
1590 }
1591 break;
1592
1593 case 4:
1594 /* We can only get to RAID4 from RAID5 */
1595 if (re->level != 4 && re->level != 5)
1596 return "Cannot convert to RAID4 from this level";
1597
1598 switch (re->level) {
1599 case 4:
1600 re->after.layout = 0;
1601 break;
1602 case 5:
1603 re->after.layout = ALGORITHM_PARITY_N;
1604 break;
1605 }
1606 break;
1607
1608 case 5:
1609 /* We get to RAID5 from RAID5 or RAID6 */
1610 if (re->level != 5 && re->level != 6)
1611 return "Cannot convert to RAID5 from this level";
1612
1613 switch (re->level) {
1614 case 5:
1615 if (info->new_layout == UnSet)
1616 re->after.layout = re->before.layout;
1617 else
1618 re->after.layout = info->new_layout;
1619 break;
1620 case 6:
1621 if (info->new_layout == UnSet)
1622 info->new_layout = re->before.layout;
1623
1624 /* after.layout needs to be raid6 version of new_layout */
1625 if (info->new_layout == ALGORITHM_PARITY_N)
1626 re->after.layout = ALGORITHM_PARITY_N;
1627 else {
1628 char layout[40];
1629 char *ls = map_num(r5layout, info->new_layout);
1630 int l;
1631 if (ls) {
1632 /* Current RAID6 layout has a RAID5
1633 * equivalent - good
1634 */
1635 strcat(strcpy(layout, ls), "-6");
1636 l = map_name(r6layout, layout);
1637 if (l == UnSet)
1638 return "Cannot find RAID6 layout to convert to";
1639 } else {
1640 /* Current RAID6 has no equivalent.
1641 * If it is already a '-6' layout we
1642 * can leave it unchanged, else we must
1643 * fail
1644 */
1645 ls = map_num(r6layout,
1646 info->new_layout);
1647 if (!ls ||
1648 strcmp(ls+strlen(ls)-2, "-6") != 0)
1649 return "Please specify new layout";
1650 l = info->new_layout;
1651 }
1652 re->after.layout = l;
1653 }
1654 }
1655 break;
1656
1657 case 6:
1658 /* We must already be at level 6 */
1659 if (re->level != 6)
1660 return "Impossible level change";
1661 if (info->new_layout == UnSet)
1662 re->after.layout = info->array.layout;
1663 else
1664 re->after.layout = info->new_layout;
1665 break;
1666 default:
1667 return "Impossible level change requested";
1668 }
1669 if (info->delta_disks == UnSet)
1670 info->delta_disks = delta_parity;
1671
1672 re->after.data_disks =
1673 (re->before.data_disks + info->delta_disks - delta_parity);
1674
1675 switch (re->level) {
1676 case 6:
1677 re->parity = 2;
1678 break;
1679 case 4:
1680 case 5:
1681 re->parity = 1;
1682 break;
1683 default:
1684 re->parity = 0;
1685 break;
1686 }
1687 /* So we have a restripe operation, we need to calculate the number
1688 * of blocks per reshape operation.
1689 */
1690 re->new_size = info->component_size * re->before.data_disks;
1691 if (info->new_chunk == 0)
1692 info->new_chunk = info->array.chunk_size;
1693 if (re->after.data_disks == re->before.data_disks &&
1694 re->after.layout == re->before.layout &&
1695 info->new_chunk == info->array.chunk_size) {
1696 /* Nothing to change, can change level immediately. */
1697 re->level = info->new_level;
1698 re->backup_blocks = 0;
1699 return NULL;
1700 }
1701 if (re->after.data_disks == 1 && re->before.data_disks == 1) {
1702 /* chunk and layout changes make no difference */
1703 re->level = info->new_level;
1704 re->backup_blocks = 0;
1705 return NULL;
1706 }
1707
1708 if (re->after.data_disks == re->before.data_disks &&
1709 get_linux_version() < 2006032)
1710 return "in-place reshape is not safe before 2.6.32 - sorry.";
1711
1712 if (re->after.data_disks < re->before.data_disks &&
1713 get_linux_version() < 2006030)
1714 return "reshape to fewer devices is not supported before 2.6.30 - sorry.";
1715
1716 re->backup_blocks = compute_backup_blocks(
1717 info->new_chunk, info->array.chunk_size,
1718 re->after.data_disks, re->before.data_disks);
1719 re->min_offset_change = re->backup_blocks / re->before.data_disks;
1720
1721 re->new_size = info->component_size * re->after.data_disks;
1722 return NULL;
1723 }
1724
1725 static int set_array_size(struct supertype *st, struct mdinfo *sra,
1726 char *text_version)
1727 {
1728 struct mdinfo *info;
1729 char *subarray;
1730 int ret_val = -1;
1731
1732 if ((st == NULL) || (sra == NULL))
1733 return ret_val;
1734
1735 if (text_version == NULL)
1736 text_version = sra->text_version;
1737 subarray = strchr(text_version + 1, '/')+1;
1738 info = st->ss->container_content(st, subarray);
1739 if (info) {
1740 unsigned long long current_size = 0;
1741 unsigned long long new_size = info->custom_array_size/2;
1742
1743 if (sysfs_get_ll(sra, NULL, "array_size", &current_size) == 0 &&
1744 new_size > current_size) {
1745 if (sysfs_set_num(sra, NULL, "array_size", new_size)
1746 < 0)
1747 dprintf("Error: Cannot set array size");
1748 else {
1749 ret_val = 0;
1750 dprintf("Array size changed");
1751 }
1752 dprintf_cont(" from %llu to %llu.\n",
1753 current_size, new_size);
1754 }
1755 sysfs_free(info);
1756 } else
1757 dprintf("Error: set_array_size(): info pointer in NULL\n");
1758
1759 return ret_val;
1760 }
1761
1762 static int reshape_array(char *container, int fd, char *devname,
1763 struct supertype *st, struct mdinfo *info,
1764 int force, struct mddev_dev *devlist,
1765 unsigned long long data_offset,
1766 char *backup_file, int verbose, int forked,
1767 int restart, int freeze_reshape);
1768 static int reshape_container(char *container, char *devname,
1769 int mdfd,
1770 struct supertype *st,
1771 struct mdinfo *info,
1772 int force,
1773 char *backup_file, int verbose,
1774 int forked, int restart, int freeze_reshape);
1775
1776 int Grow_reshape(char *devname, int fd,
1777 struct mddev_dev *devlist,
1778 unsigned long long data_offset,
1779 struct context *c, struct shape *s)
1780 {
1781 /* Make some changes in the shape of an array.
1782 * The kernel must support the change.
1783 *
1784 * There are three different changes. Each can trigger
1785 * a resync or recovery so we freeze that until we have
1786 * requested everything (if kernel supports freezing - 2.6.30).
1787 * The steps are:
1788 * - change size (i.e. component_size)
1789 * - change level
1790 * - change layout/chunksize/ndisks
1791 *
1792 * The last can require a reshape. It is different on different
1793 * levels so we need to check the level before actioning it.
1794 * Some times the level change needs to be requested after the
1795 * reshape (e.g. raid6->raid5, raid5->raid0)
1796 *
1797 */
1798 struct mdu_array_info_s array;
1799 int rv = 0;
1800 struct supertype *st;
1801 char *subarray = NULL;
1802
1803 int frozen;
1804 int changed = 0;
1805 char *container = NULL;
1806 int cfd = -1;
1807
1808 struct mddev_dev *dv;
1809 int added_disks;
1810
1811 struct mdinfo info;
1812 struct mdinfo *sra;
1813
1814 if (md_get_array_info(fd, &array) < 0) {
1815 pr_err("%s is not an active md array - aborting\n",
1816 devname);
1817 return 1;
1818 }
1819 if (s->level != UnSet && s->chunk) {
1820 pr_err("Cannot change array level in the same operation as changing chunk size.\n");
1821 return 1;
1822 }
1823
1824 if (data_offset != INVALID_SECTORS && array.level != 10 &&
1825 (array.level < 4 || array.level > 6)) {
1826 pr_err("--grow --data-offset not yet supported\n");
1827 return 1;
1828 }
1829
1830 if (s->size > 0 &&
1831 (s->chunk || s->level!= UnSet || s->layout_str || s->raiddisks)) {
1832 pr_err("cannot change component size at the same time as other changes.\n"
1833 " Change size first, then check data is intact before making other changes.\n");
1834 return 1;
1835 }
1836
1837 if (s->raiddisks && s->raiddisks < array.raid_disks &&
1838 array.level > 1 && get_linux_version() < 2006032 &&
1839 !check_env("MDADM_FORCE_FEWER")) {
1840 pr_err("reducing the number of devices is not safe before Linux 2.6.32\n"
1841 " Please use a newer kernel\n");
1842 return 1;
1843 }
1844
1845 if (array.level > 1 && s->size > 1 &&
1846 (unsigned long long) (array.chunk_size / 1024) > s->size) {
1847 pr_err("component size must be larger than chunk size.\n");
1848 return 1;
1849 }
1850
1851 st = super_by_fd(fd, &subarray);
1852 if (!st) {
1853 pr_err("Unable to determine metadata format for %s\n", devname);
1854 return 1;
1855 }
1856 if (s->raiddisks > st->max_devs) {
1857 pr_err("Cannot increase raid-disks on this array beyond %d\n", st->max_devs);
1858 return 1;
1859 }
1860 if (s->level == 0 && (array.state & (1 << MD_SB_BITMAP_PRESENT)) &&
1861 !(array.state & (1 << MD_SB_CLUSTERED)) && !st->ss->external) {
1862 array.state &= ~(1 << MD_SB_BITMAP_PRESENT);
1863 if (md_set_array_info(fd, &array) != 0) {
1864 pr_err("failed to remove internal bitmap.\n");
1865 return 1;
1866 }
1867 }
1868
1869 /* in the external case we need to check that the requested reshape is
1870 * supported, and perform an initial check that the container holds the
1871 * pre-requisite spare devices (mdmon owns final validation)
1872 */
1873 if (st->ss->external) {
1874 int retval;
1875
1876 if (subarray) {
1877 container = st->container_devnm;
1878 cfd = open_dev_excl(st->container_devnm);
1879 } else {
1880 container = st->devnm;
1881 close(fd);
1882 cfd = open_dev_excl(st->devnm);
1883 fd = cfd;
1884 }
1885 if (cfd < 0) {
1886 pr_err("Unable to open container for %s\n", devname);
1887 free(subarray);
1888 return 1;
1889 }
1890
1891 retval = st->ss->load_container(st, cfd, NULL);
1892
1893 if (retval) {
1894 pr_err("Cannot read superblock for %s\n", devname);
1895 free(subarray);
1896 return 1;
1897 }
1898
1899 /* check if operation is supported for metadata handler */
1900 if (st->ss->container_content) {
1901 struct mdinfo *cc = NULL;
1902 struct mdinfo *content = NULL;
1903
1904 cc = st->ss->container_content(st, subarray);
1905 for (content = cc; content ; content = content->next) {
1906 int allow_reshape = 1;
1907
1908 /* check if reshape is allowed based on metadata
1909 * indications stored in content.array.status
1910 */
1911 if (content->array.state &
1912 (1 << MD_SB_BLOCK_VOLUME))
1913 allow_reshape = 0;
1914 if (content->array.state &
1915 (1 << MD_SB_BLOCK_CONTAINER_RESHAPE))
1916 allow_reshape = 0;
1917 if (!allow_reshape) {
1918 pr_err("cannot reshape arrays in container with unsupported metadata: %s(%s)\n",
1919 devname, container);
1920 sysfs_free(cc);
1921 free(subarray);
1922 return 1;
1923 }
1924 if (content->consistency_policy ==
1925 CONSISTENCY_POLICY_PPL) {
1926 pr_err("Operation not supported when ppl consistency policy is enabled\n");
1927 sysfs_free(cc);
1928 free(subarray);
1929 return 1;
1930 }
1931 if (content->consistency_policy ==
1932 CONSISTENCY_POLICY_BITMAP) {
1933 pr_err("Operation not supported when write-intent bitmap is enabled\n");
1934 sysfs_free(cc);
1935 free(subarray);
1936 return 1;
1937 }
1938 }
1939 sysfs_free(cc);
1940 }
1941 if (mdmon_running(container))
1942 st->update_tail = &st->updates;
1943 }
1944
1945 added_disks = 0;
1946 for (dv = devlist; dv; dv = dv->next)
1947 added_disks++;
1948 if (s->raiddisks > array.raid_disks &&
1949 array.spare_disks + added_disks <
1950 (s->raiddisks - array.raid_disks) &&
1951 !c->force) {
1952 pr_err("Need %d spare%s to avoid degraded array, and only have %d.\n"
1953 " Use --force to over-ride this check.\n",
1954 s->raiddisks - array.raid_disks,
1955 s->raiddisks - array.raid_disks == 1 ? "" : "s",
1956 array.spare_disks + added_disks);
1957 return 1;
1958 }
1959
1960 sra = sysfs_read(fd, NULL, GET_LEVEL | GET_DISKS | GET_DEVS |
1961 GET_STATE | GET_VERSION);
1962 if (sra) {
1963 if (st->ss->external && subarray == NULL) {
1964 array.level = LEVEL_CONTAINER;
1965 sra->array.level = LEVEL_CONTAINER;
1966 }
1967 } else {
1968 pr_err("failed to read sysfs parameters for %s\n",
1969 devname);
1970 return 1;
1971 }
1972 frozen = freeze(st);
1973 if (frozen < -1) {
1974 /* freeze() already spewed the reason */
1975 sysfs_free(sra);
1976 return 1;
1977 } else if (frozen < 0) {
1978 pr_err("%s is performing resync/recovery and cannot be reshaped\n", devname);
1979 sysfs_free(sra);
1980 return 1;
1981 }
1982
1983 /* ========= set size =============== */
1984 if (s->size > 0 &&
1985 (s->size == MAX_SIZE || s->size != (unsigned)array.size)) {
1986 unsigned long long orig_size = get_component_size(fd)/2;
1987 unsigned long long min_csize;
1988 struct mdinfo *mdi;
1989 int raid0_takeover = 0;
1990
1991 if (orig_size == 0)
1992 orig_size = (unsigned) array.size;
1993
1994 if (orig_size == 0) {
1995 pr_err("Cannot set device size in this type of array.\n");
1996 rv = 1;
1997 goto release;
1998 }
1999
2000 if (array.level == 0) {
2001 pr_err("Component size change is not supported for RAID0\n");
2002 rv = 1;
2003 goto release;
2004 }
2005
2006 if (reshape_super(st, s->size, UnSet, UnSet, 0, 0, UnSet, NULL,
2007 devname, APPLY_METADATA_CHANGES,
2008 c->verbose > 0)) {
2009 rv = 1;
2010 goto release;
2011 }
2012 sync_metadata(st);
2013 if (st->ss->external) {
2014 /* metadata can have size limitation
2015 * update size value according to metadata information
2016 */
2017 struct mdinfo *sizeinfo =
2018 st->ss->container_content(st, subarray);
2019 if (sizeinfo) {
2020 unsigned long long new_size =
2021 sizeinfo->custom_array_size/2;
2022 int data_disks = get_data_disks(
2023 sizeinfo->array.level,
2024 sizeinfo->array.layout,
2025 sizeinfo->array.raid_disks);
2026 new_size /= data_disks;
2027 dprintf("Metadata size correction from %llu to %llu (%llu)\n",
2028 orig_size, new_size,
2029 new_size * data_disks);
2030 s->size = new_size;
2031 sysfs_free(sizeinfo);
2032 }
2033 }
2034
2035 /* Update the size of each member device in case
2036 * they have been resized. This will never reduce
2037 * below the current used-size. The "size" attribute
2038 * understands '0' to mean 'max'.
