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>
28 #if ! defined(__BIG_ENDIAN) && ! defined(__LITTLE_ENDIAN)
29 #error no endian defined
35 #define offsetof(t,f) ((size_t)&(((t*)0)->f))
38 int restore_backup(struct supertype
*st
,
39 struct mdinfo
*content
,
49 int disk_count
= next_spare
+ working_disks
;
51 dprintf("Called restore_backup()\n");
52 fdlist
= xmalloc(sizeof(int) * disk_count
);
54 for (i
= 0; i
< next_spare
; i
++)
56 for (dev
= content
->devs
; dev
; dev
= dev
->next
) {
62 fd
= dev_open(buf
, O_RDWR
);
64 if (dev
->disk
.raid_disk
>= 0)
65 fdlist
[dev
->disk
.raid_disk
] = fd
;
67 fdlist
[next_spare
++] = fd
;
70 if (st
->ss
->external
&& st
->ss
->recover_backup
)
71 err
= st
->ss
->recover_backup(st
, content
);
73 err
= Grow_restart(st
, content
, fdlist
, next_spare
,
74 backup_file
, verbose
> 0);
76 while (next_spare
> 0) {
78 if (fdlist
[next_spare
] >= 0)
79 close(fdlist
[next_spare
]);
83 pr_err("Failed to restore critical"
84 " section for reshape - sorry.\n");
86 pr_err("Possibly you need"
87 " to specify a --backup-file\n");
91 dprintf("restore_backup() returns status OK.\n");
95 int Grow_Add_device(char *devname
, int fd
, char *newdev
)
97 /* Add a device to an active array.
98 * Currently, just extend a linear array.
99 * This requires writing a new superblock on the
100 * new device, calling the kernel to add the device,
101 * and if that succeeds, update the superblock on
103 * This means that we need to *find* all other devices.
110 struct supertype
*st
= NULL
;
111 char *subarray
= NULL
;
113 if (ioctl(fd
, GET_ARRAY_INFO
, &info
.array
) < 0) {
114 pr_err("cannot get array info for %s\n", devname
);
118 if (info
.array
.level
!= -1) {
119 pr_err("can only add devices to linear arrays\n");
123 st
= super_by_fd(fd
, &subarray
);
125 pr_err("cannot handle arrays with superblock version %d\n",
126 info
.array
.major_version
);
131 pr_err("Cannot grow linear sub-arrays yet\n");
137 nfd
= open(newdev
, O_RDWR
|O_EXCL
|O_DIRECT
);
139 pr_err("cannot open %s\n", newdev
);
144 if ((stb
.st_mode
& S_IFMT
) != S_IFBLK
) {
145 pr_err("%s is not a block device!\n", newdev
);
150 /* now check out all the devices and make sure we can read the
152 for (d
=0 ; d
< info
.array
.raid_disks
; d
++) {
153 mdu_disk_info_t disk
;
156 st
->ss
->free_super(st
);
159 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0) {
160 pr_err("cannot get device detail for device %d\n",
166 dv
= map_dev(disk
.major
, disk
.minor
, 1);
168 pr_err("cannot find device file for device %d\n",
174 fd2
= dev_open(dv
, O_RDWR
);
176 pr_err("cannot open device file %s\n", dv
);
182 if (st
->ss
->load_super(st
, fd2
, NULL
)) {
183 pr_err("cannot find super block on %s\n", dv
);
191 /* Ok, looks good. Lets update the superblock and write it out to
195 info
.disk
.number
= d
;
196 info
.disk
.major
= major(stb
.st_rdev
);
197 info
.disk
.minor
= minor(stb
.st_rdev
);
198 info
.disk
.raid_disk
= d
;
199 info
.disk
.state
= (1 << MD_DISK_SYNC
) | (1 << MD_DISK_ACTIVE
);
200 st
->ss
->update_super(st
, &info
, "linear-grow-new", newdev
,
203 if (st
->ss
->store_super(st
, nfd
)) {
204 pr_err("Cannot store new superblock on %s\n",
211 if (ioctl(fd
, ADD_NEW_DISK
, &info
.disk
) != 0) {
212 pr_err("Cannot add new disk to this array\n");
215 /* Well, that seems to have worked.
216 * Now go through and update all superblocks
219 if (ioctl(fd
, GET_ARRAY_INFO
, &info
.array
) < 0) {
220 pr_err("cannot get array info for %s\n", devname
);
225 for (d
=0 ; d
< info
.array
.raid_disks
; d
++) {
226 mdu_disk_info_t disk
;
230 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0) {
231 pr_err("cannot get device detail for device %d\n",
235 dv
= map_dev(disk
.major
, disk
.minor
, 1);
237 pr_err("cannot find device file for device %d\n",
241 fd2
= dev_open(dv
, O_RDWR
);
243 pr_err("cannot open device file %s\n", dv
);
246 if (st
->ss
->load_super(st
, fd2
, NULL
)) {
247 pr_err("cannot find super block on %s\n", dv
);
251 info
.array
.raid_disks
= nd
+1;
252 info
.array
.nr_disks
= nd
+1;
253 info
.array
.active_disks
= nd
+1;
254 info
.array
.working_disks
= nd
+1;
256 st
->ss
->update_super(st
, &info
, "linear-grow-update", dv
,
259 if (st
->ss
->store_super(st
, fd2
)) {
260 pr_err("Cannot store new superblock on %s\n", dv
);
270 int Grow_addbitmap(char *devname
, int fd
, struct context
*c
, struct shape
*s
)
273 * First check that array doesn't have a bitmap
274 * Then create the bitmap
277 * For internal bitmaps, we need to check the version,
278 * find all the active devices, and write the bitmap block
281 mdu_bitmap_file_t bmf
;
282 mdu_array_info_t array
;
283 struct supertype
*st
;
284 char *subarray
= NULL
;
285 int major
= BITMAP_MAJOR_HI
;
286 int vers
= md_get_version(fd
);
287 unsigned long long bitmapsize
, array_size
;
290 major
= BITMAP_MAJOR_HOSTENDIAN
;
291 pr_err("Warning - bitmaps created on this kernel"
292 " are not portable\n"
293 " between different architectures. Consider upgrading"
294 " the Linux kernel.\n");
297 if (ioctl(fd
, GET_BITMAP_FILE
, &bmf
) != 0) {
299 pr_err("Memory allocation failure.\n");
301 pr_err("bitmaps not supported by this kernel.\n");
304 if (bmf
.pathname
[0]) {
305 if (strcmp(s
->bitmap_file
,"none")==0) {
306 if (ioctl(fd
, SET_BITMAP_FILE
, -1)!= 0) {
307 pr_err("failed to remove bitmap %s\n",
313 pr_err("%s already has a bitmap (%s)\n",
314 devname
, bmf
.pathname
);
317 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
318 pr_err("cannot get array status for %s\n", devname
);
321 if (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)) {
322 if (strcmp(s
->bitmap_file
, "none")==0) {
323 array
.state
&= ~(1<<MD_SB_BITMAP_PRESENT
);
324 if (ioctl(fd
, SET_ARRAY_INFO
, &array
)!= 0) {
325 pr_err("failed to remove internal bitmap.\n");
330 pr_err("Internal bitmap already present on %s\n",
335 if (strcmp(s
->bitmap_file
, "none") == 0) {
336 pr_err("no bitmap found on %s\n", devname
);
339 if (array
.level
<= 0) {
340 pr_err("Bitmaps not meaningful with level %s\n",
341 map_num(pers
, array
.level
)?:"of this array");
344 bitmapsize
= array
.size
;
346 if (get_dev_size(fd
, NULL
, &array_size
) &&
347 array_size
> (0x7fffffffULL
<<9)) {
348 /* Array is big enough that we cannot trust array.size
349 * try other approaches
351 bitmapsize
= get_component_size(fd
);
353 if (bitmapsize
== 0) {
354 pr_err("Cannot reliably determine size of array to create bitmap - sorry.\n");
358 if (array
.level
== 10) {
359 int ncopies
= (array
.layout
&255)*((array
.layout
>>8)&255);
360 bitmapsize
= bitmapsize
* array
.raid_disks
/ ncopies
;
363 st
= super_by_fd(fd
, &subarray
);
365 pr_err("Cannot understand version %d.%d\n",
366 array
.major_version
, array
.minor_version
);
370 pr_err("Cannot add bitmaps to sub-arrays yet\n");
375 if (strcmp(s
->bitmap_file
, "internal") == 0) {
378 int offset_setable
= 0;
380 if (st
->ss
->add_internal_bitmap
== NULL
) {
381 pr_err("Internal bitmaps not supported "
382 "with %s metadata\n", st
->ss
->name
);
385 mdi
= sysfs_read(fd
, NULL
, GET_BITMAP_LOCATION
);
388 for (d
=0; d
< st
->max_devs
; d
++) {
389 mdu_disk_info_t disk
;
392 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0)
394 if (disk
.major
== 0 &&
397 if ((disk
.state
& (1<<MD_DISK_SYNC
))==0)
399 dv
= map_dev(disk
.major
, disk
.minor
, 1);
401 int fd2
= dev_open(dv
, O_RDWR
);
404 if (st
->ss
->load_super(st
, fd2
, NULL
)==0) {
405 if (st
->ss
->add_internal_bitmap(
407 &s
->bitmap_chunk
, c
->delay
, s
->write_behind
,
408 bitmapsize
, offset_setable
,
411 st
->ss
->write_bitmap(st
, fd2
);
413 pr_err("failed to create internal bitmap"
414 " - chunksize problem.\n");
422 if (offset_setable
) {
423 st
->ss
->getinfo_super(st
, mdi
, NULL
);
424 sysfs_init(mdi
, fd
, NULL
);
425 rv
= sysfs_set_num_signed(mdi
, NULL
, "bitmap/location",
428 array
.state
|= (1<<MD_SB_BITMAP_PRESENT
);
429 rv
= ioctl(fd
, SET_ARRAY_INFO
, &array
);
433 pr_err("Cannot add bitmap while array is"
434 " resyncing or reshaping etc.\n");
435 pr_err("failed to set internal bitmap.\n");
442 int max_devs
= st
->max_devs
;
444 /* try to load a superblock */
445 for (d
= 0; d
< max_devs
; d
++) {
446 mdu_disk_info_t disk
;
450 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0)
452 if ((disk
.major
==0 && disk
.minor
==0) ||
453 (disk
.state
& (1<<MD_DISK_REMOVED
)))
455 dv
= map_dev(disk
.major
, disk
.minor
, 1);
458 fd2
= dev_open(dv
, O_RDONLY
);
460 if (st
->ss
->load_super(st
, fd2
, NULL
) == 0) {
462 st
->ss
->uuid_from_super(st
, uuid
);
469 pr_err("cannot find UUID for array!\n");
472 if (CreateBitmap(s
->bitmap_file
, c
->force
, (char*)uuid
, s
->bitmap_chunk
,
473 c
->delay
, s
->write_behind
, bitmapsize
, major
)) {
476 bitmap_fd
= open(s
->bitmap_file
, O_RDWR
);
478 pr_err("weird: %s cannot be opened\n",
482 if (ioctl(fd
, SET_BITMAP_FILE
, bitmap_fd
) < 0) {
485 pr_err("Cannot add bitmap while array is"
486 " resyncing or reshaping etc.\n");
487 pr_err("Cannot set bitmap file for %s: %s\n",
488 devname
, strerror(err
));
497 * When reshaping an array we might need to backup some data.
498 * This is written to all spares with a 'super_block' describing it.
499 * The superblock goes 4K from the end of the used space on the
501 * It if written after the backup is complete.
502 * It has the following structure.
505 static struct mdp_backup_super
{
506 char magic
[16]; /* md_backup_data-1 or -2 */
509 /* start/sizes in 512byte sectors */
510 __u64 devstart
; /* address on backup device/file of data */
513 __u32 sb_csum
; /* csum of preceeding bytes. */
515 __u64 devstart2
; /* offset in to data of second section */
518 __u32 sb_csum2
; /* csum of preceeding bytes. */
520 } __attribute__((aligned(512))) bsb
, bsb2
;
522 static __u32
bsb_csum(char *buf
, int len
)
526 for (i
= 0; i
< len
; i
++)
527 csum
= (csum
<<3) + buf
[0];
528 return __cpu_to_le32(csum
);
531 static int check_idle(struct supertype
*st
)
533 /* Check that all member arrays for this container, or the
534 * container of this array, are idle
536 char *container
= (st
->container_devnm
[0]
537 ? st
->container_devnm
: st
->devnm
);
538 struct mdstat_ent
*ent
, *e
;
541 ent
= mdstat_read(0, 0);
542 for (e
= ent
; e
; e
= e
->next
) {
543 if (!is_container_member(e
, container
))
545 if (e
->percent
>= 0) {
554 static int freeze_container(struct supertype
*st
)
556 char *container
= (st
->container_devnm
[0]
557 ? st
->container_devnm
: st
->devnm
);
562 if (block_monitor(container
, 1)) {
563 pr_err("failed to freeze container\n");
570 static void unfreeze_container(struct supertype
*st
)
572 char *container
= (st
->container_devnm
[0]
573 ? st
->container_devnm
: st
->devnm
);
575 unblock_monitor(container
, 1);
578 static int freeze(struct supertype
*st
)
580 /* Try to freeze resync/rebuild on this array/container.
581 * Return -1 if the array is busy,
582 * return -2 container cannot be frozen,
583 * return 0 if this kernel doesn't support 'frozen'
584 * return 1 if it worked.
