]> 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
Make sure NOFILE resource limit is big enough.
[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(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 return "New chunk size does not"
1009 " divide component size";
1010 break;
1011 default:
1012 return "chunk size not meaningful for this level";
1013 }
1014 else
1015 info->new_chunk = info->array.chunk_size;
1016
1017 switch (info->array.level) {
1018 default:
1019 return "Cannot understand this RAID level";
1020 case 1:
1021 /* RAID1 can convert to RAID1 with different disks, or
1022 * raid5 with 2 disks, or
1023 * raid0 with 1 disk
1024 */
1025 if (info->new_level > 1 &&
1026 (info->component_size & 7))
1027 return "Cannot convert RAID1 of this size - "
1028 "reduce size to multiple of 4K first.";
1029 if (info->new_level == 0) {
1030 if (info->delta_disks != UnSet &&
1031 info->delta_disks != 0)
1032 return "Cannot change number of disks "
1033 "with RAID1->RAID0 conversion";
1034 re->level = 0;
1035 re->before.data_disks = 1;
1036 re->after.data_disks = 1;
1037 return NULL;
1038 }
1039 if (info->new_level == 1) {
1040 if (info->delta_disks == UnSet)
1041 /* Don't know what to do */
1042 return "no change requested for Growing RAID1";
1043 re->level = 1;
1044 return NULL;
1045 }
1046 if (info->array.raid_disks == 2 &&
1047 info->new_level == 5) {
1048
1049 re->level = 5;
1050 re->before.data_disks = 1;
1051 if (info->delta_disks != UnSet &&
1052 info->delta_disks != 0)
1053 re->after.data_disks = 1 + info->delta_disks;
1054 else
1055 re->after.data_disks = 1;
1056 if (re->after.data_disks < 1)
1057 return "Number of disks too small for RAID5";
1058
1059 re->before.layout = ALGORITHM_LEFT_SYMMETRIC;
1060 info->array.chunk_size = 65536;
1061 break;
1062 }
1063 /* Could do some multi-stage conversions, but leave that to
1064 * later.
1065 */
1066 return "Impossibly level change request for RAID1";
1067
1068 case 10:
1069 /* RAID10 can be converted from near mode to
1070 * RAID0 by removing some devices.
1071 * It can also be reshaped if the kernel supports
1072 * new_data_offset.
1073 */
1074 switch (info->new_level) {
1075 case 0:
1076 if ((info->array.layout & ~0xff) != 0x100)
1077 return "Cannot Grow RAID10 with far/offset layout";
1078 /* number of devices must be multiple of number of copies */
1079 if (info->array.raid_disks % (info->array.layout & 0xff))
1080 return "RAID10 layout too complex for Grow operation";
1081
1082 new_disks = (info->array.raid_disks
1083 / (info->array.layout & 0xff));
1084 if (info->delta_disks == UnSet)
1085 info->delta_disks = (new_disks
1086 - info->array.raid_disks);
1087
1088 if (info->delta_disks != new_disks - info->array.raid_disks)
1089 return "New number of raid-devices impossible for RAID10";
1090 if (info->new_chunk &&
1091 info->new_chunk != info->array.chunk_size)
1092 return "Cannot change chunk-size with RAID10 Grow";
1093
1094 /* looks good */
1095 re->level = 0;
1096 re->before.data_disks = new_disks;
1097 re->after.data_disks = re->before.data_disks;
1098 return NULL;
1099
1100 case 10:
1101 near = info->array.layout & 0xff;
1102 far = (info->array.layout >> 8) & 0xff;
1103 offset = info->array.layout & 0x10000;
1104 if (far > 1 && !offset)
1105 return "Cannot reshape RAID10 in far-mode";
1106 copies = near * far;
1107
1108 old_chunk = info->array.chunk_size * far;
1109
1110 if (info->new_layout == UnSet)
1111 info->new_layout = info->array.layout;
1112 else {
1113 near = info->new_layout & 0xff;
1114 far = (info->new_layout >> 8) & 0xff;
1115 offset = info->new_layout & 0x10000;
1116 if (far > 1 && !offset)
1117 return "Cannot reshape RAID10 to far-mode";
1118 if (near * far != copies)
1119 return "Cannot change number of copies"
1120 " when reshaping RAID10";
1121 }
1122 if (info->delta_disks == UnSet)
1123 info->delta_disks = 0;
1124 new_disks = (info->array.raid_disks +
1125 info->delta_disks);
1126
1127 new_chunk = info->new_chunk * far;
1128
1129 re->level = 10;
1130 re->before.layout = info->array.layout;
1131 re->before.data_disks = info->array.raid_disks;
1132 re->after.layout = info->new_layout;
1133 re->after.data_disks = new_disks;
1134 /* For RAID10 we don't do backup but do allow reshape,
1135 * so set backup_blocks to INVALID_SECTORS rather than
1136 * zero.
1137 * And there is no need to synchronise stripes on both
1138 * 'old' and 'new'. So the important
1139 * number is the minimum data_offset difference
1140 * which is the larger of (offset copies * chunk).
1141 */
1142 re->backup_blocks = INVALID_SECTORS;
1143 re->min_offset_change = max(old_chunk, new_chunk) / 512;
1144 if (new_disks < re->before.data_disks &&
1145 info->space_after < re->min_offset_change)
1146 /* Reduce component size by one chunk */
1147 re->new_size = (info->component_size -
1148 re->min_offset_change);
1149 else
1150 re->new_size = info->component_size;
1151 re->new_size = re->new_size * new_disks / copies;
1152 return NULL;
1153
1154 default:
1155 return "RAID10 can only be changed to RAID0";
1156 }
1157 case 0:
1158 /* RAID0 can be converted to RAID10, or to RAID456 */
1159 if (info->new_level == 10) {
1160 if (info->new_layout == UnSet && info->delta_disks == UnSet) {
1161 /* Assume near=2 layout */
1162 info->new_layout = 0x102;
1163 info->delta_disks = info->array.raid_disks;
1164 }
1165 if (info->new_layout == UnSet) {
1166 int copies = 1 + (info->delta_disks
1167 / info->array.raid_disks);
1168 if (info->array.raid_disks * (copies-1)
1169 != info->delta_disks)
1170 return "Impossible number of devices"
1171 " for RAID0->RAID10";
1172 info->new_layout = 0x100 + copies;
1173 }
1174 if (info->delta_disks == UnSet) {
1175 int copies = info->new_layout & 0xff;
1176 if (info->new_layout != 0x100 + copies)
1177 return "New layout impossible"
1178 " for RAID0->RAID10";;
1179 info->delta_disks = (copies - 1) *
1180 info->array.raid_disks;
1181 }
1182 if (info->new_chunk &&
1183 info->new_chunk != info->array.chunk_size)
1184 return "Cannot change chunk-size with RAID0->RAID10";
1185 /* looks good */
1186 re->level = 10;
1187 re->before.data_disks = (info->array.raid_disks +
1188 info->delta_disks);
1189 re->after.data_disks = re->before.data_disks;
1190 re->before.layout = info->new_layout;
1191 return NULL;
1192 }
1193
1194 /* RAID0 can also covert to RAID0/4/5/6 by first converting to
1195 * a raid4 style layout of the final level.
1196 */
1197 switch (info->new_level) {
1198 case 4:
1199 delta_parity = 1;
1200 case 0:
1201 re->level = 4;
1202 re->before.layout = 0;
1203 break;
1204 case 5:
1205 delta_parity = 1;
1206 re->level = 5;
1207 re->before.layout = ALGORITHM_PARITY_N;
1208 break;
1209 case 6:
1210 delta_parity = 2;
1211 re->level = 6;
1212 re->before.layout = ALGORITHM_PARITY_N;
1213 break;
1214 default:
1215 return "Impossible level change requested";
1216 }
1217 re->before.data_disks = info->array.raid_disks;
1218 /* determining 'after' layout happens outside this 'switch' */
1219 break;
1220
1221 case 4:
1222 info->array.layout = ALGORITHM_PARITY_N;
1223 case 5:
1224 switch (info->new_level) {
1225 case 0:
1226 delta_parity = -1;
1227 case 4:
1228 re->level = info->array.level;
1229 re->before.data_disks = info->array.raid_disks - 1;
1230 re->before.layout = info->array.layout;
1231 break;
1232 case 5:
1233 re->level = 5;
1234 re->before.data_disks = info->array.raid_disks - 1;
1235 re->before.layout = info->array.layout;
1236 break;
1237 case 6:
1238 delta_parity = 1;
1239 re->level = 6;
1240 re->before.data_disks = info->array.raid_disks - 1;
1241 switch (info->array.layout) {
1242 case ALGORITHM_LEFT_ASYMMETRIC:
1243 re->before.layout = ALGORITHM_LEFT_ASYMMETRIC_6;
1244 break;
1245 case ALGORITHM_RIGHT_ASYMMETRIC:
1246 re->before.layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
1247 break;
1248 case ALGORITHM_LEFT_SYMMETRIC:
1249 re->before.layout = ALGORITHM_LEFT_SYMMETRIC_6;
1250 break;
1251 case ALGORITHM_RIGHT_SYMMETRIC:
1252 re->before.layout = ALGORITHM_RIGHT_SYMMETRIC_6;
1253 break;
1254 case ALGORITHM_PARITY_0:
1255 re->before.layout = ALGORITHM_PARITY_0_6;
1256 break;
1257 case ALGORITHM_PARITY_N:
1258 re->before.layout = ALGORITHM_PARITY_N_6;
1259 break;
1260 default:
1261 return "Cannot convert an array with this layout";
1262 }
1263 break;
1264 case 1:
1265 if (info->array.raid_disks != 2)
1266 return "Can only convert a 2-device array to RAID1";
1267 if (info->delta_disks != UnSet &&
1268 info->delta_disks != 0)
1269 return "Cannot set raid_disk when "
1270 "converting RAID5->RAID1";
1271 re->level = 1;
1272 info->new_chunk = 0;
1273 return NULL;
1274 default:
1275 return "Impossible level change requested";
1276 }
1277 break;
1278 case 6:
1279 switch (info->new_level) {
1280 case 4:
1281 case 5:
1282 delta_parity = -1;
1283 case 6:
1284 re->level = 6;
1285 re->before.data_disks = info->array.raid_disks - 2;
1286 re->before.layout = info->array.layout;
1287 break;
1288 default:
1289 return "Impossible level change requested";
1290 }
1291 break;
1292 }
1293
1294 /* If we reached here then it looks like a re-stripe is
1295 * happening. We have determined the intermediate level
1296 * and initial raid_disks/layout and stored these in 're'.
1297 *
1298 * We need to deduce the final layout that can be atomically
1299 * converted to the end state.
1300 */
1301 switch (info->new_level) {
1302 case 0:
1303 /* We can only get to RAID0 from RAID4 or RAID5
1304 * with appropriate layout and one extra device
1305 */
1306 if (re->level != 4 && re->level != 5)
1307 return "Cannot covert to RAID0 from this level";
1308
1309 switch (re->level) {
1310 case 4:
1311 re->before.layout = 0;
1312 re->after.layout = 0;
1313 break;
1314 case 5:
1315 re->after.layout = ALGORITHM_PARITY_N;
1316 break;
1317 }
1318 break;
1319
1320 case 4:
1321 /* We can only get to RAID4 from RAID5 */
1322 if (re->level != 4 && re->level != 5)
1323 return "Cannot convert to RAID4 from this level";
1324
1325 switch (re->level) {
1326 case 4:
1327 re->before.layout = 0;
1328 re->after.layout = 0;
1329 break;
1330 case 5:
1331 re->after.layout = ALGORITHM_PARITY_N;
1332 break;
1333 }
1334 break;
1335
1336 case 5:
1337 /* We get to RAID5 from RAID5 or RAID6 */
1338 if (re->level != 5 && re->level != 6)
1339 return "Cannot convert to RAID5 from this level";
1340
1341 switch (re->level) {
1342 case 5:
1343 if (info->new_layout == UnSet)
1344 re->after.layout = re->before.layout;
1345 else
1346 re->after.layout = info->new_layout;
1347 break;
1348 case 6:
1349 if (info->new_layout == UnSet)
1350 info->new_layout = re->before.layout;
1351
1352 /* after.layout needs to be raid6 version of new_layout */
1353 if (info->new_layout == ALGORITHM_PARITY_N)
1354 re->after.layout = ALGORITHM_PARITY_N;
1355 else {
1356 char layout[40];
1357 char *ls = map_num(r5layout, info->new_layout);
1358 int l;
1359 if (ls) {
1360 /* Current RAID6 layout has a RAID5
1361 * equivalent - good
1362 */
1363 strcat(strcpy(layout, ls), "-6");
1364 l = map_name(r6layout, layout);
1365 if (l == UnSet)
1366 return "Cannot find RAID6 layout"
1367 " to convert to";
1368 } else {
1369 /* Current RAID6 has no equivalent.
1370 * If it is already a '-6' layout we
1371 * can leave it unchanged, else we must
1372 * fail
1373 */
1374 ls = map_num(r6layout, info->new_layout);
1375 if (!ls ||
1376 strcmp(ls+strlen(ls)-2, "-6") != 0)
1377 return "Please specify new layout";
1378 l = info->new_layout;
1379 }
1380 re->after.layout = l;
1381 }
1382 }
1383 break;
1384
1385 case 6:
1386 /* We must already be at level 6 */
1387 if (re->level != 6)
1388 return "Impossible level change";
1389 if (info->new_layout == UnSet)
1390 re->after.layout = info->array.layout;
1391 else
1392 re->after.layout = info->new_layout;
1393 break;
1394 default:
1395 return "Impossible level change requested";
1396 }
1397 if (info->delta_disks == UnSet)
1398 info->delta_disks = delta_parity;
1399
1400 re->after.data_disks = (re->before.data_disks
1401 + info->delta_disks
1402 - delta_parity);
1403 switch (re->level) {
1404 case 6: re->parity = 2;
1405 break;
1406 case 4:
1407 case 5: re->parity = 1;
1408 break;
1409 default: re->parity = 0;
1410 break;
1411 }
1412 /* So we have a restripe operation, we need to calculate the number
1413 * of blocks per reshape operation.
1414 */
1415 if (info->new_chunk == 0)
1416 info->new_chunk = info->array.chunk_size;
1417 if (re->after.data_disks == re->before.data_disks &&
1418 re->after.layout == re->before.layout &&
1419 info->new_chunk == info->array.chunk_size) {
1420 /* Nothing to change, can change level immediately. */
1421 re->level = info->new_level;
1422 re->backup_blocks = 0;
1423 return NULL;
1424 }
1425 if (re->after.data_disks == 1 && re->before.data_disks == 1) {
1426 /* chunk and layout changes make no difference */
1427 re->level = info->new_level;
1428 re->backup_blocks = 0;
1429 return NULL;
1430 }
1431
1432 if (re->after.data_disks == re->before.data_disks &&
1433 get_linux_version() < 2006032)
1434 return "in-place reshape is not safe before 2.6.32 - sorry.";
1435
1436 if (re->after.data_disks < re->before.data_disks &&
1437 get_linux_version() < 2006030)
1438 return "reshape to fewer devices is not supported before 2.6.30 - sorry.";
1439
1440 re->backup_blocks = compute_backup_blocks(
1441 info->new_chunk, info->array.chunk_size,
1442 re->after.data_disks,
1443 re->before.data_disks);
1444 re->min_offset_change = re->backup_blocks / re->before.data_disks;
1445
1446 re->new_size = info->component_size * re->after.data_disks;
1447 return NULL;
1448 }
1449
1450 static int set_array_size(struct supertype *st, struct mdinfo *sra,
1451 char *text_version)
1452 {
1453 struct mdinfo *info;
1454 char *subarray;
1455 int ret_val = -1;
1456
1457 if ((st == NULL) || (sra == NULL))
1458 return ret_val;
1459
1460 if (text_version == NULL)
1461 text_version = sra->text_version;
1462 subarray = strchr(text_version+1, '/')+1;
1463 info = st->ss->container_content(st, subarray);
1464 if (info) {
1465 unsigned long long current_size = 0;
1466 unsigned long long new_size =
1467 info->custom_array_size/2;
1468
1469 if (sysfs_get_ll(sra, NULL, "array_size", &current_size) == 0 &&
1470 new_size > current_size) {
1471 if (sysfs_set_num(sra, NULL, "array_size", new_size)
1472 < 0)
1473 dprintf("Error: Cannot set array size");
1474 else {
1475 ret_val = 0;
1476 dprintf("Array size changed");
1477 }
1478 dprintf(" from %llu to %llu.\n",
1479 current_size, new_size);
1480 }
1481 sysfs_free(info);
1482 } else
1483 dprintf("Error: set_array_size(): info pointer in NULL\n");
1484
1485 return ret_val;
1486 }
1487
1488 static int reshape_array(char *container, int fd, char *devname,
1489 struct supertype *st, struct mdinfo *info,
1490 int force, struct mddev_dev *devlist,
1491 unsigned long long data_offset,
1492 char *backup_file, int verbose, int forked,
1493 int restart, int freeze_reshape);
1494 static int reshape_container(char *container, char *devname,
1495 int mdfd,
1496 struct supertype *st,
1497 struct mdinfo *info,
1498 int force,
1499 char *backup_file,
1500 int verbose, int restart, int freeze_reshape);
1501
1502 int Grow_reshape(char *devname, int fd,
1503 struct mddev_dev *devlist,
1504 unsigned long long data_offset,
1505 struct context *c, struct shape *s)
1506 {
1507 /* Make some changes in the shape of an array.
