2 * mdadm - manage Linux "md" devices aka RAID arrays.
4 * Copyright (C) 2001-2009 Neil Brown <neilb@suse.de>
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.
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.
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
22 * Email: <neilb@suse.de>
28 #if ! defined(__BIG_ENDIAN) && ! defined(__LITTLE_ENDIAN)
29 #error no endian defined
35 #define offsetof(t,f) ((size_t)&(((t*)0)->f))
38 int restore_backup(struct supertype
*st
,
39 struct mdinfo
*content
,
49 int disk_count
= next_spare
+ working_disks
;
51 dprintf("Called restore_backup()\n");
52 fdlist
= xmalloc(sizeof(int) * disk_count
);
54 for (i
= 0; i
< next_spare
; i
++)
56 for (dev
= content
->devs
; dev
; dev
= dev
->next
) {
62 fd
= dev_open(buf
, O_RDWR
);
64 if (dev
->disk
.raid_disk
>= 0)
65 fdlist
[dev
->disk
.raid_disk
] = fd
;
67 fdlist
[next_spare
++] = fd
;
70 if (st
->ss
->external
&& st
->ss
->recover_backup
)
71 err
= st
->ss
->recover_backup(st
, content
);
73 err
= Grow_restart(st
, content
, fdlist
, next_spare
,
74 backup_file
, verbose
> 0);
76 while (next_spare
> 0) {
78 if (fdlist
[next_spare
] >= 0)
79 close(fdlist
[next_spare
]);
83 pr_err("Failed to restore critical"
84 " section for reshape - sorry.\n");
86 pr_err("Possibly you need"
87 " to specify a --backup-file\n");
91 dprintf("restore_backup() returns status OK.\n");
95 int Grow_Add_device(char *devname
, int fd
, char *newdev
)
97 /* Add a device to an active array.
98 * Currently, just extend a linear array.
99 * This requires writing a new superblock on the
100 * new device, calling the kernel to add the device,
101 * and if that succeeds, update the superblock on
103 * This means that we need to *find* all other devices.
110 struct supertype
*st
= NULL
;
111 char *subarray
= NULL
;
113 if (ioctl(fd
, GET_ARRAY_INFO
, &info
.array
) < 0) {
114 pr_err("cannot get array info for %s\n", devname
);
118 if (info
.array
.level
!= -1) {
119 pr_err("can only add devices to linear arrays\n");
123 st
= super_by_fd(fd
, &subarray
);
125 pr_err("cannot handle arrays with superblock version %d\n", info
.array
.major_version
);
130 pr_err("Cannot grow linear sub-arrays yet\n");
136 nfd
= open(newdev
, O_RDWR
|O_EXCL
|O_DIRECT
);
138 pr_err("cannot open %s\n", newdev
);
143 if ((stb
.st_mode
& S_IFMT
) != S_IFBLK
) {
144 pr_err("%s is not a block device!\n", newdev
);
149 /* now check out all the devices and make sure we can read the superblock */
150 for (d
=0 ; d
< info
.array
.raid_disks
; d
++) {
151 mdu_disk_info_t disk
;
154 st
->ss
->free_super(st
);
157 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0) {
158 pr_err("cannot get device detail for device %d\n",
164 dv
= map_dev(disk
.major
, disk
.minor
, 1);
166 pr_err("cannot find device file for device %d\n",
172 fd2
= dev_open(dv
, O_RDWR
);
174 pr_err("cannot open device file %s\n", dv
);
180 if (st
->ss
->load_super(st
, fd2
, NULL
)) {
181 pr_err("cannot find super block on %s\n", dv
);
189 /* Ok, looks good. Lets update the superblock and write it out to
193 info
.disk
.number
= d
;
194 info
.disk
.major
= major(stb
.st_rdev
);
195 info
.disk
.minor
= minor(stb
.st_rdev
);
196 info
.disk
.raid_disk
= d
;
197 info
.disk
.state
= (1 << MD_DISK_SYNC
) | (1 << MD_DISK_ACTIVE
);
198 st
->ss
->update_super(st
, &info
, "linear-grow-new", newdev
,
201 if (st
->ss
->store_super(st
, nfd
)) {
202 pr_err("Cannot store new superblock on %s\n",
209 if (ioctl(fd
, ADD_NEW_DISK
, &info
.disk
) != 0) {
210 pr_err("Cannot add new disk to this array\n");
213 /* Well, that seems to have worked.
214 * Now go through and update all superblocks
217 if (ioctl(fd
, GET_ARRAY_INFO
, &info
.array
) < 0) {
218 pr_err("cannot get array info for %s\n", devname
);
223 for (d
=0 ; d
< info
.array
.raid_disks
; d
++) {
224 mdu_disk_info_t disk
;
228 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0) {
229 pr_err("cannot get device detail for device %d\n",
233 dv
= map_dev(disk
.major
, disk
.minor
, 1);
235 pr_err("cannot find device file for device %d\n",
239 fd2
= dev_open(dv
, O_RDWR
);
241 pr_err("cannot open device file %s\n", dv
);
244 if (st
->ss
->load_super(st
, fd2
, NULL
)) {
245 pr_err("cannot find super block on %s\n", dv
);
249 info
.array
.raid_disks
= nd
+1;
250 info
.array
.nr_disks
= nd
+1;
251 info
.array
.active_disks
= nd
+1;
252 info
.array
.working_disks
= nd
+1;
254 st
->ss
->update_super(st
, &info
, "linear-grow-update", dv
,
257 if (st
->ss
->store_super(st
, fd2
)) {
258 pr_err("Cannot store new superblock on %s\n", dv
);
268 int Grow_addbitmap(char *devname
, int fd
, char *file
, int chunk
, int delay
, int write_behind
, int force
)
271 * First check that array doesn't have a bitmap
272 * Then create the bitmap
275 * For internal bitmaps, we need to check the version,
276 * find all the active devices, and write the bitmap block
279 mdu_bitmap_file_t bmf
;
280 mdu_array_info_t array
;
281 struct supertype
*st
;
282 char *subarray
= NULL
;
283 int major
= BITMAP_MAJOR_HI
;
284 int vers
= md_get_version(fd
);
285 unsigned long long bitmapsize
, array_size
;
288 major
= BITMAP_MAJOR_HOSTENDIAN
;
289 pr_err("Warning - bitmaps created on this kernel"
290 " are not portable\n"
291 " between different architectures. Consider upgrading"
292 " the Linux kernel.\n");
295 if (ioctl(fd
, GET_BITMAP_FILE
, &bmf
) != 0) {
297 pr_err("Memory allocation failure.\n");
299 pr_err("bitmaps not supported by this kernel.\n");
302 if (bmf
.pathname
[0]) {
303 if (strcmp(file
,"none")==0) {
304 if (ioctl(fd
, SET_BITMAP_FILE
, -1)!= 0) {
305 pr_err("failed to remove bitmap %s\n",
311 pr_err("%s already has a bitmap (%s)\n",
312 devname
, bmf
.pathname
);
315 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
316 pr_err("cannot get array status for %s\n", devname
);
319 if (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)) {
320 if (strcmp(file
, "none")==0) {
321 array
.state
&= ~(1<<MD_SB_BITMAP_PRESENT
);
322 if (ioctl(fd
, SET_ARRAY_INFO
, &array
)!= 0) {
323 pr_err("failed to remove internal bitmap.\n");
328 pr_err("Internal bitmap already present on %s\n",
333 if (strcmp(file
, "none") == 0) {
334 pr_err("no bitmap found on %s\n", devname
);
337 if (array
.level
<= 0) {
338 pr_err("Bitmaps not meaningful with level %s\n",
339 map_num(pers
, array
.level
)?:"of this array");
342 bitmapsize
= array
.size
;
344 if (get_dev_size(fd
, NULL
, &array_size
) &&
345 array_size
> (0x7fffffffULL
<<9)) {
346 /* Array is big enough that we cannot trust array.size
347 * try other approaches
349 bitmapsize
= get_component_size(fd
);
351 if (bitmapsize
== 0) {
352 pr_err("Cannot reliably determine size of array to create bitmap - sorry.\n");
356 if (array
.level
== 10) {
357 int ncopies
= (array
.layout
&255)*((array
.layout
>>8)&255);
358 bitmapsize
= bitmapsize
* array
.raid_disks
/ ncopies
;
361 st
= super_by_fd(fd
, &subarray
);
363 pr_err("Cannot understand version %d.%d\n",
364 array
.major_version
, array
.minor_version
);
368 pr_err("Cannot add bitmaps to sub-arrays yet\n");
373 if (strcmp(file
, "internal") == 0) {
376 int offset_setable
= 0;
378 if (st
->ss
->add_internal_bitmap
== NULL
) {
379 pr_err("Internal bitmaps not supported "
380 "with %s metadata\n", st
->ss
->name
);
383 mdi
= sysfs_read(fd
, -1, GET_BITMAP_LOCATION
);
386 for (d
=0; d
< st
->max_devs
; d
++) {
387 mdu_disk_info_t disk
;
390 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0)
392 if (disk
.major
== 0 &&
395 if ((disk
.state
& (1<<MD_DISK_SYNC
))==0)
397 dv
= map_dev(disk
.major
, disk
.minor
, 1);
399 int fd2
= dev_open(dv
, O_RDWR
);
402 if (st
->ss
->load_super(st
, fd2
, NULL
)==0) {
403 if (st
->ss
->add_internal_bitmap(
405 &chunk
, delay
, write_behind
,
406 bitmapsize
, offset_setable
,
409 st
->ss
->write_bitmap(st
, fd2
);
412 "to create internal bitmap - chunksize problem.\n");
420 if (offset_setable
) {
421 st
->ss
->getinfo_super(st
, mdi
, NULL
);
422 sysfs_init(mdi
, fd
, -1);
423 rv
= sysfs_set_num_signed(mdi
, NULL
, "bitmap/location",
426 array
.state
|= (1<<MD_SB_BITMAP_PRESENT
);
427 rv
= ioctl(fd
, SET_ARRAY_INFO
, &array
);
431 pr_err("Cannot add bitmap while array is"
432 " resyncing or reshaping etc.\n");
433 pr_err("failed to set internal bitmap.\n");
440 int max_devs
= st
->max_devs
;
442 /* try to load a superblock */
443 for (d
=0; d
<max_devs
; d
++) {
444 mdu_disk_info_t disk
;
448 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0)
450 if ((disk
.major
==0 && disk
.minor
==0) ||
451 (disk
.state
& (1<<MD_DISK_REMOVED
)))
453 dv
= map_dev(disk
.major
, disk
.minor
, 1);
455 fd2
= dev_open(dv
, O_RDONLY
);
457 if (st
->ss
->load_super(st
, fd2
, NULL
) == 0) {
459 st
->ss
->uuid_from_super(st
, uuid
);
466 pr_err("cannot find UUID for array!\n");
469 if (CreateBitmap(file
, force
, (char*)uuid
, chunk
,
470 delay
, write_behind
, bitmapsize
, major
)) {
473 bitmap_fd
= open(file
, O_RDWR
);
475 pr_err("weird: %s cannot be opened\n",
479 if (ioctl(fd
, SET_BITMAP_FILE
, bitmap_fd
) < 0) {
482 pr_err("Cannot add bitmap while array is"
483 " resyncing or reshaping etc.\n");
484 pr_err("Cannot set bitmap file for %s: %s\n",
485 devname
, strerror(err
));
494 * When reshaping an array we might need to backup some data.
495 * This is written to all spares with a 'super_block' describing it.
496 * The superblock goes 4K from the end of the used space on the
498 * It if written after the backup is complete.
499 * It has the following structure.
502 static struct mdp_backup_super
{
503 char magic
[16]; /* md_backup_data-1 or -2 */
506 /* start/sizes in 512byte sectors */
507 __u64 devstart
; /* address on backup device/file of data */
510 __u32 sb_csum
; /* csum of preceeding bytes. */
512 __u64 devstart2
; /* offset in to data of second section */
515 __u32 sb_csum2
; /* csum of preceeding bytes. */
517 } __attribute__((aligned(512))) bsb
, bsb2
;
519 static __u32
bsb_csum(char *buf
, int len
)
523 for (i
=0; i
<len
; i
++)
524 csum
= (csum
<<3) + buf
[0];
525 return __cpu_to_le32(csum
);
528 static int check_idle(struct supertype
*st
)
530 /* Check that all member arrays for this container, or the
531 * container of this array, are idle
533 int container_dev
= (st
->container_dev
!= NoMdDev
534 ? st
->container_dev
: st
->devnum
);
536 struct mdstat_ent
*ent
, *e
;
539 fmt_devname(container
, container_dev
);
540 ent
= mdstat_read(0, 0);
541 for (e
= ent
; e
; e
= e
->next
) {
542 if (!is_container_member(e
, container
))
544 if (e
->percent
>= 0) {
553 static int freeze_container(struct supertype
*st
)
555 int container_dev
= (st
->container_dev
!= NoMdDev
556 ? st
->container_dev
: st
->devnum
);
562 fmt_devname(container
, container_dev
);
564 if (block_monitor(container
, 1)) {
565 pr_err("failed to freeze container\n");
572 static void unfreeze_container(struct supertype
*st
)
574 int container_dev
= (st
->container_dev
!= NoMdDev
575 ? st
->container_dev
: st
->devnum
);
578 fmt_devname(container
, container_dev
);
580 unblock_monitor(container
, 1);
583 static int freeze(struct supertype
*st
)
585 /* Try to freeze resync/rebuild on this array/container.
586 * Return -1 if the array is busy,
587 * return -2 container cannot be frozen,
588 * return 0 if this kernel doesn't support 'frozen'
589 * return 1 if it worked.
