]> git.ipfire.org Git - thirdparty/mdadm.git/blob - Grow.c
projects / thirdparty / mdadm.git / blob
? search:


868bdc3a3be4fe8134a342d0dbcd43b52136015c
[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_for(1, 0, true);
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 struct context *c, struct shape *s)
1779 {
1780 /* Make some changes in the shape of an array.
1781 * The kernel must support the change.
1782 *
1783 * There are three different changes. Each can trigger
1784 * a resync or recovery so we freeze that until we have
1785 * requested everything (if kernel supports freezing - 2.6.30).
1786 * The steps are:
1787 * - change size (i.e. component_size)
1788 * - change level
1789 * - change layout/chunksize/ndisks
1790 *
1791 * The last can require a reshape. It is different on different
1792 * levels so we need to check the level before actioning it.
1793 * Some times the level change needs to be requested after the
1794 * reshape (e.g. raid6->raid5, raid5->raid0)
1795 *
1796 */
1797 struct mdu_array_info_s array;
1798 int rv = 0;
1799 struct supertype *st;
1800 char *subarray = NULL;
1801
1802 int frozen;
1803 int changed = 0;
1804 char *container = NULL;
1805 int cfd = -1;
1806
1807 struct mddev_dev *dv;
1808 int added_disks;
1809
1810 struct mdinfo info;
1811 struct mdinfo *sra;
1812
1813 if (md_get_array_info(fd, &array) < 0) {
1814 pr_err("%s is not an active md array - aborting\n",
1815 devname);
1816 return 1;
1817 }
1818 if (s->level != UnSet && s->chunk) {
1819 pr_err("Cannot change array level in the same operation as changing chunk size.\n");
1820 return 1;
1821 }
1822
1823 if (s->data_offset != INVALID_SECTORS && array.level != 10 &&
1824 (array.level < 4 || array.level > 6)) {
1825 pr_err("--grow --data-offset not yet supported\n");
1826 return 1;
1827 }
1828
1829 if (s->size > 0 &&
1830 (s->chunk || s->level!= UnSet || s->layout_str || s->raiddisks)) {
1831 pr_err("cannot change component size at the same time as other changes.\n"
1832 " Change size first, then check data is intact before making other changes.\n");
1833 return 1;
1834 }
1835
1836 if (s->raiddisks && s->raiddisks < array.raid_disks &&
1837 array.level > 1 && get_linux_version() < 2006032 &&
1838 !check_env("MDADM_FORCE_FEWER")) {
1839 pr_err("reducing the number of devices is not safe before Linux 2.6.32\n"
1840 " Please use a newer kernel\n");
1841 return 1;
1842 }
1843
1844 if (array.level > 1 && s->size > 1 &&
1845 (unsigned long long) (array.chunk_size / 1024) > s->size) {
1846 pr_err("component size must be larger than chunk size.\n");
1847 return 1;
1848 }
1849
1850 st = super_by_fd(fd, &subarray);
1851 if (!st) {
1852 pr_err("Unable to determine metadata format for %s\n", devname);
1853 return 1;
1854 }
1855 if (s->raiddisks > st->max_devs) {
1856 pr_err("Cannot increase raid-disks on this array beyond %d\n", st->max_devs);
1857 return 1;
1858 }
1859 if (s->level == 0 && (array.state & (1 << MD_SB_BITMAP_PRESENT)) &&
1860 !(array.state & (1 << MD_SB_CLUSTERED)) && !st->ss->external) {
1861 array.state &= ~(1 << MD_SB_BITMAP_PRESENT);
1862 if (md_set_array_info(fd, &array) != 0) {
1863 pr_err("failed to remove internal bitmap.\n");
1864 return 1;
1865 }
1866 }
1867
1868 /* in the external case we need to check that the requested reshape is
1869 * supported, and perform an initial check that the container holds the
1870 * pre-requisite spare devices (mdmon owns final validation)
1871 */
1872 if (st->ss->external) {
1873 int retval;
1874
1875 if (subarray) {
1876 container = st->container_devnm;
1877 cfd = open_dev_excl(st->container_devnm);
1878 } else {
1879 container = st->devnm;
1880 close(fd);
1881 cfd = open_dev_excl(st->devnm);
1882 fd = cfd;
1883 }
1884 if (cfd < 0) {
1885 pr_err("Unable to open container for %s\n", devname);
1886 free(subarray);
1887 return 1;
1888 }
1889
1890 retval = st->ss->load_container(st, cfd, NULL);
1891
1892 if (retval) {
1893 pr_err("Cannot read superblock for %s\n", devname);
1894 close(cfd);
1895 free(subarray);
1896 return 1;
1897 }
1898
1899 if (s->raiddisks && subarray) {
1900 pr_err("--raid-devices operation can be performed on a container only\n");
1901 close(cfd);
1902 free(subarray);
1903 return 1;
1904 }
1905
1906 /* check if operation is supported for metadata handler */
1907 if (st->ss->container_content) {
1908 struct mdinfo *cc = NULL;
1909 struct mdinfo *content = NULL;
1910
1911 cc = st->ss->container_content(st, subarray);
1912 for (content = cc; content ; content = content->next) {
1913 int allow_reshape = 1;
1914
1915 /* check if reshape is allowed based on metadata
1916 * indications stored in content.array.status
1917 */
1918 if (content->array.state &
1919 (1 << MD_SB_BLOCK_VOLUME))
1920 allow_reshape = 0;
1921 if (content->array.state &
1922 (1 << MD_SB_BLOCK_CONTAINER_RESHAPE))
1923 allow_reshape = 0;
1924 if (!allow_reshape) {
1925 pr_err("cannot reshape arrays in container with unsupported metadata: %s(%s)\n",
1926 devname, container);
1927 sysfs_free(cc);
1928 free(subarray);
1929 return 1;
1930 }
1931 if (content->consistency_policy ==
1932 CONSISTENCY_POLICY_PPL) {
1933 pr_err("Operation not supported when ppl consistency policy is enabled\n");
1934 sysfs_free(cc);
1935 free(subarray);
1936 return 1;
1937 }
1938 if (content->consistency_policy ==
1939 CONSISTENCY_POLICY_BITMAP) {
1940 pr_err("Operation not supported when write-intent bitmap is enabled\n");
1941 sysfs_free(cc);
1942 free(subarray);
1943 return 1;
1944 }
1945 }
1946 sysfs_free(cc);
1947 }
1948 if (mdmon_running(container))
1949 st->update_tail = &st->updates;
1950 }
1951
1952 added_disks = 0;
1953 for (dv = devlist; dv; dv = dv->next)
1954 added_disks++;
1955 if (s->raiddisks > array.raid_disks &&
1956 array.spare_disks + added_disks <
1957 (s->raiddisks - array.raid_disks) &&
1958 !c->force) {
1959 pr_err("Need %d spare%s to avoid degraded array, and only have %d.\n"
1960 " Use --force to over-ride this check.\n",
1961 s->raiddisks - array.raid_disks,
1962 s->raiddisks - array.raid_disks == 1 ? "" : "s",
1963 array.spare_disks + added_disks);
1964 return 1;
1965 }
1966
1967 sra = sysfs_read(fd, NULL, GET_LEVEL | GET_DISKS | GET_DEVS |
1968 GET_STATE | GET_VERSION);
1969 if (sra) {
1970 if (st->ss->external && subarray == NULL) {
1971 array.level = LEVEL_CONTAINER;
1972 sra->array.level = LEVEL_CONTAINER;
1973 }
1974 } else {
1975 pr_err("failed to read sysfs parameters for %s\n",
1976 devname);
1977 return 1;
1978 }
1979 frozen = freeze(st);
1980 if (frozen < -1) {
1981 /* freeze() already spewed the reason */
1982 sysfs_free(sra);
1983 return 1;
1984 } else if (frozen < 0) {
1985 pr_err("%s is performing resync/recovery and cannot be reshaped\n", devname);
1986 sysfs_free(sra);
1987 return 1;
1988 }
1989
1990 /* ========= set size =============== */
1991 if (s->size > 0 &&
1992 (s->size == MAX_SIZE || s->size != (unsigned)array.size)) {
1993 unsigned long long orig_size = get_component_size(fd)/2;
1994 unsigned long long min_csize;
1995 struct mdinfo *mdi;
1996 int raid0_takeover = 0;
1997
1998 if (orig_size == 0)
1999 orig_size = (unsigned) array.size;
2000
2001 if (orig_size == 0) {
2002 pr_err("Cannot set device size in this type of array.\n");
2003 rv = 1;
2004 goto release;
2005 }
2006
2007 if (array.level == 0) {
2008 pr_err("Component size change is not supported for RAID0\n");
2009 rv = 1;
2010 goto release;
2011 }
2012
2013 if (reshape_super(st, s->size, UnSet, UnSet, 0, 0, UnSet, NULL,
2014 devname, APPLY_METADATA_CHANGES,
2015 c->verbose > 0)) {
2016 rv = 1;
2017 goto release;
2018 }
2019 sync_metadata(st);
2020 if (st->ss->external) {
2021 /* metadata can have size limitation
2022 * update size value according to metadata information
2023 */
2024 struct mdinfo *sizeinfo =
2025 st->ss->container_content(st, subarray);
2026 if (sizeinfo) {
2027 unsigned long long new_size =
2028 sizeinfo->custom_array_size/2;
2029 int data_disks = get_data_disks(
2030 sizeinfo->array.level,
2031 sizeinfo->array.layout,
2032 sizeinfo->array.raid_disks);
2033 new_size /= data_disks;
2034 dprintf("Metadata size correction from %llu to %llu (%llu)\n",
2035 orig_size, new_size,
2036 new_size * data_disks);
2037 s->size = new_size;
2038 sysfs_free(sizeinfo);
2039 }
2040 }
2041
2042 /* Update the size of each member device in case
2043 * they have been resized. This will never reduce
2044 * below the current used-size. The "size" attribute
2045 * understands '0' to mean 'max'.
2046 */
2047 min_csize = 0;
2048 for (mdi = sra->devs; mdi; mdi = mdi->next) {
2049 sysfs_set_num(sra, mdi, "size",
2050 s->size == MAX_SIZE ? 0 : s->size);
2051 if (array.not_persistent == 0 &&
2052 array.major_version == 0 &&
2053 get_linux_version() < 3001000) {
2054 /* Dangerous to allow size to exceed 2TB */
2055 unsigned long long csize;
2056 if (sysfs_get_ll(sra, mdi, "size",
2057 &csize) == 0) {
2058 if (csize >= 2ULL*1024*1024*1024)
2059 csize = 2ULL*1024*1024*1024;
2060 if ((min_csize == 0 ||
2061 (min_csize > csize)))
2062 min_csize = csize;
2063 }
2064 }
2065 }
2066 if (min_csize && s->size > min_csize) {
2067 pr_err("Cannot safely make this array use more than 2TB per device on this kernel.\n");
2068 rv = 1;
2069 goto size_change_error;
2070 }
2071 if (min_csize && s->size == MAX_SIZE) {
2072 /* Don't let the kernel choose a size - it will get
2073 * it wrong
2074 */
2075 pr_err("Limited v0.90 array to 2TB per device\n");
2076 s->size = min_csize;
2077 }
2078 if (st->ss->external) {
2079 if (sra->array.level == 0) {
2080 rv = sysfs_set_str(sra, NULL, "level", "raid5");
2081 if (!rv) {
2082 raid0_takeover = 1;
2083 /* get array parameters after takeover
2084 * to change one parameter at time only
2085 */
2086 rv = md_get_array_info(fd, &array);
2087 }
2088 }
2089 /* make sure mdmon is
2090 * aware of the new level */
2091 if (!mdmon_running(st->container_devnm))
2092 start_mdmon(st->container_devnm);
2093 ping_monitor(container);
2094 if (mdmon_running(st->container_devnm) &&
2095 st->update_tail == NULL)
2096 st->update_tail = &st->updates;
2097 }
2098
2099 if (s->size == MAX_SIZE)
2100 s->size = 0;
2101 array.size = s->size;
2102 if (s->size & ~INT32_MAX) {
2103 /* got truncated to 32bit, write to
2104 * component_size instead
2105 */
2106 if (sra)
2107 rv = sysfs_set_num(sra, NULL,
2108 "component_size", s->size);
2109 else
2110 rv = -1;
2111 } else {
2112 rv = md_set_array_info(fd, &array);
2113
2114 /* manage array size when it is managed externally
2115 */
2116 if ((rv == 0) && st->ss->external)
2117 rv = set_array_size(st, sra, sra->text_version);
2118 }
2119
2120 if (raid0_takeover) {
2121 /* do not recync non-existing parity,
2122 * we will drop it anyway
2123 */
2124 sysfs_set_str(sra, NULL, "sync_action", "frozen");
2125 /* go back to raid0, drop parity disk
2126 */
2127 sysfs_set_str(sra, NULL, "level", "raid0");
2128 md_get_array_info(fd, &array);
2129 }
2130
2131 size_change_error:
2132 if (rv != 0) {
2133 int err = errno;
2134
2135 /* restore metadata */
2136 if (reshape_super(st, orig_size, UnSet, UnSet, 0, 0,
2137 UnSet, NULL, devname,
2138 ROLLBACK_METADATA_CHANGES,
2139 c->verbose) == 0)
2140 sync_metadata(st);
2141 pr_err("Cannot set device size for %s: %s\n",
2142 devname, strerror(err));
2143 if (err == EBUSY &&
2144 (array.state & (1<<MD_SB_BITMAP_PRESENT)))
2145 cont_err("Bitmap must be removed before size can be changed\n");
2146 rv = 1;
2147 goto release;
2148 }
2149 if (s->assume_clean) {
2150 /* This will fail on kernels older than 3.0 unless
2151 * a backport has been arranged.
2152 */
2153 if (sra == NULL ||
2154 sysfs_set_str(sra, NULL, "resync_start",
2155 "none") < 0)
2156 pr_err("--assume-clean not supported with --grow on this kernel\n");
2157 }
2158 md_get_array_info(fd, &array);
2159 s->size = get_component_size(fd)/2;
2160 if (s->size == 0)
2161 s->size = array.size;
2162 if (c->verbose >= 0) {
2163 if (s->size == orig_size)
2164 pr_err("component size of %s unchanged at %lluK\n",
2165 devname, s->size);
2166 else
2167 pr_err("component size of %s has been set to %lluK\n",
2168 devname, s->size);
2169 }
2170 changed = 1;
2171 } else if (array.level != LEVEL_CONTAINER) {
2172 s->size = get_component_size(fd)/2;
2173 if (s->size == 0)
2174 s->size = array.size;
2175 }
2176
2177 /* See if there is anything else to do */
2178 if ((s->level == UnSet || s->level == array.level) &&
2179 (s->layout_str == NULL) &&
2180 (s->chunk == 0 || s->chunk == array.chunk_size) &&
2181 s->data_offset == INVALID_SECTORS &&
2182 (s->raiddisks == 0 || s->raiddisks == array.raid_disks)) {
2183 /* Nothing more to do */
2184 if (!changed && c->verbose >= 0)
2185 pr_err("%s: no change requested\n", devname);
2186 goto release;
2187 }
2188
2189 /* ========= check for Raid10/Raid1 -> Raid0 conversion ===============
2190 * current implementation assumes that following conditions must be met:
2191 * - RAID10:
2192 * - far_copies == 1
2193 * - near_copies == 2
2194 */
2195 if ((s->level == 0 && array.level == 10 && sra &&
2196 array.layout == ((1 << 8) + 2) && !(array.raid_disks & 1)) ||
2197 (s->level == 0 && array.level == 1 && sra)) {
2198 int err;
2199
2200 err = remove_disks_for_takeover(st, sra, array.layout);
2201 if (err) {
2202 dprintf("Array cannot be reshaped\n");
2203 if (cfd > -1)
2204 close(cfd);
2205 rv = 1;
2206 goto release;
2207 }
2208 /* Make sure mdmon has seen the device removal
2209 * and updated metadata before we continue with
2210 * level change
2211 */
2212 if (container)
2213 ping_monitor(container);
2214 }
2215
2216 memset(&info, 0, sizeof(info));
2217 info.array = array;
2218 if (sysfs_init(&info, fd, NULL)) {
2219 pr_err("failed to initialize sysfs.\n");
2220 rv = 1;
2221 goto release;
2222 }
2223 strcpy(info.text_version, sra->text_version);
2224 info.component_size = s->size*2;
2225 info.new_level = s->level;
2226 info.new_chunk = s->chunk * 1024;
2227 if (info.array.level == LEVEL_CONTAINER) {
2228 info.delta_disks = UnSet;
2229 info.array.raid_disks = s->raiddisks;
2230 } else if (s->raiddisks)
2231 info.delta_disks = s->raiddisks - info.array.raid_disks;
2232 else
2233 info.delta_disks = UnSet;
2234 if (s->layout_str == NULL) {
2235 info.new_layout = UnSet;
2236 if (info.array.level == 6 &&
2237 (info.new_level == 6 || info.new_level == UnSet) &&
2238 info.array.layout >= 16) {
2239 pr_err("%s has a non-standard layout. If you wish to preserve this\n", devname);
2240 cont_err("during the reshape, please specify --layout=preserve\n");
2241 cont_err("If you want to change it, specify a layout or use --layout=normalise\n");
2242 rv = 1;
2243 goto release;
2244 }
2245 } else if (strcmp(s->layout_str, "normalise") == 0 ||
2246 strcmp(s->layout_str, "normalize") == 0) {
2247 /* If we have a -6 RAID6 layout, remove the '-6'. */
2248 info.new_layout = UnSet;
2249 if (info.array.level == 6 && info.new_level == UnSet) {
2250 char l[40], *h;
2251 strcpy(l, map_num_s(r6layout, info.array.layout));
2252 h = strrchr(l, '-');
2253 if (h && strcmp(h, "-6") == 0) {
2254 *h = 0;
2255 info.new_layout = map_name(r6layout, l);
2256 }
2257 } else {
2258 pr_err("%s is only meaningful when reshaping a RAID6 array.\n", s->layout_str);
2259 rv = 1;
2260 goto release;
2261 }
2262 } else if (strcmp(s->layout_str, "preserve") == 0) {
2263 /* This means that a non-standard RAID6 layout
2264 * is OK.
