]> git.ipfire.org Git - thirdparty/mdadm.git/blob - Grow.c
projects / thirdparty / mdadm.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Grow: pass INVALID_SECTORS to reshape_array, not 0.
[thirdparty/mdadm.git] / Grow.c
1 /*
2 * mdadm - manage Linux "md" devices aka RAID arrays.
3 *
4 * Copyright (C) 2001-2012 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 <stdint.h>
28
29 #if ! defined(__BIG_ENDIAN) && ! defined(__LITTLE_ENDIAN)
30 #error no endian defined
31 #endif
32 #include "md_u.h"
33 #include "md_p.h"
34
35 #ifndef offsetof
36 #define offsetof(t,f) ((size_t)&(((t*)0)->f))
37 #endif
38
39 int restore_backup(struct supertype *st,
40 struct mdinfo *content,
41 int working_disks,
42 int next_spare,
43 char *backup_file,
44 int verbose)
45 {
46 int i;
47 int *fdlist;
48 struct mdinfo *dev;
49 int err;
50 int disk_count = next_spare + working_disks;
51
52 dprintf("Called restore_backup()\n");
53 fdlist = xmalloc(sizeof(int) * disk_count);
54
55 enable_fds(next_spare);
56 for (i = 0; i < next_spare; i++)
57 fdlist[i] = -1;
58 for (dev = content->devs; dev; dev = dev->next) {
59 char buf[22];
60 int fd;
61 sprintf(buf, "%d:%d",
62 dev->disk.major,
63 dev->disk.minor);
64 fd = dev_open(buf, O_RDWR);
65
66 if (dev->disk.raid_disk >= 0)
67 fdlist[dev->disk.raid_disk] = fd;
68 else
69 fdlist[next_spare++] = fd;
70 }
71
72 if (st->ss->external && st->ss->recover_backup)
73 err = st->ss->recover_backup(st, content);
74 else
75 err = Grow_restart(st, content, fdlist, next_spare,
76 backup_file, verbose > 0);
77
78 while (next_spare > 0) {
79 next_spare--;
80 if (fdlist[next_spare] >= 0)
81 close(fdlist[next_spare]);
82 }
83 free(fdlist);
84 if (err) {
85 pr_err("Failed to restore critical"
86 " section for reshape - sorry.\n");
87 if (!backup_file)
88 pr_err("Possibly you need"
89 " to specify a --backup-file\n");
90 return 1;
91 }
92
93 dprintf("restore_backup() returns status OK.\n");
94 return 0;
95 }
96
97 int Grow_Add_device(char *devname, int fd, char *newdev)
98 {
99 /* Add a device to an active array.
100 * Currently, just extend a linear array.
101 * This requires writing a new superblock on the
102 * new device, calling the kernel to add the device,
103 * and if that succeeds, update the superblock on
104 * all other devices.
105 * This means that we need to *find* all other devices.
106 */
107 struct mdinfo info;
108
109 struct stat stb;
110 int nfd, fd2;
111 int d, nd;
112 struct supertype *st = NULL;
113 char *subarray = NULL;
114
115 if (ioctl(fd, GET_ARRAY_INFO, &info.array) < 0) {
116 pr_err("cannot get array info for %s\n", devname);
117 return 1;
118 }
119
120 if (info.array.level != -1) {
121 pr_err("can only add devices to linear arrays\n");
122 return 1;
123 }
124
125 st = super_by_fd(fd, &subarray);
126 if (!st) {
127 pr_err("cannot handle arrays with superblock version %d\n",
128 info.array.major_version);
129 return 1;
130 }
131
132 if (subarray) {
133 pr_err("Cannot grow linear sub-arrays yet\n");
134 free(subarray);
135 free(st);
136 return 1;
137 }
138
139 nfd = open(newdev, O_RDWR|O_EXCL|O_DIRECT);
140 if (nfd < 0) {
141 pr_err("cannot open %s\n", newdev);
142 free(st);
143 return 1;
144 }
145 fstat(nfd, &stb);
146 if ((stb.st_mode & S_IFMT) != S_IFBLK) {
147 pr_err("%s is not a block device!\n", newdev);
148 close(nfd);
149 free(st);
150 return 1;
151 }
152 /* now check out all the devices and make sure we can read the
153 * superblock */
154 for (d=0 ; d < info.array.raid_disks ; d++) {
155 mdu_disk_info_t disk;
156 char *dv;
157
158 st->ss->free_super(st);
159
160 disk.number = d;
161 if (ioctl(fd, GET_DISK_INFO, &disk) < 0) {
162 pr_err("cannot get device detail for device %d\n",
163 d);
164 close(nfd);
165 free(st);
166 return 1;
167 }
168 dv = map_dev(disk.major, disk.minor, 1);
169 if (!dv) {
170 pr_err("cannot find device file for device %d\n",
171 d);
172 close(nfd);
173 free(st);
174 return 1;
175 }
176 fd2 = dev_open(dv, O_RDWR);
177 if (fd2 < 0) {
178 pr_err("cannot open device file %s\n", dv);
179 close(nfd);
180 free(st);
181 return 1;
182 }
183
184 if (st->ss->load_super(st, fd2, NULL)) {
185 pr_err("cannot find super block on %s\n", dv);
186 close(nfd);
187 close(fd2);
188 free(st);
189 return 1;
190 }
191 close(fd2);
192 }
193 /* Ok, looks good. Lets update the superblock and write it out to
194 * newdev.
195 */
196
197 info.disk.number = d;
198 info.disk.major = major(stb.st_rdev);
199 info.disk.minor = minor(stb.st_rdev);
200 info.disk.raid_disk = d;
201 info.disk.state = (1 << MD_DISK_SYNC) | (1 << MD_DISK_ACTIVE);
202 st->ss->update_super(st, &info, "linear-grow-new", newdev,
203 0, 0, NULL);
204
205 if (st->ss->store_super(st, nfd)) {
206 pr_err("Cannot store new superblock on %s\n",
207 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 (ioctl(fd, GET_ARRAY_INFO, &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 (ioctl(fd, GET_DISK_INFO, &disk) < 0) {
233 pr_err("cannot get device detail for device %d\n",
234 d);
235 return 1;
236 }
237 dv = map_dev(disk.major, disk.minor, 1);
238 if (!dv) {
239 pr_err("cannot find device file for device %d\n",
240 d);
241 return 1;
242 }
243 fd2 = dev_open(dv, O_RDWR);
244 if (fd2 < 0) {
245 pr_err("cannot open device file %s\n", dv);
246 return 1;
247 }
248 if (st->ss->load_super(st, fd2, NULL)) {
249 pr_err("cannot find super block on %s\n", dv);
250 close(fd);
251 return 1;
252 }
253 info.array.raid_disks = nd+1;
254 info.array.nr_disks = nd+1;
255 info.array.active_disks = nd+1;
256 info.array.working_disks = nd+1;
257
258 st->ss->update_super(st, &info, "linear-grow-update", dv,
259 0, 0, NULL);
260
261 if (st->ss->store_super(st, fd2)) {
262 pr_err("Cannot store new superblock on %s\n", dv);
263 close(fd2);
264 return 1;
265 }
266 close(fd2);
267 }
268
269 return 0;
270 }
271
272 int Grow_addbitmap(char *devname, int fd, struct context *c, struct shape *s)
273 {
274 /*
275 * First check that array doesn't have a bitmap
276 * Then create the bitmap
277 * Then add it
278 *
279 * For internal bitmaps, we need to check the version,
280 * find all the active devices, and write the bitmap block
281 * to all devices
282 */
283 mdu_bitmap_file_t bmf;
284 mdu_array_info_t array;
285 struct supertype *st;
286 char *subarray = NULL;
287 int major = BITMAP_MAJOR_HI;
288 int vers = md_get_version(fd);
289 unsigned long long bitmapsize, array_size;
290
291 if (vers < 9003) {
292 major = BITMAP_MAJOR_HOSTENDIAN;
293 pr_err("Warning - bitmaps created on this kernel"
294 " are not portable\n"
295 " between different architectures. Consider upgrading"
296 " the Linux kernel.\n");
297 }
298
299 if (ioctl(fd, GET_BITMAP_FILE, &bmf) != 0) {
300 if (errno == ENOMEM)
301 pr_err("Memory allocation failure.\n");
302 else
303 pr_err("bitmaps not supported by this kernel.\n");
304 return 1;
305 }
306 if (bmf.pathname[0]) {
307 if (strcmp(s->bitmap_file,"none")==0) {
308 if (ioctl(fd, SET_BITMAP_FILE, -1)!= 0) {
309 pr_err("failed to remove bitmap %s\n",
310 bmf.pathname);
311 return 1;
312 }
313 return 0;
314 }
315 pr_err("%s already has a bitmap (%s)\n",
316 devname, bmf.pathname);
317 return 1;
318 }
319 if (ioctl(fd, GET_ARRAY_INFO, &array) != 0) {
320 pr_err("cannot get array status for %s\n", devname);
321 return 1;
322 }
323 if (array.state & (1<<MD_SB_BITMAP_PRESENT)) {
324 if (strcmp(s->bitmap_file, "none")==0) {
325 array.state &= ~(1<<MD_SB_BITMAP_PRESENT);
326 if (ioctl(fd, SET_ARRAY_INFO, &array)!= 0) {
327 pr_err("failed to remove internal bitmap.\n");
328 return 1;
329 }
330 return 0;
331 }
332 pr_err("Internal bitmap already present on %s\n",
333 devname);
334 return 1;
335 }
336
337 if (strcmp(s->bitmap_file, "none") == 0) {
338 pr_err("no bitmap found on %s\n", devname);
339 return 1;
340 }
341 if (array.level <= 0) {
342 pr_err("Bitmaps not meaningful with level %s\n",
343 map_num(pers, array.level)?:"of this array");
344 return 1;
345 }
346 bitmapsize = array.size;
347 bitmapsize <<= 1;
348 if (get_dev_size(fd, NULL, &array_size) &&
349 array_size > (0x7fffffffULL<<9)) {
350 /* Array is big enough that we cannot trust array.size
351 * try other approaches
352 */
353 bitmapsize = get_component_size(fd);
354 }
355 if (bitmapsize == 0) {
356 pr_err("Cannot reliably determine size of array to create bitmap - sorry.\n");
357 return 1;
358 }
359
360 if (array.level == 10) {
361 int ncopies = (array.layout&255)*((array.layout>>8)&255);
362 bitmapsize = bitmapsize * array.raid_disks / ncopies;
363 }
364
365 st = super_by_fd(fd, &subarray);
366 if (!st) {
367 pr_err("Cannot understand version %d.%d\n",
368 array.major_version, array.minor_version);
369 return 1;
370 }
371 if (subarray) {
372 pr_err("Cannot add bitmaps to sub-arrays yet\n");
373 free(subarray);
374 free(st);
375 return 1;
376 }
377 if (strcmp(s->bitmap_file, "internal") == 0) {
378 int rv;
379 int d;
380 int offset_setable = 0;
381 struct mdinfo *mdi;
382 if (st->ss->add_internal_bitmap == NULL) {
383 pr_err("Internal bitmaps not supported "
384 "with %s metadata\n", st->ss->name);
385 return 1;
386 }
387 mdi = sysfs_read(fd, NULL, GET_BITMAP_LOCATION);
388 if (mdi)
389 offset_setable = 1;
390 for (d=0; d< st->max_devs; d++) {
391 mdu_disk_info_t disk;
392 char *dv;
393 disk.number = d;
394 if (ioctl(fd, GET_DISK_INFO, &disk) < 0)
395 continue;
396 if (disk.major == 0 &&
397 disk.minor == 0)
398 continue;
399 if ((disk.state & (1<<MD_DISK_SYNC))==0)
400 continue;
401 dv = map_dev(disk.major, disk.minor, 1);
402 if (dv) {
403 int fd2 = dev_open(dv, O_RDWR);
404 if (fd2 < 0)
405 continue;
406 if (st->ss->load_super(st, fd2, NULL)==0) {
407 if (st->ss->add_internal_bitmap(
408 st,
409 &s->bitmap_chunk, c->delay, s->write_behind,
410 bitmapsize, offset_setable,
411 major)
412 )
413 st->ss->write_bitmap(st, fd2);
414 else {
415 pr_err("failed to create internal bitmap"
416 " - chunksize problem.\n");
417 close(fd2);
418 return 1;
419 }
420 }
421 close(fd2);
422 }
423 }
424 if (offset_setable) {
425 st->ss->getinfo_super(st, mdi, NULL);
426 sysfs_init(mdi, fd, NULL);
427 rv = sysfs_set_num_signed(mdi, NULL, "bitmap/location",
428 mdi->bitmap_offset);
429 } else {
430 array.state |= (1<<MD_SB_BITMAP_PRESENT);
431 rv = ioctl(fd, SET_ARRAY_INFO, &array);
432 }
433 if (rv < 0) {
434 if (errno == EBUSY)
435 pr_err("Cannot add bitmap while array is"
436 " resyncing or reshaping etc.\n");
437 pr_err("failed to set internal bitmap.\n");
438 return 1;
439 }
440 } else {
441 int uuid[4];
442 int bitmap_fd;
443 int d;
444 int max_devs = st->max_devs;
445
446 /* try to load a superblock */
447 for (d = 0; d < max_devs; d++) {
448 mdu_disk_info_t disk;
449 char *dv;
450 int fd2;
451 disk.number = d;
452 if (ioctl(fd, GET_DISK_INFO, &disk) < 0)
453 continue;
454 if ((disk.major==0 && disk.minor==0) ||
455 (disk.state & (1<<MD_DISK_REMOVED)))
456 continue;
457 dv = map_dev(disk.major, disk.minor, 1);
458 if (!dv)
459 continue;
460 fd2 = dev_open(dv, O_RDONLY);
461 if (fd2 >= 0) {
462 if (st->ss->load_super(st, fd2, NULL) == 0) {
463 close(fd2);
464 st->ss->uuid_from_super(st, uuid);
465 break;
466 }
467 close(fd2);
468 }
469 }
470 if (d == max_devs) {
471 pr_err("cannot find UUID for array!\n");
472 return 1;
473 }
474 if (CreateBitmap(s->bitmap_file, c->force, (char*)uuid, s->bitmap_chunk,
475 c->delay, s->write_behind, bitmapsize, major)) {
476 return 1;
477 }
478 bitmap_fd = open(s->bitmap_file, O_RDWR);
479 if (bitmap_fd < 0) {
480 pr_err("weird: %s cannot be opened\n",
481 s->bitmap_file);
482 return 1;
483 }
484 if (ioctl(fd, SET_BITMAP_FILE, bitmap_fd) < 0) {
485 int err = errno;
486 if (errno == EBUSY)
487 pr_err("Cannot add bitmap while array is"
488 " resyncing or reshaping etc.\n");
489 pr_err("Cannot set bitmap file for %s: %s\n",
490 devname, strerror(err));
491 return 1;
492 }
493 }
494
495 return 0;
496 }
497
498 /*
499 * When reshaping an array we might need to backup some data.
500 * This is written to all spares with a 'super_block' describing it.
501 * The superblock goes 4K from the end of the used space on the
502 * device.
503 * It if written after the backup is complete.
504 * It has the following structure.
505 */
506
507 static struct mdp_backup_super {
508 char magic[16]; /* md_backup_data-1 or -2 */
509 __u8 set_uuid[16];
510 __u64 mtime;
511 /* start/sizes in 512byte sectors */
512 __u64 devstart; /* address on backup device/file of data */
513 __u64 arraystart;
514 __u64 length;
515 __u32 sb_csum; /* csum of preceeding bytes. */
516 __u32 pad1;
517 __u64 devstart2; /* offset in to data of second section */
518 __u64 arraystart2;
519 __u64 length2;
520 __u32 sb_csum2; /* csum of preceeding bytes. */
521 __u8 pad[512-68-32];
522 } __attribute__((aligned(512))) bsb, bsb2;
523
524 static __u32 bsb_csum(char *buf, int len)
525 {
526 int i;
527 int csum = 0;
528 for (i = 0; i < len; i++)
529 csum = (csum<<3) + buf[0];
530 return __cpu_to_le32(csum);
531 }
532
533 static int check_idle(struct supertype *st)
534 {
535 /* Check that all member arrays for this container, or the
536 * container of this array, are idle
537 */
538 char *container = (st->container_devnm[0]
539 ? st->container_devnm : st->devnm);
540 struct mdstat_ent *ent, *e;
541 int is_idle = 1;
542
543 ent = mdstat_read(0, 0);
544 for (e = ent ; e; e = e->next) {
545 if (!is_container_member(e, container))
546 continue;
547 if (e->percent >= 0) {
548 is_idle = 0;
549 break;
550 }
551 }
552 free_mdstat(ent);
553 return is_idle;
554 }
555
556 static int freeze_container(struct supertype *st)
557 {
558 char *container = (st->container_devnm[0]
559 ? st->container_devnm : st->devnm);
560
561 if (!check_idle(st))
562 return -1;
563
564 if (block_monitor(container, 1)) {
565 pr_err("failed to freeze container\n");
566 return -2;
567 }
568
569 return 1;
570 }
571
572 static void unfreeze_container(struct supertype *st)
573 {
574 char *container = (st->container_devnm[0]
575 ? st->container_devnm : st->devnm);
576
577 unblock_monitor(container, 1);
578 }
579
580 static int freeze(struct supertype *st)
581 {
582 /* Try to freeze resync/rebuild on this array/container.
583 * Return -1 if the array is busy,
584 * return -2 container cannot be frozen,
585 * return 0 if this kernel doesn't support 'frozen'
586 * return 1 if it worked.