2039 */
2040 min_csize = 0;
2041 for (mdi = sra->devs; mdi; mdi = mdi->next) {
2042 sysfs_set_num(sra, mdi, "size",
2043 s->size == MAX_SIZE ? 0 : s->size);
2044 if (array.not_persistent == 0 &&
2045 array.major_version == 0 &&
2046 get_linux_version() < 3001000) {
2047 /* Dangerous to allow size to exceed 2TB */
2048 unsigned long long csize;
2049 if (sysfs_get_ll(sra, mdi, "size",
2050 &csize) == 0) {
2051 if (csize >= 2ULL*1024*1024*1024)
2052 csize = 2ULL*1024*1024*1024;
2053 if ((min_csize == 0 ||
2054 (min_csize > csize)))
2055 min_csize = csize;
2056 }
2057 }
2058 }
2059 if (min_csize && s->size > min_csize) {
2060 pr_err("Cannot safely make this array use more than 2TB per device on this kernel.\n");
2061 rv = 1;
2062 goto size_change_error;
2063 }
2064 if (min_csize && s->size == MAX_SIZE) {
2065 /* Don't let the kernel choose a size - it will get
2066 * it wrong
2067 */
2068 pr_err("Limited v0.90 array to 2TB per device\n");
2069 s->size = min_csize;
2070 }
2071 if (st->ss->external) {
2072 if (sra->array.level == 0) {
2073 rv = sysfs_set_str(sra, NULL, "level", "raid5");
2074 if (!rv) {
2075 raid0_takeover = 1;
2076 /* get array parameters after takeover
2077 * to change one parameter at time only
2078 */
2079 rv = md_get_array_info(fd, &array);
2080 }
2081 }
2082 /* make sure mdmon is
2083 * aware of the new level */
2084 if (!mdmon_running(st->container_devnm))
2085 start_mdmon(st->container_devnm);
2086 ping_monitor(container);
2087 if (mdmon_running(st->container_devnm) &&
2088 st->update_tail == NULL)
2089 st->update_tail = &st->updates;
2090 }
2091
2092 if (s->size == MAX_SIZE)
2093 s->size = 0;
2094 array.size = s->size;
2095 if (s->size & ~INT32_MAX) {
2096 /* got truncated to 32bit, write to
2097 * component_size instead
2098 */
2099 if (sra)
2100 rv = sysfs_set_num(sra, NULL,
2101 "component_size", s->size);
2102 else
2103 rv = -1;
2104 } else {
2105 rv = md_set_array_info(fd, &array);
2106
2107 /* manage array size when it is managed externally
2108 */
2109 if ((rv == 0) && st->ss->external)
2110 rv = set_array_size(st, sra, sra->text_version);
2111 }
2112
2113 if (raid0_takeover) {
2114 /* do not recync non-existing parity,
2115 * we will drop it anyway
2116 */
2117 sysfs_set_str(sra, NULL, "sync_action", "frozen");
2118 /* go back to raid0, drop parity disk
2119 */
2120 sysfs_set_str(sra, NULL, "level", "raid0");
2121 md_get_array_info(fd, &array);
2122 }
2123
2124 size_change_error:
2125 if (rv != 0) {
2126 int err = errno;
2127
2128 /* restore metadata */
2129 if (reshape_super(st, orig_size, UnSet, UnSet, 0, 0,
2130 UnSet, NULL, devname,
2131 ROLLBACK_METADATA_CHANGES,
2132 c->verbose) == 0)
2133 sync_metadata(st);
2134 pr_err("Cannot set device size for %s: %s\n",
2135 devname, strerror(err));
2136 if (err == EBUSY &&
2137 (array.state & (1<<MD_SB_BITMAP_PRESENT)))
2138 cont_err("Bitmap must be removed before size can be changed\n");
2139 rv = 1;
2140 goto release;
2141 }
2142 if (s->assume_clean) {
2143 /* This will fail on kernels older than 3.0 unless
2144 * a backport has been arranged.
2145 */
2146 if (sra == NULL ||
2147 sysfs_set_str(sra, NULL, "resync_start",
2148 "none") < 0)
2149 pr_err("--assume-clean not supported with --grow on this kernel\n");
2150 }
2151 md_get_array_info(fd, &array);
2152 s->size = get_component_size(fd)/2;
2153 if (s->size == 0)
2154 s->size = array.size;
2155 if (c->verbose >= 0) {
2156 if (s->size == orig_size)
2157 pr_err("component size of %s unchanged at %lluK\n",
2158 devname, s->size);
2159 else
2160 pr_err("component size of %s has been set to %lluK\n",
2161 devname, s->size);
2162 }
2163 changed = 1;
2164 } else if (array.level != LEVEL_CONTAINER) {
2165 s->size = get_component_size(fd)/2;
2166 if (s->size == 0)
2167 s->size = array.size;
2168 }
2169
2170 /* See if there is anything else to do */
2171 if ((s->level == UnSet || s->level == array.level) &&
2172 (s->layout_str == NULL) &&
2173 (s->chunk == 0 || s->chunk == array.chunk_size) &&
2174 data_offset == INVALID_SECTORS &&
2175 (s->raiddisks == 0 || s->raiddisks == array.raid_disks)) {
2176 /* Nothing more to do */
2177 if (!changed && c->verbose >= 0)
2178 pr_err("%s: no change requested\n", devname);
2179 goto release;
2180 }
2181
2182 /* ========= check for Raid10/Raid1 -> Raid0 conversion ===============
2183 * current implementation assumes that following conditions must be met:
2184 * - RAID10:
2185 * - far_copies == 1
2186 * - near_copies == 2
2187 */
2188 if ((s->level == 0 && array.level == 10 && sra &&
2189 array.layout == ((1 << 8) + 2) && !(array.raid_disks & 1)) ||
2190 (s->level == 0 && array.level == 1 && sra)) {
2191 int err;
2192
2193 err = remove_disks_for_takeover(st, sra, array.layout);
2194 if (err) {
2195 dprintf("Array cannot be reshaped\n");
2196 if (cfd > -1)
2197 close(cfd);
2198 rv = 1;
2199 goto release;
2200 }
2201 /* Make sure mdmon has seen the device removal
2202 * and updated metadata before we continue with
2203 * level change
2204 */
2205 if (container)
2206 ping_monitor(container);
2207 }
2208
2209 memset(&info, 0, sizeof(info));
2210 info.array = array;
2211 if (sysfs_init(&info, fd, NULL)) {
2212 pr_err("failed to initialize sysfs.\n");
2213 rv = 1;
2214 goto release;
2215 }
2216 strcpy(info.text_version, sra->text_version);
2217 info.component_size = s->size*2;
2218 info.new_level = s->level;
2219 info.new_chunk = s->chunk * 1024;
2220 if (info.array.level == LEVEL_CONTAINER) {
2221 info.delta_disks = UnSet;
2222 info.array.raid_disks = s->raiddisks;
2223 } else if (s->raiddisks)
2224 info.delta_disks = s->raiddisks - info.array.raid_disks;
2225 else
2226 info.delta_disks = UnSet;
2227 if (s->layout_str == NULL) {
2228 info.new_layout = UnSet;
2229 if (info.array.level == 6 &&
2230 (info.new_level == 6 || info.new_level == UnSet) &&
2231 info.array.layout >= 16) {
2232 pr_err("%s has a non-standard layout. If you wish to preserve this\n", devname);
2233 cont_err("during the reshape, please specify --layout=preserve\n");
2234 cont_err("If you want to change it, specify a layout or use --layout=normalise\n");
2235 rv = 1;
2236 goto release;
2237 }
2238 } else if (strcmp(s->layout_str, "normalise") == 0 ||
2239 strcmp(s->layout_str, "normalize") == 0) {
2240 /* If we have a -6 RAID6 layout, remove the '-6'. */
2241 info.new_layout = UnSet;
2242 if (info.array.level == 6 && info.new_level == UnSet) {
2243 char l[40], *h;
2244 strcpy(l, map_num_s(r6layout, info.array.layout));
2245 h = strrchr(l, '-');
2246 if (h && strcmp(h, "-6") == 0) {
2247 *h = 0;
2248 info.new_layout = map_name(r6layout, l);
2249 }
2250 } else {
2251 pr_err("%s is only meaningful when reshaping a RAID6 array.\n", s->layout_str);
2252 rv = 1;
2253 goto release;
2254 }
2255 } else if (strcmp(s->layout_str, "preserve") == 0) {
2256 /* This means that a non-standard RAID6 layout
2257 * is OK.
2258 * In particular:
2259 * - When reshape a RAID6 (e.g. adding a device)
2260 * which is in a non-standard layout, it is OK
2261 * to preserve that layout.
2262 * - When converting a RAID5 to RAID6, leave it in
2263 * the XXX-6 layout, don't re-layout.
2264 */
2265 if (info.array.level == 6 && info.new_level == UnSet)
2266 info.new_layout = info.array.layout;
2267 else if (info.array.level == 5 && info.new_level == 6) {
2268 char l[40];
2269 strcpy(l, map_num_s(r5layout, info.array.layout));
2270 strcat(l, "-6");
2271 info.new_layout = map_name(r6layout, l);
2272 } else {
2273 pr_err("%s in only meaningful when reshaping to RAID6\n", s->layout_str);
2274 rv = 1;
2275 goto release;
2276 }
2277 } else {
2278 int l = info.new_level;
2279 if (l == UnSet)
2280 l = info.array.level;
2281 switch (l) {
2282 case 5:
2283 info.new_layout = map_name(r5layout, s->layout_str);
2284 break;
2285 case 6:
2286 info.new_layout = map_name(r6layout, s->layout_str);
2287 break;
2288 case 10:
2289 info.new_layout = parse_layout_10(s->layout_str);
2290 break;
2291 case LEVEL_FAULTY:
2292 info.new_layout = parse_layout_faulty(s->layout_str);
2293 break;
2294 default:
2295 pr_err("layout not meaningful with this level\n");
2296 rv = 1;
2297 goto release;
2298 }
2299 if (info.new_layout == UnSet) {
2300 pr_err("layout %s not understood for this level\n",
2301 s->layout_str);
2302 rv = 1;
2303 goto release;
2304 }
2305 }
2306
2307 if (array.level == LEVEL_FAULTY) {
2308 if (s->level != UnSet && s->level != array.level) {
2309 pr_err("cannot change level of Faulty device\n");
2310 rv =1 ;
2311 }
2312 if (s->chunk) {
2313 pr_err("cannot set chunksize of Faulty device\n");
2314 rv =1 ;
2315 }
2316 if (s->raiddisks && s->raiddisks != 1) {
2317 pr_err("cannot set raid_disks of Faulty device\n");
2318 rv =1 ;
2319 }
2320 if (s->layout_str) {
2321 if (md_get_array_info(fd, &array) != 0) {
2322 dprintf("Cannot get array information.\n");
2323 goto release;
2324 }
2325 array.layout = info.new_layout;
2326 if (md_set_array_info(fd, &array) != 0) {
2327 pr_err("failed to set new layout\n");
2328 rv = 1;
2329 } else if (c->verbose >= 0)
2330 printf("layout for %s set to %d\n",
2331 devname, array.layout);
2332 }
2333 } else if (array.level == LEVEL_CONTAINER) {
2334 /* This change is to be applied to every array in the
2335 * container. This is only needed when the metadata imposes
2336 * restraints of the various arrays in the container.
2337 * Currently we only know that IMSM requires all arrays
2338 * to have the same number of devices so changing the
2339 * number of devices (On-Line Capacity Expansion) must be
2340 * performed at the level of the container
2341 */
2342 close_fd(&fd);
2343 rv = reshape_container(container, devname, -1, st, &info,
2344 c->force, c->backup_file, c->verbose,
2345 0, 0, 0);
2346 frozen = 0;
2347 } else {
2348 /* get spare devices from external metadata
2349 */
2350 if (st->ss->external) {
2351 struct mdinfo *info2;
2352
2353 info2 = st->ss->container_content(st, subarray);
2354 if (info2) {
2355 info.array.spare_disks =
2356 info2->array.spare_disks;
2357 sysfs_free(info2);
2358 }
2359 }
2360
2361 /* Impose these changes on a single array. First
2362 * check that the metadata is OK with the change. */
2363
2364 if (reshape_super(st, 0, info.new_level,
2365 info.new_layout, info.new_chunk,
2366 info.array.raid_disks, info.delta_disks,
2367 c->backup_file, devname,
2368 APPLY_METADATA_CHANGES, c->verbose)) {
2369 rv = 1;
2370 goto release;
2371 }
2372 sync_metadata(st);
2373 rv = reshape_array(container, fd, devname, st, &info, c->force,
2374 devlist, data_offset, c->backup_file,
2375 c->verbose, 0, 0, 0);
2376 frozen = 0;
2377 }
2378 release:
2379 sysfs_free(sra);
2380 if (frozen > 0)
2381 unfreeze(st);
2382 return rv;
2383 }
2384
2385 /* verify_reshape_position()
2386 * Function checks if reshape position in metadata is not farther
2387 * than position in md.
2388 * Return value:
2389 * 0 : not valid sysfs entry
2390 * it can be caused by not started reshape, it should be started
2391 * by reshape array or raid0 array is before takeover
2392 * -1 : error, reshape position is obviously wrong
2393 * 1 : success, reshape progress correct or updated
2394 */
2395 static int verify_reshape_position(struct mdinfo *info, int level)
2396 {
2397 int ret_val = 0;
2398 char buf[40];
2399 int rv;
2400
2401 /* read sync_max, failure can mean raid0 array */
2402 rv = sysfs_get_str(info, NULL, "sync_max", buf, 40);
2403
2404 if (rv > 0) {
2405 char *ep;
2406 unsigned long long position = strtoull(buf, &ep, 0);
2407
2408 dprintf("Read sync_max sysfs entry is: %s\n", buf);
2409 if (!(ep == buf || (*ep != 0 && *ep != '\n' && *ep != ' '))) {
2410 position *= get_data_disks(level,
2411 info->new_layout,
2412 info->array.raid_disks);
2413 if (info->reshape_progress < position) {
2414 dprintf("Corrected reshape progress (%llu) to md position (%llu)\n",
2415 info->reshape_progress, position);
2416 info->reshape_progress = position;
2417 ret_val = 1;
2418 } else if (info->reshape_progress > position) {
2419 pr_err("Fatal error: array reshape was not properly frozen (expected reshape position is %llu, but reshape progress is %llu.\n",
2420 position, info->reshape_progress);
2421 ret_val = -1;
2422 } else {
2423 dprintf("Reshape position in md and metadata are the same;");
2424 ret_val = 1;
2425 }
2426 }
2427 } else if (rv == 0) {
2428 /* for valid sysfs entry, 0-length content
2429 * should be indicated as error
2430 */
2431 ret_val = -1;
2432 }
2433
2434 return ret_val;
2435 }
2436
2437 static unsigned long long choose_offset(unsigned long long lo,
2438 unsigned long long hi,
2439 unsigned long long min,
2440 unsigned long long max)
2441 {
2442 /* Choose a new offset between hi and lo.
2443 * It must be between min and max, but
2444 * we would prefer something near the middle of hi/lo, and also
2445 * prefer to be aligned to a big power of 2.
2446 *
2447 * So we start with the middle, then for each bit,
2448 * starting at '1' and increasing, if it is set, we either
2449 * add it or subtract it if possible, preferring the option
2450 * which is furthest from the boundary.
2451 *
2452 * We stop once we get a 1MB alignment. As units are in sectors,
2453 * 1MB = 2*1024 sectors.
