2 * mdadm - manage Linux "md" devices aka RAID arrays.
4 * Copyright (C) 2001-2013 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>
30 #if ! defined(__BIG_ENDIAN) && ! defined(__LITTLE_ENDIAN)
31 #error no endian defined
37 #define offsetof(t,f) ((size_t)&(((t*)0)->f))
40 int restore_backup(struct supertype
*st
,
41 struct mdinfo
*content
,
51 int disk_count
= next_spare
+ working_disks
;
53 dprintf("Called restore_backup()\n");
54 fdlist
= xmalloc(sizeof(int) * disk_count
);
56 enable_fds(next_spare
);
57 for (i
= 0; i
< next_spare
; i
++)
59 for (dev
= content
->devs
; dev
; dev
= dev
->next
) {
65 fd
= dev_open(buf
, O_RDWR
);
67 if (dev
->disk
.raid_disk
>= 0)
68 fdlist
[dev
->disk
.raid_disk
] = fd
;
70 fdlist
[next_spare
++] = fd
;
73 if (st
->ss
->external
&& st
->ss
->recover_backup
)
74 err
= st
->ss
->recover_backup(st
, content
);
76 err
= Grow_restart(st
, content
, fdlist
, next_spare
,
77 backup_file
, verbose
> 0);
79 while (next_spare
> 0) {
81 if (fdlist
[next_spare
] >= 0)
82 close(fdlist
[next_spare
]);
86 pr_err("Failed to restore critical"
87 " section for reshape - sorry.\n");
89 pr_err("Possibly you need"
90 " to specify a --backup-file\n");
94 dprintf("restore_backup() returns status OK.\n");
98 int Grow_Add_device(char *devname
, int fd
, char *newdev
)
100 /* Add a device to an active array.
101 * Currently, just extend a linear array.
102 * This requires writing a new superblock on the
103 * new device, calling the kernel to add the device,
104 * and if that succeeds, update the superblock on
106 * This means that we need to *find* all other devices.
113 struct supertype
*st
= NULL
;
114 char *subarray
= NULL
;
116 if (ioctl(fd
, GET_ARRAY_INFO
, &info
.array
) < 0) {
117 pr_err("cannot get array info for %s\n", devname
);
121 if (info
.array
.level
!= -1) {
122 pr_err("can only add devices to linear arrays\n");
126 st
= super_by_fd(fd
, &subarray
);
128 pr_err("cannot handle arrays with superblock version %d\n",
129 info
.array
.major_version
);
134 pr_err("Cannot grow linear sub-arrays yet\n");
140 nfd
= open(newdev
, O_RDWR
|O_EXCL
|O_DIRECT
);
142 pr_err("cannot open %s\n", newdev
);
147 if ((stb
.st_mode
& S_IFMT
) != S_IFBLK
) {
148 pr_err("%s is not a block device!\n", newdev
);
153 /* now check out all the devices and make sure we can read the
155 for (d
=0 ; d
< info
.array
.raid_disks
; d
++) {
156 mdu_disk_info_t disk
;
159 st
->ss
->free_super(st
);
162 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0) {
163 pr_err("cannot get device detail for device %d\n",
169 dv
= map_dev(disk
.major
, disk
.minor
, 1);
171 pr_err("cannot find device file for device %d\n",
177 fd2
= dev_open(dv
, O_RDWR
);
179 pr_err("cannot open device file %s\n", dv
);
185 if (st
->ss
->load_super(st
, fd2
, NULL
)) {
186 pr_err("cannot find super block on %s\n", dv
);
194 /* Ok, looks good. Lets update the superblock and write it out to
198 info
.disk
.number
= d
;
199 info
.disk
.major
= major(stb
.st_rdev
);
200 info
.disk
.minor
= minor(stb
.st_rdev
);
201 info
.disk
.raid_disk
= d
;
202 info
.disk
.state
= (1 << MD_DISK_SYNC
) | (1 << MD_DISK_ACTIVE
);
203 st
->ss
->update_super(st
, &info
, "linear-grow-new", newdev
,
206 if (st
->ss
->store_super(st
, nfd
)) {
207 pr_err("Cannot store new superblock on %s\n",
214 if (ioctl(fd
, ADD_NEW_DISK
, &info
.disk
) != 0) {
215 pr_err("Cannot add new disk to this array\n");
218 /* Well, that seems to have worked.
219 * Now go through and update all superblocks
222 if (ioctl(fd
, GET_ARRAY_INFO
, &info
.array
) < 0) {
223 pr_err("cannot get array info for %s\n", devname
);
228 for (d
=0 ; d
< info
.array
.raid_disks
; d
++) {
229 mdu_disk_info_t disk
;
233 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0) {
234 pr_err("cannot get device detail for device %d\n",
238 dv
= map_dev(disk
.major
, disk
.minor
, 1);
240 pr_err("cannot find device file for device %d\n",
244 fd2
= dev_open(dv
, O_RDWR
);
246 pr_err("cannot open device file %s\n", dv
);
249 if (st
->ss
->load_super(st
, fd2
, NULL
)) {
250 pr_err("cannot find super block on %s\n", dv
);
254 info
.array
.raid_disks
= nd
+1;
255 info
.array
.nr_disks
= nd
+1;
256 info
.array
.active_disks
= nd
+1;
257 info
.array
.working_disks
= nd
+1;
259 st
->ss
->update_super(st
, &info
, "linear-grow-update", dv
,
262 if (st
->ss
->store_super(st
, fd2
)) {
263 pr_err("Cannot store new superblock on %s\n", dv
);
273 int Grow_addbitmap(char *devname
, int fd
, struct context
*c
, struct shape
*s
)
276 * First check that array doesn't have a bitmap
277 * Then create the bitmap
280 * For internal bitmaps, we need to check the version,
281 * find all the active devices, and write the bitmap block
284 mdu_bitmap_file_t bmf
;
285 mdu_array_info_t array
;
286 struct supertype
*st
;
287 char *subarray
= NULL
;
288 int major
= BITMAP_MAJOR_HI
;
289 int vers
= md_get_version(fd
);
290 unsigned long long bitmapsize
, array_size
;
293 major
= BITMAP_MAJOR_HOSTENDIAN
;
294 pr_err("Warning - bitmaps created on this kernel"
295 " are not portable\n"
296 " between different architectures. Consider upgrading"
297 " the Linux kernel.\n");
300 if (ioctl(fd
, GET_BITMAP_FILE
, &bmf
) != 0) {
302 pr_err("Memory allocation failure.\n");
304 pr_err("bitmaps not supported by this kernel.\n");
307 if (bmf
.pathname
[0]) {
308 if (strcmp(s
->bitmap_file
,"none")==0) {
309 if (ioctl(fd
, SET_BITMAP_FILE
, -1)!= 0) {
310 pr_err("failed to remove bitmap %s\n",
316 pr_err("%s already has a bitmap (%s)\n",
317 devname
, bmf
.pathname
);
320 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
321 pr_err("cannot get array status for %s\n", devname
);
324 if (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)) {
325 if (strcmp(s
->bitmap_file
, "none")==0) {
326 array
.state
&= ~(1<<MD_SB_BITMAP_PRESENT
);
327 if (ioctl(fd
, SET_ARRAY_INFO
, &array
)!= 0) {
328 pr_err("failed to remove internal bitmap.\n");
333 pr_err("Internal bitmap already present on %s\n",
338 if (strcmp(s
->bitmap_file
, "none") == 0) {
339 pr_err("no bitmap found on %s\n", devname
);
342 if (array
.level
<= 0) {
343 pr_err("Bitmaps not meaningful with level %s\n",
344 map_num(pers
, array
.level
)?:"of this array");
347 bitmapsize
= array
.size
;
349 if (get_dev_size(fd
, NULL
, &array_size
) &&
350 array_size
> (0x7fffffffULL
<<9)) {
351 /* Array is big enough that we cannot trust array.size
352 * try other approaches
354 bitmapsize
= get_component_size(fd
);
356 if (bitmapsize
== 0) {
357 pr_err("Cannot reliably determine size of array to create bitmap - sorry.\n");
361 if (array
.level
== 10) {
362 int ncopies
= (array
.layout
&255)*((array
.layout
>>8)&255);
363 bitmapsize
= bitmapsize
* array
.raid_disks
/ ncopies
;
366 st
= super_by_fd(fd
, &subarray
);
368 pr_err("Cannot understand version %d.%d\n",
369 array
.major_version
, array
.minor_version
);
373 pr_err("Cannot add bitmaps to sub-arrays yet\n");
378 if (strcmp(s
->bitmap_file
, "internal") == 0) {
381 int offset_setable
= 0;
383 if (st
->ss
->add_internal_bitmap
== NULL
) {
384 pr_err("Internal bitmaps not supported "
385 "with %s metadata\n", st
->ss
->name
);
388 mdi
= sysfs_read(fd
, NULL
, GET_BITMAP_LOCATION
);
391 for (d
=0; d
< st
->max_devs
; d
++) {
392 mdu_disk_info_t disk
;
395 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0)
397 if (disk
.major
== 0 &&
400 if ((disk
.state
& (1<<MD_DISK_SYNC
))==0)
402 dv
= map_dev(disk
.major
, disk
.minor
, 1);
404 int fd2
= dev_open(dv
, O_RDWR
);
407 if (st
->ss
->load_super(st
, fd2
, NULL
)==0) {
408 if (st
->ss
->add_internal_bitmap(
410 &s
->bitmap_chunk
, c
->delay
, s
->write_behind
,
411 bitmapsize
, offset_setable
,
414 st
->ss
->write_bitmap(st
, fd2
);
416 pr_err("failed to create internal bitmap"
417 " - chunksize problem.\n");
425 if (offset_setable
) {
426 st
->ss
->getinfo_super(st
, mdi
, NULL
);
427 sysfs_init(mdi
, fd
, NULL
);
428 rv
= sysfs_set_num_signed(mdi
, NULL
, "bitmap/location",
431 array
.state
|= (1<<MD_SB_BITMAP_PRESENT
);
432 rv
= ioctl(fd
, SET_ARRAY_INFO
, &array
);
436 pr_err("Cannot add bitmap while array is"
437 " resyncing or reshaping etc.\n");
438 pr_err("failed to set internal bitmap.\n");
445 int max_devs
= st
->max_devs
;
447 /* try to load a superblock */
448 for (d
= 0; d
< max_devs
; d
++) {
449 mdu_disk_info_t disk
;
453 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0)
455 if ((disk
.major
==0 && disk
.minor
==0) ||
456 (disk
.state
& (1<<MD_DISK_REMOVED
)))
458 dv
= map_dev(disk
.major
, disk
.minor
, 1);
461 fd2
= dev_open(dv
, O_RDONLY
);
463 if (st
->ss
->load_super(st
, fd2
, NULL
) == 0) {
465 st
->ss
->uuid_from_super(st
, uuid
);
472 pr_err("cannot find UUID for array!\n");
475 if (CreateBitmap(s
->bitmap_file
, c
->force
, (char*)uuid
, s
->bitmap_chunk
,
476 c
->delay
, s
->write_behind
, bitmapsize
, major
)) {
479 bitmap_fd
= open(s
->bitmap_file
, O_RDWR
);
481 pr_err("weird: %s cannot be opened\n",
485 if (ioctl(fd
, SET_BITMAP_FILE
, bitmap_fd
) < 0) {
488 pr_err("Cannot add bitmap while array is"
489 " resyncing or reshaping etc.\n");
490 pr_err("Cannot set bitmap file for %s: %s\n",
491 devname
, strerror(err
));
500 * When reshaping an array we might need to backup some data.
501 * This is written to all spares with a 'super_block' describing it.
502 * The superblock goes 4K from the end of the used space on the
504 * It if written after the backup is complete.
505 * It has the following structure.
508 static struct mdp_backup_super
{
509 char magic
[16]; /* md_backup_data-1 or -2 */
512 /* start/sizes in 512byte sectors */
513 __u64 devstart
; /* address on backup device/file of data */
516 __u32 sb_csum
; /* csum of preceeding bytes. */
518 __u64 devstart2
; /* offset in to data of second section */
521 __u32 sb_csum2
; /* csum of preceeding bytes. */
523 } __attribute__((aligned(512))) bsb
, bsb2
;
525 static __u32
bsb_csum(char *buf
, int len
)
529 for (i
= 0; i
< len
; i
++)
530 csum
= (csum
<<3) + buf
[0];
531 return __cpu_to_le32(csum
);
534 static int check_idle(struct supertype
*st
)
536 /* Check that all member arrays for this container, or the
537 * container of this array, are idle
539 char *container
= (st
->container_devnm
[0]
540 ? st
->container_devnm
: st
->devnm
);
541 struct mdstat_ent
*ent
, *e
;
544 ent
= mdstat_read(0, 0);
545 for (e
= ent
; e
; e
= e
->next
) {
546 if (!is_container_member(e
, container
))
548 if (e
->percent
>= 0) {
557 static int freeze_container(struct supertype
*st
)
559 char *container
= (st
->container_devnm
[0]
560 ? st
->container_devnm
: st
->devnm
);
565 if (block_monitor(container
, 1)) {
566 pr_err("failed to freeze container\n");
573 static void unfreeze_container(struct supertype
*st
)
575 char *container
= (st
->container_devnm
[0]
576 ? st
->container_devnm
: st
->devnm
);
578 unblock_monitor(container
, 1);
581 static int freeze(struct supertype
*st
)
583 /* Try to freeze resync/rebuild on this array/container.
584 * Return -1 if the array is busy,
585 * return -2 container cannot be frozen,
586 * return 0 if this kernel doesn't support 'frozen'
587 * return 1 if it worked.
589 if (st
->ss
->external
)
590 return freeze_container(st
);
592 struct mdinfo
*sra
= sysfs_read(-1, st
->devnm
, GET_VERSION
);
598 /* Need to clear any 'read-auto' status */
599 if (sysfs_get_str(sra
, NULL
, "array_state", buf
, 20) > 0 &&
600 strncmp(buf
, "read-auto", 9) == 0)
601 sysfs_set_str(sra
, NULL
, "array_state", "clean");
603 err
= sysfs_freeze_array(sra
);
609 static void unfreeze(struct supertype
*st
)
611 if (st
->ss
->external
)
612 return unfreeze_container(st
);
614 struct mdinfo
*sra
= sysfs_read(-1, st
->devnm
, GET_VERSION
);
618 sysfs_get_str(sra
, NULL
, "sync_action", buf
, 20) > 0
619 && strcmp(buf
, "frozen\n") == 0) {
620 printf("unfreeze\n");
621 sysfs_set_str(sra
, NULL
, "sync_action", "idle");
627 static void wait_reshape(struct mdinfo
*sra
)
629 int fd
= sysfs_get_fd(sra
, NULL
, "sync_action");
635 while (sysfs_fd_get_str(fd
, action
, 20) > 0 &&
636 strncmp(action
, "reshape", 7) == 0)
637 sysfs_wait(fd
, NULL
);
641 static int reshape_super(struct supertype
*st
, unsigned long long size
,
642 int level
, int layout
, int chunksize
, int raid_disks
,
643 int delta_disks
, char *backup_file
, char *dev
,
644 int direction
, int verbose
)
646 /* nothing extra to check in the native case */
647 if (!st
->ss
->external
)
649 if (!st
->ss
->reshape_super
||
650 !st
->ss
->manage_reshape
) {
651 pr_err("%s metadata does not support reshape\n",
656 return st
->ss
->reshape_super(st
, size
, level
, layout
, chunksize
,
657 raid_disks
, delta_disks
, backup_file
, dev
,
661 static void sync_metadata(struct supertype
*st
)
663 if (st
->ss
->external
) {
664 if (st
->update_tail
) {
665 flush_metadata_updates(st
);
666 st
->update_tail
= &st
->updates
;
668 st
->ss
->sync_metadata(st
);
672 static int subarray_set_num(char *container
, struct mdinfo
*sra
, char *name
, int n
)
674 /* when dealing with external metadata subarrays we need to be
675 * prepared to handle EAGAIN. The kernel may need to wait for
676 * mdmon to mark the array active so the kernel can handle
677 * allocations/writeback when preparing the reshape action
678 * (md_allow_write()). We temporarily disable safe_mode_delay
679 * to close a race with the array_state going clean before the
680 * next write to raid_disks / stripe_cache_size
685 /* only 'raid_disks' and 'stripe_cache_size' trigger md_allow_write */
687 (strcmp(name
, "raid_disks") != 0 &&
688 strcmp(name
, "stripe_cache_size") != 0))
689 return sysfs_set_num(sra
, NULL
, name
, n
);
691 rc
= sysfs_get_str(sra
, NULL
, "safe_mode_delay", safe
, sizeof(safe
));
694 sysfs_set_num(sra
, NULL
, "safe_mode_delay", 0);
695 rc
= sysfs_set_num(sra
, NULL
, name
, n
);
696 if (rc
< 0 && errno
== EAGAIN
) {
697 ping_monitor(container
);
698 /* if we get EAGAIN here then the monitor is not active
701 rc
= sysfs_set_num(sra
, NULL
, name
, n
);
703 sysfs_set_str(sra
, NULL
, "safe_mode_delay", safe
);
707 int start_reshape(struct mdinfo
*sra
, int already_running
,
708 int before_data_disks
, int data_disks
)
711 unsigned long long sync_max_to_set
;
713 sysfs_set_num(sra
, NULL
, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL
);
714 err
= sysfs_set_num(sra
, NULL
, "suspend_hi", sra
->reshape_progress
);
715 err
= err
?: sysfs_set_num(sra
, NULL
, "suspend_lo",
716 sra
->reshape_progress
);
717 if (before_data_disks
<= data_disks
)
718 sync_max_to_set
= sra
->reshape_progress
/ data_disks
;
720 sync_max_to_set
= (sra
->component_size
* data_disks
721 - sra
->reshape_progress
) / data_disks
;
722 if (!already_running
)
723 sysfs_set_num(sra
, NULL
, "sync_min", sync_max_to_set
);
724 err
= err
?: sysfs_set_num(sra
, NULL
, "sync_max", sync_max_to_set
);
725 if (!already_running
)
726 err
= err
?: sysfs_set_str(sra
, NULL
, "sync_action", "reshape");
731 void abort_reshape(struct mdinfo
*sra
)
733 sysfs_set_str(sra
, NULL
, "sync_action", "idle");
734 sysfs_set_num(sra
, NULL
, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL
);
735 sysfs_set_num(sra
, NULL
, "suspend_hi", 0);
736 sysfs_set_num(sra
, NULL
, "suspend_lo", 0);
737 sysfs_set_num(sra
, NULL
, "sync_min", 0);
738 // It isn't safe to reset sync_max as we aren't monitoring.
