2 * mdadm - manage Linux "md" devices aka RAID arrays.
4 * Copyright (C) 2001-2012 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>
29 #if ! defined(__BIG_ENDIAN) && ! defined(__LITTLE_ENDIAN)
30 #error no endian defined
36 #define offsetof(t,f) ((size_t)&(((t*)0)->f))
39 int restore_backup(struct supertype
*st
,
40 struct mdinfo
*content
,
50 int disk_count
= next_spare
+ working_disks
;
52 dprintf("Called restore_backup()\n");
53 fdlist
= xmalloc(sizeof(int) * disk_count
);
55 enable_fds(next_spare
);
56 for (i
= 0; i
< next_spare
; i
++)
58 for (dev
= content
->devs
; dev
; dev
= dev
->next
) {
64 fd
= dev_open(buf
, O_RDWR
);
66 if (dev
->disk
.raid_disk
>= 0)
67 fdlist
[dev
->disk
.raid_disk
] = fd
;
69 fdlist
[next_spare
++] = fd
;
72 if (st
->ss
->external
&& st
->ss
->recover_backup
)
73 err
= st
->ss
->recover_backup(st
, content
);
75 err
= Grow_restart(st
, content
, fdlist
, next_spare
,
76 backup_file
, verbose
> 0);
78 while (next_spare
> 0) {
80 if (fdlist
[next_spare
] >= 0)
81 close(fdlist
[next_spare
]);
85 pr_err("Failed to restore critical"
86 " section for reshape - sorry.\n");
88 pr_err("Possibly you need"
89 " to specify a --backup-file\n");
93 dprintf("restore_backup() returns status OK.\n");
97 int Grow_Add_device(char *devname
, int fd
, char *newdev
)
99 /* Add a device to an active array.
100 * Currently, just extend a linear array.
101 * This requires writing a new superblock on the
102 * new device, calling the kernel to add the device,
103 * and if that succeeds, update the superblock on
105 * This means that we need to *find* all other devices.
112 struct supertype
*st
= NULL
;
113 char *subarray
= NULL
;
115 if (ioctl(fd
, GET_ARRAY_INFO
, &info
.array
) < 0) {
116 pr_err("cannot get array info for %s\n", devname
);
120 if (info
.array
.level
!= -1) {
121 pr_err("can only add devices to linear arrays\n");
125 st
= super_by_fd(fd
, &subarray
);
127 pr_err("cannot handle arrays with superblock version %d\n",
128 info
.array
.major_version
);
133 pr_err("Cannot grow linear sub-arrays yet\n");
139 nfd
= open(newdev
, O_RDWR
|O_EXCL
|O_DIRECT
);
141 pr_err("cannot open %s\n", newdev
);
146 if ((stb
.st_mode
& S_IFMT
) != S_IFBLK
) {
147 pr_err("%s is not a block device!\n", newdev
);
152 /* now check out all the devices and make sure we can read the
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 pr_err("cannot get device detail for device %d\n",
168 dv
= map_dev(disk
.major
, disk
.minor
, 1);
170 pr_err("cannot find device file for device %d\n",
176 fd2
= dev_open(dv
, O_RDWR
);
178 pr_err("cannot open device file %s\n", dv
);
184 if (st
->ss
->load_super(st
, fd2
, NULL
)) {
185 pr_err("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 pr_err("Cannot store new superblock on %s\n",
213 if (ioctl(fd
, ADD_NEW_DISK
, &info
.disk
) != 0) {
214 pr_err("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 pr_err("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 pr_err("cannot get device detail for device %d\n",
237 dv
= map_dev(disk
.major
, disk
.minor
, 1);
239 pr_err("cannot find device file for device %d\n",
243 fd2
= dev_open(dv
, O_RDWR
);
245 pr_err("cannot open device file %s\n", dv
);
248 if (st
->ss
->load_super(st
, fd2
, NULL
)) {
249 pr_err("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 pr_err("Cannot store new superblock on %s\n", dv
);
272 int Grow_addbitmap(char *devname
, int fd
, struct context
*c
, struct shape
*s
)
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 pr_err("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 pr_err("Memory allocation failure.\n");
303 pr_err("bitmaps not supported by this kernel.\n");
306 if (bmf
.pathname
[0]) {
307 if (strcmp(s
->bitmap_file
,"none")==0) {
308 if (ioctl(fd
, SET_BITMAP_FILE
, -1)!= 0) {
309 pr_err("failed to remove bitmap %s\n",
315 pr_err("%s already has a bitmap (%s)\n",
316 devname
, bmf
.pathname
);
319 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
320 pr_err("cannot get array status for %s\n", devname
);
323 if (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)) {
324 if (strcmp(s
->bitmap_file
, "none")==0) {
325 array
.state
&= ~(1<<MD_SB_BITMAP_PRESENT
);
326 if (ioctl(fd
, SET_ARRAY_INFO
, &array
)!= 0) {
327 pr_err("failed to remove internal bitmap.\n");
332 pr_err("Internal bitmap already present on %s\n",
337 if (strcmp(s
->bitmap_file
, "none") == 0) {
338 pr_err("no bitmap found on %s\n", devname
);
341 if (array
.level
<= 0) {
342 pr_err("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 pr_err("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 pr_err("Cannot understand version %d.%d\n",
368 array
.major_version
, array
.minor_version
);
372 pr_err("Cannot add bitmaps to sub-arrays yet\n");
377 if (strcmp(s
->bitmap_file
, "internal") == 0) {
380 int offset_setable
= 0;
382 if (st
->ss
->add_internal_bitmap
== NULL
) {
383 pr_err("Internal bitmaps not supported "
384 "with %s metadata\n", st
->ss
->name
);
387 mdi
= sysfs_read(fd
, NULL
, 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 &s
->bitmap_chunk
, c
->delay
, s
->write_behind
,
410 bitmapsize
, offset_setable
,
413 st
->ss
->write_bitmap(st
, fd2
);
415 pr_err("failed to create internal bitmap"
416 " - chunksize problem.\n");
424 if (offset_setable
) {
425 st
->ss
->getinfo_super(st
, mdi
, NULL
);
426 sysfs_init(mdi
, fd
, NULL
);
427 rv
= sysfs_set_num_signed(mdi
, NULL
, "bitmap/location",
430 array
.state
|= (1<<MD_SB_BITMAP_PRESENT
);
431 rv
= ioctl(fd
, SET_ARRAY_INFO
, &array
);
435 pr_err("Cannot add bitmap while array is"
436 " resyncing or reshaping etc.\n");
437 pr_err("failed to set internal bitmap.\n");
444 int max_devs
= st
->max_devs
;
446 /* try to load a superblock */
447 for (d
= 0; d
< max_devs
; d
++) {
448 mdu_disk_info_t disk
;
452 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0)
454 if ((disk
.major
==0 && disk
.minor
==0) ||
455 (disk
.state
& (1<<MD_DISK_REMOVED
)))
457 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 pr_err("cannot find UUID for array!\n");
474 if (CreateBitmap(s
->bitmap_file
, c
->force
, (char*)uuid
, s
->bitmap_chunk
,
475 c
->delay
, s
->write_behind
, bitmapsize
, major
)) {
478 bitmap_fd
= open(s
->bitmap_file
, O_RDWR
);
480 pr_err("weird: %s cannot be opened\n",
484 if (ioctl(fd
, SET_BITMAP_FILE
, bitmap_fd
) < 0) {
487 pr_err("Cannot add bitmap while array is"
488 " resyncing or reshaping etc.\n");
489 pr_err("Cannot set bitmap file for %s: %s\n",
490 devname
, strerror(err
));
499 * When reshaping an array we might need to backup some data.
500 * This is written to all spares with a 'super_block' describing it.
501 * The superblock goes 4K from the end of the used space on the
503 * It if written after the backup is complete.
504 * It has the following structure.
507 static struct mdp_backup_super
{
508 char magic
[16]; /* md_backup_data-1 or -2 */
511 /* start/sizes in 512byte sectors */
512 __u64 devstart
; /* address on backup device/file of data */
515 __u32 sb_csum
; /* csum of preceeding bytes. */
517 __u64 devstart2
; /* offset in to data of second section */
520 __u32 sb_csum2
; /* csum of preceeding bytes. */
522 } __attribute__((aligned(512))) bsb
, bsb2
;
524 static __u32
bsb_csum(char *buf
, int len
)
528 for (i
= 0; i
< len
; i
++)
529 csum
= (csum
<<3) + buf
[0];
530 return __cpu_to_le32(csum
);
533 static int check_idle(struct supertype
*st
)
535 /* Check that all member arrays for this container, or the
536 * container of this array, are idle
538 char *container
= (st
->container_devnm
[0]
539 ? st
->container_devnm
: st
->devnm
);
540 struct mdstat_ent
*ent
, *e
;
543 ent
= mdstat_read(0, 0);
544 for (e
= ent
; e
; e
= e
->next
) {
545 if (!is_container_member(e
, container
))
547 if (e
->percent
>= 0) {
556 static int freeze_container(struct supertype
*st
)
558 char *container
= (st
->container_devnm
[0]
559 ? st
->container_devnm
: st
->devnm
);
564 if (block_monitor(container
, 1)) {
565 pr_err("failed to freeze container\n");
572 static void unfreeze_container(struct supertype
*st
)
574 char *container
= (st
->container_devnm
[0]
575 ? st
->container_devnm
: st
->devnm
);
577 unblock_monitor(container
, 1);
580 static int freeze(struct supertype
*st
)
582 /* Try to freeze resync/rebuild on this array/container.
583 * Return -1 if the array is busy,
584 * return -2 container cannot be frozen,
585 * return 0 if this kernel doesn't support 'frozen'
586 * return 1 if it worked.
588 if (st
->ss
->external
)
589 return freeze_container(st
);
591 struct mdinfo
*sra
= sysfs_read(-1, st
->devnm
, GET_VERSION
);
597 /* Need to clear any 'read-auto' status */
598 if (sysfs_get_str(sra
, NULL
, "array_state", buf
, 20) > 0 &&
599 strncmp(buf
, "read-auto", 9) == 0)
600 sysfs_set_str(sra
, NULL
, "array_state", "clean");
602 err
= sysfs_freeze_array(sra
);
608 static void unfreeze(struct supertype
*st
)
610 if (st
->ss
->external
)
611 return unfreeze_container(st
);
613 struct mdinfo
*sra
= sysfs_read(-1, st
->devnm
, GET_VERSION
);
616 sysfs_set_str(sra
, NULL
, "sync_action", "idle");
621 static void wait_reshape(struct mdinfo
*sra
)
623 int fd
= sysfs_get_fd(sra
, NULL
, "sync_action");
629 while (sysfs_fd_get_str(fd
, action
, 20) > 0 &&
630 strncmp(action
, "reshape", 7) == 0)
631 sysfs_wait(fd
, NULL
);
635 static int reshape_super(struct supertype
*st
, unsigned long long size
,
636 int level
, int layout
, int chunksize
, int raid_disks
,
637 int delta_disks
, char *backup_file
, char *dev
,
638 int direction
, int verbose
)
640 /* nothing extra to check in the native case */
641 if (!st
->ss
->external
)
643 if (!st
->ss
->reshape_super
||
644 !st
->ss
->manage_reshape
) {
645 pr_err("%s metadata does not support reshape\n",
650 return st
->ss
->reshape_super(st
, size
, level
, layout
, chunksize
,
651 raid_disks
, delta_disks
, backup_file
, dev
,
655 static void sync_metadata(struct supertype
*st
)
657 if (st
->ss
->external
) {
658 if (st
->update_tail
) {
659 flush_metadata_updates(st
);
660 st
->update_tail
= &st
->updates
;
662 st
->ss
->sync_metadata(st
);
666 static int subarray_set_num(char *container
, struct mdinfo
*sra
, char *name
, int n
)
668 /* when dealing with external metadata subarrays we need to be
669 * prepared to handle EAGAIN. The kernel may need to wait for
670 * mdmon to mark the array active so the kernel can handle
671 * allocations/writeback when preparing the reshape action
672 * (md_allow_write()). We temporarily disable safe_mode_delay
673 * to close a race with the array_state going clean before the
674 * next write to raid_disks / stripe_cache_size
679 /* only 'raid_disks' and 'stripe_cache_size' trigger md_allow_write */
681 (strcmp(name
, "raid_disks") != 0 &&
682 strcmp(name
, "stripe_cache_size") != 0))
683 return sysfs_set_num(sra
, NULL
, name
, n
);
685 rc
= sysfs_get_str(sra
, NULL
, "safe_mode_delay", safe
, sizeof(safe
));
688 sysfs_set_num(sra
, NULL
, "safe_mode_delay", 0);
689 rc
= sysfs_set_num(sra
, NULL
, name
, n
);
690 if (rc
< 0 && errno
== EAGAIN
) {
691 ping_monitor(container
);
692 /* if we get EAGAIN here then the monitor is not active
695 rc
= sysfs_set_num(sra
, NULL
, name
, n
);
697 sysfs_set_str(sra
, NULL
, "safe_mode_delay", safe
);
701 int start_reshape(struct mdinfo
*sra
, int already_running
,
702 int before_data_disks
, int data_disks
)
705 unsigned long long sync_max_to_set
;
707 sysfs_set_num(sra
, NULL
, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL
);
708 err
= sysfs_set_num(sra
, NULL
, "suspend_hi", sra
->reshape_progress
);
709 err
= err
?: sysfs_set_num(sra
, NULL
, "suspend_lo",
710 sra
->reshape_progress
);
711 if (before_data_disks
<= data_disks
)
712 sync_max_to_set
= sra
->reshape_progress
/ data_disks
;
714 sync_max_to_set
= (sra
->component_size
* data_disks
715 - sra
->reshape_progress
) / data_disks
;
716 if (!already_running
)
717 sysfs_set_num(sra
, NULL
, "sync_min", sync_max_to_set
);
718 err
= err
?: sysfs_set_num(sra
, NULL
, "sync_max", sync_max_to_set
);
719 if (!already_running
)
720 err
= err
?: sysfs_set_str(sra
, NULL
, "sync_action", "reshape");
725 void abort_reshape(struct mdinfo
*sra
)
727 sysfs_set_str(sra
, NULL
, "sync_action", "idle");
728 sysfs_set_num(sra
, NULL
, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL
);
729 sysfs_set_num(sra
, NULL
, "suspend_hi", 0);
730 sysfs_set_num(sra
, NULL
, "suspend_lo", 0);
731 sysfs_set_num(sra
, NULL
, "sync_min", 0);
732 sysfs_set_str(sra
, NULL
, "sync_max", "max");
735 int remove_disks_for_takeover(struct supertype
*st
,
740 struct mdinfo
*remaining
;
743 if (sra
->array
.level
== 10)
744 nr_of_copies
= layout
& 0xff;
745 else if (sra
->array
.level
== 1)
746 nr_of_copies
= sra
->array
.raid_disks
;
750 remaining
= sra
->devs
;
752 /* for each 'copy', select one device and remove from the list. */
753 for (slot
= 0; slot
< sra
->array
.raid_disks
; slot
+= nr_of_copies
) {
754 struct mdinfo
**diskp
;
757 /* Find a working device to keep */
758 for (diskp
= &remaining
; *diskp
; diskp
= &(*diskp
)->next
) {
759 struct mdinfo
*disk
= *diskp
;
761 if (disk
->disk
.raid_disk
< slot
)
763 if (disk
->disk
.raid_disk
>= slot
+ nr_of_copies
)
765 if (disk
->disk
.state
& (1<<MD_DISK_REMOVED
))
767 if (disk
->disk
.state
& (1<<MD_DISK_FAULTY
))
769 if (!(disk
->disk
.state
& (1<<MD_DISK_SYNC
)))
772 /* We have found a good disk to use! */
774 disk
->next
= sra
->devs
;
783 if (slot
< sra
->array
.raid_disks
) {
784 /* didn't find all slots */
790 sra
->devs
= remaining
;
794 /* Remove all 'remaining' devices from the array */
796 struct mdinfo
*sd
= remaining
;
797 remaining
= sd
->next
;
799 sysfs_set_str(sra
, sd
, "state", "faulty");
800 sysfs_set_str(sra
, sd
, "slot", "none");
801 /* for external metadata disks should be removed in mdmon */
802 if (!st
->ss
->external
)
803 sysfs_set_str(sra
, sd
, "state", "remove");
804 sd
->disk
.state
|= (1<<MD_DISK_REMOVED
);
805 sd
->disk
.state
&= ~(1<<MD_DISK_SYNC
);
806 sd
->next
= sra
->devs
;
812 void reshape_free_fdlist(int *fdlist
,
813 unsigned long long *offsets
,
818 for (i
= 0; i
< size
; i
++)
826 int reshape_prepare_fdlist(char *devname
,
830 unsigned long blocks
,
833 unsigned long long *offsets
)
839 for (d
= 0; d
<= nrdisks
; d
++)
842 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
843 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
845 if (sd
->disk
.state
& (1<<MD_DISK_SYNC
)) {
846 char *dn
= map_dev(sd
->disk
.major
,
848 fdlist
[sd
->disk
.raid_disk
]
849 = dev_open(dn
, O_RDONLY
);
850 offsets
[sd
->disk
.raid_disk
] = sd
->data_offset
*512;
851 if (fdlist
[sd
->disk
.raid_disk
] < 0) {
852 pr_err("%s: cannot open component %s\n",
853 devname
, dn
? dn
: "-unknown-");
857 } else if (backup_file
== NULL
) {
859 char *dn
= map_dev(sd
->disk
.major
,
861 fdlist
[d
] = dev_open(dn
, O_RDWR
);
862 offsets
[d
] = (sd
->data_offset
+ sra
->component_size
- blocks
- 8)*512;
864 pr_err("%s: cannot open component %s\n",
865 devname
, dn
? dn
: "-unknown-");
876 int reshape_open_backup_file(char *backup_file
,
881 unsigned long long *offsets
,
884 /* Return 1 on success, 0 on any form of failure */
885 /* need to check backup file is large enough */
891 *fdlist
= open(backup_file
, O_RDWR
|O_CREAT
|(restart
? O_TRUNC
: O_EXCL
),
895 pr_err("%s: cannot create backup file %s: %s\n",
896 devname
, backup_file
, strerror(errno
));
899 /* Guard against backup file being on array device.
