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
= malloc(sizeof(int) * disk_count
);
55 Name
": cannot allocate memory for disk list\n");
58 for (i
= 0; i
< next_spare
; i
++)
60 for (dev
= content
->devs
; dev
; dev
= dev
->next
) {
66 fd
= dev_open(buf
, O_RDWR
);
68 if (dev
->disk
.raid_disk
>= 0)
69 fdlist
[dev
->disk
.raid_disk
] = fd
;
71 fdlist
[next_spare
++] = fd
;
74 if (st
->ss
->external
&& st
->ss
->recover_backup
)
75 err
= st
->ss
->recover_backup(st
, content
);
77 err
= Grow_restart(st
, content
, fdlist
, next_spare
,
78 backup_file
, verbose
> 0);
80 while (next_spare
> 0) {
82 if (fdlist
[next_spare
] >= 0)
83 close(fdlist
[next_spare
]);
87 fprintf(stderr
, Name
": Failed to restore critical"
88 " section for reshape - sorry.\n");
90 fprintf(stderr
, Name
": Possibly you need"
91 " to specify a --backup-file\n");
95 dprintf("restore_backup() returns status OK.\n");
99 int Grow_Add_device(char *devname
, int fd
, char *newdev
)
101 /* Add a device to an active array.
102 * Currently, just extend a linear array.
103 * This requires writing a new superblock on the
104 * new device, calling the kernel to add the device,
105 * and if that succeeds, update the superblock on
107 * This means that we need to *find* all other devices.
114 struct supertype
*st
= NULL
;
115 char *subarray
= NULL
;
117 if (ioctl(fd
, GET_ARRAY_INFO
, &info
.array
) < 0) {
118 fprintf(stderr
, Name
": cannot get array info for %s\n", devname
);
122 if (info
.array
.level
!= -1) {
123 fprintf(stderr
, Name
": can only add devices to linear arrays\n");
127 st
= super_by_fd(fd
, &subarray
);
129 fprintf(stderr
, Name
": cannot handle arrays with superblock version %d\n", info
.array
.major_version
);
134 fprintf(stderr
, Name
": Cannot grow linear sub-arrays yet\n");
140 nfd
= open(newdev
, O_RDWR
|O_EXCL
|O_DIRECT
);
142 fprintf(stderr
, Name
": cannot open %s\n", newdev
);
147 if ((stb
.st_mode
& S_IFMT
) != S_IFBLK
) {
148 fprintf(stderr
, Name
": %s is not a block device!\n", newdev
);
153 /* now check out all the devices and make sure we can read the superblock */
154 for (d
=0 ; d
< info
.array
.raid_disks
; d
++) {
155 mdu_disk_info_t disk
;
158 st
->ss
->free_super(st
);
161 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0) {
162 fprintf(stderr
, Name
": cannot get device detail for device %d\n",
168 dv
= map_dev(disk
.major
, disk
.minor
, 1);
170 fprintf(stderr
, Name
": cannot find device file for device %d\n",
176 fd2
= dev_open(dv
, O_RDWR
);
178 fprintf(stderr
, Name
": cannot open device file %s\n", dv
);
184 if (st
->ss
->load_super(st
, fd2
, NULL
)) {
185 fprintf(stderr
, Name
": cannot find super block on %s\n", dv
);
193 /* Ok, looks good. Lets update the superblock and write it out to
197 info
.disk
.number
= d
;
198 info
.disk
.major
= major(stb
.st_rdev
);
199 info
.disk
.minor
= minor(stb
.st_rdev
);
200 info
.disk
.raid_disk
= d
;
201 info
.disk
.state
= (1 << MD_DISK_SYNC
) | (1 << MD_DISK_ACTIVE
);
202 st
->ss
->update_super(st
, &info
, "linear-grow-new", newdev
,
205 if (st
->ss
->store_super(st
, nfd
)) {
206 fprintf(stderr
, Name
": Cannot store new superblock on %s\n",
213 if (ioctl(fd
, ADD_NEW_DISK
, &info
.disk
) != 0) {
214 fprintf(stderr
, Name
": Cannot add new disk to this array\n");
217 /* Well, that seems to have worked.
218 * Now go through and update all superblocks
221 if (ioctl(fd
, GET_ARRAY_INFO
, &info
.array
) < 0) {
222 fprintf(stderr
, Name
": cannot get array info for %s\n", devname
);
227 for (d
=0 ; d
< info
.array
.raid_disks
; d
++) {
228 mdu_disk_info_t disk
;
232 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0) {
233 fprintf(stderr
, Name
": cannot get device detail for device %d\n",
237 dv
= map_dev(disk
.major
, disk
.minor
, 1);
239 fprintf(stderr
, Name
": cannot find device file for device %d\n",
243 fd2
= dev_open(dv
, O_RDWR
);
245 fprintf(stderr
, Name
": cannot open device file %s\n", dv
);
248 if (st
->ss
->load_super(st
, fd2
, NULL
)) {
249 fprintf(stderr
, Name
": cannot find super block on %s\n", dv
);
253 info
.array
.raid_disks
= nd
+1;
254 info
.array
.nr_disks
= nd
+1;
255 info
.array
.active_disks
= nd
+1;
256 info
.array
.working_disks
= nd
+1;
258 st
->ss
->update_super(st
, &info
, "linear-grow-update", dv
,
261 if (st
->ss
->store_super(st
, fd2
)) {
262 fprintf(stderr
, Name
": Cannot store new superblock on %s\n", dv
);
272 int Grow_addbitmap(char *devname
, int fd
, char *file
, int chunk
, int delay
, int write_behind
, int force
)
275 * First check that array doesn't have a bitmap
276 * Then create the bitmap
279 * For internal bitmaps, we need to check the version,
280 * find all the active devices, and write the bitmap block
283 mdu_bitmap_file_t bmf
;
284 mdu_array_info_t array
;
285 struct supertype
*st
;
286 char *subarray
= NULL
;
287 int major
= BITMAP_MAJOR_HI
;
288 int vers
= md_get_version(fd
);
289 unsigned long long bitmapsize
, array_size
;
292 major
= BITMAP_MAJOR_HOSTENDIAN
;
293 fprintf(stderr
, Name
": Warning - bitmaps created on this kernel"
294 " are not portable\n"
295 " between different architectures. Consider upgrading"
296 " the Linux kernel.\n");
299 if (ioctl(fd
, GET_BITMAP_FILE
, &bmf
) != 0) {
301 fprintf(stderr
, Name
": Memory allocation failure.\n");
303 fprintf(stderr
, Name
": bitmaps not supported by this kernel.\n");
306 if (bmf
.pathname
[0]) {
307 if (strcmp(file
,"none")==0) {
308 if (ioctl(fd
, SET_BITMAP_FILE
, -1)!= 0) {
309 fprintf(stderr
, Name
": failed to remove bitmap %s\n",
315 fprintf(stderr
, Name
": %s already has a bitmap (%s)\n",
316 devname
, bmf
.pathname
);
319 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
320 fprintf(stderr
, Name
": cannot get array status for %s\n", devname
);
323 if (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)) {
324 if (strcmp(file
, "none")==0) {
325 array
.state
&= ~(1<<MD_SB_BITMAP_PRESENT
);
326 if (ioctl(fd
, SET_ARRAY_INFO
, &array
)!= 0) {
327 fprintf(stderr
, Name
": failed to remove internal bitmap.\n");
332 fprintf(stderr
, Name
": Internal bitmap already present on %s\n",
337 if (strcmp(file
, "none") == 0) {
338 fprintf(stderr
, Name
": no bitmap found on %s\n", devname
);
341 if (array
.level
<= 0) {
342 fprintf(stderr
, Name
": Bitmaps not meaningful with level %s\n",
343 map_num(pers
, array
.level
)?:"of this array");
346 bitmapsize
= array
.size
;
348 if (get_dev_size(fd
, NULL
, &array_size
) &&
349 array_size
> (0x7fffffffULL
<<9)) {
350 /* Array is big enough that we cannot trust array.size
351 * try other approaches
353 bitmapsize
= get_component_size(fd
);
355 if (bitmapsize
== 0) {
356 fprintf(stderr
, Name
": Cannot reliably determine size of array to create bitmap - sorry.\n");
360 if (array
.level
== 10) {
361 int ncopies
= (array
.layout
&255)*((array
.layout
>>8)&255);
362 bitmapsize
= bitmapsize
* array
.raid_disks
/ ncopies
;
365 st
= super_by_fd(fd
, &subarray
);
367 fprintf(stderr
, Name
": Cannot understand version %d.%d\n",
368 array
.major_version
, array
.minor_version
);
372 fprintf(stderr
, Name
": Cannot add bitmaps to sub-arrays yet\n");
377 if (strcmp(file
, "internal") == 0) {
380 int offset_setable
= 0;
382 if (st
->ss
->add_internal_bitmap
== NULL
) {
383 fprintf(stderr
, Name
": Internal bitmaps not supported "
384 "with %s metadata\n", st
->ss
->name
);
387 mdi
= sysfs_read(fd
, -1, GET_BITMAP_LOCATION
);
390 for (d
=0; d
< st
->max_devs
; d
++) {
391 mdu_disk_info_t disk
;
394 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0)
396 if (disk
.major
== 0 &&
399 if ((disk
.state
& (1<<MD_DISK_SYNC
))==0)
401 dv
= map_dev(disk
.major
, disk
.minor
, 1);
403 int fd2
= dev_open(dv
, O_RDWR
);
406 if (st
->ss
->load_super(st
, fd2
, NULL
)==0) {
407 if (st
->ss
->add_internal_bitmap(
409 &chunk
, delay
, write_behind
,
410 bitmapsize
, offset_setable
,
413 st
->ss
->write_bitmap(st
, fd2
);
415 fprintf(stderr
, Name
": failed "
416 "to create internal bitmap - chunksize problem.\n");
424 if (offset_setable
) {
425 st
->ss
->getinfo_super(st
, mdi
, NULL
);
426 sysfs_init(mdi
, fd
, -1);
427 rv
= sysfs_set_num(mdi
, NULL
, "bitmap/location",
430 array
.state
|= (1<<MD_SB_BITMAP_PRESENT
);
431 rv
= ioctl(fd
, SET_ARRAY_INFO
, &array
);
436 ": Cannot add bitmap while array is"
437 " resyncing or reshaping etc.\n");
438 fprintf(stderr
, Name
": failed to set internal bitmap.\n");
445 int max_devs
= st
->max_devs
;
447 /* try to load a superblock */
448 for (d
=0; d
<max_devs
; d
++) {
449 mdu_disk_info_t disk
;
453 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0)
455 if ((disk
.major
==0 && disk
.minor
==0) ||
456 (disk
.state
& (1<<MD_DISK_REMOVED
)))
458 dv
= map_dev(disk
.major
, disk
.minor
, 1);
460 fd2
= dev_open(dv
, O_RDONLY
);
462 if (st
->ss
->load_super(st
, fd2
, NULL
) == 0) {
464 st
->ss
->uuid_from_super(st
, uuid
);
471 fprintf(stderr
, Name
": cannot find UUID for array!\n");
474 if (CreateBitmap(file
, force
, (char*)uuid
, chunk
,
475 delay
, write_behind
, bitmapsize
, major
)) {
478 bitmap_fd
= open(file
, O_RDWR
);
480 fprintf(stderr
, Name
": weird: %s cannot be opened\n",
484 if (ioctl(fd
, SET_BITMAP_FILE
, bitmap_fd
) < 0) {
488 ": Cannot add bitmap while array is"
489 " resyncing or reshaping etc.\n");
490 fprintf(stderr
, Name
": Cannot set bitmap file for %s: %s\n",
491 devname
, strerror(err
));
500 * When reshaping an array we might need to backup some data.
501 * This is written to all spares with a 'super_block' describing it.
502 * The superblock goes 4K from the end of the used space on the
504 * It if written after the backup is complete.
505 * It has the following structure.
508 static struct mdp_backup_super
{
509 char magic
[16]; /* md_backup_data-1 or -2 */
512 /* start/sizes in 512byte sectors */
513 __u64 devstart
; /* address on backup device/file of data */
516 __u32 sb_csum
; /* csum of preceeding bytes. */
518 __u64 devstart2
; /* offset in to data of second section */
521 __u32 sb_csum2
; /* csum of preceeding bytes. */
523 } __attribute__((aligned(512))) bsb
, bsb2
;
525 static __u32
bsb_csum(char *buf
, int len
)
529 for (i
=0; i
<len
; i
++)
530 csum
= (csum
<<3) + buf
[0];
531 return __cpu_to_le32(csum
);
534 static int check_idle(struct supertype
*st
)
536 /* Check that all member arrays for this container, or the
537 * container of this array, are idle
539 int container_dev
= (st
->container_dev
!= NoMdDev
540 ? st
->container_dev
: st
->devnum
);
542 struct mdstat_ent
*ent
, *e
;
545 fmt_devname(container
, container_dev
);
546 ent
= mdstat_read(0, 0);
547 for (e
= ent
; e
; e
= e
->next
) {
548 if (!is_container_member(e
, container
))
550 if (e
->percent
>= 0) {
559 static int freeze_container(struct supertype
*st
)
561 int container_dev
= (st
->container_dev
!= NoMdDev
562 ? st
->container_dev
: st
->devnum
);
568 fmt_devname(container
, container_dev
);
570 if (block_monitor(container
, 1)) {
571 fprintf(stderr
, Name
": failed to freeze container\n");
578 static void unfreeze_container(struct supertype
*st
)
580 int container_dev
= (st
->container_dev
!= NoMdDev
581 ? st
->container_dev
: st
->devnum
);
584 fmt_devname(container
, container_dev
);
586 unblock_monitor(container
, 1);
589 static int freeze(struct supertype
*st
)
591 /* Try to freeze resync/rebuild on this array/container.
592 * Return -1 if the array is busy,
593 * return -2 container cannot be frozen,
594 * return 0 if this kernel doesn't support 'frozen'
595 * return 1 if it worked.
