2 * mdadm - manage Linux "md" devices aka RAID arrays.
4 * Copyright (C) 2001-2009 Neil Brown <neilb@suse.de>
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License as published by
9 * the Free Software Foundation; either version 2 of the License, or
10 * (at your option) any later version.
12 * This program is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 * GNU General Public License for more details.
17 * You should have received a copy of the GNU General Public License
18 * along with this program; if not, write to the Free Software
19 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
22 * Email: <neilb@suse.de>
28 #if ! defined(__BIG_ENDIAN) && ! defined(__LITTLE_ENDIAN)
29 #error no endian defined
35 #define offsetof(t,f) ((size_t)&(((t*)0)->f))
38 int Grow_Add_device(char *devname
, int fd
, char *newdev
)
40 /* Add a device to an active array.
41 * Currently, just extend a linear array.
42 * This requires writing a new superblock on the
43 * new device, calling the kernel to add the device,
44 * and if that succeeds, update the superblock on
46 * This means that we need to *find* all other devices.
53 struct supertype
*st
= NULL
;
54 char *subarray
= NULL
;
56 if (ioctl(fd
, GET_ARRAY_INFO
, &info
.array
) < 0) {
57 fprintf(stderr
, Name
": cannot get array info for %s\n", devname
);
61 if (info
.array
.level
!= -1) {
62 fprintf(stderr
, Name
": can only add devices to linear arrays\n");
66 st
= super_by_fd(fd
, &subarray
);
68 fprintf(stderr
, Name
": cannot handle arrays with superblock version %d\n", info
.array
.major_version
);
73 fprintf(stderr
, Name
": Cannot grow linear sub-arrays yet\n");
78 nfd
= open(newdev
, O_RDWR
|O_EXCL
|O_DIRECT
);
80 fprintf(stderr
, Name
": cannot open %s\n", newdev
);
85 if ((stb
.st_mode
& S_IFMT
) != S_IFBLK
) {
86 fprintf(stderr
, Name
": %s is not a block device!\n", newdev
);
91 /* now check out all the devices and make sure we can read the superblock */
92 for (d
=0 ; d
< info
.array
.raid_disks
; d
++) {
96 st
->ss
->free_super(st
);
99 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0) {
100 fprintf(stderr
, Name
": cannot get device detail for device %d\n",
106 dv
= map_dev(disk
.major
, disk
.minor
, 1);
108 fprintf(stderr
, Name
": cannot find device file for device %d\n",
114 fd2
= dev_open(dv
, O_RDWR
);
116 fprintf(stderr
, Name
": cannot open device file %s\n", dv
);
122 if (st
->ss
->load_super(st
, fd2
, NULL
)) {
123 fprintf(stderr
, Name
": cannot find super block on %s\n", dv
);
131 /* Ok, looks good. Lets update the superblock and write it out to
135 info
.disk
.number
= d
;
136 info
.disk
.major
= major(stb
.st_rdev
);
137 info
.disk
.minor
= minor(stb
.st_rdev
);
138 info
.disk
.raid_disk
= d
;
139 info
.disk
.state
= (1 << MD_DISK_SYNC
) | (1 << MD_DISK_ACTIVE
);
140 st
->ss
->update_super(st
, &info
, "linear-grow-new", newdev
,
143 if (st
->ss
->store_super(st
, nfd
)) {
144 fprintf(stderr
, Name
": Cannot store new superblock on %s\n",
151 if (ioctl(fd
, ADD_NEW_DISK
, &info
.disk
) != 0) {
152 fprintf(stderr
, Name
": Cannot add new disk to this array\n");
155 /* Well, that seems to have worked.
156 * Now go through and update all superblocks
159 if (ioctl(fd
, GET_ARRAY_INFO
, &info
.array
) < 0) {
160 fprintf(stderr
, Name
": cannot get array info for %s\n", devname
);
165 for (d
=0 ; d
< info
.array
.raid_disks
; d
++) {
166 mdu_disk_info_t disk
;
170 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0) {
171 fprintf(stderr
, Name
": cannot get device detail for device %d\n",
175 dv
= map_dev(disk
.major
, disk
.minor
, 1);
177 fprintf(stderr
, Name
": cannot find device file for device %d\n",
181 fd2
= dev_open(dv
, O_RDWR
);
183 fprintf(stderr
, Name
": cannot open device file %s\n", dv
);
186 if (st
->ss
->load_super(st
, fd2
, NULL
)) {
187 fprintf(stderr
, Name
": cannot find super block on %s\n", dv
);
191 info
.array
.raid_disks
= nd
+1;
192 info
.array
.nr_disks
= nd
+1;
193 info
.array
.active_disks
= nd
+1;
194 info
.array
.working_disks
= nd
+1;
196 st
->ss
->update_super(st
, &info
, "linear-grow-update", dv
,
199 if (st
->ss
->store_super(st
, fd2
)) {
200 fprintf(stderr
, Name
": Cannot store new superblock on %s\n", dv
);
210 int Grow_addbitmap(char *devname
, int fd
, char *file
, int chunk
, int delay
, int write_behind
, int force
)
213 * First check that array doesn't have a bitmap
214 * Then create the bitmap
217 * For internal bitmaps, we need to check the version,
218 * find all the active devices, and write the bitmap block
221 mdu_bitmap_file_t bmf
;
222 mdu_array_info_t array
;
223 struct supertype
