]> git.ipfire.org Git - thirdparty/mdadm.git/blob - Grow.c
projects / thirdparty / mdadm.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Grow: report better message when --grow --chunk cannot work.
[thirdparty/mdadm.git] / Grow.c
1 /*
2 * mdadm - manage Linux "md" devices aka RAID arrays.
3 *
4 * Copyright (C) 2001-2012 Neil Brown <neilb@suse.de>
5 *
6 *
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License as published by
9 * the Free Software Foundation; either version 2 of the License, or
10 * (at your option) any later version.
11 *
12 * This program is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 * GNU General Public License for more details.
16 *
17 * You should have received a copy of the GNU General Public License
18 * along with this program; if not, write to the Free Software
19 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
20 *
21 * Author: Neil Brown
22 * Email: <neilb@suse.de>
23 */
24 #include "mdadm.h"
25 #include "dlink.h"
26 #include <sys/mman.h>
27 #include <stdint.h>
28
29 #if ! defined(__BIG_ENDIAN) && ! defined(__LITTLE_ENDIAN)
30 #error no endian defined
31 #endif
32 #include "md_u.h"
33 #include "md_p.h"
34
35 #ifndef offsetof
36 #define offsetof(t,f) ((size_t)&(((t*)0)->f))
37 #endif
38
39 int restore_backup(struct supertype *st,
40 struct mdinfo *content,
41 int working_disks,
42 int next_spare,
43 char *backup_file,
44 int verbose)
45 {
46 int i;
47 int *fdlist;
48 struct mdinfo *dev;
49 int err;
50 int disk_count = next_spare + working_disks;
51
52 dprintf("Called restore_backup()\n");
53 fdlist = xmalloc(sizeof(int) * disk_count);
54
55 enable_fds(next_spare);
56 for (i = 0; i < next_spare; i++)
57 fdlist[i] = -1;
58 for (dev = content->devs; dev; dev = dev->next) {
59 char buf[22];
60 int fd;
61 sprintf(buf, "%d:%d",
62 dev->disk.major,
63 dev->disk.minor);
64 fd = dev_open(buf, O_RDWR);
65
66 if (dev->disk.raid_disk >= 0)
67 fdlist[dev->disk.raid_disk] = fd;
68 else
69 fdlist[next_spare++] = fd;
70 }
71
72 if (st->ss->external && st->ss->recover_backup)
73 err = st->ss->recover_backup(st, content);
74 else
75 err = Grow_restart(st, content, fdlist, next_spare,
76 backup_file, verbose > 0);
77
78 while (next_spare > 0) {
79 next_spare--;
80 if (fdlist[next_spare] >= 0)
81 close(fdlist[next_spare]);
82 }
83 free(fdlist);
84 if (err) {
85 pr_err("Failed to restore critical"
86 " section for reshape - sorry.\n");
87 if (!backup_file)
88 pr_err("Possibly you need"
89 " to specify a --backup-file\n");
90 return 1;
91 }
92
93 dprintf("restore_backup() returns status OK.\n");
94 return 0;
95 }
96
97 int Grow_Add_device(char *devname, int fd, char *newdev)
98 {
99 /* Add a device to an active array.
100 * Currently, just extend a linear array.
101 * This requires writing a new superblock on the
102 * new device, calling the kernel to add the device,
103 * and if that succeeds, update the superblock on
104 * all other devices.
105 * This means that we need to *find* all other devices.
106 */
107 struct mdinfo info;
108
109 struct stat stb;
110 int nfd, fd2;
111 int d, nd;
112 struct supertype *st = NULL;
113 char *subarray = NULL;
114
115 if (ioctl(fd, GET_ARRAY_INFO, &info.array) < 0) {
116 pr_err("cannot get array info for %s\n", devname);
117 return 1;
118 }
119
120 if (info.array.level != -1) {
121 pr_err("can only add devices to linear arrays\n");
122 return 1;
123 }
124
125 st = super_by_fd(fd, &subarray);
126 if (!st) {
127 pr_err("cannot handle arrays with superblock version %d\n",
128 info.array.major_version);
129 return 1;
130 }
131
132 if (subarray) {
133 pr_err("Cannot grow linear sub-arrays yet\n");
134 free(subarray);
135 free(st);
136 return 1;
137 }
138
139 nfd = open(newdev, O_RDWR|O_EXCL|O_DIRECT);
140 if (nfd < 0) {
141 pr_err("cannot open %s\n", newdev);
142 free(st);
143 return 1;
144 }
145 fstat(nfd, &stb);
146 if ((stb.st_mode & S_IFMT) != S_IFBLK) {
147 pr_err("%s is not a block device!\n", newdev);
148 close(nfd);
149 free(st);
150 return 1;
151 }
152 /* now check out all the devices and make sure we can read the
153 * superblock */
154 for (d=0 ; d < info.array.raid_disks ; d++) {
155 mdu_disk_info_t disk;
156 char *dv;
157
158 st->ss->free_super(st);
159
160 disk.number = d;
161 if (ioctl(fd, GET_DISK_INFO, &disk) < 0) {
162 pr_err("cannot get device detail for device %d\n",
163 d);
164 close(nfd);
165 free(st);
166 return 1;
167 }
168 dv = map_dev(disk.major, disk.minor, 1);
169 if (!dv) {
170 pr_err("cannot find device file for device %d\n",
171 d);
172 close(nfd);
173 free(st);
174 return 1;
175 }
176 fd2 = dev_open(dv, O_RDWR);
177 if (fd2 < 0) {
178 pr_err("cannot open device file %s\n", dv);
179 close(nfd);
180 free(st);
181 return 1;
182 }
183
184 if (st->ss->load_super(st, fd2, NULL)) {
185 pr_err("cannot find super block on %s\n", dv);
186 close(nfd);
187 close(fd2);
188 free(st);
189 return 1;
190 }
191 close(fd2);
192 }
193 /* Ok, looks good. Lets update the superblock and write it out to
194 * newdev.
195 */
196
197 info.disk.number = d;
198 info.disk.major = major(stb.st_rdev);
199 info.disk.minor = minor(stb.st_rdev);
200 info.disk.raid_disk = d;
201 info.disk.state = (1 << MD_DISK_SYNC) | (1 << MD_DISK_ACTIVE);
202 st->ss->update_super(st, &info, "linear-grow-new", newdev,
203 0, 0, NULL);
204
205 if (st->ss->store_super(st, nfd)) {
206 pr_err("Cannot store new superblock on %s\n",
207 newdev);
208 close(nfd);
209 return 1;
210 }
211 close(nfd);
212
213 if (ioctl(fd, ADD_NEW_DISK, &info.disk) != 0) {
214 pr_err("Cannot add new disk to this array\n");
215 return 1;
216 }
217 /* Well, that seems to have worked.
218 * Now go through and update all superblocks
219 */
220
221 if (ioctl(fd, GET_ARRAY_INFO, &info.array) < 0) {
222 pr_err("cannot get array info for %s\n", devname);
223 return 1;
224 }
225
226 nd = d;
227 for (d=0 ; d < info.array.raid_disks ; d++) {
228 mdu_disk_info_t disk;
229 char *dv;
230
231 disk.number = d;
232 if (ioctl(fd, GET_DISK_INFO, &disk) < 0) {
233 pr_err("cannot get device detail for device %d\n",
234 d);
235 return 1;
236 }
237 dv = map_dev(disk.major, disk.minor, 1);
238 if (!dv) {
239 pr_err("cannot find device file for device %d\n",
240 d);
241 return 1;
242 }
243 fd2 = dev_open(dv, O_RDWR);
244 if (fd2 < 0) {
245 pr_err("cannot open device file %s\n", dv);
246 return 1;
247 }
248 if (st->ss->load_super(st, fd2, NULL)) {
249 pr_err("cannot find super block on %s\n", dv);
250 close(fd);
251 return 1;
252 }
253 info.array.raid_disks = nd+1;
254 info.array.nr_disks = nd+1;
255 info.array.active_disks = nd+1;
256 info.array.working_disks = nd+1;
257
258 st->ss->update_super(st, &info, "linear-grow-update", dv,
259 0, 0, NULL);
260
261 if (st->ss->store_super(st, fd2)) {
262 pr_err("Cannot store new superblock on %s\n", dv);
263 close(fd2);
264 return 1;
265 }
266 close(fd2);
267 }
268
269 return 0;
270 }
271
272 int Grow_addbitmap(char *devname, int fd, struct context *c, struct shape *s)
273 {
274 /*
275 * First check that array doesn't have a bitmap
276 * Then create the bitmap
277 * Then add it
278 *
279 * For internal bitmaps, we need to check the version,
280 * find all the active devices, and write the bitmap block
281 * to all devices
282 */
283 mdu_bitmap_file_t bmf;
284 mdu_array_info_t array;
285 struct supertype *st;
286 char *subarray = NULL;
287 int major = BITMAP_MAJOR_HI;
288 int vers = md_get_version(fd);
289 unsigned long long bitmapsize, array_size;
290
291 if (vers < 9003) {
292 major = BITMAP_MAJOR_HOSTENDIAN;
293 pr_err("Warning - bitmaps created on this kernel"
294 " are not portable\n"
295 " between different architectures. Consider upgrading"
296 " the Linux kernel.\n");
297 }
298
299 if (ioctl(fd, GET_BITMAP_FILE, &bmf) != 0) {
300 if (errno == ENOMEM)
301 pr_err("Memory allocation failure.\n");
302 else
303 pr_err("bitmaps not supported by this kernel.\n");
304 return 1;
305 }
306 if (bmf.pathname[0]) {
307 if (strcmp(s->bitmap_file,"none")==0) {
308 if (ioctl(fd, SET_BITMAP_FILE, -1)!= 0) {
309 pr_err("failed to remove bitmap %s\n",
310 bmf.pathname);
311 return 1;
312 }
313 return 0;
314 }
315 pr_err("%s already has a bitmap (%s)\n",
316 devname, bmf.pathname);
317 return 1;
318 }
319 if (ioctl(fd, GET_ARRAY_INFO, &array) != 0) {
320 pr_err("cannot get array status for %s\n", devname);
321 return 1;
322 }
323 if (array.state & (1<<MD_SB_BITMAP_PRESENT)) {
324 if (strcmp(s->bitmap_file, "none")==0) {
325 array.state &= ~(1<<MD_SB_BITMAP_PRESENT);
326 if (ioctl(fd, SET_ARRAY_INFO, &array)!= 0) {
327 pr_err("failed to remove internal bitmap.\n");
328 return 1;
329 }
330 return 0;
331 }
332 pr_err("Internal bitmap already present on %s\n",
333 devname);
334 return 1;
335 }
336
337 if (strcmp(s->bitmap_file, "none") == 0) {
338 pr_err("no bitmap found on %s\n", devname);
339 return 1;
340 }
341 if (array.level <= 0) {
342 pr_err("Bitmaps not meaningful with level %s\n",
343 map_num(pers, array.level)?:"of this array");
344 return 1;
345 }
346 bitmapsize = array.size;
347 bitmapsize <<= 1;
348 if (get_dev_size(fd, NULL, &array_size) &&
349 array_size > (0x7fffffffULL<<9)) {
350 /* Array is big enough that we cannot trust array.size
351 * try other approaches
352 */
353 bitmapsize = get_component_size(fd);
354 }
355 if (bitmapsize == 0) {
356 pr_err("Cannot reliably determine size of array to create bitmap - sorry.\n");
357 return 1;
358 }
359
360 if (array.level == 10) {
361 int ncopies = (array.layout&255)*((array.layout>>8)&255);
362 bitmapsize = bitmapsize * array.raid_disks / ncopies;
363 }
364
365 st = super_by_fd(fd, &subarray);
366 if (!st) {
367 pr_err("Cannot understand version %d.%d\n",
368 array.major_version, array.minor_version);
369 return 1;
370 }
371 if (subarray) {
372 pr_err("Cannot add bitmaps to sub-arrays yet\n");
373 free(subarray);
374 free(st);
375 return 1;
376 }
377 if (strcmp(s->bitmap_file, "internal") == 0) {
378 int rv;
379 int d;
380 int offset_setable = 0;
381 struct mdinfo *mdi;
382 if (st->ss->add_internal_bitmap == NULL) {
383 pr_err("Internal bitmaps not supported "
384 "with %s metadata\n", st->ss->name);
385 return 1;
386 }
387 mdi = sysfs_read(fd, NULL, GET_BITMAP_LOCATION);
388 if (mdi)
389 offset_setable = 1;
390 for (d=0; d< st->max_devs; d++) {
391 mdu_disk_info_t disk;
392 char *dv;
393 disk.number = d;
394 if (ioctl(fd, GET_DISK_INFO, &disk) < 0)
395 continue;
396 if (disk.major == 0 &&
397 disk.minor == 0)
398 continue;
399 if ((disk.state & (1<<MD_DISK_SYNC))==0)
400 continue;
401 dv = map_dev(disk.major, disk.minor, 1);
402 if (dv) {
403 int fd2 = dev_open(dv, O_RDWR);
404 if (fd2 < 0)
405 continue;
406 if (st->ss->load_super(st, fd2, NULL)==0) {
407 if (st->ss->add_internal_bitmap(
408 st,
409 &s->bitmap_chunk, c->delay, s->write_behind,
410 bitmapsize, offset_setable,
411 major)
412 )
413 st->ss->write_bitmap(st, fd2);
414 else {
415 pr_err("failed to create internal bitmap"
416 " - chunksize problem.\n");
417 close(fd2);
418 return 1;
419 }
420 }
421 close(fd2);
422 }
423 }
424 if (offset_setable) {
425 st->ss->getinfo_super(st, mdi, NULL);
426 sysfs_init(mdi, fd, NULL);
427 rv = sysfs_set_num_signed(mdi, NULL, "bitmap/location",
428 mdi->bitmap_offset);
429 } else {
430 array.state |= (1<<MD_SB_BITMAP_PRESENT);
431 rv = ioctl(fd, SET_ARRAY_INFO, &array);
432 }
433 if (rv < 0) {
434 if (errno == EBUSY)
435 pr_err("Cannot add bitmap while array is"
436 " resyncing or reshaping etc.\n");
437 pr_err("failed to set internal bitmap.\n");
438 return 1;
439 }
440 } else {
441 int uuid[4];
442 int bitmap_fd;
443 int d;
444 int max_devs = st->max_devs;
445
446 /* try to load a superblock */
447 for (d = 0; d < max_devs; d++) {
448 mdu_disk_info_t disk;
449 char *dv;
450 int fd2;
451 disk.number = d;
452 if (ioctl(fd, GET_DISK_INFO, &disk) < 0)
453 continue;
454 if ((disk.major==0 && disk.minor==0) ||
455 (disk.state & (1<<MD_DISK_REMOVED)))
456 continue;
457 dv = map_dev(disk.major, disk.minor, 1);
458 if (!dv)
459 continue;
460 fd2 = dev_open(dv, O_RDONLY);
461 if (fd2 >= 0) {
462 if (st->ss->load_super(st, fd2, NULL) == 0) {
463 close(fd2);
464 st->ss->uuid_from_super(st, uuid);
465 break;
466 }
467 close(fd2);
468 }
469 }
470 if (d == max_devs) {
471 pr_err("cannot find UUID for array!\n");
472 return 1;
473 }
474 if (CreateBitmap(s->bitmap_file, c->force, (char*)uuid, s->bitmap_chunk,
475 c->delay, s->write_behind, bitmapsize, major)) {
476 return 1;
477 }
478 bitmap_fd = open(s->bitmap_file, O_RDWR);
479 if (bitmap_fd < 0) {
480 pr_err("weird: %s cannot be opened\n",
481 s->bitmap_file);
482 return 1;
483 }
484 if (ioctl(fd, SET_BITMAP_FILE, bitmap_fd) < 0) {
485 int err = errno;
486 if (errno == EBUSY)
487 pr_err("Cannot add bitmap while array is"
488 " resyncing or reshaping etc.\n");
489 pr_err("Cannot set bitmap file for %s: %s\n",
490 devname, strerror(err));
491 return 1;
492 }
493 }
494
495 return 0;
496 }
497
498 /*
499 * When reshaping an array we might need to backup some data.
500 * This is written to all spares with a 'super_block' describing it.
501 * The superblock goes 4K from the end of the used space on the
502 * device.
503 * It if written after the backup is complete.
504 * It has the following structure.
505 */
506
507 static struct mdp_backup_super {
508 char magic[16]; /* md_backup_data-1 or -2 */
509 __u8 set_uuid[16];
510 __u64 mtime;
511 /* start/sizes in 512byte sectors */
512 __u64 devstart; /* address on backup device/file of data */
513 __u64 arraystart;
514 __u64 length;
515 __u32 sb_csum; /* csum of preceeding bytes. */
516 __u32 pad1;
517 __u64 devstart2; /* offset in to data of second section */
518 __u64 arraystart2;
519 __u64 length2;
520 __u32 sb_csum2; /* csum of preceeding bytes. */
521 __u8 pad[512-68-32];
522 } __attribute__((aligned(512))) bsb, bsb2;
523
524 static __u32 bsb_csum(char *buf, int len)
525 {
526 int i;
527 int csum = 0;
528 for (i = 0; i < len; i++)
529 csum = (csum<<3) + buf[0];
530 return __cpu_to_le32(csum);
531 }
532
533 static int check_idle(struct supertype *st)
534 {
535 /* Check that all member arrays for this container, or the
536 * container of this array, are idle
537 */
538 char *container = (st->container_devnm[0]
539 ? st->container_devnm : st->devnm);
540 struct mdstat_ent *ent, *e;
541 int is_idle = 1;
542
543 ent = mdstat_read(0, 0);
544 for (e = ent ; e; e = e->next) {
545 if (!is_container_member(e, container))
546 continue;
547 if (e->percent >= 0) {
548 is_idle = 0;
549 break;
550 }
551 }
552 free_mdstat(ent);
553 return is_idle;
554 }
555
556 static int freeze_container(struct supertype *st)
557 {
558 char *container = (st->container_devnm[0]
559 ? st->container_devnm : st->devnm);
560
561 if (!check_idle(st))
562 return -1;
563
564 if (block_monitor(container, 1)) {
565 pr_err("failed to freeze container\n");
566 return -2;
567 }
568
569 return 1;
570 }
571
572 static void unfreeze_container(struct supertype *st)
573 {
574 char *container = (st->container_devnm[0]
575 ? st->container_devnm : st->devnm);
576
577 unblock_monitor(container, 1);
578 }
579
580 static int freeze(struct supertype *st)
581 {
582 /* Try to freeze resync/rebuild on this array/container.
583 * Return -1 if the array is busy,
584 * return -2 container cannot be frozen,
585 * return 0 if this kernel doesn't support 'frozen'
586 * return 1 if it worked.