739 // Array really should be stopped at this point.
742 int remove_disks_for_takeover(struct supertype
*st
,
747 struct mdinfo
*remaining
;
750 if (sra
->array
.level
== 10)
751 nr_of_copies
= layout
& 0xff;
752 else if (sra
->array
.level
== 1)
753 nr_of_copies
= sra
->array
.raid_disks
;
757 remaining
= sra
->devs
;
759 /* for each 'copy', select one device and remove from the list. */
760 for (slot
= 0; slot
< sra
->array
.raid_disks
; slot
+= nr_of_copies
) {
761 struct mdinfo
**diskp
;
764 /* Find a working device to keep */
765 for (diskp
= &remaining
; *diskp
; diskp
= &(*diskp
)->next
) {
766 struct mdinfo
*disk
= *diskp
;
768 if (disk
->disk
.raid_disk
< slot
)
770 if (disk
->disk
.raid_disk
>= slot
+ nr_of_copies
)
772 if (disk
->disk
.state
& (1<<MD_DISK_REMOVED
))
774 if (disk
->disk
.state
& (1<<MD_DISK_FAULTY
))
776 if (!(disk
->disk
.state
& (1<<MD_DISK_SYNC
)))
779 /* We have found a good disk to use! */
781 disk
->next
= sra
->devs
;
790 if (slot
< sra
->array
.raid_disks
) {
791 /* didn't find all slots */
797 sra
->devs
= remaining
;
801 /* Remove all 'remaining' devices from the array */
803 struct mdinfo
*sd
= remaining
;
804 remaining
= sd
->next
;
806 sysfs_set_str(sra
, sd
, "state", "faulty");
807 sysfs_set_str(sra
, sd
, "slot", "none");
808 /* for external metadata disks should be removed in mdmon */
809 if (!st
->ss
->external
)
810 sysfs_set_str(sra
, sd
, "state", "remove");
811 sd
->disk
.state
|= (1<<MD_DISK_REMOVED
);
812 sd
->disk
.state
&= ~(1<<MD_DISK_SYNC
);
813 sd
->next
= sra
->devs
;
819 void reshape_free_fdlist(int *fdlist
,
820 unsigned long long *offsets
,
825 for (i
= 0; i
< size
; i
++)
833 int reshape_prepare_fdlist(char *devname
,
837 unsigned long blocks
,
840 unsigned long long *offsets
)
846 for (d
= 0; d
<= nrdisks
; d
++)
849 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
850 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
852 if (sd
->disk
.state
& (1<<MD_DISK_SYNC
)) {
853 char *dn
= map_dev(sd
->disk
.major
,
855 fdlist
[sd
->disk
.raid_disk
]
856 = dev_open(dn
, O_RDONLY
);
857 offsets
[sd
->disk
.raid_disk
] = sd
->data_offset
*512;
858 if (fdlist
[sd
->disk
.raid_disk
] < 0) {
859 pr_err("%s: cannot open component %s\n",
860 devname
, dn
? dn
: "-unknown-");
864 } else if (backup_file
== NULL
) {
866 char *dn
= map_dev(sd
->disk
.major
,
868 fdlist
[d
] = dev_open(dn
, O_RDWR
);
869 offsets
[d
] = (sd
->data_offset
+ sra
->component_size
- blocks
- 8)*512;
871 pr_err("%s: cannot open component %s\n",
872 devname
, dn
? dn
: "-unknown-");
883 int reshape_open_backup_file(char *backup_file
,
888 unsigned long long *offsets
,
891 /* Return 1 on success, 0 on any form of failure */
892 /* need to check backup file is large enough */
898 *fdlist
= open(backup_file
, O_RDWR
|O_CREAT
|(restart
? O_TRUNC
: O_EXCL
),
902 pr_err("%s: cannot create backup file %s: %s\n",
903 devname
, backup_file
, strerror(errno
));
906 /* Guard against backup file being on array device.
907 * If array is partitioned or if LVM etc is in the
908 * way this will not notice, but it is better than
911 fstat(*fdlist
, &stb
);
914 if (stb
.st_rdev
== dev
) {
915 pr_err("backup file must NOT be"
916 " on the array being reshaped.\n");
922 for (i
=0; i
< blocks
+ 8 ; i
++) {
923 if (write(*fdlist
, buf
, 512) != 512) {
924 pr_err("%s: cannot create"
925 " backup file %s: %s\n",
926 devname
, backup_file
, strerror(errno
));
930 if (fsync(*fdlist
) != 0) {
931 pr_err("%s: cannot create backup file %s: %s\n",
932 devname
, backup_file
, strerror(errno
));
939 unsigned long compute_backup_blocks(int nchunk
, int ochunk
,
940 unsigned int ndata
, unsigned int odata
)
942 unsigned long a
, b
, blocks
;
943 /* So how much do we need to backup.
944 * We need an amount of data which is both a whole number of
945 * old stripes and a whole number of new stripes.
946 * So LCM for (chunksize*datadisks).
948 a
= (ochunk
/512) * odata
;
949 b
= (nchunk
/512) * ndata
;
952 /* LCM == product / GCD */
953 blocks
= (ochunk
/512) * (nchunk
/512) * odata
* ndata
/ a
;
958 char *analyse_change(char *devname
, struct mdinfo
*info
, struct reshape
*re
)
960 /* Based on the current array state in info->array and
961 * the changes in info->new_* etc, determine:
962 * - whether the change is possible
963 * - Intermediate level/raid_disks/layout
964 * - whether a restriping reshape is needed
965 * - number of sectors in minimum change unit. This
966 * will cover a whole number of stripes in 'before' and
969 * Return message if the change should be rejected
970 * NULL if the change can be achieved
972 * This can be called as part of starting a reshape, or
973 * when assembling an array that is undergoing reshape.
975 int near
, far
, offset
, copies
;
977 int old_chunk
, new_chunk
;
978 /* delta_parity records change in number of devices
979 * caused by level change
981 int delta_parity
= 0;
983 memset(re
, 0, sizeof(*re
));
985 /* If a new level not explicitly given, we assume no-change */
986 if (info
->new_level
== UnSet
)
987 info
->new_level
= info
->array
.level
;
990 switch (info
->new_level
) {
996 /* chunk size is meaningful, must divide component_size
999 if (info
->component_size
% (info
->new_chunk
/512)) {
1000 unsigned long long shrink
= info
->component_size
;
1001 shrink
&= ~(unsigned long long)(info
->new_chunk
/512-1);
1002 pr_err("New chunk size (%dK) does not evenly divide device size (%lluk)\n",
1003 info
->new_chunk
/1024, info
->component_size
/2);
1004 pr_err("After shrinking any filesystem, \"mdadm --grow %s --size %llu\"\n",
1006 pr_err("will shrink the array so the given chunk size would work.\n");
1011 return "chunk size not meaningful for this level";
1014 info
->new_chunk
= info
->array
.chunk_size
;
1016 switch (info
->array
.level
) {
1018 return "Cannot understand this RAID level";
1020 /* RAID1 can convert to RAID1 with different disks, or
1021 * raid5 with 2 disks, or
1024 if (info
->new_level
> 1 &&
1025 (info
->component_size
& 7))
1026 return "Cannot convert RAID1 of this size - "
1027 "reduce size to multiple of 4K first.";
1028 if (info
->new_level
== 0) {
1029 if (info
->delta_disks
!= UnSet
&&
1030 info
->delta_disks
!= 0)
1031 return "Cannot change number of disks "
1032 "with RAID1->RAID0 conversion";
1034 re
->before
.data_disks
= 1;
1035 re
->after
.data_disks
= 1;
1038 if (info
->new_level
== 1) {
1039 if (info
->delta_disks
== UnSet
)
1040 /* Don't know what to do */
1041 return "no change requested for Growing RAID1";
1045 if (info
->array
.raid_disks
== 2 &&
1046 info
->new_level
== 5) {
1049 re
->before
.data_disks
= 1;
1050 if (info
->delta_disks
!= UnSet
&&
1051 info
->delta_disks
!= 0)
1052 re
->after
.data_disks
= 1 + info
->delta_disks
;
1054 re
->after
.data_disks
= 1;
1055 if (re
->after
.data_disks
< 1)
1056 return "Number of disks too small for RAID5";
1058 re
->before
.layout
= ALGORITHM_LEFT_SYMMETRIC
;
1059 info
->array
.chunk_size
= 65536;
1062 /* Could do some multi-stage conversions, but leave that to
1065 return "Impossibly level change request for RAID1";
1068 /* RAID10 can be converted from near mode to
1069 * RAID0 by removing some devices.
1070 * It can also be reshaped if the kernel supports
1073 switch (info
->new_level
) {
1075 if ((info
->array
.layout
& ~0xff) != 0x100)
1076 return "Cannot Grow RAID10 with far/offset layout";
1077 /* number of devices must be multiple of number of copies */
1078 if (info
->array
.raid_disks
% (info
->array
.layout
& 0xff))
1079 return "RAID10 layout too complex for Grow operation";
1081 new_disks
= (info
->array
.raid_disks
1082 / (info
->array
.layout
& 0xff));
1083 if (info
->delta_disks
== UnSet
)
1084 info
->delta_disks
= (new_disks
1085 - info
->array
.raid_disks
);
1087 if (info
->delta_disks
!= new_disks
- info
->array
.raid_disks
)
1088 return "New number of raid-devices impossible for RAID10";
1089 if (info
->new_chunk
&&
1090 info
->new_chunk
!= info
->array
.chunk_size
)
1091 return "Cannot change chunk-size with RAID10 Grow";
1095 re
->before
.data_disks
= new_disks
;
1096 re
->after
.data_disks
= re
->before
.data_disks
;
1100 near
= info
->array
.layout
& 0xff;
1101 far
= (info
->array
.layout
>> 8) & 0xff;
1102 offset
= info
->array
.layout
& 0x10000;
1103 if (far
> 1 && !offset
)
1104 return "Cannot reshape RAID10 in far-mode";
1105 copies
= near
* far
;
1107 old_chunk
= info
->array
.chunk_size
* far
;
1109 if (info
->new_layout
== UnSet
)
1110 info
->new_layout
= info
->array
.layout
;
1112 near
= info
->new_layout
& 0xff;
1113 far
= (info
->new_layout
>> 8) & 0xff;
1114 offset
= info
->new_layout
& 0x10000;
1115 if (far
> 1 && !offset
)
1116 return "Cannot reshape RAID10 to far-mode";
1117 if (near
* far
!= copies
)
1118 return "Cannot change number of copies"
1119 " when reshaping RAID10";
1121 if (info
->delta_disks
== UnSet
)
1122 info
->delta_disks
= 0;
1123 new_disks
= (info
->array
.raid_disks
+
1126 new_chunk
= info
->new_chunk
* far
;
1129 re
->before
.layout
= info
->array
.layout
;
1130 re
->before
.data_disks
= info
->array
.raid_disks
;
1131 re
->after
.layout
= info
->new_layout
;
1132 re
->after
.data_disks
= new_disks
;
1133 /* For RAID10 we don't do backup but do allow reshape,
1134 * so set backup_blocks to INVALID_SECTORS rather than
1136 * And there is no need to synchronise stripes on both
1137 * 'old' and 'new'. So the important
1138 * number is the minimum data_offset difference
1139 * which is the larger of (offset copies * chunk).
1141 re
->backup_blocks
= INVALID_SECTORS
;
1142 re
->min_offset_change
= max(old_chunk
, new_chunk
) / 512;
1143 if (new_disks
< re
->before
.data_disks
&&
1144 info
->space_after
< re
->min_offset_change
)
1145 /* Reduce component size by one chunk */
1146 re
->new_size
= (info
->component_size
-
1147 re
->min_offset_change
);
1149 re
->new_size
= info
->component_size
;
1150 re
->new_size
= re
->new_size
* new_disks
/ copies
;
1154 return "RAID10 can only be changed to RAID0";
1157 /* RAID0 can be converted to RAID10, or to RAID456 */
1158 if (info
->new_level
== 10) {
1159 if (info
->new_layout
== UnSet
&& info
->delta_disks
== UnSet
) {
1160 /* Assume near=2 layout */
1161 info
->new_layout
= 0x102;
1162 info
->delta_disks
= info
->array
.raid_disks
;
1164 if (info
->new_layout
== UnSet
) {
1165 int copies
= 1 + (info
->delta_disks
1166 / info
->array
.raid_disks
);
1167 if (info
->array
.raid_disks
* (copies
-1)
1168 != info
->delta_disks
)
1169 return "Impossible number of devices"
1170 " for RAID0->RAID10";
1171 info
->new_layout
= 0x100 + copies
;
1173 if (info
->delta_disks
== UnSet
) {
1174 int copies
= info
->new_layout
& 0xff;
1175 if (info
->new_layout
!= 0x100 + copies
)
1176 return "New layout impossible"
1177 " for RAID0->RAID10";;
1178 info
->delta_disks
= (copies
- 1) *
1179 info
->array
.raid_disks
;
1181 if (info
->new_chunk
&&
1182 info
->new_chunk
!= info
->array
.chunk_size
)
1183 return "Cannot change chunk-size with RAID0->RAID10";
1186 re
->before
.data_disks
= (info
->array
.raid_disks
+
1188 re
->after
.data_disks
= re
->before
.data_disks
;
1189 re
->before
.layout
= info
->new_layout
;
1193 /* RAID0 can also covert to RAID0/4/5/6 by first converting to
1194 * a raid4 style layout of the final level.
1196 switch (info
->new_level
) {
1201 re
->before
.layout
= 0;
1206 re
->before
.layout
= ALGORITHM_PARITY_N
;
1207 if (info
->new_layout
== UnSet
)
1208 info
->new_layout
= map_name(r5layout
, "default");
1213 re
->before
.layout
= ALGORITHM_PARITY_N
;
1214 if (info
->new_layout
== UnSet
)
1215 info
->new_layout
= map_name(r6layout
, "default");
1218 return "Impossible level change requested";
1220 re
->before
.data_disks
= info
->array
.raid_disks
;
1221 /* determining 'after' layout happens outside this 'switch' */
1225 info
->array
.layout
= ALGORITHM_PARITY_N
;
1227 switch (info
->new_level
) {
1231 re
->level
= info
->array
.level
;
1232 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1233 re
->before
.layout
= info
->array
.layout
;
1237 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1238 re
->before
.layout
= info
->array
.layout
;
1243 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1244 switch (info
->array
.layout
) {
1245 case ALGORITHM_LEFT_ASYMMETRIC
:
1246 re
->before
.layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
1248 case ALGORITHM_RIGHT_ASYMMETRIC
:
1249 re
->before
.layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
1251 case ALGORITHM_LEFT_SYMMETRIC
:
1252 re
->before
.layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
1254 case ALGORITHM_RIGHT_SYMMETRIC
:
1255 re
->before
.layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
1257 case ALGORITHM_PARITY_0
:
1258 re
->before
.layout
= ALGORITHM_PARITY_0_6
;
1260 case ALGORITHM_PARITY_N
:
1261 re
->before
.layout
= ALGORITHM_PARITY_N_6
;
1264 return "Cannot convert an array with this layout";
1268 if (info
->array
.raid_disks
!= 2)
1269 return "Can only convert a 2-device array to RAID1";
1270 if (info
->delta_disks
!= UnSet
&&
1271 info
->delta_disks
!= 0)
1272 return "Cannot set raid_disk when "
1273 "converting RAID5->RAID1";
1275 info
->new_chunk
= 0;
1278 return "Impossible level change requested";
1282 switch (info
->new_level
) {
1288 re
->before
.data_disks
= info
->array
.raid_disks
- 2;
1289 re
->before
.layout
= info
->array
.layout
;
1292 return "Impossible level change requested";
1297 /* If we reached here then it looks like a re-stripe is
1298 * happening. We have determined the intermediate level
1299 * and initial raid_disks/layout and stored these in 're'.
1301 * We need to deduce the final layout that can be atomically
1302 * converted to the end state.
1304 switch (info
->new_level
) {
1306 /* We can only get to RAID0 from RAID4 or RAID5
1307 * with appropriate layout and one extra device
1309 if (re
->level
!= 4 && re
->level
!= 5)
1310 return "Cannot covert to RAID0 from this level";
1312 switch (re
->level
) {
1314 re
->before
.layout
= 0;
1315 re
->after
.layout
= 0;
1318 re
->after
.layout
= ALGORITHM_PARITY_N
;
1324 /* We can only get to RAID4 from RAID5 */
1325 if (re
->level
!= 4 && re
->level
!= 5)
1326 return "Cannot convert to RAID4 from this level";
1328 switch (re
->level
) {
1330 re
->before
.layout
= 0;
1331 re
->after
.layout
= 0;
1334 re
->after
.layout
= ALGORITHM_PARITY_N
;
1340 /* We get to RAID5 from RAID5 or RAID6 */
1341 if (re
->level
!= 5 && re
->level
!= 6)
1342 return "Cannot convert to RAID5 from this level";
1344 switch (re
->level
) {
1346 if (info
->new_layout
== UnSet
)
1347 re
->after
.layout
= re
->before
.layout
;
1349 re
->after
.layout
= info
->new_layout
;
1352 if (info
->new_layout
== UnSet
)
1353 info
->new_layout
= re
->before
.layout
;
1355 /* after.layout needs to be raid6 version of new_layout */
1356 if (info
->new_layout
== ALGORITHM_PARITY_N
)
1357 re
->after
.layout
= ALGORITHM_PARITY_N
;
1360 char *ls
= map_num(r5layout
, info
->new_layout
);
1363 /* Current RAID6 layout has a RAID5
1366 strcat(strcpy(layout
, ls
), "-6");
1367 l
= map_name(r6layout
, layout
);
1369 return "Cannot find RAID6 layout"
1372 /* Current RAID6 has no equivalent.