*st
;
224 char *subarray
= NULL
;
225 int major
= BITMAP_MAJOR_HI
;
226 int vers
= md_get_version(fd
);
227 unsigned long long bitmapsize
, array_size
;
230 major
= BITMAP_MAJOR_HOSTENDIAN
;
231 fprintf(stderr
, Name
": Warning - bitmaps created on this kernel"
232 " are not portable\n"
233 " between different architectures. Consider upgrading"
234 " the Linux kernel.\n");
237 if (ioctl(fd
, GET_BITMAP_FILE
, &bmf
) != 0) {
239 fprintf(stderr
, Name
": Memory allocation failure.\n");
241 fprintf(stderr
, Name
": bitmaps not supported by this kernel.\n");
244 if (bmf
.pathname
[0]) {
245 if (strcmp(file
,"none")==0) {
246 if (ioctl(fd
, SET_BITMAP_FILE
, -1)!= 0) {
247 fprintf(stderr
, Name
": failed to remove bitmap %s\n",
253 fprintf(stderr
, Name
": %s already has a bitmap (%s)\n",
254 devname
, bmf
.pathname
);
257 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
258 fprintf(stderr
, Name
": cannot get array status for %s\n", devname
);
261 if (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)) {
262 if (strcmp(file
, "none")==0) {
263 array
.state
&= ~(1<<MD_SB_BITMAP_PRESENT
);
264 if (ioctl(fd
, SET_ARRAY_INFO
, &array
)!= 0) {
265 fprintf(stderr
, Name
": failed to remove internal bitmap.\n");
270 fprintf(stderr
, Name
": Internal bitmap already present on %s\n",
275 if (strcmp(file
, "none") == 0) {
276 fprintf(stderr
, Name
": no bitmap found on %s\n", devname
);
279 if (array
.level
<= 0) {
280 fprintf(stderr
, Name
": Bitmaps not meaningful with level %s\n",
281 map_num(pers
, array
.level
)?:"of this array");
284 bitmapsize
= array
.size
;
286 if (get_dev_size(fd
, NULL
, &array_size
) &&
287 array_size
> (0x7fffffffULL
<<9)) {
288 /* Array is big enough that we cannot trust array.size
289 * try other approaches
291 bitmapsize
= get_component_size(fd
);
293 if (bitmapsize
== 0) {
294 fprintf(stderr
, Name
": Cannot reliably determine size of array to create bitmap - sorry.\n");
298 if (array
.level
== 10) {
299 int ncopies
= (array
.layout
&255)*((array
.layout
>>8)&255);
300 bitmapsize
= bitmapsize
* array
.raid_disks
/ ncopies
;
303 st
= super_by_fd(fd
, &subarray
);
305 fprintf(stderr
, Name
": Cannot understand version %d.%d\n",
306 array
.major_version
, array
.minor_version
);
310 fprintf(stderr
, Name
": Cannot add bitmaps to sub-arrays yet\n");
315 if (strcmp(file
, "internal") == 0) {
317 if (st
->ss
->add_internal_bitmap
== NULL
) {
318 fprintf(stderr
, Name
": Internal bitmaps not supported "
319 "with %s metadata\n", st
->ss
->name
);
322 for (d
=0; d
< st
->max_devs
; d
++) {
323 mdu_disk_info_t disk
;
326 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0)
328 if (disk
.major
== 0 &&
331 if ((disk
.state
& (1<<MD_DISK_SYNC
))==0)
333 dv
= map_dev(disk
.major
, disk
.minor
, 1);
335 int fd2
= dev_open(dv
, O_RDWR
);
338 if (st
->ss
->load_super(st
, fd2
, NULL
)==0) {
339 if (st
->ss
->add_internal_bitmap(
341 &chunk
, delay
, write_behind
,
342 bitmapsize
, 0, major
)
344 st
->ss
->write_bitmap(st
, fd2
);
346 fprintf(stderr
, Name
": failed to create internal bitmap - chunksize problem.\n");
354 array
.state
|= (1<<MD_SB_BITMAP_PRESENT
);
355 if (ioctl(fd
, SET_ARRAY_INFO
, &array
)!= 0) {
358 ": Cannot add bitmap while array is"
359 " resyncing or reshaping etc.\n");
360 fprintf(stderr
, Name
": failed to set internal bitmap.\n");
367 int max_devs
= st
->max_devs
;
369 /* try to load a superblock */
370 for (d
=0; d
<max_devs
; d
++) {
371 mdu_disk_info_t disk
;
375 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0)
377 if ((disk
.major
==0 && disk
.minor
==0) ||
378 (disk
.state
& (1<<MD_DISK_REMOVED
)))
380 dv
= map_dev(disk
.major
, disk
.minor
, 1);
382 fd2
= dev_open(dv
, O_RDONLY
);
384 st
->ss
->load_super(st
, fd2
, NULL
) == 0) {
386 st
->ss
->uuid_from_super(st
, uuid
);
392 fprintf(stderr
, Name
": cannot find UUID for array!\n");
395 if (CreateBitmap(file
, force
, (char*)uuid
, chunk
,
396 delay
, write_behind
, bitmapsize
, major
)) {
399 bitmap_fd
= open(file
, O_RDWR
);
401 fprintf(stderr
, Name
": weird: %s cannot be opened\n",
405 if (ioctl(fd
, SET_BITMAP_FILE
, bitmap_fd
) < 0) {
409 ": Cannot add bitmap while array is"
410 " resyncing or reshaping etc.\n");
411 fprintf(stderr
, Name
": Cannot set bitmap file for %s: %s\n",
412 devname
, strerror(err
));
422 * When reshaping an array we might need to backup some data.
423 * This is written to all spares with a 'super_block' describing it.
424 * The superblock goes 4K from the end of the used space on the
426 * It if written after the backup is complete.
427 * It has the following structure.