591 if (st
->ss
->external
)
592 return freeze_container(st
);
594 struct mdinfo
*sra
= sysfs_read(-1, st
->devnum
, GET_VERSION
);
600 /* Need to clear any 'read-auto' status */
601 if (sysfs_get_str(sra
, NULL
, "array_state", buf
, 20) > 0 &&
602 strncmp(buf
, "read-auto", 9) == 0)
603 sysfs_set_str(sra
, NULL
, "array_state", "clean");
605 err
= sysfs_freeze_array(sra
);
611 static void unfreeze(struct supertype
*st
)
613 if (st
->ss
->external
)
614 return unfreeze_container(st
);
616 struct mdinfo
*sra
= sysfs_read(-1, st
->devnum
, GET_VERSION
);
619 sysfs_set_str(sra
, NULL
, "sync_action", "idle");
624 static void wait_reshape(struct mdinfo
*sra
)
626 int fd
= sysfs_get_fd(sra
, NULL
, "sync_action");
632 while (sysfs_fd_get_str(fd
, action
, 20) > 0 &&
633 strncmp(action
, "reshape", 7) == 0) {
637 select(fd
+1, NULL
, NULL
, &rfds
, NULL
);
642 static int reshape_super(struct supertype
*st
, long long size
, int level
,
643 int layout
, int chunksize
, int raid_disks
,
644 int delta_disks
, char *backup_file
, char *dev
,
645 int direction
, int verbose
)
647 /* nothing extra to check in the native case */
648 if (!st
->ss
->external
)
650 if (!st
->ss
->reshape_super
||
651 !st
->ss
->manage_reshape
) {
652 pr_err("%s metadata does not support reshape\n",
657 return st
->ss
->reshape_super(st
, size
, level
, layout
, chunksize
,
658 raid_disks
, delta_disks
, backup_file
, dev
,
662 static void sync_metadata(struct supertype
*st
)
664 if (st
->ss
->external
) {
665 if (st
->update_tail
) {
666 flush_metadata_updates(st
);
667 st
->update_tail
= &st
->updates
;
669 st
->ss
->sync_metadata(st
);
673 static int subarray_set_num(char *container
, struct mdinfo
*sra
, char *name
, int n
)
675 /* when dealing with external metadata subarrays we need to be
676 * prepared to handle EAGAIN. The kernel may need to wait for
677 * mdmon to mark the array active so the kernel can handle
678 * allocations/writeback when preparing the reshape action
679 * (md_allow_write()). We temporarily disable safe_mode_delay
680 * to close a race with the array_state going clean before the
681 * next write to raid_disks / stripe_cache_size
686 /* only 'raid_disks' and 'stripe_cache_size' trigger md_allow_write */
688 (strcmp(name
, "raid_disks") != 0 &&
689 strcmp(name
, "stripe_cache_size") != 0))
690 return sysfs_set_num(sra
, NULL
, name
, n
);
692 rc
= sysfs_get_str(sra
, NULL
, "safe_mode_delay", safe
, sizeof(safe
));
695 sysfs_set_num(sra
, NULL
, "safe_mode_delay", 0);
696 rc
= sysfs_set_num(sra
, NULL
, name
, n
);
697 if (rc
< 0 && errno
== EAGAIN
) {
698 ping_monitor(container
);
699 /* if we get EAGAIN here then the monitor is not active
702 rc
= sysfs_set_num(sra
, NULL
, name
, n
);
704 sysfs_set_str(sra
, NULL
, "safe_mode_delay", safe
);
708 int start_reshape(struct mdinfo
*sra
, int already_running
,
709 int before_data_disks
, int data_disks
)
712 unsigned long long sync_max_to_set
;
714 sysfs_set_num(sra
, NULL
, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL
);
715 err
= sysfs_set_num(sra
, NULL
, "suspend_hi", sra
->reshape_progress
);
716 err
= err
?: sysfs_set_num(sra
, NULL
, "suspend_lo",
717 sra
->reshape_progress
);
718 if (before_data_disks
<= data_disks
)
719 sync_max_to_set
= sra
->reshape_progress
/ data_disks
;
721 sync_max_to_set
= (sra
->component_size
* data_disks
722 - sra
->reshape_progress
) / data_disks
;
723 if (!already_running
)
724 sysfs_set_num(sra
, NULL
, "sync_min", sync_max_to_set
);
725 err
= err
?: sysfs_set_num(sra
, NULL
, "sync_max", sync_max_to_set
);
726 if (!already_running
)
727 err
= err
?: sysfs_set_str(sra
, NULL
, "sync_action", "reshape");
732 void abort_reshape(struct mdinfo
*sra
)
734 sysfs_set_str(sra
, NULL
, "sync_action", "idle");
735 sysfs_set_num(sra
, NULL
, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL
);
736 sysfs_set_num(sra
, NULL
, "suspend_hi", 0);
737 sysfs_set_num(sra
, NULL
, "suspend_lo", 0);
738 sysfs_set_num(sra
, NULL
, "sync_min", 0);
739 sysfs_set_str(sra
, NULL
, "sync_max", "max");
742 int remove_disks_for_takeover(struct supertype
*st
,
747 struct mdinfo
*remaining
;
750 if (sra
->array
.level
== 10)
751 nr_of_copies
= layout
& 0xff;
752 else if (sra
->array
.level
== 1)
753 nr_of_copies
= sra
->array
.raid_disks
;
757 remaining
= sra
->devs
;
759 /* for each 'copy', select one device and remove from the list. */
760 for (slot
= 0; slot
< sra
->array
.raid_disks
; slot
+= nr_of_copies
) {
761 struct mdinfo
**diskp
;
764 /* Find a working device to keep */
765 for (diskp
= &remaining
; *diskp
; diskp
= &(*diskp
)->next
) {
766 struct mdinfo
*disk
= *diskp
;
768 if (disk
->disk
.raid_disk
< slot
)
770 if (disk
->disk
.raid_disk
>= slot
+ nr_of_copies
)
772 if (disk
->disk
.state
& (1<<MD_DISK_REMOVED
))
774 if (disk
->disk
.state
& (1<<MD_DISK_FAULTY
))
776 if (!(disk
->disk
.state
& (1<<MD_DISK_SYNC
)))
779 /* We have found a good disk to use! */
781 disk
->next
= sra
->devs
;
790 if (slot
< sra
->array
.raid_disks
) {
791 /* didn't find all slots */
797 sra
->devs
= remaining
;
801 /* Remove all 'remaining' devices from the array */
803 struct mdinfo
*sd
= remaining
;
804 remaining
= sd
->next
;
806 sysfs_set_str(sra
, sd
, "state", "faulty");
807 sysfs_set_str(sra
, sd
, "slot", "none");
808 /* for external metadata disks should be removed in mdmon */
809 if (!st
->ss
->external
)
810 sysfs_set_str(sra
, sd
, "state", "remove");
811 sd
->disk
.state
|= (1<<MD_DISK_REMOVED
);
812 sd
->disk
.state
&= ~(1<<MD_DISK_SYNC
);
813 sd
->next
= sra
->devs
;
819 void reshape_free_fdlist(int *fdlist
,
820 unsigned long long *offsets
,
825 for (i
= 0; i
< size
; i
++)
833 int reshape_prepare_fdlist(char *devname
,
837 unsigned long blocks
,
840 unsigned long long *offsets
)
845 for (d
= 0; d
<= nrdisks
; d
++)
848 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
849 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
851 if (sd
->disk
.state
& (1<<MD_DISK_SYNC
)) {
852 char *dn
= map_dev(sd
->disk
.major
,
854 fdlist
[sd
->disk
.raid_disk
]
855 = dev_open(dn
, O_RDONLY
);
856 offsets
[sd
->disk
.raid_disk
] = sd
->data_offset
*512;
857 if (fdlist
[sd
->disk
.raid_disk
] < 0) {
858 pr_err("%s: cannot open component %s\n",
859 devname
, dn
? dn
: "-unknown-");
863 } else if (backup_file
== NULL
) {
865 char *dn
= map_dev(sd
->disk
.major
,
867 fdlist
[d
] = dev_open(dn
, O_RDWR
);
868 offsets
[d
] = (sd
->data_offset
+ sra
->component_size
- blocks
- 8)*512;
870 pr_err("%s: cannot open component %s\n",
871 devname
, dn
? dn
: "-unknown-");
882 int reshape_open_backup_file(char *backup_file
,
887 unsigned long long *offsets
,
890 /* Return 1 on success, 0 on any form of failure */
891 /* need to check backup file is large enough */
897 *fdlist
= open(backup_file
, O_RDWR
|O_CREAT
|(restart
? O_TRUNC
: O_EXCL
),
901 pr_err("%s: cannot create backup file %s: %s\n",
902 devname
, backup_file
, strerror(errno
));
905 /* Guard against backup file being on array device.
906 * If array is partitioned or if LVM etc is in the
907 * way this will not notice, but it is better than
910 fstat(*fdlist
, &stb
);
913 if (stb
.st_rdev
== dev
) {
914 pr_err("backup file must NOT be"
915 " on the array being reshaped.\n");
921 for (i
=0; i
< blocks
+ 8 ; i
++) {
922 if (write(*fdlist
, buf
, 512) != 512) {
923 pr_err("%s: cannot create"
924 " backup file %s: %s\n",
925 devname
, backup_file
, strerror(errno
));
929 if (fsync(*fdlist
) != 0) {
930 pr_err("%s: cannot create backup file %s: %s\n",
931 devname
, backup_file
, strerror(errno
));
938 unsigned long compute_backup_blocks(int nchunk
, int ochunk
,
939 unsigned int ndata
, unsigned int odata
)
941 unsigned long a
, b
, blocks
;
942 /* So how much do we need to backup.
943 * We need an amount of data which is both a whole number of
944 * old stripes and a whole number of new stripes.
945 * So LCM for (chunksize*datadisks).
947 a
= (ochunk
/512) * odata
;
948 b
= (nchunk
/512) * ndata
;
956 /* LCM == product / GCD */
957 blocks
= (ochunk
/512) * (nchunk
/512) * odata
* ndata
/ a
;
962 char *analyse_change(struct mdinfo
*info
, struct reshape
*re
)
964 /* Based on the current array state in info->array and
965 * the changes in info->new_* etc, determine:
966 * - whether the change is possible
967 * - Intermediate level/raid_disks/layout
968 * - whether a restriping reshape is needed
969 * - number of sectors in minimum change unit. This
970 * will cover a whole number of stripes in 'before' and
973 * Return message if the change should be rejected
974 * NULL if the change can be achieved
976 * This can be called as part of starting a reshape, or
977 * when assembling an array that is undergoing reshape.
980 /* delta_parity records change in number of devices
981 * caused by level change
983 int delta_parity
= 0;
985 /* If a new level not explicitly given, we assume no-change */
986 if (info
->new_level
== UnSet
)
987 info
->new_level
= info
->array
.level
;
990 switch (info
->new_level
) {
996 /* chunk size is meaningful, must divide component_size
999 if (info
->component_size
% (info
->new_chunk
/512))
1000 return "New chunk size does not"
1001 " divide component size";
1004 return "chunk size not meaningful for this level";
1007 info
->new_chunk
= info
->array
.chunk_size
;
1009 switch (info
->array
.level
) {
1011 return "Cannot understand this RAID level";
1013 /* RAID1 can convert to RAID1 with different disks, or
1014 * raid5 with 2 disks, or
1017 if (info
->new_level
> 1 &&
1018 (info
->component_size
& 7))
1019 return "Cannot convert RAID1 of this size - "
1020 "reduce size to multiple of 4K first.";
1021 if (info
->new_level
== 0) {
1022 if (info
->delta_disks
!= UnSet
&&
1023 info
->delta_disks
!= 0)
1024 return "Cannot change number of disks "
1025 "with RAID1->RAID0 conversion";
1027 re
->before
.data_disks
= 1;
1028 re
->after
.data_disks
= 1;
1029 re
->before
.layout
= 0;
1030 re
->backup_blocks
= 0;
1034 if (info
->new_level
== 1) {
1035 if (info
->delta_disks
== UnSet
)
1036 /* Don't know what to do */
1037 return "no change requested for Growing RAID1";
1039 re
->backup_blocks
= 0;
1043 if (info
->array
.raid_disks
== 2 &&
1044 info
->new_level
== 5) {
1047 re
->before
.data_disks
= 1;
1048 if (info
->delta_disks
!= UnSet
&&
1049 info
->delta_disks
!= 0)
1050 re
->after
.data_disks
= 1 + info
->delta_disks
;
1052 re
->after
.data_disks
= 1;
1053 if (re
->after
.data_disks
< 1)
1054 return "Number of disks too small for RAID5";
1056 re
->before
.layout
= ALGORITHM_LEFT_SYMMETRIC
;
1057 info
->array
.chunk_size
= 65536;
1060 /* Could do some multi-stage conversions, but leave that to
1063 return "Impossibly level change request for RAID1";
1066 /* RAID10 can only be converted from near mode to
1067 * RAID0 by removing some devices
1069 if ((info
->array
.layout
& ~0xff) != 0x100)
1070 return "Cannot Grow RAID10 with far/offset layout";
1071 /* number of devices must be multiple of number of copies */
1072 if (info
->array
.raid_disks
% (info
->array
.layout
& 0xff))
1073 return "RAID10 layout too complex for Grow operation";
1075 if (info
->new_level
!= 0)
1076 return "RAID10 can only be changed to RAID0";
1077 new_disks
= (info
->array
.raid_disks
1078 / (info
->array
.layout
& 0xff));
1079 if (info
->delta_disks
== UnSet
)
1080 info
->delta_disks
= (new_disks
1081 - info
->array
.raid_disks
);
1083 if (info
->delta_disks
!= new_disks
- info
->array
.raid_disks
)
1084 return "New number of raid-devices impossible for RAID10";
1085 if (info
->new_chunk
&&
1086 info
->new_chunk
!= info
->array
.chunk_size
)
1087 return "Cannot change chunk-size with RAID10 Grow";
1092 re
->before
.data_disks
= new_disks
;
1093 re
->after
.data_disks
= re
->before
.data_disks
;
1094 re
->before
.layout
= 0;
1095 re
->backup_blocks
= 0;
1099 /* RAID0 can be converted to RAID10, or to RAID456 */
1100 if (info
->new_level
== 10) {
1101 if (info
->new_layout
== UnSet
&& info
->delta_disks
== UnSet
) {
1102 /* Assume near=2 layout */
1103 info
->new_layout
= 0x102;
1104 info
->delta_disks
= info
->array
.raid_disks
;
1106 if (info
->new_layout
== UnSet
) {
1107 int copies
= 1 + (info
->delta_disks
1108 / info
->array
.raid_disks
);
1109 if (info
->array
.raid_disks
* (copies
-1)
1110 != info
->delta_disks
)
1111 return "Impossible number of devices"
1112 " for RAID0->RAID10";
1113 info
->new_layout
= 0x100 + copies
;
1115 if (info
->delta_disks
== UnSet
) {
1116 int copies
= info
->new_layout
& 0xff;
1117 if (info
->new_layout
!= 0x100 + copies
)
1118 return "New layout impossible"
1119 " for RAID0->RAID10";;
1120 info
->delta_disks
= (copies
- 1) *
1121 info
->array
.raid_disks
;
1123 if (info
->new_chunk
&&
1124 info
->new_chunk
!= info
->array
.chunk_size
)
1125 return "Cannot change chunk-size with RAID0->RAID10";
1129 re
->before
.data_disks
= (info
->array
.raid_disks
+
1131 re
->after
.data_disks
= re
->before
.data_disks
;
1132 re
->before
.layout
= info
->new_layout
;
1133 re
->backup_blocks
= 0;
1137 /* RAID0 can also covert to RAID0/4/5/6 by first converting to
1138 * a raid4 style layout of the final level.