1373 * If it is already a '-6' layout we
1374 * can leave it unchanged, else we must
1377 ls
= map_num(r6layout
, info
->new_layout
);
1379 strcmp(ls
+strlen(ls
)-2, "-6") != 0)
1380 return "Please specify new layout";
1381 l
= info
->new_layout
;
1383 re
->after
.layout
= l
;
1389 /* We must already be at level 6 */
1391 return "Impossible level change";
1392 if (info
->new_layout
== UnSet
)
1393 re
->after
.layout
= info
->array
.layout
;
1395 re
->after
.layout
= info
->new_layout
;
1398 return "Impossible level change requested";
1400 if (info
->delta_disks
== UnSet
)
1401 info
->delta_disks
= delta_parity
;
1403 re
->after
.data_disks
= (re
->before
.data_disks
1406 switch (re
->level
) {
1407 case 6: re
->parity
= 2;
1410 case 5: re
->parity
= 1;
1412 default: re
->parity
= 0;
1415 /* So we have a restripe operation, we need to calculate the number
1416 * of blocks per reshape operation.
1418 re
->new_size
= info
->component_size
* re
->before
.data_disks
;
1419 if (info
->new_chunk
== 0)
1420 info
->new_chunk
= info
->array
.chunk_size
;
1421 if (re
->after
.data_disks
== re
->before
.data_disks
&&
1422 re
->after
.layout
== re
->before
.layout
&&
1423 info
->new_chunk
== info
->array
.chunk_size
) {
1424 /* Nothing to change, can change level immediately. */
1425 re
->level
= info
->new_level
;
1426 re
->backup_blocks
= 0;
1429 if (re
->after
.data_disks
== 1 && re
->before
.data_disks
== 1) {
1430 /* chunk and layout changes make no difference */
1431 re
->level
= info
->new_level
;
1432 re
->backup_blocks
= 0;
1436 if (re
->after
.data_disks
== re
->before
.data_disks
&&
1437 get_linux_version() < 2006032)
1438 return "in-place reshape is not safe before 2.6.32 - sorry.";
1440 if (re
->after
.data_disks
< re
->before
.data_disks
&&
1441 get_linux_version() < 2006030)
1442 return "reshape to fewer devices is not supported before 2.6.30 - sorry.";
1444 re
->backup_blocks
= compute_backup_blocks(
1445 info
->new_chunk
, info
->array
.chunk_size
,
1446 re
->after
.data_disks
,
1447 re
->before
.data_disks
);
1448 re
->min_offset_change
= re
->backup_blocks
/ re
->before
.data_disks
;
1450 re
->new_size
= info
->component_size
* re
->after
.data_disks
;
1454 static int set_array_size(struct supertype
*st
, struct mdinfo
*sra
,
1457 struct mdinfo
*info
;
1461 if ((st
== NULL
) || (sra
== NULL
))
1464 if (text_version
== NULL
)
1465 text_version
= sra
->text_version
;
1466 subarray
= strchr(text_version
+1, '/')+1;
1467 info
= st
->ss
->container_content(st
, subarray
);
1469 unsigned long long current_size
= 0;
1470 unsigned long long new_size
=
1471 info
->custom_array_size
/2;
1473 if (sysfs_get_ll(sra
, NULL
, "array_size", ¤t_size
) == 0 &&
1474 new_size
> current_size
) {
1475 if (sysfs_set_num(sra
, NULL
, "array_size", new_size
)
1477 dprintf("Error: Cannot set array size");
1480 dprintf("Array size changed");
1482 dprintf(" from %llu to %llu.\n",
1483 current_size
, new_size
);
1487 dprintf("Error: set_array_size(): info pointer in NULL\n");
1492 static int reshape_array(char *container
, int fd
, char *devname
,
1493 struct supertype
*st
, struct mdinfo
*info
,
1494 int force
, struct mddev_dev
*devlist
,
1495 unsigned long long data_offset
,
1496 char *backup_file
, int verbose
, int forked
,
1497 int restart
, int freeze_reshape
);
1498 static int reshape_container(char *container
, char *devname
,
1500 struct supertype
*st
,
1501 struct mdinfo
*info
,
1504 int verbose
, int restart
, int freeze_reshape
);
1506 int Grow_reshape(char *devname
, int fd
,
1507 struct mddev_dev
*devlist
,
1508 unsigned long long data_offset
,
1509 struct context
*c
, struct shape
*s
)
1511 /* Make some changes in the shape of an array.
1512 * The kernel must support the change.
1514 * There are three different changes. Each can trigger
1515 * a resync or recovery so we freeze that until we have
1516 * requested everything (if kernel supports freezing - 2.6.30).
1518 * - change size (i.e. component_size)
1520 * - change layout/chunksize/ndisks
1522 * The last can require a reshape. It is different on different
1523 * levels so we need to check the level before actioning it.
1524 * Some times the level change needs to be requested after the
1525 * reshape (e.g. raid6->raid5, raid5->raid0)
1528 struct mdu_array_info_s array
;
1530 struct supertype
*st
;
1531 char *subarray
= NULL
;
1535 char *container
= NULL
;
1538 struct mddev_dev
*dv
;
1544 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) < 0) {
1545 pr_err("%s is not an active md array - aborting\n",
1549 if (data_offset
!= INVALID_SECTORS
&& array
.level
!= 10
1550 && (array
.level
< 4 || array
.level
> 6)) {
1551 pr_err("--grow --data-offset not yet supported\n");
1556 (s
->chunk
|| s
->level
!= UnSet
|| s
->layout_str
|| s
->raiddisks
)) {
1557 pr_err("cannot change component size at the same time "
1558 "as other changes.\n"
1559 " Change size first, then check data is intact before "
1560 "making other changes.\n");
1564 if (s
->raiddisks
&& s
->raiddisks
< array
.raid_disks
&& array
.level
> 1 &&
1565 get_linux_version() < 2006032 &&
1566 !check_env("MDADM_FORCE_FEWER")) {
1567 pr_err("reducing the number of devices is not safe before Linux 2.6.32\n"
1568 " Please use a newer kernel\n");
1572 st
= super_by_fd(fd
, &subarray
);
1574 pr_err("Unable to determine metadata format for %s\n", devname
);
1577 if (s
->raiddisks
> st
->max_devs
) {
1578 pr_err("Cannot increase raid-disks on this array"
1579 " beyond %d\n", st
->max_devs
);
1583 /* in the external case we need to check that the requested reshape is
1584 * supported, and perform an initial check that the container holds the
1585 * pre-requisite spare devices (mdmon owns final validation)
1587 if (st
->ss
->external
) {
1591 container
= st
->container_devnm
;
1592 cfd
= open_dev_excl(st
->container_devnm
);
1594 container
= st
->devnm
;
1596 cfd
= open_dev_excl(st
->devnm
);
1600 pr_err("Unable to open container for %s\n",
1606 rv
= st
->ss
->load_container(st
, cfd
, NULL
);
1609 pr_err("Cannot read superblock for %s\n",
1615 /* check if operation is supported for metadata handler */
1616 if (st
->ss
->container_content
) {
1617 struct mdinfo
*cc
= NULL
;
1618 struct mdinfo
*content
= NULL
;
1620 cc
= st
->ss
->container_content(st
, subarray
);
1621 for (content
= cc
; content
; content
= content
->next
) {
1622 int allow_reshape
= 1;
1624 /* check if reshape is allowed based on metadata
1625 * indications stored in content.array.status
1627 if (content
->array
.state
& (1<<MD_SB_BLOCK_VOLUME
))
1629 if (content
->array
.state
1630 & (1<<MD_SB_BLOCK_CONTAINER_RESHAPE
))
1632 if (!allow_reshape
) {
1633 pr_err("cannot reshape arrays in"
1634 " container with unsupported"
1635 " metadata: %s(%s)\n",
1636 devname
, container
);
1644 if (mdmon_running(container
))
1645 st
->update_tail
= &st
->updates
;
1649 for (dv
= devlist
; dv
; dv
= dv
->next
)
1651 if (s
->raiddisks
> array
.raid_disks
&&
1652 array
.spare_disks
+added_disks
< (s
->raiddisks
- array
.raid_disks
) &&
1654 pr_err("Need %d spare%s to avoid degraded array,"
1655 " and only have %d.\n"
1656 " Use --force to over-ride this check.\n",
1657 s
->raiddisks
- array
.raid_disks
,
1658 s
->raiddisks
- array
.raid_disks
== 1 ? "" : "s",
1659 array
.spare_disks
+ added_disks
);
1663 sra
= sysfs_read(fd
, NULL
, GET_LEVEL
| GET_DISKS
| GET_DEVS
1664 | GET_STATE
| GET_VERSION
);
1666 if (st
->ss
->external
&& subarray
== NULL
) {
1667 array
.level
= LEVEL_CONTAINER
;
1668 sra
->array
.level
= LEVEL_CONTAINER
;
1671 pr_err("failed to read sysfs parameters for %s\n",
1675 frozen
= freeze(st
);
1677 /* freeze() already spewed the reason */
1680 } else if (frozen
< 0) {
1681 pr_err("%s is performing resync/recovery and cannot"
1682 " be reshaped\n", devname
);
1687 /* ========= set size =============== */
1688 if (s
->size
> 0 && (s
->size
== MAX_SIZE
|| s
->size
!= (unsigned)array
.size
)) {
1689 unsigned long long orig_size
= get_component_size(fd
)/2;
1690 unsigned long long min_csize
;
1692 int raid0_takeover
= 0;
1695 orig_size
= (unsigned) array
.size
;
1697 if (orig_size
== 0) {
1698 pr_err("Cannot set device size in this type of array.\n");
1703 if (reshape_super(st
, s
->size
, UnSet
, UnSet
, 0, 0, UnSet
, NULL
,
1704 devname
, APPLY_METADATA_CHANGES
, c
->verbose
> 0)) {
1709 if (st
->ss
->external
) {
1710 /* metadata can have size limitation
1711 * update size value according to metadata information
1713 struct mdinfo
*sizeinfo
=
1714 st
->ss
->container_content(st
, subarray
);
1716 unsigned long long new_size
=
1717 sizeinfo
->custom_array_size
/2;
1718 int data_disks
= get_data_disks(
1719 sizeinfo
->array
.level
,
1720 sizeinfo
->array
.layout
,
1721 sizeinfo
->array
.raid_disks
);
1722 new_size
/= data_disks
;
1723 dprintf("Metadata size correction from %llu to "
1724 "%llu (%llu)\n", orig_size
, new_size
,
1725 new_size
* data_disks
);
1727 sysfs_free(sizeinfo
);
1731 /* Update the size of each member device in case
1732 * they have been resized. This will never reduce
1733 * below the current used-size. The "size" attribute
1734 * understands '0' to mean 'max'.
1738 for (mdi
= sra
->devs
; mdi
; mdi
= mdi
->next
) {
1739 if (sysfs_set_num(sra
, mdi
, "size",
1740 s
->size
== MAX_SIZE
? 0 : s
->size
) < 0) {
1741 /* Probably kernel refusing to let us
1742 * reduce the size - not an error.
1746 if (array
.not_persistent
== 0 &&
1747 array
.major_version
== 0 &&
1748 get_linux_version() < 3001000) {
1749 /* Dangerous to allow size to exceed 2TB */
1750 unsigned long long csize
;
1751 if (sysfs_get_ll(sra
, mdi
, "size", &csize
) == 0) {
1752 if (csize
>= 2ULL*1024*1024*1024)
1753 csize
= 2ULL*1024*1024*1024;
1754 if ((min_csize
== 0 || (min_csize
1761 pr_err("Cannot set size on "
1762 "array members.\n");
1763 goto size_change_error
;
1765 if (min_csize
&& s
->size
> min_csize
) {
1766 pr_err("Cannot safely make this array "
1767 "use more than 2TB per device on this kernel.\n");
1769 goto size_change_error
;
1771 if (min_csize
&& s
->size
== MAX_SIZE
) {
1772 /* Don't let the kernel choose a size - it will get
1775 pr_err("Limited v0.90 array to "
1776 "2TB per device\n");
1777 s
->size
= min_csize
;
1779 if (st
->ss
->external
) {
1780 if (sra
->array
.level
== 0) {
1781 rv
= sysfs_set_str(sra
, NULL
, "level",
1785 /* get array parametes after takeover
1786 * to chane one parameter at time only
1788 rv
= ioctl(fd
, GET_ARRAY_INFO
, &array
);
1791 /* make sure mdmon is
1792 * aware of the new level */
1793 if (!mdmon_running(st
->container_devnm
))
1794 start_mdmon(st
->container_devnm
);
1795 ping_monitor(container
);
1796 if (mdmon_running(st
->container_devnm
) &&
1797 st
->update_tail
== NULL
)
1798 st
->update_tail
= &st
->updates
;
1801 if (s
->size
== MAX_SIZE
)
1803 array
.size
= s
->size
;
1804 if ((unsigned)array
.size
!= s
->size
) {
1805 /* got truncated to 32bit, write to
1806 * component_size instead
1809 rv
= sysfs_set_num(sra
, NULL
,
1810 "component_size", s
->size
);
1814 rv
= ioctl(fd
, SET_ARRAY_INFO
, &array
);
1816 /* manage array size when it is managed externally
1818 if ((rv
== 0) && st
->ss
->external
)
1819 rv
= set_array_size(st
, sra
, sra
->text_version
);
1822 if (raid0_takeover
) {
1823 /* do not recync non-existing parity,
1824 * we will drop it anyway
1826 sysfs_set_str(sra
, NULL
, "sync_action", "frozen");
1827 /* go back to raid0, drop parity disk
1829 sysfs_set_str(sra
, NULL
, "level", "raid0");
1830 ioctl(fd
, GET_ARRAY_INFO
, &array
);
1837 /* restore metadata */
1838 if (reshape_super(st
, orig_size
, UnSet
, UnSet
, 0, 0,
1839 UnSet
, NULL
, devname
,
1840 ROLLBACK_METADATA_CHANGES
,
1843 pr_err("Cannot set device size for %s: %s\n",
1844 devname
, strerror(err
));
1846 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
1847 cont_err("Bitmap must be removed before size can be changed\n");
1851 if (s
->assume_clean
) {
1852 /* This will fail on kernels older than 3.0 unless
1853 * a backport has been arranged.
1856 sysfs_set_str(sra
, NULL
, "resync_start", "none") < 0)
1857 pr_err("--assume-clean not supported with --grow on this kernel\n");
1859 ioctl(fd
, GET_ARRAY_INFO
, &array
);
1860 s
->size
= get_component_size(fd
)/2;
1862 s
->size
= array
.size
;
1863 if (c
->verbose
>= 0) {
1864 if (s
->size
== orig_size
)
1865 pr_err("component size of %s "
1866 "unchanged at %lluK\n",
1869 pr_err("component size of %s "
1870 "has been set to %lluK\n",
1874 } else if (array
.level
!= LEVEL_CONTAINER
) {
1875 s
->size
= get_component_size(fd
)/2;
1877 s
->size
= array
.size
;
1880 /* See if there is anything else to do */
1881 if ((s
->level
== UnSet
|| s
->level
== array
.level
) &&
1882 (s
->layout_str
== NULL
) &&
1883 (s
->chunk
== 0 || s
->chunk
== array
.chunk_size
) &&
1884 data_offset
== INVALID_SECTORS
&&
1885 (s
->raiddisks
== 0 || s
->raiddisks
== array
.raid_disks
)) {
1886 /* Nothing more to do */
1887 if (!changed
&& c
->verbose
>= 0)
1888 pr_err("%s: no change requested\n",
1893 /* ========= check for Raid10/Raid1 -> Raid0 conversion ===============
1894 * current implementation assumes that following conditions must be met:
1897 * - near_copies == 2
1899 if ((s
->level
== 0 && array
.level
== 10 && sra
&&
1900 array
.layout
== ((1 << 8) + 2) && !(array
.raid_disks
& 1)) ||
1901 (s
->level
== 0 && array
.level
== 1 && sra
)) {
1903 err
= remove_disks_for_takeover(st
, sra
, array
.layout
);
1905 dprintf(Name
": Array cannot be reshaped\n");
1911 /* Make sure mdmon has seen the device removal
1912 * and updated metadata before we continue with
1916 ping_monitor(container
);
1919 memset(&info
, 0, sizeof(info
));
1921 sysfs_init(&info
, fd
, NULL
);
1922 strcpy(info
.text_version
, sra
->text_version
);
1923 info
.component_size
= s
->size
*2;
1924 info
.new_level
= s
->level
;
1925 info
.new_chunk
= s
->chunk
* 1024;
1926 if (info
.array
.level
== LEVEL_CONTAINER
) {
1927 info
.delta_disks
= UnSet
;
1928 info
.array
.raid_disks
= s
->raiddisks
;
1929 } else if (s
->raiddisks
)
1930 info
.delta_disks
= s
->raiddisks
- info
.array
.raid_disks
;
1932 info
.delta_disks
= UnSet
;
1933 if (s
->layout_str
== NULL
) {
1934 info
.new_layout
= UnSet
;
1935 if (info
.array
.level
== 6 &&
1936 (info
.new_level
== 6 || info
.new_level
== UnSet
) &&
1937 info
.array
.layout
>= 16) {
1938 pr_err("%s has a non-standard layout. If you"
1939 " wish to preserve this\n", devname
);
1940 cont_err("during the reshape, please specify"
1941 " --layout=preserve\n");
1942 cont_err("If you want to change it, specify a"
1943 " layout or use --layout=normalise\n");
1947 } else if (strcmp(s
->layout_str
, "normalise") == 0 ||
1948 strcmp(s
->layout_str
, "normalize") == 0) {
1949 /* If we have a -6 RAID6 layout, remove the '-6'. */
1950 info
.new_layout
= UnSet
;
1951 if (info
.array
.level
== 6 && info
.new_level
== UnSet
) {
1953 strcpy(l
, map_num(r6layout
, info
.array
.layout
));
1954 h
= strrchr(l
, '-');
1955 if (h
&& strcmp(h
, "-6") == 0) {
1957 info
.new_layout
= map_name(r6layout
, l
);
1960 pr_err("%s is only meaningful when reshaping"
1961 " a RAID6 array.\n", s
->layout_str
);
1965 } else if (strcmp(s
->layout_str
, "preserve") == 0) {
1966 /* This means that a non-standard RAID6 layout
1969 * - When reshape a RAID6 (e.g. adding a device)
1970 * which is in a non-standard layout, it is OK
1971 * to preserve that layout.