900 * If array is partitioned or if LVM etc is in the
901 * way this will not notice, but it is better than
904 fstat(*fdlist
, &stb
);
907 if (stb
.st_rdev
== dev
) {
908 pr_err("backup file must NOT be"
909 " on the array being reshaped.\n");
915 for (i
=0; i
< blocks
+ 8 ; i
++) {
916 if (write(*fdlist
, buf
, 512) != 512) {
917 pr_err("%s: cannot create"
918 " backup file %s: %s\n",
919 devname
, backup_file
, strerror(errno
));
923 if (fsync(*fdlist
) != 0) {
924 pr_err("%s: cannot create backup file %s: %s\n",
925 devname
, backup_file
, strerror(errno
));
932 unsigned long compute_backup_blocks(int nchunk
, int ochunk
,
933 unsigned int ndata
, unsigned int odata
)
935 unsigned long a
, b
, blocks
;
936 /* So how much do we need to backup.
937 * We need an amount of data which is both a whole number of
938 * old stripes and a whole number of new stripes.
939 * So LCM for (chunksize*datadisks).
941 a
= (ochunk
/512) * odata
;
942 b
= (nchunk
/512) * ndata
;
945 /* LCM == product / GCD */
946 blocks
= (ochunk
/512) * (nchunk
/512) * odata
* ndata
/ a
;
951 char *analyse_change(char *devname
, struct mdinfo
*info
, struct reshape
*re
)
953 /* Based on the current array state in info->array and
954 * the changes in info->new_* etc, determine:
955 * - whether the change is possible
956 * - Intermediate level/raid_disks/layout
957 * - whether a restriping reshape is needed
958 * - number of sectors in minimum change unit. This
959 * will cover a whole number of stripes in 'before' and
962 * Return message if the change should be rejected
963 * NULL if the change can be achieved
965 * This can be called as part of starting a reshape, or
966 * when assembling an array that is undergoing reshape.
968 int near
, far
, offset
, copies
;
970 int old_chunk
, new_chunk
;
971 /* delta_parity records change in number of devices
972 * caused by level change
974 int delta_parity
= 0;
976 memset(re
, 0, sizeof(*re
));
978 /* If a new level not explicitly given, we assume no-change */
979 if (info
->new_level
== UnSet
)
980 info
->new_level
= info
->array
.level
;
983 switch (info
->new_level
) {
989 /* chunk size is meaningful, must divide component_size
992 if (info
->component_size
% (info
->new_chunk
/512)) {
993 unsigned long long shrink
= info
->component_size
;
994 shrink
&= ~(unsigned long long)(info
->new_chunk
/512-1);
995 pr_err("New chunk size (%dK) does not evenly divide device size (%lluk)\n",
996 info
->new_chunk
/1024, info
->component_size
/2);
997 pr_err("After shrinking any filesystem, \"mdadm --grow %s --size %llu\"\n",
999 pr_err("will shrink the array so the given chunk size would work.\n");
1004 return "chunk size not meaningful for this level";
1007 info
->new_chunk
= info
->array
.chunk_size
;
1009 switch (info
->array
.level
) {
1011 return "Cannot understand this RAID level";
1013 /* RAID1 can convert to RAID1 with different disks, or
1014 * raid5 with 2 disks, or
1017 if (info
->new_level
> 1 &&
1018 (info
->component_size
& 7))
1019 return "Cannot convert RAID1 of this size - "
1020 "reduce size to multiple of 4K first.";
1021 if (info
->new_level
== 0) {
1022 if (info
->delta_disks
!= UnSet
&&
1023 info
->delta_disks
!= 0)
1024 return "Cannot change number of disks "
1025 "with RAID1->RAID0 conversion";
1027 re
->before
.data_disks
= 1;
1028 re
->after
.data_disks
= 1;
1031 if (info
->new_level
== 1) {
1032 if (info
->delta_disks
== UnSet
)
1033 /* Don't know what to do */
1034 return "no change requested for Growing RAID1";
1038 if (info
->array
.raid_disks
== 2 &&
1039 info
->new_level
== 5) {
1042 re
->before
.data_disks
= 1;
1043 if (info
->delta_disks
!= UnSet
&&
1044 info
->delta_disks
!= 0)
1045 re
->after
.data_disks
= 1 + info
->delta_disks
;
1047 re
->after
.data_disks
= 1;
1048 if (re
->after
.data_disks
< 1)
1049 return "Number of disks too small for RAID5";
1051 re
->before
.layout
= ALGORITHM_LEFT_SYMMETRIC
;
1052 info
->array
.chunk_size
= 65536;
1055 /* Could do some multi-stage conversions, but leave that to
1058 return "Impossibly level change request for RAID1";
1061 /* RAID10 can be converted from near mode to
1062 * RAID0 by removing some devices.
1063 * It can also be reshaped if the kernel supports
1066 switch (info
->new_level
) {
1068 if ((info
->array
.layout
& ~0xff) != 0x100)
1069 return "Cannot Grow RAID10 with far/offset layout";
1070 /* number of devices must be multiple of number of copies */
1071 if (info
->array
.raid_disks
% (info
->array
.layout
& 0xff))
1072 return "RAID10 layout too complex for Grow operation";
1074 new_disks
= (info
->array
.raid_disks
1075 / (info
->array
.layout
& 0xff));
1076 if (info
->delta_disks
== UnSet
)
1077 info
->delta_disks
= (new_disks
1078 - info
->array
.raid_disks
);
1080 if (info
->delta_disks
!= new_disks
- info
->array
.raid_disks
)
1081 return "New number of raid-devices impossible for RAID10";
1082 if (info
->new_chunk
&&
1083 info
->new_chunk
!= info
->array
.chunk_size
)
1084 return "Cannot change chunk-size with RAID10 Grow";
1088 re
->before
.data_disks
= new_disks
;
1089 re
->after
.data_disks
= re
->before
.data_disks
;
1093 near
= info
->array
.layout
& 0xff;
1094 far
= (info
->array
.layout
>> 8) & 0xff;
1095 offset
= info
->array
.layout
& 0x10000;
1096 if (far
> 1 && !offset
)
1097 return "Cannot reshape RAID10 in far-mode";
1098 copies
= near
* far
;
1100 old_chunk
= info
->array
.chunk_size
* far
;
1102 if (info
->new_layout
== UnSet
)
1103 info
->new_layout
= info
->array
.layout
;
1105 near
= info
->new_layout
& 0xff;
1106 far
= (info
->new_layout
>> 8) & 0xff;
1107 offset
= info
->new_layout
& 0x10000;
1108 if (far
> 1 && !offset
)
1109 return "Cannot reshape RAID10 to far-mode";
1110 if (near
* far
!= copies
)
1111 return "Cannot change number of copies"
1112 " when reshaping RAID10";
1114 if (info
->delta_disks
== UnSet
)
1115 info
->delta_disks
= 0;
1116 new_disks
= (info
->array
.raid_disks
+
1119 new_chunk
= info
->new_chunk
* far
;
1122 re
->before
.layout
= info
->array
.layout
;
1123 re
->before
.data_disks
= info
->array
.raid_disks
;
1124 re
->after
.layout
= info
->new_layout
;
1125 re
->after
.data_disks
= new_disks
;
1126 /* For RAID10 we don't do backup but do allow reshape,
1127 * so set backup_blocks to INVALID_SECTORS rather than
1129 * And there is no need to synchronise stripes on both
1130 * 'old' and 'new'. So the important
1131 * number is the minimum data_offset difference
1132 * which is the larger of (offset copies * chunk).
1134 re
->backup_blocks
= INVALID_SECTORS
;
1135 re
->min_offset_change
= max(old_chunk
, new_chunk
) / 512;
1136 if (new_disks
< re
->before
.data_disks
&&
1137 info
->space_after
< re
->min_offset_change
)
1138 /* Reduce component size by one chunk */
1139 re
->new_size
= (info
->component_size
-
1140 re
->min_offset_change
);
1142 re
->new_size
= info
->component_size
;
1143 re
->new_size
= re
->new_size
* new_disks
/ copies
;
1147 return "RAID10 can only be changed to RAID0";
1150 /* RAID0 can be converted to RAID10, or to RAID456 */
1151 if (info
->new_level
== 10) {
1152 if (info
->new_layout
== UnSet
&& info
->delta_disks
== UnSet
) {
1153 /* Assume near=2 layout */
1154 info
->new_layout
= 0x102;
1155 info
->delta_disks
= info
->array
.raid_disks
;
1157 if (info
->new_layout
== UnSet
) {
1158 int copies
= 1 + (info
->delta_disks
1159 / info
->array
.raid_disks
);
1160 if (info
->array
.raid_disks
* (copies
-1)
1161 != info
->delta_disks
)
1162 return "Impossible number of devices"
1163 " for RAID0->RAID10";
1164 info
->new_layout
= 0x100 + copies
;
1166 if (info
->delta_disks
== UnSet
) {
1167 int copies
= info
->new_layout
& 0xff;
1168 if (info
->new_layout
!= 0x100 + copies
)
1169 return "New layout impossible"
1170 " for RAID0->RAID10";;
1171 info
->delta_disks
= (copies
- 1) *
1172 info
->array
.raid_disks
;
1174 if (info
->new_chunk
&&
1175 info
->new_chunk
!= info
->array
.chunk_size
)
1176 return "Cannot change chunk-size with RAID0->RAID10";
1179 re
->before
.data_disks
= (info
->array
.raid_disks
+
1181 re
->after
.data_disks
= re
->before
.data_disks
;
1182 re
->before
.layout
= info
->new_layout
;
1186 /* RAID0 can also covert to RAID0/4/5/6 by first converting to
1187 * a raid4 style layout of the final level.
1189 switch (info
->new_level
) {
1194 re
->before
.layout
= 0;
1199 re
->before
.layout
= ALGORITHM_PARITY_N
;
1200 if (info
->new_layout
== UnSet
)
1201 info
->new_layout
= map_name(r5layout
, "default");
1206 re
->before
.layout
= ALGORITHM_PARITY_N
;
1207 if (info
->new_layout
== UnSet
)
1208 info
->new_layout
= map_name(r6layout
, "default");
1211 return "Impossible level change requested";
1213 re
->before
.data_disks
= info
->array
.raid_disks
;
1214 /* determining 'after' layout happens outside this 'switch' */
1218 info
->array
.layout
= ALGORITHM_PARITY_N
;
1220 switch (info
->new_level
) {
1224 re
->level
= info
->array
.level
;
1225 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1226 re
->before
.layout
= info
->array
.layout
;
1230 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1231 re
->before
.layout
= info
->array
.layout
;
1236 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1237 switch (info
->array
.layout
) {
1238 case ALGORITHM_LEFT_ASYMMETRIC
:
1239 re
->before
.layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
1241 case ALGORITHM_RIGHT_ASYMMETRIC
:
1242 re
->before
.layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
1244 case ALGORITHM_LEFT_SYMMETRIC
:
1245 re
->before
.layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
1247 case ALGORITHM_RIGHT_SYMMETRIC
:
1248 re
->before
.layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
1250 case ALGORITHM_PARITY_0
:
1251 re
->before
.layout
= ALGORITHM_PARITY_0_6
;
1253 case ALGORITHM_PARITY_N
:
1254 re
->before
.layout
= ALGORITHM_PARITY_N_6
;
1257 return "Cannot convert an array with this layout";
1261 if (info
->array
.raid_disks
!= 2)
1262 return "Can only convert a 2-device array to RAID1";
1263 if (info
->delta_disks
!= UnSet
&&
1264 info
->delta_disks
!= 0)
1265 return "Cannot set raid_disk when "
1266 "converting RAID5->RAID1";
1268 info
->new_chunk
= 0;
1271 return "Impossible level change requested";
1275 switch (info
->new_level
) {
1281 re
->before
.data_disks
= info
->array
.raid_disks
- 2;
1282 re
->before
.layout
= info
->array
.layout
;
1285 return "Impossible level change requested";
1290 /* If we reached here then it looks like a re-stripe is
1291 * happening. We have determined the intermediate level
1292 * and initial raid_disks/layout and stored these in 're'.
1294 * We need to deduce the final layout that can be atomically
1295 * converted to the end state.
1297 switch (info
->new_level
) {
1299 /* We can only get to RAID0 from RAID4 or RAID5
1300 * with appropriate layout and one extra device
1302 if (re
->level
!= 4 && re
->level
!= 5)
1303 return "Cannot covert to RAID0 from this level";
1305 switch (re
->level
) {
1307 re
->before
.layout
= 0;
1308 re
->after
.layout
= 0;
1311 re
->after
.layout
= ALGORITHM_PARITY_N
;
1317 /* We can only get to RAID4 from RAID5 */
1318 if (re
->level
!= 4 && re
->level
!= 5)
1319 return "Cannot convert to RAID4 from this level";
1321 switch (re
->level
) {
1323 re
->before
.layout
= 0;
1324 re
->after
.layout
= 0;
1327 re
->after
.layout
= ALGORITHM_PARITY_N
;
1333 /* We get to RAID5 from RAID5 or RAID6 */
1334 if (re
->level
!= 5 && re
->level
!= 6)
1335 return "Cannot convert to RAID5 from this level";
1337 switch (re
->level
) {
1339 if (info
->new_layout
== UnSet
)
1340 re
->after
.layout
= re
->before
.layout
;
1342 re
->after
.layout
= info
->new_layout
;
1345 if (info
->new_layout
== UnSet
)
1346 info
->new_layout
= re
->before
.layout
;
1348 /* after.layout needs to be raid6 version of new_layout */
1349 if (info
->new_layout
== ALGORITHM_PARITY_N
)
1350 re
->after
.layout
= ALGORITHM_PARITY_N
;
1353 char *ls
= map_num(r5layout
, info
->new_layout
);
1356 /* Current RAID6 layout has a RAID5
1359 strcat(strcpy(layout
, ls
), "-6");
1360 l
= map_name(r6layout
, layout
);
1362 return "Cannot find RAID6 layout"
1365 /* Current RAID6 has no equivalent.
1366 * If it is already a '-6' layout we
1367 * can leave it unchanged, else we must
1370 ls
= map_num(r6layout
, info
->new_layout
);
1372 strcmp(ls
+strlen(ls
)-2, "-6") != 0)
1373 return "Please specify new layout";
1374 l
= info
->new_layout
;
1376 re
->after
.layout
= l
;
1382 /* We must already be at level 6 */
1384 return "Impossible level change";
1385 if (info
->new_layout
== UnSet
)
1386 re
->after
.layout
= info
->array
.layout
;
1388 re
->after
.layout
= info
->new_layout
;
1391 return "Impossible level change requested";
1393 if (info
->delta_disks
== UnSet
)
1394 info
->delta_disks
= delta_parity
;
1396 re
->after
.data_disks
= (re
->before
.data_disks
1399 switch (re
->level
) {
1400 case 6: re
->parity
= 2;
1403 case 5: re
->parity
= 1;
1405 default: re
->parity
= 0;
1408 /* So we have a restripe operation, we need to calculate the number
1409 * of blocks per reshape operation.
