]> git.ipfire.org Git - thirdparty/mdadm.git/blob - Grow.c
Correctly abort level change when reshape_array fails.
[thirdparty/mdadm.git] / Grow.c
1 /*
2 * mdadm - manage Linux "md" devices aka RAID arrays.
3 *
4 * Copyright (C) 2001-2009 Neil Brown <neilb@suse.de>
5 *
6 *
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License as published by
9 * the Free Software Foundation; either version 2 of the License, or
10 * (at your option) any later version.
11 *
12 * This program is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 * GNU General Public License for more details.
16 *
17 * You should have received a copy of the GNU General Public License
18 * along with this program; if not, write to the Free Software
19 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
20 *
21 * Author: Neil Brown
22 * Email: <neilb@suse.de>
23 */
24 #include "mdadm.h"
25 #include "dlink.h"
26 #include <sys/mman.h>
27
28 #if ! defined(__BIG_ENDIAN) && ! defined(__LITTLE_ENDIAN)
29 #error no endian defined
30 #endif
31 #include "md_u.h"
32 #include "md_p.h"
33
34 #ifndef offsetof
35 #define offsetof(t,f) ((size_t)&(((t*)0)->f))
36 #endif
37
38 int Grow_Add_device(char *devname, int fd, char *newdev)
39 {
40 /* Add a device to an active array.
41 * Currently, just extend a linear array.
42 * This requires writing a new superblock on the
43 * new device, calling the kernel to add the device,
44 * and if that succeeds, update the superblock on
45 * all other devices.
46 * This means that we need to *find* all other devices.
47 */
48 struct mdinfo info;
49
50 struct stat stb;
51 int nfd, fd2;
52 int d, nd;
53 struct supertype *st = NULL;
54 char *subarray = NULL;
55
56 if (ioctl(fd, GET_ARRAY_INFO, &info.array) < 0) {
57 fprintf(stderr, Name ": cannot get array info for %s\n", devname);
58 return 1;
59 }
60
61 if (info.array.level != -1) {
62 fprintf(stderr, Name ": can only add devices to linear arrays\n");
63 return 1;
64 }
65
66 st = super_by_fd(fd, &subarray);
67 if (!st) {
68 fprintf(stderr, Name ": cannot handle arrays with superblock version %d\n", info.array.major_version);
69 return 1;
70 }
71
72 if (subarray) {
73 fprintf(stderr, Name ": Cannot grow linear sub-arrays yet\n");
74 free(subarray);
75 free(st);
76 }
77
78 nfd = open(newdev, O_RDWR|O_EXCL|O_DIRECT);
79 if (nfd < 0) {
80 fprintf(stderr, Name ": cannot open %s\n", newdev);
81 free(st);
82 return 1;
83 }
84 fstat(nfd, &stb);
85 if ((stb.st_mode & S_IFMT) != S_IFBLK) {
86 fprintf(stderr, Name ": %s is not a block device!\n", newdev);
87 close(nfd);
88 free(st);
89 return 1;
90 }
91 /* now check out all the devices and make sure we can read the superblock */
92 for (d=0 ; d < info.array.raid_disks ; d++) {
93 mdu_disk_info_t disk;
94 char *dv;
95
96 st->ss->free_super(st);
97
98 disk.number = d;
99 if (ioctl(fd, GET_DISK_INFO, &disk) < 0) {
100 fprintf(stderr, Name ": cannot get device detail for device %d\n",
101 d);
102 close(nfd);
103 free(st);
104 return 1;
105 }
106 dv = map_dev(disk.major, disk.minor, 1);
107 if (!dv) {
108 fprintf(stderr, Name ": cannot find device file for device %d\n",
109 d);
110 close(nfd);
111 free(st);
112 return 1;
113 }
114 fd2 = dev_open(dv, O_RDWR);
115 if (!fd2) {
116 fprintf(stderr, Name ": cannot open device file %s\n", dv);
117 close(nfd);
118 free(st);
119 return 1;
120 }
121
122 if (st->ss->load_super(st, fd2, NULL)) {
123 fprintf(stderr, Name ": cannot find super block on %s\n", dv);
124 close(nfd);
125 close(fd2);
126 free(st);
127 return 1;
128 }
129 close(fd2);
130 }
131 /* Ok, looks good. Lets update the superblock and write it out to
132 * newdev.
133 */
134
135 info.disk.number = d;
136 info.disk.major = major(stb.st_rdev);
137 info.disk.minor = minor(stb.st_rdev);
138 info.disk.raid_disk = d;
139 info.disk.state = (1 << MD_DISK_SYNC) | (1 << MD_DISK_ACTIVE);
140 st->ss->update_super(st, &info, "linear-grow-new", newdev,
141 0, 0, NULL);
142
143 if (st->ss->store_super(st, nfd)) {
144 fprintf(stderr, Name ": Cannot store new superblock on %s\n",
145 newdev);
146 close(nfd);
147 return 1;
148 }
149 close(nfd);
150
151 if (ioctl(fd, ADD_NEW_DISK, &info.disk) != 0) {
152 fprintf(stderr, Name ": Cannot add new disk to this array\n");
153 return 1;
154 }
155 /* Well, that seems to have worked.
156 * Now go through and update all superblocks
157 */
158
159 if (ioctl(fd, GET_ARRAY_INFO, &info.array) < 0) {
160 fprintf(stderr, Name ": cannot get array info for %s\n", devname);
161 return 1;
162 }
163
164 nd = d;
165 for (d=0 ; d < info.array.raid_disks ; d++) {
166 mdu_disk_info_t disk;
167 char *dv;
168
169 disk.number = d;
170 if (ioctl(fd, GET_DISK_INFO, &disk) < 0) {
171 fprintf(stderr, Name ": cannot get device detail for device %d\n",
172 d);
173 return 1;
174 }
175 dv = map_dev(disk.major, disk.minor, 1);
176 if (!dv) {
177 fprintf(stderr, Name ": cannot find device file for device %d\n",
178 d);
179 return 1;
180 }
181 fd2 = dev_open(dv, O_RDWR);
182 if (fd2 < 0) {
183 fprintf(stderr, Name ": cannot open device file %s\n", dv);
184 return 1;
185 }
186 if (st->ss->load_super(st, fd2, NULL)) {
187 fprintf(stderr, Name ": cannot find super block on %s\n", dv);
188 close(fd);
189 return 1;
190 }
191 info.array.raid_disks = nd+1;
192 info.array.nr_disks = nd+1;
193 info.array.active_disks = nd+1;
194 info.array.working_disks = nd+1;
195
196 st->ss->update_super(st, &info, "linear-grow-update", dv,
197 0, 0, NULL);
198
199 if (st->ss->store_super(st, fd2)) {
200 fprintf(stderr, Name ": Cannot store new superblock on %s\n", dv);
201 close(fd2);
202 return 1;
203 }
204 close(fd2);
205 }
206
207 return 0;
208 }
209
210 int Grow_addbitmap(char *devname, int fd, char *file, int chunk, int delay, int write_behind, int force)
211 {
212 /*
213 * First check that array doesn't have a bitmap
214 * Then create the bitmap
215 * Then add it
216 *
217 * For internal bitmaps, we need to check the version,
218 * find all the active devices, and write the bitmap block
219 * to all devices
220 */
221 mdu_bitmap_file_t bmf;
222 mdu_array_info_t array;
223 struct supertype *st;
224 char *subarray = NULL;
225 int major = BITMAP_MAJOR_HI;
226 int vers = md_get_version(fd);
227 unsigned long long bitmapsize, array_size;
228
229 if (vers < 9003) {
230 major = BITMAP_MAJOR_HOSTENDIAN;
231 fprintf(stderr, Name ": Warning - bitmaps created on this kernel"
232 " are not portable\n"
233 " between different architectures. Consider upgrading"
234 " the Linux kernel.\n");
235 }
236
237 if (ioctl(fd, GET_BITMAP_FILE, &bmf) != 0) {
238 if (errno == ENOMEM)
239 fprintf(stderr, Name ": Memory allocation failure.\n");
240 else
241 fprintf(stderr, Name ": bitmaps not supported by this kernel.\n");
242 return 1;
243 }
244 if (bmf.pathname[0]) {
245 if (strcmp(file,"none")==0) {
246 if (ioctl(fd, SET_BITMAP_FILE, -1)!= 0) {
247 fprintf(stderr, Name ": failed to remove bitmap %s\n",
248 bmf.pathname);
249 return 1;
250 }
251 return 0;
252 }
253 fprintf(stderr, Name ": %s already has a bitmap (%s)\n",
254 devname, bmf.pathname);
255 return 1;
256 }
257 if (ioctl(fd, GET_ARRAY_INFO, &array) != 0) {
258 fprintf(stderr, Name ": cannot get array status for %s\n", devname);
259 return 1;
260 }
261 if (array.state & (1<<MD_SB_BITMAP_PRESENT)) {
262 if (strcmp(file, "none")==0) {
263 array.state &= ~(1<<MD_SB_BITMAP_PRESENT);
264 if (ioctl(fd, SET_ARRAY_INFO, &array)!= 0) {
265 fprintf(stderr, Name ": failed to remove internal bitmap.\n");
266 return 1;
267 }
268 return 0;
269 }
270 fprintf(stderr, Name ": Internal bitmap already present on %s\n",
271 devname);
272 return 1;
273 }
274
275 if (strcmp(file, "none") == 0) {
276 fprintf(stderr, Name ": no bitmap found on %s\n", devname);
277 return 1;
278 }
279 if (array.level <= 0) {
280 fprintf(stderr, Name ": Bitmaps not meaningful with level %s\n",
281 map_num(pers, array.level)?:"of this array");
282 return 1;
283 }
284 bitmapsize = array.size;
285 bitmapsize <<= 1;
286 if (get_dev_size(fd, NULL, &array_size) &&
287 array_size > (0x7fffffffULL<<9)) {
288 /* Array is big enough that we cannot trust array.size
289 * try other approaches
290 */
291 bitmapsize = get_component_size(fd);
292 }
293 if (bitmapsize == 0) {
294 fprintf(stderr, Name ": Cannot reliably determine size of array to create bitmap - sorry.\n");
295 return 1;
296 }
297
298 if (array.level == 10) {
299 int ncopies = (array.layout&255)*((array.layout>>8)&255);
300 bitmapsize = bitmapsize * array.raid_disks / ncopies;
301 }
302
303 st = super_by_fd(fd, &subarray);
304 if (!st) {
305 fprintf(stderr, Name ": Cannot understand version %d.%d\n",
306 array.major_version, array.minor_version);
307 return 1;
308 }
309 if (subarray) {
310 fprintf(stderr, Name ": Cannot add bitmaps to sub-arrays yet\n");
311 free(subarray);
312 free(st);
313 return 1;
314 }
315 if (strcmp(file, "internal") == 0) {
316 int d;
317 if (st->ss->add_internal_bitmap == NULL) {
318 fprintf(stderr, Name ": Internal bitmaps not supported "
319 "with %s metadata\n", st->ss->name);
320 return 1;
321 }
322 for (d=0; d< st->max_devs; d++) {
323 mdu_disk_info_t disk;
324 char *dv;
325 disk.number = d;
326 if (ioctl(fd, GET_DISK_INFO, &disk) < 0)
327 continue;
328 if (disk.major == 0 &&
329 disk.minor == 0)
330 continue;
331 if ((disk.state & (1<<MD_DISK_SYNC))==0)
332 continue;
333 dv = map_dev(disk.major, disk.minor, 1);
334 if (dv) {
335 int fd2 = dev_open(dv, O_RDWR);
336 if (fd2 < 0)
337 continue;
338 if (st->ss->load_super(st, fd2, NULL)==0) {
339 if (st->ss->add_internal_bitmap(
340 st,
341 &chunk, delay, write_behind,
342 bitmapsize, 0, major)
343 )
344 st->ss->write_bitmap(st, fd2);
345 else {
346 fprintf(stderr, Name ": failed to create internal bitmap - chunksize problem.\n");
347 close(fd2);
348 return 1;
349 }
350 }
351 close(fd2);
352 }
353 }
354 array.state |= (1<<MD_SB_BITMAP_PRESENT);
355 if (ioctl(fd, SET_ARRAY_INFO, &array)!= 0) {
356 if (errno == EBUSY)
357 fprintf(stderr, Name
358 ": Cannot add bitmap while array is"
359 " resyncing or reshaping etc.\n");
360 fprintf(stderr, Name ": failed to set internal bitmap.\n");
361 return 1;
362 }
363 } else {
364 int uuid[4];
365 int bitmap_fd;
366 int d;
367 int max_devs = st->max_devs;
368
369 /* try to load a superblock */
370 for (d=0; d<max_devs; d++) {
371 mdu_disk_info_t disk;
372 char *dv;
373 int fd2;
374 disk.number = d;
375 if (ioctl(fd, GET_DISK_INFO, &disk) < 0)
376 continue;
377 if ((disk.major==0 && disk.minor==0) ||
378 (disk.state & (1<<MD_DISK_REMOVED)))
379 continue;
380 dv = map_dev(disk.major, disk.minor, 1);
381 if (!dv) continue;
382 fd2 = dev_open(dv, O_RDONLY);
383 if (fd2 >= 0 &&
384 st->ss->load_super(st, fd2, NULL) == 0) {
385 close(fd2);
386 st->ss->uuid_from_super(st, uuid);
387 break;
388 }
389 close(fd2);
390 }
391 if (d == max_devs) {
392 fprintf(stderr, Name ": cannot find UUID for array!\n");
393 return 1;
394 }
395 if (CreateBitmap(file, force, (char*)uuid, chunk,
396 delay, write_behind, bitmapsize, major)) {
397 return 1;
398 }
399 bitmap_fd = open(file, O_RDWR);
400 if (bitmap_fd < 0) {
401 fprintf(stderr, Name ": weird: %s cannot be opened\n",
402 file);
403 return 1;
404 }
405 if (ioctl(fd, SET_BITMAP_FILE, bitmap_fd) < 0) {
406 int err = errno;
407 if (errno == EBUSY)
408 fprintf(stderr, Name
409 ": Cannot add bitmap while array is"
410 " resyncing or reshaping etc.\n");
411 fprintf(stderr, Name ": Cannot set bitmap file for %s: %s\n",
412 devname, strerror(err));
413 return 1;
414 }
415 }
416
417 return 0;
418 }
419
420
421 /*
422 * When reshaping an array we might need to backup some data.