587 */
588 if (st->ss->external)
589 return freeze_container(st);
590 else {
591 struct mdinfo *sra = sysfs_read(-1, st->devnm, GET_VERSION);
592 int err;
593 char buf[20];
594
595 if (!sra)
596 return -1;
597 /* Need to clear any 'read-auto' status */
598 if (sysfs_get_str(sra, NULL, "array_state", buf, 20) > 0 &&
599 strncmp(buf, "read-auto", 9) == 0)
600 sysfs_set_str(sra, NULL, "array_state", "clean");
601
602 err = sysfs_freeze_array(sra);
603 sysfs_free(sra);
604 return err;
605 }
606 }
607
608 static void unfreeze(struct supertype *st)
609 {
610 if (st->ss->external)
611 return unfreeze_container(st);
612 else {
613 struct mdinfo *sra = sysfs_read(-1, st->devnm, GET_VERSION);
614
615 if (sra)
616 sysfs_set_str(sra, NULL, "sync_action", "idle");
617 sysfs_free(sra);
618 }
619 }
620
621 static void wait_reshape(struct mdinfo *sra)
622 {
623 int fd = sysfs_get_fd(sra, NULL, "sync_action");
624 char action[20];
625
626 if (fd < 0)
627 return;
628
629 while (sysfs_fd_get_str(fd, action, 20) > 0 &&
630 strncmp(action, "reshape", 7) == 0)
631 sysfs_wait(fd, NULL);
632 close(fd);
633 }
634
635 static int reshape_super(struct supertype *st, unsigned long long size,
636 int level, int layout, int chunksize, int raid_disks,
637 int delta_disks, char *backup_file, char *dev,
638 int direction, int verbose)
639 {
640 /* nothing extra to check in the native case */
641 if (!st->ss->external)
642 return 0;
643 if (!st->ss->reshape_super ||
644 !st->ss->manage_reshape) {
645 pr_err("%s metadata does not support reshape\n",
646 st->ss->name);
647 return 1;
648 }
649
650 return st->ss->reshape_super(st, size, level, layout, chunksize,
651 raid_disks, delta_disks, backup_file, dev,
652 direction, verbose);
653 }
654
655 static void sync_metadata(struct supertype *st)
656 {
657 if (st->ss->external) {
658 if (st->update_tail) {
659 flush_metadata_updates(st);
660 st->update_tail = &st->updates;
661 } else
662 st->ss->sync_metadata(st);
663 }
664 }
665
666 static int subarray_set_num(char *container, struct mdinfo *sra, char *name, int n)
667 {
668 /* when dealing with external metadata subarrays we need to be
669 * prepared to handle EAGAIN. The kernel may need to wait for
670 * mdmon to mark the array active so the kernel can handle
671 * allocations/writeback when preparing the reshape action
672 * (md_allow_write()). We temporarily disable safe_mode_delay
673 * to close a race with the array_state going clean before the
674 * next write to raid_disks / stripe_cache_size
675 */
676 char safe[50];
677 int rc;
678
679 /* only 'raid_disks' and 'stripe_cache_size' trigger md_allow_write */
680 if (!container ||
681 (strcmp(name, "raid_disks") != 0 &&
682 strcmp(name, "stripe_cache_size") != 0))
683 return sysfs_set_num(sra, NULL, name, n);
684
685 rc = sysfs_get_str(sra, NULL, "safe_mode_delay", safe, sizeof(safe));
686 if (rc <= 0)
687 return -1;
688 sysfs_set_num(sra, NULL, "safe_mode_delay", 0);
689 rc = sysfs_set_num(sra, NULL, name, n);
690 if (rc < 0 && errno == EAGAIN) {
691 ping_monitor(container);
692 /* if we get EAGAIN here then the monitor is not active
693 * so stop trying
694 */
695 rc = sysfs_set_num(sra, NULL, name, n);
696 }
697 sysfs_set_str(sra, NULL, "safe_mode_delay", safe);
698 return rc;
699 }
700
701 int start_reshape(struct mdinfo *sra, int already_running,
702 int before_data_disks, int data_disks)
703 {
704 int err;
705 unsigned long long sync_max_to_set;
706
707 sysfs_set_num(sra, NULL, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL);
708 err = sysfs_set_num(sra, NULL, "suspend_hi", sra->reshape_progress);
709 err = err ?: sysfs_set_num(sra, NULL, "suspend_lo",
710 sra->reshape_progress);
711 if (before_data_disks <= data_disks)
712 sync_max_to_set = sra->reshape_progress / data_disks;
713 else
714 sync_max_to_set = (sra->component_size * data_disks
715 - sra->reshape_progress) / data_disks;
716 if (!already_running)
717 sysfs_set_num(sra, NULL, "sync_min", sync_max_to_set);
718 err = err ?: sysfs_set_num(sra, NULL, "sync_max", sync_max_to_set);
719 if (!already_running)
720 err = err ?: sysfs_set_str(sra, NULL, "sync_action", "reshape");
721
722 return err;
723 }
724
725 void abort_reshape(struct mdinfo *sra)
726 {
727 sysfs_set_str(sra, NULL, "sync_action", "idle");
728 sysfs_set_num(sra, NULL, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL);
729 sysfs_set_num(sra, NULL, "suspend_hi", 0);
730 sysfs_set_num(sra, NULL, "suspend_lo", 0);
731 sysfs_set_num(sra, NULL, "sync_min", 0);
732 sysfs_set_str(sra, NULL, "sync_max", "max");
733 }
734
735 int remove_disks_for_takeover(struct supertype *st,
736 struct mdinfo *sra,
737 int layout)
738 {
739 int nr_of_copies;
740 struct mdinfo *remaining;
741 int slot;
742
743 if (sra->array.level == 10)
744 nr_of_copies = layout & 0xff;
745 else if (sra->array.level == 1)
746 nr_of_copies = sra->array.raid_disks;
747 else
748 return 1;
749
750 remaining = sra->devs;
751 sra->devs = NULL;
752 /* for each 'copy', select one device and remove from the list. */
753 for (slot = 0; slot < sra->array.raid_disks; slot += nr_of_copies) {
754 struct mdinfo **diskp;
755 int found = 0;
756
757 /* Find a working device to keep */
758 for (diskp = &remaining; *diskp ; diskp = &(*diskp)->next) {
759 struct mdinfo *disk = *diskp;
760
761 if (disk->disk.raid_disk < slot)
762 continue;
763 if (disk->disk.raid_disk >= slot + nr_of_copies)
764 continue;
765 if (disk->disk.state & (1<<MD_DISK_REMOVED))
766 continue;
767 if (disk->disk.state & (1<<MD_DISK_FAULTY))
768 continue;
769 if (!(disk->disk.state & (1<<MD_DISK_SYNC)))
770 continue;
771
772 /* We have found a good disk to use! */
773 *diskp = disk->next;
774 disk->next = sra->devs;
775 sra->devs = disk;
776 found = 1;
777 break;
778 }
779 if (!found)
780 break;
781 }
782
783 if (slot < sra->array.raid_disks) {
784 /* didn't find all slots */
785 struct mdinfo **e;
786 e = &remaining;
787 while (*e)
788 e = &(*e)->next;
789 *e = sra->devs;
790 sra->devs = remaining;
791 return 1;
792 }
793
794 /* Remove all 'remaining' devices from the array */
795 while (remaining) {
796 struct mdinfo *sd = remaining;
797 remaining = sd->next;
798
799 sysfs_set_str(sra, sd, "state", "faulty");
800 sysfs_set_str(sra, sd, "slot", "none");
801 /* for external metadata disks should be removed in mdmon */
802 if (!st->ss->external)
803 sysfs_set_str(sra, sd, "state", "remove");
804 sd->disk.state |= (1<<MD_DISK_REMOVED);
805 sd->disk.state &= ~(1<<MD_DISK_SYNC);
806 sd->next = sra->devs;
807 sra->devs = sd;
808 }
809 return 0;
810 }
811
812 void reshape_free_fdlist(int *fdlist,
813 unsigned long long *offsets,
814 int size)
815 {
816 int i;
817
818 for (i = 0; i < size; i++)
819 if (fdlist[i] >= 0)
820 close(fdlist[i]);
821
822 free(fdlist);
823 free(offsets);
824 }
825
826 int reshape_prepare_fdlist(char *devname,
827 struct mdinfo *sra,
828 int raid_disks,
829 int nrdisks,
830 unsigned long blocks,
831 char *backup_file,
832 int *fdlist,
833 unsigned long long *offsets)
834 {
835 int d = 0;
836 struct mdinfo *sd;
837
838 enable_fds(nrdisks);
839 for (d = 0; d <= nrdisks; d++)
840 fdlist[d] = -1;
841 d = raid_disks;
842 for (sd = sra->devs; sd; sd = sd->next) {
843 if (sd->disk.state & (1<<MD_DISK_FAULTY))
844 continue;
845 if (sd->disk.state & (1<<MD_DISK_SYNC)) {
846 char *dn = map_dev(sd->disk.major,
847 sd->disk.minor, 1);
848 fdlist[sd->disk.raid_disk]
849 = dev_open(dn, O_RDONLY);
850 offsets[sd->disk.raid_disk] = sd->data_offset*512;
851 if (fdlist[sd->disk.raid_disk] < 0) {
852 pr_err("%s: cannot open component %s\n",
853 devname, dn ? dn : "-unknown-");
854 d = -1;
855 goto release;
856 }
857 } else if (backup_file == NULL) {
858 /* spare */
859 char *dn = map_dev(sd->disk.major,
860 sd->disk.minor, 1);
861 fdlist[d] = dev_open(dn, O_RDWR);
862 offsets[d] = (sd->data_offset + sra->component_size - blocks - 8)*512;
863 if (fdlist[d] < 0) {
864 pr_err("%s: cannot open component %s\n",
865 devname, dn ? dn : "-unknown-");
866 d = -1;
867 goto release;
868 }
869 d++;
870 }
871 }
872 release:
873 return d;
874 }
875
876 int reshape_open_backup_file(char *backup_file,
877 int fd,
878 char *devname,
879 long blocks,
880 int *fdlist,
881 unsigned long long *offsets,
882 int restart)
883 {
884 /* Return 1 on success, 0 on any form of failure */
885 /* need to check backup file is large enough */
886 char buf[512];
887 struct stat stb;
888 unsigned int dev;
889 int i;
890
891 *fdlist = open(backup_file, O_RDWR|O_CREAT|(restart ? O_TRUNC : O_EXCL),
892 S_IRUSR | S_IWUSR);
893 *offsets = 8 * 512;
894 if (*fdlist < 0) {
895 pr_err("%s: cannot create backup file %s: %s\n",
896 devname, backup_file, strerror(errno));
897 return 0;
898 }
899 /* Guard against backup file being on array device.
900 * If array is partitioned or if LVM etc is in the
901 * way this will not notice, but it is better than
902 * nothing.
903 */
904 fstat(*fdlist, &stb);
905 dev = stb.st_dev;
906 fstat(fd, &stb);
907 if (stb.st_rdev == dev) {
908 pr_err("backup file must NOT be"
909 " on the array being reshaped.\n");
910 close(*fdlist);
911 return 0;
912 }
913
914 memset(buf, 0, 512);
915 for (i=0; i < blocks + 8 ; i++) {
916 if (write(*fdlist, buf, 512) != 512) {
917 pr_err("%s: cannot create"
918 " backup file %s: %s\n",
919 devname, backup_file, strerror(errno));
920 return 0;
921 }
922 }
923 if (fsync(*fdlist) != 0) {
924 pr_err("%s: cannot create backup file %s: %s\n",
925 devname, backup_file, strerror(errno));
926 return 0;
927 }
928
929 return 1;
930 }
931
932 unsigned long compute_backup_blocks(int nchunk, int ochunk,
933 unsigned int ndata, unsigned int odata)
934 {
935 unsigned long a, b, blocks;
936 /* So how much do we need to backup.
937 * We need an amount of data which is both a whole number of
938 * old stripes and a whole number of new stripes.
939 * So LCM for (chunksize*datadisks).
940 */
941 a = (ochunk/512) * odata;
942 b = (nchunk/512) * ndata;
943 /* Find GCD */
944 a = GCD(a, b);
945 /* LCM == product / GCD */
946 blocks = (ochunk/512) * (nchunk/512) * odata * ndata / a;
947
948 return blocks;
949 }
950
951 char *analyse_change(char *devname, struct mdinfo *info, struct reshape *re)
952 {
953 /* Based on the current array state in info->array and
954 * the changes in info->new_* etc, determine:
955 * - whether the change is possible
956 * - Intermediate level/raid_disks/layout
957 * - whether a restriping reshape is needed
958 * - number of sectors in minimum change unit. This
959 * will cover a whole number of stripes in 'before' and
960 * 'after'.
961 *
962 * Return message if the change should be rejected
963 * NULL if the change can be achieved
964 *
965 * This can be called as part of starting a reshape, or
966 * when assembling an array that is undergoing reshape.
967 */
968 int near, far, offset, copies;
969 int new_disks;
970 int old_chunk, new_chunk;
971 /* delta_parity records change in number of devices
972 * caused by level change
973 */
974 int delta_parity = 0;
975
976 memset(re, 0, sizeof(*re));
977
978 /* If a new level not explicitly given, we assume no-change */
979 if (info->new_level == UnSet)
980 info->new_level = info->array.level;
981
982 if (info->new_chunk)
983 switch (info->new_level) {
984 case 0:
985 case 4:
986 case 5:
987 case 6:
988 case 10:
989 /* chunk size is meaningful, must divide component_size
990 * evenly
991 */
992 if (info->component_size % (info->new_chunk/512)) {
993 unsigned long long shrink = info->component_size;
994 shrink &= ~(unsigned long long)(info->new_chunk/512-1);
995 pr_err("New chunk size (%dK) does not evenly divide device size (%lluk)\n",
996 info->new_chunk/1024, info->component_size/2);
997 pr_err("After shrinking any filesystem, \"mdadm --grow %s --size %llu\"\n",
998 devname, shrink/2);
999 pr_err("will shrink the array so the given chunk size would work.\n");
1000 return "";
1001 }
1002 break;
1003 default:
1004 return "chunk size not meaningful for this level";
1005 }
1006 else
1007 info->new_chunk = info->array.chunk_size;
1008
1009 switch (info->array.level) {
1010 default:
1011 return "Cannot understand this RAID level";
1012 case 1:
1013 /* RAID1 can convert to RAID1 with different disks, or
1014 * raid5 with 2 disks, or
1015 * raid0 with 1 disk
1016 */
1017 if (info->new_level > 1 &&
1018 (info->component_size & 7))
1019 return "Cannot convert RAID1 of this size - "
1020 "reduce size to multiple of 4K first.";
1021 if (info->new_level == 0) {
1022 if (info->delta_disks != UnSet &&
1023 info->delta_disks != 0)
1024 return "Cannot change number of disks "
1025 "with RAID1->RAID0 conversion";
1026 re->level = 0;
1027 re->before.data_disks = 1;
1028 re->after.data_disks = 1;
1029 return NULL;
1030 }
1031 if (info->new_level == 1) {
1032 if (info->delta_disks == UnSet)
1033 /* Don't know what to do */
1034 return "no change requested for Growing RAID1";
1035 re->level = 1;
1036 return NULL;
1037 }
1038 if (info->array.raid_disks == 2 &&
1039 info->new_level == 5) {
1040
1041 re->level = 5;
1042 re->before.data_disks = 1;
1043 if (info->delta_disks != UnSet &&
1044 info->delta_disks != 0)
1045 re->after.data_disks = 1 + info->delta_disks;
1046 else
1047 re->after.data_disks = 1;
1048 if (re->after.data_disks < 1)
1049 return "Number of disks too small for RAID5";
1050
1051 re->before.layout = ALGORITHM_LEFT_SYMMETRIC;
1052 info->array.chunk_size = 65536;
1053 break;
1054 }
1055 /* Could do some multi-stage conversions, but leave that to
1056 * later.
1057 */
1058 return "Impossibly level change request for RAID1";
1059
1060 case 10:
1061 /* RAID10 can be converted from near mode to
1062 * RAID0 by removing some devices.
1063 * It can also be reshaped if the kernel supports
1064 * new_data_offset.
1065 */
1066 switch (info->new_level) {
1067 case 0:
1068 if ((info->array.layout & ~0xff) != 0x100)
1069 return "Cannot Grow RAID10 with far/offset layout";
1070 /* number of devices must be multiple of number of copies */
1071 if (info->array.raid_disks % (info->array.layout & 0xff))
1072 return "RAID10 layout too complex for Grow operation";
1073
1074 new_disks = (info->array.raid_disks
1075 / (info->array.layout & 0xff));
1076 if (info->delta_disks == UnSet)
1077 info->delta_disks = (new_disks
1078 - info->array.raid_disks);
1079
1080 if (info->delta_disks != new_disks - info->array.raid_disks)
1081 return "New number of raid-devices impossible for RAID10";
1082 if (info->new_chunk &&
1083 info->new_chunk != info->array.chunk_size)
1084 return "Cannot change chunk-size with RAID10 Grow";
1085
1086 /* looks good */
1087 re->level = 0;
1088 re->before.data_disks = new_disks;
1089 re->after.data_disks = re->before.data_disks;
1090 return NULL;
1091
1092 case 10:
1093 near = info->array.layout & 0xff;
1094 far = (info->array.layout >> 8) & 0xff;
1095 offset = info->array.layout & 0x10000;
1096 if (far > 1 && !offset)
1097 return "Cannot reshape RAID10 in far-mode";
1098 copies = near * far;
1099
1100 old_chunk = info->array.chunk_size * far;
1101
1102 if (info->new_layout == UnSet)
1103 info->new_layout = info->array.layout;
1104 else {
1105 near = info->new_layout & 0xff;
1106 far = (info->new_layout >> 8) & 0xff;
1107 offset = info->new_layout & 0x10000;
1108 if (far > 1 && !offset)
1109 return "Cannot reshape RAID10 to far-mode";
1110 if (near * far != copies)
1111 return "Cannot change number of copies"
1112 " when reshaping RAID10";
1113 }
1114 if (info->delta_disks == UnSet)
1115 info->delta_disks = 0;
1116 new_disks = (info->array.raid_disks +
1117 info->delta_disks);
1118
1119 new_chunk = info->new_chunk * far;
1120
1121 re->level = 10;
1122 re->before.layout = info->array.layout;
1123 re->before.data_disks = info->array.raid_disks;
1124 re->after.layout = info->new_layout;
1125 re->after.data_disks = new_disks;
1126 /* For RAID10 we don't do backup but do allow reshape,
1127 * so set backup_blocks to INVALID_SECTORS rather than
1128 * zero.
1129 * And there is no need to synchronise stripes on both
1130 * 'old' and 'new'. So the important
1131 * number is the minimum data_offset difference
1132 * which is the larger of (offset copies * chunk).
1133 */
1134 re->backup_blocks = INVALID_SECTORS;
1135 re->min_offset_change = max(old_chunk, new_chunk) / 512;
1136 if (new_disks < re->before.data_disks &&
1137 info->space_after < re->min_offset_change)
1138 /* Reduce component size by one chunk */
1139 re->new_size = (info->component_size -
1140 re->min_offset_change);
1141 else
1142 re->new_size = info->component_size;
1143 re->new_size = re->new_size * new_disks / copies;
1144 return NULL;
1145
1146 default:
1147 return "RAID10 can only be changed to RAID0";
1148 }
1149 case 0:
1150 /* RAID0 can be converted to RAID10, or to RAID456 */
1151 if (info->new_level == 10) {
1152 if (info->new_layout == UnSet && info->delta_disks == UnSet) {
1153 /* Assume near=2 layout */
1154 info->new_layout = 0x102;
1155 info->delta_disks = info->array.raid_disks;
1156 }
1157 if (info->new_layout == UnSet) {
1158 int copies = 1 + (info->delta_disks
1159 / info->array.raid_disks);
1160 if (info->array.raid_disks * (copies-1)
1161 != info->delta_disks)
1162 return "Impossible number of devices"
1163 " for RAID0->RAID10";
1164 info->new_layout = 0x100 + copies;
1165 }
1166 if (info->delta_disks == UnSet) {
1167 int copies = info->new_layout & 0xff;
1168 if (info->new_layout != 0x100 + copies)
1169 return "New layout impossible"
1170 " for RAID0->RAID10";;
1171 info->delta_disks = (copies - 1) *
1172 info->array.raid_disks;
1173 }
1174 if (info->new_chunk &&
1175 info->new_chunk != info->array.chunk_size)
1176 return "Cannot change chunk-size with RAID0->RAID10";
1177 /* looks good */
1178 re->level = 10;
1179 re->before.data_disks = (info->array.raid_disks +
1180 info->delta_disks);
1181 re->after.data_disks = re->before.data_disks;
1182 re->before.layout = info->new_layout;
1183 return NULL;
1184 }
1185
1186 /* RAID0 can also covert to RAID0/4/5/6 by first converting to
1187 * a raid4 style layout of the final level.
1188 */
1189 switch (info->new_level) {
1190 case 4:
1191 delta_parity = 1;
1192 case 0:
1193 re->level = 4;
1194 re->before.layout = 0;
1195 break;
1196 case 5:
1197 delta_parity = 1;
1198 re->level = 5;
1199 re->before.layout = ALGORITHM_PARITY_N;
1200 if (info->new_layout == UnSet)
1201 info->new_layout = map_name(r5layout, "default");
1202 break;
1203 case 6:
1204 delta_parity = 2;
1205 re->level = 6;
1206 re->before.layout = ALGORITHM_PARITY_N;
1207 if (info->new_layout == UnSet)
1208 info->new_layout = map_name(r6layout, "default");
1209 break;
1210 default:
1211 return "Impossible level change requested";
1212 }
1213 re->before.data_disks = info->array.raid_disks;
1214 /* determining 'after' layout happens outside this 'switch' */
1215 break;
1216
1217 case 4:
1218 info->array.layout = ALGORITHM_PARITY_N;
1219 case 5:
1220 switch (info->new_level) {
1221 case 0:
1222 delta_parity = -1;
1223 case 4:
1224 re->level = info->array.level;
1225 re->before.data_disks = info->array.raid_disks - 1;
1226 re->before.layout = info->array.layout;
1227 break;
1228 case 5:
1229 re->level = 5;
1230 re->before.data_disks = info->array.raid_disks - 1;
1231 re->before.layout = info->array.layout;
1232 break;
1233 case 6:
1234 delta_parity = 1;
1235 re->level = 6;
1236 re->before.data_disks = info->array.raid_disks - 1;
1237 switch (info->array.layout) {
1238 case ALGORITHM_LEFT_ASYMMETRIC:
1239 re->before.layout = ALGORITHM_LEFT_ASYMMETRIC_6;
1240 break;
1241 case ALGORITHM_RIGHT_ASYMMETRIC:
1242 re->before.layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
1243 break;
1244 case ALGORITHM_LEFT_SYMMETRIC:
1245 re->before.layout = ALGORITHM_LEFT_SYMMETRIC_6;
1246 break;
1247 case ALGORITHM_RIGHT_SYMMETRIC:
1248 re->before.layout = ALGORITHM_RIGHT_SYMMETRIC_6;
1249 break;
1250 case ALGORITHM_PARITY_0:
1251 re->before.layout = ALGORITHM_PARITY_0_6;
1252 break;
1253 case ALGORITHM_PARITY_N:
1254 re->before.layout = ALGORITHM_PARITY_N_6;
1255 break;
1256 default:
1257 return "Cannot convert an array with this layout";
1258 }
1259 break;
1260 case 1:
1261 if (info->array.raid_disks != 2)
1262 return "Can only convert a 2-device array to RAID1";
1263 if (info->delta_disks != UnSet &&
1264 info->delta_disks != 0)
1265 return "Cannot set raid_disk when "
1266 "converting RAID5->RAID1";
1267 re->level = 1;
1268 info->new_chunk = 0;
1269 return NULL;
1270 default:
1271 return "Impossible level change requested";
1272 }
1273 break;
1274 case 6:
1275 switch (info->new_level) {
1276 case 4:
1277 case 5:
1278 delta_parity = -1;
1279 case 6:
1280 re->level = 6;
1281 re->before.data_disks = info->array.raid_disks - 2;
1282 re->before.layout = info->array.layout;
1283 break;
1284 default:
1285 return "Impossible level change requested";
1286 }
1287 break;
1288 }
1289
1290 /* If we reached here then it looks like a re-stripe is
1291 * happening. We have determined the intermediate level
1292 * and initial raid_disks/layout and stored these in 're'.
1293 *
1294 * We need to deduce the final layout that can be atomically
1295 * converted to the end state.