586 if (st
->ss
->external
)
587 return freeze_container(st
);
589 struct mdinfo
*sra
= sysfs_read(-1, st
->devnm
, GET_VERSION
);
595 /* Need to clear any 'read-auto' status */
596 if (sysfs_get_str(sra
, NULL
, "array_state", buf
, 20) > 0 &&
597 strncmp(buf
, "read-auto", 9) == 0)
598 sysfs_set_str(sra
, NULL
, "array_state", "clean");
600 err
= sysfs_freeze_array(sra
);
606 static void unfreeze(struct supertype
*st
)
608 if (st
->ss
->external
)
609 return unfreeze_container(st
);
611 struct mdinfo
*sra
= sysfs_read(-1, st
->devnm
, GET_VERSION
);
614 sysfs_set_str(sra
, NULL
, "sync_action", "idle");
619 static void wait_reshape(struct mdinfo
*sra
)
621 int fd
= sysfs_get_fd(sra
, NULL
, "sync_action");
627 while (sysfs_fd_get_str(fd
, action
, 20) > 0 &&
628 strncmp(action
, "reshape", 7) == 0) {
632 select(fd
+1, NULL
, NULL
, &rfds
, NULL
);
637 static int reshape_super(struct supertype
*st
, unsigned long long size
,
638 int level
, int layout
, int chunksize
, int raid_disks
,
639 int delta_disks
, char *backup_file
, char *dev
,
640 int direction
, int verbose
)
642 /* nothing extra to check in the native case */
643 if (!st
->ss
->external
)
645 if (!st
->ss
->reshape_super
||
646 !st
->ss
->manage_reshape
) {
647 pr_err("%s metadata does not support reshape\n",
652 return st
->ss
->reshape_super(st
, size
, level
, layout
, chunksize
,
653 raid_disks
, delta_disks
, backup_file
, dev
,
657 static void sync_metadata(struct supertype
*st
)
659 if (st
->ss
->external
) {
660 if (st
->update_tail
) {
661 flush_metadata_updates(st
);
662 st
->update_tail
= &st
->updates
;
664 st
->ss
->sync_metadata(st
);
668 static int subarray_set_num(char *container
, struct mdinfo
*sra
, char *name
, int n
)
670 /* when dealing with external metadata subarrays we need to be
671 * prepared to handle EAGAIN. The kernel may need to wait for
672 * mdmon to mark the array active so the kernel can handle
673 * allocations/writeback when preparing the reshape action
674 * (md_allow_write()). We temporarily disable safe_mode_delay
675 * to close a race with the array_state going clean before the
676 * next write to raid_disks / stripe_cache_size
681 /* only 'raid_disks' and 'stripe_cache_size' trigger md_allow_write */
683 (strcmp(name
, "raid_disks") != 0 &&
684 strcmp(name
, "stripe_cache_size") != 0))
685 return sysfs_set_num(sra
, NULL
, name
, n
);
687 rc
= sysfs_get_str(sra
, NULL
, "safe_mode_delay", safe
, sizeof(safe
));
690 sysfs_set_num(sra
, NULL
, "safe_mode_delay", 0);
691 rc
= sysfs_set_num(sra
, NULL
, name
, n
);
692 if (rc
< 0 && errno
== EAGAIN
) {
693 ping_monitor(container
);
694 /* if we get EAGAIN here then the monitor is not active
697 rc
= sysfs_set_num(sra
, NULL
, name
, n
);
699 sysfs_set_str(sra
, NULL
, "safe_mode_delay", safe
);
703 int start_reshape(struct mdinfo
*sra
, int already_running
,
704 int before_data_disks
, int data_disks
)
707 unsigned long long sync_max_to_set
;
709 sysfs_set_num(sra
, NULL
, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL
);
710 err
= sysfs_set_num(sra
, NULL
, "suspend_hi", sra
->reshape_progress
);
711 err
= err
?: sysfs_set_num(sra
, NULL
, "suspend_lo",
712 sra
->reshape_progress
);
713 if (before_data_disks
<= data_disks
)
714 sync_max_to_set
= sra
->reshape_progress
/ data_disks
;
716 sync_max_to_set
= (sra
->component_size
* data_disks
717 - sra
->reshape_progress
) / data_disks
;
718 if (!already_running
)
719 sysfs_set_num(sra
, NULL
, "sync_min", sync_max_to_set
);
720 err
= err
?: sysfs_set_num(sra
, NULL
, "sync_max", sync_max_to_set
);
721 if (!already_running
)
722 err
= err
?: sysfs_set_str(sra
, NULL
, "sync_action", "reshape");
727 void abort_reshape(struct mdinfo
*sra
)
729 sysfs_set_str(sra
, NULL
, "sync_action", "idle");
730 sysfs_set_num(sra
, NULL
, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL
);
731 sysfs_set_num(sra
, NULL
, "suspend_hi", 0);
732 sysfs_set_num(sra
, NULL
, "suspend_lo", 0);
733 sysfs_set_num(sra
, NULL
, "sync_min", 0);
734 sysfs_set_str(sra
, NULL
, "sync_max", "max");
737 int remove_disks_for_takeover(struct supertype
*st
,
742 struct mdinfo
*remaining
;
745 if (sra
->array
.level
== 10)
746 nr_of_copies
= layout
& 0xff;
747 else if (sra
->array
.level
== 1)
748 nr_of_copies
= sra
->array
.raid_disks
;
752 remaining
= sra
->devs
;
754 /* for each 'copy', select one device and remove from the list. */
755 for (slot
= 0; slot
< sra
->array
.raid_disks
; slot
+= nr_of_copies
) {
756 struct mdinfo
**diskp
;
759 /* Find a working device to keep */
760 for (diskp
= &remaining
; *diskp
; diskp
= &(*diskp
)->next
) {
761 struct mdinfo
*disk
= *diskp
;
763 if (disk
->disk
.raid_disk
< slot
)
765 if (disk
->disk
.raid_disk
>= slot
+ nr_of_copies
)
767 if (disk
->disk
.state
& (1<<MD_DISK_REMOVED
))
769 if (disk
->disk
.state
& (1<<MD_DISK_FAULTY
))
771 if (!(disk
->disk
.state
& (1<<MD_DISK_SYNC
)))
774 /* We have found a good disk to use! */
776 disk
->next
= sra
->devs
;
785 if (slot
< sra
->array
.raid_disks
) {
786 /* didn't find all slots */
792 sra
->devs
= remaining
;
796 /* Remove all 'remaining' devices from the array */
798 struct mdinfo
*sd
= remaining
;
799 remaining
= sd
->next
;
801 sysfs_set_str(sra
, sd
, "state", "faulty");
802 sysfs_set_str(sra
, sd
, "slot", "none");
803 /* for external metadata disks should be removed in mdmon */
804 if (!st
->ss
->external
)
805 sysfs_set_str(sra
, sd
, "state", "remove");
806 sd
->disk
.state
|= (1<<MD_DISK_REMOVED
);
807 sd
->disk
.state
&= ~(1<<MD_DISK_SYNC
);
808 sd
->next
= sra
->devs
;
814 void reshape_free_fdlist(int *fdlist
,
815 unsigned long long *offsets
,
820 for (i
= 0; i
< size
; i
++)
828 int reshape_prepare_fdlist(char *devname
,
832 unsigned long blocks
,
835 unsigned long long *offsets
)
840 for (d
= 0; d
<= nrdisks
; d
++)
843 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
844 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
846 if (sd
->disk
.state
& (1<<MD_DISK_SYNC
)) {
847 char *dn
= map_dev(sd
->disk
.major
,
849 fdlist
[sd
->disk
.raid_disk
]
850 = dev_open(dn
, O_RDONLY
);
851 offsets
[sd
->disk
.raid_disk
] = sd
->data_offset
*512;
852 if (fdlist
[sd
->disk
.raid_disk
] < 0) {
853 pr_err("%s: cannot open component %s\n",
854 devname
, dn
? dn
: "-unknown-");
858 } else if (backup_file
== NULL
) {
860 char *dn
= map_dev(sd
->disk
.major
,
862 fdlist
[d
] = dev_open(dn
, O_RDWR
);
863 offsets
[d
] = (sd
->data_offset
+ sra
->component_size
- blocks
- 8)*512;
865 pr_err("%s: cannot open component %s\n",
866 devname
, dn
? dn
: "-unknown-");
877 int reshape_open_backup_file(char *backup_file
,
882 unsigned long long *offsets
,
885 /* Return 1 on success, 0 on any form of failure */
886 /* need to check backup file is large enough */
892 *fdlist
= open(backup_file
, O_RDWR
|O_CREAT
|(restart
? O_TRUNC
: O_EXCL
),
896 pr_err("%s: cannot create backup file %s: %s\n",
897 devname
, backup_file
, strerror(errno
));
900 /* Guard against backup file being on array device.
901 * If array is partitioned or if LVM etc is in the
902 * way this will not notice, but it is better than
905 fstat(*fdlist
, &stb
);
908 if (stb
.st_rdev
== dev
) {
909 pr_err("backup file must NOT be"
910 " on the array being reshaped.\n");
916 for (i
=0; i
< blocks
+ 8 ; i
++) {
917 if (write(*fdlist
, buf
, 512) != 512) {
918 pr_err("%s: cannot create"
919 " backup file %s: %s\n",
920 devname
, backup_file
, strerror(errno
));
924 if (fsync(*fdlist
) != 0) {
925 pr_err("%s: cannot create backup file %s: %s\n",
926 devname
, backup_file
, strerror(errno
));
933 unsigned long GCD(unsigned long a
, unsigned long b
)
944 unsigned long compute_backup_blocks(int nchunk
, int ochunk
,
945 unsigned int ndata
, unsigned int odata
)
947 unsigned long a
, b
, blocks
;
948 /* So how much do we need to backup.
949 * We need an amount of data which is both a whole number of
950 * old stripes and a whole number of new stripes.
951 * So LCM for (chunksize*datadisks).
953 a
= (ochunk
/512) * odata
;
954 b
= (nchunk
/512) * ndata
;
957 /* LCM == product / GCD */
958 blocks
= (ochunk
/512) * (nchunk
/512) * odata
* ndata
/ a
;
963 char *analyse_change(struct mdinfo
*info
, struct reshape
*re
)
965 /* Based on the current array state in info->array and
966 * the changes in info->new_* etc, determine:
967 * - whether the change is possible
968 * - Intermediate level/raid_disks/layout
969 * - whether a restriping reshape is needed
970 * - number of sectors in minimum change unit. This
971 * will cover a whole number of stripes in 'before' and
974 * Return message if the change should be rejected
975 * NULL if the change can be achieved
977 * This can be called as part of starting a reshape, or
978 * when assembling an array that is undergoing reshape.
980 int near
, far
, offset
, copies
;
982 int old_chunk
, new_chunk
;
983 /* delta_parity records change in number of devices
984 * caused by level change
986 int delta_parity
= 0;
988 memset(re
, 0, sizeof(*re
));
990 /* If a new level not explicitly given, we assume no-change */
991 if (info
->new_level
== UnSet
)
992 info
->new_level
= info
->array
.level
;
995 switch (info
->new_level
) {
1001 /* chunk size is meaningful, must divide component_size
1004 if (info
->component_size
% (info
->new_chunk
/512))
1005 return "New chunk size does not"
1006 " divide component size";
1009 return "chunk size not meaningful for this level";
1012 info
->new_chunk
= info
->array
.chunk_size
;
1014 switch (info
->array
.level
) {
1016 return "Cannot understand this RAID level";
1018 /* RAID1 can convert to RAID1 with different disks, or
1019 * raid5 with 2 disks, or
1022 if (info
->new_level
> 1 &&
1023 (info
->component_size
& 7))
1024 return "Cannot convert RAID1 of this size - "
1025 "reduce size to multiple of 4K first.";
1026 if (info
->new_level
== 0) {
1027 if (info
->delta_disks
!= UnSet
&&
1028 info
->delta_disks
!= 0)
1029 return "Cannot change number of disks "
1030 "with RAID1->RAID0 conversion";
1032 re
->before
.data_disks
= 1;
1033 re
->after
.data_disks
= 1;
1036 if (info
->new_level
== 1) {
1037 if (info
->delta_disks
== UnSet
)
1038 /* Don't know what to do */
1039 return "no change requested for Growing RAID1";
1043 if (info
->array
.raid_disks
== 2 &&
1044 info
->new_level
== 5) {
1047 re
->before
.data_disks
= 1;
1048 if (info
->delta_disks
!= UnSet
&&
1049 info
->delta_disks
!= 0)
1050 re
->after
.data_disks
= 1 + info
->delta_disks
;
1052 re
->after
.data_disks
= 1;
1053 if (re
->after
.data_disks
< 1)
1054 return "Number of disks too small for RAID5";
1056 re
->before
.layout
= ALGORITHM_LEFT_SYMMETRIC
;
1057 info
->array
.chunk_size
= 65536;
1060 /* Could do some multi-stage conversions, but leave that to
1063 return "Impossibly level change request for RAID1";
1066 /* RAID10 can be converted from near mode to
1067 * RAID0 by removing some devices.
1068 * It can also be reshaped if the kernel supports
1071 switch (info
->new_level
) {
1073 if ((info
->array
.layout
& ~0xff) != 0x100)
1074 return "Cannot Grow RAID10 with far/offset layout";
1075 /* number of devices must be multiple of number of copies */
1076 if (info
->array
.raid_disks
% (info
->array
.layout
& 0xff))
1077 return "RAID10 layout too complex for Grow operation";
1079 new_disks
= (info
->array
.raid_disks
1080 / (info
->array
.layout
& 0xff));
1081 if (info
->delta_disks
== UnSet
)
1082 info
->delta_disks
= (new_disks
1083 - info
->array
.raid_disks
);
1085 if (info
->delta_disks
!= new_disks
- info
->array
.raid_disks
)
1086 return "New number of raid-devices impossible for RAID10";
1087 if (info
->new_chunk
&&
1088 info
->new_chunk
!= info
->array
.chunk_size
)
1089 return "Cannot change chunk-size with RAID10 Grow";
1093 re
->before
.data_disks
= new_disks
;
1094 re
->after
.data_disks
= re
->before
.data_disks
;
1098 near
= info
->array
.layout
& 0xff;
1099 far
= (info
->array
.layout
>> 8) & 0xff;
1100 offset
= info
->array
.layout
& 0x10000;
1101 if (far
> 1 && !offset
)
1102 return "Cannot reshape RAID10 in far-mode";
1103 copies
= near
* far
;
1105 old_chunk
= info
->array
.chunk_size
* far
;
1107 if (info
->new_layout
== UnSet
)
1108 info
->new_layout
= info
->array
.layout
;
1110 near
= info
->new_layout
& 0xff;
1111 far
= (info
->new_layout
>> 8) & 0xff;
1112 offset
= info
->new_layout
& 0x10000;
1113 if (far
> 1 && !offset
)
1114 return "Cannot reshape RAID10 to far-mode";
1115 if (near
* far
!= copies
)
1116 return "Cannot change number of copies"
1117 " when reshaping RAID10";
1119 if (info
->delta_disks
== UnSet
)
1120 info
->delta_disks
= 0;
1121 new_disks
= (info
->array
.raid_disks
+
1124 new_chunk
= info
->new_chunk
* far
;
1127 re
->before
.layout
= info
->array
.layout
;
1128 re
->before
.data_disks
= info
->array
.raid_disks
;
1129 re
->after
.layout
= info
->new_layout
;
1130 re
->after
.data_disks
= new_disks
;
1131 /* For RAID10 we don't do backup but do allow reshape,
1132 * so set backup_blocks to INVALID_SECTORS rather than
1134 * And there is no need to synchronise stripes on both
1135 * 'old' and 'new'. So the important
1136 * number is the minimum data_offset difference
1137 * which is the larger of (offset copies * chunk).
1139 re
->backup_blocks
= INVALID_SECTORS
;
1140 re
->min_offset_change
= max(old_chunk
, new_chunk
) / 512;
1141 if (new_disks
< re
->before
.data_disks
&&
1142 info
->space_after
< re
->min_offset_change
)
1143 /* Reduce component size by one chunk */
1144 re
->new_size
= (info
->component_size
-
1145 re
->min_offset_change
);
1147 re
->new_size
= info
->component_size
;
1148 re
->new_size
= re
->new_size
* new_disks
/ copies
;
1152 return "RAID10 can only be changed to RAID0";
1155 /* RAID0 can be converted to RAID10, or to RAID456 */
1156 if (info
->new_level
== 10) {
1157 if (info
->new_layout
== UnSet
&& info
->delta_disks
== UnSet
) {
1158 /* Assume near=2 layout */
1159 info
->new_layout
= 0x102;
1160 info
->delta_disks
= info
->array
.raid_disks
;
1162 if (info
->new_layout
== UnSet
) {
1163 int copies
= 1 + (info
->delta_disks
1164 / info
->array
.raid_disks
);
1165 if (info
->array
.raid_disks
* (copies
-1)
1166 != info
->delta_disks
)
1167 return "Impossible number of devices"
1168 " for RAID0->RAID10";
1169 info
->new_layout
= 0x100 + copies
;
1171 if (info
->delta_disks
== UnSet
) {
1172 int copies
= info
->new_layout
& 0xff;
1173 if (info
->new_layout
!= 0x100 + copies
)
1174 return "New layout impossible"
1175 " for RAID0->RAID10";;
1176 info
->delta_disks
= (copies
- 1) *
1177 info
->array
.raid_disks
;
1179 if (info
->new_chunk
&&
1180 info
->new_chunk
!= info
->array
.chunk_size
)
1181 return "Cannot change chunk-size with RAID0->RAID10";
1184 re
->before
.data_disks
= (info
->array
.raid_disks
+
1186 re
->after
.data_disks
= re
->before
.data_disks
;
1187 re
->before
.layout
= info
->new_layout
;
1191 /* RAID0 can also covert to RAID0/4/5/6 by first converting to
1192 * a raid4 style layout of the final level.