1972 * - When converting a RAID5 to RAID6, leave it in
1973 * the XXX-6 layout, don't re-layout.
1975 if (info
.array
.level
== 6 && info
.new_level
== UnSet
)
1976 info
.new_layout
= info
.array
.layout
;
1977 else if (info
.array
.level
== 5 && info
.new_level
== 6) {
1979 strcpy(l
, map_num(r5layout
, info
.array
.layout
));
1981 info
.new_layout
= map_name(r6layout
, l
);
1983 pr_err("%s in only meaningful when reshaping"
1984 " to RAID6\n", s
->layout_str
);
1989 int l
= info
.new_level
;
1991 l
= info
.array
.level
;
1994 info
.new_layout
= map_name(r5layout
, s
->layout_str
);
1997 info
.new_layout
= map_name(r6layout
, s
->layout_str
);
2000 info
.new_layout
= parse_layout_10(s
->layout_str
);
2003 info
.new_layout
= parse_layout_faulty(s
->layout_str
);
2006 pr_err("layout not meaningful"
2007 " with this level\n");
2011 if (info
.new_layout
== UnSet
) {
2012 pr_err("layout %s not understood"
2013 " for this level\n",
2020 if (array
.level
== LEVEL_FAULTY
) {
2021 if (s
->level
!= UnSet
&& s
->level
!= array
.level
) {
2022 pr_err("cannot change level of Faulty device\n");
2026 pr_err("cannot set chunksize of Faulty device\n");
2029 if (s
->raiddisks
&& s
->raiddisks
!= 1) {
2030 pr_err("cannot set raid_disks of Faulty device\n");
2033 if (s
->layout_str
) {
2034 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
2035 dprintf("Cannot get array information.\n");
2038 array
.layout
= info
.new_layout
;
2039 if (ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
2040 pr_err("failed to set new layout\n");
2042 } else if (c
->verbose
>= 0)
2043 printf("layout for %s set to %d\n",
2044 devname
, array
.layout
);
2046 } else if (array
.level
== LEVEL_CONTAINER
) {
2047 /* This change is to be applied to every array in the
2048 * container. This is only needed when the metadata imposes
2049 * restraints of the various arrays in the container.
2050 * Currently we only know that IMSM requires all arrays
2051 * to have the same number of devices so changing the
2052 * number of devices (On-Line Capacity Expansion) must be
2053 * performed at the level of the container
2055 rv
= reshape_container(container
, devname
, -1, st
, &info
,
2056 c
->force
, c
->backup_file
, c
->verbose
, 0, 0);
2059 /* get spare devices from external metadata
2061 if (st
->ss
->external
) {
2062 struct mdinfo
*info2
;
2064 info2
= st
->ss
->container_content(st
, subarray
);
2066 info
.array
.spare_disks
=
2067 info2
->array
.spare_disks
;
2072 /* Impose these changes on a single array. First
2073 * check that the metadata is OK with the change. */
2075 if (reshape_super(st
, 0, info
.new_level
,
2076 info
.new_layout
, info
.new_chunk
,
2077 info
.array
.raid_disks
, info
.delta_disks
,
2078 c
->backup_file
, devname
, APPLY_METADATA_CHANGES
,
2084 rv
= reshape_array(container
, fd
, devname
, st
, &info
, c
->force
,
2085 devlist
, data_offset
, c
->backup_file
, c
->verbose
,
2096 /* verify_reshape_position()
2097 * Function checks if reshape position in metadata is not farther
2098 * than position in md.
2100 * 0 : not valid sysfs entry
2101 * it can be caused by not started reshape, it should be started
2102 * by reshape array or raid0 array is before takeover
2103 * -1 : error, reshape position is obviously wrong
2104 * 1 : success, reshape progress correct or updated
2106 static int verify_reshape_position(struct mdinfo
*info
, int level
)
2112 /* read sync_max, failure can mean raid0 array */
2113 rv
= sysfs_get_str(info
, NULL
, "sync_max", buf
, 40);
2117 unsigned long long position
= strtoull(buf
, &ep
, 0);
2119 dprintf(Name
": Read sync_max sysfs entry is: %s\n", buf
);
2120 if (!(ep
== buf
|| (*ep
!= 0 && *ep
!= '\n' && *ep
!= ' '))) {
2121 position
*= get_data_disks(level
,
2123 info
->array
.raid_disks
);
2124 if (info
->reshape_progress
< position
) {
2125 dprintf("Corrected reshape progress (%llu) to "
2126 "md position (%llu)\n",
2127 info
->reshape_progress
, position
);
2128 info
->reshape_progress
= position
;
2130 } else if (info
->reshape_progress
> position
) {
2131 pr_err("Fatal error: array "
2132 "reshape was not properly frozen "
2133 "(expected reshape position is %llu, "
2134 "but reshape progress is %llu.\n",
2135 position
, info
->reshape_progress
);
2138 dprintf("Reshape position in md and metadata "
2143 } else if (rv
== 0) {
2144 /* for valid sysfs entry, 0-length content
2145 * should be indicated as error
2153 static unsigned long long choose_offset(unsigned long long lo
,
2154 unsigned long long hi
,
2155 unsigned long long min
,
2156 unsigned long long max
)
2158 /* Choose a new offset between hi and lo.
2159 * It must be between min and max, but
2160 * we would prefer something near the middle of hi/lo, and also
2161 * prefer to be aligned to a big power of 2.
2163 * So we start with the middle, then for each bit,
2164 * starting at '1' and increasing, if it is set, we either
2165 * add it or subtract it if possible, preferring the option
2166 * which is furthest from the boundary.
2168 * We stop once we get a 1MB alignment. As units are in sectors,
2169 * 1MB = 2*1024 sectors.
2171 unsigned long long choice
= (lo
+ hi
) / 2;
2172 unsigned long long bit
= 1;
2174 for (bit
= 1; bit
< 2*1024; bit
= bit
<< 1) {
2175 unsigned long long bigger
, smaller
;
2176 if (! (bit
& choice
))
2178 bigger
= choice
+ bit
;
2179 smaller
= choice
- bit
;
2180 if (bigger
> max
&& smaller
< min
)
2184 else if (smaller
< min
)
2186 else if (hi
- bigger
> smaller
- lo
)
2194 static int set_new_data_offset(struct mdinfo
*sra
, struct supertype
*st
,
2195 char *devname
, int delta_disks
,
2196 unsigned long long data_offset
,
2197 unsigned long long min
,
2203 unsigned long long before
, after
;
2205 /* Need to find min space before and after so same is used
2208 before
= UINT64_MAX
;
2210 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
2214 struct supertype
*st2
;
2215 struct mdinfo info2
;
2217 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
2219 dn
= map_dev(sd
->disk
.major
, sd
->disk
.minor
, 0);
2220 dfd
= dev_open(dn
, O_RDONLY
);
2222 pr_err("%s: cannot open component %s\n",
2223 devname
, dn
? dn
: "-unknown-");
2226 st2
= dup_super(st
);
2227 rv
= st2
->ss
->load_super(st2
,dfd
, NULL
);
2231 pr_err("%s: cannot get superblock from %s\n",
2235 st2
->ss
->getinfo_super(st2
, &info2
, NULL
);
2236 st2
->ss
->free_super(st2
);
2238 if (info2
.space_before
== 0 &&
2239 info2
.space_after
== 0) {
2240 /* Metadata doesn't support data_offset changes */
2243 if (before
> info2
.space_before
)
2244 before
= info2
.space_before
;
2245 if (after
> info2
.space_after
)
2246 after
= info2
.space_after
;
2248 if (data_offset
!= INVALID_SECTORS
) {
2250 if (info2
.data_offset
== data_offset
) {
2251 pr_err("%s: already has that data_offset\n",
2255 if (data_offset
< info2
.data_offset
)
2259 } else if ((data_offset
<= info2
.data_offset
&& dir
== 1) ||
2260 (data_offset
>= info2
.data_offset
&& dir
== -1)) {
2261 pr_err("%s: differing data offsets on devices make this --data-offset setting impossible\n",
2267 if (before
== UINT64_MAX
)
2268 /* impossible really, there must be no devices */
2271 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
2272 char *dn
= map_dev(sd
->disk
.major
, sd
->disk
.minor
, 0);
2273 unsigned long long new_data_offset
;
2275 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
2277 if (delta_disks
< 0) {
2278 /* Don't need any space as array is shrinking
2279 * just move data_offset up by min
2281 if (data_offset
== INVALID_SECTORS
)
2282 new_data_offset
= sd
->data_offset
+ min
;
2284 if (data_offset
< sd
->data_offset
+ min
) {
2285 pr_err("--data-offset too small for %s\n",
2289 new_data_offset
= data_offset
;
2291 } else if (delta_disks
> 0) {
2292 /* need space before */
2296 pr_err("Insufficient head-space for reshape on %s\n",
2300 if (data_offset
== INVALID_SECTORS
)
2301 new_data_offset
= sd
->data_offset
- min
;
2303 if (data_offset
> sd
->data_offset
- min
) {
2304 pr_err("--data-offset too large for %s\n",
2308 new_data_offset
= data_offset
;
2312 /* can move up or down. If 'data_offset'
2313 * was set we would have already decided,
2314 * so just choose direction with most space.
2321 sysfs_set_str(sra
, NULL
, "reshape_direction",
2322 dir
== 1 ? "backwards" : "forwards");
2324 /* Increase data offset */
2328 pr_err("Insufficient tail-space for reshape on %s\n",
2332 if (data_offset
!= INVALID_SECTORS
&&
2333 data_offset
< sd
->data_offset
+ min
) {
2334 pr_err("--data-offset too small on %s\n",
2338 if (data_offset
!= INVALID_SECTORS
)
2339 new_data_offset
= data_offset
;
2341 new_data_offset
= choose_offset(sd
->data_offset
,
2342 sd
->data_offset
+ after
,
2343 sd
->data_offset
+ min
,
2344 sd
->data_offset
+ after
);
2346 /* Decrease data offset */
2350 pr_err("insufficient head-room on %s\n",
2354 if (data_offset
!= INVALID_SECTORS
&&
2355 data_offset
< sd
->data_offset
- min
) {
2356 pr_err("--data-offset too small on %s\n",
2360 if (data_offset
!= INVALID_SECTORS
)
2361 new_data_offset
= data_offset
;
2363 new_data_offset
= choose_offset(sd
->data_offset
- before
,
2365 sd
->data_offset
- before
,
2366 sd
->data_offset
- min
);
2369 err
= sysfs_set_num(sra
, sd
, "new_offset", new_data_offset
);
2370 if (err
< 0 && errno
== E2BIG
) {
2371 /* try again after increasing data size to max */
2372 err
= sysfs_set_num(sra
, sd
, "size", 0);
2373 if (err
< 0 && errno
== EINVAL
&&
2374 !(sd
->disk
.state
& (1<<MD_DISK_SYNC
))) {
2375 /* some kernels have a bug where you cannot
2376 * use '0' on spare devices. */
2377 sysfs_set_num(sra
, sd
, "size",
2378 (sra
->component_size
+ after
)/2);
2380 err
= sysfs_set_num(sra
, sd
, "new_offset",
2384 if (errno
== E2BIG
&& data_offset
!= INVALID_SECTORS
) {
2385 pr_err("data-offset is too big for %s\n",
2389 if (sd
== sra
->devs
&&
2390 (errno
== ENOENT
|| errno
== E2BIG
))
2391 /* Early kernel, no 'new_offset' file,
2392 * or kernel doesn't like us.
2393 * For RAID5/6 this is not fatal
2396 pr_err("Cannot set new_offset for %s\n",
2405 /* Just use a backup file */
2409 static int raid10_reshape(char *container
, int fd
, char *devname
,
2410 struct supertype
*st
, struct mdinfo
*info
,
2411 struct reshape
*reshape
,
2412 unsigned long long data_offset
,
2413 int force
, int verbose
)
2415 /* Changing raid_disks, layout, chunksize or possibly
2416 * just data_offset for a RAID10.
2417 * We must always change data_offset. We change by at least
2418 * ->min_offset_change which is the largest of the old and new
2420 * If raid_disks is increasing, then data_offset must decrease
2421 * by at least this copy size.
2422 * If raid_disks is unchanged, data_offset must increase or
2423 * decrease by at least min_offset_change but preferably by much more.
2424 * We choose half of the available space.
2425 * If raid_disks is decreasing, data_offset must increase by
2426 * at least min_offset_change. To allow of this, component_size
2427 * must be decreased by the same amount.
2429 * So we calculate the required minimum and direction, possibly
2430 * reduce the component_size, then iterate through the devices
2431 * and set the new_data_offset.
2432 * If that all works, we set chunk_size, layout, raid_disks, and start
2436 unsigned long long min
;
2439 sra
= sysfs_read(fd
, NULL
,
2440 GET_COMPONENT
|GET_DEVS
|GET_OFFSET
|GET_STATE
|GET_CHUNK
2443 pr_err("%s: Cannot get array details from sysfs\n",
2447 min
= reshape
->min_offset_change
;
2449 if (info
->delta_disks
)
2450 sysfs_set_str(sra
, NULL
, "reshape_direction",
2451 info
->delta_disks
< 0 ? "backwards" : "forwards");
2452 if (info
->delta_disks
< 0 &&
2453 info
->space_after
< min
) {
2454 int rv
= sysfs_set_num(sra
, NULL
, "component_size",
2455 (sra
->component_size
-
2458 pr_err("cannot reduce component size\n");
2462 err
= set_new_data_offset(sra
, st
, devname
, info
->delta_disks
, data_offset
,
2465 pr_err("Cannot set new_data_offset: RAID10 reshape not\n");
2466 cont_err("supported on this kernel\n");
2472 if (!err
&& sysfs_set_num(sra
, NULL
, "chunk_size", info
->new_chunk
) < 0)
2474 if (!err
&& sysfs_set_num(sra
, NULL
, "layout", reshape
->after
.layout
) < 0)
2476 if (!err
&& sysfs_set_num(sra
, NULL
, "raid_disks",
2477 info
->array
.raid_disks
+ info
->delta_disks
) < 0)
2479 if (!err
&& sysfs_set_str(sra
, NULL
, "sync_action", "reshape") < 0)
2482 pr_err("Cannot set array shape for %s\n",
2485 (info
->array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
2486 cont_err(" Bitmap must be removed before"
2487 " shape can be changed\n");
2497 static void get_space_after(int fd
, struct supertype
*st
, struct mdinfo
*info
)
2499 struct mdinfo
*sra
, *sd
;
2500 /* Initialisation to silence compiler warning */
2501 unsigned long long min_space_before
= 0, min_space_after
= 0;
2504 sra
= sysfs_read(fd
, NULL
, GET_DEVS
);
2507 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
2510 struct supertype
*st2
;
2511 struct mdinfo info2
;
2513 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
2515 dn
= map_dev(sd
->disk
.major
, sd
->disk
.minor
, 0);
2516 dfd
= dev_open(dn
, O_RDONLY
);
2519 st2
= dup_super(st
);
2520 if (st2
->ss
->load_super(st2
,dfd
, NULL
)) {
2526 st2
->ss
->getinfo_super(st2
, &info2
, NULL
);
2527 st2
->ss
->free_super(st2
);
2530 min_space_before
> info2
.space_before
)
2531 min_space_before
= info2
.space_before
;
2533 min_space_after
> info2
.space_after
)
2534 min_space_after
= info2
.space_after
;
2537 if (sd
== NULL
&& !first
) {
2538 info
->space_after
= min_space_after
;
2539 info
->space_before
= min_space_before
;
2544 static void update_cache_size(char *container
, struct mdinfo
*sra
,
2545 struct mdinfo
*info
,
2546 int disks
, unsigned long long blocks
)
2548 /* Check that the internal stripe cache is
2549 * large enough, or it won't work.
2550 * It must hold at least 4 stripes of the larger
2553 unsigned long cache
;
2554 cache
= max(info
->array
.chunk_size
, info
->new_chunk
);
2555 cache
*= 4; /* 4 stripes minimum */
2556 cache
/= 512; /* convert to sectors */
2557 /* make sure there is room for 'blocks' with a bit to spare */
2558 if (cache
< 16 + blocks
/ disks
)
2559 cache
= 16 + blocks
/ disks
;
2560 cache
/= (4096/512); /* Covert from sectors to pages */
2562 if (sra
->cache_size
< cache
)
2563 subarray_set_num(container
, sra
, "stripe_cache_size",
2567 static int impose_reshape(struct mdinfo
*sra
,
2568 struct mdinfo
*info
,
2569 struct supertype
*st
,
2572 char *devname
, char *container
,
2573 struct reshape
*reshape
)
2575 struct mdu_array_info_s array
;
2577 sra
->new_chunk
= info
->new_chunk
;
2580 /* for external metadata checkpoint saved by mdmon can be lost
2581 * or missed /due to e.g. crash/. Check if md is not during
2582 * restart farther than metadata points to.