1140 switch (info
->new_level
) {
1145 re
->before
.layout
= 0;
1150 re
->before
.layout
= ALGORITHM_PARITY_N
;
1155 re
->before
.layout
= ALGORITHM_PARITY_N
;
1158 return "Impossible level change requested";
1160 re
->before
.data_disks
= info
->array
.raid_disks
;
1161 /* determining 'after' layout happens outside this 'switch' */
1165 info
->array
.layout
= ALGORITHM_PARITY_N
;
1167 switch (info
->new_level
) {
1171 re
->level
= info
->array
.level
;
1172 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1173 re
->before
.layout
= info
->array
.layout
;
1177 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1178 re
->before
.layout
= info
->array
.layout
;
1183 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1184 switch (info
->array
.layout
) {
1185 case ALGORITHM_LEFT_ASYMMETRIC
:
1186 re
->before
.layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
1188 case ALGORITHM_RIGHT_ASYMMETRIC
:
1189 re
->before
.layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
1191 case ALGORITHM_LEFT_SYMMETRIC
:
1192 re
->before
.layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
1194 case ALGORITHM_RIGHT_SYMMETRIC
:
1195 re
->before
.layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
1197 case ALGORITHM_PARITY_0
:
1198 re
->before
.layout
= ALGORITHM_PARITY_0_6
;
1200 case ALGORITHM_PARITY_N
:
1201 re
->before
.layout
= ALGORITHM_PARITY_N_6
;
1204 return "Cannot convert an array with this layout";
1208 if (info
->array
.raid_disks
!= 2)
1209 return "Can only convert a 2-device array to RAID1";
1210 if (info
->delta_disks
!= UnSet
&&
1211 info
->delta_disks
!= 0)
1212 return "Cannot set raid_disk when "
1213 "converting RAID5->RAID1";
1217 return "Impossible level change requested";
1221 switch (info
->new_level
) {
1227 re
->before
.data_disks
= info
->array
.raid_disks
- 2;
1228 re
->before
.layout
= info
->array
.layout
;
1231 return "Impossible level change requested";
1236 /* If we reached here then it looks like a re-stripe is
1237 * happening. We have determined the intermediate level
1238 * and initial raid_disks/layout and stored these in 're'.
1240 * We need to deduce the final layout that can be atomically
1241 * converted to the end state.
1243 switch (info
->new_level
) {
1245 /* We can only get to RAID0 from RAID4 or RAID5
1246 * with appropriate layout and one extra device
1248 if (re
->level
!= 4 && re
->level
!= 5)
1249 return "Cannot covert to RAID0 from this level";
1251 switch (re
->level
) {
1253 re
->after
.layout
= 0 ; break;
1255 re
->after
.layout
= ALGORITHM_PARITY_N
; break;
1260 /* We can only get to RAID4 from RAID5 */
1261 if (re
->level
!= 4 && re
->level
!= 5)
1262 return "Cannot convert to RAID4 from this level";
1264 switch (re
->level
) {
1266 re
->after
.layout
= 0 ; break;
1268 re
->after
.layout
= ALGORITHM_PARITY_N
; break;
1273 /* We get to RAID5 from RAID5 or RAID6 */
1274 if (re
->level
!= 5 && re
->level
!= 6)
1275 return "Cannot convert to RAID5 from this level";
1277 switch (re
->level
) {
1279 if (info
->new_layout
== UnSet
)
1280 re
->after
.layout
= re
->before
.layout
;
1282 re
->after
.layout
= info
->new_layout
;
1285 if (info
->new_layout
== UnSet
)
1286 info
->new_layout
= re
->before
.layout
;
1288 /* after.layout needs to be raid6 version of new_layout */
1289 if (info
->new_layout
== ALGORITHM_PARITY_N
)
1290 re
->after
.layout
= ALGORITHM_PARITY_N
;
1293 char *ls
= map_num(r5layout
, info
->new_layout
);
1296 /* Current RAID6 layout has a RAID5
1299 strcat(strcpy(layout
, ls
), "-6");
1300 l
= map_name(r6layout
, layout
);
1302 return "Cannot find RAID6 layout"
1305 /* Current RAID6 has no equivalent.
1306 * If it is already a '-6' layout we
1307 * can leave it unchanged, else we must
1310 ls
= map_num(r6layout
, info
->new_layout
);
1312 strcmp(ls
+strlen(ls
)-2, "-6") != 0)
1313 return "Please specify new layout";
1314 l
= info
->new_layout
;
1316 re
->after
.layout
= l
;
1322 /* We must already be at level 6 */
1324 return "Impossible level change";
1325 if (info
->new_layout
== UnSet
)
1326 re
->after
.layout
= info
->array
.layout
;
1328 re
->after
.layout
= info
->new_layout
;
1331 return "Impossible level change requested";
1333 if (info
->delta_disks
== UnSet
)
1334 info
->delta_disks
= delta_parity
;
1336 re
->after
.data_disks
= (re
->before
.data_disks
1339 switch (re
->level
) {
1340 case 6: re
->parity
= 2; break;
1342 case 5: re
->parity
= 1; break;
1343 default: re
->parity
= 0; break;
1345 /* So we have a restripe operation, we need to calculate the number
1346 * of blocks per reshape operation.
1348 if (info
->new_chunk
== 0)
1349 info
->new_chunk
= info
->array
.chunk_size
;
1350 if (re
->after
.data_disks
== re
->before
.data_disks
&&
1351 re
->after
.layout
== re
->before
.layout
&&
1352 info
->new_chunk
== info
->array
.chunk_size
) {
1353 /* Nothing to change */
1354 re
->backup_blocks
= 0;
1357 if (re
->after
.data_disks
== 1 && re
->before
.data_disks
== 1) {
1358 /* chunk and layout changes make no difference */
1359 re
->backup_blocks
= 0;
1363 if (re
->after
.data_disks
== re
->before
.data_disks
&&
1364 get_linux_version() < 2006032)
1365 return "in-place reshape is not safe before 2.6.32 - sorry.";
1367 if (re
->after
.data_disks
< re
->before
.data_disks
&&
1368 get_linux_version() < 2006030)
1369 return "reshape to fewer devices is not supported before 2.6.30 - sorry.";
1371 re
->backup_blocks
= compute_backup_blocks(
1372 info
->new_chunk
, info
->array
.chunk_size
,
1373 re
->after
.data_disks
,
1374 re
->before
.data_disks
);
1376 re
->new_size
= info
->component_size
* re
->after
.data_disks
;
1380 static int set_array_size(struct supertype
*st
, struct mdinfo
*sra
,
1383 struct mdinfo
*info
;
1387 if ((st
== NULL
) || (sra
== NULL
))
1390 if (text_version
== NULL
)
1391 text_version
= sra
->text_version
;
1392 subarray
= strchr(text_version
+1, '/')+1;
1393 info
= st
->ss
->container_content(st
, subarray
);
1395 unsigned long long current_size
= 0;
1396 unsigned long long new_size
=
1397 info
->custom_array_size
/2;
1399 if (sysfs_get_ll(sra
, NULL
, "array_size", ¤t_size
) == 0 &&
1400 new_size
> current_size
) {
1401 if (sysfs_set_num(sra
, NULL
, "array_size", new_size
)
1403 dprintf("Error: Cannot set array size");
1406 dprintf("Array size changed");
1408 dprintf(" from %llu to %llu.\n",
1409 current_size
, new_size
);
1413 dprintf("Error: set_array_size(): info pointer in NULL\n");
1418 static int reshape_array(char *container
, int fd
, char *devname
,
1419 struct supertype
*st
, struct mdinfo
*info
,
1420 int force
, struct mddev_dev
*devlist
,
1421 char *backup_file
, int quiet
, int forked
,
1422 int restart
, int freeze_reshape
);
1423 static int reshape_container(char *container
, char *devname
,
1425 struct supertype
*st
,
1426 struct mdinfo
*info
,
1429 int quiet
, int restart
, int freeze_reshape
);
1431 int Grow_reshape(char *devname
, int fd
, int quiet
, char *backup_file
,
1433 int level
, char *layout_str
, int chunksize
, int raid_disks
,
1434 struct mddev_dev
*devlist
,
1435 int assume_clean
, int force
)
1437 /* Make some changes in the shape of an array.
1438 * The kernel must support the change.
1440 * There are three different changes. Each can trigger
1441 * a resync or recovery so we freeze that until we have
1442 * requested everything (if kernel supports freezing - 2.6.30).
1444 * - change size (i.e. component_size)
1446 * - change layout/chunksize/ndisks
1448 * The last can require a reshape. It is different on different
1449 * levels so we need to check the level before actioning it.
1450 * Some times the level change needs to be requested after the
1451 * reshape (e.g. raid6->raid5, raid5->raid0)
1454 struct mdu_array_info_s array
;
1456 struct supertype
*st
;
1457 char *subarray
= NULL
;
1461 char *container
= NULL
;
1462 char container_buf
[20];
1465 struct mddev_dev
*dv
;
1471 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) < 0) {
1472 pr_err("%s is not an active md array - aborting\n",
1478 (chunksize
|| level
!= UnSet
|| layout_str
|| raid_disks
)) {
1479 pr_err("cannot change component size at the same time "
1480 "as other changes.\n"
1481 " Change size first, then check data is intact before "
1482 "making other changes.\n");
1486 if (raid_disks
&& raid_disks
< array
.raid_disks
&& array
.level
> 1 &&
1487 get_linux_version() < 2006032 &&
1488 !check_env("MDADM_FORCE_FEWER")) {
1489 pr_err("reducing the number of devices is not safe before Linux 2.6.32\n"
1490 " Please use a newer kernel\n");
1494 st
= super_by_fd(fd
, &subarray
);
1496 pr_err("Unable to determine metadata format for %s\n", devname
);
1499 if (raid_disks
> st
->max_devs
) {
1500 pr_err("Cannot increase raid-disks on this array"
1501 " beyond %d\n", st
->max_devs
);
1505 /* in the external case we need to check that the requested reshape is
1506 * supported, and perform an initial check that the container holds the
1507 * pre-requisite spare devices (mdmon owns final validation)
1509 if (st
->ss
->external
) {
1514 container_dev
= st
->container_dev
;
1515 cfd
= open_dev_excl(st
->container_dev
);
1517 container_dev
= st
->devnum
;
1519 cfd
= open_dev_excl(st
->devnum
);
1523 pr_err("Unable to open container for %s\n",
1529 fmt_devname(container_buf
, container_dev
);
1530 container
= container_buf
;
1532 rv
= st
->ss
->load_container(st
, cfd
, NULL
);
1535 pr_err("Cannot read superblock for %s\n",
1541 /* check if operation is supported for metadata handler */
1542 if (st
->ss
->container_content
) {
1543 struct mdinfo
*cc
= NULL
;
1544 struct mdinfo
*content
= NULL
;
1546 cc
= st
->ss
->container_content(st
, subarray
);
1547 for (content
= cc
; content
; content
= content
->next
) {
1548 int allow_reshape
= 1;
1550 /* check if reshape is allowed based on metadata
1551 * indications stored in content.array.status
1553 if (content
->array
.state
& (1<<MD_SB_BLOCK_VOLUME
))
1555 if (content
->array
.state
1556 & (1<<MD_SB_BLOCK_CONTAINER_RESHAPE
))
1558 if (!allow_reshape
) {
1559 pr_err("cannot reshape arrays in"
1560 " container with unsupported"
1561 " metadata: %s(%s)\n",
1562 devname
, container_buf
);
1570 if (mdmon_running(container_dev
))
1571 st
->update_tail
= &st
->updates
;
1575 for (dv
= devlist
; dv
; dv
= dv
->next
)
1577 if (raid_disks
> array
.raid_disks
&&
1578 array
.spare_disks
+added_disks
< (raid_disks
- array
.raid_disks
) &&
1580 pr_err("Need %d spare%s to avoid degraded array,"
1581 " and only have %d.\n"
1582 " Use --force to over-ride this check.\n",
1583 raid_disks
- array
.raid_disks
,
1584 raid_disks
- array
.raid_disks
== 1 ? "" : "s",
1585 array
.spare_disks
+ added_disks
);
1589 sra
= sysfs_read(fd
, 0, GET_LEVEL
| GET_DISKS
| GET_DEVS
1590 | GET_STATE
| GET_VERSION
);
1592 if (st
->ss
->external
&& subarray
== NULL
) {
1593 array
.level
= LEVEL_CONTAINER
;
1594 sra
->array
.level
= LEVEL_CONTAINER
;
1597 pr_err("failed to read sysfs parameters for %s\n",
1601 frozen
= freeze(st
);
1603 /* freeze() already spewed the reason */
1606 } else if (frozen
< 0) {
1607 pr_err("%s is performing resync/recovery and cannot"
1608 " be reshaped\n", devname
);
1613 /* ========= set size =============== */
1614 if (size
>= 0 && (size
== 0 || size
!= array
.size
)) {
1615 long long orig_size
= get_component_size(fd
)/2;
1616 long long min_csize
;
1618 int raid0_takeover
= 0;
1621 orig_size
= array
.size
;
1623 if (reshape_super(st
, size
, UnSet
, UnSet
, 0, 0, UnSet
, NULL
,
1624 devname
, APPLY_METADATA_CHANGES
, !quiet
)) {
1629 if (st
->ss
->external
) {
1630 /* metadata can have size limitation
1631 * update size value according to metadata information
1633 struct mdinfo
*sizeinfo
=
1634 st
->ss
->container_content(st
, subarray
);
1636 unsigned long long new_size
=
1637 sizeinfo
->custom_array_size
/2;
1638 int data_disks
= get_data_disks(
1639 sizeinfo
->array
.level
,
1640 sizeinfo
->array
.layout
,
1641 sizeinfo
->array
.raid_disks
);
1642 new_size
/= data_disks
;
1643 dprintf("Metadata size correction from %llu to "
1644 "%llu (%llu)\n", orig_size
, new_size
,
1645 new_size
* data_disks
);
1647 sysfs_free(sizeinfo
);
1651 /* Update the size of each member device in case
1652 * they have been resized. This will never reduce
1653 * below the current used-size. The "size" attribute
1654 * understands '0' to mean 'max'.
1658 for (mdi
= sra
->devs
; mdi
; mdi
= mdi
->next
) {
1659 if (sysfs_set_num(sra
, mdi
, "size", size
) < 0) {
1660 /* Probably kernel refusing to let us
1661 * reduce the size - not an error.
