2 * mdadm - manage Linux "md" devices aka RAID arrays.
4 * Copyright (C) 2001-2012 Neil Brown <neilb@suse.de>
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License as published by
9 * the Free Software Foundation; either version 2 of the License, or
10 * (at your option) any later version.
12 * This program is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 * GNU General Public License for more details.
17 * You should have received a copy of the GNU General Public License
18 * along with this program; if not, write to the Free Software
19 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
22 * Email: <neilb@suse.de>
29 #if ! defined(__BIG_ENDIAN) && ! defined(__LITTLE_ENDIAN)
30 #error no endian defined
36 #define offsetof(t,f) ((size_t)&(((t*)0)->f))
39 int restore_backup(struct supertype
*st
,
40 struct mdinfo
*content
,
50 int disk_count
= next_spare
+ working_disks
;
52 dprintf("Called restore_backup()\n");
53 fdlist
= xmalloc(sizeof(int) * disk_count
);
55 enable_fds(next_spare
);
56 for (i
= 0; i
< next_spare
; i
++)
58 for (dev
= content
->devs
; dev
; dev
= dev
->next
) {
64 fd
= dev_open(buf
, O_RDWR
);
66 if (dev
->disk
.raid_disk
>= 0)
67 fdlist
[dev
->disk
.raid_disk
] = fd
;
69 fdlist
[next_spare
++] = fd
;
72 if (st
->ss
->external
&& st
->ss
->recover_backup
)
73 err
= st
->ss
->recover_backup(st
, content
);
75 err
= Grow_restart(st
, content
, fdlist
, next_spare
,
76 backup_file
, verbose
> 0);
78 while (next_spare
> 0) {
80 if (fdlist
[next_spare
] >= 0)
81 close(fdlist
[next_spare
]);
85 pr_err("Failed to restore critical"
86 " section for reshape - sorry.\n");
88 pr_err("Possibly you need"
89 " to specify a --backup-file\n");
93 dprintf("restore_backup() returns status OK.\n");
97 int Grow_Add_device(char *devname
, int fd
, char *newdev
)
99 /* Add a device to an active array.
100 * Currently, just extend a linear array.
101 * This requires writing a new superblock on the
102 * new device, calling the kernel to add the device,
103 * and if that succeeds, update the superblock on
105 * This means that we need to *find* all other devices.
112 struct supertype
*st
= NULL
;
113 char *subarray
= NULL
;
115 if (ioctl(fd
, GET_ARRAY_INFO
, &info
.array
) < 0) {
116 pr_err("cannot get array info for %s\n", devname
);
120 if (info
.array
.level
!= -1) {
121 pr_err("can only add devices to linear arrays\n");
125 st
= super_by_fd(fd
, &subarray
);
127 pr_err("cannot handle arrays with superblock version %d\n",
128 info
.array
.major_version
);
133 pr_err("Cannot grow linear sub-arrays yet\n");
139 nfd
= open(newdev
, O_RDWR
|O_EXCL
|O_DIRECT
);
141 pr_err("cannot open %s\n", newdev
);
146 if ((stb
.st_mode
& S_IFMT
) != S_IFBLK
) {
147 pr_err("%s is not a block device!\n", newdev
);
152 /* now check out all the devices and make sure we can read the
154 for (d
=0 ; d
< info
.array
.raid_disks
; d
++) {
155 mdu_disk_info_t disk
;
158 st
->ss
->free_super(st
);
161 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0) {
162 pr_err("cannot get device detail for device %d\n",
168 dv
= map_dev(disk
.major
, disk
.minor
, 1);
170 pr_err("cannot find device file for device %d\n",
176 fd2
= dev_open(dv
, O_RDWR
);
178 pr_err("cannot open device file %s\n", dv
);
184 if (st
->ss
->load_super(st
, fd2
, NULL
)) {
185 pr_err("cannot find super block on %s\n", dv
);
193 /* Ok, looks good. Lets update the superblock and write it out to
197 info
.disk
.number
= d
;
198 info
.disk
.major
= major(stb
.st_rdev
);
199 info
.disk
.minor
= minor(stb
.st_rdev
);
200 info
.disk
.raid_disk
= d
;
201 info
.disk
.state
= (1 << MD_DISK_SYNC
) | (1 << MD_DISK_ACTIVE
);
202 st
->ss
->update_super(st
, &info
, "linear-grow-new", newdev
,
205 if (st
->ss
->store_super(st
, nfd
)) {
206 pr_err("Cannot store new superblock on %s\n",
213 if (ioctl(fd
, ADD_NEW_DISK
, &info
.disk
) != 0) {
214 pr_err("Cannot add new disk to this array\n");
217 /* Well, that seems to have worked.
218 * Now go through and update all superblocks
221 if (ioctl(fd
, GET_ARRAY_INFO
, &info
.array
) < 0) {
222 pr_err("cannot get array info for %s\n", devname
);
227 for (d
=0 ; d
< info
.array
.raid_disks
; d
++) {
228 mdu_disk_info_t disk
;
232 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0) {
233 pr_err("cannot get device detail for device %d\n",
237 dv
= map_dev(disk
.major
, disk
.minor
, 1);
239 pr_err("cannot find device file for device %d\n",
243 fd2
= dev_open(dv
, O_RDWR
);
245 pr_err("cannot open device file %s\n", dv
);
248 if (st
->ss
->load_super(st
, fd2
, NULL
)) {
249 pr_err("cannot find super block on %s\n", dv
);
253 info
.array
.raid_disks
= nd
+1;
254 info
.array
.nr_disks
= nd
+1;
255 info
.array
.active_disks
= nd
+1;
256 info
.array
.working_disks
= nd
+1;
258 st
->ss
->update_super(st
, &info
, "linear-grow-update", dv
,
261 if (st
->ss
->store_super(st
, fd2
)) {
262 pr_err("Cannot store new superblock on %s\n", dv
);
272 int Grow_addbitmap(char *devname
, int fd
, struct context
*c
, struct shape
*s
)
275 * First check that array doesn't have a bitmap
276 * Then create the bitmap
279 * For internal bitmaps, we need to check the version,
280 * find all the active devices, and write the bitmap block
283 mdu_bitmap_file_t bmf
;
284 mdu_array_info_t array
;
285 struct supertype
*st
;
286 char *subarray
= NULL
;
287 int major
= BITMAP_MAJOR_HI
;
288 int vers
= md_get_version(fd
);
289 unsigned long long bitmapsize
, array_size
;
292 major
= BITMAP_MAJOR_HOSTENDIAN
;
293 pr_err("Warning - bitmaps created on this kernel"
294 " are not portable\n"
295 " between different architectures. Consider upgrading"
296 " the Linux kernel.\n");
299 if (ioctl(fd
, GET_BITMAP_FILE
, &bmf
) != 0) {
301 pr_err("Memory allocation failure.\n");
303 pr_err("bitmaps not supported by this kernel.\n");
306 if (bmf
.pathname
[0]) {
307 if (strcmp(s
->bitmap_file
,"none")==0) {
308 if (ioctl(fd
, SET_BITMAP_FILE
, -1)!= 0) {
309 pr_err("failed to remove bitmap %s\n",
315 pr_err("%s already has a bitmap (%s)\n",
316 devname
, bmf
.pathname
);
319 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
320 pr_err("cannot get array status for %s\n", devname
);
323 if (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)) {
324 if (strcmp(s
->bitmap_file
, "none")==0) {
325 array
.state
&= ~(1<<MD_SB_BITMAP_PRESENT
);
326 if (ioctl(fd
, SET_ARRAY_INFO
, &array
)!= 0) {
327 pr_err("failed to remove internal bitmap.\n");
332 pr_err("Internal bitmap already present on %s\n",
337 if (strcmp(s
->bitmap_file
, "none") == 0) {
338 pr_err("no bitmap found on %s\n", devname
);
341 if (array
.level
<= 0) {
342 pr_err("Bitmaps not meaningful with level %s\n",
343 map_num(pers
, array
.level
)?:"of this array");
346 bitmapsize
= array
.size
;
348 if (get_dev_size(fd
, NULL
, &array_size
) &&
349 array_size
> (0x7fffffffULL
<<9)) {
350 /* Array is big enough that we cannot trust array.size
351 * try other approaches
353 bitmapsize
= get_component_size(fd
);
355 if (bitmapsize
== 0) {
356 pr_err("Cannot reliably determine size of array to create bitmap - sorry.\n");
360 if (array
.level
== 10) {
361 int ncopies
= (array
.layout
&255)*((array
.layout
>>8)&255);
362 bitmapsize
= bitmapsize
* array
.raid_disks
/ ncopies
;
365 st
= super_by_fd(fd
, &subarray
);
367 pr_err("Cannot understand version %d.%d\n",
368 array
.major_version
, array
.minor_version
);
372 pr_err("Cannot add bitmaps to sub-arrays yet\n");
377 if (strcmp(s
->bitmap_file
, "internal") == 0) {
380 int offset_setable
= 0;
382 if (st
->ss
->add_internal_bitmap
== NULL
) {
383 pr_err("Internal bitmaps not supported "
384 "with %s metadata\n", st
->ss
->name
);
387 mdi
= sysfs_read(fd
, NULL
, GET_BITMAP_LOCATION
);
390 for (d
=0; d
< st
->max_devs
; d
++) {
391 mdu_disk_info_t disk
;
394 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0)
396 if (disk
.major
== 0 &&
399 if ((disk
.state
& (1<<MD_DISK_SYNC
))==0)
401 dv
= map_dev(disk
.major
, disk
.minor
, 1);
403 int fd2
= dev_open(dv
, O_RDWR
);
406 if (st
->ss
->load_super(st
, fd2
, NULL
)==0) {
407 if (st
->ss
->add_internal_bitmap(
409 &s
->bitmap_chunk
, c
->delay
, s
->write_behind
,
410 bitmapsize
, offset_setable
,
413 st
->ss
->write_bitmap(st
, fd2
);
415 pr_err("failed to create internal bitmap"
416 " - chunksize problem.\n");
424 if (offset_setable
) {
425 st
->ss
->getinfo_super(st
, mdi
, NULL
);
426 sysfs_init(mdi
, fd
, NULL
);
427 rv
= sysfs_set_num_signed(mdi
, NULL
, "bitmap/location",
430 array
.state
|= (1<<MD_SB_BITMAP_PRESENT
);
431 rv
= ioctl(fd
, SET_ARRAY_INFO
, &array
);
435 pr_err("Cannot add bitmap while array is"
436 " resyncing or reshaping etc.\n");
437 pr_err("failed to set internal bitmap.\n");
444 int max_devs
= st
->max_devs
;
446 /* try to load a superblock */
447 for (d
= 0; d
< max_devs
; d
++) {
448 mdu_disk_info_t disk
;
452 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0)
454 if ((disk
.major
==0 && disk
.minor
==0) ||
455 (disk
.state
& (1<<MD_DISK_REMOVED
)))
457 dv
= map_dev(disk
.major
, disk
.minor
, 1);
460 fd2
= dev_open(dv
, O_RDONLY
);
462 if (st
->ss
->load_super(st
, fd2
, NULL
) == 0) {
464 st
->ss
->uuid_from_super(st
, uuid
);
471 pr_err("cannot find UUID for array!\n");
474 if (CreateBitmap(s
->bitmap_file
, c
->force
, (char*)uuid
, s
->bitmap_chunk
,
475 c
->delay
, s
->write_behind
, bitmapsize
, major
)) {
478 bitmap_fd
= open(s
->bitmap_file
, O_RDWR
);
480 pr_err("weird: %s cannot be opened\n",
484 if (ioctl(fd
, SET_BITMAP_FILE
, bitmap_fd
) < 0) {
487 pr_err("Cannot add bitmap while array is"
488 " resyncing or reshaping etc.\n");
489 pr_err("Cannot set bitmap file for %s: %s\n",
490 devname
, strerror(err
));
499 * When reshaping an array we might need to backup some data.
500 * This is written to all spares with a 'super_block' describing it.
501 * The superblock goes 4K from the end of the used space on the
503 * It if written after the backup is complete.
504 * It has the following structure.
507 static struct mdp_backup_super
{
508 char magic
[16]; /* md_backup_data-1 or -2 */
511 /* start/sizes in 512byte sectors */
512 __u64 devstart
; /* address on backup device/file of data */
515 __u32 sb_csum
; /* csum of preceeding bytes. */
517 __u64 devstart2
; /* offset in to data of second section */
520 __u32 sb_csum2
; /* csum of preceeding bytes. */
522 } __attribute__((aligned(512))) bsb
, bsb2
;
524 static __u32
bsb_csum(char *buf
, int len
)
528 for (i
= 0; i
< len
; i
++)
529 csum
= (csum
<<3) + buf
[0];
530 return __cpu_to_le32(csum
);
533 static int check_idle(struct supertype
*st
)
535 /* Check that all member arrays for this container, or the
536 * container of this array, are idle
538 char *container
= (st
->container_devnm
[0]
539 ? st
->container_devnm
: st
->devnm
);
540 struct mdstat_ent
*ent
, *e
;
543 ent
= mdstat_read(0, 0);
544 for (e
= ent
; e
; e
= e
->next
) {
545 if (!is_container_member(e
, container
))
547 if (e
->percent
>= 0) {
556 static int freeze_container(struct supertype
*st
)
558 char *container
= (st
->container_devnm
[0]
559 ? st
->container_devnm
: st
->devnm
);
564 if (block_monitor(container
, 1)) {
565 pr_err("failed to freeze container\n");
572 static void unfreeze_container(struct supertype
*st
)
574 char *container
= (st
->container_devnm
[0]
575 ? st
->container_devnm
: st
->devnm
);
577 unblock_monitor(container
, 1);
580 static int freeze(struct supertype
*st
)
582 /* Try to freeze resync/rebuild on this array/container.
583 * Return -1 if the array is busy,
584 * return -2 container cannot be frozen,
585 * return 0 if this kernel doesn't support 'frozen'
586 * return 1 if it worked.