2265 * In particular:
2266 * - When reshape a RAID6 (e.g. adding a device)
2267 * which is in a non-standard layout, it is OK
2268 * to preserve that layout.
2269 * - When converting a RAID5 to RAID6, leave it in
2270 * the XXX-6 layout, don't re-layout.
2271 */
2272 if (info.array.level == 6 && info.new_level == UnSet)
2273 info.new_layout = info.array.layout;
2274 else if (info.array.level == 5 && info.new_level == 6) {
2275 char l[40];
2276 strcpy(l, map_num_s(r5layout, info.array.layout));
2277 strcat(l, "-6");
2278 info.new_layout = map_name(r6layout, l);
2279 } else {
2280 pr_err("%s in only meaningful when reshaping to RAID6\n", s->layout_str);
2281 rv = 1;
2282 goto release;
2283 }
2284 } else {
2285 int l = info.new_level;
2286 if (l == UnSet)
2287 l = info.array.level;
2288 switch (l) {
2289 case 5:
2290 info.new_layout = map_name(r5layout, s->layout_str);
2291 break;
2292 case 6:
2293 info.new_layout = map_name(r6layout, s->layout_str);
2294 break;
2295 case 10:
2296 info.new_layout = parse_layout_10(s->layout_str);
2297 break;
2298 case LEVEL_FAULTY:
2299 info.new_layout = parse_layout_faulty(s->layout_str);
2300 break;
2301 default:
2302 pr_err("layout not meaningful with this level\n");
2303 rv = 1;
2304 goto release;
2305 }
2306 if (info.new_layout == UnSet) {
2307 pr_err("layout %s not understood for this level\n",
2308 s->layout_str);
2309 rv = 1;
2310 goto release;
2311 }
2312 }
2313
2314 if (array.level == LEVEL_FAULTY) {
2315 if (s->level != UnSet && s->level != array.level) {
2316 pr_err("cannot change level of Faulty device\n");
2317 rv =1 ;
2318 }
2319 if (s->chunk) {
2320 pr_err("cannot set chunksize of Faulty device\n");
2321 rv =1 ;
2322 }
2323 if (s->raiddisks && s->raiddisks != 1) {
2324 pr_err("cannot set raid_disks of Faulty device\n");
2325 rv =1 ;
2326 }
2327 if (s->layout_str) {
2328 if (md_get_array_info(fd, &array) != 0) {
2329 dprintf("Cannot get array information.\n");
2330 goto release;
2331 }
2332 array.layout = info.new_layout;
2333 if (md_set_array_info(fd, &array) != 0) {
2334 pr_err("failed to set new layout\n");
2335 rv = 1;
2336 } else if (c->verbose >= 0)
2337 printf("layout for %s set to %d\n",
2338 devname, array.layout);
2339 }
2340 } else if (array.level == LEVEL_CONTAINER) {
2341 /* This change is to be applied to every array in the
2342 * container. This is only needed when the metadata imposes
2343 * restraints of the various arrays in the container.
2344 * Currently we only know that IMSM requires all arrays
2345 * to have the same number of devices so changing the
2346 * number of devices (On-Line Capacity Expansion) must be
2347 * performed at the level of the container
2348 */
2349 close_fd(&fd);
2350 rv = reshape_container(container, devname, -1, st, &info,
2351 c->force, c->backup_file, c->verbose,
2352 0, 0, 0);
2353 frozen = 0;
2354 } else {
2355 /* get spare devices from external metadata
2356 */
2357 if (st->ss->external) {
2358 struct mdinfo *info2;
2359
2360 info2 = st->ss->container_content(st, subarray);
2361 if (info2) {
2362 info.array.spare_disks =
2363 info2->array.spare_disks;
2364 sysfs_free(info2);
2365 }
2366 }
2367
2368 /* Impose these changes on a single array. First
2369 * check that the metadata is OK with the change. */
2370
2371 if (reshape_super(st, 0, info.new_level,
2372 info.new_layout, info.new_chunk,
2373 info.array.raid_disks, info.delta_disks,
2374 c->backup_file, devname,
2375 APPLY_METADATA_CHANGES, c->verbose)) {
2376 rv = 1;
2377 goto release;
2378 }
2379 sync_metadata(st);
2380 rv = reshape_array(container, fd, devname, st, &info, c->force,
2381 devlist, s->data_offset, c->backup_file,
2382 c->verbose, 0, 0, 0);
2383 frozen = 0;
2384 }
2385 release:
2386 sysfs_free(sra);
2387 if (frozen > 0)
2388 unfreeze(st);
2389 return rv;
2390 }
2391
2392 /* verify_reshape_position()
2393 * Function checks if reshape position in metadata is not farther
2394 * than position in md.
2395 * Return value:
2396 * 0 : not valid sysfs entry
2397 * it can be caused by not started reshape, it should be started
2398 * by reshape array or raid0 array is before takeover
2399 * -1 : error, reshape position is obviously wrong
2400 * 1 : success, reshape progress correct or updated
2401 */
2402 static int verify_reshape_position(struct mdinfo *info, int level)
2403 {
2404 int ret_val = 0;
2405 char buf[40];
2406 int rv;
2407
2408 /* read sync_max, failure can mean raid0 array */
2409 rv = sysfs_get_str(info, NULL, "sync_max", buf, 40);
2410
2411 if (rv > 0) {
2412 char *ep;
2413 unsigned long long position = strtoull(buf, &ep, 0);
2414
2415 dprintf("Read sync_max sysfs entry is: %s\n", buf);
2416 if (!(ep == buf || (*ep != 0 && *ep != '\n' && *ep != ' '))) {
2417 position *= get_data_disks(level,
2418 info->new_layout,
2419 info->array.raid_disks);
2420 if (info->reshape_progress < position) {
2421 dprintf("Corrected reshape progress (%llu) to md position (%llu)\n",
2422 info->reshape_progress, position);
2423 info->reshape_progress = position;
2424 ret_val = 1;
2425 } else if (info->reshape_progress > position) {
2426 pr_err("Fatal error: array reshape was not properly frozen (expected reshape position is %llu, but reshape progress is %llu.\n",
2427 position, info->reshape_progress);
2428 ret_val = -1;
2429 } else {
2430 dprintf("Reshape position in md and metadata are the same;");
2431 ret_val = 1;
2432 }
2433 }
2434 } else if (rv == 0) {
2435 /* for valid sysfs entry, 0-length content
2436 * should be indicated as error
2437 */
2438 ret_val = -1;
2439 }
2440
2441 return ret_val;
2442 }
2443
2444 static unsigned long long choose_offset(unsigned long long lo,
2445 unsigned long long hi,
2446 unsigned long long min,
2447 unsigned long long max)
2448 {
2449 /* Choose a new offset between hi and lo.
2450 * It must be between min and max, but
2451 * we would prefer something near the middle of hi/lo, and also
2452 * prefer to be aligned to a big power of 2.
2453 *
2454 * So we start with the middle, then for each bit,
2455 * starting at '1' and increasing, if it is set, we either
2456 * add it or subtract it if possible, preferring the option
2457 * which is furthest from the boundary.
2458 *
2459 * We stop once we get a 1MB alignment. As units are in sectors,
2460 * 1MB = 2*1024 sectors.
2461 */
2462 unsigned long long choice = (lo + hi) / 2;
2463 unsigned long long bit = 1;
2464
2465 for (bit = 1; bit < 2*1024; bit = bit << 1) {
2466 unsigned long long bigger, smaller;
2467 if (! (bit & choice))
2468 continue;
2469 bigger = choice + bit;
2470 smaller = choice - bit;
2471 if (bigger > max && smaller < min)
2472 break;
2473 if (bigger > max)
2474 choice = smaller;
2475 else if (smaller < min)
2476 choice = bigger;
2477 else if (hi - bigger > smaller - lo)
2478 choice = bigger;
2479 else
2480 choice = smaller;
2481 }
2482 return choice;
2483 }
2484
2485 static int set_new_data_offset(struct mdinfo *sra, struct supertype *st,
2486 char *devname, int delta_disks,
2487 unsigned long long data_offset,
2488 unsigned long long min,
2489 int can_fallback)
2490 {
2491 struct mdinfo *sd;
2492 int dir = 0;
2493 int err = 0;
2494 unsigned long long before, after;
2495
2496 /* Need to find min space before and after so same is used
2497 * on all devices
2498 */
2499 before = UINT64_MAX;
2500 after = UINT64_MAX;
2501 for (sd = sra->devs; sd; sd = sd->next) {
2502 char *dn;
2503 int dfd;
2504 int rv;
2505 struct supertype *st2;
2506 struct mdinfo info2;
2507
2508 if (sd->disk.state & (1<<MD_DISK_FAULTY))
2509 continue;
2510 dn = map_dev(sd->disk.major, sd->disk.minor, 0);
2511 dfd = dev_open(dn, O_RDONLY);
2512 if (dfd < 0) {
2513 pr_err("%s: cannot open component %s\n",
2514 devname, dn ? dn : "-unknown-");
2515 goto release;
2516 }
2517 st2 = dup_super(st);
2518 rv = st2->ss->load_super(st2,dfd, NULL);
2519 close(dfd);
2520 if (rv) {
2521 free(st2);
2522 pr_err("%s: cannot get superblock from %s\n",
2523 devname, dn);
2524 goto release;
2525 }
2526 st2->ss->getinfo_super(st2, &info2, NULL);
2527 st2->ss->free_super(st2);
2528 free(st2);
2529 if (info2.space_before == 0 &&
2530 info2.space_after == 0) {
2531 /* Metadata doesn't support data_offset changes */
2532 if (!can_fallback)
2533 pr_err("%s: Metadata version doesn't support data_offset changes\n",
2534 devname);
2535 goto fallback;
2536 }
2537 if (before > info2.space_before)
2538 before = info2.space_before;
2539 if (after > info2.space_after)
2540 after = info2.space_after;
2541
2542 if (data_offset != INVALID_SECTORS) {
2543 if (dir == 0) {
2544 if (info2.data_offset == data_offset) {
2545 pr_err("%s: already has that data_offset\n",
2546 dn);
2547 goto release;
2548 }
2549 if (data_offset < info2.data_offset)
2550 dir = -1;
2551 else
2552 dir = 1;
2553 } else if ((data_offset <= info2.data_offset &&
2554 dir == 1) ||
2555 (data_offset >= info2.data_offset &&
2556 dir == -1)) {
2557 pr_err("%s: differing data offsets on devices make this --data-offset setting impossible\n",
2558 dn);
2559 goto release;
2560 }
2561 }
2562 }
2563 if (before == UINT64_MAX)
2564 /* impossible really, there must be no devices */
2565 return 1;
2566
2567 for (sd = sra->devs; sd; sd = sd->next) {
2568 char *dn = map_dev(sd->disk.major, sd->disk.minor, 0);
2569 unsigned long long new_data_offset;
2570
2571 if (sd->disk.state & (1<<MD_DISK_FAULTY))
2572 continue;
2573 if (delta_disks < 0) {
2574 /* Don't need any space as array is shrinking
2575 * just move data_offset up by min
2576 */
2577 if (data_offset == INVALID_SECTORS)
2578 new_data_offset = sd->data_offset + min;
2579 else {
2580 if (data_offset < sd->data_offset + min) {
2581 pr_err("--data-offset too small for %s\n",
2582 dn);
2583 goto release;
2584 }
2585 new_data_offset = data_offset;
2586 }
2587 } else if (delta_disks > 0) {
2588 /* need space before */
2589 if (before < min) {
2590 if (can_fallback)
2591 goto fallback;
2592 pr_err("Insufficient head-space for reshape on %s\n",
2593 dn);
2594 goto release;
2595 }
2596 if (data_offset == INVALID_SECTORS)
2597 new_data_offset = sd->data_offset - min;
2598 else {
2599 if (data_offset > sd->data_offset - min) {
2600 pr_err("--data-offset too large for %s\n",
2601 dn);
2602 goto release;
2603 }
2604 new_data_offset = data_offset;
2605 }
2606 } else {
2607 if (dir == 0) {
2608 /* can move up or down. If 'data_offset'
2609 * was set we would have already decided,
2610 * so just choose direction with most space.
2611 */
2612 if (before > after)
2613 dir = -1;
2614 else
2615 dir = 1;
2616 }
2617 sysfs_set_str(sra, NULL, "reshape_direction",
2618 dir == 1 ? "backwards" : "forwards");
2619 if (dir > 0) {
2620 /* Increase data offset */
2621 if (after < min) {
2622 if (can_fallback)
2623 goto fallback;
2624 pr_err("Insufficient tail-space for reshape on %s\n",
2625 dn);
2626 goto release;
2627 }
2628 if (data_offset != INVALID_SECTORS &&
2629 data_offset < sd->data_offset + min) {
2630 pr_err("--data-offset too small on %s\n",
2631 dn);
2632 goto release;
2633 }
2634 if (data_offset != INVALID_SECTORS)
2635 new_data_offset = data_offset;
2636 else
2637 new_data_offset = choose_offset(sd->data_offset,
2638 sd->data_offset + after,
2639 sd->data_offset + min,
2640 sd->data_offset + after);
2641 } else {
2642 /* Decrease data offset */
2643 if (before < min) {
2644 if (can_fallback)
2645 goto fallback;
2646 pr_err("insufficient head-room on %s\n",
2647 dn);
2648 goto release;
2649 }
2650 if (data_offset != INVALID_SECTORS &&
2651 data_offset > sd->data_offset - min) {
2652 pr_err("--data-offset too large on %s\n",
2653 dn);
2654 goto release;
2655 }
2656 if (data_offset != INVALID_SECTORS)
2657 new_data_offset = data_offset;
2658 else
2659 new_data_offset = choose_offset(sd->data_offset - before,
2660 sd->data_offset,
2661 sd->data_offset - before,
2662 sd->data_offset - min);
2663 }
2664 }
2665 err = sysfs_set_num(sra, sd, "new_offset", new_data_offset);
2666 if (err < 0 && errno == E2BIG) {
2667 /* try again after increasing data size to max */
2668 err = sysfs_set_num(sra, sd, "size", 0);
2669 if (err < 0 && errno == EINVAL &&
2670 !(sd->disk.state & (1<<MD_DISK_SYNC))) {
2671 /* some kernels have a bug where you cannot
2672 * use '0' on spare devices. */
2673 sysfs_set_num(sra, sd, "size",
2674 (sra->component_size + after)/2);
2675 }
2676 err = sysfs_set_num(sra, sd, "new_offset",
2677 new_data_offset);
2678 }
2679 if (err < 0) {
2680 if (errno == E2BIG && data_offset != INVALID_SECTORS) {
2681 pr_err("data-offset is too big for %s\n", dn);
2682 goto release;
2683 }
2684 if (sd == sra->devs &&
2685 (errno == ENOENT || errno == E2BIG))
2686 /* Early kernel, no 'new_offset' file,
2687 * or kernel doesn't like us.