1508 * The kernel must support the change.
1509 *
1510 * There are three different changes. Each can trigger
1511 * a resync or recovery so we freeze that until we have
1512 * requested everything (if kernel supports freezing - 2.6.30).
1513 * The steps are:
1514 * - change size (i.e. component_size)
1515 * - change level
1516 * - change layout/chunksize/ndisks
1517 *
1518 * The last can require a reshape. It is different on different
1519 * levels so we need to check the level before actioning it.
1520 * Some times the level change needs to be requested after the
1521 * reshape (e.g. raid6->raid5, raid5->raid0)
1522 *
1523 */
1524 struct mdu_array_info_s array;
1525 int rv = 0;
1526 struct supertype *st;
1527 char *subarray = NULL;
1528
1529 int frozen;
1530 int changed = 0;
1531 char *container = NULL;
1532 int cfd = -1;
1533
1534 struct mddev_dev *dv;
1535 int added_disks;
1536
1537 struct mdinfo info;
1538 struct mdinfo *sra;
1539
1540
1541 if (ioctl(fd, GET_ARRAY_INFO, &array) < 0) {
1542 pr_err("%s is not an active md array - aborting\n",
1543 devname);
1544 return 1;
1545 }
1546 if (data_offset != INVALID_SECTORS && array.level != 10
1547 && (array.level < 4 || array.level > 6)) {
1548 pr_err("--grow --data-offset not yet supported\n");
1549 return 1;
1550 }
1551
1552 if (s->size > 0 &&
1553 (s->chunk || s->level!= UnSet || s->layout_str || s->raiddisks)) {
1554 pr_err("cannot change component size at the same time "
1555 "as other changes.\n"
1556 " Change size first, then check data is intact before "
1557 "making other changes.\n");
1558 return 1;
1559 }
1560
1561 if (s->raiddisks && s->raiddisks < array.raid_disks && array.level > 1 &&
1562 get_linux_version() < 2006032 &&
1563 !check_env("MDADM_FORCE_FEWER")) {
1564 pr_err("reducing the number of devices is not safe before Linux 2.6.32\n"
1565 " Please use a newer kernel\n");
1566 return 1;
1567 }
1568
1569 st = super_by_fd(fd, &subarray);
1570 if (!st) {
1571 pr_err("Unable to determine metadata format for %s\n", devname);
1572 return 1;
1573 }
1574 if (s->raiddisks > st->max_devs) {
1575 pr_err("Cannot increase raid-disks on this array"
1576 " beyond %d\n", st->max_devs);
1577 return 1;
1578 }
1579
1580 /* in the external case we need to check that the requested reshape is
1581 * supported, and perform an initial check that the container holds the
1582 * pre-requisite spare devices (mdmon owns final validation)
1583 */
1584 if (st->ss->external) {
1585 int rv;
1586
1587 if (subarray) {
1588 container = st->container_devnm;
1589 cfd = open_dev_excl(st->container_devnm);
1590 } else {
1591 container = st->devnm;
1592 close(fd);
1593 cfd = open_dev_excl(st->devnm);
1594 fd = cfd;
1595 }
1596 if (cfd < 0) {
1597 pr_err("Unable to open container for %s\n",
1598 devname);
1599 free(subarray);
1600 return 1;
1601 }
1602
1603 rv = st->ss->load_container(st, cfd, NULL);
1604
1605 if (rv) {
1606 pr_err("Cannot read superblock for %s\n",
1607 devname);
1608 free(subarray);
1609 return 1;
1610 }
1611
1612 /* check if operation is supported for metadata handler */
1613 if (st->ss->container_content) {
1614 struct mdinfo *cc = NULL;
1615 struct mdinfo *content = NULL;
1616
1617 cc = st->ss->container_content(st, subarray);
1618 for (content = cc; content ; content = content->next) {
1619 int allow_reshape = 1;
1620
1621 /* check if reshape is allowed based on metadata
1622 * indications stored in content.array.status
1623 */
1624 if (content->array.state & (1<<MD_SB_BLOCK_VOLUME))
1625 allow_reshape = 0;
1626 if (content->array.state
1627 & (1<<MD_SB_BLOCK_CONTAINER_RESHAPE))
1628 allow_reshape = 0;
1629 if (!allow_reshape) {
1630 pr_err("cannot reshape arrays in"
1631 " container with unsupported"
1632 " metadata: %s(%s)\n",
1633 devname, container);
1634 sysfs_free(cc);
1635 free(subarray);
1636 return 1;
1637 }
1638 }
1639 sysfs_free(cc);
1640 }
1641 if (mdmon_running(container))
1642 st->update_tail = &st->updates;
1643 }
1644
1645 added_disks = 0;
1646 for (dv = devlist; dv; dv = dv->next)
1647 added_disks++;
1648 if (s->raiddisks > array.raid_disks &&
1649 array.spare_disks +added_disks < (s->raiddisks - array.raid_disks) &&
1650 !c->force) {
1651 pr_err("Need %d spare%s to avoid degraded array,"
1652 " and only have %d.\n"
1653 " Use --force to over-ride this check.\n",
1654 s->raiddisks - array.raid_disks,
1655 s->raiddisks - array.raid_disks == 1 ? "" : "s",
1656 array.spare_disks + added_disks);
1657 return 1;
1658 }
1659
1660 sra = sysfs_read(fd, NULL, GET_LEVEL | GET_DISKS | GET_DEVS
1661 | GET_STATE | GET_VERSION);
1662 if (sra) {
1663 if (st->ss->external && subarray == NULL) {
1664 array.level = LEVEL_CONTAINER;
1665 sra->array.level = LEVEL_CONTAINER;
1666 }
1667 } else {
1668 pr_err("failed to read sysfs parameters for %s\n",
1669 devname);
1670 return 1;
1671 }
1672 frozen = freeze(st);
1673 if (frozen < -1) {
1674 /* freeze() already spewed the reason */
1675 sysfs_free(sra);
1676 return 1;
1677 } else if (frozen < 0) {
1678 pr_err("%s is performing resync/recovery and cannot"
1679 " be reshaped\n", devname);
1680 sysfs_free(sra);
1681 return 1;
1682 }
1683
1684 /* ========= set size =============== */
1685 if (s->size > 0 && (s->size == MAX_SIZE || s->size != (unsigned)array.size)) {
1686 unsigned long long orig_size = get_component_size(fd)/2;
1687 unsigned long long min_csize;
1688 struct mdinfo *mdi;
1689 int raid0_takeover = 0;
1690
1691 if (orig_size == 0)
1692 orig_size = (unsigned) array.size;
1693
1694 if (orig_size == 0) {
1695 pr_err("Cannot set device size in this type of array.\n");
1696 rv = 1;
1697 goto release;
1698 }
1699
1700 if (reshape_super(st, s->size, UnSet, UnSet, 0, 0, UnSet, NULL,
1701 devname, APPLY_METADATA_CHANGES, c->verbose > 0)) {
1702 rv = 1;
1703 goto release;
1704 }
1705 sync_metadata(st);
1706 if (st->ss->external) {
1707 /* metadata can have size limitation
1708 * update size value according to metadata information
1709 */
1710 struct mdinfo *sizeinfo =
1711 st->ss->container_content(st, subarray);
1712 if (sizeinfo) {
1713 unsigned long long new_size =
1714 sizeinfo->custom_array_size/2;
1715 int data_disks = get_data_disks(
1716 sizeinfo->array.level,
1717 sizeinfo->array.layout,
1718 sizeinfo->array.raid_disks);
1719 new_size /= data_disks;
1720 dprintf("Metadata size correction from %llu to "
1721 "%llu (%llu)\n", orig_size, new_size,
1722 new_size * data_disks);
1723 s->size = new_size;
1724 sysfs_free(sizeinfo);
1725 }
1726 }
1727
1728 /* Update the size of each member device in case
1729 * they have been resized. This will never reduce
1730 * below the current used-size. The "size" attribute
1731 * understands '0' to mean 'max'.
1732 */
1733 min_csize = 0;
1734 rv = 0;
1735 for (mdi = sra->devs; mdi; mdi = mdi->next) {
1736 if (sysfs_set_num(sra, mdi, "size",
1737 s->size == MAX_SIZE ? 0 : s->size) < 0) {
1738 /* Probably kernel refusing to let us
1739 * reduce the size - not an error.
1740 */
1741 break;
1742 }
1743 if (array.not_persistent == 0 &&
1744 array.major_version == 0 &&
1745 get_linux_version() < 3001000) {
1746 /* Dangerous to allow size to exceed 2TB */
1747 unsigned long long csize;
1748 if (sysfs_get_ll(sra, mdi, "size", &csize) == 0) {
1749 if (csize >= 2ULL*1024*1024*1024)
1750 csize = 2ULL*1024*1024*1024;
1751 if ((min_csize == 0 || (min_csize
1752 > csize)))
1753 min_csize = csize;
1754 }
1755 }
1756 }
1757 if (rv) {
1758 pr_err("Cannot set size on "
1759 "array members.\n");
1760 goto size_change_error;
1761 }
1762 if (min_csize && s->size > min_csize) {
1763 pr_err("Cannot safely make this array "
1764 "use more than 2TB per device on this kernel.\n");
1765 rv = 1;
1766 goto size_change_error;
1767 }
1768 if (min_csize && s->size == MAX_SIZE) {
1769 /* Don't let the kernel choose a size - it will get
1770 * it wrong
1771 */
1772 pr_err("Limited v0.90 array to "
1773 "2TB per device\n");
1774 s->size = min_csize;
1775 }
1776 if (st->ss->external) {
1777 if (sra->array.level == 0) {
1778 rv = sysfs_set_str(sra, NULL, "level",
1779 "raid5");
1780 if (!rv) {
1781 raid0_takeover = 1;
1782 /* get array parametes after takeover
1783 * to chane one parameter at time only
1784 */
1785 rv = ioctl(fd, GET_ARRAY_INFO, &array);
1786 }
1787 }
1788 /* make sure mdmon is
1789 * aware of the new level */
1790 if (!mdmon_running(st->container_devnm))
1791 start_mdmon(st->container_devnm);
1792 ping_monitor(container);
1793 if (mdmon_running(st->container_devnm) &&
1794 st->update_tail == NULL)
1795 st->update_tail = &st->updates;
1796 }
1797
1798 if (s->size == MAX_SIZE)
1799 s->size = 0;
1800 array.size = s->size;
1801 if ((unsigned)array.size != s->size) {
1802 /* got truncated to 32bit, write to
1803 * component_size instead
1804 */
1805 if (sra)
1806 rv = sysfs_set_num(sra, NULL,
1807 "component_size", s->size);
1808 else
1809 rv = -1;
1810 } else {
1811 rv = ioctl(fd, SET_ARRAY_INFO, &array);
1812
1813 /* manage array size when it is managed externally
1814 */
1815 if ((rv == 0) && st->ss->external)
1816 rv = set_array_size(st, sra, sra->text_version);
1817 }
1818
1819 if (raid0_takeover) {
1820 /* do not recync non-existing parity,
1821 * we will drop it anyway
1822 */
1823 sysfs_set_str(sra, NULL, "sync_action", "frozen");
1824 /* go back to raid0, drop parity disk
1825 */
1826 sysfs_set_str(sra, NULL, "level", "raid0");
1827 ioctl(fd, GET_ARRAY_INFO, &array);
1828 }
1829
1830 size_change_error:
1831 if (rv != 0) {
1832 int err = errno;
1833
1834 /* restore metadata */
1835 if (reshape_super(st, orig_size, UnSet, UnSet, 0, 0,
1836 UnSet, NULL, devname,
1837 ROLLBACK_METADATA_CHANGES,
1838 c->verbose) == 0)
1839 sync_metadata(st);
1840 pr_err("Cannot set device size for %s: %s\n",
1841 devname, strerror(err));
1842 if (err == EBUSY &&
1843 (array.state & (1<<MD_SB_BITMAP_PRESENT)))
1844 cont_err("Bitmap must be removed before size can be changed\n");
1845 rv = 1;
1846 goto release;
1847 }
1848 if (s->assume_clean) {
1849 /* This will fail on kernels older than 3.0 unless
1850 * a backport has been arranged.
1851 */
1852 if (sra == NULL ||
1853 sysfs_set_str(sra, NULL, "resync_start", "none") < 0)
1854 pr_err("--assume-clean not supported with --grow on this kernel\n");
1855 }
1856 ioctl(fd, GET_ARRAY_INFO, &array);
1857 s->size = get_component_size(fd)/2;
1858 if (s->size == 0)
1859 s->size = array.size;
1860 if (c->verbose >= 0) {
1861 if (s->size == orig_size)
1862 pr_err("component size of %s "
1863 "unchanged at %lluK\n",
1864 devname, s->size);
1865 else
1866 pr_err("component size of %s "
1867 "has been set to %lluK\n",
1868 devname, s->size);
1869 }
1870 changed = 1;
1871 } else if (array.level != LEVEL_CONTAINER) {
1872 s->size = get_component_size(fd)/2;
1873 if (s->size == 0)
1874 s->size = array.size;
1875 }
1876
1877 /* See if there is anything else to do */
1878 if ((s->level == UnSet || s->level == array.level) &&
1879 (s->layout_str == NULL) &&
1880 (s->chunk == 0 || s->chunk == array.chunk_size) &&
1881 data_offset == INVALID_SECTORS &&
1882 (s->raiddisks == 0 || s->raiddisks == array.raid_disks)) {
1883 /* Nothing more to do */
1884 if (!changed && c->verbose >= 0)
1885 pr_err("%s: no change requested\n",
1886 devname);
1887 goto release;
1888 }
1889
1890 /* ========= check for Raid10/Raid1 -> Raid0 conversion ===============
1891 * current implementation assumes that following conditions must be met:
1892 * - RAID10:
1893 * - far_copies == 1
1894 * - near_copies == 2
1895 */
1896 if ((s->level == 0 && array.level == 10 && sra &&
1897 array.layout == ((1 << 8) + 2) && !(array.raid_disks & 1)) ||
1898 (s->level == 0 && array.level == 1 && sra)) {
1899 int err;
1900 err = remove_disks_for_takeover(st, sra, array.layout);
1901 if (err) {
1902 dprintf(Name": Array cannot be reshaped\n");
1903 if (cfd > -1)
1904 close(cfd);
1905 rv = 1;
1906 goto release;
1907 }
1908 /* Make sure mdmon has seen the device removal
1909 * and updated metadata before we continue with
1910 * level change
1911 */
1912 if (container)
1913 ping_monitor(container);
1914 }
1915
1916 memset(&info, 0, sizeof(info));
1917 info.array = array;
1918 sysfs_init(&info, fd, NULL);
1919 strcpy(info.text_version, sra->text_version);
1920 info.component_size = s->size*2;
1921 info.new_level = s->level;
1922 info.new_chunk = s->chunk * 1024;
1923 if (info.array.level == LEVEL_CONTAINER) {
1924 info.delta_disks = UnSet;
1925 info.array.raid_disks = s->raiddisks;
1926 } else if (s->raiddisks)
1927 info.delta_disks = s->raiddisks - info.array.raid_disks;
1928 else
1929 info.delta_disks = UnSet;
1930 if (s->layout_str == NULL) {
1931 info.new_layout = UnSet;
1932 if (info.array.level == 6 &&
1933 (info.new_level == 6 || info.new_level == UnSet) &&
1934 info.array.layout >= 16) {
1935 pr_err("%s has a non-standard layout. If you"
1936 " wish to preserve this\n", devname);
1937 cont_err("during the reshape, please specify"
1938 " --layout=preserve\n");
1939 cont_err("If you want to change it, specify a"
1940 " layout or use --layout=normalise\n");
1941 rv = 1;
1942 goto release;
1943 }
1944 } else if (strcmp(s->layout_str, "normalise") == 0 ||
1945 strcmp(s->layout_str, "normalize") == 0) {
1946 /* If we have a -6 RAID6 layout, remove the '-6'. */
1947 info.new_layout = UnSet;
1948 if (info.array.level == 6 && info.new_level == UnSet) {
1949 char l[40], *h;
1950 strcpy(l, map_num(r6layout, info.array.layout));
1951 h = strrchr(l, '-');
1952 if (h && strcmp(h, "-6") == 0) {
1953 *h = 0;
1954 info.new_layout = map_name(r6layout, l);
1955 }
1956 } else {
1957 pr_err("%s is only meaningful when reshaping"
1958 " a RAID6 array.\n", s->layout_str);
1959 rv = 1;
1960 goto release;
1961 }
1962 } else if (strcmp(s->layout_str, "preserve") == 0) {
1963 /* This means that a non-standard RAID6 layout
1964 * is OK.