588 if (st
->ss
->external
)
589 return freeze_container(st
);
591 struct mdinfo
*sra
= sysfs_read(-1, st
->devnm
, GET_VERSION
);
597 /* Need to clear any 'read-auto' status */
598 if (sysfs_get_str(sra
, NULL
, "array_state", buf
, 20) > 0 &&
599 strncmp(buf
, "read-auto", 9) == 0)
600 sysfs_set_str(sra
, NULL
, "array_state", "clean");
602 err
= sysfs_freeze_array(sra
);
608 static void unfreeze(struct supertype
*st
)
610 if (st
->ss
->external
)
611 return unfreeze_container(st
);
613 struct mdinfo
*sra
= sysfs_read(-1, st
->devnm
, GET_VERSION
);
616 sysfs_set_str(sra
, NULL
, "sync_action", "idle");
621 static void wait_reshape(struct mdinfo
*sra
)
623 int fd
= sysfs_get_fd(sra
, NULL
, "sync_action");
629 while (sysfs_fd_get_str(fd
, action
, 20) > 0 &&
630 strncmp(action
, "reshape", 7) == 0) {
634 select(fd
+1, NULL
, NULL
, &rfds
, NULL
);
639 static int reshape_super(struct supertype
*st
, unsigned long long size
,
640 int level
, int layout
, int chunksize
, int raid_disks
,
641 int delta_disks
, char *backup_file
, char *dev
,
642 int direction
, int verbose
)
644 /* nothing extra to check in the native case */
645 if (!st
->ss
->external
)
647 if (!st
->ss
->reshape_super
||
648 !st
->ss
->manage_reshape
) {
649 pr_err("%s metadata does not support reshape\n",
654 return st
->ss
->reshape_super(st
, size
, level
, layout
, chunksize
,
655 raid_disks
, delta_disks
, backup_file
, dev
,
659 static void sync_metadata(struct supertype
*st
)
661 if (st
->ss
->external
) {
662 if (st
->update_tail
) {
663 flush_metadata_updates(st
);
664 st
->update_tail
= &st
->updates
;
666 st
->ss
->sync_metadata(st
);
670 static int subarray_set_num(char *container
, struct mdinfo
*sra
, char *name
, int n
)
672 /* when dealing with external metadata subarrays we need to be
673 * prepared to handle EAGAIN. The kernel may need to wait for
674 * mdmon to mark the array active so the kernel can handle
675 * allocations/writeback when preparing the reshape action
676 * (md_allow_write()). We temporarily disable safe_mode_delay
677 * to close a race with the array_state going clean before the
678 * next write to raid_disks / stripe_cache_size
683 /* only 'raid_disks' and 'stripe_cache_size' trigger md_allow_write */
685 (strcmp(name
, "raid_disks") != 0 &&
686 strcmp(name
, "stripe_cache_size") != 0))
687 return sysfs_set_num(sra
, NULL
, name
, n
);
689 rc
= sysfs_get_str(sra
, NULL
, "safe_mode_delay", safe
, sizeof(safe
));
692 sysfs_set_num(sra
, NULL
, "safe_mode_delay", 0);
693 rc
= sysfs_set_num(sra
, NULL
, name
, n
);
694 if (rc
< 0 && errno
== EAGAIN
) {
695 ping_monitor(container
);
696 /* if we get EAGAIN here then the monitor is not active
699 rc
= sysfs_set_num(sra
, NULL
, name
, n
);
701 sysfs_set_str(sra
, NULL
, "safe_mode_delay", safe
);
705 int start_reshape(struct mdinfo
*sra
, int already_running
,
706 int before_data_disks
, int data_disks
)
709 unsigned long long sync_max_to_set
;
711 sysfs_set_num(sra
, NULL
, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL
);
712 err
= sysfs_set_num(sra
, NULL
, "suspend_hi", sra
->reshape_progress
);
713 err
= err
?: sysfs_set_num(sra
, NULL
, "suspend_lo",
714 sra
->reshape_progress
);
715 if (before_data_disks
<= data_disks
)
716 sync_max_to_set
= sra
->reshape_progress
/ data_disks
;
718 sync_max_to_set
= (sra
->component_size
* data_disks
719 - sra
->reshape_progress
) / data_disks
;
720 if (!already_running
)
721 sysfs_set_num(sra
, NULL
, "sync_min", sync_max_to_set
);
722 err
= err
?: sysfs_set_num(sra
, NULL
, "sync_max", sync_max_to_set
);
723 if (!already_running
)
724 err
= err
?: sysfs_set_str(sra
, NULL
, "sync_action", "reshape");
729 void abort_reshape(struct mdinfo
*sra
)
731 sysfs_set_str(sra
, NULL
, "sync_action", "idle");
732 sysfs_set_num(sra
, NULL
, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL
);
733 sysfs_set_num(sra
, NULL
, "suspend_hi", 0);
734 sysfs_set_num(sra
, NULL
, "suspend_lo", 0);
735 sysfs_set_num(sra
, NULL
, "sync_min", 0);
736 sysfs_set_str(sra
, NULL
, "sync_max", "max");
739 int remove_disks_for_takeover(struct supertype
*st
,
744 struct mdinfo
*remaining
;
747 if (sra
->array
.level
== 10)
748 nr_of_copies
= layout
& 0xff;
749 else if (sra
->array
.level
== 1)
750 nr_of_copies
= sra
->array
.raid_disks
;
754 remaining
= sra
->devs
;
756 /* for each 'copy', select one device and remove from the list. */
757 for (slot
= 0; slot
< sra
->array
.raid_disks
; slot
+= nr_of_copies
) {
758 struct mdinfo
**diskp
;
761 /* Find a working device to keep */
762 for (diskp
= &remaining
; *diskp
; diskp
= &(*diskp
)->next
) {
763 struct mdinfo
*disk
= *diskp
;
765 if (disk
->disk
.raid_disk
< slot
)
767 if (disk
->disk
.raid_disk
>= slot
+ nr_of_copies
)
769 if (disk
->disk
.state
& (1<<MD_DISK_REMOVED
))
771 if (disk
->disk
.state
& (1<<MD_DISK_FAULTY
))
773 if (!(disk
->disk
.state
& (1<<MD_DISK_SYNC
)))
776 /* We have found a good disk to use! */
778 disk
->next
= sra
->devs
;
787 if (slot
< sra
->array
.raid_disks
) {
788 /* didn't find all slots */
794 sra
->devs
= remaining
;
798 /* Remove all 'remaining' devices from the array */
800 struct mdinfo
*sd
= remaining
;
801 remaining
= sd
->next
;
803 sysfs_set_str(sra
, sd
, "state", "faulty");
804 sysfs_set_str(sra
, sd
, "slot", "none");
805 /* for external metadata disks should be removed in mdmon */
806 if (!st
->ss
->external
)
807 sysfs_set_str(sra
, sd
, "state", "remove");
808 sd
->disk
.state
|= (1<<MD_DISK_REMOVED
);
809 sd
->disk
.state
&= ~(1<<MD_DISK_SYNC
);
810 sd
->next
= sra
->devs
;
816 void reshape_free_fdlist(int *fdlist
,
817 unsigned long long *offsets
,
822 for (i
= 0; i
< size
; i
++)
830 int reshape_prepare_fdlist(char *devname
,
834 unsigned long blocks
,
837 unsigned long long *offsets
)
843 for (d
= 0; d
<= nrdisks
; d
++)
846 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
847 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
849 if (sd
->disk
.state
& (1<<MD_DISK_SYNC
)) {
850 char *dn
= map_dev(sd
->disk
.major
,
852 fdlist
[sd
->disk
.raid_disk
]
853 = dev_open(dn
, O_RDONLY
);
854 offsets
[sd
->disk
.raid_disk
] = sd
->data_offset
*512;
855 if (fdlist
[sd
->disk
.raid_disk
] < 0) {
856 pr_err("%s: cannot open component %s\n",
857 devname
, dn
? dn
: "-unknown-");
861 } else if (backup_file
== NULL
) {
863 char *dn
= map_dev(sd
->disk
.major
,
865 fdlist
[d
] = dev_open(dn
, O_RDWR
);
866 offsets
[d
] = (sd
->data_offset
+ sra
->component_size
- blocks
- 8)*512;
868 pr_err("%s: cannot open component %s\n",
869 devname
, dn
? dn
: "-unknown-");
880 int reshape_open_backup_file(char *backup_file
,
885 unsigned long long *offsets
,
888 /* Return 1 on success, 0 on any form of failure */
889 /* need to check backup file is large enough */
895 *fdlist
= open(backup_file
, O_RDWR
|O_CREAT
|(restart
? O_TRUNC
: O_EXCL
),
899 pr_err("%s: cannot create backup file %s: %s\n",
900 devname
, backup_file
, strerror(errno
));
903 /* Guard against backup file being on array device.
904 * If array is partitioned or if LVM etc is in the
905 * way this will not notice, but it is better than
908 fstat(*fdlist
, &stb
);
911 if (stb
.st_rdev
== dev
) {
912 pr_err("backup file must NOT be"
913 " on the array being reshaped.\n");
919 for (i
=0; i
< blocks
+ 8 ; i
++) {
920 if (write(*fdlist
, buf
, 512) != 512) {
921 pr_err("%s: cannot create"
922 " backup file %s: %s\n",
923 devname
, backup_file
, strerror(errno
));
927 if (fsync(*fdlist
) != 0) {
928 pr_err("%s: cannot create backup file %s: %s\n",
929 devname
, backup_file
, strerror(errno
));
936 unsigned long GCD(unsigned long a
, unsigned long b
)
947 unsigned long compute_backup_blocks(int nchunk
, int ochunk
,
948 unsigned int ndata
, unsigned int odata
)
950 unsigned long a
, b
, blocks
;
951 /* So how much do we need to backup.
952 * We need an amount of data which is both a whole number of
953 * old stripes and a whole number of new stripes.
954 * So LCM for (chunksize*datadisks).
956 a
= (ochunk
/512) * odata
;
957 b
= (nchunk
/512) * ndata
;
960 /* LCM == product / GCD */
961 blocks
= (ochunk
/512) * (nchunk
/512) * odata
* ndata
/ a
;
966 char *analyse_change(struct mdinfo
*info
, struct reshape
*re
)
968 /* Based on the current array state in info->array and
969 * the changes in info->new_* etc, determine:
970 * - whether the change is possible
971 * - Intermediate level/raid_disks/layout
972 * - whether a restriping reshape is needed
973 * - number of sectors in minimum change unit. This
974 * will cover a whole number of stripes in 'before' and
977 * Return message if the change should be rejected
978 * NULL if the change can be achieved
980 * This can be called as part of starting a reshape, or
981 * when assembling an array that is undergoing reshape.
983 int near
, far
, offset
, copies
;
985 int old_chunk
, new_chunk
;
986 /* delta_parity records change in number of devices
987 * caused by level change
989 int delta_parity
= 0;
991 memset(re
, 0, sizeof(*re
));
993 /* If a new level not explicitly given, we assume no-change */
994 if (info
->new_level
== UnSet
)
995 info
->new_level
= info
->array
.level
;
998 switch (info
->new_level
) {
1004 /* chunk size is meaningful, must divide component_size
1007 if (info
->component_size
% (info
->new_chunk
/512))
1008 return "New chunk size does not"
1009 " divide component size";
1012 return "chunk size not meaningful for this level";
1015 info
->new_chunk
= info
->array
.chunk_size
;
1017 switch (info
->array
.level
) {
1019 return "Cannot understand this RAID level";
1021 /* RAID1 can convert to RAID1 with different disks, or
1022 * raid5 with 2 disks, or
1025 if (info
->new_level
> 1 &&
1026 (info
->component_size
& 7))
1027 return "Cannot convert RAID1 of this size - "
1028 "reduce size to multiple of 4K first.";
1029 if (info
->new_level
== 0) {
1030 if (info
->delta_disks
!= UnSet
&&
1031 info
->delta_disks
!= 0)
1032 return "Cannot change number of disks "
1033 "with RAID1->RAID0 conversion";
1035 re
->before
.data_disks
= 1;
1036 re
->after
.data_disks
= 1;
1039 if (info
->new_level
== 1) {
1040 if (info
->delta_disks
== UnSet
)
1041 /* Don't know what to do */
1042 return "no change requested for Growing RAID1";
1046 if (info
->array
.raid_disks
== 2 &&
1047 info
->new_level
== 5) {
1050 re
->before
.data_disks
= 1;
1051 if (info
->delta_disks
!= UnSet
&&
1052 info
->delta_disks
!= 0)
1053 re
->after
.data_disks
= 1 + info
->delta_disks
;
1055 re
->after
.data_disks
= 1;
1056 if (re
->after
.data_disks
< 1)
1057 return "Number of disks too small for RAID5";
1059 re
->before
.layout
= ALGORITHM_LEFT_SYMMETRIC
;
1060 info
->array
.chunk_size
= 65536;
1063 /* Could do some multi-stage conversions, but leave that to
1066 return "Impossibly level change request for RAID1";
1069 /* RAID10 can be converted from near mode to
1070 * RAID0 by removing some devices.
1071 * It can also be reshaped if the kernel supports
1074 switch (info
->new_level
) {
1076 if ((info
->array
.layout
& ~0xff) != 0x100)
1077 return "Cannot Grow RAID10 with far/offset layout";
1078 /* number of devices must be multiple of number of copies */
1079 if (info
->array
.raid_disks
% (info
->array
.layout
& 0xff))
1080 return "RAID10 layout too complex for Grow operation";
1082 new_disks
= (info
->array
.raid_disks
1083 / (info
->array
.layout
& 0xff));
1084 if (info
->delta_disks
== UnSet
)
1085 info
->delta_disks
= (new_disks
1086 - info
->array
.raid_disks
);
1088 if (info
->delta_disks
!= new_disks
- info
->array
.raid_disks
)
1089 return "New number of raid-devices impossible for RAID10";
1090 if (info
->new_chunk
&&
1091 info
->new_chunk
!= info
->array
.chunk_size
)
1092 return "Cannot change chunk-size with RAID10 Grow";
1096 re
->before
.data_disks
= new_disks
;
1097 re
->after
.data_disks
= re
->before
.data_disks
;
1101 near
= info
->array
.layout
& 0xff;
1102 far
= (info
->array
.layout
>> 8) & 0xff;
1103 offset
= info
->array
.layout
& 0x10000;
1104 if (far
> 1 && !offset
)
1105 return "Cannot reshape RAID10 in far-mode";
1106 copies
= near
* far
;
1108 old_chunk
= info
->array
.chunk_size
* far
;
1110 if (info
->new_layout
== UnSet
)
1111 info
->new_layout
= info
->array
.layout
;
1113 near
= info
->new_layout
& 0xff;
1114 far
= (info
->new_layout
>> 8) & 0xff;
1115 offset
= info
->new_layout
& 0x10000;
1116 if (far
> 1 && !offset
)
1117 return "Cannot reshape RAID10 to far-mode";
1118 if (near
* far
!= copies
)
1119 return "Cannot change number of copies"
1120 " when reshaping RAID10";
1122 if (info
->delta_disks
== UnSet
)
1123 info
->delta_disks
= 0;
1124 new_disks
= (info
->array
.raid_disks
+
1127 new_chunk
= info
->new_chunk
* far
;
1130 re
->before
.layout
= info
->array
.layout
;
1131 re
->before
.data_disks
= info
->array
.raid_disks
;
1132 re
->after
.layout
= info
->new_layout
;
1133 re
->after
.data_disks
= new_disks
;
1134 /* For RAID10 we don't do backup but do allow reshape,
1135 * so set backup_blocks to INVALID_SECTORS rather than
1137 * And there is no need to synchronise stripes on both
1138 * 'old' and 'new'. So the important
1139 * number is the minimum data_offset difference
1140 * which is the larger of (offset copies * chunk).
1142 re
->backup_blocks
= INVALID_SECTORS
;
1143 re
->min_offset_change
= max(old_chunk
, new_chunk
) / 512;
1144 if (new_disks
< re
->before
.data_disks
&&
1145 info
->space_after
< re
->min_offset_change
)
1146 /* Reduce component size by one chunk */
1147 re
->new_size
= (info
->component_size
-
1148 re
->min_offset_change
);
1150 re
->new_size
= info
->component_size
;
1151 re
->new_size
= re
->new_size
* new_disks
/ copies
;
1155 return "RAID10 can only be changed to RAID0";
1158 /* RAID0 can be converted to RAID10, or to RAID456 */
1159 if (info
->new_level
== 10) {
1160 if (info
->new_layout
== UnSet
&& info
->delta_disks
== UnSet
) {
1161 /* Assume near=2 layout */
1162 info
->new_layout
= 0x102;
1163 info
->delta_disks
= info
->array
.raid_disks
;
1165 if (info
->new_layout
== UnSet
) {
1166 int copies
= 1 + (info
->delta_disks
1167 / info
->array
.raid_disks
);
1168 if (info
->array
.raid_disks
* (copies
-1)
1169 != info
->delta_disks
)
1170 return "Impossible number of devices"
1171 " for RAID0->RAID10";
1172 info
->new_layout
= 0x100 + copies
;
1174 if (info
->delta_disks
== UnSet
) {
1175 int copies
= info
->new_layout
& 0xff;
1176 if (info
->new_layout
!= 0x100 + copies
)
1177 return "New layout impossible"
1178 " for RAID0->RAID10";;
1179 info
->delta_disks
= (copies
- 1) *
1180 info
->array
.raid_disks
;
1182 if (info
->new_chunk
&&
1183 info
->new_chunk
!= info
->array
.chunk_size
)
1184 return "Cannot change chunk-size with RAID0->RAID10";
1187 re
->before
.data_disks
= (info
->array
.raid_disks
+
1189 re
->after
.data_disks
= re
->before
.data_disks
;
1190 re
->before
.layout
= info
->new_layout
;
1194 /* RAID0 can also covert to RAID0/4/5/6 by first converting to
1195 * a raid4 style layout of the final level.
1197 switch (info
->new_level
) {
1202 re
->before
.layout
= 0;
1207 re
->before
.layout
= ALGORITHM_PARITY_N
;
1212 re
->before
.layout
= ALGORITHM_PARITY_N
;
1215 return "Impossible level change requested";
1217 re
->before
.data_disks
= info
->array
.raid_disks
;
1218 /* determining 'after' layout happens outside this 'switch' */
1222 info
->array
.layout
= ALGORITHM_PARITY_N
;
1224 switch (info
->new_level
) {
1228 re
->level
= info
->array
.level
;
1229 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1230 re
->before
.layout
= info
->array
.layout
;
1234 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1235 re
->before
.layout
= info
->array
.layout
;
1240 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1241 switch (info
->array
.layout
) {
1242 case ALGORITHM_LEFT_ASYMMETRIC
:
1243 re
->before
.layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
1245 case ALGORITHM_RIGHT_ASYMMETRIC
:
1246 re
->before
.layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
1248 case ALGORITHM_LEFT_SYMMETRIC
:
1249 re
->before
.layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
1251 case ALGORITHM_RIGHT_SYMMETRIC
:
1252 re
->before
.layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
1254 case ALGORITHM_PARITY_0
:
1255 re
->before
.layout
= ALGORITHM_PARITY_0_6
;
1257 case ALGORITHM_PARITY_N
:
1258 re
->before
.layout
= ALGORITHM_PARITY_N_6
;
1261 return "Cannot convert an array with this layout";
1265 if (info
->array
.raid_disks
!= 2)
1266 return "Can only convert a 2-device array to RAID1";
1267 if (info
->delta_disks
!= UnSet
&&
1268 info
->delta_disks
!= 0)
1269 return "Cannot set raid_disk when "
1270 "converting RAID5->RAID1";
1272 info
->new_chunk
= 0;
1275 return "Impossible level change requested";
1279 switch (info
->new_level
) {
1285 re
->before
.data_disks
= info
->array
.raid_disks
- 2;
1286 re
->before
.layout
= info
->array
.layout
;
1289 return "Impossible level change requested";
1294 /* If we reached here then it looks like a re-stripe is
1295 * happening. We have determined the intermediate level
1296 * and initial raid_disks/layout and stored these in 're'.
1298 * We need to deduce the final layout that can be atomically
1299 * converted to the end state.