2583 * If so, this means metadata information is obsolete.
2585 if (st
->ss
->external
)
2586 verify_reshape_position(info
, reshape
->level
);
2587 sra
->reshape_progress
= info
->reshape_progress
;
2589 sra
->reshape_progress
= 0;
2590 if (reshape
->after
.data_disks
< reshape
->before
.data_disks
)
2591 /* start from the end of the new array */
2592 sra
->reshape_progress
= (sra
->component_size
2593 * reshape
->after
.data_disks
);
2596 ioctl(fd
, GET_ARRAY_INFO
, &array
);
2597 if (info
->array
.chunk_size
== info
->new_chunk
&&
2598 reshape
->before
.layout
== reshape
->after
.layout
&&
2599 st
->ss
->external
== 0) {
2600 /* use SET_ARRAY_INFO but only if reshape hasn't started */
2601 array
.raid_disks
= reshape
->after
.data_disks
+ reshape
->parity
;
2603 ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
2606 pr_err("Cannot set device shape for %s: %s\n",
2607 devname
, strerror(errno
));
2610 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
2611 cont_err("Bitmap must be removed before"
2612 " shape can be changed\n");
2616 } else if (!restart
) {
2617 /* set them all just in case some old 'new_*' value
2618 * persists from some earlier problem.
2621 if (sysfs_set_num(sra
, NULL
, "chunk_size", info
->new_chunk
) < 0)
2623 if (!err
&& sysfs_set_num(sra
, NULL
, "layout",
2624 reshape
->after
.layout
) < 0)
2626 if (!err
&& subarray_set_num(container
, sra
, "raid_disks",
2627 reshape
->after
.data_disks
+
2628 reshape
->parity
) < 0)
2631 pr_err("Cannot set device shape for %s\n",
2635 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
2636 cont_err("Bitmap must be removed before"
2637 " shape can be changed\n");
2646 static int impose_level(int fd
, int level
, char *devname
, int verbose
)
2649 struct mdu_array_info_s array
;
2651 sysfs_init(&info
, fd
, NULL
);
2653 ioctl(fd
, GET_ARRAY_INFO
, &array
);
2655 (array
.level
>= 4 && array
.level
<= 6)) {
2656 /* To convert to RAID0 we need to fail and
2657 * remove any non-data devices. */
2660 int data_disks
= array
.raid_disks
- 1;
2661 if (array
.level
== 6)
2663 if (array
.level
== 5 &&
2664 array
.layout
!= ALGORITHM_PARITY_N
)
2666 if (array
.level
== 6 &&
2667 array
.layout
!= ALGORITHM_PARITY_N_6
)
2669 sysfs_set_str(&info
, NULL
,"sync_action", "idle");
2670 /* First remove any spares so no recovery starts */
2671 for (d
= 0, found
= 0;
2672 d
< MAX_DISKS
&& found
< array
.nr_disks
;
2674 mdu_disk_info_t disk
;
2676 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0)
2678 if (disk
.major
== 0 && disk
.minor
== 0)
2681 if ((disk
.state
& (1 << MD_DISK_ACTIVE
))
2682 && disk
.raid_disk
< data_disks
)
2685 ioctl(fd
, HOT_REMOVE_DISK
,
2686 makedev(disk
.major
, disk
.minor
));
2688 /* Now fail anything left */
2689 ioctl(fd
, GET_ARRAY_INFO
, &array
);
2690 for (d
= 0, found
= 0;
2691 d
< MAX_DISKS
&& found
< array
.nr_disks
;
2694 mdu_disk_info_t disk
;
2696 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0)
2698 if (disk
.major
== 0 && disk
.minor
== 0)
2701 if ((disk
.state
& (1 << MD_DISK_ACTIVE
))
2702 && disk
.raid_disk
< data_disks
)
2705 ioctl(fd
, SET_DISK_FAULTY
,
2706 makedev(disk
.major
, disk
.minor
));
2708 while (ioctl(fd
, HOT_REMOVE_DISK
,
2709 makedev(disk
.major
, disk
.minor
)) < 0
2716 c
= map_num(pers
, level
);
2718 int err
= sysfs_set_str(&info
, NULL
, "level", c
);
2721 pr_err("%s: could not set level to %s\n",
2724 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
2725 cont_err("Bitmap must be removed"
2726 " before level can be changed\n");
2730 pr_err("level of %s changed to %s\n",
2737 static void catch_term(int sig
)
2742 static int reshape_array(char *container
, int fd
, char *devname
,
2743 struct supertype
*st
, struct mdinfo
*info
,
2744 int force
, struct mddev_dev
*devlist
,
2745 unsigned long long data_offset
,
2746 char *backup_file
, int verbose
, int forked
,
2747 int restart
, int freeze_reshape
)
2749 struct reshape reshape
;
2752 int orig_level
= UnSet
;
2756 struct mdu_array_info_s array
;
2759 struct mddev_dev
*dv
;
2763 unsigned long long *offsets
= NULL
;
2767 unsigned long blocks
;
2768 unsigned long long array_size
;
2770 struct mdinfo
*sra
= NULL
;
2772 /* when reshaping a RAID0, the component_size might be zero.
2773 * So try to fix that up.
2775 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
2776 dprintf("Cannot get array information.\n");
2779 if (array
.level
== 0 && info
->component_size
== 0) {
2780 get_dev_size(fd
, NULL
, &array_size
);
2781 info
->component_size
= array_size
/ array
.raid_disks
;
2784 if (array
.level
== 10)
2785 /* Need space_after info */
2786 get_space_after(fd
, st
, info
);
2788 if (info
->reshape_active
) {
2789 int new_level
= info
->new_level
;
2790 info
->new_level
= UnSet
;
2791 if (info
->delta_disks
> 0)
2792 info
->array
.raid_disks
-= info
->delta_disks
;
2793 msg
= analyse_change(devname
, info
, &reshape
);
2794 info
->new_level
= new_level
;
2795 if (info
->delta_disks
> 0)
2796 info
->array
.raid_disks
+= info
->delta_disks
;
2798 /* Make sure the array isn't read-only */
2799 ioctl(fd
, RESTART_ARRAY_RW
, 0);
2801 msg
= analyse_change(devname
, info
, &reshape
);
2803 /* if msg == "", error has already been printed */
2805 pr_err("%s\n", msg
);
2809 (reshape
.level
!= info
->array
.level
||
2810 reshape
.before
.layout
!= info
->array
.layout
||
2811 reshape
.before
.data_disks
+ reshape
.parity
2812 != info
->array
.raid_disks
- max(0, info
->delta_disks
))) {
2813 pr_err("reshape info is not in native format -"
2814 " cannot continue.\n");
2818 if (st
->ss
->external
&& restart
&& (info
->reshape_progress
== 0)) {
2819 /* When reshape is restarted from '0', very begin of array
2820 * it is possible that for external metadata reshape and array
2821 * configuration doesn't happen.
2822 * Check if md has the same opinion, and reshape is restarted
2823 * from 0. If so, this is regular reshape start after reshape
2824 * switch in metadata to next array only.
2826 if ((verify_reshape_position(info
, reshape
.level
) >= 0) &&
2827 (info
->reshape_progress
== 0))
2831 /* reshape already started. just skip to monitoring the reshape */
2832 if (reshape
.backup_blocks
== 0)
2834 if (restart
& RESHAPE_NO_BACKUP
)
2837 /* Need 'sra' down at 'started:' */
2838 sra
= sysfs_read(fd
, NULL
,
2839 GET_COMPONENT
|GET_DEVS
|GET_OFFSET
|GET_STATE
|GET_CHUNK
|
2842 pr_err("%s: Cannot get array details from sysfs\n",
2848 /* The container is frozen but the array may not be.
2849 * So freeze the array so spares don't get put to the wrong use
2850 * FIXME there should probably be a cleaner separation between
2851 * freeze_array and freeze_container.
2853 sysfs_freeze_array(info
);
2854 /* Check we have enough spares to not be degraded */
2856 for (dv
= devlist
; dv
; dv
=dv
->next
)
2858 spares_needed
= max(reshape
.before
.data_disks
,
2859 reshape
.after
.data_disks
)
2860 + reshape
.parity
- array
.raid_disks
;
2863 info
->new_level
> 1 && info
->array
.level
> 1 &&
2864 spares_needed
> info
->array
.spare_disks
+ added_disks
) {
2865 pr_err("Need %d spare%s to avoid degraded array,"
2866 " and only have %d.\n"
2867 " Use --force to over-ride this check.\n",
2869 spares_needed
== 1 ? "" : "s",
2870 info
->array
.spare_disks
+ added_disks
);
2873 /* Check we have enough spares to not fail */
2874 spares_needed
= max(reshape
.before
.data_disks
,
2875 reshape
.after
.data_disks
)
2877 if ((info
->new_level
> 1 || info
->new_level
== 0) &&
2878 spares_needed
> info
->array
.spare_disks
+added_disks
) {
2879 pr_err("Need %d spare%s to create working array,"
2880 " and only have %d.\n",
2882 spares_needed
== 1 ? "" : "s",
2883 info
->array
.spare_disks
+ added_disks
);
2887 if (reshape
.level
!= array
.level
) {
2888 int err
= impose_level(fd
, reshape
.level
, devname
, verbose
);
2891 info
->new_layout
= UnSet
; /* after level change,
2892 * layout is meaningless */
2893 orig_level
= array
.level
;
2894 sysfs_freeze_array(info
);
2896 if (reshape
.level
> 0 && st
->ss
->external
) {
2897 /* make sure mdmon is aware of the new level */
2898 if (mdmon_running(container
))
2899 flush_mdmon(container
);
2901 if (!mdmon_running(container
))
2902 start_mdmon(container
);
2903 ping_monitor(container
);
2904 if (mdmon_running(container
) &&
2905 st
->update_tail
== NULL
)
2906 st
->update_tail
= &st
->updates
;
2909 /* ->reshape_super might have chosen some spares from the
2910 * container that it wants to be part of the new array.
2911 * We can collect them with ->container_content and give
2912 * them to the kernel.
2914 if (st
->ss
->reshape_super
&& st
->ss
->container_content
) {
2915 char *subarray
= strchr(info
->text_version
+1, '/')+1;
2916 struct mdinfo
*info2
=
2917 st
->ss
->container_content(st
, subarray
);
2921 sysfs_init(info2
, fd
, st
->devnm
);
2922 /* When increasing number of devices, we need to set
2923 * new raid_disks before adding these, or they might
2926 if (reshape
.backup_blocks
&&
2927 reshape
.after
.data_disks
> reshape
.before
.data_disks
)
2928 subarray_set_num(container
, info2
, "raid_disks",
2929 reshape
.after
.data_disks
+
2931 for (d
= info2
->devs
; d
; d
= d
->next
) {
2932 if (d
->disk
.state
== 0 &&
2933 d
->disk
.raid_disk
>= 0) {
2934 /* This is a spare that wants to
2935 * be part of the array.
2937 add_disk(fd
, st
, info2
, d
);
2943 /* We might have been given some devices to add to the
2944 * array. Now that the array has been changed to the right
2945 * level and frozen, we can safely add them.
2948 Manage_subdevs(devname
, fd
, devlist
, verbose
,
2951 if (reshape
.backup_blocks
== 0 && data_offset
!= INVALID_SECTORS
)
2952 reshape
.backup_blocks
= reshape
.before
.data_disks
* info
->array
.chunk_size
/512;
2953 if (reshape
.backup_blocks
== 0) {
2954 /* No restriping needed, but we might need to impose
2955 * some more changes: layout, raid_disks, chunk_size
2957 /* read current array info */
2958 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
2959 dprintf("Cannot get array information.\n");
2962 /* compare current array info with new values and if
2963 * it is different update them to new */
2964 if (info
->new_layout
!= UnSet
&&
2965 info
->new_layout
!= array
.layout
) {
2966 array
.layout
= info
->new_layout
;
2967 if (ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
2968 pr_err("failed to set new layout\n");
2970 } else if (verbose
>= 0)
2971 printf("layout for %s set to %d\n",
2972 devname
, array
.layout
);
2974 if (info
->delta_disks
!= UnSet
&&
2975 info
->delta_disks
!= 0 &&
2976 array
.raid_disks
!= (info
->array
.raid_disks
+ info
->delta_disks
)) {
2977 array
.raid_disks
+= info
->delta_disks
;
2978 if (ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
2979 pr_err("failed to set raid disks\n");
2981 } else if (verbose
>= 0) {
2982 printf("raid_disks for %s set to %d\n",
2983 devname
, array
.raid_disks
);
2986 if (info
->new_chunk
!= 0 &&
2987 info
->new_chunk
!= array
.chunk_size
) {
2988 if (sysfs_set_num(info
, NULL
,
2989 "chunk_size", info
->new_chunk
) != 0) {
2990 pr_err("failed to set chunk size\n");
2992 } else if (verbose
>= 0)
2993 printf("chunk size for %s set to %d\n",
2994 devname
, array
.chunk_size
);
3001 * There are three possibilities.
3002 * 1/ The array will shrink.
3003 * We need to ensure the reshape will pause before reaching
3004 * the 'critical section'. We also need to fork and wait for
3005 * that to happen. When it does we
3006 * suspend/backup/complete/unfreeze
3008 * 2/ The array will not change size.
3009 * This requires that we keep a backup of a sliding window
3010 * so that we can restore data after a crash. So we need
3011 * to fork and monitor progress.
3012 * In future we will allow the data_offset to change, so
3013 * a sliding backup becomes unnecessary.
3015 * 3/ The array will grow. This is relatively easy.
3016 * However the kernel's restripe routines will cheerfully
3017 * overwrite some early data before it is safe. So we
3018 * need to make a backup of the early parts of the array
3019 * and be ready to restore it if rebuild aborts very early.
3020 * For externally managed metadata, we still need a forked
3021 * child to monitor the reshape and suspend IO over the region
3022 * that is being reshaped.
3024 * We backup data by writing it to one spare, or to a
3025 * file which was given on command line.
3027 * In each case, we first make sure that storage is available
3028 * for the required backup.
3030 * - request the shape change.
3031 * - fork to handle backup etc.
3033 /* Check that we can hold all the data */
3034 get_dev_size(fd
, NULL
, &array_size
);
3035 if (reshape
.new_size
< (array_size
/512)) {
3036 pr_err("this change will reduce the size of the array.\n"
3037 " use --grow --array-size first to truncate array.\n"
3038 " e.g. mdadm --grow %s --array-size %llu\n",
3039 devname
, reshape
.new_size
/2);
3043 if (array
.level
== 10) {
3044 /* Reshaping RAID10 does not require any data backup by
3045 * user-space. Instead it requires that the data_offset
3046 * is changed to avoid the need for backup.
3047 * So this is handled very separately
3050 /* Nothing to do. */
3052 return raid10_reshape(container
, fd
, devname
, st
, info
,
3053 &reshape
, data_offset
,
3056 sra
= sysfs_read(fd
, NULL
,
3057 GET_COMPONENT
|GET_DEVS
|GET_OFFSET
|GET_STATE
|GET_CHUNK
|
3060 pr_err("%s: Cannot get array details from sysfs\n",
3066 switch(set_new_data_offset(sra
, st
, devname
,
3067 reshape
.after
.data_disks
- reshape
.before
.data_disks
,
3069 reshape
.min_offset_change
, 1)) {
3073 /* Updated data_offset, so it's easy now */
3074 update_cache_size(container
, sra
, info
,
3075 min(reshape
.before
.data_disks
,
3076 reshape
.after
.data_disks
),
3077 reshape
.backup_blocks
);
3079 /* Right, everything seems fine. Let's kick things off.
3083 if (impose_reshape(sra
, info
, st
, fd
, restart
,
3084 devname
, container
, &reshape
) < 0)
3086 if (sysfs_set_str(sra
, NULL
, "sync_action", "reshape") < 0) {
3087 pr_err("Failed to initiate reshape!\n");
3090 if (info
->new_level
== reshape
.level
)
3092 /* need to adjust level when reshape completes */
3094 case -1: /* ignore error, but don't wait */
3096 default: /* parent */
3104 fd
= open_dev(sra
->sys_name
);
3106 impose_level(fd
, info
->new_level
, devname
, verbose
);
3108 case 1: /* Couldn't set data_offset, try the old way */
3109 if (data_offset
!= INVALID_SECTORS
) {
3110 pr_err("Cannot update data_offset on this array\n");
3117 /* Decide how many blocks (sectors) for a reshape
3118 * unit. The number we have so far is just a minimum
3120 blocks
= reshape
.backup_blocks
;
3121 if (reshape
.before
.data_disks
==
3122 reshape
.after
.data_disks
) {
3123 /* Make 'blocks' bigger for better throughput, but
3124 * not so big that we reject it below.
3125 * Try for 16 megabytes
3127 while (blocks
* 32 < sra
->component_size
&&
3131 pr_err("Need to backup %luK of critical "
3132 "section..\n", blocks
/2);
3134 if (blocks
>= sra
->component_size
/2) {
3135 pr_err("%s: Something wrong"
3136 " - reshape aborted\n",
3141 /* Now we need to open all these devices so we can read/write.