1965 * In particular:
1966 * - When reshape a RAID6 (e.g. adding a device)
1967 * which is in a non-standard layout, it is OK
1968 * to preserve that layout.
1969 * - When converting a RAID5 to RAID6, leave it in
1970 * the XXX-6 layout, don't re-layout.
1971 */
1972 if (info.array.level == 6 && info.new_level == UnSet)
1973 info.new_layout = info.array.layout;
1974 else if (info.array.level == 5 && info.new_level == 6) {
1975 char l[40];
1976 strcpy(l, map_num(r5layout, info.array.layout));
1977 strcat(l, "-6");
1978 info.new_layout = map_name(r6layout, l);
1979 } else {
1980 pr_err("%s in only meaningful when reshaping"
1981 " to RAID6\n", s->layout_str);
1982 rv = 1;
1983 goto release;
1984 }
1985 } else {
1986 int l = info.new_level;
1987 if (l == UnSet)
1988 l = info.array.level;
1989 switch (l) {
1990 case 5:
1991 info.new_layout = map_name(r5layout, s->layout_str);
1992 break;
1993 case 6:
1994 info.new_layout = map_name(r6layout, s->layout_str);
1995 break;
1996 case 10:
1997 info.new_layout = parse_layout_10(s->layout_str);
1998 break;
1999 case LEVEL_FAULTY:
2000 info.new_layout = parse_layout_faulty(s->layout_str);
2001 break;
2002 default:
2003 pr_err("layout not meaningful"
2004 " with this level\n");
2005 rv = 1;
2006 goto release;
2007 }
2008 if (info.new_layout == UnSet) {
2009 pr_err("layout %s not understood"
2010 " for this level\n",
2011 s->layout_str);
2012 rv = 1;
2013 goto release;
2014 }
2015 }
2016
2017 if (array.level == LEVEL_FAULTY) {
2018 if (s->level != UnSet && s->level != array.level) {
2019 pr_err("cannot change level of Faulty device\n");
2020 rv =1 ;
2021 }
2022 if (s->chunk) {
2023 pr_err("cannot set chunksize of Faulty device\n");
2024 rv =1 ;
2025 }
2026 if (s->raiddisks && s->raiddisks != 1) {
2027 pr_err("cannot set raid_disks of Faulty device\n");
2028 rv =1 ;
2029 }
2030 if (s->layout_str) {
2031 if (ioctl(fd, GET_ARRAY_INFO, &array) != 0) {
2032 dprintf("Cannot get array information.\n");
2033 goto release;
2034 }
2035 array.layout = info.new_layout;
2036 if (ioctl(fd, SET_ARRAY_INFO, &array) != 0) {
2037 pr_err("failed to set new layout\n");
2038 rv = 1;
2039 } else if (c->verbose >= 0)
2040 printf("layout for %s set to %d\n",
2041 devname, array.layout);
2042 }
2043 } else if (array.level == LEVEL_CONTAINER) {
2044 /* This change is to be applied to every array in the
2045 * container. This is only needed when the metadata imposes
2046 * restraints of the various arrays in the container.
2047 * Currently we only know that IMSM requires all arrays
2048 * to have the same number of devices so changing the
2049 * number of devices (On-Line Capacity Expansion) must be
2050 * performed at the level of the container
2051 */
2052 rv = reshape_container(container, devname, -1, st, &info,
2053 c->force, c->backup_file, c->verbose, 0, 0);
2054 frozen = 0;
2055 } else {
2056 /* get spare devices from external metadata
2057 */
2058 if (st->ss->external) {
2059 struct mdinfo *info2;
2060
2061 info2 = st->ss->container_content(st, subarray);
2062 if (info2) {
2063 info.array.spare_disks =
2064 info2->array.spare_disks;
2065 sysfs_free(info2);
2066 }
2067 }
2068
2069 /* Impose these changes on a single array. First
2070 * check that the metadata is OK with the change. */
2071
2072 if (reshape_super(st, 0, info.new_level,
2073 info.new_layout, info.new_chunk,
2074 info.array.raid_disks, info.delta_disks,
2075 c->backup_file, devname, APPLY_METADATA_CHANGES,
2076 c->verbose)) {
2077 rv = 1;
2078 goto release;
2079 }
2080 sync_metadata(st);
2081 rv = reshape_array(container, fd, devname, st, &info, c->force,
2082 devlist, data_offset, c->backup_file, c->verbose,
2083 0, 0, 0);
2084 frozen = 0;
2085 }
2086 release:
2087 sysfs_free(sra);
2088 if (frozen > 0)
2089 unfreeze(st);
2090 return rv;
2091 }
2092
2093 /* verify_reshape_position()
2094 * Function checks if reshape position in metadata is not farther
2095 * than position in md.
2096 * Return value:
2097 * 0 : not valid sysfs entry
2098 * it can be caused by not started reshape, it should be started
2099 * by reshape array or raid0 array is before takeover
2100 * -1 : error, reshape position is obviously wrong
2101 * 1 : success, reshape progress correct or updated
2102 */
2103 static int verify_reshape_position(struct mdinfo *info, int level)
2104 {
2105 int ret_val = 0;
2106 char buf[40];
2107 int rv;
2108
2109 /* read sync_max, failure can mean raid0 array */
2110 rv = sysfs_get_str(info, NULL, "sync_max", buf, 40);
2111
2112 if (rv > 0) {
2113 char *ep;
2114 unsigned long long position = strtoull(buf, &ep, 0);
2115
2116 dprintf(Name": Read sync_max sysfs entry is: %s\n", buf);
2117 if (!(ep == buf || (*ep != 0 && *ep != '\n' && *ep != ' '))) {
2118 position *= get_data_disks(level,
2119 info->new_layout,
2120 info->array.raid_disks);
2121 if (info->reshape_progress < position) {
2122 dprintf("Corrected reshape progress (%llu) to "
2123 "md position (%llu)\n",
2124 info->reshape_progress, position);
2125 info->reshape_progress = position;
2126 ret_val = 1;
2127 } else if (info->reshape_progress > position) {
2128 pr_err("Fatal error: array "
2129 "reshape was not properly frozen "
2130 "(expected reshape position is %llu, "
2131 "but reshape progress is %llu.\n",
2132 position, info->reshape_progress);
2133 ret_val = -1;
2134 } else {
2135 dprintf("Reshape position in md and metadata "
2136 "are the same;");
2137 ret_val = 1;
2138 }
2139 }
2140 } else if (rv == 0) {
2141 /* for valid sysfs entry, 0-length content
2142 * should be indicated as error
2143 */
2144 ret_val = -1;
2145 }
2146
2147 return ret_val;
2148 }
2149
2150 static int set_new_data_offset(struct mdinfo *sra, struct supertype *st,
2151 char *devname, int delta_disks,
2152 unsigned long long data_offset,
2153 unsigned long long min)
2154 {
2155 struct mdinfo *sd;
2156 int dir = 0;
2157 int err = 0;
2158 unsigned long long before, after;
2159
2160 /* Need to find min space before and after so same is used
2161 * on all devices
2162 */
2163 before = UINT64_MAX;
2164 after = UINT64_MAX;
2165 for (sd = sra->devs; sd; sd = sd->next) {
2166 char *dn;
2167 int dfd;
2168 int rv;
2169 struct supertype *st2;
2170 struct mdinfo info2;
2171
2172 if (sd->disk.state & (1<<MD_DISK_FAULTY))
2173 continue;
2174 dn = map_dev(sd->disk.major, sd->disk.minor, 0);
2175 dfd = dev_open(dn, O_RDONLY);
2176 if (dfd < 0) {
2177 pr_err("%s: cannot open component %s\n",
2178 devname, dn ? dn : "-unknown-");
2179 goto release;
2180 }
2181 st2 = dup_super(st);
2182 rv = st2->ss->load_super(st2,dfd, NULL);
2183 close(dfd);
2184 if (rv) {
2185 free(st2);
2186 pr_err("%s: cannot get superblock from %s\n",
2187 devname, dn);
2188 goto release;
2189 }
2190 st2->ss->getinfo_super(st2, &info2, NULL);
2191 st2->ss->free_super(st2);
2192 free(st2);
2193 if (info2.space_before == 0 &&
2194 info2.space_after == 0) {
2195 /* Metadata doesn't support data_offset changes */
2196 return 1;
2197 }
2198 if (before > info2.space_before)
2199 before = info2.space_before;
2200 if (after > info2.space_after)
2201 after = info2.space_after;
2202
2203 if (data_offset != INVALID_SECTORS) {
2204 if (dir == 0) {
2205 if (info2.data_offset == data_offset) {
2206 pr_err("%s: already has that data_offset\n",
2207 dn);
2208 goto release;
2209 }
2210 if (data_offset < info2.data_offset)
2211 dir = -1;
2212 else
2213 dir = 1;
2214 } else if ((data_offset <= info2.data_offset && dir == 1) ||
2215 (data_offset >= info2.data_offset && dir == -1)) {
2216 pr_err("%s: differing data offsets on devices make this --data-offset setting impossible\n",
2217 dn);
2218 goto release;
2219 }
2220 }
2221 }
2222 if (before == UINT64_MAX)
2223 /* impossible really, there must be no devices */
2224 return 1;
2225
2226 for (sd = sra->devs; sd; sd = sd->next) {
2227 char *dn = map_dev(sd->disk.major, sd->disk.minor, 0);
2228
2229 struct mdinfo info2;
2230 if (delta_disks < 0) {
2231 /* Don't need any space as array is shrinking
2232 * just move data_offset up by min
2233 */
2234 if (data_offset == INVALID_SECTORS)
2235 info2.new_data_offset = info2.data_offset + min;
2236 else {
2237 if (data_offset < info2.data_offset + min) {
2238 pr_err("--data-offset too small for %s\n",
2239 dn);
2240 goto release;
2241 }
2242 info2.new_data_offset = data_offset;
2243 }
2244 } else if (delta_disks > 0) {
2245 /* need space before */
2246 if (before < min) {
2247 pr_err("Insufficient head-space for reshape on %s\n",
2248 dn);
2249 goto release;
2250 }
2251 if (data_offset == INVALID_SECTORS)
2252 info2.new_data_offset = info2.data_offset - min;
2253 else {
2254 if (data_offset > info2.data_offset - min) {
2255 pr_err("--data-offset too large for %s\n",
2256 dn);
2257 goto release;
2258 }
2259 info2.new_data_offset = data_offset;
2260 }
2261 } else {
2262 if (dir == 0) {
2263 /* can move up or down. If 'data_offset'
2264 * was set we would have already decided,
2265 * so just choose direction with most space.
2266 */
2267 if (info2.space_before > info2.space_after)
2268 dir = -1;
2269 else
2270 dir = 1;
2271 sysfs_set_str(sra, NULL, "reshape_direction",
2272 dir == 1 ? "backwards" : "forwards");
2273 }
2274
2275 switch (dir) {
2276 case 1: /* Increase data offset */
2277 if (after < min) {
2278 pr_err("Insufficient tail-space for reshape on %s\n",
2279 dn);
2280 goto release;
2281 }
2282 if (data_offset != INVALID_SECTORS &&
2283 data_offset < info2.data_offset + min) {
2284 pr_err("--data-offset too small on %s\n",
2285 dn);
2286 goto release;
2287 }
2288 if (data_offset != INVALID_SECTORS)
2289 info2.new_data_offset = data_offset;
2290 else {
2291 unsigned long long off = after / 2;
2292 off &= ~7ULL;
2293 if (off < min)
2294 off = min;
2295 info2.new_data_offset =
2296 info2.data_offset + off;
2297 }
2298 break;
2299 case -1: /* Decrease data offset */
2300 if (before < min) {
2301 pr_err("insufficient head-room on %s\n",
2302 dn);
2303 goto release;
2304 }
2305 if (data_offset != INVALID_SECTORS &&
2306 data_offset < info2.data_offset - min) {
2307 pr_err("--data-offset too small on %s\n",
2308 dn);
2309 goto release;
2310 }
2311 if (data_offset != INVALID_SECTORS)
2312 info2.new_data_offset = data_offset;
2313 else {
2314 unsigned long long off = before / 2;
2315 off &= ~7ULL;
2316 if (off < min)
2317 off = min;
2318 info2.new_data_offset =
2319 info2.data_offset - off;
2320 }
2321 break;
2322 }
2323 }
2324 if (sysfs_set_num(sra, sd, "new_offset",
2325 info2.new_data_offset) < 0) {
2326 err = errno;
2327 err = -1;
2328 if (errno == E2BIG && data_offset != INVALID_SECTORS) {
2329 pr_err("data-offset is too big for %s\n",
2330 dn);
2331 goto release;
2332 }
2333 if (sd == sra->devs &&
2334 (errno == ENOENT || errno == E2BIG))
2335 /* Early kernel, no 'new_offset' file,
2336 * or kernel doesn't like us.
2337 * For RAID5/6 this is not fatal
2338 */
2339 return 1;
2340 pr_err("Cannot set new_offset for %s\n",
2341 dn);
2342 break;
2343 }
2344 }
2345 return err;
2346 release:
2347 return -1;
2348 }
2349
2350 static int raid10_reshape(char *container, int fd, char *devname,
2351 struct supertype *st, struct mdinfo *info,
2352 struct reshape *reshape,
2353 unsigned long long data_offset,
2354 int force, int verbose)
2355 {
2356 /* Changing raid_disks, layout, chunksize or possibly
2357 * just data_offset for a RAID10.
2358 * We must always change data_offset. We change by at least
2359 * ->min_offset_change which is the largest of the old and new
2360 * chunk sizes.
2361 * If raid_disks is increasing, then data_offset must decrease
2362 * by at least this copy size.
2363 * If raid_disks is unchanged, data_offset must increase or
2364 * decrease by at least min_offset_change but preferably by much more.
2365 * We choose half of the available space.
2366 * If raid_disks is decreasing, data_offset must increase by
2367 * at least min_offset_change. To allow of this, component_size
2368 * must be decreased by the same amount.
2369 *
2370 * So we calculate the required minimum and direction, possibly
2371 * reduce the component_size, then iterate through the devices
2372 * and set the new_data_offset.
2373 * If that all works, we set chunk_size, layout, raid_disks, and start
2374 * 'reshape'
2375 */
2376 struct mdinfo *sra;
2377 unsigned long long min;
2378 int err = 0;
2379
2380 sra = sysfs_read(fd, NULL,
2381 GET_COMPONENT|GET_DEVS|GET_OFFSET|GET_STATE|GET_CHUNK
2382 );
2383 if (!sra) {
2384 pr_err("%s: Cannot get array details from sysfs\n",
2385 devname);
2386 goto release;
2387 }
2388 min = reshape->min_offset_change;
2389
2390 if (info->delta_disks)
2391 sysfs_set_str(sra, NULL, "reshape_direction",
2392 info->delta_disks < 0 ? "backwards" : "forwards");
2393 if (info->delta_disks < 0 &&
2394 info->space_after < min) {
2395 int rv = sysfs_set_num(sra, NULL, "component_size",
2396 (sra->component_size -
2397 min)/2);
2398 if (rv) {
2399 pr_err("cannot reduce component size\n");
2400 goto release;
2401 }
2402 }
2403 err = set_new_data_offset(sra, st, devname, info->delta_disks, data_offset,
2404 min);
2405 if (err == 1) {
2406 pr_err("Cannot set new_data_offset: RAID10 reshape not\n");
2407 cont_err("supported on this kernel\n");
2408 err = -1;
2409 }
2410 if (err < 0)
2411 goto release;
2412
2413 if (!err && sysfs_set_num(sra, NULL, "chunk_size", info->new_chunk) < 0)
2414 err = errno;
2415 if (!err && sysfs_set_num(sra, NULL, "layout", reshape->after.layout) < 0)
2416 err = errno;
2417 if (!err && sysfs_set_num(sra, NULL, "raid_disks",
2418 info->array.raid_disks + info->delta_disks) < 0)
2419 err = errno;
2420 if (!err && sysfs_set_str(sra, NULL, "sync_action", "reshape") < 0)
2421 err = errno;
2422 if (err) {
2423 pr_err("Cannot set array shape for %s\n",
2424 devname);
2425 if (err == EBUSY &&
2426 (info->array.state & (1<<MD_SB_BITMAP_PRESENT)))
2427 cont_err(" Bitmap must be removed before"
2428 " shape can be changed\n");
2429 goto release;
2430 }
2431 sysfs_free(sra);
2432 return 0;
2433 release:
2434 sysfs_free(sra);
2435 return 1;
2436 }
2437
2438 static void get_space_after(int fd, struct supertype *st, struct mdinfo *info)
2439 {
2440 struct mdinfo *sra, *sd;
2441 /* Initialisation to silence compiler warning */
2442 unsigned long long min_space_before = 0, min_space_after = 0;
2443 int first = 1;
2444
2445 sra = sysfs_read(fd, NULL, GET_DEVS);
2446 if (!sra)
2447 return;
2448 for (sd = sra->devs; sd; sd = sd->next) {
2449 char *dn;
2450 int dfd;
2451 struct supertype *st2;
2452 struct mdinfo info2;
2453
2454 if (sd->disk.state & (1<<MD_DISK_FAULTY))
2455 continue;
2456 dn = map_dev(sd->disk.major, sd->disk.minor, 0);
2457 dfd = dev_open(dn, O_RDONLY);
2458 if (dfd < 0)
2459 break;
2460 st2 = dup_super(st);
2461 if (st2->ss->load_super(st2,dfd, NULL)) {
2462 close(dfd);
2463 free(st2);
2464 break;
2465 }
2466 close(dfd);
2467 st2->ss->getinfo_super(st2, &info2, NULL);
2468 st2->ss->free_super(st2);
2469 free(st2);
2470 if (first ||
2471 min_space_before > info2.space_before)
2472 min_space_before = info2.space_before;
2473 if (first ||
2474 min_space_after > info2.space_after)
2475 min_space_after = info2.space_after;
2476 first = 0;
2477 }
2478 if (sd == NULL && !first) {
2479 info->space_after = min_space_after;
2480 info->space_before = min_space_before;
2481 }
2482 sysfs_free(sra);
2483 }
2484
2485 static void update_cache_size(char *container, struct mdinfo *sra,
2486 struct mdinfo *info,
2487 int disks, unsigned long long blocks)
2488 {
2489 /* Check that the internal stripe cache is
2490 * large enough, or it won't work.