2454 */
2455 unsigned long long choice = (lo + hi) / 2;
2456 unsigned long long bit = 1;
2457
2458 for (bit = 1; bit < 2*1024; bit = bit << 1) {
2459 unsigned long long bigger, smaller;
2460 if (! (bit & choice))
2461 continue;
2462 bigger = choice + bit;
2463 smaller = choice - bit;
2464 if (bigger > max && smaller < min)
2465 break;
2466 if (bigger > max)
2467 choice = smaller;
2468 else if (smaller < min)
2469 choice = bigger;
2470 else if (hi - bigger > smaller - lo)
2471 choice = bigger;
2472 else
2473 choice = smaller;
2474 }
2475 return choice;
2476 }
2477
2478 static int set_new_data_offset(struct mdinfo *sra, struct supertype *st,
2479 char *devname, int delta_disks,
2480 unsigned long long data_offset,
2481 unsigned long long min,
2482 int can_fallback)
2483 {
2484 struct mdinfo *sd;
2485 int dir = 0;
2486 int err = 0;
2487 unsigned long long before, after;
2488
2489 /* Need to find min space before and after so same is used
2490 * on all devices
2491 */
2492 before = UINT64_MAX;
2493 after = UINT64_MAX;
2494 for (sd = sra->devs; sd; sd = sd->next) {
2495 char *dn;
2496 int dfd;
2497 int rv;
2498 struct supertype *st2;
2499 struct mdinfo info2;
2500
2501 if (sd->disk.state & (1<<MD_DISK_FAULTY))
2502 continue;
2503 dn = map_dev(sd->disk.major, sd->disk.minor, 0);
2504 dfd = dev_open(dn, O_RDONLY);
2505 if (dfd < 0) {
2506 pr_err("%s: cannot open component %s\n",
2507 devname, dn ? dn : "-unknown-");
2508 goto release;
2509 }
2510 st2 = dup_super(st);
2511 rv = st2->ss->load_super(st2,dfd, NULL);
2512 close(dfd);
2513 if (rv) {
2514 free(st2);
2515 pr_err("%s: cannot get superblock from %s\n",
2516 devname, dn);
2517 goto release;
2518 }
2519 st2->ss->getinfo_super(st2, &info2, NULL);
2520 st2->ss->free_super(st2);
2521 free(st2);
2522 if (info2.space_before == 0 &&
2523 info2.space_after == 0) {
2524 /* Metadata doesn't support data_offset changes */
2525 if (!can_fallback)
2526 pr_err("%s: Metadata version doesn't support data_offset changes\n",
2527 devname);
2528 goto fallback;
2529 }
2530 if (before > info2.space_before)
2531 before = info2.space_before;
2532 if (after > info2.space_after)
2533 after = info2.space_after;
2534
2535 if (data_offset != INVALID_SECTORS) {
2536 if (dir == 0) {
2537 if (info2.data_offset == data_offset) {
2538 pr_err("%s: already has that data_offset\n",
2539 dn);
2540 goto release;
2541 }
2542 if (data_offset < info2.data_offset)
2543 dir = -1;
2544 else
2545 dir = 1;
2546 } else if ((data_offset <= info2.data_offset &&
2547 dir == 1) ||
2548 (data_offset >= info2.data_offset &&
2549 dir == -1)) {
2550 pr_err("%s: differing data offsets on devices make this --data-offset setting impossible\n",
2551 dn);
2552 goto release;
2553 }
2554 }
2555 }
2556 if (before == UINT64_MAX)
2557 /* impossible really, there must be no devices */
2558 return 1;
2559
2560 for (sd = sra->devs; sd; sd = sd->next) {
2561 char *dn = map_dev(sd->disk.major, sd->disk.minor, 0);
2562 unsigned long long new_data_offset;
2563
2564 if (sd->disk.state & (1<<MD_DISK_FAULTY))
2565 continue;
2566 if (delta_disks < 0) {
2567 /* Don't need any space as array is shrinking
2568 * just move data_offset up by min
2569 */
2570 if (data_offset == INVALID_SECTORS)
2571 new_data_offset = sd->data_offset + min;
2572 else {
2573 if (data_offset < sd->data_offset + min) {
2574 pr_err("--data-offset too small for %s\n",
2575 dn);
2576 goto release;
2577 }
2578 new_data_offset = data_offset;
2579 }
2580 } else if (delta_disks > 0) {
2581 /* need space before */
2582 if (before < min) {
2583 if (can_fallback)
2584 goto fallback;
2585 pr_err("Insufficient head-space for reshape on %s\n",
2586 dn);
2587 goto release;
2588 }
2589 if (data_offset == INVALID_SECTORS)
2590 new_data_offset = sd->data_offset - min;
2591 else {
2592 if (data_offset > sd->data_offset - min) {
2593 pr_err("--data-offset too large for %s\n",
2594 dn);
2595 goto release;
2596 }
2597 new_data_offset = data_offset;
2598 }
2599 } else {
2600 if (dir == 0) {
2601 /* can move up or down. If 'data_offset'
2602 * was set we would have already decided,
2603 * so just choose direction with most space.
2604 */
2605 if (before > after)
2606 dir = -1;
2607 else
2608 dir = 1;
2609 }
2610 sysfs_set_str(sra, NULL, "reshape_direction",
2611 dir == 1 ? "backwards" : "forwards");
2612 if (dir > 0) {
2613 /* Increase data offset */
2614 if (after < min) {
2615 if (can_fallback)
2616 goto fallback;
2617 pr_err("Insufficient tail-space for reshape on %s\n",
2618 dn);
2619 goto release;
2620 }
2621 if (data_offset != INVALID_SECTORS &&
2622 data_offset < sd->data_offset + min) {
2623 pr_err("--data-offset too small on %s\n",
2624 dn);
2625 goto release;
2626 }
2627 if (data_offset != INVALID_SECTORS)
2628 new_data_offset = data_offset;
2629 else
2630 new_data_offset = choose_offset(sd->data_offset,
2631 sd->data_offset + after,
2632 sd->data_offset + min,
2633 sd->data_offset + after);
2634 } else {
2635 /* Decrease data offset */
2636 if (before < min) {
2637 if (can_fallback)
2638 goto fallback;
2639 pr_err("insufficient head-room on %s\n",
2640 dn);
2641 goto release;
2642 }
2643 if (data_offset != INVALID_SECTORS &&
2644 data_offset > sd->data_offset - min) {
2645 pr_err("--data-offset too large on %s\n",
2646 dn);
2647 goto release;
2648 }
2649 if (data_offset != INVALID_SECTORS)
2650 new_data_offset = data_offset;
2651 else
2652 new_data_offset = choose_offset(sd->data_offset - before,
2653 sd->data_offset,
2654 sd->data_offset - before,
2655 sd->data_offset - min);
2656 }
2657 }
2658 err = sysfs_set_num(sra, sd, "new_offset", new_data_offset);
2659 if (err < 0 && errno == E2BIG) {
2660 /* try again after increasing data size to max */
2661 err = sysfs_set_num(sra, sd, "size", 0);
2662 if (err < 0 && errno == EINVAL &&
2663 !(sd->disk.state & (1<<MD_DISK_SYNC))) {
2664 /* some kernels have a bug where you cannot
2665 * use '0' on spare devices. */
2666 sysfs_set_num(sra, sd, "size",
2667 (sra->component_size + after)/2);
2668 }
2669 err = sysfs_set_num(sra, sd, "new_offset",
2670 new_data_offset);
2671 }
2672 if (err < 0) {
2673 if (errno == E2BIG && data_offset != INVALID_SECTORS) {
2674 pr_err("data-offset is too big for %s\n", dn);
2675 goto release;
2676 }
2677 if (sd == sra->devs &&
2678 (errno == ENOENT || errno == E2BIG))
2679 /* Early kernel, no 'new_offset' file,
2680 * or kernel doesn't like us.
2681 * For RAID5/6 this is not fatal
2682 */
2683 return 1;
2684 pr_err("Cannot set new_offset for %s\n", dn);
2685 break;
2686 }
2687 }
2688 return err;
2689 release:
2690 return -1;
2691 fallback:
2692 /* Just use a backup file */
2693 return 1;
2694 }
2695
2696 static int raid10_reshape(char *container, int fd, char *devname,
2697 struct supertype *st, struct mdinfo *info,
2698 struct reshape *reshape,
2699 unsigned long long data_offset,
2700 int force, int verbose)
2701 {
2702 /* Changing raid_disks, layout, chunksize or possibly
2703 * just data_offset for a RAID10.
2704 * We must always change data_offset. We change by at least
2705 * ->min_offset_change which is the largest of the old and new
2706 * chunk sizes.
2707 * If raid_disks is increasing, then data_offset must decrease
2708 * by at least this copy size.
2709 * If raid_disks is unchanged, data_offset must increase or
2710 * decrease by at least min_offset_change but preferably by much more.
2711 * We choose half of the available space.
2712 * If raid_disks is decreasing, data_offset must increase by
2713 * at least min_offset_change. To allow of this, component_size
2714 * must be decreased by the same amount.
2715 *
2716 * So we calculate the required minimum and direction, possibly
2717 * reduce the component_size, then iterate through the devices
2718 * and set the new_data_offset.
2719 * If that all works, we set chunk_size, layout, raid_disks, and start
2720 * 'reshape'
2721 */
2722 struct mdinfo *sra;
2723 unsigned long long min;
2724 int err = 0;
2725
2726 sra = sysfs_read(fd, NULL,
2727 GET_COMPONENT|GET_DEVS|GET_OFFSET|GET_STATE|GET_CHUNK
2728 );
2729 if (!sra) {
2730 pr_err("%s: Cannot get array details from sysfs\n", devname);
2731 goto release;
2732 }
2733 min = reshape->min_offset_change;
2734
2735 if (info->delta_disks)
2736 sysfs_set_str(sra, NULL, "reshape_direction",
2737 info->delta_disks < 0 ? "backwards" : "forwards");
2738 if (info->delta_disks < 0 && info->space_after < min) {
2739 int rv = sysfs_set_num(sra, NULL, "component_size",
2740 (sra->component_size - min)/2);
2741 if (rv) {
2742 pr_err("cannot reduce component size\n");
2743 goto release;
2744 }
2745 }
2746 err = set_new_data_offset(sra, st, devname, info->delta_disks,
2747 data_offset, min, 0);
2748 if (err == 1) {
2749 pr_err("Cannot set new_data_offset: RAID10 reshape not\n");
2750 cont_err("supported on this kernel\n");
2751 err = -1;
2752 }
2753 if (err < 0)
2754 goto release;
2755
2756 if (!err && sysfs_set_num(sra, NULL, "chunk_size", info->new_chunk) < 0)
2757 err = errno;
2758 if (!err && sysfs_set_num(sra, NULL, "layout",
2759 reshape->after.layout) < 0)
2760 err = errno;
2761 if (!err &&
2762 sysfs_set_num(sra, NULL, "raid_disks",
2763 info->array.raid_disks + info->delta_disks) < 0)
2764 err = errno;
2765 if (!err && sysfs_set_str(sra, NULL, "sync_action", "reshape") < 0)
2766 err = errno;
2767 if (err) {
2768 pr_err("Cannot set array shape for %s\n",
2769 devname);
2770 if (err == EBUSY &&
2771 (info->array.state & (1<<MD_SB_BITMAP_PRESENT)))
2772 cont_err(" Bitmap must be removed before shape can be changed\n");
2773 goto release;
2774 }
2775 sysfs_free(sra);
2776 return 0;
2777 release:
2778 sysfs_free(sra);
2779 return 1;
2780 }
2781
2782 static void get_space_after(int fd, struct supertype *st, struct mdinfo *info)
2783 {
2784 struct mdinfo *sra, *sd;
2785 /* Initialisation to silence compiler warning */
2786 unsigned long long min_space_before = 0, min_space_after = 0;
2787 int first = 1;
2788
2789 sra = sysfs_read(fd, NULL, GET_DEVS);
2790 if (!sra)
2791 return;
2792 for (sd = sra->devs; sd; sd = sd->next) {
2793 char *dn;
2794 int dfd;
2795 struct supertype *st2;
2796 struct mdinfo info2;
2797
2798 if (sd->disk.state & (1<<MD_DISK_FAULTY))
2799 continue;
2800 dn = map_dev(sd->disk.major, sd->disk.minor, 0);
2801 dfd = dev_open(dn, O_RDONLY);
2802 if (dfd < 0)
2803 break;
2804 st2 = dup_super(st);
2805 if (st2->ss->load_super(st2,dfd, NULL)) {
2806 close(dfd);
2807 free(st2);
2808 break;
2809 }
2810 close(dfd);
2811 st2->ss->getinfo_super(st2, &info2, NULL);
2812 st2->ss->free_super(st2);
2813 free(st2);
2814 if (first ||
2815 min_space_before > info2.space_before)
2816 min_space_before = info2.space_before;
2817 if (first ||
2818 min_space_after > info2.space_after)
2819 min_space_after = info2.space_after;
2820 first = 0;
2821 }
2822 if (sd == NULL && !first) {
2823 info->space_after = min_space_after;
2824 info->space_before = min_space_before;
2825 }
2826 sysfs_free(sra);
2827 }
2828
2829 static void update_cache_size(char *container, struct mdinfo *sra,
2830 struct mdinfo *info,
2831 int disks, unsigned long long blocks)
2832 {
2833 /* Check that the internal stripe cache is
2834 * large enough, or it won't work.
2835 * It must hold at least 4 stripes of the larger
2836 * chunk size
2837 */
2838 unsigned long cache;
2839 cache = max(info->array.chunk_size, info->new_chunk);
2840 cache *= 4; /* 4 stripes minimum */
2841 cache /= 512; /* convert to sectors */
2842 /* make sure there is room for 'blocks' with a bit to spare */
2843 if (cache < 16 + blocks / disks)
2844 cache = 16 + blocks / disks;
2845 cache /= (4096/512); /* Convert from sectors to pages */
2846
2847 if (sra->cache_size < cache)
2848 subarray_set_num(container, sra, "stripe_cache_size",
2849 cache+1);
2850 }
2851
2852 static int impose_reshape(struct mdinfo *sra,
2853 struct mdinfo *info,
2854 struct supertype *st,
2855 int fd,
2856 int restart,
2857 char *devname, char *container,
2858 struct reshape *reshape)
2859 {
2860 struct mdu_array_info_s array;
2861
2862 sra->new_chunk = info->new_chunk;
2863
2864 if (restart) {
2865 /* for external metadata checkpoint saved by mdmon can be lost
2866 * or missed /due to e.g. crash/. Check if md is not during
2867 * restart farther than metadata points to.
2868 * If so, this means metadata information is obsolete.
2869 */
2870 if (st->ss->external)
2871 verify_reshape_position(info, reshape->level);
2872 sra->reshape_progress = info->reshape_progress;
2873 } else {
2874 sra->reshape_progress = 0;
2875 if (reshape->after.data_disks < reshape->before.data_disks)
2876 /* start from the end of the new array */
2877 sra->reshape_progress = (sra->component_size
2878 * reshape->after.data_disks);
2879 }
2880
2881 md_get_array_info(fd, &array);
2882 if (info->array.chunk_size == info->new_chunk &&
2883 reshape->before.layout == reshape->after.layout &&
2884 st->ss->external == 0) {
2885 /* use SET_ARRAY_INFO but only if reshape hasn't started */
2886 array.raid_disks = reshape->after.data_disks + reshape->parity;
2887 if (!restart && md_set_array_info(fd, &array) != 0) {
2888 int err = errno;
2889
2890 pr_err("Cannot set device shape for %s: %s\n",
2891 devname, strerror(errno));
2892
2893 if (err == EBUSY &&
2894 (array.state & (1<<MD_SB_BITMAP_PRESENT)))
2895 cont_err("Bitmap must be removed before shape can be changed\n");
2896
2897 goto release;
2898 }
2899 } else if (!restart) {
2900 /* set them all just in case some old 'new_*' value
2901 * persists from some earlier problem.
2902 */
2903 int err = 0;
2904 if (sysfs_set_num(sra, NULL, "chunk_size", info->new_chunk) < 0)
2905 err = errno;
2906 if (!err && sysfs_set_num(sra, NULL, "layout",
2907 reshape->after.layout) < 0)
2908 err = errno;
2909 if (!err && subarray_set_num(container, sra, "raid_disks",
2910 reshape->after.data_disks +
2911 reshape->parity) < 0)
2912 err = errno;
2913 if (err) {
2914 pr_err("Cannot set device shape for %s\n", devname);
2915
2916 if (err == EBUSY &&
2917 (array.state & (1<<MD_SB_BITMAP_PRESENT)))
2918 cont_err("Bitmap must be removed before shape can be changed\n");
2919 goto release;
2920 }
2921 }
2922 return 0;
2923 release:
2924 return -1;
2925 }
2926
2927 static int impose_level(int fd, int level, char *devname, int verbose)
2928 {
2929 char *c;
2930 struct mdu_array_info_s array;
2931 struct mdinfo info;
2932
2933 if (sysfs_init(&info, fd, NULL)) {
2934 pr_err("failed to initialize sysfs.\n");
2935 return 1;
2936 }
2937
2938 md_get_array_info(fd, &array);
2939 if (level == 0 && (array.level >= 4 && array.level <= 6)) {
2940 /* To convert to RAID0 we need to fail and
2941 * remove any non-data devices. */
2942 int found = 0;
2943 int d;
2944 int data_disks = array.raid_disks - 1;
2945 if (array.level == 6)
2946 data_disks -= 1;
2947 if (array.level == 5 && array.layout != ALGORITHM_PARITY_N)
2948 return -1;
2949 if (array.level == 6 && array.layout != ALGORITHM_PARITY_N_6)
2950 return -1;
2951 sysfs_set_str(&info, NULL,"sync_action", "idle");
2952 /* First remove any spares so no recovery starts */
2953 for (d = 0, found = 0;
2954 d < MAX_DISKS && found < array.nr_disks; d++) {
2955 mdu_disk_info_t disk;
2956 disk.number = d;
2957 if (md_get_disk_info(fd, &disk) < 0)
2958 continue;
2959 if (disk.major == 0 && disk.minor == 0)
2960 continue;
2961 found++;
2962 if ((disk.state & (1 << MD_DISK_ACTIVE)) &&
2963 disk.raid_disk < data_disks)
2964 /* keep this */
2965 continue;
2966 ioctl(fd, HOT_REMOVE_DISK,
2967 makedev(disk.major, disk.minor));
2968 }
2969 /* Now fail anything left */
2970 md_get_array_info(fd, &array);
2971 for (d = 0, found = 0;
2972 d < MAX_DISKS && found < array.nr_disks; d++) {
2973 mdu_disk_info_t disk;
2974 disk.number = d;
2975 if (md_get_disk_info(fd, &disk) < 0)
2976 continue;
2977 if (disk.major == 0 && disk.minor == 0)
2978 continue;
2979 found++;
2980 if ((disk.state & (1 << MD_DISK_ACTIVE)) &&
2981 disk.raid_disk < data_disks)
2982 /* keep this */
2983 continue;
2984 ioctl(fd, SET_DISK_FAULTY,
2985 makedev(disk.major, disk.minor));
2986 hot_remove_disk(fd, makedev(disk.major, disk.minor), 1);
2987 }
2988 }
2989 c = map_num(pers, level);
2990 if (c) {
2991 int err = sysfs_set_str(&info, NULL, "level", c);
2992 if (err) {
2993 err = errno;
2994 pr_err("%s: could not set level to %s\n",
2995 devname, c);
2996 if (err == EBUSY &&
2997 (array.state & (1<<MD_SB_BITMAP_PRESENT)))
2998 cont_err("Bitmap must be removed before level can be changed\n");
2999 return err;
3000 }
3001 if (verbose >= 0)
3002 pr_err("level of %s changed to %s\n", devname, c);
3003 }
3004 return 0;
3005 }
3006
3007 int sigterm = 0;
3008 static void catch_term(int sig)
3009 {
3010 sigterm = 1;
3011 }
3012
3013 static int reshape_array(char *container, int fd, char *devname,
3014 struct supertype *st, struct mdinfo *info,
3015 int force, struct mddev_dev *devlist,
3016 unsigned long long data_offset,
3017 char *backup_file, int verbose, int forked,
3018 int restart, int freeze_reshape)
3019 {
3020 struct reshape reshape;
3021 int spares_needed;
3022 char *msg;
3023 int orig_level = UnSet;
3024 int odisks;
3025 int delayed;
3026
3027 struct mdu_array_info_s array;
3028 char *c;
3029
3030 struct mddev_dev *dv;
3031 int added_disks;
3032
3033 int *fdlist = NULL;
3034 unsigned long long *offsets = NULL;
3035 int d;
3036 int nrdisks;
3037 int err;
3038 unsigned long blocks;
3039 unsigned long long array_size;
3040 int done;
3041 struct mdinfo *sra = NULL;
3042 char buf[20];
3043
3044 /* when reshaping a RAID0, the component_size might be zero.