1665 if (array
.not_persistent
== 0 &&
1666 array
.major_version
== 0 &&
1667 get_linux_version() < 3001000) {
1668 /* Dangerous to allow size to exceed 2TB */
1669 unsigned long long csize
;
1670 if (sysfs_get_ll(sra
, mdi
, "size", &csize
) == 0) {
1671 if (csize
>= 2ULL*1024*1024*1024)
1672 csize
= 2ULL*1024*1024*1024;
1673 if ((min_csize
== 0 || (min_csize
1674 > (long long)csize
)))
1680 pr_err("Cannot set size on "
1681 "array members.\n");
1682 goto size_change_error
;
1684 if (min_csize
&& size
> min_csize
) {
1685 pr_err("Cannot safely make this array "
1686 "use more than 2TB per device on this kernel.\n");
1688 goto size_change_error
;
1690 if (min_csize
&& size
== 0) {
1691 /* Don't let the kernel choose a size - it will get
1694 pr_err("Limited v0.90 array to "
1695 "2TB per device\n");
1698 if (st
->ss
->external
) {
1699 if (sra
->array
.level
== 0) {
1700 rv
= sysfs_set_str(sra
, NULL
, "level",
1704 /* get array parametes after takeover
1705 * to chane one parameter at time only
1707 rv
= ioctl(fd
, GET_ARRAY_INFO
, &array
);
1710 /* make sure mdmon is
1711 * aware of the new level */
1712 if (!mdmon_running(st
->container_dev
))
1713 start_mdmon(st
->container_dev
);
1714 ping_monitor(container
);
1715 if (mdmon_running(st
->container_dev
) &&
1716 st
->update_tail
== NULL
)
1717 st
->update_tail
= &st
->updates
;
1721 if (array
.size
!= size
) {
1722 /* got truncated to 32bit, write to
1723 * component_size instead
1726 rv
= sysfs_set_num(sra
, NULL
,
1727 "component_size", size
);
1731 rv
= ioctl(fd
, SET_ARRAY_INFO
, &array
);
1733 /* manage array size when it is managed externally
1735 if ((rv
== 0) && st
->ss
->external
)
1736 rv
= set_array_size(st
, sra
, sra
->text_version
);
1739 if (raid0_takeover
) {
1740 /* do not recync non-existing parity,
1741 * we will drop it anyway
1743 sysfs_set_str(sra
, NULL
, "sync_action", "frozen");
1744 /* go back to raid0, drop parity disk
1746 sysfs_set_str(sra
, NULL
, "level", "raid0");
1747 ioctl(fd
, GET_ARRAY_INFO
, &array
);
1754 /* restore metadata */
1755 if (reshape_super(st
, orig_size
, UnSet
, UnSet
, 0, 0,
1756 UnSet
, NULL
, devname
,
1757 ROLLBACK_METADATA_CHANGES
,
1760 pr_err("Cannot set device size for %s: %s\n",
1761 devname
, strerror(err
));
1763 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
1764 cont_err("Bitmap must be removed before size can be changed\n");
1769 /* This will fail on kernels newer than 3.0 unless
1770 * a backport has been arranged.
1773 sysfs_set_str(sra
, NULL
, "resync_start", "none") < 0)
1774 pr_err("--assume-clean not support with --grow on this kernel\n");
1776 ioctl(fd
, GET_ARRAY_INFO
, &array
);
1777 size
= get_component_size(fd
)/2;
1781 if (size
== orig_size
)
1782 pr_err("component size of %s "
1783 "unchanged at %lluK\n",
1786 pr_err("component size of %s "
1787 "has been set to %lluK\n",
1791 } else if (array
.level
!= LEVEL_CONTAINER
) {
1792 size
= get_component_size(fd
)/2;
1797 /* See if there is anything else to do */
1798 if ((level
== UnSet
|| level
== array
.level
) &&
1799 (layout_str
== NULL
) &&
1800 (chunksize
== 0 || chunksize
== array
.chunk_size
) &&
1801 (raid_disks
== 0 || raid_disks
== array
.raid_disks
)) {
1802 /* Nothing more to do */
1803 if (!changed
&& !quiet
)
1804 pr_err("%s: no change requested\n",
1809 /* ========= check for Raid10/Raid1 -> Raid0 conversion ===============
1810 * current implementation assumes that following conditions must be met:
1813 * - near_copies == 2
1815 if ((level
== 0 && array
.level
== 10 && sra
&&
1816 array
.layout
== ((1 << 8) + 2) && !(array
.raid_disks
& 1)) ||
1817 (level
== 0 && array
.level
== 1 && sra
)) {
1819 err
= remove_disks_for_takeover(st
, sra
, array
.layout
);
1821 dprintf(Name
": Array cannot be reshaped\n");
1827 /* Make sure mdmon has seen the device removal
1828 * and updated metadata before we continue with
1832 ping_monitor(container
);
1835 memset(&info
, 0, sizeof(info
));
1837 sysfs_init(&info
, fd
, NoMdDev
);
1838 strcpy(info
.text_version
, sra
->text_version
);
1839 info
.component_size
= size
*2;
1840 info
.new_level
= level
;
1841 info
.new_chunk
= chunksize
* 1024;
1842 if (info
.array
.level
== LEVEL_CONTAINER
) {
1843 info
.delta_disks
= UnSet
;
1844 info
.array
.raid_disks
= raid_disks
;
1845 } else if (raid_disks
)
1846 info
.delta_disks
= raid_disks
- info
.array
.raid_disks
;
1848 info
.delta_disks
= UnSet
;
1849 if (layout_str
== NULL
) {
1850 info
.new_layout
= UnSet
;
1851 if (info
.array
.level
== 6 &&
1852 (info
.new_level
== 6 || info
.new_level
== UnSet
) &&
1853 info
.array
.layout
>= 16) {
1854 pr_err("%s has a non-standard layout. If you"
1855 " wish to preserve this\n", devname
);
1856 cont_err("during the reshape, please specify"
1857 " --layout=preserve\n");
1858 cont_err("If you want to change it, specify a"
1859 " layout or use --layout=normalise\n");
1863 } else if (strcmp(layout_str
, "normalise") == 0 ||
1864 strcmp(layout_str
, "normalize") == 0) {
1865 /* If we have a -6 RAID6 layout, remove the '-6'. */
1866 info
.new_layout
= UnSet
;
1867 if (info
.array
.level
== 6 && info
.new_level
== UnSet
) {
1869 strcpy(l
, map_num(r6layout
, info
.array
.layout
));
1870 h
= strrchr(l
, '-');
1871 if (h
&& strcmp(h
, "-6") == 0) {
1873 info
.new_layout
= map_name(r6layout
, l
);
1876 pr_err("%s is only meaningful when reshaping"
1877 " a RAID6 array.\n", layout_str
);
1881 } else if (strcmp(layout_str
, "preserve") == 0) {
1882 /* This means that a non-standard RAID6 layout
1885 * - When reshape a RAID6 (e.g. adding a device)
1886 * which is in a non-standard layout, it is OK
1887 * to preserve that layout.
1888 * - When converting a RAID5 to RAID6, leave it in
1889 * the XXX-6 layout, don't re-layout.
1891 if (info
.array
.level
== 6 && info
.new_level
== UnSet
)
1892 info
.new_layout
= info
.array
.layout
;
1893 else if (info
.array
.level
== 5 && info
.new_level
== 6) {
1895 strcpy(l
, map_num(r5layout
, info
.array
.layout
));
1897 info
.new_layout
= map_name(r6layout
, l
);
1899 pr_err("%s in only meaningful when reshaping"
1900 " to RAID6\n", layout_str
);
1905 int l
= info
.new_level
;
1907 l
= info
.array
.level
;
1910 info
.new_layout
= map_name(r5layout
, layout_str
);
1913 info
.new_layout
= map_name(r6layout
, layout_str
);
1916 info
.new_layout
= parse_layout_10(layout_str
);
1919 info
.new_layout
= parse_layout_faulty(layout_str
);
1922 pr_err("layout not meaningful"
1923 " with this level\n");
1927 if (info
.new_layout
== UnSet
) {
1928 pr_err("layout %s not understood"
1929 " for this level\n",
1936 if (array
.level
== LEVEL_FAULTY
) {
1937 if (level
!= UnSet
&& level
!= array
.level
) {
1938 pr_err("cannot change level of Faulty device\n");
1942 pr_err("cannot set chunksize of Faulty device\n");
1945 if (raid_disks
&& raid_disks
!= 1) {
1946 pr_err("cannot set raid_disks of Faulty device\n");
1950 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
1951 dprintf("Cannot get array information.\n");
1954 array
.layout
= info
.new_layout
;
1955 if (ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
1956 pr_err("failed to set new layout\n");
1959 printf("layout for %s set to %d\n",
1960 devname
, array
.layout
);
1962 } else if (array
.level
== LEVEL_CONTAINER
) {
1963 /* This change is to be applied to every array in the
1964 * container. This is only needed when the metadata imposes
1965 * restraints of the various arrays in the container.
1966 * Currently we only know that IMSM requires all arrays
1967 * to have the same number of devices so changing the
1968 * number of devices (On-Line Capacity Expansion) must be
1969 * performed at the level of the container
1971 rv
= reshape_container(container
, devname
, -1, st
, &info
,
1972 force
, backup_file
, quiet
, 0, 0);
1975 /* get spare devices from external metadata
1977 if (st
->ss
->external
) {
1978 struct mdinfo
*info2
;
1980 info2
= st
->ss
->container_content(st
, subarray
);
1982 info
.array
.spare_disks
=
1983 info2
->array
.spare_disks
;
1988 /* Impose these changes on a single array. First
1989 * check that the metadata is OK with the change. */
1991 if (reshape_super(st
, -1, info
.new_level
,
1992 info
.new_layout
, info
.new_chunk
,
1993 info
.array
.raid_disks
, info
.delta_disks
,
1994 backup_file
, devname
, APPLY_METADATA_CHANGES
,
2000 rv
= reshape_array(container
, fd
, devname
, st
, &info
, force
,
2001 devlist
, backup_file
, quiet
, 0, 0, 0);
2011 /* verify_reshape_position()
2012 * Function checks if reshape position in metadata is not farther
2013 * than position in md.
2015 * 0 : not valid sysfs entry
2016 * it can be caused by not started reshape, it should be started
2017 * by reshape array or raid0 array is before takeover
2018 * -1 : error, reshape position is obviously wrong
2019 * 1 : success, reshape progress correct or updated
2021 static int verify_reshape_position(struct mdinfo
*info
, int level
)
2027 /* read sync_max, failure can mean raid0 array */
2028 rv
= sysfs_get_str(info
, NULL
, "sync_max", buf
, 40);
2032 unsigned long long position
= strtoull(buf
, &ep
, 0);
2034 dprintf(Name
": Read sync_max sysfs entry is: %s\n", buf
);
2035 if (!(ep
== buf
|| (*ep
!= 0 && *ep
!= '\n' && *ep
!= ' '))) {
2036 position
*= get_data_disks(level
,
2038 info
->array
.raid_disks
);
2039 if (info
->reshape_progress
< position
) {
2040 dprintf("Corrected reshape progress (%llu) to "
2041 "md position (%llu)\n",
2042 info
->reshape_progress
, position
);
2043 info
->reshape_progress
= position
;
2045 } else if (info
->reshape_progress
> position
) {
2046 pr_err("Fatal error: array "
2047 "reshape was not properly frozen "
2048 "(expected reshape position is %llu, "
2049 "but reshape progress is %llu.\n",
2050 position
, info
->reshape_progress
);
2053 dprintf("Reshape position in md and metadata "
2058 } else if (rv
== 0) {
2059 /* for valid sysfs entry, 0-length content
2060 * should be indicated as error
2068 static int reshape_array(char *container
, int fd
, char *devname
,
2069 struct supertype
*st
, struct mdinfo
*info
,
2070 int force
, struct mddev_dev
*devlist
,
2071 char *backup_file
, int quiet
, int forked
,
2072 int restart
, int freeze_reshape
)
2074 struct reshape reshape
;
2077 int orig_level
= UnSet
;
2080 struct mdu_array_info_s array
;
2083 struct mddev_dev
*dv
;
2087 unsigned long long *offsets
= NULL
;
2091 unsigned long blocks
;
2092 unsigned long cache
;
2093 unsigned long long array_size
;
2095 struct mdinfo
*sra
= NULL
;
2097 /* when reshaping a RAID0, the component_size might be zero.
2098 * So try to fix that up.
2100 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
2101 dprintf("Cannot get array information.\n");
2104 if (array
.level
== 0 && info
->component_size
== 0) {
2105 get_dev_size(fd
, NULL
, &array_size
);
2106 info
->component_size
= array_size
/ array
.raid_disks
;
2109 if (info
->reshape_active
) {
2110 int new_level
= info
->new_level
;
2111 info
->new_level
= UnSet
;
2112 if (info
->delta_disks
> 0)
2113 info
->array
.raid_disks
-= info
->delta_disks
;
2114 msg
= analyse_change(info
, &reshape
);
2115 info
->new_level
= new_level
;
2116 if (info
->delta_disks
> 0)
2117 info
->array
.raid_disks
+= info
->delta_disks
;
2119 /* Make sure the array isn't read-only */
2120 ioctl(fd
, RESTART_ARRAY_RW
, 0);
2122 msg
= analyse_change(info
, &reshape
);
2124 pr_err("%s\n", msg
);
2128 (reshape
.level
!= info
->array
.level
||
2129 reshape
.before
.layout
!= info
->array
.layout
||
2130 reshape
.before
.data_disks
+ reshape
.parity
2131 != info
->array
.raid_disks
- max(0, info
->delta_disks
))) {
2132 pr_err("reshape info is not in native format -"
2133 " cannot continue.\n");
2137 if (st
->ss
->external
&& restart
&& (info
->reshape_progress
== 0)) {
2138 /* When reshape is restarted from '0', very begin of array
2139 * it is possible that for external metadata reshape and array
2140 * configuration doesn't happen.
2141 * Check if md has the same opinion, and reshape is restarted
2142 * from 0. If so, this is regular reshape start after reshape
2143 * switch in metadata to next array only.
2145 if ((verify_reshape_position(info
, reshape
.level
) >= 0) &&
2146 (info
->reshape_progress
== 0))
2150 /* reshape already started. just skip to monitoring the reshape */
2151 if (reshape
.backup_blocks
== 0)
2155 /* The container is frozen but the array may not be.
2156 * So freeze the array so spares don't get put to the wrong use
2157 * FIXME there should probably be a cleaner separation between
2158 * freeze_array and freeze_container.
2160 sysfs_freeze_array(info
);
2161 /* Check we have enough spares to not be degraded */
2163 for (dv
= devlist
; dv
; dv
=dv
->next
)
2165 spares_needed
= max(reshape
.before
.data_disks
,
2166 reshape
.after
.data_disks
)
2167 + reshape
.parity
- array
.raid_disks
;
2170 info
->new_level
> 1 && info
->array
.level
> 1 &&
2171 spares_needed
> info
->array
.spare_disks
+ added_disks
) {
2172 pr_err("Need %d spare%s to avoid degraded array,"
2173 " and only have %d.\n"
2174 " Use --force to over-ride this check.\n",
2176 spares_needed
== 1 ? "" : "s",
2177 info
->array
.spare_disks
+ added_disks
);
2180 /* Check we have enough spares to not fail */
2181 spares_needed
= max(reshape
.before
.data_disks
,
2182 reshape
.after
.data_disks
)
2184 if ((info
->new_level
> 1 || info
->new_level
== 0) &&
2185 spares_needed
> info
->array
.spare_disks
+added_disks
) {
2186 pr_err("Need %d spare%s to create working array,"
2187 " and only have %d.\n",
2189 spares_needed
== 1 ? "" : "s",
2190 info
->array
.spare_disks
+ added_disks
);
2194 if (reshape
.level
!= array
.level
) {
2195 char *c
= map_num(pers
, reshape
.level
);
2200 err
= sysfs_set_str(info
, NULL
, "level", c
);
2203 pr_err("%s: could not set level to %s\n",
2206 (info
->array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
2207 cont_err("Bitmap must be removed"
2208 " before level can be changed\n");
2212 pr_err("level of %s changed to %s\n",
2214 orig_level
= array
.level
;
2215 sysfs_freeze_array(info
);
2217 if (reshape
.level
> 0 && st
->ss
->external
) {
2218 /* make sure mdmon is aware of the new level */
2219 if (mdmon_running(st
->container_dev
))
2220 flush_mdmon(container
);
2222 if (!mdmon_running(st
->container_dev
))
2223 start_mdmon(st
->container_dev
);
2224 ping_monitor(container
);
2225 if (mdmon_running(st
->container_dev
) &&
2226 st
->update_tail
== NULL
)
2227 st
->update_tail
= &st
->updates
;
2230 /* ->reshape_super might have chosen some spares from the
2231 * container that it wants to be part of the new array.