423 * This is written to all spares with a 'super_block' describing it.
424 * The superblock goes 4K from the end of the used space on the
425 * device.
426 * It if written after the backup is complete.
427 * It has the following structure.
428 */
429
430 static struct mdp_backup_super {
431 char magic[16]; /* md_backup_data-1 or -2 */
432 __u8 set_uuid[16];
433 __u64 mtime;
434 /* start/sizes in 512byte sectors */
435 __u64 devstart; /* address on backup device/file of data */
436 __u64 arraystart;
437 __u64 length;
438 __u32 sb_csum; /* csum of preceeding bytes. */
439 __u32 pad1;
440 __u64 devstart2; /* offset in to data of second section */
441 __u64 arraystart2;
442 __u64 length2;
443 __u32 sb_csum2; /* csum of preceeding bytes. */
444 __u8 pad[512-68-32];
445 } __attribute__((aligned(512))) bsb, bsb2;
446
447 static __u32 bsb_csum(char *buf, int len)
448 {
449 int i;
450 int csum = 0;
451 for (i=0; i<len; i++)
452 csum = (csum<<3) + buf[0];
453 return __cpu_to_le32(csum);
454 }
455
456 static int check_idle(struct supertype *st)
457 {
458 /* Check that all member arrays for this container, or the
459 * container of this array, are idle
460 */
461 int container_dev = (st->container_dev != NoMdDev
462 ? st->container_dev : st->devnum);
463 char container[40];
464 struct mdstat_ent *ent, *e;
465 int is_idle = 1;
466
467 fmt_devname(container, container_dev);
468 ent = mdstat_read(0, 0);
469 for (e = ent ; e; e = e->next) {
470 if (!is_container_member(e, container))
471 continue;
472 if (e->percent >= 0) {
473 is_idle = 0;
474 break;
475 }
476 }
477 free_mdstat(ent);
478 return is_idle;
479 }
480
481 static int freeze_container(struct supertype *st)
482 {
483 int container_dev = (st->container_dev != NoMdDev
484 ? st->container_dev : st->devnum);
485 char container[40];
486
487 if (!check_idle(st))
488 return -1;
489
490 fmt_devname(container, container_dev);
491
492 if (block_monitor(container, 1)) {
493 fprintf(stderr, Name ": failed to freeze container\n");
494 return -2;
495 }
496
497 return 1;
498 }
499
500 static void unfreeze_container(struct supertype *st)
501 {
502 int container_dev = (st->container_dev != NoMdDev
503 ? st->container_dev : st->devnum);
504 char container[40];
505
506 fmt_devname(container, container_dev);
507
508 unblock_monitor(container, 1);
509 }
510
511 static int freeze(struct supertype *st)
512 {
513 /* Try to freeze resync/rebuild on this array/container.
514 * Return -1 if the array is busy,
515 * return -2 container cannot be frozen,
516 * return 0 if this kernel doesn't support 'frozen'
517 * return 1 if it worked.
518 */
519 if (st->ss->external)
520 return freeze_container(st);
521 else {
522 struct mdinfo *sra = sysfs_read(-1, st->devnum, GET_VERSION);
523 int err;
524
525 if (!sra)
526 return -1;
527 err = sysfs_freeze_array(sra);
528 sysfs_free(sra);
529 return err;
530 }
531 }
532
533 static void unfreeze(struct supertype *st, int frozen)
534 {
535 /* If 'frozen' is 1, unfreeze the array */
536 if (frozen <= 0)
537 return;
538
539 if (st->ss->external)
540 return unfreeze_container(st);
541 else {
542 struct mdinfo *sra = sysfs_read(-1, st->devnum, GET_VERSION);
543
544 if (sra)
545 sysfs_set_str(sra, NULL, "sync_action", "idle");
546 else
547 fprintf(stderr, Name ": failed to unfreeze array\n");
548 sysfs_free(sra);
549 }
550 }
551
552 static void wait_reshape(struct mdinfo *sra)
553 {
554 int fd = sysfs_get_fd(sra, NULL, "sync_action");
555 char action[20];
556
557 if (fd < 0)
558 return;
559
560 while (sysfs_fd_get_str(fd, action, 20) > 0 &&
561 strncmp(action, "reshape", 7) == 0) {
562 fd_set rfds;
563 FD_ZERO(&rfds);
564 FD_SET(fd, &rfds);
565 select(fd+1, NULL, NULL, &rfds, NULL);
566 }
567 close(fd);
568 }
569
570 static int reshape_super(struct supertype *st, long long size, int level,
571 int layout, int chunksize, int raid_disks,
572 char *backup_file, char *dev, int verbose)
573 {
574 /* nothing extra to check in the native case */
575 if (!st->ss->external)
576 return 0;
577 if (!st->ss->reshape_super ||
578 !st->ss->manage_reshape) {
579 fprintf(stderr, Name ": %s metadata does not support reshape\n",
580 st->ss->name);
581 return 1;
582 }
583
584 return st->ss->reshape_super(st, size, level, layout, chunksize,
585 raid_disks, backup_file, dev, verbose);
586 }
587
588 static void sync_metadata(struct supertype *st)
589 {
590 if (st->ss->external) {
591 if (st->update_tail) {
592 flush_metadata_updates(st);
593 st->update_tail = &st->updates;
594 } else
595 st->ss->sync_metadata(st);
596 }
597 }
598
599 static int subarray_set_num(char *container, struct mdinfo *sra, char *name, int n)
600 {
601 /* when dealing with external metadata subarrays we need to be
602 * prepared to handle EAGAIN. The kernel may need to wait for
603 * mdmon to mark the array active so the kernel can handle
604 * allocations/writeback when preparing the reshape action
605 * (md_allow_write()). We temporarily disable safe_mode_delay
606 * to close a race with the array_state going clean before the
607 * next write to raid_disks / stripe_cache_size
608 */
609 char safe[50];
610 int rc;
611
612 /* only 'raid_disks' and 'stripe_cache_size' trigger md_allow_write */
613 if (!container ||
614 (strcmp(name, "raid_disks") != 0 &&
615 strcmp(name, "stripe_cache_size") != 0))
616 return sysfs_set_num(sra, NULL, name, n);
617
618 rc = sysfs_get_str(sra, NULL, "safe_mode_delay", safe, sizeof(safe));
619 if (rc <= 0)
620 return -1;
621 sysfs_set_num(sra, NULL, "safe_mode_delay", 0);
622 rc = sysfs_set_num(sra, NULL, name, n);
623 if (rc < 0 && errno == EAGAIN) {
624 ping_monitor(container);
625 /* if we get EAGAIN here then the monitor is not active
626 * so stop trying
627 */
628 rc = sysfs_set_num(sra, NULL, name, n);
629 }
630 sysfs_set_str(sra, NULL, "safe_mode_delay", safe);
631 return rc;
632 }
633
634 int start_reshape(struct mdinfo *sra)
635 {
636 int err;
637 sysfs_set_num(sra, NULL, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL);
638 err = sysfs_set_num(sra, NULL, "suspend_hi", 0);
639 err = err ?: sysfs_set_num(sra, NULL, "suspend_lo", 0);
640 err = err ?: sysfs_set_num(sra, NULL, "sync_min", 0);
641 err = err ?: sysfs_set_num(sra, NULL, "sync_max", 0);
642 err = err ?: sysfs_set_str(sra, NULL, "sync_action", "reshape");
643
644 return err;
645 }
646
647 void abort_reshape(struct mdinfo *sra)
648 {
649 sysfs_set_str(sra, NULL, "sync_action", "idle");
650 sysfs_set_num(sra, NULL, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL);
651 sysfs_set_num(sra, NULL, "suspend_hi", 0);
652 sysfs_set_num(sra, NULL, "suspend_lo", 0);
653 sysfs_set_num(sra, NULL, "sync_min", 0);
654 sysfs_set_str(sra, NULL, "sync_max", "max");
655 }
656
657 int remove_disks_on_raid10_to_raid0_takeover(struct supertype *st,
658 struct mdinfo *sra,
659 int layout)
660 {
661 int nr_of_copies;
662 struct mdinfo *remaining;
663 int slot;
664
665 nr_of_copies = layout & 0xff;
666
667 remaining = sra->devs;
668 sra->devs = NULL;
669 /* for each 'copy', select one device and remove from the list. */
670 for (slot = 0; slot < sra->array.raid_disks; slot += nr_of_copies) {
671 struct mdinfo **diskp;
672 int found = 0;
673
674 /* Find a working device to keep */
675 for (diskp = &remaining; *diskp ; diskp = &(*diskp)->next) {
676 struct mdinfo *disk = *diskp;
677
678 if (disk->disk.raid_disk < slot)
679 continue;
680 if (disk->disk.raid_disk >= slot + nr_of_copies)
681 continue;
682 if (disk->disk.state & (1<<MD_DISK_REMOVED))
683 continue;
684 if (disk->disk.state & (1<<MD_DISK_FAULTY))
685 continue;
686 if (!(disk->disk.state & (1<<MD_DISK_SYNC)))
687 continue;
688
689 /* We have found a good disk to use! */
690 *diskp = disk->next;
691 disk->next = sra->devs;
692 sra->devs = disk;
693 found = 1;
694 break;
695 }
696 if (!found)
697 break;
698 }
699
700 if (slot < sra->array.raid_disks) {
701 /* didn't find all slots */
702 struct mdinfo **e;
703 e = &remaining;
704 while (*e)
705 e = &(*e)->next;
706 *e = sra->devs;
707 sra->devs = remaining;
708 return 1;
709 }
710
711 /* Remove all 'remaining' devices from the array */
712 while (remaining) {
713 struct mdinfo *sd = remaining;
714 remaining = sd->next;
715
716 sysfs_set_str(sra, sd, "state", "faulty");
717 sysfs_set_str(sra, sd, "slot", "none");
718 sysfs_set_str(sra, sd, "state", "remove");
719 sd->disk.state |= (1<<MD_DISK_REMOVED);
720 sd->disk.state &= ~(1<<MD_DISK_SYNC);
721 sd->next = sra->devs;
722 sra->devs = sd;
723 }
724 return 0;
725 }
726
727 void reshape_free_fdlist(int *fdlist,
728 unsigned long long *offsets,
729 int size)
730 {
731 int i;
732
733 for (i = 0; i < size; i++)
734 if (fdlist[i] >= 0)
735 close(fdlist[i]);
736
737 free(fdlist);
738 free(offsets);
739 }
740
741 int reshape_prepare_fdlist(char *devname,
742 struct mdinfo *sra,
743 int raid_disks,
744 int nrdisks,
745 unsigned long blocks,
746 char *backup_file,
747 int *fdlist,
748 unsigned long long *offsets)
749 {
750 int d = 0;
751 struct mdinfo *sd;
752
753 for (d = 0; d <= nrdisks; d++)
754 fdlist[d] = -1;
755 d = raid_disks;
756 for (sd = sra->devs; sd; sd = sd->next) {
757 if (sd->disk.state & (1<<MD_DISK_FAULTY))
758 continue;
759 if (sd->disk.state & (1<<MD_DISK_SYNC)) {
760 char *dn = map_dev(sd->disk.major,
761 sd->disk.minor, 1);
762 fdlist[sd->disk.raid_disk]
763 = dev_open(dn, O_RDONLY);
764 offsets[sd->disk.raid_disk] = sd->data_offset*512;
765 if (fdlist[sd->disk.raid_disk] < 0) {
766 fprintf(stderr,
767 Name ": %s: cannot open component %s\n",
768 devname, dn ? dn : "-unknown-");
769 d = -1;
770 goto release;
771 }
772 } else if (backup_file == NULL) {
773 /* spare */
774 char *dn = map_dev(sd->disk.major,
775 sd->disk.minor, 1);
776 fdlist[d] = dev_open(dn, O_RDWR);
777 offsets[d] = (sd->data_offset + sra->component_size - blocks - 8)*512;
778 if (fdlist[d] < 0) {
779 fprintf(stderr, Name ": %s: cannot open component %s\n",
780 devname, dn ? dn : "-unknown-");
781 d = -1;
782 goto release;
783 }
784 d++;
785 }
786 }
787 release:
788 return d;
789 }
790
791 int reshape_open_backup_file(char *backup_file,
792 int fd,
793 char *devname,
794 long blocks,
795 int *fdlist,
796 unsigned long long *offsets)
797 {
798 /* Return 1 on success, 0 on any form of failure */
799 /* need to check backup file is large enough */
800 char buf[512];
801 struct stat stb;
802 unsigned int dev;
803 int i;
804
805 *fdlist = open(backup_file, O_RDWR|O_CREAT|O_EXCL,
806 S_IRUSR | S_IWUSR);
807 *offsets = 8 * 512;
808 if (*fdlist < 0) {
809 fprintf(stderr, Name ": %s: cannot create backup file %s: %s\n",
810 devname, backup_file, strerror(errno));
811 return 0;
812 }
813 /* Guard against backup file being on array device.
814 * If array is partitioned or if LVM etc is in the
815 * way this will not notice, but it is better than
816 * nothing.
817 */
818 fstat(*fdlist, &stb);
819 dev = stb.st_dev;
820 fstat(fd, &stb);
821 if (stb.st_rdev == dev) {
822 fprintf(stderr, Name ": backup file must NOT be"
823 " on the array being reshaped.\n");
824 close(*fdlist);
825 return 0;
826 }
827
828 memset(buf, 0, 512);
829 for (i=0; i < blocks + 1 ; i++) {
830 if (write(*fdlist, buf, 512) != 512) {
831 fprintf(stderr, Name ": %s: cannot create"
832 " backup file %s: %s\n",
833 devname, backup_file, strerror(errno));
834 return 0;
835 }
836 }
837 if (fsync(*fdlist) != 0) {
838 fprintf(stderr, Name ": %s: cannot create backup file %s: %s\n",
839 devname, backup_file, strerror(errno));
840 return 0;
841 }
842
843 return 1;
844 }
845
846 unsigned long compute_backup_blocks(int nchunk, int ochunk,
847 unsigned int ndata, unsigned int odata)
848 {
849 unsigned long a, b, blocks;
850 /* So how much do we need to backup.