1194 switch (info
->new_level
) {
1199 re
->before
.layout
= 0;
1204 re
->before
.layout
= ALGORITHM_PARITY_N
;
1209 re
->before
.layout
= ALGORITHM_PARITY_N
;
1212 return "Impossible level change requested";
1214 re
->before
.data_disks
= info
->array
.raid_disks
;
1215 /* determining 'after' layout happens outside this 'switch' */
1219 info
->array
.layout
= ALGORITHM_PARITY_N
;
1221 switch (info
->new_level
) {
1225 re
->level
= info
->array
.level
;
1226 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1227 re
->before
.layout
= info
->array
.layout
;
1231 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1232 re
->before
.layout
= info
->array
.layout
;
1237 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1238 switch (info
->array
.layout
) {
1239 case ALGORITHM_LEFT_ASYMMETRIC
:
1240 re
->before
.layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
1242 case ALGORITHM_RIGHT_ASYMMETRIC
:
1243 re
->before
.layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
1245 case ALGORITHM_LEFT_SYMMETRIC
:
1246 re
->before
.layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
1248 case ALGORITHM_RIGHT_SYMMETRIC
:
1249 re
->before
.layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
1251 case ALGORITHM_PARITY_0
:
1252 re
->before
.layout
= ALGORITHM_PARITY_0_6
;
1254 case ALGORITHM_PARITY_N
:
1255 re
->before
.layout
= ALGORITHM_PARITY_N_6
;
1258 return "Cannot convert an array with this layout";
1262 if (info
->array
.raid_disks
!= 2)
1263 return "Can only convert a 2-device array to RAID1";
1264 if (info
->delta_disks
!= UnSet
&&
1265 info
->delta_disks
!= 0)
1266 return "Cannot set raid_disk when "
1267 "converting RAID5->RAID1";
1269 info
->new_chunk
= 0;
1272 return "Impossible level change requested";
1276 switch (info
->new_level
) {
1282 re
->before
.data_disks
= info
->array
.raid_disks
- 2;
1283 re
->before
.layout
= info
->array
.layout
;
1286 return "Impossible level change requested";
1291 /* If we reached here then it looks like a re-stripe is
1292 * happening. We have determined the intermediate level
1293 * and initial raid_disks/layout and stored these in 're'.
1295 * We need to deduce the final layout that can be atomically
1296 * converted to the end state.
1298 switch (info
->new_level
) {
1300 /* We can only get to RAID0 from RAID4 or RAID5
1301 * with appropriate layout and one extra device
1303 if (re
->level
!= 4 && re
->level
!= 5)
1304 return "Cannot covert to RAID0 from this level";
1306 switch (re
->level
) {
1308 re
->before
.layout
= 0;
1309 re
->after
.layout
= 0;
1312 re
->after
.layout
= ALGORITHM_PARITY_N
;
1318 /* We can only get to RAID4 from RAID5 */
1319 if (re
->level
!= 4 && re
->level
!= 5)
1320 return "Cannot convert to RAID4 from this level";
1322 switch (re
->level
) {
1324 re
->before
.layout
= 0;
1325 re
->after
.layout
= 0;
1328 re
->after
.layout
= ALGORITHM_PARITY_N
;
1334 /* We get to RAID5 from RAID5 or RAID6 */
1335 if (re
->level
!= 5 && re
->level
!= 6)
1336 return "Cannot convert to RAID5 from this level";
1338 switch (re
->level
) {
1340 if (info
->new_layout
== UnSet
)
1341 re
->after
.layout
= re
->before
.layout
;
1343 re
->after
.layout
= info
->new_layout
;
1346 if (info
->new_layout
== UnSet
)
1347 info
->new_layout
= re
->before
.layout
;
1349 /* after.layout needs to be raid6 version of new_layout */
1350 if (info
->new_layout
== ALGORITHM_PARITY_N
)
1351 re
->after
.layout
= ALGORITHM_PARITY_N
;
1354 char *ls
= map_num(r5layout
, info
->new_layout
);
1357 /* Current RAID6 layout has a RAID5
1360 strcat(strcpy(layout
, ls
), "-6");
1361 l
= map_name(r6layout
, layout
);
1363 return "Cannot find RAID6 layout"
1366 /* Current RAID6 has no equivalent.
1367 * If it is already a '-6' layout we
1368 * can leave it unchanged, else we must
1371 ls
= map_num(r6layout
, info
->new_layout
);
1373 strcmp(ls
+strlen(ls
)-2, "-6") != 0)
1374 return "Please specify new layout";
1375 l
= info
->new_layout
;
1377 re
->after
.layout
= l
;
1383 /* We must already be at level 6 */
1385 return "Impossible level change";
1386 if (info
->new_layout
== UnSet
)
1387 re
->after
.layout
= info
->array
.layout
;
1389 re
->after
.layout
= info
->new_layout
;
1392 return "Impossible level change requested";
1394 if (info
->delta_disks
== UnSet
)
1395 info
->delta_disks
= delta_parity
;
1397 re
->after
.data_disks
= (re
->before
.data_disks
1400 switch (re
->level
) {
1401 case 6: re
->parity
= 2;
1404 case 5: re
->parity
= 1;
1406 default: re
->parity
= 0;
1409 /* So we have a restripe operation, we need to calculate the number
1410 * of blocks per reshape operation.
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
;
1538 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) < 0) {
1539 pr_err("%s is not an active md array - aborting\n",
1543 if (data_offset
!= INVALID_SECTORS
&& array
.level
!= 10
1544 && (array
.level
< 4 || array
.level
> 6)) {
1545 pr_err("--grow --data-offset not yet supported\n");
1550 (s
->chunk
|| s
->level
!= UnSet
|| s
->layout_str
|| s
->raiddisks
)) {
1551 pr_err("cannot change component size at the same time "
1552 "as other changes.\n"
1553 " Change size first, then check data is intact before "
1554 "making other changes.\n");
1558 if (s
->raiddisks
&& s
->raiddisks
< array
.raid_disks
&& array
.level
> 1 &&
1559 get_linux_version() < 2006032 &&
1560 !check_env("MDADM_FORCE_FEWER")) {
1561 pr_err("reducing the number of devices is not safe before Linux 2.6.32\n"
1562 " Please use a newer kernel\n");
1566 st
= super_by_fd(fd
, &subarray
);
1568 pr_err("Unable to determine metadata format for %s\n", devname
);
1571 if (s
->raiddisks
> st
->max_devs
) {
1572 pr_err("Cannot increase raid-disks on this array"
1573 " beyond %d\n", st
->max_devs
);
1577 /* in the external case we need to check that the requested reshape is
1578 * supported, and perform an initial check that the container holds the
1579 * pre-requisite spare devices (mdmon owns final validation)
1581 if (st
->ss
->external
) {
1585 container
= st
->container_devnm
;
1586 cfd
= open_dev_excl(st
->container_devnm
);
1588 container
= st
->devnm
;
1590 cfd
= open_dev_excl(st
->devnm
);
1594 pr_err("Unable to open container for %s\n",
1600 rv
= st
->ss
->load_container(st
, cfd
, NULL
);
1603 pr_err("Cannot read superblock for %s\n",
1609 /* check if operation is supported for metadata handler */
1610 if (st
->ss
->container_content
) {
1611 struct mdinfo
*cc
= NULL
;
1612 struct mdinfo
*content
= NULL
;
1614 cc
= st
->ss
->container_content(st
, subarray
);
1615 for (content
= cc
; content
; content
= content
->next
) {
1616 int allow_reshape
= 1;
1618 /* check if reshape is allowed based on metadata
1619 * indications stored in content.array.status
1621 if (content
->array
.state
& (1<<MD_SB_BLOCK_VOLUME
))
1623 if (content
->array
.state
1624 & (1<<MD_SB_BLOCK_CONTAINER_RESHAPE
))
1626 if (!allow_reshape
) {
1627 pr_err("cannot reshape arrays in"
1628 " container with unsupported"
1629 " metadata: %s(%s)\n",
1630 devname
, container
);
1638 if (mdmon_running(container
))
1639 st
->update_tail
= &st
->updates
;
1643 for (dv
= devlist
; dv
; dv
= dv
->next
)
1645 if (s
->raiddisks
> array
.raid_disks
&&
1646 array
.spare_disks
+added_disks
< (s
->raiddisks
- array
.raid_disks
) &&
1648 pr_err("Need %d spare%s to avoid degraded array,"
1649 " and only have %d.\n"
1650 " Use --force to over-ride this check.\n",
1651 s
->raiddisks
- array
.raid_disks
,
1652 s
->raiddisks
- array
.raid_disks
== 1 ? "" : "s",
1653 array
.spare_disks
+ added_disks
);
1657 sra
= sysfs_read(fd
, NULL
, GET_LEVEL
| GET_DISKS
| GET_DEVS
1658 | GET_STATE
| GET_VERSION
);
1660 if (st
->ss
->external
&& subarray
== NULL
) {
1661 array
.level
= LEVEL_CONTAINER
;
1662 sra
->array
.level
= LEVEL_CONTAINER
;
1665 pr_err("failed to read sysfs parameters for %s\n",
1669 frozen
= freeze(st
);
1671 /* freeze() already spewed the reason */
1674 } else if (frozen
< 0) {
1675 pr_err("%s is performing resync/recovery and cannot"
1676 " be reshaped\n", devname
);
1681 /* ========= set size =============== */
1682 if (s
->size
> 0 && (s
->size
== MAX_SIZE
|| s
->size
!= (unsigned)array
.size
)) {
1683 unsigned long long orig_size
= get_component_size(fd
)/2;
1684 unsigned long long min_csize
;
1686 int raid0_takeover
= 0;
1689 orig_size
= (unsigned) array
.size
;
1691 if (orig_size
== 0) {
1692 pr_err("Cannot set device size in this type of array.\n");
1697 if (reshape_super(st
, s
->size
, UnSet
, UnSet
, 0, 0, UnSet
, NULL
,
1698 devname
, APPLY_METADATA_CHANGES
, c
->verbose
> 0)) {
1703 if (st
->ss
->external
) {
1704 /* metadata can have size limitation
1705 * update size value according to metadata information
1707 struct mdinfo
*sizeinfo
=
1708 st
->ss
->container_content(st
, subarray
);
1710 unsigned long long new_size
=
1711 sizeinfo
->custom_array_size
/2;
1712 int data_disks
= get_data_disks(
1713 sizeinfo
->array
.level
,
1714 sizeinfo
->array
.layout
,
1715 sizeinfo
->array
.raid_disks
);
1716 new_size
/= data_disks
;
1717 dprintf("Metadata size correction from %llu to "
1718 "%llu (%llu)\n", orig_size
, new_size
,
1719 new_size
* data_disks
);
1721 sysfs_free(sizeinfo
);
1725 /* Update the size of each member device in case
1726 * they have been resized. This will never reduce
1727 * below the current used-size. The "size" attribute
1728 * understands '0' to mean 'max'.
1732 for (mdi
= sra
->devs
; mdi
; mdi
= mdi
->next
) {
1733 if (sysfs_set_num(sra
, mdi
, "size",
1734 s
->size
== MAX_SIZE
? 0 : s
->size
) < 0) {
1735 /* Probably kernel refusing to let us
1736 * reduce the size - not an error.
1740 if (array
.not_persistent
== 0 &&
1741 array
.major_version
== 0 &&
1742 get_linux_version() < 3001000) {
1743 /* Dangerous to allow size to exceed 2TB */
1744 unsigned long long csize
;
1745 if (sysfs_get_ll(sra
, mdi
, "size", &csize
) == 0) {
1746 if (csize
>= 2ULL*1024*1024*1024)
1747 csize
= 2ULL*1024*1024*1024;
1748 if ((min_csize
== 0 || (min_csize
1755 pr_err("Cannot set size on "
1756 "array members.\n");
1757 goto size_change_error
;
1759 if (min_csize
&& s
->size
> min_csize
) {
1760 pr_err("Cannot safely make this array "
1761 "use more than 2TB per device on this kernel.\n");
1763 goto size_change_error
;
1765 if (min_csize
&& s
->size
== MAX_SIZE
) {
1766 /* Don't let the kernel choose a size - it will get
1769 pr_err("Limited v0.90 array to "
1770 "2TB per device\n");
1771 s
->size
= min_csize
;
1773 if (st
->ss
->external
) {
1774 if (sra
->array
.level
== 0) {
1775 rv
= sysfs_set_str(sra
, NULL
, "level",
1779 /* get array parametes after takeover
1780 * to chane one parameter at time only
1782 rv
= ioctl(fd
, GET_ARRAY_INFO
, &array
);
1785 /* make sure mdmon is
1786 * aware of the new level */
1787 if (!mdmon_running(st
->container_devnm
))
1788 start_mdmon(st
->container_devnm
);
1789 ping_monitor(container
);
1790 if (mdmon_running(st
->container_devnm
) &&
1791 st
->update_tail
== NULL
)
1792 st
->update_tail
= &st
->updates
;
1795 if (s
->size
== MAX_SIZE
)
1797 array
.size
= s
->size
;
1798 if ((unsigned)array
.size
!= s
->size
) {
1799 /* got truncated to 32bit, write to
1800 * component_size instead
1803 rv
= sysfs_set_num(sra
, NULL
,
1804 "component_size", s
->size
);
1808 rv
= ioctl(fd
, SET_ARRAY_INFO
, &array
);
1810 /* manage array size when it is managed externally
1812 if ((rv
== 0) && st
->ss
->external
)
1813 rv
= set_array_size(st
, sra
, sra
->text_version
);
1816 if (raid0_takeover
) {
1817 /* do not recync non-existing parity,
1818 * we will drop it anyway
1820 sysfs_set_str(sra
, NULL
, "sync_action", "frozen");
1821 /* go back to raid0, drop parity disk
1823 sysfs_set_str(sra
, NULL
, "level", "raid0");
1824 ioctl(fd
, GET_ARRAY_INFO
, &array
);
1831 /* restore metadata */
1832 if (reshape_super(st
, orig_size
, UnSet
, UnSet
, 0, 0,
1833 UnSet
, NULL
, devname
,
1834 ROLLBACK_METADATA_CHANGES
,
1837 pr_err("Cannot set device size for %s: %s\n",
1838 devname
, strerror(err
));
1840 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
1841 cont_err("Bitmap must be removed before size can be changed\n");
1845 if (s
->assume_clean
) {
1846 /* This will fail on kernels older than 3.0 unless
1847 * a backport has been arranged.