1411 re
->new_size
= info
->component_size
* re
->before
.data_disks
;
1412 if (info
->new_chunk
== 0)
1413 info
->new_chunk
= info
->array
.chunk_size
;
1414 if (re
->after
.data_disks
== re
->before
.data_disks
&&
1415 re
->after
.layout
== re
->before
.layout
&&
1416 info
->new_chunk
== info
->array
.chunk_size
) {
1417 /* Nothing to change, can change level immediately. */
1418 re
->level
= info
->new_level
;
1419 re
->backup_blocks
= 0;
1422 if (re
->after
.data_disks
== 1 && re
->before
.data_disks
== 1) {
1423 /* chunk and layout changes make no difference */
1424 re
->level
= info
->new_level
;
1425 re
->backup_blocks
= 0;
1429 if (re
->after
.data_disks
== re
->before
.data_disks
&&
1430 get_linux_version() < 2006032)
1431 return "in-place reshape is not safe before 2.6.32 - sorry.";
1433 if (re
->after
.data_disks
< re
->before
.data_disks
&&
1434 get_linux_version() < 2006030)
1435 return "reshape to fewer devices is not supported before 2.6.30 - sorry.";
1437 re
->backup_blocks
= compute_backup_blocks(
1438 info
->new_chunk
, info
->array
.chunk_size
,
1439 re
->after
.data_disks
,
1440 re
->before
.data_disks
);
1441 re
->min_offset_change
= re
->backup_blocks
/ re
->before
.data_disks
;
1443 re
->new_size
= info
->component_size
* re
->after
.data_disks
;
1447 static int set_array_size(struct supertype
*st
, struct mdinfo
*sra
,
1450 struct mdinfo
*info
;
1454 if ((st
== NULL
) || (sra
== NULL
))
1457 if (text_version
== NULL
)
1458 text_version
= sra
->text_version
;
1459 subarray
= strchr(text_version
+1, '/')+1;
1460 info
= st
->ss
->container_content(st
, subarray
);
1462 unsigned long long current_size
= 0;
1463 unsigned long long new_size
=
1464 info
->custom_array_size
/2;
1466 if (sysfs_get_ll(sra
, NULL
, "array_size", ¤t_size
) == 0 &&
1467 new_size
> current_size
) {
1468 if (sysfs_set_num(sra
, NULL
, "array_size", new_size
)
1470 dprintf("Error: Cannot set array size");
1473 dprintf("Array size changed");
1475 dprintf(" from %llu to %llu.\n",
1476 current_size
, new_size
);
1480 dprintf("Error: set_array_size(): info pointer in NULL\n");
1485 static int reshape_array(char *container
, int fd
, char *devname
,
1486 struct supertype
*st
, struct mdinfo
*info
,
1487 int force
, struct mddev_dev
*devlist
,
1488 unsigned long long data_offset
,
1489 char *backup_file
, int verbose
, int forked
,
1490 int restart
, int freeze_reshape
);
1491 static int reshape_container(char *container
, char *devname
,
1493 struct supertype
*st
,
1494 struct mdinfo
*info
,
1497 int verbose
, int restart
, int freeze_reshape
);
1499 int Grow_reshape(char *devname
, int fd
,
1500 struct mddev_dev
*devlist
,
1501 unsigned long long data_offset
,
1502 struct context
*c
, struct shape
*s
)
1504 /* Make some changes in the shape of an array.
1505 * The kernel must support the change.
1507 * There are three different changes. Each can trigger
1508 * a resync or recovery so we freeze that until we have
1509 * requested everything (if kernel supports freezing - 2.6.30).
1511 * - change size (i.e. component_size)
1513 * - change layout/chunksize/ndisks
1515 * The last can require a reshape. It is different on different
1516 * levels so we need to check the level before actioning it.
1517 * Some times the level change needs to be requested after the
1518 * reshape (e.g. raid6->raid5, raid5->raid0)
1521 struct mdu_array_info_s array
;
1523 struct supertype
*st
;
1524 char *subarray
= NULL
;
1528 char *container
= NULL
;
1531 struct mddev_dev
*dv
;
1537 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) < 0) {
1538 pr_err("%s is not an active md array - aborting\n",
1542 if (data_offset
!= INVALID_SECTORS
&& array
.level
!= 10
1543 && (array
.level
< 4 || array
.level
> 6)) {
1544 pr_err("--grow --data-offset not yet supported\n");
1549 (s
->chunk
|| s
->level
!= UnSet
|| s
->layout_str
|| s
->raiddisks
)) {
1550 pr_err("cannot change component size at the same time "
1551 "as other changes.\n"
1552 " Change size first, then check data is intact before "
1553 "making other changes.\n");
1557 if (s
->raiddisks
&& s
->raiddisks
< array
.raid_disks
&& array
.level
> 1 &&
1558 get_linux_version() < 2006032 &&
1559 !check_env("MDADM_FORCE_FEWER")) {
1560 pr_err("reducing the number of devices is not safe before Linux 2.6.32\n"
1561 " Please use a newer kernel\n");
1565 st
= super_by_fd(fd
, &subarray
);
1567 pr_err("Unable to determine metadata format for %s\n", devname
);
1570 if (s
->raiddisks
> st
->max_devs
) {
1571 pr_err("Cannot increase raid-disks on this array"
1572 " beyond %d\n", st
->max_devs
);
1576 /* in the external case we need to check that the requested reshape is
1577 * supported, and perform an initial check that the container holds the
1578 * pre-requisite spare devices (mdmon owns final validation)
1580 if (st
->ss
->external
) {
1584 container
= st
->container_devnm
;
1585 cfd
= open_dev_excl(st
->container_devnm
);
1587 container
= st
->devnm
;
1589 cfd
= open_dev_excl(st
->devnm
);
1593 pr_err("Unable to open container for %s\n",
1599 rv
= st
->ss
->load_container(st
, cfd
, NULL
);
1602 pr_err("Cannot read superblock for %s\n",
1608 /* check if operation is supported for metadata handler */
1609 if (st
->ss
->container_content
) {
1610 struct mdinfo
*cc
= NULL
;
1611 struct mdinfo
*content
= NULL
;
1613 cc
= st
->ss
->container_content(st
, subarray
);
1614 for (content
= cc
; content
; content
= content
->next
) {
1615 int allow_reshape
= 1;
1617 /* check if reshape is allowed based on metadata
1618 * indications stored in content.array.status
1620 if (content
->array
.state
& (1<<MD_SB_BLOCK_VOLUME
))
1622 if (content
->array
.state
1623 & (1<<MD_SB_BLOCK_CONTAINER_RESHAPE
))
1625 if (!allow_reshape
) {
1626 pr_err("cannot reshape arrays in"
1627 " container with unsupported"
1628 " metadata: %s(%s)\n",
1629 devname
, container
);
1637 if (mdmon_running(container
))
1638 st
->update_tail
= &st
->updates
;
1642 for (dv
= devlist
; dv
; dv
= dv
->next
)
1644 if (s
->raiddisks
> array
.raid_disks
&&
1645 array
.spare_disks
+added_disks
< (s
->raiddisks
- array
.raid_disks
) &&
1647 pr_err("Need %d spare%s to avoid degraded array,"
1648 " and only have %d.\n"
1649 " Use --force to over-ride this check.\n",
1650 s
->raiddisks
- array
.raid_disks
,
1651 s
->raiddisks
- array
.raid_disks
== 1 ? "" : "s",
1652 array
.spare_disks
+ added_disks
);
1656 sra
= sysfs_read(fd
, NULL
, GET_LEVEL
| GET_DISKS
| GET_DEVS
1657 | GET_STATE
| GET_VERSION
);
1659 if (st
->ss
->external
&& subarray
== NULL
) {
1660 array
.level
= LEVEL_CONTAINER
;
1661 sra
->array
.level
= LEVEL_CONTAINER
;
1664 pr_err("failed to read sysfs parameters for %s\n",
1668 frozen
= freeze(st
);
1670 /* freeze() already spewed the reason */
1673 } else if (frozen
< 0) {
1674 pr_err("%s is performing resync/recovery and cannot"
1675 " be reshaped\n", devname
);
1680 /* ========= set size =============== */
1681 if (s
->size
> 0 && (s
->size
== MAX_SIZE
|| s
->size
!= (unsigned)array
.size
)) {
1682 unsigned long long orig_size
= get_component_size(fd
)/2;
1683 unsigned long long min_csize
;
1685 int raid0_takeover
= 0;
1688 orig_size
= (unsigned) array
.size
;
1690 if (orig_size
== 0) {
1691 pr_err("Cannot set device size in this type of array.\n");
1696 if (reshape_super(st
, s
->size
, UnSet
, UnSet
, 0, 0, UnSet
, NULL
,
1697 devname
, APPLY_METADATA_CHANGES
, c
->verbose
> 0)) {
1702 if (st
->ss
->external
) {
1703 /* metadata can have size limitation
1704 * update size value according to metadata information
1706 struct mdinfo
*sizeinfo
=
1707 st
->ss
->container_content(st
, subarray
);
1709 unsigned long long new_size
=
1710 sizeinfo
->custom_array_size
/2;
1711 int data_disks
= get_data_disks(
1712 sizeinfo
->array
.level
,
1713 sizeinfo
->array
.layout
,
1714 sizeinfo
->array
.raid_disks
);
1715 new_size
/= data_disks
;
1716 dprintf("Metadata size correction from %llu to "
1717 "%llu (%llu)\n", orig_size
, new_size
,
1718 new_size
* data_disks
);
1720 sysfs_free(sizeinfo
);
1724 /* Update the size of each member device in case
1725 * they have been resized. This will never reduce
1726 * below the current used-size. The "size" attribute
1727 * understands '0' to mean 'max'.
1731 for (mdi
= sra
->devs
; mdi
; mdi
= mdi
->next
) {
1732 if (sysfs_set_num(sra
, mdi
, "size",
1733 s
->size
== MAX_SIZE
? 0 : s
->size
) < 0) {
1734 /* Probably kernel refusing to let us
1735 * reduce the size - not an error.
1739 if (array
.not_persistent
== 0 &&
1740 array
.major_version
== 0 &&
1741 get_linux_version() < 3001000) {
1742 /* Dangerous to allow size to exceed 2TB */
1743 unsigned long long csize
;
1744 if (sysfs_get_ll(sra
, mdi
, "size", &csize
) == 0) {
1745 if (csize
>= 2ULL*1024*1024*1024)
1746 csize
= 2ULL*1024*1024*1024;
1747 if ((min_csize
== 0 || (min_csize
1754 pr_err("Cannot set size on "
1755 "array members.\n");
1756 goto size_change_error
;
1758 if (min_csize
&& s
->size
> min_csize
) {
1759 pr_err("Cannot safely make this array "
1760 "use more than 2TB per device on this kernel.\n");
1762 goto size_change_error
;
1764 if (min_csize
&& s
->size
== MAX_SIZE
) {
1765 /* Don't let the kernel choose a size - it will get
1768 pr_err("Limited v0.90 array to "
1769 "2TB per device\n");
1770 s
->size
= min_csize
;
1772 if (st
->ss
->external
) {
1773 if (sra
->array
.level
== 0) {
1774 rv
= sysfs_set_str(sra
, NULL
, "level",
1778 /* get array parametes after takeover
1779 * to chane one parameter at time only
1781 rv
= ioctl(fd
, GET_ARRAY_INFO
, &array
);
1784 /* make sure mdmon is
1785 * aware of the new level */
1786 if (!mdmon_running(st
->container_devnm
))
1787 start_mdmon(st
->container_devnm
);
1788 ping_monitor(container
);
1789 if (mdmon_running(st
->container_devnm
) &&
1790 st
->update_tail
== NULL
)
1791 st
->update_tail
= &st
->updates
;
1794 if (s
->size
== MAX_SIZE
)
1796 array
.size
= s
->size
;
1797 if ((unsigned)array
.size
!= s
->size
) {
1798 /* got truncated to 32bit, write to
1799 * component_size instead
1802 rv
= sysfs_set_num(sra
, NULL
,
1803 "component_size", s
->size
);
1807 rv
= ioctl(fd
, SET_ARRAY_INFO
, &array
);
1809 /* manage array size when it is managed externally
1811 if ((rv
== 0) && st
->ss
->external
)
1812 rv
= set_array_size(st
, sra
, sra
->text_version
);
1815 if (raid0_takeover
) {
1816 /* do not recync non-existing parity,
1817 * we will drop it anyway
1819 sysfs_set_str(sra
, NULL
, "sync_action", "frozen");
1820 /* go back to raid0, drop parity disk
1822 sysfs_set_str(sra
, NULL
, "level", "raid0");
1823 ioctl(fd
, GET_ARRAY_INFO
, &array
);
1830 /* restore metadata */
1831 if (reshape_super(st
, orig_size
, UnSet
, UnSet
, 0, 0,
1832 UnSet
, NULL
, devname
,
1833 ROLLBACK_METADATA_CHANGES
,
1836 pr_err("Cannot set device size for %s: %s\n",
1837 devname
, strerror(err
));
1839 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
1840 cont_err("Bitmap must be removed before size can be changed\n");
1844 if (s
->assume_clean
) {
1845 /* This will fail on kernels older than 3.0 unless
1846 * a backport has been arranged.
1849 sysfs_set_str(sra
, NULL
, "resync_start", "none") < 0)
1850 pr_err("--assume-clean not supported with --grow on this kernel\n");
1852 ioctl(fd
, GET_ARRAY_INFO
, &array
);
1853 s
->size
= get_component_size(fd
)/2;
1855 s
->size
= array
.size
;
1856 if (c
->verbose
>= 0) {
1857 if (s
->size
== orig_size
)
1858 pr_err("component size of %s "
1859 "unchanged at %lluK\n",
1862 pr_err("component size of %s "
1863 "has been set to %lluK\n",
1867 } else if (array
.level
!= LEVEL_CONTAINER
) {
1868 s
->size
= get_component_size(fd
)/2;
1870 s
->size
= array
.size
;
1873 /* See if there is anything else to do */
1874 if ((s
->level
== UnSet
|| s
->level
== array
.level
) &&
1875 (s
->layout_str
== NULL
) &&
1876 (s
->chunk
== 0 || s
->chunk
== array
.chunk_size
) &&
1877 data_offset
== INVALID_SECTORS
&&
1878 (s
->raiddisks
== 0 || s
->raiddisks
== array
.raid_disks
)) {
1879 /* Nothing more to do */
1880 if (!changed
&& c
->verbose
>= 0)
1881 pr_err("%s: no change requested\n",
1886 /* ========= check for Raid10/Raid1 -> Raid0 conversion ===============
1887 * current implementation assumes that following conditions must be met:
1890 * - near_copies == 2
1892 if ((s
->level
== 0 && array
.level
== 10 && sra
&&
1893 array
.layout
== ((1 << 8) + 2) && !(array
.raid_disks
& 1)) ||
1894 (s
->level
== 0 && array
.level
== 1 && sra
)) {
1896 err
= remove_disks_for_takeover(st
, sra
, array
.layout
);
1898 dprintf(Name
": Array cannot be reshaped\n");
1904 /* Make sure mdmon has seen the device removal
1905 * and updated metadata before we continue with
1909 ping_monitor(container
);
1912 memset(&info
, 0, sizeof(info
));
1914 sysfs_init(&info
, fd
, NULL
);
1915 strcpy(info
.text_version
, sra
->text_version
);
1916 info
.component_size
= s
->size
*2;
1917 info
.new_level
= s
->level
;
1918 info
.new_chunk
= s
->chunk
* 1024;
1919 if (info
.array
.level
== LEVEL_CONTAINER
) {
1920 info
.delta_disks
= UnSet
;
1921 info
.array
.raid_disks
= s
->raiddisks
;
1922 } else if (s
->raiddisks
)
1923 info
.delta_disks
= s
->raiddisks
- info
.array
.raid_disks
;
1925 info
.delta_disks
= UnSet
;
1926 if (s
->layout_str
== NULL
) {
1927 info
.new_layout
= UnSet
;
1928 if (info
.array
.level
== 6 &&
1929 (info
.new_level
== 6 || info
.new_level
== UnSet
) &&
1930 info
.array
.layout
>= 16) {
1931 pr_err("%s has a non-standard layout. If you"
1932 " wish to preserve this\n", devname
);
1933 cont_err("during the reshape, please specify"
1934 " --layout=preserve\n");
1935 cont_err("If you want to change it, specify a"
1936 " layout or use --layout=normalise\n");
1940 } else if (strcmp(s
->layout_str
, "normalise") == 0 ||
1941 strcmp(s
->layout_str
, "normalize") == 0) {
1942 /* If we have a -6 RAID6 layout, remove the '-6'. */
1943 info
.new_layout
= UnSet
;
1944 if (info
.array
.level
== 6 && info
.new_level
== UnSet
) {
1946 strcpy(l
, map_num(r6layout
, info
.array
.layout
));
1947 h
= strrchr(l
, '-');
1948 if (h
&& strcmp(h
, "-6") == 0) {
1950 info
.new_layout
= map_name(r6layout
, l
);
1953 pr_err("%s is only meaningful when reshaping"
1954 " a RAID6 array.\n", s
->layout_str
);
1958 } else if (strcmp(s
->layout_str
, "preserve") == 0) {
1959 /* This means that a non-standard RAID6 layout
1962 * - When reshape a RAID6 (e.g. adding a device)
1963 * which is in a non-standard layout, it is OK
1964 * to preserve that layout.
1965 * - When converting a RAID5 to RAID6, leave it in
1966 * the XXX-6 layout, don't re-layout.