2491 * It must hold at least 4 stripes of the larger
2492 * chunk size
2493 */
2494 unsigned long cache;
2495 cache = max(info->array.chunk_size, info->new_chunk);
2496 cache *= 4; /* 4 stripes minimum */
2497 cache /= 512; /* convert to sectors */
2498 /* make sure there is room for 'blocks' with a bit to spare */
2499 if (cache < 16 + blocks / disks)
2500 cache = 16 + blocks / disks;
2501 cache /= (4096/512); /* Covert from sectors to pages */
2502
2503 if (sra->cache_size < cache)
2504 subarray_set_num(container, sra, "stripe_cache_size",
2505 cache+1);
2506 }
2507
2508 static int impose_reshape(struct mdinfo *sra,
2509 struct mdinfo *info,
2510 struct supertype *st,
2511 int fd,
2512 int restart,
2513 char *devname, char *container,
2514 struct reshape *reshape)
2515 {
2516 struct mdu_array_info_s array;
2517
2518 sra->new_chunk = info->new_chunk;
2519
2520 if (restart) {
2521 /* for external metadata checkpoint saved by mdmon can be lost
2522 * or missed /due to e.g. crash/. Check if md is not during
2523 * restart farther than metadata points to.
2524 * If so, this means metadata information is obsolete.
2525 */
2526 if (st->ss->external)
2527 verify_reshape_position(info, reshape->level);
2528 sra->reshape_progress = info->reshape_progress;
2529 } else {
2530 sra->reshape_progress = 0;
2531 if (reshape->after.data_disks < reshape->before.data_disks)
2532 /* start from the end of the new array */
2533 sra->reshape_progress = (sra->component_size
2534 * reshape->after.data_disks);
2535 }
2536
2537 ioctl(fd, GET_ARRAY_INFO, &array);
2538 if (info->array.chunk_size == info->new_chunk &&
2539 reshape->before.layout == reshape->after.layout &&
2540 st->ss->external == 0) {
2541 /* use SET_ARRAY_INFO but only if reshape hasn't started */
2542 array.raid_disks = reshape->after.data_disks + reshape->parity;
2543 if (!restart &&
2544 ioctl(fd, SET_ARRAY_INFO, &array) != 0) {
2545 int err = errno;
2546
2547 pr_err("Cannot set device shape for %s: %s\n",
2548 devname, strerror(errno));
2549
2550 if (err == EBUSY &&
2551 (array.state & (1<<MD_SB_BITMAP_PRESENT)))
2552 cont_err("Bitmap must be removed before"
2553 " shape can be changed\n");
2554
2555 goto release;
2556 }
2557 } else if (!restart) {
2558 /* set them all just in case some old 'new_*' value
2559 * persists from some earlier problem.
2560 */
2561 int err = 0;
2562 if (sysfs_set_num(sra, NULL, "chunk_size", info->new_chunk) < 0)
2563 err = errno;
2564 if (!err && sysfs_set_num(sra, NULL, "layout",
2565 reshape->after.layout) < 0)
2566 err = errno;
2567 if (!err && subarray_set_num(container, sra, "raid_disks",
2568 reshape->after.data_disks +
2569 reshape->parity) < 0)
2570 err = errno;
2571 if (err) {
2572 pr_err("Cannot set device shape for %s\n",
2573 devname);
2574
2575 if (err == EBUSY &&
2576 (array.state & (1<<MD_SB_BITMAP_PRESENT)))
2577 cont_err("Bitmap must be removed before"
2578 " shape can be changed\n");
2579 goto release;
2580 }
2581 }
2582 return 0;
2583 release:
2584 return -1;
2585 }
2586
2587 static int reshape_array(char *container, int fd, char *devname,
2588 struct supertype *st, struct mdinfo *info,
2589 int force, struct mddev_dev *devlist,
2590 unsigned long long data_offset,
2591 char *backup_file, int verbose, int forked,
2592 int restart, int freeze_reshape)
2593 {
2594 struct reshape reshape;
2595 int spares_needed;
2596 char *msg;
2597 int orig_level = UnSet;
2598 int odisks;
2599 int delayed;
2600
2601 struct mdu_array_info_s array;
2602 char *c;
2603
2604 struct mddev_dev *dv;
2605 int added_disks;
2606
2607 int *fdlist = NULL;
2608 unsigned long long *offsets = NULL;
2609 int d;
2610 int nrdisks;
2611 int err;
2612 unsigned long blocks;
2613 unsigned long long array_size;
2614 int done;
2615 struct mdinfo *sra = NULL;
2616
2617 /* when reshaping a RAID0, the component_size might be zero.
2618 * So try to fix that up.
2619 */
2620 if (ioctl(fd, GET_ARRAY_INFO, &array) != 0) {
2621 dprintf("Cannot get array information.\n");
2622 goto release;
2623 }
2624 if (array.level == 0 && info->component_size == 0) {
2625 get_dev_size(fd, NULL, &array_size);
2626 info->component_size = array_size / array.raid_disks;
2627 }
2628
2629 if (array.level == 10)
2630 /* Need space_after info */
2631 get_space_after(fd, st, info);
2632
2633 if (info->reshape_active) {
2634 int new_level = info->new_level;
2635 info->new_level = UnSet;
2636 if (info->delta_disks > 0)
2637 info->array.raid_disks -= info->delta_disks;
2638 msg = analyse_change(info, &reshape);
2639 info->new_level = new_level;
2640 if (info->delta_disks > 0)
2641 info->array.raid_disks += info->delta_disks;
2642 if (!restart)
2643 /* Make sure the array isn't read-only */
2644 ioctl(fd, RESTART_ARRAY_RW, 0);
2645 } else
2646 msg = analyse_change(info, &reshape);
2647 if (msg) {
2648 pr_err("%s\n", msg);
2649 goto release;
2650 }
2651 if (restart &&
2652 (reshape.level != info->array.level ||
2653 reshape.before.layout != info->array.layout ||
2654 reshape.before.data_disks + reshape.parity
2655 != info->array.raid_disks - max(0, info->delta_disks))) {
2656 pr_err("reshape info is not in native format -"
2657 " cannot continue.\n");
2658 goto release;
2659 }
2660
2661 if (st->ss->external && restart && (info->reshape_progress == 0)) {
2662 /* When reshape is restarted from '0', very begin of array
2663 * it is possible that for external metadata reshape and array
2664 * configuration doesn't happen.
2665 * Check if md has the same opinion, and reshape is restarted
2666 * from 0. If so, this is regular reshape start after reshape
2667 * switch in metadata to next array only.
2668 */
2669 if ((verify_reshape_position(info, reshape.level) >= 0) &&
2670 (info->reshape_progress == 0))
2671 restart = 0;
2672 }
2673 if (restart) {
2674 /* reshape already started. just skip to monitoring the reshape */
2675 if (reshape.backup_blocks == 0)
2676 return 0;
2677 if (restart & RESHAPE_NO_BACKUP)
2678 return 0;
2679 goto started;
2680 }
2681 /* The container is frozen but the array may not be.
2682 * So freeze the array so spares don't get put to the wrong use
2683 * FIXME there should probably be a cleaner separation between
2684 * freeze_array and freeze_container.
2685 */
2686 sysfs_freeze_array(info);
2687 /* Check we have enough spares to not be degraded */
2688 added_disks = 0;
2689 for (dv = devlist; dv ; dv=dv->next)
2690 added_disks++;
2691 spares_needed = max(reshape.before.data_disks,
2692 reshape.after.data_disks)
2693 + reshape.parity - array.raid_disks;
2694
2695 if (!force &&
2696 info->new_level > 1 && info->array.level > 1 &&
2697 spares_needed > info->array.spare_disks + added_disks) {
2698 pr_err("Need %d spare%s to avoid degraded array,"
2699 " and only have %d.\n"
2700 " Use --force to over-ride this check.\n",
2701 spares_needed,
2702 spares_needed == 1 ? "" : "s",
2703 info->array.spare_disks + added_disks);
2704 goto release;
2705 }
2706 /* Check we have enough spares to not fail */
2707 spares_needed = max(reshape.before.data_disks,
2708 reshape.after.data_disks)
2709 - array.raid_disks;
2710 if ((info->new_level > 1 || info->new_level == 0) &&
2711 spares_needed > info->array.spare_disks +added_disks) {
2712 pr_err("Need %d spare%s to create working array,"
2713 " and only have %d.\n",
2714 spares_needed,
2715 spares_needed == 1 ? "" : "s",
2716 info->array.spare_disks + added_disks);
2717 goto release;
2718 }
2719
2720 if (reshape.level != array.level) {
2721 char *c = map_num(pers, reshape.level);
2722 int err;
2723 if (c == NULL)
2724 goto release;
2725
2726 err = sysfs_set_str(info, NULL, "level", c);
2727 if (err) {
2728 err = errno;
2729 pr_err("%s: could not set level to %s\n",
2730 devname, c);
2731 if (err == EBUSY &&
2732 (info->array.state & (1<<MD_SB_BITMAP_PRESENT)))
2733 cont_err("Bitmap must be removed"
2734 " before level can be changed\n");
2735 goto release;
2736 }
2737 if (verbose >= 0)
2738 pr_err("level of %s changed to %s\n",
2739 devname, c);
2740 orig_level = array.level;
2741 sysfs_freeze_array(info);
2742
2743 if (reshape.level > 0 && st->ss->external) {
2744 /* make sure mdmon is aware of the new level */
2745 if (mdmon_running(container))
2746 flush_mdmon(container);
2747
2748 if (!mdmon_running(container))
2749 start_mdmon(container);
2750 ping_monitor(container);
2751 if (mdmon_running(container) &&
2752 st->update_tail == NULL)
2753 st->update_tail = &st->updates;
2754 }
2755 }
2756 /* ->reshape_super might have chosen some spares from the
2757 * container that it wants to be part of the new array.
2758 * We can collect them with ->container_content and give
2759 * them to the kernel.
2760 */
2761 if (st->ss->reshape_super && st->ss->container_content) {
2762 char *subarray = strchr(info->text_version+1, '/')+1;
2763 struct mdinfo *info2 =
2764 st->ss->container_content(st, subarray);
2765 struct mdinfo *d;
2766
2767 if (info2) {
2768 sysfs_init(info2, fd, st->devnm);
2769 /* When increasing number of devices, we need to set
2770 * new raid_disks before adding these, or they might
2771 * be rejected.
2772 */
2773 if (reshape.backup_blocks &&
2774 reshape.after.data_disks > reshape.before.data_disks)
2775 subarray_set_num(container, info2, "raid_disks",
2776 reshape.after.data_disks +
2777 reshape.parity);
2778 for (d = info2->devs; d; d = d->next) {
2779 if (d->disk.state == 0 &&
2780 d->disk.raid_disk >= 0) {
2781 /* This is a spare that wants to
2782 * be part of the array.
2783 */
2784 add_disk(fd, st, info2, d);
2785 }
2786 }
2787 sysfs_free(info2);
2788 }
2789 }
2790 /* We might have been given some devices to add to the
2791 * array. Now that the array has been changed to the right
2792 * level and frozen, we can safely add them.
2793 */
2794 if (devlist)
2795 Manage_subdevs(devname, fd, devlist, verbose,
2796 0,NULL, 0);
2797
2798
2799 if (reshape.backup_blocks == 0 && data_offset)
2800 reshape.backup_blocks = reshape.before.data_disks * info->array.chunk_size/512;
2801 if (reshape.backup_blocks == 0) {
2802 /* No restriping needed, but we might need to impose
2803 * some more changes: layout, raid_disks, chunk_size
2804 */
2805 /* read current array info */
2806 if (ioctl(fd, GET_ARRAY_INFO, &array) != 0) {
2807 dprintf("Cannot get array information.\n");
2808 goto release;
2809 }
2810 /* compare current array info with new values and if
2811 * it is different update them to new */
2812 if (info->new_layout != UnSet &&
2813 info->new_layout != array.layout) {
2814 array.layout = info->new_layout;
2815 if (ioctl(fd, SET_ARRAY_INFO, &array) != 0) {
2816 pr_err("failed to set new layout\n");
2817 goto release;
2818 } else if (verbose >= 0)
2819 printf("layout for %s set to %d\n",
2820 devname, array.layout);
2821 }
2822 if (info->delta_disks != UnSet &&
2823 info->delta_disks != 0 &&
2824 array.raid_disks != (info->array.raid_disks + info->delta_disks)) {
2825 array.raid_disks += info->delta_disks;
2826 if (ioctl(fd, SET_ARRAY_INFO, &array) != 0) {
2827 pr_err("failed to set raid disks\n");
2828 goto release;
2829 } else if (verbose >= 0) {
2830 printf("raid_disks for %s set to %d\n",
2831 devname, array.raid_disks);
2832 }
2833 }
2834 if (info->new_chunk != 0 &&
2835 info->new_chunk != array.chunk_size) {
2836 if (sysfs_set_num(info, NULL,
2837 "chunk_size", info->new_chunk) != 0) {
2838 pr_err("failed to set chunk size\n");
2839 goto release;
2840 } else if (verbose >= 0)
2841 printf("chunk size for %s set to %d\n",
2842 devname, array.chunk_size);
2843 }
2844 unfreeze(st);
2845 return 0;
2846 }
2847
2848 /*
2849 * There are three possibilities.
2850 * 1/ The array will shrink.
2851 * We need to ensure the reshape will pause before reaching
2852 * the 'critical section'. We also need to fork and wait for
2853 * that to happen. When it does we
2854 * suspend/backup/complete/unfreeze
2855 *
2856 * 2/ The array will not change size.
2857 * This requires that we keep a backup of a sliding window
2858 * so that we can restore data after a crash. So we need
2859 * to fork and monitor progress.
2860 * In future we will allow the data_offset to change, so
2861 * a sliding backup becomes unnecessary.
2862 *
2863 * 3/ The array will grow. This is relatively easy.
2864 * However the kernel's restripe routines will cheerfully
2865 * overwrite some early data before it is safe. So we
2866 * need to make a backup of the early parts of the array
2867 * and be ready to restore it if rebuild aborts very early.
2868 * For externally managed metadata, we still need a forked
2869 * child to monitor the reshape and suspend IO over the region
2870 * that is being reshaped.
2871 *
2872 * We backup data by writing it to one spare, or to a
2873 * file which was given on command line.
2874 *
2875 * In each case, we first make sure that storage is available
2876 * for the required backup.
2877 * Then we:
2878 * - request the shape change.
2879 * - fork to handle backup etc.
2880 */
2881 /* Check that we can hold all the data */
2882 get_dev_size(fd, NULL, &array_size);
2883 if (reshape.new_size < (array_size/512)) {
2884 pr_err("this change will reduce the size of the array.\n"
2885 " use --grow --array-size first to truncate array.\n"
2886 " e.g. mdadm --grow %s --array-size %llu\n",
2887 devname, reshape.new_size/2);
2888 goto release;
2889 }
2890
2891 if (array.level == 10) {
2892 /* Reshaping RAID10 does not require any data backup by
2893 * user-space. Instead it requires that the data_offset
2894 * is changed to avoid the need for backup.