1850 sysfs_set_str(sra
, NULL
, "resync_start", "none") < 0)
1851 pr_err("--assume-clean not supported with --grow on this kernel\n");
1853 ioctl(fd
, GET_ARRAY_INFO
, &array
);
1854 s
->size
= get_component_size(fd
)/2;
1856 s
->size
= array
.size
;
1857 if (c
->verbose
>= 0) {
1858 if (s
->size
== orig_size
)
1859 pr_err("component size of %s "
1860 "unchanged at %lluK\n",
1863 pr_err("component size of %s "
1864 "has been set to %lluK\n",
1868 } else if (array
.level
!= LEVEL_CONTAINER
) {
1869 s
->size
= get_component_size(fd
)/2;
1871 s
->size
= array
.size
;
1874 /* See if there is anything else to do */
1875 if ((s
->level
== UnSet
|| s
->level
== array
.level
) &&
1876 (s
->layout_str
== NULL
) &&
1877 (s
->chunk
== 0 || s
->chunk
== array
.chunk_size
) &&
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 int set_new_data_offset(struct mdinfo
*sra
, struct supertype
*st
,
2147 char *devname
, int delta_disks
,
2148 unsigned long long data_offset
,
2149 unsigned long long min
)
2155 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
2159 struct supertype
*st2
;
2160 struct mdinfo info2
;
2162 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
2164 dn
= map_dev(sd
->disk
.major
, sd
->disk
.minor
, 0);
2165 dfd
= dev_open(dn
, O_RDONLY
);
2167 pr_err("%s: cannot open component %s\n",
2168 devname
, dn
? dn
: "-unknown-");
2171 st2
= dup_super(st
);
2172 rv
= st2
->ss
->load_super(st2
,dfd
, NULL
);
2176 pr_err("%s: cannot get superblock from %s\n",
2180 st2
->ss
->getinfo_super(st2
, &info2
, NULL
);
2181 st2
->ss
->free_super(st2
);
2183 if (info2
.space_before
== 0 &&
2184 info2
.space_after
== 0) {
2185 /* Metadata doesn't support data_offset changes */
2188 if (delta_disks
< 0) {
2189 /* Don't need any space as array is shrinking
2190 * just move data_offset up by min
2192 if (data_offset
== INVALID_SECTORS
)
2193 info2
.new_data_offset
= info2
.data_offset
+ min
;
2195 if (data_offset
< info2
.data_offset
+ min
) {
2196 pr_err("--data-offset too small for %s\n",
2200 info2
.new_data_offset
= data_offset
;
2202 } else if (delta_disks
> 0) {
2203 /* need space before */
2204 if (info2
.space_before
< min
) {
2205 pr_err("Insufficient head-space for reshape on %s\n",
2209 if (data_offset
== INVALID_SECTORS
)
2210 info2
.new_data_offset
= info2
.data_offset
- min
;
2212 if (data_offset
> info2
.data_offset
- min
) {
2213 pr_err("--data-offset too large for %s\n",
2217 info2
.new_data_offset
= data_offset
;
2221 /* can move up or down. 'data_offset'
2222 * might guide us, otherwise choose
2223 * direction with most space
2225 if (data_offset
== INVALID_SECTORS
) {
2226 if (info2
.space_before
> info2
.space_after
)
2230 } else if (data_offset
< info2
.data_offset
)
2234 sysfs_set_str(sra
, NULL
, "reshape_direction",
2235 dir
== 1 ? "backwards" : "forwards");
2238 case 1: /* Increase data offset */
2239 if (info2
.space_after
< min
) {
2240 pr_err("Insufficient tail-space for reshape on %s\n",
2244 if (data_offset
!= INVALID_SECTORS
&&
2245 data_offset
< info2
.data_offset
+ min
) {
2246 pr_err("--data-offset too small on %s\n",
2250 if (data_offset
!= INVALID_SECTORS
)
2251 info2
.new_data_offset
= data_offset
;
2253 unsigned long long off
=
2254 info2
.space_after
/ 2;
2258 info2
.new_data_offset
=
2259 info2
.data_offset
+ off
;
2262 case -1: /* Decrease data offset */
2263 if (info2
.space_before
< min
) {
2264 pr_err("insufficient head-room on %s\n",
2268 if (data_offset
!= INVALID_SECTORS
&&
2269 data_offset
< info2
.data_offset
- min
) {
2270 pr_err("--data-offset too small on %s\n",
2274 if (data_offset
!= INVALID_SECTORS
)
2275 info2
.new_data_offset
= data_offset
;
2277 unsigned long long off
=
2278 info2
.space_before
/ 2;
2282 info2
.new_data_offset
=
2283 info2
.data_offset
- off
;
2288 if (sysfs_set_num(sra
, sd
, "new_offset",
2289 info2
.new_data_offset
) < 0) {
2291 if (sd
== sra
->devs
&& err
== ENOENT
)
2292 /* Early kernel, no 'new_offset' file.
2293 * For RAID5/6 this is not fatal
2296 pr_err("Cannot set new_offset for %s\n",
2306 static int raid10_reshape(char *container
, int fd
, char *devname
,
2307 struct supertype
*st
, struct mdinfo
*info
,
2308 struct reshape
*reshape
,
2309 unsigned long long data_offset
,
2310 int force
, int verbose
)
2312 /* Changing raid_disks, layout, chunksize or possibly
2313 * just data_offset for a RAID10.
2314 * We must always change data_offset. We change by at least
2315 * ->min_offset_change which is the largest of the old and new
2317 * If raid_disks is increasing, then data_offset must decrease
2318 * by at least this copy size.
2319 * If raid_disks is unchanged, data_offset must increase or
2320 * decrease by at least min_offset_change but preferably by much more.
2321 * We choose half of the available space.
2322 * If raid_disks is decreasing, data_offset must increase by
2323 * at least min_offset_change. To allow of this, component_size
2324 * must be decreased by the same amount.
2326 * So we calculate the required minimum and direction, possibly
2327 * reduce the component_size, then iterate through the devices
2328 * and set the new_data_offset.
2329 * If that all works, we set chunk_size, layout, raid_disks, and start
2333 unsigned long long min
;
2336 sra
= sysfs_read(fd
, NULL
,
2337 GET_COMPONENT
|GET_DEVS
|GET_OFFSET
|GET_STATE
|GET_CHUNK
2340 pr_err("%s: Cannot get array details from sysfs\n",
2344 min
= reshape
->min_offset_change
;
2346 if (info
->delta_disks
)
2347 sysfs_set_str(sra
, NULL
, "reshape_direction",
2348 info
->delta_disks
< 0 ? "backwards" : "forwards");
2349 if (info
->delta_disks
< 0 &&
2350 info
->space_after
< min
) {
2351 int rv
= sysfs_set_num(sra
, NULL
, "component_size",
2352 (sra
->component_size
-
2355 pr_err("cannot reduce component size\n");
2359 err
= set_new_data_offset(sra
, st
, devname
, info
->delta_disks
, data_offset
,
2362 pr_err("Cannot set new_data_offset: RAID10 reshape not\n");
2363 cont_err("supported on this kernel\n");
2369 if (!err
&& sysfs_set_num(sra
, NULL
, "chunk_size", info
->new_chunk
) < 0)
2371 if (!err
&& sysfs_set_num(sra
, NULL
, "layout", reshape
->after
.layout
) < 0)
2373 if (!err
&& sysfs_set_num(sra
, NULL
, "raid_disks",
2374 info
->array
.raid_disks
+ info
->delta_disks
) < 0)
2376 if (!err
&& sysfs_set_str(sra
, NULL
, "sync_action", "reshape") < 0)
2379 pr_err("Cannot set array shape for %s\n",
2382 (info
->array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
2383 cont_err(" Bitmap must be removed before"
2384 " shape can be changed\n");
2394 static void get_space_after(int fd
, struct supertype
*st
, struct mdinfo
*info
)
2396 struct mdinfo
*sra
, *sd
;
2397 /* Initialisation to silence compiler warning */
2398 unsigned long long min_space_before
= 0, min_space_after
= 0;
2401 sra
= sysfs_read(fd
, NULL
, GET_DEVS
);
2404 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
2407 struct supertype
*st2
;
2408 struct mdinfo info2
;
2410 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
2412 dn
= map_dev(sd
->disk
.major
, sd
->disk
.minor
, 0);
2413 dfd
= dev_open(dn
, O_RDONLY
);
2416 st2
= dup_super(st
);
2417 if (st2
->ss
->load_super(st2
,dfd
, NULL
)) {
2423 st2
->ss
->getinfo_super(st2
, &info2
, NULL
);
2424 st2
->ss
->free_super(st2
);
2427 min_space_before
> info2
.space_before
)
2428 min_space_before
= info2
.space_before
;
2430 min_space_after
> info2
.space_after
)
2431 min_space_after
= info2
.space_after
;
2434 if (sd
== NULL
&& !first
) {
2435 info
->space_after
= min_space_after
;
2436 info
->space_before
= min_space_before
;
2441 static void update_cache_size(char *container
, struct mdinfo
*sra
,
2442 struct mdinfo
*info
,
2443 int disks
, unsigned long long blocks
)
2445 /* Check that the internal stripe cache is
2446 * large enough, or it won't work.
2447 * It must hold at least 4 stripes of the larger
2450 unsigned long cache
;
2451 cache
= max(info
->array
.chunk_size
, info
->new_chunk
);
2452 cache
*= 4; /* 4 stripes minimum */
2453 cache
/= 512; /* convert to sectors */
2454 /* make sure there is room for 'blocks' with a bit to spare */
2455 if (cache
< 16 + blocks
/ disks
)
2456 cache
= 16 + blocks
/ disks
;
2457 cache
/= (4096/512); /* Covert from sectors to pages */
2459 if (sra
->cache_size
< cache
)
2460 subarray_set_num(container
, sra
, "stripe_cache_size",
2464 static int impose_reshape(struct mdinfo
*sra
,
2465 struct mdinfo
*info
,
2466 struct supertype
*st
,
2469 char *devname
, char *container
,
2470 struct reshape
*reshape
)
2472 struct mdu_array_info_s array
;
2474 sra
->new_chunk
= info
->new_chunk
;
2477 /* for external metadata checkpoint saved by mdmon can be lost
2478 * or missed /due to e.g. crash/. Check if md is not during
2479 * restart farther than metadata points to.
2480 * If so, this means metadata information is obsolete.
2482 if (st
->ss
->external
)
2483 verify_reshape_position(info
, reshape
->level
);
2484 sra
->reshape_progress
= info
->reshape_progress
;
2486 sra
->reshape_progress
= 0;
2487 if (reshape
->after
.data_disks
< reshape
->before
.data_disks
)
2488 /* start from the end of the new array */
2489 sra
->reshape_progress
= (sra
->component_size
2490 * reshape
->after
.data_disks
);
2493 ioctl(fd
, GET_ARRAY_INFO
, &array
);
2494 if (info
->array
.chunk_size
== info
->new_chunk
&&
2495 reshape
->before
.layout
== reshape
->after
.layout
&&
2496 st
->ss
->external
== 0) {
2497 /* use SET_ARRAY_INFO but only if reshape hasn't started */
2498 array
.raid_disks
= reshape
->after
.data_disks
+ reshape
->parity
;
2500 ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
2503 pr_err("Cannot set device shape for %s: %s\n",
2504 devname
, strerror(errno
));
2507 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
2508 cont_err("Bitmap must be removed before"
2509 " shape can be changed\n");
2513 } else if (!restart
) {
2514 /* set them all just in case some old 'new_*' value
2515 * persists from some earlier problem.
2518 if (sysfs_set_num(sra
, NULL
, "chunk_size", info
->new_chunk
) < 0)
2520 if (!err
&& sysfs_set_num(sra
, NULL
, "layout",
2521 reshape
->after
.layout
) < 0)
2523 if (!err
&& subarray_set_num(container
, sra
, "raid_disks",
2524 reshape
->after
.data_disks
+
2525 reshape
->parity
) < 0)
2528 pr_err("Cannot set device shape for %s\n",
2532 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
2533 cont_err("Bitmap must be removed before"
2534 " shape can be changed\n");
2543 static int reshape_array(char *container
, int fd
, char *devname
,
2544 struct supertype
*st
, struct mdinfo
*info
,
2545 int force
, struct mddev_dev
*devlist
,
2546 unsigned long long data_offset
,
2547 char *backup_file
, int verbose
, int forked
,
2548 int restart
, int freeze_reshape
)
2550 struct reshape reshape
;
2553 int orig_level
= UnSet
;
2557 struct mdu_array_info_s array
;
2560 struct mddev_dev
*dv
;
2564 unsigned long long *offsets
= NULL
;
2568 unsigned long blocks
;
2569 unsigned long long array_size
;
2571 struct mdinfo
*sra
= NULL
;
2573 /* when reshaping a RAID0, the component_size might be zero.
2574 * So try to fix that up.
2576 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
2577 dprintf("Cannot get array information.\n");
2580 if (array
.level
== 0 && info
->component_size
== 0) {
2581 get_dev_size(fd
, NULL
, &array_size
);
2582 info
->component_size
= array_size
/ array
.raid_disks
;
2585 if (array
.level
== 10)
2586 /* Need space_after info */
2587 get_space_after(fd
, st
, info
);
2589 if (info
->reshape_active
) {
2590 int new_level
= info
->new_level
;
2591 info
->new_level
= UnSet
;
2592 if (info
->delta_disks
> 0)
2593 info
->array
.raid_disks
-= info
->delta_disks
;
2594 msg
= analyse_change(info
, &reshape
);
2595 info
->new_level
= new_level
;
2596 if (info
->delta_disks
> 0)
2597 info
->array
.raid_disks
+= info
->delta_disks
;
2599 /* Make sure the array isn't read-only */
2600 ioctl(fd
, RESTART_ARRAY_RW
, 0);
2602 msg
= analyse_change(info
, &reshape
);
2604 pr_err("%s\n", msg
);
2608 (reshape
.level
!= info
->array
.level
||
2609 reshape
.before
.layout
!= info
->array
.layout
||
2610 reshape
.before
.data_disks
+ reshape
.parity
2611 != info
->array
.raid_disks
- max(0, info
->delta_disks
))) {
2612 pr_err("reshape info is not in native format -"
2613 " cannot continue.\n");
2617 if (st
->ss
->external
&& restart
&& (info
->reshape_progress
== 0)) {
2618 /* When reshape is restarted from '0', very begin of array
2619 * it is possible that for external metadata reshape and array
2620 * configuration doesn't happen.
2621 * Check if md has the same opinion, and reshape is restarted
2622 * from 0. If so, this is regular reshape start after reshape
2623 * switch in metadata to next array only.
2625 if ((verify_reshape_position(info
, reshape
.level
) >= 0) &&
2626 (info
->reshape_progress
== 0))
2630 /* reshape already started. just skip to monitoring the reshape */
2631 if (reshape
.backup_blocks
== 0)
2635 /* The container is frozen but the array may not be.
2636 * So freeze the array so spares don't get put to the wrong use
2637 * FIXME there should probably be a cleaner separation between
2638 * freeze_array and freeze_container.