851 * We need an amount of data which is both a whole number of
852 * old stripes and a whole number of new stripes.
853 * So LCM for (chunksize*datadisks).
854 */
855 a = (ochunk/512) * odata;
856 b = (nchunk/512) * ndata;
857 /* Find GCD */
858 while (a != b) {
859 if (a < b)
860 b -= a;
861 if (b < a)
862 a -= b;
863 }
864 /* LCM == product / GCD */
865 blocks = (ochunk/512) * (nchunk/512) * odata * ndata / a;
866
867 return blocks;
868 }
869
870 /* 'struct reshape' records the intermediate states
871 * a general reshape.
872 * The starting geometry is converted to the 'before' geometry
873 * by at most an atomic level change. They could be the same.
874 * Similarly the 'after' geometry is converted to the final
875 * geometry by at most a level change.
876 * Note that 'before' and 'after' must have the same level.
877 * 'blocks' is the minimum number of sectors for a reshape unit.
878 * This will be a multiple of the stripe size in each of the
879 * 'before' and 'after' geometries.
880 * If 'blocks' is 0, no restriping is necessary.
881 */
882 struct reshape {
883 int level;
884 int parity; /* number of parity blocks/devices */
885 struct {
886 int layout;
887 int data_disks;
888 } before, after;
889 unsigned long long blocks;
890 unsigned long long stripes; /* number of old stripes that comprise 'blocks'*/
891 unsigned long long new_size; /* New size of array in sectors */
892 };
893
894 char *analyse_change(struct mdinfo *info, struct reshape *re)
895 {
896 /* Based on the current array state in info->array and
897 * the changes in info->new_* etc, determine:
898 * - whether the change is possible
899 * - Intermediate level/raid_disks/layout
900 * - whether a restriping reshape is needed
901 * - number of sectors in minimum change unit. This
902 * will cover a whole number of stripes in 'before' and
903 * 'after'.
904 *
905 * Return message if the change should be rejected
906 * NULL if the change can be achieved
907 *
908 * This can be called as part of starting a reshape, or
909 * when assembling an array that is undergoing reshape.
910 */
911 int new_disks;
912
913 /* If a new level not explicitly given, we assume no-change */
914 if (info->new_level == UnSet)
915 info->new_level = info->array.level;
916
917 if (info->new_chunk)
918 switch (info->new_level) {
919 case 0:
920 case 4:
921 case 5:
922 case 6:
923 case 10:
924 /* chunk size is meaningful, must divide component_size
925 * evenly
926 */
927 if (info->component_size % (info->new_chunk/512))
928 return "New chunk size does not"
929 " divide component size";
930 break;
931 default:
932 return "chunk size not meaningful for this level";
933 }
934 else
935 info->new_chunk = info->array.chunk_size;
936
937 switch (info->array.level) {
938 case 1:
939 /* RAID1 can convert to RAID1 with different disks, or
940 * raid5 with 2 disks
941 */
942 if (info->new_level == 1) {
943 if (info->delta_disks == UnSet)
944 /* Don't know what to do */
945 return "no change requested for Growing RAID1";
946 re->level = 1;
947 re->before.data_disks = (info->array.raid_disks +
948 info->delta_disks);
949 re->before.layout = 0;
950 re->blocks = 0;
951 re->parity = 0;
952 return NULL;
953 }
954 if (info->array.raid_disks == 2 &&
955 info->array.raid_disks == 5) {
956 /* simple in-place conversion */
957 re->level = 5;
958 re->parity = 1;
959 re->before.data_disks = 1;
960 re->before.layout = ALGORITHM_LEFT_SYMMETRIC;
961 re->blocks = 0;
962 return NULL;
963 }
964 /* Could do some multi-stage conversions, but leave that to
965 * later.
966 */
967 return "Impossibly level change request for RAID1";
968
969 case 10:
970 /* RAID10 can only be converted from near mode to
971 * RAID0 by removing some devices
972 */
973 if ((info->array.layout & ~0xff) != 0x100)
974 return "Cannot Grow RAID10 with far/offset layout";
975 /* number of devices must be multiple of number of copies */
976 if (info->array.raid_disks % (info->array.layout & 0xff))
977 return "RAID10 layout too complex for Grow operation";
978
979 if (info->new_level != 0)
980 return "RAID10 can only be changed to RAID0";
981 new_disks = (info->array.raid_disks
982 / (info->array.layout & 0xff));
983 if (info->delta_disks != UnSet) {
984 info->delta_disks = (new_disks
985 - info->array.raid_disks);
986 }
987 if (info->delta_disks != new_disks - info->array.raid_disks)
988 return "New number of raid-devices impossible for RAID10";
989 if (info->new_chunk &&
990 info->new_chunk != info->array.chunk_size)
991 return "Cannot change chunk-size with RAID10 Grow";
992
993 /* looks good */
994 re->level = 0;
995 re->parity = 0;
996 re->before.data_disks = new_disks;
997 re->before.layout = 0;
998 re->blocks = 0;
999 return NULL;
1000
1001 case 0:
1002 /* RAID0 can be converted to RAID10, or to RAID456 */
1003 if (info->new_level == 10) {
1004 if (info->new_layout == UnSet && info->delta_disks == UnSet) {
1005 /* Assume near=2 layout */
1006 info->new_layout = 0x102;
1007 info->delta_disks = info->array.raid_disks;
1008 }
1009 if (info->new_layout == UnSet) {
1010 int copies = 1 + (info->delta_disks
1011 / info->array.raid_disks);
1012 if (info->array.raid_disks * (copies-1)
1013 != info->delta_disks)
1014 return "Impossible number of devices"
1015 " for RAID0->RAID10";
1016 info->new_layout = 0x100 + copies;
1017 }
1018 if (info->delta_disks == UnSet) {
1019 int copies = info->new_layout & 0xff;
1020 if (info->new_layout != 0x100 + copies)
1021 return "New layout impossible"
1022 " for RAID0->RAID10";;
1023 info->delta_disks = (copies - 1) *
1024 info->array.raid_disks;
1025 }
1026 if (info->new_chunk &&
1027 info->new_chunk != info->array.chunk_size)
1028 return "Cannot change chunk-size with RAID0->RAID10";
1029 /* looks good */
1030 re->level = 10;
1031 re->parity = 0;
1032 re->before.data_disks = (info->array.raid_disks +
1033 info->delta_disks);
1034 re->before.layout = info->new_layout;
1035 re->blocks = 0;
1036 return NULL;
1037 }
1038
1039 /* RAID0 can also covert to RAID0/4/5/6 by first converting to
1040 * a raid4 style layout of the final level.
1041 */
1042 switch (info->new_level) {
1043 case 0:
1044 case 4:
1045 re->level = 4;
1046 re->before.layout = 0;
1047 break;
1048 case 5:
1049 re->level = 5;
1050 re->before.layout = ALGORITHM_PARITY_N;
1051 break;
1052 case 6:
1053 re->level = 6;
1054 re->before.layout = ALGORITHM_PARITY_N;
1055 break;
1056 default:
1057 return "Impossible level change requested";
1058 }
1059 re->before.data_disks = info->array.raid_disks;
1060 /* determining 'after' layout happens outside this 'switch' */
1061 break;
1062
1063 case 4:
1064 info->array.layout = ALGORITHM_PARITY_N;
1065 case 5:
1066 switch (info->new_level) {
1067 case 4:
1068 re->level = info->array.level;
1069 re->before.data_disks = info->array.raid_disks - 1;
1070 re->before.layout = info->array.layout;
1071 break;
1072 case 5:
1073 re->level = 5;
1074 re->before.data_disks = info->array.raid_disks - 1;
1075 re->before.layout = info->array.layout;
1076 break;
1077 case 6:
1078 re->level = 6;
1079 re->before.data_disks = info->array.raid_disks - 1;
1080 switch (info->array.layout) {
1081 case ALGORITHM_LEFT_ASYMMETRIC:
1082 re->before.layout = ALGORITHM_LEFT_ASYMMETRIC_6;
1083 break;
1084 case ALGORITHM_RIGHT_ASYMMETRIC:
1085 re->before.layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
1086 break;
1087 case ALGORITHM_LEFT_SYMMETRIC:
1088 re->before.layout = ALGORITHM_LEFT_SYMMETRIC_6;
1089 break;
1090 case ALGORITHM_RIGHT_SYMMETRIC:
1091 re->before.layout = ALGORITHM_RIGHT_SYMMETRIC_6;
1092 break;
1093 case ALGORITHM_PARITY_0:
1094 re->before.layout = ALGORITHM_PARITY_0_6;
1095 break;
1096 case ALGORITHM_PARITY_N:
1097 re->before.layout = ALGORITHM_PARITY_N_6;
1098 break;
1099 default:
1100 return "Cannot convert an array with this layout";
1101 }
1102 break;
1103 case 1:
1104 if (info->array.raid_disks != 2)
1105 return "Can only convert a 2-device array to RAID1";
1106 re->level = 1;
1107 re->before.data_disks = 2;
1108 re->before.layout = 0;
1109 break;
1110 default:
1111 return "Impossible level change requested";
1112 }
1113 break;
1114 case 6:
1115 switch (info->new_level) {
1116 case 4:
1117 case 5:
1118 case 6:
1119 re->level = 6;
1120 re->before.data_disks = info->array.raid_disks - 2;
1121 re->before.layout = info->array.layout;
1122 break;
1123 default:
1124 return "Impossible level change requested";
1125 }
1126 break;
1127 }
1128
1129 /* If we reached here then it looks like a re-stripe is
1130 * happening. We have determined the intermediate level
1131 * and initial raid_disks/layout and stored these in 're'.
1132 *
1133 * We need to deduce the final layout that can be atomically
1134 * converted to the end state.
1135 */
1136 switch (info->new_level) {
1137 case 0:
1138 /* We can only get to RAID0 from RAID4 or RAID5
1139 * with appropriate layout and one extra device
1140 */
1141 if (re->level != 4 && re->level != 5)
1142 return "Cannot covert to RAID0 from this level";
1143 if (info->delta_disks == UnSet)
1144 re->after.data_disks = re->before.data_disks;
1145 else
1146 re->after.data_disks =
1147 info->array.raid_disks + info->delta_disks;
1148 switch (re->level) {
1149 case 4:
1150 re->after.layout = 0 ; break;
1151 case 5:
1152 re->after.layout = ALGORITHM_PARITY_N; break;
1153 }
1154 break;
1155
1156 case 4:
1157 /* We can only get to RAID4 from RAID5 */
1158 if (re->level != 4 && re->level != 5)
1159 return "Cannot convert to RAID4 from this level";
1160 if (info->delta_disks == UnSet)
1161 re->after.data_disks = re->before.data_disks;
1162 else
1163 re->after.data_disks =
1164 re->before.data_disks + info->delta_disks;
1165 switch (re->level) {
1166 case 4:
1167 re->after.layout = 0 ; break;
1168 case 5:
1169 re->after.layout = ALGORITHM_PARITY_N; break;
1170 }
1171 break;
1172
1173 case 5:
1174 /* We get to RAID5 for RAID5 or RAID6 */
1175 if (re->level != 5 && re->level != 6)
1176 return "Cannot convert to RAID5 from this level";
1177 if (info->delta_disks == UnSet)
1178 re->after.data_disks = re->before.data_disks;
1179 else if (re->level == 5)
1180 re->after.data_disks =
1181 re->before.data_disks + info->delta_disks;
1182 else
1183 re->after.data_disks =
1184 info->array.raid_disks + info->delta_disks - 1;
1185 switch (re->level) {
1186 case 5:
1187 if (info->new_layout == UnSet)
1188 re->after.layout = re->before.layout;
1189 else
1190 re->after.layout = info->new_layout;
1191 break;
1192 case 6:
1193 if (info->new_layout == UnSet)
1194 info->new_layout = re->before.layout;
1195
1196 /* after.layout needs to be raid6 version of new_layout */
1197 if (info->new_layout == ALGORITHM_PARITY_N)
1198 re->after.layout = ALGORITHM_PARITY_N;
1199 else {
1200 char layout[40];
1201 char *ls = map_num(r5layout, info->new_layout);
1202 int l;
1203 strcat(strcpy(layout, ls), "-6");
1204 l = map_name(r6layout, layout);
1205 if (l == UnSet)
1206 return "Cannot find RAID6 layout"
1207 " to convert to";
1208 re->after.layout = l;
1209 }
1210 }
1211 break;
1212
1213 case 6:
1214 /* We must already be at level 6 */
1215 if (re->level != 6)
1216 return "Impossible level change";
1217 if (info->delta_disks == UnSet)
1218 re->after.data_disks = re->before.data_disks;
1219 else
1220 re->after.data_disks = (info->array.raid_disks +
1221 info->delta_disks) - 2;
1222 if (info->new_layout == UnSet)
1223 re->after.layout = info->array.layout;
1224 else
1225 re->after.layout = info->new_layout;
1226 break;
1227 default:
1228 return "Impossible level change requested";
1229 }
1230 switch (re->level) {
1231 case 6: re->parity = 2; break;
1232 case 4:
1233 case 5: re->parity = 1; break;
1234 default: re->parity = 0; break;
1235 }
1236 /* So we have a restripe operation, we need to calculate the number
1237 * of blocks per reshape operation.