597 if (st
->ss
->external
)
598 return freeze_container(st
);
600 struct mdinfo
*sra
= sysfs_read(-1, st
->devnum
, GET_VERSION
);
606 /* Need to clear any 'read-auto' status */
607 if (sysfs_get_str(sra
, NULL
, "array_state", buf
, 20) > 0 &&
608 strncmp(buf
, "read-auto", 9) == 0)
609 sysfs_set_str(sra
, NULL
, "array_state", "clean");
611 err
= sysfs_freeze_array(sra
);
617 static void unfreeze(struct supertype
*st
)
619 if (st
->ss
->external
)
620 return unfreeze_container(st
);
622 struct mdinfo
*sra
= sysfs_read(-1, st
->devnum
, GET_VERSION
);
625 sysfs_set_str(sra
, NULL
, "sync_action", "idle");
627 fprintf(stderr
, Name
": failed to unfreeze array\n");
632 static void wait_reshape(struct mdinfo
*sra
)
634 int fd
= sysfs_get_fd(sra
, NULL
, "sync_action");
640 while (sysfs_fd_get_str(fd
, action
, 20) > 0 &&
641 strncmp(action
, "reshape", 7) == 0) {
645 select(fd
+1, NULL
, NULL
, &rfds
, NULL
);
650 static int reshape_super(struct supertype
*st
, long long size
, int level
,
651 int layout
, int chunksize
, int raid_disks
,
652 int delta_disks
, char *backup_file
, char *dev
,
655 /* nothing extra to check in the native case */
656 if (!st
->ss
->external
)
658 if (!st
->ss
->reshape_super
||
659 !st
->ss
->manage_reshape
) {
660 fprintf(stderr
, Name
": %s metadata does not support reshape\n",
665 return st
->ss
->reshape_super(st
, size
, level
, layout
, chunksize
,
666 raid_disks
, delta_disks
, backup_file
, dev
,
670 static void sync_metadata(struct supertype
*st
)
672 if (st
->ss
->external
) {
673 if (st
->update_tail
) {
674 flush_metadata_updates(st
);
675 st
->update_tail
= &st
->updates
;
677 st
->ss
->sync_metadata(st
);
681 static int subarray_set_num(char *container
, struct mdinfo
*sra
, char *name
, int n
)
683 /* when dealing with external metadata subarrays we need to be
684 * prepared to handle EAGAIN. The kernel may need to wait for
685 * mdmon to mark the array active so the kernel can handle
686 * allocations/writeback when preparing the reshape action
687 * (md_allow_write()). We temporarily disable safe_mode_delay
688 * to close a race with the array_state going clean before the
689 * next write to raid_disks / stripe_cache_size
694 /* only 'raid_disks' and 'stripe_cache_size' trigger md_allow_write */
696 (strcmp(name
, "raid_disks") != 0 &&
697 strcmp(name
, "stripe_cache_size") != 0))
698 return sysfs_set_num(sra
, NULL
, name
, n
);
700 rc
= sysfs_get_str(sra
, NULL
, "safe_mode_delay", safe
, sizeof(safe
));
703 sysfs_set_num(sra
, NULL
, "safe_mode_delay", 0);
704 rc
= sysfs_set_num(sra
, NULL
, name
, n
);
705 if (rc
< 0 && errno
== EAGAIN
) {
706 ping_monitor(container
);
707 /* if we get EAGAIN here then the monitor is not active
710 rc
= sysfs_set_num(sra
, NULL
, name
, n
);
712 sysfs_set_str(sra
, NULL
, "safe_mode_delay", safe
);
716 int start_reshape(struct mdinfo
*sra
, int already_running
,
717 int before_data_disks
, int data_disks
)
720 unsigned long long sync_max_to_set
;
722 sysfs_set_num(sra
, NULL
, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL
);
723 err
= sysfs_set_num(sra
, NULL
, "suspend_hi", sra
->reshape_progress
);
724 err
= err
?: sysfs_set_num(sra
, NULL
, "suspend_lo",
725 sra
->reshape_progress
);
726 if (before_data_disks
<= data_disks
)
727 sync_max_to_set
= sra
->reshape_progress
/ data_disks
;
729 sync_max_to_set
= (sra
->component_size
* data_disks
730 - sra
->reshape_progress
) / data_disks
;
731 if (!already_running
)
732 sysfs_set_num(sra
, NULL
, "sync_min", sync_max_to_set
);
733 err
= err
?: sysfs_set_num(sra
, NULL
, "sync_max", sync_max_to_set
);
734 if (!already_running
)
735 err
= err
?: sysfs_set_str(sra
, NULL
, "sync_action", "reshape");
740 void abort_reshape(struct mdinfo
*sra
)
742 sysfs_set_str(sra
, NULL
, "sync_action", "idle");
743 sysfs_set_num(sra
, NULL
, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL
);
744 sysfs_set_num(sra
, NULL
, "suspend_hi", 0);
745 sysfs_set_num(sra
, NULL
, "suspend_lo", 0);
746 sysfs_set_num(sra
, NULL
, "sync_min", 0);
747 sysfs_set_str(sra
, NULL
, "sync_max", "max");
750 int remove_disks_for_takeover(struct supertype
*st
,
755 struct mdinfo
*remaining
;
758 if (sra
->array
.level
== 10)
759 nr_of_copies
= layout
& 0xff;
760 else if (sra
->array
.level
== 1)
761 nr_of_copies
= sra
->array
.raid_disks
;
765 remaining
= sra
->devs
;
767 /* for each 'copy', select one device and remove from the list. */
768 for (slot
= 0; slot
< sra
->array
.raid_disks
; slot
+= nr_of_copies
) {
769 struct mdinfo
**diskp
;
772 /* Find a working device to keep */
773 for (diskp
= &remaining
; *diskp
; diskp
= &(*diskp
)->next
) {
774 struct mdinfo
*disk
= *diskp
;
776 if (disk
->disk
.raid_disk
< slot
)
778 if (disk
->disk
.raid_disk
>= slot
+ nr_of_copies
)
780 if (disk
->disk
.state
& (1<<MD_DISK_REMOVED
))
782 if (disk
->disk
.state
& (1<<MD_DISK_FAULTY
))
784 if (!(disk
->disk
.state
& (1<<MD_DISK_SYNC
)))
787 /* We have found a good disk to use! */
789 disk
->next
= sra
->devs
;
798 if (slot
< sra
->array
.raid_disks
) {
799 /* didn't find all slots */
805 sra
->devs
= remaining
;
809 /* Remove all 'remaining' devices from the array */
811 struct mdinfo
*sd
= remaining
;
812 remaining
= sd
->next
;
814 sysfs_set_str(sra
, sd
, "state", "faulty");
815 sysfs_set_str(sra
, sd
, "slot", "none");
816 /* for external metadata disks should be removed in mdmon */
817 if (!st
->ss
->external
)
818 sysfs_set_str(sra
, sd
, "state", "remove");
819 sd
->disk
.state
|= (1<<MD_DISK_REMOVED
);
820 sd
->disk
.state
&= ~(1<<MD_DISK_SYNC
);
821 sd
->next
= sra
->devs
;
827 void reshape_free_fdlist(int *fdlist
,
828 unsigned long long *offsets
,
833 for (i
= 0; i
< size
; i
++)
841 int reshape_prepare_fdlist(char *devname
,
845 unsigned long blocks
,
848 unsigned long long *offsets
)
853 for (d
= 0; d
<= nrdisks
; d
++)
856 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
857 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
859 if (sd
->disk
.state
& (1<<MD_DISK_SYNC
)) {
860 char *dn
= map_dev(sd
->disk
.major
,
862 fdlist
[sd
->disk
.raid_disk
]
863 = dev_open(dn
, O_RDONLY
);
864 offsets
[sd
->disk
.raid_disk
] = sd
->data_offset
*512;
865 if (fdlist
[sd
->disk
.raid_disk
] < 0) {
867 Name
": %s: cannot open component %s\n",
868 devname
, dn
? dn
: "-unknown-");
872 } else if (backup_file
== NULL
) {
874 char *dn
= map_dev(sd
->disk
.major
,
876 fdlist
[d
] = dev_open(dn
, O_RDWR
);
877 offsets
[d
] = (sd
->data_offset
+ sra
->component_size
- blocks
- 8)*512;
879 fprintf(stderr
, Name
": %s: cannot open component %s\n",
880 devname
, dn
? dn
: "-unknown-");
891 int reshape_open_backup_file(char *backup_file
,
896 unsigned long long *offsets
,
899 /* Return 1 on success, 0 on any form of failure */
900 /* need to check backup file is large enough */
906 *fdlist
= open(backup_file
, O_RDWR
|O_CREAT
|(restart
? O_TRUNC
: O_EXCL
),
910 fprintf(stderr
, Name
": %s: cannot create backup file %s: %s\n",
911 devname
, backup_file
, strerror(errno
));
914 /* Guard against backup file being on array device.
915 * If array is partitioned or if LVM etc is in the
916 * way this will not notice, but it is better than
919 fstat(*fdlist
, &stb
);
922 if (stb
.st_rdev
== dev
) {
923 fprintf(stderr
, Name
": backup file must NOT be"
924 " on the array being reshaped.\n");
930 for (i
=0; i
< blocks
+ 8 ; i
++) {
931 if (write(*fdlist
, buf
, 512) != 512) {
932 fprintf(stderr
, Name
": %s: cannot create"
933 " backup file %s: %s\n",
934 devname
, backup_file
, strerror(errno
));
938 if (fsync(*fdlist
) != 0) {
939 fprintf(stderr
, Name
": %s: cannot create backup file %s: %s\n",
940 devname
, backup_file
, strerror(errno
));
947 unsigned long compute_backup_blocks(int nchunk
, int ochunk
,
948 unsigned int ndata
, unsigned int odata
)
950 unsigned long a
, b
, blocks
;
951 /* So how much do we need to backup.
952 * We need an amount of data which is both a whole number of
953 * old stripes and a whole number of new stripes.
954 * So LCM for (chunksize*datadisks).
956 a
= (ochunk
/512) * odata
;
957 b
= (nchunk
/512) * ndata
;
965 /* LCM == product / GCD */
966 blocks
= (ochunk
/512) * (nchunk
/512) * odata
* ndata
/ a
;
971 char *analyse_change(struct mdinfo
*info
, struct reshape
*re
)
973 /* Based on the current array state in info->array and
974 * the changes in info->new_* etc, determine:
975 * - whether the change is possible
976 * - Intermediate level/raid_disks/layout
977 * - whether a restriping reshape is needed
978 * - number of sectors in minimum change unit. This
979 * will cover a whole number of stripes in 'before' and
982 * Return message if the change should be rejected
983 * NULL if the change can be achieved
985 * This can be called as part of starting a reshape, or
986 * when assembling an array that is undergoing reshape.
989 /* delta_parity records change in number of devices
990 * caused by level change
992 int delta_parity
= 0;
994 /* If a new level not explicitly given, we assume no-change */
995 if (info
->new_level
== UnSet
)
996 info
->new_level
= info
->array
.level
;
999 switch (info
->new_level
) {
1005 /* chunk size is meaningful, must divide component_size
1008 if (info
->component_size
% (info
->new_chunk
/512))
1009 return "New chunk size does not"
1010 " divide component size";
1013 return "chunk size not meaningful for this level";
1016 info
->new_chunk
= info
->array
.chunk_size
;
1018 switch (info
->array
.level
) {
1020 return "Cannot understand this RAID level";
1022 /* RAID1 can convert to RAID1 with different disks, or
1023 * raid5 with 2 disks, or
1026 if (info
->new_level
> 1 &&
1027 (info
->component_size
& 7))
1028 return "Cannot convert RAID1 of this size - "
1029 "reduce size to multiple of 4K first.";
1030 if (info
->new_level
== 0) {
1031 if (info
->delta_disks
!= UnSet
&&
1032 info
->delta_disks
!= 0)
1033 return "Cannot change number of disks "
1034 "with RAID1->RAID0 conversion";
1036 re
->before
.data_disks
= 1;
1037 re
->after
.data_disks
= 1;
1038 re
->before
.layout
= 0;
1039 re
->backup_blocks
= 0;
1043 if (info
->new_level
== 1) {
1044 if (info
->delta_disks
== UnSet
)
1045 /* Don't know what to do */
1046 return "no change requested for Growing RAID1";
1048 re
->backup_blocks
= 0;
1052 if (info
->array
.raid_disks
== 2 &&
1053 info
->new_level
== 5) {
1056 re
->before
.data_disks
= 1;
1057 if (info
->delta_disks
!= UnSet
&&
1058 info
->delta_disks
!= 0)
1059 re
->after
.data_disks
= 1 + info
->delta_disks
;
1061 re
->after
.data_disks
= 1;
1062 if (re
->after
.data_disks
< 1)
1063 return "Number of disks too small for RAID5";
1065 re
->before
.layout
= ALGORITHM_LEFT_SYMMETRIC
;
1066 info
->array
.chunk_size
= 65536;
1069 /* Could do some multi-stage conversions, but leave that to
1072 return "Impossibly level change request for RAID1";
1075 /* RAID10 can only be converted from near mode to
1076 * RAID0 by removing some devices
1078 if ((info
->array
.layout
& ~0xff) != 0x100)
1079 return "Cannot Grow RAID10 with far/offset layout";
1080 /* number of devices must be multiple of number of copies */
1081 if (info
->array
.raid_disks
% (info
->array
.layout
& 0xff))
1082 return "RAID10 layout too complex for Grow operation";
1084 if (info
->new_level
!= 0)
1085 return "RAID10 can only be changed to RAID0";
1086 new_disks
= (info
->array
.raid_disks
1087 / (info
->array
.layout
& 0xff));
1088 if (info
->delta_disks
== UnSet
)
1089 info
->delta_disks
= (new_disks
1090 - info
->array
.raid_disks
);
1092 if (info
->delta_disks
!= new_disks
- info
->array
.raid_disks
)
1093 return "New number of raid-devices impossible for RAID10";
1094 if (info
->new_chunk
&&
1095 info
->new_chunk
!= info
->array
.chunk_size
)
1096 return "Cannot change chunk-size with RAID10 Grow";
1101 re
->before
.data_disks
= new_disks
;
1102 re
->after
.data_disks
= re
->before
.data_disks
;
1103 re
->before
.layout
= 0;
1104 re
->backup_blocks
= 0;
1108 /* RAID0 can be converted to RAID10, or to RAID456 */
1109 if (info
->new_level
== 10) {
1110 if (info
->new_layout
== UnSet
&& info
->delta_disks
== UnSet
) {
1111 /* Assume near=2 layout */
1112 info
->new_layout
= 0x102;
1113 info
->delta_disks
= info
->array
.raid_disks
;
1115 if (info
->new_layout
== UnSet
) {
1116 int copies
= 1 + (info
->delta_disks
1117 / info
->array
.raid_disks
);
1118 if (info
->array
.raid_disks
* (copies
-1)
1119 != info
->delta_disks
)
1120 return "Impossible number of devices"
1121 " for RAID0->RAID10";
1122 info
->new_layout
= 0x100 + copies
;
1124 if (info
->delta_disks
== UnSet
) {
1125 int copies
= info
->new_layout
& 0xff;
1126 if (info
->new_layout
!= 0x100 + copies
)
1127 return "New layout impossible"
1128 " for RAID0->RAID10";;
1129 info
->delta_disks
= (copies
- 1) *
1130 info
->array
.raid_disks
;
1132 if (info
->new_chunk
&&
1133 info
->new_chunk
!= info
->array
.chunk_size
)
1134 return "Cannot change chunk-size with RAID0->RAID10";
1138 re
->before
.data_disks
= (info
->array
.raid_disks
+
1140 re
->after
.data_disks
= re
->before
.data_disks
;
1141 re
->before
.layout
= info
->new_layout
;
1142 re
->backup_blocks
= 0;
1146 /* RAID0 can also covert to RAID0/4/5/6 by first converting to
1147 * a raid4 style layout of the final level.