3045 * So try to fix that up.
3046 */
3047 if (md_get_array_info(fd, &array) != 0) {
3048 dprintf("Cannot get array information.\n");
3049 goto release;
3050 }
3051 if (array.level == 0 && info->component_size == 0) {
3052 get_dev_size(fd, NULL, &array_size);
3053 info->component_size = array_size / array.raid_disks;
3054 }
3055
3056 if (array.level == 10)
3057 /* Need space_after info */
3058 get_space_after(fd, st, info);
3059
3060 if (info->reshape_active) {
3061 int new_level = info->new_level;
3062 info->new_level = UnSet;
3063 if (info->delta_disks > 0)
3064 info->array.raid_disks -= info->delta_disks;
3065 msg = analyse_change(devname, info, &reshape);
3066 info->new_level = new_level;
3067 if (info->delta_disks > 0)
3068 info->array.raid_disks += info->delta_disks;
3069 if (!restart)
3070 /* Make sure the array isn't read-only */
3071 ioctl(fd, RESTART_ARRAY_RW, 0);
3072 } else
3073 msg = analyse_change(devname, info, &reshape);
3074 if (msg) {
3075 /* if msg == "", error has already been printed */
3076 if (msg[0])
3077 pr_err("%s\n", msg);
3078 goto release;
3079 }
3080 if (restart && (reshape.level != info->array.level ||
3081 reshape.before.layout != info->array.layout ||
3082 reshape.before.data_disks + reshape.parity !=
3083 info->array.raid_disks - max(0, info->delta_disks))) {
3084 pr_err("reshape info is not in native format - cannot continue.\n");
3085 goto release;
3086 }
3087
3088 if (st->ss->external && restart && (info->reshape_progress == 0) &&
3089 !((sysfs_get_str(info, NULL, "sync_action",
3090 buf, sizeof(buf)) > 0) &&
3091 (strncmp(buf, "reshape", 7) == 0))) {
3092 /* When reshape is restarted from '0', very begin of array
3093 * it is possible that for external metadata reshape and array
3094 * configuration doesn't happen.
3095 * Check if md has the same opinion, and reshape is restarted
3096 * from 0. If so, this is regular reshape start after reshape
3097 * switch in metadata to next array only.
3098 */
3099 if ((verify_reshape_position(info, reshape.level) >= 0) &&
3100 (info->reshape_progress == 0))
3101 restart = 0;
3102 }
3103 if (restart) {
3104 /*
3105 * reshape already started. just skip to monitoring
3106 * the reshape
3107 */
3108 if (reshape.backup_blocks == 0)
3109 return 0;
3110 if (restart & RESHAPE_NO_BACKUP)
3111 return 0;
3112
3113 /* Need 'sra' down at 'started:' */
3114 sra = sysfs_read(fd, NULL,
3115 GET_COMPONENT|GET_DEVS|GET_OFFSET|GET_STATE|
3116 GET_CHUNK|GET_CACHE);
3117 if (!sra) {
3118 pr_err("%s: Cannot get array details from sysfs\n",
3119 devname);
3120 goto release;
3121 }
3122
3123 if (!backup_file)
3124 backup_file = locate_backup(sra->sys_name);
3125
3126 goto started;
3127 }
3128 /* The container is frozen but the array may not be.
3129 * So freeze the array so spares don't get put to the wrong use
3130 * FIXME there should probably be a cleaner separation between
3131 * freeze_array and freeze_container.
3132 */
3133 sysfs_freeze_array(info);
3134 /* Check we have enough spares to not be degraded */
3135 added_disks = 0;
3136 for (dv = devlist; dv ; dv=dv->next)
3137 added_disks++;
3138 spares_needed = max(reshape.before.data_disks,
3139 reshape.after.data_disks) +
3140 reshape.parity - array.raid_disks;
3141
3142 if (!force && info->new_level > 1 && info->array.level > 1 &&
3143 spares_needed > info->array.spare_disks + added_disks) {
3144 pr_err("Need %d spare%s to avoid degraded array, and only have %d.\n"
3145 " Use --force to over-ride this check.\n",
3146 spares_needed,
3147 spares_needed == 1 ? "" : "s",
3148 info->array.spare_disks + added_disks);
3149 goto release;
3150 }
3151 /* Check we have enough spares to not fail */
3152 spares_needed = max(reshape.before.data_disks,
3153 reshape.after.data_disks)
3154 - array.raid_disks;
3155 if ((info->new_level > 1 || info->new_level == 0) &&
3156 spares_needed > info->array.spare_disks +added_disks) {
3157 pr_err("Need %d spare%s to create working array, and only have %d.\n",
3158 spares_needed, spares_needed == 1 ? "" : "s",
3159 info->array.spare_disks + added_disks);
3160 goto release;
3161 }
3162
3163 if (reshape.level != array.level) {
3164 int err = impose_level(fd, reshape.level, devname, verbose);
3165 if (err)
3166 goto release;
3167 info->new_layout = UnSet; /* after level change,
3168 * layout is meaningless */
3169 orig_level = array.level;
3170 sysfs_freeze_array(info);
3171
3172 if (reshape.level > 0 && st->ss->external) {
3173 /* make sure mdmon is aware of the new level */
3174 if (mdmon_running(container))
3175 flush_mdmon(container);
3176
3177 if (!mdmon_running(container))
3178 start_mdmon(container);
3179 ping_monitor(container);
3180 if (mdmon_running(container) && st->update_tail == NULL)
3181 st->update_tail = &st->updates;
3182 }
3183 }
3184 /* ->reshape_super might have chosen some spares from the
3185 * container that it wants to be part of the new array.
3186 * We can collect them with ->container_content and give
3187 * them to the kernel.
3188 */
3189 if (st->ss->reshape_super && st->ss->container_content) {
3190 char *subarray = strchr(info->text_version+1, '/')+1;
3191 struct mdinfo *info2 =
3192 st->ss->container_content(st, subarray);
3193 struct mdinfo *d;
3194
3195 if (info2) {
3196 if (sysfs_init(info2, fd, st->devnm)) {
3197 pr_err("unable to initialize sysfs for %s\n",
3198 st->devnm);
3199 free(info2);
3200 goto release;
3201 }
3202 /* When increasing number of devices, we need to set
3203 * new raid_disks before adding these, or they might
3204 * be rejected.
3205 */
3206 if (reshape.backup_blocks &&
3207 reshape.after.data_disks >
3208 reshape.before.data_disks)
3209 subarray_set_num(container, info2, "raid_disks",
3210 reshape.after.data_disks +
3211 reshape.parity);
3212 for (d = info2->devs; d; d = d->next) {
3213 if (d->disk.state == 0 &&
3214 d->disk.raid_disk >= 0) {
3215 /* This is a spare that wants to
3216 * be part of the array.
3217 */
3218 add_disk(fd, st, info2, d);
3219 }
3220 }
3221 sysfs_free(info2);
3222 }
3223 }
3224 /* We might have been given some devices to add to the
3225 * array. Now that the array has been changed to the right
3226 * level and frozen, we can safely add them.
3227 */
3228 if (devlist) {
3229 if (Manage_subdevs(devname, fd, devlist, verbose, 0, NULL, 0))
3230 goto release;
3231 }
3232
3233 if (reshape.backup_blocks == 0 && data_offset != INVALID_SECTORS)
3234 reshape.backup_blocks = reshape.before.data_disks * info->array.chunk_size/512;
3235 if (reshape.backup_blocks == 0) {
3236 /* No restriping needed, but we might need to impose
3237 * some more changes: layout, raid_disks, chunk_size
3238 */
3239 /* read current array info */
3240 if (md_get_array_info(fd, &array) != 0) {
3241 dprintf("Cannot get array information.\n");
3242 goto release;
3243 }
3244 /* compare current array info with new values and if
3245 * it is different update them to new */
3246 if (info->new_layout != UnSet &&
3247 info->new_layout != array.layout) {
3248 array.layout = info->new_layout;
3249 if (md_set_array_info(fd, &array) != 0) {
3250 pr_err("failed to set new layout\n");
3251 goto release;
3252 } else if (verbose >= 0)
3253 printf("layout for %s set to %d\n",
3254 devname, array.layout);
3255 }
3256 if (info->delta_disks != UnSet && info->delta_disks != 0 &&
3257 array.raid_disks !=
3258 (info->array.raid_disks + info->delta_disks)) {
3259 array.raid_disks += info->delta_disks;
3260 if (md_set_array_info(fd, &array) != 0) {
3261 pr_err("failed to set raid disks\n");
3262 goto release;
3263 } else if (verbose >= 0) {
3264 printf("raid_disks for %s set to %d\n",
3265 devname, array.raid_disks);
3266 }
3267 }
3268 if (info->new_chunk != 0 &&
3269 info->new_chunk != array.chunk_size) {
3270 if (sysfs_set_num(info, NULL,
3271 "chunk_size", info->new_chunk) != 0) {
3272 pr_err("failed to set chunk size\n");
3273 goto release;
3274 } else if (verbose >= 0)
3275 printf("chunk size for %s set to %d\n",
3276 devname, info->new_chunk);
3277 }
3278 unfreeze(st);
3279 return 0;
3280 }
3281
3282 /*
3283 * There are three possibilities.
3284 * 1/ The array will shrink.
3285 * We need to ensure the reshape will pause before reaching
3286 * the 'critical section'. We also need to fork and wait for
3287 * that to happen. When it does we
3288 * suspend/backup/complete/unfreeze
3289 *
3290 * 2/ The array will not change size.
3291 * This requires that we keep a backup of a sliding window
3292 * so that we can restore data after a crash. So we need
3293 * to fork and monitor progress.
3294 * In future we will allow the data_offset to change, so
3295 * a sliding backup becomes unnecessary.
3296 *
3297 * 3/ The array will grow. This is relatively easy.
3298 * However the kernel's restripe routines will cheerfully
3299 * overwrite some early data before it is safe. So we
3300 * need to make a backup of the early parts of the array
3301 * and be ready to restore it if rebuild aborts very early.
3302 * For externally managed metadata, we still need a forked
3303 * child to monitor the reshape and suspend IO over the region
3304 * that is being reshaped.
3305 *
3306 * We backup data by writing it to one spare, or to a
3307 * file which was given on command line.
3308 *
3309 * In each case, we first make sure that storage is available
3310 * for the required backup.
3311 * Then we:
3312 * - request the shape change.
3313 * - fork to handle backup etc.
3314 */
3315 /* Check that we can hold all the data */
3316 get_dev_size(fd, NULL, &array_size);
3317 if (reshape.new_size < (array_size/512)) {
3318 pr_err("this change will reduce the size of the array.\n"
3319 " use --grow --array-size first to truncate array.\n"
3320 " e.g. mdadm --grow %s --array-size %llu\n",
3321 devname, reshape.new_size/2);
3322 goto release;
3323 }
3324
3325 if (array.level == 10) {
3326 /* Reshaping RAID10 does not require any data backup by
3327 * user-space. Instead it requires that the data_offset
3328 * is changed to avoid the need for backup.
3329 * So this is handled very separately
3330 */
3331 if (restart)
3332 /* Nothing to do. */
3333 return 0;
3334 return raid10_reshape(container, fd, devname, st, info,
3335 &reshape, data_offset, force, verbose);
3336 }
3337 sra = sysfs_read(fd, NULL,
3338 GET_COMPONENT|GET_DEVS|GET_OFFSET|GET_STATE|GET_CHUNK|
3339 GET_CACHE);
3340 if (!sra) {
3341 pr_err("%s: Cannot get array details from sysfs\n",
3342 devname);
3343 goto release;
3344 }
3345
3346 if (!backup_file)
3347 switch(set_new_data_offset(sra, st, devname,
3348 reshape.after.data_disks - reshape.before.data_disks,
3349 data_offset,
3350 reshape.min_offset_change, 1)) {
3351 case -1:
3352 goto release;
3353 case 0:
3354 /* Updated data_offset, so it's easy now */
3355 update_cache_size(container, sra, info,
3356 min(reshape.before.data_disks,
3357 reshape.after.data_disks),
3358 reshape.backup_blocks);
3359
3360 /* Right, everything seems fine. Let's kick things off.
3361 */
3362 sync_metadata(st);
3363
3364 if (impose_reshape(sra, info, st, fd, restart,
3365 devname, container, &reshape) < 0)
3366 goto release;
3367 if (sysfs_set_str(sra, NULL, "sync_action", "reshape") < 0) {
3368 struct mdinfo *sd;
3369 if (errno != EINVAL) {
3370 pr_err("Failed to initiate reshape!\n");
3371 goto release;
3372 }
3373 /* revert data_offset and try the old way */
3374 for (sd = sra->devs; sd; sd = sd->next) {
3375 sysfs_set_num(sra, sd, "new_offset",
3376 sd->data_offset);
3377 sysfs_set_str(sra, NULL, "reshape_direction",
3378 "forwards");
3379 }
3380 break;
3381 }
3382 if (info->new_level == reshape.level)
3383 return 0;
3384 /* need to adjust level when reshape completes */
3385 switch(fork()) {
3386 case -1: /* ignore error, but don't wait */
3387 return 0;
3388 default: /* parent */
3389 return 0;
3390 case 0:
3391 manage_fork_fds(0);
3392 map_fork();
3393 break;
3394 }
3395 close(fd);
3396 wait_reshape(sra);
3397 fd = open_dev(sra->sys_name);
3398 if (fd >= 0)
3399 impose_level(fd, info->new_level, devname, verbose);
3400 return 0;
3401 case 1: /* Couldn't set data_offset, try the old way */
3402 if (data_offset != INVALID_SECTORS) {
3403 pr_err("Cannot update data_offset on this array\n");
3404 goto release;
3405 }
3406 break;
3407 }
3408
3409 started:
3410 /* Decide how many blocks (sectors) for a reshape
3411 * unit. The number we have so far is just a minimum
3412 */
3413 blocks = reshape.backup_blocks;
3414 if (reshape.before.data_disks ==
3415 reshape.after.data_disks) {
3416 /* Make 'blocks' bigger for better throughput, but
3417 * not so big that we reject it below.
3418 * Try for 16 megabytes
3419 */
3420 while (blocks * 32 < sra->component_size && blocks < 16*1024*2)
3421 blocks *= 2;
3422 } else
3423 pr_err("Need to backup %luK of critical section..\n", blocks/2);
3424
3425 if (blocks >= sra->component_size/2) {
3426 pr_err("%s: Something wrong - reshape aborted\n", devname);
3427 goto release;
3428 }
3429
3430 /* Now we need to open all these devices so we can read/write.