2688 * For RAID5/6 this is not fatal
2689 */
2690 return 1;
2691 pr_err("Cannot set new_offset for %s\n", dn);
2692 break;
2693 }
2694 }
2695 return err;
2696 release:
2697 return -1;
2698 fallback:
2699 /* Just use a backup file */
2700 return 1;
2701 }
2702
2703 static int raid10_reshape(char *container, int fd, char *devname,
2704 struct supertype *st, struct mdinfo *info,
2705 struct reshape *reshape,
2706 unsigned long long data_offset,
2707 int force, int verbose)
2708 {
2709 /* Changing raid_disks, layout, chunksize or possibly
2710 * just data_offset for a RAID10.
2711 * We must always change data_offset. We change by at least
2712 * ->min_offset_change which is the largest of the old and new
2713 * chunk sizes.
2714 * If raid_disks is increasing, then data_offset must decrease
2715 * by at least this copy size.
2716 * If raid_disks is unchanged, data_offset must increase or
2717 * decrease by at least min_offset_change but preferably by much more.
2718 * We choose half of the available space.
2719 * If raid_disks is decreasing, data_offset must increase by
2720 * at least min_offset_change. To allow of this, component_size
2721 * must be decreased by the same amount.
2722 *
2723 * So we calculate the required minimum and direction, possibly
2724 * reduce the component_size, then iterate through the devices
2725 * and set the new_data_offset.
2726 * If that all works, we set chunk_size, layout, raid_disks, and start
2727 * 'reshape'
2728 */
2729 struct mdinfo *sra;
2730 unsigned long long min;
2731 int err = 0;
2732
2733 sra = sysfs_read(fd, NULL,
2734 GET_COMPONENT|GET_DEVS|GET_OFFSET|GET_STATE|GET_CHUNK
2735 );
2736 if (!sra) {
2737 pr_err("%s: Cannot get array details from sysfs\n", devname);
2738 goto release;
2739 }
2740 min = reshape->min_offset_change;
2741
2742 if (info->delta_disks)
2743 sysfs_set_str(sra, NULL, "reshape_direction",
2744 info->delta_disks < 0 ? "backwards" : "forwards");
2745 if (info->delta_disks < 0 && info->space_after < min) {
2746 int rv = sysfs_set_num(sra, NULL, "component_size",
2747 (sra->component_size - min)/2);
2748 if (rv) {
2749 pr_err("cannot reduce component size\n");
2750 goto release;
2751 }
2752 }
2753 err = set_new_data_offset(sra, st, devname, info->delta_disks,
2754 data_offset, min, 0);
2755 if (err == 1) {
2756 pr_err("Cannot set new_data_offset: RAID10 reshape not\n");
2757 cont_err("supported on this kernel\n");
2758 err = -1;
2759 }
2760 if (err < 0)
2761 goto release;
2762
2763 if (!err && sysfs_set_num(sra, NULL, "chunk_size", info->new_chunk) < 0)
2764 err = errno;
2765 if (!err && sysfs_set_num(sra, NULL, "layout",
2766 reshape->after.layout) < 0)
2767 err = errno;
2768 if (!err &&
2769 sysfs_set_num(sra, NULL, "raid_disks",
2770 info->array.raid_disks + info->delta_disks) < 0)
2771 err = errno;
2772 if (!err && sysfs_set_str(sra, NULL, "sync_action", "reshape") < 0)
2773 err = errno;
2774 if (err) {
2775 pr_err("Cannot set array shape for %s\n",
2776 devname);
2777 if (err == EBUSY &&
2778 (info->array.state & (1<<MD_SB_BITMAP_PRESENT)))
2779 cont_err(" Bitmap must be removed before shape can be changed\n");
2780 goto release;
2781 }
2782 sysfs_free(sra);
2783 return 0;
2784 release:
2785 sysfs_free(sra);
2786 return 1;
2787 }
2788
2789 static void get_space_after(int fd, struct supertype *st, struct mdinfo *info)
2790 {
2791 struct mdinfo *sra, *sd;
2792 /* Initialisation to silence compiler warning */
2793 unsigned long long min_space_before = 0, min_space_after = 0;
2794 int first = 1;
2795
2796 sra = sysfs_read(fd, NULL, GET_DEVS);
2797 if (!sra)
2798 return;
2799 for (sd = sra->devs; sd; sd = sd->next) {
2800 char *dn;
2801 int dfd;
2802 struct supertype *st2;
2803 struct mdinfo info2;
2804
2805 if (sd->disk.state & (1<<MD_DISK_FAULTY))
2806 continue;
2807 dn = map_dev(sd->disk.major, sd->disk.minor, 0);
2808 dfd = dev_open(dn, O_RDONLY);
2809 if (dfd < 0)
2810 break;
2811 st2 = dup_super(st);
2812 if (st2->ss->load_super(st2,dfd, NULL)) {
2813 close(dfd);
2814 free(st2);
2815 break;
2816 }
2817 close(dfd);
2818 st2->ss->getinfo_super(st2, &info2, NULL);
2819 st2->ss->free_super(st2);
2820 free(st2);
2821 if (first ||
2822 min_space_before > info2.space_before)
2823 min_space_before = info2.space_before;
2824 if (first ||
2825 min_space_after > info2.space_after)
2826 min_space_after = info2.space_after;
2827 first = 0;
2828 }
2829 if (sd == NULL && !first) {
2830 info->space_after = min_space_after;
2831 info->space_before = min_space_before;
2832 }
2833 sysfs_free(sra);
2834 }
2835
2836 static void update_cache_size(char *container, struct mdinfo *sra,
2837 struct mdinfo *info,
2838 int disks, unsigned long long blocks)
2839 {
2840 /* Check that the internal stripe cache is
2841 * large enough, or it won't work.
2842 * It must hold at least 4 stripes of the larger
2843 * chunk size
2844 */
2845 unsigned long cache;
2846 cache = max(info->array.chunk_size, info->new_chunk);
2847 cache *= 4; /* 4 stripes minimum */
2848 cache /= 512; /* convert to sectors */
2849 /* make sure there is room for 'blocks' with a bit to spare */
2850 if (cache < 16 + blocks / disks)
2851 cache = 16 + blocks / disks;
2852 cache /= (4096/512); /* Convert from sectors to pages */
2853
2854 if (sra->cache_size < cache)
2855 subarray_set_num(container, sra, "stripe_cache_size",
2856 cache+1);
2857 }
2858
2859 static int impose_reshape(struct mdinfo *sra,
2860 struct mdinfo *info,
2861 struct supertype *st,
2862 int fd,
2863 int restart,
2864 char *devname, char *container,
2865 struct reshape *reshape)
2866 {
2867 struct mdu_array_info_s array;
2868
2869 sra->new_chunk = info->new_chunk;
2870
2871 if (restart) {
2872 /* for external metadata checkpoint saved by mdmon can be lost
2873 * or missed /due to e.g. crash/. Check if md is not during
2874 * restart farther than metadata points to.
2875 * If so, this means metadata information is obsolete.
2876 */
2877 if (st->ss->external)
2878 verify_reshape_position(info, reshape->level);
2879 sra->reshape_progress = info->reshape_progress;
2880 } else {
2881 sra->reshape_progress = 0;
2882 if (reshape->after.data_disks < reshape->before.data_disks)
2883 /* start from the end of the new array */
2884 sra->reshape_progress = (sra->component_size
2885 * reshape->after.data_disks);
2886 }
2887
2888 md_get_array_info(fd, &array);
2889 if (info->array.chunk_size == info->new_chunk &&
2890 reshape->before.layout == reshape->after.layout &&
2891 st->ss->external == 0) {
2892 /* use SET_ARRAY_INFO but only if reshape hasn't started */
2893 array.raid_disks = reshape->after.data_disks + reshape->parity;
2894 if (!restart && md_set_array_info(fd, &array) != 0) {
2895 int err = errno;
2896
2897 pr_err("Cannot set device shape for %s: %s\n",
2898 devname, strerror(errno));
2899
2900 if (err == EBUSY &&
2901 (array.state & (1<<MD_SB_BITMAP_PRESENT)))
2902 cont_err("Bitmap must be removed before shape can be changed\n");
2903
2904 goto release;
2905 }
2906 } else if (!restart) {
2907 /* set them all just in case some old 'new_*' value
2908 * persists from some earlier problem.
2909 */
2910 int err = 0;
2911 if (sysfs_set_num(sra, NULL, "chunk_size", info->new_chunk) < 0)
2912 err = errno;
2913 if (!err && sysfs_set_num(sra, NULL, "layout",
2914 reshape->after.layout) < 0)
2915 err = errno;
2916 if (!err && subarray_set_num(container, sra, "raid_disks",
2917 reshape->after.data_disks +
2918 reshape->parity) < 0)
2919 err = errno;
2920 if (err) {
2921 pr_err("Cannot set device shape for %s\n", devname);
2922
2923 if (err == EBUSY &&
2924 (array.state & (1<<MD_SB_BITMAP_PRESENT)))
2925 cont_err("Bitmap must be removed before shape can be changed\n");
2926 goto release;
2927 }
2928 }
2929 return 0;
2930 release:
2931 return -1;
2932 }
2933
2934 static int impose_level(int fd, int level, char *devname, int verbose)
2935 {
2936 char *c;
2937 struct mdu_array_info_s array;
2938 struct mdinfo info;
2939
2940 if (sysfs_init(&info, fd, NULL)) {
2941 pr_err("failed to initialize sysfs.\n");
2942 return 1;
2943 }
2944
2945 md_get_array_info(fd, &array);
2946 if (level == 0 && is_level456(array.level)) {
2947 /* To convert to RAID0 we need to fail and
2948 * remove any non-data devices. */
2949 int found = 0;
2950 int d;
2951 int data_disks = array.raid_disks - 1;
2952 if (array.level == 6)
2953 data_disks -= 1;
2954 if (array.level == 5 && array.layout != ALGORITHM_PARITY_N)
2955 return -1;
2956 if (array.level == 6 && array.layout != ALGORITHM_PARITY_N_6)
2957 return -1;
2958 sysfs_set_str(&info, NULL,"sync_action", "idle");
2959 /* First remove any spares so no recovery starts */
2960 for (d = 0, found = 0;
2961 d < MAX_DISKS && found < array.nr_disks; d++) {
2962 mdu_disk_info_t disk;
2963 disk.number = d;
2964 if (md_get_disk_info(fd, &disk) < 0)
2965 continue;
2966 if (disk.major == 0 && disk.minor == 0)
2967 continue;
2968 found++;
2969 if ((disk.state & (1 << MD_DISK_ACTIVE)) &&
2970 disk.raid_disk < data_disks)
2971 /* keep this */
2972 continue;
2973 ioctl(fd, HOT_REMOVE_DISK,
2974 makedev(disk.major, disk.minor));
2975 }
2976 /* Now fail anything left */
2977 md_get_array_info(fd, &array);
2978 for (d = 0, found = 0;
2979 d < MAX_DISKS && found < array.nr_disks; d++) {
2980 mdu_disk_info_t disk;
2981 disk.number = d;
2982 if (md_get_disk_info(fd, &disk) < 0)
2983 continue;
2984 if (disk.major == 0 && disk.minor == 0)
2985 continue;
2986 found++;
2987 if ((disk.state & (1 << MD_DISK_ACTIVE)) &&
2988 disk.raid_disk < data_disks)
2989 /* keep this */
2990 continue;
2991 ioctl(fd, SET_DISK_FAULTY,
2992 makedev(disk.major, disk.minor));
2993 hot_remove_disk(fd, makedev(disk.major, disk.minor), 1);
2994 }
2995 }
2996 c = map_num(pers, level);
2997 if (c) {
2998 int err = sysfs_set_str(&info, NULL, "level", c);
2999 if (err) {
3000 err = errno;
3001 pr_err("%s: could not set level to %s\n",
3002 devname, c);
3003 if (err == EBUSY &&
3004 (array.state & (1<<MD_SB_BITMAP_PRESENT)))
3005 cont_err("Bitmap must be removed before level can be changed\n");
3006 return err;
3007 }
3008 if (verbose >= 0)
3009 pr_err("level of %s changed to %s\n", devname, c);
3010 }
3011 return 0;
3012 }
3013
3014 int sigterm = 0;
3015 static void catch_term(int sig)
3016 {
3017 sigterm = 1;
3018 }
3019
3020 static int reshape_array(char *container, int fd, char *devname,
3021 struct supertype *st, struct mdinfo *info,
3022 int force, struct mddev_dev *devlist,
3023 unsigned long long data_offset,
3024 char *backup_file, int verbose, int forked,
3025 int restart, int freeze_reshape)
3026 {
3027 struct reshape reshape;
3028 int spares_needed;
3029 char *msg;
3030 int orig_level = UnSet;
3031 int odisks;
3032 int delayed;
3033
3034 struct mdu_array_info_s array;
3035 char *c;
3036
3037 struct mddev_dev *dv;
3038 int added_disks;
3039
3040 int *fdlist = NULL;
3041 unsigned long long *offsets = NULL;
3042 int d;
3043 int nrdisks;
3044 int err;
3045 unsigned long blocks;
3046 unsigned long long array_size;
3047 int done;
3048 struct mdinfo *sra = NULL;
3049 char buf[20];
3050
3051 /* when reshaping a RAID0, the component_size might be zero.
3052 * So try to fix that up.
3053 */
3054 if (md_get_array_info(fd, &array) != 0) {
3055 dprintf("Cannot get array information.\n");
3056 goto release;
3057 }
3058 if (array.level == 0 && info->component_size == 0) {
3059 get_dev_size(fd, NULL, &array_size);
3060 info->component_size = array_size / array.raid_disks;
3061 }
3062
3063 if (array.level == 10)
3064 /* Need space_after info */
3065 get_space_after(fd, st, info);
3066
3067 if (info->reshape_active) {
3068 int new_level = info->new_level;
3069 info->new_level = UnSet;
3070 if (info->delta_disks > 0)
3071 info->array.raid_disks -= info->delta_disks;
3072 msg = analyse_change(devname, info, &reshape);
3073 info->new_level = new_level;
3074 if (info->delta_disks > 0)
3075 info->array.raid_disks += info->delta_disks;
3076 if (!restart)
3077 /* Make sure the array isn't read-only */
3078 ioctl(fd, RESTART_ARRAY_RW, 0);
3079 } else
3080 msg = analyse_change(devname, info, &reshape);
3081 if (msg) {
3082 /* if msg == "", error has already been printed */
3083 if (msg[0])
3084 pr_err("%s\n", msg);
3085 goto release;
3086 }
3087 if (restart && (reshape.level != info->array.level ||
3088 reshape.before.layout != info->array.layout ||
3089 reshape.before.data_disks + reshape.parity !=
3090 info->array.raid_disks - max(0, info->delta_disks))) {
3091 pr_err("reshape info is not in native format - cannot continue.\n");
3092 goto release;
3093 }
3094
3095 if (st->ss->external && restart && (info->reshape_progress == 0) &&
3096 !((sysfs_get_str(info, NULL, "sync_action",
3097 buf, sizeof(buf)) > 0) &&
3098 (strncmp(buf, "reshape", 7) == 0))) {
3099 /* When reshape is restarted from '0', very begin of array
3100 * it is possible that for external metadata reshape and array
3101 * configuration doesn't happen.