430 static struct mdp_backup_super
{
431 char magic
[16]; /* md_backup_data-1 or -2 */
434 /* start/sizes in 512byte sectors */
435 __u64 devstart
; /* address on backup device/file of data */
438 __u32 sb_csum
; /* csum of preceeding bytes. */
440 __u64 devstart2
; /* offset in to data of second section */
443 __u32 sb_csum2
; /* csum of preceeding bytes. */
445 } __attribute__((aligned(512))) bsb
, bsb2
;
447 static __u32
bsb_csum(char *buf
, int len
)
451 for (i
=0; i
<len
; i
++)
452 csum
= (csum
<<3) + buf
[0];
453 return __cpu_to_le32(csum
);
456 static int check_idle(struct supertype
*st
)
458 /* Check that all member arrays for this container, or the
459 * container of this array, are idle
461 int container_dev
= (st
->container_dev
!= NoMdDev
462 ? st
->container_dev
: st
->devnum
);
464 struct mdstat_ent
*ent
, *e
;
467 fmt_devname(container
, container_dev
);
468 ent
= mdstat_read(0, 0);
469 for (e
= ent
; e
; e
= e
->next
) {
470 if (!is_container_member(e
, container
))
472 if (e
->percent
>= 0) {
481 static int freeze_container(struct supertype
*st
)
483 int container_dev
= (st
->container_dev
!= NoMdDev
484 ? st
->container_dev
: st
->devnum
);
490 fmt_devname(container
, container_dev
);
492 if (block_monitor(container
, 1)) {
493 fprintf(stderr
, Name
": failed to freeze container\n");
500 static void unfreeze_container(struct supertype
*st
)
502 int container_dev
= (st
->container_dev
!= NoMdDev
503 ? st
->container_dev
: st
->devnum
);
506 fmt_devname(container
, container_dev
);
508 unblock_monitor(container
, 1);
511 static int freeze(struct supertype
*st
)
513 /* Try to freeze resync/rebuild on this array/container.
514 * Return -1 if the array is busy,
515 * return -2 container cannot be frozen,
516 * return 0 if this kernel doesn't support 'frozen'
517 * return 1 if it worked.
519 if (st
->ss
->external
)
520 return freeze_container(st
);
522 struct mdinfo
*sra
= sysfs_read(-1, st
->devnum
, GET_VERSION
);
527 err
= sysfs_freeze_array(sra
);
533 static void unfreeze(struct supertype
*st
, int frozen
)
535 /* If 'frozen' is 1, unfreeze the array */
539 if (st
->ss
->external
)
540 return unfreeze_container(st
);
542 struct mdinfo
*sra
= sysfs_read(-1, st
->devnum
, GET_VERSION
);
545 sysfs_set_str(sra
, NULL
, "sync_action", "idle");
547 fprintf(stderr
, Name
": failed to unfreeze array\n");
552 static void wait_reshape(struct mdinfo
*sra
)
554 int fd
= sysfs_get_fd(sra
, NULL
, "sync_action");
560 while (sysfs_fd_get_str(fd
, action
, 20) > 0 &&
561 strncmp(action
, "reshape", 7) == 0) {
565 select(fd
+1, NULL
, NULL
, &rfds
, NULL
);
570 static int reshape_super(struct supertype
*st
, long long size
, int level
,
571 int layout
, int chunksize
, int raid_disks
,
572 char *backup_file
, char *dev
, int verbose
)
574 /* nothing extra to check in the native case */
575 if (!st
->ss
->external
)
577 if (!st
->ss
->reshape_super
||
578 !st
->ss
->manage_reshape
) {
579 fprintf(stderr
, Name
": %s metadata does not support reshape\n",
584 return st
->ss
->reshape_super(st
, size
, level
, layout
, chunksize
,
585 raid_disks
, backup_file
, dev
, verbose
);
588 static void sync_metadata(struct supertype
*st
)
590 if (st
->ss
->external
) {
591 if (st
->update_tail
) {
592 flush_metadata_updates(st
);
593 st
->update_tail
= &st
->updates
;
595 st
->ss
->sync_metadata(st
);
599 static int subarray_set_num(char *container
, struct mdinfo
*sra
, char *name
, int n
)
601 /* when dealing with external metadata subarrays we need to be
602 * prepared to handle EAGAIN. The kernel may need to wait for
603 * mdmon to mark the array active so the kernel can handle
604 * allocations/writeback when preparing the reshape action
605 * (md_allow_write()). We temporarily disable safe_mode_delay
606 * to close a race with the array_state going clean before the
607 * next write to raid_disks / stripe_cache_size
612 /* only 'raid_disks' and 'stripe_cache_size' trigger md_allow_write */
614 (strcmp(name
, "raid_disks") != 0 &&
615 strcmp(name
, "stripe_cache_size") != 0))
616 return sysfs_set_num(sra
, NULL
, name
, n
);
618 rc
= sysfs_get_str(sra
, NULL
, "safe_mode_delay", safe
, sizeof(safe
));
621 sysfs_set_num(sra
, NULL
, "safe_mode_delay", 0);
622 rc
= sysfs_set_num(sra
, NULL
, name
, n
);
623 if (rc
< 0 && errno
== EAGAIN
) {
624 ping_monitor(container
);
625 /* if we get EAGAIN here then the monitor is not active
628 rc
= sysfs_set_num(sra
, NULL
, name
, n
);
630 sysfs_set_str(sra
, NULL
, "safe_mode_delay", safe
);
634 int start_reshape(struct mdinfo
*sra
)
637 sysfs_set_num(sra
, NULL
, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL
);
638 err
= sysfs_set_num(sra
, NULL
, "suspend_hi", 0);
639 err
= err
?: sysfs_set_num(sra
, NULL
, "suspend_lo", 0);
640 err
= err
?: sysfs_set_num(sra
, NULL
, "sync_min", 0);
641 err
= err
?: sysfs_set_num(sra
, NULL
, "sync_max", 0);
642 err
= err