1301 switch (info
->new_level
) {
1303 /* We can only get to RAID0 from RAID4 or RAID5
1304 * with appropriate layout and one extra device
1306 if (re
->level
!= 4 && re
->level
!= 5)
1307 return "Cannot covert to RAID0 from this level";
1309 switch (re
->level
) {
1311 re
->before
.layout
= 0;
1312 re
->after
.layout
= 0;
1315 re
->after
.layout
= ALGORITHM_PARITY_N
;
1321 /* We can only get to RAID4 from RAID5 */
1322 if (re
->level
!= 4 && re
->level
!= 5)
1323 return "Cannot convert to RAID4 from this level";
1325 switch (re
->level
) {
1327 re
->before
.layout
= 0;
1328 re
->after
.layout
= 0;
1331 re
->after
.layout
= ALGORITHM_PARITY_N
;
1337 /* We get to RAID5 from RAID5 or RAID6 */
1338 if (re
->level
!= 5 && re
->level
!= 6)
1339 return "Cannot convert to RAID5 from this level";
1341 switch (re
->level
) {
1343 if (info
->new_layout
== UnSet
)
1344 re
->after
.layout
= re
->before
.layout
;
1346 re
->after
.layout
= info
->new_layout
;
1349 if (info
->new_layout
== UnSet
)
1350 info
->new_layout
= re
->before
.layout
;
1352 /* after.layout needs to be raid6 version of new_layout */
1353 if (info
->new_layout
== ALGORITHM_PARITY_N
)
1354 re
->after
.layout
= ALGORITHM_PARITY_N
;
1357 char *ls
= map_num(r5layout
, info
->new_layout
);
1360 /* Current RAID6 layout has a RAID5
1363 strcat(strcpy(layout
, ls
), "-6");
1364 l
= map_name(r6layout
, layout
);
1366 return "Cannot find RAID6 layout"
1369 /* Current RAID6 has no equivalent.
1370 * If it is already a '-6' layout we
1371 * can leave it unchanged, else we must
1374 ls
= map_num(r6layout
, info
->new_layout
);
1376 strcmp(ls
+strlen(ls
)-2, "-6") != 0)
1377 return "Please specify new layout";
1378 l
= info
->new_layout
;
1380 re
->after
.layout
= l
;
1386 /* We must already be at level 6 */
1388 return "Impossible level change";
1389 if (info
->new_layout
== UnSet
)
1390 re
->after
.layout
= info
->array
.layout
;
1392 re
->after
.layout
= info
->new_layout
;
1395 return "Impossible level change requested";
1397 if (info
->delta_disks
== UnSet
)
1398 info
->delta_disks
= delta_parity
;
1400 re
->after
.data_disks
= (re
->before
.data_disks
1403 switch (re
->level
) {
1404 case 6: re
->parity
= 2;
1407 case 5: re
->parity
= 1;
1409 default: re
->parity
= 0;
1412 /* So we have a restripe operation, we need to calculate the number
1413 * of blocks per reshape operation.
1415 if (info
->new_chunk
== 0)
1416 info
->new_chunk
= info
->array
.chunk_size
;
1417 if (re
->after
.data_disks
== re
->before
.data_disks
&&
1418 re
->after
.layout
== re
->before
.layout
&&
1419 info
->new_chunk
== info
->array
.chunk_size
) {
1420 /* Nothing to change, can change level immediately. */
1421 re
->level
= info
->new_level
;
1422 re
->backup_blocks
= 0;
1425 if (re
->after
.data_disks
== 1 && re
->before
.data_disks
== 1) {
1426 /* chunk and layout changes make no difference */
1427 re
->level
= info
->new_level
;
1428 re
->backup_blocks
= 0;
1432 if (re
->after
.data_disks
== re
->before
.data_disks
&&
1433 get_linux_version() < 2006032)
1434 return "in-place reshape is not safe before 2.6.32 - sorry.";
1436 if (re
->after
.data_disks
< re
->before
.data_disks
&&
1437 get_linux_version() < 2006030)
1438 return "reshape to fewer devices is not supported before 2.6.30 - sorry.";
1440 re
->backup_blocks
= compute_backup_blocks(
1441 info
->new_chunk
, info
->array
.chunk_size
,
1442 re
->after
.data_disks
,
1443 re
->before
.data_disks
);
1444 re
->min_offset_change
= re
->backup_blocks
/ re
->before
.data_disks
;
1446 re
->new_size
= info
->component_size
* re
->after
.data_disks
;
1450 static int set_array_size(struct supertype
*st
, struct mdinfo
*sra
,
1453 struct mdinfo
*info
;
1457 if ((st
== NULL
) || (sra
== NULL
))
1460 if (text_version
== NULL
)
1461 text_version
= sra
->text_version
;
1462 subarray
= strchr(text_version
+1, '/')+1;
1463 info
= st
->ss
->container_content(st
, subarray
);
1465 unsigned long long current_size
= 0;
1466 unsigned long long new_size
=
1467 info
->custom_array_size
/2;
1469 if (sysfs_get_ll(sra
, NULL
, "array_size", ¤t_size
) == 0 &&
1470 new_size
> current_size
) {
1471 if (sysfs_set_num(sra
, NULL
, "array_size", new_size
)
1473 dprintf("Error: Cannot set array size");
1476 dprintf("Array size changed");
1478 dprintf(" from %llu to %llu.\n",
1479 current_size
, new_size
);
1483 dprintf("Error: set_array_size(): info pointer in NULL\n");
1488 static int reshape_array(char *container
, int fd
, char *devname
,
1489 struct supertype
*st
, struct mdinfo
*info
,
1490 int force
, struct mddev_dev
*devlist
,
1491 unsigned long long data_offset
,
1492 char *backup_file
, int verbose
, int forked
,
1493 int restart
, int freeze_reshape
);
1494 static int reshape_container(char *container
, char *devname
,
1496 struct supertype
*st
,
1497 struct mdinfo
*info
,
1500 int verbose
, int restart
, int freeze_reshape
);
1502 int Grow_reshape(char *devname
, int fd
,
1503 struct mddev_dev
*devlist
,
1504 unsigned long long data_offset
,
1505 struct context
*c
, struct shape
*s
)
1507 /* Make some changes in the shape of an array.
1508 * The kernel must support the change.
1510 * There are three different changes. Each can trigger
1511 * a resync or recovery so we freeze that until we have
1512 * requested everything (if kernel supports freezing - 2.6.30).
1514 * - change size (i.e. component_size)
1516 * - change layout/chunksize/ndisks
1518 * The last can require a reshape. It is different on different
1519 * levels so we need to check the level before actioning it.
1520 * Some times the level change needs to be requested after the
1521 * reshape (e.g. raid6->raid5, raid5->raid0)
1524 struct mdu_array_info_s array
;
1526 struct supertype
*st
;
1527 char *subarray
= NULL
;
1531 char *container
= NULL
;
1534 struct mddev_dev
*dv
;
1540 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) < 0) {
1541 pr_err("%s is not an active md array - aborting\n",
1545 if (data_offset
!= INVALID_SECTORS
&& array
.level
!= 10
1546 && (array
.level
< 4 || array
.level
> 6)) {
1547 pr_err("--grow --data-offset not yet supported\n");
1552 (s
->chunk
|| s
->level
!= UnSet
|| s
->layout_str
|| s
->raiddisks
)) {
1553 pr_err("cannot change component size at the same time "
1554 "as other changes.\n"
1555 " Change size first, then check data is intact before "
1556 "making other changes.\n");
1560 if (s
->raiddisks
&& s
->raiddisks
< array
.raid_disks
&& array
.level
> 1 &&
1561 get_linux_version() < 2006032 &&
1562 !check_env("MDADM_FORCE_FEWER")) {
1563 pr_err("reducing the number of devices is not safe before Linux 2.6.32\n"
1564 " Please use a newer kernel\n");
1568 st
= super_by_fd(fd
, &subarray
);
1570 pr_err("Unable to determine metadata format for %s\n", devname
);
1573 if (s
->raiddisks
> st
->max_devs
) {
1574 pr_err("Cannot increase raid-disks on this array"
1575 " beyond %d\n", st
->max_devs
);
1579 /* in the external case we need to check that the requested reshape is
1580 * supported, and perform an initial check that the container holds the
1581 * pre-requisite spare devices (mdmon owns final validation)
1583 if (st
->ss
->external
) {
1587 container
= st
->container_devnm
;
1588 cfd
= open_dev_excl(st
->container_devnm
);
1590 container
= st
->devnm
;
1592 cfd
= open_dev_excl(st
->devnm
);
1596 pr_err("Unable to open container for %s\n",
1602 rv
= st
->ss
->load_container(st
, cfd
, NULL
);
1605 pr_err("Cannot read superblock for %s\n",
1611 /* check if operation is supported for metadata handler */
1612 if (st
->ss
->container_content
) {
1613 struct mdinfo
*cc
= NULL
;
1614 struct mdinfo
*content
= NULL
;
1616 cc
= st
->ss
->container_content(st
, subarray
);
1617 for (content
= cc
; content
; content
= content
->next
) {
1618 int allow_reshape
= 1;
1620 /* check if reshape is allowed based on metadata
1621 * indications stored in content.array.status
1623 if (content
->array
.state
& (1<<MD_SB_BLOCK_VOLUME
))
1625 if (content
->array
.state
1626 & (1<<MD_SB_BLOCK_CONTAINER_RESHAPE
))
1628 if (!allow_reshape
) {
1629 pr_err("cannot reshape arrays in"
1630 " container with unsupported"
1631 " metadata: %s(%s)\n",
1632 devname
, container
);
1640 if (mdmon_running(container
))
1641 st
->update_tail
= &st
->updates
;
1645 for (dv
= devlist
; dv
; dv
= dv
->next
)
1647 if (s
->raiddisks
> array
.raid_disks
&&
1648 array
.spare_disks
+added_disks
< (s
->raiddisks
- array
.raid_disks
) &&
1650 pr_err("Need %d spare%s to avoid degraded array,"
1651 " and only have %d.\n"
1652 " Use --force to over-ride this check.\n",
1653 s
->raiddisks
- array
.raid_disks
,
1654 s
->raiddisks
- array
.raid_disks
== 1 ? "" : "s",
1655 array
.spare_disks
+ added_disks
);
1659 sra
= sysfs_read(fd
, NULL
, GET_LEVEL
| GET_DISKS
| GET_DEVS
1660 | GET_STATE
| GET_VERSION
);
1662 if (st
->ss
->external
&& subarray
== NULL
) {
1663 array
.level
= LEVEL_CONTAINER
;
1664 sra
->array
.level
= LEVEL_CONTAINER
;
1667 pr_err("failed to read sysfs parameters for %s\n",
1671 frozen
= freeze(st
);
1673 /* freeze() already spewed the reason */
1676 } else if (frozen
< 0) {
1677 pr_err("%s is performing resync/recovery and cannot"
1678 " be reshaped\n", devname
);
1683 /* ========= set size =============== */
1684 if (s
->size
> 0 && (s
->size
== MAX_SIZE
|| s
->size
!= (unsigned)array
.size
)) {
1685 unsigned long long orig_size
= get_component_size(fd
)/2;
1686 unsigned long long min_csize
;
1688 int raid0_takeover
= 0;
1691 orig_size
= (unsigned) array
.size
;
1693 if (orig_size
== 0) {
1694 pr_err("Cannot set device size in this type of array.\n");
1699 if (reshape_super(st
, s
->size
, UnSet
, UnSet
, 0, 0, UnSet
, NULL
,
1700 devname
, APPLY_METADATA_CHANGES
, c
->verbose
> 0)) {
1705 if (st
->ss
->external
) {
1706 /* metadata can have size limitation
1707 * update size value according to metadata information
1709 struct mdinfo
*sizeinfo
=
1710 st
->ss
->container_content(st
, subarray
);
1712 unsigned long long new_size
=
1713 sizeinfo
->custom_array_size
/2;
1714 int data_disks
= get_data_disks(
1715 sizeinfo
->array
.level
,
1716 sizeinfo
->array
.layout
,
1717 sizeinfo
->array
.raid_disks
);
1718 new_size
/= data_disks
;
1719 dprintf("Metadata size correction from %llu to "
1720 "%llu (%llu)\n", orig_size
, new_size
,
1721 new_size
* data_disks
);
1723 sysfs_free(sizeinfo
);
1727 /* Update the size of each member device in case
1728 * they have been resized. This will never reduce
1729 * below the current used-size. The "size" attribute
1730 * understands '0' to mean 'max'.
1734 for (mdi
= sra
->devs
; mdi
; mdi
= mdi
->next
) {
1735 if (sysfs_set_num(sra
, mdi
, "size",
1736 s
->size
== MAX_SIZE
? 0 : s
->size
) < 0) {
1737 /* Probably kernel refusing to let us
1738 * reduce the size - not an error.
1742 if (array
.not_persistent
== 0 &&
1743 array
.major_version
== 0 &&
1744 get_linux_version() < 3001000) {
1745 /* Dangerous to allow size to exceed 2TB */
1746 unsigned long long csize
;
1747 if (sysfs_get_ll(sra
, mdi
, "size", &csize
) == 0) {
1748 if (csize
>= 2ULL*1024*1024*1024)
1749 csize
= 2ULL*1024*1024*1024;
1750 if ((min_csize
== 0 || (min_csize
1757 pr_err("Cannot set size on "
1758 "array members.\n");
1759 goto size_change_error
;
1761 if (min_csize
&& s
->size
> min_csize
) {
1762 pr_err("Cannot safely make this array "
1763 "use more than 2TB per device on this kernel.\n");
1765 goto size_change_error
;
1767 if (min_csize
&& s
->size
== MAX_SIZE
) {
1768 /* Don't let the kernel choose a size - it will get
1771 pr_err("Limited v0.90 array to "
1772 "2TB per device\n");
1773 s
->size
= min_csize
;
1775 if (st
->ss
->external
) {
1776 if (sra
->array
.level
== 0) {
1777 rv
= sysfs_set_str(sra
, NULL
, "level",
1781 /* get array parametes after takeover
1782 * to chane one parameter at time only
1784 rv
= ioctl(fd
, GET_ARRAY_INFO
, &array
);
1787 /* make sure mdmon is
1788 * aware of the new level */
1789 if (!mdmon_running(st
->container_devnm
))
1790 start_mdmon(st
->container_devnm
);
1791 ping_monitor(container
);
1792 if (mdmon_running(st
->container_devnm
) &&
1793 st
->update_tail
== NULL
)
1794 st
->update_tail
= &st
->updates
;
1797 if (s
->size
== MAX_SIZE
)
1799 array
.size
= s
->size
;
1800 if ((unsigned)array
.size
!= s
->size
) {
1801 /* got truncated to 32bit, write to
1802 * component_size instead
1805 rv
= sysfs_set_num(sra
, NULL
,
1806 "component_size", s
->size
);
1810 rv
= ioctl(fd
, SET_ARRAY_INFO
, &array
);
1812 /* manage array size when it is managed externally
1814 if ((rv
== 0) && st
->ss
->external
)
1815 rv
= set_array_size(st
, sra
, sra
->text_version
);
1818 if (raid0_takeover
) {
1819 /* do not recync non-existing parity,
1820 * we will drop it anyway
1822 sysfs_set_str(sra
, NULL
, "sync_action", "frozen");
1823 /* go back to raid0, drop parity disk
1825 sysfs_set_str(sra
, NULL
, "level", "raid0");
1826 ioctl(fd
, GET_ARRAY_INFO
, &array
);
1833 /* restore metadata */
1834 if (reshape_super(st
, orig_size
, UnSet
, UnSet
, 0, 0,
1835 UnSet
, NULL
, devname
,
1836 ROLLBACK_METADATA_CHANGES
,
1839 pr_err("Cannot set device size for %s: %s\n",
1840 devname
, strerror(err
));
1842 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
1843 cont_err("Bitmap must be removed before size can be changed\n");
1847 if (s
->assume_clean
) {
1848 /* This will fail on kernels older than 3.0 unless
1849 * a backport has been arranged.