587 */
588 if (st->ss->external)
589 return freeze_container(st);
590 else {
591 struct mdinfo *sra = sysfs_read(-1, st->devnm, GET_VERSION);
592 int err;
593 char buf[20];
594
595 if (!sra)
596 return -1;
597 /* Need to clear any 'read-auto' status */
598 if (sysfs_get_str(sra, NULL, "array_state", buf, 20) > 0 &&
599 strncmp(buf, "read-auto", 9) == 0)
600 sysfs_set_str(sra, NULL, "array_state", "clean");
601
602 err = sysfs_freeze_array(sra);
603 sysfs_free(sra);
604 return err;
605 }
606 }
607
608 static void unfreeze(struct supertype *st)
609 {
610 if (st->ss->external)
611 return unfreeze_container(st);
612 else {
613 struct mdinfo *sra = sysfs_read(-1, st->devnm, GET_VERSION);
614
615 if (sra)
616 sysfs_set_str(sra, NULL, "sync_action", "idle");
617 sysfs_free(sra);
618 }
619 }
620
621 static void wait_reshape(struct mdinfo *sra)
622 {
623 int fd = sysfs_get_fd(sra, NULL, "sync_action");
624 char action[20];
625
626 if (fd < 0)
627 return;
628
629 while (sysfs_fd_get_str(fd, action, 20) > 0 &&
630 strncmp(action, "reshape", 7) == 0) {
631 fd_set rfds;
632 FD_ZERO(&rfds);
633 FD_SET(fd, &rfds);
634 select(fd+1, NULL, NULL, &rfds, NULL);
635 }
636 close(fd);
637 }
638
639 static int reshape_super(struct supertype *st, unsigned long long size,
640 int level, int layout, int chunksize, int raid_disks,
641 int delta_disks, char *backup_file, char *dev,
642 int direction, int verbose)
643 {
644 /* nothing extra to check in the native case */
645 if (!st->ss->external)
646 return 0;
647 if (!st->ss->reshape_super ||
648 !st->ss->manage_reshape) {
649 pr_err("%s metadata does not support reshape\n",
650 st->ss->name);
651 return 1;
652 }
653
654 return st->ss->reshape_super(st, size, level, layout, chunksize,
655 raid_disks, delta_disks, backup_file, dev,
656 direction, verbose);
657 }
658
659 static void sync_metadata(struct supertype *st)
660 {
661 if (st->ss->external) {
662 if (st->update_tail) {
663 flush_metadata_updates(st);
664 st->update_tail = &st->updates;
665 } else
666 st->ss->sync_metadata(st);
667 }
668 }
669
670 static int subarray_set_num(char *container, struct mdinfo *sra, char *name, int n)
671 {
672 /* when dealing with external metadata subarrays we need to be
673 * prepared to handle EAGAIN. The kernel may need to wait for
674 * mdmon to mark the array active so the kernel can handle
675 * allocations/writeback when preparing the reshape action
676 * (md_allow_write()). We temporarily disable safe_mode_delay
677 * to close a race with the array_state going clean before the
678 * next write to raid_disks / stripe_cache_size
679 */
680 char safe[50];
681 int rc;
682
683 /* only 'raid_disks' and 'stripe_cache_size' trigger md_allow_write */
684 if (!container ||
685 (strcmp(name, "raid_disks") != 0 &&
686 strcmp(name, "stripe_cache_size") != 0))
687 return sysfs_set_num(sra, NULL, name, n);
688
689 rc = sysfs_get_str(sra, NULL, "safe_mode_delay", safe, sizeof(safe));
690 if (rc <= 0)
691 return -1;
692 sysfs_set_num(sra, NULL, "safe_mode_delay", 0);
693 rc = sysfs_set_num(sra, NULL, name, n);
694 if (rc < 0 && errno == EAGAIN) {
695 ping_monitor(container);
696 /* if we get EAGAIN here then the monitor is not active
697 * so stop trying
698 */
699 rc = sysfs_set_num(sra, NULL, name, n);
700 }
701 sysfs_set_str(sra, NULL, "safe_mode_delay", safe);
702 return rc;
703 }
704
705 int start_reshape(struct mdinfo *sra, int already_running,
706 int before_data_disks, int data_disks)
707 {
708 int err;
709 unsigned long long sync_max_to_set;
710
711 sysfs_set_num(sra, NULL, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL);
712 err = sysfs_set_num(sra, NULL, "suspend_hi", sra->reshape_progress);
713 err = err ?: sysfs_set_num(sra, NULL, "suspend_lo",
714 sra->reshape_progress);
715 if (before_data_disks <= data_disks)
716 sync_max_to_set = sra->reshape_progress / data_disks;
717 else
718 sync_max_to_set = (sra->component_size * data_disks
719 - sra->reshape_progress) / data_disks;
720 if (!already_running)
721 sysfs_set_num(sra, NULL, "sync_min", sync_max_to_set);
722 err = err ?: sysfs_set_num(sra, NULL, "sync_max", sync_max_to_set);
723 if (!already_running)
724 err = err ?: sysfs_set_str(sra, NULL, "sync_action", "reshape");
725
726 return err;
727 }
728
729 void abort_reshape(struct mdinfo *sra)
730 {
731 sysfs_set_str(sra, NULL, "sync_action", "idle");
732 sysfs_set_num(sra, NULL, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL);
733 sysfs_set_num(sra, NULL, "suspend_hi", 0);
734 sysfs_set_num(sra, NULL, "suspend_lo", 0);
735 sysfs_set_num(sra, NULL, "sync_min", 0);
736 sysfs_set_str(sra, NULL, "sync_max", "max");
737 }
738
739 int remove_disks_for_takeover(struct supertype *st,
740 struct mdinfo *sra,
741 int layout)
742 {
743 int nr_of_copies;
744 struct mdinfo *remaining;
745 int slot;
746
747 if (sra->array.level == 10)
748 nr_of_copies = layout & 0xff;
749 else if (sra->array.level == 1)
750 nr_of_copies = sra->array.raid_disks;
751 else
752 return 1;
753
754 remaining = sra->devs;
755 sra->devs = NULL;
756 /* for each 'copy', select one device and remove from the list. */
757 for (slot = 0; slot < sra->array.raid_disks; slot += nr_of_copies) {
758 struct mdinfo **diskp;
759 int found = 0;
760
761 /* Find a working device to keep */
762 for (diskp = &remaining; *diskp ; diskp = &(*diskp)->next) {
763 struct mdinfo *disk = *diskp;
764
765 if (disk->disk.raid_disk < slot)
766 continue;
767 if (disk->disk.raid_disk >= slot + nr_of_copies)
768 continue;
769 if (disk->disk.state & (1<<MD_DISK_REMOVED))
770 continue;
771 if (disk->disk.state & (1<<MD_DISK_FAULTY))
772 continue;
773 if (!(disk->disk.state & (1<<MD_DISK_SYNC)))
774 continue;
775
776 /* We have found a good disk to use! */
777 *diskp = disk->next;
778 disk->next = sra->devs;
779 sra->devs = disk;
780 found = 1;
781 break;
782 }
783 if (!found)
784 break;
785 }
786
787 if (slot < sra->array.raid_disks) {
788 /* didn't find all slots */
789 struct mdinfo **e;
790 e = &remaining;
791 while (*e)
792 e = &(*e)->next;
793 *e = sra->devs;
794 sra->devs = remaining;
795 return 1;
796 }
797
798 /* Remove all 'remaining' devices from the array */
799 while (remaining) {
800 struct mdinfo *sd = remaining;
801 remaining = sd->next;
802
803 sysfs_set_str(sra, sd, "state", "faulty");
804 sysfs_set_str(sra, sd, "slot", "none");
805 /* for external metadata disks should be removed in mdmon */
806 if (!st->ss->external)
807 sysfs_set_str(sra, sd, "state", "remove");
808 sd->disk.state |= (1<<MD_DISK_REMOVED);
809 sd->disk.state &= ~(1<<MD_DISK_SYNC);
810 sd->next = sra->devs;
811 sra->devs = sd;
812 }
813 return 0;
814 }
815
816 void reshape_free_fdlist(int *fdlist,
817 unsigned long long *offsets,
818 int size)
819 {
820 int i;
821
822 for (i = 0; i < size; i++)
823 if (fdlist[i] >= 0)
824 close(fdlist[i]);
825
826 free(fdlist);
827 free(offsets);
828 }
829
830 int reshape_prepare_fdlist(char *devname,
831 struct mdinfo *sra,
832 int raid_disks,
833 int nrdisks,
834 unsigned long blocks,
835 char *backup_file,
836 int *fdlist,
837 unsigned long long *offsets)
838 {
839 int d = 0;
840 struct mdinfo *sd;
841
842 enable_fds(nrdisks);
843 for (d = 0; d <= nrdisks; d++)
844 fdlist[d] = -1;
845 d = raid_disks;
846 for (sd = sra->devs; sd; sd = sd->next) {
847 if (sd->disk.state & (1<<MD_DISK_FAULTY))
848 continue;
849 if (sd->disk.state & (1<<MD_DISK_SYNC)) {
850 char *dn = map_dev(sd->disk.major,
851 sd->disk.minor, 1);
852 fdlist[sd->disk.raid_disk]
853 = dev_open(dn, O_RDONLY);
854 offsets[sd->disk.raid_disk] = sd->data_offset*512;
855 if (fdlist[sd->disk.raid_disk] < 0) {
856 pr_err("%s: cannot open component %s\n",
857 devname, dn ? dn : "-unknown-");
858 d = -1;
859 goto release;
860 }
861 } else if (backup_file == NULL) {
862 /* spare */
863 char *dn = map_dev(sd->disk.major,
864 sd->disk.minor, 1);
865 fdlist[d] = dev_open(dn, O_RDWR);
866 offsets[d] = (sd->data_offset + sra->component_size - blocks - 8)*512;
867 if (fdlist[d] < 0) {
868 pr_err("%s: cannot open component %s\n",
869 devname, dn ? dn : "-unknown-");
870 d = -1;
871 goto release;
872 }
873 d++;
874 }
875 }
876 release:
877 return d;
878 }
879
880 int reshape_open_backup_file(char *backup_file,
881 int fd,
882 char *devname,
883 long blocks,
884 int *fdlist,
885 unsigned long long *offsets,
886 int restart)
887 {
888 /* Return 1 on success, 0 on any form of failure */
889 /* need to check backup file is large enough */
890 char buf[512];
891 struct stat stb;
892 unsigned int dev;
893 int i;
894
895 *fdlist = open(backup_file, O_RDWR|O_CREAT|(restart ? O_TRUNC : O_EXCL),
896 S_IRUSR | S_IWUSR);
897 *offsets = 8 * 512;
898 if (*fdlist < 0) {
899 pr_err("%s: cannot create backup file %s: %s\n",
900 devname, backup_file, strerror(errno));
901 return 0;
902 }
903 /* Guard against backup file being on array device.
904 * If array is partitioned or if LVM etc is in the
905 * way this will not notice, but it is better than
906 * nothing.
907 */
908 fstat(*fdlist, &stb);
909 dev = stb.st_dev;
910 fstat(fd, &stb);
911 if (stb.st_rdev == dev) {
912 pr_err("backup file must NOT be"
913 " on the array being reshaped.\n");
914 close(*fdlist);
915 return 0;
916 }
917
918 memset(buf, 0, 512);
919 for (i=0; i < blocks + 8 ; i++) {
920 if (write(*fdlist, buf, 512) != 512) {
921 pr_err("%s: cannot create"
922 " backup file %s: %s\n",
923 devname, backup_file, strerror(errno));
924 return 0;
925 }
926 }
927 if (fsync(*fdlist) != 0) {
928 pr_err("%s: cannot create backup file %s: %s\n",
929 devname, backup_file, strerror(errno));
930 return 0;
931 }
932
933 return 1;
934 }
935
936 unsigned long GCD(unsigned long a, unsigned long b)
937 {
938 while (a != b) {
939 if (a < b)
940 b -= a;
941 if (b < a)
942 a -= b;
943 }
944 return a;
945 }
946
947 unsigned long compute_backup_blocks(int nchunk, int ochunk,
948 unsigned int ndata, unsigned int odata)
949 {
950 unsigned long a, b, blocks;
951 /* So how much do we need to backup.
952 * We need an amount of data which is both a whole number of
953 * old stripes and a whole number of new stripes.
954 * So LCM for (chunksize*datadisks).
955 */
956 a = (ochunk/512) * odata;
957 b = (nchunk/512) * ndata;
958 /* Find GCD */
959 a = GCD(a, b);
960 /* LCM == product / GCD */
961 blocks = (ochunk/512) * (nchunk/512) * odata * ndata / a;
962
963 return blocks;
964 }
965
966 char *analyse_change(char *devname, struct mdinfo *info, struct reshape *re)
967 {
968 /* Based on the current array state in info->array and
969 * the changes in info->new_* etc, determine:
970 * - whether the change is possible
971 * - Intermediate level/raid_disks/layout
972 * - whether a restriping reshape is needed
973 * - number of sectors in minimum change unit. This
974 * will cover a whole number of stripes in 'before' and
975 * 'after'.
976 *
977 * Return message if the change should be rejected
978 * NULL if the change can be achieved
979 *
980 * This can be called as part of starting a reshape, or
981 * when assembling an array that is undergoing reshape.
982 */
983 int near, far, offset, copies;
984 int new_disks;
985 int old_chunk, new_chunk;
986 /* delta_parity records change in number of devices
987 * caused by level change
988 */
989 int delta_parity = 0;
990
991 memset(re, 0, sizeof(*re));
992
993 /* If a new level not explicitly given, we assume no-change */
994 if (info->new_level == UnSet)
995 info->new_level = info->array.level;
996
997 if (info->new_chunk)
998 switch (info->new_level) {
999 case 0:
1000 case 4:
1001 case 5:
1002 case 6:
1003 case 10:
1004 /* chunk size is meaningful, must divide component_size
1005 * evenly
1006 */
1007 if (info->component_size % (info->new_chunk/512)) {
1008 unsigned long long shrink = info->component_size;
1009 shrink &= ~(unsigned long long)(info->new_chunk/512-1);
1010 pr_err("New chunk size (%dK) does not evenly divide device size (%lluk)\n",
1011 info->new_chunk/1024, info->component_size/2);
1012 pr_err("After shrinking any filesystem, \"mdadm --grow %s --size %llu\"\n",
1013 devname, shrink/2);
1014 pr_err("will shrink the array so the given chunk size would work.\n");
1015 return "";
1016 }
1017 break;
1018 default:
1019 return "chunk size not meaningful for this level";
1020 }
1021 else
1022 info->new_chunk = info->array.chunk_size;
1023
1024 switch (info->array.level) {
1025 default:
1026 return "Cannot understand this RAID level";
1027 case 1:
1028 /* RAID1 can convert to RAID1 with different disks, or
1029 * raid5 with 2 disks, or
1030 * raid0 with 1 disk
1031 */
1032 if (info->new_level > 1 &&
1033 (info->component_size & 7))
1034 return "Cannot convert RAID1 of this size - "
1035 "reduce size to multiple of 4K first.";
1036 if (info->new_level == 0) {
1037 if (info->delta_disks != UnSet &&
1038 info->delta_disks != 0)
1039 return "Cannot change number of disks "
1040 "with RAID1->RAID0 conversion";
1041 re->level = 0;
1042 re->before.data_disks = 1;
1043 re->after.data_disks = 1;
1044 return NULL;
1045 }
1046 if (info->new_level == 1) {
1047 if (info->delta_disks == UnSet)
1048 /* Don't know what to do */
1049 return "no change requested for Growing RAID1";
1050 re->level = 1;
1051 return NULL;
1052 }
1053 if (info->array.raid_disks == 2 &&
1054 info->new_level == 5) {
1055
1056 re->level = 5;
1057 re->before.data_disks = 1;
1058 if (info->delta_disks != UnSet &&
1059 info->delta_disks != 0)
1060 re->after.data_disks = 1 + info->delta_disks;
1061 else
1062 re->after.data_disks = 1;
1063 if (re->after.data_disks < 1)
1064 return "Number of disks too small for RAID5";
1065
1066 re->before.layout = ALGORITHM_LEFT_SYMMETRIC;
1067 info->array.chunk_size = 65536;
1068 break;
1069 }
1070 /* Could do some multi-stage conversions, but leave that to
1071 * later.
1072 */
1073 return "Impossibly level change request for RAID1";
1074
1075 case 10:
1076 /* RAID10 can be converted from near mode to
1077 * RAID0 by removing some devices.
1078 * It can also be reshaped if the kernel supports
1079 * new_data_offset.
1080 */
1081 switch (info->new_level) {
1082 case 0:
1083 if ((info->array.layout & ~0xff) != 0x100)
1084 return "Cannot Grow RAID10 with far/offset layout";
1085 /* number of devices must be multiple of number of copies */
1086 if (info->array.raid_disks % (info->array.layout & 0xff))
1087 return "RAID10 layout too complex for Grow operation";
1088
1089 new_disks = (info->array.raid_disks
1090 / (info->array.layout & 0xff));
1091 if (info->delta_disks == UnSet)
1092 info->delta_disks = (new_disks
1093 - info->array.raid_disks);
1094
1095 if (info->delta_disks != new_disks - info->array.raid_disks)
1096 return "New number of raid-devices impossible for RAID10";
1097 if (info->new_chunk &&
1098 info->new_chunk != info->array.chunk_size)
1099 return "Cannot change chunk-size with RAID10 Grow";
1100
1101 /* looks good */
1102 re->level = 0;
1103 re->before.data_disks = new_disks;
1104 re->after.data_disks = re->before.data_disks;
1105 return NULL;
1106
1107 case 10:
1108 near = info->array.layout & 0xff;
1109 far = (info->array.layout >> 8) & 0xff;
1110 offset = info->array.layout & 0x10000;
1111 if (far > 1 && !offset)
1112 return "Cannot reshape RAID10 in far-mode";
1113 copies = near * far;
1114
1115 old_chunk = info->array.chunk_size * far;
1116
1117 if (info->new_layout == UnSet)
1118 info->new_layout = info->array.layout;
1119 else {
1120 near = info->new_layout & 0xff;
1121 far = (info->new_layout >> 8) & 0xff;
1122 offset = info->new_layout & 0x10000;
1123 if (far > 1 && !offset)
1124 return "Cannot reshape RAID10 to far-mode";
1125 if (near * far != copies)
1126 return "Cannot change number of copies"
1127 " when reshaping RAID10";
1128 }
1129 if (info->delta_disks == UnSet)
1130 info->delta_disks = 0;
1131 new_disks = (info->array.raid_disks +
1132 info->delta_disks);
1133
1134 new_chunk = info->new_chunk * far;
1135
1136 re->level = 10;
1137 re->before.layout = info->array.layout;
1138 re->before.data_disks = info->array.raid_disks;
1139 re->after.layout = info->new_layout;
1140 re->after.data_disks = new_disks;
1141 /* For RAID10 we don't do backup but do allow reshape,
1142 * so set backup_blocks to INVALID_SECTORS rather than
1143 * zero.
1144 * And there is no need to synchronise stripes on both
1145 * 'old' and 'new'. So the important
1146 * number is the minimum data_offset difference
1147 * which is the larger of (offset copies * chunk).
1148 */
1149 re->backup_blocks = INVALID_SECTORS;
1150 re->min_offset_change = max(old_chunk, new_chunk) / 512;
1151 if (new_disks < re->before.data_disks &&
1152 info->space_after < re->min_offset_change)
1153 /* Reduce component size by one chunk */
1154 re->new_size = (info->component_size -
1155 re->min_offset_change);
1156 else
1157 re->new_size = info->component_size;
1158 re->new_size = re->new_size * new_disks / copies;
1159 return NULL;
1160
1161 default:
1162 return "RAID10 can only be changed to RAID0";
1163 }
1164 case 0:
1165 /* RAID0 can be converted to RAID10, or to RAID456 */
1166 if (info->new_level == 10) {
1167 if (info->new_layout == UnSet && info->delta_disks == UnSet) {
1168 /* Assume near=2 layout */
1169 info->new_layout = 0x102;
1170 info->delta_disks = info->array.raid_disks;
1171 }
1172 if (info->new_layout == UnSet) {
1173 int copies = 1 + (info->delta_disks
1174 / info->array.raid_disks);
1175 if (info->array.raid_disks * (copies-1)
1176 != info->delta_disks)
1177 return "Impossible number of devices"
1178 " for RAID0->RAID10";
1179 info->new_layout = 0x100 + copies;
1180 }
1181 if (info->delta_disks == UnSet) {
1182 int copies = info->new_layout & 0xff;
1183 if (info->new_layout != 0x100 + copies)
1184 return "New layout impossible"
1185 " for RAID0->RAID10";;
1186 info->delta_disks = (copies - 1) *
1187 info->array.raid_disks;
1188 }
1189 if (info->new_chunk &&
1190 info->new_chunk != info->array.chunk_size)
1191 return "Cannot change chunk-size with RAID0->RAID10";
1192 /* looks good */
1193 re->level = 10;
1194 re->before.data_disks = (info->array.raid_disks +
1195 info->delta_disks);
1196 re->after.data_disks = re->before.data_disks;
1197 re->before.layout = info->new_layout;
1198 return NULL;
1199 }
1200
1201 /* RAID0 can also covert to RAID0/4/5/6 by first converting to
1202 * a raid4 style layout of the final level.
1203 */
1204 switch (info->new_level) {
1205 case 4:
1206 delta_parity = 1;
1207 case 0:
1208 re->level = 4;
1209 re->before.layout = 0;
1210 break;
1211 case 5:
1212 delta_parity = 1;
1213 re->level = 5;
1214 re->before.layout = ALGORITHM_PARITY_N;
1215 if (info->new_layout == UnSet)
1216 info->new_layout = map_name(r5layout, "default");
1217 break;
1218 case 6:
1219 delta_parity = 2;
1220 re->level = 6;
1221 re->before.layout = ALGORITHM_PARITY_N;
1222 if (info->new_layout == UnSet)
1223 info->new_layout = map_name(r6layout, "default");
1224 break;
1225 default:
1226 return "Impossible level change requested";
1227 }
1228 re->before.data_disks = info->array.raid_disks;
1229 /* determining 'after' layout happens outside this 'switch' */
1230 break;
1231
1232 case 4:
1233 info->array.layout = ALGORITHM_PARITY_N;
1234 case 5:
1235 switch (info->new_level) {
1236 case 0:
1237 delta_parity = -1;
1238 case 4:
1239 re->level = info->array.level;
1240 re->before.data_disks = info->array.raid_disks - 1;
1241 re->before.layout = info->array.layout;
1242 break;
1243 case 5:
1244 re->level = 5;
1245 re->before.data_disks = info->array.raid_disks - 1;
1246 re->before.layout = info->array.layout;
1247 break;
1248 case 6:
1249 delta_parity = 1;
1250 re->level = 6;
1251 re->before.data_disks = info->array.raid_disks - 1;
1252 switch (info->array.layout) {
1253 case ALGORITHM_LEFT_ASYMMETRIC:
1254 re->before.layout = ALGORITHM_LEFT_ASYMMETRIC_6;
1255 break;
1256 case ALGORITHM_RIGHT_ASYMMETRIC:
1257 re->before.layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
1258 break;
1259 case ALGORITHM_LEFT_SYMMETRIC:
1260 re->before.layout = ALGORITHM_LEFT_SYMMETRIC_6;
1261 break;
1262 case ALGORITHM_RIGHT_SYMMETRIC:
1263 re->before.layout = ALGORITHM_RIGHT_SYMMETRIC_6;
1264 break;
1265 case ALGORITHM_PARITY_0:
1266 re->before.layout = ALGORITHM_PARITY_0_6;
1267 break;
1268 case ALGORITHM_PARITY_N:
1269 re->before.layout = ALGORITHM_PARITY_N_6;
1270 break;
1271 default:
1272 return "Cannot convert an array with this layout";
1273 }
1274 break;
1275 case 1:
1276 if (info->array.raid_disks != 2)
1277 return "Can only convert a 2-device array to RAID1";
1278 if (info->delta_disks != UnSet &&
1279 info->delta_disks != 0)
1280 return "Cannot set raid_disk when "
1281 "converting RAID5->RAID1";
1282 re->level = 1;
1283 info->new_chunk = 0;
1284 return NULL;
1285 default:
1286 return "Impossible level change requested";
1287 }
1288 break;
1289 case 6:
1290 switch (info->new_level) {
1291 case 4:
1292 case 5:
1293 delta_parity = -1;
1294 case 6:
1295 re->level = 6;
1296 re->before.data_disks = info->array.raid_disks - 2;
1297 re->before.layout = info->array.layout;
1298 break;
1299 default:
1300 return "Impossible level change requested";
1301 }
1302 break;
1303 }
1304
1305 /* If we reached here then it looks like a re-stripe is
1306 * happening. We have determined the intermediate level
1307 * and initial raid_disks/layout and stored these in 're'.