?: sysfs_set_str(sra
, NULL
, "sync_action", "reshape");
647 void abort_reshape(struct mdinfo
*sra
)
649 sysfs_set_str(sra
, NULL
, "sync_action", "idle");
650 sysfs_set_num(sra
, NULL
, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL
);
651 sysfs_set_num(sra
, NULL
, "suspend_hi", 0);
652 sysfs_set_num(sra
, NULL
, "suspend_lo", 0);
653 sysfs_set_num(sra
, NULL
, "sync_min", 0);
654 sysfs_set_str(sra
, NULL
, "sync_max", "max");
657 int remove_disks_on_raid10_to_raid0_takeover(struct supertype
*st
,
662 struct mdinfo
*remaining
;
665 nr_of_copies
= layout
& 0xff;
667 remaining
= sra
->devs
;
669 /* for each 'copy', select one device and remove from the list. */
670 for (slot
= 0; slot
< sra
->array
.raid_disks
; slot
+= nr_of_copies
) {
671 struct mdinfo
**diskp
;
674 /* Find a working device to keep */
675 for (diskp
= &remaining
; *diskp
; diskp
= &(*diskp
)->next
) {
676 struct mdinfo
*disk
= *diskp
;
678 if (disk
->disk
.raid_disk
< slot
)
680 if (disk
->disk
.raid_disk
>= slot
+ nr_of_copies
)
682 if (disk
->disk
.state
& (1<<MD_DISK_REMOVED
))
684 if (disk
->disk
.state
& (1<<MD_DISK_FAULTY
))
686 if (!(disk
->disk
.state
& (1<<MD_DISK_SYNC
)))
689 /* We have found a good disk to use! */
691 disk
->next
= sra
->devs
;
700 if (slot
< sra
->array
.raid_disks
) {
701 /* didn't find all slots */
707 sra
->devs
= remaining
;
711 /* Remove all 'remaining' devices from the array */
713 struct mdinfo
*sd
= remaining
;
714 remaining
= sd
->next
;
716 sysfs_set_str(sra
, sd
, "state", "faulty");
717 sysfs_set_str(sra
, sd
, "slot", "none");
718 sysfs_set_str(sra
, sd
, "state", "remove");
719 sd
->disk
.state
|= (1<<MD_DISK_REMOVED
);
720 sd
->disk
.state
&= ~(1<<MD_DISK_SYNC
);
721 sd
->next
= sra
->devs
;
727 void reshape_free_fdlist(int *fdlist
,
728 unsigned long long *offsets
,
733 for (i
= 0; i
< size
; i
++)
741 int reshape_prepare_fdlist(char *devname
,
745 unsigned long blocks
,
748 unsigned long long *offsets
)
753 for (d
= 0; d
<= nrdisks
; d
++)
756 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
757 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
759 if (sd
->disk
.state
& (1<<MD_DISK_SYNC
)) {
760 char *dn
= map_dev(sd
->disk
.major
,
762 fdlist
[sd
->disk
.raid_disk
]
763 = dev_open(dn
, O_RDONLY
);
764 offsets
[sd
->disk
.raid_disk
] = sd
->data_offset
*512;
765 if (fdlist
[sd
->disk
.raid_disk
] < 0) {
767 Name
": %s: cannot open component %s\n",
768 devname
, dn
? dn
: "-unknown-");
772 } else if (backup_file
== NULL
) {
774 char *dn
= map_dev(sd
->disk
.major
,
776 fdlist
[d
] = dev_open(dn
, O_RDWR
);
777 offsets
[d
] = (sd
->data_offset
+ sra
->component_size
- blocks
- 8)*512;
779 fprintf(stderr
, Name
": %s: cannot open component %s\n",
780 devname
, dn
? dn
: "-unknown-");
791 int reshape_open_backup_file(char *backup_file
,
796 unsigned long long *offsets
)
798 /* Return 1 on success, 0 on any form of failure */
799 /* need to check backup file is large enough */
805 *fdlist
= open(backup_file
, O_RDWR
|O_CREAT
|O_EXCL
,
809 fprintf(stderr
, Name
": %s: cannot create backup file %s: %s\n",
810 devname
, backup_file
, strerror(errno
));
813 /* Guard against backup file being on array device.
814 * If array is partitioned or if LVM etc is in the
815 * way this will not notice, but it is better than
818 fstat(*fdlist
, &stb
);
821 if (stb
.st_rdev
== dev
) {
822 fprintf(stderr
, Name
": backup file must NOT be"
823 " on the array being reshaped.\n");
829 for (i
=0; i
< blocks
+ 1 ; i
++) {
830 if (write(*fdlist
, buf
, 512) != 512) {
831 fprintf(stderr
, Name
": %s: cannot create"
832 " backup file %s: %s\n",
833 devname
, backup_file
, strerror(errno
));
837 if (fsync(*fdlist
) != 0) {
838 fprintf(stderr
, Name
": %s: cannot create backup file %s: %s\n",
839 devname
, backup_file
, strerror(errno
));
846 unsigned long compute_backup_blocks(int nchunk
, int ochunk
,
847 unsigned int ndata
, unsigned int odata
)
849 unsigned long a
, b
, blocks
;
850 /* So how much do we need to backup.
851 * We need an amount of data which is both a whole number of
852 * old stripes and a whole number of new stripes.
853 * So LCM for (chunksize*datadisks).
855 a
= (ochunk
/512) * odata
;
856 b
= (nchunk
/512) * ndata
;
864 /* LCM == product / GCD */
865 blocks
= (ochunk
/512) * (nchunk
/512) * odata
* ndata
/ a
;
870 /* 'struct reshape' records the intermediate states
872 * The starting geometry is converted to the 'before' geometry
873 * by at most an atomic level change. They could be the same.
874 * Similarly the 'after' geometry is converted to the final
875 * geometry by at most a level change.
876 * Note that 'before' and 'after' must have the same level.
877 * 'blocks' is the minimum number of sectors for a reshape unit.
878 * This will be a multiple of the stripe size in each of the
879 * 'before' and 'after' geometries.
880 * If 'blocks' is 0, no restriping is necessary.