1238 */
1239 if (info->new_chunk == 0)
1240 info->new_chunk = info->array.chunk_size;
1241 if (re->after.data_disks == re->before.data_disks &&
1242 re->after.layout == re->before.layout &&
1243 info->new_chunk == info->array.chunk_size) {
1244 /* Nothing to change */
1245 re->blocks = 0;
1246 return NULL;
1247 }
1248 if (re->after.data_disks == 1 && re->before.data_disks == 1) {
1249 /* chunks can layout changes make no difference */
1250 re->blocks = 0;
1251 return NULL;
1252 }
1253
1254 if (re->after.data_disks == re->before.data_disks &&
1255 get_linux_version() < 2006032)
1256 return "in-place reshape is not safe before 2.6.32 - sorry.";
1257
1258 if (re->after.data_disks < re->before.data_disks &&
1259 get_linux_version() < 2006030)
1260 return "reshape to fewer devices is not supported before 2.6.32 - sorry.";
1261
1262 re->blocks = compute_backup_blocks(
1263 info->new_chunk, info->array.chunk_size,
1264 re->after.data_disks,
1265 re->before.data_disks);
1266
1267 re->new_size = info->component_size * re->after.data_disks;
1268 return NULL;
1269 }
1270
1271 static int reshape_array(char *container, int fd, char *devname,
1272 struct supertype *st, struct mdinfo *info,
1273 int force, char *backup_file, int quiet, int forked);
1274 static int reshape_container(char *container, int cfd, char *devname,
1275 struct supertype *st,
1276 struct mdinfo *info,
1277 int force,
1278 char *backup_file,
1279 int quiet);
1280 static int child_monitor(int afd, struct mdinfo *sra, struct reshape *reshape,
1281 unsigned long stripes,
1282 int *fds, unsigned long long *offsets,
1283 int dests, int *destfd, unsigned long long *destoffsets);
1284
1285
1286 int Grow_reshape(char *devname, int fd, int quiet, char *backup_file,
1287 long long size,
1288 int level, char *layout_str, int chunksize, int raid_disks,
1289 int force)
1290 {
1291 /* Make some changes in the shape of an array.
1292 * The kernel must support the change.
1293 *
1294 * There are three different changes. Each can trigger
1295 * a resync or recovery so we freeze that until we have
1296 * requested everything (if kernel supports freezing - 2.6.30).
1297 * The steps are:
1298 * - change size (i.e. component_size)
1299 * - change level
1300 * - change layout/chunksize/ndisks
1301 *
1302 * The last can require a reshape. It is different on different
1303 * levels so we need to check the level before actioning it.
1304 * Some times the level change needs to be requested after the
1305 * reshape (e.g. raid6->raid5, raid5->raid0)
1306 *
1307 */
1308 struct mdu_array_info_s array;
1309 int rv = 0;
1310 struct supertype *st;
1311 char *subarray = NULL;
1312
1313 int frozen;
1314 int changed = 0;
1315 char *container = NULL;
1316 char container_buf[20];
1317 int cfd = -1;
1318
1319 struct mdinfo info;
1320 struct mdinfo *sra;
1321
1322 if (ioctl(fd, GET_ARRAY_INFO, &array) < 0) {
1323 fprintf(stderr, Name ": %s is not an active md array - aborting\n",
1324 devname);
1325 return 1;
1326 }
1327
1328 if (size >= 0 &&
1329 (chunksize || level!= UnSet || layout_str || raid_disks)) {
1330 fprintf(stderr, Name ": cannot change component size at the same time "
1331 "as other changes.\n"
1332 " Change size first, then check data is intact before "
1333 "making other changes.\n");
1334 return 1;
1335 }
1336
1337 if (raid_disks && raid_disks < array.raid_disks && array.level > 1 &&
1338 get_linux_version() < 2006032 &&
1339 !check_env("MDADM_FORCE_FEWER")) {
1340 fprintf(stderr, Name ": reducing the number of devices is not safe before Linux 2.6.32\n"
1341 " Please use a newer kernel\n");
1342 return 1;
1343 }
1344
1345 st = super_by_fd(fd, &subarray);
1346 if (!st) {
1347 fprintf(stderr, Name ": Unable to determine metadata format for %s\n", devname);
1348 return 1;
1349 }
1350 if (raid_disks > st->max_devs) {
1351 fprintf(stderr, Name ": Cannot increase raid-disks on this array"
1352 " beyond %d\n", st->max_devs);
1353 return 1;
1354 }
1355
1356 /* in the external case we need to check that the requested reshape is
1357 * supported, and perform an initial check that the container holds the
1358 * pre-requisite spare devices (mdmon owns final validation)
1359 */
1360 if (st->ss->external) {
1361 int container_dev;
1362 int rv;
1363
1364 if (subarray) {
1365 container_dev = st->container_dev;
1366 cfd = open_dev_excl(st->container_dev);
1367 } else {
1368 container_dev = st->devnum;
1369 close(fd);
1370 cfd = open_dev_excl(st->devnum);
1371 fd = cfd;
1372 }
1373 if (cfd < 0) {
1374 fprintf(stderr, Name ": Unable to open container for %s\n",
1375 devname);
1376 free(subarray);
1377 return 1;
1378 }
1379
1380 fmt_devname(container_buf, container_dev);
1381 container = container_buf;
1382
1383 if (subarray)
1384 rv = st->ss->load_container(st, cfd, NULL);
1385 else
1386 rv = st->ss->load_super(st, cfd, NULL);
1387 if (rv) {
1388 fprintf(stderr, Name ": Cannot read superblock for %s\n",
1389 devname);
1390 free(subarray);
1391 return 1;
1392 }
1393
1394 if (mdmon_running(container_dev))
1395 st->update_tail = &st->updates;
1396 }
1397
1398 if (raid_disks > array.raid_disks &&
1399 array.spare_disks < (raid_disks - array.raid_disks) &&
1400 !force) {
1401 fprintf(stderr,
1402 Name ": Need %d spare%s to avoid degraded array,"
1403 " and only have %d.\n"
1404 " Use --force to over-ride this check.\n",
1405 raid_disks - array.raid_disks,
1406 raid_disks - array.raid_disks == 1 ? "" : "s",
1407 array.spare_disks);
1408 return 1;
1409 }
1410
1411 sra = sysfs_read(fd, 0, GET_LEVEL | GET_DISKS | GET_DEVS | GET_STATE);
1412 if (sra) {
1413 if (st->ss->external && subarray == NULL) {
1414 array.level = LEVEL_CONTAINER;
1415 sra->array.level = LEVEL_CONTAINER;
1416 }
1417 } else {
1418 fprintf(stderr, Name ": failed to read sysfs parameters for %s\n",
1419 devname);
1420 return 1;
1421 }
1422 frozen = freeze(st);
1423 if (frozen < -1) {
1424 /* freeze() already spewed the reason */
1425 return 1;
1426 } else if (frozen < 0) {
1427 fprintf(stderr, Name ": %s is performing resync/recovery and cannot"
1428 " be reshaped\n", devname);
1429 return 1;
1430 }
1431
1432 /* ========= set size =============== */
1433 if (size >= 0 && (size == 0 || size != array.size)) {
1434 long long orig_size = array.size;
1435
1436 if (reshape_super(st, size, UnSet, UnSet, 0, 0, NULL, devname, !quiet)) {
1437 rv = 1;
1438 goto release;
1439 }
1440 sync_metadata(st);
1441 array.size = size;
1442 if (array.size != size) {
1443 /* got truncated to 32bit, write to
1444 * component_size instead
1445 */
1446 if (sra)
1447 rv = sysfs_set_num(sra, NULL,
1448 "component_size", size);
1449 else
1450 rv = -1;
1451 } else
1452 rv = ioctl(fd, SET_ARRAY_INFO, &array);
1453 if (rv != 0) {
1454 int err = errno;
1455
1456 /* restore metadata */
1457 if (reshape_super(st, orig_size, UnSet, UnSet, 0, 0,
1458 NULL, devname, !quiet) == 0)
1459 sync_metadata(st);
1460 fprintf(stderr, Name ": Cannot set device size for %s: %s\n",
1461 devname, strerror(err));
1462 if (err == EBUSY &&
1463 (array.state & (1<<MD_SB_BITMAP_PRESENT)))
1464 fprintf(stderr, " Bitmap must be removed before size can be changed\n");
1465 rv = 1;
1466 goto release;
1467 }
1468 ioctl(fd, GET_ARRAY_INFO, &array);
1469 size = get_component_size(fd)/2;
1470 if (size == 0)
1471 size = array.size;
1472 if (!quiet)
1473 fprintf(stderr, Name ": component size of %s has been set to %lluK\n",
1474 devname, size);
1475 changed = 1;
1476 } else if (array.level != LEVEL_CONTAINER) {
1477 size = get_component_size(fd)/2;
1478 if (size == 0)
1479 size = array.size;
1480 }
1481
1482 /* ========= check for Raid10 -> Raid0 conversion ===============
1483 * current implementation assumes that following conditions must be met:
1484 * - far_copies == 1
1485 * - near_copies == 2
1486 */
1487 if (level == 0 && array.level == 10 && sra &&
1488 array.layout == ((1 << 8) + 2) && !(array.raid_disks & 1)) {
1489 int err;
1490 err = remove_disks_on_raid10_to_raid0_takeover(st, sra, array.layout);
1491 if (err) {
1492 dprintf(Name": Array cannot be reshaped\n");
1493 if (cfd > -1)
1494 close(cfd);
1495 rv = 1;
1496 goto release;
1497 }
1498 }
1499
1500 info.array = array;
1501 sysfs_init(&info, fd, NoMdDev);
1502 info.component_size = size*2;
1503 info.new_level = level;
1504 info.new_chunk = chunksize * 1024;
1505 if (raid_disks)
1506 info.delta_disks = raid_disks - info.array.raid_disks;
1507 else
1508 info.delta_disks = UnSet;
1509 if (layout_str == NULL) {
1510 info.new_layout = UnSet;
1511 if (info.array.level == 6 &&
1512 (info.new_level == 6 || info.new_level == UnSet) &&
1513 info.array.layout >= 16) {
1514 fprintf(stderr, Name
1515 ": %s has a non-standard layout. If you"
1516 " wish to preserve this\n"
1517 " during the reshape, please specify"
1518 " --layout=preserve\n"
1519 " If you want to change it, specify a"
1520 " layout or use --layout=normalise\n",
1521 devname);
1522 rv = 1;
1523 goto release;
1524 }
1525 } else if (strcmp(layout_str, "normalise") == 0 ||
1526 strcmp(layout_str, "normalize") == 0) {
1527 /* If we have a -6 RAID6 layout, remove the '-6'. */
1528 info.new_layout = UnSet;
1529 if (info.array.level == 6 && info.new_level == UnSet) {
1530 char l[40], *h;
1531 strcpy(l, map_num(r6layout, info.array.layout));
1532 h = strrchr(l, '-');
1533 if (h && strcmp(h, "-6") == 0) {
1534 *h = 0;
1535 info.new_layout = map_name(r6layout, l);
1536 }
1537 }
1538 } else if (strcmp(layout_str, "preserve") == 0) {
1539 info.new_layout = UnSet;
1540 } else {
1541 int l = info.new_level;
1542 if (l == UnSet)
1543 l = info.array.level;
1544 switch (l) {
1545 case 5:
1546 info.new_layout = map_name(r5layout, layout_str);
1547 break;
1548 case 6:
1549 info.new_layout = map_name(r6layout, layout_str);
1550 break;
1551 case 10:
1552 info.new_layout = parse_layout_10(layout_str);
1553 break;
1554 case LEVEL_FAULTY:
1555 info.new_layout = parse_layout_faulty(layout_str);
1556 break;
1557 default:
1558 fprintf(stderr, Name ": layout not meaningful"
1559 " with this level\n");
1560 rv = 1;
1561 goto release;
1562 }
1563 if (info.new_layout == UnSet) {
1564 fprintf(stderr, Name ": layout %s not understood"
1565 " for this level\n",
1566 layout_str);
1567 rv = 1;
1568 goto release;
1569 }
1570 }
1571
1572 if (array.level == LEVEL_CONTAINER) {
1573 /* This change is to be applied to every array in the
1574 * container. This is only needed when the metadata imposes
1575 * restraints of the various arrays in the container.
1576 * Currently we only know that IMSM requires all arrays
1577 * to have the same number of devices so changing the
1578 * number of devices (On-Line Capacity Expansion) must be
1579 * performed at the level of the container
1580 */
1581 rv = reshape_container(container, fd, devname, st, &info,
1582 force, backup_file, quiet);
1583 } else {
1584 /* Impose these changes on a single array. First
1585 * check that the metadata is OK with the change. */
1586
1587 if (reshape_super(st, info.component_size, info.new_level,
1588 info.new_layout, info.new_chunk,
1589 info.array.raid_disks + info.delta_disks,
1590 backup_file, devname, quiet)) {
1591 rv = 1;
1592 goto release;
1593 }
1594 sync_metadata(st);
1595 rv = reshape_array(container, fd, devname, st, &info, force,
1596 backup_file, quiet, 0);
1597 }
1598 /* reshape_* released the array */
1599 return rv;
1600 release:
1601 unfreeze(st, frozen);
1602 return rv;
1603 }
1604
1605 static int reshape_array(char *container, int fd, char *devname,
1606 struct supertype *st, struct mdinfo *info,
1607 int force,
1608 char *backup_file, int quiet, int forked)
1609 {
1610 struct reshape reshape;
1611 int spares_needed;
1612 char *msg;
1613 int orig_level = UnSet;
1614 int disks, odisks;
1615
1616 struct mdu_array_info_s array;
1617 char *c;
1618 int rv = 0;
1619
1620 int *fdlist;
1621 unsigned long long *offsets;
1622 int d;
1623 int nrdisks;
1624 int err;
1625 int frozen;
1626 unsigned long blocks, stripes;
1627 unsigned long cache;
1628 unsigned long long array_size;
1629 int done;
1630 struct mdinfo *sra, *sd;
1631
1632 msg = analyse_change(info, &reshape);
1633 if (msg) {
1634 fprintf(stderr, Name ": %s\n", msg);
1635 return 1;
1636 }
1637 if (ioctl(fd, GET_ARRAY_INFO, &array) != 0) {
1638 dprintf("Canot get array information.\n");
1639 return 1;
1640 }
1641 spares_needed = max(reshape.before.data_disks,
1642 reshape.after.data_disks)
1643 + reshape.parity - array.raid_disks;
1644
1645 if (!force && spares_needed > info->array.spare_disks) {
1646 fprintf(stderr,
1647 Name ": Need %d spare%s to avoid degraded array,"
1648 " and only have %d.\n"
1649 " Use --force to over-ride this check.\n",
1650 spares_needed,
1651 spares_needed == 1 ? "" : "s",
1652 info->array.spare_disks);
1653 return 1;
1654 }
1655
1656 if (reshape.level != info->array.level) {
1657 char *c = map_num(pers, reshape.level);
1658 int err;
1659 if (c == NULL)
1660 return 1; /* This should not be possible */
1661
1662 err = sysfs_set_str(info, NULL, "level", c);
1663 if (err) {
1664 err = errno;
1665 fprintf(stderr, Name ": %s: could not set level to %s\n",
1666 devname, c);
1667 if (err == EBUSY &&
1668 (info->array.state & (1<<MD_SB_BITMAP_PRESENT)))
1669 fprintf(stderr, " Bitmap must be removed"
1670 " before level can be changed\n");
1671 return 1;
1672 }
1673 if (!quiet)
1674 fprintf(stderr, Name ": level of %s changed to %s\n",
1675 devname, c);
1676 orig_level = info->array.level;
1677 }
1678
1679 if (reshape.level > 0 && st->ss->external &&
1680 !mdmon_running(st->container_dev)) {
1681 start_mdmon(st->container_dev);
1682 ping_monitor(container);
1683 }
1684
1685 /* ->reshape_super might have chosen some spares from the
1686 * container that it wants to be part of the new array.