1296 */
1297 switch (info->new_level) {
1298 case 0:
1299 /* We can only get to RAID0 from RAID4 or RAID5
1300 * with appropriate layout and one extra device
1301 */
1302 if (re->level != 4 && re->level != 5)
1303 return "Cannot covert to RAID0 from this level";
1304
1305 switch (re->level) {
1306 case 4:
1307 re->before.layout = 0;
1308 re->after.layout = 0;
1309 break;
1310 case 5:
1311 re->after.layout = ALGORITHM_PARITY_N;
1312 break;
1313 }
1314 break;
1315
1316 case 4:
1317 /* We can only get to RAID4 from RAID5 */
1318 if (re->level != 4 && re->level != 5)
1319 return "Cannot convert to RAID4 from this level";
1320
1321 switch (re->level) {
1322 case 4:
1323 re->before.layout = 0;
1324 re->after.layout = 0;
1325 break;
1326 case 5:
1327 re->after.layout = ALGORITHM_PARITY_N;
1328 break;
1329 }
1330 break;
1331
1332 case 5:
1333 /* We get to RAID5 from RAID5 or RAID6 */
1334 if (re->level != 5 && re->level != 6)
1335 return "Cannot convert to RAID5 from this level";
1336
1337 switch (re->level) {
1338 case 5:
1339 if (info->new_layout == UnSet)
1340 re->after.layout = re->before.layout;
1341 else
1342 re->after.layout = info->new_layout;
1343 break;
1344 case 6:
1345 if (info->new_layout == UnSet)
1346 info->new_layout = re->before.layout;
1347
1348 /* after.layout needs to be raid6 version of new_layout */
1349 if (info->new_layout == ALGORITHM_PARITY_N)
1350 re->after.layout = ALGORITHM_PARITY_N;
1351 else {
1352 char layout[40];
1353 char *ls = map_num(r5layout, info->new_layout);
1354 int l;
1355 if (ls) {
1356 /* Current RAID6 layout has a RAID5
1357 * equivalent - good
1358 */
1359 strcat(strcpy(layout, ls), "-6");
1360 l = map_name(r6layout, layout);
1361 if (l == UnSet)
1362 return "Cannot find RAID6 layout"
1363 " to convert to";
1364 } else {
1365 /* Current RAID6 has no equivalent.
1366 * If it is already a '-6' layout we
1367 * can leave it unchanged, else we must
1368 * fail
1369 */
1370 ls = map_num(r6layout, info->new_layout);
1371 if (!ls ||
1372 strcmp(ls+strlen(ls)-2, "-6") != 0)
1373 return "Please specify new layout";
1374 l = info->new_layout;
1375 }
1376 re->after.layout = l;
1377 }
1378 }
1379 break;
1380
1381 case 6:
1382 /* We must already be at level 6 */
1383 if (re->level != 6)
1384 return "Impossible level change";
1385 if (info->new_layout == UnSet)
1386 re->after.layout = info->array.layout;
1387 else
1388 re->after.layout = info->new_layout;
1389 break;
1390 default:
1391 return "Impossible level change requested";
1392 }
1393 if (info->delta_disks == UnSet)
1394 info->delta_disks = delta_parity;
1395
1396 re->after.data_disks = (re->before.data_disks
1397 + info->delta_disks
1398 - delta_parity);
1399 switch (re->level) {
1400 case 6: re->parity = 2;
1401 break;
1402 case 4:
1403 case 5: re->parity = 1;
1404 break;
1405 default: re->parity = 0;
1406 break;
1407 }
1408 /* So we have a restripe operation, we need to calculate the number
1409 * of blocks per reshape operation.
1410 */
1411 re->new_size = info->component_size * re->before.data_disks;
1412 if (info->new_chunk == 0)
1413 info->new_chunk = info->array.chunk_size;
1414 if (re->after.data_disks == re->before.data_disks &&
1415 re->after.layout == re->before.layout &&
1416 info->new_chunk == info->array.chunk_size) {
1417 /* Nothing to change, can change level immediately. */
1418 re->level = info->new_level;
1419 re->backup_blocks = 0;
1420 return NULL;
1421 }
1422 if (re->after.data_disks == 1 && re->before.data_disks == 1) {
1423 /* chunk and layout changes make no difference */
1424 re->level = info->new_level;
1425 re->backup_blocks = 0;
1426 return NULL;
1427 }
1428
1429 if (re->after.data_disks == re->before.data_disks &&
1430 get_linux_version() < 2006032)
1431 return "in-place reshape is not safe before 2.6.32 - sorry.";
1432
1433 if (re->after.data_disks < re->before.data_disks &&
1434 get_linux_version() < 2006030)
1435 return "reshape to fewer devices is not supported before 2.6.30 - sorry.";
1436
1437 re->backup_blocks = compute_backup_blocks(
1438 info->new_chunk, info->array.chunk_size,
1439 re->after.data_disks,
1440 re->before.data_disks);
1441 re->min_offset_change = re->backup_blocks / re->before.data_disks;
1442
1443 re->new_size = info->component_size * re->after.data_disks;
1444 return NULL;
1445 }
1446
1447 static int set_array_size(struct supertype *st, struct mdinfo *sra,
1448 char *text_version)
1449 {
1450 struct mdinfo *info;
1451 char *subarray;
1452 int ret_val = -1;
1453
1454 if ((st == NULL) || (sra == NULL))
1455 return ret_val;
1456
1457 if (text_version == NULL)
1458 text_version = sra->text_version;
1459 subarray = strchr(text_version+1, '/')+1;
1460 info = st->ss->container_content(st, subarray);
1461 if (info) {
1462 unsigned long long current_size = 0;
1463 unsigned long long new_size =
1464 info->custom_array_size/2;
1465
1466 if (sysfs_get_ll(sra, NULL, "array_size", &current_size) == 0 &&
1467 new_size > current_size) {
1468 if (sysfs_set_num(sra, NULL, "array_size", new_size)
1469 < 0)
1470 dprintf("Error: Cannot set array size");
1471 else {
1472 ret_val = 0;
1473 dprintf("Array size changed");
1474 }
1475 dprintf(" from %llu to %llu.\n",
1476 current_size, new_size);
1477 }
1478 sysfs_free(info);
1479 } else
1480 dprintf("Error: set_array_size(): info pointer in NULL\n");
1481
1482 return ret_val;
1483 }
1484
1485 static int reshape_array(char *container, int fd, char *devname,
1486 struct supertype *st, struct mdinfo *info,
1487 int force, struct mddev_dev *devlist,
1488 unsigned long long data_offset,
1489 char *backup_file, int verbose, int forked,
1490 int restart, int freeze_reshape);
1491 static int reshape_container(char *container, char *devname,
1492 int mdfd,
1493 struct supertype *st,
1494 struct mdinfo *info,
1495 int force,
1496 char *backup_file,
1497 int verbose, int restart, int freeze_reshape);
1498
1499 int Grow_reshape(char *devname, int fd,
1500 struct mddev_dev *devlist,
1501 unsigned long long data_offset,
1502 struct context *c, struct shape *s)
1503 {
1504 /* Make some changes in the shape of an array.
1505 * The kernel must support the change.
1506 *
1507 * There are three different changes. Each can trigger
1508 * a resync or recovery so we freeze that until we have
1509 * requested everything (if kernel supports freezing - 2.6.30).
1510 * The steps are:
1511 * - change size (i.e. component_size)
1512 * - change level
1513 * - change layout/chunksize/ndisks
1514 *
1515 * The last can require a reshape. It is different on different
1516 * levels so we need to check the level before actioning it.
1517 * Some times the level change needs to be requested after the
1518 * reshape (e.g. raid6->raid5, raid5->raid0)
1519 *
1520 */
1521 struct mdu_array_info_s array;
1522 int rv = 0;
1523 struct supertype *st;
1524 char *subarray = NULL;
1525
1526 int frozen;
1527 int changed = 0;
1528 char *container = NULL;
1529 int cfd = -1;
1530
1531 struct mddev_dev *dv;
1532 int added_disks;
1533
1534 struct mdinfo info;
1535 struct mdinfo *sra;
1536
1537 if (ioctl(fd, GET_ARRAY_INFO, &array) < 0) {
1538 pr_err("%s is not an active md array - aborting\n",
1539 devname);
1540 return 1;
1541 }
1542 if (data_offset != INVALID_SECTORS && array.level != 10
1543 && (array.level < 4 || array.level > 6)) {
1544 pr_err("--grow --data-offset not yet supported\n");
1545 return 1;
1546 }
1547
1548 if (s->size > 0 &&
1549 (s->chunk || s->level!= UnSet || s->layout_str || s->raiddisks)) {
1550 pr_err("cannot change component size at the same time "
1551 "as other changes.\n"
1552 " Change size first, then check data is intact before "
1553 "making other changes.\n");
1554 return 1;
1555 }
1556
1557 if (s->raiddisks && s->raiddisks < array.raid_disks && array.level > 1 &&
1558 get_linux_version() < 2006032 &&
1559 !check_env("MDADM_FORCE_FEWER")) {
1560 pr_err("reducing the number of devices is not safe before Linux 2.6.32\n"
1561 " Please use a newer kernel\n");
1562 return 1;
1563 }
1564
1565 st = super_by_fd(fd, &subarray);
1566 if (!st) {
1567 pr_err("Unable to determine metadata format for %s\n", devname);
1568 return 1;
1569 }
1570 if (s->raiddisks > st->max_devs) {
1571 pr_err("Cannot increase raid-disks on this array"
1572 " beyond %d\n", st->max_devs);
1573 return 1;
1574 }
1575
1576 /* in the external case we need to check that the requested reshape is
1577 * supported, and perform an initial check that the container holds the
1578 * pre-requisite spare devices (mdmon owns final validation)
1579 */
1580 if (st->ss->external) {
1581 int rv;
1582
1583 if (subarray) {
1584 container = st->container_devnm;
1585 cfd = open_dev_excl(st->container_devnm);
1586 } else {
1587 container = st->devnm;
1588 close(fd);
1589 cfd = open_dev_excl(st->devnm);
1590 fd = cfd;
1591 }
1592 if (cfd < 0) {
1593 pr_err("Unable to open container for %s\n",
1594 devname);
1595 free(subarray);
1596 return 1;
1597 }
1598
1599 rv = st->ss->load_container(st, cfd, NULL);
1600
1601 if (rv) {
1602 pr_err("Cannot read superblock for %s\n",
1603 devname);
1604 free(subarray);
1605 return 1;
1606 }
1607
1608 /* check if operation is supported for metadata handler */
1609 if (st->ss->container_content) {
1610 struct mdinfo *cc = NULL;
1611 struct mdinfo *content = NULL;
1612
1613 cc = st->ss->container_content(st, subarray);
1614 for (content = cc; content ; content = content->next) {
1615 int allow_reshape = 1;
1616
1617 /* check if reshape is allowed based on metadata
1618 * indications stored in content.array.status
1619 */
1620 if (content->array.state & (1<<MD_SB_BLOCK_VOLUME))
1621 allow_reshape = 0;
1622 if (content->array.state
1623 & (1<<MD_SB_BLOCK_CONTAINER_RESHAPE))
1624 allow_reshape = 0;
1625 if (!allow_reshape) {
1626 pr_err("cannot reshape arrays in"
1627 " container with unsupported"
1628 " metadata: %s(%s)\n",
1629 devname, container);
1630 sysfs_free(cc);
1631 free(subarray);
1632 return 1;
1633 }
1634 }
1635 sysfs_free(cc);
1636 }
1637 if (mdmon_running(container))
1638 st->update_tail = &st->updates;
1639 }
1640
1641 added_disks = 0;
1642 for (dv = devlist; dv; dv = dv->next)
1643 added_disks++;
1644 if (s->raiddisks > array.raid_disks &&
1645 array.spare_disks +added_disks < (s->raiddisks - array.raid_disks) &&
1646 !c->force) {
1647 pr_err("Need %d spare%s to avoid degraded array,"
1648 " and only have %d.\n"
1649 " Use --force to over-ride this check.\n",
1650 s->raiddisks - array.raid_disks,
1651 s->raiddisks - array.raid_disks == 1 ? "" : "s",
1652 array.spare_disks + added_disks);
1653 return 1;
1654 }
1655
1656 sra = sysfs_read(fd, NULL, GET_LEVEL | GET_DISKS | GET_DEVS
1657 | GET_STATE | GET_VERSION);
1658 if (sra) {
1659 if (st->ss->external && subarray == NULL) {
1660 array.level = LEVEL_CONTAINER;
1661 sra->array.level = LEVEL_CONTAINER;
1662 }
1663 } else {
1664 pr_err("failed to read sysfs parameters for %s\n",
1665 devname);
1666 return 1;
1667 }
1668 frozen = freeze(st);
1669 if (frozen < -1) {
1670 /* freeze() already spewed the reason */
1671 sysfs_free(sra);
1672 return 1;
1673 } else if (frozen < 0) {
1674 pr_err("%s is performing resync/recovery and cannot"
1675 " be reshaped\n", devname);
1676 sysfs_free(sra);
1677 return 1;
1678 }
1679
1680 /* ========= set size =============== */
1681 if (s->size > 0 && (s->size == MAX_SIZE || s->size != (unsigned)array.size)) {
1682 unsigned long long orig_size = get_component_size(fd)/2;
1683 unsigned long long min_csize;
1684 struct mdinfo *mdi;
1685 int raid0_takeover = 0;
1686
1687 if (orig_size == 0)
1688 orig_size = (unsigned) array.size;
1689
1690 if (orig_size == 0) {
1691 pr_err("Cannot set device size in this type of array.\n");
1692 rv = 1;
1693 goto release;
1694 }
1695
1696 if (reshape_super(st, s->size, UnSet, UnSet, 0, 0, UnSet, NULL,
1697 devname, APPLY_METADATA_CHANGES, c->verbose > 0)) {
1698 rv = 1;
1699 goto release;
1700 }
1701 sync_metadata(st);
1702 if (st->ss->external) {
1703 /* metadata can have size limitation
1704 * update size value according to metadata information
1705 */
1706 struct mdinfo *sizeinfo =
1707 st->ss->container_content(st, subarray);
1708 if (sizeinfo) {
1709 unsigned long long new_size =
1710 sizeinfo->custom_array_size/2;
1711 int data_disks = get_data_disks(
1712 sizeinfo->array.level,
1713 sizeinfo->array.layout,
1714 sizeinfo->array.raid_disks);
1715 new_size /= data_disks;
1716 dprintf("Metadata size correction from %llu to "
1717 "%llu (%llu)\n", orig_size, new_size,
1718 new_size * data_disks);
1719 s->size = new_size;
1720 sysfs_free(sizeinfo);
1721 }
1722 }
1723
1724 /* Update the size of each member device in case
1725 * they have been resized. This will never reduce
1726 * below the current used-size. The "size" attribute
1727 * understands '0' to mean 'max'.
1728 */
1729 min_csize = 0;
1730 rv = 0;
1731 for (mdi = sra->devs; mdi; mdi = mdi->next) {
1732 if (sysfs_set_num(sra, mdi, "size",
1733 s->size == MAX_SIZE ? 0 : s->size) < 0) {
1734 /* Probably kernel refusing to let us
1735 * reduce the size - not an error.
1736 */
1737 break;
1738 }
1739 if (array.not_persistent == 0 &&
1740 array.major_version == 0 &&
1741 get_linux_version() < 3001000) {
1742 /* Dangerous to allow size to exceed 2TB */
1743 unsigned long long csize;
1744 if (sysfs_get_ll(sra, mdi, "size", &csize) == 0) {
1745 if (csize >= 2ULL*1024*1024*1024)
1746 csize = 2ULL*1024*1024*1024;
1747 if ((min_csize == 0 || (min_csize
1748 > csize)))
1749 min_csize = csize;
1750 }
1751 }
1752 }
1753 if (rv) {
1754 pr_err("Cannot set size on "
1755 "array members.\n");
1756 goto size_change_error;
1757 }
1758 if (min_csize && s->size > min_csize) {
1759 pr_err("Cannot safely make this array "
1760 "use more than 2TB per device on this kernel.\n");
1761 rv = 1;
1762 goto size_change_error;
1763 }
1764 if (min_csize && s->size == MAX_SIZE) {
1765 /* Don't let the kernel choose a size - it will get
1766 * it wrong
1767 */
1768 pr_err("Limited v0.90 array to "
1769 "2TB per device\n");
1770 s->size = min_csize;
1771 }
1772 if (st->ss->external) {
1773 if (sra->array.level == 0) {
1774 rv = sysfs_set_str(sra, NULL, "level",
1775 "raid5");
1776 if (!rv) {
1777 raid0_takeover = 1;
1778 /* get array parametes after takeover
1779 * to chane one parameter at time only
1780 */
1781 rv = ioctl(fd, GET_ARRAY_INFO, &array);
1782 }
1783 }
1784 /* make sure mdmon is
1785 * aware of the new level */
1786 if (!mdmon_running(st->container_devnm))
1787 start_mdmon(st->container_devnm);
1788 ping_monitor(container);
1789 if (mdmon_running(st->container_devnm) &&
1790 st->update_tail == NULL)
1791 st->update_tail = &st->updates;
1792 }
1793
1794 if (s->size == MAX_SIZE)
1795 s->size = 0;
1796 array.size = s->size;
1797 if ((unsigned)array.size != s->size) {
1798 /* got truncated to 32bit, write to
1799 * component_size instead
1800 */
1801 if (sra)
1802 rv = sysfs_set_num(sra, NULL,
1803 "component_size", s->size);
1804 else
1805 rv = -1;
1806 } else {
1807 rv = ioctl(fd, SET_ARRAY_INFO, &array);
1808
1809 /* manage array size when it is managed externally
1810 */
1811 if ((rv == 0) && st->ss->external)
1812 rv = set_array_size(st, sra, sra->text_version);
1813 }
1814
1815 if (raid0_takeover) {
1816 /* do not recync non-existing parity,
1817 * we will drop it anyway
1818 */
1819 sysfs_set_str(sra, NULL, "sync_action", "frozen");
1820 /* go back to raid0, drop parity disk
1821 */
1822 sysfs_set_str(sra, NULL, "level", "raid0");
1823 ioctl(fd, GET_ARRAY_INFO, &array);
1824 }
1825
1826 size_change_error:
1827 if (rv != 0) {
1828 int err = errno;
1829
1830 /* restore metadata */
1831 if (reshape_super(st, orig_size, UnSet, UnSet, 0, 0,
1832 UnSet, NULL, devname,
1833 ROLLBACK_METADATA_CHANGES,
1834 c->verbose) == 0)
1835 sync_metadata(st);
1836 pr_err("Cannot set device size for %s: %s\n",
1837 devname, strerror(err));
1838 if (err == EBUSY &&
1839 (array.state & (1<<MD_SB_BITMAP_PRESENT)))
1840 cont_err("Bitmap must be removed before size can be changed\n");
1841 rv = 1;
1842 goto release;
1843 }
1844 if (s->assume_clean) {
1845 /* This will fail on kernels older than 3.0 unless
1846 * a backport has been arranged.
1847 */
1848 if (sra == NULL ||
1849 sysfs_set_str(sra, NULL, "resync_start", "none") < 0)
1850 pr_err("--assume-clean not supported with --grow on this kernel\n");
1851 }
1852 ioctl(fd, GET_ARRAY_INFO, &array);
1853 s->size = get_component_size(fd)/2;
1854 if (s->size == 0)
1855 s->size = array.size;
1856 if (c->verbose >= 0) {
1857 if (s->size == orig_size)
1858 pr_err("component size of %s "
1859 "unchanged at %lluK\n",
1860 devname, s->size);
1861 else
1862 pr_err("component size of %s "
1863 "has been set to %lluK\n",
1864 devname, s->size);
1865 }
1866 changed = 1;
1867 } else if (array.level != LEVEL_CONTAINER) {
1868 s->size = get_component_size(fd)/2;
1869 if (s->size == 0)
1870 s->size = array.size;
1871 }
1872
1873 /* See if there is anything else to do */
1874 if ((s->level == UnSet || s->level == array.level) &&
1875 (s->layout_str == NULL) &&
1876 (s->chunk == 0 || s->chunk == array.chunk_size) &&
1877 data_offset == INVALID_SECTORS &&
1878 (s->raiddisks == 0 || s->raiddisks == array.raid_disks)) {
1879 /* Nothing more to do */
1880 if (!changed && c->verbose >= 0)
1881 pr_err("%s: no change requested\n",
1882 devname);
1883 goto release;
1884 }
1885
1886 /* ========= check for Raid10/Raid1 -> Raid0 conversion ===============
1887 * current implementation assumes that following conditions must be met:
1888 * - RAID10:
1889 * - far_copies == 1
1890 * - near_copies == 2
1891 */
1892 if ((s->level == 0 && array.level == 10 && sra &&
1893 array.layout == ((1 << 8) + 2) && !(array.raid_disks & 1)) ||
1894 (s->level == 0 && array.level == 1 && sra)) {
1895 int err;
1896 err = remove_disks_for_takeover(st, sra, array.layout);
1897 if (err) {
1898 dprintf(Name": Array cannot be reshaped\n");
1899 if (cfd > -1)
1900 close(cfd);
1901 rv = 1;
1902 goto release;
1903 }
1904 /* Make sure mdmon has seen the device removal
1905 * and updated metadata before we continue with
1906 * level change
1907 */
1908 if (container)
1909 ping_monitor(container);
1910 }
1911
1912 memset(&info, 0, sizeof(info));
1913 info.array = array;
1914 sysfs_init(&info, fd, NULL);
1915 strcpy(info.text_version, sra->text_version);
1916 info.component_size = s->size*2;
1917 info.new_level = s->level;
1918 info.new_chunk = s->chunk * 1024;
1919 if (info.array.level == LEVEL_CONTAINER) {
1920 info.delta_disks = UnSet;
1921 info.array.raid_disks = s->raiddisks;
1922 } else if (s->raiddisks)
1923 info.delta_disks = s->raiddisks - info.array.raid_disks;
1924 else
1925 info.delta_disks = UnSet;
1926 if (s->layout_str == NULL) {
1927 info.new_layout = UnSet;
1928 if (info.array.level == 6 &&
1929 (info.new_level == 6 || info.new_level == UnSet) &&
1930 info.array.layout >= 16) {
1931 pr_err("%s has a non-standard layout. If you"
1932 " wish to preserve this\n", devname);
1933 cont_err("during the reshape, please specify"
1934 " --layout=preserve\n");
1935 cont_err("If you want to change it, specify a"
1936 " layout or use --layout=normalise\n");
1937 rv = 1;
1938 goto release;
1939 }
1940 } else if (strcmp(s->layout_str, "normalise") == 0 ||
1941 strcmp(s->layout_str, "normalize") == 0) {
1942 /* If we have a -6 RAID6 layout, remove the '-6'. */
1943 info.new_layout = UnSet;
1944 if (info.array.level == 6 && info.new_level == UnSet) {
1945 char l[40], *h;
1946 strcpy(l, map_num(r6layout, info.array.layout));
1947 h = strrchr(l, '-');
1948 if (h && strcmp(h, "-6") == 0) {
1949 *h = 0;
1950 info.new_layout = map_name(r6layout, l);
1951 }
1952 } else {
1953 pr_err("%s is only meaningful when reshaping"
1954 " a RAID6 array.\n", s->layout_str);
1955 rv = 1;
1956 goto release;
1957 }
1958 } else if (strcmp(s->layout_str, "preserve") == 0) {
1959 /* This means that a non-standard RAID6 layout
1960 * is OK.