1968 if (info
.array
.level
== 6 && info
.new_level
== UnSet
)
1969 info
.new_layout
= info
.array
.layout
;
1970 else if (info
.array
.level
== 5 && info
.new_level
== 6) {
1972 strcpy(l
, map_num(r5layout
, info
.array
.layout
));
1974 info
.new_layout
= map_name(r6layout
, l
);
1976 pr_err("%s in only meaningful when reshaping"
1977 " to RAID6\n", s
->layout_str
);
1982 int l
= info
.new_level
;
1984 l
= info
.array
.level
;
1987 info
.new_layout
= map_name(r5layout
, s
->layout_str
);
1990 info
.new_layout
= map_name(r6layout
, s
->layout_str
);
1993 info
.new_layout
= parse_layout_10(s
->layout_str
);
1996 info
.new_layout
= parse_layout_faulty(s
->layout_str
);
1999 pr_err("layout not meaningful"
2000 " with this level\n");
2004 if (info
.new_layout
== UnSet
) {
2005 pr_err("layout %s not understood"
2006 " for this level\n",
2013 if (array
.level
== LEVEL_FAULTY
) {
2014 if (s
->level
!= UnSet
&& s
->level
!= array
.level
) {
2015 pr_err("cannot change level of Faulty device\n");
2019 pr_err("cannot set chunksize of Faulty device\n");
2022 if (s
->raiddisks
&& s
->raiddisks
!= 1) {
2023 pr_err("cannot set raid_disks of Faulty device\n");
2026 if (s
->layout_str
) {
2027 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
2028 dprintf("Cannot get array information.\n");
2031 array
.layout
= info
.new_layout
;
2032 if (ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
2033 pr_err("failed to set new layout\n");
2035 } else if (c
->verbose
>= 0)
2036 printf("layout for %s set to %d\n",
2037 devname
, array
.layout
);
2039 } else if (array
.level
== LEVEL_CONTAINER
) {
2040 /* This change is to be applied to every array in the
2041 * container. This is only needed when the metadata imposes
2042 * restraints of the various arrays in the container.
2043 * Currently we only know that IMSM requires all arrays
2044 * to have the same number of devices so changing the
2045 * number of devices (On-Line Capacity Expansion) must be
2046 * performed at the level of the container
2048 rv
= reshape_container(container
, devname
, -1, st
, &info
,
2049 c
->force
, c
->backup_file
, c
->verbose
, 0, 0);
2052 /* get spare devices from external metadata
2054 if (st
->ss
->external
) {
2055 struct mdinfo
*info2
;
2057 info2
= st
->ss
->container_content(st
, subarray
);
2059 info
.array
.spare_disks
=
2060 info2
->array
.spare_disks
;
2065 /* Impose these changes on a single array. First
2066 * check that the metadata is OK with the change. */
2068 if (reshape_super(st
, 0, info
.new_level
,
2069 info
.new_layout
, info
.new_chunk
,
2070 info
.array
.raid_disks
, info
.delta_disks
,
2071 c
->backup_file
, devname
, APPLY_METADATA_CHANGES
,
2077 rv
= reshape_array(container
, fd
, devname
, st
, &info
, c
->force
,
2078 devlist
, data_offset
, c
->backup_file
, c
->verbose
,
2089 /* verify_reshape_position()
2090 * Function checks if reshape position in metadata is not farther
2091 * than position in md.
2093 * 0 : not valid sysfs entry
2094 * it can be caused by not started reshape, it should be started
2095 * by reshape array or raid0 array is before takeover
2096 * -1 : error, reshape position is obviously wrong
2097 * 1 : success, reshape progress correct or updated
2099 static int verify_reshape_position(struct mdinfo
*info
, int level
)
2105 /* read sync_max, failure can mean raid0 array */
2106 rv
= sysfs_get_str(info
, NULL
, "sync_max", buf
, 40);
2110 unsigned long long position
= strtoull(buf
, &ep
, 0);
2112 dprintf(Name
": Read sync_max sysfs entry is: %s\n", buf
);
2113 if (!(ep
== buf
|| (*ep
!= 0 && *ep
!= '\n' && *ep
!= ' '))) {
2114 position
*= get_data_disks(level
,
2116 info
->array
.raid_disks
);
2117 if (info
->reshape_progress
< position
) {
2118 dprintf("Corrected reshape progress (%llu) to "
2119 "md position (%llu)\n",
2120 info
->reshape_progress
, position
);
2121 info
->reshape_progress
= position
;
2123 } else if (info
->reshape_progress
> position
) {
2124 pr_err("Fatal error: array "
2125 "reshape was not properly frozen "
2126 "(expected reshape position is %llu, "
2127 "but reshape progress is %llu.\n",
2128 position
, info
->reshape_progress
);
2131 dprintf("Reshape position in md and metadata "
2136 } else if (rv
== 0) {
2137 /* for valid sysfs entry, 0-length content
2138 * should be indicated as error
2146 static unsigned long long choose_offset(unsigned long long lo
,
2147 unsigned long long hi
,
2148 unsigned long long min
,
2149 unsigned long long max
)
2151 /* Choose a new offset between hi and lo.
2152 * It must be between min and max, but
2153 * we would prefer something near the middle of hi/lo, and also
2154 * prefer to be aligned to a big power of 2.
2156 * So we start with the middle, then for each bit,
2157 * starting at '1' and increasing, if it is set, we either
2158 * add it or subtract it if possible, preferring the option
2159 * which is furthest from the boundary.
2161 * We stop once we get a 1MB alignment. As units are in sectors,
2162 * 1MB = 2*1024 sectors.
2164 unsigned long long choice
= (lo
+ hi
) / 2;
2165 unsigned long long bit
= 1;
2167 for (bit
= 1; bit
< 2*1024; bit
= bit
<< 1) {
2168 unsigned long long bigger
, smaller
;
2169 if (! (bit
& choice
))
2171 bigger
= choice
+ bit
;
2172 smaller
= choice
- bit
;
2173 if (bigger
> max
&& smaller
< min
)
2177 else if (smaller
< min
)
2179 else if (hi
- bigger
> smaller
- lo
)
2187 static int set_new_data_offset(struct mdinfo
*sra
, struct supertype
*st
,
2188 char *devname
, int delta_disks
,
2189 unsigned long long data_offset
,
2190 unsigned long long min
,
2196 unsigned long long before
, after
;
2198 /* Need to find min space before and after so same is used
2201 before
= UINT64_MAX
;
2203 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
2207 struct supertype
*st2
;
2208 struct mdinfo info2
;
2210 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
2212 dn
= map_dev(sd
->disk
.major
, sd
->disk
.minor
, 0);
2213 dfd
= dev_open(dn
, O_RDONLY
);
2215 pr_err("%s: cannot open component %s\n",
2216 devname
, dn
? dn
: "-unknown-");
2219 st2
= dup_super(st
);
2220 rv
= st2
->ss
->load_super(st2
,dfd
, NULL
);
2224 pr_err("%s: cannot get superblock from %s\n",
2228 st2
->ss
->getinfo_super(st2
, &info2
, NULL
);
2229 st2
->ss
->free_super(st2
);
2231 if (info2
.space_before
== 0 &&
2232 info2
.space_after
== 0) {
2233 /* Metadata doesn't support data_offset changes */
2236 if (before
> info2
.space_before
)
2237 before
= info2
.space_before
;
2238 if (after
> info2
.space_after
)
2239 after
= info2
.space_after
;
2241 if (data_offset
!= INVALID_SECTORS
) {
2243 if (info2
.data_offset
== data_offset
) {
2244 pr_err("%s: already has that data_offset\n",
2248 if (data_offset
< info2
.data_offset
)
2252 } else if ((data_offset
<= info2
.data_offset
&& dir
== 1) ||
2253 (data_offset
>= info2
.data_offset
&& dir
== -1)) {
2254 pr_err("%s: differing data offsets on devices make this --data-offset setting impossible\n",
2260 if (before
== UINT64_MAX
)
2261 /* impossible really, there must be no devices */
2264 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
2265 char *dn
= map_dev(sd
->disk
.major
, sd
->disk
.minor
, 0);
2266 unsigned long long new_data_offset
;
2268 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
2270 if (delta_disks
< 0) {
2271 /* Don't need any space as array is shrinking
2272 * just move data_offset up by min
2274 if (data_offset
== INVALID_SECTORS
)
2275 new_data_offset
= sd
->data_offset
+ min
;
2277 if (data_offset
< sd
->data_offset
+ min
) {
2278 pr_err("--data-offset too small for %s\n",
2282 new_data_offset
= data_offset
;
2284 } else if (delta_disks
> 0) {
2285 /* need space before */
2289 pr_err("Insufficient head-space for reshape on %s\n",
2293 if (data_offset
== INVALID_SECTORS
)
2294 new_data_offset
= sd
->data_offset
- min
;
2296 if (data_offset
> sd
->data_offset
- min
) {
2297 pr_err("--data-offset too large for %s\n",
2301 new_data_offset
= data_offset
;
2305 /* can move up or down. If 'data_offset'
2306 * was set we would have already decided,
2307 * so just choose direction with most space.
2314 sysfs_set_str(sra
, NULL
, "reshape_direction",
2315 dir
== 1 ? "backwards" : "forwards");
2317 /* Increase data offset */
2321 pr_err("Insufficient tail-space for reshape on %s\n",
2325 if (data_offset
!= INVALID_SECTORS
&&
2326 data_offset
< sd
->data_offset
+ min
) {
2327 pr_err("--data-offset too small on %s\n",
2331 if (data_offset
!= INVALID_SECTORS
)
2332 new_data_offset
= data_offset
;
2334 new_data_offset
= choose_offset(sd
->data_offset
,
2335 sd
->data_offset
+ after
,
2336 sd
->data_offset
+ min
,
2337 sd
->data_offset
+ after
);
2339 /* Decrease data offset */
2343 pr_err("insufficient head-room on %s\n",
2347 if (data_offset
!= INVALID_SECTORS
&&
2348 data_offset
< sd
->data_offset
- min
) {
2349 pr_err("--data-offset too small on %s\n",
2353 if (data_offset
!= INVALID_SECTORS
)
2354 new_data_offset
= data_offset
;
2356 new_data_offset
= choose_offset(sd
->data_offset
- before
,
2358 sd
->data_offset
- before
,
2359 sd
->data_offset
- min
);
2362 err
= sysfs_set_num(sra
, sd
, "new_offset", new_data_offset
);
2363 if (err
< 0 && errno
== E2BIG
) {
2364 /* try again after increasing data size to max */
2365 err
= sysfs_set_num(sra
, sd
, "size", 0);
2366 if (err
< 0 && errno
== EINVAL
&&
2367 !(sd
->disk
.state
& (1<<MD_DISK_SYNC
))) {
2368 /* some kernels have a bug where you cannot
2369 * use '0' on spare devices. */
2370 sysfs_set_num(sra
, sd
, "size",
2371 (sra
->component_size
+ after
)/2);
2373 err
= sysfs_set_num(sra
, sd
, "new_offset",
2377 if (errno
== E2BIG
&& data_offset
!= INVALID_SECTORS
) {
2378 pr_err("data-offset is too big for %s\n",
2382 if (sd
== sra
->devs
&&
2383 (errno
== ENOENT
|| errno
== E2BIG
))
2384 /* Early kernel, no 'new_offset' file,
2385 * or kernel doesn't like us.
2386 * For RAID5/6 this is not fatal
2389 pr_err("Cannot set new_offset for %s\n",
2398 /* Just use a backup file */
2402 static int raid10_reshape(char *container
, int fd
, char *devname
,
2403 struct supertype
*st
, struct mdinfo
*info
,
2404 struct reshape
*reshape
,
2405 unsigned long long data_offset
,
2406 int force
, int verbose
)
2408 /* Changing raid_disks, layout, chunksize or possibly
2409 * just data_offset for a RAID10.
2410 * We must always change data_offset. We change by at least
2411 * ->min_offset_change which is the largest of the old and new
2413 * If raid_disks is increasing, then data_offset must decrease
2414 * by at least this copy size.
2415 * If raid_disks is unchanged, data_offset must increase or
2416 * decrease by at least min_offset_change but preferably by much more.
2417 * We choose half of the available space.
2418 * If raid_disks is decreasing, data_offset must increase by
2419 * at least min_offset_change. To allow of this, component_size
2420 * must be decreased by the same amount.
2422 * So we calculate the required minimum and direction, possibly
2423 * reduce the component_size, then iterate through the devices
2424 * and set the new_data_offset.
2425 * If that all works, we set chunk_size, layout, raid_disks, and start
2429 unsigned long long min
;
2432 sra
= sysfs_read(fd
, NULL
,
2433 GET_COMPONENT
|GET_DEVS
|GET_OFFSET
|GET_STATE
|GET_CHUNK
2436 pr_err("%s: Cannot get array details from sysfs\n",
2440 min
= reshape
->min_offset_change
;
2442 if (info
->delta_disks
)
2443 sysfs_set_str(sra
, NULL
, "reshape_direction",
2444 info
->delta_disks
< 0 ? "backwards" : "forwards");
2445 if (info
->delta_disks
< 0 &&
2446 info
->space_after
< min
) {
2447 int rv
= sysfs_set_num(sra
, NULL
, "component_size",
2448 (sra
->component_size
-
2451 pr_err("cannot reduce component size\n");
2455 err
= set_new_data_offset(sra
, st
, devname
, info
->delta_disks
, data_offset
,
2458 pr_err("Cannot set new_data_offset: RAID10 reshape not\n");
2459 cont_err("supported on this kernel\n");
2465 if (!err
&& sysfs_set_num(sra
, NULL
, "chunk_size", info
->new_chunk
) < 0)
2467 if (!err
&& sysfs_set_num(sra
, NULL
, "layout", reshape
->after
.layout
) < 0)
2469 if (!err
&& sysfs_set_num(sra
, NULL
, "raid_disks",
2470 info
->array
.raid_disks
+ info
->delta_disks
) < 0)
2472 if (!err
&& sysfs_set_str(sra
, NULL
, "sync_action", "reshape") < 0)
2475 pr_err("Cannot set array shape for %s\n",
2478 (info
->array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
2479 cont_err(" Bitmap must be removed before"
2480 " shape can be changed\n");
2490 static void get_space_after(int fd
, struct supertype
*st
, struct mdinfo
*info
)
2492 struct mdinfo
*sra
, *sd
;
2493 /* Initialisation to silence compiler warning */
2494 unsigned long long min_space_before
= 0, min_space_after
= 0;
2497 sra
= sysfs_read(fd
, NULL
, GET_DEVS
);
2500 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
2503 struct supertype
*st2
;
2504 struct mdinfo info2
;
2506 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
2508 dn
= map_dev(sd
->disk
.major
, sd
->disk
.minor
, 0);
2509 dfd
= dev_open(dn
, O_RDONLY
);
2512 st2
= dup_super(st
);
2513 if (st2
->ss
->load_super(st2
,dfd
, NULL
)) {
2519 st2
->ss
->getinfo_super(st2
, &info2
, NULL
);
2520 st2
->ss
->free_super(st2
);
2523 min_space_before
> info2
.space_before
)
2524 min_space_before
= info2
.space_before
;
2526 min_space_after
> info2
.space_after
)
2527 min_space_after
= info2
.space_after
;
2530 if (sd
== NULL
&& !first
) {
2531 info
->space_after
= min_space_after
;
2532 info
->space_before
= min_space_before
;
2537 static void update_cache_size(char *container
, struct mdinfo
*sra
,
2538 struct mdinfo
*info
,
2539 int disks
, unsigned long long blocks
)
2541 /* Check that the internal stripe cache is
2542 * large enough, or it won't work.
2543 * It must hold at least 4 stripes of the larger
2546 unsigned long cache
;
2547 cache
= max(info
->array
.chunk_size
, info
->new_chunk
);
2548 cache
*= 4; /* 4 stripes minimum */
2549 cache
/= 512; /* convert to sectors */
2550 /* make sure there is room for 'blocks' with a bit to spare */
2551 if (cache
< 16 + blocks
/ disks
)
2552 cache
= 16 + blocks
/ disks
;
2553 cache
/= (4096/512); /* Covert from sectors to pages */
2555 if (sra
->cache_size
< cache
)
2556 subarray_set_num(container
, sra
, "stripe_cache_size",
2560 static int impose_reshape(struct mdinfo
*sra
,
2561 struct mdinfo
*info
,
2562 struct supertype
*st
,
2565 char *devname
, char *container
,
2566 struct reshape
*reshape
)
2568 struct mdu_array_info_s array
;
2570 sra
->new_chunk
= info
->new_chunk
;
2573 /* for external metadata checkpoint saved by mdmon can be lost
2574 * or missed /due to e.g. crash/. Check if md is not during
2575 * restart farther than metadata points to.
2576 * If so, this means metadata information is obsolete.
2578 if (st
->ss
->external
)
2579 verify_reshape_position(info
, reshape
->level
);
2580 sra
->reshape_progress
= info
->reshape_progress
;
2582 sra
->reshape_progress
= 0;
2583 if (reshape
->after
.data_disks
< reshape
->before
.data_disks
)
2584 /* start from the end of the new array */
2585 sra
->reshape_progress
= (sra
->component_size
2586 * reshape
->after
.data_disks
);
2589 ioctl(fd
, GET_ARRAY_INFO
, &array
);
2590 if (info
->array
.chunk_size
== info
->new_chunk
&&
2591 reshape
->before
.layout
== reshape
->after
.layout
&&
2592 st
->ss
->external
== 0) {
2593 /* use SET_ARRAY_INFO but only if reshape hasn't started */
2594 array
.raid_disks
= reshape
->after
.data_disks
+ reshape
->parity
;
2596 ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
2599 pr_err("Cannot set device shape for %s: %s\n",
2600 devname
, strerror(errno
));
2603 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
2604 cont_err("Bitmap must be removed before"
2605 " shape can be changed\n");
2609 } else if (!restart
) {
2610 /* set them all just in case some old 'new_*' value
2611 * persists from some earlier problem.