3143 nrdisks
= max(reshape
.before
.data_disks
,
3144 reshape
.after
.data_disks
) + reshape
.parity
3145 + sra
->array
.spare_disks
;
3146 fdlist
= xcalloc((1+nrdisks
), sizeof(int));
3147 offsets
= xcalloc((1+nrdisks
), sizeof(offsets
[0]));
3149 odisks
= reshape
.before
.data_disks
+ reshape
.parity
;
3150 d
= reshape_prepare_fdlist(devname
, sra
, odisks
,
3151 nrdisks
, blocks
, backup_file
,
3156 if ((st
->ss
->manage_reshape
== NULL
) ||
3157 (st
->ss
->recover_backup
== NULL
)) {
3158 if (backup_file
== NULL
) {
3159 if (reshape
.after
.data_disks
<=
3160 reshape
.before
.data_disks
) {
3161 pr_err("%s: Cannot grow - need backup-file\n",
3163 pr_err(" Please provide one with \"--backup=...\"\n");
3165 } else if (sra
->array
.spare_disks
== 0) {
3166 pr_err("%s: Cannot grow - "
3167 "need a spare or backup-file to backup "
3168 "critical section\n", devname
);
3172 if (!reshape_open_backup_file(backup_file
, fd
, devname
,
3174 fdlist
+d
, offsets
+d
,
3182 update_cache_size(container
, sra
, info
,
3183 min(reshape
.before
.data_disks
, reshape
.after
.data_disks
),
3186 /* Right, everything seems fine. Let's kick things off.
3187 * If only changing raid_disks, use ioctl, else use
3192 if (impose_reshape(sra
, info
, st
, fd
, restart
,
3193 devname
, container
, &reshape
) < 0)
3196 err
= start_reshape(sra
, restart
, reshape
.before
.data_disks
,
3197 reshape
.after
.data_disks
);
3199 pr_err("Cannot %s reshape for %s\n",
3200 restart
? "continue" : "start",
3205 sysfs_set_str(sra
, NULL
, "array_state", "active");
3206 if (freeze_reshape
) {
3210 pr_err("Reshape has to be continued from"
3211 " location %llu when root filesystem has been mounted.\n",
3212 sra
->reshape_progress
);
3216 /* Now we just need to kick off the reshape and watch, while
3217 * handling backups of the data...
3218 * This is all done by a forked background process.
3220 switch(forked
? 0 : fork()) {
3222 pr_err("Cannot run child to monitor reshape: %s\n",
3236 /* If another array on the same devices is busy, the
3237 * reshape will wait for them. This would mean that
3238 * the first section that we suspend will stay suspended
3239 * for a long time. So check on that possibility
3240 * by looking for "DELAYED" in /proc/mdstat, and if found,
3244 struct mdstat_ent
*mds
, *m
;
3246 mds
= mdstat_read(1, 0);
3247 for (m
= mds
; m
; m
= m
->next
)
3248 if (strcmp(m
->devnm
, sra
->sys_name
) == 0) {
3250 m
->percent
== RESYNC_DELAYED
)
3253 /* Haven't started the reshape thread
3260 if (delayed
== 1 && get_linux_version() < 3007000) {
3261 pr_err("Reshape is delayed, but cannot wait carefully with this kernel.\n"
3262 " You might experience problems until other reshapes complete.\n");
3266 mdstat_wait(30 - (delayed
-1) * 25);
3270 if (check_env("MDADM_GROW_VERIFY"))
3271 fd
= open(devname
, O_RDONLY
| O_DIRECT
);
3274 mlockall(MCL_FUTURE
);
3276 signal(SIGTERM
, catch_term
);
3278 if (st
->ss
->external
) {
3279 /* metadata handler takes it from here */
3280 done
= st
->ss
->manage_reshape(
3281 fd
, sra
, &reshape
, st
, blocks
,
3283 d
- odisks
, fdlist
+odisks
,
3286 done
= child_monitor(
3287 fd
, sra
, &reshape
, st
, blocks
,
3289 d
- odisks
, fdlist
+odisks
,
3295 if (backup_file
&& done
)
3296 unlink(backup_file
);
3302 if (!st
->ss
->external
&&
3303 !(reshape
.before
.data_disks
!= reshape
.after
.data_disks
3304 && info
->custom_array_size
) &&
3305 info
->new_level
== reshape
.level
&&
3307 /* no need to wait for the reshape to finish as
3308 * there is nothing more to do.
3315 if (st
->ss
->external
) {
3316 /* Re-load the metadata as much could have changed */
3317 int cfd
= open_dev(st
->container_devnm
);
3319 flush_mdmon(container
);
3320 st
->ss
->free_super(st
);
3321 st
->ss
->load_container(st
, cfd
, container
);
3326 /* set new array size if required customer_array_size is used
3329 if (reshape
.before
.data_disks
!=
3330 reshape
.after
.data_disks
&&
3331 info
->custom_array_size
)
3332 set_array_size(st
, info
, info
->text_version
);
3334 if (info
->new_level
!= reshape
.level
) {
3336 fd
= open(devname
, O_RDONLY
);
3337 impose_level(fd
, info
->new_level
, devname
, verbose
);
3339 if (info
->new_level
== 0)
3340 st
->update_tail
= NULL
;
3352 if (orig_level
!= UnSet
&& sra
) {
3353 c
= map_num(pers
, orig_level
);
3354 if (c
&& sysfs_set_str(sra
, NULL
, "level", c
) == 0)
3355 pr_err("aborting level change\n");
3363 /* mdfd handle is passed to be closed in child process (after fork).
3365 int reshape_container(char *container
, char *devname
,
3367 struct supertype
*st
,
3368 struct mdinfo
*info
,
3371 int verbose
, int restart
, int freeze_reshape
)
3373 struct mdinfo
*cc
= NULL
;
3375 char last_devnm
[32] = "";
3377 /* component_size is not meaningful for a container,
3378 * so pass '0' meaning 'no change'
3381 reshape_super(st
, 0, info
->new_level
,
3382 info
->new_layout
, info
->new_chunk
,
3383 info
->array
.raid_disks
, info
->delta_disks
,
3384 backup_file
, devname
, APPLY_METADATA_CHANGES
,
3392 /* ping monitor to be sure that update is on disk
3394 ping_monitor(container
);
3397 case -1: /* error */
3398 perror("Cannot fork to complete reshape\n");
3401 default: /* parent */
3402 if (!freeze_reshape
)
3403 printf(Name
": multi-array reshape continues"
3404 " in background\n");
3411 /* close unused handle in child process
3417 /* For each member array with reshape_active,
3418 * we need to perform the reshape.
3419 * We pick the first array that needs reshaping and
3420 * reshape it. reshape_array() will re-read the metadata
3421 * so the next time through a different array should be
3422 * ready for reshape.
3423 * It is possible that the 'different' array will not
3424 * be assembled yet. In that case we simple exit.
3425 * When it is assembled, the mdadm which assembles it
3426 * will take over the reshape.
3428 struct mdinfo
*content
;
3430 struct mdstat_ent
*mdstat
;
3436 cc
= st
->ss
->container_content(st
, NULL
);
3438 for (content
= cc
; content
; content
= content
->next
) {
3440 if (!content
->reshape_active
)
3443 subarray
= strchr(content
->text_version
+1, '/')+1;
3444 mdstat
= mdstat_by_subdev(subarray
, container
);
3447 if (mdstat
->active
== 0) {
3448 pr_err("Skipping inactive array %s.\n",
3450 free_mdstat(mdstat
);
3459 devid
= devnm2devid(mdstat
->devnm
);
3460 adev
= map_dev(major(devid
), minor(devid
), 0);
3462 adev
= content
->text_version
;
3464 fd
= open_dev(mdstat
->devnm
);
3466 printf(Name
": Device %s cannot be opened for reshape.",
3471 if (strcmp(last_devnm
, mdstat
->devnm
) == 0) {
3472 /* Do not allow for multiple reshape_array() calls for
3474 * It can happen when reshape_array() returns without
3475 * error, when reshape is not finished (wrong reshape
3476 * starting/continuation conditions). Mdmon doesn't
3477 * switch to next array in container and reentry
3478 * conditions for the same array occur.
3479 * This is possibly interim until the behaviour of
3480 * reshape_array is resolved().
3482 printf(Name
": Multiple reshape execution detected for "
3483 "device %s.", adev
);
3487 strcpy(last_devnm
, mdstat
->devnm
);
3489 sysfs_init(content
, fd
, mdstat
->devnm
);
3491 if (mdmon_running(container
))
3492 flush_mdmon(container
);
3494 rv
= reshape_array(container
, fd
, adev
, st
,
3495 content
, force
, NULL
, INVALID_SECTORS
,
3496 backup_file
, verbose
, 1, restart
,
3500 if (freeze_reshape
) {
3509 if (mdmon_running(container
))
3510 flush_mdmon(container
);
3519 * We run a child process in the background which performs the following
3521 * - wait for resync to reach a certain point
3522 * - suspend io to the following section
3523 * - backup that section
3524 * - allow resync to proceed further
3526 * - discard the backup.
3528 * When are combined in slightly different ways in the three cases.
3530 * - suspend/backup/allow/wait/resume/discard
3532 * - allow/wait/suspend/backup/allow/wait/resume/discard
3534 * - wait/resume/discard/suspend/backup/allow
3536 * suspend/backup/allow always come together
3537 * wait/resume/discard do too.
3538 * For the same-size case we have two backups to improve flow.
3542 int progress_reshape(struct mdinfo
*info
, struct reshape
*reshape
,
3543 unsigned long long backup_point
,
3544 unsigned long long wait_point
,
3545 unsigned long long *suspend_point
,
3546 unsigned long long *reshape_completed
, int *frozen
)
3548 /* This function is called repeatedly by the reshape manager.
3549 * It determines how much progress can safely be made and allows
3551 * - 'info' identifies the array and particularly records in
3552 * ->reshape_progress the metadata's knowledge of progress
3553 * This is a sector offset from the start of the array
3554 * of the next array block to be relocated. This number
3555 * may increase from 0 or decrease from array_size, depending
3556 * on the type of reshape that is happening.
3557 * Note that in contrast, 'sync_completed' is a block count of the
3558 * reshape so far. It gives the distance between the start point
3559 * (head or tail of device) and the next place that data will be
3560 * written. It always increases.
3561 * - 'reshape' is the structure created by analyse_change
3562 * - 'backup_point' shows how much the metadata manager has backed-up
3563 * data. For reshapes with increasing progress, it is the next address
3564 * to be backed up, previous addresses have been backed-up. For
3565 * decreasing progress, it is the earliest address that has been
3566 * backed up - later address are also backed up.
3567 * So addresses between reshape_progress and backup_point are
3568 * backed up providing those are in the 'correct' order.
3569 * - 'wait_point' is an array address. When reshape_completed
3570 * passes this point, progress_reshape should return. It might
3571 * return earlier if it determines that ->reshape_progress needs
3572 * to be updated or further backup is needed.
3573 * - suspend_point is maintained by progress_reshape and the caller
3574 * should not touch it except to initialise to zero.
3575 * It is an array address and it only increases in 2.6.37 and earlier.
3576 * This makes it difficult to handle reducing reshapes with
3577 * external metadata.
3578 * However: it is similar to backup_point in that it records the
3579 * other end of a suspended region from reshape_progress.
3580 * it is moved to extend the region that is safe to backup and/or
3582 * - reshape_completed is read from sysfs and returned. The caller
3583 * should copy this into ->reshape_progress when it has reason to
3584 * believe that the metadata knows this, and any backup outside this
3588 * 1 if more data from backup_point - but only as far as suspend_point,
3589 * should be backed up
3590 * 0 if things are progressing smoothly
3591 * -1 if the reshape is finished because it is all done,
3592 * -2 if the reshape is finished due to an error.
3595 int advancing
= (reshape
->after
.data_disks
3596 >= reshape
->before
.data_disks
);
3597 unsigned long long need_backup
; /* All data between start of array and
3598 * here will at some point need to
3601 unsigned long long read_offset
, write_offset
;
3602 unsigned long long write_range
;
3603 unsigned long long max_progress
, target
, completed
;
3604 unsigned long long array_size
= (info
->component_size
3605 * reshape
->before
.data_disks
);
3609 /* First, we unsuspend any region that is now known to be safe.
3610 * If suspend_point is on the 'wrong' side of reshape_progress, then
3611 * we don't have or need suspension at the moment. This is true for
3612 * native metadata when we don't need to back-up.
3615 if (info
->reshape_progress
<= *suspend_point
)
3616 sysfs_set_num(info
, NULL
, "suspend_lo",
3617 info
->reshape_progress
);
3619 /* Note: this won't work in 2.6.37 and before.
3620 * Something somewhere should make sure we don't need it!
3622 if (info
->reshape_progress
>= *suspend_point
)
3623 sysfs_set_num(info
, NULL
, "suspend_hi",
3624 info
->reshape_progress
);
3627 /* Now work out how far it is safe to progress.
3628 * If the read_offset for ->reshape_progress is less than
3629 * 'blocks' beyond the write_offset, we can only progress as far
3631 * Otherwise we can progress until the write_offset for the new location
3632 * reaches (within 'blocks' of) the read_offset at the current location.
3633 * However that region must be suspended unless we are using native
3635 * If we need to suspend more, we limit it to 128M per device, which is
3636 * rather arbitrary and should be some time-based calculation.
3638 read_offset
= info
->reshape_progress
/ reshape
->before
.data_disks
;
3639 write_offset
= info
->reshape_progress
/ reshape
->after
.data_disks
;
3640 write_range
= info
->new_chunk
/512;
3641 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
)
3642 need_backup
= array_size
;
3644 need_backup
= reshape
->backup_blocks
;
3646 if (read_offset
< write_offset
+ write_range
)
3647 max_progress
= backup_point
;
3651 reshape
->after
.data_disks
;
3653 if (read_offset
> write_offset
- write_range
)
3654 /* Can only progress as far as has been backed up,
3655 * which must be suspended */
3656 max_progress
= backup_point
;
3657 else if (info
->reshape_progress
<= need_backup
)
3658 max_progress
= backup_point
;
3660 if (info
->array
.major_version
>= 0)
3661 /* Can progress until backup is needed */
3662 max_progress
= need_backup
;
3664 /* Can progress until metadata update is required */
3667 reshape
->after
.data_disks
;
3668 /* but data must be suspended */
3669 if (max_progress
< *suspend_point
)
3670 max_progress
= *suspend_point
;
3675 /* We know it is safe to progress to 'max_progress' providing
3676 * it is suspended or we are using native metadata.
3677 * Consider extending suspend_point 128M per device if it
3678 * is less than 64M per device beyond reshape_progress.
3679 * But always do a multiple of 'blocks'
3680 * FIXME this is too big - it takes to long to complete
3683 target
= 64*1024*2 * min(reshape
->before
.data_disks
,
3684 reshape
->after
.data_disks
);
3685 target
/= reshape
->backup_blocks
;
3688 target
*= reshape
->backup_blocks
;
3690 /* For externally managed metadata we always need to suspend IO to
3691 * the area being reshaped so we regularly push suspend_point forward.
3692 * For native metadata we only need the suspend if we are going to do
3696 if ((need_backup
> info
->reshape_progress
3697 || info
->array
.major_version
< 0) &&
3698 *suspend_point
< info
->reshape_progress
+ target
) {
3699 if (need_backup
< *suspend_point
+ 2 * target
)
3700 *suspend_point
= need_backup
;
3701 else if (*suspend_point
+ 2 * target
< array_size
)
3702 *suspend_point
+= 2 * target
;
3704 *suspend_point
= array_size
;
3705 sysfs_set_num(info
, NULL
, "suspend_hi", *suspend_point
);
3706 if (max_progress
> *suspend_point
)
3707 max_progress
= *suspend_point
;
3710 if (info
->array
.major_version
>= 0) {
3711 /* Only need to suspend when about to backup */
3712 if (info
->reshape_progress
< need_backup
* 2 &&
3713 *suspend_point
> 0) {
3715 sysfs_set_num(info
, NULL
, "suspend_lo", 0);
3716 sysfs_set_num(info
, NULL
, "suspend_hi", need_backup
);
3719 /* Need to suspend continually */
3720 if (info
->reshape_progress
< *suspend_point
)
3721 *suspend_point
= info
->reshape_progress
;
3722 if (*suspend_point
+ target
< info
->reshape_progress
)
3723 /* No need to move suspend region yet */;
3725 if (*suspend_point
>= 2 * target
)
3726 *suspend_point
-= 2 * target
;
3729 sysfs_set_num(info
, NULL
, "suspend_lo",
3732 if (max_progress
< *suspend_point
)
3733 max_progress
= *suspend_point
;
3737 /* now set sync_max to allow that progress. sync_max, like
3738 * sync_completed is a count of sectors written per device, so
3739 * we find the difference between max_progress and the start point,
3740 * and divide that by after.data_disks to get a sync_max
3742 * At the same time we convert wait_point to a similar number
3743 * for comparing against sync_completed.
3745 /* scale down max_progress to per_disk */
3746 max_progress
/= reshape
->after
.data_disks
;
3747 /* Round to chunk size as some kernels give an erroneously high number */
3748 max_progress
/= info
->new_chunk
/512;
3749 max_progress
*= info
->new_chunk
/512;
3750 /* And round to old chunk size as the kernel wants that */
3751 max_progress
/= info
->array
.chunk_size
/512;
3752 max_progress
*= info
->array
.chunk_size
/512;
3753 /* Limit progress to the whole device */
3754 if (max_progress
> info
->component_size
)
3755 max_progress
= info
->component_size
;
3756 wait_point
/= reshape
->after
.data_disks
;
3758 /* switch from 'device offset' to 'processed block count' */
3759 max_progress
= info
->component_size
- max_progress
;
3760 wait_point
= info
->component_size
- wait_point
;
3764 sysfs_set_num(info
, NULL
, "sync_max", max_progress
);
3766 /* Now wait. If we have already reached the point that we were
3767 * asked to wait to, don't wait at all, else wait for any change.