2640 sysfs_freeze_array(info
);
2641 /* Check we have enough spares to not be degraded */
2643 for (dv
= devlist
; dv
; dv
=dv
->next
)
2645 spares_needed
= max(reshape
.before
.data_disks
,
2646 reshape
.after
.data_disks
)
2647 + reshape
.parity
- array
.raid_disks
;
2650 info
->new_level
> 1 && info
->array
.level
> 1 &&
2651 spares_needed
> info
->array
.spare_disks
+ added_disks
) {
2652 pr_err("Need %d spare%s to avoid degraded array,"
2653 " and only have %d.\n"
2654 " Use --force to over-ride this check.\n",
2656 spares_needed
== 1 ? "" : "s",
2657 info
->array
.spare_disks
+ added_disks
);
2660 /* Check we have enough spares to not fail */
2661 spares_needed
= max(reshape
.before
.data_disks
,
2662 reshape
.after
.data_disks
)
2664 if ((info
->new_level
> 1 || info
->new_level
== 0) &&
2665 spares_needed
> info
->array
.spare_disks
+added_disks
) {
2666 pr_err("Need %d spare%s to create working array,"
2667 " and only have %d.\n",
2669 spares_needed
== 1 ? "" : "s",
2670 info
->array
.spare_disks
+ added_disks
);
2674 if (reshape
.level
!= array
.level
) {
2675 char *c
= map_num(pers
, reshape
.level
);
2680 err
= sysfs_set_str(info
, NULL
, "level", c
);
2683 pr_err("%s: could not set level to %s\n",
2686 (info
->array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
2687 cont_err("Bitmap must be removed"
2688 " before level can be changed\n");
2692 pr_err("level of %s changed to %s\n",
2694 orig_level
= array
.level
;
2695 sysfs_freeze_array(info
);
2697 if (reshape
.level
> 0 && st
->ss
->external
) {
2698 /* make sure mdmon is aware of the new level */
2699 if (mdmon_running(container
))
2700 flush_mdmon(container
);
2702 if (!mdmon_running(container
))
2703 start_mdmon(container
);
2704 ping_monitor(container
);
2705 if (mdmon_running(container
) &&
2706 st
->update_tail
== NULL
)
2707 st
->update_tail
= &st
->updates
;
2710 /* ->reshape_super might have chosen some spares from the
2711 * container that it wants to be part of the new array.
2712 * We can collect them with ->container_content and give
2713 * them to the kernel.
2715 if (st
->ss
->reshape_super
&& st
->ss
->container_content
) {
2716 char *subarray
= strchr(info
->text_version
+1, '/')+1;
2717 struct mdinfo
*info2
=
2718 st
->ss
->container_content(st
, subarray
);
2722 sysfs_init(info2
, fd
, st
->devnm
);
2723 /* When increasing number of devices, we need to set
2724 * new raid_disks before adding these, or they might
2727 if (reshape
.backup_blocks
&&
2728 reshape
.after
.data_disks
> reshape
.before
.data_disks
)
2729 subarray_set_num(container
, info2
, "raid_disks",
2730 reshape
.after
.data_disks
+
2732 for (d
= info2
->devs
; d
; d
= d
->next
) {
2733 if (d
->disk
.state
== 0 &&
2734 d
->disk
.raid_disk
>= 0) {
2735 /* This is a spare that wants to
2736 * be part of the array.
2738 add_disk(fd
, st
, info2
, d
);
2744 /* We might have been given some devices to add to the
2745 * array. Now that the array has been changed to the right
2746 * level and frozen, we can safely add them.
2749 Manage_subdevs(devname
, fd
, devlist
, verbose
,
2752 if (reshape
.backup_blocks
== 0) {
2753 /* No restriping needed, but we might need to impose
2754 * some more changes: layout, raid_disks, chunk_size
2756 /* read current array info */
2757 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
2758 dprintf("Cannot get array information.\n");
2761 /* compare current array info with new values and if
2762 * it is different update them to new */
2763 if (info
->new_layout
!= UnSet
&&
2764 info
->new_layout
!= array
.layout
) {
2765 array
.layout
= info
->new_layout
;
2766 if (ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
2767 pr_err("failed to set new layout\n");
2769 } else if (verbose
>= 0)
2770 printf("layout for %s set to %d\n",
2771 devname
, array
.layout
);
2773 if (info
->delta_disks
!= UnSet
&&
2774 info
->delta_disks
!= 0 &&
2775 array
.raid_disks
!= (info
->array
.raid_disks
+ info
->delta_disks
)) {
2776 array
.raid_disks
+= info
->delta_disks
;
2777 if (ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
2778 pr_err("failed to set raid disks\n");
2780 } else if (verbose
>= 0) {
2781 printf("raid_disks for %s set to %d\n",
2782 devname
, array
.raid_disks
);
2785 if (info
->new_chunk
!= 0 &&
2786 info
->new_chunk
!= array
.chunk_size
) {
2787 if (sysfs_set_num(info
, NULL
,
2788 "chunk_size", info
->new_chunk
) != 0) {
2789 pr_err("failed to set chunk size\n");
2791 } else if (verbose
>= 0)
2792 printf("chunk size for %s set to %d\n",
2793 devname
, array
.chunk_size
);
2800 * There are three possibilities.
2801 * 1/ The array will shrink.
2802 * We need to ensure the reshape will pause before reaching
2803 * the 'critical section'. We also need to fork and wait for
2804 * that to happen. When it does we
2805 * suspend/backup/complete/unfreeze
2807 * 2/ The array will not change size.
2808 * This requires that we keep a backup of a sliding window
2809 * so that we can restore data after a crash. So we need
2810 * to fork and monitor progress.
2811 * In future we will allow the data_offset to change, so
2812 * a sliding backup becomes unnecessary.
2814 * 3/ The array will grow. This is relatively easy.
2815 * However the kernel's restripe routines will cheerfully
2816 * overwrite some early data before it is safe. So we
2817 * need to make a backup of the early parts of the array
2818 * and be ready to restore it if rebuild aborts very early.
2819 * For externally managed metadata, we still need a forked
2820 * child to monitor the reshape and suspend IO over the region
2821 * that is being reshaped.
2823 * We backup data by writing it to one spare, or to a
2824 * file which was given on command line.
2826 * In each case, we first make sure that storage is available
2827 * for the required backup.
2829 * - request the shape change.
2830 * - fork to handle backup etc.
2832 /* Check that we can hold all the data */
2833 get_dev_size(fd
, NULL
, &array_size
);
2834 if (reshape
.new_size
< (array_size
/512)) {
2835 pr_err("this change will reduce the size of the array.\n"
2836 " use --grow --array-size first to truncate array.\n"
2837 " e.g. mdadm --grow %s --array-size %llu\n",
2838 devname
, reshape
.new_size
/2);
2844 if (array
.level
== 10) {
2845 /* Reshaping RAID10 does not require any data backup by
2846 * user-space. Instead it requires that the data_offset
2847 * is changed to avoid the need for backup.
2848 * So this is handled very separately
2851 /* Nothing to do. */
2853 return raid10_reshape(container
, fd
, devname
, st
, info
,
2854 &reshape
, data_offset
,
2857 sra
= sysfs_read(fd
, NULL
,
2858 GET_COMPONENT
|GET_DEVS
|GET_OFFSET
|GET_STATE
|GET_CHUNK
|
2861 pr_err("%s: Cannot get array details from sysfs\n",
2866 switch(set_new_data_offset(sra
, st
, devname
, info
->delta_disks
,
2868 reshape
.min_offset_change
)) {
2872 /* Updated data_offset, so it's easy now */
2873 update_cache_size(container
, sra
, info
,
2874 min(reshape
.before
.data_disks
,
2875 reshape
.after
.data_disks
),
2876 reshape
.backup_blocks
);
2878 /* Right, everything seems fine. Let's kick things off.
2882 if (impose_reshape(sra
, info
, st
, fd
, restart
,
2883 devname
, container
, &reshape
) < 0)
2885 if (sysfs_set_str(sra
, NULL
, "sync_action", "reshape") < 0) {
2886 pr_err("Failed to initiate reshape!\n");
2891 case 1: /* Couldn't set data_offset, try the old way */
2892 if (data_offset
!= INVALID_SECTORS
) {
2893 pr_err("Cannot update data_offset on this array\n");
2899 /* Decide how many blocks (sectors) for a reshape
2900 * unit. The number we have so far is just a minimum
2902 blocks
= reshape
.backup_blocks
;
2903 if (reshape
.before
.data_disks
==
2904 reshape
.after
.data_disks
) {
2905 /* Make 'blocks' bigger for better throughput, but
2906 * not so big that we reject it below.
2907 * Try for 16 megabytes
2909 while (blocks
* 32 < sra
->component_size
&&
2913 pr_err("Need to backup %luK of critical "
2914 "section..\n", blocks
/2);
2916 if (blocks
>= sra
->component_size
/2) {
2917 pr_err("%s: Something wrong"
2918 " - reshape aborted\n",
2923 /* Now we need to open all these devices so we can read/write.
2925 nrdisks
= max(reshape
.before
.data_disks
,
2926 reshape
.after
.data_disks
) + reshape
.parity
2927 + sra
->array
.spare_disks
;
2928 fdlist
= xcalloc((1+nrdisks
), sizeof(int));
2929 offsets
= xcalloc((1+nrdisks
), sizeof(offsets
[0]));
2931 odisks
= reshape
.before
.data_disks
+ reshape
.parity
;
2932 d
= reshape_prepare_fdlist(devname
, sra
, odisks
,
2933 nrdisks
, blocks
, backup_file
,
2938 if ((st
->ss
->manage_reshape
== NULL
) ||
2939 (st
->ss
->recover_backup
== NULL
)) {
2940 if (backup_file
== NULL
) {
2941 if (reshape
.after
.data_disks
<=
2942 reshape
.before
.data_disks
) {
2943 pr_err("%s: Cannot grow - "
2944 "need backup-file\n", devname
);
2946 } else if (sra
->array
.spare_disks
== 0) {
2947 pr_err("%s: Cannot grow - "
2948 "need a spare or backup-file to backup "
2949 "critical section\n", devname
);
2953 if (!reshape_open_backup_file(backup_file
, fd
, devname
,
2955 fdlist
+d
, offsets
+d
,
2963 update_cache_size(container
, sra
, info
,
2964 min(reshape
.before
.data_disks
, reshape
.after
.data_disks
),
2967 /* Right, everything seems fine. Let's kick things off.
2968 * If only changing raid_disks, use ioctl, else use
2973 if (impose_reshape(sra
, info
, st
, fd
, restart
,
2974 devname
, container
, &reshape
) < 0)
2977 err
= start_reshape(sra
, restart
, reshape
.before
.data_disks
,
2978 reshape
.after
.data_disks
);
2980 pr_err("Cannot %s reshape for %s\n",
2981 restart
? "continue" : "start",
2986 sysfs_set_str(sra
, NULL
, "array_state", "active");
2987 if (freeze_reshape
) {
2991 pr_err("Reshape has to be continued from"
2992 " location %llu when root filesystem has been mounted.\n",
2993 sra
->reshape_progress
);
2997 /* Now we just need to kick off the reshape and watch, while
2998 * handling backups of the data...
2999 * This is all done by a forked background process.
3001 switch(forked
? 0 : fork()) {
3003 pr_err("Cannot run child to monitor reshape: %s\n",
3017 /* If another array on the same devices is busy, the
3018 * reshape will wait for them. This would mean that
3019 * the first section that we suspend will stay suspended
3020 * for a long time. So check on that possibility
3021 * by looking for "DELAYED" in /proc/mdstat, and if found,
3025 struct mdstat_ent
*mds
, *m
;
3027 mds
= mdstat_read(0, 0);
3028 for (m
= mds
; m
; m
= m
->next
)
3029 if (strcmp(m
->devnm
, sra
->sys_name
) == 0) {
3031 m
->percent
== RESYNC_DELAYED
)
3034 /* Haven't started the reshape thread
3041 if (delayed
== 1 && get_linux_version() < 3007000) {
3042 pr_err("Reshape is delayed, but cannot wait carefully with this kernel.\n"
3043 " You might experience problems until other reshapes complete.\n");
3047 sleep(30 - (delayed
-1) * 25);
3051 if (check_env("MDADM_GROW_VERIFY"))
3052 fd
= open(devname
, O_RDONLY
| O_DIRECT
);
3055 mlockall(MCL_FUTURE
);
3057 if (st
->ss
->external
) {
3058 /* metadata handler takes it from here */
3059 done
= st
->ss
->manage_reshape(
3060 fd
, sra
, &reshape
, st
, blocks
,
3062 d
- odisks
, fdlist
+odisks
,
3065 done
= child_monitor(
3066 fd
, sra
, &reshape
, st
, blocks
,
3068 d
- odisks
, fdlist
+odisks
,
3074 if (backup_file
&& done
)
3075 unlink(backup_file
);
3081 if (!st
->ss
->external
&&
3082 !(reshape
.before
.data_disks
!= reshape
.after
.data_disks
3083 && info
->custom_array_size
) &&
3084 info
->new_level
== reshape
.level
&&
3086 /* no need to wait for the reshape to finish as
3087 * there is nothing more to do.
3094 if (st
->ss
->external
) {
3095 /* Re-load the metadata as much could have changed */
3096 int cfd
= open_dev(st
->container_devnm
);
3098 flush_mdmon(container
);
3099 st
->ss
->free_super(st
);
3100 st
->ss
->load_container(st
, cfd
, container
);
3105 /* set new array size if required customer_array_size is used
3108 if (reshape
.before
.data_disks
!=
3109 reshape
.after
.data_disks
&&
3110 info
->custom_array_size
)
3111 set_array_size(st
, info
, info
->text_version
);
3113 if (info
->new_level
!= reshape
.level
) {
3115 c
= map_num(pers
, info
->new_level
);
3117 err
= sysfs_set_str(sra
, NULL
, "level", c
);
3119 pr_err("%s: could not set level "
3120 "to %s\n", devname
, c
);
3122 if (info
->new_level
== 0)
3123 st
->update_tail
= NULL
;
3135 if (orig_level
!= UnSet
&& sra
) {
3136 c
= map_num(pers
, orig_level
);
3137 if (c
&& sysfs_set_str(sra
, NULL
, "level", c
) == 0)
3138 pr_err("aborting level change\n");
3146 /* mdfd handle is passed to be closed in child process (after fork).
3148 int reshape_container(char *container
, char *devname
,
3150 struct supertype
*st
,
3151 struct mdinfo
*info
,
3154 int verbose
, int restart
, int freeze_reshape
)
3156 struct mdinfo
*cc
= NULL
;
3158 char last_devnm
[32] = "";
3160 /* component_size is not meaningful for a container,
3161 * so pass '0' meaning 'no change'
3164 reshape_super(st
, 0, info
->new_level
,
3165 info
->new_layout
, info
->new_chunk
,
3166 info
->array
.raid_disks
, info
->delta_disks
,
3167 backup_file
, devname
, APPLY_METADATA_CHANGES
,
3175 /* ping monitor to be sure that update is on disk
3177 ping_monitor(container
);
3180 case -1: /* error */
3181 perror("Cannot fork to complete reshape\n");
3184 default: /* parent */
3185 if (!freeze_reshape
)
3186 printf(Name
": multi-array reshape continues"
3187 " in background\n");
3194 /* close unused handle in child process
3200 /* For each member array with reshape_active,
3201 * we need to perform the reshape.
3202 * We pick the first array that needs reshaping and
3203 * reshape it. reshape_array() will re-read the metadata
3204 * so the next time through a different array should be
3205 * ready for reshape.
3206 * It is possible that the 'different' array will not
3207 * be assembled yet. In that case we simple exit.
3208 * When it is assembled, the mdadm which assembles it
3209 * will take over the reshape.