2614 if (sysfs_set_num(sra
, NULL
, "chunk_size", info
->new_chunk
) < 0)
2616 if (!err
&& sysfs_set_num(sra
, NULL
, "layout",
2617 reshape
->after
.layout
) < 0)
2619 if (!err
&& subarray_set_num(container
, sra
, "raid_disks",
2620 reshape
->after
.data_disks
+
2621 reshape
->parity
) < 0)
2624 pr_err("Cannot set device shape for %s\n",
2628 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
2629 cont_err("Bitmap must be removed before"
2630 " shape can be changed\n");
2639 static int impose_level(int fd
, int level
, char *devname
, int verbose
)
2642 struct mdu_array_info_s array
;
2644 sysfs_init(&info
, fd
, NULL
);
2646 ioctl(fd
, GET_ARRAY_INFO
, &array
);
2648 (array
.level
>= 4 && array
.level
<= 6)) {
2649 /* To convert to RAID0 we need to fail and
2650 * remove any non-data devices. */
2653 int data_disks
= array
.raid_disks
- 1;
2654 if (array
.level
== 6)
2656 if (array
.level
== 5 &&
2657 array
.layout
!= ALGORITHM_PARITY_N
)
2659 if (array
.level
== 6 &&
2660 array
.layout
!= ALGORITHM_PARITY_N_6
)
2662 sysfs_set_str(&info
, NULL
,"sync_action", "idle");
2663 /* First remove any spares so no recovery starts */
2664 for (d
= 0, found
= 0;
2665 d
< MAX_DISKS
&& found
< array
.nr_disks
;
2667 mdu_disk_info_t disk
;
2669 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0)
2671 if (disk
.major
== 0 && disk
.minor
== 0)
2674 if ((disk
.state
& (1 << MD_DISK_ACTIVE
))
2675 && disk
.raid_disk
< data_disks
)
2678 ioctl(fd
, HOT_REMOVE_DISK
,
2679 makedev(disk
.major
, disk
.minor
));
2681 /* Now fail anything left */
2682 ioctl(fd
, GET_ARRAY_INFO
, &array
);
2683 for (d
= 0, found
= 0;
2684 d
< MAX_DISKS
&& found
< array
.nr_disks
;
2687 mdu_disk_info_t disk
;
2689 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0)
2691 if (disk
.major
== 0 && disk
.minor
== 0)
2694 if ((disk
.state
& (1 << MD_DISK_ACTIVE
))
2695 && disk
.raid_disk
< data_disks
)
2698 ioctl(fd
, SET_DISK_FAULTY
,
2699 makedev(disk
.major
, disk
.minor
));
2701 while (ioctl(fd
, HOT_REMOVE_DISK
,
2702 makedev(disk
.major
, disk
.minor
)) < 0
2709 c
= map_num(pers
, level
);
2711 int err
= sysfs_set_str(&info
, NULL
, "level", c
);
2714 pr_err("%s: could not set level to %s\n",
2717 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
2718 cont_err("Bitmap must be removed"
2719 " before level can be changed\n");
2723 pr_err("level of %s changed to %s\n",
2729 static int reshape_array(char *container
, int fd
, char *devname
,
2730 struct supertype
*st
, struct mdinfo
*info
,
2731 int force
, struct mddev_dev
*devlist
,
2732 unsigned long long data_offset
,
2733 char *backup_file
, int verbose
, int forked
,
2734 int restart
, int freeze_reshape
)
2736 struct reshape reshape
;
2739 int orig_level
= UnSet
;
2743 struct mdu_array_info_s array
;
2746 struct mddev_dev
*dv
;
2750 unsigned long long *offsets
= NULL
;
2754 unsigned long blocks
;
2755 unsigned long long array_size
;
2757 struct mdinfo
*sra
= NULL
;
2759 /* when reshaping a RAID0, the component_size might be zero.
2760 * So try to fix that up.
2762 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
2763 dprintf("Cannot get array information.\n");
2766 if (array
.level
== 0 && info
->component_size
== 0) {
2767 get_dev_size(fd
, NULL
, &array_size
);
2768 info
->component_size
= array_size
/ array
.raid_disks
;
2771 if (array
.level
== 10)
2772 /* Need space_after info */
2773 get_space_after(fd
, st
, info
);
2775 if (info
->reshape_active
) {
2776 int new_level
= info
->new_level
;
2777 info
->new_level
= UnSet
;
2778 if (info
->delta_disks
> 0)
2779 info
->array
.raid_disks
-= info
->delta_disks
;
2780 msg
= analyse_change(devname
, info
, &reshape
);
2781 info
->new_level
= new_level
;
2782 if (info
->delta_disks
> 0)
2783 info
->array
.raid_disks
+= info
->delta_disks
;
2785 /* Make sure the array isn't read-only */
2786 ioctl(fd
, RESTART_ARRAY_RW
, 0);
2788 msg
= analyse_change(devname
, info
, &reshape
);
2790 /* if msg == "", error has already been printed */
2792 pr_err("%s\n", msg
);
2796 (reshape
.level
!= info
->array
.level
||
2797 reshape
.before
.layout
!= info
->array
.layout
||
2798 reshape
.before
.data_disks
+ reshape
.parity
2799 != info
->array
.raid_disks
- max(0, info
->delta_disks
))) {
2800 pr_err("reshape info is not in native format -"
2801 " cannot continue.\n");
2805 if (st
->ss
->external
&& restart
&& (info
->reshape_progress
== 0)) {
2806 /* When reshape is restarted from '0', very begin of array
2807 * it is possible that for external metadata reshape and array
2808 * configuration doesn't happen.
2809 * Check if md has the same opinion, and reshape is restarted
2810 * from 0. If so, this is regular reshape start after reshape
2811 * switch in metadata to next array only.
2813 if ((verify_reshape_position(info
, reshape
.level
) >= 0) &&
2814 (info
->reshape_progress
== 0))
2818 /* reshape already started. just skip to monitoring the reshape */
2819 if (reshape
.backup_blocks
== 0)
2821 if (restart
& RESHAPE_NO_BACKUP
)
2824 /* Need 'sra' down at 'started:' */
2825 sra
= sysfs_read(fd
, NULL
,
2826 GET_COMPONENT
|GET_DEVS
|GET_OFFSET
|GET_STATE
|GET_CHUNK
|
2829 pr_err("%s: Cannot get array details from sysfs\n",
2835 /* The container is frozen but the array may not be.
2836 * So freeze the array so spares don't get put to the wrong use
2837 * FIXME there should probably be a cleaner separation between
2838 * freeze_array and freeze_container.
2840 sysfs_freeze_array(info
);
2841 /* Check we have enough spares to not be degraded */
2843 for (dv
= devlist
; dv
; dv
=dv
->next
)
2845 spares_needed
= max(reshape
.before
.data_disks
,
2846 reshape
.after
.data_disks
)
2847 + reshape
.parity
- array
.raid_disks
;
2850 info
->new_level
> 1 && info
->array
.level
> 1 &&
2851 spares_needed
> info
->array
.spare_disks
+ added_disks
) {
2852 pr_err("Need %d spare%s to avoid degraded array,"
2853 " and only have %d.\n"
2854 " Use --force to over-ride this check.\n",
2856 spares_needed
== 1 ? "" : "s",
2857 info
->array
.spare_disks
+ added_disks
);
2860 /* Check we have enough spares to not fail */
2861 spares_needed
= max(reshape
.before
.data_disks
,
2862 reshape
.after
.data_disks
)
2864 if ((info
->new_level
> 1 || info
->new_level
== 0) &&
2865 spares_needed
> info
->array
.spare_disks
+added_disks
) {
2866 pr_err("Need %d spare%s to create working array,"
2867 " and only have %d.\n",
2869 spares_needed
== 1 ? "" : "s",
2870 info
->array
.spare_disks
+ added_disks
);
2874 if (reshape
.level
!= array
.level
) {
2875 int err
= impose_level(fd
, reshape
.level
, devname
, verbose
);
2878 info
->new_layout
= UnSet
; /* after level change,
2879 * layout is meaningless */
2880 orig_level
= array
.level
;
2881 sysfs_freeze_array(info
);
2883 if (reshape
.level
> 0 && st
->ss
->external
) {
2884 /* make sure mdmon is aware of the new level */
2885 if (mdmon_running(container
))
2886 flush_mdmon(container
);
2888 if (!mdmon_running(container
))
2889 start_mdmon(container
);
2890 ping_monitor(container
);
2891 if (mdmon_running(container
) &&
2892 st
->update_tail
== NULL
)
2893 st
->update_tail
= &st
->updates
;
2896 /* ->reshape_super might have chosen some spares from the
2897 * container that it wants to be part of the new array.
2898 * We can collect them with ->container_content and give
2899 * them to the kernel.
2901 if (st
->ss
->reshape_super
&& st
->ss
->container_content
) {
2902 char *subarray
= strchr(info
->text_version
+1, '/')+1;
2903 struct mdinfo
*info2
=
2904 st
->ss
->container_content(st
, subarray
);
2908 sysfs_init(info2
, fd
, st
->devnm
);
2909 /* When increasing number of devices, we need to set
2910 * new raid_disks before adding these, or they might
2913 if (reshape
.backup_blocks
&&
2914 reshape
.after
.data_disks
> reshape
.before
.data_disks
)
2915 subarray_set_num(container
, info2
, "raid_disks",
2916 reshape
.after
.data_disks
+
2918 for (d
= info2
->devs
; d
; d
= d
->next
) {
2919 if (d
->disk
.state
== 0 &&
2920 d
->disk
.raid_disk
>= 0) {
2921 /* This is a spare that wants to
2922 * be part of the array.
2924 add_disk(fd
, st
, info2
, d
);
2930 /* We might have been given some devices to add to the
2931 * array. Now that the array has been changed to the right
2932 * level and frozen, we can safely add them.
2935 Manage_subdevs(devname
, fd
, devlist
, verbose
,
2938 if (reshape
.backup_blocks
== 0 && data_offset
!= INVALID_SECTORS
)
2939 reshape
.backup_blocks
= reshape
.before
.data_disks
* info
->array
.chunk_size
/512;
2940 if (reshape
.backup_blocks
== 0) {
2941 /* No restriping needed, but we might need to impose
2942 * some more changes: layout, raid_disks, chunk_size
2944 /* read current array info */
2945 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
2946 dprintf("Cannot get array information.\n");
2949 /* compare current array info with new values and if
2950 * it is different update them to new */
2951 if (info
->new_layout
!= UnSet
&&
2952 info
->new_layout
!= array
.layout
) {
2953 array
.layout
= info
->new_layout
;
2954 if (ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
2955 pr_err("failed to set new layout\n");
2957 } else if (verbose
>= 0)
2958 printf("layout for %s set to %d\n",
2959 devname
, array
.layout
);
2961 if (info
->delta_disks
!= UnSet
&&
2962 info
->delta_disks
!= 0 &&
2963 array
.raid_disks
!= (info
->array
.raid_disks
+ info
->delta_disks
)) {
2964 array
.raid_disks
+= info
->delta_disks
;
2965 if (ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
2966 pr_err("failed to set raid disks\n");
2968 } else if (verbose
>= 0) {
2969 printf("raid_disks for %s set to %d\n",
2970 devname
, array
.raid_disks
);
2973 if (info
->new_chunk
!= 0 &&
2974 info
->new_chunk
!= array
.chunk_size
) {
2975 if (sysfs_set_num(info
, NULL
,
2976 "chunk_size", info
->new_chunk
) != 0) {
2977 pr_err("failed to set chunk size\n");
2979 } else if (verbose
>= 0)
2980 printf("chunk size for %s set to %d\n",
2981 devname
, array
.chunk_size
);
2988 * There are three possibilities.
2989 * 1/ The array will shrink.
2990 * We need to ensure the reshape will pause before reaching
2991 * the 'critical section'. We also need to fork and wait for
2992 * that to happen. When it does we
2993 * suspend/backup/complete/unfreeze
2995 * 2/ The array will not change size.
2996 * This requires that we keep a backup of a sliding window
2997 * so that we can restore data after a crash. So we need
2998 * to fork and monitor progress.
2999 * In future we will allow the data_offset to change, so
3000 * a sliding backup becomes unnecessary.
3002 * 3/ The array will grow. This is relatively easy.
3003 * However the kernel's restripe routines will cheerfully
3004 * overwrite some early data before it is safe. So we
3005 * need to make a backup of the early parts of the array
3006 * and be ready to restore it if rebuild aborts very early.
3007 * For externally managed metadata, we still need a forked
3008 * child to monitor the reshape and suspend IO over the region
3009 * that is being reshaped.
3011 * We backup data by writing it to one spare, or to a
3012 * file which was given on command line.
3014 * In each case, we first make sure that storage is available
3015 * for the required backup.
3017 * - request the shape change.
3018 * - fork to handle backup etc.
3020 /* Check that we can hold all the data */
3021 get_dev_size(fd
, NULL
, &array_size
);
3022 if (reshape
.new_size
< (array_size
/512)) {
3023 pr_err("this change will reduce the size of the array.\n"
3024 " use --grow --array-size first to truncate array.\n"
3025 " e.g. mdadm --grow %s --array-size %llu\n",
3026 devname
, reshape
.new_size
/2);
3030 if (array
.level
== 10) {
3031 /* Reshaping RAID10 does not require any data backup by
3032 * user-space. Instead it requires that the data_offset
3033 * is changed to avoid the need for backup.
3034 * So this is handled very separately
3037 /* Nothing to do. */
3039 return raid10_reshape(container
, fd
, devname
, st
, info
,
3040 &reshape
, data_offset
,
3043 sra
= sysfs_read(fd
, NULL
,
3044 GET_COMPONENT
|GET_DEVS
|GET_OFFSET
|GET_STATE
|GET_CHUNK
|
3047 pr_err("%s: Cannot get array details from sysfs\n",
3053 switch(set_new_data_offset(sra
, st
, devname
,
3054 reshape
.after
.data_disks
- reshape
.before
.data_disks
,
3056 reshape
.min_offset_change
, 1)) {
3060 /* Updated data_offset, so it's easy now */
3061 update_cache_size(container
, sra
, info
,
3062 min(reshape
.before
.data_disks
,
3063 reshape
.after
.data_disks
),
3064 reshape
.backup_blocks
);
3066 /* Right, everything seems fine. Let's kick things off.
3070 if (impose_reshape(sra
, info
, st
, fd
, restart
,
3071 devname
, container
, &reshape
) < 0)
3073 if (sysfs_set_str(sra
, NULL
, "sync_action", "reshape") < 0) {
3074 pr_err("Failed to initiate reshape!\n");
3077 if (info
->new_level
== reshape
.level
)
3079 /* need to adjust level when reshape completes */
3081 case -1: /* ignore error, but don't wait */
3083 default: /* parent */
3090 impose_level(fd
, info
->new_level
, devname
, verbose
);
3093 case 1: /* Couldn't set data_offset, try the old way */
3094 if (data_offset
!= INVALID_SECTORS
) {
3095 pr_err("Cannot update data_offset on this array\n");
3102 /* Decide how many blocks (sectors) for a reshape
3103 * unit. The number we have so far is just a minimum
3105 blocks
= reshape
.backup_blocks
;
3106 if (reshape
.before
.data_disks
==
3107 reshape
.after
.data_disks
) {
3108 /* Make 'blocks' bigger for better throughput, but
3109 * not so big that we reject it below.
3110 * Try for 16 megabytes
3112 while (blocks
* 32 < sra
->component_size
&&
3116 pr_err("Need to backup %luK of critical "
3117 "section..\n", blocks
/2);
3119 if (blocks
>= sra
->component_size
/2) {
3120 pr_err("%s: Something wrong"
3121 " - reshape aborted\n",
3126 /* Now we need to open all these devices so we can read/write.
3128 nrdisks
= max(reshape
.before
.data_disks
,
3129 reshape
.after
.data_disks
) + reshape
.parity
3130 + sra
->array
.spare_disks
;
3131 fdlist
= xcalloc((1+nrdisks
), sizeof(int));
3132 offsets
= xcalloc((1+nrdisks
), sizeof(offsets
[0]));
3134 odisks
= reshape
.before
.data_disks
+ reshape
.parity
;
3135 d
= reshape_prepare_fdlist(devname
, sra
, odisks
,
3136 nrdisks
, blocks
, backup_file
,
3141 if ((st
->ss
->manage_reshape
== NULL
) ||
3142 (st
->ss
->recover_backup
== NULL
)) {
3143 if (backup_file
== NULL
) {
3144 if (reshape
.after
.data_disks
<=
3145 reshape
.before
.data_disks
) {
3146 pr_err("%s: Cannot grow - "
3147 "need backup-file\n", devname
);
3149 } else if (sra
->array
.spare_disks
== 0) {
3150 pr_err("%s: Cannot grow - "
3151 "need a spare or backup-file to backup "
3152 "critical section\n", devname
);
3156 if (!reshape_open_backup_file(backup_file
, fd
, devname
,
3158 fdlist
+d
, offsets
+d
,
3166 update_cache_size(container
, sra
, info
,
3167 min(reshape
.before
.data_disks
, reshape
.after
.data_disks
),
3170 /* Right, everything seems fine. Let's kick things off.