1852 sysfs_set_str(sra
, NULL
, "resync_start", "none") < 0)
1853 pr_err("--assume-clean not supported with --grow on this kernel\n");
1855 ioctl(fd
, GET_ARRAY_INFO
, &array
);
1856 s
->size
= get_component_size(fd
)/2;
1858 s
->size
= array
.size
;
1859 if (c
->verbose
>= 0) {
1860 if (s
->size
== orig_size
)
1861 pr_err("component size of %s "
1862 "unchanged at %lluK\n",
1865 pr_err("component size of %s "
1866 "has been set to %lluK\n",
1870 } else if (array
.level
!= LEVEL_CONTAINER
) {
1871 s
->size
= get_component_size(fd
)/2;
1873 s
->size
= array
.size
;
1876 /* See if there is anything else to do */
1877 if ((s
->level
== UnSet
|| s
->level
== array
.level
) &&
1878 (s
->layout_str
== NULL
) &&
1879 (s
->chunk
== 0 || s
->chunk
== array
.chunk_size
) &&
1880 data_offset
== INVALID_SECTORS
&&
1881 (s
->raiddisks
== 0 || s
->raiddisks
== array
.raid_disks
)) {
1882 /* Nothing more to do */
1883 if (!changed
&& c
->verbose
>= 0)
1884 pr_err("%s: no change requested\n",
1889 /* ========= check for Raid10/Raid1 -> Raid0 conversion ===============
1890 * current implementation assumes that following conditions must be met:
1893 * - near_copies == 2
1895 if ((s
->level
== 0 && array
.level
== 10 && sra
&&
1896 array
.layout
== ((1 << 8) + 2) && !(array
.raid_disks
& 1)) ||
1897 (s
->level
== 0 && array
.level
== 1 && sra
)) {
1899 err
= remove_disks_for_takeover(st
, sra
, array
.layout
);
1901 dprintf(Name
": Array cannot be reshaped\n");
1907 /* Make sure mdmon has seen the device removal
1908 * and updated metadata before we continue with
1912 ping_monitor(container
);
1915 memset(&info
, 0, sizeof(info
));
1917 sysfs_init(&info
, fd
, NULL
);
1918 strcpy(info
.text_version
, sra
->text_version
);
1919 info
.component_size
= s
->size
*2;
1920 info
.new_level
= s
->level
;
1921 info
.new_chunk
= s
->chunk
* 1024;
1922 if (info
.array
.level
== LEVEL_CONTAINER
) {
1923 info
.delta_disks
= UnSet
;
1924 info
.array
.raid_disks
= s
->raiddisks
;
1925 } else if (s
->raiddisks
)
1926 info
.delta_disks
= s
->raiddisks
- info
.array
.raid_disks
;
1928 info
.delta_disks
= UnSet
;
1929 if (s
->layout_str
== NULL
) {
1930 info
.new_layout
= UnSet
;
1931 if (info
.array
.level
== 6 &&
1932 (info
.new_level
== 6 || info
.new_level
== UnSet
) &&
1933 info
.array
.layout
>= 16) {
1934 pr_err("%s has a non-standard layout. If you"
1935 " wish to preserve this\n", devname
);
1936 cont_err("during the reshape, please specify"
1937 " --layout=preserve\n");
1938 cont_err("If you want to change it, specify a"
1939 " layout or use --layout=normalise\n");
1943 } else if (strcmp(s
->layout_str
, "normalise") == 0 ||
1944 strcmp(s
->layout_str
, "normalize") == 0) {
1945 /* If we have a -6 RAID6 layout, remove the '-6'. */
1946 info
.new_layout
= UnSet
;
1947 if (info
.array
.level
== 6 && info
.new_level
== UnSet
) {
1949 strcpy(l
, map_num(r6layout
, info
.array
.layout
));
1950 h
= strrchr(l
, '-');
1951 if (h
&& strcmp(h
, "-6") == 0) {
1953 info
.new_layout
= map_name(r6layout
, l
);
1956 pr_err("%s is only meaningful when reshaping"
1957 " a RAID6 array.\n", s
->layout_str
);
1961 } else if (strcmp(s
->layout_str
, "preserve") == 0) {
1962 /* This means that a non-standard RAID6 layout
1965 * - When reshape a RAID6 (e.g. adding a device)
1966 * which is in a non-standard layout, it is OK
1967 * to preserve that layout.
1968 * - When converting a RAID5 to RAID6, leave it in
1969 * the XXX-6 layout, don't re-layout.
1971 if (info
.array
.level
== 6 && info
.new_level
== UnSet
)
1972 info
.new_layout
= info
.array
.layout
;
1973 else if (info
.array
.level
== 5 && info
.new_level
== 6) {
1975 strcpy(l
, map_num(r5layout
, info
.array
.layout
));
1977 info
.new_layout
= map_name(r6layout
, l
);
1979 pr_err("%s in only meaningful when reshaping"
1980 " to RAID6\n", s
->layout_str
);
1985 int l
= info
.new_level
;
1987 l
= info
.array
.level
;
1990 info
.new_layout
= map_name(r5layout
, s
->layout_str
);
1993 info
.new_layout
= map_name(r6layout
, s
->layout_str
);
1996 info
.new_layout
= parse_layout_10(s
->layout_str
);
1999 info
.new_layout
= parse_layout_faulty(s
->layout_str
);
2002 pr_err("layout not meaningful"
2003 " with this level\n");
2007 if (info
.new_layout
== UnSet
) {
2008 pr_err("layout %s not understood"
2009 " for this level\n",
2016 if (array
.level
== LEVEL_FAULTY
) {
2017 if (s
->level
!= UnSet
&& s
->level
!= array
.level
) {
2018 pr_err("cannot change level of Faulty device\n");
2022 pr_err("cannot set chunksize of Faulty device\n");
2025 if (s
->raiddisks
&& s
->raiddisks
!= 1) {
2026 pr_err("cannot set raid_disks of Faulty device\n");
2029 if (s
->layout_str
) {
2030 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
2031 dprintf("Cannot get array information.\n");
2034 array
.layout
= info
.new_layout
;
2035 if (ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
2036 pr_err("failed to set new layout\n");
2038 } else if (c
->verbose
>= 0)
2039 printf("layout for %s set to %d\n",
2040 devname
, array
.layout
);
2042 } else if (array
.level
== LEVEL_CONTAINER
) {
2043 /* This change is to be applied to every array in the
2044 * container. This is only needed when the metadata imposes
2045 * restraints of the various arrays in the container.
2046 * Currently we only know that IMSM requires all arrays
2047 * to have the same number of devices so changing the
2048 * number of devices (On-Line Capacity Expansion) must be
2049 * performed at the level of the container
2051 rv
= reshape_container(container
, devname
, -1, st
, &info
,
2052 c
->force
, c
->backup_file
, c
->verbose
, 0, 0);
2055 /* get spare devices from external metadata
2057 if (st
->ss
->external
) {
2058 struct mdinfo
*info2
;
2060 info2
= st
->ss
->container_content(st
, subarray
);
2062 info
.array
.spare_disks
=
2063 info2
->array
.spare_disks
;
2068 /* Impose these changes on a single array. First
2069 * check that the metadata is OK with the change. */
2071 if (reshape_super(st
, 0, info
.new_level
,
2072 info
.new_layout
, info
.new_chunk
,
2073 info
.array
.raid_disks
, info
.delta_disks
,
2074 c
->backup_file
, devname
, APPLY_METADATA_CHANGES
,
2080 rv
= reshape_array(container
, fd
, devname
, st
, &info
, c
->force
,
2081 devlist
, data_offset
, c
->backup_file
, c
->verbose
,
2092 /* verify_reshape_position()
2093 * Function checks if reshape position in metadata is not farther
2094 * than position in md.
2096 * 0 : not valid sysfs entry
2097 * it can be caused by not started reshape, it should be started
2098 * by reshape array or raid0 array is before takeover
2099 * -1 : error, reshape position is obviously wrong
2100 * 1 : success, reshape progress correct or updated
2102 static int verify_reshape_position(struct mdinfo
*info
, int level
)
2108 /* read sync_max, failure can mean raid0 array */
2109 rv
= sysfs_get_str(info
, NULL
, "sync_max", buf
, 40);
2113 unsigned long long position
= strtoull(buf
, &ep
, 0);
2115 dprintf(Name
": Read sync_max sysfs entry is: %s\n", buf
);
2116 if (!(ep
== buf
|| (*ep
!= 0 && *ep
!= '\n' && *ep
!= ' '))) {
2117 position
*= get_data_disks(level
,
2119 info
->array
.raid_disks
);
2120 if (info
->reshape_progress
< position
) {
2121 dprintf("Corrected reshape progress (%llu) to "
2122 "md position (%llu)\n",
2123 info
->reshape_progress
, position
);
2124 info
->reshape_progress
= position
;
2126 } else if (info
->reshape_progress
> position
) {
2127 pr_err("Fatal error: array "
2128 "reshape was not properly frozen "
2129 "(expected reshape position is %llu, "
2130 "but reshape progress is %llu.\n",
2131 position
, info
->reshape_progress
);
2134 dprintf("Reshape position in md and metadata "
2139 } else if (rv
== 0) {
2140 /* for valid sysfs entry, 0-length content
2141 * should be indicated as error
2149 static unsigned long long choose_offset(unsigned long long lo
,
2150 unsigned long long hi
,
2151 unsigned long long min
,
2152 unsigned long long max
)
2154 /* Choose a new offset between hi and lo.
2155 * It must be between min and max, but
2156 * we would prefer something near the middle of hi/lo, and also
2157 * prefer to be aligned to a big power of 2.
2159 * So we start with the middle, then for each bit,
2160 * starting at '1' and increasing, if it is set, we either
2161 * add it or subtract it if possible, preferring the option
2162 * which is furthest from the boundary.
2164 * We stop once we get a 1MB alignment. As units are in sectors,
2165 * 1MB = 2*1024 sectors.
2167 unsigned long long choice
= (lo
+ hi
) / 2;
2168 unsigned long long bit
= 1;
2170 for (bit
= 1; bit
< 2*1024; bit
= bit
<< 1) {
2171 unsigned long long bigger
, smaller
;
2172 if (! (bit
& choice
))
2174 bigger
= choice
+ bit
;
2175 smaller
= choice
- bit
;
2176 if (bigger
> max
&& smaller
< min
)
2180 else if (smaller
< min
)
2182 else if (hi
- bigger
> smaller
- lo
)
2190 static int set_new_data_offset(struct mdinfo
*sra
, struct supertype
*st
,
2191 char *devname
, int delta_disks
,
2192 unsigned long long data_offset
,
2193 unsigned long long min
)
2198 unsigned long long before
, after
;
2200 /* Need to find min space before and after so same is used
2203 before
= UINT64_MAX
;
2205 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
2209 struct supertype
*st2
;
2210 struct mdinfo info2
;
2212 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
2214 dn
= map_dev(sd
->disk
.major
, sd
->disk
.minor
, 0);
2215 dfd
= dev_open(dn
, O_RDONLY
);
2217 pr_err("%s: cannot open component %s\n",
2218 devname
, dn
? dn
: "-unknown-");
2221 st2
= dup_super(st
);
2222 rv
= st2
->ss
->load_super(st2
,dfd
, NULL
);
2226 pr_err("%s: cannot get superblock from %s\n",
2230 st2
->ss
->getinfo_super(st2
, &info2
, NULL
);
2231 st2
->ss
->free_super(st2
);
2233 if (info2
.space_before
== 0 &&
2234 info2
.space_after
== 0) {
2235 /* Metadata doesn't support data_offset changes */
2238 if (before
> info2
.space_before
)
2239 before
= info2
.space_before
;
2240 if (after
> info2
.space_after
)
2241 after
= info2
.space_after
;
2243 if (data_offset
!= INVALID_SECTORS
) {
2245 if (info2
.data_offset
== data_offset
) {
2246 pr_err("%s: already has that data_offset\n",
2250 if (data_offset
< info2
.data_offset
)
2254 } else if ((data_offset
<= info2
.data_offset
&& dir
== 1) ||
2255 (data_offset
>= info2
.data_offset
&& dir
== -1)) {
2256 pr_err("%s: differing data offsets on devices make this --data-offset setting impossible\n",
2262 if (before
== UINT64_MAX
)
2263 /* impossible really, there must be no devices */
2266 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
2267 char *dn
= map_dev(sd
->disk
.major
, sd
->disk
.minor
, 0);
2268 unsigned long long new_data_offset
;
2270 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
2272 if (delta_disks
< 0) {
2273 /* Don't need any space as array is shrinking
2274 * just move data_offset up by min
2276 if (data_offset
== INVALID_SECTORS
)
2277 new_data_offset
= sd
->data_offset
+ min
;
2279 if (data_offset
< sd
->data_offset
+ min
) {
2280 pr_err("--data-offset too small for %s\n",
2284 new_data_offset
= data_offset
;
2286 } else if (delta_disks
> 0) {
2287 /* need space before */
2289 pr_err("Insufficient head-space for reshape on %s\n",
2293 if (data_offset
== INVALID_SECTORS
)
2294 new_data_offset
= sd
->data_offset
- min
;
2296 if (data_offset
> sd
->data_offset
- min
) {
2297 pr_err("--data-offset too large for %s\n",
2301 new_data_offset
= data_offset
;
2305 /* can move up or down. If 'data_offset'
2306 * was set we would have already decided,
2307 * so just choose direction with most space.
2314 sysfs_set_str(sra
, NULL
, "reshape_direction",
2315 dir
== 1 ? "backwards" : "forwards");
2317 /* Increase data offset */
2319 pr_err("Insufficient tail-space for reshape on %s\n",
2323 if (data_offset
!= INVALID_SECTORS
&&
2324 data_offset
< sd
->data_offset
+ min
) {
2325 pr_err("--data-offset too small on %s\n",
2329 if (data_offset
!= INVALID_SECTORS
)
2330 new_data_offset
= data_offset
;
2332 new_data_offset
= choose_offset(sd
->data_offset
,
2333 sd
->data_offset
+ after
,
2334 sd
->data_offset
+ min
,
2335 sd
->data_offset
+ after
);
2337 /* Decrease data offset */
2339 pr_err("insufficient head-room on %s\n",
2343 if (data_offset
!= INVALID_SECTORS
&&
2344 data_offset
< sd
->data_offset
- min
) {
2345 pr_err("--data-offset too small on %s\n",
2349 if (data_offset
!= INVALID_SECTORS
)
2350 new_data_offset
= data_offset
;
2352 new_data_offset
= choose_offset(sd
->data_offset
- before
,
2354 sd
->data_offset
- before
,
2355 sd
->data_offset
- min
);
2358 err
= sysfs_set_num(sra
, sd
, "new_offset", new_data_offset
);
2359 if (err
< 0 && errno
== E2BIG
) {
2360 /* try again after increasing data size to max */
2361 err
= sysfs_set_num(sra
, sd
, "size", 0);
2362 if (err
< 0 && errno
== EINVAL
&&
2363 !(sd
->disk
.state
& (1<<MD_DISK_SYNC
))) {
2364 /* some kernels have a bug where you cannot
2365 * use '0' on spare devices. */
2366 sysfs_set_num(sra
, sd
, "size",
2367 (sra
->component_size
+ after
)/2);
2369 err
= sysfs_set_num(sra
, sd
, "new_offset",
2373 if (errno
== E2BIG
&& data_offset
!= INVALID_SECTORS
) {
2374 pr_err("data-offset is too big for %s\n",
2378 if (sd
== sra
->devs
&&
2379 (errno
== ENOENT
|| errno
== E2BIG
))
2380 /* Early kernel, no 'new_offset' file,
2381 * or kernel doesn't like us.
2382 * For RAID5/6 this is not fatal
2385 pr_err("Cannot set new_offset for %s\n",
2395 static int raid10_reshape(char *container
, int fd
, char *devname
,
2396 struct supertype
*st
, struct mdinfo
*info
,
2397 struct reshape
*reshape
,
2398 unsigned long long data_offset
,
2399 int force
, int verbose
)
2401 /* Changing raid_disks, layout, chunksize or possibly
2402 * just data_offset for a RAID10.
2403 * We must always change data_offset. We change by at least
2404 * ->min_offset_change which is the largest of the old and new
2406 * If raid_disks is increasing, then data_offset must decrease
2407 * by at least this copy size.
2408 * If raid_disks is unchanged, data_offset must increase or
2409 * decrease by at least min_offset_change but preferably by much more.
2410 * We choose half of the available space.
2411 * If raid_disks is decreasing, data_offset must increase by
2412 * at least min_offset_change. To allow of this, component_size
2413 * must be decreased by the same amount.
2415 * So we calculate the required minimum and direction, possibly
2416 * reduce the component_size, then iterate through the devices
2417 * and set the new_data_offset.
2418 * If that all works, we set chunk_size, layout, raid_disks, and start
2422 unsigned long long min
;
2425 sra
= sysfs_read(fd
, NULL
,
2426 GET_COMPONENT
|GET_DEVS
|GET_OFFSET
|GET_STATE
|GET_CHUNK
2429 pr_err("%s: Cannot get array details from sysfs\n",
2433 min
= reshape
->min_offset_change
;
2435 if (info
->delta_disks
)
2436 sysfs_set_str(sra
, NULL
, "reshape_direction",
2437 info
->delta_disks
< 0 ? "backwards" : "forwards");
2438 if (info
->delta_disks
< 0 &&
2439 info
->space_after
< min
) {
2440 int rv
= sysfs_set_num(sra
, NULL
, "component_size",
2441 (sra
->component_size
-
2444 pr_err("cannot reduce component size\n");
2448 err
= set_new_data_offset(sra
, st
, devname
, info
->delta_disks
, data_offset
,
2451 pr_err("Cannot set new_data_offset: RAID10 reshape not\n");
2452 cont_err("supported on this kernel\n");
2458 if (!err
&& sysfs_set_num(sra
, NULL
, "chunk_size", info
->new_chunk
) < 0)
2460 if (!err
&& sysfs_set_num(sra
, NULL
, "layout", reshape
->after
.layout
) < 0)
2462 if (!err
&& sysfs_set_num(sra
, NULL
, "raid_disks",
2463 info
->array
.raid_disks
+ info
->delta_disks
) < 0)
2465 if (!err
&& sysfs_set_str(sra
, NULL
, "sync_action", "reshape") < 0)
2468 pr_err("Cannot set array shape for %s\n",
2471 (info
->array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
2472 cont_err(" Bitmap must be removed before"
2473 " shape can be changed\n");
2483 static void get_space_after(int fd
, struct supertype
*st
, struct mdinfo
*info
)
2485 struct mdinfo
*sra
, *sd
;
2486 /* Initialisation to silence compiler warning */
2487 unsigned long long min_space_before
= 0, min_space_after
= 0;
2490 sra
= sysfs_read(fd
, NULL
, GET_DEVS
);
2493 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
2496 struct supertype
*st2
;
2497 struct mdinfo info2
;
2499 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
2501 dn
= map_dev(sd
->disk
.major
, sd
->disk
.minor
, 0);
2502 dfd
= dev_open(dn
, O_RDONLY
);
2505 st2
= dup_super(st
);
2506 if (st2
->ss
->load_super(st2
,dfd
, NULL
)) {
2512 st2
->ss
->getinfo_super(st2
, &info2
, NULL
);
2513 st2
->ss
->free_super(st2
);
2516 min_space_before
> info2
.space_before
)
2517 min_space_before
= info2
.space_before
;
2519 min_space_after
> info2
.space_after
)
2520 min_space_after
= info2
.space_after
;
2523 if (sd
== NULL
&& !first
) {
2524 info
->space_after
= min_space_after
;
2525 info
->space_before
= min_space_before
;
2530 static void update_cache_size(char *container
, struct mdinfo
*sra
,
2531 struct mdinfo
*info
,
2532 int disks
, unsigned long long blocks
)
2534 /* Check that the internal stripe cache is
2535 * large enough, or it won't work.