3102 * Check if md has the same opinion, and reshape is restarted
3103 * from 0. If so, this is regular reshape start after reshape
3104 * switch in metadata to next array only.
3105 */
3106 if ((verify_reshape_position(info, reshape.level) >= 0) &&
3107 (info->reshape_progress == 0))
3108 restart = 0;
3109 }
3110 if (restart) {
3111 /*
3112 * reshape already started. just skip to monitoring
3113 * the reshape
3114 */
3115 if (reshape.backup_blocks == 0)
3116 return 0;
3117 if (restart & RESHAPE_NO_BACKUP)
3118 return 0;
3119
3120 /* Need 'sra' down at 'started:' */
3121 sra = sysfs_read(fd, NULL,
3122 GET_COMPONENT|GET_DEVS|GET_OFFSET|GET_STATE|
3123 GET_CHUNK|GET_CACHE);
3124 if (!sra) {
3125 pr_err("%s: Cannot get array details from sysfs\n",
3126 devname);
3127 goto release;
3128 }
3129
3130 if (!backup_file)
3131 backup_file = locate_backup(sra->sys_name);
3132
3133 goto started;
3134 }
3135 /* The container is frozen but the array may not be.
3136 * So freeze the array so spares don't get put to the wrong use
3137 * FIXME there should probably be a cleaner separation between
3138 * freeze_array and freeze_container.
3139 */
3140 sysfs_freeze_array(info);
3141 /* Check we have enough spares to not be degraded */
3142 added_disks = 0;
3143 for (dv = devlist; dv ; dv=dv->next)
3144 added_disks++;
3145 spares_needed = max(reshape.before.data_disks,
3146 reshape.after.data_disks) +
3147 reshape.parity - array.raid_disks;
3148
3149 if (!force && info->new_level > 1 && info->array.level > 1 &&
3150 spares_needed > info->array.spare_disks + added_disks) {
3151 pr_err("Need %d spare%s to avoid degraded array, and only have %d.\n"
3152 " Use --force to over-ride this check.\n",
3153 spares_needed,
3154 spares_needed == 1 ? "" : "s",
3155 info->array.spare_disks + added_disks);
3156 goto release;
3157 }
3158 /* Check we have enough spares to not fail */
3159 spares_needed = max(reshape.before.data_disks,
3160 reshape.after.data_disks)
3161 - array.raid_disks;
3162 if ((info->new_level > 1 || info->new_level == 0) &&
3163 spares_needed > info->array.spare_disks +added_disks) {
3164 pr_err("Need %d spare%s to create working array, and only have %d.\n",
3165 spares_needed, spares_needed == 1 ? "" : "s",
3166 info->array.spare_disks + added_disks);
3167 goto release;
3168 }
3169
3170 if (reshape.level != array.level) {
3171 int err = impose_level(fd, reshape.level, devname, verbose);
3172 if (err)
3173 goto release;
3174 info->new_layout = UnSet; /* after level change,
3175 * layout is meaningless */
3176 orig_level = array.level;
3177 sysfs_freeze_array(info);
3178
3179 if (reshape.level > 0 && st->ss->external) {
3180 /* make sure mdmon is aware of the new level */
3181 if (mdmon_running(container))
3182 flush_mdmon(container);
3183
3184 if (!mdmon_running(container))
3185 start_mdmon(container);
3186 ping_monitor(container);
3187 if (mdmon_running(container) && st->update_tail == NULL)
3188 st->update_tail = &st->updates;
3189 }
3190 }
3191 /* ->reshape_super might have chosen some spares from the
3192 * container that it wants to be part of the new array.
3193 * We can collect them with ->container_content and give
3194 * them to the kernel.
3195 */
3196 if (st->ss->reshape_super && st->ss->container_content) {
3197 char *subarray = strchr(info->text_version+1, '/')+1;
3198 struct mdinfo *info2 =
3199 st->ss->container_content(st, subarray);
3200 struct mdinfo *d;
3201
3202 if (info2) {
3203 if (sysfs_init(info2, fd, st->devnm)) {
3204 pr_err("unable to initialize sysfs for %s\n",
3205 st->devnm);
3206 free(info2);
3207 goto release;
3208 }
3209 /* When increasing number of devices, we need to set
3210 * new raid_disks before adding these, or they might
3211 * be rejected.
3212 */
3213 if (reshape.backup_blocks &&
3214 reshape.after.data_disks >
3215 reshape.before.data_disks)
3216 subarray_set_num(container, info2, "raid_disks",
3217 reshape.after.data_disks +
3218 reshape.parity);
3219 for (d = info2->devs; d; d = d->next) {
3220 if (d->disk.state == 0 &&
3221 d->disk.raid_disk >= 0) {
3222 /* This is a spare that wants to
3223 * be part of the array.
3224 */
3225 add_disk(fd, st, info2, d);
3226 }
3227 }
3228 sysfs_free(info2);
3229 }
3230 }
3231 /* We might have been given some devices to add to the
3232 * array. Now that the array has been changed to the right
3233 * level and frozen, we can safely add them.
3234 */
3235 if (devlist) {
3236 if (Manage_subdevs(devname, fd, devlist, verbose, 0, NULL, 0))
3237 goto release;
3238 }
3239
3240 if (reshape.backup_blocks == 0 && data_offset != INVALID_SECTORS)
3241 reshape.backup_blocks = reshape.before.data_disks * info->array.chunk_size/512;
3242 if (reshape.backup_blocks == 0) {
3243 /* No restriping needed, but we might need to impose
3244 * some more changes: layout, raid_disks, chunk_size
3245 */
3246 /* read current array info */
3247 if (md_get_array_info(fd, &array) != 0) {
3248 dprintf("Cannot get array information.\n");
3249 goto release;
3250 }
3251 /* compare current array info with new values and if
3252 * it is different update them to new */
3253 if (info->new_layout != UnSet &&
3254 info->new_layout != array.layout) {
3255 array.layout = info->new_layout;
3256 if (md_set_array_info(fd, &array) != 0) {
3257 pr_err("failed to set new layout\n");
3258 goto release;
3259 } else if (verbose >= 0)
3260 printf("layout for %s set to %d\n",
3261 devname, array.layout);
3262 }
3263 if (info->delta_disks != UnSet && info->delta_disks != 0 &&
3264 array.raid_disks !=
3265 (info->array.raid_disks + info->delta_disks)) {
3266 array.raid_disks += info->delta_disks;
3267 if (md_set_array_info(fd, &array) != 0) {
3268 pr_err("failed to set raid disks\n");
3269 goto release;
3270 } else if (verbose >= 0) {
3271 printf("raid_disks for %s set to %d\n",
3272 devname, array.raid_disks);
3273 }
3274 }
3275 if (info->new_chunk != 0 &&
3276 info->new_chunk != array.chunk_size) {
3277 if (sysfs_set_num(info, NULL,
3278 "chunk_size", info->new_chunk) != 0) {
3279 pr_err("failed to set chunk size\n");
3280 goto release;
3281 } else if (verbose >= 0)
3282 printf("chunk size for %s set to %d\n",
3283 devname, info->new_chunk);
3284 }
3285 unfreeze(st);
3286 return 0;
3287 }
3288
3289 /*
3290 * There are three possibilities.
3291 * 1/ The array will shrink.
3292 * We need to ensure the reshape will pause before reaching
3293 * the 'critical section'. We also need to fork and wait for
3294 * that to happen. When it does we
3295 * suspend/backup/complete/unfreeze
3296 *
3297 * 2/ The array will not change size.
3298 * This requires that we keep a backup of a sliding window
3299 * so that we can restore data after a crash. So we need
3300 * to fork and monitor progress.
3301 * In future we will allow the data_offset to change, so
3302 * a sliding backup becomes unnecessary.
3303 *
3304 * 3/ The array will grow. This is relatively easy.
3305 * However the kernel's restripe routines will cheerfully
3306 * overwrite some early data before it is safe. So we
3307 * need to make a backup of the early parts of the array
3308 * and be ready to restore it if rebuild aborts very early.
3309 * For externally managed metadata, we still need a forked
3310 * child to monitor the reshape and suspend IO over the region
3311 * that is being reshaped.
3312 *
3313 * We backup data by writing it to one spare, or to a
3314 * file which was given on command line.
3315 *
3316 * In each case, we first make sure that storage is available
3317 * for the required backup.
3318 * Then we:
3319 * - request the shape change.
3320 * - fork to handle backup etc.
3321 */
3322 /* Check that we can hold all the data */
3323 get_dev_size(fd, NULL, &array_size);
3324 if (reshape.new_size < (array_size/512)) {
3325 pr_err("this change will reduce the size of the array.\n"
3326 " use --grow --array-size first to truncate array.\n"
3327 " e.g. mdadm --grow %s --array-size %llu\n",
3328 devname, reshape.new_size/2);
3329 goto release;
3330 }
3331
3332 if (array.level == 10) {
3333 /* Reshaping RAID10 does not require any data backup by
3334 * user-space. Instead it requires that the data_offset
3335 * is changed to avoid the need for backup.
3336 * So this is handled very separately
3337 */
3338 if (restart)
3339 /* Nothing to do. */
3340 return 0;
3341 return raid10_reshape(container, fd, devname, st, info,
3342 &reshape, data_offset, force, verbose);
3343 }
3344 sra = sysfs_read(fd, NULL,
3345 GET_COMPONENT|GET_DEVS|GET_OFFSET|GET_STATE|GET_CHUNK|
3346 GET_CACHE);
3347 if (!sra) {
3348 pr_err("%s: Cannot get array details from sysfs\n",
3349 devname);
3350 goto release;
3351 }
3352
3353 if (!backup_file)
3354 switch(set_new_data_offset(sra, st, devname,
3355 reshape.after.data_disks - reshape.before.data_disks,
3356 data_offset,
3357 reshape.min_offset_change, 1)) {
3358 case -1:
3359 goto release;
3360 case 0:
3361 /* Updated data_offset, so it's easy now */
3362 update_cache_size(container, sra, info,
3363 min(reshape.before.data_disks,
3364 reshape.after.data_disks),
3365 reshape.backup_blocks);
3366
3367 /* Right, everything seems fine. Let's kick things off.
3368 */
3369 sync_metadata(st);
3370
3371 if (impose_reshape(sra, info, st, fd, restart,
3372 devname, container, &reshape) < 0)
3373 goto release;
3374 if (sysfs_set_str(sra, NULL, "sync_action", "reshape") < 0) {
3375 struct mdinfo *sd;
3376 if (errno != EINVAL) {
3377 pr_err("Failed to initiate reshape!\n");
3378 goto release;
3379 }
3380 /* revert data_offset and try the old way */
3381 for (sd = sra->devs; sd; sd = sd->next) {
3382 sysfs_set_num(sra, sd, "new_offset",
3383 sd->data_offset);
3384 sysfs_set_str(sra, NULL, "reshape_direction",
3385 "forwards");
3386 }
3387 break;
3388 }
3389 if (info->new_level == reshape.level)
3390 return 0;
3391 /* need to adjust level when reshape completes */
3392 switch(fork()) {
3393 case -1: /* ignore error, but don't wait */
3394 return 0;
3395 default: /* parent */
3396 return 0;
3397 case 0:
3398 manage_fork_fds(0);
3399 map_fork();
3400 break;
3401 }
3402 close(fd);
3403 wait_reshape(sra);
3404 fd = open_dev(sra->sys_name);
3405 if (fd >= 0)
3406 impose_level(fd, info->new_level, devname, verbose);
3407 return 0;
3408 case 1: /* Couldn't set data_offset, try the old way */
3409 if (data_offset != INVALID_SECTORS) {
3410 pr_err("Cannot update data_offset on this array\n");
3411 goto release;
3412 }
3413 break;
3414 }
3415
3416 started:
3417 /* Decide how many blocks (sectors) for a reshape
3418 * unit. The number we have so far is just a minimum
3419 */
3420 blocks = reshape.backup_blocks;
3421 if (reshape.before.data_disks ==
3422 reshape.after.data_disks) {
3423 /* Make 'blocks' bigger for better throughput, but
3424 * not so big that we reject it below.
3425 * Try for 16 megabytes
3426 */
3427 while (blocks * 32 < sra->component_size && blocks < 16*1024*2)
3428 blocks *= 2;
3429 } else
3430 pr_err("Need to backup %luK of critical section..\n", blocks/2);
3431
3432 if (blocks >= sra->component_size/2) {
3433 pr_err("%s: Something wrong - reshape aborted\n", devname);
3434 goto release;
3435 }
3436
3437 /* Now we need to open all these devices so we can read/write.
3438 */
3439 nrdisks = max(reshape.before.data_disks,
3440 reshape.after.data_disks) + reshape.parity
3441 + sra->array.spare_disks;
3442 fdlist = xcalloc((1+nrdisks), sizeof(int));
3443 offsets = xcalloc((1+nrdisks), sizeof(offsets[0]));
3444
3445 odisks = reshape.before.data_disks + reshape.parity;
3446 d = reshape_prepare_fdlist(devname, sra, odisks, nrdisks, blocks,
3447 backup_file, fdlist, offsets);
3448 if (d < odisks) {
3449 goto release;
3450 }
3451 if ((st->ss->manage_reshape == NULL) ||
3452 (st->ss->recover_backup == NULL)) {
3453 if (backup_file == NULL) {
3454 if (reshape.after.data_disks <=
3455 reshape.before.data_disks) {
3456 pr_err("%s: Cannot grow - need backup-file\n",
3457 devname);
3458 pr_err(" Please provide one with \"--backup=...\"\n");
3459 goto release;
3460 } else if (d == odisks) {
3461 pr_err("%s: Cannot grow - need a spare or backup-file to backup critical section\n", devname);
3462 goto release;
3463 }
3464 } else {
3465 if (!reshape_open_backup_file(backup_file, fd, devname,
3466 (signed)blocks,
3467 fdlist+d, offsets+d,
3468 sra->sys_name, restart)) {
3469 goto release;
3470 }
3471 d++;
3472 }
3473 }
3474
3475 update_cache_size(container, sra, info,
3476 min(reshape.before.data_disks,
3477 reshape.after.data_disks), blocks);
3478
3479 /* Right, everything seems fine. Let's kick things off.
3480 * If only changing raid_disks, use ioctl, else use
3481 * sysfs.
3482 */
3483 sync_metadata(st);
3484
3485 if (impose_reshape(sra, info, st, fd, restart,
3486 devname, container, &reshape) < 0)
3487 goto release;
3488
3489 err = start_reshape(sra, restart, reshape.before.data_disks,
3490 reshape.after.data_disks, st);
3491 if (err) {
3492 pr_err("Cannot %s reshape for %s\n",
3493 restart ? "continue" : "start", devname);
3494 goto release;
3495 }
3496 if (restart)
3497 sysfs_set_str(sra, NULL, "array_state", "active");
3498 if (freeze_reshape) {
3499 free(fdlist);
3500 free(offsets);
3501 sysfs_free(sra);
3502 pr_err("Reshape has to be continued from location %llu when root filesystem has been mounted.\n",
3503 sra->reshape_progress);
3504 return 1;
3505 }
3506
3507 if (!forked)
3508 if (continue_via_systemd(container ?: sra->sys_name,
3509 GROW_SERVICE)) {
3510 free(fdlist);
3511 free(offsets);
3512 sysfs_free(sra);
3513 return 0;
3514 }
3515
3516 /* Now we just need to kick off the reshape and watch, while
3517 * handling backups of the data...
3518 * This is all done by a forked background process.