884 int parity
; /* number of parity blocks/devices */
889 unsigned long long blocks
;
890 unsigned long long stripes
; /* number of old stripes that comprise 'blocks'*/
891 unsigned long long new_size
; /* New size of array in sectors */
894 char *analyse_change(struct mdinfo
*info
, struct reshape
*re
)
896 /* Based on the current array state in info->array and
897 * the changes in info->new_* etc, determine:
898 * - whether the change is possible
899 * - Intermediate level/raid_disks/layout
900 * - whether a restriping reshape is needed
901 * - number of sectors in minimum change unit. This
902 * will cover a whole number of stripes in 'before' and
905 * Return message if the change should be rejected
906 * NULL if the change can be achieved
908 * This can be called as part of starting a reshape, or
909 * when assembling an array that is undergoing reshape.
913 /* If a new level not explicitly given, we assume no-change */
914 if (info
->new_level
== UnSet
)
915 info
->new_level
= info
->array
.level
;
918 switch (info
->new_level
) {
924 /* chunk size is meaningful, must divide component_size
927 if (info
->component_size
% (info
->new_chunk
/512))
928 return "New chunk size does not"
929 " divide component size";
932 return "chunk size not meaningful for this level";
935 info
->new_chunk
= info
->array
.chunk_size
;
937 switch (info
->array
.level
) {
939 /* RAID1 can convert to RAID1 with different disks, or
942 if (info
->new_level
== 1) {
943 if (info
->delta_disks
== UnSet
)
944 /* Don't know what to do */
945 return "no change requested for Growing RAID1";
947 re
->before
.data_disks
= (info
->array
.raid_disks
+
949 re
->before
.layout
= 0;
954 if (info
->array
.raid_disks
== 2 &&
955 info
->array
.raid_disks
== 5) {
956 /* simple in-place conversion */
959 re
->before
.data_disks
= 1;
960 re
->before
.layout
= ALGORITHM_LEFT_SYMMETRIC
;
964 /* Could do some multi-stage conversions, but leave that to
967 return "Impossibly level change request for RAID1";
970 /* RAID10 can only be converted from near mode to
971 * RAID0 by removing some devices
973 if ((info
->array
.layout
& ~0xff) != 0x100)
974 return "Cannot Grow RAID10 with far/offset layout";
975 /* number of devices must be multiple of number of copies */
976 if (info
->array
.raid_disks
% (info
->array
.layout
& 0xff))
977 return "RAID10 layout too complex for Grow operation";
979 if (info
->new_level
!= 0)
980 return "RAID10 can only be changed to RAID0";
981 new_disks
= (info
->array
.raid_disks
982 / (info
->array
.layout
& 0xff));
983 if (info
->delta_disks
!= UnSet
) {
984 info
->delta_disks
= (new_disks
985 - info
->array
.raid_disks
);
987 if (info
->delta_disks
!= new_disks
- info
->array
.raid_disks
)
988 return "New number of raid-devices impossible for RAID10";
989 if (info
->new_chunk
&&
990 info
->new_chunk
!= info
->array
.chunk_size
)
991 return "Cannot change chunk-size with RAID10 Grow";
996 re
->before
.data_disks
= new_disks
;
997 re
->before
.layout
= 0;
1002 /* RAID0 can be converted to RAID10, or to RAID456 */
1003 if (info
->new_level
== 10) {
1004 if (info
->new_layout
== UnSet
&& info
->delta_disks
== UnSet
) {
1005 /* Assume near=2 layout */
1006 info
->new_layout
= 0x102;
1007 info
->delta_disks
= info
->array
.raid_disks
;
1009 if (info
->new_layout
== UnSet
) {
1010 int copies
= 1 + (info
->delta_disks
1011 / info
->array
.raid_disks
);
1012 if (info
->array
.raid_disks
* (copies
-1)
1013 != info
->delta_disks
)
1014 return "Impossible number of devices"
1015 " for RAID0->RAID10";
1016 info
->new_layout
= 0x100 + copies
;
1018 if (info
->delta_disks
== UnSet
) {
1019 int copies
= info
->new_layout
& 0xff;
1020 if (info
->new_layout
!= 0x100 + copies
)
1021 return "New layout impossible"
1022 " for RAID0->RAID10";;
1023 info
->delta_disks
= (copies
- 1) *
1024 info
->array
.raid_disks
;
1026 if (info
->new_chunk
&&
1027 info
->new_chunk
!= info
->array
.chunk_size
)
1028 return "Cannot change chunk-size with RAID0->RAID10";
1032 re
->before
.data_disks
= (info
->array
.raid_disks
+
1034 re
->before
.layout
= info
->new_layout
;
1039 /* RAID0 can also covert to RAID0/4/5/6 by first converting to
1040 * a raid4 style layout of the final level.
1042 switch (info
->new_level
) {
1046 re
->before
.layout
= 0;
1050 re
->before
.layout
= ALGORITHM_PARITY_N
;
1054 re
->before
.layout
= ALGORITHM_PARITY_N
;
1057 return "Impossible level change requested";
1059 re
->before
.data_disks
= info
->array
.raid_disks
;
1060 /* determining 'after' layout happens outside this 'switch' */
1064 info
->array
.layout
= ALGORITHM_PARITY_N
;
1066 switch (info
->new_level
) {
1068 re
->level
= info
->array
.level
;
1069 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1070 re
->before
.layout
= info
->array
.layout
;
1074 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1075 re
->before
.layout
= info
->array
.layout
;
1079 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1080 switch (info
->array
.layout
) {
1081 case ALGORITHM_LEFT_ASYMMETRIC
:
1082 re
->before
.layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
1084 case ALGORITHM_RIGHT_ASYMMETRIC
:
1085 re
->before
.layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
1087 case ALGORITHM_LEFT_SYMMETRIC
:
1088 re
->before
.layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
1090 case ALGORITHM_RIGHT_SYMMETRIC
:
1091 re
->before
.layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
1093 case ALGORITHM_PARITY_0
:
1094 re
->before
.layout
= ALGORITHM_PARITY_0_6
;
1096 case ALGORITHM_PARITY_N
:
1097 re
->before
.layout
= ALGORITHM_PARITY_N_6
;
1100 return "Cannot convert an array with this layout";
1104 if (info
->array
.raid_disks
!= 2)
1105 return "Can only convert a 2-device array to RAID1";
1107 re
->before
.data_disks
= 2;
1108 re
->before
.layout
= 0;
1111 return "Impossible level change requested";
1115 switch (info
->new_level
) {
1120 re
->before
.data_disks
= info
->array
.raid_disks
- 2;
1121 re
->before
.layout
= info
->array
.layout
;
1124 return "Impossible level change requested";
1129 /* If we reached here then it looks like a re-stripe is
1130 * happening. We have determined the intermediate level
1131 * and initial raid_disks/layout and stored these in 're'.