1961 * In particular:
1962 * - When reshape a RAID6 (e.g. adding a device)
1963 * which is in a non-standard layout, it is OK
1964 * to preserve that layout.
1965 * - When converting a RAID5 to RAID6, leave it in
1966 * the XXX-6 layout, don't re-layout.
1967 */
1968 if (info.array.level == 6 && info.new_level == UnSet)
1969 info.new_layout = info.array.layout;
1970 else if (info.array.level == 5 && info.new_level == 6) {
1971 char l[40];
1972 strcpy(l, map_num(r5layout, info.array.layout));
1973 strcat(l, "-6");
1974 info.new_layout = map_name(r6layout, l);
1975 } else {
1976 pr_err("%s in only meaningful when reshaping"
1977 " to RAID6\n", s->layout_str);
1978 rv = 1;
1979 goto release;
1980 }
1981 } else {
1982 int l = info.new_level;
1983 if (l == UnSet)
1984 l = info.array.level;
1985 switch (l) {
1986 case 5:
1987 info.new_layout = map_name(r5layout, s->layout_str);
1988 break;
1989 case 6:
1990 info.new_layout = map_name(r6layout, s->layout_str);
1991 break;
1992 case 10:
1993 info.new_layout = parse_layout_10(s->layout_str);
1994 break;
1995 case LEVEL_FAULTY:
1996 info.new_layout = parse_layout_faulty(s->layout_str);
1997 break;
1998 default:
1999 pr_err("layout not meaningful"
2000 " with this level\n");
2001 rv = 1;
2002 goto release;
2003 }
2004 if (info.new_layout == UnSet) {
2005 pr_err("layout %s not understood"
2006 " for this level\n",
2007 s->layout_str);
2008 rv = 1;
2009 goto release;
2010 }
2011 }
2012
2013 if (array.level == LEVEL_FAULTY) {
2014 if (s->level != UnSet && s->level != array.level) {
2015 pr_err("cannot change level of Faulty device\n");
2016 rv =1 ;
2017 }
2018 if (s->chunk) {
2019 pr_err("cannot set chunksize of Faulty device\n");
2020 rv =1 ;
2021 }
2022 if (s->raiddisks && s->raiddisks != 1) {
2023 pr_err("cannot set raid_disks of Faulty device\n");
2024 rv =1 ;
2025 }
2026 if (s->layout_str) {
2027 if (ioctl(fd, GET_ARRAY_INFO, &array) != 0) {
2028 dprintf("Cannot get array information.\n");
2029 goto release;
2030 }
2031 array.layout = info.new_layout;
2032 if (ioctl(fd, SET_ARRAY_INFO, &array) != 0) {
2033 pr_err("failed to set new layout\n");
2034 rv = 1;
2035 } else if (c->verbose >= 0)
2036 printf("layout for %s set to %d\n",
2037 devname, array.layout);
2038 }
2039 } else if (array.level == LEVEL_CONTAINER) {
2040 /* This change is to be applied to every array in the
2041 * container. This is only needed when the metadata imposes
2042 * restraints of the various arrays in the container.
2043 * Currently we only know that IMSM requires all arrays
2044 * to have the same number of devices so changing the
2045 * number of devices (On-Line Capacity Expansion) must be
2046 * performed at the level of the container
2047 */
2048 rv = reshape_container(container, devname, -1, st, &info,
2049 c->force, c->backup_file, c->verbose, 0, 0);
2050 frozen = 0;
2051 } else {
2052 /* get spare devices from external metadata
2053 */
2054 if (st->ss->external) {
2055 struct mdinfo *info2;
2056
2057 info2 = st->ss->container_content(st, subarray);
2058 if (info2) {
2059 info.array.spare_disks =
2060 info2->array.spare_disks;
2061 sysfs_free(info2);
2062 }
2063 }
2064
2065 /* Impose these changes on a single array. First
2066 * check that the metadata is OK with the change. */
2067
2068 if (reshape_super(st, 0, info.new_level,
2069 info.new_layout, info.new_chunk,
2070 info.array.raid_disks, info.delta_disks,
2071 c->backup_file, devname, APPLY_METADATA_CHANGES,
2072 c->verbose)) {
2073 rv = 1;
2074 goto release;
2075 }
2076 sync_metadata(st);
2077 rv = reshape_array(container, fd, devname, st, &info, c->force,
2078 devlist, data_offset, c->backup_file, c->verbose,
2079 0, 0, 0);
2080 frozen = 0;
2081 }
2082 release:
2083 sysfs_free(sra);
2084 if (frozen > 0)
2085 unfreeze(st);
2086 return rv;
2087 }
2088
2089 /* verify_reshape_position()
2090 * Function checks if reshape position in metadata is not farther
2091 * than position in md.
2092 * Return value:
2093 * 0 : not valid sysfs entry
2094 * it can be caused by not started reshape, it should be started
2095 * by reshape array or raid0 array is before takeover
2096 * -1 : error, reshape position is obviously wrong
2097 * 1 : success, reshape progress correct or updated
2098 */
2099 static int verify_reshape_position(struct mdinfo *info, int level)
2100 {
2101 int ret_val = 0;
2102 char buf[40];
2103 int rv;
2104
2105 /* read sync_max, failure can mean raid0 array */
2106 rv = sysfs_get_str(info, NULL, "sync_max", buf, 40);
2107
2108 if (rv > 0) {
2109 char *ep;
2110 unsigned long long position = strtoull(buf, &ep, 0);
2111
2112 dprintf(Name": Read sync_max sysfs entry is: %s\n", buf);
2113 if (!(ep == buf || (*ep != 0 && *ep != '\n' && *ep != ' '))) {
2114 position *= get_data_disks(level,
2115 info->new_layout,
2116 info->array.raid_disks);
2117 if (info->reshape_progress < position) {
2118 dprintf("Corrected reshape progress (%llu) to "
2119 "md position (%llu)\n",
2120 info->reshape_progress, position);
2121 info->reshape_progress = position;
2122 ret_val = 1;
2123 } else if (info->reshape_progress > position) {
2124 pr_err("Fatal error: array "
2125 "reshape was not properly frozen "
2126 "(expected reshape position is %llu, "
2127 "but reshape progress is %llu.\n",
2128 position, info->reshape_progress);
2129 ret_val = -1;
2130 } else {
2131 dprintf("Reshape position in md and metadata "
2132 "are the same;");
2133 ret_val = 1;
2134 }
2135 }
2136 } else if (rv == 0) {
2137 /* for valid sysfs entry, 0-length content
2138 * should be indicated as error
2139 */
2140 ret_val = -1;
2141 }
2142
2143 return ret_val;
2144 }
2145
2146 static unsigned long long choose_offset(unsigned long long lo,
2147 unsigned long long hi,
2148 unsigned long long min,
2149 unsigned long long max)
2150 {
2151 /* Choose a new offset between hi and lo.
2152 * It must be between min and max, but
2153 * we would prefer something near the middle of hi/lo, and also
2154 * prefer to be aligned to a big power of 2.
2155 *
2156 * So we start with the middle, then for each bit,
2157 * starting at '1' and increasing, if it is set, we either
2158 * add it or subtract it if possible, preferring the option
2159 * which is furthest from the boundary.
2160 *
2161 * We stop once we get a 1MB alignment. As units are in sectors,
2162 * 1MB = 2*1024 sectors.
2163 */
2164 unsigned long long choice = (lo + hi) / 2;
2165 unsigned long long bit = 1;
2166
2167 for (bit = 1; bit < 2*1024; bit = bit << 1) {
2168 unsigned long long bigger, smaller;
2169 if (! (bit & choice))
2170 continue;
2171 bigger = choice + bit;
2172 smaller = choice - bit;
2173 if (bigger > max && smaller < min)
2174 break;
2175 if (bigger > max)
2176 choice = smaller;
2177 else if (smaller < min)
2178 choice = bigger;
2179 else if (hi - bigger > smaller - lo)
2180 choice = bigger;
2181 else
2182 choice = smaller;
2183 }
2184 return choice;
2185 }
2186
2187 static int set_new_data_offset(struct mdinfo *sra, struct supertype *st,
2188 char *devname, int delta_disks,
2189 unsigned long long data_offset,
2190 unsigned long long min,
2191 int can_fallback)
2192 {
2193 struct mdinfo *sd;
2194 int dir = 0;
2195 int err = 0;
2196 unsigned long long before, after;
2197
2198 /* Need to find min space before and after so same is used
2199 * on all devices
2200 */
2201 before = UINT64_MAX;
2202 after = UINT64_MAX;
2203 for (sd = sra->devs; sd; sd = sd->next) {
2204 char *dn;
2205 int dfd;
2206 int rv;
2207 struct supertype *st2;
2208 struct mdinfo info2;
2209
2210 if (sd->disk.state & (1<<MD_DISK_FAULTY))
2211 continue;
2212 dn = map_dev(sd->disk.major, sd->disk.minor, 0);
2213 dfd = dev_open(dn, O_RDONLY);
2214 if (dfd < 0) {
2215 pr_err("%s: cannot open component %s\n",
2216 devname, dn ? dn : "-unknown-");
2217 goto release;
2218 }
2219 st2 = dup_super(st);
2220 rv = st2->ss->load_super(st2,dfd, NULL);
2221 close(dfd);
2222 if (rv) {
2223 free(st2);
2224 pr_err("%s: cannot get superblock from %s\n",
2225 devname, dn);
2226 goto release;
2227 }
2228 st2->ss->getinfo_super(st2, &info2, NULL);
2229 st2->ss->free_super(st2);
2230 free(st2);
2231 if (info2.space_before == 0 &&
2232 info2.space_after == 0) {
2233 /* Metadata doesn't support data_offset changes */
2234 return 1;
2235 }
2236 if (before > info2.space_before)
2237 before = info2.space_before;
2238 if (after > info2.space_after)
2239 after = info2.space_after;
2240
2241 if (data_offset != INVALID_SECTORS) {
2242 if (dir == 0) {
2243 if (info2.data_offset == data_offset) {
2244 pr_err("%s: already has that data_offset\n",
2245 dn);
2246 goto release;
2247 }
2248 if (data_offset < info2.data_offset)
2249 dir = -1;
2250 else
2251 dir = 1;
2252 } else if ((data_offset <= info2.data_offset && dir == 1) ||
2253 (data_offset >= info2.data_offset && dir == -1)) {
2254 pr_err("%s: differing data offsets on devices make this --data-offset setting impossible\n",
2255 dn);
2256 goto release;
2257 }
2258 }
2259 }
2260 if (before == UINT64_MAX)
2261 /* impossible really, there must be no devices */
2262 return 1;
2263
2264 for (sd = sra->devs; sd; sd = sd->next) {
2265 char *dn = map_dev(sd->disk.major, sd->disk.minor, 0);
2266 unsigned long long new_data_offset;
2267
2268 if (sd->disk.state & (1<<MD_DISK_FAULTY))
2269 continue;
2270 if (delta_disks < 0) {
2271 /* Don't need any space as array is shrinking
2272 * just move data_offset up by min
2273 */
2274 if (data_offset == INVALID_SECTORS)
2275 new_data_offset = sd->data_offset + min;
2276 else {
2277 if (data_offset < sd->data_offset + min) {
2278 pr_err("--data-offset too small for %s\n",
2279 dn);
2280 goto release;
2281 }
2282 new_data_offset = data_offset;
2283 }
2284 } else if (delta_disks > 0) {
2285 /* need space before */
2286 if (before < min) {
2287 if (can_fallback)
2288 goto fallback;
2289 pr_err("Insufficient head-space for reshape on %s\n",
2290 dn);
2291 goto release;
2292 }
2293 if (data_offset == INVALID_SECTORS)
2294 new_data_offset = sd->data_offset - min;
2295 else {
2296 if (data_offset > sd->data_offset - min) {
2297 pr_err("--data-offset too large for %s\n",
2298 dn);
2299 goto release;
2300 }
2301 new_data_offset = data_offset;
2302 }
2303 } else {
2304 if (dir == 0) {
2305 /* can move up or down. If 'data_offset'
2306 * was set we would have already decided,
2307 * so just choose direction with most space.
2308 */
2309 if (before > after)
2310 dir = -1;
2311 else
2312 dir = 1;
2313 }
2314 sysfs_set_str(sra, NULL, "reshape_direction",
2315 dir == 1 ? "backwards" : "forwards");
2316 if (dir > 0) {
2317 /* Increase data offset */
2318 if (after < min) {
2319 if (can_fallback)
2320 goto fallback;
2321 pr_err("Insufficient tail-space for reshape on %s\n",
2322 dn);
2323 goto release;
2324 }
2325 if (data_offset != INVALID_SECTORS &&
2326 data_offset < sd->data_offset + min) {
2327 pr_err("--data-offset too small on %s\n",
2328 dn);
2329 goto release;
2330 }
2331 if (data_offset != INVALID_SECTORS)
2332 new_data_offset = data_offset;
2333 else
2334 new_data_offset = choose_offset(sd->data_offset,
2335 sd->data_offset + after,
2336 sd->data_offset + min,
2337 sd->data_offset + after);
2338 } else {
2339 /* Decrease data offset */
2340 if (before < min) {
2341 if (can_fallback)
2342 goto fallback;
2343 pr_err("insufficient head-room on %s\n",
2344 dn);
2345 goto release;
2346 }
2347 if (data_offset != INVALID_SECTORS &&
2348 data_offset < sd->data_offset - min) {
2349 pr_err("--data-offset too small on %s\n",
2350 dn);
2351 goto release;
2352 }
2353 if (data_offset != INVALID_SECTORS)
2354 new_data_offset = data_offset;
2355 else
2356 new_data_offset = choose_offset(sd->data_offset - before,
2357 sd->data_offset,
2358 sd->data_offset - before,
2359 sd->data_offset - min);
2360 }
2361 }
2362 err = sysfs_set_num(sra, sd, "new_offset", new_data_offset);
2363 if (err < 0 && errno == E2BIG) {
2364 /* try again after increasing data size to max */
2365 err = sysfs_set_num(sra, sd, "size", 0);
2366 if (err < 0 && errno == EINVAL &&
2367 !(sd->disk.state & (1<<MD_DISK_SYNC))) {
2368 /* some kernels have a bug where you cannot
2369 * use '0' on spare devices. */
2370 sysfs_set_num(sra, sd, "size",
2371 (sra->component_size + after)/2);
2372 }
2373 err = sysfs_set_num(sra, sd, "new_offset",
2374 new_data_offset);
2375 }
2376 if (err < 0) {
2377 if (errno == E2BIG && data_offset != INVALID_SECTORS) {
2378 pr_err("data-offset is too big for %s\n",
2379 dn);
2380 goto release;
2381 }
2382 if (sd == sra->devs &&
2383 (errno == ENOENT || errno == E2BIG))
2384 /* Early kernel, no 'new_offset' file,
2385 * or kernel doesn't like us.
2386 * For RAID5/6 this is not fatal
2387 */
2388 return 1;
2389 pr_err("Cannot set new_offset for %s\n",
2390 dn);
2391 break;
2392 }
2393 }
2394 return err;
2395 release:
2396 return -1;
2397 fallback:
2398 /* Just use a backup file */
2399 return 1;
2400 }
2401
2402 static int raid10_reshape(char *container, int fd, char *devname,
2403 struct supertype *st, struct mdinfo *info,
2404 struct reshape *reshape,
2405 unsigned long long data_offset,
2406 int force, int verbose)
2407 {
2408 /* Changing raid_disks, layout, chunksize or possibly
2409 * just data_offset for a RAID10.
2410 * We must always change data_offset. We change by at least
2411 * ->min_offset_change which is the largest of the old and new
2412 * chunk sizes.
2413 * If raid_disks is increasing, then data_offset must decrease
2414 * by at least this copy size.
2415 * If raid_disks is unchanged, data_offset must increase or
2416 * decrease by at least min_offset_change but preferably by much more.
2417 * We choose half of the available space.
2418 * If raid_disks is decreasing, data_offset must increase by
2419 * at least min_offset_change. To allow of this, component_size
2420 * must be decreased by the same amount.
2421 *
2422 * So we calculate the required minimum and direction, possibly
2423 * reduce the component_size, then iterate through the devices
2424 * and set the new_data_offset.
2425 * If that all works, we set chunk_size, layout, raid_disks, and start
2426 * 'reshape'
2427 */
2428 struct mdinfo *sra;
2429 unsigned long long min;
2430 int err = 0;
2431
2432 sra = sysfs_read(fd, NULL,
2433 GET_COMPONENT|GET_DEVS|GET_OFFSET|GET_STATE|GET_CHUNK
2434 );
2435 if (!sra) {
2436 pr_err("%s: Cannot get array details from sysfs\n",
2437 devname);
2438 goto release;
2439 }
2440 min = reshape->min_offset_change;
2441
2442 if (info->delta_disks)
2443 sysfs_set_str(sra, NULL, "reshape_direction",
2444 info->delta_disks < 0 ? "backwards" : "forwards");
2445 if (info->delta_disks < 0 &&
2446 info->space_after < min) {
2447 int rv = sysfs_set_num(sra, NULL, "component_size",
2448 (sra->component_size -
2449 min)/2);
2450 if (rv) {
2451 pr_err("cannot reduce component size\n");
2452 goto release;
2453 }
2454 }
2455 err = set_new_data_offset(sra, st, devname, info->delta_disks, data_offset,
2456 min, 0);
2457 if (err == 1) {
2458 pr_err("Cannot set new_data_offset: RAID10 reshape not\n");
2459 cont_err("supported on this kernel\n");
2460 err = -1;
2461 }
2462 if (err < 0)
2463 goto release;
2464
2465 if (!err && sysfs_set_num(sra, NULL, "chunk_size", info->new_chunk) < 0)
2466 err = errno;
2467 if (!err && sysfs_set_num(sra, NULL, "layout", reshape->after.layout) < 0)
2468 err = errno;
2469 if (!err && sysfs_set_num(sra, NULL, "raid_disks",
2470 info->array.raid_disks + info->delta_disks) < 0)
2471 err = errno;
2472 if (!err && sysfs_set_str(sra, NULL, "sync_action", "reshape") < 0)
2473 err = errno;
2474 if (err) {
2475 pr_err("Cannot set array shape for %s\n",
2476 devname);
2477 if (err == EBUSY &&
2478 (info->array.state & (1<<MD_SB_BITMAP_PRESENT)))
2479 cont_err(" Bitmap must be removed before"
2480 " shape can be changed\n");
2481 goto release;
2482 }
2483 sysfs_free(sra);
2484 return 0;
2485 release:
2486 sysfs_free(sra);
2487 return 1;
2488 }
2489
2490 static void get_space_after(int fd, struct supertype *st, struct mdinfo *info)
2491 {
2492 struct mdinfo *sra, *sd;
2493 /* Initialisation to silence compiler warning */
2494 unsigned long long min_space_before = 0, min_space_after = 0;
2495 int first = 1;
2496
2497 sra = sysfs_read(fd, NULL, GET_DEVS);
2498 if (!sra)
2499 return;
2500 for (sd = sra->devs; sd; sd = sd->next) {
2501 char *dn;
2502 int dfd;
2503 struct supertype *st2;
2504 struct mdinfo info2;
2505
2506 if (sd->disk.state & (1<<MD_DISK_FAULTY))
2507 continue;
2508 dn = map_dev(sd->disk.major, sd->disk.minor, 0);
2509 dfd = dev_open(dn, O_RDONLY);
2510 if (dfd < 0)
2511 break;
2512 st2 = dup_super(st);
2513 if (st2->ss->load_super(st2,dfd, NULL)) {
2514 close(dfd);
2515 free(st2);
2516 break;
2517 }
2518 close(dfd);
2519 st2->ss->getinfo_super(st2, &info2, NULL);
2520 st2->ss->free_super(st2);
2521 free(st2);
2522 if (first ||
2523 min_space_before > info2.space_before)
2524 min_space_before = info2.space_before;
2525 if (first ||
2526 min_space_after > info2.space_after)
2527 min_space_after = info2.space_after;
2528 first = 0;
2529 }
2530 if (sd == NULL && !first) {
2531 info->space_after = min_space_after;
2532 info->space_before = min_space_before;
2533 }
2534 sysfs_free(sra);
2535 }
2536
2537 static void update_cache_size(char *container, struct mdinfo *sra,
2538 struct mdinfo *info,
2539 int disks, unsigned long long blocks)
2540 {
2541 /* Check that the internal stripe cache is
2542 * large enough, or it won't work.