3519 */
3520 switch(forked ? 0 : fork()) {
3521 case -1:
3522 pr_err("Cannot run child to monitor reshape: %s\n",
3523 strerror(errno));
3524 abort_reshape(sra);
3525 goto release;
3526 default:
3527 free(fdlist);
3528 free(offsets);
3529 sysfs_free(sra);
3530 return 0;
3531 case 0:
3532 map_fork();
3533 break;
3534 }
3535
3536 /* Close unused file descriptor in the forked process */
3537 close_fd(&fd);
3538
3539 /* If another array on the same devices is busy, the
3540 * reshape will wait for them. This would mean that
3541 * the first section that we suspend will stay suspended
3542 * for a long time. So check on that possibility
3543 * by looking for "DELAYED" in /proc/mdstat, and if found,
3544 * wait a while
3545 */
3546 do {
3547 struct mdstat_ent *mds, *m;
3548 delayed = 0;
3549 mds = mdstat_read(1, 0);
3550 for (m = mds; m; m = m->next)
3551 if (strcmp(m->devnm, sra->sys_name) == 0) {
3552 if (m->resync && m->percent == RESYNC_DELAYED)
3553 delayed = 1;
3554 if (m->resync == 0)
3555 /* Haven't started the reshape thread
3556 * yet, wait a bit
3557 */
3558 delayed = 2;
3559 break;
3560 }
3561 free_mdstat(mds);
3562 if (delayed == 1 && get_linux_version() < 3007000) {
3563 pr_err("Reshape is delayed, but cannot wait carefully with this kernel.\n"
3564 " You might experience problems until other reshapes complete.\n");
3565 delayed = 0;
3566 }
3567 if (delayed)
3568 mdstat_wait(30 - (delayed-1) * 25);
3569 } while (delayed);
3570 mdstat_close();
3571 if (check_env("MDADM_GROW_VERIFY"))
3572 fd = open(devname, O_RDONLY | O_DIRECT);
3573 else
3574 fd = -1;
3575 mlockall(MCL_FUTURE);
3576
3577 if (signal_s(SIGTERM, catch_term) == SIG_ERR)
3578 goto release;
3579
3580 if (st->ss->external) {
3581 /* metadata handler takes it from here */
3582 done = st->ss->manage_reshape(
3583 fd, sra, &reshape, st, blocks,
3584 fdlist, offsets, d - odisks, fdlist + odisks,
3585 offsets + odisks);
3586 } else
3587 done = child_monitor(
3588 fd, sra, &reshape, st, blocks, fdlist, offsets,
3589 d - odisks, fdlist + odisks, offsets + odisks);
3590
3591 free(fdlist);
3592 free(offsets);
3593
3594 if (backup_file && done) {
3595 char *bul;
3596 bul = make_backup(sra->sys_name);
3597 if (bul) {
3598 char buf[1024];
3599 int l = readlink(bul, buf, sizeof(buf) - 1);
3600 if (l > 0) {
3601 buf[l]=0;
3602 unlink(buf);
3603 }
3604 unlink(bul);
3605 free(bul);
3606 }
3607 unlink(backup_file);
3608 }
3609 if (!done) {
3610 abort_reshape(sra);
3611 goto out;
3612 }
3613
3614 if (!st->ss->external &&
3615 !(reshape.before.data_disks != reshape.after.data_disks &&
3616 info->custom_array_size) && info->new_level == reshape.level &&
3617 !forked) {
3618 /* no need to wait for the reshape to finish as
3619 * there is nothing more to do.
3620 */
3621 sysfs_free(sra);
3622 exit(0);
3623 }
3624 wait_reshape(sra);
3625
3626 if (st->ss->external) {
3627 /* Re-load the metadata as much could have changed */
3628 int cfd = open_dev(st->container_devnm);
3629 if (cfd >= 0) {
3630 flush_mdmon(container);
3631 st->ss->free_super(st);
3632 st->ss->load_container(st, cfd, container);
3633 close(cfd);
3634 }
3635 }
3636
3637 /* set new array size if required customer_array_size is used
3638 * by this metadata.
3639 */
3640 if (reshape.before.data_disks != reshape.after.data_disks &&
3641 info->custom_array_size)
3642 set_array_size(st, info, info->text_version);
3643
3644 if (info->new_level != reshape.level) {
3645 if (fd < 0)
3646 fd = open(devname, O_RDONLY);
3647 impose_level(fd, info->new_level, devname, verbose);
3648 close(fd);
3649 if (info->new_level == 0)
3650 st->update_tail = NULL;
3651 }
3652 out:
3653 sysfs_free(sra);
3654 if (forked)
3655 return 0;
3656 unfreeze(st);
3657 exit(0);
3658
3659 release:
3660 free(fdlist);
3661 free(offsets);
3662 if (orig_level != UnSet && sra) {
3663 c = map_num(pers, orig_level);
3664 if (c && sysfs_set_str(sra, NULL, "level", c) == 0)
3665 pr_err("aborting level change\n");
3666 }
3667 sysfs_free(sra);
3668 if (!forked)
3669 unfreeze(st);
3670 return 1;
3671 }
3672
3673 /* mdfd handle is passed to be closed in child process (after fork).
3674 */
3675 int reshape_container(char *container, char *devname,
3676 int mdfd,
3677 struct supertype *st,
3678 struct mdinfo *info,
3679 int force,
3680 char *backup_file, int verbose,
3681 int forked, int restart, int freeze_reshape)
3682 {
3683 struct mdinfo *cc = NULL;
3684 int rv = restart;
3685 char last_devnm[32] = "";
3686
3687 /* component_size is not meaningful for a container,
3688 * so pass '0' meaning 'no change'
3689 */
3690 if (!restart &&
3691 reshape_super(st, 0, info->new_level,
3692 info->new_layout, info->new_chunk,
3693 info->array.raid_disks, info->delta_disks,
3694 backup_file, devname, APPLY_METADATA_CHANGES,
3695 verbose)) {
3696 unfreeze(st);
3697 return 1;
3698 }
3699
3700 sync_metadata(st);
3701
3702 /* ping monitor to be sure that update is on disk
3703 */
3704 ping_monitor(container);
3705
3706 if (!forked && !freeze_reshape)
3707 if (continue_via_systemd(container, GROW_SERVICE))
3708 return 0;
3709
3710 switch (forked ? 0 : fork()) {
3711 case -1: /* error */
3712 perror("Cannot fork to complete reshape\n");
3713 unfreeze(st);
3714 return 1;
3715 default: /* parent */
3716 if (!freeze_reshape)
3717 printf("%s: multi-array reshape continues in background\n", Name);
3718 return 0;
3719 case 0: /* child */
3720 manage_fork_fds(0);
3721 map_fork();
3722 break;
3723 }
3724
3725 /* close unused handle in child process
3726 */
3727 if (mdfd > -1)
3728 close(mdfd);
3729
3730 while(1) {
3731 /* For each member array with reshape_active,
3732 * we need to perform the reshape.
3733 * We pick the first array that needs reshaping and
3734 * reshape it. reshape_array() will re-read the metadata
3735 * so the next time through a different array should be
3736 * ready for reshape.
3737 * It is possible that the 'different' array will not
3738 * be assembled yet. In that case we simple exit.
3739 * When it is assembled, the mdadm which assembles it
3740 * will take over the reshape.
3741 */
3742 struct mdinfo *content;
3743 int fd;
3744 struct mdstat_ent *mdstat;
3745 char *adev;
3746 dev_t devid;
3747
3748 sysfs_free(cc);
3749
3750 cc = st->ss->container_content(st, NULL);
3751
3752 for (content = cc; content ; content = content->next) {
3753 char *subarray;
3754 if (!content->reshape_active)
3755 continue;
3756
3757 subarray = strchr(content->text_version+1, '/')+1;
3758 mdstat = mdstat_by_subdev(subarray, container);
3759 if (!mdstat)
3760 continue;
3761 if (mdstat->active == 0) {
3762 pr_err("Skipping inactive array %s.\n",
3763 mdstat->devnm);
3764 free_mdstat(mdstat);
3765 mdstat = NULL;
3766 continue;
3767 }
3768 break;
3769 }
3770 if (!content)
3771 break;
3772
3773 devid = devnm2devid(mdstat->devnm);
3774 adev = map_dev(major(devid), minor(devid), 0);
3775 if (!adev)
3776 adev = content->text_version;
3777
3778 fd = open_dev(mdstat->devnm);
3779 if (fd < 0) {
3780 pr_err("Device %s cannot be opened for reshape.\n",
3781 adev);
3782 break;
3783 }
3784
3785 if (strcmp(last_devnm, mdstat->devnm) == 0) {
3786 /* Do not allow for multiple reshape_array() calls for
3787 * the same array.
3788 * It can happen when reshape_array() returns without
3789 * error, when reshape is not finished (wrong reshape
3790 * starting/continuation conditions). Mdmon doesn't
3791 * switch to next array in container and reentry
3792 * conditions for the same array occur.
3793 * This is possibly interim until the behaviour of
3794 * reshape_array is resolved().
3795 */
3796 printf("%s: Multiple reshape execution detected for device %s.\n", Name, adev);
3797 close(fd);
3798 break;
3799 }
3800 strcpy(last_devnm, mdstat->devnm);
3801
3802 if (sysfs_init(content, fd, mdstat->devnm)) {
3803 pr_err("Unable to initialize sysfs for %s\n",
3804 mdstat->devnm);
3805 rv = 1;
3806 break;
3807 }
3808
3809 if (mdmon_running(container))
3810 flush_mdmon(container);
3811
3812 rv = reshape_array(container, fd, adev, st,
3813 content, force, NULL, INVALID_SECTORS,
3814 backup_file, verbose, 1, restart,
3815 freeze_reshape);
3816 close(fd);
3817
3818 if (freeze_reshape) {
3819 sysfs_free(cc);
3820 exit(0);
3821 }
3822
3823 restart = 0;
3824 if (rv)
3825 break;
3826
3827 if (mdmon_running(container))
3828 flush_mdmon(container);
3829 }
3830 if (!rv)
3831 unfreeze(st);
3832 sysfs_free(cc);
3833 exit(0);
3834 }
3835
3836 /*
3837 * We run a child process in the background which performs the following
3838 * steps:
3839 * - wait for resync to reach a certain point
3840 * - suspend io to the following section
3841 * - backup that section
3842 * - allow resync to proceed further
3843 * - resume io
3844 * - discard the backup.
3845 *
3846 * When are combined in slightly different ways in the three cases.
3847 * Grow:
3848 * - suspend/backup/allow/wait/resume/discard
3849 * Shrink:
3850 * - allow/wait/suspend/backup/allow/wait/resume/discard
3851 * same-size:
3852 * - wait/resume/discard/suspend/backup/allow
3853 *
3854 * suspend/backup/allow always come together
3855 * wait/resume/discard do too.
3856 * For the same-size case we have two backups to improve flow.
3857 *
3858 */
3859
3860 int progress_reshape(struct mdinfo *info, struct reshape *reshape,
3861 unsigned long long backup_point,
3862 unsigned long long wait_point,
3863 unsigned long long *suspend_point,
3864 unsigned long long *reshape_completed, int *frozen)
3865 {
3866 /* This function is called repeatedly by the reshape manager.
3867 * It determines how much progress can safely be made and allows
3868 * that progress.
3869 * - 'info' identifies the array and particularly records in
3870 * ->reshape_progress the metadata's knowledge of progress
3871 * This is a sector offset from the start of the array
3872 * of the next array block to be relocated. This number
3873 * may increase from 0 or decrease from array_size, depending
3874 * on the type of reshape that is happening.
3875 * Note that in contrast, 'sync_completed' is a block count of the
3876 * reshape so far. It gives the distance between the start point
3877 * (head or tail of device) and the next place that data will be
3878 * written. It always increases.
3879 * - 'reshape' is the structure created by analyse_change
3880 * - 'backup_point' shows how much the metadata manager has backed-up
3881 * data. For reshapes with increasing progress, it is the next address
3882 * to be backed up, previous addresses have been backed-up. For
3883 * decreasing progress, it is the earliest address that has been
3884 * backed up - later address are also backed up.
3885 * So addresses between reshape_progress and backup_point are
3886 * backed up providing those are in the 'correct' order.
3887 * - 'wait_point' is an array address. When reshape_completed
3888 * passes this point, progress_reshape should return. It might
3889 * return earlier if it determines that ->reshape_progress needs
3890 * to be updated or further backup is needed.
3891 * - suspend_point is maintained by progress_reshape and the caller
3892 * should not touch it except to initialise to zero.
3893 * It is an array address and it only increases in 2.6.37 and earlier.
3894 * This makes it difficult to handle reducing reshapes with
3895 * external metadata.
3896 * However: it is similar to backup_point in that it records the
3897 * other end of a suspended region from reshape_progress.
3898 * it is moved to extend the region that is safe to backup and/or
3899 * reshape
3900 * - reshape_completed is read from sysfs and returned. The caller
3901 * should copy this into ->reshape_progress when it has reason to
3902 * believe that the metadata knows this, and any backup outside this
3903 * has been erased.
3904 *
3905 * Return value is:
3906 * 1 if more data from backup_point - but only as far as suspend_point,
3907 * should be backed up
3908 * 0 if things are progressing smoothly
3909 * -1 if the reshape is finished because it is all done,
3910 * -2 if the reshape is finished due to an error.
3911 */
3912
3913 int advancing = (reshape->after.data_disks
3914 >= reshape->before.data_disks);
3915 unsigned long long need_backup; /* All data between start of array and
3916 * here will at some point need to
3917 * be backed up.
3918 */
3919 unsigned long long read_offset, write_offset;
3920 unsigned long long write_range;
3921 unsigned long long max_progress, target, completed;
3922 unsigned long long array_size = (info->component_size
3923 * reshape->before.data_disks);
3924 int fd;
3925 char buf[20];
3926
3927 /* First, we unsuspend any region that is now known to be safe.
3928 * If suspend_point is on the 'wrong' side of reshape_progress, then
3929 * we don't have or need suspension at the moment. This is true for
3930 * native metadata when we don't need to back-up.
3931 */
3932 if (advancing) {
3933 if (info->reshape_progress <= *suspend_point)
3934 sysfs_set_num(info, NULL, "suspend_lo",
3935 info->reshape_progress);
3936 } else {
3937 /* Note: this won't work in 2.6.37 and before.
3938 * Something somewhere should make sure we don't need it!
3939 */
3940 if (info->reshape_progress >= *suspend_point)
3941 sysfs_set_num(info, NULL, "suspend_hi",
3942 info->reshape_progress);
3943 }
3944
3945 /* Now work out how far it is safe to progress.
3946 * If the read_offset for ->reshape_progress is less than
3947 * 'blocks' beyond the write_offset, we can only progress as far
3948 * as a backup.
3949 * Otherwise we can progress until the write_offset for the new location
3950 * reaches (within 'blocks' of) the read_offset at the current location.
3951 * However that region must be suspended unless we are using native
3952 * metadata.
3953 * If we need to suspend more, we limit it to 128M per device, which is
3954 * rather arbitrary and should be some time-based calculation.
3955 */
3956 read_offset = info->reshape_progress / reshape->before.data_disks;
3957 write_offset = info->reshape_progress / reshape->after.data_disks;
3958 write_range = info->new_chunk/512;
3959 if (reshape->before.data_disks == reshape->after.data_disks)
3960 need_backup = array_size;
3961 else
3962 need_backup = reshape->backup_blocks;
3963 if (advancing) {
3964 if (read_offset < write_offset + write_range)
3965 max_progress = backup_point;
3966 else
3967 max_progress =
3968 read_offset * reshape->after.data_disks;
3969 } else {
3970 if (read_offset > write_offset - write_range)
3971 /* Can only progress as far as has been backed up,
3972 * which must be suspended */
3973 max_progress = backup_point;
3974 else if (info->reshape_progress <= need_backup)
3975 max_progress = backup_point;
3976 else {
3977 if (info->array.major_version >= 0)
3978 /* Can progress until backup is needed */
3979 max_progress = need_backup;
3980 else {
3981 /* Can progress until metadata update is required */
3982 max_progress =
3983 read_offset * reshape->after.data_disks;
3984 /* but data must be suspended */
3985 if (max_progress < *suspend_point)
3986 max_progress = *suspend_point;
3987 }
3988 }
3989 }
3990
3991 /* We know it is safe to progress to 'max_progress' providing
3992 * it is suspended or we are using native metadata.
3993 * Consider extending suspend_point 128M per device if it
3994 * is less than 64M per device beyond reshape_progress.
3995 * But always do a multiple of 'blocks'
3996 * FIXME this is too big - it takes to long to complete
3997 * this much.