1149 switch (info
->new_level
) {
1154 re
->before
.layout
= 0;
1159 re
->before
.layout
= ALGORITHM_PARITY_N
;
1164 re
->before
.layout
= ALGORITHM_PARITY_N
;
1167 return "Impossible level change requested";
1169 re
->before
.data_disks
= info
->array
.raid_disks
;
1170 /* determining 'after' layout happens outside this 'switch' */
1174 info
->array
.layout
= ALGORITHM_PARITY_N
;
1176 switch (info
->new_level
) {
1180 re
->level
= info
->array
.level
;
1181 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1182 re
->before
.layout
= info
->array
.layout
;
1186 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1187 re
->before
.layout
= info
->array
.layout
;
1192 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1193 switch (info
->array
.layout
) {
1194 case ALGORITHM_LEFT_ASYMMETRIC
:
1195 re
->before
.layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
1197 case ALGORITHM_RIGHT_ASYMMETRIC
:
1198 re
->before
.layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
1200 case ALGORITHM_LEFT_SYMMETRIC
:
1201 re
->before
.layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
1203 case ALGORITHM_RIGHT_SYMMETRIC
:
1204 re
->before
.layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
1206 case ALGORITHM_PARITY_0
:
1207 re
->before
.layout
= ALGORITHM_PARITY_0_6
;
1209 case ALGORITHM_PARITY_N
:
1210 re
->before
.layout
= ALGORITHM_PARITY_N_6
;
1213 return "Cannot convert an array with this layout";
1217 if (info
->array
.raid_disks
!= 2)
1218 return "Can only convert a 2-device array to RAID1";
1219 if (info
->delta_disks
!= UnSet
&&
1220 info
->delta_disks
!= 0)
1221 return "Cannot set raid_disk when "
1222 "converting RAID5->RAID1";
1226 return "Impossible level change requested";
1230 switch (info
->new_level
) {
1236 re
->before
.data_disks
= info
->array
.raid_disks
- 2;
1237 re
->before
.layout
= info
->array
.layout
;
1240 return "Impossible level change requested";
1245 /* If we reached here then it looks like a re-stripe is
1246 * happening. We have determined the intermediate level
1247 * and initial raid_disks/layout and stored these in 're'.
1249 * We need to deduce the final layout that can be atomically
1250 * converted to the end state.
1252 switch (info
->new_level
) {
1254 /* We can only get to RAID0 from RAID4 or RAID5
1255 * with appropriate layout and one extra device
1257 if (re
->level
!= 4 && re
->level
!= 5)
1258 return "Cannot covert to RAID0 from this level";
1260 switch (re
->level
) {
1262 re
->after
.layout
= 0 ; break;
1264 re
->after
.layout
= ALGORITHM_PARITY_N
; break;
1269 /* We can only get to RAID4 from RAID5 */
1270 if (re
->level
!= 4 && re
->level
!= 5)
1271 return "Cannot convert to RAID4 from this level";
1273 switch (re
->level
) {
1275 re
->after
.layout
= 0 ; break;
1277 re
->after
.layout
= ALGORITHM_PARITY_N
; break;
1282 /* We get to RAID5 from RAID5 or RAID6 */
1283 if (re
->level
!= 5 && re
->level
!= 6)
1284 return "Cannot convert to RAID5 from this level";
1286 switch (re
->level
) {
1288 if (info
->new_layout
== UnSet
)
1289 re
->after
.layout
= re
->before
.layout
;
1291 re
->after
.layout
= info
->new_layout
;
1294 if (info
->new_layout
== UnSet
)
1295 info
->new_layout
= re
->before
.layout
;
1297 /* after.layout needs to be raid6 version of new_layout */
1298 if (info
->new_layout
== ALGORITHM_PARITY_N
)
1299 re
->after
.layout
= ALGORITHM_PARITY_N
;
1302 char *ls
= map_num(r5layout
, info
->new_layout
);
1305 /* Current RAID6 layout has a RAID5
1308 strcat(strcpy(layout
, ls
), "-6");
1309 l
= map_name(r6layout
, layout
);
1311 return "Cannot find RAID6 layout"
1314 /* Current RAID6 has no equivalent.
1315 * If it is already a '-6' layout we
1316 * can leave it unchanged, else we must
1319 ls
= map_num(r6layout
, info
->new_layout
);
1321 strcmp(ls
+strlen(ls
)-2, "-6") != 0)
1322 return "Please specify new layout";
1323 l
= info
->new_layout
;
1325 re
->after
.layout
= l
;
1331 /* We must already be at level 6 */
1333 return "Impossible level change";
1334 if (info
->new_layout
== UnSet
)
1335 re
->after
.layout
= info
->array
.layout
;
1337 re
->after
.layout
= info
->new_layout
;
1340 return "Impossible level change requested";
1342 if (info
->delta_disks
== UnSet
)
1343 info
->delta_disks
= delta_parity
;
1345 re
->after
.data_disks
= (re
->before
.data_disks
1348 switch (re
->level
) {
1349 case 6: re
->parity
= 2; break;
1351 case 5: re
->parity
= 1; break;
1352 default: re
->parity
= 0; break;
1354 /* So we have a restripe operation, we need to calculate the number
1355 * of blocks per reshape operation.
1357 if (info
->new_chunk
== 0)
1358 info
->new_chunk
= info
->array
.chunk_size
;
1359 if (re
->after
.data_disks
== re
->before
.data_disks
&&
1360 re
->after
.layout
== re
->before
.layout
&&
1361 info
->new_chunk
== info
->array
.chunk_size
) {
1362 /* Nothing to change */
1363 re
->backup_blocks
= 0;
1366 if (re
->after
.data_disks
== 1 && re
->before
.data_disks
== 1) {
1367 /* chunk and layout changes make no difference */
1368 re
->backup_blocks
= 0;
1372 if (re
->after
.data_disks
== re
->before
.data_disks
&&
1373 get_linux_version() < 2006032)
1374 return "in-place reshape is not safe before 2.6.32 - sorry.";
1376 if (re
->after
.data_disks
< re
->before
.data_disks
&&
1377 get_linux_version() < 2006030)
1378 return "reshape to fewer devices is not supported before 2.6.30 - sorry.";
1380 re
->backup_blocks
= compute_backup_blocks(
1381 info
->new_chunk
, info
->array
.chunk_size
,
1382 re
->after
.data_disks
,
1383 re
->before
.data_disks
);
1385 re
->new_size
= info
->component_size
* re
->after
.data_disks
;
1389 static int reshape_array(char *container
, int fd
, char *devname
,
1390 struct supertype
*st
, struct mdinfo
*info
,
1391 int force
, struct mddev_dev
*devlist
,
1392 char *backup_file
, int quiet
, int forked
,
1393 int restart
, int freeze_reshape
);
1394 static int reshape_container(char *container
, char *devname
,
1396 struct supertype
*st
,
1397 struct mdinfo
*info
,
1400 int quiet
, int restart
, int freeze_reshape
);
1402 int Grow_reshape(char *devname
, int fd
, int quiet
, char *backup_file
,
1404 int level
, char *layout_str
, int chunksize
, int raid_disks
,
1405 struct mddev_dev
*devlist
,
1406 int assume_clean
, int force
)
1408 /* Make some changes in the shape of an array.
1409 * The kernel must support the change.
1411 * There are three different changes. Each can trigger
1412 * a resync or recovery so we freeze that until we have
1413 * requested everything (if kernel supports freezing - 2.6.30).
1415 * - change size (i.e. component_size)
1417 * - change layout/chunksize/ndisks
1419 * The last can require a reshape. It is different on different
1420 * levels so we need to check the level before actioning it.
1421 * Some times the level change needs to be requested after the
1422 * reshape (e.g. raid6->raid5, raid5->raid0)
1425 struct mdu_array_info_s array
;
1427 struct supertype
*st
;
1428 char *subarray
= NULL
;
1432 char *container
= NULL
;
1433 char container_buf
[20];
1436 struct mddev_dev
*dv
;
1442 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) < 0) {
1443 fprintf(stderr
, Name
": %s is not an active md array - aborting\n",
1449 (chunksize
|| level
!= UnSet
|| layout_str
|| raid_disks
)) {
1450 fprintf(stderr
, Name
": cannot change component size at the same time "
1451 "as other changes.\n"
1452 " Change size first, then check data is intact before "
1453 "making other changes.\n");
1457 if (raid_disks
&& raid_disks
< array
.raid_disks
&& array
.level
> 1 &&
1458 get_linux_version() < 2006032 &&
1459 !check_env("MDADM_FORCE_FEWER")) {
1460 fprintf(stderr
, Name
": reducing the number of devices is not safe before Linux 2.6.32\n"
1461 " Please use a newer kernel\n");
1465 st
= super_by_fd(fd
, &subarray
);
1467 fprintf(stderr
, Name
": Unable to determine metadata format for %s\n", devname
);
1470 if (raid_disks
> st
->max_devs
) {
1471 fprintf(stderr
, Name
": Cannot increase raid-disks on this array"
1472 " beyond %d\n", st
->max_devs
);
1476 /* in the external case we need to check that the requested reshape is
1477 * supported, and perform an initial check that the container holds the
1478 * pre-requisite spare devices (mdmon owns final validation)
1480 if (st
->ss
->external
) {
1485 container_dev
= st
->container_dev
;
1486 cfd
= open_dev_excl(st
->container_dev
);
1488 container_dev
= st
->devnum
;
1490 cfd
= open_dev_excl(st
->devnum
);
1494 fprintf(stderr
, Name
": Unable to open container for %s\n",
1500 fmt_devname(container_buf
, container_dev
);
1501 container
= container_buf
;
1503 rv
= st
->ss
->load_container(st
, cfd
, NULL
);
1506 fprintf(stderr
, Name
": Cannot read superblock for %s\n",
1512 /* check if operation is supported for metadata handler */
1513 if (st
->ss
->container_content
) {
1514 struct mdinfo
*cc
= NULL
;
1515 struct mdinfo
*content
= NULL
;
1517 cc
= st
->ss
->container_content(st
, subarray
);
1518 for (content
= cc
; content
; content
= content
->next
) {
1519 int allow_reshape
= 1;
1521 /* check if reshape is allowed based on metadata
1522 * indications stored in content.array.status
1524 if (content
->array
.state
& (1<<MD_SB_BLOCK_VOLUME
))
1526 if (content
->array
.state
1527 & (1<<MD_SB_BLOCK_CONTAINER_RESHAPE
))
1529 if (!allow_reshape
) {
1530 fprintf(stderr
, Name
1531 " cannot reshape arrays in"
1532 " container with unsupported"
1533 " metadata: %s(%s)\n",
1534 devname
, container_buf
);
1542 if (mdmon_running(container_dev
))
1543 st
->update_tail
= &st
->updates
;
1547 for (dv
= devlist
; dv
; dv
= dv
->next
)
1549 if (raid_disks
> array
.raid_disks
&&
1550 array
.spare_disks
+added_disks
< (raid_disks
- array
.raid_disks
) &&
1553 Name
": Need %d spare%s to avoid degraded array,"
1554 " and only have %d.\n"
1555 " Use --force to over-ride this check.\n",
1556 raid_disks
- array
.raid_disks
,
1557 raid_disks
- array
.raid_disks
== 1 ? "" : "s",
1558 array
.spare_disks
+ added_disks
);
1562 sra
= sysfs_read(fd
, 0, GET_LEVEL
| GET_DISKS
| GET_DEVS
1563 | GET_STATE
| GET_VERSION
);
1565 if (st
->ss
->external
&& subarray
== NULL
) {
1566 array
.level
= LEVEL_CONTAINER
;
1567 sra
->array
.level
= LEVEL_CONTAINER
;
1570 fprintf(stderr
, Name
": failed to read sysfs parameters for %s\n",
1574 frozen
= freeze(st
);
1576 /* freeze() already spewed the reason */
1579 } else if (frozen
< 0) {
1580 fprintf(stderr
, Name
": %s is performing resync/recovery and cannot"
1581 " be reshaped\n", devname
);
1586 /* ========= set size =============== */
1587 if (size
>= 0 && (size
== 0 || size
!= array
.size
)) {
1588 long long orig_size
= get_component_size(fd
)/2;
1589 long long min_csize
;
1593 orig_size
= array
.size
;
1595 if (reshape_super(st
, size
, UnSet
, UnSet
, 0, 0, UnSet
, NULL
,
1602 /* Update the size of each member device in case
1603 * they have been resized. This will never reduce
1604 * below the current used-size. The "size" attribute
1605 * understands '0' to mean 'max'.
1608 for (mdi
= sra
->devs
; mdi
; mdi
= mdi
->next
) {
1609 if (sysfs_set_num(sra
, mdi
, "size", size
) < 0)
1611 if (array
.not_persistent
== 0 &&
1612 array
.major_version
== 0 &&
1613 get_linux_version() < 3001000) {
1614 /* Dangerous to allow size to exceed 2TB */
1615 unsigned long long csize
;
1616 if (sysfs_get_ll(sra
, mdi
, "size", &csize
) == 0) {
1617 if (csize
>= 2ULL*1024*1024*1024)
1618 csize
= 2ULL*1024*1024*1024;
1619 if ((min_csize
== 0 || (min_csize
1620 > (long long)csize
)))
1625 if (min_csize
&& size
> min_csize
) {
1626 fprintf(stderr
, Name
": Cannot safely make this array "
1627 "use more than 2TB per device on this kernel.\n");
1631 if (min_csize
&& size
== 0) {
1632 /* Don't let the kernel choose a size - it will get
1635 fprintf(stderr
, Name
": Limited v0.90 array to "
1636 "2TB per device\n");
1641 if (array
.size
!= size
) {
1642 /* got truncated to 32bit, write to
1643 * component_size instead
1646 rv
= sysfs_set_num(sra
, NULL
,
1647 "component_size", size
);
1651 rv
= ioctl(fd
, SET_ARRAY_INFO
, &array
);
1655 /* restore metadata */
1656 if (reshape_super(st
, orig_size
, UnSet
, UnSet
, 0, 0,
1657 UnSet
, NULL
, devname
, !quiet
) == 0)
1659 fprintf(stderr
, Name
": Cannot set device size for %s: %s\n",
1660 devname
, strerror(err
));
1662 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
1663 fprintf(stderr
, " Bitmap must be removed before size can be changed\n");
1668 /* This will fail on kernels newer than 3.0 unless
1669 * a backport has been arranged.