2536 * It must hold at least 4 stripes of the larger
2539 unsigned long cache
;
2540 cache
= max(info
->array
.chunk_size
, info
->new_chunk
);
2541 cache
*= 4; /* 4 stripes minimum */
2542 cache
/= 512; /* convert to sectors */
2543 /* make sure there is room for 'blocks' with a bit to spare */
2544 if (cache
< 16 + blocks
/ disks
)
2545 cache
= 16 + blocks
/ disks
;
2546 cache
/= (4096/512); /* Covert from sectors to pages */
2548 if (sra
->cache_size
< cache
)
2549 subarray_set_num(container
, sra
, "stripe_cache_size",
2553 static int impose_reshape(struct mdinfo
*sra
,
2554 struct mdinfo
*info
,
2555 struct supertype
*st
,
2558 char *devname
, char *container
,
2559 struct reshape
*reshape
)
2561 struct mdu_array_info_s array
;
2563 sra
->new_chunk
= info
->new_chunk
;
2566 /* for external metadata checkpoint saved by mdmon can be lost
2567 * or missed /due to e.g. crash/. Check if md is not during
2568 * restart farther than metadata points to.
2569 * If so, this means metadata information is obsolete.
2571 if (st
->ss
->external
)
2572 verify_reshape_position(info
, reshape
->level
);
2573 sra
->reshape_progress
= info
->reshape_progress
;
2575 sra
->reshape_progress
= 0;
2576 if (reshape
->after
.data_disks
< reshape
->before
.data_disks
)
2577 /* start from the end of the new array */
2578 sra
->reshape_progress
= (sra
->component_size
2579 * reshape
->after
.data_disks
);
2582 ioctl(fd
, GET_ARRAY_INFO
, &array
);
2583 if (info
->array
.chunk_size
== info
->new_chunk
&&
2584 reshape
->before
.layout
== reshape
->after
.layout
&&
2585 st
->ss
->external
== 0) {
2586 /* use SET_ARRAY_INFO but only if reshape hasn't started */
2587 array
.raid_disks
= reshape
->after
.data_disks
+ reshape
->parity
;
2589 ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
2592 pr_err("Cannot set device shape for %s: %s\n",
2593 devname
, strerror(errno
));
2596 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
2597 cont_err("Bitmap must be removed before"
2598 " shape can be changed\n");
2602 } else if (!restart
) {
2603 /* set them all just in case some old 'new_*' value
2604 * persists from some earlier problem.
2607 if (sysfs_set_num(sra
, NULL
, "chunk_size", info
->new_chunk
) < 0)
2609 if (!err
&& sysfs_set_num(sra
, NULL
, "layout",
2610 reshape
->after
.layout
) < 0)
2612 if (!err
&& subarray_set_num(container
, sra
, "raid_disks",
2613 reshape
->after
.data_disks
+
2614 reshape
->parity
) < 0)
2617 pr_err("Cannot set device shape for %s\n",
2621 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
2622 cont_err("Bitmap must be removed before"
2623 " shape can be changed\n");
2632 static int reshape_array(char *container
, int fd
, char *devname
,
2633 struct supertype
*st
, struct mdinfo
*info
,
2634 int force
, struct mddev_dev
*devlist
,
2635 unsigned long long data_offset
,
2636 char *backup_file
, int verbose
, int forked
,
2637 int restart
, int freeze_reshape
)
2639 struct reshape reshape
;
2642 int orig_level
= UnSet
;
2646 struct mdu_array_info_s array
;
2649 struct mddev_dev
*dv
;
2653 unsigned long long *offsets
= NULL
;
2657 unsigned long blocks
;
2658 unsigned long long array_size
;
2660 struct mdinfo
*sra
= NULL
;
2662 /* when reshaping a RAID0, the component_size might be zero.
2663 * So try to fix that up.
2665 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
2666 dprintf("Cannot get array information.\n");
2669 if (array
.level
== 0 && info
->component_size
== 0) {
2670 get_dev_size(fd
, NULL
, &array_size
);
2671 info
->component_size
= array_size
/ array
.raid_disks
;
2674 if (array
.level
== 10)
2675 /* Need space_after info */
2676 get_space_after(fd
, st
, info
);
2678 if (info
->reshape_active
) {
2679 int new_level
= info
->new_level
;
2680 info
->new_level
= UnSet
;
2681 if (info
->delta_disks
> 0)
2682 info
->array
.raid_disks
-= info
->delta_disks
;
2683 msg
= analyse_change(info
, &reshape
);
2684 info
->new_level
= new_level
;
2685 if (info
->delta_disks
> 0)
2686 info
->array
.raid_disks
+= info
->delta_disks
;
2688 /* Make sure the array isn't read-only */
2689 ioctl(fd
, RESTART_ARRAY_RW
, 0);
2691 msg
= analyse_change(info
, &reshape
);
2693 pr_err("%s\n", msg
);
2697 (reshape
.level
!= info
->array
.level
||
2698 reshape
.before
.layout
!= info
->array
.layout
||
2699 reshape
.before
.data_disks
+ reshape
.parity
2700 != info
->array
.raid_disks
- max(0, info
->delta_disks
))) {
2701 pr_err("reshape info is not in native format -"
2702 " cannot continue.\n");
2706 if (st
->ss
->external
&& restart
&& (info
->reshape_progress
== 0)) {
2707 /* When reshape is restarted from '0', very begin of array
2708 * it is possible that for external metadata reshape and array
2709 * configuration doesn't happen.
2710 * Check if md has the same opinion, and reshape is restarted
2711 * from 0. If so, this is regular reshape start after reshape
2712 * switch in metadata to next array only.
2714 if ((verify_reshape_position(info
, reshape
.level
) >= 0) &&
2715 (info
->reshape_progress
== 0))
2719 /* reshape already started. just skip to monitoring the reshape */
2720 if (reshape
.backup_blocks
== 0)
2722 if (restart
& RESHAPE_NO_BACKUP
)
2726 /* The container is frozen but the array may not be.
2727 * So freeze the array so spares don't get put to the wrong use
2728 * FIXME there should probably be a cleaner separation between
2729 * freeze_array and freeze_container.
2731 sysfs_freeze_array(info
);
2732 /* Check we have enough spares to not be degraded */
2734 for (dv
= devlist
; dv
; dv
=dv
->next
)
2736 spares_needed
= max(reshape
.before
.data_disks
,
2737 reshape
.after
.data_disks
)
2738 + reshape
.parity
- array
.raid_disks
;
2741 info
->new_level
> 1 && info
->array
.level
> 1 &&
2742 spares_needed
> info
->array
.spare_disks
+ added_disks
) {
2743 pr_err("Need %d spare%s to avoid degraded array,"
2744 " and only have %d.\n"
2745 " Use --force to over-ride this check.\n",
2747 spares_needed
== 1 ? "" : "s",
2748 info
->array
.spare_disks
+ added_disks
);
2751 /* Check we have enough spares to not fail */
2752 spares_needed
= max(reshape
.before
.data_disks
,
2753 reshape
.after
.data_disks
)
2755 if ((info
->new_level
> 1 || info
->new_level
== 0) &&
2756 spares_needed
> info
->array
.spare_disks
+added_disks
) {
2757 pr_err("Need %d spare%s to create working array,"
2758 " and only have %d.\n",
2760 spares_needed
== 1 ? "" : "s",
2761 info
->array
.spare_disks
+ added_disks
);
2765 if (reshape
.level
!= array
.level
) {
2766 char *c
= map_num(pers
, reshape
.level
);
2771 err
= sysfs_set_str(info
, NULL
, "level", c
);
2774 pr_err("%s: could not set level to %s\n",
2777 (info
->array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
2778 cont_err("Bitmap must be removed"
2779 " before level can be changed\n");
2783 pr_err("level of %s changed to %s\n",
2785 orig_level
= array
.level
;
2786 sysfs_freeze_array(info
);
2788 if (reshape
.level
> 0 && st
->ss
->external
) {
2789 /* make sure mdmon is aware of the new level */
2790 if (mdmon_running(container
))
2791 flush_mdmon(container
);
2793 if (!mdmon_running(container
))
2794 start_mdmon(container
);
2795 ping_monitor(container
);
2796 if (mdmon_running(container
) &&
2797 st
->update_tail
== NULL
)
2798 st
->update_tail
= &st
->updates
;
2801 /* ->reshape_super might have chosen some spares from the
2802 * container that it wants to be part of the new array.
2803 * We can collect them with ->container_content and give
2804 * them to the kernel.
2806 if (st
->ss
->reshape_super
&& st
->ss
->container_content
) {
2807 char *subarray
= strchr(info
->text_version
+1, '/')+1;
2808 struct mdinfo
*info2
=
2809 st
->ss
->container_content(st
, subarray
);
2813 sysfs_init(info2
, fd
, st
->devnm
);
2814 /* When increasing number of devices, we need to set
2815 * new raid_disks before adding these, or they might
2818 if (reshape
.backup_blocks
&&
2819 reshape
.after
.data_disks
> reshape
.before
.data_disks
)
2820 subarray_set_num(container
, info2
, "raid_disks",
2821 reshape
.after
.data_disks
+
2823 for (d
= info2
->devs
; d
; d
= d
->next
) {
2824 if (d
->disk
.state
== 0 &&
2825 d
->disk
.raid_disk
>= 0) {
2826 /* This is a spare that wants to
2827 * be part of the array.
2829 add_disk(fd
, st
, info2
, d
);
2835 /* We might have been given some devices to add to the
2836 * array. Now that the array has been changed to the right
2837 * level and frozen, we can safely add them.
2840 Manage_subdevs(devname
, fd
, devlist
, verbose
,
2843 if (reshape
.backup_blocks
== 0 && data_offset
!= INVALID_SECTORS
)
2844 reshape
.backup_blocks
= reshape
.before
.data_disks
* info
->array
.chunk_size
/512;
2845 if (reshape
.backup_blocks
== 0) {
2846 /* No restriping needed, but we might need to impose
2847 * some more changes: layout, raid_disks, chunk_size
2849 /* read current array info */
2850 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
2851 dprintf("Cannot get array information.\n");
2854 /* compare current array info with new values and if
2855 * it is different update them to new */
2856 if (info
->new_layout
!= UnSet
&&
2857 info
->new_layout
!= array
.layout
) {
2858 array
.layout
= info
->new_layout
;
2859 if (ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
2860 pr_err("failed to set new layout\n");
2862 } else if (verbose
>= 0)
2863 printf("layout for %s set to %d\n",
2864 devname
, array
.layout
);
2866 if (info
->delta_disks
!= UnSet
&&
2867 info
->delta_disks
!= 0 &&
2868 array
.raid_disks
!= (info
->array
.raid_disks
+ info
->delta_disks
)) {
2869 array
.raid_disks
+= info
->delta_disks
;
2870 if (ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
2871 pr_err("failed to set raid disks\n");
2873 } else if (verbose
>= 0) {
2874 printf("raid_disks for %s set to %d\n",
2875 devname
, array
.raid_disks
);
2878 if (info
->new_chunk
!= 0 &&
2879 info
->new_chunk
!= array
.chunk_size
) {
2880 if (sysfs_set_num(info
, NULL
,
2881 "chunk_size", info
->new_chunk
) != 0) {
2882 pr_err("failed to set chunk size\n");
2884 } else if (verbose
>= 0)
2885 printf("chunk size for %s set to %d\n",
2886 devname
, array
.chunk_size
);
2893 * There are three possibilities.
2894 * 1/ The array will shrink.
2895 * We need to ensure the reshape will pause before reaching
2896 * the 'critical section'. We also need to fork and wait for
2897 * that to happen. When it does we
2898 * suspend/backup/complete/unfreeze
2900 * 2/ The array will not change size.
2901 * This requires that we keep a backup of a sliding window
2902 * so that we can restore data after a crash. So we need
2903 * to fork and monitor progress.
2904 * In future we will allow the data_offset to change, so
2905 * a sliding backup becomes unnecessary.
2907 * 3/ The array will grow. This is relatively easy.
2908 * However the kernel's restripe routines will cheerfully
2909 * overwrite some early data before it is safe. So we
2910 * need to make a backup of the early parts of the array
2911 * and be ready to restore it if rebuild aborts very early.
2912 * For externally managed metadata, we still need a forked
2913 * child to monitor the reshape and suspend IO over the region
2914 * that is being reshaped.
2916 * We backup data by writing it to one spare, or to a
2917 * file which was given on command line.
2919 * In each case, we first make sure that storage is available
2920 * for the required backup.
2922 * - request the shape change.
2923 * - fork to handle backup etc.
2925 /* Check that we can hold all the data */
2926 get_dev_size(fd
, NULL
, &array_size
);
2927 if (reshape
.new_size
< (array_size
/512)) {
2928 pr_err("this change will reduce the size of the array.\n"
2929 " use --grow --array-size first to truncate array.\n"
2930 " e.g. mdadm --grow %s --array-size %llu\n",
2931 devname
, reshape
.new_size
/2);
2935 if (array
.level
== 10) {
2936 /* Reshaping RAID10 does not require any data backup by
2937 * user-space. Instead it requires that the data_offset
2938 * is changed to avoid the need for backup.
2939 * So this is handled very separately
2942 /* Nothing to do. */
2944 return raid10_reshape(container
, fd
, devname
, st
, info
,
2945 &reshape
, data_offset
,
2948 sra
= sysfs_read(fd
, NULL
,
2949 GET_COMPONENT
|GET_DEVS
|GET_OFFSET
|GET_STATE
|GET_CHUNK
|
2952 pr_err("%s: Cannot get array details from sysfs\n",
2958 switch(set_new_data_offset(sra
, st
, devname
,
2959 reshape
.after
.data_disks
- reshape
.before
.data_disks
,
2961 reshape
.min_offset_change
)) {
2965 /* Updated data_offset, so it's easy now */
2966 update_cache_size(container
, sra
, info
,
2967 min(reshape
.before
.data_disks
,
2968 reshape
.after
.data_disks
),
2969 reshape
.backup_blocks
);
2971 /* Right, everything seems fine. Let's kick things off.
2975 if (impose_reshape(sra
, info
, st
, fd
, restart
,
2976 devname
, container
, &reshape
) < 0)
2978 if (sysfs_set_str(sra
, NULL
, "sync_action", "reshape") < 0) {
2979 pr_err("Failed to initiate reshape!\n");
2984 case 1: /* Couldn't set data_offset, try the old way */
2985 if (data_offset
!= INVALID_SECTORS
) {
2986 pr_err("Cannot update data_offset on this array\n");
2993 /* Decide how many blocks (sectors) for a reshape
2994 * unit. The number we have so far is just a minimum
2996 blocks
= reshape
.backup_blocks
;
2997 if (reshape
.before
.data_disks
==
2998 reshape
.after
.data_disks
) {
2999 /* Make 'blocks' bigger for better throughput, but
3000 * not so big that we reject it below.
3001 * Try for 16 megabytes
3003 while (blocks
* 32 < sra
->component_size
&&
3007 pr_err("Need to backup %luK of critical "
3008 "section..\n", blocks
/2);
3010 if (blocks
>= sra
->component_size
/2) {
3011 pr_err("%s: Something wrong"
3012 " - reshape aborted\n",
3017 /* Now we need to open all these devices so we can read/write.
3019 nrdisks
= max(reshape
.before
.data_disks
,
3020 reshape
.after
.data_disks
) + reshape
.parity
3021 + sra
->array
.spare_disks
;
3022 fdlist
= xcalloc((1+nrdisks
), sizeof(int));
3023 offsets
= xcalloc((1+nrdisks
), sizeof(offsets
[0]));
3025 odisks
= reshape
.before
.data_disks
+ reshape
.parity
;
3026 d
= reshape_prepare_fdlist(devname
, sra
, odisks
,
3027 nrdisks
, blocks
, backup_file
,
3032 if ((st
->ss
->manage_reshape
== NULL
) ||
3033 (st
->ss
->recover_backup
== NULL
)) {
3034 if (backup_file
== NULL
) {
3035 if (reshape
.after
.data_disks
<=
3036 reshape
.before
.data_disks
) {
3037 pr_err("%s: Cannot grow - "
3038 "need backup-file\n", devname
);
3040 } else if (sra
->array
.spare_disks
== 0) {
3041 pr_err("%s: Cannot grow - "
3042 "need a spare or backup-file to backup "
3043 "critical section\n", devname
);
3047 if (!reshape_open_backup_file(backup_file
, fd
, devname
,
3049 fdlist
+d
, offsets
+d
,
3057 update_cache_size(container
, sra
, info
,
3058 min(reshape
.before
.data_disks
, reshape
.after
.data_disks
),
3061 /* Right, everything seems fine. Let's kick things off.