3431 */
3432 nrdisks = max(reshape.before.data_disks,
3433 reshape.after.data_disks) + reshape.parity
3434 + sra->array.spare_disks;
3435 fdlist = xcalloc((1+nrdisks), sizeof(int));
3436 offsets = xcalloc((1+nrdisks), sizeof(offsets[0]));
3437
3438 odisks = reshape.before.data_disks + reshape.parity;
3439 d = reshape_prepare_fdlist(devname, sra, odisks, nrdisks, blocks,
3440 backup_file, fdlist, offsets);
3441 if (d < odisks) {
3442 goto release;
3443 }
3444 if ((st->ss->manage_reshape == NULL) ||
3445 (st->ss->recover_backup == NULL)) {
3446 if (backup_file == NULL) {
3447 if (reshape.after.data_disks <=
3448 reshape.before.data_disks) {
3449 pr_err("%s: Cannot grow - need backup-file\n",
3450 devname);
3451 pr_err(" Please provide one with \"--backup=...\"\n");
3452 goto release;
3453 } else if (d == odisks) {
3454 pr_err("%s: Cannot grow - need a spare or backup-file to backup critical section\n", devname);
3455 goto release;
3456 }
3457 } else {
3458 if (!reshape_open_backup_file(backup_file, fd, devname,
3459 (signed)blocks,
3460 fdlist+d, offsets+d,
3461 sra->sys_name, restart)) {
3462 goto release;
3463 }
3464 d++;
3465 }
3466 }
3467
3468 update_cache_size(container, sra, info,
3469 min(reshape.before.data_disks,
3470 reshape.after.data_disks), blocks);
3471
3472 /* Right, everything seems fine. Let's kick things off.
3473 * If only changing raid_disks, use ioctl, else use
3474 * sysfs.
3475 */
3476 sync_metadata(st);
3477
3478 if (impose_reshape(sra, info, st, fd, restart,
3479 devname, container, &reshape) < 0)
3480 goto release;
3481
3482 err = start_reshape(sra, restart, reshape.before.data_disks,
3483 reshape.after.data_disks, st);
3484 if (err) {
3485 pr_err("Cannot %s reshape for %s\n",
3486 restart ? "continue" : "start", devname);
3487 goto release;
3488 }
3489 if (restart)
3490 sysfs_set_str(sra, NULL, "array_state", "active");
3491 if (freeze_reshape) {
3492 free(fdlist);
3493 free(offsets);
3494 sysfs_free(sra);
3495 pr_err("Reshape has to be continued from location %llu when root filesystem has been mounted.\n",
3496 sra->reshape_progress);
3497 return 1;
3498 }
3499
3500 if (!forked)
3501 if (continue_via_systemd(container ?: sra->sys_name,
3502 GROW_SERVICE)) {
3503 free(fdlist);
3504 free(offsets);
3505 sysfs_free(sra);
3506 return 0;
3507 }
3508
3509 close(fd);
3510 /* Now we just need to kick off the reshape and watch, while
3511 * handling backups of the data...
3512 * This is all done by a forked background process.
3513 */
3514 switch(forked ? 0 : fork()) {
3515 case -1:
3516 pr_err("Cannot run child to monitor reshape: %s\n",
3517 strerror(errno));
3518 abort_reshape(sra);
3519 goto release;
3520 default:
3521 free(fdlist);
3522 free(offsets);
3523 sysfs_free(sra);
3524 return 0;
3525 case 0:
3526 map_fork();
3527 break;
3528 }
3529
3530 /* If another array on the same devices is busy, the
3531 * reshape will wait for them. This would mean that
3532 * the first section that we suspend will stay suspended
3533 * for a long time. So check on that possibility
3534 * by looking for "DELAYED" in /proc/mdstat, and if found,
3535 * wait a while
3536 */
3537 do {
3538 struct mdstat_ent *mds, *m;
3539 delayed = 0;
3540 mds = mdstat_read(1, 0);
3541 for (m = mds; m; m = m->next)
3542 if (strcmp(m->devnm, sra->sys_name) == 0) {
3543 if (m->resync && m->percent == RESYNC_DELAYED)
3544 delayed = 1;
3545 if (m->resync == 0)
3546 /* Haven't started the reshape thread
3547 * yet, wait a bit
3548 */
3549 delayed = 2;
3550 break;
3551 }
3552 free_mdstat(mds);
3553 if (delayed == 1 && get_linux_version() < 3007000) {
3554 pr_err("Reshape is delayed, but cannot wait carefully with this kernel.\n"
3555 " You might experience problems until other reshapes complete.\n");
3556 delayed = 0;
3557 }
3558 if (delayed)
3559 mdstat_wait(30 - (delayed-1) * 25);
3560 } while (delayed);
3561 mdstat_close();
3562 if (check_env("MDADM_GROW_VERIFY"))
3563 fd = open(devname, O_RDONLY | O_DIRECT);
3564 else
3565 fd = -1;
3566 mlockall(MCL_FUTURE);
3567
3568 if (signal_s(SIGTERM, catch_term) == SIG_ERR)
3569 goto release;
3570
3571 if (st->ss->external) {
3572 /* metadata handler takes it from here */
3573 done = st->ss->manage_reshape(
3574 fd, sra, &reshape, st, blocks,
3575 fdlist, offsets, d - odisks, fdlist + odisks,
3576 offsets + odisks);
3577 } else
3578 done = child_monitor(
3579 fd, sra, &reshape, st, blocks, fdlist, offsets,
3580 d - odisks, fdlist + odisks, offsets + odisks);
3581
3582 free(fdlist);
3583 free(offsets);
3584
3585 if (backup_file && done) {
3586 char *bul;
3587 bul = make_backup(sra->sys_name);
3588 if (bul) {
3589 char buf[1024];
3590 int l = readlink(bul, buf, sizeof(buf) - 1);
3591 if (l > 0) {
3592 buf[l]=0;
3593 unlink(buf);
3594 }
3595 unlink(bul);
3596 free(bul);
3597 }
3598 unlink(backup_file);
3599 }
3600 if (!done) {
3601 abort_reshape(sra);
3602 goto out;
3603 }
3604
3605 if (!st->ss->external &&
3606 !(reshape.before.data_disks != reshape.after.data_disks &&
3607 info->custom_array_size) && info->new_level == reshape.level &&
3608 !forked) {
3609 /* no need to wait for the reshape to finish as
3610 * there is nothing more to do.
3611 */
3612 sysfs_free(sra);
3613 exit(0);
3614 }
3615 wait_reshape(sra);
3616
3617 if (st->ss->external) {
3618 /* Re-load the metadata as much could have changed */
3619 int cfd = open_dev(st->container_devnm);
3620 if (cfd >= 0) {
3621 flush_mdmon(container);
3622 st->ss->free_super(st);
3623 st->ss->load_container(st, cfd, container);
3624 close(cfd);
3625 }
3626 }
3627
3628 /* set new array size if required customer_array_size is used
3629 * by this metadata.
3630 */
3631 if (reshape.before.data_disks != reshape.after.data_disks &&
3632 info->custom_array_size)
3633 set_array_size(st, info, info->text_version);
3634
3635 if (info->new_level != reshape.level) {
3636 if (fd < 0)
3637 fd = open(devname, O_RDONLY);
3638 impose_level(fd, info->new_level, devname, verbose);
3639 close(fd);
3640 if (info->new_level == 0)
3641 st->update_tail = NULL;
3642 }
3643 out:
3644 sysfs_free(sra);
3645 if (forked)
3646 return 0;
3647 unfreeze(st);
3648 exit(0);
3649
3650 release:
3651 free(fdlist);
3652 free(offsets);
3653 if (orig_level != UnSet && sra) {
3654 c = map_num(pers, orig_level);
3655 if (c && sysfs_set_str(sra, NULL, "level", c) == 0)
3656 pr_err("aborting level change\n");
3657 }
3658 sysfs_free(sra);
3659 if (!forked)
3660 unfreeze(st);
3661 return 1;
3662 }
3663
3664 /* mdfd handle is passed to be closed in child process (after fork).
3665 */
3666 int reshape_container(char *container, char *devname,
3667 int mdfd,
3668 struct supertype *st,
3669 struct mdinfo *info,
3670 int force,
3671 char *backup_file, int verbose,
3672 int forked, int restart, int freeze_reshape)
3673 {
3674 struct mdinfo *cc = NULL;
3675 int rv = restart;
3676 char last_devnm[32] = "";
3677
3678 /* component_size is not meaningful for a container,
3679 * so pass '0' meaning 'no change'
3680 */
3681 if (!restart &&
3682 reshape_super(st, 0, info->new_level,
3683 info->new_layout, info->new_chunk,
3684 info->array.raid_disks, info->delta_disks,
3685 backup_file, devname, APPLY_METADATA_CHANGES,
3686 verbose)) {
3687 unfreeze(st);
3688 return 1;
3689 }
3690
3691 sync_metadata(st);
3692
3693 /* ping monitor to be sure that update is on disk
3694 */
3695 ping_monitor(container);
3696
3697 if (!forked && !freeze_reshape)
3698 if (continue_via_systemd(container, GROW_SERVICE))
3699 return 0;
3700
3701 switch (forked ? 0 : fork()) {
3702 case -1: /* error */
3703 perror("Cannot fork to complete reshape\n");
3704 unfreeze(st);
3705 return 1;
3706 default: /* parent */
3707 if (!freeze_reshape)
3708 printf("%s: multi-array reshape continues in background\n", Name);
3709 return 0;
3710 case 0: /* child */
3711 manage_fork_fds(0);
3712 map_fork();
3713 break;
3714 }
3715
3716 /* close unused handle in child process
3717 */
3718 if (mdfd > -1)
3719 close(mdfd);
3720
3721 while(1) {
3722 /* For each member array with reshape_active,
3723 * we need to perform the reshape.
3724 * We pick the first array that needs reshaping and
3725 * reshape it. reshape_array() will re-read the metadata
3726 * so the next time through a different array should be
3727 * ready for reshape.
3728 * It is possible that the 'different' array will not
3729 * be assembled yet. In that case we simple exit.
3730 * When it is assembled, the mdadm which assembles it
3731 * will take over the reshape.
3732 */
3733 struct mdinfo *content;
3734 int fd;
3735 struct mdstat_ent *mdstat;
3736 char *adev;
3737 dev_t devid;
3738
3739 sysfs_free(cc);
3740
3741 cc = st->ss->container_content(st, NULL);
3742
3743 for (content = cc; content ; content = content->next) {
3744 char *subarray;
3745 if (!content->reshape_active)
3746 continue;
3747
3748 subarray = strchr(content->text_version+1, '/')+1;
3749 mdstat = mdstat_by_subdev(subarray, container);
3750 if (!mdstat)
3751 continue;
3752 if (mdstat->active == 0) {
3753 pr_err("Skipping inactive array %s.\n",
3754 mdstat->devnm);
3755 free_mdstat(mdstat);
3756 mdstat = NULL;
3757 continue;
3758 }
3759 break;
3760 }
3761 if (!content)
3762 break;
3763
3764 devid = devnm2devid(mdstat->devnm);
3765 adev = map_dev(major(devid), minor(devid), 0);
3766 if (!adev)
3767 adev = content->text_version;
3768
3769 fd = open_dev(mdstat->devnm);
3770 if (fd < 0) {
3771 pr_err("Device %s cannot be opened for reshape.\n",
3772 adev);
3773 break;
3774 }
3775
3776 if (strcmp(last_devnm, mdstat->devnm) == 0) {
3777 /* Do not allow for multiple reshape_array() calls for
3778 * the same array.
3779 * It can happen when reshape_array() returns without
3780 * error, when reshape is not finished (wrong reshape
3781 * starting/continuation conditions). Mdmon doesn't
3782 * switch to next array in container and reentry
3783 * conditions for the same array occur.
3784 * This is possibly interim until the behaviour of
3785 * reshape_array is resolved().
3786 */
3787 printf("%s: Multiple reshape execution detected for device %s.\n", Name, adev);
3788 close(fd);
3789 break;
3790 }
3791 strcpy(last_devnm, mdstat->devnm);
3792
3793 if (sysfs_init(content, fd, mdstat->devnm)) {
3794 pr_err("Unable to initialize sysfs for %s\n",
3795 mdstat->devnm);
3796 rv = 1;
3797 break;
3798 }
3799
3800 if (mdmon_running(container))
3801 flush_mdmon(container);
3802
3803 rv = reshape_array(container, fd, adev, st,
3804 content, force, NULL, INVALID_SECTORS,
3805 backup_file, verbose, 1, restart,
3806 freeze_reshape);
3807 close(fd);
3808
3809 if (freeze_reshape) {
3810 sysfs_free(cc);
3811 exit(0);
3812 }
3813
3814 restart = 0;
3815 if (rv)
3816 break;
3817
3818 if (mdmon_running(container))
3819 flush_mdmon(container);
3820 }
3821 if (!rv)
3822 unfreeze(st);
3823 sysfs_free(cc);
3824 exit(0);
3825 }
3826
3827 /*
3828 * We run a child process in the background which performs the following
3829 * steps:
3830 * - wait for resync to reach a certain point
3831 * - suspend io to the following section
3832 * - backup that section
3833 * - allow resync to proceed further
3834 * - resume io
3835 * - discard the backup.
3836 *
3837 * When are combined in slightly different ways in the three cases.
3838 * Grow:
3839 * - suspend/backup/allow/wait/resume/discard
3840 * Shrink:
3841 * - allow/wait/suspend/backup/allow/wait/resume/discard
3842 * same-size:
3843 * - wait/resume/discard/suspend/backup/allow
3844 *
3845 * suspend/backup/allow always come together
3846 * wait/resume/discard do too.
3847 * For the same-size case we have two backups to improve flow.
3848 *
3849 */
3850
3851 int progress_reshape(struct mdinfo *info, struct reshape *reshape,
3852 unsigned long long backup_point,
3853 unsigned long long wait_point,
3854 unsigned long long *suspend_point,
3855 unsigned long long *reshape_completed, int *frozen)
3856 {
3857 /* This function is called repeatedly by the reshape manager.
3858 * It determines how much progress can safely be made and allows
3859 * that progress.
3860 * - 'info' identifies the array and particularly records in
3861 * ->reshape_progress the metadata's knowledge of progress
3862 * This is a sector offset from the start of the array
3863 * of the next array block to be relocated. This number
3864 * may increase from 0 or decrease from array_size, depending
3865 * on the type of reshape that is happening.
3866 * Note that in contrast, 'sync_completed' is a block count of the
3867 * reshape so far. It gives the distance between the start point
3868 * (head or tail of device) and the next place that data will be
3869 * written. It always increases.
3870 * - 'reshape' is the structure created by analyse_change
3871 * - 'backup_point' shows how much the metadata manager has backed-up
3872 * data. For reshapes with increasing progress, it is the next address
3873 * to be backed up, previous addresses have been backed-up. For
3874 * decreasing progress, it is the earliest address that has been
3875 * backed up - later address are also backed up.
3876 * So addresses between reshape_progress and backup_point are
3877 * backed up providing those are in the 'correct' order.
3878 * - 'wait_point' is an array address. When reshape_completed
3879 * passes this point, progress_reshape should return. It might
3880 * return earlier if it determines that ->reshape_progress needs
3881 * to be updated or further backup is needed.
3882 * - suspend_point is maintained by progress_reshape and the caller
3883 * should not touch it except to initialise to zero.
3884 * It is an array address and it only increases in 2.6.37 and earlier.
3885 * This makes it difficult to handle reducing reshapes with
3886 * external metadata.
3887 * However: it is similar to backup_point in that it records the
3888 * other end of a suspended region from reshape_progress.
3889 * it is moved to extend the region that is safe to backup and/or
3890 * reshape
3891 * - reshape_completed is read from sysfs and returned. The caller
3892 * should copy this into ->reshape_progress when it has reason to
3893 * believe that the metadata knows this, and any backup outside this
3894 * has been erased.
3895 *
3896 * Return value is:
3897 * 1 if more data from backup_point - but only as far as suspend_point,
3898 * should be backed up
3899 * 0 if things are progressing smoothly
3900 * -1 if the reshape is finished because it is all done,
3901 * -2 if the reshape is finished due to an error.
3902 */
3903
3904 int advancing = (reshape->after.data_disks
3905 >= reshape->before.data_disks);
3906 unsigned long long need_backup; /* All data between start of array and
3907 * here will at some point need to
3908 * be backed up.
3909 */
3910 unsigned long long read_offset, write_offset;
3911 unsigned long long write_range;
3912 unsigned long long max_progress, target, completed;
3913 unsigned long long array_size = (info->component_size
3914 * reshape->before.data_disks);
3915 int fd;
3916 char buf[20];
3917
3918 /* First, we unsuspend any region that is now known to be safe.
3919 * If suspend_point is on the 'wrong' side of reshape_progress, then
3920 * we don't have or need suspension at the moment. This is true for
3921 * native metadata when we don't need to back-up.
3922 */
3923 if (advancing) {
3924 if (info->reshape_progress <= *suspend_point)
3925 sysfs_set_num(info, NULL, "suspend_lo",
3926 info->reshape_progress);
3927 } else {
3928 /* Note: this won't work in 2.6.37 and before.
3929 * Something somewhere should make sure we don't need it!
3930 */
3931 if (info->reshape_progress >= *suspend_point)
3932 sysfs_set_num(info, NULL, "suspend_hi",
3933 info->reshape_progress);
3934 }
3935
3936 /* Now work out how far it is safe to progress.