1687 * We can collect them with ->container_content and give
1688 * them to the kernel.
1689 */
1690 if (st->ss->reshape_super && st->ss->container_content) {
1691 char *subarray = strchr(info->text_version+1, '/')+1;
1692 struct mdinfo *info2 =
1693 st->ss->container_content(st, subarray);
1694 struct mdinfo *d;
1695
1696 if (info2)
1697 for (d = info2->devs; d; d = d->next) {
1698 if (d->disk.state == 0 &&
1699 d->disk.raid_disk >= 0) {
1700 /* This is a spare that wants to
1701 * be part of the array.
1702 */
1703 add_disk(fd, st, info2, d);
1704 }
1705 }
1706 sysfs_free(info2);
1707 }
1708
1709 if (reshape.blocks == 0) {
1710 /* No restriping needed, but we might need to impose
1711 * some more changes: layout, raid_disks, chunk_size
1712 */
1713 if (info->new_layout != UnSet &&
1714 info->new_layout != info->array.layout) {
1715 info->array.layout = info->new_layout;
1716 if (ioctl(fd, SET_ARRAY_INFO, &info->array) != 0) {
1717 fprintf(stderr, Name ": failed to set new layout\n");
1718 rv = 1;
1719 } else if (!quiet)
1720 printf("layout for %s set to %d\n",
1721 devname, info->array.layout);
1722 }
1723 if (info->delta_disks != UnSet &&
1724 info->delta_disks != 0) {
1725 info->array.raid_disks += info->delta_disks;
1726 if (ioctl(fd, SET_ARRAY_INFO, &info->array) != 0) {
1727 fprintf(stderr, Name ": failed to set raid disks\n");
1728 rv = 1;
1729 } else if (!quiet)
1730 printf("raid_disks for %s set to %d\n",
1731 devname, info->array.raid_disks);
1732 }
1733 if (info->new_chunk != 0 &&
1734 info->new_chunk != info->array.chunk_size) {
1735 if (sysfs_set_num(info, NULL,
1736 "chunk_size", info->new_chunk) != 0) {
1737 fprintf(stderr, Name ": failed to set chunk size\n");
1738 rv = 1;
1739 } else if (!quiet)
1740 printf("chunk size for %s set to %d\n",
1741 devname, info->array.chunk_size);
1742 }
1743
1744 return rv;
1745 }
1746
1747 /*
1748 * There are three possibilities.
1749 * 1/ The array will shrink.
1750 * We need to ensure the reshape will pause before reaching
1751 * the 'critical section'. We also need to fork and wait for
1752 * that to happen. When it does we
1753 * suspend/backup/complete/unfreeze
1754 *
1755 * 2/ The array will not change size.
1756 * This requires that we keep a backup of a sliding window
1757 * so that we can restore data after a crash. So we need
1758 * to fork and monitor progress.
1759 * In future we will allow the data_offset to change, so
1760 * a sliding backup becomes unnecessary.
1761 *
1762 * 3/ The array will grow. This is relatively easy.
1763 * However the kernel's restripe routines will cheerfully
1764 * overwrite some early data before it is safe. So we
1765 * need to make a backup of the early parts of the array
1766 * and be ready to restore it if rebuild aborts very early.
1767 * For externally managed metadata, we still need a forked
1768 * child to monitor the reshape and suspend IO over the region
1769 * that is being reshaped.
1770 *
1771 * We backup data by writing it to one spare, or to a
1772 * file which was given on command line.
1773 *
1774 * In each case, we first make sure that storage is available
1775 * for the required backup.
1776 * Then we:
1777 * - request the shape change.
1778 * - fork to handle backup etc.
1779 */
1780
1781 /* Check that we can hold all the data */
1782 get_dev_size(fd, NULL, &array_size);
1783 if (reshape.new_size < (array_size/512)) {
1784 fprintf(stderr,
1785 Name ": this change will reduce the size of the array.\n"
1786 " use --grow --array-size first to truncate array.\n"
1787 " e.g. mdadm --grow %s --array-size %llu\n",
1788 devname, reshape.new_size/2);
1789 rv = 1;
1790 goto release;
1791 }
1792
1793 sra = sysfs_read(fd, 0,
1794 GET_COMPONENT|GET_DEVS|GET_OFFSET|GET_STATE|GET_CHUNK|
1795 GET_CACHE);
1796
1797 if (!sra) {
1798 fprintf(stderr, Name ": %s: Cannot get array details from sysfs\n",
1799 devname);
1800 rv = 1;
1801 goto release;
1802 }
1803
1804 /* Decide how many blocks (sectors) for a reshape
1805 * unit. The number we have so far is just a minimum
1806 */
1807 blocks = reshape.blocks;
1808 if (reshape.before.data_disks ==
1809 reshape.after.data_disks) {
1810 /* Make 'blocks' bigger for better throughput, but
1811 * not so big that we reject it below.
1812 * Try for 16 megabytes
1813 */
1814 while (blocks * 32 < sra->component_size &&
1815 blocks < 16*1024*2)
1816 blocks *= 2;
1817 } else
1818 fprintf(stderr, Name ": Need to backup %luK of critical "
1819 "section..\n", blocks/2);
1820
1821 if (blocks >= sra->component_size/2) {
1822 fprintf(stderr, Name ": %s: Something wrong"
1823 " - reshape aborted\n",
1824 devname);
1825 rv = 1;
1826 goto release;
1827 }
1828
1829 /* Now we need to open all these devices so we can read/write.
1830 */
1831 nrdisks = array.raid_disks + sra->array.spare_disks;
1832 fdlist = malloc((1+nrdisks) * sizeof(int));
1833 offsets = malloc((1+nrdisks) * sizeof(offsets[0]));
1834 if (!fdlist || !offsets) {
1835 fprintf(stderr, Name ": malloc failed: grow aborted\n");
1836 rv = 1;
1837 goto release;
1838 }
1839
1840 d = reshape_prepare_fdlist(devname, sra, array.raid_disks,
1841 nrdisks, blocks, backup_file,
1842 fdlist, offsets);
1843 if (d < 0) {
1844 rv = 1;
1845 goto release;
1846 }
1847 if (backup_file == NULL) {
1848 if (reshape.after.data_disks <= reshape.before.data_disks) {
1849 fprintf(stderr,
1850 Name ": %s: Cannot grow - need backup-file\n",
1851 devname);
1852 rv = 1;
1853 goto release;
1854 } else if (sra->array.spare_disks == 0) {
1855 fprintf(stderr, Name ": %s: Cannot grow - need a spare or "
1856 "backup-file to backup critical section\n",
1857 devname);
1858 rv = 1;
1859 goto release;
1860 }
1861 } else {
1862 if (!reshape_open_backup_file(backup_file, fd, devname,
1863 (signed)blocks,
1864 fdlist+d, offsets+d)) {
1865 rv = 1;
1866 goto release;
1867 }
1868 d++;
1869 }
1870
1871 /* lastly, check that the internal stripe cache is
1872 * large enough, or it won't work.
1873 * It must hold at least 4 stripes of the larger
1874 * chunk size
1875 */
1876 cache = max(info->array.chunk_size, info->new_chunk);
1877 cache *= 4; /* 4 stripes minimum */
1878 cache /= 512; /* convert to sectors */
1879 disks = min(reshape.before.data_disks, reshape.after.data_disks);
1880 /* make sure there is room for 'blocks' with a bit to spare */
1881 if (cache < 16 + blocks / disks)
1882 cache = 16 + blocks / disks;
1883 cache /= (4096/512); /* Covert from sectors to pages */
1884
1885 if (sra->cache_size < cache)
1886 subarray_set_num(container, sra, "stripe_cache_size",
1887 cache+1);
1888
1889 /* Right, everything seems fine. Let's kick things off.
1890 * If only changing raid_disks, use ioctl, else use
1891 * sysfs.
1892 */
1893 sync_metadata(st);
1894
1895 if (info->array.chunk_size == info->new_chunk &&
1896 reshape.before.layout == reshape.after.layout &&
1897 st->ss->external == 0) {
1898 array.raid_disks = reshape.after.data_disks + reshape.parity;
1899 if (ioctl(fd, SET_ARRAY_INFO, &array) != 0) {
1900 int err = errno;
1901 rv = 1;
1902 fprintf(stderr,
1903 Name ": Cannot set device shape for %s: %s\n",
1904 devname, strerror(errno));
1905
1906 if (err == EBUSY &&
1907 (array.state & (1<<MD_SB_BITMAP_PRESENT)))
1908 fprintf(stderr,
1909 " Bitmap must be removed before"
1910 " shape can be changed\n");
1911
1912 goto release;
1913 }
1914 } else {
1915 /* set them all just in case some old 'new_*' value
1916 * persists from some earlier problem
1917 */
1918 int err = err; /* only used if rv==1, and always set if
1919 * rv==1, so initialisation not needed,
1920 * despite gcc warning
1921 */
1922 if (sysfs_set_num(sra, NULL, "chunk_size", info->new_chunk) < 0)
1923 rv = 1, err = errno;
1924 if (!rv && sysfs_set_num(sra, NULL, "layout",
1925 reshape.after.layout) < 0)
1926 rv = 1, err = errno;
1927 if (!rv && subarray_set_num(container, sra, "raid_disks",
1928 reshape.after.data_disks +
1929 reshape.parity) < 0)
1930 rv = 1, err = errno;
1931 if (rv) {
1932 fprintf(stderr, Name ": Cannot set device shape for %s\n",
1933 devname);
1934
1935 if (err == EBUSY &&
1936 (array.state & (1<<MD_SB_BITMAP_PRESENT)))
1937 fprintf(stderr,
1938 " Bitmap must be removed before"
1939 " shape can be changed\n");
1940 goto release;
1941 }
1942 }
1943
1944 start_reshape(sra);
1945 if (st->ss->external) {
1946 /* metadata handler takes it from here */
1947 ping_manager(container);
1948 st->ss->manage_reshape(st, backup_file);
1949 frozen = 0;
1950 goto release;
1951 }
1952
1953 /* set up the backup-super-block. This requires the
1954 * uuid from the array.
1955 */
1956 /* Find a superblock */
1957 for (sd = sra->devs; sd; sd = sd->next) {
1958 char *dn;
1959 int devfd;
1960 int ok;
1961 if (sd->disk.state & (1<<MD_DISK_FAULTY))
1962 continue;
1963 dn = map_dev(sd->disk.major, sd->disk.minor, 1);
1964 devfd = dev_open(dn, O_RDONLY);
1965 if (devfd < 0)
1966 continue;
1967 ok = st->ss->load_super(st, devfd, NULL);
1968 close(devfd);
1969 if (ok >= 0)
1970 break;
1971 }
1972 if (!sd) {
1973 fprintf(stderr, Name ": %s: Cannot find a superblock\n",
1974 devname);
1975 rv = 1;
1976 abort_reshape(sra);
1977 goto release;
1978 }
1979
1980 memset(&bsb, 0, 512);
1981 memcpy(bsb.magic, "md_backup_data-1", 16);
1982 st->ss->uuid_from_super(st, (int*)&bsb.set_uuid);
1983 bsb.mtime = __cpu_to_le64(time(0));
1984 bsb.devstart2 = blocks;
1985
1986 stripes = reshape.blocks / (info->array.chunk_size/512) /
1987 reshape.before.data_disks;
1988
1989 /* Now we just need to kick off the reshape and watch, while
1990 * handling backups of the data...
1991 * This is all done by a forked background process.
1992 */
1993 switch(forked ? 0 : fork()) {
1994 case 0:
1995 close(fd);
1996 if (check_env("MDADM_GROW_VERIFY"))
1997 fd = open(devname, O_RDONLY | O_DIRECT);
1998 else
1999 fd = -1;
2000 mlockall(MCL_FUTURE);
2001
2002 odisks = reshape.before.data_disks + reshape.parity;
2003
2004 done = child_monitor(fd, sra, &reshape, stripes,
2005 fdlist, offsets,
2006 d - odisks, fdlist+odisks, offsets+odisks);
2007
2008 if (backup_file && done)
2009 unlink(backup_file);
2010 if (!done) {
2011 abort_reshape(sra);
2012 goto out;
2013 }
2014 /* set new array size if required customer_array_size is used
2015 * by this metadata.