3211 struct mdinfo
*content
;
3213 struct mdstat_ent
*mdstat
;
3219 cc
= st
->ss
->container_content(st
, NULL
);
3221 for (content
= cc
; content
; content
= content
->next
) {
3223 if (!content
->reshape_active
)
3226 subarray
= strchr(content
->text_version
+1, '/')+1;
3227 mdstat
= mdstat_by_subdev(subarray
, container
);
3230 if (mdstat
->active
== 0) {
3231 pr_err("Skipping inactive array %s.\n",
3233 free_mdstat(mdstat
);
3242 devid
= devnm2devid(mdstat
->devnm
);
3243 adev
= map_dev(major(devid
), minor(devid
), 0);
3245 adev
= content
->text_version
;
3247 fd
= open_dev(mdstat
->devnm
);
3249 printf(Name
": Device %s cannot be opened for reshape.",
3254 if (strcmp(last_devnm
, mdstat
->devnm
) == 0) {
3255 /* Do not allow for multiple reshape_array() calls for
3257 * It can happen when reshape_array() returns without
3258 * error, when reshape is not finished (wrong reshape
3259 * starting/continuation conditions). Mdmon doesn't
3260 * switch to next array in container and reentry
3261 * conditions for the same array occur.
3262 * This is possibly interim until the behaviour of
3263 * reshape_array is resolved().
3265 printf(Name
": Multiple reshape execution detected for "
3266 "device %s.", adev
);
3270 strcpy(last_devnm
, mdstat
->devnm
);
3272 sysfs_init(content
, fd
, mdstat
->devnm
);
3274 if (mdmon_running(container
))
3275 flush_mdmon(container
);
3277 rv
= reshape_array(container
, fd
, adev
, st
,
3278 content
, force
, NULL
, 0ULL,
3279 backup_file
, verbose
, 1, restart
,
3283 if (freeze_reshape
) {
3292 if (mdmon_running(container
))
3293 flush_mdmon(container
);
3302 * We run a child process in the background which performs the following
3304 * - wait for resync to reach a certain point
3305 * - suspend io to the following section
3306 * - backup that section
3307 * - allow resync to proceed further
3309 * - discard the backup.
3311 * When are combined in slightly different ways in the three cases.
3313 * - suspend/backup/allow/wait/resume/discard
3315 * - allow/wait/suspend/backup/allow/wait/resume/discard
3317 * - wait/resume/discard/suspend/backup/allow
3319 * suspend/backup/allow always come together
3320 * wait/resume/discard do too.
3321 * For the same-size case we have two backups to improve flow.
3325 int progress_reshape(struct mdinfo
*info
, struct reshape
*reshape
,
3326 unsigned long long backup_point
,
3327 unsigned long long wait_point
,
3328 unsigned long long *suspend_point
,
3329 unsigned long long *reshape_completed
)
3331 /* This function is called repeatedly by the reshape manager.
3332 * It determines how much progress can safely be made and allows
3334 * - 'info' identifies the array and particularly records in
3335 * ->reshape_progress the metadata's knowledge of progress
3336 * This is a sector offset from the start of the array
3337 * of the next array block to be relocated. This number
3338 * may increase from 0 or decrease from array_size, depending
3339 * on the type of reshape that is happening.
3340 * Note that in contrast, 'sync_completed' is a block count of the
3341 * reshape so far. It gives the distance between the start point
3342 * (head or tail of device) and the next place that data will be
3343 * written. It always increases.
3344 * - 'reshape' is the structure created by analyse_change
3345 * - 'backup_point' shows how much the metadata manager has backed-up
3346 * data. For reshapes with increasing progress, it is the next address
3347 * to be backed up, previous addresses have been backed-up. For
3348 * decreasing progress, it is the earliest address that has been
3349 * backed up - later address are also backed up.
3350 * So addresses between reshape_progress and backup_point are
3351 * backed up providing those are in the 'correct' order.
3352 * - 'wait_point' is an array address. When reshape_completed
3353 * passes this point, progress_reshape should return. It might
3354 * return earlier if it determines that ->reshape_progress needs
3355 * to be updated or further backup is needed.
3356 * - suspend_point is maintained by progress_reshape and the caller
3357 * should not touch it except to initialise to zero.
3358 * It is an array address and it only increases in 2.6.37 and earlier.
3359 * This makes it difficult to handle reducing reshapes with
3360 * external metadata.
3361 * However: it is similar to backup_point in that it records the
3362 * other end of a suspended region from reshape_progress.
3363 * it is moved to extend the region that is safe to backup and/or
3365 * - reshape_completed is read from sysfs and returned. The caller
3366 * should copy this into ->reshape_progress when it has reason to
3367 * believe that the metadata knows this, and any backup outside this
3371 * 1 if more data from backup_point - but only as far as suspend_point,
3372 * should be backed up
3373 * 0 if things are progressing smoothly
3374 * -1 if the reshape is finished because it is all done,
3375 * -2 if the reshape is finished due to an error.
3378 int advancing
= (reshape
->after
.data_disks
3379 >= reshape
->before
.data_disks
);
3380 unsigned long long need_backup
; /* All data between start of array and
3381 * here will at some point need to
3384 unsigned long long read_offset
, write_offset
;
3385 unsigned long long write_range
;
3386 unsigned long long max_progress
, target
, completed
;
3387 unsigned long long array_size
= (info
->component_size
3388 * reshape
->before
.data_disks
);
3392 /* First, we unsuspend any region that is now known to be safe.
3393 * If suspend_point is on the 'wrong' side of reshape_progress, then
3394 * we don't have or need suspension at the moment. This is true for
3395 * native metadata when we don't need to back-up.
3398 if (info
->reshape_progress
<= *suspend_point
)
3399 sysfs_set_num(info
, NULL
, "suspend_lo",
3400 info
->reshape_progress
);
3402 /* Note: this won't work in 2.6.37 and before.
3403 * Something somewhere should make sure we don't need it!
3405 if (info
->reshape_progress
>= *suspend_point
)
3406 sysfs_set_num(info
, NULL
, "suspend_hi",
3407 info
->reshape_progress
);
3410 /* Now work out how far it is safe to progress.
3411 * If the read_offset for ->reshape_progress is less than
3412 * 'blocks' beyond the write_offset, we can only progress as far
3414 * Otherwise we can progress until the write_offset for the new location
3415 * reaches (within 'blocks' of) the read_offset at the current location.
3416 * However that region must be suspended unless we are using native
3418 * If we need to suspend more, we limit it to 128M per device, which is
3419 * rather arbitrary and should be some time-based calculation.
3421 read_offset
= info
->reshape_progress
/ reshape
->before
.data_disks
;
3422 write_offset
= info
->reshape_progress
/ reshape
->after
.data_disks
;
3423 write_range
= info
->new_chunk
/512;
3424 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
)
3425 need_backup
= array_size
;
3427 need_backup
= reshape
->backup_blocks
;
3429 if (read_offset
< write_offset
+ write_range
)
3430 max_progress
= backup_point
;
3434 reshape
->after
.data_disks
;
3436 if (read_offset
> write_offset
- write_range
)
3437 /* Can only progress as far as has been backed up,
3438 * which must be suspended */
3439 max_progress
= backup_point
;
3440 else if (info
->reshape_progress
<= need_backup
)
3441 max_progress
= backup_point
;
3443 if (info
->array
.major_version
>= 0)
3444 /* Can progress until backup is needed */
3445 max_progress
= need_backup
;
3447 /* Can progress until metadata update is required */
3450 reshape
->after
.data_disks
;
3451 /* but data must be suspended */
3452 if (max_progress
< *suspend_point
)
3453 max_progress
= *suspend_point
;
3458 /* We know it is safe to progress to 'max_progress' providing
3459 * it is suspended or we are using native metadata.
3460 * Consider extending suspend_point 128M per device if it
3461 * is less than 64M per device beyond reshape_progress.
3462 * But always do a multiple of 'blocks'
3463 * FIXME this is too big - it takes to long to complete
3466 target
= 64*1024*2 * min(reshape
->before
.data_disks
,
3467 reshape
->after
.data_disks
);
3468 target
/= reshape
->backup_blocks
;
3471 target
*= reshape
->backup_blocks
;
3473 /* For externally managed metadata we always need to suspend IO to
3474 * the area being reshaped so we regularly push suspend_point forward.
3475 * For native metadata we only need the suspend if we are going to do
3479 if ((need_backup
> info
->reshape_progress
3480 || info
->array
.major_version
< 0) &&
3481 *suspend_point
< info
->reshape_progress
+ target
) {
3482 if (need_backup
< *suspend_point
+ 2 * target
)
3483 *suspend_point
= need_backup
;
3484 else if (*suspend_point
+ 2 * target
< array_size
)
3485 *suspend_point
+= 2 * target
;
3487 *suspend_point
= array_size
;
3488 sysfs_set_num(info
, NULL
, "suspend_hi", *suspend_point
);
3489 if (max_progress
> *suspend_point
)
3490 max_progress
= *suspend_point
;
3493 if (info
->array
.major_version
>= 0) {
3494 /* Only need to suspend when about to backup */
3495 if (info
->reshape_progress
< need_backup
* 2 &&
3496 *suspend_point
> 0) {
3498 sysfs_set_num(info
, NULL
, "suspend_lo", 0);
3499 sysfs_set_num(info
, NULL
, "suspend_hi", need_backup
);
3502 /* Need to suspend continually */
3503 if (info
->reshape_progress
< *suspend_point
)
3504 *suspend_point
= info
->reshape_progress
;
3505 if (*suspend_point
+ target
< info
->reshape_progress
)
3506 /* No need to move suspend region yet */;
3508 if (*suspend_point
>= 2 * target
)
3509 *suspend_point
-= 2 * target
;
3512 sysfs_set_num(info
, NULL
, "suspend_lo",
3515 if (max_progress
< *suspend_point
)
3516 max_progress
= *suspend_point
;
3520 /* now set sync_max to allow that progress. sync_max, like
3521 * sync_completed is a count of sectors written per device, so
3522 * we find the difference between max_progress and the start point,
3523 * and divide that by after.data_disks to get a sync_max
3525 * At the same time we convert wait_point to a similar number
3526 * for comparing against sync_completed.
3528 /* scale down max_progress to per_disk */
3529 max_progress
/= reshape
->after
.data_disks
;
3530 /* Round to chunk size as some kernels give an erroneously high number */
3531 max_progress
/= info
->new_chunk
/512;
3532 max_progress
*= info
->new_chunk
/512;
3533 /* And round to old chunk size as the kernel wants that */
3534 max_progress
/= info
->array
.chunk_size
/512;
3535 max_progress
*= info
->array
.chunk_size
/512;
3536 /* Limit progress to the whole device */
3537 if (max_progress
> info
->component_size
)
3538 max_progress
= info
->component_size
;
3539 wait_point
/= reshape
->after
.data_disks
;
3541 /* switch from 'device offset' to 'processed block count' */
3542 max_progress
= info
->component_size
- max_progress
;
3543 wait_point
= info
->component_size
- wait_point
;
3546 sysfs_set_num(info
, NULL
, "sync_max", max_progress
);
3548 /* Now wait. If we have already reached the point that we were
3549 * asked to wait to, don't wait at all, else wait for any change.
3550 * We need to select on 'sync_completed' as that is the place that
3551 * notifications happen, but we are really interested in
3552 * 'reshape_position'
3554 fd
= sysfs_get_fd(info
, NULL
, "sync_completed");
3556 goto check_progress
;
3558 if (sysfs_fd_get_ll(fd
, &completed
) < 0)
3559 goto check_progress
;
3561 while (completed
< max_progress
&& completed
< wait_point
) {
3562 /* Check that sync_action is still 'reshape' to avoid
3563 * waiting forever on a dead array
3567 if (sysfs_get_str(info
, NULL
, "sync_action",
3569 strncmp(action
, "reshape", 7) != 0)
3571 /* Some kernels reset 'sync_completed' to zero
3572 * before setting 'sync_action' to 'idle'.
3573 * So we need these extra tests.
3575 if (completed
== 0 && advancing
3576 && info
->reshape_progress
> 0)
3578 if (completed
== 0 && !advancing
3579 && info
->reshape_progress
< (info
->component_size
3580 * reshape
->after
.data_disks
))
3584 select(fd
+1, NULL
, NULL
, &rfds
, NULL
);
3585 if (sysfs_fd_get_ll(fd
, &completed
) < 0)
3586 goto check_progress
;
3588 /* Some kernels reset 'sync_completed' to zero,
3589 * we need to have real point we are in md
3592 completed
= max_progress
;
3594 /* some kernels can give an incorrectly high 'completed' number */
3595 completed
/= (info
->new_chunk
/512);
3596 completed
*= (info
->new_chunk
/512);
3597 /* Convert 'completed' back in to a 'progress' number */
3598 completed
*= reshape
->after
.data_disks
;
3600 completed
= info
->component_size
* reshape
->after
.data_disks
3603 *reshape_completed
= completed
;
3607 /* We return the need_backup flag. Caller will decide
3608 * how much - a multiple of ->backup_blocks up to *suspend_point
3611 return need_backup
> info
->reshape_progress
;
3613 return need_backup
>= info
->reshape_progress
;
3616 /* if we couldn't read a number from sync_completed, then
3617 * either the reshape did complete, or it aborted.
3618 * We can tell which by checking for 'none' in reshape_position.
3619 * If it did abort, then it might immediately restart if it
3620 * it was just a device failure that leaves us degraded but
3624 if (sysfs_get_str(info
, NULL
, "reshape_position", buf
, sizeof(buf
)) < 0
3625 || strncmp(buf
, "none", 4) != 0) {
3626 /* The abort might only be temporary. Wait up to 10
3627 * seconds for fd to contain a valid number again.
3633 while (fd
>= 0 && rv
< 0 && tv
.tv_sec
> 0) {
3637 if (select(fd
+1, NULL
, NULL
, &rfds
, &tv
) != 1)
3639 switch (sysfs_fd_get_ll(fd
, &completed
)) {
3641 /* all good again */
3644 case -2: /* read error - abort */
3651 return rv
; /* abort */
3653 /* Maybe racing with array shutdown - check state */
3656 if (sysfs_get_str(info
, NULL
, "array_state", buf
, sizeof(buf
)) < 0
3657 || strncmp(buf
, "inactive", 8) == 0
3658 || strncmp(buf
, "clear",5) == 0)
3659 return -2; /* abort */
3660 return -1; /* complete */
3664 /* FIXME return status is never checked */
3665 static int grow_backup(struct mdinfo
*sra
,
3666 unsigned long long offset
, /* per device */
3667 unsigned long stripes
, /* per device, in old chunks */
3668 int *sources
, unsigned long long *offsets
,
3669 int disks
, int chunk
, int level
, int layout
,
3670 int dests
, int *destfd
, unsigned long long *destoffsets
,
3671 int part
, int *degraded
,
3674 /* Backup 'blocks' sectors at 'offset' on each device of the array,
3675 * to storage 'destfd' (offset 'destoffsets'), after first
3676 * suspending IO. Then allow resync to continue
3677 * over the suspended section.
3678 * Use part 'part' of the backup-super-block.