2232 * We can collect them with ->container_content and give
2233 * them to the kernel.
2235 if (st
->ss
->reshape_super
&& st
->ss
->container_content
) {
2236 char *subarray
= strchr(info
->text_version
+1, '/')+1;
2237 struct mdinfo
*info2
=
2238 st
->ss
->container_content(st
, subarray
);
2242 sysfs_init(info2
, fd
, st
->devnum
);
2243 /* When increasing number of devices, we need to set
2244 * new raid_disks before adding these, or they might
2247 if (reshape
.backup_blocks
&&
2248 reshape
.after
.data_disks
> reshape
.before
.data_disks
)
2249 subarray_set_num(container
, info2
, "raid_disks",
2250 reshape
.after
.data_disks
+
2252 for (d
= info2
->devs
; d
; d
= d
->next
) {
2253 if (d
->disk
.state
== 0 &&
2254 d
->disk
.raid_disk
>= 0) {
2255 /* This is a spare that wants to
2256 * be part of the array.
2258 add_disk(fd
, st
, info2
, d
);
2264 /* We might have been given some devices to add to the
2265 * array. Now that the array has been changed to the right
2266 * level and frozen, we can safely add them.
2269 Manage_subdevs(devname
, fd
, devlist
, !quiet
,
2272 if (reshape
.backup_blocks
== 0) {
2273 /* No restriping needed, but we might need to impose
2274 * some more changes: layout, raid_disks, chunk_size
2276 /* read current array info */
2277 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
2278 dprintf("Cannot get array information.\n");
2281 /* compare current array info with new values and if
2282 * it is different update them to new */
2283 if (info
->new_layout
!= UnSet
&&
2284 info
->new_layout
!= array
.layout
) {
2285 array
.layout
= info
->new_layout
;
2286 if (ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
2287 pr_err("failed to set new layout\n");
2290 printf("layout for %s set to %d\n",
2291 devname
, array
.layout
);
2293 if (info
->delta_disks
!= UnSet
&&
2294 info
->delta_disks
!= 0 &&
2295 array
.raid_disks
!= (info
->array
.raid_disks
+ info
->delta_disks
)) {
2296 array
.raid_disks
+= info
->delta_disks
;
2297 if (ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
2298 pr_err("failed to set raid disks\n");
2300 } else if (!quiet
) {
2301 printf("raid_disks for %s set to %d\n",
2302 devname
, array
.raid_disks
);
2305 if (info
->new_chunk
!= 0 &&
2306 info
->new_chunk
!= array
.chunk_size
) {
2307 if (sysfs_set_num(info
, NULL
,
2308 "chunk_size", info
->new_chunk
) != 0) {
2309 pr_err("failed to set chunk size\n");
2312 printf("chunk size for %s set to %d\n",
2313 devname
, array
.chunk_size
);
2320 * There are three possibilities.
2321 * 1/ The array will shrink.
2322 * We need to ensure the reshape will pause before reaching
2323 * the 'critical section'. We also need to fork and wait for
2324 * that to happen. When it does we
2325 * suspend/backup/complete/unfreeze
2327 * 2/ The array will not change size.
2328 * This requires that we keep a backup of a sliding window
2329 * so that we can restore data after a crash. So we need
2330 * to fork and monitor progress.
2331 * In future we will allow the data_offset to change, so
2332 * a sliding backup becomes unnecessary.
2334 * 3/ The array will grow. This is relatively easy.
2335 * However the kernel's restripe routines will cheerfully
2336 * overwrite some early data before it is safe. So we
2337 * need to make a backup of the early parts of the array
2338 * and be ready to restore it if rebuild aborts very early.
2339 * For externally managed metadata, we still need a forked
2340 * child to monitor the reshape and suspend IO over the region
2341 * that is being reshaped.
2343 * We backup data by writing it to one spare, or to a
2344 * file which was given on command line.
2346 * In each case, we first make sure that storage is available
2347 * for the required backup.
2349 * - request the shape change.
2350 * - fork to handle backup etc.
2353 /* Check that we can hold all the data */
2354 get_dev_size(fd
, NULL
, &array_size
);
2355 if (reshape
.new_size
< (array_size
/512)) {
2356 pr_err("this change will reduce the size of the array.\n"
2357 " use --grow --array-size first to truncate array.\n"
2358 " e.g. mdadm --grow %s --array-size %llu\n",
2359 devname
, reshape
.new_size
/2);
2363 sra
= sysfs_read(fd
, 0,
2364 GET_COMPONENT
|GET_DEVS
|GET_OFFSET
|GET_STATE
|GET_CHUNK
|
2367 pr_err("%s: Cannot get array details from sysfs\n",
2372 /* Decide how many blocks (sectors) for a reshape
2373 * unit. The number we have so far is just a minimum
2375 blocks
= reshape
.backup_blocks
;
2376 if (reshape
.before
.data_disks
==
2377 reshape
.after
.data_disks
) {
2378 /* Make 'blocks' bigger for better throughput, but
2379 * not so big that we reject it below.
2380 * Try for 16 megabytes
2382 while (blocks
* 32 < sra
->component_size
&&
2386 pr_err("Need to backup %luK of critical "
2387 "section..\n", blocks
/2);
2389 if (blocks
>= sra
->component_size
/2) {
2390 pr_err("%s: Something wrong"
2391 " - reshape aborted\n",
2396 /* Now we need to open all these devices so we can read/write.
2398 nrdisks
= max(reshape
.before
.data_disks
,
2399 reshape
.after
.data_disks
) + reshape
.parity
2400 + sra
->array
.spare_disks
;
2401 fdlist
= xcalloc((1+nrdisks
), sizeof(int));
2402 offsets
= xcalloc((1+nrdisks
), sizeof(offsets
[0]));
2404 odisks
= reshape
.before
.data_disks
+ reshape
.parity
;
2405 d
= reshape_prepare_fdlist(devname
, sra
, odisks
,
2406 nrdisks
, blocks
, backup_file
,
2411 if ((st
->ss
->manage_reshape
== NULL
) ||
2412 (st
->ss
->recover_backup
== NULL
)) {
2413 if (backup_file
== NULL
) {
2414 if (reshape
.after
.data_disks
<=
2415 reshape
.before
.data_disks
) {
2416 pr_err("%s: Cannot grow - "
2417 "need backup-file\n", devname
);
2419 } else if (sra
->array
.spare_disks
== 0) {
2420 pr_err("%s: Cannot grow - "
2421 "need a spare or backup-file to backup "
2422 "critical section\n", devname
);
2426 if (!reshape_open_backup_file(backup_file
, fd
, devname
,
2428 fdlist
+d
, offsets
+d
,
2436 /* lastly, check that the internal stripe cache is
2437 * large enough, or it won't work.
2438 * It must hold at least 4 stripes of the larger
2441 cache
= max(info
->array
.chunk_size
, info
->new_chunk
);
2442 cache
*= 4; /* 4 stripes minimum */
2443 cache
/= 512; /* convert to sectors */
2444 disks
= min(reshape
.before
.data_disks
, reshape
.after
.data_disks
);
2445 /* make sure there is room for 'blocks' with a bit to spare */
2446 if (cache
< 16 + blocks
/ disks
)
2447 cache
= 16 + blocks
/ disks
;
2448 cache
/= (4096/512); /* Covert from sectors to pages */
2450 if (sra
->cache_size
< cache
)
2451 subarray_set_num(container
, sra
, "stripe_cache_size",
2454 /* Right, everything seems fine. Let's kick things off.
2455 * If only changing raid_disks, use ioctl, else use
2460 sra
->new_chunk
= info
->new_chunk
;
2463 /* for external metadata checkpoint saved by mdmon can be lost
2464 * or missed /due to e.g. crash/. Check if md is not during
2465 * restart farther than metadata points to.
2466 * If so, this means metadata information is obsolete.
2468 if (st
->ss
->external
)
2469 verify_reshape_position(info
, reshape
.level
);
2470 sra
->reshape_progress
= info
->reshape_progress
;
2472 sra
->reshape_progress
= 0;
2473 if (reshape
.after
.data_disks
< reshape
.before
.data_disks
)
2474 /* start from the end of the new array */
2475 sra
->reshape_progress
= (sra
->component_size
2476 * reshape
.after
.data_disks
);
2479 if (info
->array
.chunk_size
== info
->new_chunk
&&
2480 reshape
.before
.layout
== reshape
.after
.layout
&&
2481 st
->ss
->external
== 0) {
2482 /* use SET_ARRAY_INFO but only if reshape hasn't started */
2483 ioctl(fd
, GET_ARRAY_INFO
, &array
);
2484 array
.raid_disks
= reshape
.after
.data_disks
+ reshape
.parity
;
2486 ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
2489 pr_err("Cannot set device shape for %s: %s\n",
2490 devname
, strerror(errno
));
2493 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
2494 cont_err("Bitmap must be removed before"
2495 " shape can be changed\n");
2499 } else if (!restart
) {
2500 /* set them all just in case some old 'new_*' value
2501 * persists from some earlier problem.
2504 if (sysfs_set_num(sra
, NULL
, "chunk_size", info
->new_chunk
) < 0)
2506 if (!err
&& sysfs_set_num(sra
, NULL
, "layout",
2507 reshape
.after
.layout
) < 0)
2509 if (!err
&& subarray_set_num(container
, sra
, "raid_disks",
2510 reshape
.after
.data_disks
+
2511 reshape
.parity
) < 0)
2514 pr_err("Cannot set device shape for %s\n",
2518 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
2519 cont_err("Bitmap must be removed before"
2520 " shape can be changed\n");
2525 err
= start_reshape(sra
, restart
, reshape
.before
.data_disks
,
2526 reshape
.after
.data_disks
);
2528 pr_err("Cannot %s reshape for %s\n",
2529 restart
? "continue" : "start",
2534 sysfs_set_str(sra
, NULL
, "array_state", "active");
2535 if (freeze_reshape
) {
2539 pr_err("Reshape has to be continued from"
2540 " location %llu when root filesystem has been mounted.\n",
2541 sra
->reshape_progress
);
2545 /* Now we just need to kick off the reshape and watch, while
2546 * handling backups of the data...
2547 * This is all done by a forked background process.
2549 switch(forked
? 0 : fork()) {
2551 pr_err("Cannot run child to monitor reshape: %s\n",
2566 if (check_env("MDADM_GROW_VERIFY"))
2567 fd
= open(devname
, O_RDONLY
| O_DIRECT
);
2570 mlockall(MCL_FUTURE
);
2572 if (st
->ss
->external
) {
2573 /* metadata handler takes it from here */
2574 done
= st
->ss
->manage_reshape(
2575 fd
, sra
, &reshape
, st
, blocks
,
2577 d
- odisks
, fdlist
+odisks
,
2580 done
= child_monitor(
2581 fd
, sra
, &reshape
, st
, blocks
,
2583 d
- odisks
, fdlist
+odisks
,
2589 if (backup_file
&& done
)
2590 unlink(backup_file
);
2596 if (!st
->ss
->external
&&
2597 !(reshape
.before
.data_disks
!= reshape
.after
.data_disks
2598 && info
->custom_array_size
) &&
2599 info
->new_level
== reshape
.level
&&
2601 /* no need to wait for the reshape to finish as
2602 * there is nothing more to do.
2609 if (st
->ss
->external
) {
2610 /* Re-load the metadata as much could have changed */
2611 int cfd
= open_dev(st
->container_dev
);
2613 flush_mdmon(container
);
2614 st
->ss
->free_super(st
);
2615 st
->ss
->load_container(st
, cfd
, container
);
2620 /* set new array size if required customer_array_size is used
2623 if (reshape
.before
.data_disks
!=
2624 reshape
.after
.data_disks
&&
2625 info
->custom_array_size
)
2626 set_array_size(st
, info
, info
->text_version
);
2628 if (info
->new_level
!= reshape
.level
) {
2630 c
= map_num(pers
, info
->new_level
);
2632 err
= sysfs_set_str(sra
, NULL
, "level", c
);
2634 pr_err("%s: could not set level "
2635 "to %s\n", devname
, c
);
2637 if (info
->new_level
== 0)
2638 st
->update_tail
= NULL
;
2650 if (orig_level
!= UnSet
&& sra
) {
2651 c
= map_num(pers
, orig_level
);
2652 if (c
&& sysfs_set_str(sra
, NULL
, "level", c
) == 0)
2653 pr_err("aborting level change\n");
2661 /* mdfd handle is passed to be closed in child process (after fork).
2663 int reshape_container(char *container
, char *devname
,
2665 struct supertype
*st
,
2666 struct mdinfo
*info
,
2669 int quiet
, int restart
, int freeze_reshape
)
2671 struct mdinfo
*cc
= NULL
;
2673 int last_devnum
= -1;
2675 /* component_size is not meaningful for a container,
2676 * so pass '-1' meaning 'no change'
2679 reshape_super(st
, -1, info
->new_level
,
2680 info
->new_layout
, info
->new_chunk
,
2681 info
->array
.raid_disks
, info
->delta_disks
,
2682 backup_file
, devname
, APPLY_METADATA_CHANGES
,
2690 /* ping monitor to be sure that update is on disk
2692 ping_monitor(container
);
2695 case -1: /* error */
2696 perror("Cannot fork to complete reshape\n");
2699 default: /* parent */
2700 if (!freeze_reshape
)
2701 printf(Name
": multi-array reshape continues"
2702 " in background\n");
2709 /* close unused handle in child process
2715 /* For each member array with reshape_active,
2716 * we need to perform the reshape.
2717 * We pick the first array that needs reshaping and
2718 * reshape it. reshape_array() will re-read the metadata
2719 * so the next time through a different array should be
2720 * ready for reshape.
2721 * It is possible that the 'different' array will not
2722 * be assembled yet. In that case we simple exit.
2723 * When it is assembled, the mdadm which assembles it
2724 * will take over the reshape.