2016 */
2017 if (reshape.before.data_disks !=
2018 reshape.after.data_disks &&
2019 info->custom_array_size) {
2020 struct mdinfo *info2;
2021 char *subarray = strchr(info->text_version+1, '/')+1;
2022
2023 wait_reshape(sra);
2024 ping_monitor(container);
2025
2026 info2 = st->ss->container_content(st, subarray);
2027 if (info2) {
2028 unsigned long long current_size = 0;
2029 unsigned long long new_size =
2030 info2->custom_array_size/2;
2031
2032 if (sysfs_get_ll(sra,
2033 NULL,
2034 "array_size",
2035 &current_size) == 0 &&
2036 new_size > current_size) {
2037 if (sysfs_set_num(sra, NULL,
2038 "array_size", new_size)
2039 < 0)
2040 dprintf("Error: Cannot"
2041 " set array size");
2042 else
2043 dprintf("Array size "
2044 "changed");
2045 dprintf(" from %llu to %llu.\n",
2046 current_size, new_size);
2047 }
2048 sysfs_free(info2);
2049 }
2050 }
2051
2052 if (info->new_level != reshape.level) {
2053 /* We need to wait for the reshape to finish
2054 * (which will have happened unless
2055 * odata < ndata) and then set the level
2056 */
2057
2058 if (reshape.before.data_disks <
2059 reshape.after.data_disks)
2060 wait_reshape(sra);
2061
2062 c = map_num(pers, info->new_level);
2063 if (c == NULL)
2064 goto out;/* not possible */
2065
2066 err = sysfs_set_str(sra, NULL, "level", c);
2067 if (err)
2068 fprintf(stderr, Name\
2069 ": %s: could not set level "
2070 "to %s\n", devname, c);
2071 }
2072 out:
2073 if (forked)
2074 return 0;
2075 exit(0);
2076 case -1:
2077 fprintf(stderr, Name ": Cannot run child to monitor reshape: %s\n",
2078 strerror(errno));
2079 rv = 1;
2080 abort_reshape(sra);
2081 break;
2082 default:
2083 /* The child will take care of unfreezing the array */
2084 frozen = 0;
2085 break;
2086 }
2087
2088
2089 release:
2090 if (!rv) {
2091 if (container)
2092 ping_monitor(container);
2093 if (st->ss->external) {
2094 /* Re-load the metadata as much could have changed */
2095 int cfd = open_dev(st->container_dev);
2096 if (cfd >= 0) {
2097 st->ss->free_super(st);
2098 st->ss->load_container(st, cfd, container);
2099 close(cfd);
2100 }
2101 }
2102 }
2103 if (rv && orig_level != UnSet && sra) {
2104 c = map_num(pers, orig_level);
2105 if (c && sysfs_set_str(sra, NULL, "level", c) == 0)
2106 fprintf(stderr, Name ": aborting level change\n");
2107 }
2108 unfreeze(st, frozen);
2109 return rv;
2110 }
2111
2112 int reshape_container(char *container, int cfd, char *devname,
2113 struct supertype *st,
2114 struct mdinfo *info,
2115 int force,
2116 char *backup_file,
2117 int quiet)
2118 {
2119 struct mdinfo *cc = NULL;
2120
2121 if (reshape_super(st, info->component_size, info->new_level,
2122 info->new_layout, info->new_chunk,
2123 info->array.raid_disks + info->delta_disks,
2124 backup_file, devname, quiet))
2125 return 1;
2126
2127 sync_metadata(st);
2128
2129 /* ping monitor to be sure that update is on disk
2130 */
2131 ping_monitor(container);
2132
2133 switch (fork()) {
2134 case -1: /* error */
2135 perror("Cannot fork to complete reshape\n");
2136 return 1;
2137 default: /* parent */
2138 printf(Name ": multi-array reshape continues in background\n");
2139 return 0;
2140 case 0: /* child */
2141 break;
2142 }
2143
2144 while(1) {
2145 /* For each member array with reshape_active,
2146 * we need to perform the reshape.
2147 * We pick the first array that needs reshaping and
2148 * reshape it. reshape_array() will re-read the metadata
2149 * so the next time through a different array should be
2150 * ready for reshape.
2151 */
2152 struct mdinfo *content;
2153 int rv;
2154 int fd;
2155 struct mdstat_ent *mdstat;
2156 char *adev;
2157
2158 sysfs_free(cc);
2159
2160 cc = st->ss->container_content(st, NULL);
2161
2162 for (content = cc; content ; content = content->next) {
2163 char *subarray;
2164 if (!content->reshape_active)
2165 continue;
2166
2167 subarray = strchr(content->text_version+1, '/')+1;
2168 mdstat = mdstat_by_subdev(subarray,
2169 devname2devnum(container));
2170 if (!mdstat)
2171 continue;
2172 break;
2173 }
2174 if (!content)
2175 break;
2176
2177 fd = open_dev_excl(mdstat->devnum);
2178 if (fd < 0)
2179 break;
2180 adev = map_dev(dev2major(mdstat->devnum),
2181 dev2minor(mdstat->devnum),
2182 0);
2183 if (!adev)
2184 adev = content->text_version;
2185
2186 sysfs_init(content, fd, mdstat->devnum);
2187
2188 rv = reshape_array(container, fd, adev, st,
2189 content, force,
2190 backup_file, quiet, 1);
2191 close(fd);
2192 if (rv)
2193 break;
2194 }
2195 sysfs_free(cc);
2196 exit(0);
2197 }
2198
2199 /*
2200 * We run a child process in the background which performs the following
2201 * steps:
2202 * - wait for resync to reach a certain point
2203 * - suspend io to the following section
2204 * - backup that section
2205 * - allow resync to proceed further
2206 * - resume io
2207 * - discard the backup.
2208 *
2209 * When are combined in slightly different ways in the three cases.
2210 * Grow:
2211 * - suspend/backup/allow/wait/resume/discard
2212 * Shrink:
2213 * - allow/wait/suspend/backup/allow/wait/resume/discard
2214 * same-size:
2215 * - wait/resume/discard/suspend/backup/allow
2216 *
2217 * suspend/backup/allow always come together
2218 * wait/resume/discard do too.
2219 * For the same-size case we have two backups to improve flow.
2220 *
2221 */
2222
2223 int progress_reshape(struct mdinfo *info, struct reshape *reshape,
2224 unsigned long long backup_point,
2225 unsigned long long wait_point,
2226 unsigned long long *suspend_point,
2227 unsigned long long *reshape_completed)
2228 {
2229 /* This function is called repeatedly by the reshape manager.
2230 * It determines how much progress can safely be made and allows
2231 * that progress.
2232 * - 'info' identifies the array and particularly records in
2233 * ->reshape_progress the metadata's knowledge of progress
2234 * This is a sector offset from the start of the array
2235 * of the next array block to be relocated. This number
2236 * may increase from 0 or decrease from array_size, depending
2237 * on the type of reshape that is happening.
2238 * Note that in contrast, 'sync_completed' is a block count of the
2239 * reshape so far. It gives the distance between the start point
2240 * (head or tail of device) and the next place that data will be
2241 * written. It always increases.
2242 * - 'reshape' is the structure created by analyse_change
2243 * - 'backup_point' shows how much the metadata manager has backed-up
2244 * data. For reshapes with increasing progress, it is the next address
2245 * to be backed up, previous addresses have been backed-up. For
2246 * decreasing progress, it is the earliest address that has been
2247 * backed up - later address are also backed up.
2248 * So addresses between reshape_progress and backup_point are
2249 * backed up providing those are in the 'correct' order.
2250 * - 'wait_point' is an array address. When reshape_completed
2251 * passes this point, progress_reshape should return. It might
2252 * return earlier if it determines that ->reshape_progress needs
2253 * to be updated or further backup is needed.
2254 * - suspend_point is maintained by progress_reshape and the caller
2255 * should not touch it except to initialise to zero.
2256 * It is an array address and it only increases in 2.6.37 and earlier.
2257 * This makes it difficult to handle reducing reshapes with
2258 * external metadata.
2259 * However: it is similar to backup_point in that it records the
2260 * other end of a suspended region from reshape_progress.
2261 * it is moved to extend the region that is safe to backup and/or
2262 * reshape
2263 * - reshape_completed is read from sysfs and returned. The caller
2264 * should copy this into ->reshape_progress when it has reason to
2265 * believe that the metadata knows this, and any backup outside this
2266 * has been erased.
2267 *
2268 * Return value is:
2269 * 1 if more data from backup_point - but only as far as suspend_point,
2270 * should be backed up
2271 * 0 if things are progressing smoothly
2272 * -1 if the reshape is finished, either because it is all done,
2273 * or due to an error.
2274 */
2275
2276 int advancing = (reshape->after.data_disks
2277 >= reshape->before.data_disks);
2278 int need_backup = (reshape->after.data_disks
2279 == reshape->before.data_disks);
2280 unsigned long long read_offset, write_offset;
2281 unsigned long long read_range, write_range;
2282 unsigned long long max_progress, target, completed;
2283 int fd;
2284
2285 /* First, we unsuspend any region that is now known to be safe.
2286 * If suspend_point is on the 'wrong' side of reshape_progress, then
2287 * we don't have or need suspension at the moment. This is true for
2288 * native metadata when we don't need to back-up.
2289 */
2290 if (advancing) {
2291 if (info->reshape_progress < *suspend_point)
2292 sysfs_set_num(info, NULL, "suspend_lo",
2293 info->reshape_progress);
2294 } else {
2295 /* Note: this won't work in 2.6.37 and before.
2296 * Something somewhere should make sure we don't need it!
2297 */
2298 if (info->reshape_progress > *suspend_point)
2299 sysfs_set_num(info, NULL, "suspend_hi",
2300 info->reshape_progress);
2301 }
2302
2303 /* Now work out how far it is safe to progress.
2304 * If the read_offset for ->reshape_progress is less than
2305 * 'blocks' beyond the write_offset, we can only progress as far
2306 * as a backup.
2307 * Otherwise we can progress until the write_offset for the new location
2308 * reaches (within 'blocks' of) the read_offset at the current location.
2309 * However that region must be suspended unless we are using native
2310 * metadata.
2311 * If we need to suspend more, we limit it to 128M per device, which is
2312 * rather arbitrary and should be some time-based calculation.
2313 */
2314 write_offset = info->reshape_progress / reshape->before.data_disks;
2315 read_offset = info->reshape_progress / reshape->after.data_disks;
2316 write_range = reshape->blocks / reshape->before.data_disks;
2317 read_range = reshape->blocks / reshape->after.data_disks;
2318 if (advancing) {
2319 if (read_offset < write_offset + write_range) {
2320 max_progress = backup_point;
2321 if (max_progress <= info->reshape_progress)
2322 need_backup = 1;
2323 } else {
2324 max_progress =
2325 (read_offset - write_range) *
2326 reshape->before.data_disks;
2327 }
2328 } else {
2329 if (read_offset > write_offset - write_range) {
2330 max_progress = backup_point;
2331 if (max_progress >= info->reshape_progress)
2332 need_backup = 1;
2333 } else {
2334 max_progress =
2335 (read_offset + write_range) *
2336 reshape->before.data_disks;
2337 /* If we are using internal metadata, then we can
2338 * progress all the way to the suspend_point without
2339 * worrying about backing-up/suspending along the
2340 * way.
2341 */
2342 if (max_progress < *suspend_point &&
2343 info->array.major_version >= 0)
2344 max_progress = *suspend_point;
2345 }
2346 }
2347
2348 /* We know it is safe to progress to 'max_progress' providing
2349 * it is suspended or we are using native metadata.
2350 * Consider extending suspend_point 128M per device if it
2351 * is less than 64M per device beyond reshape_progress.
2352 * But always do a multiple of 'blocks'
2353 */
2354 target = 64*1024*2 * min(reshape->before.data_disks,
2355 reshape->after.data_disks);
2356 target /= reshape->blocks;
2357 if (target < 2)
2358 target = 2;
2359 target *= reshape->blocks;
2360
2361 /* For externally managed metadata we always need to suspend IO to
2362 * the area being reshaped so we regularly push suspend_point forward.
2363 * For native metadata we only need the suspend if we are going to do
2364 * a backup.
2365 */
2366 if (advancing) {
2367 if ((need_backup || info->array.major_version < 0) &&
2368 *suspend_point < info->reshape_progress + target) {
2369 if (max_progress < *suspend_point + 2 * target)
2370 *suspend_point = max_progress;
2371 else
2372 *suspend_point += 2 * target;
2373 sysfs_set_num(info, NULL, "suspend_hi", *suspend_point);
2374 max_progress = *suspend_point;
2375 }
2376 } else {
2377 if ((need_backup || info->array.major_version < 0) &&
2378 *suspend_point > info->reshape_progress - target) {
2379 if (max_progress > *suspend_point - 2 * target)
2380 *suspend_point = max_progress;
2381 else
2382 *suspend_point -= 2 * target;
2383 sysfs_set_num(info, NULL, "suspend_lo", *suspend_point);
2384 max_progress = *suspend_point;
2385 }
2386 }
2387
2388 /* now set sync_max to allow that progress. sync_max, like
2389 * sync_completed is a count of sectors written per device, so
2390 * we find the difference between max_progress and the start point,
2391 * and divide that by after.data_disks to get a sync_max
2392 * number.
2393 * At the same time we convert wait_point to a similar number
2394 * for comparing against sync_completed.
2395 */
2396 if (!advancing) {
2397 max_progress = info->component_size * reshape->after.data_disks
2398 - max_progress;
2399 wait_point = info->component_size * reshape->after.data_disks
2400 - wait_point;
2401 }
2402 max_progress /= reshape->after.data_disks;
2403 wait_point /= reshape->after.data_disks;
2404
2405 sysfs_set_num(info, NULL, "sync_max", max_progress);
2406
2407 /* Now wait. If we have already reached the point that we were
2408 * asked to wait to, don't wait at all, else wait for any change.
2409 * We need to select on 'sync_completed' as that is the place that
2410 * notifications happen, but we are really interested in
2411 * 'reshape_position'
2412 */
2413 fd = sysfs_get_fd(info, NULL, "sync_completed");
2414 if (fd < 0)
2415 return -1;
2416
2417 if (sysfs_fd_get_ll(fd, &completed) < 0) {
2418 close(fd);
2419 return -1;
2420 }
2421 while (completed < max_progress && completed < wait_point) {
2422 /* Check that sync_action is still 'reshape' to avoid
2423 * waiting forever on a dead array
2424 */
2425 char action[20];
2426 fd_set rfds;
2427 if (sysfs_get_str(info, NULL, "sync_action",
2428 action, 20) <= 0 ||
2429 strncmp(action, "reshape", 7) != 0)
2430 break;
2431 FD_ZERO(&rfds);
2432 FD_SET(fd, &rfds);
2433 select(fd+1, NULL, NULL, &rfds, NULL);
2434 if (sysfs_fd_get_ll(fd, &completed) < 0) {
2435 close(fd);
2436 return -1;
2437 }
2438 }
2439 /* Convert 'completed' back in to a 'progress' number */
2440 completed *= reshape->after.data_disks;
2441 if (!advancing) {
2442 completed = info->component_size * reshape->after.data_disks
2443 - completed;
2444 }
2445 *reshape_completed = completed;
2446
2447 close(fd);
2448
2449 /* We return the need_backup flag. Caller will decide
2450 * how much (a multiple of ->blocks) and will adjust
2451 * suspend_{lo,hi} and suspend_point.