2895 * So this is handled very separately
2896 */
2897 if (restart)
2898 /* Nothing to do. */
2899 return 0;
2900 return raid10_reshape(container, fd, devname, st, info,
2901 &reshape, data_offset,
2902 force, verbose);
2903 }
2904 sra = sysfs_read(fd, NULL,
2905 GET_COMPONENT|GET_DEVS|GET_OFFSET|GET_STATE|GET_CHUNK|
2906 GET_CACHE);
2907 if (!sra) {
2908 pr_err("%s: Cannot get array details from sysfs\n",
2909 devname);
2910 goto release;
2911 }
2912
2913 if (!backup_file)
2914 switch(set_new_data_offset(sra, st, devname,
2915 reshape.after.data_disks - reshape.before.data_disks,
2916 data_offset,
2917 reshape.min_offset_change)) {
2918 case -1:
2919 goto release;
2920 case 0:
2921 /* Updated data_offset, so it's easy now */
2922 update_cache_size(container, sra, info,
2923 min(reshape.before.data_disks,
2924 reshape.after.data_disks),
2925 reshape.backup_blocks);
2926
2927 /* Right, everything seems fine. Let's kick things off.
2928 */
2929 sync_metadata(st);
2930
2931 if (impose_reshape(sra, info, st, fd, restart,
2932 devname, container, &reshape) < 0)
2933 goto release;
2934 if (sysfs_set_str(sra, NULL, "sync_action", "reshape") < 0) {
2935 pr_err("Failed to initiate reshape!\n");
2936 goto release;
2937 }
2938
2939 return 0;
2940 case 1: /* Couldn't set data_offset, try the old way */
2941 if (data_offset != INVALID_SECTORS) {
2942 pr_err("Cannot update data_offset on this array\n");
2943 goto release;
2944 }
2945 break;
2946 }
2947
2948 started:
2949 /* Decide how many blocks (sectors) for a reshape
2950 * unit. The number we have so far is just a minimum
2951 */
2952 blocks = reshape.backup_blocks;
2953 if (reshape.before.data_disks ==
2954 reshape.after.data_disks) {
2955 /* Make 'blocks' bigger for better throughput, but
2956 * not so big that we reject it below.
2957 * Try for 16 megabytes
2958 */
2959 while (blocks * 32 < sra->component_size &&
2960 blocks < 16*1024*2)
2961 blocks *= 2;
2962 } else
2963 pr_err("Need to backup %luK of critical "
2964 "section..\n", blocks/2);
2965
2966 if (blocks >= sra->component_size/2) {
2967 pr_err("%s: Something wrong"
2968 " - reshape aborted\n",
2969 devname);
2970 goto release;
2971 }
2972
2973 /* Now we need to open all these devices so we can read/write.
2974 */
2975 nrdisks = max(reshape.before.data_disks,
2976 reshape.after.data_disks) + reshape.parity
2977 + sra->array.spare_disks;
2978 fdlist = xcalloc((1+nrdisks), sizeof(int));
2979 offsets = xcalloc((1+nrdisks), sizeof(offsets[0]));
2980
2981 odisks = reshape.before.data_disks + reshape.parity;
2982 d = reshape_prepare_fdlist(devname, sra, odisks,
2983 nrdisks, blocks, backup_file,
2984 fdlist, offsets);
2985 if (d < 0) {
2986 goto release;
2987 }
2988 if ((st->ss->manage_reshape == NULL) ||
2989 (st->ss->recover_backup == NULL)) {
2990 if (backup_file == NULL) {
2991 if (reshape.after.data_disks <=
2992 reshape.before.data_disks) {
2993 pr_err("%s: Cannot grow - "
2994 "need backup-file\n", devname);
2995 goto release;
2996 } else if (sra->array.spare_disks == 0) {
2997 pr_err("%s: Cannot grow - "
2998 "need a spare or backup-file to backup "
2999 "critical section\n", devname);
3000 goto release;
3001 }
3002 } else {
3003 if (!reshape_open_backup_file(backup_file, fd, devname,
3004 (signed)blocks,
3005 fdlist+d, offsets+d,
3006 restart)) {
3007 goto release;
3008 }
3009 d++;
3010 }
3011 }
3012
3013 update_cache_size(container, sra, info,
3014 min(reshape.before.data_disks, reshape.after.data_disks),
3015 blocks);
3016
3017 /* Right, everything seems fine. Let's kick things off.
3018 * If only changing raid_disks, use ioctl, else use
3019 * sysfs.
3020 */
3021 sync_metadata(st);
3022
3023 if (impose_reshape(sra, info, st, fd, restart,
3024 devname, container, &reshape) < 0)
3025 goto release;
3026
3027 err = start_reshape(sra, restart, reshape.before.data_disks,
3028 reshape.after.data_disks);
3029 if (err) {
3030 pr_err("Cannot %s reshape for %s\n",
3031 restart ? "continue" : "start",
3032 devname);
3033 goto release;
3034 }
3035 if (restart)
3036 sysfs_set_str(sra, NULL, "array_state", "active");
3037 if (freeze_reshape) {
3038 free(fdlist);
3039 free(offsets);
3040 sysfs_free(sra);
3041 pr_err("Reshape has to be continued from"
3042 " location %llu when root filesystem has been mounted.\n",
3043 sra->reshape_progress);
3044 return 1;
3045 }
3046
3047 /* Now we just need to kick off the reshape and watch, while
3048 * handling backups of the data...
3049 * This is all done by a forked background process.
3050 */
3051 switch(forked ? 0 : fork()) {
3052 case -1:
3053 pr_err("Cannot run child to monitor reshape: %s\n",
3054 strerror(errno));
3055 abort_reshape(sra);
3056 goto release;
3057 default:
3058 free(fdlist);
3059 free(offsets);
3060 sysfs_free(sra);
3061 return 0;
3062 case 0:
3063 map_fork();
3064 break;
3065 }
3066
3067 /* If another array on the same devices is busy, the
3068 * reshape will wait for them. This would mean that
3069 * the first section that we suspend will stay suspended
3070 * for a long time. So check on that possibility
3071 * by looking for "DELAYED" in /proc/mdstat, and if found,
3072 * wait a while
3073 */
3074 do {
3075 struct mdstat_ent *mds, *m;
3076 delayed = 0;
3077 mds = mdstat_read(0, 0);
3078 for (m = mds; m; m = m->next)
3079 if (strcmp(m->devnm, sra->sys_name) == 0) {
3080 if (m->resync &&
3081 m->percent == RESYNC_DELAYED)
3082 delayed = 1;
3083 if (m->resync == 0)
3084 /* Haven't started the reshape thread
3085 * yet, wait a bit
3086 */
3087 delayed = 2;
3088 break;
3089 }
3090 free_mdstat(mds);
3091 if (delayed == 1 && get_linux_version() < 3007000) {
3092 pr_err("Reshape is delayed, but cannot wait carefully with this kernel.\n"
3093 " You might experience problems until other reshapes complete.\n");
3094 delayed = 0;
3095 }
3096 if (delayed)
3097 sleep(30 - (delayed-1) * 25);
3098 } while (delayed);
3099
3100 close(fd);
3101 if (check_env("MDADM_GROW_VERIFY"))
3102 fd = open(devname, O_RDONLY | O_DIRECT);
3103 else
3104 fd = -1;
3105 mlockall(MCL_FUTURE);
3106
3107 if (st->ss->external) {
3108 /* metadata handler takes it from here */
3109 done = st->ss->manage_reshape(
3110 fd, sra, &reshape, st, blocks,
3111 fdlist, offsets,
3112 d - odisks, fdlist+odisks,
3113 offsets+odisks);
3114 } else
3115 done = child_monitor(
3116 fd, sra, &reshape, st, blocks,
3117 fdlist, offsets,
3118 d - odisks, fdlist+odisks,
3119 offsets+odisks);
3120
3121 free(fdlist);
3122 free(offsets);
3123
3124 if (backup_file && done)
3125 unlink(backup_file);
3126 if (!done) {
3127 abort_reshape(sra);
3128 goto out;
3129 }
3130
3131 if (!st->ss->external &&
3132 !(reshape.before.data_disks != reshape.after.data_disks
3133 && info->custom_array_size) &&
3134 info->new_level == reshape.level &&
3135 !forked) {
3136 /* no need to wait for the reshape to finish as
3137 * there is nothing more to do.
3138 */
3139 sysfs_free(sra);
3140 exit(0);
3141 }
3142 wait_reshape(sra);
3143
3144 if (st->ss->external) {
3145 /* Re-load the metadata as much could have changed */
3146 int cfd = open_dev(st->container_devnm);
3147 if (cfd >= 0) {
3148 flush_mdmon(container);
3149 st->ss->free_super(st);
3150 st->ss->load_container(st, cfd, container);
3151 close(cfd);
3152 }
3153 }
3154
3155 /* set new array size if required customer_array_size is used
3156 * by this metadata.
3157 */
3158 if (reshape.before.data_disks !=
3159 reshape.after.data_disks &&
3160 info->custom_array_size)
3161 set_array_size(st, info, info->text_version);
3162
3163 if (info->new_level != reshape.level) {
3164
3165 c = map_num(pers, info->new_level);
3166 if (c) {
3167 err = sysfs_set_str(sra, NULL, "level", c);
3168 if (err)
3169 pr_err("%s: could not set level "
3170 "to %s\n", devname, c);
3171 }
3172 if (info->new_level == 0)
3173 st->update_tail = NULL;
3174 }
3175 out:
3176 sysfs_free(sra);
3177 if (forked)
3178 return 0;
3179 unfreeze(st);
3180 exit(0);
3181
3182 release:
3183 free(fdlist);
3184 free(offsets);
3185 if (orig_level != UnSet && sra) {
3186 c = map_num(pers, orig_level);
3187 if (c && sysfs_set_str(sra, NULL, "level", c) == 0)
3188 pr_err("aborting level change\n");
3189 }
3190 sysfs_free(sra);
3191 if (!forked)
3192 unfreeze(st);
3193 return 1;
3194 }
3195
3196 /* mdfd handle is passed to be closed in child process (after fork).
3197 */
3198 int reshape_container(char *container, char *devname,
3199 int mdfd,
3200 struct supertype *st,
3201 struct mdinfo *info,
3202 int force,
3203 char *backup_file,
3204 int verbose, int restart, int freeze_reshape)
3205 {
3206 struct mdinfo *cc = NULL;
3207 int rv = restart;
3208 char last_devnm[32] = "";
3209
3210 /* component_size is not meaningful for a container,
3211 * so pass '0' meaning 'no change'
3212 */
3213 if (!restart &&
3214 reshape_super(st, 0, info->new_level,
3215 info->new_layout, info->new_chunk,
3216 info->array.raid_disks, info->delta_disks,
3217 backup_file, devname, APPLY_METADATA_CHANGES,
3218 verbose)) {
3219 unfreeze(st);
3220 return 1;
3221 }
3222
3223 sync_metadata(st);
3224
3225 /* ping monitor to be sure that update is on disk
3226 */
3227 ping_monitor(container);
3228
3229 switch (fork()) {
3230 case -1: /* error */
3231 perror("Cannot fork to complete reshape\n");
3232 unfreeze(st);
3233 return 1;
3234 default: /* parent */
3235 if (!freeze_reshape)
3236 printf(Name ": multi-array reshape continues"
3237 " in background\n");
3238 return 0;
3239 case 0: /* child */
3240 map_fork();
3241 break;
3242 }
3243
3244 /* close unused handle in child process
3245 */
3246 if (mdfd > -1)
3247 close(mdfd);
3248
3249 while(1) {
3250 /* For each member array with reshape_active,
3251 * we need to perform the reshape.
3252 * We pick the first array that needs reshaping and
3253 * reshape it. reshape_array() will re-read the metadata
3254 * so the next time through a different array should be
3255 * ready for reshape.
3256 * It is possible that the 'different' array will not
3257 * be assembled yet. In that case we simple exit.
3258 * When it is assembled, the mdadm which assembles it
3259 * will take over the reshape.
3260 */
3261 struct mdinfo *content;
3262 int fd;
3263 struct mdstat_ent *mdstat;
3264 char *adev;
3265 int devid;
3266
3267 sysfs_free(cc);
3268
3269 cc = st->ss->container_content(st, NULL);
3270
3271 for (content = cc; content ; content = content->next) {
3272 char *subarray;
3273 if (!content->reshape_active)
3274 continue;
3275
3276 subarray = strchr(content->text_version+1, '/')+1;
3277 mdstat = mdstat_by_subdev(subarray, container);
3278 if (!mdstat)
3279 continue;
3280 if (mdstat->active == 0) {
3281 pr_err("Skipping inactive array %s.\n",
3282 mdstat->devnm);
3283 free_mdstat(mdstat);
3284 mdstat = NULL;
3285 continue;
3286 }
3287 break;
3288 }
3289 if (!content)
3290 break;
3291
3292 devid = devnm2devid(mdstat->devnm);
3293 adev = map_dev(major(devid), minor(devid), 0);
3294 if (!adev)
3295 adev = content->text_version;
3296
3297 fd = open_dev(mdstat->devnm);
3298 if (fd < 0) {
3299 printf(Name ": Device %s cannot be opened for reshape.",
3300 adev);
3301 break;
3302 }
3303
3304 if (strcmp(last_devnm, mdstat->devnm) == 0) {
3305 /* Do not allow for multiple reshape_array() calls for
3306 * the same array.
3307 * It can happen when reshape_array() returns without
3308 * error, when reshape is not finished (wrong reshape
3309 * starting/continuation conditions). Mdmon doesn't
3310 * switch to next array in container and reentry
3311 * conditions for the same array occur.
3312 * This is possibly interim until the behaviour of
3313 * reshape_array is resolved().
3314 */
3315 printf(Name ": Multiple reshape execution detected for "
3316 "device %s.", adev);
3317 close(fd);
3318 break;
3319 }
3320 strcpy(last_devnm, mdstat->devnm);
3321
3322 sysfs_init(content, fd, mdstat->devnm);
3323
3324 if (mdmon_running(container))
3325 flush_mdmon(container);
3326
3327 rv = reshape_array(container, fd, adev, st,
3328 content, force, NULL, 0ULL,
3329 backup_file, verbose, 1, restart,
3330 freeze_reshape);
3331 close(fd);
3332
3333 if (freeze_reshape) {
3334 sysfs_free(cc);
3335 exit(0);
3336 }
3337
3338 restart = 0;
3339 if (rv)
3340 break;
3341
3342 if (mdmon_running(container))
3343 flush_mdmon(container);
3344 }
3345 if (!rv)
3346 unfreeze(st);
3347 sysfs_free(cc);
3348 exit(0);
3349 }
3350
3351 /*
3352 * We run a child process in the background which performs the following
3353 * steps:
3354 * - wait for resync to reach a certain point
3355 * - suspend io to the following section
3356 * - backup that section
3357 * - allow resync to proceed further
3358 * - resume io
3359 * - discard the backup.
3360 *
3361 * When are combined in slightly different ways in the three cases.
3362 * Grow:
3363 * - suspend/backup/allow/wait/resume/discard
3364 * Shrink:
3365 * - allow/wait/suspend/backup/allow/wait/resume/discard
3366 * same-size:
3367 * - wait/resume/discard/suspend/backup/allow
3368 *
3369 * suspend/backup/allow always come together
3370 * wait/resume/discard do too.
3371 * For the same-size case we have two backups to improve flow.
3372 *
3373 */
3374
3375 int progress_reshape(struct mdinfo *info, struct reshape *reshape,
3376 unsigned long long backup_point,
3377 unsigned long long wait_point,
3378 unsigned long long *suspend_point,
3379 unsigned long long *reshape_completed)
3380 {
3381 /* This function is called repeatedly by the reshape manager.
3382 * It determines how much progress can safely be made and allows
3383 * that progress.
3384 * - 'info' identifies the array and particularly records in
3385 * ->reshape_progress the metadata's knowledge of progress
3386 * This is a sector offset from the start of the array
3387 * of the next array block to be relocated. This number
3388 * may increase from 0 or decrease from array_size, depending
3389 * on the type of reshape that is happening.
3390 * Note that in contrast, 'sync_completed' is a block count of the
3391 * reshape so far. It gives the distance between the start point
3392 * (head or tail of device) and the next place that data will be
3393 * written. It always increases.
3394 * - 'reshape' is the structure created by analyse_change
3395 * - 'backup_point' shows how much the metadata manager has backed-up
3396 * data. For reshapes with increasing progress, it is the next address
3397 * to be backed up, previous addresses have been backed-up. For
3398 * decreasing progress, it is the earliest address that has been
3399 * backed up - later address are also backed up.
3400 * So addresses between reshape_progress and backup_point are
3401 * backed up providing those are in the 'correct' order.