1308 *
1309 * We need to deduce the final layout that can be atomically
1310 * converted to the end state.
1311 */
1312 switch (info->new_level) {
1313 case 0:
1314 /* We can only get to RAID0 from RAID4 or RAID5
1315 * with appropriate layout and one extra device
1316 */
1317 if (re->level != 4 && re->level != 5)
1318 return "Cannot covert to RAID0 from this level";
1319
1320 switch (re->level) {
1321 case 4:
1322 re->before.layout = 0;
1323 re->after.layout = 0;
1324 break;
1325 case 5:
1326 re->after.layout = ALGORITHM_PARITY_N;
1327 break;
1328 }
1329 break;
1330
1331 case 4:
1332 /* We can only get to RAID4 from RAID5 */
1333 if (re->level != 4 && re->level != 5)
1334 return "Cannot convert to RAID4 from this level";
1335
1336 switch (re->level) {
1337 case 4:
1338 re->before.layout = 0;
1339 re->after.layout = 0;
1340 break;
1341 case 5:
1342 re->after.layout = ALGORITHM_PARITY_N;
1343 break;
1344 }
1345 break;
1346
1347 case 5:
1348 /* We get to RAID5 from RAID5 or RAID6 */
1349 if (re->level != 5 && re->level != 6)
1350 return "Cannot convert to RAID5 from this level";
1351
1352 switch (re->level) {
1353 case 5:
1354 if (info->new_layout == UnSet)
1355 re->after.layout = re->before.layout;
1356 else
1357 re->after.layout = info->new_layout;
1358 break;
1359 case 6:
1360 if (info->new_layout == UnSet)
1361 info->new_layout = re->before.layout;
1362
1363 /* after.layout needs to be raid6 version of new_layout */
1364 if (info->new_layout == ALGORITHM_PARITY_N)
1365 re->after.layout = ALGORITHM_PARITY_N;
1366 else {
1367 char layout[40];
1368 char *ls = map_num(r5layout, info->new_layout);
1369 int l;
1370 if (ls) {
1371 /* Current RAID6 layout has a RAID5
1372 * equivalent - good
1373 */
1374 strcat(strcpy(layout, ls), "-6");
1375 l = map_name(r6layout, layout);
1376 if (l == UnSet)
1377 return "Cannot find RAID6 layout"
1378 " to convert to";
1379 } else {
1380 /* Current RAID6 has no equivalent.
1381 * If it is already a '-6' layout we
1382 * can leave it unchanged, else we must
1383 * fail
1384 */
1385 ls = map_num(r6layout, info->new_layout);
1386 if (!ls ||
1387 strcmp(ls+strlen(ls)-2, "-6") != 0)
1388 return "Please specify new layout";
1389 l = info->new_layout;
1390 }
1391 re->after.layout = l;
1392 }
1393 }
1394 break;
1395
1396 case 6:
1397 /* We must already be at level 6 */
1398 if (re->level != 6)
1399 return "Impossible level change";
1400 if (info->new_layout == UnSet)
1401 re->after.layout = info->array.layout;
1402 else
1403 re->after.layout = info->new_layout;
1404 break;
1405 default:
1406 return "Impossible level change requested";
1407 }
1408 if (info->delta_disks == UnSet)
1409 info->delta_disks = delta_parity;
1410
1411 re->after.data_disks = (re->before.data_disks
1412 + info->delta_disks
1413 - delta_parity);
1414 switch (re->level) {
1415 case 6: re->parity = 2;
1416 break;
1417 case 4:
1418 case 5: re->parity = 1;
1419 break;
1420 default: re->parity = 0;
1421 break;
1422 }
1423 /* So we have a restripe operation, we need to calculate the number
1424 * of blocks per reshape operation.
1425 */
1426 re->new_size = info->component_size * re->before.data_disks;
1427 if (info->new_chunk == 0)
1428 info->new_chunk = info->array.chunk_size;
1429 if (re->after.data_disks == re->before.data_disks &&
1430 re->after.layout == re->before.layout &&
1431 info->new_chunk == info->array.chunk_size) {
1432 /* Nothing to change, can change level immediately. */
1433 re->level = info->new_level;
1434 re->backup_blocks = 0;
1435 return NULL;
1436 }
1437 if (re->after.data_disks == 1 && re->before.data_disks == 1) {
1438 /* chunk and layout changes make no difference */
1439 re->level = info->new_level;
1440 re->backup_blocks = 0;
1441 return NULL;
1442 }
1443
1444 if (re->after.data_disks == re->before.data_disks &&
1445 get_linux_version() < 2006032)
1446 return "in-place reshape is not safe before 2.6.32 - sorry.";
1447
1448 if (re->after.data_disks < re->before.data_disks &&
1449 get_linux_version() < 2006030)
1450 return "reshape to fewer devices is not supported before 2.6.30 - sorry.";
1451
1452 re->backup_blocks = compute_backup_blocks(
1453 info->new_chunk, info->array.chunk_size,
1454 re->after.data_disks,
1455 re->before.data_disks);
1456 re->min_offset_change = re->backup_blocks / re->before.data_disks;
1457
1458 re->new_size = info->component_size * re->after.data_disks;
1459 return NULL;
1460 }
1461
1462 static int set_array_size(struct supertype *st, struct mdinfo *sra,
1463 char *text_version)
1464 {
1465 struct mdinfo *info;
1466 char *subarray;
1467 int ret_val = -1;
1468
1469 if ((st == NULL) || (sra == NULL))
1470 return ret_val;
1471
1472 if (text_version == NULL)
1473 text_version = sra->text_version;
1474 subarray = strchr(text_version+1, '/')+1;
1475 info = st->ss->container_content(st, subarray);
1476 if (info) {
1477 unsigned long long current_size = 0;
1478 unsigned long long new_size =
1479 info->custom_array_size/2;
1480
1481 if (sysfs_get_ll(sra, NULL, "array_size", &current_size) == 0 &&
1482 new_size > current_size) {
1483 if (sysfs_set_num(sra, NULL, "array_size", new_size)
1484 < 0)
1485 dprintf("Error: Cannot set array size");
1486 else {
1487 ret_val = 0;
1488 dprintf("Array size changed");
1489 }
1490 dprintf(" from %llu to %llu.\n",
1491 current_size, new_size);
1492 }
1493 sysfs_free(info);
1494 } else
1495 dprintf("Error: set_array_size(): info pointer in NULL\n");
1496
1497 return ret_val;
1498 }
1499
1500 static int reshape_array(char *container, int fd, char *devname,
1501 struct supertype *st, struct mdinfo *info,
1502 int force, struct mddev_dev *devlist,
1503 unsigned long long data_offset,
1504 char *backup_file, int verbose, int forked,
1505 int restart, int freeze_reshape);
1506 static int reshape_container(char *container, char *devname,
1507 int mdfd,
1508 struct supertype *st,
1509 struct mdinfo *info,
1510 int force,
1511 char *backup_file,
1512 int verbose, int restart, int freeze_reshape);
1513
1514 int Grow_reshape(char *devname, int fd,
1515 struct mddev_dev *devlist,
1516 unsigned long long data_offset,
1517 struct context *c, struct shape *s)
1518 {
1519 /* Make some changes in the shape of an array.
1520 * The kernel must support the change.
1521 *
1522 * There are three different changes. Each can trigger
1523 * a resync or recovery so we freeze that until we have
1524 * requested everything (if kernel supports freezing - 2.6.30).
1525 * The steps are:
1526 * - change size (i.e. component_size)
1527 * - change level
1528 * - change layout/chunksize/ndisks
1529 *
1530 * The last can require a reshape. It is different on different
1531 * levels so we need to check the level before actioning it.
1532 * Some times the level change needs to be requested after the
1533 * reshape (e.g. raid6->raid5, raid5->raid0)
1534 *
1535 */
1536 struct mdu_array_info_s array;
1537 int rv = 0;
1538 struct supertype *st;
1539 char *subarray = NULL;
1540
1541 int frozen;
1542 int changed = 0;
1543 char *container = NULL;
1544 int cfd = -1;
1545
1546 struct mddev_dev *dv;
1547 int added_disks;
1548
1549 struct mdinfo info;
1550 struct mdinfo *sra;
1551
1552 if (ioctl(fd, GET_ARRAY_INFO, &array) < 0) {
1553 pr_err("%s is not an active md array - aborting\n",
1554 devname);
1555 return 1;
1556 }
1557 if (data_offset != INVALID_SECTORS && array.level != 10
1558 && (array.level < 4 || array.level > 6)) {
1559 pr_err("--grow --data-offset not yet supported\n");
1560 return 1;
1561 }
1562
1563 if (s->size > 0 &&
1564 (s->chunk || s->level!= UnSet || s->layout_str || s->raiddisks)) {
1565 pr_err("cannot change component size at the same time "
1566 "as other changes.\n"
1567 " Change size first, then check data is intact before "
1568 "making other changes.\n");
1569 return 1;
1570 }
1571
1572 if (s->raiddisks && s->raiddisks < array.raid_disks && array.level > 1 &&
1573 get_linux_version() < 2006032 &&
1574 !check_env("MDADM_FORCE_FEWER")) {
1575 pr_err("reducing the number of devices is not safe before Linux 2.6.32\n"
1576 " Please use a newer kernel\n");
1577 return 1;
1578 }
1579
1580 st = super_by_fd(fd, &subarray);
1581 if (!st) {
1582 pr_err("Unable to determine metadata format for %s\n", devname);
1583 return 1;
1584 }
1585 if (s->raiddisks > st->max_devs) {
1586 pr_err("Cannot increase raid-disks on this array"
1587 " beyond %d\n", st->max_devs);
1588 return 1;
1589 }
1590
1591 /* in the external case we need to check that the requested reshape is
1592 * supported, and perform an initial check that the container holds the
1593 * pre-requisite spare devices (mdmon owns final validation)
1594 */
1595 if (st->ss->external) {
1596 int rv;
1597
1598 if (subarray) {
1599 container = st->container_devnm;
1600 cfd = open_dev_excl(st->container_devnm);
1601 } else {
1602 container = st->devnm;
1603 close(fd);
1604 cfd = open_dev_excl(st->devnm);
1605 fd = cfd;
1606 }
1607 if (cfd < 0) {
1608 pr_err("Unable to open container for %s\n",
1609 devname);
1610 free(subarray);
1611 return 1;
1612 }
1613
1614 rv = st->ss->load_container(st, cfd, NULL);
1615
1616 if (rv) {
1617 pr_err("Cannot read superblock for %s\n",
1618 devname);
1619 free(subarray);
1620 return 1;
1621 }
1622
1623 /* check if operation is supported for metadata handler */
1624 if (st->ss->container_content) {
1625 struct mdinfo *cc = NULL;
1626 struct mdinfo *content = NULL;
1627
1628 cc = st->ss->container_content(st, subarray);
1629 for (content = cc; content ; content = content->next) {
1630 int allow_reshape = 1;
1631
1632 /* check if reshape is allowed based on metadata
1633 * indications stored in content.array.status
1634 */
1635 if (content->array.state & (1<<MD_SB_BLOCK_VOLUME))
1636 allow_reshape = 0;
1637 if (content->array.state
1638 & (1<<MD_SB_BLOCK_CONTAINER_RESHAPE))
1639 allow_reshape = 0;
1640 if (!allow_reshape) {
1641 pr_err("cannot reshape arrays in"
1642 " container with unsupported"
1643 " metadata: %s(%s)\n",
1644 devname, container);
1645 sysfs_free(cc);
1646 free(subarray);
1647 return 1;
1648 }
1649 }
1650 sysfs_free(cc);
1651 }
1652 if (mdmon_running(container))
1653 st->update_tail = &st->updates;
1654 }
1655
1656 added_disks = 0;
1657 for (dv = devlist; dv; dv = dv->next)
1658 added_disks++;
1659 if (s->raiddisks > array.raid_disks &&
1660 array.spare_disks +added_disks < (s->raiddisks - array.raid_disks) &&
1661 !c->force) {
1662 pr_err("Need %d spare%s to avoid degraded array,"
1663 " and only have %d.\n"
1664 " Use --force to over-ride this check.\n",
1665 s->raiddisks - array.raid_disks,
1666 s->raiddisks - array.raid_disks == 1 ? "" : "s",
1667 array.spare_disks + added_disks);
1668 return 1;
1669 }
1670
1671 sra = sysfs_read(fd, NULL, GET_LEVEL | GET_DISKS | GET_DEVS
1672 | GET_STATE | GET_VERSION);
1673 if (sra) {
1674 if (st->ss->external && subarray == NULL) {
1675 array.level = LEVEL_CONTAINER;
1676 sra->array.level = LEVEL_CONTAINER;
1677 }
1678 } else {
1679 pr_err("failed to read sysfs parameters for %s\n",
1680 devname);
1681 return 1;
1682 }
1683 frozen = freeze(st);
1684 if (frozen < -1) {
1685 /* freeze() already spewed the reason */
1686 sysfs_free(sra);
1687 return 1;
1688 } else if (frozen < 0) {
1689 pr_err("%s is performing resync/recovery and cannot"
1690 " be reshaped\n", devname);
1691 sysfs_free(sra);
1692 return 1;
1693 }
1694
1695 /* ========= set size =============== */
1696 if (s->size > 0 && (s->size == MAX_SIZE || s->size != (unsigned)array.size)) {
1697 unsigned long long orig_size = get_component_size(fd)/2;
1698 unsigned long long min_csize;
1699 struct mdinfo *mdi;
1700 int raid0_takeover = 0;
1701
1702 if (orig_size == 0)
1703 orig_size = (unsigned) array.size;
1704
1705 if (orig_size == 0) {
1706 pr_err("Cannot set device size in this type of array.\n");
1707 rv = 1;
1708 goto release;
1709 }
1710
1711 if (reshape_super(st, s->size, UnSet, UnSet, 0, 0, UnSet, NULL,
1712 devname, APPLY_METADATA_CHANGES, c->verbose > 0)) {
1713 rv = 1;
1714 goto release;
1715 }
1716 sync_metadata(st);
1717 if (st->ss->external) {
1718 /* metadata can have size limitation
1719 * update size value according to metadata information
1720 */
1721 struct mdinfo *sizeinfo =
1722 st->ss->container_content(st, subarray);
1723 if (sizeinfo) {
1724 unsigned long long new_size =
1725 sizeinfo->custom_array_size/2;
1726 int data_disks = get_data_disks(
1727 sizeinfo->array.level,
1728 sizeinfo->array.layout,
1729 sizeinfo->array.raid_disks);
1730 new_size /= data_disks;
1731 dprintf("Metadata size correction from %llu to "
1732 "%llu (%llu)\n", orig_size, new_size,
1733 new_size * data_disks);
1734 s->size = new_size;
1735 sysfs_free(sizeinfo);
1736 }
1737 }
1738
1739 /* Update the size of each member device in case
1740 * they have been resized. This will never reduce
1741 * below the current used-size. The "size" attribute
1742 * understands '0' to mean 'max'.
1743 */
1744 min_csize = 0;
1745 rv = 0;
1746 for (mdi = sra->devs; mdi; mdi = mdi->next) {
1747 if (sysfs_set_num(sra, mdi, "size",
1748 s->size == MAX_SIZE ? 0 : s->size) < 0) {
1749 /* Probably kernel refusing to let us
1750 * reduce the size - not an error.
1751 */
1752 break;
1753 }
1754 if (array.not_persistent == 0 &&
1755 array.major_version == 0 &&
1756 get_linux_version() < 3001000) {
1757 /* Dangerous to allow size to exceed 2TB */
1758 unsigned long long csize;
1759 if (sysfs_get_ll(sra, mdi, "size", &csize) == 0) {
1760 if (csize >= 2ULL*1024*1024*1024)
1761 csize = 2ULL*1024*1024*1024;
1762 if ((min_csize == 0 || (min_csize
1763 > csize)))
1764 min_csize = csize;
1765 }
1766 }
1767 }
1768 if (rv) {
1769 pr_err("Cannot set size on "
1770 "array members.\n");
1771 goto size_change_error;
1772 }
1773 if (min_csize && s->size > min_csize) {
1774 pr_err("Cannot safely make this array "
1775 "use more than 2TB per device on this kernel.\n");
1776 rv = 1;
1777 goto size_change_error;
1778 }
1779 if (min_csize && s->size == MAX_SIZE) {
1780 /* Don't let the kernel choose a size - it will get
1781 * it wrong
1782 */
1783 pr_err("Limited v0.90 array to "
1784 "2TB per device\n");
1785 s->size = min_csize;
1786 }
1787 if (st->ss->external) {
1788 if (sra->array.level == 0) {
1789 rv = sysfs_set_str(sra, NULL, "level",
1790 "raid5");
1791 if (!rv) {
1792 raid0_takeover = 1;
1793 /* get array parametes after takeover
1794 * to chane one parameter at time only
1795 */
1796 rv = ioctl(fd, GET_ARRAY_INFO, &array);
1797 }
1798 }
1799 /* make sure mdmon is
1800 * aware of the new level */
1801 if (!mdmon_running(st->container_devnm))
1802 start_mdmon(st->container_devnm);
1803 ping_monitor(container);
1804 if (mdmon_running(st->container_devnm) &&
1805 st->update_tail == NULL)
1806 st->update_tail = &st->updates;
1807 }
1808
1809 if (s->size == MAX_SIZE)
1810 s->size = 0;
1811 array.size = s->size;
1812 if ((unsigned)array.size != s->size) {
1813 /* got truncated to 32bit, write to
1814 * component_size instead
1815 */
1816 if (sra)
1817 rv = sysfs_set_num(sra, NULL,
1818 "component_size", s->size);
1819 else
1820 rv = -1;
1821 } else {
1822 rv = ioctl(fd, SET_ARRAY_INFO, &array);
1823
1824 /* manage array size when it is managed externally
1825 */
1826 if ((rv == 0) && st->ss->external)
1827 rv = set_array_size(st, sra, sra->text_version);
1828 }
1829
1830 if (raid0_takeover) {
1831 /* do not recync non-existing parity,
1832 * we will drop it anyway
1833 */
1834 sysfs_set_str(sra, NULL, "sync_action", "frozen");
1835 /* go back to raid0, drop parity disk
1836 */
1837 sysfs_set_str(sra, NULL, "level", "raid0");
1838 ioctl(fd, GET_ARRAY_INFO, &array);
1839 }
1840
1841 size_change_error:
1842 if (rv != 0) {
1843 int err = errno;
1844
1845 /* restore metadata */
1846 if (reshape_super(st, orig_size, UnSet, UnSet, 0, 0,
1847 UnSet, NULL, devname,
1848 ROLLBACK_METADATA_CHANGES,
1849 c->verbose) == 0)
1850 sync_metadata(st);
1851 pr_err("Cannot set device size for %s: %s\n",
1852 devname, strerror(err));
1853 if (err == EBUSY &&
1854 (array.state & (1<<MD_SB_BITMAP_PRESENT)))
1855 cont_err("Bitmap must be removed before size can be changed\n");
1856 rv = 1;
1857 goto release;
1858 }
1859 if (s->assume_clean) {
1860 /* This will fail on kernels older than 3.0 unless
1861 * a backport has been arranged.