3937 * If the read_offset for ->reshape_progress is less than
3938 * 'blocks' beyond the write_offset, we can only progress as far
3939 * as a backup.
3940 * Otherwise we can progress until the write_offset for the new location
3941 * reaches (within 'blocks' of) the read_offset at the current location.
3942 * However that region must be suspended unless we are using native
3943 * metadata.
3944 * If we need to suspend more, we limit it to 128M per device, which is
3945 * rather arbitrary and should be some time-based calculation.
3946 */
3947 read_offset = info->reshape_progress / reshape->before.data_disks;
3948 write_offset = info->reshape_progress / reshape->after.data_disks;
3949 write_range = info->new_chunk/512;
3950 if (reshape->before.data_disks == reshape->after.data_disks)
3951 need_backup = array_size;
3952 else
3953 need_backup = reshape->backup_blocks;
3954 if (advancing) {
3955 if (read_offset < write_offset + write_range)
3956 max_progress = backup_point;
3957 else
3958 max_progress =
3959 read_offset * reshape->after.data_disks;
3960 } else {
3961 if (read_offset > write_offset - write_range)
3962 /* Can only progress as far as has been backed up,
3963 * which must be suspended */
3964 max_progress = backup_point;
3965 else if (info->reshape_progress <= need_backup)
3966 max_progress = backup_point;
3967 else {
3968 if (info->array.major_version >= 0)
3969 /* Can progress until backup is needed */
3970 max_progress = need_backup;
3971 else {
3972 /* Can progress until metadata update is required */
3973 max_progress =
3974 read_offset * reshape->after.data_disks;
3975 /* but data must be suspended */
3976 if (max_progress < *suspend_point)
3977 max_progress = *suspend_point;
3978 }
3979 }
3980 }
3981
3982 /* We know it is safe to progress to 'max_progress' providing
3983 * it is suspended or we are using native metadata.
3984 * Consider extending suspend_point 128M per device if it
3985 * is less than 64M per device beyond reshape_progress.
3986 * But always do a multiple of 'blocks'
3987 * FIXME this is too big - it takes to long to complete
3988 * this much.
3989 */
3990 target = 64*1024*2 * min(reshape->before.data_disks,
3991 reshape->after.data_disks);
3992 target /= reshape->backup_blocks;
3993 if (target < 2)
3994 target = 2;
3995 target *= reshape->backup_blocks;
3996
3997 /* For externally managed metadata we always need to suspend IO to
3998 * the area being reshaped so we regularly push suspend_point forward.
3999 * For native metadata we only need the suspend if we are going to do
4000 * a backup.
4001 */
4002 if (advancing) {
4003 if ((need_backup > info->reshape_progress ||
4004 info->array.major_version < 0) &&
4005 *suspend_point < info->reshape_progress + target) {
4006 if (need_backup < *suspend_point + 2 * target)
4007 *suspend_point = need_backup;
4008 else if (*suspend_point + 2 * target < array_size)
4009 *suspend_point += 2 * target;
4010 else
4011 *suspend_point = array_size;
4012 sysfs_set_num(info, NULL, "suspend_hi", *suspend_point);
4013 if (max_progress > *suspend_point)
4014 max_progress = *suspend_point;
4015 }
4016 } else {
4017 if (info->array.major_version >= 0) {
4018 /* Only need to suspend when about to backup */
4019 if (info->reshape_progress < need_backup * 2 &&
4020 *suspend_point > 0) {
4021 *suspend_point = 0;
4022 sysfs_set_num(info, NULL, "suspend_lo", 0);
4023 sysfs_set_num(info, NULL, "suspend_hi",
4024 need_backup);
4025 }
4026 } else {
4027 /* Need to suspend continually */
4028 if (info->reshape_progress < *suspend_point)
4029 *suspend_point = info->reshape_progress;
4030 if (*suspend_point + target < info->reshape_progress)
4031 /* No need to move suspend region yet */;
4032 else {
4033 if (*suspend_point >= 2 * target)
4034 *suspend_point -= 2 * target;
4035 else
4036 *suspend_point = 0;
4037 sysfs_set_num(info, NULL, "suspend_lo",
4038 *suspend_point);
4039 }
4040 if (max_progress < *suspend_point)
4041 max_progress = *suspend_point;
4042 }
4043 }
4044
4045 /* now set sync_max to allow that progress. sync_max, like
4046 * sync_completed is a count of sectors written per device, so
4047 * we find the difference between max_progress and the start point,
4048 * and divide that by after.data_disks to get a sync_max
4049 * number.
4050 * At the same time we convert wait_point to a similar number
4051 * for comparing against sync_completed.
4052 */
4053 /* scale down max_progress to per_disk */
4054 max_progress /= reshape->after.data_disks;
4055 /*
4056 * Round to chunk size as some kernels give an erroneously
4057 * high number
4058 */
4059 max_progress /= info->new_chunk/512;
4060 max_progress *= info->new_chunk/512;
4061 /* And round to old chunk size as the kernel wants that */
4062 max_progress /= info->array.chunk_size/512;
4063 max_progress *= info->array.chunk_size/512;
4064 /* Limit progress to the whole device */
4065 if (max_progress > info->component_size)
4066 max_progress = info->component_size;
4067 wait_point /= reshape->after.data_disks;
4068 if (!advancing) {
4069 /* switch from 'device offset' to 'processed block count' */
4070 max_progress = info->component_size - max_progress;
4071 wait_point = info->component_size - wait_point;
4072 }
4073
4074 if (!*frozen)
4075 sysfs_set_num(info, NULL, "sync_max", max_progress);
4076
4077 /* Now wait. If we have already reached the point that we were
4078 * asked to wait to, don't wait at all, else wait for any change.
4079 * We need to select on 'sync_completed' as that is the place that
4080 * notifications happen, but we are really interested in
4081 * 'reshape_position'
4082 */
4083 fd = sysfs_get_fd(info, NULL, "sync_completed");
4084 if (fd < 0)
4085 goto check_progress;
4086
4087 if (sysfs_fd_get_ll(fd, &completed) < 0)
4088 goto check_progress;
4089
4090 while (completed < max_progress && completed < wait_point) {
4091 /* Check that sync_action is still 'reshape' to avoid
4092 * waiting forever on a dead array
4093 */
4094 char action[20];
4095 if (sysfs_get_str(info, NULL, "sync_action", action, 20) <= 0 ||
4096 strncmp(action, "reshape", 7) != 0)
4097 break;
4098 /* Some kernels reset 'sync_completed' to zero
4099 * before setting 'sync_action' to 'idle'.
4100 * So we need these extra tests.
4101 */
4102 if (completed == 0 && advancing &&
4103 strncmp(action, "idle", 4) == 0 &&
4104 info->reshape_progress > 0)
4105 break;
4106 if (completed == 0 && !advancing &&
4107 strncmp(action, "idle", 4) == 0 &&
4108 info->reshape_progress <
4109 (info->component_size * reshape->after.data_disks))
4110 break;
4111 sysfs_wait(fd, NULL);
4112 if (sysfs_fd_get_ll(fd, &completed) < 0)
4113 goto check_progress;
4114 }
4115 /* Some kernels reset 'sync_completed' to zero,
4116 * we need to have real point we are in md.
4117 * So in that case, read 'reshape_position' from sysfs.
4118 */
4119 if (completed == 0) {
4120 unsigned long long reshapep;
4121 char action[20];
4122 if (sysfs_get_str(info, NULL, "sync_action", action, 20) > 0 &&
4123 strncmp(action, "idle", 4) == 0 &&
4124 sysfs_get_ll(info, NULL,
4125 "reshape_position", &reshapep) == 0)
4126 *reshape_completed = reshapep;
4127 } else {
4128 /* some kernels can give an incorrectly high
4129 * 'completed' number, so round down */
4130 completed /= (info->new_chunk/512);
4131 completed *= (info->new_chunk/512);
4132 /* Convert 'completed' back in to a 'progress' number */
4133 completed *= reshape->after.data_disks;
4134 if (!advancing)
4135 completed = (info->component_size
4136 * reshape->after.data_disks
4137 - completed);
4138 *reshape_completed = completed;
4139 }
4140
4141 close(fd);
4142
4143 /* We return the need_backup flag. Caller will decide
4144 * how much - a multiple of ->backup_blocks up to *suspend_point
4145 */
4146 if (advancing)
4147 return need_backup > info->reshape_progress;
4148 else
4149 return need_backup >= info->reshape_progress;
4150
4151 check_progress:
4152 /* if we couldn't read a number from sync_completed, then
4153 * either the reshape did complete, or it aborted.
4154 * We can tell which by checking for 'none' in reshape_position.
4155 * If it did abort, then it might immediately restart if it
4156 * it was just a device failure that leaves us degraded but
4157 * functioning.
4158 */
4159 if (sysfs_get_str(info, NULL, "reshape_position", buf,
4160 sizeof(buf)) < 0 || strncmp(buf, "none", 4) != 0) {
4161 /* The abort might only be temporary. Wait up to 10
4162 * seconds for fd to contain a valid number again.
4163 */
4164 int wait = 10000;
4165 int rv = -2;
4166 unsigned long long new_sync_max;
4167 while (fd >= 0 && rv < 0 && wait > 0) {
4168 if (sysfs_wait(fd, &wait) != 1)
4169 break;
4170 switch (sysfs_fd_get_ll(fd, &completed)) {
4171 case 0:
4172 /* all good again */
4173 rv = 1;
4174 /* If "sync_max" is no longer max_progress
4175 * we need to freeze things
4176 */
4177 sysfs_get_ll(info, NULL, "sync_max",
4178 &new_sync_max);
4179 *frozen = (new_sync_max != max_progress);
4180 break;
4181 case -2: /* read error - abort */
4182 wait = 0;
4183 break;
4184 }
4185 }
4186 if (fd >= 0)
4187 close(fd);
4188 return rv; /* abort */
4189 } else {
4190 /* Maybe racing with array shutdown - check state */
4191 if (fd >= 0)
4192 close(fd);
4193 if (sysfs_get_str(info, NULL, "array_state", buf,
4194 sizeof(buf)) < 0 ||
4195 strncmp(buf, "inactive", 8) == 0 ||
4196 strncmp(buf, "clear",5) == 0)
4197 return -2; /* abort */
4198 return -1; /* complete */
4199 }
4200 }
4201
4202 /* FIXME return status is never checked */
4203 static int grow_backup(struct mdinfo *sra,
4204 unsigned long long offset, /* per device */
4205 unsigned long stripes, /* per device, in old chunks */
4206 int *sources, unsigned long long *offsets,
4207 int disks, int chunk, int level, int layout,
4208 int dests, int *destfd, unsigned long long *destoffsets,
4209 int part, int *degraded,
4210 char *buf)
4211 {
4212 /* Backup 'blocks' sectors at 'offset' on each device of the array,
4213 * to storage 'destfd' (offset 'destoffsets'), after first
4214 * suspending IO. Then allow resync to continue
4215 * over the suspended section.
4216 * Use part 'part' of the backup-super-block.
4217 */
4218 int odata = disks;
4219 int rv = 0;
4220 int i;
4221 unsigned long long ll;
4222 int new_degraded;
4223 //printf("offset %llu\n", offset);
4224 if (level >= 4)
4225 odata--;
4226 if (level == 6)
4227 odata--;
4228
4229 /* Check that array hasn't become degraded, else we might backup the wrong data */
4230 if (sysfs_get_ll(sra, NULL, "degraded", &ll) < 0)
4231 return -1; /* FIXME this error is ignored */
4232 new_degraded = (int)ll;
4233 if (new_degraded != *degraded) {
4234 /* check each device to ensure it is still working */
4235 struct mdinfo *sd;
4236 for (sd = sra->devs ; sd ; sd = sd->next) {
4237 if (sd->disk.state & (1<<MD_DISK_FAULTY))
4238 continue;
4239 if (sd->disk.state & (1<<MD_DISK_SYNC)) {
4240 char sbuf[100];
4241
4242 if (sysfs_get_str(sra, sd, "state",
4243 sbuf, sizeof(sbuf)) < 0 ||
4244 strstr(sbuf, "faulty") ||
4245 strstr(sbuf, "in_sync") == NULL) {
4246 /* this device is dead */
4247 sd->disk.state = (1<<MD_DISK_FAULTY);
4248 if (sd->disk.raid_disk >= 0 &&
4249 sources[sd->disk.raid_disk] >= 0) {
4250 close(sources[sd->disk.raid_disk]);
4251 sources[sd->disk.raid_disk] = -1;
4252 }
4253 }
4254 }
4255 }
4256 *degraded = new_degraded;
4257 }
4258 if (part) {
4259 bsb.arraystart2 = __cpu_to_le64(offset * odata);
4260 bsb.length2 = __cpu_to_le64(stripes * (chunk/512) * odata);
4261 } else {
4262 bsb.arraystart = __cpu_to_le64(offset * odata);
4263 bsb.length = __cpu_to_le64(stripes * (chunk/512) * odata);
4264 }
4265 if (part)
4266 bsb.magic[15] = '2';
4267 for (i = 0; i < dests; i++)
4268 if (part)
4269 lseek64(destfd[i], destoffsets[i] +
4270 __le64_to_cpu(bsb.devstart2)*512, 0);
4271 else
4272 lseek64(destfd[i], destoffsets[i], 0);
4273
4274 rv = save_stripes(sources, offsets, disks, chunk, level, layout,
4275 dests, destfd, offset * 512 * odata,
4276 stripes * chunk * odata, buf);
4277
4278 if (rv)
4279 return rv;
4280 bsb.mtime = __cpu_to_le64(time(0));
4281 for (i = 0; i < dests; i++) {
4282 bsb.devstart = __cpu_to_le64(destoffsets[i]/512);
4283
4284 bsb.sb_csum = bsb_csum((char*)&bsb,
4285 ((char*)&bsb.sb_csum)-((char*)&bsb));
4286 if (memcmp(bsb.magic, "md_backup_data-2", 16) == 0)
4287 bsb.sb_csum2 = bsb_csum((char*)&bsb,
4288 ((char*)&bsb.sb_csum2)-((char*)&bsb));
4289
4290 rv = -1;
4291 if ((unsigned long long)lseek64(destfd[i],
4292 destoffsets[i] - 4096, 0) !=
4293 destoffsets[i] - 4096)
4294 break;
4295 if (write(destfd[i], &bsb, 512) != 512)
4296 break;
4297 if (destoffsets[i] > 4096) {
4298 if ((unsigned long long)lseek64(destfd[i], destoffsets[i]+stripes*chunk*odata, 0) !=
4299 destoffsets[i]+stripes*chunk*odata)
4300 break;
4301 if (write(destfd[i], &bsb, 512) != 512)
4302 break;
4303 }
4304 fsync(destfd[i]);
4305 rv = 0;
4306 }
4307
4308 return rv;
4309 }
4310
4311 /* in 2.6.30, the value reported by sync_completed can be
4312 * less that it should be by one stripe.
4313 * This only happens when reshape hits sync_max and pauses.
4314 * So allow wait_backup to either extent sync_max further
4315 * than strictly necessary, or return before the
4316 * sync has got quite as far as we would really like.
4317 * This is what 'blocks2' is for.
4318 * The various caller give appropriate values so that
4319 * every works.
4320 */
4321 /* FIXME return value is often ignored */
4322 static int forget_backup(int dests, int *destfd,
4323 unsigned long long *destoffsets,
4324 int part)
4325 {
4326 /*
4327 * Erase backup 'part' (which is 0 or 1)
4328 */
4329 int i;
4330 int rv;
4331
4332 if (part) {
4333 bsb.arraystart2 = __cpu_to_le64(0);
4334 bsb.length2 = __cpu_to_le64(0);
4335 } else {
4336 bsb.arraystart = __cpu_to_le64(0);
4337 bsb.length = __cpu_to_le64(0);
4338 }
4339 bsb.mtime = __cpu_to_le64(time(0));
4340 rv = 0;
4341 for (i = 0; i < dests; i++) {
4342 bsb.devstart = __cpu_to_le64(destoffsets[i]/512);
4343 bsb.sb_csum = bsb_csum((char*)&bsb,
4344 ((char*)&bsb.sb_csum)-((char*)&bsb));
4345 if (memcmp(bsb.magic, "md_backup_data-2", 16) == 0)
4346 bsb.sb_csum2 = bsb_csum((char*)&bsb,
4347 ((char*)&bsb.sb_csum2)-((char*)&bsb));
4348 if ((unsigned long long)lseek64(destfd[i], destoffsets[i]-4096, 0) !=
4349 destoffsets[i]-4096)
4350 rv = -1;
4351 if (rv == 0 && write(destfd[i], &bsb, 512) != 512)
4352 rv = -1;
4353 fsync(destfd[i]);
4354 }
4355 return rv;
4356 }
4357
4358 static void fail(char *msg)
4359 {
4360 int rv;
4361 rv = (write(2, msg, strlen(msg)) != (int)strlen(msg));
4362 rv |= (write(2, "\n", 1) != 1);
4363 exit(rv ? 1 : 2);
4364 }
4365
4366 static char *abuf, *bbuf;
4367 static unsigned long long abuflen;
4368 static void validate(int afd, int bfd, unsigned long long offset)
4369 {
4370 /* check that the data in the backup against the array.