3402 * - 'wait_point' is an array address. When reshape_completed
3403 * passes this point, progress_reshape should return. It might
3404 * return earlier if it determines that ->reshape_progress needs
3405 * to be updated or further backup is needed.
3406 * - suspend_point is maintained by progress_reshape and the caller
3407 * should not touch it except to initialise to zero.
3408 * It is an array address and it only increases in 2.6.37 and earlier.
3409 * This makes it difficult to handle reducing reshapes with
3410 * external metadata.
3411 * However: it is similar to backup_point in that it records the
3412 * other end of a suspended region from reshape_progress.
3413 * it is moved to extend the region that is safe to backup and/or
3414 * reshape
3415 * - reshape_completed is read from sysfs and returned. The caller
3416 * should copy this into ->reshape_progress when it has reason to
3417 * believe that the metadata knows this, and any backup outside this
3418 * has been erased.
3419 *
3420 * Return value is:
3421 * 1 if more data from backup_point - but only as far as suspend_point,
3422 * should be backed up
3423 * 0 if things are progressing smoothly
3424 * -1 if the reshape is finished because it is all done,
3425 * -2 if the reshape is finished due to an error.
3426 */
3427
3428 int advancing = (reshape->after.data_disks
3429 >= reshape->before.data_disks);
3430 unsigned long long need_backup; /* All data between start of array and
3431 * here will at some point need to
3432 * be backed up.
3433 */
3434 unsigned long long read_offset, write_offset;
3435 unsigned long long write_range;
3436 unsigned long long max_progress, target, completed;
3437 unsigned long long array_size = (info->component_size
3438 * reshape->before.data_disks);
3439 int fd;
3440 char buf[20];
3441
3442 /* First, we unsuspend any region that is now known to be safe.
3443 * If suspend_point is on the 'wrong' side of reshape_progress, then
3444 * we don't have or need suspension at the moment. This is true for
3445 * native metadata when we don't need to back-up.
3446 */
3447 if (advancing) {
3448 if (info->reshape_progress <= *suspend_point)
3449 sysfs_set_num(info, NULL, "suspend_lo",
3450 info->reshape_progress);
3451 } else {
3452 /* Note: this won't work in 2.6.37 and before.
3453 * Something somewhere should make sure we don't need it!
3454 */
3455 if (info->reshape_progress >= *suspend_point)
3456 sysfs_set_num(info, NULL, "suspend_hi",
3457 info->reshape_progress);
3458 }
3459
3460 /* Now work out how far it is safe to progress.
3461 * If the read_offset for ->reshape_progress is less than
3462 * 'blocks' beyond the write_offset, we can only progress as far
3463 * as a backup.
3464 * Otherwise we can progress until the write_offset for the new location
3465 * reaches (within 'blocks' of) the read_offset at the current location.
3466 * However that region must be suspended unless we are using native
3467 * metadata.
3468 * If we need to suspend more, we limit it to 128M per device, which is
3469 * rather arbitrary and should be some time-based calculation.
3470 */
3471 read_offset = info->reshape_progress / reshape->before.data_disks;
3472 write_offset = info->reshape_progress / reshape->after.data_disks;
3473 write_range = info->new_chunk/512;
3474 if (reshape->before.data_disks == reshape->after.data_disks)
3475 need_backup = array_size;
3476 else
3477 need_backup = reshape->backup_blocks;
3478 if (advancing) {
3479 if (read_offset < write_offset + write_range)
3480 max_progress = backup_point;
3481 else
3482 max_progress =
3483 read_offset *
3484 reshape->after.data_disks;
3485 } else {
3486 if (read_offset > write_offset - write_range)
3487 /* Can only progress as far as has been backed up,
3488 * which must be suspended */
3489 max_progress = backup_point;
3490 else if (info->reshape_progress <= need_backup)
3491 max_progress = backup_point;
3492 else {
3493 if (info->array.major_version >= 0)
3494 /* Can progress until backup is needed */
3495 max_progress = need_backup;
3496 else {
3497 /* Can progress until metadata update is required */
3498 max_progress =
3499 read_offset *
3500 reshape->after.data_disks;
3501 /* but data must be suspended */
3502 if (max_progress < *suspend_point)
3503 max_progress = *suspend_point;
3504 }
3505 }
3506 }
3507
3508 /* We know it is safe to progress to 'max_progress' providing
3509 * it is suspended or we are using native metadata.
3510 * Consider extending suspend_point 128M per device if it
3511 * is less than 64M per device beyond reshape_progress.
3512 * But always do a multiple of 'blocks'
3513 * FIXME this is too big - it takes to long to complete
3514 * this much.
3515 */
3516 target = 64*1024*2 * min(reshape->before.data_disks,
3517 reshape->after.data_disks);
3518 target /= reshape->backup_blocks;
3519 if (target < 2)
3520 target = 2;
3521 target *= reshape->backup_blocks;
3522
3523 /* For externally managed metadata we always need to suspend IO to
3524 * the area being reshaped so we regularly push suspend_point forward.
3525 * For native metadata we only need the suspend if we are going to do
3526 * a backup.
3527 */
3528 if (advancing) {
3529 if ((need_backup > info->reshape_progress
3530 || info->array.major_version < 0) &&
3531 *suspend_point < info->reshape_progress + target) {
3532 if (need_backup < *suspend_point + 2 * target)
3533 *suspend_point = need_backup;
3534 else if (*suspend_point + 2 * target < array_size)
3535 *suspend_point += 2 * target;
3536 else
3537 *suspend_point = array_size;
3538 sysfs_set_num(info, NULL, "suspend_hi", *suspend_point);
3539 if (max_progress > *suspend_point)
3540 max_progress = *suspend_point;
3541 }
3542 } else {
3543 if (info->array.major_version >= 0) {
3544 /* Only need to suspend when about to backup */
3545 if (info->reshape_progress < need_backup * 2 &&
3546 *suspend_point > 0) {
3547 *suspend_point = 0;
3548 sysfs_set_num(info, NULL, "suspend_lo", 0);
3549 sysfs_set_num(info, NULL, "suspend_hi", need_backup);
3550 }
3551 } else {
3552 /* Need to suspend continually */
3553 if (info->reshape_progress < *suspend_point)
3554 *suspend_point = info->reshape_progress;
3555 if (*suspend_point + target < info->reshape_progress)
3556 /* No need to move suspend region yet */;
3557 else {
3558 if (*suspend_point >= 2 * target)
3559 *suspend_point -= 2 * target;
3560 else
3561 *suspend_point = 0;
3562 sysfs_set_num(info, NULL, "suspend_lo",
3563 *suspend_point);
3564 }
3565 if (max_progress < *suspend_point)
3566 max_progress = *suspend_point;
3567 }
3568 }
3569
3570 /* now set sync_max to allow that progress. sync_max, like
3571 * sync_completed is a count of sectors written per device, so
3572 * we find the difference between max_progress and the start point,
3573 * and divide that by after.data_disks to get a sync_max
3574 * number.
3575 * At the same time we convert wait_point to a similar number
3576 * for comparing against sync_completed.
3577 */
3578 /* scale down max_progress to per_disk */
3579 max_progress /= reshape->after.data_disks;
3580 /* Round to chunk size as some kernels give an erroneously high number */
3581 max_progress /= info->new_chunk/512;
3582 max_progress *= info->new_chunk/512;
3583 /* And round to old chunk size as the kernel wants that */
3584 max_progress /= info->array.chunk_size/512;
3585 max_progress *= info->array.chunk_size/512;
3586 /* Limit progress to the whole device */
3587 if (max_progress > info->component_size)
3588 max_progress = info->component_size;
3589 wait_point /= reshape->after.data_disks;
3590 if (!advancing) {
3591 /* switch from 'device offset' to 'processed block count' */
3592 max_progress = info->component_size - max_progress;
3593 wait_point = info->component_size - wait_point;
3594 }
3595
3596 sysfs_set_num(info, NULL, "sync_max", max_progress);
3597
3598 /* Now wait. If we have already reached the point that we were
3599 * asked to wait to, don't wait at all, else wait for any change.
3600 * We need to select on 'sync_completed' as that is the place that
3601 * notifications happen, but we are really interested in
3602 * 'reshape_position'
3603 */
3604 fd = sysfs_get_fd(info, NULL, "sync_completed");
3605 if (fd < 0)
3606 goto check_progress;
3607
3608 if (sysfs_fd_get_ll(fd, &completed) < 0)
3609 goto check_progress;
3610
3611 while (completed < max_progress && completed < wait_point) {
3612 /* Check that sync_action is still 'reshape' to avoid
3613 * waiting forever on a dead array
3614 */
3615 char action[20];
3616 fd_set rfds;
3617 if (sysfs_get_str(info, NULL, "sync_action",
3618 action, 20) <= 0 ||
3619 strncmp(action, "reshape", 7) != 0)
3620 break;
3621 /* Some kernels reset 'sync_completed' to zero
3622 * before setting 'sync_action' to 'idle'.
3623 * So we need these extra tests.
3624 */
3625 if (completed == 0 && advancing
3626 && info->reshape_progress > 0)
3627 break;
3628 if (completed == 0 && !advancing
3629 && info->reshape_progress < (info->component_size
3630 * reshape->after.data_disks))
3631 break;
3632 FD_ZERO(&rfds);
3633 FD_SET(fd, &rfds);
3634 select(fd+1, NULL, NULL, &rfds, NULL);
3635 if (sysfs_fd_get_ll(fd, &completed) < 0)
3636 goto check_progress;
3637 }
3638 /* Some kernels reset 'sync_completed' to zero,
3639 * we need to have real point we are in md
3640 */
3641 if (completed == 0)
3642 completed = max_progress;
3643
3644 /* some kernels can give an incorrectly high 'completed' number */
3645 completed /= (info->new_chunk/512);
3646 completed *= (info->new_chunk/512);
3647 /* Convert 'completed' back in to a 'progress' number */
3648 completed *= reshape->after.data_disks;
3649 if (!advancing) {
3650 completed = info->component_size * reshape->after.data_disks
3651 - completed;
3652 }
3653 *reshape_completed = completed;
3654
3655 close(fd);
3656
3657 /* We return the need_backup flag. Caller will decide
3658 * how much - a multiple of ->backup_blocks up to *suspend_point
3659 */
3660 if (advancing)
3661 return need_backup > info->reshape_progress;
3662 else
3663 return need_backup >= info->reshape_progress;
3664
3665 check_progress:
3666 /* if we couldn't read a number from sync_completed, then
3667 * either the reshape did complete, or it aborted.
3668 * We can tell which by checking for 'none' in reshape_position.
3669 * If it did abort, then it might immediately restart if it
3670 * it was just a device failure that leaves us degraded but
3671 * functioning.
3672 */
3673 strcpy(buf, "hi");
3674 if (sysfs_get_str(info, NULL, "reshape_position", buf, sizeof(buf)) < 0
3675 || strncmp(buf, "none", 4) != 0) {
3676 /* The abort might only be temporary. Wait up to 10
3677 * seconds for fd to contain a valid number again.
3678 */
3679 struct timeval tv;
3680 int rv = -2;
3681 tv.tv_sec = 10;
3682 tv.tv_usec = 0;
3683 while (fd >= 0 && rv < 0 && tv.tv_sec > 0) {
3684 fd_set rfds;
3685 FD_ZERO(&rfds);
3686 FD_SET(fd, &rfds);
3687 if (select(fd+1, NULL, NULL, &rfds, &tv) != 1)
3688 break;
3689 switch (sysfs_fd_get_ll(fd, &completed)) {
3690 case 0:
3691 /* all good again */
3692 rv = 1;
3693 break;
3694 case -2: /* read error - abort */
3695 tv.tv_sec = 0;
3696 break;
3697 }
3698 }
3699 if (fd >= 0)
3700 close(fd);
3701 return rv; /* abort */
3702 } else {
3703 /* Maybe racing with array shutdown - check state */
3704 if (fd >= 0)
3705 close(fd);
3706 if (sysfs_get_str(info, NULL, "array_state", buf, sizeof(buf)) < 0
3707 || strncmp(buf, "inactive", 8) == 0
3708 || strncmp(buf, "clear",5) == 0)
3709 return -2; /* abort */
3710 return -1; /* complete */
3711 }
3712 }
3713
3714 /* FIXME return status is never checked */
3715 static int grow_backup(struct mdinfo *sra,
3716 unsigned long long offset, /* per device */
3717 unsigned long stripes, /* per device, in old chunks */
3718 int *sources, unsigned long long *offsets,
3719 int disks, int chunk, int level, int layout,
3720 int dests, int *destfd, unsigned long long *destoffsets,
3721 int part, int *degraded,
3722 char *buf)
3723 {
3724 /* Backup 'blocks' sectors at 'offset' on each device of the array,
3725 * to storage 'destfd' (offset 'destoffsets'), after first
3726 * suspending IO. Then allow resync to continue
3727 * over the suspended section.
3728 * Use part 'part' of the backup-super-block.
3729 */
3730 int odata = disks;
3731 int rv = 0;
3732 int i;
3733 unsigned long long ll;
3734 int new_degraded;
3735 //printf("offset %llu\n", offset);
3736 if (level >= 4)
3737 odata--;
3738 if (level == 6)
3739 odata--;
3740
3741 /* Check that array hasn't become degraded, else we might backup the wrong data */
3742 if (sysfs_get_ll(sra, NULL, "degraded", &ll) < 0)
3743 return -1; /* FIXME this error is ignored */
3744 new_degraded = (int)ll;
3745 if (new_degraded != *degraded) {
3746 /* check each device to ensure it is still working */
3747 struct mdinfo *sd;
3748 for (sd = sra->devs ; sd ; sd = sd->next) {
3749 if (sd->disk.state & (1<<MD_DISK_FAULTY))
3750 continue;
3751 if (sd->disk.state & (1<<MD_DISK_SYNC)) {
3752 char sbuf[20];
3753 if (sysfs_get_str(sra, sd, "state", sbuf, 20) < 0 ||
3754 strstr(sbuf, "faulty") ||
3755 strstr(sbuf, "in_sync") == NULL) {
3756 /* this device is dead */
3757 sd->disk.state = (1<<MD_DISK_FAULTY);
3758 if (sd->disk.raid_disk >= 0 &&
3759 sources[sd->disk.raid_disk] >= 0) {
3760 close(sources[sd->disk.raid_disk]);
3761 sources[sd->disk.raid_disk] = -1;
3762 }
3763 }
3764 }
3765 }
3766 *degraded = new_degraded;
3767 }
3768 if (part) {
3769 bsb.arraystart2 = __cpu_to_le64(offset * odata);
3770 bsb.length2 = __cpu_to_le64(stripes * (chunk/512) * odata);
3771 } else {
3772 bsb.arraystart = __cpu_to_le64(offset * odata);
3773 bsb.length = __cpu_to_le64(stripes * (chunk/512) * odata);
3774 }
3775 if (part)
3776 bsb.magic[15] = '2';
3777 for (i = 0; i < dests; i++)
3778 if (part)
3779 lseek64(destfd[i], destoffsets[i] + __le64_to_cpu(bsb.devstart2)*512, 0);
3780 else
3781 lseek64(destfd[i], destoffsets[i], 0);
3782
3783 rv = save_stripes(sources, offsets,
3784 disks, chunk, level, layout,
3785 dests, destfd,
3786 offset*512*odata, stripes * chunk * odata,
3787 buf);
3788
3789 if (rv)
3790 return rv;
3791 bsb.mtime = __cpu_to_le64(time(0));
3792 for (i = 0; i < dests; i++) {
3793 bsb.devstart = __cpu_to_le64(destoffsets[i]/512);
3794
3795 bsb.sb_csum = bsb_csum((char*)&bsb, ((char*)&bsb.sb_csum)-((char*)&bsb));
3796 if (memcmp(bsb.magic, "md_backup_data-2", 16) == 0)
3797 bsb.sb_csum2 = bsb_csum((char*)&bsb,
3798 ((char*)&bsb.sb_csum2)-((char*)&bsb));
3799
3800 rv = -1;
3801 if ((unsigned long long)lseek64(destfd[i], destoffsets[i] - 4096, 0)
3802 != destoffsets[i] - 4096)
3803 break;
3804 if (write(destfd[i], &bsb, 512) != 512)
3805 break;
3806 if (destoffsets[i] > 4096) {
3807 if ((unsigned long long)lseek64(destfd[i], destoffsets[i]+stripes*chunk*odata, 0) !=
3808 destoffsets[i]+stripes*chunk*odata)
3809 break;
3810 if (write(destfd[i], &bsb, 512) != 512)
3811 break;
3812 }
3813 fsync(destfd[i]);
3814 rv = 0;
3815 }
3816
3817 return rv;
3818 }
3819
3820 /* in 2.6.30, the value reported by sync_completed can be
3821 * less that it should be by one stripe.
3822 * This only happens when reshape hits sync_max and pauses.
3823 * So allow wait_backup to either extent sync_max further
3824 * than strictly necessary, or return before the
3825 * sync has got quite as far as we would really like.