2543 * It must hold at least 4 stripes of the larger
2544 * chunk size
2545 */
2546 unsigned long cache;
2547 cache = max(info->array.chunk_size, info->new_chunk);
2548 cache *= 4; /* 4 stripes minimum */
2549 cache /= 512; /* convert to sectors */
2550 /* make sure there is room for 'blocks' with a bit to spare */
2551 if (cache < 16 + blocks / disks)
2552 cache = 16 + blocks / disks;
2553 cache /= (4096/512); /* Covert from sectors to pages */
2554
2555 if (sra->cache_size < cache)
2556 subarray_set_num(container, sra, "stripe_cache_size",
2557 cache+1);
2558 }
2559
2560 static int impose_reshape(struct mdinfo *sra,
2561 struct mdinfo *info,
2562 struct supertype *st,
2563 int fd,
2564 int restart,
2565 char *devname, char *container,
2566 struct reshape *reshape)
2567 {
2568 struct mdu_array_info_s array;
2569
2570 sra->new_chunk = info->new_chunk;
2571
2572 if (restart) {
2573 /* for external metadata checkpoint saved by mdmon can be lost
2574 * or missed /due to e.g. crash/. Check if md is not during
2575 * restart farther than metadata points to.
2576 * If so, this means metadata information is obsolete.
2577 */
2578 if (st->ss->external)
2579 verify_reshape_position(info, reshape->level);
2580 sra->reshape_progress = info->reshape_progress;
2581 } else {
2582 sra->reshape_progress = 0;
2583 if (reshape->after.data_disks < reshape->before.data_disks)
2584 /* start from the end of the new array */
2585 sra->reshape_progress = (sra->component_size
2586 * reshape->after.data_disks);
2587 }
2588
2589 ioctl(fd, GET_ARRAY_INFO, &array);
2590 if (info->array.chunk_size == info->new_chunk &&
2591 reshape->before.layout == reshape->after.layout &&
2592 st->ss->external == 0) {
2593 /* use SET_ARRAY_INFO but only if reshape hasn't started */
2594 array.raid_disks = reshape->after.data_disks + reshape->parity;
2595 if (!restart &&
2596 ioctl(fd, SET_ARRAY_INFO, &array) != 0) {
2597 int err = errno;
2598
2599 pr_err("Cannot set device shape for %s: %s\n",
2600 devname, strerror(errno));
2601
2602 if (err == EBUSY &&
2603 (array.state & (1<<MD_SB_BITMAP_PRESENT)))
2604 cont_err("Bitmap must be removed before"
2605 " shape can be changed\n");
2606
2607 goto release;
2608 }
2609 } else if (!restart) {
2610 /* set them all just in case some old 'new_*' value
2611 * persists from some earlier problem.
2612 */
2613 int err = 0;
2614 if (sysfs_set_num(sra, NULL, "chunk_size", info->new_chunk) < 0)
2615 err = errno;
2616 if (!err && sysfs_set_num(sra, NULL, "layout",
2617 reshape->after.layout) < 0)
2618 err = errno;
2619 if (!err && subarray_set_num(container, sra, "raid_disks",
2620 reshape->after.data_disks +
2621 reshape->parity) < 0)
2622 err = errno;
2623 if (err) {
2624 pr_err("Cannot set device shape for %s\n",
2625 devname);
2626
2627 if (err == EBUSY &&
2628 (array.state & (1<<MD_SB_BITMAP_PRESENT)))
2629 cont_err("Bitmap must be removed before"
2630 " shape can be changed\n");
2631 goto release;
2632 }
2633 }
2634 return 0;
2635 release:
2636 return -1;
2637 }
2638
2639 static int impose_level(int fd, int level, char *devname, int verbose)
2640 {
2641 char *c;
2642 struct mdu_array_info_s array;
2643 struct mdinfo info;
2644 sysfs_init(&info, fd, NULL);
2645
2646 ioctl(fd, GET_ARRAY_INFO, &array);
2647 if (level == 0 &&
2648 (array.level >= 4 && array.level <= 6)) {
2649 /* To convert to RAID0 we need to fail and
2650 * remove any non-data devices. */
2651 int found = 0;
2652 int d;
2653 int data_disks = array.raid_disks - 1;
2654 if (array.level == 6)
2655 data_disks -= 1;
2656 if (array.level == 5 &&
2657 array.layout != ALGORITHM_PARITY_N)
2658 return -1;
2659 if (array.level == 6 &&
2660 array.layout != ALGORITHM_PARITY_N_6)
2661 return -1;
2662 sysfs_set_str(&info, NULL,"sync_action", "idle");
2663 /* First remove any spares so no recovery starts */
2664 for (d = 0, found = 0;
2665 d < MAX_DISKS && found < array.nr_disks;
2666 d++) {
2667 mdu_disk_info_t disk;
2668 disk.number = d;
2669 if (ioctl(fd, GET_DISK_INFO, &disk) < 0)
2670 continue;
2671 if (disk.major == 0 && disk.minor == 0)
2672 continue;
2673 found++;
2674 if ((disk.state & (1 << MD_DISK_ACTIVE))
2675 && disk.raid_disk < data_disks)
2676 /* keep this */
2677 continue;
2678 ioctl(fd, HOT_REMOVE_DISK,
2679 makedev(disk.major, disk.minor));
2680 }
2681 /* Now fail anything left */
2682 ioctl(fd, GET_ARRAY_INFO, &array);
2683 for (d = 0, found = 0;
2684 d < MAX_DISKS && found < array.nr_disks;
2685 d++) {
2686 int cnt;
2687 mdu_disk_info_t disk;
2688 disk.number = d;
2689 if (ioctl(fd, GET_DISK_INFO, &disk) < 0)
2690 continue;
2691 if (disk.major == 0 && disk.minor == 0)
2692 continue;
2693 found++;
2694 if ((disk.state & (1 << MD_DISK_ACTIVE))
2695 && disk.raid_disk < data_disks)
2696 /* keep this */
2697 continue;
2698 ioctl(fd, SET_DISK_FAULTY,
2699 makedev(disk.major, disk.minor));
2700 cnt = 5;
2701 while (ioctl(fd, HOT_REMOVE_DISK,
2702 makedev(disk.major, disk.minor)) < 0
2703 && errno == EBUSY
2704 && cnt--) {
2705 usleep(10000);
2706 }
2707 }
2708 }
2709 c = map_num(pers, level);
2710 if (c) {
2711 int err = sysfs_set_str(&info, NULL, "level", c);
2712 if (err) {
2713 err = errno;
2714 pr_err("%s: could not set level to %s\n",
2715 devname, c);
2716 if (err == EBUSY &&
2717 (array.state & (1<<MD_SB_BITMAP_PRESENT)))
2718 cont_err("Bitmap must be removed"
2719 " before level can be changed\n");
2720 return err;
2721 }
2722 if (verbose >= 0)
2723 pr_err("level of %s changed to %s\n",
2724 devname, c);
2725 }
2726 return 0;
2727 }
2728
2729 static int reshape_array(char *container, int fd, char *devname,
2730 struct supertype *st, struct mdinfo *info,
2731 int force, struct mddev_dev *devlist,
2732 unsigned long long data_offset,
2733 char *backup_file, int verbose, int forked,
2734 int restart, int freeze_reshape)
2735 {
2736 struct reshape reshape;
2737 int spares_needed;
2738 char *msg;
2739 int orig_level = UnSet;
2740 int odisks;
2741 int delayed;
2742
2743 struct mdu_array_info_s array;
2744 char *c;
2745
2746 struct mddev_dev *dv;
2747 int added_disks;
2748
2749 int *fdlist = NULL;
2750 unsigned long long *offsets = NULL;
2751 int d;
2752 int nrdisks;
2753 int err;
2754 unsigned long blocks;
2755 unsigned long long array_size;
2756 int done;
2757 struct mdinfo *sra = NULL;
2758
2759 /* when reshaping a RAID0, the component_size might be zero.
2760 * So try to fix that up.
2761 */
2762 if (ioctl(fd, GET_ARRAY_INFO, &array) != 0) {
2763 dprintf("Cannot get array information.\n");
2764 goto release;
2765 }
2766 if (array.level == 0 && info->component_size == 0) {
2767 get_dev_size(fd, NULL, &array_size);
2768 info->component_size = array_size / array.raid_disks;
2769 }
2770
2771 if (array.level == 10)
2772 /* Need space_after info */
2773 get_space_after(fd, st, info);
2774
2775 if (info->reshape_active) {
2776 int new_level = info->new_level;
2777 info->new_level = UnSet;
2778 if (info->delta_disks > 0)
2779 info->array.raid_disks -= info->delta_disks;
2780 msg = analyse_change(devname, info, &reshape);
2781 info->new_level = new_level;
2782 if (info->delta_disks > 0)
2783 info->array.raid_disks += info->delta_disks;
2784 if (!restart)
2785 /* Make sure the array isn't read-only */
2786 ioctl(fd, RESTART_ARRAY_RW, 0);
2787 } else
2788 msg = analyse_change(devname, info, &reshape);
2789 if (msg) {
2790 /* if msg == "", error has already been printed */
2791 if (msg[0])
2792 pr_err("%s\n", msg);
2793 goto release;
2794 }
2795 if (restart &&
2796 (reshape.level != info->array.level ||
2797 reshape.before.layout != info->array.layout ||
2798 reshape.before.data_disks + reshape.parity
2799 != info->array.raid_disks - max(0, info->delta_disks))) {
2800 pr_err("reshape info is not in native format -"
2801 " cannot continue.\n");
2802 goto release;
2803 }
2804
2805 if (st->ss->external && restart && (info->reshape_progress == 0)) {
2806 /* When reshape is restarted from '0', very begin of array
2807 * it is possible that for external metadata reshape and array
2808 * configuration doesn't happen.
2809 * Check if md has the same opinion, and reshape is restarted
2810 * from 0. If so, this is regular reshape start after reshape
2811 * switch in metadata to next array only.
2812 */
2813 if ((verify_reshape_position(info, reshape.level) >= 0) &&
2814 (info->reshape_progress == 0))
2815 restart = 0;
2816 }
2817 if (restart) {
2818 /* reshape already started. just skip to monitoring the reshape */
2819 if (reshape.backup_blocks == 0)
2820 return 0;
2821 if (restart & RESHAPE_NO_BACKUP)
2822 return 0;
2823
2824 /* Need 'sra' down at 'started:' */
2825 sra = sysfs_read(fd, NULL,
2826 GET_COMPONENT|GET_DEVS|GET_OFFSET|GET_STATE|GET_CHUNK|
2827 GET_CACHE);
2828 if (!sra) {
2829 pr_err("%s: Cannot get array details from sysfs\n",
2830 devname);
2831 goto release;
2832 }
2833 goto started;
2834 }
2835 /* The container is frozen but the array may not be.
2836 * So freeze the array so spares don't get put to the wrong use
2837 * FIXME there should probably be a cleaner separation between
2838 * freeze_array and freeze_container.
2839 */
2840 sysfs_freeze_array(info);
2841 /* Check we have enough spares to not be degraded */
2842 added_disks = 0;
2843 for (dv = devlist; dv ; dv=dv->next)
2844 added_disks++;
2845 spares_needed = max(reshape.before.data_disks,
2846 reshape.after.data_disks)
2847 + reshape.parity - array.raid_disks;
2848
2849 if (!force &&
2850 info->new_level > 1 && info->array.level > 1 &&
2851 spares_needed > info->array.spare_disks + added_disks) {
2852 pr_err("Need %d spare%s to avoid degraded array,"
2853 " and only have %d.\n"
2854 " Use --force to over-ride this check.\n",
2855 spares_needed,
2856 spares_needed == 1 ? "" : "s",
2857 info->array.spare_disks + added_disks);
2858 goto release;
2859 }
2860 /* Check we have enough spares to not fail */
2861 spares_needed = max(reshape.before.data_disks,
2862 reshape.after.data_disks)
2863 - array.raid_disks;
2864 if ((info->new_level > 1 || info->new_level == 0) &&
2865 spares_needed > info->array.spare_disks +added_disks) {
2866 pr_err("Need %d spare%s to create working array,"
2867 " and only have %d.\n",
2868 spares_needed,
2869 spares_needed == 1 ? "" : "s",
2870 info->array.spare_disks + added_disks);
2871 goto release;
2872 }
2873
2874 if (reshape.level != array.level) {
2875 int err = impose_level(fd, reshape.level, devname, verbose);
2876 if (err)
2877 goto release;
2878 info->new_layout = UnSet; /* after level change,
2879 * layout is meaningless */
2880 orig_level = array.level;
2881 sysfs_freeze_array(info);
2882
2883 if (reshape.level > 0 && st->ss->external) {
2884 /* make sure mdmon is aware of the new level */
2885 if (mdmon_running(container))
2886 flush_mdmon(container);
2887
2888 if (!mdmon_running(container))
2889 start_mdmon(container);
2890 ping_monitor(container);
2891 if (mdmon_running(container) &&
2892 st->update_tail == NULL)
2893 st->update_tail = &st->updates;
2894 }
2895 }
2896 /* ->reshape_super might have chosen some spares from the
2897 * container that it wants to be part of the new array.
2898 * We can collect them with ->container_content and give
2899 * them to the kernel.
2900 */
2901 if (st->ss->reshape_super && st->ss->container_content) {
2902 char *subarray = strchr(info->text_version+1, '/')+1;
2903 struct mdinfo *info2 =
2904 st->ss->container_content(st, subarray);
2905 struct mdinfo *d;
2906
2907 if (info2) {
2908 sysfs_init(info2, fd, st->devnm);
2909 /* When increasing number of devices, we need to set
2910 * new raid_disks before adding these, or they might
2911 * be rejected.
2912 */
2913 if (reshape.backup_blocks &&
2914 reshape.after.data_disks > reshape.before.data_disks)
2915 subarray_set_num(container, info2, "raid_disks",
2916 reshape.after.data_disks +
2917 reshape.parity);
2918 for (d = info2->devs; d; d = d->next) {
2919 if (d->disk.state == 0 &&
2920 d->disk.raid_disk >= 0) {
2921 /* This is a spare that wants to
2922 * be part of the array.
2923 */
2924 add_disk(fd, st, info2, d);
2925 }
2926 }
2927 sysfs_free(info2);
2928 }
2929 }
2930 /* We might have been given some devices to add to the
2931 * array. Now that the array has been changed to the right
2932 * level and frozen, we can safely add them.
2933 */
2934 if (devlist)
2935 Manage_subdevs(devname, fd, devlist, verbose,
2936 0,NULL, 0);
2937
2938 if (reshape.backup_blocks == 0 && data_offset != INVALID_SECTORS)
2939 reshape.backup_blocks = reshape.before.data_disks * info->array.chunk_size/512;
2940 if (reshape.backup_blocks == 0) {
2941 /* No restriping needed, but we might need to impose
2942 * some more changes: layout, raid_disks, chunk_size
2943 */
2944 /* read current array info */
2945 if (ioctl(fd, GET_ARRAY_INFO, &array) != 0) {
2946 dprintf("Cannot get array information.\n");
2947 goto release;
2948 }
2949 /* compare current array info with new values and if
2950 * it is different update them to new */
2951 if (info->new_layout != UnSet &&
2952 info->new_layout != array.layout) {
2953 array.layout = info->new_layout;
2954 if (ioctl(fd, SET_ARRAY_INFO, &array) != 0) {
2955 pr_err("failed to set new layout\n");
2956 goto release;
2957 } else if (verbose >= 0)
2958 printf("layout for %s set to %d\n",
2959 devname, array.layout);
2960 }
2961 if (info->delta_disks != UnSet &&
2962 info->delta_disks != 0 &&
2963 array.raid_disks != (info->array.raid_disks + info->delta_disks)) {
2964 array.raid_disks += info->delta_disks;
2965 if (ioctl(fd, SET_ARRAY_INFO, &array) != 0) {
2966 pr_err("failed to set raid disks\n");
2967 goto release;
2968 } else if (verbose >= 0) {
2969 printf("raid_disks for %s set to %d\n",
2970 devname, array.raid_disks);
2971 }
2972 }
2973 if (info->new_chunk != 0 &&
2974 info->new_chunk != array.chunk_size) {
2975 if (sysfs_set_num(info, NULL,
2976 "chunk_size", info->new_chunk) != 0) {
2977 pr_err("failed to set chunk size\n");
2978 goto release;
2979 } else if (verbose >= 0)
2980 printf("chunk size for %s set to %d\n",
2981 devname, array.chunk_size);
2982 }
2983 unfreeze(st);
2984 return 0;
2985 }
2986
2987 /*
2988 * There are three possibilities.
2989 * 1/ The array will shrink.
2990 * We need to ensure the reshape will pause before reaching
2991 * the 'critical section'. We also need to fork and wait for
2992 * that to happen. When it does we
2993 * suspend/backup/complete/unfreeze
2994 *
2995 * 2/ The array will not change size.
2996 * This requires that we keep a backup of a sliding window
2997 * so that we can restore data after a crash. So we need
2998 * to fork and monitor progress.
2999 * In future we will allow the data_offset to change, so
3000 * a sliding backup becomes unnecessary.
3001 *
3002 * 3/ The array will grow. This is relatively easy.
3003 * However the kernel's restripe routines will cheerfully
3004 * overwrite some early data before it is safe. So we
3005 * need to make a backup of the early parts of the array
3006 * and be ready to restore it if rebuild aborts very early.
3007 * For externally managed metadata, we still need a forked
3008 * child to monitor the reshape and suspend IO over the region
3009 * that is being reshaped.
3010 *
3011 * We backup data by writing it to one spare, or to a
3012 * file which was given on command line.
3013 *
3014 * In each case, we first make sure that storage is available
3015 * for the required backup.
3016 * Then we:
3017 * - request the shape change.
3018 * - fork to handle backup etc.
3019 */
3020 /* Check that we can hold all the data */
3021 get_dev_size(fd, NULL, &array_size);
3022 if (reshape.new_size < (array_size/512)) {
3023 pr_err("this change will reduce the size of the array.\n"
3024 " use --grow --array-size first to truncate array.\n"
3025 " e.g. mdadm --grow %s --array-size %llu\n",
3026 devname, reshape.new_size/2);
3027 goto release;
3028 }
3029
3030 if (array.level == 10) {
3031 /* Reshaping RAID10 does not require any data backup by
3032 * user-space. Instead it requires that the data_offset
3033 * is changed to avoid the need for backup.
3034 * So this is handled very separately
3035 */
3036 if (restart)
3037 /* Nothing to do. */
3038 return 0;
3039 return raid10_reshape(container, fd, devname, st, info,
3040 &reshape, data_offset,
3041 force, verbose);
3042 }
3043 sra = sysfs_read(fd, NULL,
3044 GET_COMPONENT|GET_DEVS|GET_OFFSET|GET_STATE|GET_CHUNK|
3045 GET_CACHE);
3046 if (!sra) {
3047 pr_err("%s: Cannot get array details from sysfs\n",
3048 devname);
3049 goto release;
3050 }
3051
3052 if (!backup_file)
3053 switch(set_new_data_offset(sra, st, devname,
3054 reshape.after.data_disks - reshape.before.data_disks,
3055 data_offset,
3056 reshape.min_offset_change, 1)) {
3057 case -1:
3058 goto release;
3059 case 0:
3060 /* Updated data_offset, so it's easy now */
3061 update_cache_size(container, sra, info,
3062 min(reshape.before.data_disks,
3063 reshape.after.data_disks),
3064 reshape.backup_blocks);
3065
3066 /* Right, everything seems fine. Let's kick things off.
3067 */
3068 sync_metadata(st);
3069
3070 if (impose_reshape(sra, info, st, fd, restart,
3071 devname, container, &reshape) < 0)
3072 goto release;
3073 if (sysfs_set_str(sra, NULL, "sync_action", "reshape") < 0) {
3074 pr_err("Failed to initiate reshape!\n");
3075 goto release;
3076 }
3077 if (info->new_level == reshape.level)
3078 return 0;
3079 /* need to adjust level when reshape completes */
3080 switch(fork()) {
3081 case -1: /* ignore error, but don't wait */
3082 return 0;
3083 default: /* parent */
3084 return 0;
3085 case 0:
3086 map_fork();
3087 break;
3088 }
3089 wait_reshape(sra);
3090 impose_level(fd, info->new_level, devname, verbose);
3091
3092 return 0;
3093 case 1: /* Couldn't set data_offset, try the old way */
3094 if (data_offset != INVALID_SECTORS) {
3095 pr_err("Cannot update data_offset on this array\n");
3096 goto release;
3097 }
3098 break;
3099 }
3100
3101 started:
3102 /* Decide how many blocks (sectors) for a reshape
3103 * unit. The number we have so far is just a minimum
3104 */
3105 blocks = reshape.backup_blocks;
3106 if (reshape.before.data_disks ==
3107 reshape.after.data_disks) {
3108 /* Make 'blocks' bigger for better throughput, but
3109 * not so big that we reject it below.
3110 * Try for 16 megabytes
3111 */
3112 while (blocks * 32 < sra->component_size &&
3113 blocks < 16*1024*2)
3114 blocks *= 2;
3115 } else
3116 pr_err("Need to backup %luK of critical "
3117 "section..\n", blocks/2);
3118
3119 if (blocks >= sra->component_size/2) {
3120 pr_err("%s: Something wrong"
3121 " - reshape aborted\n",
3122 devname);
3123 goto release;
3124 }
3125
3126 /* Now we need to open all these devices so we can read/write.