1862 */
1863 if (sra == NULL ||
1864 sysfs_set_str(sra, NULL, "resync_start", "none") < 0)
1865 pr_err("--assume-clean not supported with --grow on this kernel\n");
1866 }
1867 ioctl(fd, GET_ARRAY_INFO, &array);
1868 s->size = get_component_size(fd)/2;
1869 if (s->size == 0)
1870 s->size = array.size;
1871 if (c->verbose >= 0) {
1872 if (s->size == orig_size)
1873 pr_err("component size of %s "
1874 "unchanged at %lluK\n",
1875 devname, s->size);
1876 else
1877 pr_err("component size of %s "
1878 "has been set to %lluK\n",
1879 devname, s->size);
1880 }
1881 changed = 1;
1882 } else if (array.level != LEVEL_CONTAINER) {
1883 s->size = get_component_size(fd)/2;
1884 if (s->size == 0)
1885 s->size = array.size;
1886 }
1887
1888 /* See if there is anything else to do */
1889 if ((s->level == UnSet || s->level == array.level) &&
1890 (s->layout_str == NULL) &&
1891 (s->chunk == 0 || s->chunk == array.chunk_size) &&
1892 data_offset == INVALID_SECTORS &&
1893 (s->raiddisks == 0 || s->raiddisks == array.raid_disks)) {
1894 /* Nothing more to do */
1895 if (!changed && c->verbose >= 0)
1896 pr_err("%s: no change requested\n",
1897 devname);
1898 goto release;
1899 }
1900
1901 /* ========= check for Raid10/Raid1 -> Raid0 conversion ===============
1902 * current implementation assumes that following conditions must be met:
1903 * - RAID10:
1904 * - far_copies == 1
1905 * - near_copies == 2
1906 */
1907 if ((s->level == 0 && array.level == 10 && sra &&
1908 array.layout == ((1 << 8) + 2) && !(array.raid_disks & 1)) ||
1909 (s->level == 0 && array.level == 1 && sra)) {
1910 int err;
1911 err = remove_disks_for_takeover(st, sra, array.layout);
1912 if (err) {
1913 dprintf(Name": Array cannot be reshaped\n");
1914 if (cfd > -1)
1915 close(cfd);
1916 rv = 1;
1917 goto release;
1918 }
1919 /* Make sure mdmon has seen the device removal
1920 * and updated metadata before we continue with
1921 * level change
1922 */
1923 if (container)
1924 ping_monitor(container);
1925 }
1926
1927 memset(&info, 0, sizeof(info));
1928 info.array = array;
1929 sysfs_init(&info, fd, NULL);
1930 strcpy(info.text_version, sra->text_version);
1931 info.component_size = s->size*2;
1932 info.new_level = s->level;
1933 info.new_chunk = s->chunk * 1024;
1934 if (info.array.level == LEVEL_CONTAINER) {
1935 info.delta_disks = UnSet;
1936 info.array.raid_disks = s->raiddisks;
1937 } else if (s->raiddisks)
1938 info.delta_disks = s->raiddisks - info.array.raid_disks;
1939 else
1940 info.delta_disks = UnSet;
1941 if (s->layout_str == NULL) {
1942 info.new_layout = UnSet;
1943 if (info.array.level == 6 &&
1944 (info.new_level == 6 || info.new_level == UnSet) &&
1945 info.array.layout >= 16) {
1946 pr_err("%s has a non-standard layout. If you"
1947 " wish to preserve this\n", devname);
1948 cont_err("during the reshape, please specify"
1949 " --layout=preserve\n");
1950 cont_err("If you want to change it, specify a"
1951 " layout or use --layout=normalise\n");
1952 rv = 1;
1953 goto release;
1954 }
1955 } else if (strcmp(s->layout_str, "normalise") == 0 ||
1956 strcmp(s->layout_str, "normalize") == 0) {
1957 /* If we have a -6 RAID6 layout, remove the '-6'. */
1958 info.new_layout = UnSet;
1959 if (info.array.level == 6 && info.new_level == UnSet) {
1960 char l[40], *h;
1961 strcpy(l, map_num(r6layout, info.array.layout));
1962 h = strrchr(l, '-');
1963 if (h && strcmp(h, "-6") == 0) {
1964 *h = 0;
1965 info.new_layout = map_name(r6layout, l);
1966 }
1967 } else {
1968 pr_err("%s is only meaningful when reshaping"
1969 " a RAID6 array.\n", s->layout_str);
1970 rv = 1;
1971 goto release;
1972 }
1973 } else if (strcmp(s->layout_str, "preserve") == 0) {
1974 /* This means that a non-standard RAID6 layout
1975 * is OK.
1976 * In particular:
1977 * - When reshape a RAID6 (e.g. adding a device)
1978 * which is in a non-standard layout, it is OK
1979 * to preserve that layout.
1980 * - When converting a RAID5 to RAID6, leave it in
1981 * the XXX-6 layout, don't re-layout.
1982 */
1983 if (info.array.level == 6 && info.new_level == UnSet)
1984 info.new_layout = info.array.layout;
1985 else if (info.array.level == 5 && info.new_level == 6) {
1986 char l[40];
1987 strcpy(l, map_num(r5layout, info.array.layout));
1988 strcat(l, "-6");
1989 info.new_layout = map_name(r6layout, l);
1990 } else {
1991 pr_err("%s in only meaningful when reshaping"
1992 " to RAID6\n", s->layout_str);
1993 rv = 1;
1994 goto release;
1995 }
1996 } else {
1997 int l = info.new_level;
1998 if (l == UnSet)
1999 l = info.array.level;
2000 switch (l) {
2001 case 5:
2002 info.new_layout = map_name(r5layout, s->layout_str);
2003 break;
2004 case 6:
2005 info.new_layout = map_name(r6layout, s->layout_str);
2006 break;
2007 case 10:
2008 info.new_layout = parse_layout_10(s->layout_str);
2009 break;
2010 case LEVEL_FAULTY:
2011 info.new_layout = parse_layout_faulty(s->layout_str);
2012 break;
2013 default:
2014 pr_err("layout not meaningful"
2015 " with this level\n");
2016 rv = 1;
2017 goto release;
2018 }
2019 if (info.new_layout == UnSet) {
2020 pr_err("layout %s not understood"
2021 " for this level\n",
2022 s->layout_str);
2023 rv = 1;
2024 goto release;
2025 }
2026 }
2027
2028 if (array.level == LEVEL_FAULTY) {
2029 if (s->level != UnSet && s->level != array.level) {
2030 pr_err("cannot change level of Faulty device\n");
2031 rv =1 ;
2032 }
2033 if (s->chunk) {
2034 pr_err("cannot set chunksize of Faulty device\n");
2035 rv =1 ;
2036 }
2037 if (s->raiddisks && s->raiddisks != 1) {
2038 pr_err("cannot set raid_disks of Faulty device\n");
2039 rv =1 ;
2040 }
2041 if (s->layout_str) {
2042 if (ioctl(fd, GET_ARRAY_INFO, &array) != 0) {
2043 dprintf("Cannot get array information.\n");
2044 goto release;
2045 }
2046 array.layout = info.new_layout;
2047 if (ioctl(fd, SET_ARRAY_INFO, &array) != 0) {
2048 pr_err("failed to set new layout\n");
2049 rv = 1;
2050 } else if (c->verbose >= 0)
2051 printf("layout for %s set to %d\n",
2052 devname, array.layout);
2053 }
2054 } else if (array.level == LEVEL_CONTAINER) {
2055 /* This change is to be applied to every array in the
2056 * container. This is only needed when the metadata imposes
2057 * restraints of the various arrays in the container.
2058 * Currently we only know that IMSM requires all arrays
2059 * to have the same number of devices so changing the
2060 * number of devices (On-Line Capacity Expansion) must be
2061 * performed at the level of the container
2062 */
2063 rv = reshape_container(container, devname, -1, st, &info,
2064 c->force, c->backup_file, c->verbose, 0, 0);
2065 frozen = 0;
2066 } else {
2067 /* get spare devices from external metadata
2068 */
2069 if (st->ss->external) {
2070 struct mdinfo *info2;
2071
2072 info2 = st->ss->container_content(st, subarray);
2073 if (info2) {
2074 info.array.spare_disks =
2075 info2->array.spare_disks;
2076 sysfs_free(info2);
2077 }
2078 }
2079
2080 /* Impose these changes on a single array. First
2081 * check that the metadata is OK with the change. */
2082
2083 if (reshape_super(st, 0, info.new_level,
2084 info.new_layout, info.new_chunk,
2085 info.array.raid_disks, info.delta_disks,
2086 c->backup_file, devname, APPLY_METADATA_CHANGES,
2087 c->verbose)) {
2088 rv = 1;
2089 goto release;
2090 }
2091 sync_metadata(st);
2092 rv = reshape_array(container, fd, devname, st, &info, c->force,
2093 devlist, data_offset, c->backup_file, c->verbose,
2094 0, 0, 0);
2095 frozen = 0;
2096 }
2097 release:
2098 sysfs_free(sra);
2099 if (frozen > 0)
2100 unfreeze(st);
2101 return rv;
2102 }
2103
2104 /* verify_reshape_position()
2105 * Function checks if reshape position in metadata is not farther
2106 * than position in md.
2107 * Return value:
2108 * 0 : not valid sysfs entry
2109 * it can be caused by not started reshape, it should be started
2110 * by reshape array or raid0 array is before takeover
2111 * -1 : error, reshape position is obviously wrong
2112 * 1 : success, reshape progress correct or updated
2113 */
2114 static int verify_reshape_position(struct mdinfo *info, int level)
2115 {
2116 int ret_val = 0;
2117 char buf[40];
2118 int rv;
2119
2120 /* read sync_max, failure can mean raid0 array */
2121 rv = sysfs_get_str(info, NULL, "sync_max", buf, 40);
2122
2123 if (rv > 0) {
2124 char *ep;
2125 unsigned long long position = strtoull(buf, &ep, 0);
2126
2127 dprintf(Name": Read sync_max sysfs entry is: %s\n", buf);
2128 if (!(ep == buf || (*ep != 0 && *ep != '\n' && *ep != ' '))) {
2129 position *= get_data_disks(level,
2130 info->new_layout,
2131 info->array.raid_disks);
2132 if (info->reshape_progress < position) {
2133 dprintf("Corrected reshape progress (%llu) to "
2134 "md position (%llu)\n",
2135 info->reshape_progress, position);
2136 info->reshape_progress = position;
2137 ret_val = 1;
2138 } else if (info->reshape_progress > position) {
2139 pr_err("Fatal error: array "
2140 "reshape was not properly frozen "
2141 "(expected reshape position is %llu, "
2142 "but reshape progress is %llu.\n",
2143 position, info->reshape_progress);
2144 ret_val = -1;
2145 } else {
2146 dprintf("Reshape position in md and metadata "
2147 "are the same;");
2148 ret_val = 1;
2149 }
2150 }
2151 } else if (rv == 0) {
2152 /* for valid sysfs entry, 0-length content
2153 * should be indicated as error
2154 */
2155 ret_val = -1;
2156 }
2157
2158 return ret_val;
2159 }
2160
2161 static unsigned long long choose_offset(unsigned long long lo,
2162 unsigned long long hi,
2163 unsigned long long min,
2164 unsigned long long max)
2165 {
2166 /* Choose a new offset between hi and lo.
2167 * It must be between min and max, but
2168 * we would prefer something near the middle of hi/lo, and also
2169 * prefer to be aligned to a big power of 2.
2170 *
2171 * So we start with the middle, then for each bit,
2172 * starting at '1' and increasing, if it is set, we either
2173 * add it or subtract it if possible, preferring the option
2174 * which is furthest from the boundary.
2175 *
2176 * We stop once we get a 1MB alignment. As units are in sectors,
2177 * 1MB = 2*1024 sectors.
2178 */
2179 unsigned long long choice = (lo + hi) / 2;
2180 unsigned long long bit = 1;
2181
2182 for (bit = 1; bit < 2*1024; bit = bit << 1) {
2183 unsigned long long bigger, smaller;
2184 if (! (bit & choice))
2185 continue;
2186 bigger = choice + bit;
2187 smaller = choice - bit;
2188 if (bigger > max && smaller < min)
2189 break;
2190 if (bigger > max)
2191 choice = smaller;
2192 else if (smaller < min)
2193 choice = bigger;
2194 else if (hi - bigger > smaller - lo)
2195 choice = bigger;
2196 else
2197 choice = smaller;
2198 }
2199 return choice;
2200 }
2201
2202 static int set_new_data_offset(struct mdinfo *sra, struct supertype *st,
2203 char *devname, int delta_disks,
2204 unsigned long long data_offset,
2205 unsigned long long min)
2206 {
2207 struct mdinfo *sd;
2208 int dir = 0;
2209 int err = 0;
2210 unsigned long long before, after;
2211
2212 /* Need to find min space before and after so same is used
2213 * on all devices
2214 */
2215 before = UINT64_MAX;
2216 after = UINT64_MAX;
2217 for (sd = sra->devs; sd; sd = sd->next) {
2218 char *dn;
2219 int dfd;
2220 int rv;
2221 struct supertype *st2;
2222 struct mdinfo info2;
2223
2224 if (sd->disk.state & (1<<MD_DISK_FAULTY))
2225 continue;
2226 dn = map_dev(sd->disk.major, sd->disk.minor, 0);
2227 dfd = dev_open(dn, O_RDONLY);
2228 if (dfd < 0) {
2229 pr_err("%s: cannot open component %s\n",
2230 devname, dn ? dn : "-unknown-");
2231 goto release;
2232 }
2233 st2 = dup_super(st);
2234 rv = st2->ss->load_super(st2,dfd, NULL);
2235 close(dfd);
2236 if (rv) {
2237 free(st2);
2238 pr_err("%s: cannot get superblock from %s\n",
2239 devname, dn);
2240 goto release;
2241 }
2242 st2->ss->getinfo_super(st2, &info2, NULL);
2243 st2->ss->free_super(st2);
2244 free(st2);
2245 if (info2.space_before == 0 &&
2246 info2.space_after == 0) {
2247 /* Metadata doesn't support data_offset changes */
2248 return 1;
2249 }
2250 if (before > info2.space_before)
2251 before = info2.space_before;
2252 if (after > info2.space_after)
2253 after = info2.space_after;
2254
2255 if (data_offset != INVALID_SECTORS) {
2256 if (dir == 0) {
2257 if (info2.data_offset == data_offset) {
2258 pr_err("%s: already has that data_offset\n",
2259 dn);
2260 goto release;
2261 }
2262 if (data_offset < info2.data_offset)
2263 dir = -1;
2264 else
2265 dir = 1;
2266 } else if ((data_offset <= info2.data_offset && dir == 1) ||
2267 (data_offset >= info2.data_offset && dir == -1)) {
2268 pr_err("%s: differing data offsets on devices make this --data-offset setting impossible\n",
2269 dn);
2270 goto release;
2271 }
2272 }
2273 }
2274 if (before == UINT64_MAX)
2275 /* impossible really, there must be no devices */
2276 return 1;
2277
2278 for (sd = sra->devs; sd; sd = sd->next) {
2279 char *dn = map_dev(sd->disk.major, sd->disk.minor, 0);
2280 unsigned long long new_data_offset;
2281
2282 if (sd->disk.state & (1<<MD_DISK_FAULTY))
2283 continue;
2284 if (delta_disks < 0) {
2285 /* Don't need any space as array is shrinking
2286 * just move data_offset up by min
2287 */
2288 if (data_offset == INVALID_SECTORS)
2289 new_data_offset = sd->data_offset + min;
2290 else {
2291 if (data_offset < sd->data_offset + min) {
2292 pr_err("--data-offset too small for %s\n",
2293 dn);
2294 goto release;
2295 }
2296 new_data_offset = data_offset;
2297 }
2298 } else if (delta_disks > 0) {
2299 /* need space before */
2300 if (before < min) {
2301 pr_err("Insufficient head-space for reshape on %s\n",
2302 dn);
2303 goto release;
2304 }
2305 if (data_offset == INVALID_SECTORS)
2306 new_data_offset = sd->data_offset - min;
2307 else {
2308 if (data_offset > sd->data_offset - min) {
2309 pr_err("--data-offset too large for %s\n",
2310 dn);
2311 goto release;
2312 }
2313 new_data_offset = data_offset;
2314 }
2315 } else {
2316 if (dir == 0) {
2317 /* can move up or down. If 'data_offset'
2318 * was set we would have already decided,
2319 * so just choose direction with most space.
2320 */
2321 if (before > after)
2322 dir = -1;
2323 else
2324 dir = 1;
2325 }
2326 sysfs_set_str(sra, NULL, "reshape_direction",
2327 dir == 1 ? "backwards" : "forwards");
2328 if (dir > 0) {
2329 /* Increase data offset */
2330 if (after < min) {
2331 pr_err("Insufficient tail-space for reshape on %s\n",
2332 dn);
2333 goto release;
2334 }
2335 if (data_offset != INVALID_SECTORS &&
2336 data_offset < sd->data_offset + min) {
2337 pr_err("--data-offset too small on %s\n",
2338 dn);
2339 goto release;
2340 }
2341 if (data_offset != INVALID_SECTORS)
2342 new_data_offset = data_offset;
2343 else
2344 new_data_offset = choose_offset(sd->data_offset,
2345 sd->data_offset + after,
2346 sd->data_offset + min,
2347 sd->data_offset + after);
2348 } else {
2349 /* Decrease data offset */
2350 if (before < min) {
2351 pr_err("insufficient head-room on %s\n",
2352 dn);
2353 goto release;
2354 }
2355 if (data_offset != INVALID_SECTORS &&
2356 data_offset < sd->data_offset - min) {
2357 pr_err("--data-offset too small on %s\n",
2358 dn);
2359 goto release;
2360 }
2361 if (data_offset != INVALID_SECTORS)
2362 new_data_offset = data_offset;
2363 else
2364 new_data_offset = choose_offset(sd->data_offset - before,
2365 sd->data_offset,
2366 sd->data_offset - before,
2367 sd->data_offset - min);
2368 }
2369 }
2370 err = sysfs_set_num(sra, sd, "new_offset", new_data_offset);
2371 if (err < 0 && errno == E2BIG) {
2372 /* try again after increasing data size to max */
2373 err = sysfs_set_num(sra, sd, "size", 0);
2374 if (err < 0 && errno == EINVAL &&
2375 !(sd->disk.state & (1<<MD_DISK_SYNC))) {
2376 /* some kernels have a bug where you cannot
2377 * use '0' on spare devices. */
2378 sysfs_set_num(sra, sd, "size",
2379 (sra->component_size + after)/2);
2380 }
2381 err = sysfs_set_num(sra, sd, "new_offset",
2382 new_data_offset);
2383 }
2384 if (err < 0) {
2385 if (errno == E2BIG && data_offset != INVALID_SECTORS) {
2386 pr_err("data-offset is too big for %s\n",
2387 dn);
2388 goto release;
2389 }
2390 if (sd == sra->devs &&
2391 (errno == ENOENT || errno == E2BIG))
2392 /* Early kernel, no 'new_offset' file,
2393 * or kernel doesn't like us.
2394 * For RAID5/6 this is not fatal
2395 */
2396 return 1;
2397 pr_err("Cannot set new_offset for %s\n",
2398 dn);
2399 break;
2400 }
2401 }
2402 return err;
2403 release:
2404 return -1;
2405 }
2406
2407 static int raid10_reshape(char *container, int fd, char *devname,
2408 struct supertype *st, struct mdinfo *info,
2409 struct reshape *reshape,
2410 unsigned long long data_offset,
2411 int force, int verbose)
2412 {
2413 /* Changing raid_disks, layout, chunksize or possibly
2414 * just data_offset for a RAID10.
2415 * We must always change data_offset. We change by at least
2416 * ->min_offset_change which is the largest of the old and new
2417 * chunk sizes.
2418 * If raid_disks is increasing, then data_offset must decrease
2419 * by at least this copy size.
2420 * If raid_disks is unchanged, data_offset must increase or
2421 * decrease by at least min_offset_change but preferably by much more.
2422 * We choose half of the available space.
2423 * If raid_disks is decreasing, data_offset must increase by
2424 * at least min_offset_change. To allow of this, component_size
2425 * must be decreased by the same amount.
2426 *
2427 * So we calculate the required minimum and direction, possibly
2428 * reduce the component_size, then iterate through the devices
2429 * and set the new_data_offset.
2430 * If that all works, we set chunk_size, layout, raid_disks, and start
2431 * 'reshape'
2432 */
2433 struct mdinfo *sra;
2434 unsigned long long min;
2435 int err = 0;
2436
2437 sra = sysfs_read(fd, NULL,
2438 GET_COMPONENT|GET_DEVS|GET_OFFSET|GET_STATE|GET_CHUNK
2439 );
2440 if (!sra) {
2441 pr_err("%s: Cannot get array details from sysfs\n",
2442 devname);
2443 goto release;
2444 }
2445 min = reshape->min_offset_change;
2446
2447 if (info->delta_disks)
2448 sysfs_set_str(sra, NULL, "reshape_direction",
2449 info->delta_disks < 0 ? "backwards" : "forwards");
2450 if (info->delta_disks < 0 &&
2451 info->space_after < min) {
2452 int rv = sysfs_set_num(sra, NULL, "component_size",
2453 (sra->component_size -
2454 min)/2);
2455 if (rv) {
2456 pr_err("cannot reduce component size\n");
2457 goto release;
2458 }
2459 }
2460 err = set_new_data_offset(sra, st, devname, info->delta_disks, data_offset,
2461 min);
2462 if (err == 1) {
2463 pr_err("Cannot set new_data_offset: RAID10 reshape not\n");
2464 cont_err("supported on this kernel\n");
2465 err = -1;
2466 }
2467 if (err < 0)
2468 goto release;
2469
2470 if (!err && sysfs_set_num(sra, NULL, "chunk_size", info->new_chunk) < 0)
2471 err = errno;
2472 if (!err && sysfs_set_num(sra, NULL, "layout", reshape->after.layout) < 0)
2473 err = errno;
2474 if (!err && sysfs_set_num(sra, NULL, "raid_disks",
2475 info->array.raid_disks + info->delta_disks) < 0)
2476 err = errno;
2477 if (!err && sysfs_set_str(sra, NULL, "sync_action", "reshape") < 0)
2478 err = errno;
2479 if (err) {
2480 pr_err("Cannot set array shape for %s\n",
2481 devname);
2482 if (err == EBUSY &&
2483 (info->array.state & (1<<MD_SB_BITMAP_PRESENT)))
2484 cont_err(" Bitmap must be removed before"
2485 " shape can be changed\n");
2486 goto release;
2487 }
2488 sysfs_free(sra);
2489 return 0;
2490 release:
2491 sysfs_free(sra);
2492 return 1;
2493 }
2494
2495 static void get_space_after(int fd, struct supertype *st, struct mdinfo *info)
2496 {
2497 struct mdinfo *sra, *sd;
2498 /* Initialisation to silence compiler warning */
2499 unsigned long long min_space_before = 0, min_space_after = 0;
2500 int first = 1;
2501
2502 sra = sysfs_read(fd, NULL, GET_DEVS);
2503 if (!sra)
2504 return;
2505 for (sd = sra->devs; sd; sd = sd->next) {
2506 char *dn;
2507 int dfd;
2508 struct supertype *st2;
2509 struct mdinfo info2;
2510
2511 if (sd->disk.state & (1<<MD_DISK_FAULTY))
2512 continue;
2513 dn = map_dev(sd->disk.major, sd->disk.minor, 0);
2514 dfd = dev_open(dn, O_RDONLY);
2515 if (dfd < 0)
2516 break;
2517 st2 = dup_super(st);
2518 if (st2->ss->load_super(st2,dfd, NULL)) {
2519 close(dfd);
2520 free(st2);
2521 break;
2522 }
2523 close(dfd);
2524 st2->ss->getinfo_super(st2, &info2, NULL);
2525 st2->ss->free_super(st2);
2526 free(st2);
2527 if (first ||
2528 min_space_before > info2.space_before)
2529 min_space_before = info2.space_before;
2530 if (first ||
2531 min_space_after > info2.space_after)
2532 min_space_after = info2.space_after;
2533 first = 0;
2534 }
2535 if (sd == NULL && !first) {
2536 info->space_after = min_space_after;
2537 info->space_before = min_space_before;
2538 }
2539 sysfs_free(sra);
2540 }
2541
2542 static void update_cache_size(char *container, struct mdinfo *sra,
2543 struct mdinfo *info,
2544 int disks, unsigned long long blocks)
2545 {
2546 /* Check that the internal stripe cache is
2547 * large enough, or it won't work.