3826 * This is what 'blocks2' is for.
3827 * The various caller give appropriate values so that
3828 * every works.
3829 */
3830 /* FIXME return value is often ignored */
3831 static int forget_backup(int dests, int *destfd,
3832 unsigned long long *destoffsets,
3833 int part)
3834 {
3835 /*
3836 * Erase backup 'part' (which is 0 or 1)
3837 */
3838 int i;
3839 int rv;
3840
3841 if (part) {
3842 bsb.arraystart2 = __cpu_to_le64(0);
3843 bsb.length2 = __cpu_to_le64(0);
3844 } else {
3845 bsb.arraystart = __cpu_to_le64(0);
3846 bsb.length = __cpu_to_le64(0);
3847 }
3848 bsb.mtime = __cpu_to_le64(time(0));
3849 rv = 0;
3850 for (i = 0; i < dests; i++) {
3851 bsb.devstart = __cpu_to_le64(destoffsets[i]/512);
3852 bsb.sb_csum = bsb_csum((char*)&bsb, ((char*)&bsb.sb_csum)-((char*)&bsb));
3853 if (memcmp(bsb.magic, "md_backup_data-2", 16) == 0)
3854 bsb.sb_csum2 = bsb_csum((char*)&bsb,
3855 ((char*)&bsb.sb_csum2)-((char*)&bsb));
3856 if ((unsigned long long)lseek64(destfd[i], destoffsets[i]-4096, 0) !=
3857 destoffsets[i]-4096)
3858 rv = -1;
3859 if (rv == 0 &&
3860 write(destfd[i], &bsb, 512) != 512)
3861 rv = -1;
3862 fsync(destfd[i]);
3863 }
3864 return rv;
3865 }
3866
3867 static void fail(char *msg)
3868 {
3869 int rv;
3870 rv = (write(2, msg, strlen(msg)) != (int)strlen(msg));
3871 rv |= (write(2, "\n", 1) != 1);
3872 exit(rv ? 1 : 2);
3873 }
3874
3875 static char *abuf, *bbuf;
3876 static unsigned long long abuflen;
3877 static void validate(int afd, int bfd, unsigned long long offset)
3878 {
3879 /* check that the data in the backup against the array.
3880 * This is only used for regression testing and should not
3881 * be used while the array is active
3882 */
3883 if (afd < 0)
3884 return;
3885 lseek64(bfd, offset - 4096, 0);
3886 if (read(bfd, &bsb2, 512) != 512)
3887 fail("cannot read bsb");
3888 if (bsb2.sb_csum != bsb_csum((char*)&bsb2,
3889 ((char*)&bsb2.sb_csum)-((char*)&bsb2)))
3890 fail("first csum bad");
3891 if (memcmp(bsb2.magic, "md_backup_data", 14) != 0)
3892 fail("magic is bad");
3893 if (memcmp(bsb2.magic, "md_backup_data-2", 16) == 0 &&
3894 bsb2.sb_csum2 != bsb_csum((char*)&bsb2,
3895 ((char*)&bsb2.sb_csum2)-((char*)&bsb2)))
3896 fail("second csum bad");
3897
3898 if (__le64_to_cpu(bsb2.devstart)*512 != offset)
3899 fail("devstart is wrong");
3900
3901 if (bsb2.length) {
3902 unsigned long long len = __le64_to_cpu(bsb2.length)*512;
3903
3904 if (abuflen < len) {
3905 free(abuf);
3906 free(bbuf);
3907 abuflen = len;
3908 if (posix_memalign((void**)&abuf, 4096, abuflen) ||
3909 posix_memalign((void**)&bbuf, 4096, abuflen)) {
3910 abuflen = 0;
3911 /* just stop validating on mem-alloc failure */
3912 return;
3913 }
3914 }
3915
3916 lseek64(bfd, offset, 0);
3917 if ((unsigned long long)read(bfd, bbuf, len) != len) {
3918 //printf("len %llu\n", len);
3919 fail("read first backup failed");
3920 }
3921 lseek64(afd, __le64_to_cpu(bsb2.arraystart)*512, 0);
3922 if ((unsigned long long)read(afd, abuf, len) != len)
3923 fail("read first from array failed");
3924 if (memcmp(bbuf, abuf, len) != 0) {
3925 #if 0
3926 int i;
3927 printf("offset=%llu len=%llu\n",
3928 (unsigned long long)__le64_to_cpu(bsb2.arraystart)*512, len);
3929 for (i=0; i<len; i++)
3930 if (bbuf[i] != abuf[i]) {
3931 printf("first diff byte %d\n", i);
3932 break;
3933 }
3934 #endif
3935 fail("data1 compare failed");
3936 }
3937 }
3938 if (bsb2.length2) {
3939 unsigned long long len = __le64_to_cpu(bsb2.length2)*512;
3940
3941 if (abuflen < len) {
3942 free(abuf);
3943 free(bbuf);
3944 abuflen = len;
3945 abuf = xmalloc(abuflen);
3946 bbuf = xmalloc(abuflen);
3947 }
3948
3949 lseek64(bfd, offset+__le64_to_cpu(bsb2.devstart2)*512, 0);
3950 if ((unsigned long long)read(bfd, bbuf, len) != len)
3951 fail("read second backup failed");
3952 lseek64(afd, __le64_to_cpu(bsb2.arraystart2)*512, 0);
3953 if ((unsigned long long)read(afd, abuf, len) != len)
3954 fail("read second from array failed");
3955 if (memcmp(bbuf, abuf, len) != 0)
3956 fail("data2 compare failed");
3957 }
3958 }
3959
3960 int child_monitor(int afd, struct mdinfo *sra, struct reshape *reshape,
3961 struct supertype *st, unsigned long blocks,
3962 int *fds, unsigned long long *offsets,
3963 int dests, int *destfd, unsigned long long *destoffsets)
3964 {
3965 /* Monitor a reshape where backup is being performed using
3966 * 'native' mechanism - either to a backup file, or
3967 * to some space in a spare.
3968 */
3969 char *buf;
3970 int degraded = -1;
3971 unsigned long long speed;
3972 unsigned long long suspend_point, array_size;
3973 unsigned long long backup_point, wait_point;
3974 unsigned long long reshape_completed;
3975 int done = 0;
3976 int increasing = reshape->after.data_disks >= reshape->before.data_disks;
3977 int part = 0; /* The next part of the backup area to fill. It may already
3978 * be full, so we need to check */
3979 int level = reshape->level;
3980 int layout = reshape->before.layout;
3981 int data = reshape->before.data_disks;
3982 int disks = reshape->before.data_disks + reshape->parity;
3983 int chunk = sra->array.chunk_size;
3984 struct mdinfo *sd;
3985 unsigned long stripes;
3986 int uuid[4];
3987
3988 /* set up the backup-super-block. This requires the
3989 * uuid from the array.
3990 */
3991 /* Find a superblock */
3992 for (sd = sra->devs; sd; sd = sd->next) {
3993 char *dn;
3994 int devfd;
3995 int ok;
3996 if (sd->disk.state & (1<<MD_DISK_FAULTY))
3997 continue;
3998 dn = map_dev(sd->disk.major, sd->disk.minor, 1);
3999 devfd = dev_open(dn, O_RDONLY);
4000 if (devfd < 0)
4001 continue;
4002 ok = st->ss->load_super(st, devfd, NULL);
4003 close(devfd);
4004 if (ok == 0)
4005 break;
4006 }
4007 if (!sd) {
4008 pr_err("Cannot find a superblock\n");
4009 return 0;
4010 }
4011
4012 memset(&bsb, 0, 512);
4013 memcpy(bsb.magic, "md_backup_data-1", 16);
4014 st->ss->uuid_from_super(st, uuid);
4015 memcpy(bsb.set_uuid, uuid, 16);
4016 bsb.mtime = __cpu_to_le64(time(0));
4017 bsb.devstart2 = blocks;
4018
4019 stripes = blocks / (sra->array.chunk_size/512) /
4020 reshape->before.data_disks;
4021
4022 if (posix_memalign((void**)&buf, 4096, disks * chunk))
4023 /* Don't start the 'reshape' */
4024 return 0;
4025 if (reshape->before.data_disks == reshape->after.data_disks) {
4026 sysfs_get_ll(sra, NULL, "sync_speed_min", &speed);
4027 sysfs_set_num(sra, NULL, "sync_speed_min", 200000);
4028 }
4029
4030 if (increasing) {
4031 array_size = sra->component_size * reshape->after.data_disks;
4032 backup_point = sra->reshape_progress;
4033 suspend_point = 0;
4034 } else {
4035 array_size = sra->component_size * reshape->before.data_disks;
4036 backup_point = reshape->backup_blocks;
4037 suspend_point = array_size;
4038 }
4039
4040 while (!done) {
4041 int rv;
4042
4043 /* Want to return as soon the oldest backup slot can
4044 * be released as that allows us to start backing up
4045 * some more, providing suspend_point has been
4046 * advanced, which it should have.
4047 */
4048 if (increasing) {
4049 wait_point = array_size;
4050 if (part == 0 && __le64_to_cpu(bsb.length) > 0)
4051 wait_point = (__le64_to_cpu(bsb.arraystart) +
4052 __le64_to_cpu(bsb.length));
4053 if (part == 1 && __le64_to_cpu(bsb.length2) > 0)
4054 wait_point = (__le64_to_cpu(bsb.arraystart2) +
4055 __le64_to_cpu(bsb.length2));
4056 } else {
4057 wait_point = 0;
4058 if (part == 0 && __le64_to_cpu(bsb.length) > 0)
4059 wait_point = __le64_to_cpu(bsb.arraystart);
4060 if (part == 1 && __le64_to_cpu(bsb.length2) > 0)
4061 wait_point = __le64_to_cpu(bsb.arraystart2);
4062 }
4063
4064 rv = progress_reshape(sra, reshape,
4065 backup_point, wait_point,
4066 &suspend_point, &reshape_completed);
4067 /* external metadata would need to ping_monitor here */
4068 sra->reshape_progress = reshape_completed;
4069
4070 /* Clear any backup region that is before 'here' */
4071 if (increasing) {
4072 if (__le64_to_cpu(bsb.length) > 0 &&
4073 reshape_completed >= (__le64_to_cpu(bsb.arraystart) +
4074 __le64_to_cpu(bsb.length)))
4075 forget_backup(dests, destfd,
4076 destoffsets, 0);
4077 if (__le64_to_cpu(bsb.length2) > 0 &&
4078 reshape_completed >= (__le64_to_cpu(bsb.arraystart2) +
4079 __le64_to_cpu(bsb.length2)))
4080 forget_backup(dests, destfd,
4081 destoffsets, 1);
4082 } else {
4083 if (__le64_to_cpu(bsb.length) > 0 &&
4084 reshape_completed <= (__le64_to_cpu(bsb.arraystart)))
4085 forget_backup(dests, destfd,
4086 destoffsets, 0);
4087 if (__le64_to_cpu(bsb.length2) > 0 &&
4088 reshape_completed <= (__le64_to_cpu(bsb.arraystart2)))
4089 forget_backup(dests, destfd,
4090 destoffsets, 1);
4091 }
4092
4093 if (rv < 0) {
4094 if (rv == -1)
4095 done = 1;
4096 break;
4097 }
4098 if (rv == 0 && increasing && !st->ss->external) {
4099 /* No longer need to monitor this reshape */
4100 done = 1;
4101 break;
4102 }
4103
4104 while (rv) {
4105 unsigned long long offset;
4106 unsigned long actual_stripes;
4107 /* Need to backup some data.
4108 * If 'part' is not used and the desired
4109 * backup size is suspended, do a backup,
4110 * then consider the next part.
4111 */
4112 /* Check that 'part' is unused */
4113 if (part == 0 && __le64_to_cpu(bsb.length) != 0)
4114 break;
4115 if (part == 1 && __le64_to_cpu(bsb.length2) != 0)
4116 break;
4117
4118 offset = backup_point / data;
4119 actual_stripes = stripes;
4120 if (increasing) {
4121 if (offset + actual_stripes * (chunk/512) >
4122 sra->component_size)
4123 actual_stripes = ((sra->component_size - offset)
4124 / (chunk/512));
4125 if (offset + actual_stripes * (chunk/512) >
4126 suspend_point/data)
4127 break;
4128 } else {
4129 if (offset < actual_stripes * (chunk/512))
4130 actual_stripes = offset / (chunk/512);
4131 offset -= actual_stripes * (chunk/512);
4132 if (offset < suspend_point/data)
4133 break;
4134 }
4135 if (actual_stripes == 0)
4136 break;
4137 grow_backup(sra, offset, actual_stripes,
4138 fds, offsets,
4139 disks, chunk, level, layout,
4140 dests, destfd, destoffsets,
4141 part, &degraded, buf);
4142 validate(afd, destfd[0], destoffsets[0]);
4143 /* record where 'part' is up to */
4144 part = !part;
4145 if (increasing)
4146 backup_point += actual_stripes * (chunk/512) * data;
4147 else
4148 backup_point -= actual_stripes * (chunk/512) * data;
4149 }
4150 }
4151
4152 /* FIXME maybe call progress_reshape one more time instead */
4153 abort_reshape(sra); /* remove any remaining suspension */
4154 if (reshape->before.data_disks == reshape->after.data_disks)
4155 sysfs_set_num(sra, NULL, "sync_speed_min", speed);
4156 free(buf);
4157 return done;
4158 }
4159
4160 /*
4161 * If any spare contains md_back_data-1 which is recent wrt mtime,
4162 * write that data into the array and update the super blocks with
4163 * the new reshape_progress
4164 */
4165 int Grow_restart(struct supertype *st, struct mdinfo *info, int *fdlist, int cnt,
4166 char *backup_file, int verbose)
4167 {
4168 int i, j;
4169 int old_disks;
4170 unsigned long long *offsets;
4171 unsigned long long nstripe, ostripe;
4172 int ndata, odata;
4173
4174 odata = info->array.raid_disks - info->delta_disks - 1;
4175 if (info->array.level == 6) odata--; /* number of data disks */
4176 ndata = info->array.raid_disks - 1;
4177 if (info->new_level == 6) ndata--;
4178
4179 old_disks = info->array.raid_disks - info->delta_disks;
4180
4181 if (info->delta_disks <= 0)
4182 /* Didn't grow, so the backup file must have
4183 * been used
4184 */
4185 old_disks = cnt;
4186 for (i=old_disks-(backup_file?1:0); i<cnt; i++) {
4187 struct mdinfo dinfo;
4188 int fd;
4189 int bsbsize;
4190 char *devname, namebuf[20];
4191 unsigned long long lo, hi;
4192
4193 /* This was a spare and may have some saved data on it.
4194 * Load the superblock, find and load the
4195 * backup_super_block.
4196 * If either fail, go on to next device.
4197 * If the backup contains no new info, just return
4198 * else restore data and update all superblocks
4199 */
4200 if (i == old_disks-1) {
4201 fd = open(backup_file, O_RDONLY);
4202 if (fd<0) {
4203 pr_err("backup file %s inaccessible: %s\n",
4204 backup_file, strerror(errno));
4205 continue;
4206 }
4207 devname = backup_file;
4208 } else {
4209 fd = fdlist[i];
4210 if (fd < 0)
4211 continue;
4212 if (st->ss->load_super(st, fd, NULL))
4213 continue;
4214
4215 st->ss->getinfo_super(st, &dinfo, NULL);
4216 st->ss->free_super(st);
4217
4218 if (lseek64(fd,
4219 (dinfo.data_offset + dinfo.component_size - 8) <<9,
4220 0) < 0) {
4221 pr_err("Cannot seek on device %d\n", i);
4222 continue; /* Cannot seek */
4223 }
4224 sprintf(namebuf, "device-%d", i);
4225 devname = namebuf;
4226 }
4227 if (read(fd, &bsb, sizeof(bsb)) != sizeof(bsb)) {
4228 if (verbose)
4229 pr_err("Cannot read from %s\n", devname);
4230 continue; /* Cannot read */
4231 }
4232 if (memcmp(bsb.magic, "md_backup_data-1", 16) != 0 &&
4233 memcmp(bsb.magic, "md_backup_data-2", 16) != 0) {
4234 if (verbose)
4235 pr_err("No backup metadata on %s\n", devname);
4236 continue;
4237 }
4238 if (bsb.sb_csum != bsb_csum((char*)&bsb, ((char*)&bsb.sb_csum)-((char*)&bsb))) {
4239 if (verbose)
4240 pr_err("Bad backup-metadata checksum on %s\n", devname);
4241 continue; /* bad checksum */
4242 }
4243 if (memcmp(bsb.magic, "md_backup_data-2", 16) == 0 &&
4244 bsb.sb_csum2 != bsb_csum((char*)&bsb, ((char*)&bsb.sb_csum2)-((char*)&bsb))) {
4245 if (verbose)
4246 pr_err("Bad backup-metadata checksum2 on %s\n", devname);
4247 continue; /* Bad second checksum */
4248 }
4249 if (memcmp(bsb.set_uuid,info->uuid, 16) != 0) {
4250 if (verbose)
4251 pr_err("Wrong uuid on backup-metadata on %s\n", devname);
4252 continue; /* Wrong uuid */
4253 }
4254
4255 /* array utime and backup-mtime should be updated at much the same time, but it seems that
4256 * sometimes they aren't... So allow considerable flexability in matching, and allow
4257 * this test to be overridden by an environment variable.