3127 */
3128 nrdisks = max(reshape.before.data_disks,
3129 reshape.after.data_disks) + reshape.parity
3130 + sra->array.spare_disks;
3131 fdlist = xcalloc((1+nrdisks), sizeof(int));
3132 offsets = xcalloc((1+nrdisks), sizeof(offsets[0]));
3133
3134 odisks = reshape.before.data_disks + reshape.parity;
3135 d = reshape_prepare_fdlist(devname, sra, odisks,
3136 nrdisks, blocks, backup_file,
3137 fdlist, offsets);
3138 if (d < 0) {
3139 goto release;
3140 }
3141 if ((st->ss->manage_reshape == NULL) ||
3142 (st->ss->recover_backup == NULL)) {
3143 if (backup_file == NULL) {
3144 if (reshape.after.data_disks <=
3145 reshape.before.data_disks) {
3146 pr_err("%s: Cannot grow - need backup-file\n",
3147 devname);
3148 pr_err(" Please provide one with \"--backup=...\"\n");
3149 goto release;
3150 } else if (sra->array.spare_disks == 0) {
3151 pr_err("%s: Cannot grow - "
3152 "need a spare or backup-file to backup "
3153 "critical section\n", devname);
3154 goto release;
3155 }
3156 } else {
3157 if (!reshape_open_backup_file(backup_file, fd, devname,
3158 (signed)blocks,
3159 fdlist+d, offsets+d,
3160 restart)) {
3161 goto release;
3162 }
3163 d++;
3164 }
3165 }
3166
3167 update_cache_size(container, sra, info,
3168 min(reshape.before.data_disks, reshape.after.data_disks),
3169 blocks);
3170
3171 /* Right, everything seems fine. Let's kick things off.
3172 * If only changing raid_disks, use ioctl, else use
3173 * sysfs.
3174 */
3175 sync_metadata(st);
3176
3177 if (impose_reshape(sra, info, st, fd, restart,
3178 devname, container, &reshape) < 0)
3179 goto release;
3180
3181 err = start_reshape(sra, restart, reshape.before.data_disks,
3182 reshape.after.data_disks);
3183 if (err) {
3184 pr_err("Cannot %s reshape for %s\n",
3185 restart ? "continue" : "start",
3186 devname);
3187 goto release;
3188 }
3189 if (restart)
3190 sysfs_set_str(sra, NULL, "array_state", "active");
3191 if (freeze_reshape) {
3192 free(fdlist);
3193 free(offsets);
3194 sysfs_free(sra);
3195 pr_err("Reshape has to be continued from"
3196 " location %llu when root filesystem has been mounted.\n",
3197 sra->reshape_progress);
3198 return 1;
3199 }
3200
3201 /* Now we just need to kick off the reshape and watch, while
3202 * handling backups of the data...
3203 * This is all done by a forked background process.
3204 */
3205 switch(forked ? 0 : fork()) {
3206 case -1:
3207 pr_err("Cannot run child to monitor reshape: %s\n",
3208 strerror(errno));
3209 abort_reshape(sra);
3210 goto release;
3211 default:
3212 free(fdlist);
3213 free(offsets);
3214 sysfs_free(sra);
3215 return 0;
3216 case 0:
3217 map_fork();
3218 break;
3219 }
3220
3221 /* If another array on the same devices is busy, the
3222 * reshape will wait for them. This would mean that
3223 * the first section that we suspend will stay suspended
3224 * for a long time. So check on that possibility
3225 * by looking for "DELAYED" in /proc/mdstat, and if found,
3226 * wait a while
3227 */
3228 do {
3229 struct mdstat_ent *mds, *m;
3230 delayed = 0;
3231 mds = mdstat_read(1, 0);
3232 for (m = mds; m; m = m->next)
3233 if (strcmp(m->devnm, sra->sys_name) == 0) {
3234 if (m->resync &&
3235 m->percent == RESYNC_DELAYED)
3236 delayed = 1;
3237 if (m->resync == 0)
3238 /* Haven't started the reshape thread
3239 * yet, wait a bit
3240 */
3241 delayed = 2;
3242 break;
3243 }
3244 free_mdstat(mds);
3245 if (delayed == 1 && get_linux_version() < 3007000) {
3246 pr_err("Reshape is delayed, but cannot wait carefully with this kernel.\n"
3247 " You might experience problems until other reshapes complete.\n");
3248 delayed = 0;
3249 }
3250 if (delayed)
3251 mdstat_wait(30 - (delayed-1) * 25);
3252 } while (delayed);
3253 mdstat_close();
3254 close(fd);
3255 if (check_env("MDADM_GROW_VERIFY"))
3256 fd = open(devname, O_RDONLY | O_DIRECT);
3257 else
3258 fd = -1;
3259 mlockall(MCL_FUTURE);
3260
3261 if (st->ss->external) {
3262 /* metadata handler takes it from here */
3263 done = st->ss->manage_reshape(
3264 fd, sra, &reshape, st, blocks,
3265 fdlist, offsets,
3266 d - odisks, fdlist+odisks,
3267 offsets+odisks);
3268 } else
3269 done = child_monitor(
3270 fd, sra, &reshape, st, blocks,
3271 fdlist, offsets,
3272 d - odisks, fdlist+odisks,
3273 offsets+odisks);
3274
3275 free(fdlist);
3276 free(offsets);
3277
3278 if (backup_file && done)
3279 unlink(backup_file);
3280 if (!done) {
3281 abort_reshape(sra);
3282 goto out;
3283 }
3284
3285 if (!st->ss->external &&
3286 !(reshape.before.data_disks != reshape.after.data_disks
3287 && info->custom_array_size) &&
3288 info->new_level == reshape.level &&
3289 !forked) {
3290 /* no need to wait for the reshape to finish as
3291 * there is nothing more to do.
3292 */
3293 sysfs_free(sra);
3294 exit(0);
3295 }
3296 wait_reshape(sra);
3297
3298 if (st->ss->external) {
3299 /* Re-load the metadata as much could have changed */
3300 int cfd = open_dev(st->container_devnm);
3301 if (cfd >= 0) {
3302 flush_mdmon(container);
3303 st->ss->free_super(st);
3304 st->ss->load_container(st, cfd, container);
3305 close(cfd);
3306 }
3307 }
3308
3309 /* set new array size if required customer_array_size is used
3310 * by this metadata.
3311 */
3312 if (reshape.before.data_disks !=
3313 reshape.after.data_disks &&
3314 info->custom_array_size)
3315 set_array_size(st, info, info->text_version);
3316
3317 if (info->new_level != reshape.level) {
3318 if (fd < 0)
3319 fd = open(devname, O_RDONLY);
3320 impose_level(fd, info->new_level, devname, verbose);
3321 close(fd);
3322 if (info->new_level == 0)
3323 st->update_tail = NULL;
3324 }
3325 out:
3326 sysfs_free(sra);
3327 if (forked)
3328 return 0;
3329 unfreeze(st);
3330 exit(0);
3331
3332 release:
3333 free(fdlist);
3334 free(offsets);
3335 if (orig_level != UnSet && sra) {
3336 c = map_num(pers, orig_level);
3337 if (c && sysfs_set_str(sra, NULL, "level", c) == 0)
3338 pr_err("aborting level change\n");
3339 }
3340 sysfs_free(sra);
3341 if (!forked)
3342 unfreeze(st);
3343 return 1;
3344 }
3345
3346 /* mdfd handle is passed to be closed in child process (after fork).
3347 */
3348 int reshape_container(char *container, char *devname,
3349 int mdfd,
3350 struct supertype *st,
3351 struct mdinfo *info,
3352 int force,
3353 char *backup_file,
3354 int verbose, int restart, int freeze_reshape)
3355 {
3356 struct mdinfo *cc = NULL;
3357 int rv = restart;
3358 char last_devnm[32] = "";
3359
3360 /* component_size is not meaningful for a container,
3361 * so pass '0' meaning 'no change'
3362 */
3363 if (!restart &&
3364 reshape_super(st, 0, info->new_level,
3365 info->new_layout, info->new_chunk,
3366 info->array.raid_disks, info->delta_disks,
3367 backup_file, devname, APPLY_METADATA_CHANGES,
3368 verbose)) {
3369 unfreeze(st);
3370 return 1;
3371 }
3372
3373 sync_metadata(st);
3374
3375 /* ping monitor to be sure that update is on disk
3376 */
3377 ping_monitor(container);
3378
3379 switch (fork()) {
3380 case -1: /* error */
3381 perror("Cannot fork to complete reshape\n");
3382 unfreeze(st);
3383 return 1;
3384 default: /* parent */
3385 if (!freeze_reshape)
3386 printf(Name ": multi-array reshape continues"
3387 " in background\n");
3388 return 0;
3389 case 0: /* child */
3390 map_fork();
3391 break;
3392 }
3393
3394 /* close unused handle in child process
3395 */
3396 if (mdfd > -1)
3397 close(mdfd);
3398
3399 while(1) {
3400 /* For each member array with reshape_active,
3401 * we need to perform the reshape.
3402 * We pick the first array that needs reshaping and
3403 * reshape it. reshape_array() will re-read the metadata
3404 * so the next time through a different array should be
3405 * ready for reshape.
3406 * It is possible that the 'different' array will not
3407 * be assembled yet. In that case we simple exit.
3408 * When it is assembled, the mdadm which assembles it
3409 * will take over the reshape.
3410 */
3411 struct mdinfo *content;
3412 int fd;
3413 struct mdstat_ent *mdstat;
3414 char *adev;
3415 int devid;
3416
3417 sysfs_free(cc);
3418
3419 cc = st->ss->container_content(st, NULL);
3420
3421 for (content = cc; content ; content = content->next) {
3422 char *subarray;
3423 if (!content->reshape_active)
3424 continue;
3425
3426 subarray = strchr(content->text_version+1, '/')+1;
3427 mdstat = mdstat_by_subdev(subarray, container);
3428 if (!mdstat)
3429 continue;
3430 if (mdstat->active == 0) {
3431 pr_err("Skipping inactive array %s.\n",
3432 mdstat->devnm);
3433 free_mdstat(mdstat);
3434 mdstat = NULL;
3435 continue;
3436 }
3437 break;
3438 }
3439 if (!content)
3440 break;
3441
3442 devid = devnm2devid(mdstat->devnm);
3443 adev = map_dev(major(devid), minor(devid), 0);
3444 if (!adev)
3445 adev = content->text_version;
3446
3447 fd = open_dev(mdstat->devnm);
3448 if (fd < 0) {
3449 printf(Name ": Device %s cannot be opened for reshape.",
3450 adev);
3451 break;
3452 }
3453
3454 if (strcmp(last_devnm, mdstat->devnm) == 0) {
3455 /* Do not allow for multiple reshape_array() calls for
3456 * the same array.
3457 * It can happen when reshape_array() returns without
3458 * error, when reshape is not finished (wrong reshape
3459 * starting/continuation conditions). Mdmon doesn't
3460 * switch to next array in container and reentry
3461 * conditions for the same array occur.
3462 * This is possibly interim until the behaviour of
3463 * reshape_array is resolved().
3464 */
3465 printf(Name ": Multiple reshape execution detected for "
3466 "device %s.", adev);
3467 close(fd);
3468 break;
3469 }
3470 strcpy(last_devnm, mdstat->devnm);
3471
3472 sysfs_init(content, fd, mdstat->devnm);
3473
3474 if (mdmon_running(container))
3475 flush_mdmon(container);
3476
3477 rv = reshape_array(container, fd, adev, st,
3478 content, force, NULL, INVALID_SECTORS,
3479 backup_file, verbose, 1, restart,
3480 freeze_reshape);
3481 close(fd);
3482
3483 if (freeze_reshape) {
3484 sysfs_free(cc);
3485 exit(0);
3486 }
3487
3488 restart = 0;
3489 if (rv)
3490 break;
3491
3492 if (mdmon_running(container))
3493 flush_mdmon(container);
3494 }
3495 if (!rv)
3496 unfreeze(st);
3497 sysfs_free(cc);
3498 exit(0);
3499 }
3500
3501 /*
3502 * We run a child process in the background which performs the following
3503 * steps:
3504 * - wait for resync to reach a certain point
3505 * - suspend io to the following section
3506 * - backup that section
3507 * - allow resync to proceed further
3508 * - resume io
3509 * - discard the backup.
3510 *
3511 * When are combined in slightly different ways in the three cases.
3512 * Grow:
3513 * - suspend/backup/allow/wait/resume/discard
3514 * Shrink:
3515 * - allow/wait/suspend/backup/allow/wait/resume/discard
3516 * same-size:
3517 * - wait/resume/discard/suspend/backup/allow
3518 *
3519 * suspend/backup/allow always come together
3520 * wait/resume/discard do too.
3521 * For the same-size case we have two backups to improve flow.
3522 *
3523 */
3524
3525 int progress_reshape(struct mdinfo *info, struct reshape *reshape,
3526 unsigned long long backup_point,
3527 unsigned long long wait_point,
3528 unsigned long long *suspend_point,
3529 unsigned long long *reshape_completed, int *frozen)
3530 {
3531 /* This function is called repeatedly by the reshape manager.
3532 * It determines how much progress can safely be made and allows
3533 * that progress.
3534 * - 'info' identifies the array and particularly records in
3535 * ->reshape_progress the metadata's knowledge of progress
3536 * This is a sector offset from the start of the array
3537 * of the next array block to be relocated. This number
3538 * may increase from 0 or decrease from array_size, depending
3539 * on the type of reshape that is happening.
3540 * Note that in contrast, 'sync_completed' is a block count of the
3541 * reshape so far. It gives the distance between the start point
3542 * (head or tail of device) and the next place that data will be
3543 * written. It always increases.
3544 * - 'reshape' is the structure created by analyse_change
3545 * - 'backup_point' shows how much the metadata manager has backed-up
3546 * data. For reshapes with increasing progress, it is the next address
3547 * to be backed up, previous addresses have been backed-up. For
3548 * decreasing progress, it is the earliest address that has been
3549 * backed up - later address are also backed up.
3550 * So addresses between reshape_progress and backup_point are
3551 * backed up providing those are in the 'correct' order.
3552 * - 'wait_point' is an array address. When reshape_completed
3553 * passes this point, progress_reshape should return. It might
3554 * return earlier if it determines that ->reshape_progress needs
3555 * to be updated or further backup is needed.
3556 * - suspend_point is maintained by progress_reshape and the caller
3557 * should not touch it except to initialise to zero.
3558 * It is an array address and it only increases in 2.6.37 and earlier.
3559 * This makes it difficult to handle reducing reshapes with
3560 * external metadata.
3561 * However: it is similar to backup_point in that it records the
3562 * other end of a suspended region from reshape_progress.
3563 * it is moved to extend the region that is safe to backup and/or
3564 * reshape
3565 * - reshape_completed is read from sysfs and returned. The caller
3566 * should copy this into ->reshape_progress when it has reason to
3567 * believe that the metadata knows this, and any backup outside this
3568 * has been erased.
3569 *
3570 * Return value is:
3571 * 1 if more data from backup_point - but only as far as suspend_point,
3572 * should be backed up
3573 * 0 if things are progressing smoothly
3574 * -1 if the reshape is finished because it is all done,
3575 * -2 if the reshape is finished due to an error.
3576 */
3577
3578 int advancing = (reshape->after.data_disks
3579 >= reshape->before.data_disks);
3580 unsigned long long need_backup; /* All data between start of array and
3581 * here will at some point need to
3582 * be backed up.
3583 */
3584 unsigned long long read_offset, write_offset;
3585 unsigned long long write_range;
3586 unsigned long long max_progress, target, completed;
3587 unsigned long long array_size = (info->component_size
3588 * reshape->before.data_disks);
3589 int fd;
3590 char buf[20];
3591
3592 /* First, we unsuspend any region that is now known to be safe.
3593 * If suspend_point is on the 'wrong' side of reshape_progress, then
3594 * we don't have or need suspension at the moment. This is true for
3595 * native metadata when we don't need to back-up.
3596 */
3597 if (advancing) {
3598 if (info->reshape_progress <= *suspend_point)
3599 sysfs_set_num(info, NULL, "suspend_lo",
3600 info->reshape_progress);
3601 } else {
3602 /* Note: this won't work in 2.6.37 and before.
3603 * Something somewhere should make sure we don't need it!
3604 */
3605 if (info->reshape_progress >= *suspend_point)
3606 sysfs_set_num(info, NULL, "suspend_hi",
3607 info->reshape_progress);
3608 }
3609
3610 /* Now work out how far it is safe to progress.
3611 * If the read_offset for ->reshape_progress is less than
3612 * 'blocks' beyond the write_offset, we can only progress as far
3613 * as a backup.
3614 * Otherwise we can progress until the write_offset for the new location
3615 * reaches (within 'blocks' of) the read_offset at the current location.
3616 * However that region must be suspended unless we are using native
3617 * metadata.
3618 * If we need to suspend more, we limit it to 128M per device, which is
3619 * rather arbitrary and should be some time-based calculation.
3620 */
3621 read_offset = info->reshape_progress / reshape->before.data_disks;
3622 write_offset = info->reshape_progress / reshape->after.data_disks;
3623 write_range = info->new_chunk/512;
3624 if (reshape->before.data_disks == reshape->after.data_disks)
3625 need_backup = array_size;
3626 else
3627 need_backup = reshape->backup_blocks;
3628 if (advancing) {
3629 if (read_offset < write_offset + write_range)
3630 max_progress = backup_point;
3631 else
3632 max_progress =
3633 read_offset *
3634 reshape->after.data_disks;
3635 } else {
3636 if (read_offset > write_offset - write_range)
3637 /* Can only progress as far as has been backed up,
3638 * which must be suspended */
3639 max_progress = backup_point;
3640 else if (info->reshape_progress <= need_backup)
3641 max_progress = backup_point;
3642 else {
3643 if (info->array.major_version >= 0)
3644 /* Can progress until backup is needed */
3645 max_progress = need_backup;
3646 else {
3647 /* Can progress until metadata update is required */
3648 max_progress =
3649 read_offset *
3650 reshape->after.data_disks;
3651 /* but data must be suspended */
3652 if (max_progress < *suspend_point)
3653 max_progress = *suspend_point;
3654 }
3655 }
3656 }
3657
3658 /* We know it is safe to progress to 'max_progress' providing
3659 * it is suspended or we are using native metadata.
3660 * Consider extending suspend_point 128M per device if it
3661 * is less than 64M per device beyond reshape_progress.
3662 * But always do a multiple of 'blocks'
3663 * FIXME this is too big - it takes to long to complete
3664 * this much.
3665 */
3666 target = 64*1024*2 * min(reshape->before.data_disks,
3667 reshape->after.data_disks);
3668 target /= reshape->backup_blocks;
3669 if (target < 2)
3670 target = 2;
3671 target *= reshape->backup_blocks;
3672
3673 /* For externally managed metadata we always need to suspend IO to
3674 * the area being reshaped so we regularly push suspend_point forward.
3675 * For native metadata we only need the suspend if we are going to do
3676 * a backup.
3677 */
3678 if (advancing) {
3679 if ((need_backup > info->reshape_progress
3680 || info->array.major_version < 0) &&
3681 *suspend_point < info->reshape_progress + target) {
3682 if (need_backup < *suspend_point + 2 * target)
3683 *suspend_point = need_backup;
3684 else if (*suspend_point + 2 * target < array_size)
3685 *suspend_point += 2 * target;
3686 else
3687 *suspend_point = array_size;
3688 sysfs_set_num(info, NULL, "suspend_hi", *suspend_point);
3689 if (max_progress > *suspend_point)
3690 max_progress = *suspend_point;
3691 }
3692 } else {
3693 if (info->array.major_version >= 0) {
3694 /* Only need to suspend when about to backup */
3695 if (info->reshape_progress < need_backup * 2 &&
3696 *suspend_point > 0) {
3697 *suspend_point = 0;
3698 sysfs_set_num(info, NULL, "suspend_lo", 0);
3699 sysfs_set_num(info, NULL, "suspend_hi", need_backup);
3700 }
3701 } else {
3702 /* Need to suspend continually */
3703 if (info->reshape_progress < *suspend_point)
3704 *suspend_point = info->reshape_progress;
3705 if (*suspend_point + target < info->reshape_progress)
3706 /* No need to move suspend region yet */;
3707 else {
3708 if (*suspend_point >= 2 * target)
3709 *suspend_point -= 2 * target;
3710 else
3711 *suspend_point = 0;
3712 sysfs_set_num(info, NULL, "suspend_lo",
3713 *suspend_point);
3714 }
3715 if (max_progress < *suspend_point)
3716 max_progress = *suspend_point;
3717 }
3718 }
3719
3720 /* now set sync_max to allow that progress. sync_max, like
3721 * sync_completed is a count of sectors written per device, so
3722 * we find the difference between max_progress and the start point,
3723 * and divide that by after.data_disks to get a sync_max
3724 * number.
3725 * At the same time we convert wait_point to a similar number
3726 * for comparing against sync_completed.
3727 */
3728 /* scale down max_progress to per_disk */
3729 max_progress /= reshape->after.data_disks;
3730 /* Round to chunk size as some kernels give an erroneously high number */
3731 max_progress /= info->new_chunk/512;
3732 max_progress *= info->new_chunk/512;
3733 /* And round to old chunk size as the kernel wants that */
3734 max_progress /= info->array.chunk_size/512;
3735 max_progress *= info->array.chunk_size/512;
3736 /* Limit progress to the whole device */
3737 if (max_progress > info->component_size)
3738 max_progress = info->component_size;
3739 wait_point /= reshape->after.data_disks;
3740 if (!advancing) {
3741 /* switch from 'device offset' to 'processed block count' */
3742 max_progress = info->component_size - max_progress;
3743 wait_point = info->component_size - wait_point;
3744 }
3745
3746 if (!*frozen)
3747 sysfs_set_num(info, NULL, "sync_max", max_progress);
3748
3749 /* Now wait. If we have already reached the point that we were
3750 * asked to wait to, don't wait at all, else wait for any change.