2726 struct mdinfo
*content
;
2728 struct mdstat_ent
*mdstat
;
2733 cc
= st
->ss
->container_content(st
, NULL
);
2735 for (content
= cc
; content
; content
= content
->next
) {
2737 if (!content
->reshape_active
)
2740 subarray
= strchr(content
->text_version
+1, '/')+1;
2741 mdstat
= mdstat_by_subdev(subarray
,
2742 devname2devnum(container
));
2745 if (mdstat
->active
== 0) {
2746 pr_err("Skipping inactive "
2747 "array md%i.\n", mdstat
->devnum
);
2748 free_mdstat(mdstat
);
2757 adev
= map_dev(dev2major(mdstat
->devnum
),
2758 dev2minor(mdstat
->devnum
),
2761 adev
= content
->text_version
;
2763 fd
= open_dev(mdstat
->devnum
);
2765 printf(Name
": Device %s cannot be opened for reshape.",
2770 if (last_devnum
== mdstat
->devnum
) {
2771 /* Do not allow for multiple reshape_array() calls for
2773 * It can happen when reshape_array() returns without
2774 * error, when reshape is not finished (wrong reshape
2775 * starting/continuation conditions). Mdmon doesn't
2776 * switch to next array in container and reentry
2777 * conditions for the same array occur.
2778 * This is possibly interim until the behaviour of
2779 * reshape_array is resolved().
2781 printf(Name
": Multiple reshape execution detected for "
2782 "device %s.", adev
);
2786 last_devnum
= mdstat
->devnum
;
2788 sysfs_init(content
, fd
, mdstat
->devnum
);
2790 if (mdmon_running(devname2devnum(container
)))
2791 flush_mdmon(container
);
2793 rv
= reshape_array(container
, fd
, adev
, st
,
2794 content
, force
, NULL
,
2795 backup_file
, quiet
, 1, restart
,
2799 if (freeze_reshape
) {
2808 if (mdmon_running(devname2devnum(container
)))
2809 flush_mdmon(container
);
2818 * We run a child process in the background which performs the following
2820 * - wait for resync to reach a certain point
2821 * - suspend io to the following section
2822 * - backup that section
2823 * - allow resync to proceed further
2825 * - discard the backup.
2827 * When are combined in slightly different ways in the three cases.
2829 * - suspend/backup/allow/wait/resume/discard
2831 * - allow/wait/suspend/backup/allow/wait/resume/discard
2833 * - wait/resume/discard/suspend/backup/allow
2835 * suspend/backup/allow always come together
2836 * wait/resume/discard do too.
2837 * For the same-size case we have two backups to improve flow.
2841 int progress_reshape(struct mdinfo
*info
, struct reshape
*reshape
,
2842 unsigned long long backup_point
,
2843 unsigned long long wait_point
,
2844 unsigned long long *suspend_point
,
2845 unsigned long long *reshape_completed
)
2847 /* This function is called repeatedly by the reshape manager.
2848 * It determines how much progress can safely be made and allows
2850 * - 'info' identifies the array and particularly records in
2851 * ->reshape_progress the metadata's knowledge of progress
2852 * This is a sector offset from the start of the array
2853 * of the next array block to be relocated. This number
2854 * may increase from 0 or decrease from array_size, depending
2855 * on the type of reshape that is happening.
2856 * Note that in contrast, 'sync_completed' is a block count of the
2857 * reshape so far. It gives the distance between the start point
2858 * (head or tail of device) and the next place that data will be
2859 * written. It always increases.
2860 * - 'reshape' is the structure created by analyse_change
2861 * - 'backup_point' shows how much the metadata manager has backed-up
2862 * data. For reshapes with increasing progress, it is the next address
2863 * to be backed up, previous addresses have been backed-up. For
2864 * decreasing progress, it is the earliest address that has been
2865 * backed up - later address are also backed up.
2866 * So addresses between reshape_progress and backup_point are
2867 * backed up providing those are in the 'correct' order.
2868 * - 'wait_point' is an array address. When reshape_completed
2869 * passes this point, progress_reshape should return. It might
2870 * return earlier if it determines that ->reshape_progress needs
2871 * to be updated or further backup is needed.
2872 * - suspend_point is maintained by progress_reshape and the caller
2873 * should not touch it except to initialise to zero.
2874 * It is an array address and it only increases in 2.6.37 and earlier.
2875 * This makes it difficult to handle reducing reshapes with
2876 * external metadata.
2877 * However: it is similar to backup_point in that it records the
2878 * other end of a suspended region from reshape_progress.
2879 * it is moved to extend the region that is safe to backup and/or
2881 * - reshape_completed is read from sysfs and returned. The caller
2882 * should copy this into ->reshape_progress when it has reason to
2883 * believe that the metadata knows this, and any backup outside this
2887 * 1 if more data from backup_point - but only as far as suspend_point,
2888 * should be backed up
2889 * 0 if things are progressing smoothly
2890 * -1 if the reshape is finished because it is all done,
2891 * -2 if the reshape is finished due to an error.
2894 int advancing
= (reshape
->after
.data_disks
2895 >= reshape
->before
.data_disks
);
2896 unsigned long long need_backup
; /* All data between start of array and
2897 * here will at some point need to
2900 unsigned long long read_offset
, write_offset
;
2901 unsigned long long write_range
;
2902 unsigned long long max_progress
, target
, completed
;
2903 unsigned long long array_size
= (info
->component_size
2904 * reshape
->before
.data_disks
);
2908 /* First, we unsuspend any region that is now known to be safe.
2909 * If suspend_point is on the 'wrong' side of reshape_progress, then
2910 * we don't have or need suspension at the moment. This is true for
2911 * native metadata when we don't need to back-up.
2914 if (info
->reshape_progress
<= *suspend_point
)
2915 sysfs_set_num(info
, NULL
, "suspend_lo",
2916 info
->reshape_progress
);
2918 /* Note: this won't work in 2.6.37 and before.
2919 * Something somewhere should make sure we don't need it!
2921 if (info
->reshape_progress
>= *suspend_point
)
2922 sysfs_set_num(info
, NULL
, "suspend_hi",
2923 info
->reshape_progress
);
2926 /* Now work out how far it is safe to progress.
2927 * If the read_offset for ->reshape_progress is less than
2928 * 'blocks' beyond the write_offset, we can only progress as far
2930 * Otherwise we can progress until the write_offset for the new location
2931 * reaches (within 'blocks' of) the read_offset at the current location.
2932 * However that region must be suspended unless we are using native
2934 * If we need to suspend more, we limit it to 128M per device, which is
2935 * rather arbitrary and should be some time-based calculation.
2937 read_offset
= info
->reshape_progress
/ reshape
->before
.data_disks
;
2938 write_offset
= info
->reshape_progress
/ reshape
->after
.data_disks
;
2939 write_range
= info
->new_chunk
/512;
2940 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
)
2941 need_backup
= array_size
;
2943 need_backup
= reshape
->backup_blocks
;
2945 if (read_offset
< write_offset
+ write_range
)
2946 max_progress
= backup_point
;
2950 reshape
->after
.data_disks
;
2952 if (read_offset
> write_offset
- write_range
)
2953 /* Can only progress as far as has been backed up,
2954 * which must be suspended */
2955 max_progress
= backup_point
;
2956 else if (info
->reshape_progress
<= need_backup
)
2957 max_progress
= backup_point
;
2959 if (info
->array
.major_version
>= 0)
2960 /* Can progress until backup is needed */
2961 max_progress
= need_backup
;
2963 /* Can progress until metadata update is required */
2966 reshape
->after
.data_disks
;
2967 /* but data must be suspended */
2968 if (max_progress
< *suspend_point
)
2969 max_progress
= *suspend_point
;
2974 /* We know it is safe to progress to 'max_progress' providing
2975 * it is suspended or we are using native metadata.
2976 * Consider extending suspend_point 128M per device if it
2977 * is less than 64M per device beyond reshape_progress.
2978 * But always do a multiple of 'blocks'
2979 * FIXME this is too big - it takes to long to complete
2982 target
= 64*1024*2 * min(reshape
->before
.data_disks
,
2983 reshape
->after
.data_disks
);
2984 target
/= reshape
->backup_blocks
;
2987 target
*= reshape
->backup_blocks
;
2989 /* For externally managed metadata we always need to suspend IO to
2990 * the area being reshaped so we regularly push suspend_point forward.
2991 * For native metadata we only need the suspend if we are going to do
2995 if ((need_backup
> info
->reshape_progress
2996 || info
->array
.major_version
< 0) &&
2997 *suspend_point
< info
->reshape_progress
+ target
) {
2998 if (need_backup
< *suspend_point
+ 2 * target
)
2999 *suspend_point
= need_backup
;
3000 else if (*suspend_point
+ 2 * target
< array_size
)
3001 *suspend_point
+= 2 * target
;
3003 *suspend_point
= array_size
;
3004 sysfs_set_num(info
, NULL
, "suspend_hi", *suspend_point
);
3005 if (max_progress
> *suspend_point
)
3006 max_progress
= *suspend_point
;
3009 if (info
->array
.major_version
>= 0) {
3010 /* Only need to suspend when about to backup */
3011 if (info
->reshape_progress
< need_backup
* 2 &&
3012 *suspend_point
> 0) {
3014 sysfs_set_num(info
, NULL
, "suspend_lo", 0);
3015 sysfs_set_num(info
, NULL
, "suspend_hi", need_backup
);
3018 /* Need to suspend continually */
3019 if (info
->reshape_progress
< *suspend_point
)
3020 *suspend_point
= info
->reshape_progress
;
3021 if (*suspend_point
+ target
< info
->reshape_progress
)
3022 /* No need to move suspend region yet */;
3024 if (*suspend_point
>= 2 * target
)
3025 *suspend_point
-= 2 * target
;
3028 sysfs_set_num(info
, NULL
, "suspend_lo",
3031 if (max_progress
< *suspend_point
)
3032 max_progress
= *suspend_point
;
3036 /* now set sync_max to allow that progress. sync_max, like
3037 * sync_completed is a count of sectors written per device, so
3038 * we find the difference between max_progress and the start point,
3039 * and divide that by after.data_disks to get a sync_max
3041 * At the same time we convert wait_point to a similar number
3042 * for comparing against sync_completed.
3044 /* scale down max_progress to per_disk */
3045 max_progress
/= reshape
->after
.data_disks
;
3046 /* Round to chunk size as some kernels give an erroneously high number */
3047 max_progress
/= info
->new_chunk
/512;
3048 max_progress
*= info
->new_chunk
/512;
3049 /* And round to old chunk size as the kernel wants that */
3050 max_progress
/= info
->array
.chunk_size
/512;
3051 max_progress
*= info
->array
.chunk_size
/512;
3052 /* Limit progress to the whole device */
3053 if (max_progress
> info
->component_size
)
3054 max_progress
= info
->component_size
;
3055 wait_point
/= reshape
->after
.data_disks
;
3057 /* switch from 'device offset' to 'processed block count' */
3058 max_progress
= info
->component_size
- max_progress
;
3059 wait_point
= info
->component_size
- wait_point
;
3062 sysfs_set_num(info
, NULL
, "sync_max", max_progress
);
3064 /* Now wait. If we have already reached the point that we were
3065 * asked to wait to, don't wait at all, else wait for any change.
3066 * We need to select on 'sync_completed' as that is the place that
3067 * notifications happen, but we are really interested in
3068 * 'reshape_position'
3070 fd
= sysfs_get_fd(info
, NULL
, "sync_completed");
3072 goto check_progress
;
3074 if (sysfs_fd_get_ll(fd
, &completed
) < 0)
3075 goto check_progress
;
3077 while (completed
< max_progress
&& completed
< wait_point
) {
3078 /* Check that sync_action is still 'reshape' to avoid
3079 * waiting forever on a dead array
3083 if (sysfs_get_str(info
, NULL
, "sync_action",
3085 strncmp(action
, "reshape", 7) != 0)
3087 /* Some kernels reset 'sync_completed' to zero
3088 * before setting 'sync_action' to 'idle'.
3089 * So we need these extra tests.
3091 if (completed
== 0 && advancing
3092 && info
->reshape_progress
> 0)
3094 if (completed
== 0 && !advancing
3095 && info
->reshape_progress
< (info
->component_size
3096 * reshape
->after
.data_disks
))
3100 select(fd
+1, NULL
, NULL
, &rfds
, NULL
);
3101 if (sysfs_fd_get_ll(fd
, &completed
) < 0)
3102 goto check_progress
;
3104 /* Some kernels reset 'sync_completed' to zero,
3105 * we need to have real point we are in md
3108 completed
= max_progress
;
3110 /* some kernels can give an incorrectly high 'completed' number */
3111 completed
/= (info
->new_chunk
/512);
3112 completed
*= (info
->new_chunk
/512);
3113 /* Convert 'completed' back in to a 'progress' number */
3114 completed
*= reshape
->after
.data_disks
;
3116 completed
= info
->component_size
* reshape
->after
.data_disks
3119 *reshape_completed
= completed
;
3123 /* We return the need_backup flag. Caller will decide
3124 * how much - a multiple of ->backup_blocks up to *suspend_point
3127 return need_backup
> info
->reshape_progress
;
3129 return need_backup
>= info
->reshape_progress
;
3132 /* if we couldn't read a number from sync_completed, then
3133 * either the reshape did complete, or it aborted.
3134 * We can tell which by checking for 'none' in reshape_position.
3135 * If it did abort, then it might immediately restart if it
3136 * it was just a device failure that leaves us degraded but
3140 if (sysfs_get_str(info
, NULL
, "reshape_position", buf
, sizeof(buf
)) < 0
3141 || strncmp(buf
, "none", 4) != 0) {
3142 /* The abort might only be temporary. Wait up to 10
3143 * seconds for fd to contain a valid number again.
3149 while (fd
>= 0 && rv
< 0 && tv
.tv_sec
> 0) {
3153 if (select(fd
+1, NULL
, NULL
, &rfds
, &tv
) != 1)
3155 switch (sysfs_fd_get_ll(fd
, &completed
)) {
3157 /* all good again */
3160 case -2: /* read error - abort */
3167 return rv
; /* abort */
3169 /* Maybe racing with array shutdown - check state */
3172 if (sysfs_get_str(info
, NULL
, "array_state", buf
, sizeof(buf
)) < 0
3173 || strncmp(buf
, "inactive", 8) == 0
3174 || strncmp(buf
, "clear",5) == 0)
3175 return -2; /* abort */
3176 return -1; /* complete */
3180 /* FIXME return status is never checked */
3181 static int grow_backup(struct mdinfo
*sra
,
3182 unsigned long long offset
, /* per device */
3183 unsigned long stripes
, /* per device, in old chunks */
3184 int *sources
, unsigned long long *offsets
,
3185 int disks
, int chunk
, int level
, int layout
,
3186 int dests
, int *destfd
, unsigned long long *destoffsets
,
3187 int part
, int *degraded
,
3190 /* Backup 'blocks' sectors at 'offset' on each device of the array,
3191 * to storage 'destfd' (offset 'destoffsets'), after first
3192 * suspending IO. Then allow resync to continue
3193 * over the suspended section.
3194 * Use part 'part' of the backup-super-block.