1672 sysfs_set_str(sra
, NULL
, "resync_start", "none") < 0)
1673 fprintf(stderr
, Name
": --assume-clean not support with --grow on this kernel\n");
1675 ioctl(fd
, GET_ARRAY_INFO
, &array
);
1676 size
= get_component_size(fd
)/2;
1680 if (size
== orig_size
)
1681 fprintf(stderr
, Name
": component size of %s "
1682 "unchanged at %lluK\n",
1685 fprintf(stderr
, Name
": component size of %s "
1686 "has been set to %lluK\n",
1690 } else if (array
.level
!= LEVEL_CONTAINER
) {
1691 size
= get_component_size(fd
)/2;
1696 /* See if there is anything else to do */
1697 if ((level
== UnSet
|| level
== array
.level
) &&
1698 (layout_str
== NULL
) &&
1699 (chunksize
== 0 || chunksize
== array
.chunk_size
) &&
1700 (raid_disks
== 0 || raid_disks
== array
.raid_disks
)) {
1701 /* Nothing more to do */
1702 if (!changed
&& !quiet
)
1703 fprintf(stderr
, Name
": %s: no change requested\n",
1708 /* ========= check for Raid10/Raid1 -> Raid0 conversion ===============
1709 * current implementation assumes that following conditions must be met:
1712 * - near_copies == 2
1714 if ((level
== 0 && array
.level
== 10 && sra
&&
1715 array
.layout
== ((1 << 8) + 2) && !(array
.raid_disks
& 1)) ||
1716 (level
== 0 && array
.level
== 1 && sra
)) {
1718 err
= remove_disks_for_takeover(st
, sra
, array
.layout
);
1720 dprintf(Name
": Array cannot be reshaped\n");
1726 /* Make sure mdmon has seen the device removal
1727 * and updated metadata before we continue with
1731 ping_monitor(container
);
1734 memset(&info
, 0, sizeof(info
));
1736 sysfs_init(&info
, fd
, NoMdDev
);
1737 strcpy(info
.text_version
, sra
->text_version
);
1738 info
.component_size
= size
*2;
1739 info
.new_level
= level
;
1740 info
.new_chunk
= chunksize
* 1024;
1741 if (info
.array
.level
== LEVEL_CONTAINER
) {
1742 info
.delta_disks
= UnSet
;
1743 info
.array
.raid_disks
= raid_disks
;
1744 } else if (raid_disks
)
1745 info
.delta_disks
= raid_disks
- info
.array
.raid_disks
;
1747 info
.delta_disks
= UnSet
;
1748 if (layout_str
== NULL
) {
1749 info
.new_layout
= UnSet
;
1750 if (info
.array
.level
== 6 &&
1751 (info
.new_level
== 6 || info
.new_level
== UnSet
) &&
1752 info
.array
.layout
>= 16) {
1753 fprintf(stderr
, Name
1754 ": %s has a non-standard layout. If you"
1755 " wish to preserve this\n"
1756 " during the reshape, please specify"
1757 " --layout=preserve\n"
1758 " If you want to change it, specify a"
1759 " layout or use --layout=normalise\n",
1764 } else if (strcmp(layout_str
, "normalise") == 0 ||
1765 strcmp(layout_str
, "normalize") == 0) {
1766 /* If we have a -6 RAID6 layout, remove the '-6'. */
1767 info
.new_layout
= UnSet
;
1768 if (info
.array
.level
== 6 && info
.new_level
== UnSet
) {
1770 strcpy(l
, map_num(r6layout
, info
.array
.layout
));
1771 h
= strrchr(l
, '-');
1772 if (h
&& strcmp(h
, "-6") == 0) {
1774 info
.new_layout
= map_name(r6layout
, l
);
1777 } else if (strcmp(layout_str
, "preserve") == 0) {
1778 info
.new_layout
= UnSet
;
1780 int l
= info
.new_level
;
1782 l
= info
.array
.level
;
1785 info
.new_layout
= map_name(r5layout
, layout_str
);
1788 info
.new_layout
= map_name(r6layout
, layout_str
);
1791 info
.new_layout
= parse_layout_10(layout_str
);
1794 info
.new_layout
= parse_layout_faulty(layout_str
);
1797 fprintf(stderr
, Name
": layout not meaningful"
1798 " with this level\n");
1802 if (info
.new_layout
== UnSet
) {
1803 fprintf(stderr
, Name
": layout %s not understood"
1804 " for this level\n",
1811 if (array
.level
== LEVEL_FAULTY
) {
1812 if (level
!= UnSet
&& level
!= array
.level
) {
1813 fprintf(stderr
, Name
": cannot change level of Faulty device\n");
1817 fprintf(stderr
, Name
": cannot set chunksize of Faulty device\n");
1820 if (raid_disks
&& raid_disks
!= 1) {
1821 fprintf(stderr
, Name
": cannot set raid_disks of Faulty device\n");
1825 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
1826 dprintf("Cannot get array information.\n");
1829 array
.layout
= info
.new_layout
;
1830 if (ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
1831 fprintf(stderr
, Name
": failed to set new layout\n");
1834 printf("layout for %s set to %d\n",
1835 devname
, array
.layout
);
1837 } else if (array
.level
== LEVEL_CONTAINER
) {
1838 /* This change is to be applied to every array in the
1839 * container. This is only needed when the metadata imposes
1840 * restraints of the various arrays in the container.
1841 * Currently we only know that IMSM requires all arrays
1842 * to have the same number of devices so changing the
1843 * number of devices (On-Line Capacity Expansion) must be
1844 * performed at the level of the container
1846 rv
= reshape_container(container
, devname
, -1, st
, &info
,
1847 force
, backup_file
, quiet
, 0, 0);
1850 /* get spare devices from external metadata
1852 if (st
->ss
->external
) {
1853 struct mdinfo
*info2
;
1855 info2
= st
->ss
->container_content(st
, subarray
);
1857 info
.array
.spare_disks
=
1858 info2
->array
.spare_disks
;
1863 /* Impose these changes on a single array. First
1864 * check that the metadata is OK with the change. */
1866 if (reshape_super(st
, info
.component_size
, info
.new_level
,
1867 info
.new_layout
, info
.new_chunk
,
1868 info
.array
.raid_disks
, info
.delta_disks
,
1869 backup_file
, devname
, quiet
)) {
1874 rv
= reshape_array(container
, fd
, devname
, st
, &info
, force
,
1875 devlist
, backup_file
, quiet
, 0, 0, 0);
1885 /* verify_reshape_position()
1886 * Function checks if reshape position in metadata is not farther
1887 * than position in md.
1889 * 0 : not valid sysfs entry
1890 * it can be caused by not started reshape, it should be started
1891 * by reshape array or raid0 array is before takeover
1892 * -1 : error, reshape position is obviously wrong
1893 * 1 : success, reshape progress correct or updated
1895 static int verify_reshape_position(struct mdinfo
*info
, int level
)
1901 /* read sync_max, failure can mean raid0 array */
1902 rv
= sysfs_get_str(info
, NULL
, "sync_max", buf
, 40);
1906 unsigned long long position
= strtoull(buf
, &ep
, 0);
1908 dprintf(Name
": Read sync_max sysfs entry is: %s\n", buf
);
1909 if (!(ep
== buf
|| (*ep
!= 0 && *ep
!= '\n' && *ep
!= ' '))) {
1910 position
*= get_data_disks(level
,
1912 info
->array
.raid_disks
);
1913 if (info
->reshape_progress
< position
) {
1914 dprintf("Corrected reshape progress (%llu) to "
1915 "md position (%llu)\n",
1916 info
->reshape_progress
, position
);
1917 info
->reshape_progress
= position
;
1919 } else if (info
->reshape_progress
> position
) {
1920 fprintf(stderr
, Name
": Fatal error: array "
1921 "reshape was not properly frozen "
1922 "(expected reshape position is %llu, "
1923 "but reshape progress is %llu.\n",
1924 position
, info
->reshape_progress
);
1927 dprintf("Reshape position in md and metadata "
1932 } else if (rv
== 0) {
1933 /* for valid sysfs entry, 0-length content
1934 * should be indicated as error
1942 static int reshape_array(char *container
, int fd
, char *devname
,
1943 struct supertype
*st
, struct mdinfo
*info
,
1944 int force
, struct mddev_dev
*devlist
,
1945 char *backup_file
, int quiet
, int forked
,
1946 int restart
, int freeze_reshape
)
1948 struct reshape reshape
;
1951 int orig_level
= UnSet
;
1954 struct mdu_array_info_s array
;
1957 struct mddev_dev
*dv
;
1961 unsigned long long *offsets
= NULL
;
1965 unsigned long blocks
;
1966 unsigned long cache
;
1967 unsigned long long array_size
;
1969 struct mdinfo
*sra
= NULL
;
1971 /* when reshaping a RAID0, the component_size might be zero.
1972 * So try to fix that up.
1974 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
1975 dprintf("Cannot get array information.\n");
1978 if (array
.level
== 0 && info
->component_size
== 0) {
1979 get_dev_size(fd
, NULL
, &array_size
);
1980 info
->component_size
= array_size
/ array
.raid_disks
;
1983 if (info
->reshape_active
) {
1984 int new_level
= info
->new_level
;
1985 info
->new_level
= UnSet
;
1986 if (info
->delta_disks
> 0)
1987 info
->array
.raid_disks
-= info
->delta_disks
;
1988 msg
= analyse_change(info
, &reshape
);
1989 info
->new_level
= new_level
;
1990 if (info
->delta_disks
> 0)
1991 info
->array
.raid_disks
+= info
->delta_disks
;
1993 /* Make sure the array isn't read-only */
1994 ioctl(fd
, RESTART_ARRAY_RW
, 0);
1996 msg
= analyse_change(info
, &reshape
);
1998 fprintf(stderr
, Name
": %s\n", msg
);
2002 (reshape
.level
!= info
->array
.level
||
2003 reshape
.before
.layout
!= info
->array
.layout
||
2004 reshape
.before
.data_disks
+ reshape
.parity
2005 != info
->array
.raid_disks
- max(0, info
->delta_disks
))) {
2006 fprintf(stderr
, Name
": reshape info is not in native format -"
2007 " cannot continue.\n");
2011 if (st
->ss
->external
&& restart
&& (info
->reshape_progress
== 0)) {
2012 /* When reshape is restarted from '0', very begin of array
2013 * it is possible that for external metadata reshape and array
2014 * configuration doesn't happen.
2015 * Check if md has the same opinion, and reshape is restarted
2016 * from 0. If so, this is regular reshape start after reshape
2017 * switch in metadata to next array only.
2019 if ((verify_reshape_position(info
, reshape
.level
) >= 0) &&
2020 (info
->reshape_progress
== 0))
2024 /* reshape already started. just skip to monitoring the reshape */
2025 if (reshape
.backup_blocks
== 0)
2029 /* The container is frozen but the array may not be.
2030 * So freeze the array so spares don't get put to the wrong use
2031 * FIXME there should probably be a cleaner separation between
2032 * freeze_array and freeze_container.
2034 sysfs_freeze_array(info
);
2035 /* Check we have enough spares to not be degraded */
2037 for (dv
= devlist
; dv
; dv
=dv
->next
)
2039 spares_needed
= max(reshape
.before
.data_disks
,
2040 reshape
.after
.data_disks
)
2041 + reshape
.parity
- array
.raid_disks
;
2044 info
->new_level
> 1 && info
->array
.level
> 1 &&
2045 spares_needed
> info
->array
.spare_disks
+ added_disks
) {
2047 Name
": Need %d spare%s to avoid degraded array,"
2048 " and only have %d.\n"
2049 " Use --force to over-ride this check.\n",
2051 spares_needed
== 1 ? "" : "s",
2052 info
->array
.spare_disks
+ added_disks
);
2055 /* Check we have enough spares to not fail */
2056 spares_needed
= max(reshape
.before
.data_disks
,
2057 reshape
.after
.data_disks
)
2059 if ((info
->new_level
> 1 || info
->new_level
== 0) &&
2060 spares_needed
> info
->array
.spare_disks
+added_disks
) {
2062 Name
": Need %d spare%s to create working array,"
2063 " and only have %d.\n",
2065 spares_needed
== 1 ? "" : "s",
2066 info
->array
.spare_disks
+ added_disks
);
2070 if (reshape
.level
!= array
.level
) {
2071 char *c
= map_num(pers
, reshape
.level
);
2076 err
= sysfs_set_str(info
, NULL
, "level", c
);
2079 fprintf(stderr
, Name
": %s: could not set level to %s\n",
2082 (info
->array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
2083 fprintf(stderr
, " Bitmap must be removed"
2084 " before level can be changed\n");
2088 fprintf(stderr
, Name
": level of %s changed to %s\n",
2090 orig_level
= array
.level
;
2091 sysfs_freeze_array(info
);
2093 if (reshape
.level
> 0 && st
->ss
->external
) {
2094 /* make sure mdmon is aware of the new level */
2095 if (mdmon_running(st
->container_dev
))
2096 flush_mdmon(container
);
2098 if (!mdmon_running(st
->container_dev
))
2099 start_mdmon(st
->container_dev
);
2100 ping_monitor(container
);
2101 if (mdmon_running(st
->container_dev
) &&
2102 st
->update_tail
== NULL
)
2103 st
->update_tail
= &st
->updates
;
2106 /* ->reshape_super might have chosen some spares from the
2107 * container that it wants to be part of the new array.
2108 * We can collect them with ->container_content and give
2109 * them to the kernel.
2111 if (st
->ss
->reshape_super
&& st
->ss
->container_content
) {
2112 char *subarray
= strchr(info
->text_version
+1, '/')+1;
2113 struct mdinfo
*info2
=
2114 st
->ss
->container_content(st
, subarray
);
2118 sysfs_init(info2
, fd
, st
->devnum
);
2119 /* When increasing number of devices, we need to set
2120 * new raid_disks before adding these, or they might
2123 if (reshape
.backup_blocks
&&
2124 reshape
.after
.data_disks
> reshape
.before
.data_disks
)
2125 subarray_set_num(container
, info2
, "raid_disks",
2126 reshape
.after
.data_disks
+
2128 for (d
= info2
->devs
; d
; d
= d
->next
) {
2129 if (d
->disk
.state
== 0 &&
2130 d
->disk
.raid_disk
>= 0) {
2131 /* This is a spare that wants to
2132 * be part of the array.
2134 add_disk(fd
, st
, info2
, d
);
2140 /* We might have been given some devices to add to the
2141 * array. Now that the array has been changed to the right
2142 * level and frozen, we can safely add them.