3751 * We need to select on 'sync_completed' as that is the place that
3752 * notifications happen, but we are really interested in
3753 * 'reshape_position'
3754 */
3755 fd = sysfs_get_fd(info, NULL, "sync_completed");
3756 if (fd < 0)
3757 goto check_progress;
3758
3759 if (sysfs_fd_get_ll(fd, &completed) < 0)
3760 goto check_progress;
3761
3762 while (completed < max_progress && completed < wait_point) {
3763 /* Check that sync_action is still 'reshape' to avoid
3764 * waiting forever on a dead array
3765 */
3766 char action[20];
3767 if (sysfs_get_str(info, NULL, "sync_action",
3768 action, 20) <= 0 ||
3769 strncmp(action, "reshape", 7) != 0)
3770 break;
3771 /* Some kernels reset 'sync_completed' to zero
3772 * before setting 'sync_action' to 'idle'.
3773 * So we need these extra tests.
3774 */
3775 if (completed == 0 && advancing
3776 && info->reshape_progress > 0)
3777 break;
3778 if (completed == 0 && !advancing
3779 && info->reshape_progress < (info->component_size
3780 * reshape->after.data_disks))
3781 break;
3782 sysfs_wait(fd, NULL);
3783 if (sysfs_fd_get_ll(fd, &completed) < 0)
3784 goto check_progress;
3785 }
3786 /* Some kernels reset 'sync_completed' to zero,
3787 * we need to have real point we are in md
3788 */
3789 if (completed == 0)
3790 completed = max_progress;
3791
3792 /* some kernels can give an incorrectly high 'completed' number */
3793 completed /= (info->new_chunk/512);
3794 completed *= (info->new_chunk/512);
3795 /* Convert 'completed' back in to a 'progress' number */
3796 completed *= reshape->after.data_disks;
3797 if (!advancing) {
3798 completed = info->component_size * reshape->after.data_disks
3799 - completed;
3800 }
3801 *reshape_completed = completed;
3802
3803 close(fd);
3804
3805 /* We return the need_backup flag. Caller will decide
3806 * how much - a multiple of ->backup_blocks up to *suspend_point
3807 */
3808 if (advancing)
3809 return need_backup > info->reshape_progress;
3810 else
3811 return need_backup >= info->reshape_progress;
3812
3813 check_progress:
3814 /* if we couldn't read a number from sync_completed, then
3815 * either the reshape did complete, or it aborted.
3816 * We can tell which by checking for 'none' in reshape_position.
3817 * If it did abort, then it might immediately restart if it
3818 * it was just a device failure that leaves us degraded but
3819 * functioning.
3820 */
3821 strcpy(buf, "hi");
3822 if (sysfs_get_str(info, NULL, "reshape_position", buf, sizeof(buf)) < 0
3823 || strncmp(buf, "none", 4) != 0) {
3824 /* The abort might only be temporary. Wait up to 10
3825 * seconds for fd to contain a valid number again.
3826 */
3827 int wait = 10000;
3828 int rv = -2;
3829 unsigned long long new_sync_max;
3830 while (fd >= 0 && rv < 0 && wait > 0) {
3831 if (sysfs_wait(fd, &wait) != 1)
3832 break;
3833 switch (sysfs_fd_get_ll(fd, &completed)) {
3834 case 0:
3835 /* all good again */
3836 rv = 1;
3837 /* If "sync_max" is no longer max_progress
3838 * we need to freeze things
3839 */
3840 sysfs_get_ll(info, NULL, "sync_max", &new_sync_max);
3841 *frozen = (new_sync_max != max_progress);
3842 break;
3843 case -2: /* read error - abort */
3844 wait = 0;
3845 break;
3846 }
3847 }
3848 if (fd >= 0)
3849 close(fd);
3850 return rv; /* abort */
3851 } else {
3852 /* Maybe racing with array shutdown - check state */
3853 if (fd >= 0)
3854 close(fd);
3855 if (sysfs_get_str(info, NULL, "array_state", buf, sizeof(buf)) < 0
3856 || strncmp(buf, "inactive", 8) == 0
3857 || strncmp(buf, "clear",5) == 0)
3858 return -2; /* abort */
3859 return -1; /* complete */
3860 }
3861 }
3862
3863 /* FIXME return status is never checked */
3864 static int grow_backup(struct mdinfo *sra,
3865 unsigned long long offset, /* per device */
3866 unsigned long stripes, /* per device, in old chunks */
3867 int *sources, unsigned long long *offsets,
3868 int disks, int chunk, int level, int layout,
3869 int dests, int *destfd, unsigned long long *destoffsets,
3870 int part, int *degraded,
3871 char *buf)
3872 {
3873 /* Backup 'blocks' sectors at 'offset' on each device of the array,
3874 * to storage 'destfd' (offset 'destoffsets'), after first
3875 * suspending IO. Then allow resync to continue
3876 * over the suspended section.
3877 * Use part 'part' of the backup-super-block.
3878 */
3879 int odata = disks;
3880 int rv = 0;
3881 int i;
3882 unsigned long long ll;
3883 int new_degraded;
3884 //printf("offset %llu\n", offset);
3885 if (level >= 4)
3886 odata--;
3887 if (level == 6)
3888 odata--;
3889
3890 /* Check that array hasn't become degraded, else we might backup the wrong data */
3891 if (sysfs_get_ll(sra, NULL, "degraded", &ll) < 0)
3892 return -1; /* FIXME this error is ignored */
3893 new_degraded = (int)ll;
3894 if (new_degraded != *degraded) {
3895 /* check each device to ensure it is still working */
3896 struct mdinfo *sd;
3897 for (sd = sra->devs ; sd ; sd = sd->next) {
3898 if (sd->disk.state & (1<<MD_DISK_FAULTY))
3899 continue;
3900 if (sd->disk.state & (1<<MD_DISK_SYNC)) {
3901 char sbuf[20];
3902 if (sysfs_get_str(sra, sd, "state", sbuf, 20) < 0 ||
3903 strstr(sbuf, "faulty") ||
3904 strstr(sbuf, "in_sync") == NULL) {
3905 /* this device is dead */
3906 sd->disk.state = (1<<MD_DISK_FAULTY);
3907 if (sd->disk.raid_disk >= 0 &&
3908 sources[sd->disk.raid_disk] >= 0) {
3909 close(sources[sd->disk.raid_disk]);
3910 sources[sd->disk.raid_disk] = -1;
3911 }
3912 }
3913 }
3914 }
3915 *degraded = new_degraded;
3916 }
3917 if (part) {
3918 bsb.arraystart2 = __cpu_to_le64(offset * odata);
3919 bsb.length2 = __cpu_to_le64(stripes * (chunk/512) * odata);
3920 } else {
3921 bsb.arraystart = __cpu_to_le64(offset * odata);
3922 bsb.length = __cpu_to_le64(stripes * (chunk/512) * odata);
3923 }
3924 if (part)
3925 bsb.magic[15] = '2';
3926 for (i = 0; i < dests; i++)
3927 if (part)
3928 lseek64(destfd[i], destoffsets[i] + __le64_to_cpu(bsb.devstart2)*512, 0);
3929 else
3930 lseek64(destfd[i], destoffsets[i], 0);
3931
3932 rv = save_stripes(sources, offsets,
3933 disks, chunk, level, layout,
3934 dests, destfd,
3935 offset*512*odata, stripes * chunk * odata,
3936 buf);
3937
3938 if (rv)
3939 return rv;
3940 bsb.mtime = __cpu_to_le64(time(0));
3941 for (i = 0; i < dests; i++) {
3942 bsb.devstart = __cpu_to_le64(destoffsets[i]/512);
3943
3944 bsb.sb_csum = bsb_csum((char*)&bsb, ((char*)&bsb.sb_csum)-((char*)&bsb));
3945 if (memcmp(bsb.magic, "md_backup_data-2", 16) == 0)
3946 bsb.sb_csum2 = bsb_csum((char*)&bsb,
3947 ((char*)&bsb.sb_csum2)-((char*)&bsb));
3948
3949 rv = -1;
3950 if ((unsigned long long)lseek64(destfd[i], destoffsets[i] - 4096, 0)
3951 != destoffsets[i] - 4096)
3952 break;
3953 if (write(destfd[i], &bsb, 512) != 512)
3954 break;
3955 if (destoffsets[i] > 4096) {
3956 if ((unsigned long long)lseek64(destfd[i], destoffsets[i]+stripes*chunk*odata, 0) !=
3957 destoffsets[i]+stripes*chunk*odata)
3958 break;
3959 if (write(destfd[i], &bsb, 512) != 512)
3960 break;
3961 }
3962 fsync(destfd[i]);
3963 rv = 0;
3964 }
3965
3966 return rv;
3967 }
3968
3969 /* in 2.6.30, the value reported by sync_completed can be
3970 * less that it should be by one stripe.
3971 * This only happens when reshape hits sync_max and pauses.
3972 * So allow wait_backup to either extent sync_max further
3973 * than strictly necessary, or return before the
3974 * sync has got quite as far as we would really like.
3975 * This is what 'blocks2' is for.
3976 * The various caller give appropriate values so that
3977 * every works.
3978 */
3979 /* FIXME return value is often ignored */
3980 static int forget_backup(int dests, int *destfd,
3981 unsigned long long *destoffsets,
3982 int part)
3983 {
3984 /*
3985 * Erase backup 'part' (which is 0 or 1)
3986 */
3987 int i;
3988 int rv;
3989
3990 if (part) {
3991 bsb.arraystart2 = __cpu_to_le64(0);
3992 bsb.length2 = __cpu_to_le64(0);
3993 } else {
3994 bsb.arraystart = __cpu_to_le64(0);
3995 bsb.length = __cpu_to_le64(0);
3996 }
3997 bsb.mtime = __cpu_to_le64(time(0));
3998 rv = 0;
3999 for (i = 0; i < dests; i++) {
4000 bsb.devstart = __cpu_to_le64(destoffsets[i]/512);
4001 bsb.sb_csum = bsb_csum((char*)&bsb, ((char*)&bsb.sb_csum)-((char*)&bsb));
4002 if (memcmp(bsb.magic, "md_backup_data-2", 16) == 0)
4003 bsb.sb_csum2 = bsb_csum((char*)&bsb,
4004 ((char*)&bsb.sb_csum2)-((char*)&bsb));
4005 if ((unsigned long long)lseek64(destfd[i], destoffsets[i]-4096, 0) !=
4006 destoffsets[i]-4096)
4007 rv = -1;
4008 if (rv == 0 &&
4009 write(destfd[i], &bsb, 512) != 512)
4010 rv = -1;
4011 fsync(destfd[i]);
4012 }
4013 return rv;
4014 }
4015
4016 static void fail(char *msg)
4017 {
4018 int rv;
4019 rv = (write(2, msg, strlen(msg)) != (int)strlen(msg));
4020 rv |= (write(2, "\n", 1) != 1);
4021 exit(rv ? 1 : 2);
4022 }
4023
4024 static char *abuf, *bbuf;
4025 static unsigned long long abuflen;
4026 static void validate(int afd, int bfd, unsigned long long offset)
4027 {
4028 /* check that the data in the backup against the array.
4029 * This is only used for regression testing and should not
4030 * be used while the array is active
4031 */
4032 if (afd < 0)
4033 return;
4034 lseek64(bfd, offset - 4096, 0);
4035 if (read(bfd, &bsb2, 512) != 512)
4036 fail("cannot read bsb");
4037 if (bsb2.sb_csum != bsb_csum((char*)&bsb2,
4038 ((char*)&bsb2.sb_csum)-((char*)&bsb2)))
4039 fail("first csum bad");
4040 if (memcmp(bsb2.magic, "md_backup_data", 14) != 0)
4041 fail("magic is bad");
4042 if (memcmp(bsb2.magic, "md_backup_data-2", 16) == 0 &&
4043 bsb2.sb_csum2 != bsb_csum((char*)&bsb2,
4044 ((char*)&bsb2.sb_csum2)-((char*)&bsb2)))
4045 fail("second csum bad");
4046
4047 if (__le64_to_cpu(bsb2.devstart)*512 != offset)
4048 fail("devstart is wrong");
4049
4050 if (bsb2.length) {
4051 unsigned long long len = __le64_to_cpu(bsb2.length)*512;
4052
4053 if (abuflen < len) {
4054 free(abuf);
4055 free(bbuf);
4056 abuflen = len;
4057 if (posix_memalign((void**)&abuf, 4096, abuflen) ||
4058 posix_memalign((void**)&bbuf, 4096, abuflen)) {
4059 abuflen = 0;
4060 /* just stop validating on mem-alloc failure */
4061 return;
4062 }
4063 }
4064
4065 lseek64(bfd, offset, 0);
4066 if ((unsigned long long)read(bfd, bbuf, len) != len) {
4067 //printf("len %llu\n", len);
4068 fail("read first backup failed");
4069 }
4070 lseek64(afd, __le64_to_cpu(bsb2.arraystart)*512, 0);
4071 if ((unsigned long long)read(afd, abuf, len) != len)
4072 fail("read first from array failed");
4073 if (memcmp(bbuf, abuf, len) != 0) {
4074 #if 0
4075 int i;
4076 printf("offset=%llu len=%llu\n",
4077 (unsigned long long)__le64_to_cpu(bsb2.arraystart)*512, len);
4078 for (i=0; i<len; i++)
4079 if (bbuf[i] != abuf[i]) {
4080 printf("first diff byte %d\n", i);
4081 break;
4082 }
4083 #endif
4084 fail("data1 compare failed");
4085 }
4086 }
4087 if (bsb2.length2) {
4088 unsigned long long len = __le64_to_cpu(bsb2.length2)*512;
4089
4090 if (abuflen < len) {
4091 free(abuf);
4092 free(bbuf);
4093 abuflen = len;
4094 abuf = xmalloc(abuflen);
4095 bbuf = xmalloc(abuflen);
4096 }
4097
4098 lseek64(bfd, offset+__le64_to_cpu(bsb2.devstart2)*512, 0);
4099 if ((unsigned long long)read(bfd, bbuf, len) != len)
4100 fail("read second backup failed");
4101 lseek64(afd, __le64_to_cpu(bsb2.arraystart2)*512, 0);
4102 if ((unsigned long long)read(afd, abuf, len) != len)
4103 fail("read second from array failed");
4104 if (memcmp(bbuf, abuf, len) != 0)
4105 fail("data2 compare failed");
4106 }
4107 }
4108
4109 int child_monitor(int afd, struct mdinfo *sra, struct reshape *reshape,
4110 struct supertype *st, unsigned long blocks,
4111 int *fds, unsigned long long *offsets,
4112 int dests, int *destfd, unsigned long long *destoffsets)
4113 {
4114 /* Monitor a reshape where backup is being performed using
4115 * 'native' mechanism - either to a backup file, or
4116 * to some space in a spare.
4117 */
4118 char *buf;
4119 int degraded = -1;
4120 unsigned long long speed;
4121 unsigned long long suspend_point, array_size;
4122 unsigned long long backup_point, wait_point;
4123 unsigned long long reshape_completed;
4124 int done = 0;
4125 int increasing = reshape->after.data_disks >= reshape->before.data_disks;
4126 int part = 0; /* The next part of the backup area to fill. It may already
4127 * be full, so we need to check */
4128 int level = reshape->level;
4129 int layout = reshape->before.layout;
4130 int data = reshape->before.data_disks;
4131 int disks = reshape->before.data_disks + reshape->parity;
4132 int chunk = sra->array.chunk_size;
4133 struct mdinfo *sd;
4134 unsigned long stripes;
4135 int uuid[4];
4136 int frozen = 0;
4137
4138 /* set up the backup-super-block. This requires the
4139 * uuid from the array.
4140 */
4141 /* Find a superblock */
4142 for (sd = sra->devs; sd; sd = sd->next) {
4143 char *dn;
4144 int devfd;
4145 int ok;
4146 if (sd->disk.state & (1<<MD_DISK_FAULTY))
4147 continue;
4148 dn = map_dev(sd->disk.major, sd->disk.minor, 1);
4149 devfd = dev_open(dn, O_RDONLY);
4150 if (devfd < 0)
4151 continue;
4152 ok = st->ss->load_super(st, devfd, NULL);
4153 close(devfd);
4154 if (ok == 0)
4155 break;
4156 }
4157 if (!sd) {
4158 pr_err("Cannot find a superblock\n");
4159 return 0;
4160 }
4161
4162 memset(&bsb, 0, 512);
4163 memcpy(bsb.magic, "md_backup_data-1", 16);
4164 st->ss->uuid_from_super(st, uuid);
4165 memcpy(bsb.set_uuid, uuid, 16);
4166 bsb.mtime = __cpu_to_le64(time(0));
4167 bsb.devstart2 = blocks;
4168
4169 stripes = blocks / (sra->array.chunk_size/512) /
4170 reshape->before.data_disks;
4171
4172 if (posix_memalign((void**)&buf, 4096, disks * chunk))
4173 /* Don't start the 'reshape' */
4174 return 0;
4175 if (reshape->before.data_disks == reshape->after.data_disks) {
4176 sysfs_get_ll(sra, NULL, "sync_speed_min", &speed);
4177 sysfs_set_num(sra, NULL, "sync_speed_min", 200000);
4178 }
4179
4180 if (increasing) {
4181 array_size = sra->component_size * reshape->after.data_disks;
4182 backup_point = sra->reshape_progress;
4183 suspend_point = 0;
4184 } else {
4185 array_size = sra->component_size * reshape->before.data_disks;
4186 backup_point = reshape->backup_blocks;
4187 suspend_point = array_size;
4188 }
4189
4190 while (!done) {
4191 int rv;
4192
4193 /* Want to return as soon the oldest backup slot can
4194 * be released as that allows us to start backing up
4195 * some more, providing suspend_point has been
4196 * advanced, which it should have.
4197 */
4198 if (increasing) {
4199 wait_point = array_size;
4200 if (part == 0 && __le64_to_cpu(bsb.length) > 0)
4201 wait_point = (__le64_to_cpu(bsb.arraystart) +
4202 __le64_to_cpu(bsb.length));
4203 if (part == 1 && __le64_to_cpu(bsb.length2) > 0)
4204 wait_point = (__le64_to_cpu(bsb.arraystart2) +
4205 __le64_to_cpu(bsb.length2));
4206 } else {
4207 wait_point = 0;
4208 if (part == 0 && __le64_to_cpu(bsb.length) > 0)
4209 wait_point = __le64_to_cpu(bsb.arraystart);
4210 if (part == 1 && __le64_to_cpu(bsb.length2) > 0)
4211 wait_point = __le64_to_cpu(bsb.arraystart2);
4212 }
4213
4214 reshape_completed = sra->reshape_progress;
4215 rv = progress_reshape(sra, reshape,
4216 backup_point, wait_point,
4217 &suspend_point, &reshape_completed,
4218 &frozen);
4219 /* external metadata would need to ping_monitor here */
4220 sra->reshape_progress = reshape_completed;
4221
4222 /* Clear any backup region that is before 'here' */
4223 if (increasing) {
4224 if (__le64_to_cpu(bsb.length) > 0 &&
4225 reshape_completed >= (__le64_to_cpu(bsb.arraystart) +
4226 __le64_to_cpu(bsb.length)))
4227 forget_backup(dests, destfd,
4228 destoffsets, 0);
4229 if (__le64_to_cpu(bsb.length2) > 0 &&
4230 reshape_completed >= (__le64_to_cpu(bsb.arraystart2) +
4231 __le64_to_cpu(bsb.length2)))
4232 forget_backup(dests, destfd,
4233 destoffsets, 1);
4234 } else {
4235 if (__le64_to_cpu(bsb.length) > 0 &&
4236 reshape_completed <= (__le64_to_cpu(bsb.arraystart)))
4237 forget_backup(dests, destfd,
4238 destoffsets, 0);
4239 if (__le64_to_cpu(bsb.length2) > 0 &&
4240 reshape_completed <= (__le64_to_cpu(bsb.arraystart2)))
4241 forget_backup(dests, destfd,
4242 destoffsets, 1);
4243 }
4244
4245 if (rv < 0) {
4246 if (rv == -1)
4247 done = 1;
4248 break;
4249 }
4250 if (rv == 0 && increasing && !st->ss->external) {
4251 /* No longer need to monitor this reshape */
4252 done = 1;
4253 break;
4254 }
4255
4256 while (rv) {
4257 unsigned long long offset;
4258 unsigned long actual_stripes;
4259 /* Need to backup some data.
4260 * If 'part' is not used and the desired
4261 * backup size is suspended, do a backup,
4262 * then consider the next part.