2452 */
2453 return need_backup;
2454 }
2455
2456
2457 /* FIXME return status is never checked */
2458 static int grow_backup(struct mdinfo *sra,
2459 unsigned long long offset, /* per device */
2460 unsigned long stripes, /* per device, in old chunks */
2461 int *sources, unsigned long long *offsets,
2462 int disks, int chunk, int level, int layout,
2463 int dests, int *destfd, unsigned long long *destoffsets,
2464 int part, int *degraded,
2465 char *buf)
2466 {
2467 /* Backup 'blocks' sectors at 'offset' on each device of the array,
2468 * to storage 'destfd' (offset 'destoffsets'), after first
2469 * suspending IO. Then allow resync to continue
2470 * over the suspended section.
2471 * Use part 'part' of the backup-super-block.
2472 */
2473 int odata = disks;
2474 int rv = 0;
2475 int i;
2476 unsigned long long ll;
2477 int new_degraded;
2478 //printf("offset %llu\n", offset);
2479 if (level >= 4)
2480 odata--;
2481 if (level == 6)
2482 odata--;
2483
2484 /* Check that array hasn't become degraded, else we might backup the wrong data */
2485 sysfs_get_ll(sra, NULL, "degraded", &ll);
2486 new_degraded = (int)ll;
2487 if (new_degraded != *degraded) {
2488 /* check each device to ensure it is still working */
2489 struct mdinfo *sd;
2490 for (sd = sra->devs ; sd ; sd = sd->next) {
2491 if (sd->disk.state & (1<<MD_DISK_FAULTY))
2492 continue;
2493 if (sd->disk.state & (1<<MD_DISK_SYNC)) {
2494 char sbuf[20];
2495 if (sysfs_get_str(sra, sd, "state", sbuf, 20) < 0 ||
2496 strstr(sbuf, "faulty") ||
2497 strstr(sbuf, "in_sync") == NULL) {
2498 /* this device is dead */
2499 sd->disk.state = (1<<MD_DISK_FAULTY);
2500 if (sd->disk.raid_disk >= 0 &&
2501 sources[sd->disk.raid_disk] >= 0) {
2502 close(sources[sd->disk.raid_disk]);
2503 sources[sd->disk.raid_disk] = -1;
2504 }
2505 }
2506 }
2507 }
2508 *degraded = new_degraded;
2509 }
2510 if (part) {
2511 bsb.arraystart2 = __cpu_to_le64(offset * odata);
2512 bsb.length2 = __cpu_to_le64(stripes * (chunk/512) * odata);
2513 } else {
2514 bsb.arraystart = __cpu_to_le64(offset * odata);
2515 bsb.length = __cpu_to_le64(stripes * (chunk/512) * odata);
2516 }
2517 if (part)
2518 bsb.magic[15] = '2';
2519 for (i = 0; i < dests; i++)
2520 if (part)
2521 lseek64(destfd[i], destoffsets[i] + __le64_to_cpu(bsb.devstart2)*512, 0);
2522 else
2523 lseek64(destfd[i], destoffsets[i], 0);
2524
2525 rv = save_stripes(sources, offsets,
2526 disks, chunk, level, layout,
2527 dests, destfd,
2528 offset*512*odata, stripes * chunk * odata,
2529 buf);
2530
2531 if (rv)
2532 return rv;
2533 bsb.mtime = __cpu_to_le64(time(0));
2534 for (i = 0; i < dests; i++) {
2535 bsb.devstart = __cpu_to_le64(destoffsets[i]/512);
2536
2537 bsb.sb_csum = bsb_csum((char*)&bsb, ((char*)&bsb.sb_csum)-((char*)&bsb));
2538 if (memcmp(bsb.magic, "md_backup_data-2", 16) == 0)
2539 bsb.sb_csum2 = bsb_csum((char*)&bsb,
2540 ((char*)&bsb.sb_csum2)-((char*)&bsb));
2541
2542 rv = -1;
2543 if ((unsigned long long)lseek64(destfd[i], destoffsets[i] - 4096, 0)
2544 != destoffsets[i] - 4096)
2545 break;
2546 if (write(destfd[i], &bsb, 512) != 512)
2547 break;
2548 if (destoffsets[i] > 4096) {
2549 if ((unsigned long long)lseek64(destfd[i], destoffsets[i]+stripes*chunk*odata, 0) !=
2550 destoffsets[i]+stripes*chunk*odata)
2551 break;
2552 if (write(destfd[i], &bsb, 512) != 512)
2553 break;
2554 }
2555 fsync(destfd[i]);
2556 rv = 0;
2557 }
2558
2559 return rv;
2560 }
2561
2562 /* in 2.6.30, the value reported by sync_completed can be
2563 * less that it should be by one stripe.
2564 * This only happens when reshape hits sync_max and pauses.
2565 * So allow wait_backup to either extent sync_max further
2566 * than strictly necessary, or return before the
2567 * sync has got quite as far as we would really like.
2568 * This is what 'blocks2' is for.
2569 * The various caller give appropriate values so that
2570 * every works.
2571 */
2572 /* FIXME return value is often ignored */
2573 static int forget_backup(
2574 int dests, int *destfd, unsigned long long *destoffsets,
2575 int part)
2576 {
2577 /*
2578 * Erase backup 'part' (which is 0 or 1)
2579 */
2580 int i;
2581 int rv;
2582
2583 if (part) {
2584 bsb.arraystart2 = __cpu_to_le64(0);
2585 bsb.length2 = __cpu_to_le64(0);
2586 } else {
2587 bsb.arraystart = __cpu_to_le64(0);
2588 bsb.length = __cpu_to_le64(0);
2589 }
2590 bsb.mtime = __cpu_to_le64(time(0));
2591 rv = 0;
2592 for (i = 0; i < dests; i++) {
2593 bsb.devstart = __cpu_to_le64(destoffsets[i]/512);
2594 bsb.sb_csum = bsb_csum((char*)&bsb, ((char*)&bsb.sb_csum)-((char*)&bsb));
2595 if (memcmp(bsb.magic, "md_backup_data-2", 16) == 0)
2596 bsb.sb_csum2 = bsb_csum((char*)&bsb,
2597 ((char*)&bsb.sb_csum2)-((char*)&bsb));
2598 if ((unsigned long long)lseek64(destfd[i], destoffsets[i]-4096, 0) !=
2599 destoffsets[i]-4096)
2600 rv = -1;
2601 if (rv == 0 &&
2602 write(destfd[i], &bsb, 512) != 512)
2603 rv = -1;
2604 fsync(destfd[i]);
2605 }
2606 return rv;
2607 }
2608
2609 static void fail(char *msg)
2610 {
2611 int rv;
2612 rv = (write(2, msg, strlen(msg)) != (int)strlen(msg));
2613 rv |= (write(2, "\n", 1) != 1);
2614 exit(rv ? 1 : 2);
2615 }
2616
2617 static char *abuf, *bbuf;
2618 static unsigned long long abuflen;
2619 static void validate(int afd, int bfd, unsigned long long offset)
2620 {
2621 /* check that the data in the backup against the array.
2622 * This is only used for regression testing and should not
2623 * be used while the array is active
2624 */
2625 if (afd < 0)
2626 return;
2627 lseek64(bfd, offset - 4096, 0);
2628 if (read(bfd, &bsb2, 512) != 512)
2629 fail("cannot read bsb");
2630 if (bsb2.sb_csum != bsb_csum((char*)&bsb2,
2631 ((char*)&bsb2.sb_csum)-((char*)&bsb2)))
2632 fail("first csum bad");
2633 if (memcmp(bsb2.magic, "md_backup_data", 14) != 0)
2634 fail("magic is bad");
2635 if (memcmp(bsb2.magic, "md_backup_data-2", 16) == 0 &&
2636 bsb2.sb_csum2 != bsb_csum((char*)&bsb2,
2637 ((char*)&bsb2.sb_csum2)-((char*)&bsb2)))
2638 fail("second csum bad");
2639
2640 if (__le64_to_cpu(bsb2.devstart)*512 != offset)
2641 fail("devstart is wrong");
2642
2643 if (bsb2.length) {
2644 unsigned long long len = __le64_to_cpu(bsb2.length)*512;
2645
2646 if (abuflen < len) {
2647 free(abuf);
2648 free(bbuf);
2649 abuflen = len;
2650 if (posix_memalign((void**)&abuf, 4096, abuflen) ||
2651 posix_memalign((void**)&bbuf, 4096, abuflen)) {
2652 abuflen = 0;
2653 /* just stop validating on mem-alloc failure */
2654 return;
2655 }
2656 }
2657
2658 lseek64(bfd, offset, 0);
2659 if ((unsigned long long)read(bfd, bbuf, len) != len) {
2660 //printf("len %llu\n", len);
2661 fail("read first backup failed");
2662 }
2663 lseek64(afd, __le64_to_cpu(bsb2.arraystart)*512, 0);
2664 if ((unsigned long long)read(afd, abuf, len) != len)
2665 fail("read first from array failed");
2666 if (memcmp(bbuf, abuf, len) != 0) {
2667 #if 0
2668 int i;
2669 printf("offset=%llu len=%llu\n",
2670 (unsigned long long)__le64_to_cpu(bsb2.arraystart)*512, len);
2671 for (i=0; i<len; i++)
2672 if (bbuf[i] != abuf[i]) {
2673 printf("first diff byte %d\n", i);
2674 break;
2675 }
2676 #endif
2677 fail("data1 compare failed");
2678 }
2679 }
2680 if (bsb2.length2) {
2681 unsigned long long len = __le64_to_cpu(bsb2.length2)*512;
2682
2683 if (abuflen < len) {
2684 free(abuf);
2685 free(bbuf);
2686 abuflen = len;
2687 abuf = malloc(abuflen);
2688 bbuf = malloc(abuflen);
2689 }
2690
2691 lseek64(bfd, offset+__le64_to_cpu(bsb2.devstart2)*512, 0);
2692 if ((unsigned long long)read(bfd, bbuf, len) != len)
2693 fail("read second backup failed");
2694 lseek64(afd, __le64_to_cpu(bsb2.arraystart2)*512, 0);
2695 if ((unsigned long long)read(afd, abuf, len) != len)
2696 fail("read second from array failed");
2697 if (memcmp(bbuf, abuf, len) != 0)
2698 fail("data2 compare failed");
2699 }
2700 }
2701
2702 static int child_monitor(int afd, struct mdinfo *sra, struct reshape *reshape,
2703 unsigned long stripes,
2704 int *fds, unsigned long long *offsets,
2705 int dests, int *destfd, unsigned long long *destoffsets)
2706 {
2707 /* Monitor a reshape where backup is being performed using
2708 * 'native' mechanism - either to a backup file, or
2709 * to some space in a spare.
2710 */
2711 char *buf;
2712 int degraded = -1;
2713 unsigned long long speed;
2714 unsigned long long suspend_point, array_size;
2715 unsigned long long backup_point, wait_point;
2716 unsigned long long reshape_completed;
2717 int done = 0;
2718 int increasing = reshape->after.data_disks >= reshape->before.data_disks;
2719 int part = 0; /* The next part of the backup area to fill. It may already
2720 * be full, so we need to check */
2721 int level = reshape->level;
2722 int layout = reshape->before.layout;
2723 int data = reshape->before.data_disks;
2724 int disks = reshape->before.data_disks + reshape->parity;
2725 int chunk = sra->array.chunk_size;
2726
2727 if (posix_memalign((void**)&buf, 4096, disks * chunk))
2728 /* Don't start the 'reshape' */
2729 return 0;
2730 if (reshape->before.data_disks == reshape->after.data_disks) {
2731 sysfs_get_ll(sra, NULL, "sync_speed_min", &speed);
2732 sysfs_set_num(sra, NULL, "sync_speed_min", 200000);
2733 }
2734
2735 array_size = sra->component_size * data;
2736 if (increasing) {
2737 backup_point = sra->reshape_progress;
2738 suspend_point = 0;
2739 } else {
2740 backup_point = array_size;
2741 suspend_point = array_size;
2742 }
2743
2744 while (!done) {
2745 int rv;
2746
2747 /* Want to return as soon the oldest backup slot can
2748 * be released as that allows us to start backing up
2749 * some more, providing suspend_point has been
2750 * advanced, which it should have.