3062 * If only changing raid_disks, use ioctl, else use
3067 if (impose_reshape(sra
, info
, st
, fd
, restart
,
3068 devname
, container
, &reshape
) < 0)
3071 err
= start_reshape(sra
, restart
, reshape
.before
.data_disks
,
3072 reshape
.after
.data_disks
);
3074 pr_err("Cannot %s reshape for %s\n",
3075 restart
? "continue" : "start",
3080 sysfs_set_str(sra
, NULL
, "array_state", "active");
3081 if (freeze_reshape
) {
3085 pr_err("Reshape has to be continued from"
3086 " location %llu when root filesystem has been mounted.\n",
3087 sra
->reshape_progress
);
3091 /* Now we just need to kick off the reshape and watch, while
3092 * handling backups of the data...
3093 * This is all done by a forked background process.
3095 switch(forked
? 0 : fork()) {
3097 pr_err("Cannot run child to monitor reshape: %s\n",
3111 /* If another array on the same devices is busy, the
3112 * reshape will wait for them. This would mean that
3113 * the first section that we suspend will stay suspended
3114 * for a long time. So check on that possibility
3115 * by looking for "DELAYED" in /proc/mdstat, and if found,
3119 struct mdstat_ent
*mds
, *m
;
3121 mds
= mdstat_read(0, 0);
3122 for (m
= mds
; m
; m
= m
->next
)
3123 if (strcmp(m
->devnm
, sra
->sys_name
) == 0) {
3125 m
->percent
== RESYNC_DELAYED
)
3128 /* Haven't started the reshape thread
3135 if (delayed
== 1 && get_linux_version() < 3007000) {
3136 pr_err("Reshape is delayed, but cannot wait carefully with this kernel.\n"
3137 " You might experience problems until other reshapes complete.\n");
3141 sleep(30 - (delayed
-1) * 25);
3145 if (check_env("MDADM_GROW_VERIFY"))
3146 fd
= open(devname
, O_RDONLY
| O_DIRECT
);
3149 mlockall(MCL_FUTURE
);
3151 if (st
->ss
->external
) {
3152 /* metadata handler takes it from here */
3153 done
= st
->ss
->manage_reshape(
3154 fd
, sra
, &reshape
, st
, blocks
,
3156 d
- odisks
, fdlist
+odisks
,
3159 done
= child_monitor(
3160 fd
, sra
, &reshape
, st
, blocks
,
3162 d
- odisks
, fdlist
+odisks
,
3168 if (backup_file
&& done
)
3169 unlink(backup_file
);
3175 if (!st
->ss
->external
&&
3176 !(reshape
.before
.data_disks
!= reshape
.after
.data_disks
3177 && info
->custom_array_size
) &&
3178 info
->new_level
== reshape
.level
&&
3180 /* no need to wait for the reshape to finish as
3181 * there is nothing more to do.
3188 if (st
->ss
->external
) {
3189 /* Re-load the metadata as much could have changed */
3190 int cfd
= open_dev(st
->container_devnm
);
3192 flush_mdmon(container
);
3193 st
->ss
->free_super(st
);
3194 st
->ss
->load_container(st
, cfd
, container
);
3199 /* set new array size if required customer_array_size is used
3202 if (reshape
.before
.data_disks
!=
3203 reshape
.after
.data_disks
&&
3204 info
->custom_array_size
)
3205 set_array_size(st
, info
, info
->text_version
);
3207 if (info
->new_level
!= reshape
.level
) {
3209 c
= map_num(pers
, info
->new_level
);
3211 err
= sysfs_set_str(sra
, NULL
, "level", c
);
3213 pr_err("%s: could not set level "
3214 "to %s\n", devname
, c
);
3216 if (info
->new_level
== 0)
3217 st
->update_tail
= NULL
;
3229 if (orig_level
!= UnSet
&& sra
) {
3230 c
= map_num(pers
, orig_level
);
3231 if (c
&& sysfs_set_str(sra
, NULL
, "level", c
) == 0)
3232 pr_err("aborting level change\n");
3240 /* mdfd handle is passed to be closed in child process (after fork).
3242 int reshape_container(char *container
, char *devname
,
3244 struct supertype
*st
,
3245 struct mdinfo
*info
,
3248 int verbose
, int restart
, int freeze_reshape
)
3250 struct mdinfo
*cc
= NULL
;
3252 char last_devnm
[32] = "";
3254 /* component_size is not meaningful for a container,
3255 * so pass '0' meaning 'no change'
3258 reshape_super(st
, 0, info
->new_level
,
3259 info
->new_layout
, info
->new_chunk
,
3260 info
->array
.raid_disks
, info
->delta_disks
,
3261 backup_file
, devname
, APPLY_METADATA_CHANGES
,
3269 /* ping monitor to be sure that update is on disk
3271 ping_monitor(container
);
3274 case -1: /* error */
3275 perror("Cannot fork to complete reshape\n");
3278 default: /* parent */
3279 if (!freeze_reshape
)
3280 printf(Name
": multi-array reshape continues"
3281 " in background\n");
3288 /* close unused handle in child process
3294 /* For each member array with reshape_active,
3295 * we need to perform the reshape.
3296 * We pick the first array that needs reshaping and
3297 * reshape it. reshape_array() will re-read the metadata
3298 * so the next time through a different array should be
3299 * ready for reshape.
3300 * It is possible that the 'different' array will not
3301 * be assembled yet. In that case we simple exit.
3302 * When it is assembled, the mdadm which assembles it
3303 * will take over the reshape.
3305 struct mdinfo
*content
;
3307 struct mdstat_ent
*mdstat
;
3313 cc
= st
->ss
->container_content(st
, NULL
);
3315 for (content
= cc
; content
; content
= content
->next
) {
3317 if (!content
->reshape_active
)
3320 subarray
= strchr(content
->text_version
+1, '/')+1;
3321 mdstat
= mdstat_by_subdev(subarray
, container
);
3324 if (mdstat
->active
== 0) {
3325 pr_err("Skipping inactive array %s.\n",
3327 free_mdstat(mdstat
);
3336 devid
= devnm2devid(mdstat
->devnm
);
3337 adev
= map_dev(major(devid
), minor(devid
), 0);
3339 adev
= content
->text_version
;
3341 fd
= open_dev(mdstat
->devnm
);
3343 printf(Name
": Device %s cannot be opened for reshape.",
3348 if (strcmp(last_devnm
, mdstat
->devnm
) == 0) {
3349 /* Do not allow for multiple reshape_array() calls for
3351 * It can happen when reshape_array() returns without
3352 * error, when reshape is not finished (wrong reshape
3353 * starting/continuation conditions). Mdmon doesn't
3354 * switch to next array in container and reentry
3355 * conditions for the same array occur.
3356 * This is possibly interim until the behaviour of
3357 * reshape_array is resolved().
3359 printf(Name
": Multiple reshape execution detected for "
3360 "device %s.", adev
);
3364 strcpy(last_devnm
, mdstat
->devnm
);
3366 sysfs_init(content
, fd
, mdstat
->devnm
);
3368 if (mdmon_running(container
))
3369 flush_mdmon(container
);
3371 rv
= reshape_array(container
, fd
, adev
, st
,
3372 content
, force
, NULL
, 0ULL,
3373 backup_file
, verbose
, 1, restart
,
3377 if (freeze_reshape
) {
3386 if (mdmon_running(container
))
3387 flush_mdmon(container
);
3396 * We run a child process in the background which performs the following
3398 * - wait for resync to reach a certain point
3399 * - suspend io to the following section
3400 * - backup that section
3401 * - allow resync to proceed further
3403 * - discard the backup.
3405 * When are combined in slightly different ways in the three cases.
3407 * - suspend/backup/allow/wait/resume/discard
3409 * - allow/wait/suspend/backup/allow/wait/resume/discard
3411 * - wait/resume/discard/suspend/backup/allow
3413 * suspend/backup/allow always come together
3414 * wait/resume/discard do too.
3415 * For the same-size case we have two backups to improve flow.
3419 int progress_reshape(struct mdinfo
*info
, struct reshape
*reshape
,
3420 unsigned long long backup_point
,
3421 unsigned long long wait_point
,
3422 unsigned long long *suspend_point
,
3423 unsigned long long *reshape_completed
)
3425 /* This function is called repeatedly by the reshape manager.
3426 * It determines how much progress can safely be made and allows
3428 * - 'info' identifies the array and particularly records in
3429 * ->reshape_progress the metadata's knowledge of progress
3430 * This is a sector offset from the start of the array
3431 * of the next array block to be relocated. This number
3432 * may increase from 0 or decrease from array_size, depending
3433 * on the type of reshape that is happening.
3434 * Note that in contrast, 'sync_completed' is a block count of the
3435 * reshape so far. It gives the distance between the start point
3436 * (head or tail of device) and the next place that data will be
3437 * written. It always increases.
3438 * - 'reshape' is the structure created by analyse_change
3439 * - 'backup_point' shows how much the metadata manager has backed-up
3440 * data. For reshapes with increasing progress, it is the next address
3441 * to be backed up, previous addresses have been backed-up. For
3442 * decreasing progress, it is the earliest address that has been
3443 * backed up - later address are also backed up.
3444 * So addresses between reshape_progress and backup_point are
3445 * backed up providing those are in the 'correct' order.
3446 * - 'wait_point' is an array address. When reshape_completed
3447 * passes this point, progress_reshape should return. It might
3448 * return earlier if it determines that ->reshape_progress needs
3449 * to be updated or further backup is needed.
3450 * - suspend_point is maintained by progress_reshape and the caller
3451 * should not touch it except to initialise to zero.
3452 * It is an array address and it only increases in 2.6.37 and earlier.
3453 * This makes it difficult to handle reducing reshapes with
3454 * external metadata.
3455 * However: it is similar to backup_point in that it records the
3456 * other end of a suspended region from reshape_progress.
3457 * it is moved to extend the region that is safe to backup and/or
3459 * - reshape_completed is read from sysfs and returned. The caller
3460 * should copy this into ->reshape_progress when it has reason to
3461 * believe that the metadata knows this, and any backup outside this
3465 * 1 if more data from backup_point - but only as far as suspend_point,
3466 * should be backed up
3467 * 0 if things are progressing smoothly
3468 * -1 if the reshape is finished because it is all done,
3469 * -2 if the reshape is finished due to an error.
3472 int advancing
= (reshape
->after
.data_disks
3473 >= reshape
->before
.data_disks
);
3474 unsigned long long need_backup
; /* All data between start of array and
3475 * here will at some point need to
3478 unsigned long long read_offset
, write_offset
;
3479 unsigned long long write_range
;
3480 unsigned long long max_progress
, target
, completed
;
3481 unsigned long long array_size
= (info
->component_size
3482 * reshape
->before
.data_disks
);
3486 /* First, we unsuspend any region that is now known to be safe.
3487 * If suspend_point is on the 'wrong' side of reshape_progress, then
3488 * we don't have or need suspension at the moment. This is true for
3489 * native metadata when we don't need to back-up.
3492 if (info
->reshape_progress
<= *suspend_point
)
3493 sysfs_set_num(info
, NULL
, "suspend_lo",
3494 info
->reshape_progress
);
3496 /* Note: this won't work in 2.6.37 and before.
3497 * Something somewhere should make sure we don't need it!
3499 if (info
->reshape_progress
>= *suspend_point
)
3500 sysfs_set_num(info
, NULL
, "suspend_hi",
3501 info
->reshape_progress
);
3504 /* Now work out how far it is safe to progress.
3505 * If the read_offset for ->reshape_progress is less than
3506 * 'blocks' beyond the write_offset, we can only progress as far
3508 * Otherwise we can progress until the write_offset for the new location
3509 * reaches (within 'blocks' of) the read_offset at the current location.
3510 * However that region must be suspended unless we are using native
3512 * If we need to suspend more, we limit it to 128M per device, which is
3513 * rather arbitrary and should be some time-based calculation.
3515 read_offset
= info
->reshape_progress
/ reshape
->before
.data_disks
;
3516 write_offset
= info
->reshape_progress
/ reshape
->after
.data_disks
;
3517 write_range
= info
->new_chunk
/512;
3518 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
)
3519 need_backup
= array_size
;
3521 need_backup
= reshape
->backup_blocks
;
3523 if (read_offset
< write_offset
+ write_range
)
3524 max_progress
= backup_point
;
3528 reshape
->after
.data_disks
;
3530 if (read_offset
> write_offset
- write_range
)
3531 /* Can only progress as far as has been backed up,
3532 * which must be suspended */
3533 max_progress
= backup_point
;
3534 else if (info
->reshape_progress
<= need_backup
)
3535 max_progress
= backup_point
;
3537 if (info
->array
.major_version
>= 0)
3538 /* Can progress until backup is needed */
3539 max_progress
= need_backup
;
3541 /* Can progress until metadata update is required */
3544 reshape
->after
.data_disks
;
3545 /* but data must be suspended */
3546 if (max_progress
< *suspend_point
)
3547 max_progress
= *suspend_point
;
3552 /* We know it is safe to progress to 'max_progress' providing
3553 * it is suspended or we are using native metadata.
3554 * Consider extending suspend_point 128M per device if it
3555 * is less than 64M per device beyond reshape_progress.
3556 * But always do a multiple of 'blocks'
3557 * FIXME this is too big - it takes to long to complete
3560 target
= 64*1024*2 * min(reshape
->before
.data_disks
,
3561 reshape
->after
.data_disks
);
3562 target
/= reshape
->backup_blocks
;
3565 target
*= reshape
->backup_blocks
;
3567 /* For externally managed metadata we always need to suspend IO to
3568 * the area being reshaped so we regularly push suspend_point forward.
3569 * For native metadata we only need the suspend if we are going to do
3573 if ((need_backup
> info
->reshape_progress
3574 || info
->array
.major_version
< 0) &&
3575 *suspend_point
< info
->reshape_progress
+ target
) {
3576 if (need_backup
< *suspend_point
+ 2 * target
)
3577 *suspend_point
= need_backup
;
3578 else if (*suspend_point
+ 2 * target
< array_size
)
3579 *suspend_point
+= 2 * target
;
3581 *suspend_point
= array_size
;
3582 sysfs_set_num(info
, NULL
, "suspend_hi", *suspend_point
);
3583 if (max_progress
> *suspend_point
)
3584 max_progress
= *suspend_point
;
3587 if (info
->array
.major_version
>= 0) {
3588 /* Only need to suspend when about to backup */
3589 if (info
->reshape_progress
< need_backup
* 2 &&
3590 *suspend_point
> 0) {
3592 sysfs_set_num(info
, NULL
, "suspend_lo", 0);
3593 sysfs_set_num(info
, NULL
, "suspend_hi", need_backup
);
3596 /* Need to suspend continually */
3597 if (info
->reshape_progress
< *suspend_point
)
3598 *suspend_point
= info
->reshape_progress
;
3599 if (*suspend_point
+ target
< info
->reshape_progress
)
3600 /* No need to move suspend region yet */;
3602 if (*suspend_point
>= 2 * target
)
3603 *suspend_point
-= 2 * target
;
3606 sysfs_set_num(info
, NULL
, "suspend_lo",
3609 if (max_progress
< *suspend_point
)
3610 max_progress
= *suspend_point
;
3614 /* now set sync_max to allow that progress. sync_max, like
3615 * sync_completed is a count of sectors written per device, so
3616 * we find the difference between max_progress and the start point,
3617 * and divide that by after.data_disks to get a sync_max
3619 * At the same time we convert wait_point to a similar number
3620 * for comparing against sync_completed.
3622 /* scale down max_progress to per_disk */
3623 max_progress
/= reshape
->after
.data_disks
;
3624 /* Round to chunk size as some kernels give an erroneously high number */
3625 max_progress
/= info
->new_chunk
/512;
3626 max_progress
*= info
->new_chunk
/512;
3627 /* And round to old chunk size as the kernel wants that */
3628 max_progress
/= info
->array
.chunk_size
/512;
3629 max_progress
*= info
->array
.chunk_size
/512;
3630 /* Limit progress to the whole device */
3631 if (max_progress
> info
->component_size
)
3632 max_progress
= info
->component_size
;
3633 wait_point
/= reshape
->after
.data_disks
;
3635 /* switch from 'device offset' to 'processed block count' */
3636 max_progress
= info
->component_size
- max_progress
;
3637 wait_point
= info
->component_size
- wait_point
;
3640 sysfs_set_num(info
, NULL
, "sync_max", max_progress
);
3642 /* Now wait. If we have already reached the point that we were
3643 * asked to wait to, don't wait at all, else wait for any change.
3644 * We need to select on 'sync_completed' as that is the place that
3645 * notifications happen, but we are really interested in
3646 * 'reshape_position'
3648 fd
= sysfs_get_fd(info
, NULL
, "sync_completed");
3650 goto check_progress
;
3652 if (sysfs_fd_get_ll(fd
, &completed
) < 0)
3653 goto check_progress
;
3655 while (completed
< max_progress
&& completed
< wait_point
) {
3656 /* Check that sync_action is still 'reshape' to avoid
3657 * waiting forever on a dead array
3661 if (sysfs_get_str(info
, NULL
, "sync_action",
3663 strncmp(action
, "reshape", 7) != 0)
3665 /* Some kernels reset 'sync_completed' to zero
3666 * before setting 'sync_action' to 'idle'.