4371 * This is only used for regression testing and should not
4372 * be used while the array is active
4373 */
4374 if (afd < 0)
4375 return;
4376 lseek64(bfd, offset - 4096, 0);
4377 if (read(bfd, &bsb2, 512) != 512)
4378 fail("cannot read bsb");
4379 if (bsb2.sb_csum != bsb_csum((char*)&bsb2,
4380 ((char*)&bsb2.sb_csum)-((char*)&bsb2)))
4381 fail("first csum bad");
4382 if (memcmp(bsb2.magic, "md_backup_data", 14) != 0)
4383 fail("magic is bad");
4384 if (memcmp(bsb2.magic, "md_backup_data-2", 16) == 0 &&
4385 bsb2.sb_csum2 != bsb_csum((char*)&bsb2,
4386 ((char*)&bsb2.sb_csum2)-((char*)&bsb2)))
4387 fail("second csum bad");
4388
4389 if (__le64_to_cpu(bsb2.devstart)*512 != offset)
4390 fail("devstart is wrong");
4391
4392 if (bsb2.length) {
4393 unsigned long long len = __le64_to_cpu(bsb2.length)*512;
4394
4395 if (abuflen < len) {
4396 free(abuf);
4397 free(bbuf);
4398 abuflen = len;
4399 if (posix_memalign((void**)&abuf, 4096, abuflen) ||
4400 posix_memalign((void**)&bbuf, 4096, abuflen)) {
4401 abuflen = 0;
4402 /* just stop validating on mem-alloc failure */
4403 return;
4404 }
4405 }
4406
4407 lseek64(bfd, offset, 0);
4408 if ((unsigned long long)read(bfd, bbuf, len) != len) {
4409 //printf("len %llu\n", len);
4410 fail("read first backup failed");
4411 }
4412 lseek64(afd, __le64_to_cpu(bsb2.arraystart)*512, 0);
4413 if ((unsigned long long)read(afd, abuf, len) != len)
4414 fail("read first from array failed");
4415 if (memcmp(bbuf, abuf, len) != 0) {
4416 #if 0
4417 int i;
4418 printf("offset=%llu len=%llu\n",
4419 (unsigned long long)__le64_to_cpu(bsb2.arraystart)*512, len);
4420 for (i=0; i<len; i++)
4421 if (bbuf[i] != abuf[i]) {
4422 printf("first diff byte %d\n", i);
4423 break;
4424 }
4425 #endif
4426 fail("data1 compare failed");
4427 }
4428 }
4429 if (bsb2.length2) {
4430 unsigned long long len = __le64_to_cpu(bsb2.length2)*512;
4431
4432 if (abuflen < len) {
4433 free(abuf);
4434 free(bbuf);
4435 abuflen = len;
4436 abuf = xmalloc(abuflen);
4437 bbuf = xmalloc(abuflen);
4438 }
4439
4440 lseek64(bfd, offset+__le64_to_cpu(bsb2.devstart2)*512, 0);
4441 if ((unsigned long long)read(bfd, bbuf, len) != len)
4442 fail("read second backup failed");
4443 lseek64(afd, __le64_to_cpu(bsb2.arraystart2)*512, 0);
4444 if ((unsigned long long)read(afd, abuf, len) != len)
4445 fail("read second from array failed");
4446 if (memcmp(bbuf, abuf, len) != 0)
4447 fail("data2 compare failed");
4448 }
4449 }
4450
4451 int child_monitor(int afd, struct mdinfo *sra, struct reshape *reshape,
4452 struct supertype *st, unsigned long blocks,
4453 int *fds, unsigned long long *offsets,
4454 int dests, int *destfd, unsigned long long *destoffsets)
4455 {
4456 /* Monitor a reshape where backup is being performed using
4457 * 'native' mechanism - either to a backup file, or
4458 * to some space in a spare.
4459 */
4460 char *buf;
4461 int degraded = -1;
4462 unsigned long long speed;
4463 unsigned long long suspend_point, array_size;
4464 unsigned long long backup_point, wait_point;
4465 unsigned long long reshape_completed;
4466 int done = 0;
4467 int increasing = reshape->after.data_disks >=
4468 reshape->before.data_disks;
4469 int part = 0; /* The next part of the backup area to fill. It
4470 * may already be full, so we need to check */
4471 int level = reshape->level;
4472 int layout = reshape->before.layout;
4473 int data = reshape->before.data_disks;
4474 int disks = reshape->before.data_disks + reshape->parity;
4475 int chunk = sra->array.chunk_size;
4476 struct mdinfo *sd;
4477 unsigned long stripes;
4478 int uuid[4];
4479 int frozen = 0;
4480
4481 /* set up the backup-super-block. This requires the
4482 * uuid from the array.
4483 */
4484 /* Find a superblock */
4485 for (sd = sra->devs; sd; sd = sd->next) {
4486 char *dn;
4487 int devfd;
4488 int ok;
4489 if (sd->disk.state & (1<<MD_DISK_FAULTY))
4490 continue;
4491 dn = map_dev(sd->disk.major, sd->disk.minor, 1);
4492 devfd = dev_open(dn, O_RDONLY);
4493 if (devfd < 0)
4494 continue;
4495 ok = st->ss->load_super(st, devfd, NULL);
4496 close(devfd);
4497 if (ok == 0)
4498 break;
4499 }
4500 if (!sd) {
4501 pr_err("Cannot find a superblock\n");
4502 return 0;
4503 }
4504
4505 memset(&bsb, 0, 512);
4506 memcpy(bsb.magic, "md_backup_data-1", 16);
4507 st->ss->uuid_from_super(st, uuid);
4508 memcpy(bsb.set_uuid, uuid, 16);
4509 bsb.mtime = __cpu_to_le64(time(0));
4510 bsb.devstart2 = blocks;
4511
4512 stripes = blocks / (sra->array.chunk_size/512) /
4513 reshape->before.data_disks;
4514
4515 if (posix_memalign((void**)&buf, 4096, disks * chunk))
4516 /* Don't start the 'reshape' */
4517 return 0;
4518 if (reshape->before.data_disks == reshape->after.data_disks) {
4519 sysfs_get_ll(sra, NULL, "sync_speed_min", &speed);
4520 sysfs_set_num(sra, NULL, "sync_speed_min", 200000);
4521 }
4522
4523 if (increasing) {
4524 array_size = sra->component_size * reshape->after.data_disks;
4525 backup_point = sra->reshape_progress;
4526 suspend_point = 0;
4527 } else {
4528 array_size = sra->component_size * reshape->before.data_disks;
4529 backup_point = reshape->backup_blocks;
4530 suspend_point = array_size;
4531 }
4532
4533 while (!done) {
4534 int rv;
4535
4536 /* Want to return as soon the oldest backup slot can
4537 * be released as that allows us to start backing up
4538 * some more, providing suspend_point has been
4539 * advanced, which it should have.
4540 */
4541 if (increasing) {
4542 wait_point = array_size;
4543 if (part == 0 && __le64_to_cpu(bsb.length) > 0)
4544 wait_point = (__le64_to_cpu(bsb.arraystart) +
4545 __le64_to_cpu(bsb.length));
4546 if (part == 1 && __le64_to_cpu(bsb.length2) > 0)
4547 wait_point = (__le64_to_cpu(bsb.arraystart2) +
4548 __le64_to_cpu(bsb.length2));
4549 } else {
4550 wait_point = 0;
4551 if (part == 0 && __le64_to_cpu(bsb.length) > 0)
4552 wait_point = __le64_to_cpu(bsb.arraystart);
4553 if (part == 1 && __le64_to_cpu(bsb.length2) > 0)
4554 wait_point = __le64_to_cpu(bsb.arraystart2);
4555 }
4556
4557 reshape_completed = sra->reshape_progress;
4558 rv = progress_reshape(sra, reshape,
4559 backup_point, wait_point,
4560 &suspend_point, &reshape_completed,
4561 &frozen);
4562 /* external metadata would need to ping_monitor here */
4563 sra->reshape_progress = reshape_completed;
4564
4565 /* Clear any backup region that is before 'here' */
4566 if (increasing) {
4567 if (__le64_to_cpu(bsb.length) > 0 &&
4568 reshape_completed >= (__le64_to_cpu(bsb.arraystart) +
4569 __le64_to_cpu(bsb.length)))
4570 forget_backup(dests, destfd,
4571 destoffsets, 0);
4572 if (__le64_to_cpu(bsb.length2) > 0 &&
4573 reshape_completed >= (__le64_to_cpu(bsb.arraystart2) +
4574 __le64_to_cpu(bsb.length2)))
4575 forget_backup(dests, destfd,
4576 destoffsets, 1);
4577 } else {
4578 if (__le64_to_cpu(bsb.length) > 0 &&
4579 reshape_completed <= (__le64_to_cpu(bsb.arraystart)))
4580 forget_backup(dests, destfd,
4581 destoffsets, 0);
4582 if (__le64_to_cpu(bsb.length2) > 0 &&
4583 reshape_completed <= (__le64_to_cpu(bsb.arraystart2)))
4584 forget_backup(dests, destfd,
4585 destoffsets, 1);
4586 }
4587 if (sigterm)
4588 rv = -2;
4589 if (rv < 0) {
4590 if (rv == -1)
4591 done = 1;
4592 break;
4593 }
4594 if (rv == 0 && increasing && !st->ss->external) {
4595 /* No longer need to monitor this reshape */
4596 sysfs_set_str(sra, NULL, "sync_max", "max");
4597 done = 1;
4598 break;
4599 }
4600
4601 while (rv) {
4602 unsigned long long offset;
4603 unsigned long actual_stripes;
4604 /* Need to backup some data.
4605 * If 'part' is not used and the desired
4606 * backup size is suspended, do a backup,
4607 * then consider the next part.
4608 */
4609 /* Check that 'part' is unused */
4610 if (part == 0 && __le64_to_cpu(bsb.length) != 0)
4611 break;
4612 if (part == 1 && __le64_to_cpu(bsb.length2) != 0)
4613 break;
4614
4615 offset = backup_point / data;
4616 actual_stripes = stripes;
4617 if (increasing) {
4618 if (offset + actual_stripes * (chunk/512) >
4619 sra->component_size)
4620 actual_stripes = ((sra->component_size - offset)
4621 / (chunk/512));
4622 if (offset + actual_stripes * (chunk/512) >
4623 suspend_point/data)
4624 break;
4625 } else {
4626 if (offset < actual_stripes * (chunk/512))
4627 actual_stripes = offset / (chunk/512);
4628 offset -= actual_stripes * (chunk/512);
4629 if (offset < suspend_point/data)
4630 break;
4631 }
4632 if (actual_stripes == 0)
4633 break;
4634 grow_backup(sra, offset, actual_stripes, fds, offsets,
4635 disks, chunk, level, layout, dests, destfd,
4636 destoffsets, part, &degraded, buf);
4637 validate(afd, destfd[0], destoffsets[0]);
4638 /* record where 'part' is up to */
4639 part = !part;
4640 if (increasing)
4641 backup_point += actual_stripes * (chunk/512) * data;
4642 else
4643 backup_point -= actual_stripes * (chunk/512) * data;
4644 }
4645 }
4646
4647 /* FIXME maybe call progress_reshape one more time instead */
4648 /* remove any remaining suspension */
4649 sysfs_set_num(sra, NULL, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL);
4650 sysfs_set_num(sra, NULL, "suspend_hi", 0);
4651 sysfs_set_num(sra, NULL, "suspend_lo", 0);
4652 sysfs_set_num(sra, NULL, "sync_min", 0);
4653
4654 if (reshape->before.data_disks == reshape->after.data_disks)
4655 sysfs_set_num(sra, NULL, "sync_speed_min", speed);
4656 free(buf);
4657 return done;
4658 }
4659
4660 /*
4661 * If any spare contains md_back_data-1 which is recent wrt mtime,
4662 * write that data into the array and update the super blocks with
4663 * the new reshape_progress
4664 */
4665 int Grow_restart(struct supertype *st, struct mdinfo *info, int *fdlist,
4666 int cnt, char *backup_file, int verbose)
4667 {
4668 int i, j;
4669 int old_disks;
4670 unsigned long long *offsets;
4671 unsigned long long nstripe, ostripe;
4672 int ndata, odata;
4673
4674 odata = info->array.raid_disks - info->delta_disks - 1;
4675 if (info->array.level == 6)
4676 odata--; /* number of data disks */
4677 ndata = info->array.raid_disks - 1;
4678 if (info->new_level == 6)
4679 ndata--;
4680
4681 old_disks = info->array.raid_disks - info->delta_disks;
4682
4683 if (info->delta_disks <= 0)
4684 /* Didn't grow, so the backup file must have
4685 * been used
4686 */
4687 old_disks = cnt;
4688 for (i=old_disks-(backup_file?1:0); i<cnt; i++) {
4689 struct mdinfo dinfo;
4690 int fd;
4691 int bsbsize;
4692 char *devname, namebuf[20];
4693 unsigned long long lo, hi;
4694
4695 /* This was a spare and may have some saved data on it.
4696 * Load the superblock, find and load the
4697 * backup_super_block.
4698 * If either fail, go on to next device.
4699 * If the backup contains no new info, just return
4700 * else restore data and update all superblocks
4701 */
4702 if (i == old_disks-1) {
4703 fd = open(backup_file, O_RDONLY);
4704 if (fd<0) {
4705 pr_err("backup file %s inaccessible: %s\n",
4706 backup_file, strerror(errno));
4707 continue;
4708 }
4709 devname = backup_file;
4710 } else {
4711 fd = fdlist[i];
4712 if (fd < 0)
4713 continue;
4714 if (st->ss->load_super(st, fd, NULL))
4715 continue;
4716
4717 st->ss->getinfo_super(st, &dinfo, NULL);
4718 st->ss->free_super(st);
4719
4720 if (lseek64(fd,
4721 (dinfo.data_offset + dinfo.component_size - 8) <<9,
4722 0) < 0) {
4723 pr_err("Cannot seek on device %d\n", i);
4724 continue; /* Cannot seek */
4725 }
4726 sprintf(namebuf, "device-%d", i);
4727 devname = namebuf;
4728 }
4729 if (read(fd, &bsb, sizeof(bsb)) != sizeof(bsb)) {
4730 if (verbose)
4731 pr_err("Cannot read from %s\n", devname);
4732 continue; /* Cannot read */
4733 }
4734 if (memcmp(bsb.magic, "md_backup_data-1", 16) != 0 &&
4735 memcmp(bsb.magic, "md_backup_data-2", 16) != 0) {
4736 if (verbose)
4737 pr_err("No backup metadata on %s\n", devname);
4738 continue;
4739 }
4740 if (bsb.sb_csum != bsb_csum((char*)&bsb, ((char*)&bsb.sb_csum)-((char*)&bsb))) {
4741 if (verbose)
4742 pr_err("Bad backup-metadata checksum on %s\n",
4743 devname);
4744 continue; /* bad checksum */
4745 }
4746 if (memcmp(bsb.magic, "md_backup_data-2", 16) == 0 &&
4747 bsb.sb_csum2 != bsb_csum((char*)&bsb, ((char*)&bsb.sb_csum2)-((char*)&bsb))) {
4748 if (verbose)
4749 pr_err("Bad backup-metadata checksum2 on %s\n",
4750 devname);
4751 continue; /* Bad second checksum */
4752 }
4753 if (memcmp(bsb.set_uuid,info->uuid, 16) != 0) {
4754 if (verbose)
4755 pr_err("Wrong uuid on backup-metadata on %s\n",
4756 devname);
4757 continue; /* Wrong uuid */
4758 }
4759
4760 /*
4761 * array utime and backup-mtime should be updated at
4762 * much the same time, but it seems that sometimes
4763 * they aren't... So allow considerable flexability in
4764 * matching, and allow this test to be overridden by
4765 * an environment variable.