3199 unsigned long long ll
;
3201 //printf("offset %llu\n", offset);
3207 /* Check that array hasn't become degraded, else we might backup the wrong data */
3208 if (sysfs_get_ll(sra
, NULL
, "degraded", &ll
) < 0)
3209 return -1; /* FIXME this error is ignored */
3210 new_degraded
= (int)ll
;
3211 if (new_degraded
!= *degraded
) {
3212 /* check each device to ensure it is still working */
3214 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
3215 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
3217 if (sd
->disk
.state
& (1<<MD_DISK_SYNC
)) {
3219 if (sysfs_get_str(sra
, sd
, "state", sbuf
, 20) < 0 ||
3220 strstr(sbuf
, "faulty") ||
3221 strstr(sbuf
, "in_sync") == NULL
) {
3222 /* this device is dead */
3223 sd
->disk
.state
= (1<<MD_DISK_FAULTY
);
3224 if (sd
->disk
.raid_disk
>= 0 &&
3225 sources
[sd
->disk
.raid_disk
] >= 0) {
3226 close(sources
[sd
->disk
.raid_disk
]);
3227 sources
[sd
->disk
.raid_disk
] = -1;
3232 *degraded
= new_degraded
;
3235 bsb
.arraystart2
= __cpu_to_le64(offset
* odata
);
3236 bsb
.length2
= __cpu_to_le64(stripes
* (chunk
/512) * odata
);
3238 bsb
.arraystart
= __cpu_to_le64(offset
* odata
);
3239 bsb
.length
= __cpu_to_le64(stripes
* (chunk
/512) * odata
);
3242 bsb
.magic
[15] = '2';
3243 for (i
= 0; i
< dests
; i
++)
3245 lseek64(destfd
[i
], destoffsets
[i
] + __le64_to_cpu(bsb
.devstart2
)*512, 0);
3247 lseek64(destfd
[i
], destoffsets
[i
], 0);
3249 rv
= save_stripes(sources
, offsets
,
3250 disks
, chunk
, level
, layout
,
3252 offset
*512*odata
, stripes
* chunk
* odata
,
3257 bsb
.mtime
= __cpu_to_le64(time(0));
3258 for (i
= 0; i
< dests
; i
++) {
3259 bsb
.devstart
= __cpu_to_le64(destoffsets
[i
]/512);
3261 bsb
.sb_csum
= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
));
3262 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0)
3263 bsb
.sb_csum2
= bsb_csum((char*)&bsb
,
3264 ((char*)&bsb
.sb_csum2
)-((char*)&bsb
));
3267 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
] - 4096, 0)
3268 != destoffsets
[i
] - 4096)
3270 if (write(destfd
[i
], &bsb
, 512) != 512)
3272 if (destoffsets
[i
] > 4096) {
3273 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
]+stripes
*chunk
*odata
, 0) !=
3274 destoffsets
[i
]+stripes
*chunk
*odata
)
3276 if (write(destfd
[i
], &bsb
, 512) != 512)
3286 /* in 2.6.30, the value reported by sync_completed can be
3287 * less that it should be by one stripe.
3288 * This only happens when reshape hits sync_max and pauses.
3289 * So allow wait_backup to either extent sync_max further
3290 * than strictly necessary, or return before the
3291 * sync has got quite as far as we would really like.
3292 * This is what 'blocks2' is for.
3293 * The various caller give appropriate values so that
3296 /* FIXME return value is often ignored */
3297 static int forget_backup(int dests
, int *destfd
,
3298 unsigned long long *destoffsets
,
3302 * Erase backup 'part' (which is 0 or 1)
3308 bsb
.arraystart2
= __cpu_to_le64(0);
3309 bsb
.length2
= __cpu_to_le64(0);
3311 bsb
.arraystart
= __cpu_to_le64(0);
3312 bsb
.length
= __cpu_to_le64(0);
3314 bsb
.mtime
= __cpu_to_le64(time(0));
3316 for (i
= 0; i
< dests
; i
++) {
3317 bsb
.devstart
= __cpu_to_le64(destoffsets
[i
]/512);
3318 bsb
.sb_csum
= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
));
3319 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0)
3320 bsb
.sb_csum2
= bsb_csum((char*)&bsb
,
3321 ((char*)&bsb
.sb_csum2
)-((char*)&bsb
));
3322 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
]-4096, 0) !=
3323 destoffsets
[i
]-4096)
3326 write(destfd
[i
], &bsb
, 512) != 512)
3333 static void fail(char *msg
)
3336 rv
= (write(2, msg
, strlen(msg
)) != (int)strlen(msg
));
3337 rv
|= (write(2, "\n", 1) != 1);
3341 static char *abuf
, *bbuf
;
3342 static unsigned long long abuflen
;
3343 static void validate(int afd
, int bfd
, unsigned long long offset
)
3345 /* check that the data in the backup against the array.
3346 * This is only used for regression testing and should not
3347 * be used while the array is active
3351 lseek64(bfd
, offset
- 4096, 0);
3352 if (read(bfd
, &bsb2
, 512) != 512)
3353 fail("cannot read bsb");
3354 if (bsb2
.sb_csum
!= bsb_csum((char*)&bsb2
,
3355 ((char*)&bsb2
.sb_csum
)-((char*)&bsb2
)))
3356 fail("first csum bad");
3357 if (memcmp(bsb2
.magic
, "md_backup_data", 14) != 0)
3358 fail("magic is bad");
3359 if (memcmp(bsb2
.magic
, "md_backup_data-2", 16) == 0 &&
3360 bsb2
.sb_csum2
!= bsb_csum((char*)&bsb2
,
3361 ((char*)&bsb2
.sb_csum2
)-((char*)&bsb2
)))
3362 fail("second csum bad");
3364 if (__le64_to_cpu(bsb2
.devstart
)*512 != offset
)
3365 fail("devstart is wrong");
3368 unsigned long long len
= __le64_to_cpu(bsb2
.length
)*512;
3370 if (abuflen
< len
) {
3374 if (posix_memalign((void**)&abuf
, 4096, abuflen
) ||
3375 posix_memalign((void**)&bbuf
, 4096, abuflen
)) {
3377 /* just stop validating on mem-alloc failure */
3382 lseek64(bfd
, offset
, 0);
3383 if ((unsigned long long)read(bfd
, bbuf
, len
) != len
) {
3384 //printf("len %llu\n", len);
3385 fail("read first backup failed");
3387 lseek64(afd
, __le64_to_cpu(bsb2
.arraystart
)*512, 0);
3388 if ((unsigned long long)read(afd
, abuf
, len
) != len
)
3389 fail("read first from array failed");
3390 if (memcmp(bbuf
, abuf
, len
) != 0) {
3393 printf("offset=%llu len=%llu\n",
3394 (unsigned long long)__le64_to_cpu(bsb2
.arraystart
)*512, len
);
3395 for (i
=0; i
<len
; i
++)
3396 if (bbuf
[i
] != abuf
[i
]) {
3397 printf("first diff byte %d\n", i
);
3401 fail("data1 compare failed");
3405 unsigned long long len
= __le64_to_cpu(bsb2
.length2
)*512;
3407 if (abuflen
< len
) {
3411 abuf
= xmalloc(abuflen
);
3412 bbuf
= xmalloc(abuflen
);
3415 lseek64(bfd
, offset
+__le64_to_cpu(bsb2
.devstart2
)*512, 0);
3416 if ((unsigned long long)read(bfd
, bbuf
, len
) != len
)
3417 fail("read second backup failed");
3418 lseek64(afd
, __le64_to_cpu(bsb2
.arraystart2
)*512, 0);
3419 if ((unsigned long long)read(afd
, abuf
, len
) != len
)
3420 fail("read second from array failed");
3421 if (memcmp(bbuf
, abuf
, len
) != 0)
3422 fail("data2 compare failed");
3426 int child_monitor(int afd
, struct mdinfo
*sra
, struct reshape
*reshape
,
3427 struct supertype
*st
, unsigned long blocks
,
3428 int *fds
, unsigned long long *offsets
,
3429 int dests
, int *destfd
, unsigned long long *destoffsets
)
3431 /* Monitor a reshape where backup is being performed using
3432 * 'native' mechanism - either to a backup file, or
3433 * to some space in a spare.
3437 unsigned long long speed
;
3438 unsigned long long suspend_point
, array_size
;
3439 unsigned long long backup_point
, wait_point
;
3440 unsigned long long reshape_completed
;
3442 int increasing
= reshape
->after
.data_disks
>= reshape
->before
.data_disks
;
3443 int part
= 0; /* The next part of the backup area to fill. It may already
3444 * be full, so we need to check */
3445 int level
= reshape
->level
;
3446 int layout
= reshape
->before
.layout
;
3447 int data
= reshape
->before
.data_disks
;
3448 int disks
= reshape
->before
.data_disks
+ reshape
->parity
;
3449 int chunk
= sra
->array
.chunk_size
;
3451 unsigned long stripes
;
3454 /* set up the backup-super-block. This requires the
3455 * uuid from the array.
3457 /* Find a superblock */
3458 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
3462 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
3464 dn
= map_dev(sd
->disk
.major
, sd
->disk
.minor
, 1);
3465 devfd
= dev_open(dn
, O_RDONLY
);
3468 ok
= st
->ss
->load_super(st
, devfd
, NULL
);
3474 pr_err("Cannot find a superblock\n");
3478 memset(&bsb
, 0, 512);
3479 memcpy(bsb
.magic
, "md_backup_data-1", 16);
3480 st
->ss
->uuid_from_super(st
, uuid
);
3481 memcpy(bsb
.set_uuid
, uuid
, 16);
3482 bsb
.mtime
= __cpu_to_le64(time(0));
3483 bsb
.devstart2
= blocks
;
3485 stripes
= blocks
/ (sra
->array
.chunk_size
/512) /
3486 reshape
->before
.data_disks
;
3488 if (posix_memalign((void**)&buf
, 4096, disks
* chunk
))
3489 /* Don't start the 'reshape' */
3491 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
) {
3492 sysfs_get_ll(sra
, NULL
, "sync_speed_min", &speed
);
3493 sysfs_set_num(sra
, NULL
, "sync_speed_min", 200000);
3497 array_size
= sra
->component_size
* reshape
->after
.data_disks
;
3498 backup_point
= sra
->reshape_progress
;
3501 array_size
= sra
->component_size
* reshape
->before
.data_disks
;
3502 backup_point
= reshape
->backup_blocks
;
3503 suspend_point
= array_size
;
3509 /* Want to return as soon the oldest backup slot can
3510 * be released as that allows us to start backing up
3511 * some more, providing suspend_point has been
3512 * advanced, which it should have.
3515 wait_point
= array_size
;
3516 if (part
== 0 && __le64_to_cpu(bsb
.length
) > 0)
3517 wait_point
= (__le64_to_cpu(bsb
.arraystart
) +
3518 __le64_to_cpu(bsb
.length
));
3519 if (part
== 1 && __le64_to_cpu(bsb
.length2
) > 0)
3520 wait_point
= (__le64_to_cpu(bsb
.arraystart2
) +
3521 __le64_to_cpu(bsb
.length2
));
3524 if (part
== 0 && __le64_to_cpu(bsb
.length
) > 0)
3525 wait_point
= __le64_to_cpu(bsb
.arraystart
);
3526 if (part
== 1 && __le64_to_cpu(bsb
.length2
) > 0)
3527 wait_point
= __le64_to_cpu(bsb
.arraystart2
);
3530 rv
= progress_reshape(sra
, reshape
,
3531 backup_point
, wait_point
,
3532 &suspend_point
, &reshape_completed
);
3533 /* external metadata would need to ping_monitor here */
3534 sra
->reshape_progress
= reshape_completed
;
3536 /* Clear any backup region that is before 'here' */
3538 if (__le64_to_cpu(bsb
.length
) > 0 &&
3539 reshape_completed
>= (__le64_to_cpu(bsb
.arraystart
) +
3540 __le64_to_cpu(bsb
.length
)))
3541 forget_backup(dests
, destfd
,
3543 if (__le64_to_cpu(bsb
.length2
) > 0 &&
3544 reshape_completed
>= (__le64_to_cpu(bsb
.arraystart2
) +
3545 __le64_to_cpu(bsb
.length2
)))
3546 forget_backup(dests
, destfd
,
3549 if (__le64_to_cpu(bsb
.length
) > 0 &&
3550 reshape_completed
<= (__le64_to_cpu(bsb
.arraystart
)))
3551 forget_backup(dests
, destfd
,
3553 if (__le64_to_cpu(bsb
.length2
) > 0 &&
3554 reshape_completed
<= (__le64_to_cpu(bsb
.arraystart2
)))
3555 forget_backup(dests
, destfd
,
3564 if (rv
== 0 && increasing
&& !st
->ss
->external
) {
3565 /* No longer need to monitor this reshape */
3571 unsigned long long offset
;
3572 unsigned long actual_stripes
;
3573 /* Need to backup some data.
3574 * If 'part' is not used and the desired
3575 * backup size is suspended, do a backup,
3576 * then consider the next part.
3578 /* Check that 'part' is unused */
3579 if (part
== 0 && __le64_to_cpu(bsb
.length
) != 0)
3581 if (part
== 1 && __le64_to_cpu(bsb
.length2
) != 0)
3584 offset
= backup_point
/ data
;
3585 actual_stripes
= stripes
;
3587 if (offset
+ actual_stripes
* (chunk
/512) >
3588 sra
->component_size
)
3589 actual_stripes
= ((sra
->component_size
- offset
)
3591 if (offset
+ actual_stripes
* (chunk
/512) >
3595 if (offset
< actual_stripes
* (chunk
/512))
3596 actual_stripes
= offset
/ (chunk
/512);
3597 offset
-= actual_stripes
* (chunk
/512);
3598 if (offset
< suspend_point
/data
)
3601 if (actual_stripes
== 0)
3603 grow_backup(sra
, offset
, actual_stripes
,
3605 disks
, chunk
, level
, layout
,
3606 dests
, destfd
, destoffsets
,
3607 part
, °raded
, buf
);
3608 validate(afd
, destfd
[0], destoffsets
[0]);
3609 /* record where 'part' is up to */
3612 backup_point
+= actual_stripes
* (chunk
/512) * data
;
3614 backup_point
-= actual_stripes
* (chunk
/512) * data
;
3618 /* FIXME maybe call progress_reshape one more time instead */
3619 abort_reshape(sra
); /* remove any remaining suspension */
3620 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
)
3621 sysfs_set_num(sra
, NULL
, "sync_speed_min", speed
);
3627 * If any spare contains md_back_data-1 which is recent wrt mtime,
3628 * write that data into the array and update the super blocks with
3629 * the new reshape_progress
3631 int Grow_restart(struct supertype
*st
, struct mdinfo
*info
, int *fdlist
, int cnt
,
3632 char *backup_file
, int verbose
)
3636 unsigned long long *offsets
;
3637 unsigned long long nstripe
, ostripe
;
3640 odata
= info
->array
.raid_disks
- info
->delta_disks
- 1;
3641 if (info
->array
.level
== 6) odata
--; /* number of data disks */
3642 ndata
= info
->array
.raid_disks
- 1;
3643 if (info
->new_level
== 6) ndata
--;
3645 old_disks
= info
->array
.raid_disks
- info
->delta_disks
;
3647 if (info
->delta_disks
<= 0)
3648 /* Didn't grow, so the backup file must have
3652 for (i
=old_disks
-(backup_file
?1:0); i
<cnt
; i
++) {
3653 struct mdinfo dinfo
;
3656 char *devname
, namebuf
[20];
3657 unsigned long long lo
, hi
;
3659 /* This was a spare and may have some saved data on it.