3667 * So we need these extra tests.
3669 if (completed
== 0 && advancing
3670 && info
->reshape_progress
> 0)
3672 if (completed
== 0 && !advancing
3673 && info
->reshape_progress
< (info
->component_size
3674 * reshape
->after
.data_disks
))
3678 select(fd
+1, NULL
, NULL
, &rfds
, NULL
);
3679 if (sysfs_fd_get_ll(fd
, &completed
) < 0)
3680 goto check_progress
;
3682 /* Some kernels reset 'sync_completed' to zero,
3683 * we need to have real point we are in md
3686 completed
= max_progress
;
3688 /* some kernels can give an incorrectly high 'completed' number */
3689 completed
/= (info
->new_chunk
/512);
3690 completed
*= (info
->new_chunk
/512);
3691 /* Convert 'completed' back in to a 'progress' number */
3692 completed
*= reshape
->after
.data_disks
;
3694 completed
= info
->component_size
* reshape
->after
.data_disks
3697 *reshape_completed
= completed
;
3701 /* We return the need_backup flag. Caller will decide
3702 * how much - a multiple of ->backup_blocks up to *suspend_point
3705 return need_backup
> info
->reshape_progress
;
3707 return need_backup
>= info
->reshape_progress
;
3710 /* if we couldn't read a number from sync_completed, then
3711 * either the reshape did complete, or it aborted.
3712 * We can tell which by checking for 'none' in reshape_position.
3713 * If it did abort, then it might immediately restart if it
3714 * it was just a device failure that leaves us degraded but
3718 if (sysfs_get_str(info
, NULL
, "reshape_position", buf
, sizeof(buf
)) < 0
3719 || strncmp(buf
, "none", 4) != 0) {
3720 /* The abort might only be temporary. Wait up to 10
3721 * seconds for fd to contain a valid number again.
3727 while (fd
>= 0 && rv
< 0 && tv
.tv_sec
> 0) {
3731 if (select(fd
+1, NULL
, NULL
, &rfds
, &tv
) != 1)
3733 switch (sysfs_fd_get_ll(fd
, &completed
)) {
3735 /* all good again */
3738 case -2: /* read error - abort */
3745 return rv
; /* abort */
3747 /* Maybe racing with array shutdown - check state */
3750 if (sysfs_get_str(info
, NULL
, "array_state", buf
, sizeof(buf
)) < 0
3751 || strncmp(buf
, "inactive", 8) == 0
3752 || strncmp(buf
, "clear",5) == 0)
3753 return -2; /* abort */
3754 return -1; /* complete */
3758 /* FIXME return status is never checked */
3759 static int grow_backup(struct mdinfo
*sra
,
3760 unsigned long long offset
, /* per device */
3761 unsigned long stripes
, /* per device, in old chunks */
3762 int *sources
, unsigned long long *offsets
,
3763 int disks
, int chunk
, int level
, int layout
,
3764 int dests
, int *destfd
, unsigned long long *destoffsets
,
3765 int part
, int *degraded
,
3768 /* Backup 'blocks' sectors at 'offset' on each device of the array,
3769 * to storage 'destfd' (offset 'destoffsets'), after first
3770 * suspending IO. Then allow resync to continue
3771 * over the suspended section.
3772 * Use part 'part' of the backup-super-block.
3777 unsigned long long ll
;
3779 //printf("offset %llu\n", offset);
3785 /* Check that array hasn't become degraded, else we might backup the wrong data */
3786 if (sysfs_get_ll(sra
, NULL
, "degraded", &ll
) < 0)
3787 return -1; /* FIXME this error is ignored */
3788 new_degraded
= (int)ll
;
3789 if (new_degraded
!= *degraded
) {
3790 /* check each device to ensure it is still working */
3792 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
3793 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
3795 if (sd
->disk
.state
& (1<<MD_DISK_SYNC
)) {
3797 if (sysfs_get_str(sra
, sd
, "state", sbuf
, 20) < 0 ||
3798 strstr(sbuf
, "faulty") ||
3799 strstr(sbuf
, "in_sync") == NULL
) {
3800 /* this device is dead */
3801 sd
->disk
.state
= (1<<MD_DISK_FAULTY
);
3802 if (sd
->disk
.raid_disk
>= 0 &&
3803 sources
[sd
->disk
.raid_disk
] >= 0) {
3804 close(sources
[sd
->disk
.raid_disk
]);
3805 sources
[sd
->disk
.raid_disk
] = -1;
3810 *degraded
= new_degraded
;
3813 bsb
.arraystart2
= __cpu_to_le64(offset
* odata
);
3814 bsb
.length2
= __cpu_to_le64(stripes
* (chunk
/512) * odata
);
3816 bsb
.arraystart
= __cpu_to_le64(offset
* odata
);
3817 bsb
.length
= __cpu_to_le64(stripes
* (chunk
/512) * odata
);
3820 bsb
.magic
[15] = '2';
3821 for (i
= 0; i
< dests
; i
++)
3823 lseek64(destfd
[i
], destoffsets
[i
] + __le64_to_cpu(bsb
.devstart2
)*512, 0);
3825 lseek64(destfd
[i
], destoffsets
[i
], 0);
3827 rv
= save_stripes(sources
, offsets
,
3828 disks
, chunk
, level
, layout
,
3830 offset
*512*odata
, stripes
* chunk
* odata
,
3835 bsb
.mtime
= __cpu_to_le64(time(0));
3836 for (i
= 0; i
< dests
; i
++) {
3837 bsb
.devstart
= __cpu_to_le64(destoffsets
[i
]/512);
3839 bsb
.sb_csum
= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
));
3840 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0)
3841 bsb
.sb_csum2
= bsb_csum((char*)&bsb
,
3842 ((char*)&bsb
.sb_csum2
)-((char*)&bsb
));
3845 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
] - 4096, 0)
3846 != destoffsets
[i
] - 4096)
3848 if (write(destfd
[i
], &bsb
, 512) != 512)
3850 if (destoffsets
[i
] > 4096) {
3851 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
]+stripes
*chunk
*odata
, 0) !=
3852 destoffsets
[i
]+stripes
*chunk
*odata
)
3854 if (write(destfd
[i
], &bsb
, 512) != 512)
3864 /* in 2.6.30, the value reported by sync_completed can be
3865 * less that it should be by one stripe.
3866 * This only happens when reshape hits sync_max and pauses.
3867 * So allow wait_backup to either extent sync_max further
3868 * than strictly necessary, or return before the
3869 * sync has got quite as far as we would really like.
3870 * This is what 'blocks2' is for.
3871 * The various caller give appropriate values so that
3874 /* FIXME return value is often ignored */
3875 static int forget_backup(int dests
, int *destfd
,
3876 unsigned long long *destoffsets
,
3880 * Erase backup 'part' (which is 0 or 1)
3886 bsb
.arraystart2
= __cpu_to_le64(0);
3887 bsb
.length2
= __cpu_to_le64(0);
3889 bsb
.arraystart
= __cpu_to_le64(0);
3890 bsb
.length
= __cpu_to_le64(0);
3892 bsb
.mtime
= __cpu_to_le64(time(0));
3894 for (i
= 0; i
< dests
; i
++) {
3895 bsb
.devstart
= __cpu_to_le64(destoffsets
[i
]/512);
3896 bsb
.sb_csum
= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
));
3897 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0)
3898 bsb
.sb_csum2
= bsb_csum((char*)&bsb
,
3899 ((char*)&bsb
.sb_csum2
)-((char*)&bsb
));
3900 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
]-4096, 0) !=
3901 destoffsets
[i
]-4096)
3904 write(destfd
[i
], &bsb
, 512) != 512)
3911 static void fail(char *msg
)
3914 rv
= (write(2, msg
, strlen(msg
)) != (int)strlen(msg
));
3915 rv
|= (write(2, "\n", 1) != 1);
3919 static char *abuf
, *bbuf
;
3920 static unsigned long long abuflen
;
3921 static void validate(int afd
, int bfd
, unsigned long long offset
)
3923 /* check that the data in the backup against the array.
3924 * This is only used for regression testing and should not
3925 * be used while the array is active
3929 lseek64(bfd
, offset
- 4096, 0);
3930 if (read(bfd
, &bsb2
, 512) != 512)
3931 fail("cannot read bsb");
3932 if (bsb2
.sb_csum
!= bsb_csum((char*)&bsb2
,
3933 ((char*)&bsb2
.sb_csum
)-((char*)&bsb2
)))
3934 fail("first csum bad");
3935 if (memcmp(bsb2
.magic
, "md_backup_data", 14) != 0)
3936 fail("magic is bad");
3937 if (memcmp(bsb2
.magic
, "md_backup_data-2", 16) == 0 &&
3938 bsb2
.sb_csum2
!= bsb_csum((char*)&bsb2
,
3939 ((char*)&bsb2
.sb_csum2
)-((char*)&bsb2
)))
3940 fail("second csum bad");
3942 if (__le64_to_cpu(bsb2
.devstart
)*512 != offset
)
3943 fail("devstart is wrong");
3946 unsigned long long len
= __le64_to_cpu(bsb2
.length
)*512;
3948 if (abuflen
< len
) {
3952 if (posix_memalign((void**)&abuf
, 4096, abuflen
) ||
3953 posix_memalign((void**)&bbuf
, 4096, abuflen
)) {
3955 /* just stop validating on mem-alloc failure */
3960 lseek64(bfd
, offset
, 0);
3961 if ((unsigned long long)read(bfd
, bbuf
, len
) != len
) {
3962 //printf("len %llu\n", len);
3963 fail("read first backup failed");
3965 lseek64(afd
, __le64_to_cpu(bsb2
.arraystart
)*512, 0);
3966 if ((unsigned long long)read(afd
, abuf
, len
) != len
)
3967 fail("read first from array failed");
3968 if (memcmp(bbuf
, abuf
, len
) != 0) {
3971 printf("offset=%llu len=%llu\n",
3972 (unsigned long long)__le64_to_cpu(bsb2
.arraystart
)*512, len
);
3973 for (i
=0; i
<len
; i
++)
3974 if (bbuf
[i
] != abuf
[i
]) {
3975 printf("first diff byte %d\n", i
);
3979 fail("data1 compare failed");
3983 unsigned long long len
= __le64_to_cpu(bsb2
.length2
)*512;
3985 if (abuflen
< len
) {
3989 abuf
= xmalloc(abuflen
);
3990 bbuf
= xmalloc(abuflen
);
3993 lseek64(bfd
, offset
+__le64_to_cpu(bsb2
.devstart2
)*512, 0);
3994 if ((unsigned long long)read(bfd
, bbuf
, len
) != len
)
3995 fail("read second backup failed");
3996 lseek64(afd
, __le64_to_cpu(bsb2
.arraystart2
)*512, 0);
3997 if ((unsigned long long)read(afd
, abuf
, len
) != len
)
3998 fail("read second from array failed");
3999 if (memcmp(bbuf
, abuf
, len
) != 0)
4000 fail("data2 compare failed");
4004 int child_monitor(int afd
, struct mdinfo
*sra
, struct reshape
*reshape
,
4005 struct supertype
*st
, unsigned long blocks
,
4006 int *fds
, unsigned long long *offsets
,
4007 int dests
, int *destfd
, unsigned long long *destoffsets
)
4009 /* Monitor a reshape where backup is being performed using
4010 * 'native' mechanism - either to a backup file, or
4011 * to some space in a spare.
4015 unsigned long long speed
;
4016 unsigned long long suspend_point
, array_size
;
4017 unsigned long long backup_point
, wait_point
;
4018 unsigned long long reshape_completed
;
4020 int increasing
= reshape
->after
.data_disks
>= reshape
->before
.data_disks
;
4021 int part
= 0; /* The next part of the backup area to fill. It may already
4022 * be full, so we need to check */
4023 int level
= reshape
->level
;
4024 int layout
= reshape
->before
.layout
;
4025 int data
= reshape
->before
.data_disks
;
4026 int disks
= reshape
->before
.data_disks
+ reshape
->parity
;
4027 int chunk
= sra
->array
.chunk_size
;
4029 unsigned long stripes
;
4032 /* set up the backup-super-block. This requires the
4033 * uuid from the array.
4035 /* Find a superblock */
4036 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
4040 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
4042 dn
= map_dev(sd
->disk
.major
, sd
->disk
.minor
, 1);
4043 devfd
= dev_open(dn
, O_RDONLY
);
4046 ok
= st
->ss
->load_super(st
, devfd
, NULL
);
4052 pr_err("Cannot find a superblock\n");
4056 memset(&bsb
, 0, 512);
4057 memcpy(bsb
.magic
, "md_backup_data-1", 16);
4058 st
->ss
->uuid_from_super(st
, uuid
);
4059 memcpy(bsb
.set_uuid
, uuid
, 16);
4060 bsb
.mtime
= __cpu_to_le64(time(0));
4061 bsb
.devstart2
= blocks
;
4063 stripes
= blocks
/ (sra
->array
.chunk_size
/512) /
4064 reshape
->before
.data_disks
;
4066 if (posix_memalign((void**)&buf
, 4096, disks
* chunk
))
4067 /* Don't start the 'reshape' */
4069 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
) {
4070 sysfs_get_ll(sra
, NULL
, "sync_speed_min", &speed
);
4071 sysfs_set_num(sra
, NULL
, "sync_speed_min", 200000);
4075 array_size
= sra
->component_size
* reshape
->after
.data_disks
;
4076 backup_point
= sra
->reshape_progress
;
4079 array_size
= sra
->component_size
* reshape
->before
.data_disks
;
4080 backup_point
= reshape
->backup_blocks
;
4081 suspend_point
= array_size
;
4087 /* Want to return as soon the oldest backup slot can
4088 * be released as that allows us to start backing up
4089 * some more, providing suspend_point has been
4090 * advanced, which it should have.
4093 wait_point
= array_size
;
4094 if (part
== 0 && __le64_to_cpu(bsb
.length
) > 0)
4095 wait_point
= (__le64_to_cpu(bsb
.arraystart
) +
4096 __le64_to_cpu(bsb
.length
));
4097 if (part
== 1 && __le64_to_cpu(bsb
.length2
) > 0)
4098 wait_point
= (__le64_to_cpu(bsb
.arraystart2
) +
4099 __le64_to_cpu(bsb
.length2
));
4102 if (part
== 0 && __le64_to_cpu(bsb
.length
) > 0)
4103 wait_point
= __le64_to_cpu(bsb
.arraystart
);
4104 if (part
== 1 && __le64_to_cpu(bsb
.length2
) > 0)
4105 wait_point
= __le64_to_cpu(bsb
.arraystart2
);
4108 rv
= progress_reshape(sra
, reshape
,
4109 backup_point
, wait_point
,
4110 &suspend_point
, &reshape_completed
);
4111 /* external metadata would need to ping_monitor here */
4112 sra
->reshape_progress
= reshape_completed
;
4114 /* Clear any backup region that is before 'here' */
4116 if (__le64_to_cpu(bsb
.length
) > 0 &&
4117 reshape_completed
>= (__le64_to_cpu(bsb
.arraystart
) +
4118 __le64_to_cpu(bsb
.length
)))
4119 forget_backup(dests
, destfd
,
4121 if (__le64_to_cpu(bsb
.length2
) > 0 &&
4122 reshape_completed
>= (__le64_to_cpu(bsb
.arraystart2
) +
4123 __le64_to_cpu(bsb
.length2
)))
4124 forget_backup(dests
, destfd
,
4127 if (__le64_to_cpu(bsb
.length
) > 0 &&
4128 reshape_completed
<= (__le64_to_cpu(bsb
.arraystart
)))
4129 forget_backup(dests
, destfd
,
4131 if (__le64_to_cpu(bsb
.length2
) > 0 &&
4132 reshape_completed
<= (__le64_to_cpu(bsb
.arraystart2
)))
4133 forget_backup(dests
, destfd
,
4142 if (rv
== 0 && increasing
&& !st
->ss
->external
) {
4143 /* No longer need to monitor this reshape */
4149 unsigned long long offset
;
4150 unsigned long actual_stripes
;
4151 /* Need to backup some data.
4152 * If 'part' is not used and the desired
4153 * backup size is suspended, do a backup,
4154 * then consider the next part.