2548 * It must hold at least 4 stripes of the larger
2549 * chunk size
2550 */
2551 unsigned long cache;
2552 cache = max(info->array.chunk_size, info->new_chunk);
2553 cache *= 4; /* 4 stripes minimum */
2554 cache /= 512; /* convert to sectors */
2555 /* make sure there is room for 'blocks' with a bit to spare */
2556 if (cache < 16 + blocks / disks)
2557 cache = 16 + blocks / disks;
2558 cache /= (4096/512); /* Covert from sectors to pages */
2559
2560 if (sra->cache_size < cache)
2561 subarray_set_num(container, sra, "stripe_cache_size",
2562 cache+1);
2563 }
2564
2565 static int impose_reshape(struct mdinfo *sra,
2566 struct mdinfo *info,
2567 struct supertype *st,
2568 int fd,
2569 int restart,
2570 char *devname, char *container,
2571 struct reshape *reshape)
2572 {
2573 struct mdu_array_info_s array;
2574
2575 sra->new_chunk = info->new_chunk;
2576
2577 if (restart) {
2578 /* for external metadata checkpoint saved by mdmon can be lost
2579 * or missed /due to e.g. crash/. Check if md is not during
2580 * restart farther than metadata points to.
2581 * If so, this means metadata information is obsolete.
2582 */
2583 if (st->ss->external)
2584 verify_reshape_position(info, reshape->level);
2585 sra->reshape_progress = info->reshape_progress;
2586 } else {
2587 sra->reshape_progress = 0;
2588 if (reshape->after.data_disks < reshape->before.data_disks)
2589 /* start from the end of the new array */
2590 sra->reshape_progress = (sra->component_size
2591 * reshape->after.data_disks);
2592 }
2593
2594 ioctl(fd, GET_ARRAY_INFO, &array);
2595 if (info->array.chunk_size == info->new_chunk &&
2596 reshape->before.layout == reshape->after.layout &&
2597 st->ss->external == 0) {
2598 /* use SET_ARRAY_INFO but only if reshape hasn't started */
2599 array.raid_disks = reshape->after.data_disks + reshape->parity;
2600 if (!restart &&
2601 ioctl(fd, SET_ARRAY_INFO, &array) != 0) {
2602 int err = errno;
2603
2604 pr_err("Cannot set device shape for %s: %s\n",
2605 devname, strerror(errno));
2606
2607 if (err == EBUSY &&
2608 (array.state & (1<<MD_SB_BITMAP_PRESENT)))
2609 cont_err("Bitmap must be removed before"
2610 " shape can be changed\n");
2611
2612 goto release;
2613 }
2614 } else if (!restart) {
2615 /* set them all just in case some old 'new_*' value
2616 * persists from some earlier problem.
2617 */
2618 int err = 0;
2619 if (sysfs_set_num(sra, NULL, "chunk_size", info->new_chunk) < 0)
2620 err = errno;
2621 if (!err && sysfs_set_num(sra, NULL, "layout",
2622 reshape->after.layout) < 0)
2623 err = errno;
2624 if (!err && subarray_set_num(container, sra, "raid_disks",
2625 reshape->after.data_disks +
2626 reshape->parity) < 0)
2627 err = errno;
2628 if (err) {
2629 pr_err("Cannot set device shape for %s\n",
2630 devname);
2631
2632 if (err == EBUSY &&
2633 (array.state & (1<<MD_SB_BITMAP_PRESENT)))
2634 cont_err("Bitmap must be removed before"
2635 " shape can be changed\n");
2636 goto release;
2637 }
2638 }
2639 return 0;
2640 release:
2641 return -1;
2642 }
2643
2644 static int impose_level(int fd, int level, char *devname, int verbose)
2645 {
2646 char *c;
2647 struct mdu_array_info_s array;
2648 struct mdinfo info;
2649 sysfs_init(&info, fd, NULL);
2650
2651 ioctl(fd, GET_ARRAY_INFO, &array);
2652 if (level == 0 &&
2653 (array.level >= 4 && array.level <= 6)) {
2654 /* To convert to RAID0 we need to fail and
2655 * remove any non-data devices. */
2656 int found = 0;
2657 int d;
2658 int data_disks = array.raid_disks - 1;
2659 if (array.level == 6)
2660 data_disks -= 1;
2661 if (array.level == 5 &&
2662 array.layout != ALGORITHM_PARITY_N)
2663 return -1;
2664 if (array.level == 6 &&
2665 array.layout != ALGORITHM_PARITY_N_6)
2666 return -1;
2667 sysfs_set_str(&info, NULL,"sync_action", "idle");
2668 /* First remove any spares so no recovery starts */
2669 for (d = 0, found = 0;
2670 d < MAX_DISKS && found < array.nr_disks;
2671 d++) {
2672 mdu_disk_info_t disk;
2673 disk.number = d;
2674 if (ioctl(fd, GET_DISK_INFO, &disk) < 0)
2675 continue;
2676 if (disk.major == 0 && disk.minor == 0)
2677 continue;
2678 found++;
2679 if ((disk.state & (1 << MD_DISK_ACTIVE))
2680 && disk.raid_disk < data_disks)
2681 /* keep this */
2682 continue;
2683 ioctl(fd, HOT_REMOVE_DISK,
2684 makedev(disk.major, disk.minor));
2685 }
2686 /* Now fail anything left */
2687 ioctl(fd, GET_ARRAY_INFO, &array);
2688 for (d = 0, found = 0;
2689 d < MAX_DISKS && found < array.nr_disks;
2690 d++) {
2691 int cnt;
2692 mdu_disk_info_t disk;
2693 disk.number = d;
2694 if (ioctl(fd, GET_DISK_INFO, &disk) < 0)
2695 continue;
2696 if (disk.major == 0 && disk.minor == 0)
2697 continue;
2698 found++;
2699 if ((disk.state & (1 << MD_DISK_ACTIVE))
2700 && disk.raid_disk < data_disks)
2701 /* keep this */
2702 continue;
2703 ioctl(fd, SET_DISK_FAULTY,
2704 makedev(disk.major, disk.minor));
2705 cnt = 5;
2706 while (ioctl(fd, HOT_REMOVE_DISK,
2707 makedev(disk.major, disk.minor)) < 0
2708 && errno == EBUSY
2709 && cnt--) {
2710 usleep(10000);
2711 }
2712 }
2713 }
2714 c = map_num(pers, level);
2715 if (c) {
2716 int err = sysfs_set_str(&info, NULL, "level", c);
2717 if (err) {
2718 err = errno;
2719 pr_err("%s: could not set level to %s\n",
2720 devname, c);
2721 if (err == EBUSY &&
2722 (array.state & (1<<MD_SB_BITMAP_PRESENT)))
2723 cont_err("Bitmap must be removed"
2724 " before level can be changed\n");
2725 return err;
2726 }
2727 if (verbose >= 0)
2728 pr_err("level of %s changed to %s\n",
2729 devname, c);
2730 }
2731 return 0;
2732 }
2733
2734 static int reshape_array(char *container, int fd, char *devname,
2735 struct supertype *st, struct mdinfo *info,
2736 int force, struct mddev_dev *devlist,
2737 unsigned long long data_offset,
2738 char *backup_file, int verbose, int forked,
2739 int restart, int freeze_reshape)
2740 {
2741 struct reshape reshape;
2742 int spares_needed;
2743 char *msg;
2744 int orig_level = UnSet;
2745 int odisks;
2746 int delayed;
2747
2748 struct mdu_array_info_s array;
2749 char *c;
2750
2751 struct mddev_dev *dv;
2752 int added_disks;
2753
2754 int *fdlist = NULL;
2755 unsigned long long *offsets = NULL;
2756 int d;
2757 int nrdisks;
2758 int err;
2759 unsigned long blocks;
2760 unsigned long long array_size;
2761 int done;
2762 struct mdinfo *sra = NULL;
2763
2764 /* when reshaping a RAID0, the component_size might be zero.
2765 * So try to fix that up.
2766 */
2767 if (ioctl(fd, GET_ARRAY_INFO, &array) != 0) {
2768 dprintf("Cannot get array information.\n");
2769 goto release;
2770 }
2771 if (array.level == 0 && info->component_size == 0) {
2772 get_dev_size(fd, NULL, &array_size);
2773 info->component_size = array_size / array.raid_disks;
2774 }
2775
2776 if (array.level == 10)
2777 /* Need space_after info */
2778 get_space_after(fd, st, info);
2779
2780 if (info->reshape_active) {
2781 int new_level = info->new_level;
2782 info->new_level = UnSet;
2783 if (info->delta_disks > 0)
2784 info->array.raid_disks -= info->delta_disks;
2785 msg = analyse_change(devname, info, &reshape);
2786 info->new_level = new_level;
2787 if (info->delta_disks > 0)
2788 info->array.raid_disks += info->delta_disks;
2789 if (!restart)
2790 /* Make sure the array isn't read-only */
2791 ioctl(fd, RESTART_ARRAY_RW, 0);
2792 } else
2793 msg = analyse_change(devname, info, &reshape);
2794 if (msg) {
2795 /* if msg == "", error has already been printed */
2796 if (msg[0])
2797 pr_err("%s\n", msg);
2798 goto release;
2799 }
2800 if (restart &&
2801 (reshape.level != info->array.level ||
2802 reshape.before.layout != info->array.layout ||
2803 reshape.before.data_disks + reshape.parity
2804 != info->array.raid_disks - max(0, info->delta_disks))) {
2805 pr_err("reshape info is not in native format -"
2806 " cannot continue.\n");
2807 goto release;
2808 }
2809
2810 if (st->ss->external && restart && (info->reshape_progress == 0)) {
2811 /* When reshape is restarted from '0', very begin of array
2812 * it is possible that for external metadata reshape and array
2813 * configuration doesn't happen.
2814 * Check if md has the same opinion, and reshape is restarted
2815 * from 0. If so, this is regular reshape start after reshape
2816 * switch in metadata to next array only.
2817 */
2818 if ((verify_reshape_position(info, reshape.level) >= 0) &&
2819 (info->reshape_progress == 0))
2820 restart = 0;
2821 }
2822 if (restart) {
2823 /* reshape already started. just skip to monitoring the reshape */
2824 if (reshape.backup_blocks == 0)
2825 return 0;
2826 if (restart & RESHAPE_NO_BACKUP)
2827 return 0;
2828 goto started;
2829 }
2830 /* The container is frozen but the array may not be.
2831 * So freeze the array so spares don't get put to the wrong use
2832 * FIXME there should probably be a cleaner separation between
2833 * freeze_array and freeze_container.
2834 */
2835 sysfs_freeze_array(info);
2836 /* Check we have enough spares to not be degraded */
2837 added_disks = 0;
2838 for (dv = devlist; dv ; dv=dv->next)
2839 added_disks++;
2840 spares_needed = max(reshape.before.data_disks,
2841 reshape.after.data_disks)
2842 + reshape.parity - array.raid_disks;
2843
2844 if (!force &&
2845 info->new_level > 1 && info->array.level > 1 &&
2846 spares_needed > info->array.spare_disks + added_disks) {
2847 pr_err("Need %d spare%s to avoid degraded array,"
2848 " and only have %d.\n"
2849 " Use --force to over-ride this check.\n",
2850 spares_needed,
2851 spares_needed == 1 ? "" : "s",
2852 info->array.spare_disks + added_disks);
2853 goto release;
2854 }
2855 /* Check we have enough spares to not fail */
2856 spares_needed = max(reshape.before.data_disks,
2857 reshape.after.data_disks)
2858 - array.raid_disks;
2859 if ((info->new_level > 1 || info->new_level == 0) &&
2860 spares_needed > info->array.spare_disks +added_disks) {
2861 pr_err("Need %d spare%s to create working array,"
2862 " and only have %d.\n",
2863 spares_needed,
2864 spares_needed == 1 ? "" : "s",
2865 info->array.spare_disks + added_disks);
2866 goto release;
2867 }
2868
2869 if (reshape.level != array.level) {
2870 int err = impose_level(fd, reshape.level, devname, verbose);
2871 if (err)
2872 goto release;
2873 info->new_layout = UnSet; /* after level change,
2874 * layout is meaningless */
2875 orig_level = array.level;
2876 sysfs_freeze_array(info);
2877
2878 if (reshape.level > 0 && st->ss->external) {
2879 /* make sure mdmon is aware of the new level */
2880 if (mdmon_running(container))
2881 flush_mdmon(container);
2882
2883 if (!mdmon_running(container))
2884 start_mdmon(container);
2885 ping_monitor(container);
2886 if (mdmon_running(container) &&
2887 st->update_tail == NULL)
2888 st->update_tail = &st->updates;
2889 }
2890 }
2891 /* ->reshape_super might have chosen some spares from the
2892 * container that it wants to be part of the new array.
2893 * We can collect them with ->container_content and give
2894 * them to the kernel.
2895 */
2896 if (st->ss->reshape_super && st->ss->container_content) {
2897 char *subarray = strchr(info->text_version+1, '/')+1;
2898 struct mdinfo *info2 =
2899 st->ss->container_content(st, subarray);
2900 struct mdinfo *d;
2901
2902 if (info2) {
2903 sysfs_init(info2, fd, st->devnm);
2904 /* When increasing number of devices, we need to set
2905 * new raid_disks before adding these, or they might
2906 * be rejected.
2907 */
2908 if (reshape.backup_blocks &&
2909 reshape.after.data_disks > reshape.before.data_disks)
2910 subarray_set_num(container, info2, "raid_disks",
2911 reshape.after.data_disks +
2912 reshape.parity);
2913 for (d = info2->devs; d; d = d->next) {
2914 if (d->disk.state == 0 &&
2915 d->disk.raid_disk >= 0) {
2916 /* This is a spare that wants to
2917 * be part of the array.
2918 */
2919 add_disk(fd, st, info2, d);
2920 }
2921 }
2922 sysfs_free(info2);
2923 }
2924 }
2925 /* We might have been given some devices to add to the
2926 * array. Now that the array has been changed to the right
2927 * level and frozen, we can safely add them.
2928 */
2929 if (devlist)
2930 Manage_subdevs(devname, fd, devlist, verbose,
2931 0,NULL, 0);
2932
2933 if (reshape.backup_blocks == 0 && data_offset != INVALID_SECTORS)
2934 reshape.backup_blocks = reshape.before.data_disks * info->array.chunk_size/512;
2935 if (reshape.backup_blocks == 0) {
2936 /* No restriping needed, but we might need to impose
2937 * some more changes: layout, raid_disks, chunk_size
2938 */
2939 /* read current array info */
2940 if (ioctl(fd, GET_ARRAY_INFO, &array) != 0) {
2941 dprintf("Cannot get array information.\n");
2942 goto release;
2943 }
2944 /* compare current array info with new values and if
2945 * it is different update them to new */
2946 if (info->new_layout != UnSet &&
2947 info->new_layout != array.layout) {
2948 array.layout = info->new_layout;
2949 if (ioctl(fd, SET_ARRAY_INFO, &array) != 0) {
2950 pr_err("failed to set new layout\n");
2951 goto release;
2952 } else if (verbose >= 0)
2953 printf("layout for %s set to %d\n",
2954 devname, array.layout);
2955 }
2956 if (info->delta_disks != UnSet &&
2957 info->delta_disks != 0 &&
2958 array.raid_disks != (info->array.raid_disks + info->delta_disks)) {
2959 array.raid_disks += info->delta_disks;
2960 if (ioctl(fd, SET_ARRAY_INFO, &array) != 0) {
2961 pr_err("failed to set raid disks\n");
2962 goto release;
2963 } else if (verbose >= 0) {
2964 printf("raid_disks for %s set to %d\n",
2965 devname, array.raid_disks);
2966 }
2967 }
2968 if (info->new_chunk != 0 &&
2969 info->new_chunk != array.chunk_size) {
2970 if (sysfs_set_num(info, NULL,
2971 "chunk_size", info->new_chunk) != 0) {
2972 pr_err("failed to set chunk size\n");
2973 goto release;
2974 } else if (verbose >= 0)
2975 printf("chunk size for %s set to %d\n",
2976 devname, array.chunk_size);
2977 }
2978 unfreeze(st);
2979 return 0;
2980 }
2981
2982 /*
2983 * There are three possibilities.
2984 * 1/ The array will shrink.
2985 * We need to ensure the reshape will pause before reaching
2986 * the 'critical section'. We also need to fork and wait for
2987 * that to happen. When it does we
2988 * suspend/backup/complete/unfreeze
2989 *
2990 * 2/ The array will not change size.
2991 * This requires that we keep a backup of a sliding window
2992 * so that we can restore data after a crash. So we need
2993 * to fork and monitor progress.
2994 * In future we will allow the data_offset to change, so
2995 * a sliding backup becomes unnecessary.
2996 *
2997 * 3/ The array will grow. This is relatively easy.
2998 * However the kernel's restripe routines will cheerfully
2999 * overwrite some early data before it is safe. So we
3000 * need to make a backup of the early parts of the array
3001 * and be ready to restore it if rebuild aborts very early.
3002 * For externally managed metadata, we still need a forked
3003 * child to monitor the reshape and suspend IO over the region
3004 * that is being reshaped.
3005 *
3006 * We backup data by writing it to one spare, or to a
3007 * file which was given on command line.
3008 *
3009 * In each case, we first make sure that storage is available
3010 * for the required backup.
3011 * Then we:
3012 * - request the shape change.
3013 * - fork to handle backup etc.
3014 */
3015 /* Check that we can hold all the data */
3016 get_dev_size(fd, NULL, &array_size);
3017 if (reshape.new_size < (array_size/512)) {
3018 pr_err("this change will reduce the size of the array.\n"
3019 " use --grow --array-size first to truncate array.\n"
3020 " e.g. mdadm --grow %s --array-size %llu\n",
3021 devname, reshape.new_size/2);
3022 goto release;
3023 }
3024
3025 if (array.level == 10) {
3026 /* Reshaping RAID10 does not require any data backup by
3027 * user-space. Instead it requires that the data_offset
3028 * is changed to avoid the need for backup.
3029 * So this is handled very separately
3030 */
3031 if (restart)
3032 /* Nothing to do. */
3033 return 0;
3034 return raid10_reshape(container, fd, devname, st, info,
3035 &reshape, data_offset,
3036 force, verbose);
3037 }
3038 sra = sysfs_read(fd, NULL,
3039 GET_COMPONENT|GET_DEVS|GET_OFFSET|GET_STATE|GET_CHUNK|
3040 GET_CACHE);
3041 if (!sra) {
3042 pr_err("%s: Cannot get array details from sysfs\n",
3043 devname);
3044 goto release;
3045 }
3046
3047 if (!backup_file)
3048 switch(set_new_data_offset(sra, st, devname,
3049 reshape.after.data_disks - reshape.before.data_disks,
3050 data_offset,
3051 reshape.min_offset_change)) {
3052 case -1:
3053 goto release;
3054 case 0:
3055 /* Updated data_offset, so it's easy now */
3056 update_cache_size(container, sra, info,
3057 min(reshape.before.data_disks,
3058 reshape.after.data_disks),
3059 reshape.backup_blocks);
3060
3061 /* Right, everything seems fine. Let's kick things off.
3062 */
3063 sync_metadata(st);
3064
3065 if (impose_reshape(sra, info, st, fd, restart,
3066 devname, container, &reshape) < 0)
3067 goto release;
3068 if (sysfs_set_str(sra, NULL, "sync_action", "reshape") < 0) {
3069 pr_err("Failed to initiate reshape!\n");
3070 goto release;
3071 }
3072 if (info->new_level == reshape.level)
3073 return 0;
3074 /* need to adjust level when reshape completes */
3075 switch(fork()) {
3076 case -1: /* ignore error, but don't wait */
3077 return 0;
3078 default: /* parent */
3079 return 0;
3080 case 0:
3081 map_fork();
3082 break;
3083 }
3084 wait_reshape(sra);
3085 impose_level(fd, info->new_level, devname, verbose);
3086
3087 return 0;
3088 case 1: /* Couldn't set data_offset, try the old way */
3089 if (data_offset != INVALID_SECTORS) {
3090 pr_err("Cannot update data_offset on this array\n");
3091 goto release;
3092 }
3093 break;
3094 }
3095
3096 started:
3097 /* Decide how many blocks (sectors) for a reshape
3098 * unit. The number we have so far is just a minimum
3099 */
3100 blocks = reshape.backup_blocks;
3101 if (reshape.before.data_disks ==
3102 reshape.after.data_disks) {
3103 /* Make 'blocks' bigger for better throughput, but
3104 * not so big that we reject it below.