3171 * If only changing raid_disks, use ioctl, else use
3176 if (impose_reshape(sra
, info
, st
, fd
, restart
,
3177 devname
, container
, &reshape
) < 0)
3180 err
= start_reshape(sra
, restart
, reshape
.before
.data_disks
,
3181 reshape
.after
.data_disks
);
3183 pr_err("Cannot %s reshape for %s\n",
3184 restart
? "continue" : "start",
3189 sysfs_set_str(sra
, NULL
, "array_state", "active");
3190 if (freeze_reshape
) {
3194 pr_err("Reshape has to be continued from"
3195 " location %llu when root filesystem has been mounted.\n",
3196 sra
->reshape_progress
);
3200 /* Now we just need to kick off the reshape and watch, while
3201 * handling backups of the data...
3202 * This is all done by a forked background process.
3204 switch(forked
? 0 : fork()) {
3206 pr_err("Cannot run child to monitor reshape: %s\n",
3220 /* If another array on the same devices is busy, the
3221 * reshape will wait for them. This would mean that
3222 * the first section that we suspend will stay suspended
3223 * for a long time. So check on that possibility
3224 * by looking for "DELAYED" in /proc/mdstat, and if found,
3228 struct mdstat_ent
*mds
, *m
;
3230 mds
= mdstat_read(0, 0);
3231 for (m
= mds
; m
; m
= m
->next
)
3232 if (strcmp(m
->devnm
, sra
->sys_name
) == 0) {
3234 m
->percent
== RESYNC_DELAYED
)
3237 /* Haven't started the reshape thread
3244 if (delayed
== 1 && get_linux_version() < 3007000) {
3245 pr_err("Reshape is delayed, but cannot wait carefully with this kernel.\n"
3246 " You might experience problems until other reshapes complete.\n");
3250 sleep(30 - (delayed
-1) * 25);
3254 if (check_env("MDADM_GROW_VERIFY"))
3255 fd
= open(devname
, O_RDONLY
| O_DIRECT
);
3258 mlockall(MCL_FUTURE
);
3260 if (st
->ss
->external
) {
3261 /* metadata handler takes it from here */
3262 done
= st
->ss
->manage_reshape(
3263 fd
, sra
, &reshape
, st
, blocks
,
3265 d
- odisks
, fdlist
+odisks
,
3268 done
= child_monitor(
3269 fd
, sra
, &reshape
, st
, blocks
,
3271 d
- odisks
, fdlist
+odisks
,
3277 if (backup_file
&& done
)
3278 unlink(backup_file
);
3284 if (!st
->ss
->external
&&
3285 !(reshape
.before
.data_disks
!= reshape
.after
.data_disks
3286 && info
->custom_array_size
) &&
3287 info
->new_level
== reshape
.level
&&
3289 /* no need to wait for the reshape to finish as
3290 * there is nothing more to do.
3297 if (st
->ss
->external
) {
3298 /* Re-load the metadata as much could have changed */
3299 int cfd
= open_dev(st
->container_devnm
);
3301 flush_mdmon(container
);
3302 st
->ss
->free_super(st
);
3303 st
->ss
->load_container(st
, cfd
, container
);
3308 /* set new array size if required customer_array_size is used
3311 if (reshape
.before
.data_disks
!=
3312 reshape
.after
.data_disks
&&
3313 info
->custom_array_size
)
3314 set_array_size(st
, info
, info
->text_version
);
3316 if (info
->new_level
!= reshape
.level
) {
3318 fd
= open(devname
, O_RDONLY
);
3319 impose_level(fd
, info
->new_level
, devname
, verbose
);
3321 if (info
->new_level
== 0)
3322 st
->update_tail
= NULL
;
3334 if (orig_level
!= UnSet
&& sra
) {
3335 c
= map_num(pers
, orig_level
);
3336 if (c
&& sysfs_set_str(sra
, NULL
, "level", c
) == 0)
3337 pr_err("aborting level change\n");
3345 /* mdfd handle is passed to be closed in child process (after fork).
3347 int reshape_container(char *container
, char *devname
,
3349 struct supertype
*st
,
3350 struct mdinfo
*info
,
3353 int verbose
, int restart
, int freeze_reshape
)
3355 struct mdinfo
*cc
= NULL
;
3357 char last_devnm
[32] = "";
3359 /* component_size is not meaningful for a container,
3360 * so pass '0' meaning 'no change'
3363 reshape_super(st
, 0, info
->new_level
,
3364 info
->new_layout
, info
->new_chunk
,
3365 info
->array
.raid_disks
, info
->delta_disks
,
3366 backup_file
, devname
, APPLY_METADATA_CHANGES
,
3374 /* ping monitor to be sure that update is on disk
3376 ping_monitor(container
);
3379 case -1: /* error */
3380 perror("Cannot fork to complete reshape\n");
3383 default: /* parent */
3384 if (!freeze_reshape
)
3385 printf(Name
": multi-array reshape continues"
3386 " in background\n");
3393 /* close unused handle in child process
3399 /* For each member array with reshape_active,
3400 * we need to perform the reshape.
3401 * We pick the first array that needs reshaping and
3402 * reshape it. reshape_array() will re-read the metadata
3403 * so the next time through a different array should be
3404 * ready for reshape.
3405 * It is possible that the 'different' array will not
3406 * be assembled yet. In that case we simple exit.
3407 * When it is assembled, the mdadm which assembles it
3408 * will take over the reshape.
3410 struct mdinfo
*content
;
3412 struct mdstat_ent
*mdstat
;
3418 cc
= st
->ss
->container_content(st
, NULL
);
3420 for (content
= cc
; content
; content
= content
->next
) {
3422 if (!content
->reshape_active
)
3425 subarray
= strchr(content
->text_version
+1, '/')+1;
3426 mdstat
= mdstat_by_subdev(subarray
, container
);
3429 if (mdstat
->active
== 0) {
3430 pr_err("Skipping inactive array %s.\n",
3432 free_mdstat(mdstat
);
3441 devid
= devnm2devid(mdstat
->devnm
);
3442 adev
= map_dev(major(devid
), minor(devid
), 0);
3444 adev
= content
->text_version
;
3446 fd
= open_dev(mdstat
->devnm
);
3448 printf(Name
": Device %s cannot be opened for reshape.",
3453 if (strcmp(last_devnm
, mdstat
->devnm
) == 0) {
3454 /* Do not allow for multiple reshape_array() calls for
3456 * It can happen when reshape_array() returns without
3457 * error, when reshape is not finished (wrong reshape
3458 * starting/continuation conditions). Mdmon doesn't
3459 * switch to next array in container and reentry
3460 * conditions for the same array occur.
3461 * This is possibly interim until the behaviour of
3462 * reshape_array is resolved().
3464 printf(Name
": Multiple reshape execution detected for "
3465 "device %s.", adev
);
3469 strcpy(last_devnm
, mdstat
->devnm
);
3471 sysfs_init(content
, fd
, mdstat
->devnm
);
3473 if (mdmon_running(container
))
3474 flush_mdmon(container
);
3476 rv
= reshape_array(container
, fd
, adev
, st
,
3477 content
, force
, NULL
, 0ULL,
3478 backup_file
, verbose
, 1, restart
,
3482 if (freeze_reshape
) {
3491 if (mdmon_running(container
))
3492 flush_mdmon(container
);
3501 * We run a child process in the background which performs the following
3503 * - wait for resync to reach a certain point
3504 * - suspend io to the following section
3505 * - backup that section
3506 * - allow resync to proceed further
3508 * - discard the backup.
3510 * When are combined in slightly different ways in the three cases.
3512 * - suspend/backup/allow/wait/resume/discard
3514 * - allow/wait/suspend/backup/allow/wait/resume/discard
3516 * - wait/resume/discard/suspend/backup/allow
3518 * suspend/backup/allow always come together
3519 * wait/resume/discard do too.
3520 * For the same-size case we have two backups to improve flow.
3524 int progress_reshape(struct mdinfo
*info
, struct reshape
*reshape
,
3525 unsigned long long backup_point
,
3526 unsigned long long wait_point
,
3527 unsigned long long *suspend_point
,
3528 unsigned long long *reshape_completed
)
3530 /* This function is called repeatedly by the reshape manager.
3531 * It determines how much progress can safely be made and allows
3533 * - 'info' identifies the array and particularly records in
3534 * ->reshape_progress the metadata's knowledge of progress
3535 * This is a sector offset from the start of the array
3536 * of the next array block to be relocated. This number
3537 * may increase from 0 or decrease from array_size, depending
3538 * on the type of reshape that is happening.
3539 * Note that in contrast, 'sync_completed' is a block count of the
3540 * reshape so far. It gives the distance between the start point
3541 * (head or tail of device) and the next place that data will be
3542 * written. It always increases.
3543 * - 'reshape' is the structure created by analyse_change
3544 * - 'backup_point' shows how much the metadata manager has backed-up
3545 * data. For reshapes with increasing progress, it is the next address
3546 * to be backed up, previous addresses have been backed-up. For
3547 * decreasing progress, it is the earliest address that has been
3548 * backed up - later address are also backed up.
3549 * So addresses between reshape_progress and backup_point are
3550 * backed up providing those are in the 'correct' order.
3551 * - 'wait_point' is an array address. When reshape_completed
3552 * passes this point, progress_reshape should return. It might
3553 * return earlier if it determines that ->reshape_progress needs
3554 * to be updated or further backup is needed.
3555 * - suspend_point is maintained by progress_reshape and the caller
3556 * should not touch it except to initialise to zero.
3557 * It is an array address and it only increases in 2.6.37 and earlier.
3558 * This makes it difficult to handle reducing reshapes with
3559 * external metadata.
3560 * However: it is similar to backup_point in that it records the
3561 * other end of a suspended region from reshape_progress.
3562 * it is moved to extend the region that is safe to backup and/or
3564 * - reshape_completed is read from sysfs and returned. The caller
3565 * should copy this into ->reshape_progress when it has reason to
3566 * believe that the metadata knows this, and any backup outside this
3570 * 1 if more data from backup_point - but only as far as suspend_point,
3571 * should be backed up
3572 * 0 if things are progressing smoothly
3573 * -1 if the reshape is finished because it is all done,
3574 * -2 if the reshape is finished due to an error.
3577 int advancing
= (reshape
->after
.data_disks
3578 >= reshape
->before
.data_disks
);
3579 unsigned long long need_backup
; /* All data between start of array and
3580 * here will at some point need to
3583 unsigned long long read_offset
, write_offset
;
3584 unsigned long long write_range
;
3585 unsigned long long max_progress
, target
, completed
;
3586 unsigned long long array_size
= (info
->component_size
3587 * reshape
->before
.data_disks
);
3591 /* First, we unsuspend any region that is now known to be safe.
3592 * If suspend_point is on the 'wrong' side of reshape_progress, then
3593 * we don't have or need suspension at the moment. This is true for
3594 * native metadata when we don't need to back-up.
3597 if (info
->reshape_progress
<= *suspend_point
)
3598 sysfs_set_num(info
, NULL
, "suspend_lo",
3599 info
->reshape_progress
);
3601 /* Note: this won't work in 2.6.37 and before.
3602 * Something somewhere should make sure we don't need it!
3604 if (info
->reshape_progress
>= *suspend_point
)
3605 sysfs_set_num(info
, NULL
, "suspend_hi",
3606 info
->reshape_progress
);
3609 /* Now work out how far it is safe to progress.
3610 * If the read_offset for ->reshape_progress is less than
3611 * 'blocks' beyond the write_offset, we can only progress as far
3613 * Otherwise we can progress until the write_offset for the new location
3614 * reaches (within 'blocks' of) the read_offset at the current location.
3615 * However that region must be suspended unless we are using native
3617 * If we need to suspend more, we limit it to 128M per device, which is
3618 * rather arbitrary and should be some time-based calculation.
3620 read_offset
= info
->reshape_progress
/ reshape
->before
.data_disks
;
3621 write_offset
= info
->reshape_progress
/ reshape
->after
.data_disks
;
3622 write_range
= info
->new_chunk
/512;
3623 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
)
3624 need_backup
= array_size
;
3626 need_backup
= reshape
->backup_blocks
;
3628 if (read_offset
< write_offset
+ write_range
)
3629 max_progress
= backup_point
;
3633 reshape
->after
.data_disks
;
3635 if (read_offset
> write_offset
- write_range
)
3636 /* Can only progress as far as has been backed up,
3637 * which must be suspended */
3638 max_progress
= backup_point
;
3639 else if (info
->reshape_progress
<= need_backup
)
3640 max_progress
= backup_point
;
3642 if (info
->array
.major_version
>= 0)
3643 /* Can progress until backup is needed */
3644 max_progress
= need_backup
;
3646 /* Can progress until metadata update is required */
3649 reshape
->after
.data_disks
;
3650 /* but data must be suspended */
3651 if (max_progress
< *suspend_point
)
3652 max_progress
= *suspend_point
;
3657 /* We know it is safe to progress to 'max_progress' providing
3658 * it is suspended or we are using native metadata.
3659 * Consider extending suspend_point 128M per device if it
3660 * is less than 64M per device beyond reshape_progress.
3661 * But always do a multiple of 'blocks'
3662 * FIXME this is too big - it takes to long to complete
3665 target
= 64*1024*2 * min(reshape
->before
.data_disks
,
3666 reshape
->after
.data_disks
);
3667 target
/= reshape
->backup_blocks
;
3670 target
*= reshape
->backup_blocks
;
3672 /* For externally managed metadata we always need to suspend IO to
3673 * the area being reshaped so we regularly push suspend_point forward.
3674 * For native metadata we only need the suspend if we are going to do
3678 if ((need_backup
> info
->reshape_progress
3679 || info
->array
.major_version
< 0) &&
3680 *suspend_point
< info
->reshape_progress
+ target
) {
3681 if (need_backup
< *suspend_point
+ 2 * target
)
3682 *suspend_point
= need_backup
;
3683 else if (*suspend_point
+ 2 * target
< array_size
)
3684 *suspend_point
+= 2 * target
;
3686 *suspend_point
= array_size
;
3687 sysfs_set_num(info
, NULL
, "suspend_hi", *suspend_point
);
3688 if (max_progress
> *suspend_point
)
3689 max_progress
= *suspend_point
;
3692 if (info
->array
.major_version
>= 0) {
3693 /* Only need to suspend when about to backup */
3694 if (info
->reshape_progress
< need_backup
* 2 &&
3695 *suspend_point
> 0) {
3697 sysfs_set_num(info
, NULL
, "suspend_lo", 0);
3698 sysfs_set_num(info
, NULL
, "suspend_hi", need_backup
);
3701 /* Need to suspend continually */
3702 if (info
->reshape_progress
< *suspend_point
)
3703 *suspend_point
= info
->reshape_progress
;
3704 if (*suspend_point
+ target
< info
->reshape_progress
)
3705 /* No need to move suspend region yet */;
3707 if (*suspend_point
>= 2 * target
)
3708 *suspend_point
-= 2 * target
;
3711 sysfs_set_num(info
, NULL
, "suspend_lo",
3714 if (max_progress
< *suspend_point
)
3715 max_progress
= *suspend_point
;
3719 /* now set sync_max to allow that progress. sync_max, like
3720 * sync_completed is a count of sectors written per device, so
3721 * we find the difference between max_progress and the start point,
3722 * and divide that by after.data_disks to get a sync_max
3724 * At the same time we convert wait_point to a similar number
3725 * for comparing against sync_completed.
3727 /* scale down max_progress to per_disk */
3728 max_progress
/= reshape
->after
.data_disks
;
3729 /* Round to chunk size as some kernels give an erroneously high number */
3730 max_progress
/= info
->new_chunk
/512;
3731 max_progress
*= info
->new_chunk
/512;
3732 /* And round to old chunk size as the kernel wants that */
3733 max_progress
/= info
->array
.chunk_size
/512;
3734 max_progress
*= info
->array
.chunk_size
/512;
3735 /* Limit progress to the whole device */
3736 if (max_progress
> info
->component_size
)
3737 max_progress
= info
->component_size
;
3738 wait_point
/= reshape
->after
.data_disks
;
3740 /* switch from 'device offset' to 'processed block count' */
3741 max_progress
= info
->component_size
- max_progress
;
3742 wait_point
= info
->component_size
- wait_point
;
3745 sysfs_set_num(info
, NULL
, "sync_max", max_progress
);
3747 /* Now wait. If we have already reached the point that we were
3748 * asked to wait to, don't wait at all, else wait for any change.
3749 * We need to select on 'sync_completed' as that is the place that
3750 * notifications happen, but we are really interested in
3751 * 'reshape_position'
3753 fd
= sysfs_get_fd(info
, NULL
, "sync_completed");
3755 goto check_progress
;
3757 if (sysfs_fd_get_ll(fd
, &completed
) < 0)
3758 goto check_progress
;
3760 while (completed
< max_progress
&& completed
< wait_point
) {
3761 /* Check that sync_action is still 'reshape' to avoid
3762 * waiting forever on a dead array
3765 if (sysfs_get_str(info
, NULL
, "sync_action",
3767 strncmp(action
, "reshape", 7) != 0)
3769 /* Some kernels reset 'sync_completed' to zero
3770 * before setting 'sync_action' to 'idle'.