4258 */
4259 if (info->array.utime > (int)__le64_to_cpu(bsb.mtime) + 2*60*60 ||
4260 info->array.utime < (int)__le64_to_cpu(bsb.mtime) - 10*60) {
4261 if (check_env("MDADM_GROW_ALLOW_OLD")) {
4262 pr_err("accepting backup with timestamp %lu "
4263 "for array with timestamp %lu\n",
4264 (unsigned long)__le64_to_cpu(bsb.mtime),
4265 (unsigned long)info->array.utime);
4266 } else {
4267 pr_err("too-old timestamp on backup-metadata on %s\n", devname);
4268 pr_err("If you think it is should be safe, try 'export MDADM_GROW_ALLOW_OLD=1'\n");
4269 continue; /* time stamp is too bad */
4270 }
4271 }
4272
4273 if (bsb.magic[15] == '1') {
4274 if (bsb.length == 0)
4275 continue;
4276 if (info->delta_disks >= 0) {
4277 /* reshape_progress is increasing */
4278 if (__le64_to_cpu(bsb.arraystart)
4279 + __le64_to_cpu(bsb.length)
4280 < info->reshape_progress) {
4281 nonew:
4282 if (verbose)
4283 pr_err("backup-metadata found on %s but is not needed\n", devname);
4284 continue; /* No new data here */
4285 }
4286 } else {
4287 /* reshape_progress is decreasing */
4288 if (__le64_to_cpu(bsb.arraystart) >=
4289 info->reshape_progress)
4290 goto nonew; /* No new data here */
4291 }
4292 } else {
4293 if (bsb.length == 0 && bsb.length2 == 0)
4294 continue;
4295 if (info->delta_disks >= 0) {
4296 /* reshape_progress is increasing */
4297 if ((__le64_to_cpu(bsb.arraystart)
4298 + __le64_to_cpu(bsb.length)
4299 < info->reshape_progress)
4300 &&
4301 (__le64_to_cpu(bsb.arraystart2)
4302 + __le64_to_cpu(bsb.length2)
4303 < info->reshape_progress))
4304 goto nonew; /* No new data here */
4305 } else {
4306 /* reshape_progress is decreasing */
4307 if (__le64_to_cpu(bsb.arraystart) >=
4308 info->reshape_progress &&
4309 __le64_to_cpu(bsb.arraystart2) >=
4310 info->reshape_progress)
4311 goto nonew; /* No new data here */
4312 }
4313 }
4314 if (lseek64(fd, __le64_to_cpu(bsb.devstart)*512, 0)< 0) {
4315 second_fail:
4316 if (verbose)
4317 pr_err("Failed to verify secondary backup-metadata block on %s\n",
4318 devname);
4319 continue; /* Cannot seek */
4320 }
4321 /* There should be a duplicate backup superblock 4k before here */
4322 if (lseek64(fd, -4096, 1) < 0 ||
4323 read(fd, &bsb2, sizeof(bsb2)) != sizeof(bsb2))
4324 goto second_fail; /* Cannot find leading superblock */
4325 if (bsb.magic[15] == '1')
4326 bsbsize = offsetof(struct mdp_backup_super, pad1);
4327 else
4328 bsbsize = offsetof(struct mdp_backup_super, pad);
4329 if (memcmp(&bsb2, &bsb, bsbsize) != 0)
4330 goto second_fail; /* Cannot find leading superblock */
4331
4332 /* Now need the data offsets for all devices. */
4333 offsets = xmalloc(sizeof(*offsets)*info->array.raid_disks);
4334 for(j=0; j<info->array.raid_disks; j++) {
4335 if (fdlist[j] < 0)
4336 continue;
4337 if (st->ss->load_super(st, fdlist[j], NULL))
4338 /* FIXME should be this be an error */
4339 continue;
4340 st->ss->getinfo_super(st, &dinfo, NULL);
4341 st->ss->free_super(st);
4342 offsets[j] = dinfo.data_offset * 512;
4343 }
4344 printf(Name ": restoring critical section\n");
4345
4346 if (restore_stripes(fdlist, offsets,
4347 info->array.raid_disks,
4348 info->new_chunk,
4349 info->new_level,
4350 info->new_layout,
4351 fd, __le64_to_cpu(bsb.devstart)*512,
4352 __le64_to_cpu(bsb.arraystart)*512,
4353 __le64_to_cpu(bsb.length)*512, NULL)) {
4354 /* didn't succeed, so giveup */
4355 if (verbose)
4356 pr_err("Error restoring backup from %s\n",
4357 devname);
4358 free(offsets);
4359 return 1;
4360 }
4361
4362 if (bsb.magic[15] == '2' &&
4363 restore_stripes(fdlist, offsets,
4364 info->array.raid_disks,
4365 info->new_chunk,
4366 info->new_level,
4367 info->new_layout,
4368 fd, __le64_to_cpu(bsb.devstart)*512 +
4369 __le64_to_cpu(bsb.devstart2)*512,
4370 __le64_to_cpu(bsb.arraystart2)*512,
4371 __le64_to_cpu(bsb.length2)*512, NULL)) {
4372 /* didn't succeed, so giveup */
4373 if (verbose)
4374 pr_err("Error restoring second backup from %s\n",
4375 devname);
4376 free(offsets);
4377 return 1;
4378 }
4379
4380 free(offsets);
4381
4382 /* Ok, so the data is restored. Let's update those superblocks. */
4383
4384 lo = hi = 0;
4385 if (bsb.length) {
4386 lo = __le64_to_cpu(bsb.arraystart);
4387 hi = lo + __le64_to_cpu(bsb.length);
4388 }
4389 if (bsb.magic[15] == '2' && bsb.length2) {
4390 unsigned long long lo1, hi1;
4391 lo1 = __le64_to_cpu(bsb.arraystart2);
4392 hi1 = lo1 + __le64_to_cpu(bsb.length2);
4393 if (lo == hi) {
4394 lo = lo1;
4395 hi = hi1;
4396 } else if (lo < lo1)
4397 hi = hi1;
4398 else
4399 lo = lo1;
4400 }
4401 if (lo < hi &&
4402 (info->reshape_progress < lo ||
4403 info->reshape_progress > hi))
4404 /* backup does not affect reshape_progress*/ ;
4405 else if (info->delta_disks >= 0) {
4406 info->reshape_progress = __le64_to_cpu(bsb.arraystart) +
4407 __le64_to_cpu(bsb.length);
4408 if (bsb.magic[15] == '2') {
4409 unsigned long long p2 = __le64_to_cpu(bsb.arraystart2) +
4410 __le64_to_cpu(bsb.length2);
4411 if (p2 > info->reshape_progress)
4412 info->reshape_progress = p2;
4413 }
4414 } else {
4415 info->reshape_progress = __le64_to_cpu(bsb.arraystart);
4416 if (bsb.magic[15] == '2') {
4417 unsigned long long p2 = __le64_to_cpu(bsb.arraystart2);
4418 if (p2 < info->reshape_progress)
4419 info->reshape_progress = p2;
4420 }
4421 }
4422 for (j=0; j<info->array.raid_disks; j++) {
4423 if (fdlist[j] < 0)
4424 continue;
4425 if (st->ss->load_super(st, fdlist[j], NULL))
4426 continue;
4427 st->ss->getinfo_super(st, &dinfo, NULL);
4428 dinfo.reshape_progress = info->reshape_progress;
4429 st->ss->update_super(st, &dinfo,
4430 "_reshape_progress",
4431 NULL,0, 0, NULL);
4432 st->ss->store_super(st, fdlist[j]);
4433 st->ss->free_super(st);
4434 }
4435 return 0;
4436 }
4437 /* Didn't find any backup data, try to see if any
4438 * was needed.
4439 */
4440 if (info->delta_disks < 0) {
4441 /* When shrinking, the critical section is at the end.
4442 * So see if we are before the critical section.
4443 */
4444 unsigned long long first_block;
4445 nstripe = ostripe = 0;
4446 first_block = 0;
4447 while (ostripe >= nstripe) {
4448 ostripe += info->array.chunk_size / 512;
4449 first_block = ostripe * odata;
4450 nstripe = first_block / ndata / (info->new_chunk/512) *
4451 (info->new_chunk/512);
4452 }
4453
4454 if (info->reshape_progress >= first_block)
4455 return 0;
4456 }
4457 if (info->delta_disks > 0) {
4458 /* See if we are beyond the critical section. */
4459 unsigned long long last_block;
4460 nstripe = ostripe = 0;
4461 last_block = 0;
4462 while (nstripe >= ostripe) {
4463 nstripe += info->new_chunk / 512;
4464 last_block = nstripe * ndata;
4465 ostripe = last_block / odata / (info->array.chunk_size/512) *
4466 (info->array.chunk_size/512);
4467 }
4468
4469 if (info->reshape_progress >= last_block)
4470 return 0;
4471 }
4472 /* needed to recover critical section! */
4473 if (verbose)
4474 pr_err("Failed to find backup of critical section\n");
4475 return 1;
4476 }
4477
4478 int Grow_continue_command(char *devname, int fd,
4479 char *backup_file, int verbose)
4480 {
4481 int ret_val = 0;
4482 struct supertype *st = NULL;
4483 struct mdinfo *content = NULL;
4484 struct mdinfo array;
4485 char *subarray = NULL;
4486 struct mdinfo *cc = NULL;
4487 struct mdstat_ent *mdstat = NULL;
4488 int cfd = -1;
4489 int fd2 = -1;
4490
4491 dprintf("Grow continue from command line called for %s\n",
4492 devname);
4493
4494 st = super_by_fd(fd, &subarray);
4495 if (!st || !st->ss) {
4496 pr_err("Unable to determine metadata format for %s\n",
4497 devname);
4498 return 1;
4499 }
4500 dprintf("Grow continue is run for ");
4501 if (st->ss->external == 0) {
4502 int d;
4503 dprintf("native array (%s)\n", devname);
4504 if (ioctl(fd, GET_ARRAY_INFO, &array.array) < 0) {
4505 pr_err("%s is not an active md array -"
4506 " aborting\n", devname);
4507 ret_val = 1;
4508 goto Grow_continue_command_exit;
4509 }
4510 content = &array;
4511 /* Need to load a superblock.
4512 * FIXME we should really get what we need from
4513 * sysfs
4514 */
4515 for (d = 0; d < MAX_DISKS; d++) {
4516 mdu_disk_info_t disk;
4517 char *dv;
4518 int err;
4519 disk.number = d;
4520 if (ioctl(fd, GET_DISK_INFO, &disk) < 0)
4521 continue;
4522 if (disk.major == 0 && disk.minor == 0)
4523 continue;
4524 if ((disk.state & (1 << MD_DISK_ACTIVE)) == 0)
4525 continue;
4526 dv = map_dev(disk.major, disk.minor, 1);
4527 if (!dv)
4528 continue;
4529 fd2 = dev_open(dv, O_RDONLY);
4530 if (fd2 < 0)
4531 continue;
4532 err = st->ss->load_super(st, fd2, NULL);
4533 close(fd2);
4534 if (err)
4535 continue;
4536 break;
4537 }
4538 if (d == MAX_DISKS) {
4539 pr_err("Unable to load metadata for %s\n",
4540 devname);
4541 ret_val = 1;
4542 goto Grow_continue_command_exit;
4543 }
4544 st->ss->getinfo_super(st, content, NULL);
4545 } else {
4546 char *container;
4547
4548 if (subarray) {
4549 dprintf("subarray (%s)\n", subarray);
4550 container = st->container_devnm;
4551 cfd = open_dev_excl(st->container_devnm);
4552 } else {
4553 container = st->devnm;
4554 close(fd);
4555 cfd = open_dev_excl(st->devnm);
4556 dprintf("container (%s)\n", container);
4557 fd = cfd;
4558 }
4559 if (cfd < 0) {
4560 pr_err("Unable to open container "
4561 "for %s\n", devname);
4562 ret_val = 1;
4563 goto Grow_continue_command_exit;
4564 }
4565
4566 /* find in container array under reshape
4567 */
4568 ret_val = st->ss->load_container(st, cfd, NULL);
4569 if (ret_val) {
4570 pr_err("Cannot read superblock for %s\n",
4571 devname);
4572 ret_val = 1;
4573 goto Grow_continue_command_exit;
4574 }
4575
4576 cc = st->ss->container_content(st, subarray);
4577 for (content = cc; content ; content = content->next) {
4578 char *array;
4579 int allow_reshape = 1;
4580
4581 if (content->reshape_active == 0)
4582 continue;
4583 /* The decision about array or container wide
4584 * reshape is taken in Grow_continue based
4585 * content->reshape_active state, therefore we
4586 * need to check_reshape based on
4587 * reshape_active and subarray name
4588 */
4589 if (content->array.state & (1<<MD_SB_BLOCK_VOLUME))
4590 allow_reshape = 0;
4591 if (content->reshape_active == CONTAINER_RESHAPE &&
4592 (content->array.state
4593 & (1<<MD_SB_BLOCK_CONTAINER_RESHAPE)))
4594 allow_reshape = 0;
4595
4596 if (!allow_reshape) {
4597 pr_err("cannot continue reshape of an array"
4598 " in container with unsupported"
4599 " metadata: %s(%s)\n",
4600 devname, container);
4601 ret_val = 1;
4602 goto Grow_continue_command_exit;
4603 }
4604
4605 array = strchr(content->text_version+1, '/')+1;
4606 mdstat = mdstat_by_subdev(array, container);
4607 if (!mdstat)
4608 continue;
4609 if (mdstat->active == 0) {
4610 pr_err("Skipping inactive array %s.\n",
4611 mdstat->devnm);
4612 free_mdstat(mdstat);
4613 mdstat = NULL;
4614 continue;
4615 }
4616 break;
4617 }
4618 if (!content) {
4619 pr_err("Unable to determine reshaped "
4620 "array for %s\n", devname);
4621 ret_val = 1;
4622 goto Grow_continue_command_exit;
4623 }
4624 fd2 = open_dev(mdstat->devnm);
4625 if (fd2 < 0) {
4626 pr_err("cannot open (%s)\n", mdstat->devnm);
4627 ret_val = 1;
4628 goto Grow_continue_command_exit;
4629 }
4630
4631 sysfs_init(content, fd2, mdstat->devnm);
4632
4633 /* start mdmon in case it is not running
4634 */
4635 if (!mdmon_running(container))
4636 start_mdmon(container);
4637 ping_monitor(container);
4638
4639 if (mdmon_running(container))
4640 st->update_tail = &st->updates;
4641 else {
4642 pr_err("No mdmon found. "
4643 "Grow cannot continue.\n");
4644 ret_val = 1;
4645 goto Grow_continue_command_exit;
4646 }
4647 }
4648
4649 /* verify that array under reshape is started from
4650 * correct position
4651 */
4652 if (verify_reshape_position(content, content->array.level) < 0) {
4653 ret_val = 1;
4654 goto Grow_continue_command_exit;
4655 }
4656
4657 /* continue reshape
4658 */
4659 ret_val = Grow_continue(fd, st, content, backup_file, 0);
4660
4661 Grow_continue_command_exit:
4662 if (fd2 > -1)
4663 close(fd2);
4664 if (cfd > -1)
4665 close(cfd);
4666 st->ss->free_super(st);
4667 free_mdstat(mdstat);
4668 sysfs_free(cc);
4669 free(subarray);
4670
4671 return ret_val;
4672 }
4673
4674 int Grow_continue(int mdfd, struct supertype *st, struct mdinfo *info,
4675 char *backup_file, int freeze_reshape)
4676 {
4677 int ret_val = 2;
4678
4679 if (!info->reshape_active)
4680 return ret_val;
4681
4682 if (st->ss->external) {
4683 int cfd = open_dev(st->container_devnm);
4684
4685 if (cfd < 0)
4686 return 1;
4687
4688 st->ss->load_container(st, cfd, st->container_devnm);
4689 close(cfd);
4690 ret_val = reshape_container(st->container_devnm, NULL, mdfd,
4691 st, info, 0, backup_file,
4692 0,
4693 1 | info->reshape_active,
4694 freeze_reshape);
4695 } else
4696 ret_val = reshape_array(NULL, mdfd, "array", st, info, 1,
4697 NULL, 0ULL, backup_file, 0, 0,
4698 1 | info->reshape_active,
4699 freeze_reshape);
4700
4701 return ret_val;
4702 }
Unnamed repository; edit this file 'description' to name the repository.