3105 * Try for 16 megabytes
3106 */
3107 while (blocks * 32 < sra->component_size &&
3108 blocks < 16*1024*2)
3109 blocks *= 2;
3110 } else
3111 pr_err("Need to backup %luK of critical "
3112 "section..\n", blocks/2);
3113
3114 if (blocks >= sra->component_size/2) {
3115 pr_err("%s: Something wrong"
3116 " - reshape aborted\n",
3117 devname);
3118 goto release;
3119 }
3120
3121 /* Now we need to open all these devices so we can read/write.
3122 */
3123 nrdisks = max(reshape.before.data_disks,
3124 reshape.after.data_disks) + reshape.parity
3125 + sra->array.spare_disks;
3126 fdlist = xcalloc((1+nrdisks), sizeof(int));
3127 offsets = xcalloc((1+nrdisks), sizeof(offsets[0]));
3128
3129 odisks = reshape.before.data_disks + reshape.parity;
3130 d = reshape_prepare_fdlist(devname, sra, odisks,
3131 nrdisks, blocks, backup_file,
3132 fdlist, offsets);
3133 if (d < 0) {
3134 goto release;
3135 }
3136 if ((st->ss->manage_reshape == NULL) ||
3137 (st->ss->recover_backup == NULL)) {
3138 if (backup_file == NULL) {
3139 if (reshape.after.data_disks <=
3140 reshape.before.data_disks) {
3141 pr_err("%s: Cannot grow - "
3142 "need backup-file\n", devname);
3143 goto release;
3144 } else if (sra->array.spare_disks == 0) {
3145 pr_err("%s: Cannot grow - "
3146 "need a spare or backup-file to backup "
3147 "critical section\n", devname);
3148 goto release;
3149 }
3150 } else {
3151 if (!reshape_open_backup_file(backup_file, fd, devname,
3152 (signed)blocks,
3153 fdlist+d, offsets+d,
3154 restart)) {
3155 goto release;
3156 }
3157 d++;
3158 }
3159 }
3160
3161 update_cache_size(container, sra, info,
3162 min(reshape.before.data_disks, reshape.after.data_disks),
3163 blocks);
3164
3165 /* Right, everything seems fine. Let's kick things off.
3166 * If only changing raid_disks, use ioctl, else use
3167 * sysfs.
3168 */
3169 sync_metadata(st);
3170
3171 if (impose_reshape(sra, info, st, fd, restart,
3172 devname, container, &reshape) < 0)
3173 goto release;
3174
3175 err = start_reshape(sra, restart, reshape.before.data_disks,
3176 reshape.after.data_disks);
3177 if (err) {
3178 pr_err("Cannot %s reshape for %s\n",
3179 restart ? "continue" : "start",
3180 devname);
3181 goto release;
3182 }
3183 if (restart)
3184 sysfs_set_str(sra, NULL, "array_state", "active");
3185 if (freeze_reshape) {
3186 free(fdlist);
3187 free(offsets);
3188 sysfs_free(sra);
3189 pr_err("Reshape has to be continued from"
3190 " location %llu when root filesystem has been mounted.\n",
3191 sra->reshape_progress);
3192 return 1;
3193 }
3194
3195 /* Now we just need to kick off the reshape and watch, while
3196 * handling backups of the data...
3197 * This is all done by a forked background process.
3198 */
3199 switch(forked ? 0 : fork()) {
3200 case -1:
3201 pr_err("Cannot run child to monitor reshape: %s\n",
3202 strerror(errno));
3203 abort_reshape(sra);
3204 goto release;
3205 default:
3206 free(fdlist);
3207 free(offsets);
3208 sysfs_free(sra);
3209 return 0;
3210 case 0:
3211 map_fork();
3212 break;
3213 }
3214
3215 /* If another array on the same devices is busy, the
3216 * reshape will wait for them. This would mean that
3217 * the first section that we suspend will stay suspended
3218 * for a long time. So check on that possibility
3219 * by looking for "DELAYED" in /proc/mdstat, and if found,
3220 * wait a while
3221 */
3222 do {
3223 struct mdstat_ent *mds, *m;
3224 delayed = 0;
3225 mds = mdstat_read(0, 0);
3226 for (m = mds; m; m = m->next)
3227 if (strcmp(m->devnm, sra->sys_name) == 0) {
3228 if (m->resync &&
3229 m->percent == RESYNC_DELAYED)
3230 delayed = 1;
3231 if (m->resync == 0)
3232 /* Haven't started the reshape thread
3233 * yet, wait a bit
3234 */
3235 delayed = 2;
3236 break;
3237 }
3238 free_mdstat(mds);
3239 if (delayed == 1 && get_linux_version() < 3007000) {
3240 pr_err("Reshape is delayed, but cannot wait carefully with this kernel.\n"
3241 " You might experience problems until other reshapes complete.\n");
3242 delayed = 0;
3243 }
3244 if (delayed)
3245 sleep(30 - (delayed-1) * 25);
3246 } while (delayed);
3247
3248 close(fd);
3249 if (check_env("MDADM_GROW_VERIFY"))
3250 fd = open(devname, O_RDONLY | O_DIRECT);
3251 else
3252 fd = -1;
3253 mlockall(MCL_FUTURE);
3254
3255 if (st->ss->external) {
3256 /* metadata handler takes it from here */
3257 done = st->ss->manage_reshape(
3258 fd, sra, &reshape, st, blocks,
3259 fdlist, offsets,
3260 d - odisks, fdlist+odisks,
3261 offsets+odisks);
3262 } else
3263 done = child_monitor(
3264 fd, sra, &reshape, st, blocks,
3265 fdlist, offsets,
3266 d - odisks, fdlist+odisks,
3267 offsets+odisks);
3268
3269 free(fdlist);
3270 free(offsets);
3271
3272 if (backup_file && done)
3273 unlink(backup_file);
3274 if (!done) {
3275 abort_reshape(sra);
3276 goto out;
3277 }
3278
3279 if (!st->ss->external &&
3280 !(reshape.before.data_disks != reshape.after.data_disks
3281 && info->custom_array_size) &&
3282 info->new_level == reshape.level &&
3283 !forked) {
3284 /* no need to wait for the reshape to finish as
3285 * there is nothing more to do.
3286 */
3287 sysfs_free(sra);
3288 exit(0);
3289 }
3290 wait_reshape(sra);
3291
3292 if (st->ss->external) {
3293 /* Re-load the metadata as much could have changed */
3294 int cfd = open_dev(st->container_devnm);
3295 if (cfd >= 0) {
3296 flush_mdmon(container);
3297 st->ss->free_super(st);
3298 st->ss->load_container(st, cfd, container);
3299 close(cfd);
3300 }
3301 }
3302
3303 /* set new array size if required customer_array_size is used
3304 * by this metadata.
3305 */
3306 if (reshape.before.data_disks !=
3307 reshape.after.data_disks &&
3308 info->custom_array_size)
3309 set_array_size(st, info, info->text_version);
3310
3311 if (info->new_level != reshape.level) {
3312 if (fd < 0)
3313 fd = open(devname, O_RDONLY);
3314 impose_level(fd, info->new_level, devname, verbose);
3315 close(fd);
3316 if (info->new_level == 0)
3317 st->update_tail = NULL;
3318 }
3319 out:
3320 sysfs_free(sra);
3321 if (forked)
3322 return 0;
3323 unfreeze(st);
3324 exit(0);
3325
3326 release:
3327 free(fdlist);
3328 free(offsets);
3329 if (orig_level != UnSet && sra) {
3330 c = map_num(pers, orig_level);
3331 if (c && sysfs_set_str(sra, NULL, "level", c) == 0)
3332 pr_err("aborting level change\n");
3333 }
3334 sysfs_free(sra);
3335 if (!forked)
3336 unfreeze(st);
3337 return 1;
3338 }
3339
3340 /* mdfd handle is passed to be closed in child process (after fork).
3341 */
3342 int reshape_container(char *container, char *devname,
3343 int mdfd,
3344 struct supertype *st,
3345 struct mdinfo *info,
3346 int force,
3347 char *backup_file,
3348 int verbose, int restart, int freeze_reshape)
3349 {
3350 struct mdinfo *cc = NULL;
3351 int rv = restart;
3352 char last_devnm[32] = "";
3353
3354 /* component_size is not meaningful for a container,
3355 * so pass '0' meaning 'no change'
3356 */
3357 if (!restart &&
3358 reshape_super(st, 0, info->new_level,
3359 info->new_layout, info->new_chunk,
3360 info->array.raid_disks, info->delta_disks,
3361 backup_file, devname, APPLY_METADATA_CHANGES,
3362 verbose)) {
3363 unfreeze(st);
3364 return 1;
3365 }
3366
3367 sync_metadata(st);
3368
3369 /* ping monitor to be sure that update is on disk
3370 */
3371 ping_monitor(container);
3372
3373 switch (fork()) {
3374 case -1: /* error */
3375 perror("Cannot fork to complete reshape\n");
3376 unfreeze(st);
3377 return 1;
3378 default: /* parent */
3379 if (!freeze_reshape)
3380 printf(Name ": multi-array reshape continues"
3381 " in background\n");
3382 return 0;
3383 case 0: /* child */
3384 map_fork();
3385 break;
3386 }
3387
3388 /* close unused handle in child process
3389 */
3390 if (mdfd > -1)
3391 close(mdfd);
3392
3393 while(1) {
3394 /* For each member array with reshape_active,
3395 * we need to perform the reshape.
3396 * We pick the first array that needs reshaping and
3397 * reshape it. reshape_array() will re-read the metadata
3398 * so the next time through a different array should be
3399 * ready for reshape.
3400 * It is possible that the 'different' array will not
3401 * be assembled yet. In that case we simple exit.
3402 * When it is assembled, the mdadm which assembles it
3403 * will take over the reshape.
3404 */
3405 struct mdinfo *content;
3406 int fd;
3407 struct mdstat_ent *mdstat;
3408 char *adev;
3409 int devid;
3410
3411 sysfs_free(cc);
3412
3413 cc = st->ss->container_content(st, NULL);
3414
3415 for (content = cc; content ; content = content->next) {
3416 char *subarray;
3417 if (!content->reshape_active)
3418 continue;
3419
3420 subarray = strchr(content->text_version+1, '/')+1;
3421 mdstat = mdstat_by_subdev(subarray, container);
3422 if (!mdstat)
3423 continue;
3424 if (mdstat->active == 0) {
3425 pr_err("Skipping inactive array %s.\n",
3426 mdstat->devnm);
3427 free_mdstat(mdstat);
3428 mdstat = NULL;
3429 continue;
3430 }
3431 break;
3432 }
3433 if (!content)
3434 break;
3435
3436 devid = devnm2devid(mdstat->devnm);
3437 adev = map_dev(major(devid), minor(devid), 0);
3438 if (!adev)
3439 adev = content->text_version;
3440
3441 fd = open_dev(mdstat->devnm);
3442 if (fd < 0) {
3443 printf(Name ": Device %s cannot be opened for reshape.",
3444 adev);
3445 break;
3446 }
3447
3448 if (strcmp(last_devnm, mdstat->devnm) == 0) {
3449 /* Do not allow for multiple reshape_array() calls for
3450 * the same array.
3451 * It can happen when reshape_array() returns without
3452 * error, when reshape is not finished (wrong reshape
3453 * starting/continuation conditions). Mdmon doesn't
3454 * switch to next array in container and reentry
3455 * conditions for the same array occur.
3456 * This is possibly interim until the behaviour of
3457 * reshape_array is resolved().
3458 */
3459 printf(Name ": Multiple reshape execution detected for "
3460 "device %s.", adev);
3461 close(fd);
3462 break;
3463 }
3464 strcpy(last_devnm, mdstat->devnm);
3465
3466 sysfs_init(content, fd, mdstat->devnm);
3467
3468 if (mdmon_running(container))
3469 flush_mdmon(container);
3470
3471 rv = reshape_array(container, fd, adev, st,
3472 content, force, NULL, 0ULL,
3473 backup_file, verbose, 1, restart,
3474 freeze_reshape);
3475 close(fd);
3476
3477 if (freeze_reshape) {
3478 sysfs_free(cc);
3479 exit(0);
3480 }
3481
3482 restart = 0;
3483 if (rv)
3484 break;
3485
3486 if (mdmon_running(container))
3487 flush_mdmon(container);
3488 }
3489 if (!rv)
3490 unfreeze(st);
3491 sysfs_free(cc);
3492 exit(0);
3493 }
3494
3495 /*
3496 * We run a child process in the background which performs the following
3497 * steps:
3498 * - wait for resync to reach a certain point
3499 * - suspend io to the following section
3500 * - backup that section
3501 * - allow resync to proceed further
3502 * - resume io
3503 * - discard the backup.
3504 *
3505 * When are combined in slightly different ways in the three cases.
3506 * Grow:
3507 * - suspend/backup/allow/wait/resume/discard
3508 * Shrink:
3509 * - allow/wait/suspend/backup/allow/wait/resume/discard
3510 * same-size:
3511 * - wait/resume/discard/suspend/backup/allow
3512 *
3513 * suspend/backup/allow always come together
3514 * wait/resume/discard do too.
3515 * For the same-size case we have two backups to improve flow.
3516 *
3517 */
3518
3519 int progress_reshape(struct mdinfo *info, struct reshape *reshape,
3520 unsigned long long backup_point,
3521 unsigned long long wait_point,
3522 unsigned long long *suspend_point,
3523 unsigned long long *reshape_completed)
3524 {
3525 /* This function is called repeatedly by the reshape manager.
3526 * It determines how much progress can safely be made and allows
3527 * that progress.
3528 * - 'info' identifies the array and particularly records in
3529 * ->reshape_progress the metadata's knowledge of progress
3530 * This is a sector offset from the start of the array
3531 * of the next array block to be relocated. This number
3532 * may increase from 0 or decrease from array_size, depending
3533 * on the type of reshape that is happening.
3534 * Note that in contrast, 'sync_completed' is a block count of the
3535 * reshape so far. It gives the distance between the start point
3536 * (head or tail of device) and the next place that data will be
3537 * written. It always increases.
3538 * - 'reshape' is the structure created by analyse_change
3539 * - 'backup_point' shows how much the metadata manager has backed-up
3540 * data. For reshapes with increasing progress, it is the next address
3541 * to be backed up, previous addresses have been backed-up. For
3542 * decreasing progress, it is the earliest address that has been
3543 * backed up - later address are also backed up.
3544 * So addresses between reshape_progress and backup_point are
3545 * backed up providing those are in the 'correct' order.
3546 * - 'wait_point' is an array address. When reshape_completed
3547 * passes this point, progress_reshape should return. It might
3548 * return earlier if it determines that ->reshape_progress needs
3549 * to be updated or further backup is needed.
3550 * - suspend_point is maintained by progress_reshape and the caller
3551 * should not touch it except to initialise to zero.
3552 * It is an array address and it only increases in 2.6.37 and earlier.
3553 * This makes it difficult to handle reducing reshapes with
3554 * external metadata.
3555 * However: it is similar to backup_point in that it records the
3556 * other end of a suspended region from reshape_progress.
3557 * it is moved to extend the region that is safe to backup and/or
3558 * reshape
3559 * - reshape_completed is read from sysfs and returned. The caller
3560 * should copy this into ->reshape_progress when it has reason to
3561 * believe that the metadata knows this, and any backup outside this
3562 * has been erased.
3563 *
3564 * Return value is:
3565 * 1 if more data from backup_point - but only as far as suspend_point,
3566 * should be backed up
3567 * 0 if things are progressing smoothly
3568 * -1 if the reshape is finished because it is all done,
3569 * -2 if the reshape is finished due to an error.
3570 */
3571
3572 int advancing = (reshape->after.data_disks
3573 >= reshape->before.data_disks);
3574 unsigned long long need_backup; /* All data between start of array and
3575 * here will at some point need to
3576 * be backed up.
3577 */
3578 unsigned long long read_offset, write_offset;
3579 unsigned long long write_range;
3580 unsigned long long max_progress, target, completed;
3581 unsigned long long array_size = (info->component_size
3582 * reshape->before.data_disks);
3583 int fd;
3584 char buf[20];
3585
3586 /* First, we unsuspend any region that is now known to be safe.
3587 * If suspend_point is on the 'wrong' side of reshape_progress, then
3588 * we don't have or need suspension at the moment. This is true for
3589 * native metadata when we don't need to back-up.
3590 */
3591 if (advancing) {
3592 if (info->reshape_progress <= *suspend_point)
3593 sysfs_set_num(info, NULL, "suspend_lo",
3594 info->reshape_progress);
3595 } else {
3596 /* Note: this won't work in 2.6.37 and before.
3597 * Something somewhere should make sure we don't need it!
3598 */
3599 if (info->reshape_progress >= *suspend_point)
3600 sysfs_set_num(info, NULL, "suspend_hi",
3601 info->reshape_progress);
3602 }
3603
3604 /* Now work out how far it is safe to progress.
3605 * If the read_offset for ->reshape_progress is less than
3606 * 'blocks' beyond the write_offset, we can only progress as far
3607 * as a backup.
3608 * Otherwise we can progress until the write_offset for the new location
3609 * reaches (within 'blocks' of) the read_offset at the current location.
3610 * However that region must be suspended unless we are using native
3611 * metadata.
3612 * If we need to suspend more, we limit it to 128M per device, which is
3613 * rather arbitrary and should be some time-based calculation.
3614 */
3615 read_offset = info->reshape_progress / reshape->before.data_disks;
3616 write_offset = info->reshape_progress / reshape->after.data_disks;
3617 write_range = info->new_chunk/512;
3618 if (reshape->before.data_disks == reshape->after.data_disks)
3619 need_backup = array_size;
3620 else
3621 need_backup = reshape->backup_blocks;
3622 if (advancing) {
3623 if (read_offset < write_offset + write_range)
3624 max_progress = backup_point;
3625 else
3626 max_progress =
3627 read_offset *
3628 reshape->after.data_disks;
3629 } else {
3630 if (read_offset > write_offset - write_range)
3631 /* Can only progress as far as has been backed up,
3632 * which must be suspended */
3633 max_progress = backup_point;
3634 else if (info->reshape_progress <= need_backup)
3635 max_progress = backup_point;
3636 else {
3637 if (info->array.major_version >= 0)
3638 /* Can progress until backup is needed */
3639 max_progress = need_backup;
3640 else {
3641 /* Can progress until metadata update is required */
3642 max_progress =
3643 read_offset *
3644 reshape->after.data_disks;
3645 /* but data must be suspended */
3646 if (max_progress < *suspend_point)
3647 max_progress = *suspend_point;
3648 }
3649 }
3650 }
3651
3652 /* We know it is safe to progress to 'max_progress' providing
3653 * it is suspended or we are using native metadata.
3654 * Consider extending suspend_point 128M per device if it
3655 * is less than 64M per device beyond reshape_progress.
3656 * But always do a multiple of 'blocks'
3657 * FIXME this is too big - it takes to long to complete
3658 * this much.
3659 */
3660 target = 64*1024*2 * min(reshape->before.data_disks,
3661 reshape->after.data_disks);
3662 target /= reshape->backup_blocks;
3663 if (target < 2)
3664 target = 2;
3665 target *= reshape->backup_blocks;
3666
3667 /* For externally managed metadata we always need to suspend IO to
3668 * the area being reshaped so we regularly push suspend_point forward.
3669 * For native metadata we only need the suspend if we are going to do
3670 * a backup.
3671 */
3672 if (advancing) {
3673 if ((need_backup > info->reshape_progress
3674 || info->array.major_version < 0) &&
3675 *suspend_point < info->reshape_progress + target) {
3676 if (need_backup < *suspend_point + 2 * target)
3677 *suspend_point = need_backup;
3678 else if (*suspend_point + 2 * target < array_size)
3679 *suspend_point += 2 * target;
3680 else
3681 *suspend_point = array_size;
3682 sysfs_set_num(info, NULL, "suspend_hi", *suspend_point);
3683 if (max_progress > *suspend_point)
3684 max_progress = *suspend_point;
3685 }
3686 } else {
3687 if (info->array.major_version >= 0) {
3688 /* Only need to suspend when about to backup */
3689 if (info->reshape_progress < need_backup * 2 &&
3690 *suspend_point > 0) {
3691 *suspend_point = 0;
3692 sysfs_set_num(info, NULL, "suspend_lo", 0);
3693 sysfs_set_num(info, NULL, "suspend_hi", need_backup);
3694 }
3695 } else {
3696 /* Need to suspend continually */
3697 if (info->reshape_progress < *suspend_point)
3698 *suspend_point = info->reshape_progress;
3699 if (*suspend_point + target < info->reshape_progress)
3700 /* No need to move suspend region yet */;
3701 else {
3702 if (*suspend_point >= 2 * target)
3703 *suspend_point -= 2 * target;
3704 else
3705 *suspend_point = 0;
3706 sysfs_set_num(info, NULL, "suspend_lo",
3707 *suspend_point);
3708 }
3709 if (max_progress < *suspend_point)
3710 max_progress = *suspend_point;
3711 }
3712 }
3713
3714 /* now set sync_max to allow that progress. sync_max, like
3715 * sync_completed is a count of sectors written per device, so
3716 * we find the difference between max_progress and the start point,
3717 * and divide that by after.data_disks to get a sync_max
3718 * number.
3719 * At the same time we convert wait_point to a similar number
3720 * for comparing against sync_completed.