3771 * So we need these extra tests.
3773 if (completed
== 0 && advancing
3774 && info
->reshape_progress
> 0)
3776 if (completed
== 0 && !advancing
3777 && info
->reshape_progress
< (info
->component_size
3778 * reshape
->after
.data_disks
))
3780 sysfs_wait(fd
, NULL
);
3781 if (sysfs_fd_get_ll(fd
, &completed
) < 0)
3782 goto check_progress
;
3784 /* Some kernels reset 'sync_completed' to zero,
3785 * we need to have real point we are in md
3788 completed
= max_progress
;
3790 /* some kernels can give an incorrectly high 'completed' number */
3791 completed
/= (info
->new_chunk
/512);
3792 completed
*= (info
->new_chunk
/512);
3793 /* Convert 'completed' back in to a 'progress' number */
3794 completed
*= reshape
->after
.data_disks
;
3796 completed
= info
->component_size
* reshape
->after
.data_disks
3799 *reshape_completed
= completed
;
3803 /* We return the need_backup flag. Caller will decide
3804 * how much - a multiple of ->backup_blocks up to *suspend_point
3807 return need_backup
> info
->reshape_progress
;
3809 return need_backup
>= info
->reshape_progress
;
3812 /* if we couldn't read a number from sync_completed, then
3813 * either the reshape did complete, or it aborted.
3814 * We can tell which by checking for 'none' in reshape_position.
3815 * If it did abort, then it might immediately restart if it
3816 * it was just a device failure that leaves us degraded but
3820 if (sysfs_get_str(info
, NULL
, "reshape_position", buf
, sizeof(buf
)) < 0
3821 || strncmp(buf
, "none", 4) != 0) {
3822 /* The abort might only be temporary. Wait up to 10
3823 * seconds for fd to contain a valid number again.
3827 while (fd
>= 0 && rv
< 0 && wait
> 0) {
3828 if (sysfs_wait(fd
, &wait
) != 1)
3830 switch (sysfs_fd_get_ll(fd
, &completed
)) {
3832 /* all good again */
3835 case -2: /* read error - abort */
3842 return rv
; /* abort */
3844 /* Maybe racing with array shutdown - check state */
3847 if (sysfs_get_str(info
, NULL
, "array_state", buf
, sizeof(buf
)) < 0
3848 || strncmp(buf
, "inactive", 8) == 0
3849 || strncmp(buf
, "clear",5) == 0)
3850 return -2; /* abort */
3851 return -1; /* complete */
3855 /* FIXME return status is never checked */
3856 static int grow_backup(struct mdinfo
*sra
,
3857 unsigned long long offset
, /* per device */
3858 unsigned long stripes
, /* per device, in old chunks */
3859 int *sources
, unsigned long long *offsets
,
3860 int disks
, int chunk
, int level
, int layout
,
3861 int dests
, int *destfd
, unsigned long long *destoffsets
,
3862 int part
, int *degraded
,
3865 /* Backup 'blocks' sectors at 'offset' on each device of the array,
3866 * to storage 'destfd' (offset 'destoffsets'), after first
3867 * suspending IO. Then allow resync to continue
3868 * over the suspended section.
3869 * Use part 'part' of the backup-super-block.
3874 unsigned long long ll
;
3876 //printf("offset %llu\n", offset);
3882 /* Check that array hasn't become degraded, else we might backup the wrong data */
3883 if (sysfs_get_ll(sra
, NULL
, "degraded", &ll
) < 0)
3884 return -1; /* FIXME this error is ignored */
3885 new_degraded
= (int)ll
;
3886 if (new_degraded
!= *degraded
) {
3887 /* check each device to ensure it is still working */
3889 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
3890 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
3892 if (sd
->disk
.state
& (1<<MD_DISK_SYNC
)) {
3894 if (sysfs_get_str(sra
, sd
, "state", sbuf
, 20) < 0 ||
3895 strstr(sbuf
, "faulty") ||
3896 strstr(sbuf
, "in_sync") == NULL
) {
3897 /* this device is dead */
3898 sd
->disk
.state
= (1<<MD_DISK_FAULTY
);
3899 if (sd
->disk
.raid_disk
>= 0 &&
3900 sources
[sd
->disk
.raid_disk
] >= 0) {
3901 close(sources
[sd
->disk
.raid_disk
]);
3902 sources
[sd
->disk
.raid_disk
] = -1;
3907 *degraded
= new_degraded
;
3910 bsb
.arraystart2
= __cpu_to_le64(offset
* odata
);
3911 bsb
.length2
= __cpu_to_le64(stripes
* (chunk
/512) * odata
);
3913 bsb
.arraystart
= __cpu_to_le64(offset
* odata
);
3914 bsb
.length
= __cpu_to_le64(stripes
* (chunk
/512) * odata
);
3917 bsb
.magic
[15] = '2';
3918 for (i
= 0; i
< dests
; i
++)
3920 lseek64(destfd
[i
], destoffsets
[i
] + __le64_to_cpu(bsb
.devstart2
)*512, 0);
3922 lseek64(destfd
[i
], destoffsets
[i
], 0);
3924 rv
= save_stripes(sources
, offsets
,
3925 disks
, chunk
, level
, layout
,
3927 offset
*512*odata
, stripes
* chunk
* odata
,
3932 bsb
.mtime
= __cpu_to_le64(time(0));
3933 for (i
= 0; i
< dests
; i
++) {
3934 bsb
.devstart
= __cpu_to_le64(destoffsets
[i
]/512);
3936 bsb
.sb_csum
= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
));
3937 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0)
3938 bsb
.sb_csum2
= bsb_csum((char*)&bsb
,
3939 ((char*)&bsb
.sb_csum2
)-((char*)&bsb
));
3942 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
] - 4096, 0)
3943 != destoffsets
[i
] - 4096)
3945 if (write(destfd
[i
], &bsb
, 512) != 512)
3947 if (destoffsets
[i
] > 4096) {
3948 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
]+stripes
*chunk
*odata
, 0) !=
3949 destoffsets
[i
]+stripes
*chunk
*odata
)
3951 if (write(destfd
[i
], &bsb
, 512) != 512)
3961 /* in 2.6.30, the value reported by sync_completed can be
3962 * less that it should be by one stripe.
3963 * This only happens when reshape hits sync_max and pauses.
3964 * So allow wait_backup to either extent sync_max further
3965 * than strictly necessary, or return before the
3966 * sync has got quite as far as we would really like.
3967 * This is what 'blocks2' is for.
3968 * The various caller give appropriate values so that
3971 /* FIXME return value is often ignored */
3972 static int forget_backup(int dests
, int *destfd
,
3973 unsigned long long *destoffsets
,
3977 * Erase backup 'part' (which is 0 or 1)
3983 bsb
.arraystart2
= __cpu_to_le64(0);
3984 bsb
.length2
= __cpu_to_le64(0);
3986 bsb
.arraystart
= __cpu_to_le64(0);
3987 bsb
.length
= __cpu_to_le64(0);
3989 bsb
.mtime
= __cpu_to_le64(time(0));
3991 for (i
= 0; i
< dests
; i
++) {
3992 bsb
.devstart
= __cpu_to_le64(destoffsets
[i
]/512);
3993 bsb
.sb_csum
= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
));
3994 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0)
3995 bsb
.sb_csum2
= bsb_csum((char*)&bsb
,
3996 ((char*)&bsb
.sb_csum2
)-((char*)&bsb
));
3997 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
]-4096, 0) !=
3998 destoffsets
[i
]-4096)
4001 write(destfd
[i
], &bsb
, 512) != 512)
4008 static void fail(char *msg
)
4011 rv
= (write(2, msg
, strlen(msg
)) != (int)strlen(msg
));
4012 rv
|= (write(2, "\n", 1) != 1);
4016 static char *abuf
, *bbuf
;
4017 static unsigned long long abuflen
;
4018 static void validate(int afd
, int bfd
, unsigned long long offset
)
4020 /* check that the data in the backup against the array.
4021 * This is only used for regression testing and should not
4022 * be used while the array is active
4026 lseek64(bfd
, offset
- 4096, 0);
4027 if (read(bfd
, &bsb2
, 512) != 512)
4028 fail("cannot read bsb");
4029 if (bsb2
.sb_csum
!= bsb_csum((char*)&bsb2
,
4030 ((char*)&bsb2
.sb_csum
)-((char*)&bsb2
)))
4031 fail("first csum bad");
4032 if (memcmp(bsb2
.magic
, "md_backup_data", 14) != 0)
4033 fail("magic is bad");
4034 if (memcmp(bsb2
.magic
, "md_backup_data-2", 16) == 0 &&
4035 bsb2
.sb_csum2
!= bsb_csum((char*)&bsb2
,
4036 ((char*)&bsb2
.sb_csum2
)-((char*)&bsb2
)))
4037 fail("second csum bad");
4039 if (__le64_to_cpu(bsb2
.devstart
)*512 != offset
)
4040 fail("devstart is wrong");
4043 unsigned long long len
= __le64_to_cpu(bsb2
.length
)*512;
4045 if (abuflen
< len
) {
4049 if (posix_memalign((void**)&abuf
, 4096, abuflen
) ||
4050 posix_memalign((void**)&bbuf
, 4096, abuflen
)) {
4052 /* just stop validating on mem-alloc failure */
4057 lseek64(bfd
, offset
, 0);
4058 if ((unsigned long long)read(bfd
, bbuf
, len
) != len
) {
4059 //printf("len %llu\n", len);
4060 fail("read first backup failed");
4062 lseek64(afd
, __le64_to_cpu(bsb2
.arraystart
)*512, 0);
4063 if ((unsigned long long)read(afd
, abuf
, len
) != len
)
4064 fail("read first from array failed");
4065 if (memcmp(bbuf
, abuf
, len
) != 0) {
4068 printf("offset=%llu len=%llu\n",
4069 (unsigned long long)__le64_to_cpu(bsb2
.arraystart
)*512, len
);
4070 for (i
=0; i
<len
; i
++)
4071 if (bbuf
[i
] != abuf
[i
]) {
4072 printf("first diff byte %d\n", i
);
4076 fail("data1 compare failed");
4080 unsigned long long len
= __le64_to_cpu(bsb2
.length2
)*512;
4082 if (abuflen
< len
) {
4086 abuf
= xmalloc(abuflen
);
4087 bbuf
= xmalloc(abuflen
);
4090 lseek64(bfd
, offset
+__le64_to_cpu(bsb2
.devstart2
)*512, 0);
4091 if ((unsigned long long)read(bfd
, bbuf
, len
) != len
)
4092 fail("read second backup failed");
4093 lseek64(afd
, __le64_to_cpu(bsb2
.arraystart2
)*512, 0);
4094 if ((unsigned long long)read(afd
, abuf
, len
) != len
)
4095 fail("read second from array failed");
4096 if (memcmp(bbuf
, abuf
, len
) != 0)
4097 fail("data2 compare failed");
4101 int child_monitor(int afd
, struct mdinfo
*sra
, struct reshape
*reshape
,
4102 struct supertype
*st
, unsigned long blocks
,
4103 int *fds
, unsigned long long *offsets
,
4104 int dests
, int *destfd
, unsigned long long *destoffsets
)
4106 /* Monitor a reshape where backup is being performed using
4107 * 'native' mechanism - either to a backup file, or
4108 * to some space in a spare.
4112 unsigned long long speed
;
4113 unsigned long long suspend_point
, array_size
;
4114 unsigned long long backup_point
, wait_point
;
4115 unsigned long long reshape_completed
;
4117 int increasing
= reshape
->after
.data_disks
>= reshape
->before
.data_disks
;
4118 int part
= 0; /* The next part of the backup area to fill. It may already
4119 * be full, so we need to check */
4120 int level
= reshape
->level
;
4121 int layout
= reshape
->before
.layout
;
4122 int data
= reshape
->before
.data_disks
;
4123 int disks
= reshape
->before
.data_disks
+ reshape
->parity
;
4124 int chunk
= sra
->array
.chunk_size
;
4126 unsigned long stripes
;
4129 /* set up the backup-super-block. This requires the
4130 * uuid from the array.
4132 /* Find a superblock */
4133 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
4137 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
4139 dn
= map_dev(sd
->disk
.major
, sd
->disk
.minor
, 1);
4140 devfd
= dev_open(dn
, O_RDONLY
);
4143 ok
= st
->ss
->load_super(st
, devfd
, NULL
);
4149 pr_err("Cannot find a superblock\n");
4153 memset(&bsb
, 0, 512);
4154 memcpy(bsb
.magic
, "md_backup_data-1", 16);
4155 st
->ss
->uuid_from_super(st
, uuid
);
4156 memcpy(bsb
.set_uuid
, uuid
, 16);
4157 bsb
.mtime
= __cpu_to_le64(time(0));
4158 bsb
.devstart2
= blocks
;
4160 stripes
= blocks
/ (sra
->array
.chunk_size
/512) /
4161 reshape
->before
.data_disks
;
4163 if (posix_memalign((void**)&buf
, 4096, disks
* chunk
))
4164 /* Don't start the 'reshape' */
4166 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
) {
4167 sysfs_get_ll(sra
, NULL
, "sync_speed_min", &speed
);
4168 sysfs_set_num(sra
, NULL
, "sync_speed_min", 200000);
4172 array_size
= sra
->component_size
* reshape
->after
.data_disks
;
4173 backup_point
= sra
->reshape_progress
;
4176 array_size
= sra
->component_size
* reshape
->before
.data_disks
;
4177 backup_point
= reshape
->backup_blocks
;
4178 suspend_point
= array_size
;
4184 /* Want to return as soon the oldest backup slot can
4185 * be released as that allows us to start backing up
4186 * some more, providing suspend_point has been
4187 * advanced, which it should have.
4190 wait_point
= array_size
;
4191 if (part
== 0 && __le64_to_cpu(bsb
.length
) > 0)
4192 wait_point
= (__le64_to_cpu(bsb
.arraystart
) +
4193 __le64_to_cpu(bsb
.length
));
4194 if (part
== 1 && __le64_to_cpu(bsb
.length2
) > 0)
4195 wait_point
= (__le64_to_cpu(bsb
.arraystart2
) +
4196 __le64_to_cpu(bsb
.length2
));
4199 if (part
== 0 && __le64_to_cpu(bsb
.length
) > 0)
4200 wait_point
= __le64_to_cpu(bsb
.arraystart
);
4201 if (part
== 1 && __le64_to_cpu(bsb
.length2
) > 0)
4202 wait_point
= __le64_to_cpu(bsb
.arraystart2
);
4205 rv
= progress_reshape(sra
, reshape
,
4206 backup_point
, wait_point
,
4207 &suspend_point
, &reshape_completed
);
4208 /* external metadata would need to ping_monitor here */
4209 sra
->reshape_progress
= reshape_completed
;
4211 /* Clear any backup region that is before 'here' */
4213 if (__le64_to_cpu(bsb
.length
) > 0 &&
4214 reshape_completed
>= (__le64_to_cpu(bsb
.arraystart
) +
4215 __le64_to_cpu(bsb
.length
)))
4216 forget_backup(dests
, destfd
,
4218 if (__le64_to_cpu(bsb
.length2
) > 0 &&
4219 reshape_completed
>= (__le64_to_cpu(bsb
.arraystart2
) +
4220 __le64_to_cpu(bsb
.length2
)))
4221 forget_backup(dests
, destfd
,
4224 if (__le64_to_cpu(bsb
.length
) > 0 &&
4225 reshape_completed
<= (__le64_to_cpu(bsb
.arraystart
)))
4226 forget_backup(dests
, destfd
,
4228 if (__le64_to_cpu(bsb
.length2
) > 0 &&
4229 reshape_completed
<= (__le64_to_cpu(bsb
.arraystart2
)))
4230 forget_backup(dests
, destfd
,
4239 if (rv
== 0 && increasing
&& !st
->ss
->external
) {
4240 /* No longer need to monitor this reshape */
4246 unsigned long long offset
;
4247 unsigned long actual_stripes
;
4248 /* Need to backup some data.
4249 * If 'part' is not used and the desired
4250 * backup size is suspended, do a backup,
4251 * then consider the next part.