4766 */
4767 if(time_after(info->array.utime, (unsigned int)__le64_to_cpu(bsb.mtime) + 2*60*60) ||
4768 time_before(info->array.utime, (unsigned int)__le64_to_cpu(bsb.mtime) - 10*60)) {
4769 if (check_env("MDADM_GROW_ALLOW_OLD")) {
4770 pr_err("accepting backup with timestamp %lu for array with timestamp %lu\n",
4771 (unsigned long)__le64_to_cpu(bsb.mtime),
4772 (unsigned long)info->array.utime);
4773 } else {
4774 pr_err("too-old timestamp on backup-metadata on %s\n", devname);
4775 pr_err("If you think it is should be safe, try 'export MDADM_GROW_ALLOW_OLD=1'\n");
4776 continue; /* time stamp is too bad */
4777 }
4778 }
4779
4780 if (bsb.magic[15] == '1') {
4781 if (bsb.length == 0)
4782 continue;
4783 if (info->delta_disks >= 0) {
4784 /* reshape_progress is increasing */
4785 if (__le64_to_cpu(bsb.arraystart)
4786 + __le64_to_cpu(bsb.length)
4787 < info->reshape_progress) {
4788 nonew:
4789 if (verbose)
4790 pr_err("backup-metadata found on %s but is not needed\n", devname);
4791 continue; /* No new data here */
4792 }
4793 } else {
4794 /* reshape_progress is decreasing */
4795 if (__le64_to_cpu(bsb.arraystart) >=
4796 info->reshape_progress)
4797 goto nonew; /* No new data here */
4798 }
4799 } else {
4800 if (bsb.length == 0 && bsb.length2 == 0)
4801 continue;
4802 if (info->delta_disks >= 0) {
4803 /* reshape_progress is increasing */
4804 if ((__le64_to_cpu(bsb.arraystart)
4805 + __le64_to_cpu(bsb.length)
4806 < info->reshape_progress) &&
4807 (__le64_to_cpu(bsb.arraystart2)
4808 + __le64_to_cpu(bsb.length2)
4809 < info->reshape_progress))
4810 goto nonew; /* No new data here */
4811 } else {
4812 /* reshape_progress is decreasing */
4813 if (__le64_to_cpu(bsb.arraystart) >=
4814 info->reshape_progress &&
4815 __le64_to_cpu(bsb.arraystart2) >=
4816 info->reshape_progress)
4817 goto nonew; /* No new data here */
4818 }
4819 }
4820 if (lseek64(fd, __le64_to_cpu(bsb.devstart)*512, 0)< 0) {
4821 second_fail:
4822 if (verbose)
4823 pr_err("Failed to verify secondary backup-metadata block on %s\n",
4824 devname);
4825 continue; /* Cannot seek */
4826 }
4827 /* There should be a duplicate backup superblock 4k before here */
4828 if (lseek64(fd, -4096, 1) < 0 ||
4829 read(fd, &bsb2, sizeof(bsb2)) != sizeof(bsb2))
4830 goto second_fail; /* Cannot find leading superblock */
4831 if (bsb.magic[15] == '1')
4832 bsbsize = offsetof(struct mdp_backup_super, pad1);
4833 else
4834 bsbsize = offsetof(struct mdp_backup_super, pad);
4835 if (memcmp(&bsb2, &bsb, bsbsize) != 0)
4836 goto second_fail; /* Cannot find leading superblock */
4837
4838 /* Now need the data offsets for all devices. */
4839 offsets = xmalloc(sizeof(*offsets)*info->array.raid_disks);
4840 for(j=0; j<info->array.raid_disks; j++) {
4841 if (fdlist[j] < 0)
4842 continue;
4843 if (st->ss->load_super(st, fdlist[j], NULL))
4844 /* FIXME should be this be an error */
4845 continue;
4846 st->ss->getinfo_super(st, &dinfo, NULL);
4847 st->ss->free_super(st);
4848 offsets[j] = dinfo.data_offset * 512;
4849 }
4850 printf("%s: restoring critical section\n", Name);
4851
4852 if (restore_stripes(fdlist, offsets, info->array.raid_disks,
4853 info->new_chunk, info->new_level,
4854 info->new_layout, fd,
4855 __le64_to_cpu(bsb.devstart)*512,
4856 __le64_to_cpu(bsb.arraystart)*512,
4857 __le64_to_cpu(bsb.length)*512, NULL)) {
4858 /* didn't succeed, so giveup */
4859 if (verbose)
4860 pr_err("Error restoring backup from %s\n",
4861 devname);
4862 free(offsets);
4863 return 1;
4864 }
4865
4866 if (bsb.magic[15] == '2' &&
4867 restore_stripes(fdlist, offsets, info->array.raid_disks,
4868 info->new_chunk, info->new_level,
4869 info->new_layout, fd,
4870 __le64_to_cpu(bsb.devstart)*512 +
4871 __le64_to_cpu(bsb.devstart2)*512,
4872 __le64_to_cpu(bsb.arraystart2)*512,
4873 __le64_to_cpu(bsb.length2)*512, NULL)) {
4874 /* didn't succeed, so giveup */
4875 if (verbose)
4876 pr_err("Error restoring second backup from %s\n",
4877 devname);
4878 free(offsets);
4879 return 1;
4880 }
4881
4882 free(offsets);
4883
4884 /* Ok, so the data is restored. Let's update those superblocks. */
4885
4886 lo = hi = 0;
4887 if (bsb.length) {
4888 lo = __le64_to_cpu(bsb.arraystart);
4889 hi = lo + __le64_to_cpu(bsb.length);
4890 }
4891 if (bsb.magic[15] == '2' && bsb.length2) {
4892 unsigned long long lo1, hi1;
4893 lo1 = __le64_to_cpu(bsb.arraystart2);
4894 hi1 = lo1 + __le64_to_cpu(bsb.length2);
4895 if (lo == hi) {
4896 lo = lo1;
4897 hi = hi1;
4898 } else if (lo < lo1)
4899 hi = hi1;
4900 else
4901 lo = lo1;
4902 }
4903 if (lo < hi && (info->reshape_progress < lo ||
4904 info->reshape_progress > hi))
4905 /* backup does not affect reshape_progress*/ ;
4906 else if (info->delta_disks >= 0) {
4907 info->reshape_progress = __le64_to_cpu(bsb.arraystart) +
4908 __le64_to_cpu(bsb.length);
4909 if (bsb.magic[15] == '2') {
4910 unsigned long long p2;
4911
4912 p2 = __le64_to_cpu(bsb.arraystart2) +
4913 __le64_to_cpu(bsb.length2);
4914 if (p2 > info->reshape_progress)
4915 info->reshape_progress = p2;
4916 }
4917 } else {
4918 info->reshape_progress = __le64_to_cpu(bsb.arraystart);
4919 if (bsb.magic[15] == '2') {
4920 unsigned long long p2;
4921
4922 p2 = __le64_to_cpu(bsb.arraystart2);
4923 if (p2 < info->reshape_progress)
4924 info->reshape_progress = p2;
4925 }
4926 }
4927 for (j=0; j<info->array.raid_disks; j++) {
4928 if (fdlist[j] < 0)
4929 continue;
4930 if (st->ss->load_super(st, fdlist[j], NULL))
4931 continue;
4932 st->ss->getinfo_super(st, &dinfo, NULL);
4933 dinfo.reshape_progress = info->reshape_progress;
4934 st->ss->update_super(st, &dinfo, "_reshape_progress",
4935 NULL,0, 0, NULL);
4936 st->ss->store_super(st, fdlist[j]);
4937 st->ss->free_super(st);
4938 }
4939 return 0;
4940 }
4941 /* Didn't find any backup data, try to see if any
4942 * was needed.
4943 */
4944 if (info->delta_disks < 0) {
4945 /* When shrinking, the critical section is at the end.
4946 * So see if we are before the critical section.
4947 */
4948 unsigned long long first_block;
4949 nstripe = ostripe = 0;
4950 first_block = 0;
4951 while (ostripe >= nstripe) {
4952 ostripe += info->array.chunk_size / 512;
4953 first_block = ostripe * odata;
4954 nstripe = first_block / ndata / (info->new_chunk/512) *
4955 (info->new_chunk/512);
4956 }
4957
4958 if (info->reshape_progress >= first_block)
4959 return 0;
4960 }
4961 if (info->delta_disks > 0) {
4962 /* See if we are beyond the critical section. */
4963 unsigned long long last_block;
4964 nstripe = ostripe = 0;
4965 last_block = 0;
4966 while (nstripe >= ostripe) {
4967 nstripe += info->new_chunk / 512;
4968 last_block = nstripe * ndata;
4969 ostripe = last_block / odata / (info->array.chunk_size/512) *
4970 (info->array.chunk_size/512);
4971 }
4972
4973 if (info->reshape_progress >= last_block)
4974 return 0;
4975 }
4976 /* needed to recover critical section! */
4977 if (verbose)
4978 pr_err("Failed to find backup of critical section\n");
4979 return 1;
4980 }
4981
4982 int Grow_continue_command(char *devname, int fd,
4983 char *backup_file, int verbose)
4984 {
4985 int ret_val = 0;
4986 struct supertype *st = NULL;
4987 struct mdinfo *content = NULL;
4988 struct mdinfo array;
4989 char *subarray = NULL;
4990 struct mdinfo *cc = NULL;
4991 struct mdstat_ent *mdstat = NULL;
4992 int cfd = -1;
4993 int fd2;
4994
4995 dprintf("Grow continue from command line called for %s\n", devname);
4996
4997 st = super_by_fd(fd, &subarray);
4998 if (!st || !st->ss) {
4999 pr_err("Unable to determine metadata format for %s\n", devname);
5000 return 1;
5001 }
5002 dprintf("Grow continue is run for ");
5003 if (st->ss->external == 0) {
5004 int d;
5005 int cnt = 5;
5006 dprintf_cont("native array (%s)\n", devname);
5007 if (md_get_array_info(fd, &array.array) < 0) {
5008 pr_err("%s is not an active md array - aborting\n",
5009 devname);
5010 ret_val = 1;
5011 goto Grow_continue_command_exit;
5012 }
5013 content = &array;
5014 sysfs_init(content, fd, NULL);
5015 /* Need to load a superblock.
5016 * FIXME we should really get what we need from
5017 * sysfs
5018 */
5019 do {
5020 for (d = 0; d < MAX_DISKS; d++) {
5021 mdu_disk_info_t disk;
5022 char *dv;
5023 int err;
5024 disk.number = d;
5025 if (md_get_disk_info(fd, &disk) < 0)
5026 continue;
5027 if (disk.major == 0 && disk.minor == 0)
5028 continue;
5029 if ((disk.state & (1 << MD_DISK_ACTIVE)) == 0)
5030 continue;
5031 dv = map_dev(disk.major, disk.minor, 1);
5032 if (!dv)
5033 continue;
5034 fd2 = dev_open(dv, O_RDONLY);
5035 if (fd2 < 0)
5036 continue;
5037 err = st->ss->load_super(st, fd2, NULL);
5038 close(fd2);
5039 if (err)
5040 continue;
5041 break;
5042 }
5043 if (d == MAX_DISKS) {
5044 pr_err("Unable to load metadata for %s\n",
5045 devname);
5046 ret_val = 1;
5047 goto Grow_continue_command_exit;
5048 }
5049 st->ss->getinfo_super(st, content, NULL);
5050 if (!content->reshape_active)
5051 sleep(3);
5052 else
5053 break;
5054 } while (cnt-- > 0);
5055 } else {
5056 char *container;
5057
5058 if (subarray) {
5059 dprintf_cont("subarray (%s)\n", subarray);
5060 container = st->container_devnm;
5061 cfd = open_dev_excl(st->container_devnm);
5062 } else {
5063 container = st->devnm;
5064 close(fd);
5065 cfd = open_dev_excl(st->devnm);
5066 dprintf_cont("container (%s)\n", container);
5067 fd = cfd;
5068 }
5069 if (cfd < 0) {
5070 pr_err("Unable to open container for %s\n", devname);
5071 ret_val = 1;
5072 goto Grow_continue_command_exit;
5073 }
5074
5075 /* find in container array under reshape
5076 */
5077 ret_val = st->ss->load_container(st, cfd, NULL);
5078 if (ret_val) {
5079 pr_err("Cannot read superblock for %s\n", devname);
5080 ret_val = 1;
5081 goto Grow_continue_command_exit;
5082 }
5083
5084 cc = st->ss->container_content(st, subarray);
5085 for (content = cc; content ; content = content->next) {
5086 char *array_name;
5087 int allow_reshape = 1;
5088
5089 if (content->reshape_active == 0)
5090 continue;
5091 /* The decision about array or container wide
5092 * reshape is taken in Grow_continue based
5093 * content->reshape_active state, therefore we
5094 * need to check_reshape based on
5095 * reshape_active and subarray name
5096 */
5097 if (content->array.state & (1<<MD_SB_BLOCK_VOLUME))
5098 allow_reshape = 0;
5099 if (content->reshape_active == CONTAINER_RESHAPE &&
5100 (content->array.state
5101 & (1<<MD_SB_BLOCK_CONTAINER_RESHAPE)))
5102 allow_reshape = 0;
5103
5104 if (!allow_reshape) {
5105 pr_err("cannot continue reshape of an array in container with unsupported metadata: %s(%s)\n",
5106 devname, container);
5107 ret_val = 1;
5108 goto Grow_continue_command_exit;
5109 }
5110
5111 array_name = strchr(content->text_version+1, '/')+1;
5112 mdstat = mdstat_by_subdev(array_name, container);
5113 if (!mdstat)
5114 continue;
5115 if (mdstat->active == 0) {
5116 pr_err("Skipping inactive array %s.\n",
5117 mdstat->devnm);
5118 free_mdstat(mdstat);
5119 mdstat = NULL;
5120 continue;
5121 }
5122 break;
5123 }
5124 if (!content) {
5125 pr_err("Unable to determine reshaped array for %s\n", devname);
5126 ret_val = 1;
5127 goto Grow_continue_command_exit;
5128 }
5129 fd2 = open_dev(mdstat->devnm);
5130 if (fd2 < 0) {
5131 pr_err("cannot open (%s)\n", mdstat->devnm);
5132 ret_val = 1;
5133 goto Grow_continue_command_exit;
5134 }
5135
5136 if (sysfs_init(content, fd2, mdstat->devnm)) {
5137 pr_err("Unable to initialize sysfs for %s, Grow cannot continue.\n",
5138 mdstat->devnm);
5139 ret_val = 1;
5140 close(fd2);
5141 goto Grow_continue_command_exit;
5142 }
5143
5144 close(fd2);
5145
5146 /* start mdmon in case it is not running
5147 */
5148 if (!mdmon_running(container))
5149 start_mdmon(container);
5150 ping_monitor(container);
5151
5152 if (mdmon_running(container))
5153 st->update_tail = &st->updates;
5154 else {
5155 pr_err("No mdmon found. Grow cannot continue.\n");
5156 ret_val = 1;
5157 goto Grow_continue_command_exit;
5158 }
5159 }
5160
5161 /* verify that array under reshape is started from
5162 * correct position
5163 */
5164 if (verify_reshape_position(content, content->array.level) < 0) {
5165 ret_val = 1;
5166 goto Grow_continue_command_exit;
5167 }
5168
5169 /* continue reshape
5170 */
5171 ret_val = Grow_continue(fd, st, content, backup_file, 1, 0);
5172
5173 Grow_continue_command_exit:
5174 if (cfd > -1)
5175 close(cfd);
5176 st->ss->free_super(st);
5177 free_mdstat(mdstat);
5178 sysfs_free(cc);
5179 free(subarray);
5180
5181 return ret_val;
5182 }
5183
5184 int Grow_continue(int mdfd, struct supertype *st, struct mdinfo *info,
5185 char *backup_file, int forked, int freeze_reshape)
5186 {
5187 int ret_val = 2;
5188
5189 if (!info->reshape_active)
5190 return ret_val;
5191
5192 if (st->ss->external) {
5193 int cfd = open_dev(st->container_devnm);
5194
5195 if (cfd < 0)
5196 return 1;
5197
5198 st->ss->load_container(st, cfd, st->container_devnm);
5199 close(cfd);
5200 ret_val = reshape_container(st->container_devnm, NULL, mdfd,
5201 st, info, 0, backup_file, 0,
5202 forked, 1 | info->reshape_active,
5203 freeze_reshape);
5204 } else
5205 ret_val = reshape_array(NULL, mdfd, "array", st, info, 1,
5206 NULL, INVALID_SECTORS, backup_file,
5207 0, forked, 1 | info->reshape_active,
5208 freeze_reshape);
5209
5210 return ret_val;
5211 }
5212
5213 char *make_backup(char *name)
5214 {
5215 char *base = "backup_file-";
5216 int len;
5217 char *fname;
5218
5219 len = strlen(MAP_DIR) + 1 + strlen(base) + strlen(name)+1;
5220 fname = xmalloc(len);
5221 sprintf(fname, "%s/%s%s", MAP_DIR, base, name);
5222 return fname;
5223 }
5224
5225 char *locate_backup(char *name)
5226 {
5227 char *fl = make_backup(name);
5228 struct stat stb;
5229
5230 if (stat(fl, &stb) == 0 && S_ISREG(stb.st_mode))
5231 return fl;
5232
5233 free(fl);
5234 return NULL;
5235 }
Unnamed repository; edit this file 'description' to name the repository.