3721 */
3722 /* scale down max_progress to per_disk */
3723 max_progress /= reshape->after.data_disks;
3724 /* Round to chunk size as some kernels give an erroneously high number */
3725 max_progress /= info->new_chunk/512;
3726 max_progress *= info->new_chunk/512;
3727 /* And round to old chunk size as the kernel wants that */
3728 max_progress /= info->array.chunk_size/512;
3729 max_progress *= info->array.chunk_size/512;
3730 /* Limit progress to the whole device */
3731 if (max_progress > info->component_size)
3732 max_progress = info->component_size;
3733 wait_point /= reshape->after.data_disks;
3734 if (!advancing) {
3735 /* switch from 'device offset' to 'processed block count' */
3736 max_progress = info->component_size - max_progress;
3737 wait_point = info->component_size - wait_point;
3738 }
3739
3740 sysfs_set_num(info, NULL, "sync_max", max_progress);
3741
3742 /* Now wait. If we have already reached the point that we were
3743 * asked to wait to, don't wait at all, else wait for any change.
3744 * We need to select on 'sync_completed' as that is the place that
3745 * notifications happen, but we are really interested in
3746 * 'reshape_position'
3747 */
3748 fd = sysfs_get_fd(info, NULL, "sync_completed");
3749 if (fd < 0)
3750 goto check_progress;
3751
3752 if (sysfs_fd_get_ll(fd, &completed) < 0)
3753 goto check_progress;
3754
3755 while (completed < max_progress && completed < wait_point) {
3756 /* Check that sync_action is still 'reshape' to avoid
3757 * waiting forever on a dead array
3758 */
3759 char action[20];
3760 fd_set rfds;
3761 if (sysfs_get_str(info, NULL, "sync_action",
3762 action, 20) <= 0 ||
3763 strncmp(action, "reshape", 7) != 0)
3764 break;
3765 /* Some kernels reset 'sync_completed' to zero
3766 * before setting 'sync_action' to 'idle'.
3767 * So we need these extra tests.
3768 */
3769 if (completed == 0 && advancing
3770 && info->reshape_progress > 0)
3771 break;
3772 if (completed == 0 && !advancing
3773 && info->reshape_progress < (info->component_size
3774 * reshape->after.data_disks))
3775 break;
3776 FD_ZERO(&rfds);
3777 FD_SET(fd, &rfds);
3778 select(fd+1, NULL, NULL, &rfds, NULL);
3779 if (sysfs_fd_get_ll(fd, &completed) < 0)
3780 goto check_progress;
3781 }
3782 /* Some kernels reset 'sync_completed' to zero,
3783 * we need to have real point we are in md
3784 */
3785 if (completed == 0)
3786 completed = max_progress;
3787
3788 /* some kernels can give an incorrectly high 'completed' number */
3789 completed /= (info->new_chunk/512);
3790 completed *= (info->new_chunk/512);
3791 /* Convert 'completed' back in to a 'progress' number */
3792 completed *= reshape->after.data_disks;
3793 if (!advancing) {
3794 completed = info->component_size * reshape->after.data_disks
3795 - completed;
3796 }
3797 *reshape_completed = completed;
3798
3799 close(fd);
3800
3801 /* We return the need_backup flag. Caller will decide
3802 * how much - a multiple of ->backup_blocks up to *suspend_point
3803 */
3804 if (advancing)
3805 return need_backup > info->reshape_progress;
3806 else
3807 return need_backup >= info->reshape_progress;
3808
3809 check_progress:
3810 /* if we couldn't read a number from sync_completed, then
3811 * either the reshape did complete, or it aborted.
3812 * We can tell which by checking for 'none' in reshape_position.
3813 * If it did abort, then it might immediately restart if it
3814 * it was just a device failure that leaves us degraded but
3815 * functioning.
3816 */
3817 strcpy(buf, "hi");
3818 if (sysfs_get_str(info, NULL, "reshape_position", buf, sizeof(buf)) < 0
3819 || strncmp(buf, "none", 4) != 0) {
3820 /* The abort might only be temporary. Wait up to 10
3821 * seconds for fd to contain a valid number again.
3822 */
3823 struct timeval tv;
3824 int rv = -2;
3825 tv.tv_sec = 10;
3826 tv.tv_usec = 0;
3827 while (fd >= 0 && rv < 0 && tv.tv_sec > 0) {
3828 fd_set rfds;
3829 FD_ZERO(&rfds);
3830 FD_SET(fd, &rfds);
3831 if (select(fd+1, NULL, NULL, &rfds, &tv) != 1)
3832 break;
3833 switch (sysfs_fd_get_ll(fd, &completed)) {
3834 case 0:
3835 /* all good again */
3836 rv = 1;
3837 break;
3838 case -2: /* read error - abort */
3839 tv.tv_sec = 0;
3840 break;
3841 }
3842 }
3843 if (fd >= 0)
3844 close(fd);
3845 return rv; /* abort */
3846 } else {
3847 /* Maybe racing with array shutdown - check state */
3848 if (fd >= 0)
3849 close(fd);
3850 if (sysfs_get_str(info, NULL, "array_state", buf, sizeof(buf)) < 0
3851 || strncmp(buf, "inactive", 8) == 0
3852 || strncmp(buf, "clear",5) == 0)
3853 return -2; /* abort */
3854 return -1; /* complete */
3855 }
3856 }
3857
3858 /* FIXME return status is never checked */
3859 static int grow_backup(struct mdinfo *sra,
3860 unsigned long long offset, /* per device */
3861 unsigned long stripes, /* per device, in old chunks */
3862 int *sources, unsigned long long *offsets,
3863 int disks, int chunk, int level, int layout,
3864 int dests, int *destfd, unsigned long long *destoffsets,
3865 int part, int *degraded,
3866 char *buf)
3867 {
3868 /* Backup 'blocks' sectors at 'offset' on each device of the array,
3869 * to storage 'destfd' (offset 'destoffsets'), after first
3870 * suspending IO. Then allow resync to continue
3871 * over the suspended section.
3872 * Use part 'part' of the backup-super-block.
3873 */
3874 int odata = disks;
3875 int rv = 0;
3876 int i;
3877 unsigned long long ll;
3878 int new_degraded;
3879 //printf("offset %llu\n", offset);
3880 if (level >= 4)
3881 odata--;
3882 if (level == 6)
3883 odata--;
3884
3885 /* Check that array hasn't become degraded, else we might backup the wrong data */
3886 if (sysfs_get_ll(sra, NULL, "degraded", &ll) < 0)
3887 return -1; /* FIXME this error is ignored */
3888 new_degraded = (int)ll;
3889 if (new_degraded != *degraded) {
3890 /* check each device to ensure it is still working */
3891 struct mdinfo *sd;
3892 for (sd = sra->devs ; sd ; sd = sd->next) {
3893 if (sd->disk.state & (1<<MD_DISK_FAULTY))
3894 continue;
3895 if (sd->disk.state & (1<<MD_DISK_SYNC)) {
3896 char sbuf[20];
3897 if (sysfs_get_str(sra, sd, "state", sbuf, 20) < 0 ||
3898 strstr(sbuf, "faulty") ||
3899 strstr(sbuf, "in_sync") == NULL) {
3900 /* this device is dead */
3901 sd->disk.state = (1<<MD_DISK_FAULTY);
3902 if (sd->disk.raid_disk >= 0 &&
3903 sources[sd->disk.raid_disk] >= 0) {
3904 close(sources[sd->disk.raid_disk]);
3905 sources[sd->disk.raid_disk] = -1;
3906 }
3907 }
3908 }
3909 }
3910 *degraded = new_degraded;
3911 }
3912 if (part) {
3913 bsb.arraystart2 = __cpu_to_le64(offset * odata);
3914 bsb.length2 = __cpu_to_le64(stripes * (chunk/512) * odata);
3915 } else {
3916 bsb.arraystart = __cpu_to_le64(offset * odata);
3917 bsb.length = __cpu_to_le64(stripes * (chunk/512) * odata);
3918 }
3919 if (part)
3920 bsb.magic[15] = '2';
3921 for (i = 0; i < dests; i++)
3922 if (part)
3923 lseek64(destfd[i], destoffsets[i] + __le64_to_cpu(bsb.devstart2)*512, 0);
3924 else
3925 lseek64(destfd[i], destoffsets[i], 0);
3926
3927 rv = save_stripes(sources, offsets,
3928 disks, chunk, level, layout,
3929 dests, destfd,
3930 offset*512*odata, stripes * chunk * odata,
3931 buf);
3932
3933 if (rv)
3934 return rv;
3935 bsb.mtime = __cpu_to_le64(time(0));
3936 for (i = 0; i < dests; i++) {
3937 bsb.devstart = __cpu_to_le64(destoffsets[i]/512);
3938
3939 bsb.sb_csum = bsb_csum((char*)&bsb, ((char*)&bsb.sb_csum)-((char*)&bsb));
3940 if (memcmp(bsb.magic, "md_backup_data-2", 16) == 0)
3941 bsb.sb_csum2 = bsb_csum((char*)&bsb,
3942 ((char*)&bsb.sb_csum2)-((char*)&bsb));
3943
3944 rv = -1;
3945 if ((unsigned long long)lseek64(destfd[i], destoffsets[i] - 4096, 0)
3946 != destoffsets[i] - 4096)
3947 break;
3948 if (write(destfd[i], &bsb, 512) != 512)
3949 break;
3950 if (destoffsets[i] > 4096) {
3951 if ((unsigned long long)lseek64(destfd[i], destoffsets[i]+stripes*chunk*odata, 0) !=
3952 destoffsets[i]+stripes*chunk*odata)
3953 break;
3954 if (write(destfd[i], &bsb, 512) != 512)
3955 break;
3956 }
3957 fsync(destfd[i]);
3958 rv = 0;
3959 }
3960
3961 return rv;
3962 }
3963
3964 /* in 2.6.30, the value reported by sync_completed can be
3965 * less that it should be by one stripe.
3966 * This only happens when reshape hits sync_max and pauses.
3967 * So allow wait_backup to either extent sync_max further
3968 * than strictly necessary, or return before the
3969 * sync has got quite as far as we would really like.
3970 * This is what 'blocks2' is for.
3971 * The various caller give appropriate values so that
3972 * every works.
3973 */
3974 /* FIXME return value is often ignored */
3975 static int forget_backup(int dests, int *destfd,
3976 unsigned long long *destoffsets,
3977 int part)
3978 {
3979 /*
3980 * Erase backup 'part' (which is 0 or 1)
3981 */
3982 int i;
3983 int rv;
3984
3985 if (part) {
3986 bsb.arraystart2 = __cpu_to_le64(0);
3987 bsb.length2 = __cpu_to_le64(0);
3988 } else {
3989 bsb.arraystart = __cpu_to_le64(0);
3990 bsb.length = __cpu_to_le64(0);
3991 }
3992 bsb.mtime = __cpu_to_le64(time(0));
3993 rv = 0;
3994 for (i = 0; i < dests; i++) {
3995 bsb.devstart = __cpu_to_le64(destoffsets[i]/512);
3996 bsb.sb_csum = bsb_csum((char*)&bsb, ((char*)&bsb.sb_csum)-((char*)&bsb));
3997 if (memcmp(bsb.magic, "md_backup_data-2", 16) == 0)
3998 bsb.sb_csum2 = bsb_csum((char*)&bsb,
3999 ((char*)&bsb.sb_csum2)-((char*)&bsb));
4000 if ((unsigned long long)lseek64(destfd[i], destoffsets[i]-4096, 0) !=
4001 destoffsets[i]-4096)
4002 rv = -1;
4003 if (rv == 0 &&
4004 write(destfd[i], &bsb, 512) != 512)
4005 rv = -1;
4006 fsync(destfd[i]);
4007 }
4008 return rv;
4009 }
4010
4011 static void fail(char *msg)
4012 {
4013 int rv;
4014 rv = (write(2, msg, strlen(msg)) != (int)strlen(msg));
4015 rv |= (write(2, "\n", 1) != 1);
4016 exit(rv ? 1 : 2);
4017 }
4018
4019 static char *abuf, *bbuf;
4020 static unsigned long long abuflen;
4021 static void validate(int afd, int bfd, unsigned long long offset)
4022 {
4023 /* check that the data in the backup against the array.
4024 * This is only used for regression testing and should not
4025 * be used while the array is active
4026 */
4027 if (afd < 0)
4028 return;
4029 lseek64(bfd, offset - 4096, 0);
4030 if (read(bfd, &bsb2, 512) != 512)
4031 fail("cannot read bsb");
4032 if (bsb2.sb_csum != bsb_csum((char*)&bsb2,
4033 ((char*)&bsb2.sb_csum)-((char*)&bsb2)))
4034 fail("first csum bad");
4035 if (memcmp(bsb2.magic, "md_backup_data", 14) != 0)
4036 fail("magic is bad");
4037 if (memcmp(bsb2.magic, "md_backup_data-2", 16) == 0 &&
4038 bsb2.sb_csum2 != bsb_csum((char*)&bsb2,
4039 ((char*)&bsb2.sb_csum2)-((char*)&bsb2)))
4040 fail("second csum bad");
4041
4042 if (__le64_to_cpu(bsb2.devstart)*512 != offset)
4043 fail("devstart is wrong");
4044
4045 if (bsb2.length) {
4046 unsigned long long len = __le64_to_cpu(bsb2.length)*512;
4047
4048 if (abuflen < len) {
4049 free(abuf);
4050 free(bbuf);
4051 abuflen = len;
4052 if (posix_memalign((void**)&abuf, 4096, abuflen) ||
4053 posix_memalign((void**)&bbuf, 4096, abuflen)) {
4054 abuflen = 0;
4055 /* just stop validating on mem-alloc failure */
4056 return;
4057 }
4058 }
4059
4060 lseek64(bfd, offset, 0);
4061 if ((unsigned long long)read(bfd, bbuf, len) != len) {
4062 //printf("len %llu\n", len);
4063 fail("read first backup failed");
4064 }
4065 lseek64(afd, __le64_to_cpu(bsb2.arraystart)*512, 0);
4066 if ((unsigned long long)read(afd, abuf, len) != len)
4067 fail("read first from array failed");
4068 if (memcmp(bbuf, abuf, len) != 0) {
4069 #if 0
4070 int i;
4071 printf("offset=%llu len=%llu\n",
4072 (unsigned long long)__le64_to_cpu(bsb2.arraystart)*512, len);
4073 for (i=0; i<len; i++)
4074 if (bbuf[i] != abuf[i]) {
4075 printf("first diff byte %d\n", i);
4076 break;
4077 }
4078 #endif
4079 fail("data1 compare failed");
4080 }
4081 }
4082 if (bsb2.length2) {
4083 unsigned long long len = __le64_to_cpu(bsb2.length2)*512;
4084
4085 if (abuflen < len) {
4086 free(abuf);
4087 free(bbuf);
4088 abuflen = len;
4089 abuf = xmalloc(abuflen);
4090 bbuf = xmalloc(abuflen);
4091 }
4092
4093 lseek64(bfd, offset+__le64_to_cpu(bsb2.devstart2)*512, 0);
4094 if ((unsigned long long)read(bfd, bbuf, len) != len)
4095 fail("read second backup failed");
4096 lseek64(afd, __le64_to_cpu(bsb2.arraystart2)*512, 0);
4097 if ((unsigned long long)read(afd, abuf, len) != len)
4098 fail("read second from array failed");
4099 if (memcmp(bbuf, abuf, len) != 0)
4100 fail("data2 compare failed");
4101 }
4102 }
4103
4104 int child_monitor(int afd, struct mdinfo *sra, struct reshape *reshape,
4105 struct supertype *st, unsigned long blocks,
4106 int *fds, unsigned long long *offsets,
4107 int dests, int *destfd, unsigned long long *destoffsets)
4108 {
4109 /* Monitor a reshape where backup is being performed using
4110 * 'native' mechanism - either to a backup file, or
4111 * to some space in a spare.
4112 */
4113 char *buf;
4114 int degraded = -1;
4115 unsigned long long speed;
4116 unsigned long long suspend_point, array_size;
4117 unsigned long long backup_point, wait_point;
4118 unsigned long long reshape_completed;
4119 int done = 0;
4120 int increasing = reshape->after.data_disks >= reshape->before.data_disks;
4121 int part = 0; /* The next part of the backup area to fill. It may already
4122 * be full, so we need to check */
4123 int level = reshape->level;
4124 int layout = reshape->before.layout;
4125 int data = reshape->before.data_disks;
4126 int disks = reshape->before.data_disks + reshape->parity;
4127 int chunk = sra->array.chunk_size;
4128 struct mdinfo *sd;
4129 unsigned long stripes;
4130 int uuid[4];
4131
4132 /* set up the backup-super-block. This requires the
4133 * uuid from the array.
4134 */
4135 /* Find a superblock */
4136 for (sd = sra->devs; sd; sd = sd->next) {
4137 char *dn;
4138 int devfd;
4139 int ok;
4140 if (sd->disk.state & (1<<MD_DISK_FAULTY))
4141 continue;
4142 dn = map_dev(sd->disk.major, sd->disk.minor, 1);
4143 devfd = dev_open(dn, O_RDONLY);
4144 if (devfd < 0)
4145 continue;
4146 ok = st->ss->load_super(st, devfd, NULL);
4147 close(devfd);
4148 if (ok == 0)
4149 break;
4150 }
4151 if (!sd) {
4152 pr_err("Cannot find a superblock\n");
4153 return 0;
4154 }
4155
4156 memset(&bsb, 0, 512);
4157 memcpy(bsb.magic, "md_backup_data-1", 16);
4158 st->ss->uuid_from_super(st, uuid);
4159 memcpy(bsb.set_uuid, uuid, 16);
4160 bsb.mtime = __cpu_to_le64(time(0));
4161 bsb.devstart2 = blocks;
4162
4163 stripes = blocks / (sra->array.chunk_size/512) /
4164 reshape->before.data_disks;
4165
4166 if (posix_memalign((void**)&buf, 4096, disks * chunk))
4167 /* Don't start the 'reshape' */
4168 return 0;
4169 if (reshape->before.data_disks == reshape->after.data_disks) {
4170 sysfs_get_ll(sra, NULL, "sync_speed_min", &speed);
4171 sysfs_set_num(sra, NULL, "sync_speed_min", 200000);
4172 }
4173
4174 if (increasing) {
4175 array_size = sra->component_size * reshape->after.data_disks;
4176 backup_point = sra->reshape_progress;
4177 suspend_point = 0;
4178 } else {
4179 array_size = sra->component_size * reshape->before.data_disks;
4180 backup_point = reshape->backup_blocks;
4181 suspend_point = array_size;
4182 }
4183
4184 while (!done) {
4185 int rv;
4186
4187 /* Want to return as soon the oldest backup slot can
4188 * be released as that allows us to start backing up
4189 * some more, providing suspend_point has been
4190 * advanced, which it should have.
4191 */
4192 if (increasing) {
4193 wait_point = array_size;
4194 if (part == 0 && __le64_to_cpu(bsb.length) > 0)
4195 wait_point = (__le64_to_cpu(bsb.arraystart) +
4196 __le64_to_cpu(bsb.length));
4197 if (part == 1 && __le64_to_cpu(bsb.length2) > 0)
4198 wait_point = (__le64_to_cpu(bsb.arraystart2) +
4199 __le64_to_cpu(bsb.length2));
4200 } else {
4201 wait_point = 0;
4202 if (part == 0 && __le64_to_cpu(bsb.length) > 0)
4203 wait_point = __le64_to_cpu(bsb.arraystart);
4204 if (part == 1 && __le64_to_cpu(bsb.length2) > 0)
4205 wait_point = __le64_to_cpu(bsb.arraystart2);
4206 }
4207
4208 rv = progress_reshape(sra, reshape,
4209 backup_point, wait_point,
4210 &suspend_point, &reshape_completed);
4211 /* external metadata would need to ping_monitor here */
4212 sra->reshape_progress = reshape_completed;
4213
4214 /* Clear any backup region that is before 'here' */
4215 if (increasing) {
4216 if (__le64_to_cpu(bsb.length) > 0 &&
4217 reshape_completed >= (__le64_to_cpu(bsb.arraystart) +
4218 __le64_to_cpu(bsb.length)))
4219 forget_backup(dests, destfd,
4220 destoffsets, 0);
4221 if (__le64_to_cpu(bsb.length2) > 0 &&
4222 reshape_completed >= (__le64_to_cpu(bsb.arraystart2) +
4223 __le64_to_cpu(bsb.length2)))
4224 forget_backup(dests, destfd,
4225 destoffsets, 1);
4226 } else {
4227 if (__le64_to_cpu(bsb.length) > 0 &&
4228 reshape_completed <= (__le64_to_cpu(bsb.arraystart)))
4229 forget_backup(dests, destfd,
4230 destoffsets, 0);
4231 if (__le64_to_cpu(bsb.length2) > 0 &&
4232 reshape_completed <= (__le64_to_cpu(bsb.arraystart2)))
4233 forget_backup(dests, destfd,
4234 destoffsets, 1);
4235 }
4236
4237 if (rv < 0) {
4238 if (rv == -1)
4239 done = 1;
4240 break;
4241 }
4242 if (rv == 0 && increasing && !st->ss->external) {
4243 /* No longer need to monitor this reshape */
4244 done = 1;
4245 break;
4246 }
4247
4248 while (rv) {
4249 unsigned long long offset;
4250 unsigned long actual_stripes;
4251 /* Need to backup some data.
4252 * If 'part' is not used and the desired
4253 * backup size is suspended, do a backup,
4254 * then consider the next part.