2751 */
2752 if (increasing) {
2753 wait_point = array_size;
2754 if (part == 0 && __le64_to_cpu(bsb.length) > 0)
2755 wait_point = (__le64_to_cpu(bsb.arraystart) +
2756 __le64_to_cpu(bsb.length));
2757 if (part == 1 && __le64_to_cpu(bsb.length2) > 0)
2758 wait_point = (__le64_to_cpu(bsb.arraystart2) +
2759 __le64_to_cpu(bsb.length2));
2760 } else {
2761 wait_point = 0;
2762 if (part == 0 && __le64_to_cpu(bsb.length) > 0)
2763 wait_point = __le64_to_cpu(bsb.arraystart);
2764 if (part == 1 && __le64_to_cpu(bsb.length2) > 0)
2765 wait_point = __le64_to_cpu(bsb.arraystart2);
2766 }
2767
2768 rv = progress_reshape(sra, reshape,
2769 backup_point, wait_point,
2770 &suspend_point, &reshape_completed);
2771 if (rv < 0) {
2772 done = 1;
2773 break;
2774 }
2775
2776 /* external metadata would need to ping_monitor here */
2777 sra->reshape_progress = reshape_completed;
2778
2779 /* Clear any backup region that is before 'here' */
2780 if (increasing) {
2781 if (reshape_completed >= (__le64_to_cpu(bsb.arraystart) +
2782 __le64_to_cpu(bsb.length)))
2783 forget_backup(dests, destfd,
2784 destoffsets, 0);
2785 if (reshape_completed >= (__le64_to_cpu(bsb.arraystart2) +
2786 __le64_to_cpu(bsb.length2)))
2787 forget_backup(dests, destfd,
2788 destoffsets, 1);
2789 } else {
2790 if (reshape_completed <= (__le64_to_cpu(bsb.arraystart)))
2791 forget_backup(dests, destfd,
2792 destoffsets, 0);
2793 if (reshape_completed <= (__le64_to_cpu(bsb.arraystart2)))
2794 forget_backup(dests, destfd,
2795 destoffsets, 1);
2796 }
2797
2798 if (rv) {
2799 unsigned long long offset;
2800 /* need to backup some space... */
2801 /* Check that 'part' is unused */
2802 if (part == 0 && __le64_to_cpu(bsb.length) != 0)
2803 abort(); /* BUG here */
2804 if (part == 1 && __le64_to_cpu(bsb.length2) != 0)
2805 abort();
2806
2807 offset = backup_point / data;
2808 if (!increasing)
2809 offset -= stripes * (chunk/512);
2810 grow_backup(sra, offset, stripes,
2811 fds, offsets,
2812 disks, chunk, level, layout,
2813 dests, destfd, destoffsets,
2814 part, &degraded, buf);
2815 validate(afd, destfd[0], destoffsets[0]);
2816 /* record where 'part' is up to */
2817 part = !part;
2818 if (increasing)
2819 backup_point += stripes * (chunk/512) * data;
2820 else
2821 backup_point -= stripes * (chunk/512) * data;
2822 }
2823 }
2824
2825 if (reshape->before.data_disks == reshape->after.data_disks)
2826 sysfs_set_num(sra, NULL, "sync_speed_min", speed);
2827 free(buf);
2828 return 1; /* FIXME what does this mean? */
2829 }
2830
2831 /*
2832 * If any spare contains md_back_data-1 which is recent wrt mtime,
2833 * write that data into the array and update the super blocks with
2834 * the new reshape_progress
2835 */
2836 int Grow_restart(struct supertype *st, struct mdinfo *info, int *fdlist, int cnt,
2837 char *backup_file, int verbose)
2838 {
2839 int i, j;
2840 int old_disks;
2841 unsigned long long *offsets;
2842 unsigned long long nstripe, ostripe;
2843 int ndata, odata;
2844
2845 if (info->new_level != info->array.level)
2846 return 1; /* Cannot handle level changes (they are instantaneous) */
2847
2848 odata = info->array.raid_disks - info->delta_disks - 1;
2849 if (info->array.level == 6) odata--; /* number of data disks */
2850 ndata = info->array.raid_disks - 1;
2851 if (info->new_level == 6) ndata--;
2852
2853 old_disks = info->array.raid_disks - info->delta_disks;
2854
2855 if (info->delta_disks <= 0)
2856 /* Didn't grow, so the backup file must have
2857 * been used
2858 */
2859 old_disks = cnt;
2860 for (i=old_disks-(backup_file?1:0); i<cnt; i++) {
2861 struct mdinfo dinfo;
2862 int fd;
2863 int bsbsize;
2864 char *devname, namebuf[20];
2865
2866 /* This was a spare and may have some saved data on it.
2867 * Load the superblock, find and load the
2868 * backup_super_block.
2869 * If either fail, go on to next device.
2870 * If the backup contains no new info, just return
2871 * else restore data and update all superblocks
2872 */
2873 if (i == old_disks-1) {
2874 fd = open(backup_file, O_RDONLY);
2875 if (fd<0) {
2876 fprintf(stderr, Name ": backup file %s inaccessible: %s\n",
2877 backup_file, strerror(errno));
2878 continue;
2879 }
2880 devname = backup_file;
2881 } else {
2882 fd = fdlist[i];
2883 if (fd < 0)
2884 continue;
2885 if (st->ss->load_super(st, fd, NULL))
2886 continue;
2887
2888 st->ss->getinfo_super(st, &dinfo, NULL);
2889 st->ss->free_super(st);
2890
2891 if (lseek64(fd,
2892 (dinfo.data_offset + dinfo.component_size - 8) <<9,
2893 0) < 0) {
2894 fprintf(stderr, Name ": Cannot seek on device %d\n", i);
2895 continue; /* Cannot seek */
2896 }
2897 sprintf(namebuf, "device-%d", i);
2898 devname = namebuf;
2899 }
2900 if (read(fd, &bsb, sizeof(bsb)) != sizeof(bsb)) {
2901 if (verbose)
2902 fprintf(stderr, Name ": Cannot read from %s\n", devname);
2903 continue; /* Cannot read */
2904 }
2905 if (memcmp(bsb.magic, "md_backup_data-1", 16) != 0 &&
2906 memcmp(bsb.magic, "md_backup_data-2", 16) != 0) {
2907 if (verbose)
2908 fprintf(stderr, Name ": No backup metadata on %s\n", devname);
2909 continue;
2910 }
2911 if (bsb.sb_csum != bsb_csum((char*)&bsb, ((char*)&bsb.sb_csum)-((char*)&bsb))) {
2912 if (verbose)
2913 fprintf(stderr, Name ": Bad backup-metadata checksum on %s\n", devname);
2914 continue; /* bad checksum */
2915 }
2916 if (memcmp(bsb.magic, "md_backup_data-2", 16) == 0 &&
2917 bsb.sb_csum2 != bsb_csum((char*)&bsb, ((char*)&bsb.sb_csum2)-((char*)&bsb))) {
2918 if (verbose)
2919 fprintf(stderr, Name ": Bad backup-metadata checksum2 on %s\n", devname);
2920 continue; /* Bad second checksum */
2921 }
2922 if (memcmp(bsb.set_uuid,info->uuid, 16) != 0) {
2923 if (verbose)
2924 fprintf(stderr, Name ": Wrong uuid on backup-metadata on %s\n", devname);
2925 continue; /* Wrong uuid */
2926 }
2927
2928 /* array utime and backup-mtime should be updated at much the same time, but it seems that
2929 * sometimes they aren't... So allow considerable flexability in matching, and allow
2930 * this test to be overridden by an environment variable.
2931 */
2932 if (info->array.utime > (int)__le64_to_cpu(bsb.mtime) + 2*60*60 ||
2933 info->array.utime < (int)__le64_to_cpu(bsb.mtime) - 10*60) {
2934 if (check_env("MDADM_GROW_ALLOW_OLD")) {
2935 fprintf(stderr, Name ": accepting backup with timestamp %lu "
2936 "for array with timestamp %lu\n",
2937 (unsigned long)__le64_to_cpu(bsb.mtime),
2938 (unsigned long)info->array.utime);
2939 } else {
2940 if (verbose)
2941 fprintf(stderr, Name ": too-old timestamp on "
2942 "backup-metadata on %s\n", devname);
2943 continue; /* time stamp is too bad */
2944 }
2945 }
2946
2947 if (bsb.magic[15] == '1') {
2948 if (info->delta_disks >= 0) {
2949 /* reshape_progress is increasing */
2950 if (__le64_to_cpu(bsb.arraystart) + __le64_to_cpu(bsb.length) <
2951 info->reshape_progress) {
2952 nonew:
2953 if (verbose)
2954 fprintf(stderr, Name ": backup-metadata found on %s but is not needed\n", devname);
2955 continue; /* No new data here */
2956 }
2957 } else {
2958 /* reshape_progress is decreasing */
2959 if (__le64_to_cpu(bsb.arraystart) >=
2960 info->reshape_progress)
2961 goto nonew; /* No new data here */
2962 }
2963 } else {
2964 if (info->delta_disks >= 0) {
2965 /* reshape_progress is increasing */
2966 if (__le64_to_cpu(bsb.arraystart) + __le64_to_cpu(bsb.length) <
2967 info->reshape_progress &&
2968 __le64_to_cpu(bsb.arraystart2) + __le64_to_cpu(bsb.length2) <
2969 info->reshape_progress)
2970 goto nonew; /* No new data here */
2971 } else {
2972 /* reshape_progress is decreasing */
2973 if (__le64_to_cpu(bsb.arraystart) >=
2974 info->reshape_progress &&
2975 __le64_to_cpu(bsb.arraystart2) >=
2976 info->reshape_progress)
2977 goto nonew; /* No new data here */
2978 }
2979 }
2980 if (lseek64(fd, __le64_to_cpu(bsb.devstart)*512, 0)< 0) {
2981 second_fail:
2982 if (verbose)
2983 fprintf(stderr, Name ": Failed to verify secondary backup-metadata block on %s\n",
2984 devname);
2985 continue; /* Cannot seek */
2986 }
2987 /* There should be a duplicate backup superblock 4k before here */
2988 if (lseek64(fd, -4096, 1) < 0 ||
2989 read(fd, &bsb2, sizeof(bsb2)) != sizeof(bsb2))
2990 goto second_fail; /* Cannot find leading superblock */
2991 if (bsb.magic[15] == '1')
2992 bsbsize = offsetof(struct mdp_backup_super, pad1);
2993 else
2994 bsbsize = offsetof(struct mdp_backup_super, pad);
2995 if (memcmp(&bsb2, &bsb, bsbsize) != 0)
2996 goto second_fail; /* Cannot find leading superblock */
2997
2998 /* Now need the data offsets for all devices. */
2999 offsets = malloc(sizeof(*offsets)*info->array.raid_disks);
3000 for(j=0; j<info->array.raid_disks; j++) {
3001 if (fdlist[j] < 0)
3002 continue;
3003 if (st->ss->load_super(st, fdlist[j], NULL))
3004 /* FIXME should be this be an error */
3005 continue;
3006 st->ss->getinfo_super(st, &dinfo, NULL);
3007 st->ss->free_super(st);
3008 offsets[j] = dinfo.data_offset * 512;
3009 }
3010 printf(Name ": restoring critical section\n");
3011
3012 if (restore_stripes(fdlist, offsets,
3013 info->array.raid_disks,
3014 info->new_chunk,
3015 info->new_level,
3016 info->new_layout,
3017 fd, __le64_to_cpu(bsb.devstart)*512,
3018 __le64_to_cpu(bsb.arraystart)*512,
3019 __le64_to_cpu(bsb.length)*512)) {
3020 /* didn't succeed, so giveup */
3021 if (verbose)
3022 fprintf(stderr, Name ": Error restoring backup from %s\n",
3023 devname);
3024 return 1;
3025 }
3026
3027 if (bsb.magic[15] == '2' &&
3028 restore_stripes(fdlist, offsets,
3029 info->array.raid_disks,
3030 info->new_chunk,
3031 info->new_level,
3032 info->new_layout,
3033 fd, __le64_to_cpu(bsb.devstart)*512 +
3034 __le64_to_cpu(bsb.devstart2)*512,
3035 __le64_to_cpu(bsb.arraystart2)*512,
3036 __le64_to_cpu(bsb.length2)*512)) {
3037 /* didn't succeed, so giveup */
3038 if (verbose)
3039 fprintf(stderr, Name ": Error restoring second backup from %s\n",
3040 devname);
3041 return 1;
3042 }
3043
3044
3045 /* Ok, so the data is restored. Let's update those superblocks. */
3046
3047 if (info->delta_disks >= 0) {
3048 info->reshape_progress = __le64_to_cpu(bsb.arraystart) +
3049 __le64_to_cpu(bsb.length);
3050 if (bsb.magic[15] == '2') {
3051 unsigned long long p2 = __le64_to_cpu(bsb.arraystart2) +
3052 __le64_to_cpu(bsb.length2);
3053 if (p2 > info->reshape_progress)
3054 info->reshape_progress = p2;
3055 }
3056 } else {
3057 info->reshape_progress = __le64_to_cpu(bsb.arraystart);
3058 if (bsb.magic[15] == '2') {
3059 unsigned long long p2 = __le64_to_cpu(bsb.arraystart2);
3060 if (p2 < info->reshape_progress)
3061 info->reshape_progress = p2;
3062 }
3063 }
3064 for (j=0; j<info->array.raid_disks; j++) {
3065 if (fdlist[j] < 0) continue;
3066 if (st->ss->load_super(st, fdlist[j], NULL))
3067 continue;
3068 st->ss->getinfo_super(st, &dinfo, NULL);
3069 dinfo.reshape_progress = info->reshape_progress;
3070 st->ss->update_super(st, &dinfo,
3071 "_reshape_progress",
3072 NULL,0, 0, NULL);
3073 st->ss->store_super(st, fdlist[j]);
3074 st->ss->free_super(st);
3075 }
3076 return 0;
3077 }
3078 /* Didn't find any backup data, try to see if any
3079 * was needed.
3080 */
3081 if (info->delta_disks < 0) {
3082 /* When shrinking, the critical section is at the end.
3083 * So see if we are before the critical section.
3084 */
3085 unsigned long long first_block;
3086 nstripe = ostripe = 0;
3087 first_block = 0;
3088 while (ostripe >= nstripe) {
3089 ostripe += info->array.chunk_size / 512;
3090 first_block = ostripe * odata;
3091 nstripe = first_block / ndata / (info->new_chunk/512) *
3092 (info->new_chunk/512);
3093 }
3094
3095 if (info->reshape_progress >= first_block)
3096 return 0;
3097 }
3098 if (info->delta_disks > 0) {
3099 /* See if we are beyond the critical section. */
3100 unsigned long long last_block;
3101 nstripe = ostripe = 0;
3102 last_block = 0;
3103 while (nstripe >= ostripe) {
3104 nstripe += info->new_chunk / 512;
3105 last_block = nstripe * ndata;
3106 ostripe = last_block / odata / (info->array.chunk_size/512) *
3107 (info->array.chunk_size/512);
3108 }
3109
3110 if (info->reshape_progress >= last_block)
3111 return 0;
3112 }
3113 /* needed to recover critical section! */
3114 if (verbose)
3115 fprintf(stderr, Name ": Failed to find backup of critical section\n");
3116 return 1;
3117 }
3118
3119 int Grow_continue(int mdfd, struct supertype *st, struct mdinfo *info,
3120 char *backup_file)
3121 {
3122 int err = sysfs_set_str(info, NULL, "array_state", "readonly");
3123 if (err)
3124 return err;
3125 return reshape_array(NULL, mdfd, "array", st, info, 1, backup_file, 0, 0);
3126 }
3127
3128