3768 * We need to select on 'sync_completed' as that is the place that
3769 * notifications happen, but we are really interested in
3770 * 'reshape_position'
3772 fd
= sysfs_get_fd(info
, NULL
, "sync_completed");
3774 goto check_progress
;
3776 if (sysfs_fd_get_ll(fd
, &completed
) < 0)
3777 goto check_progress
;
3779 while (completed
< max_progress
&& completed
< wait_point
) {
3780 /* Check that sync_action is still 'reshape' to avoid
3781 * waiting forever on a dead array
3784 if (sysfs_get_str(info
, NULL
, "sync_action",
3786 strncmp(action
, "reshape", 7) != 0)
3788 /* Some kernels reset 'sync_completed' to zero
3789 * before setting 'sync_action' to 'idle'.
3790 * So we need these extra tests.
3792 if (completed
== 0 && advancing
3793 && info
->reshape_progress
> 0)
3795 if (completed
== 0 && !advancing
3796 && info
->reshape_progress
< (info
->component_size
3797 * reshape
->after
.data_disks
))
3799 sysfs_wait(fd
, NULL
);
3800 if (sysfs_fd_get_ll(fd
, &completed
) < 0)
3801 goto check_progress
;
3803 /* Some kernels reset 'sync_completed' to zero,
3804 * we need to have real point we are in md
3807 completed
= max_progress
;
3809 /* some kernels can give an incorrectly high 'completed' number */
3810 completed
/= (info
->new_chunk
/512);
3811 completed
*= (info
->new_chunk
/512);
3812 /* Convert 'completed' back in to a 'progress' number */
3813 completed
*= reshape
->after
.data_disks
;
3815 completed
= info
->component_size
* reshape
->after
.data_disks
3818 *reshape_completed
= completed
;
3822 /* We return the need_backup flag. Caller will decide
3823 * how much - a multiple of ->backup_blocks up to *suspend_point
3826 return need_backup
> info
->reshape_progress
;
3828 return need_backup
>= info
->reshape_progress
;
3831 /* if we couldn't read a number from sync_completed, then
3832 * either the reshape did complete, or it aborted.
3833 * We can tell which by checking for 'none' in reshape_position.
3834 * If it did abort, then it might immediately restart if it
3835 * it was just a device failure that leaves us degraded but
3839 if (sysfs_get_str(info
, NULL
, "reshape_position", buf
, sizeof(buf
)) < 0
3840 || strncmp(buf
, "none", 4) != 0) {
3841 /* The abort might only be temporary. Wait up to 10
3842 * seconds for fd to contain a valid number again.
3846 unsigned long long new_sync_max
;
3847 while (fd
>= 0 && rv
< 0 && wait
> 0) {
3848 if (sysfs_wait(fd
, &wait
) != 1)
3850 switch (sysfs_fd_get_ll(fd
, &completed
)) {
3852 /* all good again */
3854 /* If "sync_max" is no longer max_progress
3855 * we need to freeze things
3857 sysfs_get_ll(info
, NULL
, "sync_max", &new_sync_max
);
3858 *frozen
= (new_sync_max
!= max_progress
);
3860 case -2: /* read error - abort */
3867 return rv
; /* abort */
3869 /* Maybe racing with array shutdown - check state */
3872 if (sysfs_get_str(info
, NULL
, "array_state", buf
, sizeof(buf
)) < 0
3873 || strncmp(buf
, "inactive", 8) == 0
3874 || strncmp(buf
, "clear",5) == 0)
3875 return -2; /* abort */
3876 return -1; /* complete */
3880 /* FIXME return status is never checked */
3881 static int grow_backup(struct mdinfo
*sra
,
3882 unsigned long long offset
, /* per device */
3883 unsigned long stripes
, /* per device, in old chunks */
3884 int *sources
, unsigned long long *offsets
,
3885 int disks
, int chunk
, int level
, int layout
,
3886 int dests
, int *destfd
, unsigned long long *destoffsets
,
3887 int part
, int *degraded
,
3890 /* Backup 'blocks' sectors at 'offset' on each device of the array,
3891 * to storage 'destfd' (offset 'destoffsets'), after first
3892 * suspending IO. Then allow resync to continue
3893 * over the suspended section.
3894 * Use part 'part' of the backup-super-block.
3899 unsigned long long ll
;
3901 //printf("offset %llu\n", offset);
3907 /* Check that array hasn't become degraded, else we might backup the wrong data */
3908 if (sysfs_get_ll(sra
, NULL
, "degraded", &ll
) < 0)
3909 return -1; /* FIXME this error is ignored */
3910 new_degraded
= (int)ll
;
3911 if (new_degraded
!= *degraded
) {
3912 /* check each device to ensure it is still working */
3914 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
3915 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
3917 if (sd
->disk
.state
& (1<<MD_DISK_SYNC
)) {
3919 if (sysfs_get_str(sra
, sd
, "state", sbuf
, 20) < 0 ||
3920 strstr(sbuf
, "faulty") ||
3921 strstr(sbuf
, "in_sync") == NULL
) {
3922 /* this device is dead */
3923 sd
->disk
.state
= (1<<MD_DISK_FAULTY
);
3924 if (sd
->disk
.raid_disk
>= 0 &&
3925 sources
[sd
->disk
.raid_disk
] >= 0) {
3926 close(sources
[sd
->disk
.raid_disk
]);
3927 sources
[sd
->disk
.raid_disk
] = -1;
3932 *degraded
= new_degraded
;
3935 bsb
.arraystart2
= __cpu_to_le64(offset
* odata
);
3936 bsb
.length2
= __cpu_to_le64(stripes
* (chunk
/512) * odata
);
3938 bsb
.arraystart
= __cpu_to_le64(offset
* odata
);
3939 bsb
.length
= __cpu_to_le64(stripes
* (chunk
/512) * odata
);
3942 bsb
.magic
[15] = '2';
3943 for (i
= 0; i
< dests
; i
++)
3945 lseek64(destfd
[i
], destoffsets
[i
] + __le64_to_cpu(bsb
.devstart2
)*512, 0);
3947 lseek64(destfd
[i
], destoffsets
[i
], 0);
3949 rv
= save_stripes(sources
, offsets
,
3950 disks
, chunk
, level
, layout
,
3952 offset
*512*odata
, stripes
* chunk
* odata
,
3957 bsb
.mtime
= __cpu_to_le64(time(0));
3958 for (i
= 0; i
< dests
; i
++) {
3959 bsb
.devstart
= __cpu_to_le64(destoffsets
[i
]/512);
3961 bsb
.sb_csum
= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
));
3962 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0)
3963 bsb
.sb_csum2
= bsb_csum((char*)&bsb
,
3964 ((char*)&bsb
.sb_csum2
)-((char*)&bsb
));
3967 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
] - 4096, 0)
3968 != destoffsets
[i
] - 4096)
3970 if (write(destfd
[i
], &bsb
, 512) != 512)
3972 if (destoffsets
[i
] > 4096) {
3973 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
]+stripes
*chunk
*odata
, 0) !=
3974 destoffsets
[i
]+stripes
*chunk
*odata
)
3976 if (write(destfd
[i
], &bsb
, 512) != 512)
3986 /* in 2.6.30, the value reported by sync_completed can be
3987 * less that it should be by one stripe.
3988 * This only happens when reshape hits sync_max and pauses.
3989 * So allow wait_backup to either extent sync_max further
3990 * than strictly necessary, or return before the
3991 * sync has got quite as far as we would really like.
3992 * This is what 'blocks2' is for.
3993 * The various caller give appropriate values so that
3996 /* FIXME return value is often ignored */
3997 static int forget_backup(int dests
, int *destfd
,
3998 unsigned long long *destoffsets
,
4002 * Erase backup 'part' (which is 0 or 1)
4008 bsb
.arraystart2
= __cpu_to_le64(0);
4009 bsb
.length2
= __cpu_to_le64(0);
4011 bsb
.arraystart
= __cpu_to_le64(0);
4012 bsb
.length
= __cpu_to_le64(0);
4014 bsb
.mtime
= __cpu_to_le64(time(0));
4016 for (i
= 0; i
< dests
; i
++) {
4017 bsb
.devstart
= __cpu_to_le64(destoffsets
[i
]/512);
4018 bsb
.sb_csum
= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
));
4019 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0)
4020 bsb
.sb_csum2
= bsb_csum((char*)&bsb
,
4021 ((char*)&bsb
.sb_csum2
)-((char*)&bsb
));
4022 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
]-4096, 0) !=
4023 destoffsets
[i
]-4096)
4026 write(destfd
[i
], &bsb
, 512) != 512)
4033 static void fail(char *msg
)
4036 rv
= (write(2, msg
, strlen(msg
)) != (int)strlen(msg
));
4037 rv
|= (write(2, "\n", 1) != 1);
4041 static char *abuf
, *bbuf
;
4042 static unsigned long long abuflen
;
4043 static void validate(int afd
, int bfd
, unsigned long long offset
)
4045 /* check that the data in the backup against the array.
4046 * This is only used for regression testing and should not
4047 * be used while the array is active
4051 lseek64(bfd
, offset
- 4096, 0);
4052 if (read(bfd
, &bsb2
, 512) != 512)
4053 fail("cannot read bsb");
4054 if (bsb2
.sb_csum
!= bsb_csum((char*)&bsb2
,
4055 ((char*)&bsb2
.sb_csum
)-((char*)&bsb2
)))
4056 fail("first csum bad");
4057 if (memcmp(bsb2
.magic
, "md_backup_data", 14) != 0)
4058 fail("magic is bad");
4059 if (memcmp(bsb2
.magic
, "md_backup_data-2", 16) == 0 &&
4060 bsb2
.sb_csum2
!= bsb_csum((char*)&bsb2
,
4061 ((char*)&bsb2
.sb_csum2
)-((char*)&bsb2
)))
4062 fail("second csum bad");
4064 if (__le64_to_cpu(bsb2
.devstart
)*512 != offset
)
4065 fail("devstart is wrong");
4068 unsigned long long len
= __le64_to_cpu(bsb2
.length
)*512;
4070 if (abuflen
< len
) {
4074 if (posix_memalign((void**)&abuf
, 4096, abuflen
) ||
4075 posix_memalign((void**)&bbuf
, 4096, abuflen
)) {
4077 /* just stop validating on mem-alloc failure */
4082 lseek64(bfd
, offset
, 0);
4083 if ((unsigned long long)read(bfd
, bbuf
, len
) != len
) {
4084 //printf("len %llu\n", len);
4085 fail("read first backup failed");
4087 lseek64(afd
, __le64_to_cpu(bsb2
.arraystart
)*512, 0);
4088 if ((unsigned long long)read(afd
, abuf
, len
) != len
)
4089 fail("read first from array failed");
4090 if (memcmp(bbuf
, abuf
, len
) != 0) {
4093 printf("offset=%llu len=%llu\n",
4094 (unsigned long long)__le64_to_cpu(bsb2
.arraystart
)*512, len
);
4095 for (i
=0; i
<len
; i
++)
4096 if (bbuf
[i
] != abuf
[i
]) {
4097 printf("first diff byte %d\n", i
);
4101 fail("data1 compare failed");
4105 unsigned long long len
= __le64_to_cpu(bsb2
.length2
)*512;
4107 if (abuflen
< len
) {
4111 abuf
= xmalloc(abuflen
);
4112 bbuf
= xmalloc(abuflen
);
4115 lseek64(bfd
, offset
+__le64_to_cpu(bsb2
.devstart2
)*512, 0);
4116 if ((unsigned long long)read(bfd
, bbuf
, len
) != len
)
4117 fail("read second backup failed");
4118 lseek64(afd
, __le64_to_cpu(bsb2
.arraystart2
)*512, 0);
4119 if ((unsigned long long)read(afd
, abuf
, len
) != len
)
4120 fail("read second from array failed");
4121 if (memcmp(bbuf
, abuf
, len
) != 0)
4122 fail("data2 compare failed");
4126 int child_monitor(int afd
, struct mdinfo
*sra
, struct reshape
*reshape
,
4127 struct supertype
*st
, unsigned long blocks
,
4128 int *fds
, unsigned long long *offsets
,
4129 int dests
, int *destfd
, unsigned long long *destoffsets
)
4131 /* Monitor a reshape where backup is being performed using
4132 * 'native' mechanism - either to a backup file, or
4133 * to some space in a spare.
4137 unsigned long long speed
;
4138 unsigned long long suspend_point
, array_size
;
4139 unsigned long long backup_point
, wait_point
;
4140 unsigned long long reshape_completed
;
4142 int increasing
= reshape
->after
.data_disks
>= reshape
->before
.data_disks
;
4143 int part
= 0; /* The next part of the backup area to fill. It may already
4144 * be full, so we need to check */
4145 int level
= reshape
->level
;
4146 int layout
= reshape
->before
.layout
;
4147 int data
= reshape
->before
.data_disks
;
4148 int disks
= reshape
->before
.data_disks
+ reshape
->parity
;
4149 int chunk
= sra
->array
.chunk_size
;
4151 unsigned long stripes
;
4155 /* set up the backup-super-block. This requires the
4156 * uuid from the array.
4158 /* Find a superblock */
4159 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
4163 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
4165 dn
= map_dev(sd
->disk
.major
, sd
->disk
.minor
, 1);
4166 devfd
= dev_open(dn
, O_RDONLY
);
4169 ok
= st
->ss
->load_super(st
, devfd
, NULL
);
4175 pr_err("Cannot find a superblock\n");
4179 memset(&bsb
, 0, 512);
4180 memcpy(bsb
.magic
, "md_backup_data-1", 16);
4181 st
->ss
->uuid_from_super(st
, uuid
);
4182 memcpy(bsb
.set_uuid
, uuid
, 16);
4183 bsb
.mtime
= __cpu_to_le64(time(0));
4184 bsb
.devstart2
= blocks
;
4186 stripes
= blocks
/ (sra
->array
.chunk_size
/512) /
4187 reshape
->before
.data_disks
;
4189 if (posix_memalign((void**)&buf
, 4096, disks
* chunk
))
4190 /* Don't start the 'reshape' */
4192 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
) {
4193 sysfs_get_ll(sra
, NULL
, "sync_speed_min", &speed
);
4194 sysfs_set_num(sra
, NULL
, "sync_speed_min", 200000);
4198 array_size
= sra
->component_size
* reshape
->after
.data_disks
;
4199 backup_point
= sra
->reshape_progress
;
4202 array_size
= sra
->component_size
* reshape
->before
.data_disks
;
4203 backup_point
= reshape
->backup_blocks
;
4204 suspend_point
= array_size
;
4210 /* Want to return as soon the oldest backup slot can
4211 * be released as that allows us to start backing up
4212 * some more, providing suspend_point has been
4213 * advanced, which it should have.
4216 wait_point
= array_size
;
4217 if (part
== 0 && __le64_to_cpu(bsb
.length
) > 0)
4218 wait_point
= (__le64_to_cpu(bsb
.arraystart
) +
4219 __le64_to_cpu(bsb
.length
));
4220 if (part
== 1 && __le64_to_cpu(bsb
.length2
) > 0)
4221 wait_point
= (__le64_to_cpu(bsb
.arraystart2
) +
4222 __le64_to_cpu(bsb
.length2
));
4225 if (part
== 0 && __le64_to_cpu(bsb
.length
) > 0)
4226 wait_point
= __le64_to_cpu(bsb
.arraystart
);
4227 if (part
== 1 && __le64_to_cpu(bsb
.length2
) > 0)
4228 wait_point
= __le64_to_cpu(bsb
.arraystart2
);
4231 reshape_completed
= sra
->reshape_progress
;
4232 rv
= progress_reshape(sra
, reshape
,
4233 backup_point
, wait_point
,
4234 &suspend_point
, &reshape_completed
,
4236 /* external metadata would need to ping_monitor here */
4237 sra
->reshape_progress
= reshape_completed
;
4239 /* Clear any backup region that is before 'here' */
4241 if (__le64_to_cpu(bsb
.length
) > 0 &&
4242 reshape_completed
>= (__le64_to_cpu(bsb
.arraystart
) +
4243 __le64_to_cpu(bsb
.length
)))
4244 forget_backup(dests
, destfd
,
4246 if (__le64_to_cpu(bsb
.length2
) > 0 &&
4247 reshape_completed
>= (__le64_to_cpu(bsb
.arraystart2
) +
4248 __le64_to_cpu(bsb
.length2
)))
4249 forget_backup(dests
, destfd
,
4252 if (__le64_to_cpu(bsb
.length
) > 0 &&
4253 reshape_completed
<= (__le64_to_cpu(bsb
.arraystart
)))
4254 forget_backup(dests
, destfd
,
4256 if (__le64_to_cpu(bsb
.length2
) > 0 &&
4257 reshape_completed
<= (__le64_to_cpu(bsb
.arraystart2
)))
4258 forget_backup(dests
, destfd
,
4268 if (rv
== 0 && increasing
&& !st
->ss
->external
) {
4269 /* No longer need to monitor this reshape */
4270 sysfs_set_str(sra
, NULL
, "sync_max", "max");
4276 unsigned long long offset
;
4277 unsigned long actual_stripes
;
4278 /* Need to backup some data.
4279 * If 'part' is not used and the desired
4280 * backup size is suspended, do a backup,
4281 * then consider the next part.