4253 /* Check that 'part' is unused */
4254 if (part
== 0 && __le64_to_cpu(bsb
.length
) != 0)
4256 if (part
== 1 && __le64_to_cpu(bsb
.length2
) != 0)
4259 offset
= backup_point
/ data
;
4260 actual_stripes
= stripes
;
4262 if (offset
+ actual_stripes
* (chunk
/512) >
4263 sra
->component_size
)
4264 actual_stripes
= ((sra
->component_size
- offset
)
4266 if (offset
+ actual_stripes
* (chunk
/512) >
4270 if (offset
< actual_stripes
* (chunk
/512))
4271 actual_stripes
= offset
/ (chunk
/512);
4272 offset
-= actual_stripes
* (chunk
/512);
4273 if (offset
< suspend_point
/data
)
4276 if (actual_stripes
== 0)
4278 grow_backup(sra
, offset
, actual_stripes
,
4280 disks
, chunk
, level
, layout
,
4281 dests
, destfd
, destoffsets
,
4282 part
, °raded
, buf
);
4283 validate(afd
, destfd
[0], destoffsets
[0]);
4284 /* record where 'part' is up to */
4287 backup_point
+= actual_stripes
* (chunk
/512) * data
;
4289 backup_point
-= actual_stripes
* (chunk
/512) * data
;
4293 /* FIXME maybe call progress_reshape one more time instead */
4294 abort_reshape(sra
); /* remove any remaining suspension */
4295 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
)
4296 sysfs_set_num(sra
, NULL
, "sync_speed_min", speed
);
4302 * If any spare contains md_back_data-1 which is recent wrt mtime,
4303 * write that data into the array and update the super blocks with
4304 * the new reshape_progress
4306 int Grow_restart(struct supertype
*st
, struct mdinfo
*info
, int *fdlist
, int cnt
,
4307 char *backup_file
, int verbose
)
4311 unsigned long long *offsets
;
4312 unsigned long long nstripe
, ostripe
;
4315 odata
= info
->array
.raid_disks
- info
->delta_disks
- 1;
4316 if (info
->array
.level
== 6) odata
--; /* number of data disks */
4317 ndata
= info
->array
.raid_disks
- 1;
4318 if (info
->new_level
== 6) ndata
--;
4320 old_disks
= info
->array
.raid_disks
- info
->delta_disks
;
4322 if (info
->delta_disks
<= 0)
4323 /* Didn't grow, so the backup file must have
4327 for (i
=old_disks
-(backup_file
?1:0); i
<cnt
; i
++) {
4328 struct mdinfo dinfo
;
4331 char *devname
, namebuf
[20];
4332 unsigned long long lo
, hi
;
4334 /* This was a spare and may have some saved data on it.
4335 * Load the superblock, find and load the
4336 * backup_super_block.
4337 * If either fail, go on to next device.
4338 * If the backup contains no new info, just return
4339 * else restore data and update all superblocks
4341 if (i
== old_disks
-1) {
4342 fd
= open(backup_file
, O_RDONLY
);
4344 pr_err("backup file %s inaccessible: %s\n",
4345 backup_file
, strerror(errno
));
4348 devname
= backup_file
;
4353 if (st
->ss
->load_super(st
, fd
, NULL
))
4356 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
4357 st
->ss
->free_super(st
);
4360 (dinfo
.data_offset
+ dinfo
.component_size
- 8) <<9,
4362 pr_err("Cannot seek on device %d\n", i
);
4363 continue; /* Cannot seek */
4365 sprintf(namebuf
, "device-%d", i
);
4368 if (read(fd
, &bsb
, sizeof(bsb
)) != sizeof(bsb
)) {
4370 pr_err("Cannot read from %s\n", devname
);
4371 continue; /* Cannot read */
4373 if (memcmp(bsb
.magic
, "md_backup_data-1", 16) != 0 &&
4374 memcmp(bsb
.magic
, "md_backup_data-2", 16) != 0) {
4376 pr_err("No backup metadata on %s\n", devname
);
4379 if (bsb
.sb_csum
!= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
))) {
4381 pr_err("Bad backup-metadata checksum on %s\n", devname
);
4382 continue; /* bad checksum */
4384 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0 &&
4385 bsb
.sb_csum2
!= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum2
)-((char*)&bsb
))) {
4387 pr_err("Bad backup-metadata checksum2 on %s\n", devname
);
4388 continue; /* Bad second checksum */
4390 if (memcmp(bsb
.set_uuid
,info
->uuid
, 16) != 0) {
4392 pr_err("Wrong uuid on backup-metadata on %s\n", devname
);
4393 continue; /* Wrong uuid */
4396 /* array utime and backup-mtime should be updated at much the same time, but it seems that
4397 * sometimes they aren't... So allow considerable flexability in matching, and allow
4398 * this test to be overridden by an environment variable.
4400 if (info
->array
.utime
> (int)__le64_to_cpu(bsb
.mtime
) + 2*60*60 ||
4401 info
->array
.utime
< (int)__le64_to_cpu(bsb
.mtime
) - 10*60) {
4402 if (check_env("MDADM_GROW_ALLOW_OLD")) {
4403 pr_err("accepting backup with timestamp %lu "
4404 "for array with timestamp %lu\n",
4405 (unsigned long)__le64_to_cpu(bsb
.mtime
),
4406 (unsigned long)info
->array
.utime
);
4408 pr_err("too-old timestamp on backup-metadata on %s\n", devname
);
4409 pr_err("If you think it is should be safe, try 'export MDADM_GROW_ALLOW_OLD=1'\n");
4410 continue; /* time stamp is too bad */
4414 if (bsb
.magic
[15] == '1') {
4415 if (bsb
.length
== 0)
4417 if (info
->delta_disks
>= 0) {
4418 /* reshape_progress is increasing */
4419 if (__le64_to_cpu(bsb
.arraystart
)
4420 + __le64_to_cpu(bsb
.length
)
4421 < info
->reshape_progress
) {
4424 pr_err("backup-metadata found on %s but is not needed\n", devname
);
4425 continue; /* No new data here */
4428 /* reshape_progress is decreasing */
4429 if (__le64_to_cpu(bsb
.arraystart
) >=
4430 info
->reshape_progress
)
4431 goto nonew
; /* No new data here */
4434 if (bsb
.length
== 0 && bsb
.length2
== 0)
4436 if (info
->delta_disks
>= 0) {
4437 /* reshape_progress is increasing */
4438 if ((__le64_to_cpu(bsb
.arraystart
)
4439 + __le64_to_cpu(bsb
.length
)
4440 < info
->reshape_progress
)
4442 (__le64_to_cpu(bsb
.arraystart2
)
4443 + __le64_to_cpu(bsb
.length2
)
4444 < info
->reshape_progress
))
4445 goto nonew
; /* No new data here */
4447 /* reshape_progress is decreasing */
4448 if (__le64_to_cpu(bsb
.arraystart
) >=
4449 info
->reshape_progress
&&
4450 __le64_to_cpu(bsb
.arraystart2
) >=
4451 info
->reshape_progress
)
4452 goto nonew
; /* No new data here */
4455 if (lseek64(fd
, __le64_to_cpu(bsb
.devstart
)*512, 0)< 0) {
4458 pr_err("Failed to verify secondary backup-metadata block on %s\n",
4460 continue; /* Cannot seek */
4462 /* There should be a duplicate backup superblock 4k before here */
4463 if (lseek64(fd
, -4096, 1) < 0 ||
4464 read(fd
, &bsb2
, sizeof(bsb2
)) != sizeof(bsb2
))
4465 goto second_fail
; /* Cannot find leading superblock */
4466 if (bsb
.magic
[15] == '1')
4467 bsbsize
= offsetof(struct mdp_backup_super
, pad1
);
4469 bsbsize
= offsetof(struct mdp_backup_super
, pad
);
4470 if (memcmp(&bsb2
, &bsb
, bsbsize
) != 0)
4471 goto second_fail
; /* Cannot find leading superblock */
4473 /* Now need the data offsets for all devices. */
4474 offsets
= xmalloc(sizeof(*offsets
)*info
->array
.raid_disks
);
4475 for(j
=0; j
<info
->array
.raid_disks
; j
++) {
4478 if (st
->ss
->load_super(st
, fdlist
[j
], NULL
))
4479 /* FIXME should be this be an error */
4481 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
4482 st
->ss
->free_super(st
);
4483 offsets
[j
] = dinfo
.data_offset
* 512;
4485 printf(Name
": restoring critical section\n");
4487 if (restore_stripes(fdlist
, offsets
,
4488 info
->array
.raid_disks
,
4492 fd
, __le64_to_cpu(bsb
.devstart
)*512,
4493 __le64_to_cpu(bsb
.arraystart
)*512,
4494 __le64_to_cpu(bsb
.length
)*512, NULL
)) {
4495 /* didn't succeed, so giveup */
4497 pr_err("Error restoring backup from %s\n",
4503 if (bsb
.magic
[15] == '2' &&
4504 restore_stripes(fdlist
, offsets
,
4505 info
->array
.raid_disks
,
4509 fd
, __le64_to_cpu(bsb
.devstart
)*512 +
4510 __le64_to_cpu(bsb
.devstart2
)*512,
4511 __le64_to_cpu(bsb
.arraystart2
)*512,
4512 __le64_to_cpu(bsb
.length2
)*512, NULL
)) {
4513 /* didn't succeed, so giveup */
4515 pr_err("Error restoring second backup from %s\n",
4523 /* Ok, so the data is restored. Let's update those superblocks. */
4527 lo
= __le64_to_cpu(bsb
.arraystart
);
4528 hi
= lo
+ __le64_to_cpu(bsb
.length
);
4530 if (bsb
.magic
[15] == '2' && bsb
.length2
) {
4531 unsigned long long lo1
, hi1
;
4532 lo1
= __le64_to_cpu(bsb
.arraystart2
);
4533 hi1
= lo1
+ __le64_to_cpu(bsb
.length2
);
4537 } else if (lo
< lo1
)
4543 (info
->reshape_progress
< lo
||
4544 info
->reshape_progress
> hi
))
4545 /* backup does not affect reshape_progress*/ ;
4546 else if (info
->delta_disks
>= 0) {
4547 info
->reshape_progress
= __le64_to_cpu(bsb
.arraystart
) +
4548 __le64_to_cpu(bsb
.length
);
4549 if (bsb
.magic
[15] == '2') {
4550 unsigned long long p2
= __le64_to_cpu(bsb
.arraystart2
) +
4551 __le64_to_cpu(bsb
.length2
);
4552 if (p2
> info
->reshape_progress
)
4553 info
->reshape_progress
= p2
;
4556 info
->reshape_progress
= __le64_to_cpu(bsb
.arraystart
);
4557 if (bsb
.magic
[15] == '2') {
4558 unsigned long long p2
= __le64_to_cpu(bsb
.arraystart2
);
4559 if (p2
< info
->reshape_progress
)
4560 info
->reshape_progress
= p2
;
4563 for (j
=0; j
<info
->array
.raid_disks
; j
++) {
4566 if (st
->ss
->load_super(st
, fdlist
[j
], NULL
))
4568 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
4569 dinfo
.reshape_progress
= info
->reshape_progress
;
4570 st
->ss
->update_super(st
, &dinfo
,
4571 "_reshape_progress",
4573 st
->ss
->store_super(st
, fdlist
[j
]);
4574 st
->ss
->free_super(st
);
4578 /* Didn't find any backup data, try to see if any
4581 if (info
->delta_disks
< 0) {
4582 /* When shrinking, the critical section is at the end.
4583 * So see if we are before the critical section.
4585 unsigned long long first_block
;
4586 nstripe
= ostripe
= 0;
4588 while (ostripe
>= nstripe
) {
4589 ostripe
+= info
->array
.chunk_size
/ 512;
4590 first_block
= ostripe
* odata
;
4591 nstripe
= first_block
/ ndata
/ (info
->new_chunk
/512) *
4592 (info
->new_chunk
/512);
4595 if (info
->reshape_progress
>= first_block
)
4598 if (info
->delta_disks
> 0) {
4599 /* See if we are beyond the critical section. */
4600 unsigned long long last_block
;
4601 nstripe
= ostripe
= 0;
4603 while (nstripe
>= ostripe
) {
4604 nstripe
+= info
->new_chunk
/ 512;
4605 last_block
= nstripe
* ndata
;
4606 ostripe
= last_block
/ odata
/ (info
->array
.chunk_size
/512) *
4607 (info
->array
.chunk_size
/512);
4610 if (info
->reshape_progress
>= last_block
)
4613 /* needed to recover critical section! */
4615 pr_err("Failed to find backup of critical section\n");
4619 int Grow_continue_command(char *devname
, int fd
,
4620 char *backup_file
, int verbose
)
4623 struct supertype
*st
= NULL
;
4624 struct mdinfo
*content
= NULL
;
4625 struct mdinfo array
;
4626 char *subarray
= NULL
;
4627 struct mdinfo
*cc
= NULL
;
4628 struct mdstat_ent
*mdstat
= NULL
;
4632 dprintf("Grow continue from command line called for %s\n",
4635 st
= super_by_fd(fd
, &subarray
);
4636 if (!st
|| !st
->ss
) {
4637 pr_err("Unable to determine metadata format for %s\n",
4641 dprintf("Grow continue is run for ");
4642 if (st
->ss
->external
== 0) {
4644 dprintf("native array (%s)\n", devname
);
4645 if (ioctl(fd
, GET_ARRAY_INFO
, &array
.array
) < 0) {
4646 pr_err("%s is not an active md array -"
4647 " aborting\n", devname
);
4649 goto Grow_continue_command_exit
;
4652 /* Need to load a superblock.
4653 * FIXME we should really get what we need from
4656 for (d
= 0; d
< MAX_DISKS
; d
++) {
4657 mdu_disk_info_t disk
;
4661 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0)
4663 if (disk
.major
== 0 && disk
.minor
== 0)
4665 if ((disk
.state
& (1 << MD_DISK_ACTIVE
)) == 0)
4667 dv
= map_dev(disk
.major
, disk
.minor
, 1);
4670 fd2
= dev_open(dv
, O_RDONLY
);
4673 err
= st
->ss
->load_super(st
, fd2
, NULL
);
4679 if (d
== MAX_DISKS
) {
4680 pr_err("Unable to load metadata for %s\n",
4683 goto Grow_continue_command_exit
;
4685 st
->ss
->getinfo_super(st
, content
, NULL
);
4690 dprintf("subarray (%s)\n", subarray
);
4691 container
= st
->container_devnm
;
4692 cfd
= open_dev_excl(st
->container_devnm
);
4694 container
= st
->devnm
;
4696 cfd
= open_dev_excl(st
->devnm
);
4697 dprintf("container (%s)\n", container
);
4701 pr_err("Unable to open container "
4702 "for %s\n", devname
);
4704 goto Grow_continue_command_exit
;
4707 /* find in container array under reshape
4709 ret_val
= st
->ss
->load_container(st
, cfd
, NULL
);
4711 pr_err("Cannot read superblock for %s\n",
4714 goto Grow_continue_command_exit
;
4717 cc
= st
->ss
->container_content(st
, subarray
);
4718 for (content
= cc
; content
; content
= content
->next
) {
4720 int allow_reshape
= 1;
4722 if (content
->reshape_active
== 0)
4724 /* The decision about array or container wide
4725 * reshape is taken in Grow_continue based
4726 * content->reshape_active state, therefore we
4727 * need to check_reshape based on
4728 * reshape_active and subarray name
4730 if (content
->array
.state
& (1<<MD_SB_BLOCK_VOLUME
))
4732 if (content
->reshape_active
== CONTAINER_RESHAPE
&&
4733 (content
->array
.state
4734 & (1<<MD_SB_BLOCK_CONTAINER_RESHAPE
)))
4737 if (!allow_reshape
) {
4738 pr_err("cannot continue reshape of an array"
4739 " in container with unsupported"
4740 " metadata: %s(%s)\n",
4741 devname
, container
);
4743 goto Grow_continue_command_exit
;
4746 array
= strchr(content
->text_version
+1, '/')+1;
4747 mdstat
= mdstat_by_subdev(array
, container
);
4750 if (mdstat
->active
== 0) {
4751 pr_err("Skipping inactive array %s.\n",
4753 free_mdstat(mdstat
);
4760 pr_err("Unable to determine reshaped "
4761 "array for %s\n", devname
);
4763 goto Grow_continue_command_exit
;
4765 fd2
= open_dev(mdstat
->devnm
);
4767 pr_err("cannot open (%s)\n", mdstat
->devnm
);
4769 goto Grow_continue_command_exit
;
4772 sysfs_init(content
, fd2
, mdstat
->devnm
);
4774 /* start mdmon in case it is not running
4776 if (!mdmon_running(container
))
4777 start_mdmon(container
);
4778 ping_monitor(container
);
4780 if (mdmon_running(container
))
4781 st
->update_tail
= &st
->updates
;
4783 pr_err("No mdmon found. "
4784 "Grow cannot continue.\n");
4786 goto Grow_continue_command_exit
;
4790 /* verify that array under reshape is started from
4793 if (verify_reshape_position(content
, content
->array
.level
) < 0) {
4795 goto Grow_continue_command_exit
;
4800 ret_val
= Grow_continue(fd
, st
, content
, backup_file
, 0);
4802 Grow_continue_command_exit
:
4807 st
->ss
->free_super(st
);
4808 free_mdstat(mdstat
);
4815 int Grow_continue(int mdfd
, struct supertype
*st
, struct mdinfo
*info
,
4816 char *backup_file
, int freeze_reshape
)
4820 if (!info
->reshape_active
)
4823 if (st
->ss
->external
) {
4824 int cfd
= open_dev(st
->container_devnm
);
4829 st
->ss
->load_container(st
, cfd
, st
->container_devnm
);
4831 ret_val
= reshape_container(st
->container_devnm
, NULL
, mdfd
,
4832 st
, info
, 0, backup_file
,
4834 1 | info
->reshape_active
,
4837 ret_val
= reshape_array(NULL
, mdfd
, "array", st
, info
, 1,
4838 NULL
, 0ULL, backup_file
, 0, 0,
4839 1 | info
->reshape_active
,