3998 */
3999 target = 64*1024*2 * min(reshape->before.data_disks,
4000 reshape->after.data_disks);
4001 target /= reshape->backup_blocks;
4002 if (target < 2)
4003 target = 2;
4004 target *= reshape->backup_blocks;
4005
4006 /* For externally managed metadata we always need to suspend IO to
4007 * the area being reshaped so we regularly push suspend_point forward.
4008 * For native metadata we only need the suspend if we are going to do
4009 * a backup.
4010 */
4011 if (advancing) {
4012 if ((need_backup > info->reshape_progress ||
4013 info->array.major_version < 0) &&
4014 *suspend_point < info->reshape_progress + target) {
4015 if (need_backup < *suspend_point + 2 * target)
4016 *suspend_point = need_backup;
4017 else if (*suspend_point + 2 * target < array_size)
4018 *suspend_point += 2 * target;
4019 else
4020 *suspend_point = array_size;
4021 sysfs_set_num(info, NULL, "suspend_hi", *suspend_point);
4022 if (max_progress > *suspend_point)
4023 max_progress = *suspend_point;
4024 }
4025 } else {
4026 if (info->array.major_version >= 0) {
4027 /* Only need to suspend when about to backup */
4028 if (info->reshape_progress < need_backup * 2 &&
4029 *suspend_point > 0) {
4030 *suspend_point = 0;
4031 sysfs_set_num(info, NULL, "suspend_lo", 0);
4032 sysfs_set_num(info, NULL, "suspend_hi",
4033 need_backup);
4034 }
4035 } else {
4036 /* Need to suspend continually */
4037 if (info->reshape_progress < *suspend_point)
4038 *suspend_point = info->reshape_progress;
4039 if (*suspend_point + target < info->reshape_progress)
4040 /* No need to move suspend region yet */;
4041 else {
4042 if (*suspend_point >= 2 * target)
4043 *suspend_point -= 2 * target;
4044 else
4045 *suspend_point = 0;
4046 sysfs_set_num(info, NULL, "suspend_lo",
4047 *suspend_point);
4048 }
4049 if (max_progress < *suspend_point)
4050 max_progress = *suspend_point;
4051 }
4052 }
4053
4054 /* now set sync_max to allow that progress. sync_max, like
4055 * sync_completed is a count of sectors written per device, so
4056 * we find the difference between max_progress and the start point,
4057 * and divide that by after.data_disks to get a sync_max
4058 * number.
4059 * At the same time we convert wait_point to a similar number
4060 * for comparing against sync_completed.
4061 */
4062 /* scale down max_progress to per_disk */
4063 max_progress /= reshape->after.data_disks;
4064 /*
4065 * Round to chunk size as some kernels give an erroneously
4066 * high number
4067 */
4068 max_progress /= info->new_chunk/512;
4069 max_progress *= info->new_chunk/512;
4070 /* And round to old chunk size as the kernel wants that */
4071 max_progress /= info->array.chunk_size/512;
4072 max_progress *= info->array.chunk_size/512;
4073 /* Limit progress to the whole device */
4074 if (max_progress > info->component_size)
4075 max_progress = info->component_size;
4076 wait_point /= reshape->after.data_disks;
4077 if (!advancing) {
4078 /* switch from 'device offset' to 'processed block count' */
4079 max_progress = info->component_size - max_progress;
4080 wait_point = info->component_size - wait_point;
4081 }
4082
4083 if (!*frozen)
4084 sysfs_set_num(info, NULL, "sync_max", max_progress);
4085
4086 /* Now wait. If we have already reached the point that we were
4087 * asked to wait to, don't wait at all, else wait for any change.
4088 * We need to select on 'sync_completed' as that is the place that
4089 * notifications happen, but we are really interested in
4090 * 'reshape_position'
4091 */
4092 fd = sysfs_get_fd(info, NULL, "sync_completed");
4093 if (fd < 0)
4094 goto check_progress;
4095
4096 if (sysfs_fd_get_ll(fd, &completed) < 0)
4097 goto check_progress;
4098
4099 while (completed < max_progress && completed < wait_point) {
4100 /* Check that sync_action is still 'reshape' to avoid
4101 * waiting forever on a dead array
4102 */
4103 char action[20];
4104 if (sysfs_get_str(info, NULL, "sync_action", action, 20) <= 0 ||
4105 strncmp(action, "reshape", 7) != 0)
4106 break;
4107 /* Some kernels reset 'sync_completed' to zero
4108 * before setting 'sync_action' to 'idle'.
4109 * So we need these extra tests.
4110 */
4111 if (completed == 0 && advancing &&
4112 strncmp(action, "idle", 4) == 0 &&
4113 info->reshape_progress > 0)
4114 break;
4115 if (completed == 0 && !advancing &&
4116 strncmp(action, "idle", 4) == 0 &&
4117 info->reshape_progress <
4118 (info->component_size * reshape->after.data_disks))
4119 break;
4120 sysfs_wait(fd, NULL);
4121 if (sysfs_fd_get_ll(fd, &completed) < 0)
4122 goto check_progress;
4123 }
4124 /* Some kernels reset 'sync_completed' to zero,
4125 * we need to have real point we are in md.
4126 * So in that case, read 'reshape_position' from sysfs.
4127 */
4128 if (completed == 0) {
4129 unsigned long long reshapep;
4130 char action[20];
4131 if (sysfs_get_str(info, NULL, "sync_action", action, 20) > 0 &&
4132 strncmp(action, "idle", 4) == 0 &&
4133 sysfs_get_ll(info, NULL,
4134 "reshape_position", &reshapep) == 0)
4135 *reshape_completed = reshapep;
4136 } else {
4137 /* some kernels can give an incorrectly high
4138 * 'completed' number, so round down */
4139 completed /= (info->new_chunk/512);
4140 completed *= (info->new_chunk/512);
4141 /* Convert 'completed' back in to a 'progress' number */
4142 completed *= reshape->after.data_disks;
4143 if (!advancing)
4144 completed = (info->component_size
4145 * reshape->after.data_disks
4146 - completed);
4147 *reshape_completed = completed;
4148 }
4149
4150 close(fd);
4151
4152 /* We return the need_backup flag. Caller will decide
4153 * how much - a multiple of ->backup_blocks up to *suspend_point
4154 */
4155 if (advancing)
4156 return need_backup > info->reshape_progress;
4157 else
4158 return need_backup >= info->reshape_progress;
4159
4160 check_progress:
4161 /* if we couldn't read a number from sync_completed, then
4162 * either the reshape did complete, or it aborted.
4163 * We can tell which by checking for 'none' in reshape_position.
4164 * If it did abort, then it might immediately restart if it
4165 * it was just a device failure that leaves us degraded but
4166 * functioning.
4167 */
4168 if (sysfs_get_str(info, NULL, "reshape_position", buf,
4169 sizeof(buf)) < 0 || strncmp(buf, "none", 4) != 0) {
4170 /* The abort might only be temporary. Wait up to 10
4171 * seconds for fd to contain a valid number again.
4172 */
4173 int wait = 10000;
4174 int rv = -2;
4175 unsigned long long new_sync_max;
4176 while (fd >= 0 && rv < 0 && wait > 0) {
4177 if (sysfs_wait(fd, &wait) != 1)
4178 break;
4179 switch (sysfs_fd_get_ll(fd, &completed)) {
4180 case 0:
4181 /* all good again */
4182 rv = 1;
4183 /* If "sync_max" is no longer max_progress
4184 * we need to freeze things
4185 */
4186 sysfs_get_ll(info, NULL, "sync_max",
4187 &new_sync_max);
4188 *frozen = (new_sync_max != max_progress);
4189 break;
4190 case -2: /* read error - abort */
4191 wait = 0;
4192 break;
4193 }
4194 }
4195 if (fd >= 0)
4196 close(fd);
4197 return rv; /* abort */
4198 } else {
4199 /* Maybe racing with array shutdown - check state */
4200 if (fd >= 0)
4201 close(fd);
4202 if (sysfs_get_str(info, NULL, "array_state", buf,
4203 sizeof(buf)) < 0 ||
4204 strncmp(buf, "inactive", 8) == 0 ||
4205 strncmp(buf, "clear",5) == 0)
4206 return -2; /* abort */
4207 return -1; /* complete */
4208 }
4209 }
4210
4211 /* FIXME return status is never checked */
4212 static int grow_backup(struct mdinfo *sra,
4213 unsigned long long offset, /* per device */
4214 unsigned long stripes, /* per device, in old chunks */
4215 int *sources, unsigned long long *offsets,
4216 int disks, int chunk, int level, int layout,
4217 int dests, int *destfd, unsigned long long *destoffsets,
4218 int part, int *degraded,
4219 char *buf)
4220 {
4221 /* Backup 'blocks' sectors at 'offset' on each device of the array,
4222 * to storage 'destfd' (offset 'destoffsets'), after first
4223 * suspending IO. Then allow resync to continue
4224 * over the suspended section.
4225 * Use part 'part' of the backup-super-block.
4226 */
4227 int odata = disks;
4228 int rv = 0;
4229 int i;
4230 unsigned long long ll;
4231 int new_degraded;
4232 //printf("offset %llu\n", offset);
4233 if (level >= 4)
4234 odata--;
4235 if (level == 6)
4236 odata--;
4237
4238 /* Check that array hasn't become degraded, else we might backup the wrong data */
4239 if (sysfs_get_ll(sra, NULL, "degraded", &ll) < 0)
4240 return -1; /* FIXME this error is ignored */
4241 new_degraded = (int)ll;
4242 if (new_degraded != *degraded) {
4243 /* check each device to ensure it is still working */
4244 struct mdinfo *sd;
4245 for (sd = sra->devs ; sd ; sd = sd->next) {
4246 if (sd->disk.state & (1<<MD_DISK_FAULTY))
4247 continue;
4248 if (sd->disk.state & (1<<MD_DISK_SYNC)) {
4249 char sbuf[100];
4250
4251 if (sysfs_get_str(sra, sd, "state",
4252 sbuf, sizeof(sbuf)) < 0 ||
4253 strstr(sbuf, "faulty") ||
4254 strstr(sbuf, "in_sync") == NULL) {
4255 /* this device is dead */
4256 sd->disk.state = (1<<MD_DISK_FAULTY);
4257 if (sd->disk.raid_disk >= 0 &&
4258 sources[sd->disk.raid_disk] >= 0) {
4259 close(sources[sd->disk.raid_disk]);
4260 sources[sd->disk.raid_disk] = -1;
4261 }
4262 }
4263 }
4264 }
4265 *degraded = new_degraded;
4266 }
4267 if (part) {
4268 bsb.arraystart2 = __cpu_to_le64(offset * odata);
4269 bsb.length2 = __cpu_to_le64(stripes * (chunk/512) * odata);
4270 } else {
4271 bsb.arraystart = __cpu_to_le64(offset * odata);
4272 bsb.length = __cpu_to_le64(stripes * (chunk/512) * odata);
4273 }
4274 if (part)
4275 bsb.magic[15] = '2';
4276 for (i = 0; i < dests; i++)
4277 if (part)
4278 lseek64(destfd[i], destoffsets[i] +
4279 __le64_to_cpu(bsb.devstart2)*512, 0);
4280 else
4281 lseek64(destfd[i], destoffsets[i], 0);
4282
4283 rv = save_stripes(sources, offsets, disks, chunk, level, layout,
4284 dests, destfd, offset * 512 * odata,
4285 stripes * chunk * odata, buf);
4286
4287 if (rv)
4288 return rv;
4289 bsb.mtime = __cpu_to_le64(time(0));
4290 for (i = 0; i < dests; i++) {
4291 bsb.devstart = __cpu_to_le64(destoffsets[i]/512);
4292
4293 bsb.sb_csum = bsb_csum((char*)&bsb,
4294 ((char*)&bsb.sb_csum)-((char*)&bsb));
4295 if (memcmp(bsb.magic, "md_backup_data-2", 16) == 0)
4296 bsb.sb_csum2 = bsb_csum((char*)&bsb,
4297 ((char*)&bsb.sb_csum2)-((char*)&bsb));
4298
4299 rv = -1;
4300 if ((unsigned long long)lseek64(destfd[i],
4301 destoffsets[i] - 4096, 0) !=
4302 destoffsets[i] - 4096)
4303 break;
4304 if (write(destfd[i], &bsb, 512) != 512)
4305 break;
4306 if (destoffsets[i] > 4096) {
4307 if ((unsigned long long)lseek64(destfd[i], destoffsets[i]+stripes*chunk*odata, 0) !=
4308 destoffsets[i]+stripes*chunk*odata)
4309 break;
4310 if (write(destfd[i], &bsb, 512) != 512)
4311 break;
4312 }
4313 fsync(destfd[i]);
4314 rv = 0;
4315 }
4316
4317 return rv;
4318 }
4319
4320 /* in 2.6.30, the value reported by sync_completed can be
4321 * less that it should be by one stripe.
4322 * This only happens when reshape hits sync_max and pauses.
4323 * So allow wait_backup to either extent sync_max further
4324 * than strictly necessary, or return before the
4325 * sync has got quite as far as we would really like.
4326 * This is what 'blocks2' is for.
4327 * The various caller give appropriate values so that
4328 * every works.
4329 */
4330 /* FIXME return value is often ignored */
4331 static int forget_backup(int dests, int *destfd,
4332 unsigned long long *destoffsets,
4333 int part)
4334 {
4335 /*
4336 * Erase backup 'part' (which is 0 or 1)
4337 */
4338 int i;
4339 int rv;
4340
4341 if (part) {
4342 bsb.arraystart2 = __cpu_to_le64(0);
4343 bsb.length2 = __cpu_to_le64(0);
4344 } else {
4345 bsb.arraystart = __cpu_to_le64(0);
4346 bsb.length = __cpu_to_le64(0);
4347 }
4348 bsb.mtime = __cpu_to_le64(time(0));
4349 rv = 0;
4350 for (i = 0; i < dests; i++) {
4351 bsb.devstart = __cpu_to_le64(destoffsets[i]/512);
4352 bsb.sb_csum = bsb_csum((char*)&bsb,
4353 ((char*)&bsb.sb_csum)-((char*)&bsb));
4354 if (memcmp(bsb.magic, "md_backup_data-2", 16) == 0)
4355 bsb.sb_csum2 = bsb_csum((char*)&bsb,
4356 ((char*)&bsb.sb_csum2)-((char*)&bsb));
4357 if ((unsigned long long)lseek64(destfd[i], destoffsets[i]-4096, 0) !=
4358 destoffsets[i]-4096)
4359 rv = -1;
4360 if (rv == 0 && write(destfd[i], &bsb, 512) != 512)
4361 rv = -1;
4362 fsync(destfd[i]);
4363 }
4364 return rv;
4365 }
4366
4367 static void fail(char *msg)
4368 {
4369 int rv;
4370 rv = (write(2, msg, strlen(msg)) != (int)strlen(msg));
4371 rv |= (write(2, "\n", 1) != 1);
4372 exit(rv ? 1 : 2);
4373 }
4374
4375 static char *abuf, *bbuf;
4376 static unsigned long long abuflen;
4377 static void validate(int afd, int bfd, unsigned long long offset)
4378 {
4379 /* check that the data in the backup against the array.