4283 /* Check that 'part' is unused */
4284 if (part
== 0 && __le64_to_cpu(bsb
.length
) != 0)
4286 if (part
== 1 && __le64_to_cpu(bsb
.length2
) != 0)
4289 offset
= backup_point
/ data
;
4290 actual_stripes
= stripes
;
4292 if (offset
+ actual_stripes
* (chunk
/512) >
4293 sra
->component_size
)
4294 actual_stripes
= ((sra
->component_size
- offset
)
4296 if (offset
+ actual_stripes
* (chunk
/512) >
4300 if (offset
< actual_stripes
* (chunk
/512))
4301 actual_stripes
= offset
/ (chunk
/512);
4302 offset
-= actual_stripes
* (chunk
/512);
4303 if (offset
< suspend_point
/data
)
4306 if (actual_stripes
== 0)
4308 grow_backup(sra
, offset
, actual_stripes
,
4310 disks
, chunk
, level
, layout
,
4311 dests
, destfd
, destoffsets
,
4312 part
, °raded
, buf
);
4313 validate(afd
, destfd
[0], destoffsets
[0]);
4314 /* record where 'part' is up to */
4317 backup_point
+= actual_stripes
* (chunk
/512) * data
;
4319 backup_point
-= actual_stripes
* (chunk
/512) * data
;
4323 /* FIXME maybe call progress_reshape one more time instead */
4324 /* remove any remaining suspension */
4325 sysfs_set_num(sra
, NULL
, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL
);
4326 sysfs_set_num(sra
, NULL
, "suspend_hi", 0);
4327 sysfs_set_num(sra
, NULL
, "suspend_lo", 0);
4328 sysfs_set_num(sra
, NULL
, "sync_min", 0);
4330 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
)
4331 sysfs_set_num(sra
, NULL
, "sync_speed_min", speed
);
4337 * If any spare contains md_back_data-1 which is recent wrt mtime,
4338 * write that data into the array and update the super blocks with
4339 * the new reshape_progress
4341 int Grow_restart(struct supertype
*st
, struct mdinfo
*info
, int *fdlist
, int cnt
,
4342 char *backup_file
, int verbose
)
4346 unsigned long long *offsets
;
4347 unsigned long long nstripe
, ostripe
;
4350 odata
= info
->array
.raid_disks
- info
->delta_disks
- 1;
4351 if (info
->array
.level
== 6) odata
--; /* number of data disks */
4352 ndata
= info
->array
.raid_disks
- 1;
4353 if (info
->new_level
== 6) ndata
--;
4355 old_disks
= info
->array
.raid_disks
- info
->delta_disks
;
4357 if (info
->delta_disks
<= 0)
4358 /* Didn't grow, so the backup file must have
4362 for (i
=old_disks
-(backup_file
?1:0); i
<cnt
; i
++) {
4363 struct mdinfo dinfo
;
4366 char *devname
, namebuf
[20];
4367 unsigned long long lo
, hi
;
4369 /* This was a spare and may have some saved data on it.
4370 * Load the superblock, find and load the
4371 * backup_super_block.
4372 * If either fail, go on to next device.
4373 * If the backup contains no new info, just return
4374 * else restore data and update all superblocks
4376 if (i
== old_disks
-1) {
4377 fd
= open(backup_file
, O_RDONLY
);
4379 pr_err("backup file %s inaccessible: %s\n",
4380 backup_file
, strerror(errno
));
4383 devname
= backup_file
;
4388 if (st
->ss
->load_super(st
, fd
, NULL
))
4391 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
4392 st
->ss
->free_super(st
);
4395 (dinfo
.data_offset
+ dinfo
.component_size
- 8) <<9,
4397 pr_err("Cannot seek on device %d\n", i
);
4398 continue; /* Cannot seek */
4400 sprintf(namebuf
, "device-%d", i
);
4403 if (read(fd
, &bsb
, sizeof(bsb
)) != sizeof(bsb
)) {
4405 pr_err("Cannot read from %s\n", devname
);
4406 continue; /* Cannot read */
4408 if (memcmp(bsb
.magic
, "md_backup_data-1", 16) != 0 &&
4409 memcmp(bsb
.magic
, "md_backup_data-2", 16) != 0) {
4411 pr_err("No backup metadata on %s\n", devname
);
4414 if (bsb
.sb_csum
!= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
))) {
4416 pr_err("Bad backup-metadata checksum on %s\n", devname
);
4417 continue; /* bad checksum */
4419 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0 &&
4420 bsb
.sb_csum2
!= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum2
)-((char*)&bsb
))) {
4422 pr_err("Bad backup-metadata checksum2 on %s\n", devname
);
4423 continue; /* Bad second checksum */
4425 if (memcmp(bsb
.set_uuid
,info
->uuid
, 16) != 0) {
4427 pr_err("Wrong uuid on backup-metadata on %s\n", devname
);
4428 continue; /* Wrong uuid */
4431 /* array utime and backup-mtime should be updated at much the same time, but it seems that
4432 * sometimes they aren't... So allow considerable flexability in matching, and allow
4433 * this test to be overridden by an environment variable.
4435 if (info
->array
.utime
> (int)__le64_to_cpu(bsb
.mtime
) + 2*60*60 ||
4436 info
->array
.utime
< (int)__le64_to_cpu(bsb
.mtime
) - 10*60) {
4437 if (check_env("MDADM_GROW_ALLOW_OLD")) {
4438 pr_err("accepting backup with timestamp %lu "
4439 "for array with timestamp %lu\n",
4440 (unsigned long)__le64_to_cpu(bsb
.mtime
),
4441 (unsigned long)info
->array
.utime
);
4443 pr_err("too-old timestamp on backup-metadata on %s\n", devname
);
4444 pr_err("If you think it is should be safe, try 'export MDADM_GROW_ALLOW_OLD=1'\n");
4445 continue; /* time stamp is too bad */
4449 if (bsb
.magic
[15] == '1') {
4450 if (bsb
.length
== 0)
4452 if (info
->delta_disks
>= 0) {
4453 /* reshape_progress is increasing */
4454 if (__le64_to_cpu(bsb
.arraystart
)
4455 + __le64_to_cpu(bsb
.length
)
4456 < info
->reshape_progress
) {
4459 pr_err("backup-metadata found on %s but is not needed\n", devname
);
4460 continue; /* No new data here */
4463 /* reshape_progress is decreasing */
4464 if (__le64_to_cpu(bsb
.arraystart
) >=
4465 info
->reshape_progress
)
4466 goto nonew
; /* No new data here */
4469 if (bsb
.length
== 0 && bsb
.length2
== 0)
4471 if (info
->delta_disks
>= 0) {
4472 /* reshape_progress is increasing */
4473 if ((__le64_to_cpu(bsb
.arraystart
)
4474 + __le64_to_cpu(bsb
.length
)
4475 < info
->reshape_progress
)
4477 (__le64_to_cpu(bsb
.arraystart2
)
4478 + __le64_to_cpu(bsb
.length2
)
4479 < info
->reshape_progress
))
4480 goto nonew
; /* No new data here */
4482 /* reshape_progress is decreasing */
4483 if (__le64_to_cpu(bsb
.arraystart
) >=
4484 info
->reshape_progress
&&
4485 __le64_to_cpu(bsb
.arraystart2
) >=
4486 info
->reshape_progress
)
4487 goto nonew
; /* No new data here */
4490 if (lseek64(fd
, __le64_to_cpu(bsb
.devstart
)*512, 0)< 0) {
4493 pr_err("Failed to verify secondary backup-metadata block on %s\n",
4495 continue; /* Cannot seek */
4497 /* There should be a duplicate backup superblock 4k before here */
4498 if (lseek64(fd
, -4096, 1) < 0 ||
4499 read(fd
, &bsb2
, sizeof(bsb2
)) != sizeof(bsb2
))
4500 goto second_fail
; /* Cannot find leading superblock */
4501 if (bsb
.magic
[15] == '1')
4502 bsbsize
= offsetof(struct mdp_backup_super
, pad1
);
4504 bsbsize
= offsetof(struct mdp_backup_super
, pad
);
4505 if (memcmp(&bsb2
, &bsb
, bsbsize
) != 0)
4506 goto second_fail
; /* Cannot find leading superblock */
4508 /* Now need the data offsets for all devices. */
4509 offsets
= xmalloc(sizeof(*offsets
)*info
->array
.raid_disks
);
4510 for(j
=0; j
<info
->array
.raid_disks
; j
++) {
4513 if (st
->ss
->load_super(st
, fdlist
[j
], NULL
))
4514 /* FIXME should be this be an error */
4516 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
4517 st
->ss
->free_super(st
);
4518 offsets
[j
] = dinfo
.data_offset
* 512;
4520 printf(Name
": restoring critical section\n");
4522 if (restore_stripes(fdlist
, offsets
,
4523 info
->array
.raid_disks
,
4527 fd
, __le64_to_cpu(bsb
.devstart
)*512,
4528 __le64_to_cpu(bsb
.arraystart
)*512,
4529 __le64_to_cpu(bsb
.length
)*512, NULL
)) {
4530 /* didn't succeed, so giveup */
4532 pr_err("Error restoring backup from %s\n",
4538 if (bsb
.magic
[15] == '2' &&
4539 restore_stripes(fdlist
, offsets
,
4540 info
->array
.raid_disks
,
4544 fd
, __le64_to_cpu(bsb
.devstart
)*512 +
4545 __le64_to_cpu(bsb
.devstart2
)*512,
4546 __le64_to_cpu(bsb
.arraystart2
)*512,
4547 __le64_to_cpu(bsb
.length2
)*512, NULL
)) {
4548 /* didn't succeed, so giveup */
4550 pr_err("Error restoring second backup from %s\n",
4558 /* Ok, so the data is restored. Let's update those superblocks. */
4562 lo
= __le64_to_cpu(bsb
.arraystart
);
4563 hi
= lo
+ __le64_to_cpu(bsb
.length
);
4565 if (bsb
.magic
[15] == '2' && bsb
.length2
) {
4566 unsigned long long lo1
, hi1
;
4567 lo1
= __le64_to_cpu(bsb
.arraystart2
);
4568 hi1
= lo1
+ __le64_to_cpu(bsb
.length2
);
4572 } else if (lo
< lo1
)
4578 (info
->reshape_progress
< lo
||
4579 info
->reshape_progress
> hi
))
4580 /* backup does not affect reshape_progress*/ ;
4581 else if (info
->delta_disks
>= 0) {
4582 info
->reshape_progress
= __le64_to_cpu(bsb
.arraystart
) +
4583 __le64_to_cpu(bsb
.length
);
4584 if (bsb
.magic
[15] == '2') {
4585 unsigned long long p2
= __le64_to_cpu(bsb
.arraystart2
) +
4586 __le64_to_cpu(bsb
.length2
);
4587 if (p2
> info
->reshape_progress
)
4588 info
->reshape_progress
= p2
;
4591 info
->reshape_progress
= __le64_to_cpu(bsb
.arraystart
);
4592 if (bsb
.magic
[15] == '2') {
4593 unsigned long long p2
= __le64_to_cpu(bsb
.arraystart2
);
4594 if (p2
< info
->reshape_progress
)
4595 info
->reshape_progress
= p2
;
4598 for (j
=0; j
<info
->array
.raid_disks
; j
++) {
4601 if (st
->ss
->load_super(st
, fdlist
[j
], NULL
))
4603 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
4604 dinfo
.reshape_progress
= info
->reshape_progress
;
4605 st
->ss
->update_super(st
, &dinfo
,
4606 "_reshape_progress",
4608 st
->ss
->store_super(st
, fdlist
[j
]);
4609 st
->ss
->free_super(st
);
4613 /* Didn't find any backup data, try to see if any
4616 if (info
->delta_disks
< 0) {
4617 /* When shrinking, the critical section is at the end.
4618 * So see if we are before the critical section.
4620 unsigned long long first_block
;
4621 nstripe
= ostripe
= 0;
4623 while (ostripe
>= nstripe
) {
4624 ostripe
+= info
->array
.chunk_size
/ 512;
4625 first_block
= ostripe
* odata
;
4626 nstripe
= first_block
/ ndata
/ (info
->new_chunk
/512) *
4627 (info
->new_chunk
/512);
4630 if (info
->reshape_progress
>= first_block
)
4633 if (info
->delta_disks
> 0) {
4634 /* See if we are beyond the critical section. */
4635 unsigned long long last_block
;
4636 nstripe
= ostripe
= 0;
4638 while (nstripe
>= ostripe
) {
4639 nstripe
+= info
->new_chunk
/ 512;
4640 last_block
= nstripe
* ndata
;
4641 ostripe
= last_block
/ odata
/ (info
->array
.chunk_size
/512) *
4642 (info
->array
.chunk_size
/512);
4645 if (info
->reshape_progress
>= last_block
)
4648 /* needed to recover critical section! */
4650 pr_err("Failed to find backup of critical section\n");
4654 int Grow_continue_command(char *devname
, int fd
,
4655 char *backup_file
, int verbose
)
4658 struct supertype
*st
= NULL
;
4659 struct mdinfo
*content
= NULL
;
4660 struct mdinfo array
;
4661 char *subarray
= NULL
;
4662 struct mdinfo
*cc
= NULL
;
4663 struct mdstat_ent
*mdstat
= NULL
;
4667 dprintf("Grow continue from command line called for %s\n",
4670 st
= super_by_fd(fd
, &subarray
);
4671 if (!st
|| !st
->ss
) {
4672 pr_err("Unable to determine metadata format for %s\n",
4676 dprintf("Grow continue is run for ");
4677 if (st
->ss
->external
== 0) {
4679 dprintf("native array (%s)\n", devname
);
4680 if (ioctl(fd
, GET_ARRAY_INFO
, &array
.array
) < 0) {
4681 pr_err("%s is not an active md array -"
4682 " aborting\n", devname
);
4684 goto Grow_continue_command_exit
;
4687 /* Need to load a superblock.
4688 * FIXME we should really get what we need from
4691 for (d
= 0; d
< MAX_DISKS
; d
++) {
4692 mdu_disk_info_t disk
;
4696 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0)
4698 if (disk
.major
== 0 && disk
.minor
== 0)
4700 if ((disk
.state
& (1 << MD_DISK_ACTIVE
)) == 0)
4702 dv
= map_dev(disk
.major
, disk
.minor
, 1);
4705 fd2
= dev_open(dv
, O_RDONLY
);
4708 err
= st
->ss
->load_super(st
, fd2
, NULL
);
4710 /* invalidate fd2 to avoid possible double close() */
4716 if (d
== MAX_DISKS
) {
4717 pr_err("Unable to load metadata for %s\n",
4720 goto Grow_continue_command_exit
;
4722 st
->ss
->getinfo_super(st
, content
, NULL
);
4727 dprintf("subarray (%s)\n", subarray
);
4728 container
= st
->container_devnm
;
4729 cfd
= open_dev_excl(st
->container_devnm
);
4731 container
= st
->devnm
;
4733 cfd
= open_dev_excl(st
->devnm
);
4734 dprintf("container (%s)\n", container
);
4738 pr_err("Unable to open container "
4739 "for %s\n", devname
);
4741 goto Grow_continue_command_exit
;
4744 /* find in container array under reshape
4746 ret_val
= st
->ss
->load_container(st
, cfd
, NULL
);
4748 pr_err("Cannot read superblock for %s\n",
4751 goto Grow_continue_command_exit
;
4754 cc
= st
->ss
->container_content(st
, subarray
);
4755 for (content
= cc
; content
; content
= content
->next
) {
4757 int allow_reshape
= 1;
4759 if (content
->reshape_active
== 0)
4761 /* The decision about array or container wide
4762 * reshape is taken in Grow_continue based
4763 * content->reshape_active state, therefore we
4764 * need to check_reshape based on
4765 * reshape_active and subarray name
4767 if (content
->array
.state
& (1<<MD_SB_BLOCK_VOLUME
))
4769 if (content
->reshape_active
== CONTAINER_RESHAPE
&&
4770 (content
->array
.state
4771 & (1<<MD_SB_BLOCK_CONTAINER_RESHAPE
)))
4774 if (!allow_reshape
) {
4775 pr_err("cannot continue reshape of an array"
4776 " in container with unsupported"
4777 " metadata: %s(%s)\n",
4778 devname
, container
);
4780 goto Grow_continue_command_exit
;
4783 array
= strchr(content
->text_version
+1, '/')+1;
4784 mdstat
= mdstat_by_subdev(array
, container
);
4787 if (mdstat
->active
== 0) {
4788 pr_err("Skipping inactive array %s.\n",
4790 free_mdstat(mdstat
);
4797 pr_err("Unable to determine reshaped "
4798 "array for %s\n", devname
);
4800 goto Grow_continue_command_exit
;
4802 fd2
= open_dev(mdstat
->devnm
);
4804 pr_err("cannot open (%s)\n", mdstat
->devnm
);
4806 goto Grow_continue_command_exit
;
4809 sysfs_init(content
, fd2
, mdstat
->devnm
);
4811 /* start mdmon in case it is not running
4813 if (!mdmon_running(container
))
4814 start_mdmon(container
);
4815 ping_monitor(container
);
4817 if (mdmon_running(container
))
4818 st
->update_tail
= &st
->updates
;
4820 pr_err("No mdmon found. "
4821 "Grow cannot continue.\n");
4823 goto Grow_continue_command_exit
;
4827 /* verify that array under reshape is started from
4830 if (verify_reshape_position(content
, content
->array
.level
) < 0) {
4832 goto Grow_continue_command_exit
;
4837 ret_val
= Grow_continue(fd
, st
, content
, backup_file
, 0);
4839 Grow_continue_command_exit
:
4844 st
->ss
->free_super(st
);
4845 free_mdstat(mdstat
);
4852 int Grow_continue(int mdfd
, struct supertype
*st
, struct mdinfo
*info
,
4853 char *backup_file
, int freeze_reshape
)
4857 if (!info
->reshape_active
)
4860 if (st
->ss
->external
) {
4861 int cfd
= open_dev(st
->container_devnm
);
4866 st
->ss
->load_container(st
, cfd
, st
->container_devnm
);
4868 ret_val
= reshape_container(st
->container_devnm
, NULL
, mdfd
,
4869 st
, info
, 0, backup_file
,
4871 1 | info
->reshape_active
,
4874 ret_val
= reshape_array(NULL
, mdfd
, "array", st
, info
, 1,
4875 NULL
, INVALID_SECTORS
,
4877 1 | info
->reshape_active
,