1133 * We need to deduce the final layout that can be atomically
1134 * converted to the end state.
1136 switch (info
->new_level
) {
1138 /* We can only get to RAID0 from RAID4 or RAID5
1139 * with appropriate layout and one extra device
1141 if (re
->level
!= 4 && re
->level
!= 5)
1142 return "Cannot covert to RAID0 from this level";
1143 if (info
->delta_disks
== UnSet
)
1144 re
->after
.data_disks
= re
->before
.data_disks
;
1146 re
->after
.data_disks
=
1147 info
->array
.raid_disks
+ info
->delta_disks
;
1148 switch (re
->level
) {
1150 re
->after
.layout
= 0 ; break;
1152 re
->after
.layout
= ALGORITHM_PARITY_N
; break;
1157 /* We can only get to RAID4 from RAID5 */
1158 if (re
->level
!= 4 && re
->level
!= 5)
1159 return "Cannot convert to RAID4 from this level";
1160 if (info
->delta_disks
== UnSet
)
1161 re
->after
.data_disks
= re
->before
.data_disks
;
1163 re
->after
.data_disks
=
1164 re
->before
.data_disks
+ info
->delta_disks
;
1165 switch (re
->level
) {
1167 re
->after
.layout
= 0 ; break;
1169 re
->after
.layout
= ALGORITHM_PARITY_N
; break;
1174 /* We get to RAID5 for RAID5 or RAID6 */
1175 if (re
->level
!= 5 && re
->level
!= 6)
1176 return "Cannot convert to RAID5 from this level";
1177 if (info
->delta_disks
== UnSet
)
1178 re
->after
.data_disks
= re
->before
.data_disks
;
1179 else if (re
->level
== 5)
1180 re
->after
.data_disks
=
1181 re
->before
.data_disks
+ info
->delta_disks
;
1183 re
->after
.data_disks
=
1184 info
->array
.raid_disks
+ info
->delta_disks
- 1;
1185 switch (re
->level
) {
1187 if (info
->new_layout
== UnSet
)
1188 re
->after
.layout
= re
->before
.layout
;
1190 re
->after
.layout
= info
->new_layout
;
1193 if (info
->new_layout
== UnSet
) {
1194 re
->after
.layout
= re
->before
.layout
;
1197 /* after.layout needs to be raid6 version of new_layout */
1198 if (info
->new_layout
== ALGORITHM_PARITY_N
)
1199 re
->after
.layout
= ALGORITHM_PARITY_N
;
1202 char *ls
= map_num(r5layout
, info
->new_layout
);
1204 strcat(strcpy(layout
, ls
), "-6");
1205 l
= map_name(r6layout
, layout
);
1207 return "Cannot find RAID6 layout"
1209 re
->after
.layout
= l
;
1215 /* We must already be at level 6 */
1217 return "Impossible level change";
1218 if (info
->delta_disks
== UnSet
)
1219 re
->after
.data_disks
= re
->before
.data_disks
;
1221 re
->after
.data_disks
= (info
->array
.raid_disks
+
1222 info
->delta_disks
) - 2;
1223 if (info
->new_layout
== UnSet
)
1224 re
->after
.layout
= re
->before
.layout
;
1226 re
->after
.layout
= info
->new_layout
;
1229 return "Impossible level change requested";
1231 switch (re
->level
) {
1232 case 6: re
->parity
= 2; break;
1234 case 5: re
->parity
= 1; break;
1235 default: re
->parity
= 0; break;
1237 /* So we have a restripe operation, we need to calculate the number
1238 * of blocks per reshape operation.
1240 if (info
->new_chunk
== 0)
1241 info
->new_chunk
= info
->array
.chunk_size
;
1242 if (re
->after
.data_disks
== re
->before
.data_disks
&&
1243 re
->after
.layout
== re
->before
.layout
&&
1244 info
->new_chunk
== info
->array
.chunk_size
) {
1245 /* Nothing to change */
1249 if (re
->after
.data_disks
== 1 && re
->before
.data_disks
== 1) {
1250 /* chunks can layout changes make no difference */
1255 if (re
->after
.data_disks
== re
->before
.data_disks
&&
1256 get_linux_version() < 2006032)
1257 return "in-place reshape is not safe before 2.6.32 - sorry.";
1259 if (re
->after
.data_disks
< re
->before
.data_disks
&&
1260 get_linux_version() < 2006030)
1261 return "reshape to fewer devices is not supported before 2.6.32 - sorry.";
1263 re
->blocks
= compute_backup_blocks(
1264 info
->new_chunk
, info
->array
.chunk_size
,
1265 re
->after
.data_disks
,
1266 re
->before
.data_disks
);
1268 re
->new_size
= info
->component_size
* re
->after
.data_disks
;
1272 static int reshape_array(char *container
, int fd
, char *devname
,
1273 struct supertype
*st
, struct mdinfo
*info
,
1274 int force
, char *backup_file
, int quiet
, int forked
);
1275 static int reshape_container(char *container
, int cfd
, char *devname
,
1276 struct supertype
*st
,
1277 struct mdinfo
*info
,
1281 static int child_monitor(int afd
, struct mdinfo
*sra
, struct reshape
*reshape
,
1282 unsigned long stripes
,
1283 int *fds
, unsigned long long *offsets
,
1284 int dests
, int *destfd
, unsigned long long *destoffsets
);
1287 int Grow_reshape(char *devname
, int fd
, int quiet
, char *backup_file
,
1289 int level
, char *layout_str
, int chunksize
, int raid_disks
,
1292 /* Make some changes in the shape of an array.