4380 * This is only used for regression testing and should not
4381 * be used while the array is active
4382 */
4383 if (afd < 0)
4384 return;
4385 lseek64(bfd, offset - 4096, 0);
4386 if (read(bfd, &bsb2, 512) != 512)
4387 fail("cannot read bsb");
4388 if (bsb2.sb_csum != bsb_csum((char*)&bsb2,
4389 ((char*)&bsb2.sb_csum)-((char*)&bsb2)))
4390 fail("first csum bad");
4391 if (memcmp(bsb2.magic, "md_backup_data", 14) != 0)
4392 fail("magic is bad");
4393 if (memcmp(bsb2.magic, "md_backup_data-2", 16) == 0 &&
4394 bsb2.sb_csum2 != bsb_csum((char*)&bsb2,
4395 ((char*)&bsb2.sb_csum2)-((char*)&bsb2)))
4396 fail("second csum bad");
4397
4398 if (__le64_to_cpu(bsb2.devstart)*512 != offset)
4399 fail("devstart is wrong");
4400
4401 if (bsb2.length) {
4402 unsigned long long len = __le64_to_cpu(bsb2.length)*512;
4403
4404 if (abuflen < len) {
4405 free(abuf);
4406 free(bbuf);
4407 abuflen = len;
4408 if (posix_memalign((void**)&abuf, 4096, abuflen) ||
4409 posix_memalign((void**)&bbuf, 4096, abuflen)) {
4410 abuflen = 0;
4411 /* just stop validating on mem-alloc failure */
4412 return;
4413 }
4414 }
4415
4416 lseek64(bfd, offset, 0);
4417 if ((unsigned long long)read(bfd, bbuf, len) != len) {
4418 //printf("len %llu\n", len);
4419 fail("read first backup failed");
4420 }
4421 lseek64(afd, __le64_to_cpu(bsb2.arraystart)*512, 0);
4422 if ((unsigned long long)read(afd, abuf, len) != len)
4423 fail("read first from array failed");
4424 if (memcmp(bbuf, abuf, len) != 0) {
4425 #if 0
4426 int i;
4427 printf("offset=%llu len=%llu\n",
4428 (unsigned long long)__le64_to_cpu(bsb2.arraystart)*512, len);
4429 for (i=0; i<len; i++)
4430 if (bbuf[i] != abuf[i]) {
4431 printf("first diff byte %d\n", i);
4432 break;
4433 }
4434 #endif
4435 fail("data1 compare failed");
4436 }
4437 }
4438 if (bsb2.length2) {
4439 unsigned long long len = __le64_to_cpu(bsb2.length2)*512;
4440
4441 if (abuflen < len) {
4442 free(abuf);
4443 free(bbuf);
4444 abuflen = len;
4445 abuf = xmalloc(abuflen);
4446 bbuf = xmalloc(abuflen);
4447 }
4448
4449 lseek64(bfd, offset+__le64_to_cpu(bsb2.devstart2)*512, 0);
4450 if ((unsigned long long)read(bfd, bbuf, len) != len)
4451 fail("read second backup failed");
4452 lseek64(afd, __le64_to_cpu(bsb2.arraystart2)*512, 0);
4453 if ((unsigned long long)read(afd, abuf, len) != len)
4454 fail("read second from array failed");
4455 if (memcmp(bbuf, abuf, len) != 0)
4456 fail("data2 compare failed");
4457 }
4458 }
4459
4460 int child_monitor(int afd, struct mdinfo *sra, struct reshape *reshape,
4461 struct supertype *st, unsigned long blocks,
4462 int *fds, unsigned long long *offsets,
4463 int dests, int *destfd, unsigned long long *destoffsets)
4464 {
4465 /* Monitor a reshape where backup is being performed using
4466 * 'native' mechanism - either to a backup file, or
4467 * to some space in a spare.
4468 */
4469 char *buf;
4470 int degraded = -1;
4471 unsigned long long speed;
4472 unsigned long long suspend_point, array_size;
4473 unsigned long long backup_point, wait_point;
4474 unsigned long long reshape_completed;
4475 int done = 0;
4476 int increasing = reshape->after.data_disks >=
4477 reshape->before.data_disks;
4478 int part = 0; /* The next part of the backup area to fill. It
4479 * may already be full, so we need to check */
4480 int level = reshape->level;
4481 int layout = reshape->before.layout;
4482 int data = reshape->before.data_disks;
4483 int disks = reshape->before.data_disks + reshape->parity;
4484 int chunk = sra->array.chunk_size;
4485 struct mdinfo *sd;
4486 unsigned long stripes;
4487 int uuid[4];
4488 int frozen = 0;
4489
4490 /* set up the backup-super-block. This requires the
4491 * uuid from the array.
4492 */
4493 /* Find a superblock */
4494 for (sd = sra->devs; sd; sd = sd->next) {
4495 char *dn;
4496 int devfd;
4497 int ok;
4498 if (sd->disk.state & (1<<MD_DISK_FAULTY))
4499 continue;
4500 dn = map_dev(sd->disk.major, sd->disk.minor, 1);
4501 devfd = dev_open(dn, O_RDONLY);
4502 if (devfd < 0)
4503 continue;
4504 ok = st->ss->load_super(st, devfd, NULL);
4505 close(devfd);
4506 if (ok == 0)
4507 break;
4508 }
4509 if (!sd) {
4510 pr_err("Cannot find a superblock\n");
4511 return 0;
4512 }
4513
4514 memset(&bsb, 0, 512);
4515 memcpy(bsb.magic, "md_backup_data-1", 16);
4516 st->ss->uuid_from_super(st, uuid);
4517 memcpy(bsb.set_uuid, uuid, 16);
4518 bsb.mtime = __cpu_to_le64(time(0));
4519 bsb.devstart2 = blocks;
4520
4521 stripes = blocks / (sra->array.chunk_size/512) /
4522 reshape->before.data_disks;
4523
4524 if (posix_memalign((void**)&buf, 4096, disks * chunk))
4525 /* Don't start the 'reshape' */
4526 return 0;
4527 if (reshape->before.data_disks == reshape->after.data_disks) {
4528 sysfs_get_ll(sra, NULL, "sync_speed_min", &speed);
4529 sysfs_set_num(sra, NULL, "sync_speed_min", 200000);
4530 }
4531
4532 if (increasing) {
4533 array_size = sra->component_size * reshape->after.data_disks;
4534 backup_point = sra->reshape_progress;
4535 suspend_point = 0;
4536 } else {
4537 array_size = sra->component_size * reshape->before.data_disks;
4538 backup_point = reshape->backup_blocks;
4539 suspend_point = array_size;
4540 }
4541
4542 while (!done) {
4543 int rv;
4544
4545 /* Want to return as soon the oldest backup slot can
4546 * be released as that allows us to start backing up
4547 * some more, providing suspend_point has been
4548 * advanced, which it should have.
4549 */
4550 if (increasing) {
4551 wait_point = array_size;
4552 if (part == 0 && __le64_to_cpu(bsb.length) > 0)
4553 wait_point = (__le64_to_cpu(bsb.arraystart) +
4554 __le64_to_cpu(bsb.length));
4555 if (part == 1 && __le64_to_cpu(bsb.length2) > 0)
4556 wait_point = (__le64_to_cpu(bsb.arraystart2) +
4557 __le64_to_cpu(bsb.length2));
4558 } else {
4559 wait_point = 0;
4560 if (part == 0 && __le64_to_cpu(bsb.length) > 0)
4561 wait_point = __le64_to_cpu(bsb.arraystart);
4562 if (part == 1 && __le64_to_cpu(bsb.length2) > 0)
4563 wait_point = __le64_to_cpu(bsb.arraystart2);
4564 }
4565
4566 reshape_completed = sra->reshape_progress;
4567 rv = progress_reshape(sra, reshape,
4568 backup_point, wait_point,
4569 &suspend_point, &reshape_completed,
4570 &frozen);
4571 /* external metadata would need to ping_monitor here */
4572 sra->reshape_progress = reshape_completed;
4573
4574 /* Clear any backup region that is before 'here' */
4575 if (increasing) {
4576 if (__le64_to_cpu(bsb.length) > 0 &&
4577 reshape_completed >= (__le64_to_cpu(bsb.arraystart) +
4578 __le64_to_cpu(bsb.length)))
4579 forget_backup(dests, destfd,
4580 destoffsets, 0);
4581 if (__le64_to_cpu(bsb.length2) > 0 &&
4582 reshape_completed >= (__le64_to_cpu(bsb.arraystart2) +
4583 __le64_to_cpu(bsb.length2)))
4584 forget_backup(dests, destfd,
4585 destoffsets, 1);
4586 } else {
4587 if (__le64_to_cpu(bsb.length) > 0 &&
4588 reshape_completed <= (__le64_to_cpu(bsb.arraystart)))
4589 forget_backup(dests, destfd,
4590 destoffsets, 0);
4591 if (__le64_to_cpu(bsb.length2) > 0 &&
4592 reshape_completed <= (__le64_to_cpu(bsb.arraystart2)))
4593 forget_backup(dests, destfd,
4594 destoffsets, 1);
4595 }
4596 if (sigterm)
4597 rv = -2;
4598 if (rv < 0) {
4599 if (rv == -1)
4600 done = 1;
4601 break;
4602 }
4603 if (rv == 0 && increasing && !st->ss->external) {
4604 /* No longer need to monitor this reshape */
4605 sysfs_set_str(sra, NULL, "sync_max", "max");
4606 done = 1;
4607 break;
4608 }
4609
4610 while (rv) {
4611 unsigned long long offset;
4612 unsigned long actual_stripes;
4613 /* Need to backup some data.
4614 * If 'part' is not used and the desired
4615 * backup size is suspended, do a backup,
4616 * then consider the next part.
4617 */
4618 /* Check that 'part' is unused */
4619 if (part == 0 && __le64_to_cpu(bsb.length) != 0)
4620 break;
4621 if (part == 1 && __le64_to_cpu(bsb.length2) != 0)
4622 break;
4623
4624 offset = backup_point / data;
4625 actual_stripes = stripes;
4626 if (increasing) {
4627 if (offset + actual_stripes * (chunk/512) >
4628 sra->component_size)
4629 actual_stripes = ((sra->component_size - offset)
4630 / (chunk/512));
4631 if (offset + actual_stripes * (chunk/512) >
4632 suspend_point/data)
4633 break;
4634 } else {
4635 if (offset < actual_stripes * (chunk/512))
4636 actual_stripes = offset / (chunk/512);
4637 offset -= actual_stripes * (chunk/512);
4638 if (offset < suspend_point/data)
4639 break;
4640 }
4641 if (actual_stripes == 0)
4642 break;
4643 grow_backup(sra, offset, actual_stripes, fds, offsets,
4644 disks, chunk, level, layout, dests, destfd,
4645 destoffsets, part, &degraded, buf);
4646 validate(afd, destfd[0], destoffsets[0]);
4647 /* record where 'part' is up to */
4648 part = !part;
4649 if (increasing)
4650 backup_point += actual_stripes * (chunk/512) * data;
4651 else
4652 backup_point -= actual_stripes * (chunk/512) * data;
4653 }
4654 }
4655
4656 /* FIXME maybe call progress_reshape one more time instead */
4657 /* remove any remaining suspension */
4658 sysfs_set_num(sra, NULL, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL);
4659 sysfs_set_num(sra, NULL, "suspend_hi", 0);
4660 sysfs_set_num(sra, NULL, "suspend_lo", 0);
4661 sysfs_set_num(sra, NULL, "sync_min", 0);
4662
4663 if (reshape->before.data_disks == reshape->after.data_disks)
4664 sysfs_set_num(sra, NULL, "sync_speed_min", speed);
4665 free(buf);
4666 return done;
4667 }
4668
4669 /*
4670 * If any spare contains md_back_data-1 which is recent wrt mtime,
4671 * write that data into the array and update the super blocks with
4672 * the new reshape_progress
4673 */
4674 int Grow_restart(struct supertype *st, struct mdinfo *info, int *fdlist,
4675 int cnt, char *backup_file, int verbose)
4676 {
4677 int i, j;
4678 int old_disks;
4679 unsigned long long *offsets;
4680 unsigned long long nstripe, ostripe;
4681 int ndata, odata;
4682
4683 odata = info->array.raid_disks - info->delta_disks - 1;
4684 if (info->array.level == 6)
4685 odata--; /* number of data disks */
4686 ndata = info->array.raid_disks - 1;
4687 if (info->new_level == 6)
4688 ndata--;
4689
4690 old_disks = info->array.raid_disks - info->delta_disks;
4691
4692 if (info->delta_disks <= 0)
4693 /* Didn't grow, so the backup file must have
4694 * been used
4695 */
4696 old_disks = cnt;
4697 for (i=old_disks-(backup_file?1:0); i<cnt; i++) {
4698 struct mdinfo dinfo;
4699 int fd;
4700 int bsbsize;
4701 char *devname, namebuf[20];
4702 unsigned long long lo, hi;
4703
4704 /* This was a spare and may have some saved data on it.
4705 * Load the superblock, find and load the
4706 * backup_super_block.
4707 * If either fail, go on to next device.
4708 * If the backup contains no new info, just return
4709 * else restore data and update all superblocks
4710 */
4711 if (i == old_disks-1) {
4712 fd = open(backup_file, O_RDONLY);
4713 if (fd<0) {
4714 pr_err("backup file %s inaccessible: %s\n",
4715 backup_file, strerror(errno));
4716 continue;
4717 }
4718 devname = backup_file;
4719 } else {
4720 fd = fdlist[i];
4721 if (fd < 0)
4722 continue;
4723 if (st->ss->load_super(st, fd, NULL))
4724 continue;
4725
4726 st->ss->getinfo_super(st, &dinfo, NULL);
4727 st->ss->free_super(st);
4728
4729 if (lseek64(fd,
4730 (dinfo.data_offset + dinfo.component_size - 8) <<9,
4731 0) < 0) {
4732 pr_err("Cannot seek on device %d\n", i);
4733 continue; /* Cannot seek */
4734 }
4735 sprintf(namebuf, "device-%d", i);
4736 devname = namebuf;
4737 }
4738 if (read(fd, &bsb, sizeof(bsb)) != sizeof(bsb)) {
4739 if (verbose)
4740 pr_err("Cannot read from %s\n", devname);
4741 continue; /* Cannot read */
4742 }
4743 if (memcmp(bsb.magic, "md_backup_data-1", 16) != 0 &&
4744 memcmp(bsb.magic, "md_backup_data-2", 16) != 0) {
4745 if (verbose)
4746 pr_err("No backup metadata on %s\n", devname);
4747 continue;
4748 }
4749 if (bsb.sb_csum != bsb_csum((char*)&bsb, ((char*)&bsb.sb_csum)-((char*)&bsb))) {
4750 if (verbose)
4751 pr_err("Bad backup-metadata checksum on %s\n",
4752 devname);
4753 continue; /* bad checksum */
4754 }
4755 if (memcmp(bsb.magic, "md_backup_data-2", 16) == 0 &&
4756 bsb.sb_csum2 != bsb_csum((char*)&bsb, ((char*)&bsb.sb_csum2)-((char*)&bsb))) {
4757 if (verbose)
4758 pr_err("Bad backup-metadata checksum2 on %s\n",
4759 devname);
4760 continue; /* Bad second checksum */
4761 }
4762 if (memcmp(bsb.set_uuid,info->uuid, 16) != 0) {
4763 if (verbose)
4764 pr_err("Wrong uuid on backup-metadata on %s\n",
4765 devname);
4766 continue; /* Wrong uuid */
4767 }
4768
4769 /*
4770 * array utime and backup-mtime should be updated at
4771 * much the same time, but it seems that sometimes
4772 * they aren't... So allow considerable flexability in
4773 * matching, and allow this test to be overridden by
4774 * an environment variable.