2145 Manage_subdevs(devname
, fd
, devlist
, !quiet
,
2148 if (reshape
.backup_blocks
== 0) {
2149 /* No restriping needed, but we might need to impose
2150 * some more changes: layout, raid_disks, chunk_size
2152 /* read current array info */
2153 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
2154 dprintf("Cannot get array information.\n");
2157 /* compare current array info with new values and if
2158 * it is different update them to new */
2159 if (info
->new_layout
!= UnSet
&&
2160 info
->new_layout
!= array
.layout
) {
2161 array
.layout
= info
->new_layout
;
2162 if (ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
2163 fprintf(stderr
, Name
": failed to set new layout\n");
2166 printf("layout for %s set to %d\n",
2167 devname
, array
.layout
);
2169 if (info
->delta_disks
!= UnSet
&&
2170 info
->delta_disks
!= 0 &&
2171 array
.raid_disks
!= (info
->array
.raid_disks
+ info
->delta_disks
)) {
2172 array
.raid_disks
+= info
->delta_disks
;
2173 if (ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
2174 fprintf(stderr
, Name
": failed to set raid disks\n");
2176 } else if (!quiet
) {
2177 printf("raid_disks for %s set to %d\n",
2178 devname
, array
.raid_disks
);
2181 if (info
->new_chunk
!= 0 &&
2182 info
->new_chunk
!= array
.chunk_size
) {
2183 if (sysfs_set_num(info
, NULL
,
2184 "chunk_size", info
->new_chunk
) != 0) {
2185 fprintf(stderr
, Name
": failed to set chunk size\n");
2188 printf("chunk size for %s set to %d\n",
2189 devname
, array
.chunk_size
);
2196 * There are three possibilities.
2197 * 1/ The array will shrink.
2198 * We need to ensure the reshape will pause before reaching
2199 * the 'critical section'. We also need to fork and wait for
2200 * that to happen. When it does we
2201 * suspend/backup/complete/unfreeze
2203 * 2/ The array will not change size.
2204 * This requires that we keep a backup of a sliding window
2205 * so that we can restore data after a crash. So we need
2206 * to fork and monitor progress.
2207 * In future we will allow the data_offset to change, so
2208 * a sliding backup becomes unnecessary.
2210 * 3/ The array will grow. This is relatively easy.
2211 * However the kernel's restripe routines will cheerfully
2212 * overwrite some early data before it is safe. So we
2213 * need to make a backup of the early parts of the array
2214 * and be ready to restore it if rebuild aborts very early.
2215 * For externally managed metadata, we still need a forked
2216 * child to monitor the reshape and suspend IO over the region
2217 * that is being reshaped.
2219 * We backup data by writing it to one spare, or to a
2220 * file which was given on command line.
2222 * In each case, we first make sure that storage is available
2223 * for the required backup.
2225 * - request the shape change.
2226 * - fork to handle backup etc.
2229 /* Check that we can hold all the data */
2230 get_dev_size(fd
, NULL
, &array_size
);
2231 if (reshape
.new_size
< (array_size
/512)) {
2233 Name
": this change will reduce the size of the array.\n"
2234 " use --grow --array-size first to truncate array.\n"
2235 " e.g. mdadm --grow %s --array-size %llu\n",
2236 devname
, reshape
.new_size
/2);
2240 sra
= sysfs_read(fd
, 0,
2241 GET_COMPONENT
|GET_DEVS
|GET_OFFSET
|GET_STATE
|GET_CHUNK
|
2244 fprintf(stderr
, Name
": %s: Cannot get array details from sysfs\n",
2249 /* Decide how many blocks (sectors) for a reshape
2250 * unit. The number we have so far is just a minimum
2252 blocks
= reshape
.backup_blocks
;
2253 if (reshape
.before
.data_disks
==
2254 reshape
.after
.data_disks
) {
2255 /* Make 'blocks' bigger for better throughput, but
2256 * not so big that we reject it below.
2257 * Try for 16 megabytes
2259 while (blocks
* 32 < sra
->component_size
&&
2263 fprintf(stderr
, Name
": Need to backup %luK of critical "
2264 "section..\n", blocks
/2);
2266 if (blocks
>= sra
->component_size
/2) {
2267 fprintf(stderr
, Name
": %s: Something wrong"
2268 " - reshape aborted\n",
2273 /* Now we need to open all these devices so we can read/write.
2275 nrdisks
= max(reshape
.before
.data_disks
,
2276 reshape
.after
.data_disks
) + reshape
.parity
2277 + sra
->array
.spare_disks
;
2278 fdlist
= malloc((1+nrdisks
) * sizeof(int));
2279 offsets
= malloc((1+nrdisks
) * sizeof(offsets
[0]));
2280 if (!fdlist
|| !offsets
) {
2281 fprintf(stderr
, Name
": malloc failed: grow aborted\n");
2285 odisks
= reshape
.before
.data_disks
+ reshape
.parity
;
2286 d
= reshape_prepare_fdlist(devname
, sra
, odisks
,
2287 nrdisks
, blocks
, backup_file
,
2292 if ((st
->ss
->manage_reshape
== NULL
) ||
2293 (st
->ss
->recover_backup
== NULL
)) {
2294 if (backup_file
== NULL
) {
2295 if (reshape
.after
.data_disks
<=
2296 reshape
.before
.data_disks
) {
2297 fprintf(stderr
, Name
": %s: Cannot grow - "
2298 "need backup-file\n", devname
);
2300 } else if (sra
->array
.spare_disks
== 0) {
2301 fprintf(stderr
, Name
": %s: Cannot grow - "
2302 "need a spare or backup-file to backup "
2303 "critical section\n", devname
);
2307 if (!reshape_open_backup_file(backup_file
, fd
, devname
,
2309 fdlist
+d
, offsets
+d
,
2317 /* lastly, check that the internal stripe cache is
2318 * large enough, or it won't work.
2319 * It must hold at least 4 stripes of the larger
2322 cache
= max(info
->array
.chunk_size
, info
->new_chunk
);
2323 cache
*= 4; /* 4 stripes minimum */
2324 cache
/= 512; /* convert to sectors */
2325 disks
= min(reshape
.before
.data_disks
, reshape
.after
.data_disks
);
2326 /* make sure there is room for 'blocks' with a bit to spare */
2327 if (cache
< 16 + blocks
/ disks
)
2328 cache
= 16 + blocks
/ disks
;
2329 cache
/= (4096/512); /* Covert from sectors to pages */
2331 if (sra
->cache_size
< cache
)
2332 subarray_set_num(container
, sra
, "stripe_cache_size",
2335 /* Right, everything seems fine. Let's kick things off.
2336 * If only changing raid_disks, use ioctl, else use
2341 sra
->new_chunk
= info
->new_chunk
;
2344 /* for external metadata checkpoint saved by mdmon can be lost
2345 * or missed /due to e.g. crash/. Check if md is not during
2346 * restart farther than metadata points to.
2347 * If so, this means metadata information is obsolete.
2349 if (st
->ss
->external
)
2350 verify_reshape_position(info
, reshape
.level
);
2351 sra
->reshape_progress
= info
->reshape_progress
;
2353 sra
->reshape_progress
= 0;
2354 if (reshape
.after
.data_disks
< reshape
.before
.data_disks
)
2355 /* start from the end of the new array */
2356 sra
->reshape_progress
= (sra
->component_size
2357 * reshape
.after
.data_disks
);
2360 if (info
->array
.chunk_size
== info
->new_chunk
&&
2361 reshape
.before
.layout
== reshape
.after
.layout
&&
2362 st
->ss
->external
== 0) {
2363 /* use SET_ARRAY_INFO but only if reshape hasn't started */
2364 ioctl(fd
, GET_ARRAY_INFO
, &array
);
2365 array
.raid_disks
= reshape
.after
.data_disks
+ reshape
.parity
;
2367 ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
2371 Name
": Cannot set device shape for %s: %s\n",
2372 devname
, strerror(errno
));
2375 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
2377 " Bitmap must be removed before"
2378 " shape can be changed\n");
2382 } else if (!restart
) {
2383 /* set them all just in case some old 'new_*' value
2384 * persists from some earlier problem.
2387 if (sysfs_set_num(sra
, NULL
, "chunk_size", info
->new_chunk
) < 0)
2389 if (!err
&& sysfs_set_num(sra
, NULL
, "layout",
2390 reshape
.after
.layout
) < 0)
2392 if (!err
&& subarray_set_num(container
, sra
, "raid_disks",
2393 reshape
.after
.data_disks
+
2394 reshape
.parity
) < 0)
2397 fprintf(stderr
, Name
": Cannot set device shape for %s\n",
2401 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
2403 " Bitmap must be removed before"
2404 " shape can be changed\n");
2409 err
= start_reshape(sra
, restart
, reshape
.before
.data_disks
,
2410 reshape
.after
.data_disks
);
2413 Name
": Cannot %s reshape for %s\n",
2414 restart
? "continue" : "start",
2419 sysfs_set_str(sra
, NULL
, "array_state", "active");
2420 if (freeze_reshape
) {
2424 fprintf(stderr
, Name
": Reshape has to be continued from"
2425 " location %llu when root filesystem has been mounted.\n",
2426 sra
->reshape_progress
);
2430 /* Now we just need to kick off the reshape and watch, while
2431 * handling backups of the data...
2432 * This is all done by a forked background process.
2434 switch(forked
? 0 : fork()) {
2436 fprintf(stderr
, Name
": Cannot run child to monitor reshape: %s\n",
2451 if (check_env("MDADM_GROW_VERIFY"))
2452 fd
= open(devname
, O_RDONLY
| O_DIRECT
);
2455 mlockall(MCL_FUTURE
);
2457 if (st
->ss
->external
) {
2458 /* metadata handler takes it from here */
2459 done
= st
->ss
->manage_reshape(
2460 fd
, sra
, &reshape
, st
, blocks
,
2462 d
- odisks
, fdlist
+odisks
,
2465 done
= child_monitor(
2466 fd
, sra
, &reshape
, st
, blocks
,
2468 d
- odisks
, fdlist
+odisks
,
2474 if (backup_file
&& done
)
2475 unlink(backup_file
);
2481 if (!st
->ss
->external
&&
2482 !(reshape
.before
.data_disks
!= reshape
.after
.data_disks
2483 && info
->custom_array_size
) &&
2484 info
->new_level
== reshape
.level
&&
2486 /* no need to wait for the reshape to finish as
2487 * there is nothing more to do.
2494 if (st
->ss
->external
) {
2495 /* Re-load the metadata as much could have changed */
2496 int cfd
= open_dev(st
->container_dev
);
2498 flush_mdmon(container
);
2499 st
->ss
->free_super(st
);
2500 st
->ss
->load_container(st
, cfd
, container
);
2505 /* set new array size if required customer_array_size is used
2508 if (reshape
.before
.data_disks
!=
2509 reshape
.after
.data_disks
&&
2510 info
->custom_array_size
) {
2511 struct mdinfo
*info2
;
2512 char *subarray
= strchr(info
->text_version
+1, '/')+1;
2514 info2
= st
->ss
->container_content(st
, subarray
);
2516 unsigned long long current_size
= 0;
2517 unsigned long long new_size
=
2518 info2
->custom_array_size
/2;
2520 if (sysfs_get_ll(sra
,
2523 ¤t_size
) == 0 &&
2524 new_size
> current_size
) {
2525 if (sysfs_set_num(sra
, NULL
,
2526 "array_size", new_size
)
2528 dprintf("Error: Cannot"
2531 dprintf("Array size "
2533 dprintf(" from %llu to %llu.\n",
2534 current_size
, new_size
);
2540 if (info
->new_level
!= reshape
.level
) {
2542 c
= map_num(pers
, info
->new_level
);
2544 err
= sysfs_set_str(sra
, NULL
, "level", c
);
2546 fprintf(stderr
, Name\
2547 ": %s: could not set level "
2548 "to %s\n", devname
, c
);
2550 if (info
->new_level
== 0)
2551 st
->update_tail
= NULL
;
2563 if (orig_level
!= UnSet
&& sra
) {
2564 c
= map_num(pers
, orig_level
);
2565 if (c
&& sysfs_set_str(sra
, NULL
, "level", c
) == 0)
2566 fprintf(stderr
, Name
": aborting level change\n");
2574 /* mdfd handle is passed to be closed in child process (after fork).
2576 int reshape_container(char *container
, char *devname
,
2578 struct supertype
*st
,
2579 struct mdinfo
*info
,
2582 int quiet
, int restart
, int freeze_reshape
)
2584 struct mdinfo
*cc
= NULL
;
2586 int last_devnum
= -1;
2588 /* component_size is not meaningful for a container,
2589 * so pass '-1' meaning 'no change'
2592 reshape_super(st
, -1, info
->new_level
,
2593 info
->new_layout
, info
->new_chunk
,
2594 info
->array
.raid_disks
, info
->delta_disks
,
2595 backup_file
, devname
, quiet
)) {
2602 /* ping monitor to be sure that update is on disk
2604 ping_monitor(container
);
2607 case -1: /* error */
2608 perror("Cannot fork to complete reshape\n");
2611 default: /* parent */
2612 if (!freeze_reshape
)
2613 printf(Name
": multi-array reshape continues"
2614 " in background\n");
2621 /* close unused handle in child process
2627 /* For each member array with reshape_active,
2628 * we need to perform the reshape.
2629 * We pick the first array that needs reshaping and
2630 * reshape it. reshape_array() will re-read the metadata
2631 * so the next time through a different array should be
2632 * ready for reshape.
2633 * It is possible that the 'different' array will not
2634 * be assembled yet. In that case we simple exit.
2635 * When it is assembled, the mdadm which assembles it
2636 * will take over the reshape.
2638 struct mdinfo
*content
;
2640 struct mdstat_ent
*mdstat
;
2645 cc
= st
->ss
->container_content(st
, NULL
);
2647 for (content
= cc
; content
; content
= content
->next
) {
2649 if (!content
->reshape_active
)
2652 subarray
= strchr(content
->text_version
+1, '/')+1;
2653 mdstat
= mdstat_by_subdev(subarray
,
2654 devname2devnum(container
));
2657 if (mdstat
->active
== 0) {
2658 fprintf(stderr
, Name
": Skipping inactive "
2659 "array md%i.\n", mdstat
->devnum
);
2660 free_mdstat(mdstat
);
2669 adev
= map_dev(dev2major(mdstat
->devnum
),
2670 dev2minor(mdstat
->devnum
),
2673 adev
= content
->text_version
;
2675 fd
= open_dev(mdstat
->devnum
);
2677 printf(Name
": Device %s cannot be opened for reshape.",
2682 if (last_devnum
== mdstat
->devnum
) {
2683 /* Do not allow for multiple reshape_array() calls for
2685 * It can happen when reshape_array() returns without
2686 * error, when reshape is not finished (wrong reshape
2687 * starting/continuation conditions). Mdmon doesn't
2688 * switch to next array in container and reentry
2689 * conditions for the same array occur.
2690 * This is possibly interim until the behaviour of
2691 * reshape_array is resolved().