4263 */
4264 /* Check that 'part' is unused */
4265 if (part == 0 && __le64_to_cpu(bsb.length) != 0)
4266 break;
4267 if (part == 1 && __le64_to_cpu(bsb.length2) != 0)
4268 break;
4269
4270 offset = backup_point / data;
4271 actual_stripes = stripes;
4272 if (increasing) {
4273 if (offset + actual_stripes * (chunk/512) >
4274 sra->component_size)
4275 actual_stripes = ((sra->component_size - offset)
4276 / (chunk/512));
4277 if (offset + actual_stripes * (chunk/512) >
4278 suspend_point/data)
4279 break;
4280 } else {
4281 if (offset < actual_stripes * (chunk/512))
4282 actual_stripes = offset / (chunk/512);
4283 offset -= actual_stripes * (chunk/512);
4284 if (offset < suspend_point/data)
4285 break;
4286 }
4287 if (actual_stripes == 0)
4288 break;
4289 grow_backup(sra, offset, actual_stripes,
4290 fds, offsets,
4291 disks, chunk, level, layout,
4292 dests, destfd, destoffsets,
4293 part, &degraded, buf);
4294 validate(afd, destfd[0], destoffsets[0]);
4295 /* record where 'part' is up to */
4296 part = !part;
4297 if (increasing)
4298 backup_point += actual_stripes * (chunk/512) * data;
4299 else
4300 backup_point -= actual_stripes * (chunk/512) * data;
4301 }
4302 }
4303
4304 /* FIXME maybe call progress_reshape one more time instead */
4305 abort_reshape(sra); /* remove any remaining suspension */
4306 if (reshape->before.data_disks == reshape->after.data_disks)
4307 sysfs_set_num(sra, NULL, "sync_speed_min", speed);
4308 free(buf);
4309 return done;
4310 }
4311
4312 /*
4313 * If any spare contains md_back_data-1 which is recent wrt mtime,
4314 * write that data into the array and update the super blocks with
4315 * the new reshape_progress
4316 */
4317 int Grow_restart(struct supertype *st, struct mdinfo *info, int *fdlist, int cnt,
4318 char *backup_file, int verbose)
4319 {
4320 int i, j;
4321 int old_disks;
4322 unsigned long long *offsets;
4323 unsigned long long nstripe, ostripe;
4324 int ndata, odata;
4325
4326 odata = info->array.raid_disks - info->delta_disks - 1;
4327 if (info->array.level == 6) odata--; /* number of data disks */
4328 ndata = info->array.raid_disks - 1;
4329 if (info->new_level == 6) ndata--;
4330
4331 old_disks = info->array.raid_disks - info->delta_disks;
4332
4333 if (info->delta_disks <= 0)
4334 /* Didn't grow, so the backup file must have
4335 * been used
4336 */
4337 old_disks = cnt;
4338 for (i=old_disks-(backup_file?1:0); i<cnt; i++) {
4339 struct mdinfo dinfo;
4340 int fd;
4341 int bsbsize;
4342 char *devname, namebuf[20];
4343 unsigned long long lo, hi;
4344
4345 /* This was a spare and may have some saved data on it.
4346 * Load the superblock, find and load the
4347 * backup_super_block.
4348 * If either fail, go on to next device.
4349 * If the backup contains no new info, just return
4350 * else restore data and update all superblocks
4351 */
4352 if (i == old_disks-1) {
4353 fd = open(backup_file, O_RDONLY);
4354 if (fd<0) {
4355 pr_err("backup file %s inaccessible: %s\n",
4356 backup_file, strerror(errno));
4357 continue;
4358 }
4359 devname = backup_file;
4360 } else {
4361 fd = fdlist[i];
4362 if (fd < 0)
4363 continue;
4364 if (st->ss->load_super(st, fd, NULL))
4365 continue;
4366
4367 st->ss->getinfo_super(st, &dinfo, NULL);
4368 st->ss->free_super(st);
4369
4370 if (lseek64(fd,
4371 (dinfo.data_offset + dinfo.component_size - 8) <<9,
4372 0) < 0) {
4373 pr_err("Cannot seek on device %d\n", i);
4374 continue; /* Cannot seek */
4375 }
4376 sprintf(namebuf, "device-%d", i);
4377 devname = namebuf;
4378 }
4379 if (read(fd, &bsb, sizeof(bsb)) != sizeof(bsb)) {
4380 if (verbose)
4381 pr_err("Cannot read from %s\n", devname);
4382 continue; /* Cannot read */
4383 }
4384 if (memcmp(bsb.magic, "md_backup_data-1", 16) != 0 &&
4385 memcmp(bsb.magic, "md_backup_data-2", 16) != 0) {
4386 if (verbose)
4387 pr_err("No backup metadata on %s\n", devname);
4388 continue;
4389 }
4390 if (bsb.sb_csum != bsb_csum((char*)&bsb, ((char*)&bsb.sb_csum)-((char*)&bsb))) {
4391 if (verbose)
4392 pr_err("Bad backup-metadata checksum on %s\n", devname);
4393 continue; /* bad checksum */
4394 }
4395 if (memcmp(bsb.magic, "md_backup_data-2", 16) == 0 &&
4396 bsb.sb_csum2 != bsb_csum((char*)&bsb, ((char*)&bsb.sb_csum2)-((char*)&bsb))) {
4397 if (verbose)
4398 pr_err("Bad backup-metadata checksum2 on %s\n", devname);
4399 continue; /* Bad second checksum */
4400 }
4401 if (memcmp(bsb.set_uuid,info->uuid, 16) != 0) {
4402 if (verbose)
4403 pr_err("Wrong uuid on backup-metadata on %s\n", devname);
4404 continue; /* Wrong uuid */
4405 }
4406
4407 /* array utime and backup-mtime should be updated at much the same time, but it seems that
4408 * sometimes they aren't... So allow considerable flexability in matching, and allow
4409 * this test to be overridden by an environment variable.
4410 */
4411 if (info->array.utime > (int)__le64_to_cpu(bsb.mtime) + 2*60*60 ||
4412 info->array.utime < (int)__le64_to_cpu(bsb.mtime) - 10*60) {
4413 if (check_env("MDADM_GROW_ALLOW_OLD")) {
4414 pr_err("accepting backup with timestamp %lu "
4415 "for array with timestamp %lu\n",
4416 (unsigned long)__le64_to_cpu(bsb.mtime),
4417 (unsigned long)info->array.utime);
4418 } else {
4419 pr_err("too-old timestamp on backup-metadata on %s\n", devname);
4420 pr_err("If you think it is should be safe, try 'export MDADM_GROW_ALLOW_OLD=1'\n");
4421 continue; /* time stamp is too bad */
4422 }
4423 }
4424
4425 if (bsb.magic[15] == '1') {
4426 if (bsb.length == 0)
4427 continue;
4428 if (info->delta_disks >= 0) {
4429 /* reshape_progress is increasing */
4430 if (__le64_to_cpu(bsb.arraystart)
4431 + __le64_to_cpu(bsb.length)
4432 < info->reshape_progress) {
4433 nonew:
4434 if (verbose)
4435 pr_err("backup-metadata found on %s but is not needed\n", devname);
4436 continue; /* No new data here */
4437 }
4438 } else {
4439 /* reshape_progress is decreasing */
4440 if (__le64_to_cpu(bsb.arraystart) >=
4441 info->reshape_progress)
4442 goto nonew; /* No new data here */
4443 }
4444 } else {
4445 if (bsb.length == 0 && bsb.length2 == 0)
4446 continue;
4447 if (info->delta_disks >= 0) {
4448 /* reshape_progress is increasing */
4449 if ((__le64_to_cpu(bsb.arraystart)
4450 + __le64_to_cpu(bsb.length)
4451 < info->reshape_progress)
4452 &&
4453 (__le64_to_cpu(bsb.arraystart2)
4454 + __le64_to_cpu(bsb.length2)
4455 < info->reshape_progress))
4456 goto nonew; /* No new data here */
4457 } else {
4458 /* reshape_progress is decreasing */
4459 if (__le64_to_cpu(bsb.arraystart) >=
4460 info->reshape_progress &&
4461 __le64_to_cpu(bsb.arraystart2) >=
4462 info->reshape_progress)
4463 goto nonew; /* No new data here */
4464 }
4465 }
4466 if (lseek64(fd, __le64_to_cpu(bsb.devstart)*512, 0)< 0) {
4467 second_fail:
4468 if (verbose)
4469 pr_err("Failed to verify secondary backup-metadata block on %s\n",
4470 devname);
4471 continue; /* Cannot seek */
4472 }
4473 /* There should be a duplicate backup superblock 4k before here */
4474 if (lseek64(fd, -4096, 1) < 0 ||
4475 read(fd, &bsb2, sizeof(bsb2)) != sizeof(bsb2))
4476 goto second_fail; /* Cannot find leading superblock */
4477 if (bsb.magic[15] == '1')
4478 bsbsize = offsetof(struct mdp_backup_super, pad1);
4479 else
4480 bsbsize = offsetof(struct mdp_backup_super, pad);
4481 if (memcmp(&bsb2, &bsb, bsbsize) != 0)
4482 goto second_fail; /* Cannot find leading superblock */
4483
4484 /* Now need the data offsets for all devices. */
4485 offsets = xmalloc(sizeof(*offsets)*info->array.raid_disks);
4486 for(j=0; j<info->array.raid_disks; j++) {
4487 if (fdlist[j] < 0)
4488 continue;
4489 if (st->ss->load_super(st, fdlist[j], NULL))
4490 /* FIXME should be this be an error */
4491 continue;
4492 st->ss->getinfo_super(st, &dinfo, NULL);
4493 st->ss->free_super(st);
4494 offsets[j] = dinfo.data_offset * 512;
4495 }
4496 printf(Name ": restoring critical section\n");
4497
4498 if (restore_stripes(fdlist, offsets,
4499 info->array.raid_disks,
4500 info->new_chunk,
4501 info->new_level,
4502 info->new_layout,
4503 fd, __le64_to_cpu(bsb.devstart)*512,
4504 __le64_to_cpu(bsb.arraystart)*512,
4505 __le64_to_cpu(bsb.length)*512, NULL)) {
4506 /* didn't succeed, so giveup */
4507 if (verbose)
4508 pr_err("Error restoring backup from %s\n",
4509 devname);
4510 free(offsets);
4511 return 1;
4512 }
4513
4514 if (bsb.magic[15] == '2' &&
4515 restore_stripes(fdlist, offsets,
4516 info->array.raid_disks,
4517 info->new_chunk,
4518 info->new_level,
4519 info->new_layout,
4520 fd, __le64_to_cpu(bsb.devstart)*512 +
4521 __le64_to_cpu(bsb.devstart2)*512,
4522 __le64_to_cpu(bsb.arraystart2)*512,
4523 __le64_to_cpu(bsb.length2)*512, NULL)) {
4524 /* didn't succeed, so giveup */
4525 if (verbose)
4526 pr_err("Error restoring second backup from %s\n",
4527 devname);
4528 free(offsets);
4529 return 1;
4530 }
4531
4532 free(offsets);
4533
4534 /* Ok, so the data is restored. Let's update those superblocks. */
4535
4536 lo = hi = 0;
4537 if (bsb.length) {
4538 lo = __le64_to_cpu(bsb.arraystart);
4539 hi = lo + __le64_to_cpu(bsb.length);
4540 }
4541 if (bsb.magic[15] == '2' && bsb.length2) {
4542 unsigned long long lo1, hi1;
4543 lo1 = __le64_to_cpu(bsb.arraystart2);
4544 hi1 = lo1 + __le64_to_cpu(bsb.length2);
4545 if (lo == hi) {
4546 lo = lo1;
4547 hi = hi1;
4548 } else if (lo < lo1)
4549 hi = hi1;
4550 else
4551 lo = lo1;
4552 }
4553 if (lo < hi &&
4554 (info->reshape_progress < lo ||
4555 info->reshape_progress > hi))
4556 /* backup does not affect reshape_progress*/ ;
4557 else if (info->delta_disks >= 0) {
4558 info->reshape_progress = __le64_to_cpu(bsb.arraystart) +
4559 __le64_to_cpu(bsb.length);
4560 if (bsb.magic[15] == '2') {
4561 unsigned long long p2 = __le64_to_cpu(bsb.arraystart2) +
4562 __le64_to_cpu(bsb.length2);
4563 if (p2 > info->reshape_progress)
4564 info->reshape_progress = p2;
4565 }
4566 } else {
4567 info->reshape_progress = __le64_to_cpu(bsb.arraystart);
4568 if (bsb.magic[15] == '2') {
4569 unsigned long long p2 = __le64_to_cpu(bsb.arraystart2);
4570 if (p2 < info->reshape_progress)
4571 info->reshape_progress = p2;
4572 }
4573 }
4574 for (j=0; j<info->array.raid_disks; j++) {
4575 if (fdlist[j] < 0)
4576 continue;
4577 if (st->ss->load_super(st, fdlist[j], NULL))
4578 continue;
4579 st->ss->getinfo_super(st, &dinfo, NULL);
4580 dinfo.reshape_progress = info->reshape_progress;
4581 st->ss->update_super(st, &dinfo,
4582 "_reshape_progress",
4583 NULL,0, 0, NULL);
4584 st->ss->store_super(st, fdlist[j]);
4585 st->ss->free_super(st);
4586 }
4587 return 0;
4588 }
4589 /* Didn't find any backup data, try to see if any
4590 * was needed.
4591 */
4592 if (info->delta_disks < 0) {
4593 /* When shrinking, the critical section is at the end.
4594 * So see if we are before the critical section.
4595 */
4596 unsigned long long first_block;
4597 nstripe = ostripe = 0;
4598 first_block = 0;
4599 while (ostripe >= nstripe) {
4600 ostripe += info->array.chunk_size / 512;
4601 first_block = ostripe * odata;
4602 nstripe = first_block / ndata / (info->new_chunk/512) *
4603 (info->new_chunk/512);
4604 }
4605
4606 if (info->reshape_progress >= first_block)
4607 return 0;
4608 }
4609 if (info->delta_disks > 0) {
4610 /* See if we are beyond the critical section. */
4611 unsigned long long last_block;
4612 nstripe = ostripe = 0;
4613 last_block = 0;
4614 while (nstripe >= ostripe) {
4615 nstripe += info->new_chunk / 512;
4616 last_block = nstripe * ndata;
4617 ostripe = last_block / odata / (info->array.chunk_size/512) *
4618 (info->array.chunk_size/512);
4619 }
4620
4621 if (info->reshape_progress >= last_block)
4622 return 0;
4623 }
4624 /* needed to recover critical section! */
4625 if (verbose)
4626 pr_err("Failed to find backup of critical section\n");
4627 return 1;
4628 }
4629
4630 int Grow_continue_command(char *devname, int fd,
4631 char *backup_file, int verbose)
4632 {
4633 int ret_val = 0;
4634 struct supertype *st = NULL;
4635 struct mdinfo *content = NULL;
4636 struct mdinfo array;
4637 char *subarray = NULL;
4638 struct mdinfo *cc = NULL;
4639 struct mdstat_ent *mdstat = NULL;
4640 int cfd = -1;
4641 int fd2 = -1;
4642
4643 dprintf("Grow continue from command line called for %s\n",
4644 devname);
4645
4646 st = super_by_fd(fd, &subarray);
4647 if (!st || !st->ss) {
4648 pr_err("Unable to determine metadata format for %s\n",
4649 devname);
4650 return 1;
4651 }
4652 dprintf("Grow continue is run for ");
4653 if (st->ss->external == 0) {
4654 int d;
4655 dprintf("native array (%s)\n", devname);
4656 if (ioctl(fd, GET_ARRAY_INFO, &array.array) < 0) {
4657 pr_err("%s is not an active md array -"
4658 " aborting\n", devname);
4659 ret_val = 1;
4660 goto Grow_continue_command_exit;
4661 }
4662 content = &array;
4663 /* Need to load a superblock.
4664 * FIXME we should really get what we need from
4665 * sysfs
4666 */
4667 for (d = 0; d < MAX_DISKS; d++) {
4668 mdu_disk_info_t disk;
4669 char *dv;
4670 int err;
4671 disk.number = d;
4672 if (ioctl(fd, GET_DISK_INFO, &disk) < 0)
4673 continue;
4674 if (disk.major == 0 && disk.minor == 0)
4675 continue;
4676 if ((disk.state & (1 << MD_DISK_ACTIVE)) == 0)
4677 continue;
4678 dv = map_dev(disk.major, disk.minor, 1);
4679 if (!dv)
4680 continue;
4681 fd2 = dev_open(dv, O_RDONLY);
4682 if (fd2 < 0)
4683 continue;
4684 err = st->ss->load_super(st, fd2, NULL);
4685 close(fd2);
4686 if (err)
4687 continue;
4688 break;
4689 }
4690 if (d == MAX_DISKS) {
4691 pr_err("Unable to load metadata for %s\n",
4692 devname);
4693 ret_val = 1;
4694 goto Grow_continue_command_exit;
4695 }
4696 st->ss->getinfo_super(st, content, NULL);
4697 } else {
4698 char *container;
4699
4700 if (subarray) {
4701 dprintf("subarray (%s)\n", subarray);
4702 container = st->container_devnm;
4703 cfd = open_dev_excl(st->container_devnm);
4704 } else {
4705 container = st->devnm;
4706 close(fd);
4707 cfd = open_dev_excl(st->devnm);
4708 dprintf("container (%s)\n", container);
4709 fd = cfd;
4710 }
4711 if (cfd < 0) {
4712 pr_err("Unable to open container "
4713 "for %s\n", devname);
4714 ret_val = 1;
4715 goto Grow_continue_command_exit;
4716 }
4717
4718 /* find in container array under reshape
4719 */
4720 ret_val = st->ss->load_container(st, cfd, NULL);
4721 if (ret_val) {
4722 pr_err("Cannot read superblock for %s\n",
4723 devname);
4724 ret_val = 1;
4725 goto Grow_continue_command_exit;
4726 }
4727
4728 cc = st->ss->container_content(st, subarray);
4729 for (content = cc; content ; content = content->next) {
4730 char *array;
4731 int allow_reshape = 1;
4732
4733 if (content->reshape_active == 0)
4734 continue;
4735 /* The decision about array or container wide
4736 * reshape is taken in Grow_continue based
4737 * content->reshape_active state, therefore we
4738 * need to check_reshape based on
4739 * reshape_active and subarray name
4740 */
4741 if (content->array.state & (1<<MD_SB_BLOCK_VOLUME))
4742 allow_reshape = 0;
4743 if (content->reshape_active == CONTAINER_RESHAPE &&
4744 (content->array.state
4745 & (1<<MD_SB_BLOCK_CONTAINER_RESHAPE)))
4746 allow_reshape = 0;
4747
4748 if (!allow_reshape) {
4749 pr_err("cannot continue reshape of an array"
4750 " in container with unsupported"
4751 " metadata: %s(%s)\n",
4752 devname, container);
4753 ret_val = 1;
4754 goto Grow_continue_command_exit;
4755 }
4756
4757 array = strchr(content->text_version+1, '/')+1;
4758 mdstat = mdstat_by_subdev(array, container);
4759 if (!mdstat)
4760 continue;
4761 if (mdstat->active == 0) {
4762 pr_err("Skipping inactive array %s.\n",
4763 mdstat->devnm);
4764 free_mdstat(mdstat);
4765 mdstat = NULL;
4766 continue;
4767 }
4768 break;
4769 }
4770 if (!content) {
4771 pr_err("Unable to determine reshaped "
4772 "array for %s\n", devname);
4773 ret_val = 1;
4774 goto Grow_continue_command_exit;
4775 }
4776 fd2 = open_dev(mdstat->devnm);
4777 if (fd2 < 0) {
4778 pr_err("cannot open (%s)\n", mdstat->devnm);
4779 ret_val = 1;
4780 goto Grow_continue_command_exit;
4781 }
4782
4783 sysfs_init(content, fd2, mdstat->devnm);
4784
4785 /* start mdmon in case it is not running
4786 */
4787 if (!mdmon_running(container))
4788 start_mdmon(container);
4789 ping_monitor(container);
4790
4791 if (mdmon_running(container))
4792 st->update_tail = &st->updates;
4793 else {
4794 pr_err("No mdmon found. "
4795 "Grow cannot continue.\n");
4796 ret_val = 1;
4797 goto Grow_continue_command_exit;
4798 }
4799 }
4800
4801 /* verify that array under reshape is started from
4802 * correct position
4803 */
4804 if (verify_reshape_position(content, content->array.level) < 0) {
4805 ret_val = 1;
4806 goto Grow_continue_command_exit;
4807 }
4808
4809 /* continue reshape
4810 */
4811 ret_val = Grow_continue(fd, st, content, backup_file, 0);
4812
4813 Grow_continue_command_exit:
4814 if (fd2 > -1)
4815 close(fd2);
4816 if (cfd > -1)
4817 close(cfd);
4818 st->ss->free_super(st);
4819 free_mdstat(mdstat);
4820 sysfs_free(cc);
4821 free(subarray);
4822
4823 return ret_val;
4824 }
4825
4826 int Grow_continue(int mdfd, struct supertype *st, struct mdinfo *info,
4827 char *backup_file, int freeze_reshape)
4828 {
4829 int ret_val = 2;
4830
4831 if (!info->reshape_active)
4832 return ret_val;
4833
4834 if (st->ss->external) {
4835 int cfd = open_dev(st->container_devnm);
4836
4837 if (cfd < 0)
4838 return 1;
4839
4840 st->ss->load_container(st, cfd, st->container_devnm);
4841 close(cfd);
4842 ret_val = reshape_container(st->container_devnm, NULL, mdfd,
4843 st, info, 0, backup_file,
4844 0,
4845 1 | info->reshape_active,
4846 freeze_reshape);
4847 } else
4848 ret_val = reshape_array(NULL, mdfd, "array", st, info, 1,
4849 NULL, INVALID_SECTORS,
4850 backup_file, 0, 0,
4851 1 | info->reshape_active,
4852 freeze_reshape);
4853
4854 return ret_val;
4855 }
Unnamed repository; edit this file 'description' to name the repository.