1293 * The kernel must support the change.
1295 * There are three different changes. Each can trigger
1296 * a resync or recovery so we freeze that until we have
1297 * requested everything (if kernel supports freezing - 2.6.30).
1299 * - change size (i.e. component_size)
1301 * - change layout/chunksize/ndisks
1303 * The last can require a reshape. It is different on different
1304 * levels so we need to check the level before actioning it.
1305 * Some times the level change needs to be requested after the
1306 * reshape (e.g. raid6->raid5, raid5->raid0)
1309 struct mdu_array_info_s array
;
1311 struct supertype
*st
;
1312 char *subarray
= NULL
;
1316 char *container
= NULL
;
1317 char container_buf
[20];
1323 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) < 0) {
1324 fprintf(stderr
, Name
": %s is not an active md array - aborting\n",
1330 (chunksize
|| level
!= UnSet
|| layout_str
|| raid_disks
)) {
1331 fprintf(stderr
, Name
": cannot change component size at the same time "
1332 "as other changes.\n"
1333 " Change size first, then check data is intact before "
1334 "making other changes.\n");
1338 if (raid_disks
&& raid_disks
< array
.raid_disks
&& array
.level
> 1 &&
1339 get_linux_version() < 2006032 &&
1340 !check_env("MDADM_FORCE_FEWER")) {
1341 fprintf(stderr
, Name
": reducing the number of devices is not safe before Linux 2.6.32\n"
1342 " Please use a newer kernel\n");
1346 st
= super_by_fd(fd
, &subarray
);
1348 fprintf(stderr
, Name
": Unable to determine metadata format for %s\n", devname
);
1351 if (raid_disks
> st
->max_devs
) {
1352 fprintf(stderr
, Name
": Cannot increase raid-disks on this array"
1353 " beyond %d\n", st
->max_devs
);
1357 /* in the external case we need to check that the requested reshape is
1358 * supported, and perform an initial check that the container holds the
1359 * pre-requisite spare devices (mdmon owns final validation)
1361 if (st
->ss
->external
) {
1366 container_dev
= st
->container_dev
;
1367 cfd
= open_dev_excl(st
->container_dev
);
1369 container_dev
= st
->devnum
;
1371 cfd
= open_dev_excl(st
->devnum
);
1375 fprintf(stderr
, Name
": Unable to open container for %s\n",
1381 fmt_devname(container_buf
, container_dev
);
1382 container
= container_buf
;
1385 rv
= st
->ss
->load_container(st
, cfd
, NULL
);
1387 rv
= st
->ss
->load_super(st
, cfd
, NULL
);
1389 fprintf(stderr
, Name
": Cannot read superblock for %s\n",
1395 if (mdmon_running(container_dev
))
1396 st
->update_tail
= &st
->updates
;
1399 if (raid_disks
> array
.raid_disks
&&
1400 array
.spare_disks
< (raid_disks
- array
.raid_disks
) &&
1403 Name
": Need %d spare%s to avoid degraded array,"
1404 " and only have %d.\n"
1405 " Use --force to over-ride this check.\n",
1406 raid_disks
- array
.raid_disks
,
1407 raid_disks
- array
.raid_disks
== 1 ? "" : "s",
1412 sra
= sysfs_read(fd
, 0, GET_LEVEL
| GET_DISKS
| GET_DEVS
| GET_STATE
);
1414 if (st
->ss
->external
&& subarray
== NULL
) {
1415 array
.level
= LEVEL_CONTAINER
;
1416 sra
->array
.level
= LEVEL_CONTAINER
;
1419 fprintf(stderr
, Name
": failed to read sysfs parameters for %s\n",
1423 frozen
= freeze(st
);
1425 /* freeze() already spewed the reason */
1427 } else if (frozen
< 0) {
1428 fprintf(stderr
, Name
": %s is performing resync/recovery and cannot"
1429 " be reshaped\n", devname
);
1433 /* ========= set size =============== */
1434 if (size
>= 0 && (size
== 0 || size
!= array
.size
)) {
1435 long long orig_size
= array
.size
;
1437 if (reshape_super(st
, size
, UnSet
, UnSet
, 0, 0, NULL
, devname
, !quiet
)) {
1443 if (array
.size
!= size
) {
1444 /* got truncated to 32bit, write to
1445 * component_size instead
1448 rv
= sysfs_set_num(sra
, NULL
,
1449 "component_size", size
);
1453 rv
= ioctl(fd
, SET_ARRAY_INFO
, &array
);
1457 /* restore metadata */
1458 if (reshape_super(st
, orig_size
, UnSet
, UnSet
, 0, 0,
1459 NULL
, devname
, !quiet
) == 0)
1461 fprintf(stderr
, Name
": Cannot set device size for %s: %s\n",
1462 devname
, strerror(err
));
1464 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
1465 fprintf(stderr
, " Bitmap must be removed before size can be changed\n");
1469 ioctl(fd
, GET_ARRAY_INFO
, &array
);
1470 size
= get_component_size(fd
)/2;
1474 fprintf(stderr
, Name
": component size of %s has been set to %lluK\n",
1477 } else if (array
.level
!= LEVEL_CONTAINER
) {
1478 size
= get_component_size(fd
)/2;
1483 /* ========= check for Raid10 -> Raid0 conversion ===============
1484 * current implementation assumes that following conditions must be met:
1486 * - near_copies == 2
1488 if (level
== 0 && array
.level
== 10 && sra
&&
1489 array
.layout
== ((1 << 8) + 2) && !(array
.raid_disks
& 1)) {
1491 err
= remove_disks_on_raid10_to_raid0_takeover(st
, sra
, array
.layout
);
1493 dprintf(Name