4156 /* Check that 'part' is unused */
4157 if (part
== 0 && __le64_to_cpu(bsb
.length
) != 0)
4159 if (part
== 1 && __le64_to_cpu(bsb
.length2
) != 0)
4162 offset
= backup_point
/ data
;
4163 actual_stripes
= stripes
;
4165 if (offset
+ actual_stripes
* (chunk
/512) >
4166 sra
->component_size
)
4167 actual_stripes
= ((sra
->component_size
- offset
)
4169 if (offset
+ actual_stripes
* (chunk
/512) >
4173 if (offset
< actual_stripes
* (chunk
/512))
4174 actual_stripes
= offset
/ (chunk
/512);
4175 offset
-= actual_stripes
* (chunk
/512);
4176 if (offset
< suspend_point
/data
)
4179 if (actual_stripes
== 0)
4181 grow_backup(sra
, offset
, actual_stripes
,
4183 disks
, chunk
, level
, layout
,
4184 dests
, destfd
, destoffsets
,
4185 part
, °raded
, buf
);
4186 validate(afd
, destfd
[0], destoffsets
[0]);
4187 /* record where 'part' is up to */
4190 backup_point
+= actual_stripes
* (chunk
/512) * data
;
4192 backup_point
-= actual_stripes
* (chunk
/512) * data
;
4196 /* FIXME maybe call progress_reshape one more time instead */
4197 abort_reshape(sra
); /* remove any remaining suspension */
4198 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
)
4199 sysfs_set_num(sra
, NULL
, "sync_speed_min", speed
);
4205 * If any spare contains md_back_data-1 which is recent wrt mtime,
4206 * write that data into the array and update the super blocks with
4207 * the new reshape_progress
4209 int Grow_restart(struct supertype
*st
, struct mdinfo
*info
, int *fdlist
, int cnt
,
4210 char *backup_file
, int verbose
)
4214 unsigned long long *offsets
;
4215 unsigned long long nstripe
, ostripe
;
4218 odata
= info
->array
.raid_disks
- info
->delta_disks
- 1;
4219 if (info
->array
.level
== 6) odata
--; /* number of data disks */
4220 ndata
= info
->array
.raid_disks
- 1;
4221 if (info
->new_level
== 6) ndata
--;
4223 old_disks
= info
->array
.raid_disks
- info
->delta_disks
;
4225 if (info
->delta_disks
<= 0)
4226 /* Didn't grow, so the backup file must have
4230 for (i
=old_disks
-(backup_file
?1:0); i
<cnt
; i
++) {
4231 struct mdinfo dinfo
;
4234 char *devname
, namebuf
[20];
4235 unsigned long long lo
, hi
;
4237 /* This was a spare and may have some saved data on it.
4238 * Load the superblock, find and load the
4239 * backup_super_block.
4240 * If either fail, go on to next device.
4241 * If the backup contains no new info, just return
4242 * else restore data and update all superblocks
4244 if (i
== old_disks
-1) {
4245 fd
= open(backup_file
, O_RDONLY
);
4247 pr_err("backup file %s inaccessible: %s\n",
4248 backup_file
, strerror(errno
));
4251 devname
= backup_file
;
4256 if (st
->ss
->load_super(st
, fd
, NULL
))
4259 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
4260 st
->ss
->free_super(st
);
4263 (dinfo
.data_offset
+ dinfo
.component_size
- 8) <<9,
4265 pr_err("Cannot seek on device %d\n", i
);
4266 continue; /* Cannot seek */
4268 sprintf(namebuf
, "device-%d", i
);
4271 if (read(fd
, &bsb
, sizeof(bsb
)) != sizeof(bsb
)) {
4273 pr_err("Cannot read from %s\n", devname
);
4274 continue; /* Cannot read */
4276 if (memcmp(bsb
.magic
, "md_backup_data-1", 16) != 0 &&
4277 memcmp(bsb
.magic
, "md_backup_data-2", 16) != 0) {
4279 pr_err("No backup metadata on %s\n", devname
);
4282 if (bsb
.sb_csum
!= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
))) {
4284 pr_err("Bad backup-metadata checksum on %s\n", devname
);
4285 continue; /* bad checksum */
4287 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0 &&
4288 bsb
.sb_csum2
!= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum2
)-((char*)&bsb
))) {
4290 pr_err("Bad backup-metadata checksum2 on %s\n", devname
);
4291 continue; /* Bad second checksum */
4293 if (memcmp(bsb
.set_uuid
,info
->uuid
, 16) != 0) {
4295 pr_err("Wrong uuid on backup-metadata on %s\n", devname
);
4296 continue; /* Wrong uuid */
4299 /* array utime and backup-mtime should be updated at much the same time, but it seems that
4300 * sometimes they aren't... So allow considerable flexability in matching, and allow
4301 * this test to be overridden by an environment variable.
4303 if (info
->array
.utime
> (int)__le64_to_cpu(bsb
.mtime
) + 2*60*60 ||
4304 info
->array
.utime
< (int)__le64_to_cpu(bsb
.mtime
) - 10*60) {
4305 if (check_env("MDADM_GROW_ALLOW_OLD")) {
4306 pr_err("accepting backup with timestamp %lu "
4307 "for array with timestamp %lu\n",
4308 (unsigned long)__le64_to_cpu(bsb
.mtime
),
4309 (unsigned long)info
->array
.utime
);
4311 pr_err("too-old timestamp on backup-metadata on %s\n", devname
);
4312 pr_err("If you think it is should be safe, try 'export MDADM_GROW_ALLOW_OLD=1'\n");
4313 continue; /* time stamp is too bad */
4317 if (bsb
.magic
[15] == '1') {
4318 if (bsb
.length
== 0)
4320 if (info
->delta_disks
>= 0) {
4321 /* reshape_progress is increasing */
4322 if (__le64_to_cpu(bsb
.arraystart
)
4323 + __le64_to_cpu(bsb
.length
)
4324 < info
->reshape_progress
) {
4327 pr_err("backup-metadata found on %s but is not needed\n", devname
);
4328 continue; /* No new data here */
4331 /* reshape_progress is decreasing */
4332 if (__le64_to_cpu(bsb
.arraystart
) >=
4333 info
->reshape_progress
)
4334 goto nonew
; /* No new data here */
4337 if (bsb
.length
== 0 && bsb
.length2
== 0)
4339 if (info
->delta_disks
>= 0) {
4340 /* reshape_progress is increasing */
4341 if ((__le64_to_cpu(bsb
.arraystart
)
4342 + __le64_to_cpu(bsb
.length
)
4343 < info
->reshape_progress
)
4345 (__le64_to_cpu(bsb
.arraystart2
)
4346 + __le64_to_cpu(bsb
.length2
)
4347 < info
->reshape_progress
))
4348 goto nonew
; /* No new data here */
4350 /* reshape_progress is decreasing */
4351 if (__le64_to_cpu(bsb
.arraystart
) >=
4352 info
->reshape_progress
&&
4353 __le64_to_cpu(bsb
.arraystart2
) >=
4354 info
->reshape_progress
)
4355 goto nonew
; /* No new data here */
4358 if (lseek64(fd
, __le64_to_cpu(bsb
.devstart
)*512, 0)< 0) {
4361 pr_err("Failed to verify secondary backup-metadata block on %s\n",
4363 continue; /* Cannot seek */
4365 /* There should be a duplicate backup superblock 4k before here */
4366 if (lseek64(fd
, -4096, 1) < 0 ||
4367 read(fd
, &bsb2
, sizeof(bsb2
)) != sizeof(bsb2
))
4368 goto second_fail
; /* Cannot find leading superblock */
4369 if (bsb
.magic
[15] == '1')
4370 bsbsize
= offsetof(struct mdp_backup_super
, pad1
);
4372 bsbsize
= offsetof(struct mdp_backup_super
, pad
);
4373 if (memcmp(&bsb2
, &bsb
, bsbsize
) != 0)
4374 goto second_fail
; /* Cannot find leading superblock */
4376 /* Now need the data offsets for all devices. */
4377 offsets
= xmalloc(sizeof(*offsets
)*info
->array
.raid_disks
);
4378 for(j
=0; j
<info
->array
.raid_disks
; j
++) {
4381 if (st
->ss
->load_super(st
, fdlist
[j
], NULL
))
4382 /* FIXME should be this be an error */
4384 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
4385 st
->ss
->free_super(st
);
4386 offsets
[j
] = dinfo
.data_offset
* 512;
4388 printf(Name
": restoring critical section\n");
4390 if (restore_stripes(fdlist
, offsets
,
4391 info
->array
.raid_disks
,
4395 fd
, __le64_to_cpu(bsb
.devstart
)*512,
4396 __le64_to_cpu(bsb
.arraystart
)*512,
4397 __le64_to_cpu(bsb
.length
)*512, NULL
)) {
4398 /* didn't succeed, so giveup */
4400 pr_err("Error restoring backup from %s\n",
4406 if (bsb
.magic
[15] == '2' &&
4407 restore_stripes(fdlist
, offsets
,
4408 info
->array
.raid_disks
,
4412 fd
, __le64_to_cpu(bsb
.devstart
)*512 +
4413 __le64_to_cpu(bsb
.devstart2
)*512,
4414 __le64_to_cpu(bsb
.arraystart2
)*512,
4415 __le64_to_cpu(bsb
.length2
)*512, NULL
)) {
4416 /* didn't succeed, so giveup */
4418 pr_err("Error restoring second backup from %s\n",
4426 /* Ok, so the data is restored. Let's update those superblocks. */
4430 lo
= __le64_to_cpu(bsb
.arraystart
);
4431 hi
= lo
+ __le64_to_cpu(bsb
.length
);
4433 if (bsb
.magic
[15] == '2' && bsb
.length2
) {
4434 unsigned long long lo1
, hi1
;
4435 lo1
= __le64_to_cpu(bsb
.arraystart2
);
4436 hi1
= lo1
+ __le64_to_cpu(bsb
.length2
);
4440 } else if (lo
< lo1
)
4446 (info
->reshape_progress
< lo
||
4447 info
->reshape_progress
> hi
))
4448 /* backup does not affect reshape_progress*/ ;
4449 else if (info
->delta_disks
>= 0) {
4450 info
->reshape_progress
= __le64_to_cpu(bsb
.arraystart
) +
4451 __le64_to_cpu(bsb
.length
);
4452 if (bsb
.magic
[15] == '2') {
4453 unsigned long long p2
= __le64_to_cpu(bsb
.arraystart2
) +
4454 __le64_to_cpu(bsb
.length2
);
4455 if (p2
> info
->reshape_progress
)
4456 info
->reshape_progress
= p2
;
4459 info
->reshape_progress
= __le64_to_cpu(bsb
.arraystart
);
4460 if (bsb
.magic
[15] == '2') {
4461 unsigned long long p2
= __le64_to_cpu(bsb
.arraystart2
);
4462 if (p2
< info
->reshape_progress
)
4463 info
->reshape_progress
= p2
;
4466 for (j
=0; j
<info
->array
.raid_disks
; j
++) {
4469 if (st
->ss
->load_super(st
, fdlist
[j
], NULL
))
4471 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
4472 dinfo
.reshape_progress
= info
->reshape_progress
;
4473 st
->ss
->update_super(st
, &dinfo
,
4474 "_reshape_progress",
4476 st
->ss
->store_super(st
, fdlist
[j
]);
4477 st
->ss
->free_super(st
);
4481 /* Didn't find any backup data, try to see if any
4484 if (info
->delta_disks
< 0) {
4485 /* When shrinking, the critical section is at the end.
4486 * So see if we are before the critical section.
4488 unsigned long long first_block
;
4489 nstripe
= ostripe
= 0;
4491 while (ostripe
>= nstripe
) {
4492 ostripe
+= info
->array
.chunk_size
/ 512;
4493 first_block
= ostripe
* odata
;
4494 nstripe
= first_block
/ ndata
/ (info
->new_chunk
/512) *
4495 (info
->new_chunk
/512);
4498 if (info
->reshape_progress
>= first_block
)
4501 if (info
->delta_disks
> 0) {
4502 /* See if we are beyond the critical section. */
4503 unsigned long long last_block
;
4504 nstripe
= ostripe
= 0;
4506 while (nstripe
>= ostripe
) {
4507 nstripe
+= info
->new_chunk
/ 512;
4508 last_block
= nstripe
* ndata
;
4509 ostripe
= last_block
/ odata
/ (info
->array
.chunk_size
/512) *
4510 (info
->array
.chunk_size
/512);
4513 if (info
->reshape_progress
>= last_block
)
4516 /* needed to recover critical section! */
4518 pr_err("Failed to find backup of critical section\n");
4522 int Grow_continue_command(char *devname
, int fd
,
4523 char *backup_file
, int verbose
)
4526 struct supertype
*st
= NULL
;
4527 struct mdinfo
*content
= NULL
;
4528 struct mdinfo array
;
4529 char *subarray
= NULL
;
4530 struct mdinfo
*cc
= NULL
;
4531 struct mdstat_ent
*mdstat
= NULL
;
4535 dprintf("Grow continue from command line called for %s\n",
4538 st
= super_by_fd(fd
, &subarray
);
4539 if (!st
|| !st
->ss
) {
4540 pr_err("Unable to determine metadata format for %s\n",
4544 dprintf("Grow continue is run for ");
4545 if (st
->ss
->external
== 0) {
4547 dprintf("native array (%s)\n", devname
);
4548 if (ioctl(fd
, GET_ARRAY_INFO
, &array
.array
) < 0) {
4549 pr_err("%s is not an active md array -"
4550 " aborting\n", devname
);
4552 goto Grow_continue_command_exit
;
4555 /* Need to load a superblock.
4556 * FIXME we should really get what we need from
4559 for (d
= 0; d
< MAX_DISKS
; d
++) {
4560 mdu_disk_info_t disk
;
4564 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0)
4566 if (disk
.major
== 0 && disk
.minor
== 0)
4568 if ((disk
.state
& (1 << MD_DISK_ACTIVE
)) == 0)
4570 dv
= map_dev(disk
.major
, disk
.minor
, 1);
4573 fd2
= dev_open(dv
, O_RDONLY
);
4576 err
= st
->ss
->load_super(st
, fd2
, NULL
);
4582 if (d
== MAX_DISKS
) {
4583 pr_err("Unable to load metadata for %s\n",
4586 goto Grow_continue_command_exit
;
4588 st
->ss
->getinfo_super(st
, content
, NULL
);
4593 dprintf("subarray (%s)\n", subarray
);
4594 container
= st
->container_devnm
;
4595 cfd
= open_dev_excl(st
->container_devnm
);
4597 container
= st
->devnm
;
4599 cfd
= open_dev_excl(st
->devnm
);
4600 dprintf("container (%s)\n", container
);
4604 pr_err("Unable to open container "
4605 "for %s\n", devname
);
4607 goto Grow_continue_command_exit
;
4610 /* find in container array under reshape
4612 ret_val
= st
->ss
->load_container(st
, cfd
, NULL
);
4614 pr_err("Cannot read superblock for %s\n",
4617 goto Grow_continue_command_exit
;
4620 cc
= st
->ss
->container_content(st
, subarray
);
4621 for (content
= cc
; content
; content
= content
->next
) {
4623 int allow_reshape
= 1;
4625 if (content
->reshape_active
== 0)
4627 /* The decision about array or container wide
4628 * reshape is taken in Grow_continue based
4629 * content->reshape_active state, therefore we
4630 * need to check_reshape based on
4631 * reshape_active and subarray name
4633 if (content
->array
.state
& (1<<MD_SB_BLOCK_VOLUME
))
4635 if (content
->reshape_active
== CONTAINER_RESHAPE
&&
4636 (content
->array
.state
4637 & (1<<MD_SB_BLOCK_CONTAINER_RESHAPE
)))
4640 if (!allow_reshape
) {
4641 pr_err("cannot continue reshape of an array"
4642 " in container with unsupported"
4643 " metadata: %s(%s)\n",
4644 devname
, container
);
4646 goto Grow_continue_command_exit
;
4649 array
= strchr(content
->text_version
+1, '/')+1;
4650 mdstat
= mdstat_by_subdev(array
, container
);
4653 if (mdstat
->active
== 0) {
4654 pr_err("Skipping inactive array %s.\n",
4656 free_mdstat(mdstat
);
4663 pr_err("Unable to determine reshaped "
4664 "array for %s\n", devname
);
4666 goto Grow_continue_command_exit
;
4668 fd2
= open_dev(mdstat
->devnm
);
4670 pr_err("cannot open (%s)\n", mdstat
->devnm
);
4672 goto Grow_continue_command_exit
;
4675 sysfs_init(content
, fd2
, mdstat
->devnm
);
4677 /* start mdmon in case it is not running
4679 if (!mdmon_running(container
))
4680 start_mdmon(container
);
4681 ping_monitor(container
);
4683 if (mdmon_running(container
))
4684 st
->update_tail
= &st
->updates
;
4686 pr_err("No mdmon found. "
4687 "Grow cannot continue.\n");
4689 goto Grow_continue_command_exit
;
4693 /* verify that array under reshape is started from
4696 if (verify_reshape_position(content
, content
->array
.level
) < 0) {
4698 goto Grow_continue_command_exit
;
4703 ret_val
= Grow_continue(fd
, st
, content
, backup_file
, 0);
4705 Grow_continue_command_exit
:
4710 st
->ss
->free_super(st
);
4711 free_mdstat(mdstat
);
4718 int Grow_continue(int mdfd
, struct supertype
*st
, struct mdinfo
*info
,
4719 char *backup_file
, int freeze_reshape
)
4723 if (!info
->reshape_active
)
4726 if (st
->ss
->external
) {
4727 int cfd
= open_dev(st
->container_devnm
);
4732 st
->ss
->load_container(st
, cfd
, st
->container_devnm
);
4734 ret_val
= reshape_container(st
->container_devnm
, NULL
, mdfd
,
4735 st
, info
, 0, backup_file
,
4737 1 | info
->reshape_active
,
4740 ret_val
= reshape_array(NULL
, mdfd
, "array", st
, info
, 1,
4741 NULL
, 0ULL, backup_file
, 0, 0,
4742 1 | info
->reshape_active
,