4775 */
4776 if(time_after(info->array.utime, (unsigned int)__le64_to_cpu(bsb.mtime) + 2*60*60) ||
4777 time_before(info->array.utime, (unsigned int)__le64_to_cpu(bsb.mtime) - 10*60)) {
4778 if (check_env("MDADM_GROW_ALLOW_OLD")) {
4779 pr_err("accepting backup with timestamp %lu for array with timestamp %lu\n",
4780 (unsigned long)__le64_to_cpu(bsb.mtime),
4781 (unsigned long)info->array.utime);
4782 } else {
4783 pr_err("too-old timestamp on backup-metadata on %s\n", devname);
4784 pr_err("If you think it is should be safe, try 'export MDADM_GROW_ALLOW_OLD=1'\n");
4785 continue; /* time stamp is too bad */
4786 }
4787 }
4788
4789 if (bsb.magic[15] == '1') {
4790 if (bsb.length == 0)
4791 continue;
4792 if (info->delta_disks >= 0) {
4793 /* reshape_progress is increasing */
4794 if (__le64_to_cpu(bsb.arraystart)
4795 + __le64_to_cpu(bsb.length)
4796 < info->reshape_progress) {
4797 nonew:
4798 if (verbose)
4799 pr_err("backup-metadata found on %s but is not needed\n", devname);
4800 continue; /* No new data here */
4801 }
4802 } else {
4803 /* reshape_progress is decreasing */
4804 if (__le64_to_cpu(bsb.arraystart) >=
4805 info->reshape_progress)
4806 goto nonew; /* No new data here */
4807 }
4808 } else {
4809 if (bsb.length == 0 && bsb.length2 == 0)
4810 continue;
4811 if (info->delta_disks >= 0) {
4812 /* reshape_progress is increasing */
4813 if ((__le64_to_cpu(bsb.arraystart)
4814 + __le64_to_cpu(bsb.length)
4815 < info->reshape_progress) &&
4816 (__le64_to_cpu(bsb.arraystart2)
4817 + __le64_to_cpu(bsb.length2)
4818 < info->reshape_progress))
4819 goto nonew; /* No new data here */
4820 } else {
4821 /* reshape_progress is decreasing */
4822 if (__le64_to_cpu(bsb.arraystart) >=
4823 info->reshape_progress &&
4824 __le64_to_cpu(bsb.arraystart2) >=
4825 info->reshape_progress)
4826 goto nonew; /* No new data here */
4827 }
4828 }
4829 if (lseek64(fd, __le64_to_cpu(bsb.devstart)*512, 0)< 0) {
4830 second_fail:
4831 if (verbose)
4832 pr_err("Failed to verify secondary backup-metadata block on %s\n",
4833 devname);
4834 continue; /* Cannot seek */
4835 }
4836 /* There should be a duplicate backup superblock 4k before here */
4837 if (lseek64(fd, -4096, 1) < 0 ||
4838 read(fd, &bsb2, sizeof(bsb2)) != sizeof(bsb2))
4839 goto second_fail; /* Cannot find leading superblock */
4840 if (bsb.magic[15] == '1')
4841 bsbsize = offsetof(struct mdp_backup_super, pad1);
4842 else
4843 bsbsize = offsetof(struct mdp_backup_super, pad);
4844 if (memcmp(&bsb2, &bsb, bsbsize) != 0)
4845 goto second_fail; /* Cannot find leading superblock */
4846
4847 /* Now need the data offsets for all devices. */
4848 offsets = xmalloc(sizeof(*offsets)*info->array.raid_disks);
4849 for(j=0; j<info->array.raid_disks; j++) {
4850 if (fdlist[j] < 0)
4851 continue;
4852 if (st->ss->load_super(st, fdlist[j], NULL))
4853 /* FIXME should be this be an error */
4854 continue;
4855 st->ss->getinfo_super(st, &dinfo, NULL);
4856 st->ss->free_super(st);
4857 offsets[j] = dinfo.data_offset * 512;
4858 }
4859 printf("%s: restoring critical section\n", Name);
4860
4861 if (restore_stripes(fdlist, offsets, info->array.raid_disks,
4862 info->new_chunk, info->new_level,
4863 info->new_layout, fd,
4864 __le64_to_cpu(bsb.devstart)*512,
4865 __le64_to_cpu(bsb.arraystart)*512,
4866 __le64_to_cpu(bsb.length)*512, NULL)) {
4867 /* didn't succeed, so giveup */
4868 if (verbose)
4869 pr_err("Error restoring backup from %s\n",
4870 devname);
4871 free(offsets);
4872 return 1;
4873 }
4874
4875 if (bsb.magic[15] == '2' &&
4876 restore_stripes(fdlist, offsets, info->array.raid_disks,
4877 info->new_chunk, info->new_level,
4878 info->new_layout, fd,
4879 __le64_to_cpu(bsb.devstart)*512 +
4880 __le64_to_cpu(bsb.devstart2)*512,
4881 __le64_to_cpu(bsb.arraystart2)*512,
4882 __le64_to_cpu(bsb.length2)*512, NULL)) {
4883 /* didn't succeed, so giveup */
4884 if (verbose)
4885 pr_err("Error restoring second backup from %s\n",
4886 devname);
4887 free(offsets);
4888 return 1;
4889 }
4890
4891 free(offsets);
4892
4893 /* Ok, so the data is restored. Let's update those superblocks. */
4894
4895 lo = hi = 0;
4896 if (bsb.length) {
4897 lo = __le64_to_cpu(bsb.arraystart);
4898 hi = lo + __le64_to_cpu(bsb.length);
4899 }
4900 if (bsb.magic[15] == '2' && bsb.length2) {
4901 unsigned long long lo1, hi1;
4902 lo1 = __le64_to_cpu(bsb.arraystart2);
4903 hi1 = lo1 + __le64_to_cpu(bsb.length2);
4904 if (lo == hi) {
4905 lo = lo1;
4906 hi = hi1;
4907 } else if (lo < lo1)
4908 hi = hi1;
4909 else
4910 lo = lo1;
4911 }
4912 if (lo < hi && (info->reshape_progress < lo ||
4913 info->reshape_progress > hi))
4914 /* backup does not affect reshape_progress*/ ;
4915 else if (info->delta_disks >= 0) {
4916 info->reshape_progress = __le64_to_cpu(bsb.arraystart) +
4917 __le64_to_cpu(bsb.length);
4918 if (bsb.magic[15] == '2') {
4919 unsigned long long p2;
4920
4921 p2 = __le64_to_cpu(bsb.arraystart2) +
4922 __le64_to_cpu(bsb.length2);
4923 if (p2 > info->reshape_progress)
4924 info->reshape_progress = p2;
4925 }
4926 } else {
4927 info->reshape_progress = __le64_to_cpu(bsb.arraystart);
4928 if (bsb.magic[15] == '2') {
4929 unsigned long long p2;
4930
4931 p2 = __le64_to_cpu(bsb.arraystart2);
4932 if (p2 < info->reshape_progress)
4933 info->reshape_progress = p2;
4934 }
4935 }
4936 for (j=0; j<info->array.raid_disks; j++) {
4937 if (fdlist[j] < 0)
4938 continue;
4939 if (st->ss->load_super(st, fdlist[j], NULL))
4940 continue;
4941 st->ss->getinfo_super(st, &dinfo, NULL);
4942 dinfo.reshape_progress = info->reshape_progress;
4943 st->ss->update_super(st, &dinfo, "_reshape_progress",
4944 NULL,0, 0, NULL);
4945 st->ss->store_super(st, fdlist[j]);
4946 st->ss->free_super(st);
4947 }
4948 return 0;
4949 }
4950 /* Didn't find any backup data, try to see if any
4951 * was needed.
4952 */
4953 if (info->delta_disks < 0) {
4954 /* When shrinking, the critical section is at the end.
4955 * So see if we are before the critical section.
4956 */
4957 unsigned long long first_block;
4958 nstripe = ostripe = 0;
4959 first_block = 0;
4960 while (ostripe >= nstripe) {
4961 ostripe += info->array.chunk_size / 512;
4962 first_block = ostripe * odata;
4963 nstripe = first_block / ndata / (info->new_chunk/512) *
4964 (info->new_chunk/512);
4965 }
4966
4967 if (info->reshape_progress >= first_block)
4968 return 0;
4969 }
4970 if (info->delta_disks > 0) {
4971 /* See if we are beyond the critical section. */
4972 unsigned long long last_block;
4973 nstripe = ostripe = 0;
4974 last_block = 0;
4975 while (nstripe >= ostripe) {
4976 nstripe += info->new_chunk / 512;
4977 last_block = nstripe * ndata;
4978 ostripe = last_block / odata / (info->array.chunk_size/512) *
4979 (info->array.chunk_size/512);
4980 }
4981
4982 if (info->reshape_progress >= last_block)
4983 return 0;
4984 }
4985 /* needed to recover critical section! */
4986 if (verbose)
4987 pr_err("Failed to find backup of critical section\n");
4988 return 1;
4989 }
4990
4991 int Grow_continue_command(char *devname, int fd,
4992 char *backup_file, int verbose)
4993 {
4994 int ret_val = 0;
4995 struct supertype *st = NULL;
4996 struct mdinfo *content = NULL;
4997 struct mdinfo array;
4998 char *subarray = NULL;
4999 struct mdinfo *cc = NULL;
5000 struct mdstat_ent *mdstat = NULL;
5001 int cfd = -1;
5002 int fd2;
5003
5004 dprintf("Grow continue from command line called for %s\n", devname);
5005
5006 st = super_by_fd(fd, &subarray);
5007 if (!st || !st->ss) {
5008 pr_err("Unable to determine metadata format for %s\n", devname);
5009 return 1;
5010 }
5011 dprintf("Grow continue is run for ");
5012 if (st->ss->external == 0) {
5013 int d;
5014 int cnt = 5;
5015 dprintf_cont("native array (%s)\n", devname);
5016 if (md_get_array_info(fd, &array.array) < 0) {
5017 pr_err("%s is not an active md array - aborting\n",
5018 devname);
5019 ret_val = 1;
5020 goto Grow_continue_command_exit;
5021 }
5022 content = &array;
5023 sysfs_init(content, fd, NULL);
5024 /* Need to load a superblock.
5025 * FIXME we should really get what we need from
5026 * sysfs
5027 */
5028 do {
5029 for (d = 0; d < MAX_DISKS; d++) {
5030 mdu_disk_info_t disk;
5031 char *dv;
5032 int err;
5033 disk.number = d;
5034 if (md_get_disk_info(fd, &disk) < 0)
5035 continue;
5036 if (disk.major == 0 && disk.minor == 0)
5037 continue;
5038 if ((disk.state & (1 << MD_DISK_ACTIVE)) == 0)
5039 continue;
5040 dv = map_dev(disk.major, disk.minor, 1);
5041 if (!dv)
5042 continue;
5043 fd2 = dev_open(dv, O_RDONLY);
5044 if (fd2 < 0)
5045 continue;
5046 err = st->ss->load_super(st, fd2, NULL);
5047 close(fd2);
5048 if (err)
5049 continue;
5050 break;
5051 }
5052 if (d == MAX_DISKS) {
5053 pr_err("Unable to load metadata for %s\n",
5054 devname);
5055 ret_val = 1;
5056 goto Grow_continue_command_exit;
5057 }
5058 st->ss->getinfo_super(st, content, NULL);
5059 if (!content->reshape_active)
5060 sleep_for(3, 0, true);
5061 else
5062 break;
5063 } while (cnt-- > 0);
5064 } else {
5065 char *container;
5066
5067 if (subarray) {
5068 dprintf_cont("subarray (%s)\n", subarray);
5069 container = st->container_devnm;
5070 cfd = open_dev_excl(st->container_devnm);
5071 } else {
5072 container = st->devnm;
5073 close(fd);
5074 cfd = open_dev_excl(st->devnm);
5075 dprintf_cont("container (%s)\n", container);
5076 fd = cfd;
5077 }
5078 if (cfd < 0) {
5079 pr_err("Unable to open container for %s\n", devname);
5080 ret_val = 1;
5081 goto Grow_continue_command_exit;
5082 }
5083
5084 /* find in container array under reshape
5085 */
5086 ret_val = st->ss->load_container(st, cfd, NULL);
5087 if (ret_val) {
5088 pr_err("Cannot read superblock for %s\n", devname);
5089 ret_val = 1;
5090 goto Grow_continue_command_exit;
5091 }
5092
5093 cc = st->ss->container_content(st, subarray);
5094 for (content = cc; content ; content = content->next) {
5095 char *array_name;
5096 int allow_reshape = 1;
5097
5098 if (content->reshape_active == 0)
5099 continue;
5100 /* The decision about array or container wide
5101 * reshape is taken in Grow_continue based
5102 * content->reshape_active state, therefore we
5103 * need to check_reshape based on
5104 * reshape_active and subarray name
5105 */
5106 if (content->array.state & (1<<MD_SB_BLOCK_VOLUME))
5107 allow_reshape = 0;
5108 if (content->reshape_active == CONTAINER_RESHAPE &&
5109 (content->array.state
5110 & (1<<MD_SB_BLOCK_CONTAINER_RESHAPE)))
5111 allow_reshape = 0;
5112
5113 if (!allow_reshape) {
5114 pr_err("cannot continue reshape of an array in container with unsupported metadata: %s(%s)\n",
5115 devname, container);
5116 ret_val = 1;
5117 goto Grow_continue_command_exit;
5118 }
5119
5120 array_name = strchr(content->text_version+1, '/')+1;
5121 mdstat = mdstat_by_subdev(array_name, container);
5122 if (!mdstat)
5123 continue;
5124 if (mdstat->active == 0) {
5125 pr_err("Skipping inactive array %s.\n",
5126 mdstat->devnm);
5127 free_mdstat(mdstat);
5128 mdstat = NULL;
5129 continue;
5130 }
5131 break;
5132 }
5133 if (!content) {
5134 pr_err("Unable to determine reshaped array for %s\n", devname);
5135 ret_val = 1;
5136 goto Grow_continue_command_exit;
5137 }
5138 fd2 = open_dev(mdstat->devnm);
5139 if (fd2 < 0) {
5140 pr_err("cannot open (%s)\n", mdstat->devnm);
5141 ret_val = 1;
5142 goto Grow_continue_command_exit;
5143 }
5144
5145 if (sysfs_init(content, fd2, mdstat->devnm)) {
5146 pr_err("Unable to initialize sysfs for %s, Grow cannot continue.\n",
5147 mdstat->devnm);
5148 ret_val = 1;
5149 close(fd2);
5150 goto Grow_continue_command_exit;
5151 }
5152
5153 close(fd2);
5154
5155 /* start mdmon in case it is not running
5156 */
5157 if (!mdmon_running(container))
5158 start_mdmon(container);
5159 ping_monitor(container);
5160
5161 if (mdmon_running(container))
5162 st->update_tail = &st->updates;
5163 else {
5164 pr_err("No mdmon found. Grow cannot continue.\n");
5165 ret_val = 1;
5166 goto Grow_continue_command_exit;
5167 }
5168 }
5169
5170 /* verify that array under reshape is started from
5171 * correct position
5172 */
5173 if (verify_reshape_position(content, content->array.level) < 0) {
5174 ret_val = 1;
5175 goto Grow_continue_command_exit;
5176 }
5177
5178 /* continue reshape
5179 */
5180 ret_val = Grow_continue(fd, st, content, backup_file, 1, 0);
5181
5182 Grow_continue_command_exit:
5183 if (cfd > -1)
5184 close(cfd);
5185 st->ss->free_super(st);
5186 free_mdstat(mdstat);
5187 sysfs_free(cc);
5188 free(subarray);
5189
5190 return ret_val;
5191 }
5192
5193 int Grow_continue(int mdfd, struct supertype *st, struct mdinfo *info,
5194 char *backup_file, int forked, int freeze_reshape)
5195 {
5196 int ret_val = 2;
5197
5198 if (!info->reshape_active)
5199 return ret_val;
5200
5201 if (st->ss->external) {
5202 int cfd = open_dev(st->container_devnm);
5203
5204 if (cfd < 0)
5205 return 1;
5206
5207 st->ss->load_container(st, cfd, st->container_devnm);
5208 close(cfd);
5209 ret_val = reshape_container(st->container_devnm, NULL, mdfd,
5210 st, info, 0, backup_file, 0,
5211 forked, 1 | info->reshape_active,
5212 freeze_reshape);
5213 } else
5214 ret_val = reshape_array(NULL, mdfd, "array", st, info, 1,
5215 NULL, INVALID_SECTORS, backup_file,
5216 0, forked, 1 | info->reshape_active,
5217 freeze_reshape);
5218
5219 return ret_val;
5220 }
5221
5222 char *make_backup(char *name)
5223 {
5224 char *base = "backup_file-";
5225 int len;
5226 char *fname;
5227
5228 len = strlen(MAP_DIR) + 1 + strlen(base) + strlen(name)+1;
5229 fname = xmalloc(len);
5230 sprintf(fname, "%s/%s%s", MAP_DIR, base, name);
5231 return fname;
5232 }
5233
5234 char *locate_backup(char *name)
5235 {
5236 char *fl = make_backup(name);
5237 struct stat stb;
5238
5239 if (stat(fl, &stb) == 0 && S_ISREG(stb.st_mode))
5240 return fl;
5241
5242 free(fl);
5243 return NULL;
5244 }
Unnamed repository; edit this file 'description' to name the repository.