2693 printf(Name
": Multiple reshape execution detected for "
2694 "device %s.", adev
);
2698 last_devnum
= mdstat
->devnum
;
2700 sysfs_init(content
, fd
, mdstat
->devnum
);
2702 if (mdmon_running(devname2devnum(container
)))
2703 flush_mdmon(container
);
2705 rv
= reshape_array(container
, fd
, adev
, st
,
2706 content
, force
, NULL
,
2707 backup_file
, quiet
, 1, restart
,
2711 if (freeze_reshape
) {
2720 if (mdmon_running(devname2devnum(container
)))
2721 flush_mdmon(container
);
2730 * We run a child process in the background which performs the following
2732 * - wait for resync to reach a certain point
2733 * - suspend io to the following section
2734 * - backup that section
2735 * - allow resync to proceed further
2737 * - discard the backup.
2739 * When are combined in slightly different ways in the three cases.
2741 * - suspend/backup/allow/wait/resume/discard
2743 * - allow/wait/suspend/backup/allow/wait/resume/discard
2745 * - wait/resume/discard/suspend/backup/allow
2747 * suspend/backup/allow always come together
2748 * wait/resume/discard do too.
2749 * For the same-size case we have two backups to improve flow.
2753 int progress_reshape(struct mdinfo
*info
, struct reshape
*reshape
,
2754 unsigned long long backup_point
,
2755 unsigned long long wait_point
,
2756 unsigned long long *suspend_point
,
2757 unsigned long long *reshape_completed
)
2759 /* This function is called repeatedly by the reshape manager.
2760 * It determines how much progress can safely be made and allows
2762 * - 'info' identifies the array and particularly records in
2763 * ->reshape_progress the metadata's knowledge of progress
2764 * This is a sector offset from the start of the array
2765 * of the next array block to be relocated. This number
2766 * may increase from 0 or decrease from array_size, depending
2767 * on the type of reshape that is happening.
2768 * Note that in contrast, 'sync_completed' is a block count of the
2769 * reshape so far. It gives the distance between the start point
2770 * (head or tail of device) and the next place that data will be
2771 * written. It always increases.
2772 * - 'reshape' is the structure created by analyse_change
2773 * - 'backup_point' shows how much the metadata manager has backed-up
2774 * data. For reshapes with increasing progress, it is the next address
2775 * to be backed up, previous addresses have been backed-up. For
2776 * decreasing progress, it is the earliest address that has been
2777 * backed up - later address are also backed up.
2778 * So addresses between reshape_progress and backup_point are
2779 * backed up providing those are in the 'correct' order.
2780 * - 'wait_point' is an array address. When reshape_completed
2781 * passes this point, progress_reshape should return. It might
2782 * return earlier if it determines that ->reshape_progress needs
2783 * to be updated or further backup is needed.
2784 * - suspend_point is maintained by progress_reshape and the caller
2785 * should not touch it except to initialise to zero.
2786 * It is an array address and it only increases in 2.6.37 and earlier.
2787 * This makes it difficult to handle reducing reshapes with
2788 * external metadata.
2789 * However: it is similar to backup_point in that it records the
2790 * other end of a suspended region from reshape_progress.
2791 * it is moved to extend the region that is safe to backup and/or
2793 * - reshape_completed is read from sysfs and returned. The caller
2794 * should copy this into ->reshape_progress when it has reason to
2795 * believe that the metadata knows this, and any backup outside this
2799 * 1 if more data from backup_point - but only as far as suspend_point,
2800 * should be backed up
2801 * 0 if things are progressing smoothly
2802 * -1 if the reshape is finished because it is all done,
2803 * -2 if the reshape is finished due to an error.
2806 int advancing
= (reshape
->after
.data_disks
2807 >= reshape
->before
.data_disks
);
2808 unsigned long long need_backup
; /* All data between start of array and
2809 * here will at some point need to
2812 unsigned long long read_offset
, write_offset
;
2813 unsigned long long write_range
;
2814 unsigned long long max_progress
, target
, completed
;
2815 unsigned long long array_size
= (info
->component_size
2816 * reshape
->before
.data_disks
);
2820 /* First, we unsuspend any region that is now known to be safe.
2821 * If suspend_point is on the 'wrong' side of reshape_progress, then
2822 * we don't have or need suspension at the moment. This is true for
2823 * native metadata when we don't need to back-up.
2826 if (info
->reshape_progress
<= *suspend_point
)
2827 sysfs_set_num(info
, NULL
, "suspend_lo",
2828 info
->reshape_progress
);
2830 /* Note: this won't work in 2.6.37 and before.
2831 * Something somewhere should make sure we don't need it!
2833 if (info
->reshape_progress
>= *suspend_point
)
2834 sysfs_set_num(info
, NULL
, "suspend_hi",
2835 info
->reshape_progress
);
2838 /* Now work out how far it is safe to progress.
2839 * If the read_offset for ->reshape_progress is less than
2840 * 'blocks' beyond the write_offset, we can only progress as far
2842 * Otherwise we can progress until the write_offset for the new location
2843 * reaches (within 'blocks' of) the read_offset at the current location.
2844 * However that region must be suspended unless we are using native
2846 * If we need to suspend more, we limit it to 128M per device, which is
2847 * rather arbitrary and should be some time-based calculation.
2849 read_offset
= info
->reshape_progress
/ reshape
->before
.data_disks
;
2850 write_offset
= info
->reshape_progress
/ reshape
->after
.data_disks
;
2851 write_range
= info
->new_chunk
/512;
2852 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
)
2853 need_backup
= array_size
;
2855 need_backup
= reshape
->backup_blocks
;
2857 if (read_offset
< write_offset
+ write_range
)
2858 max_progress
= backup_point
;
2862 reshape
->after
.data_disks
;
2864 if (read_offset
> write_offset
- write_range
)
2865 /* Can only progress as far as has been backed up,
2866 * which must be suspended */
2867 max_progress
= backup_point
;
2868 else if (info
->reshape_progress
<= need_backup
)
2869 max_progress
= backup_point
;
2871 if (info
->array
.major_version
>= 0)
2872 /* Can progress until backup is needed */
2873 max_progress
= need_backup
;
2875 /* Can progress until metadata update is required */
2878 reshape
->after
.data_disks
;
2879 /* but data must be suspended */
2880 if (max_progress
< *suspend_point
)
2881 max_progress
= *suspend_point
;
2886 /* We know it is safe to progress to 'max_progress' providing
2887 * it is suspended or we are using native metadata.
2888 * Consider extending suspend_point 128M per device if it
2889 * is less than 64M per device beyond reshape_progress.
2890 * But always do a multiple of 'blocks'
2891 * FIXME this is too big - it takes to long to complete
2894 target
= 64*1024*2 * min(reshape
->before
.data_disks
,
2895 reshape
->after
.data_disks
);
2896 target
/= reshape
->backup_blocks
;
2899 target
*= reshape
->backup_blocks
;
2901 /* For externally managed metadata we always need to suspend IO to
2902 * the area being reshaped so we regularly push suspend_point forward.
2903 * For native metadata we only need the suspend if we are going to do
2907 if ((need_backup
> info
->reshape_progress
2908 || info
->array
.major_version
< 0) &&
2909 *suspend_point
< info
->reshape_progress
+ target
) {
2910 if (need_backup
< *suspend_point
+ 2 * target
)
2911 *suspend_point
= need_backup
;
2912 else if (*suspend_point
+ 2 * target
< array_size
)
2913 *suspend_point
+= 2 * target
;
2915 *suspend_point
= array_size
;
2916 sysfs_set_num(info
, NULL
, "suspend_hi", *suspend_point
);
2917 if (max_progress
> *suspend_point
)
2918 max_progress
= *suspend_point
;
2921 if (info
->array
.major_version
>= 0) {
2922 /* Only need to suspend when about to backup */
2923 if (info
->reshape_progress
< need_backup
* 2 &&
2924 *suspend_point
> 0) {
2926 sysfs_set_num(info
, NULL
, "suspend_lo", 0);
2927 sysfs_set_num(info
, NULL
, "suspend_hi", need_backup
);
2930 /* Need to suspend continually */
2931 if (info
->reshape_progress
< *suspend_point
)
2932 *suspend_point
= info
->reshape_progress
;
2933 if (*suspend_point
+ target
< info
->reshape_progress
)
2934 /* No need to move suspend region yet */;
2936 if (*suspend_point
>= 2 * target
)
2937 *suspend_point
-= 2 * target
;
2940 sysfs_set_num(info
, NULL
, "suspend_lo",
2943 if (max_progress
< *suspend_point
)
2944 max_progress
= *suspend_point
;
2948 /* now set sync_max to allow that progress. sync_max, like
2949 * sync_completed is a count of sectors written per device, so
2950 * we find the difference between max_progress and the start point,
2951 * and divide that by after.data_disks to get a sync_max
2953 * At the same time we convert wait_point to a similar number
2954 * for comparing against sync_completed.
2956 /* scale down max_progress to per_disk */
2957 max_progress
/= reshape
->after
.data_disks
;
2958 /* Round to chunk size as some kernels give an erroneously high number */
2959 max_progress
/= info
->new_chunk
/512;
2960 max_progress
*= info
->new_chunk
/512;
2961 /* And round to old chunk size as the kernel wants that */
2962 max_progress
/= info
->array
.chunk_size
/512;
2963 max_progress
*= info
->array
.chunk_size
/512;
2964 /* Limit progress to the whole device */
2965 if (max_progress
> info
->component_size
)
2966 max_progress
= info
->component_size
;
2967 wait_point
/= reshape
->after
.data_disks
;
2969 /* switch from 'device offset' to 'processed block count' */
2970 max_progress
= info
->component_size
- max_progress
;
2971 wait_point
= info
->component_size
- wait_point
;
2974 sysfs_set_num(info
, NULL
, "sync_max", max_progress
);
2976 /* Now wait. If we have already reached the point that we were
2977 * asked to wait to, don't wait at all, else wait for any change.
2978 * We need to select on 'sync_completed' as that is the place that
2979 * notifications happen, but we are really interested in
2980 * 'reshape_position'
2982 fd
= sysfs_get_fd(info
, NULL
, "sync_completed");
2984 goto check_progress
;
2986 if (sysfs_fd_get_ll(fd
, &completed
) < 0)
2987 goto check_progress
;
2989 while (completed
< max_progress
&& completed
< wait_point
) {
2990 /* Check that sync_action is still 'reshape' to avoid
2991 * waiting forever on a dead array
2995 if (sysfs_get_str(info
, NULL
, "sync_action",
2997 strncmp(action
, "reshape", 7) != 0)
2999 /* Some kernels reset 'sync_completed' to zero
3000 * before setting 'sync_action' to 'idle'.
3001 * So we need these extra tests.
3003 if (completed
== 0 && advancing
3004 && info
->reshape_progress
> 0)
3006 if (completed
== 0 && !advancing
3007 && info
->reshape_progress
< (info
->component_size
3008 * reshape
->after
.data_disks
))
3012 select(fd
+1, NULL
, NULL
, &rfds
, NULL
);
3013 if (sysfs_fd_get_ll(fd
, &completed
) < 0)
3014 goto check_progress
;
3016 /* Some kernels reset 'sync_completed' to zero,
3017 * we need to have real point we are in md
3020 completed
= max_progress
;
3022 /* some kernels can give an incorrectly high 'completed' number */
3023 completed
/= (info
->new_chunk
/512);
3024 completed
*= (info
->new_chunk
/512);
3025 /* Convert 'completed' back in to a 'progress' number */
3026 completed
*= reshape
->after
.data_disks
;
3028 completed
= info
->component_size
* reshape
->after
.data_disks
3031 *reshape_completed
= completed
;
3035 /* We return the need_backup flag. Caller will decide
3036 * how much - a multiple of ->backup_blocks up to *suspend_point
3039 return need_backup
> info
->reshape_progress
;
3041 return need_backup
>= info
->reshape_progress
;
3044 /* if we couldn't read a number from sync_completed, then
3045 * either the reshape did complete, or it aborted.
3046 * We can tell which by checking for 'none' in reshape_position.
3047 * If it did abort, then it might immediately restart if it
3048 * it was just a device failure that leaves us degraded but
3052 if (sysfs_get_str(info
, NULL
, "reshape_position", buf
, sizeof(buf
)) < 0
3053 || strncmp(buf
, "none", 4) != 0) {
3054 /* The abort might only be temporary. Wait up to 10
3055 * seconds for fd to contain a valid number again.
3061 while (fd
>= 0 && rv
< 0 && tv
.tv_sec
> 0) {
3065 if (select(fd
+1, NULL
, NULL
, &rfds
, &tv
) != 1)
3067 switch (sysfs_fd_get_ll(fd
, &completed
)) {
3069 /* all good again */
3072 case -2: /* read error - abort */
3079 return rv
; /* abort */
3081 /* Maybe racing with array shutdown - check state */
3084 if (sysfs_get_str(info
, NULL
, "array_state", buf
, sizeof(buf
)) < 0
3085 || strncmp(buf
, "inactive", 8) == 0
3086 || strncmp(buf
, "clear",5) == 0)
3087 return -2; /* abort */
3088 return -1; /* complete */
3092 /* FIXME return status is never checked */
3093 static int grow_backup(struct mdinfo
*sra
,
3094 unsigned long long offset
, /* per device */
3095 unsigned long stripes
, /* per device, in old chunks */
3096 int *sources
, unsigned long long *offsets
,
3097 int disks
, int chunk
, int level
, int layout
,
3098 int dests
, int *destfd
, unsigned long long *destoffsets
,
3099 int part
, int *degraded
,
3102 /* Backup 'blocks' sectors at 'offset' on each device of the array,
3103 * to storage 'destfd' (offset 'destoffsets'), after first
3104 * suspending IO. Then allow resync to continue
3105 * over the suspended section.
3106 * Use part 'part' of the backup-super-block.