3683 unsigned long long ll
;
3685 //printf("offset %llu\n", offset);
3691 /* Check that array hasn't become degraded, else we might backup the wrong data */
3692 if (sysfs_get_ll(sra
, NULL
, "degraded", &ll
) < 0)
3693 return -1; /* FIXME this error is ignored */
3694 new_degraded
= (int)ll
;
3695 if (new_degraded
!= *degraded
) {
3696 /* check each device to ensure it is still working */
3698 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
3699 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
3701 if (sd
->disk
.state
& (1<<MD_DISK_SYNC
)) {
3703 if (sysfs_get_str(sra
, sd
, "state", sbuf
, 20) < 0 ||
3704 strstr(sbuf
, "faulty") ||
3705 strstr(sbuf
, "in_sync") == NULL
) {
3706 /* this device is dead */
3707 sd
->disk
.state
= (1<<MD_DISK_FAULTY
);
3708 if (sd
->disk
.raid_disk
>= 0 &&
3709 sources
[sd
->disk
.raid_disk
] >= 0) {
3710 close(sources
[sd
->disk
.raid_disk
]);
3711 sources
[sd
->disk
.raid_disk
] = -1;
3716 *degraded
= new_degraded
;
3719 bsb
.arraystart2
= __cpu_to_le64(offset
* odata
);
3720 bsb
.length2
= __cpu_to_le64(stripes
* (chunk
/512) * odata
);
3722 bsb
.arraystart
= __cpu_to_le64(offset
* odata
);
3723 bsb
.length
= __cpu_to_le64(stripes
* (chunk
/512) * odata
);
3726 bsb
.magic
[15] = '2';
3727 for (i
= 0; i
< dests
; i
++)
3729 lseek64(destfd
[i
], destoffsets
[i
] + __le64_to_cpu(bsb
.devstart2
)*512, 0);
3731 lseek64(destfd
[i
], destoffsets
[i
], 0);
3733 rv
= save_stripes(sources
, offsets
,
3734 disks
, chunk
, level
, layout
,
3736 offset
*512*odata
, stripes
* chunk
* odata
,
3741 bsb
.mtime
= __cpu_to_le64(time(0));
3742 for (i
= 0; i
< dests
; i
++) {
3743 bsb
.devstart
= __cpu_to_le64(destoffsets
[i
]/512);
3745 bsb
.sb_csum
= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
));
3746 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0)
3747 bsb
.sb_csum2
= bsb_csum((char*)&bsb
,
3748 ((char*)&bsb
.sb_csum2
)-((char*)&bsb
));
3751 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
] - 4096, 0)
3752 != destoffsets
[i
] - 4096)
3754 if (write(destfd
[i
], &bsb
, 512) != 512)
3756 if (destoffsets
[i
] > 4096) {
3757 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
]+stripes
*chunk
*odata
, 0) !=
3758 destoffsets
[i
]+stripes
*chunk
*odata
)
3760 if (write(destfd
[i
], &bsb
, 512) != 512)
3770 /* in 2.6.30, the value reported by sync_completed can be
3771 * less that it should be by one stripe.
3772 * This only happens when reshape hits sync_max and pauses.
3773 * So allow wait_backup to either extent sync_max further
3774 * than strictly necessary, or return before the
3775 * sync has got quite as far as we would really like.
3776 * This is what 'blocks2' is for.
3777 * The various caller give appropriate values so that
3780 /* FIXME return value is often ignored */
3781 static int forget_backup(int dests
, int *destfd
,
3782 unsigned long long *destoffsets
,
3786 * Erase backup 'part' (which is 0 or 1)
3792 bsb
.arraystart2
= __cpu_to_le64(0);
3793 bsb
.length2
= __cpu_to_le64(0);
3795 bsb
.arraystart
= __cpu_to_le64(0);
3796 bsb
.length
= __cpu_to_le64(0);
3798 bsb
.mtime
= __cpu_to_le64(time(0));
3800 for (i
= 0; i
< dests
; i
++) {
3801 bsb
.devstart
= __cpu_to_le64(destoffsets
[i
]/512);
3802 bsb
.sb_csum
= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
));
3803 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0)
3804 bsb
.sb_csum2
= bsb_csum((char*)&bsb
,
3805 ((char*)&bsb
.sb_csum2
)-((char*)&bsb
));
3806 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
]-4096, 0) !=
3807 destoffsets
[i
]-4096)
3810 write(destfd
[i
], &bsb
, 512) != 512)
3817 static void fail(char *msg
)
3820 rv
= (write(2, msg
, strlen(msg
)) != (int)strlen(msg
));
3821 rv
|= (write(2, "\n", 1) != 1);
3825 static char *abuf
, *bbuf
;
3826 static unsigned long long abuflen
;
3827 static void validate(int afd
, int bfd
, unsigned long long offset
)
3829 /* check that the data in the backup against the array.
3830 * This is only used for regression testing and should not
3831 * be used while the array is active
3835 lseek64(bfd
, offset
- 4096, 0);
3836 if (read(bfd
, &bsb2
, 512) != 512)
3837 fail("cannot read bsb");
3838 if (bsb2
.sb_csum
!= bsb_csum((char*)&bsb2
,
3839 ((char*)&bsb2
.sb_csum
)-((char*)&bsb2
)))
3840 fail("first csum bad");
3841 if (memcmp(bsb2
.magic
, "md_backup_data", 14) != 0)
3842 fail("magic is bad");
3843 if (memcmp(bsb2
.magic
, "md_backup_data-2", 16) == 0 &&
3844 bsb2
.sb_csum2
!= bsb_csum((char*)&bsb2
,
3845 ((char*)&bsb2
.sb_csum2
)-((char*)&bsb2
)))
3846 fail("second csum bad");
3848 if (__le64_to_cpu(bsb2
.devstart
)*512 != offset
)
3849 fail("devstart is wrong");
3852 unsigned long long len
= __le64_to_cpu(bsb2
.length
)*512;
3854 if (abuflen
< len
) {
3858 if (posix_memalign((void**)&abuf
, 4096, abuflen
) ||
3859 posix_memalign((void**)&bbuf
, 4096, abuflen
)) {
3861 /* just stop validating on mem-alloc failure */
3866 lseek64(bfd
, offset
, 0);
3867 if ((unsigned long long)read(bfd
, bbuf
, len
) != len
) {
3868 //printf("len %llu\n", len);
3869 fail("read first backup failed");
3871 lseek64(afd
, __le64_to_cpu(bsb2
.arraystart
)*512, 0);
3872 if ((unsigned long long)read(afd
, abuf
, len
) != len
)
3873 fail("read first from array failed");
3874 if (memcmp(bbuf
, abuf
, len
) != 0) {
3877 printf("offset=%llu len=%llu\n",
3878 (unsigned long long)__le64_to_cpu(bsb2
.arraystart
)*512, len
);
3879 for (i
=0; i
<len
; i
++)
3880 if (bbuf
[i
] != abuf
[i
]) {
3881 printf("first diff byte %d\n", i
);
3885 fail("data1 compare failed");
3889 unsigned long long len
= __le64_to_cpu(bsb2
.length2
)*512;
3891 if (abuflen
< len
) {
3895 abuf
= xmalloc(abuflen
);
3896 bbuf
= xmalloc(abuflen
);
3899 lseek64(bfd
, offset
+__le64_to_cpu(bsb2
.devstart2
)*512, 0);
3900 if ((unsigned long long)read(bfd
, bbuf
, len
) != len
)
3901 fail("read second backup failed");
3902 lseek64(afd
, __le64_to_cpu(bsb2
.arraystart2
)*512, 0);
3903 if ((unsigned long long)read(afd
, abuf
, len
) != len
)
3904 fail("read second from array failed");
3905 if (memcmp(bbuf
, abuf
, len
) != 0)
3906 fail("data2 compare failed");
3910 int child_monitor(int afd
, struct mdinfo
*sra
, struct reshape
*reshape
,
3911 struct supertype
*st
, unsigned long blocks
,
3912 int *fds
, unsigned long long *offsets
,
3913 int dests
, int *destfd
, unsigned long long *destoffsets
)
3915 /* Monitor a reshape where backup is being performed using
3916 * 'native' mechanism - either to a backup file, or
3917 * to some space in a spare.
3921 unsigned long long speed
;
3922 unsigned long long suspend_point
, array_size
;
3923 unsigned long long backup_point
, wait_point
;
3924 unsigned long long reshape_completed
;
3926 int increasing
= reshape
->after
.data_disks
>= reshape
->before
.data_disks
;
3927 int part
= 0; /* The next part of the backup area to fill. It may already
3928 * be full, so we need to check */
3929 int level
= reshape
->level
;
3930 int layout
= reshape
->before
.layout
;
3931 int data
= reshape
->before
.data_disks
;
3932 int disks
= reshape
->before
.data_disks
+ reshape
->parity
;
3933 int chunk
= sra
->array
.chunk_size
;
3935 unsigned long stripes
;
3938 /* set up the backup-super-block. This requires the
3939 * uuid from the array.
3941 /* Find a superblock */
3942 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
3946 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
3948 dn
= map_dev(sd
->disk
.major
, sd
->disk
.minor
, 1);
3949 devfd
= dev_open(dn
, O_RDONLY
);
3952 ok
= st
->ss
->load_super(st
, devfd
, NULL
);
3958 pr_err("Cannot find a superblock\n");
3962 memset(&bsb
, 0, 512);
3963 memcpy(bsb
.magic
, "md_backup_data-1", 16);
3964 st
->ss
->uuid_from_super(st
, uuid
);
3965 memcpy(bsb
.set_uuid
, uuid
, 16);
3966 bsb
.mtime
= __cpu_to_le64(time(0));
3967 bsb
.devstart2
= blocks
;
3969 stripes
= blocks
/ (sra
->array
.chunk_size
/512) /
3970 reshape
->before
.data_disks
;
3972 if (posix_memalign((void**)&buf
, 4096, disks
* chunk
))
3973 /* Don't start the 'reshape' */
3975 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
) {
3976 sysfs_get_ll(sra
, NULL
, "sync_speed_min", &speed
);
3977 sysfs_set_num(sra
, NULL
, "sync_speed_min", 200000);
3981 array_size
= sra
->component_size
* reshape
->after
.data_disks
;
3982 backup_point
= sra
->reshape_progress
;
3985 array_size
= sra
->component_size
* reshape
->before
.data_disks
;
3986 backup_point
= reshape
->backup_blocks
;
3987 suspend_point
= array_size
;
3993 /* Want to return as soon the oldest backup slot can
3994 * be released as that allows us to start backing up
3995 * some more, providing suspend_point has been
3996 * advanced, which it should have.
3999 wait_point
= array_size
;
4000 if (part
== 0 && __le64_to_cpu(bsb
.length
) > 0)
4001 wait_point
= (__le64_to_cpu(bsb
.arraystart
) +
4002 __le64_to_cpu(bsb
.length
));
4003 if (part
== 1 && __le64_to_cpu(bsb
.length2
) > 0)
4004 wait_point
= (__le64_to_cpu(bsb
.arraystart2
) +
4005 __le64_to_cpu(bsb
.length2
));
4008 if (part
== 0 && __le64_to_cpu(bsb
.length
) > 0)
4009 wait_point
= __le64_to_cpu(bsb
.arraystart
);
4010 if (part
== 1 && __le64_to_cpu(bsb
.length2
) > 0)
4011 wait_point
= __le64_to_cpu(bsb
.arraystart2
);
4014 rv
= progress_reshape(sra
, reshape
,
4015 backup_point
, wait_point
,
4016 &suspend_point
, &reshape_completed
);
4017 /* external metadata would need to ping_monitor here */
4018 sra
->reshape_progress
= reshape_completed
;
4020 /* Clear any backup region that is before 'here' */
4022 if (__le64_to_cpu(bsb
.length
) > 0 &&
4023 reshape_completed
>= (__le64_to_cpu(bsb
.arraystart
) +
4024 __le64_to_cpu(bsb
.length
)))
4025 forget_backup(dests
, destfd
,
4027 if (__le64_to_cpu(bsb
.length2
) > 0 &&
4028 reshape_completed
>= (__le64_to_cpu(bsb
.arraystart2
) +
4029 __le64_to_cpu(bsb
.length2
)))
4030 forget_backup(dests
, destfd
,
4033 if (__le64_to_cpu(bsb
.length
) > 0 &&
4034 reshape_completed
<= (__le64_to_cpu(bsb
.arraystart
)))
4035 forget_backup(dests
, destfd
,
4037 if (__le64_to_cpu(bsb
.length2
) > 0 &&
4038 reshape_completed
<= (__le64_to_cpu(bsb
.arraystart2
)))
4039 forget_backup(dests
, destfd
,
4048 if (rv
== 0 && increasing
&& !st
->ss
->external
) {
4049 /* No longer need to monitor this reshape */
4055 unsigned long long offset
;
4056 unsigned long actual_stripes
;
4057 /* Need to backup some data.
4058 * If 'part' is not used and the desired
4059 * backup size is suspended, do a backup,
4060 * then consider the next part.
4062 /* Check that 'part' is unused */
4063 if (part
== 0 && __le64_to_cpu(bsb
.length
) != 0)
4065 if (part
== 1 && __le64_to_cpu(bsb
.length2
) != 0)
4068 offset
= backup_point
/ data
;
4069 actual_stripes
= stripes
;
4071 if (offset
+ actual_stripes
* (chunk
/512) >
4072 sra
->component_size
)
4073 actual_stripes
= ((sra
->component_size
- offset
)
4075 if (offset
+ actual_stripes
* (chunk
/512) >
4079 if (offset
< actual_stripes
* (chunk
/512))
4080 actual_stripes
= offset
/ (chunk
/512);
4081 offset
-= actual_stripes
* (chunk
/512);
4082 if (offset
< suspend_point
/data
)
4085 if (actual_stripes
== 0)
4087 grow_backup(sra
, offset
, actual_stripes
,
4089 disks
, chunk
, level
, layout
,
4090 dests
, destfd
, destoffsets
,
4091 part
, °raded
, buf
);
4092 validate(afd
, destfd
[0], destoffsets
[0]);
4093 /* record where 'part' is up to */
4096 backup_point
+= actual_stripes
* (chunk
/512) * data
;
4098 backup_point
-= actual_stripes
* (chunk
/512) * data
;
4102 /* FIXME maybe call progress_reshape one more time instead */
4103 abort_reshape(sra
); /* remove any remaining suspension */
4104 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
)
4105 sysfs_set_num(sra
, NULL
, "sync_speed_min", speed
);
4111 * If any spare contains md_back_data-1 which is recent wrt mtime,
4112 * write that data into the array and update the super blocks with
4113 * the new reshape_progress
4115 int Grow_restart(struct supertype
*st
, struct mdinfo
*info
, int *fdlist
, int cnt
,
4116 char *backup_file
, int verbose
)
4120 unsigned long long *offsets
;
4121 unsigned long long nstripe
, ostripe
;
4124 odata
= info
->array
.raid_disks
- info
->delta_disks
- 1;
4125 if (info
->array
.level
== 6) odata
--; /* number of data disks */
4126 ndata
= info
->array
.raid_disks
- 1;
4127 if (info
->new_level
== 6) ndata
--;
4129 old_disks
= info
->array
.raid_disks
- info
->delta_disks
;
4131 if (info
->delta_disks
<= 0)
4132 /* Didn't grow, so the backup file must have
4136 for (i
=old_disks
-(backup_file
?1:0); i
<cnt
; i
++) {
4137 struct mdinfo dinfo
;
4140 char *devname
, namebuf
[20];
4141 unsigned long long lo
, hi
;
4143 /* This was a spare and may have some saved data on it.
4144 * Load the superblock, find and load the
4145 * backup_super_block.
4146 * If either fail, go on to next device.