3660 * Load the superblock, find and load the
3661 * backup_super_block.
3662 * If either fail, go on to next device.
3663 * If the backup contains no new info, just return
3664 * else restore data and update all superblocks
3666 if (i
== old_disks
-1) {
3667 fd
= open(backup_file
, O_RDONLY
);
3669 pr_err("backup file %s inaccessible: %s\n",
3670 backup_file
, strerror(errno
));
3673 devname
= backup_file
;
3678 if (st
->ss
->load_super(st
, fd
, NULL
))
3681 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
3682 st
->ss
->free_super(st
);
3685 (dinfo
.data_offset
+ dinfo
.component_size
- 8) <<9,
3687 pr_err("Cannot seek on device %d\n", i
);
3688 continue; /* Cannot seek */
3690 sprintf(namebuf
, "device-%d", i
);
3693 if (read(fd
, &bsb
, sizeof(bsb
)) != sizeof(bsb
)) {
3695 pr_err("Cannot read from %s\n", devname
);
3696 continue; /* Cannot read */
3698 if (memcmp(bsb
.magic
, "md_backup_data-1", 16) != 0 &&
3699 memcmp(bsb
.magic
, "md_backup_data-2", 16) != 0) {
3701 pr_err("No backup metadata on %s\n", devname
);
3704 if (bsb
.sb_csum
!= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
))) {
3706 pr_err("Bad backup-metadata checksum on %s\n", devname
);
3707 continue; /* bad checksum */
3709 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0 &&
3710 bsb
.sb_csum2
!= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum2
)-((char*)&bsb
))) {
3712 pr_err("Bad backup-metadata checksum2 on %s\n", devname
);
3713 continue; /* Bad second checksum */
3715 if (memcmp(bsb
.set_uuid
,info
->uuid
, 16) != 0) {
3717 pr_err("Wrong uuid on backup-metadata on %s\n", devname
);
3718 continue; /* Wrong uuid */
3721 /* array utime and backup-mtime should be updated at much the same time, but it seems that
3722 * sometimes they aren't... So allow considerable flexability in matching, and allow
3723 * this test to be overridden by an environment variable.
3725 if (info
->array
.utime
> (int)__le64_to_cpu(bsb
.mtime
) + 2*60*60 ||
3726 info
->array
.utime
< (int)__le64_to_cpu(bsb
.mtime
) - 10*60) {
3727 if (check_env("MDADM_GROW_ALLOW_OLD")) {
3728 pr_err("accepting backup with timestamp %lu "
3729 "for array with timestamp %lu\n",
3730 (unsigned long)__le64_to_cpu(bsb
.mtime
),
3731 (unsigned long)info
->array
.utime
);
3734 pr_err("too-old timestamp on "
3735 "backup-metadata on %s\n", devname
);
3736 continue; /* time stamp is too bad */
3740 if (bsb
.magic
[15] == '1') {
3741 if (bsb
.length
== 0)
3743 if (info
->delta_disks
>= 0) {
3744 /* reshape_progress is increasing */
3745 if (__le64_to_cpu(bsb
.arraystart
)
3746 + __le64_to_cpu(bsb
.length
)
3747 < info
->reshape_progress
) {
3750 pr_err("backup-metadata found on %s but is not needed\n", devname
);
3751 continue; /* No new data here */
3754 /* reshape_progress is decreasing */
3755 if (__le64_to_cpu(bsb
.arraystart
) >=
3756 info
->reshape_progress
)
3757 goto nonew
; /* No new data here */
3760 if (bsb
.length
== 0 && bsb
.length2
== 0)
3762 if (info
->delta_disks
>= 0) {
3763 /* reshape_progress is increasing */
3764 if ((__le64_to_cpu(bsb
.arraystart
)
3765 + __le64_to_cpu(bsb
.length
)
3766 < info
->reshape_progress
)
3768 (__le64_to_cpu(bsb
.arraystart2
)
3769 + __le64_to_cpu(bsb
.length2
)
3770 < info
->reshape_progress
))
3771 goto nonew
; /* No new data here */
3773 /* reshape_progress is decreasing */
3774 if (__le64_to_cpu(bsb
.arraystart
) >=
3775 info
->reshape_progress
&&
3776 __le64_to_cpu(bsb
.arraystart2
) >=
3777 info
->reshape_progress
)
3778 goto nonew
; /* No new data here */
3781 if (lseek64(fd
, __le64_to_cpu(bsb
.devstart
)*512, 0)< 0) {
3784 pr_err("Failed to verify secondary backup-metadata block on %s\n",
3786 continue; /* Cannot seek */
3788 /* There should be a duplicate backup superblock 4k before here */
3789 if (lseek64(fd
, -4096, 1) < 0 ||
3790 read(fd
, &bsb2
, sizeof(bsb2
)) != sizeof(bsb2
))
3791 goto second_fail
; /* Cannot find leading superblock */
3792 if (bsb
.magic
[15] == '1')
3793 bsbsize
= offsetof(struct mdp_backup_super
, pad1
);
3795 bsbsize
= offsetof(struct mdp_backup_super
, pad
);
3796 if (memcmp(&bsb2
, &bsb
, bsbsize
) != 0)
3797 goto second_fail
; /* Cannot find leading superblock */
3799 /* Now need the data offsets for all devices. */
3800 offsets
= xmalloc(sizeof(*offsets
)*info
->array
.raid_disks
);
3801 for(j
=0; j
<info
->array
.raid_disks
; j
++) {
3804 if (st
->ss
->load_super(st
, fdlist
[j
], NULL
))
3805 /* FIXME should be this be an error */
3807 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
3808 st
->ss
->free_super(st
);
3809 offsets
[j
] = dinfo
.data_offset
* 512;
3811 printf(Name
": restoring critical section\n");
3813 if (restore_stripes(fdlist
, offsets
,
3814 info
->array
.raid_disks
,
3818 fd
, __le64_to_cpu(bsb
.devstart
)*512,
3819 __le64_to_cpu(bsb
.arraystart
)*512,
3820 __le64_to_cpu(bsb
.length
)*512, NULL
)) {
3821 /* didn't succeed, so giveup */
3823 pr_err("Error restoring backup from %s\n",
3829 if (bsb
.magic
[15] == '2' &&
3830 restore_stripes(fdlist
, offsets
,
3831 info
->array
.raid_disks
,
3835 fd
, __le64_to_cpu(bsb
.devstart
)*512 +
3836 __le64_to_cpu(bsb
.devstart2
)*512,
3837 __le64_to_cpu(bsb
.arraystart2
)*512,
3838 __le64_to_cpu(bsb
.length2
)*512, NULL
)) {
3839 /* didn't succeed, so giveup */
3841 pr_err("Error restoring second backup from %s\n",
3849 /* Ok, so the data is restored. Let's update those superblocks. */
3853 lo
= __le64_to_cpu(bsb
.arraystart
);
3854 hi
= lo
+ __le64_to_cpu(bsb
.length
);
3856 if (bsb
.magic
[15] == '2' && bsb
.length2
) {
3857 unsigned long long lo1
, hi1
;
3858 lo1
= __le64_to_cpu(bsb
.arraystart2
);
3859 hi1
= lo1
+ __le64_to_cpu(bsb
.length2
);
3863 } else if (lo
< lo1
)
3869 (info
->reshape_progress
< lo
||
3870 info
->reshape_progress
> hi
))
3871 /* backup does not affect reshape_progress*/ ;
3872 else if (info
->delta_disks
>= 0) {
3873 info
->reshape_progress
= __le64_to_cpu(bsb
.arraystart
) +
3874 __le64_to_cpu(bsb
.length
);
3875 if (bsb
.magic
[15] == '2') {
3876 unsigned long long p2
= __le64_to_cpu(bsb
.arraystart2
) +
3877 __le64_to_cpu(bsb
.length2
);
3878 if (p2
> info
->reshape_progress
)
3879 info
->reshape_progress
= p2
;
3882 info
->reshape_progress
= __le64_to_cpu(bsb
.arraystart
);
3883 if (bsb
.magic
[15] == '2') {
3884 unsigned long long p2
= __le64_to_cpu(bsb
.arraystart2
);
3885 if (p2
< info
->reshape_progress
)
3886 info
->reshape_progress
= p2
;
3889 for (j
=0; j
<info
->array
.raid_disks
; j
++) {
3890 if (fdlist
[j
] < 0) continue;
3891 if (st
->ss
->load_super(st
, fdlist
[j
], NULL
))
3893 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
3894 dinfo
.reshape_progress
= info
->reshape_progress
;
3895 st
->ss
->update_super(st
, &dinfo
,
3896 "_reshape_progress",
3898 st
->ss
->store_super(st
, fdlist
[j
]);
3899 st
->ss
->free_super(st
);
3903 /* Didn't find any backup data, try to see if any
3906 if (info
->delta_disks
< 0) {
3907 /* When shrinking, the critical section is at the end.
3908 * So see if we are before the critical section.
3910 unsigned long long first_block
;
3911 nstripe
= ostripe
= 0;
3913 while (ostripe
>= nstripe
) {
3914 ostripe
+= info
->array
.chunk_size
/ 512;
3915 first_block
= ostripe
* odata
;
3916 nstripe
= first_block
/ ndata
/ (info
->new_chunk
/512) *
3917 (info
->new_chunk
/512);
3920 if (info
->reshape_progress
>= first_block
)
3923 if (info
->delta_disks
> 0) {
3924 /* See if we are beyond the critical section. */
3925 unsigned long long last_block
;
3926 nstripe
= ostripe
= 0;
3928 while (nstripe
>= ostripe
) {
3929 nstripe
+= info
->new_chunk
/ 512;
3930 last_block
= nstripe
* ndata
;
3931 ostripe
= last_block
/ odata
/ (info
->array
.chunk_size
/512) *
3932 (info
->array
.chunk_size
/512);
3935 if (info
->reshape_progress
>= last_block
)
3938 /* needed to recover critical section! */
3940 pr_err("Failed to find backup of critical section\n");
3944 int Grow_continue_command(char *devname
, int fd
,
3945 char *backup_file
, int verbose
)
3948 struct supertype
*st
= NULL
;
3949 struct mdinfo
*content
= NULL
;
3950 struct mdinfo array
;
3951 char *subarray
= NULL
;
3952 struct mdinfo
*cc
= NULL
;
3953 struct mdstat_ent
*mdstat
= NULL
;
3958 dprintf("Grow continue from command line called for %s\n",
3961 st
= super_by_fd(fd
, &subarray
);
3962 if (!st
|| !st
->ss
) {
3963 pr_err("Unable to determine metadata format for %s\n",
3967 dprintf("Grow continue is run for ");
3968 if (st
->ss
->external
== 0) {
3969 dprintf("native array (%s)\n", devname
);
3970 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) < 0) {
3971 pr_err("%s is not an active md array -"
3972 " aborting\n", devname
);
3974 goto Grow_continue_command_exit
;
3977 sysfs_init(content
, fd
, st
->devnum
);
3982 dprintf("subarray (%s)\n", subarray
);
3983 container_dev
= st
->container_dev
;
3984 cfd
= open_dev_excl(st
->container_dev
);
3986 container_dev
= st
->devnum
;
3988 cfd
= open_dev_excl(st
->devnum
);
3989 dprintf("container (%i)\n", container_dev
);
3993 pr_err("Unable to open container "
3994 "for %s\n", devname
);
3996 goto Grow_continue_command_exit
;
3998 fmt_devname(buf
, container_dev
);
4000 /* find in container array under reshape
4002 ret_val
= st
->ss
->load_container(st
, cfd
, NULL
);
4004 pr_err("Cannot read superblock for %s\n",
4007 goto Grow_continue_command_exit
;
4010 cc
= st
->ss
->container_content(st
, subarray
);
4011 for (content
= cc
; content
; content
= content
->next
) {
4013 int allow_reshape
= 1;
4015 if (content
->reshape_active
== 0)
4017 /* The decision about array or container wide
4018 * reshape is taken in Grow_continue based
4019 * content->reshape_active state, therefore we
4020 * need to check_reshape based on
4021 * reshape_active and subarray name
4023 if (content
->array
.state
& (1<<MD_SB_BLOCK_VOLUME
))
4025 if (content
->reshape_active
== CONTAINER_RESHAPE
&&
4026 (content
->array
.state
4027 & (1<<MD_SB_BLOCK_CONTAINER_RESHAPE
)))
4030 if (!allow_reshape
) {
4031 pr_err("cannot continue reshape of an array"
4032 " in container with unsupported"
4033 " metadata: %s(%s)\n",
4036 goto Grow_continue_command_exit
;
4039 array
= strchr(content
->text_version
+1, '/')+1;
4040 mdstat
= mdstat_by_subdev(array
, container_dev
);
4043 if (mdstat
->active
== 0) {
4044 pr_err("Skipping inactive "
4045 "array md%i.\n", mdstat
->devnum
);
4046 free_mdstat(mdstat
);
4053 pr_err("Unable to determine reshaped "
4054 "array for %s\n", devname
);
4056 goto Grow_continue_command_exit
;
4058 fd2
= open_dev(mdstat
->devnum
);
4060 pr_err("cannot open (md%i)\n",
4063 goto Grow_continue_command_exit
;
4066 sysfs_init(content
, fd2
, mdstat
->devnum
);
4068 /* start mdmon in case it is not running
4070 if (!mdmon_running(container_dev
))
4071 start_mdmon(container_dev
);
4074 if (mdmon_running(container_dev
))
4075 st
->update_tail
= &st
->updates
;
4077 pr_err("No mdmon found. "
4078 "Grow cannot continue.\n");
4080 goto Grow_continue_command_exit
;
4084 /* verify that array under reshape is started from
4087 if (verify_reshape_position(content
,
4088 map_name(pers
, mdstat
->level
)) < 0) {
4090 goto Grow_continue_command_exit
;
4095 ret_val
= Grow_continue(fd
, st
, content
, backup_file
, 0);
4097 Grow_continue_command_exit
:
4102 st
->ss
->free_super(st
);
4103 free_mdstat(mdstat
);
4110 int Grow_continue(int mdfd
, struct supertype
*st
, struct mdinfo
*info
,
4111 char *backup_file
, int freeze_reshape
)
4115 if (!info
->reshape_active
)
4118 if (st
->ss
->external
) {
4120 int cfd
= open_dev(st
->container_dev
);
4125 fmt_devname(container
, st
->container_dev
);
4126 st
->ss
->load_container(st
, cfd
, container
);
4128 ret_val
= reshape_container(container
, NULL
, mdfd
,
4129 st
, info
, 0, backup_file
,
4130 0, 1, freeze_reshape
);
4132 ret_val
= reshape_array(NULL
, mdfd
, "array", st
, info
, 1,
4133 NULL
, backup_file
, 0, 0, 1,