3111 unsigned long long ll
;
3113 //printf("offset %llu\n", offset);
3119 /* Check that array hasn't become degraded, else we might backup the wrong data */
3120 if (sysfs_get_ll(sra
, NULL
, "degraded", &ll
) < 0)
3121 return -1; /* FIXME this error is ignored */
3122 new_degraded
= (int)ll
;
3123 if (new_degraded
!= *degraded
) {
3124 /* check each device to ensure it is still working */
3126 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
3127 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
3129 if (sd
->disk
.state
& (1<<MD_DISK_SYNC
)) {
3131 if (sysfs_get_str(sra
, sd
, "state", sbuf
, 20) < 0 ||
3132 strstr(sbuf
, "faulty") ||
3133 strstr(sbuf
, "in_sync") == NULL
) {
3134 /* this device is dead */
3135 sd
->disk
.state
= (1<<MD_DISK_FAULTY
);
3136 if (sd
->disk
.raid_disk
>= 0 &&
3137 sources
[sd
->disk
.raid_disk
] >= 0) {
3138 close(sources
[sd
->disk
.raid_disk
]);
3139 sources
[sd
->disk
.raid_disk
] = -1;
3144 *degraded
= new_degraded
;
3147 bsb
.arraystart2
= __cpu_to_le64(offset
* odata
);
3148 bsb
.length2
= __cpu_to_le64(stripes
* (chunk
/512) * odata
);
3150 bsb
.arraystart
= __cpu_to_le64(offset
* odata
);
3151 bsb
.length
= __cpu_to_le64(stripes
* (chunk
/512) * odata
);
3154 bsb
.magic
[15] = '2';
3155 for (i
= 0; i
< dests
; i
++)
3157 lseek64(destfd
[i
], destoffsets
[i
] + __le64_to_cpu(bsb
.devstart2
)*512, 0);
3159 lseek64(destfd
[i
], destoffsets
[i
], 0);
3161 rv
= save_stripes(sources
, offsets
,
3162 disks
, chunk
, level
, layout
,
3164 offset
*512*odata
, stripes
* chunk
* odata
,
3169 bsb
.mtime
= __cpu_to_le64(time(0));
3170 for (i
= 0; i
< dests
; i
++) {
3171 bsb
.devstart
= __cpu_to_le64(destoffsets
[i
]/512);
3173 bsb
.sb_csum
= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
));
3174 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0)
3175 bsb
.sb_csum2
= bsb_csum((char*)&bsb
,
3176 ((char*)&bsb
.sb_csum2
)-((char*)&bsb
));
3179 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
] - 4096, 0)
3180 != destoffsets
[i
] - 4096)
3182 if (write(destfd
[i
], &bsb
, 512) != 512)
3184 if (destoffsets
[i
] > 4096) {
3185 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
]+stripes
*chunk
*odata
, 0) !=
3186 destoffsets
[i
]+stripes
*chunk
*odata
)
3188 if (write(destfd
[i
], &bsb
, 512) != 512)
3198 /* in 2.6.30, the value reported by sync_completed can be
3199 * less that it should be by one stripe.
3200 * This only happens when reshape hits sync_max and pauses.
3201 * So allow wait_backup to either extent sync_max further
3202 * than strictly necessary, or return before the
3203 * sync has got quite as far as we would really like.
3204 * This is what 'blocks2' is for.
3205 * The various caller give appropriate values so that
3208 /* FIXME return value is often ignored */
3209 static int forget_backup(int dests
, int *destfd
,
3210 unsigned long long *destoffsets
,
3214 * Erase backup 'part' (which is 0 or 1)
3220 bsb
.arraystart2
= __cpu_to_le64(0);
3221 bsb
.length2
= __cpu_to_le64(0);
3223 bsb
.arraystart
= __cpu_to_le64(0);
3224 bsb
.length
= __cpu_to_le64(0);
3226 bsb
.mtime
= __cpu_to_le64(time(0));
3228 for (i
= 0; i
< dests
; i
++) {
3229 bsb
.devstart
= __cpu_to_le64(destoffsets
[i
]/512);
3230 bsb
.sb_csum
= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
));
3231 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0)
3232 bsb
.sb_csum2
= bsb_csum((char*)&bsb
,
3233 ((char*)&bsb
.sb_csum2
)-((char*)&bsb
));
3234 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
]-4096, 0) !=
3235 destoffsets
[i
]-4096)
3238 write(destfd
[i
], &bsb
, 512) != 512)
3245 static void fail(char *msg
)
3248 rv
= (write(2, msg
, strlen(msg
)) != (int)strlen(msg
));
3249 rv
|= (write(2, "\n", 1) != 1);
3253 static char *abuf
, *bbuf
;
3254 static unsigned long long abuflen
;
3255 static void validate(int afd
, int bfd
, unsigned long long offset
)
3257 /* check that the data in the backup against the array.
3258 * This is only used for regression testing and should not
3259 * be used while the array is active
3263 lseek64(bfd
, offset
- 4096, 0);
3264 if (read(bfd
, &bsb2
, 512) != 512)
3265 fail("cannot read bsb");
3266 if (bsb2
.sb_csum
!= bsb_csum((char*)&bsb2
,
3267 ((char*)&bsb2
.sb_csum
)-((char*)&bsb2
)))
3268 fail("first csum bad");
3269 if (memcmp(bsb2
.magic
, "md_backup_data", 14) != 0)
3270 fail("magic is bad");
3271 if (memcmp(bsb2
.magic
, "md_backup_data-2", 16) == 0 &&
3272 bsb2
.sb_csum2
!= bsb_csum((char*)&bsb2
,
3273 ((char*)&bsb2
.sb_csum2
)-((char*)&bsb2
)))
3274 fail("second csum bad");
3276 if (__le64_to_cpu(bsb2
.devstart
)*512 != offset
)
3277 fail("devstart is wrong");
3280 unsigned long long len
= __le64_to_cpu(bsb2
.length
)*512;
3282 if (abuflen
< len
) {
3286 if (posix_memalign((void**)&abuf
, 4096, abuflen
) ||
3287 posix_memalign((void**)&bbuf
, 4096, abuflen
)) {
3289 /* just stop validating on mem-alloc failure */
3294 lseek64(bfd
, offset
, 0);
3295 if ((unsigned long long)read(bfd
, bbuf
, len
) != len
) {
3296 //printf("len %llu\n", len);
3297 fail("read first backup failed");
3299 lseek64(afd
, __le64_to_cpu(bsb2
.arraystart
)*512, 0);
3300 if ((unsigned long long)read(afd
, abuf
, len
) != len
)
3301 fail("read first from array failed");
3302 if (memcmp(bbuf
, abuf
, len
) != 0) {
3305 printf("offset=%llu len=%llu\n",
3306 (unsigned long long)__le64_to_cpu(bsb2
.arraystart
)*512, len
);
3307 for (i
=0; i
<len
; i
++)
3308 if (bbuf
[i
] != abuf
[i
]) {
3309 printf("first diff byte %d\n", i
);
3313 fail("data1 compare failed");
3317 unsigned long long len
= __le64_to_cpu(bsb2
.length2
)*512;
3319 if (abuflen
< len
) {
3323 abuf
= malloc(abuflen
);
3324 bbuf
= malloc(abuflen
);
3327 lseek64(bfd
, offset
+__le64_to_cpu(bsb2
.devstart2
)*512, 0);
3328 if ((unsigned long long)read(bfd
, bbuf
, len
) != len
)
3329 fail("read second backup failed");
3330 lseek64(afd
, __le64_to_cpu(bsb2
.arraystart2
)*512, 0);
3331 if ((unsigned long long)read(afd
, abuf
, len
) != len
)
3332 fail("read second from array failed");
3333 if (memcmp(bbuf
, abuf
, len
) != 0)
3334 fail("data2 compare failed");
3338 int child_monitor(int afd
, struct mdinfo
*sra
, struct reshape
*reshape
,
3339 struct supertype
*st
, unsigned long blocks
,
3340 int *fds
, unsigned long long *offsets
,
3341 int dests
, int *destfd
, unsigned long long *destoffsets
)
3343 /* Monitor a reshape where backup is being performed using
3344 * 'native' mechanism - either to a backup file, or
3345 * to some space in a spare.
3349 unsigned long long speed
;
3350 unsigned long long suspend_point
, array_size
;
3351 unsigned long long backup_point
, wait_point
;
3352 unsigned long long reshape_completed
;
3354 int increasing
= reshape
->after
.data_disks
>= reshape
->before
.data_disks
;
3355 int part
= 0; /* The next part of the backup area to fill. It may already
3356 * be full, so we need to check */
3357 int level
= reshape
->level
;
3358 int layout
= reshape
->before
.layout
;
3359 int data
= reshape
->before
.data_disks
;
3360 int disks
= reshape
->before
.data_disks
+ reshape
->parity
;
3361 int chunk
= sra
->array
.chunk_size
;
3363 unsigned long stripes
;
3366 /* set up the backup-super-block. This requires the
3367 * uuid from the array.
3369 /* Find a superblock */
3370 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
3374 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
3376 dn
= map_dev(sd
->disk
.major
, sd
->disk
.minor
, 1);
3377 devfd
= dev_open(dn
, O_RDONLY
);
3380 ok
= st
->ss
->load_super(st
, devfd
, NULL
);
3386 fprintf(stderr
, Name
": Cannot find a superblock\n");
3390 memset(&bsb
, 0, 512);
3391 memcpy(bsb
.magic
, "md_backup_data-1", 16);
3392 st
->ss
->uuid_from_super(st
, uuid
);
3393 memcpy(bsb
.set_uuid
, uuid
, 16);
3394 bsb
.mtime
= __cpu_to_le64(time(0));
3395 bsb
.devstart2
= blocks
;
3397 stripes
= blocks
/ (sra
->array
.chunk_size
/512) /
3398 reshape
->before
.data_disks
;
3400 if (posix_memalign((void**)&buf
, 4096, disks
* chunk
))
3401 /* Don't start the 'reshape' */
3403 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
) {
3404 sysfs_get_ll(sra
, NULL
, "sync_speed_min", &speed
);
3405 sysfs_set_num(sra
, NULL
, "sync_speed_min", 200000);
3409 array_size
= sra
->component_size
* reshape
->after
.data_disks
;
3410 backup_point
= sra
->reshape_progress
;
3413 array_size
= sra
->component_size
* reshape
->before
.data_disks
;
3414 backup_point
= reshape
->backup_blocks
;
3415 suspend_point
= array_size
;
3421 /* Want to return as soon the oldest backup slot can
3422 * be released as that allows us to start backing up
3423 * some more, providing suspend_point has been
3424 * advanced, which it should have.
3427 wait_point
= array_size
;
3428 if (part
== 0 && __le64_to_cpu(bsb
.length
) > 0)
3429 wait_point
= (__le64_to_cpu(bsb
.arraystart
) +
3430 __le64_to_cpu(bsb
.length
));
3431 if (part
== 1 && __le64_to_cpu(bsb
.length2
) > 0)
3432 wait_point
= (__le64_to_cpu(bsb
.arraystart2
) +
3433 __le64_to_cpu(bsb
.length2
));
3436 if (part
== 0 && __le64_to_cpu(bsb
.length
) > 0)
3437 wait_point
= __le64_to_cpu(bsb
.arraystart
);
3438 if (part
== 1 && __le64_to_cpu(bsb
.length2
) > 0)
3439 wait_point
= __le64_to_cpu(bsb
.arraystart2
);
3442 rv
= progress_reshape(sra
, reshape
,
3443 backup_point
, wait_point
,
3444 &suspend_point
, &reshape_completed
);
3445 /* external metadata would need to ping_monitor here */
3446 sra
->reshape_progress
= reshape_completed
;
3448 /* Clear any backup region that is before 'here' */
3450 if (__le64_to_cpu(bsb
.length
) > 0 &&
3451 reshape_completed
>= (__le64_to_cpu(bsb
.arraystart
) +
3452 __le64_to_cpu(bsb
.length
)))
3453 forget_backup(dests
, destfd
,
3455 if (__le64_to_cpu(bsb
.length2
) > 0 &&
3456 reshape_completed
>= (__le64_to_cpu(bsb
.arraystart2
) +
3457 __le64_to_cpu(bsb
.length2
)))
3458 forget_backup(dests
, destfd
,
3461 if (__le64_to_cpu(bsb
.length
) > 0 &&
3462 reshape_completed
<= (__le64_to_cpu(bsb
.arraystart
)))
3463 forget_backup(dests
, destfd
,
3465 if (__le64_to_cpu(bsb
.length2
) > 0 &&
3466 reshape_completed
<= (__le64_to_cpu(bsb
.arraystart2
)))
3467 forget_backup(dests
, destfd
,
3476 if (rv
== 0 && increasing
&& !st
->ss
->external
) {
3477 /* No longer need to monitor this reshape */
3483 unsigned long long offset
;
3484 unsigned long actual_stripes
;
3485 /* Need to backup some data.
3486 * If 'part' is not used and the desired
3487 * backup size is suspended, do a backup,
3488 * then consider the next part.
3490 /* Check that 'part' is unused */
3491 if (part
== 0 && __le64_to_cpu(bsb
.length
) != 0)
3493 if (part
== 1 && __le64_to_cpu(bsb
.length2
) != 0)
3496 offset
= backup_point
/ data
;
3497 actual_stripes
= stripes
;
3499 if (offset
+ actual_stripes
* (chunk
/512) >
3500 sra
->component_size
)
3501 actual_stripes
= ((sra
->component_size
- offset
)
3503 if (offset
+ actual_stripes
* (chunk
/512) >
3507 if (offset
< actual_stripes
* (chunk
/512))
3508 actual_stripes
= offset
/ (chunk
/512);
3509 offset
-= actual_stripes
* (chunk
/512);
3510 if (offset
< suspend_point
/data
)
3513 if (actual_stripes
== 0)
3515 grow_backup(sra
, offset
, actual_stripes
,
3517 disks
, chunk
, level
, layout
,
3518 dests
, destfd
, destoffsets
,
3519 part
, °raded
, buf
);
3520 validate(afd
, destfd
[0], destoffsets
[0]);
3521 /* record where 'part' is up to */
3524 backup_point
+= actual_stripes
* (chunk
/512) * data
;
3526 backup_point
-= actual_stripes
* (chunk
/512) * data
;
3530 /* FIXME maybe call progress_reshape one more time instead */
3531 abort_reshape(sra
); /* remove any remaining suspension */
3532 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
)
3533 sysfs_set_num(sra
, NULL
, "sync_speed_min", speed
);
3539 * If any spare contains md_back_data-1 which is recent wrt mtime,
3540 * write that data into the array and update the super blocks with
3541 * the new reshape_progress
3543 int Grow_restart(struct supertype
*st
, struct mdinfo
*info
, int *fdlist
, int cnt
,
3544 char *backup_file
, int verbose
)
3548 unsigned long long *offsets
;
3549 unsigned long long nstripe
, ostripe
;
3552 odata
= info
->array
.raid_disks
- info
->delta_disks
- 1;
3553 if (info
->array
.level
== 6) odata
--; /* number of data disks */
3554 ndata
= info
->array
.raid_disks
- 1;
3555 if (info
->new_level
== 6) ndata
--;
3557 old_disks
= info
->array
.raid_disks
- info
->delta_disks
;
3559 if (info
->delta_disks
<= 0)
3560 /* Didn't grow, so the backup file must have
3564 for (i
=old_disks
-(backup_file
?1:0); i
<cnt
; i
++) {
3565 struct mdinfo dinfo
;
3568 char *devname
, namebuf
[20];
3569 unsigned long long lo
, hi
;
3571 /* This was a spare and may have some saved data on it.
3572 * Load the superblock, find and load the
3573 * backup_super_block.
3574 * If either fail, go on to next device.