4147 * If the backup contains no new info, just return
4148 * else restore data and update all superblocks
4150 if (i
== old_disks
-1) {
4151 fd
= open(backup_file
, O_RDONLY
);
4153 pr_err("backup file %s inaccessible: %s\n",
4154 backup_file
, strerror(errno
));
4157 devname
= backup_file
;
4162 if (st
->ss
->load_super(st
, fd
, NULL
))
4165 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
4166 st
->ss
->free_super(st
);
4169 (dinfo
.data_offset
+ dinfo
.component_size
- 8) <<9,
4171 pr_err("Cannot seek on device %d\n", i
);
4172 continue; /* Cannot seek */
4174 sprintf(namebuf
, "device-%d", i
);
4177 if (read(fd
, &bsb
, sizeof(bsb
)) != sizeof(bsb
)) {
4179 pr_err("Cannot read from %s\n", devname
);
4180 continue; /* Cannot read */
4182 if (memcmp(bsb
.magic
, "md_backup_data-1", 16) != 0 &&
4183 memcmp(bsb
.magic
, "md_backup_data-2", 16) != 0) {
4185 pr_err("No backup metadata on %s\n", devname
);
4188 if (bsb
.sb_csum
!= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
))) {
4190 pr_err("Bad backup-metadata checksum on %s\n", devname
);
4191 continue; /* bad checksum */
4193 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0 &&
4194 bsb
.sb_csum2
!= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum2
)-((char*)&bsb
))) {
4196 pr_err("Bad backup-metadata checksum2 on %s\n", devname
);
4197 continue; /* Bad second checksum */
4199 if (memcmp(bsb
.set_uuid
,info
->uuid
, 16) != 0) {
4201 pr_err("Wrong uuid on backup-metadata on %s\n", devname
);
4202 continue; /* Wrong uuid */
4205 /* array utime and backup-mtime should be updated at much the same time, but it seems that
4206 * sometimes they aren't... So allow considerable flexability in matching, and allow
4207 * this test to be overridden by an environment variable.
4209 if (info
->array
.utime
> (int)__le64_to_cpu(bsb
.mtime
) + 2*60*60 ||
4210 info
->array
.utime
< (int)__le64_to_cpu(bsb
.mtime
) - 10*60) {
4211 if (check_env("MDADM_GROW_ALLOW_OLD")) {
4212 pr_err("accepting backup with timestamp %lu "
4213 "for array with timestamp %lu\n",
4214 (unsigned long)__le64_to_cpu(bsb
.mtime
),
4215 (unsigned long)info
->array
.utime
);
4217 pr_err("too-old timestamp on backup-metadata on %s\n", devname
);
4218 pr_err("If you think it is should be safe, try 'export MDADM_GROW_ALLOW_OLD=1'\n");
4219 continue; /* time stamp is too bad */
4223 if (bsb
.magic
[15] == '1') {
4224 if (bsb
.length
== 0)
4226 if (info
->delta_disks
>= 0) {
4227 /* reshape_progress is increasing */
4228 if (__le64_to_cpu(bsb
.arraystart
)
4229 + __le64_to_cpu(bsb
.length
)
4230 < info
->reshape_progress
) {
4233 pr_err("backup-metadata found on %s but is not needed\n", devname
);
4234 continue; /* No new data here */
4237 /* reshape_progress is decreasing */
4238 if (__le64_to_cpu(bsb
.arraystart
) >=
4239 info
->reshape_progress
)
4240 goto nonew
; /* No new data here */
4243 if (bsb
.length
== 0 && bsb
.length2
== 0)
4245 if (info
->delta_disks
>= 0) {
4246 /* reshape_progress is increasing */
4247 if ((__le64_to_cpu(bsb
.arraystart
)
4248 + __le64_to_cpu(bsb
.length
)
4249 < info
->reshape_progress
)
4251 (__le64_to_cpu(bsb
.arraystart2
)
4252 + __le64_to_cpu(bsb
.length2
)
4253 < info
->reshape_progress
))
4254 goto nonew
; /* No new data here */
4256 /* reshape_progress is decreasing */
4257 if (__le64_to_cpu(bsb
.arraystart
) >=
4258 info
->reshape_progress
&&
4259 __le64_to_cpu(bsb
.arraystart2
) >=
4260 info
->reshape_progress
)
4261 goto nonew
; /* No new data here */
4264 if (lseek64(fd
, __le64_to_cpu(bsb
.devstart
)*512, 0)< 0) {
4267 pr_err("Failed to verify secondary backup-metadata block on %s\n",
4269 continue; /* Cannot seek */
4271 /* There should be a duplicate backup superblock 4k before here */
4272 if (lseek64(fd
, -4096, 1) < 0 ||
4273 read(fd
, &bsb2
, sizeof(bsb2
)) != sizeof(bsb2
))
4274 goto second_fail
; /* Cannot find leading superblock */
4275 if (bsb
.magic
[15] == '1')
4276 bsbsize
= offsetof(struct mdp_backup_super
, pad1
);
4278 bsbsize
= offsetof(struct mdp_backup_super
, pad
);
4279 if (memcmp(&bsb2
, &bsb
, bsbsize
) != 0)
4280 goto second_fail
; /* Cannot find leading superblock */
4282 /* Now need the data offsets for all devices. */
4283 offsets
= xmalloc(sizeof(*offsets
)*info
->array
.raid_disks
);
4284 for(j
=0; j
<info
->array
.raid_disks
; j
++) {
4287 if (st
->ss
->load_super(st
, fdlist
[j
], NULL
))
4288 /* FIXME should be this be an error */
4290 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
4291 st
->ss
->free_super(st
);
4292 offsets
[j
] = dinfo
.data_offset
* 512;
4294 printf(Name
": restoring critical section\n");
4296 if (restore_stripes(fdlist
, offsets
,
4297 info
->array
.raid_disks
,
4301 fd
, __le64_to_cpu(bsb
.devstart
)*512,
4302 __le64_to_cpu(bsb
.arraystart
)*512,
4303 __le64_to_cpu(bsb
.length
)*512, NULL
)) {
4304 /* didn't succeed, so giveup */
4306 pr_err("Error restoring backup from %s\n",
4312 if (bsb
.magic
[15] == '2' &&
4313 restore_stripes(fdlist
, offsets
,
4314 info
->array
.raid_disks
,
4318 fd
, __le64_to_cpu(bsb
.devstart
)*512 +
4319 __le64_to_cpu(bsb
.devstart2
)*512,
4320 __le64_to_cpu(bsb
.arraystart2
)*512,
4321 __le64_to_cpu(bsb
.length2
)*512, NULL
)) {
4322 /* didn't succeed, so giveup */
4324 pr_err("Error restoring second backup from %s\n",
4332 /* Ok, so the data is restored. Let's update those superblocks. */
4336 lo
= __le64_to_cpu(bsb
.arraystart
);
4337 hi
= lo
+ __le64_to_cpu(bsb
.length
);
4339 if (bsb
.magic
[15] == '2' && bsb
.length2
) {
4340 unsigned long long lo1
, hi1
;
4341 lo1
= __le64_to_cpu(bsb
.arraystart2
);
4342 hi1
= lo1
+ __le64_to_cpu(bsb
.length2
);
4346 } else if (lo
< lo1
)
4352 (info
->reshape_progress
< lo
||
4353 info
->reshape_progress
> hi
))
4354 /* backup does not affect reshape_progress*/ ;
4355 else if (info
->delta_disks
>= 0) {
4356 info
->reshape_progress
= __le64_to_cpu(bsb
.arraystart
) +
4357 __le64_to_cpu(bsb
.length
);
4358 if (bsb
.magic
[15] == '2') {
4359 unsigned long long p2
= __le64_to_cpu(bsb
.arraystart2
) +
4360 __le64_to_cpu(bsb
.length2
);
4361 if (p2
> info
->reshape_progress
)
4362 info
->reshape_progress
= p2
;
4365 info
->reshape_progress
= __le64_to_cpu(bsb
.arraystart
);
4366 if (bsb
.magic
[15] == '2') {
4367 unsigned long long p2
= __le64_to_cpu(bsb
.arraystart2
);
4368 if (p2
< info
->reshape_progress
)
4369 info
->reshape_progress
= p2
;
4372 for (j
=0; j
<info
->array
.raid_disks
; j
++) {
4375 if (st
->ss
->load_super(st
, fdlist
[j
], NULL
))
4377 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
4378 dinfo
.reshape_progress
= info
->reshape_progress
;
4379 st
->ss
->update_super(st
, &dinfo
,
4380 "_reshape_progress",
4382 st
->ss
->store_super(st
, fdlist
[j
]);
4383 st
->ss
->free_super(st
);
4387 /* Didn't find any backup data, try to see if any
4390 if (info
->delta_disks
< 0) {
4391 /* When shrinking, the critical section is at the end.
4392 * So see if we are before the critical section.
4394 unsigned long long first_block
;
4395 nstripe
= ostripe
= 0;
4397 while (ostripe
>= nstripe
) {
4398 ostripe
+= info
->array
.chunk_size
/ 512;
4399 first_block
= ostripe
* odata
;
4400 nstripe
= first_block
/ ndata
/ (info
->new_chunk
/512) *
4401 (info
->new_chunk
/512);
4404 if (info
->reshape_progress
>= first_block
)
4407 if (info
->delta_disks
> 0) {
4408 /* See if we are beyond the critical section. */
4409 unsigned long long last_block
;
4410 nstripe
= ostripe
= 0;
4412 while (nstripe
>= ostripe
) {
4413 nstripe
+= info
->new_chunk
/ 512;
4414 last_block
= nstripe
* ndata
;
4415 ostripe
= last_block
/ odata
/ (info
->array
.chunk_size
/512) *
4416 (info
->array
.chunk_size
/512);
4419 if (info
->reshape_progress
>= last_block
)
4422 /* needed to recover critical section! */
4424 pr_err("Failed to find backup of critical section\n");
4428 int Grow_continue_command(char *devname
, int fd
,
4429 char *backup_file
, int verbose
)
4432 struct supertype
*st
= NULL
;
4433 struct mdinfo
*content
= NULL
;
4434 struct mdinfo array
;
4435 char *subarray
= NULL
;
4436 struct mdinfo
*cc
= NULL
;
4437 struct mdstat_ent
*mdstat
= NULL
;
4441 dprintf("Grow continue from command line called for %s\n",
4444 st
= super_by_fd(fd
, &subarray
);
4445 if (!st
|| !st
->ss
) {
4446 pr_err("Unable to determine metadata format for %s\n",
4450 dprintf("Grow continue is run for ");
4451 if (st
->ss
->external
== 0) {
4453 dprintf("native array (%s)\n", devname
);
4454 if (ioctl(fd
, GET_ARRAY_INFO
, &array
.array
) < 0) {
4455 pr_err("%s is not an active md array -"
4456 " aborting\n", devname
);
4458 goto Grow_continue_command_exit
;
4461 /* Need to load a superblock.
4462 * FIXME we should really get what we need from
4465 for (d
= 0; d
< MAX_DISKS
; d
++) {
4466 mdu_disk_info_t disk
;
4470 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0)
4472 if (disk
.major
== 0 && disk
.minor
== 0)
4474 if ((disk
.state
& (1 << MD_DISK_ACTIVE
)) == 0)
4476 dv
= map_dev(disk
.major
, disk
.minor
, 1);
4479 fd2
= dev_open(dv
, O_RDONLY
);
4482 err
= st
->ss
->load_super(st
, fd2
, NULL
);
4488 if (d
== MAX_DISKS
) {
4489 pr_err("Unable to load metadata for %s\n",
4492 goto Grow_continue_command_exit
;
4494 st
->ss
->getinfo_super(st
, content
, NULL
);
4499 dprintf("subarray (%s)\n", subarray
);
4500 container
= st
->container_devnm
;
4501 cfd
= open_dev_excl(st
->container_devnm
);
4503 container
= st
->devnm
;
4505 cfd
= open_dev_excl(st
->devnm
);
4506 dprintf("container (%s)\n", container
);
4510 pr_err("Unable to open container "
4511 "for %s\n", devname
);
4513 goto Grow_continue_command_exit
;
4516 /* find in container array under reshape
4518 ret_val
= st
->ss
->load_container(st
, cfd
, NULL
);
4520 pr_err("Cannot read superblock for %s\n",
4523 goto Grow_continue_command_exit
;
4526 cc
= st
->ss
->container_content(st
, subarray
);
4527 for (content
= cc
; content
; content
= content
->next
) {
4529 int allow_reshape
= 1;
4531 if (content
->reshape_active
== 0)
4533 /* The decision about array or container wide
4534 * reshape is taken in Grow_continue based
4535 * content->reshape_active state, therefore we
4536 * need to check_reshape based on
4537 * reshape_active and subarray name
4539 if (content
->array
.state
& (1<<MD_SB_BLOCK_VOLUME
))
4541 if (content
->reshape_active
== CONTAINER_RESHAPE
&&
4542 (content
->array
.state
4543 & (1<<MD_SB_BLOCK_CONTAINER_RESHAPE
)))
4546 if (!allow_reshape
) {
4547 pr_err("cannot continue reshape of an array"
4548 " in container with unsupported"
4549 " metadata: %s(%s)\n",
4550 devname
, container
);
4552 goto Grow_continue_command_exit
;
4555 array
= strchr(content
->text_version
+1, '/')+1;
4556 mdstat
= mdstat_by_subdev(array
, container
);
4559 if (mdstat
->active
== 0) {
4560 pr_err("Skipping inactive array %s.\n",
4562 free_mdstat(mdstat
);
4569 pr_err("Unable to determine reshaped "
4570 "array for %s\n", devname
);
4572 goto Grow_continue_command_exit
;
4574 fd2
= open_dev(mdstat
->devnm
);
4576 pr_err("cannot open (%s)\n", mdstat
->devnm
);
4578 goto Grow_continue_command_exit
;
4581 sysfs_init(content
, fd2
, mdstat
->devnm
);
4583 /* start mdmon in case it is not running
4585 if (!mdmon_running(container
))
4586 start_mdmon(container
);
4587 ping_monitor(container
);
4589 if (mdmon_running(container
))
4590 st
->update_tail
= &st
->updates
;
4592 pr_err("No mdmon found. "
4593 "Grow cannot continue.\n");
4595 goto Grow_continue_command_exit
;
4599 /* verify that array under reshape is started from
4602 if (verify_reshape_position(content
, content
->array
.level
) < 0) {
4604 goto Grow_continue_command_exit
;
4609 ret_val
= Grow_continue(fd
, st
, content
, backup_file
, 0);
4611 Grow_continue_command_exit
:
4616 st
->ss
->free_super(st
);
4617 free_mdstat(mdstat
);
4624 int Grow_continue(int mdfd
, struct supertype
*st
, struct mdinfo
*info
,
4625 char *backup_file
, int freeze_reshape
)
4629 if (!info
->reshape_active
)
4632 if (st
->ss
->external
) {
4633 int cfd
= open_dev(st
->container_devnm
);
4638 st
->ss
->load_container(st
, cfd
, st
->container_devnm
);
4640 ret_val
= reshape_container(st
->container_devnm
, NULL
, mdfd
,
4641 st
, info
, 0, backup_file
,
4642 0, 1, freeze_reshape
);
4644 ret_val
= reshape_array(NULL
, mdfd
, "array", st
, info
, 1,
4645 NULL
, 0ULL, backup_file
, 0, 0, 1,