4255 */
4256 /* Check that 'part' is unused */
4257 if (part == 0 && __le64_to_cpu(bsb.length) != 0)
4258 break;
4259 if (part == 1 && __le64_to_cpu(bsb.length2) != 0)
4260 break;
4261
4262 offset = backup_point / data;
4263 actual_stripes = stripes;
4264 if (increasing) {
4265 if (offset + actual_stripes * (chunk/512) >
4266 sra->component_size)
4267 actual_stripes = ((sra->component_size - offset)
4268 / (chunk/512));
4269 if (offset + actual_stripes * (chunk/512) >
4270 suspend_point/data)
4271 break;
4272 } else {
4273 if (offset < actual_stripes * (chunk/512))
4274 actual_stripes = offset / (chunk/512);
4275 offset -= actual_stripes * (chunk/512);
4276 if (offset < suspend_point/data)
4277 break;
4278 }
4279 if (actual_stripes == 0)
4280 break;
4281 grow_backup(sra, offset, actual_stripes,
4282 fds, offsets,
4283 disks, chunk, level, layout,
4284 dests, destfd, destoffsets,
4285 part, &degraded, buf);
4286 validate(afd, destfd[0], destoffsets[0]);
4287 /* record where 'part' is up to */
4288 part = !part;
4289 if (increasing)
4290 backup_point += actual_stripes * (chunk/512) * data;
4291 else
4292 backup_point -= actual_stripes * (chunk/512) * data;
4293 }
4294 }
4295
4296 /* FIXME maybe call progress_reshape one more time instead */
4297 abort_reshape(sra); /* remove any remaining suspension */
4298 if (reshape->before.data_disks == reshape->after.data_disks)
4299 sysfs_set_num(sra, NULL, "sync_speed_min", speed);
4300 free(buf);
4301 return done;
4302 }
4303
4304 /*
4305 * If any spare contains md_back_data-1 which is recent wrt mtime,
4306 * write that data into the array and update the super blocks with
4307 * the new reshape_progress
4308 */
4309 int Grow_restart(struct supertype *st, struct mdinfo *info, int *fdlist, int cnt,
4310 char *backup_file, int verbose)
4311 {
4312 int i, j;
4313 int old_disks;
4314 unsigned long long *offsets;
4315 unsigned long long nstripe, ostripe;
4316 int ndata, odata;
4317
4318 odata = info->array.raid_disks - info->delta_disks - 1;
4319 if (info->array.level == 6) odata--; /* number of data disks */
4320 ndata = info->array.raid_disks - 1;
4321 if (info->new_level == 6) ndata--;
4322
4323 old_disks = info->array.raid_disks - info->delta_disks;
4324
4325 if (info->delta_disks <= 0)
4326 /* Didn't grow, so the backup file must have
4327 * been used
4328 */
4329 old_disks = cnt;
4330 for (i=old_disks-(backup_file?1:0); i<cnt; i++) {
4331 struct mdinfo dinfo;
4332 int fd;
4333 int bsbsize;
4334 char *devname, namebuf[20];
4335 unsigned long long lo, hi;
4336
4337 /* This was a spare and may have some saved data on it.
4338 * Load the superblock, find and load the
4339 * backup_super_block.
4340 * If either fail, go on to next device.
4341 * If the backup contains no new info, just return
4342 * else restore data and update all superblocks
4343 */
4344 if (i == old_disks-1) {
4345 fd = open(backup_file, O_RDONLY);
4346 if (fd<0) {
4347 pr_err("backup file %s inaccessible: %s\n",
4348 backup_file, strerror(errno));
4349 continue;
4350 }
4351 devname = backup_file;
4352 } else {
4353 fd = fdlist[i];
4354 if (fd < 0)
4355 continue;
4356 if (st->ss->load_super(st, fd, NULL))
4357 continue;
4358
4359 st->ss->getinfo_super(st, &dinfo, NULL);
4360 st->ss->free_super(st);
4361
4362 if (lseek64(fd,
4363 (dinfo.data_offset + dinfo.component_size - 8) <<9,
4364 0) < 0) {
4365 pr_err("Cannot seek on device %d\n", i);
4366 continue; /* Cannot seek */
4367 }
4368 sprintf(namebuf, "device-%d", i);
4369 devname = namebuf;
4370 }
4371 if (read(fd, &bsb, sizeof(bsb)) != sizeof(bsb)) {
4372 if (verbose)
4373 pr_err("Cannot read from %s\n", devname);
4374 continue; /* Cannot read */
4375 }
4376 if (memcmp(bsb.magic, "md_backup_data-1", 16) != 0 &&
4377 memcmp(bsb.magic, "md_backup_data-2", 16) != 0) {
4378 if (verbose)
4379 pr_err("No backup metadata on %s\n", devname);
4380 continue;
4381 }
4382 if (bsb.sb_csum != bsb_csum((char*)&bsb, ((char*)&bsb.sb_csum)-((char*)&bsb))) {
4383 if (verbose)
4384 pr_err("Bad backup-metadata checksum on %s\n", devname);
4385 continue; /* bad checksum */
4386 }
4387 if (memcmp(bsb.magic, "md_backup_data-2", 16) == 0 &&
4388 bsb.sb_csum2 != bsb_csum((char*)&bsb, ((char*)&bsb.sb_csum2)-((char*)&bsb))) {
4389 if (verbose)
4390 pr_err("Bad backup-metadata checksum2 on %s\n", devname);
4391 continue; /* Bad second checksum */
4392 }
4393 if (memcmp(bsb.set_uuid,info->uuid, 16) != 0) {
4394 if (verbose)
4395 pr_err("Wrong uuid on backup-metadata on %s\n", devname);
4396 continue; /* Wrong uuid */
4397 }
4398
4399 /* array utime and backup-mtime should be updated at much the same time, but it seems that
4400 * sometimes they aren't... So allow considerable flexability in matching, and allow
4401 * this test to be overridden by an environment variable.
4402 */
4403 if (info->array.utime > (int)__le64_to_cpu(bsb.mtime) + 2*60*60 ||
4404 info->array.utime < (int)__le64_to_cpu(bsb.mtime) - 10*60) {
4405 if (check_env("MDADM_GROW_ALLOW_OLD")) {
4406 pr_err("accepting backup with timestamp %lu "
4407 "for array with timestamp %lu\n",
4408 (unsigned long)__le64_to_cpu(bsb.mtime),
4409 (unsigned long)info->array.utime);
4410 } else {
4411 pr_err("too-old timestamp on backup-metadata on %s\n", devname);
4412 pr_err("If you think it is should be safe, try 'export MDADM_GROW_ALLOW_OLD=1'\n");
4413 continue; /* time stamp is too bad */
4414 }
4415 }
4416
4417 if (bsb.magic[15] == '1') {
4418 if (bsb.length == 0)
4419 continue;
4420 if (info->delta_disks >= 0) {
4421 /* reshape_progress is increasing */
4422 if (__le64_to_cpu(bsb.arraystart)
4423 + __le64_to_cpu(bsb.length)
4424 < info->reshape_progress) {
4425 nonew:
4426 if (verbose)
4427 pr_err("backup-metadata found on %s but is not needed\n", devname);
4428 continue; /* No new data here */
4429 }
4430 } else {
4431 /* reshape_progress is decreasing */
4432 if (__le64_to_cpu(bsb.arraystart) >=
4433 info->reshape_progress)
4434 goto nonew; /* No new data here */
4435 }
4436 } else {
4437 if (bsb.length == 0 && bsb.length2 == 0)
4438 continue;
4439 if (info->delta_disks >= 0) {
4440 /* reshape_progress is increasing */
4441 if ((__le64_to_cpu(bsb.arraystart)
4442 + __le64_to_cpu(bsb.length)
4443 < info->reshape_progress)
4444 &&
4445 (__le64_to_cpu(bsb.arraystart2)
4446 + __le64_to_cpu(bsb.length2)
4447 < info->reshape_progress))
4448 goto nonew; /* No new data here */
4449 } else {
4450 /* reshape_progress is decreasing */
4451 if (__le64_to_cpu(bsb.arraystart) >=
4452 info->reshape_progress &&
4453 __le64_to_cpu(bsb.arraystart2) >=
4454 info->reshape_progress)
4455 goto nonew; /* No new data here */
4456 }
4457 }
4458 if (lseek64(fd, __le64_to_cpu(bsb.devstart)*512, 0)< 0) {
4459 second_fail:
4460 if (verbose)
4461 pr_err("Failed to verify secondary backup-metadata block on %s\n",
4462 devname);
4463 continue; /* Cannot seek */
4464 }
4465 /* There should be a duplicate backup superblock 4k before here */
4466 if (lseek64(fd, -4096, 1) < 0 ||
4467 read(fd, &bsb2, sizeof(bsb2)) != sizeof(bsb2))
4468 goto second_fail; /* Cannot find leading superblock */
4469 if (bsb.magic[15] == '1')
4470 bsbsize = offsetof(struct mdp_backup_super, pad1);
4471 else
4472 bsbsize = offsetof(struct mdp_backup_super, pad);
4473 if (memcmp(&bsb2, &bsb, bsbsize) != 0)
4474 goto second_fail; /* Cannot find leading superblock */
4475
4476 /* Now need the data offsets for all devices. */
4477 offsets = xmalloc(sizeof(*offsets)*info->array.raid_disks);
4478 for(j=0; j<info->array.raid_disks; j++) {
4479 if (fdlist[j] < 0)
4480 continue;
4481 if (st->ss->load_super(st, fdlist[j], NULL))
4482 /* FIXME should be this be an error */
4483 continue;
4484 st->ss->getinfo_super(st, &dinfo, NULL);
4485 st->ss->free_super(st);
4486 offsets[j] = dinfo.data_offset * 512;
4487 }
4488 printf(Name ": restoring critical section\n");
4489
4490 if (restore_stripes(fdlist, offsets,
4491 info->array.raid_disks,
4492 info->new_chunk,
4493 info->new_level,
4494 info->new_layout,
4495 fd, __le64_to_cpu(bsb.devstart)*512,
4496 __le64_to_cpu(bsb.arraystart)*512,
4497 __le64_to_cpu(bsb.length)*512, NULL)) {
4498 /* didn't succeed, so giveup */
4499 if (verbose)
4500 pr_err("Error restoring backup from %s\n",
4501 devname);
4502 free(offsets);
4503 return 1;
4504 }
4505
4506 if (bsb.magic[15] == '2' &&
4507 restore_stripes(fdlist, offsets,
4508 info->array.raid_disks,
4509 info->new_chunk,
4510 info->new_level,
4511 info->new_layout,
4512 fd, __le64_to_cpu(bsb.devstart)*512 +
4513 __le64_to_cpu(bsb.devstart2)*512,
4514 __le64_to_cpu(bsb.arraystart2)*512,
4515 __le64_to_cpu(bsb.length2)*512, NULL)) {
4516 /* didn't succeed, so giveup */
4517 if (verbose)
4518 pr_err("Error restoring second backup from %s\n",
4519 devname);
4520 free(offsets);
4521 return 1;
4522 }
4523
4524 free(offsets);
4525
4526 /* Ok, so the data is restored. Let's update those superblocks. */
4527
4528 lo = hi = 0;
4529 if (bsb.length) {
4530 lo = __le64_to_cpu(bsb.arraystart);
4531 hi = lo + __le64_to_cpu(bsb.length);
4532 }
4533 if (bsb.magic[15] == '2' && bsb.length2) {
4534 unsigned long long lo1, hi1;
4535 lo1 = __le64_to_cpu(bsb.arraystart2);
4536 hi1 = lo1 + __le64_to_cpu(bsb.length2);
4537 if (lo == hi) {
4538 lo = lo1;
4539 hi = hi1;
4540 } else if (lo < lo1)
4541 hi = hi1;
4542 else
4543 lo = lo1;
4544 }
4545 if (lo < hi &&
4546 (info->reshape_progress < lo ||
4547 info->reshape_progress > hi))
4548 /* backup does not affect reshape_progress*/ ;
4549 else if (info->delta_disks >= 0) {
4550 info->reshape_progress = __le64_to_cpu(bsb.arraystart) +
4551 __le64_to_cpu(bsb.length);
4552 if (bsb.magic[15] == '2') {
4553 unsigned long long p2 = __le64_to_cpu(bsb.arraystart2) +
4554 __le64_to_cpu(bsb.length2);
4555 if (p2 > info->reshape_progress)
4556 info->reshape_progress = p2;
4557 }
4558 } else {
4559 info->reshape_progress = __le64_to_cpu(bsb.arraystart);
4560 if (bsb.magic[15] == '2') {
4561 unsigned long long p2 = __le64_to_cpu(bsb.arraystart2);
4562 if (p2 < info->reshape_progress)
4563 info->reshape_progress = p2;
4564 }
4565 }
4566 for (j=0; j<info->array.raid_disks; j++) {
4567 if (fdlist[j] < 0)
4568 continue;
4569 if (st->ss->load_super(st, fdlist[j], NULL))
4570 continue;
4571 st->ss->getinfo_super(st, &dinfo, NULL);
4572 dinfo.reshape_progress = info->reshape_progress;
4573 st->ss->update_super(st, &dinfo,
4574 "_reshape_progress",
4575 NULL,0, 0, NULL);
4576 st->ss->store_super(st, fdlist[j]);
4577 st->ss->free_super(st);
4578 }
4579 return 0;
4580 }
4581 /* Didn't find any backup data, try to see if any
4582 * was needed.
4583 */
4584 if (info->delta_disks < 0) {
4585 /* When shrinking, the critical section is at the end.
4586 * So see if we are before the critical section.
4587 */
4588 unsigned long long first_block;
4589 nstripe = ostripe = 0;
4590 first_block = 0;
4591 while (ostripe >= nstripe) {
4592 ostripe += info->array.chunk_size / 512;
4593 first_block = ostripe * odata;
4594 nstripe = first_block / ndata / (info->new_chunk/512) *
4595 (info->new_chunk/512);
4596 }
4597
4598 if (info->reshape_progress >= first_block)
4599 return 0;
4600 }
4601 if (info->delta_disks > 0) {
4602 /* See if we are beyond the critical section. */
4603 unsigned long long last_block;
4604 nstripe = ostripe = 0;
4605 last_block = 0;
4606 while (nstripe >= ostripe) {
4607 nstripe += info->new_chunk / 512;
4608 last_block = nstripe * ndata;
4609 ostripe = last_block / odata / (info->array.chunk_size/512) *
4610 (info->array.chunk_size/512);
4611 }
4612
4613 if (info->reshape_progress >= last_block)
4614 return 0;
4615 }
4616 /* needed to recover critical section! */
4617 if (verbose)
4618 pr_err("Failed to find backup of critical section\n");
4619 return 1;
4620 }
4621
4622 int Grow_continue_command(char *devname, int fd,
4623 char *backup_file, int verbose)
4624 {
4625 int ret_val = 0;
4626 struct supertype *st = NULL;
4627 struct mdinfo *content = NULL;
4628 struct mdinfo array;
4629 char *subarray = NULL;
4630 struct mdinfo *cc = NULL;
4631 struct mdstat_ent *mdstat = NULL;
4632 int cfd = -1;
4633 int fd2 = -1;
4634
4635 dprintf("Grow continue from command line called for %s\n",
4636 devname);
4637
4638 st = super_by_fd(fd, &subarray);
4639 if (!st || !st->ss) {
4640 pr_err("Unable to determine metadata format for %s\n",
4641 devname);
4642 return 1;
4643 }
4644 dprintf("Grow continue is run for ");
4645 if (st->ss->external == 0) {
4646 int d;
4647 dprintf("native array (%s)\n", devname);
4648 if (ioctl(fd, GET_ARRAY_INFO, &array.array) < 0) {
4649 pr_err("%s is not an active md array -"
4650 " aborting\n", devname);
4651 ret_val = 1;
4652 goto Grow_continue_command_exit;
4653 }
4654 content = &array;
4655 /* Need to load a superblock.
4656 * FIXME we should really get what we need from
4657 * sysfs
4658 */
4659 for (d = 0; d < MAX_DISKS; d++) {
4660 mdu_disk_info_t disk;
4661 char *dv;
4662 int err;
4663 disk.number = d;
4664 if (ioctl(fd, GET_DISK_INFO, &disk) < 0)
4665 continue;
4666 if (disk.major == 0 && disk.minor == 0)
4667 continue;
4668 if ((disk.state & (1 << MD_DISK_ACTIVE)) == 0)
4669 continue;
4670 dv = map_dev(disk.major, disk.minor, 1);
4671 if (!dv)
4672 continue;
4673 fd2 = dev_open(dv, O_RDONLY);
4674 if (fd2 < 0)
4675 continue;
4676 err = st->ss->load_super(st, fd2, NULL);
4677 close(fd2);
4678 if (err)
4679 continue;
4680 break;
4681 }
4682 if (d == MAX_DISKS) {
4683 pr_err("Unable to load metadata for %s\n",
4684 devname);
4685 ret_val = 1;
4686 goto Grow_continue_command_exit;
4687 }
4688 st->ss->getinfo_super(st, content, NULL);
4689 } else {
4690 char *container;
4691
4692 if (subarray) {
4693 dprintf("subarray (%s)\n", subarray);
4694 container = st->container_devnm;
4695 cfd = open_dev_excl(st->container_devnm);
4696 } else {
4697 container = st->devnm;
4698 close(fd);
4699 cfd = open_dev_excl(st->devnm);
4700 dprintf("container (%s)\n", container);
4701 fd = cfd;
4702 }
4703 if (cfd < 0) {
4704 pr_err("Unable to open container "
4705 "for %s\n", devname);
4706 ret_val = 1;
4707 goto Grow_continue_command_exit;
4708 }
4709
4710 /* find in container array under reshape
4711 */
4712 ret_val = st->ss->load_container(st, cfd, NULL);
4713 if (ret_val) {
4714 pr_err("Cannot read superblock for %s\n",
4715 devname);
4716 ret_val = 1;
4717 goto Grow_continue_command_exit;
4718 }
4719
4720 cc = st->ss->container_content(st, subarray);
4721 for (content = cc; content ; content = content->next) {
4722 char *array;
4723 int allow_reshape = 1;
4724
4725 if (content->reshape_active == 0)
4726 continue;
4727 /* The decision about array or container wide
4728 * reshape is taken in Grow_continue based
4729 * content->reshape_active state, therefore we
4730 * need to check_reshape based on
4731 * reshape_active and subarray name
4732 */
4733 if (content->array.state & (1<<MD_SB_BLOCK_VOLUME))
4734 allow_reshape = 0;
4735 if (content->reshape_active == CONTAINER_RESHAPE &&
4736 (content->array.state
4737 & (1<<MD_SB_BLOCK_CONTAINER_RESHAPE)))
4738 allow_reshape = 0;
4739
4740 if (!allow_reshape) {
4741 pr_err("cannot continue reshape of an array"
4742 " in container with unsupported"
4743 " metadata: %s(%s)\n",
4744 devname, container);
4745 ret_val = 1;
4746 goto Grow_continue_command_exit;
4747 }
4748
4749 array = strchr(content->text_version+1, '/')+1;
4750 mdstat = mdstat_by_subdev(array, container);
4751 if (!mdstat)
4752 continue;
4753 if (mdstat->active == 0) {
4754 pr_err("Skipping inactive array %s.\n",
4755 mdstat->devnm);
4756 free_mdstat(mdstat);
4757 mdstat = NULL;
4758 continue;
4759 }
4760 break;
4761 }
4762 if (!content) {
4763 pr_err("Unable to determine reshaped "
4764 "array for %s\n", devname);
4765 ret_val = 1;
4766 goto Grow_continue_command_exit;
4767 }
4768 fd2 = open_dev(mdstat->devnm);
4769 if (fd2 < 0) {
4770 pr_err("cannot open (%s)\n", mdstat->devnm);
4771 ret_val = 1;
4772 goto Grow_continue_command_exit;
4773 }
4774
4775 sysfs_init(content, fd2, mdstat->devnm);
4776
4777 /* start mdmon in case it is not running
4778 */
4779 if (!mdmon_running(container))
4780 start_mdmon(container);
4781 ping_monitor(container);
4782
4783 if (mdmon_running(container))
4784 st->update_tail = &st->updates;
4785 else {
4786 pr_err("No mdmon found. "
4787 "Grow cannot continue.\n");
4788 ret_val = 1;
4789 goto Grow_continue_command_exit;
4790 }
4791 }
4792
4793 /* verify that array under reshape is started from
4794 * correct position
4795 */
4796 if (verify_reshape_position(content, content->array.level) < 0) {
4797 ret_val = 1;
4798 goto Grow_continue_command_exit;
4799 }
4800
4801 /* continue reshape
4802 */
4803 ret_val = Grow_continue(fd, st, content, backup_file, 0);
4804
4805 Grow_continue_command_exit:
4806 if (fd2 > -1)
4807 close(fd2);
4808 if (cfd > -1)
4809 close(cfd);
4810 st->ss->free_super(st);
4811 free_mdstat(mdstat);
4812 sysfs_free(cc);
4813 free(subarray);
4814
4815 return ret_val;
4816 }
4817
4818 int Grow_continue(int mdfd, struct supertype *st, struct mdinfo *info,
4819 char *backup_file, int freeze_reshape)
4820 {
4821 int ret_val = 2;
4822
4823 if (!info->reshape_active)
4824 return ret_val;
4825
4826 if (st->ss->external) {
4827 int cfd = open_dev(st->container_devnm);
4828
4829 if (cfd < 0)
4830 return 1;
4831
4832 st->ss->load_container(st, cfd, st->container_devnm);
4833 close(cfd);
4834 ret_val = reshape_container(st->container_devnm, NULL, mdfd,
4835 st, info, 0, backup_file,
4836 0,
4837 1 | info->reshape_active,
4838 freeze_reshape);
4839 } else
4840 ret_val = reshape_array(NULL, mdfd, "array", st, info, 1,
4841 NULL, 0ULL, backup_file, 0, 0,
4842 1 | info->reshape_active,
4843 freeze_reshape);
4844
4845 return ret_val;
4846 }
Unnamed repository; edit this file 'description' to name the repository.