3575 * If the backup contains no new info, just return
3576 * else restore data and update all superblocks
3578 if (i
== old_disks
-1) {
3579 fd
= open(backup_file
, O_RDONLY
);
3581 fprintf(stderr
, Name
": backup file %s inaccessible: %s\n",
3582 backup_file
, strerror(errno
));
3585 devname
= backup_file
;
3590 if (st
->ss
->load_super(st
, fd
, NULL
))
3593 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
3594 st
->ss
->free_super(st
);
3597 (dinfo
.data_offset
+ dinfo
.component_size
- 8) <<9,
3599 fprintf(stderr
, Name
": Cannot seek on device %d\n", i
);
3600 continue; /* Cannot seek */
3602 sprintf(namebuf
, "device-%d", i
);
3605 if (read(fd
, &bsb
, sizeof(bsb
)) != sizeof(bsb
)) {
3607 fprintf(stderr
, Name
": Cannot read from %s\n", devname
);
3608 continue; /* Cannot read */
3610 if (memcmp(bsb
.magic
, "md_backup_data-1", 16) != 0 &&
3611 memcmp(bsb
.magic
, "md_backup_data-2", 16) != 0) {
3613 fprintf(stderr
, Name
": No backup metadata on %s\n", devname
);
3616 if (bsb
.sb_csum
!= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
))) {
3618 fprintf(stderr
, Name
": Bad backup-metadata checksum on %s\n", devname
);
3619 continue; /* bad checksum */
3621 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0 &&
3622 bsb
.sb_csum2
!= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum2
)-((char*)&bsb
))) {
3624 fprintf(stderr
, Name
": Bad backup-metadata checksum2 on %s\n", devname
);
3625 continue; /* Bad second checksum */
3627 if (memcmp(bsb
.set_uuid
,info
->uuid
, 16) != 0) {
3629 fprintf(stderr
, Name
": Wrong uuid on backup-metadata on %s\n", devname
);
3630 continue; /* Wrong uuid */
3633 /* array utime and backup-mtime should be updated at much the same time, but it seems that
3634 * sometimes they aren't... So allow considerable flexability in matching, and allow
3635 * this test to be overridden by an environment variable.
3637 if (info
->array
.utime
> (int)__le64_to_cpu(bsb
.mtime
) + 2*60*60 ||
3638 info
->array
.utime
< (int)__le64_to_cpu(bsb
.mtime
) - 10*60) {
3639 if (check_env("MDADM_GROW_ALLOW_OLD")) {
3640 fprintf(stderr
, Name
": accepting backup with timestamp %lu "
3641 "for array with timestamp %lu\n",
3642 (unsigned long)__le64_to_cpu(bsb
.mtime
),
3643 (unsigned long)info
->array
.utime
);
3646 fprintf(stderr
, Name
": too-old timestamp on "
3647 "backup-metadata on %s\n", devname
);
3648 continue; /* time stamp is too bad */
3652 if (bsb
.magic
[15] == '1') {
3653 if (bsb
.length
== 0)
3655 if (info
->delta_disks
>= 0) {
3656 /* reshape_progress is increasing */
3657 if (__le64_to_cpu(bsb
.arraystart
)
3658 + __le64_to_cpu(bsb
.length
)
3659 < info
->reshape_progress
) {
3662 fprintf(stderr
, Name
3663 ": backup-metadata found on %s but is not needed\n", devname
);
3664 continue; /* No new data here */
3667 /* reshape_progress is decreasing */
3668 if (__le64_to_cpu(bsb
.arraystart
) >=
3669 info
->reshape_progress
)
3670 goto nonew
; /* No new data here */
3673 if (bsb
.length
== 0 && bsb
.length2
== 0)
3675 if (info
->delta_disks
>= 0) {
3676 /* reshape_progress is increasing */
3677 if ((__le64_to_cpu(bsb
.arraystart
)
3678 + __le64_to_cpu(bsb
.length
)
3679 < info
->reshape_progress
)
3681 (__le64_to_cpu(bsb
.arraystart2
)
3682 + __le64_to_cpu(bsb
.length2
)
3683 < info
->reshape_progress
))
3684 goto nonew
; /* No new data here */
3686 /* reshape_progress is decreasing */
3687 if (__le64_to_cpu(bsb
.arraystart
) >=
3688 info
->reshape_progress
&&
3689 __le64_to_cpu(bsb
.arraystart2
) >=
3690 info
->reshape_progress
)
3691 goto nonew
; /* No new data here */
3694 if (lseek64(fd
, __le64_to_cpu(bsb
.devstart
)*512, 0)< 0) {
3697 fprintf(stderr
, Name
3698 ": Failed to verify secondary backup-metadata block on %s\n",
3700 continue; /* Cannot seek */
3702 /* There should be a duplicate backup superblock 4k before here */
3703 if (lseek64(fd
, -4096, 1) < 0 ||
3704 read(fd
, &bsb2
, sizeof(bsb2
)) != sizeof(bsb2
))
3705 goto second_fail
; /* Cannot find leading superblock */
3706 if (bsb
.magic
[15] == '1')
3707 bsbsize
= offsetof(struct mdp_backup_super
, pad1
);
3709 bsbsize
= offsetof(struct mdp_backup_super
, pad
);
3710 if (memcmp(&bsb2
, &bsb
, bsbsize
) != 0)
3711 goto second_fail
; /* Cannot find leading superblock */
3713 /* Now need the data offsets for all devices. */
3714 offsets
= malloc(sizeof(*offsets
)*info
->array
.raid_disks
);
3715 for(j
=0; j
<info
->array
.raid_disks
; j
++) {
3718 if (st
->ss
->load_super(st
, fdlist
[j
], NULL
))
3719 /* FIXME should be this be an error */
3721 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
3722 st
->ss
->free_super(st
);
3723 offsets
[j
] = dinfo
.data_offset
* 512;
3725 printf(Name
": restoring critical section\n");
3727 if (restore_stripes(fdlist
, offsets
,
3728 info
->array
.raid_disks
,
3732 fd
, __le64_to_cpu(bsb
.devstart
)*512,
3733 __le64_to_cpu(bsb
.arraystart
)*512,
3734 __le64_to_cpu(bsb
.length
)*512, NULL
)) {
3735 /* didn't succeed, so giveup */
3737 fprintf(stderr
, Name
": Error restoring backup from %s\n",
3743 if (bsb
.magic
[15] == '2' &&
3744 restore_stripes(fdlist
, offsets
,
3745 info
->array
.raid_disks
,
3749 fd
, __le64_to_cpu(bsb
.devstart
)*512 +
3750 __le64_to_cpu(bsb
.devstart2
)*512,
3751 __le64_to_cpu(bsb
.arraystart2
)*512,
3752 __le64_to_cpu(bsb
.length2
)*512, NULL
)) {
3753 /* didn't succeed, so giveup */
3755 fprintf(stderr
, Name
": Error restoring second backup from %s\n",
3763 /* Ok, so the data is restored. Let's update those superblocks. */
3767 lo
= __le64_to_cpu(bsb
.arraystart
);
3768 hi
= lo
+ __le64_to_cpu(bsb
.length
);
3770 if (bsb
.magic
[15] == '2' && bsb
.length2
) {
3771 unsigned long long lo1
, hi1
;
3772 lo1
= __le64_to_cpu(bsb
.arraystart2
);
3773 hi1
= lo1
+ __le64_to_cpu(bsb
.length2
);
3777 } else if (lo
< lo1
)
3783 (info
->reshape_progress
< lo
||
3784 info
->reshape_progress
> hi
))
3785 /* backup does not affect reshape_progress*/ ;
3786 else if (info
->delta_disks
>= 0) {
3787 info
->reshape_progress
= __le64_to_cpu(bsb
.arraystart
) +
3788 __le64_to_cpu(bsb
.length
);
3789 if (bsb
.magic
[15] == '2') {
3790 unsigned long long p2
= __le64_to_cpu(bsb
.arraystart2
) +
3791 __le64_to_cpu(bsb
.length2
);
3792 if (p2
> info
->reshape_progress
)
3793 info
->reshape_progress
= p2
;
3796 info
->reshape_progress
= __le64_to_cpu(bsb
.arraystart
);
3797 if (bsb
.magic
[15] == '2') {
3798 unsigned long long p2
= __le64_to_cpu(bsb
.arraystart2
);
3799 if (p2
< info
->reshape_progress
)
3800 info
->reshape_progress
= p2
;
3803 for (j
=0; j
<info
->array
.raid_disks
; j
++) {
3804 if (fdlist
[j
] < 0) continue;
3805 if (st
->ss
->load_super(st
, fdlist
[j
], NULL
))
3807 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
3808 dinfo
.reshape_progress
= info
->reshape_progress
;
3809 st
->ss
->update_super(st
, &dinfo
,
3810 "_reshape_progress",
3812 st
->ss
->store_super(st
, fdlist
[j
]);
3813 st
->ss
->free_super(st
);
3817 /* Didn't find any backup data, try to see if any
3820 if (info
->delta_disks
< 0) {
3821 /* When shrinking, the critical section is at the end.
3822 * So see if we are before the critical section.
3824 unsigned long long first_block
;
3825 nstripe
= ostripe
= 0;
3827 while (ostripe
>= nstripe
) {
3828 ostripe
+= info
->array
.chunk_size
/ 512;
3829 first_block
= ostripe
* odata
;
3830 nstripe
= first_block
/ ndata
/ (info
->new_chunk
/512) *
3831 (info
->new_chunk
/512);
3834 if (info
->reshape_progress
>= first_block
)
3837 if (info
->delta_disks
> 0) {
3838 /* See if we are beyond the critical section. */
3839 unsigned long long last_block
;
3840 nstripe
= ostripe
= 0;
3842 while (nstripe
>= ostripe
) {
3843 nstripe
+= info
->new_chunk
/ 512;
3844 last_block
= nstripe
* ndata
;
3845 ostripe
= last_block
/ odata
/ (info
->array
.chunk_size
/512) *
3846 (info
->array
.chunk_size
/512);
3849 if (info
->reshape_progress
>= last_block
)
3852 /* needed to recover critical section! */
3854 fprintf(stderr
, Name
": Failed to find backup of critical section\n");
3858 int Grow_continue_command(char *devname
, int fd
,
3859 char *backup_file
, int verbose
)
3862 struct supertype
*st
= NULL
;
3863 struct mdinfo
*content
= NULL
;
3864 struct mdinfo array
;
3865 char *subarray
= NULL
;
3866 struct mdinfo
*cc
= NULL
;
3867 struct mdstat_ent
*mdstat
= NULL
;
3872 dprintf("Grow continue from command line called for %s\n",
3875 st
= super_by_fd(fd
, &subarray
);
3876 if (!st
|| !st
->ss
) {
3878 Name
": Unable to determine metadata format for %s\n",
3882 dprintf("Grow continue is run for ");
3883 if (st
->ss
->external
== 0) {
3884 dprintf("native array (%s)\n", devname
);
3885 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) < 0) {
3886 fprintf(stderr
, Name
": %s is not an active md array -"
3887 " aborting\n", devname
);
3889 goto Grow_continue_command_exit
;
3892 sysfs_init(content
, fd
, st
->devnum
);
3897 dprintf("subarray (%s)\n", subarray
);
3898 container_dev
= st
->container_dev
;
3899 cfd
= open_dev_excl(st
->container_dev
);
3901 container_dev
= st
->devnum
;
3903 cfd
= open_dev_excl(st
->devnum
);
3904 dprintf("container (%i)\n", container_dev
);
3908 fprintf(stderr
, Name
": Unable to open container "
3909 "for %s\n", devname
);
3911 goto Grow_continue_command_exit
;
3913 fmt_devname(buf
, container_dev
);
3915 /* find in container array under reshape
3917 ret_val
= st
->ss
->load_container(st
, cfd
, NULL
);
3920 Name
": Cannot read superblock for %s\n",
3923 goto Grow_continue_command_exit
;
3926 cc
= st
->ss
->container_content(st
, subarray
);
3927 for (content
= cc
; content
; content
= content
->next
) {
3929 int allow_reshape
= 1;
3931 if (content
->reshape_active
== 0)
3933 /* The decision about array or container wide
3934 * reshape is taken in Grow_continue based
3935 * content->reshape_active state, therefore we
3936 * need to check_reshape based on
3937 * reshape_active and subarray name
3939 if (content
->array
.state
& (1<<MD_SB_BLOCK_VOLUME
))
3941 if (content
->reshape_active
== CONTAINER_RESHAPE
&&
3942 (content
->array
.state
3943 & (1<<MD_SB_BLOCK_CONTAINER_RESHAPE
)))
3946 if (!allow_reshape
) {
3947 fprintf(stderr
, Name
3948 ": cannot continue reshape of an array"
3949 " in container with unsupported"
3950 " metadata: %s(%s)\n",
3953 goto Grow_continue_command_exit
;
3956 array
= strchr(content
->text_version
+1, '/')+1;
3957 mdstat
= mdstat_by_subdev(array
, container_dev
);
3960 if (mdstat
->active
== 0) {
3961 fprintf(stderr
, Name
": Skipping inactive "
3962 "array md%i.\n", mdstat
->devnum
);
3963 free_mdstat(mdstat
);
3971 Name
": Unable to determine reshaped "
3972 "array for %s\n", devname
);
3974 goto Grow_continue_command_exit
;
3976 fd2
= open_dev(mdstat
->devnum
);
3978 fprintf(stderr
, Name
": cannot open (md%i)\n",
3981 goto Grow_continue_command_exit
;
3984 sysfs_init(content
, fd2
, mdstat
->devnum
);
3986 /* start mdmon in case it is not running
3988 if (!mdmon_running(container_dev
))
3989 start_mdmon(container_dev
);
3992 if (mdmon_running(container_dev
))
3993 st
->update_tail
= &st
->updates
;
3995 fprintf(stderr
, Name
": No mdmon found. "
3996 "Grow cannot continue.\n");
3998 goto Grow_continue_command_exit
;
4002 /* verify that array under reshape is started from
4005 if (verify_reshape_position(content
,
4006 map_name(pers
, mdstat
->level
)) < 0) {
4008 goto Grow_continue_command_exit
;
4013 ret_val
= Grow_continue(fd
, st
, content
, backup_file
, 0);
4015 Grow_continue_command_exit
:
4020 st
->ss
->free_super(st
);
4021 free_mdstat(mdstat
);
4028 int Grow_continue(int mdfd
, struct supertype
*st
, struct mdinfo
*info
,
4029 char *backup_file
, int freeze_reshape
)
4033 if (!info
->reshape_active
)
4036 if (st
->ss
->external
) {
4038 int cfd
= open_dev(st
->container_dev
);
4043 fmt_devname(container
, st
->container_dev
);
4044 st
->ss
->load_container(st
, cfd
, container
);
4046 ret_val
= reshape_container(container
, NULL
, mdfd
,
4047 st
, info
, 0, backup_file
,
4048 0, 1, freeze_reshape
);
4050 ret_val
= reshape_array(NULL
, mdfd
, "array", st
, info
, 1,
4051 NULL
, backup_file
, 0, 0, 1,