": Array cannot be reshaped\n");
1502 sysfs_init(&info
, fd
, NoMdDev
);
1503 info
.component_size
= size
*2;
1504 info
.new_level
= level
;
1505 info
.new_chunk
= chunksize
* 1024;
1507 info
.delta_disks
= raid_disks
- info
.array
.raid_disks
;
1509 info
.delta_disks
= UnSet
;
1510 if (layout_str
== NULL
) {
1511 info
.new_layout
= UnSet
;
1512 if (info
.array
.level
== 6 &&
1513 (info
.new_level
== 6 || info
.new_level
== UnSet
) &&
1514 info
.array
.layout
>= 16) {
1515 fprintf(stderr
, Name
1516 ": %s has a non-standard layout. If you"
1517 " wish to preserve this\n"
1518 " during the reshape, please specify"
1519 " --layout=preserve\n"
1520 " If you want to change it, specify a"
1521 " layout or use --layout=normalise\n",
1526 } else if (strcmp(layout_str
, "normalise") == 0 ||
1527 strcmp(layout_str
, "normalize") == 0) {
1528 /* If we have a -6 RAID6 layout, remove the '-6'. */
1529 info
.new_layout
= UnSet
;
1530 if (info
.array
.level
== 6 && info
.new_level
== UnSet
) {
1532 strcpy(l
, map_num(r6layout
, info
.array
.layout
));
1533 h
= strrchr(l
, '-');
1534 if (h
&& strcmp(h
, "-6") == 0) {
1536 info
.new_layout
= map_name(r6layout
, l
);
1539 } else if (strcmp(layout_str
, "preserve") == 0) {
1540 info
.new_layout
= UnSet
;
1542 int l
= info
.new_level
;
1544 l
= info
.array
.level
;
1547 info
.new_layout
= map_name(r5layout
, layout_str
);
1550 info
.new_layout
= map_name(r6layout
, layout_str
);
1553 info
.new_layout
= parse_layout_10(layout_str
);
1556 info
.new_layout
= parse_layout_faulty(layout_str
);
1559 fprintf(stderr
, Name
": layout not meaningful"
1560 " with this level\n");
1564 if (info
.new_layout
== UnSet
) {
1565 fprintf(stderr
, Name
": layout %s not understood"
1566 " for this level\n",
1573 if (array
.level
== LEVEL_CONTAINER
) {
1574 /* This change is to be applied to every array in the
1575 * container. This is only needed when the metadata imposes
1576 * restraints of the various arrays in the container.
1577 * Currently we only know that IMSM requires all arrays
1578 * to have the same number of devices so changing the
1579 * number of devices (On-Line Capacity Expansion) must be
1580 * performed at the level of the container
1582 rv
= reshape_container(container
, fd
, devname
, st
, &info
,
1583 force
, backup_file
, quiet
);
1585 /* Impose these changes on a single array. First
1586 * check that the metadata is OK with the change. */
1588 if (reshape_super(st
, info
.component_size
, info
.new_level
,
1589 info
.new_layout
, info
.new_chunk
,
1590 info
.array
.raid_disks
+ info
.delta_disks
,
1591 backup_file
, devname
, quiet
)) {
1596 rv
= reshape_array(container
, fd
, devname
, st
, &info
, force
,
1597 backup_file
, quiet
, 0);
1600 unfreeze(st
, frozen
);
1604 static int reshape_array(char *container
, int fd
, char *devname
,
1605 struct supertype
*st
, struct mdinfo
*info
,
1607 char *backup_file
, int quiet
, int forked
)
1609 struct reshape reshape
;
1612 int orig_level
= UnSet
;
1615 struct mdu_array_info_s array
;
1620 unsigned long long *offsets
;
1625 unsigned long blocks
, stripes
;
1626 unsigned long cache
;
1627 unsigned long long array_size
;
1629 struct mdinfo
*sra
, *sd
;
1631 msg
= analyse_change(info
, &reshape
);
1633 fprintf(stderr
, Name
": %s\n", msg
);
1636 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
1637 dprintf("Canot get array information.\n");
1640 spares_needed
= max(reshape
.before
.data_disks
,
1641 reshape
.after
.data_disks
)
1642 + reshape
.parity
- array
.raid_disks
;
1644 if (!force
&& spares_needed
< info
->array
.spare_disks
) {
1646 Name
": Need %d spare%s to avoid degraded array,"
1647 " and only have %d.\n"
1648 " Use --force to over-ride this check.\n",
1650 spares_needed
== 1 ? "" : "s",
1651 info
->array
.spare_disks
);
1655 if (reshape
.level
!= info
->array
.level
) {
1656 char *c
= map_num(pers
, reshape
.level
);
1659 return 1; /* This should not be possible */
1661 err
= sysfs_set_str(info
, NULL
, "level", c
);
1664 fprintf(stderr
, Name
": %s: could not set level to %s\n",
1667 (info
->array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
1668 fprintf(stderr
, " Bitmap must be removed"
1669 " before level can be changed\n");
1673 fprintf(stderr
, Name
" level of %s changed to %s\n",
1675 orig_level
= info
->array
.level
;
1678 if (reshape
.level
> 0 && st
->ss
->external
&&
1679 !mdmon_running(st
->container_dev
)) {
1680 start_mdmon(st
->container_dev
);
1681 ping_monitor(container
);
1684 /* ->reshape_super might have chosen some spares from the
1685 * container that it wants to be part of the new array.
1686 * We can collect them with ->container_content and give
1687 * them to the kernel.
1689 if (st
->ss
->reshape_super
&& st
->ss
->container_content
) {
1690 char *subarray
= strchr(info
->text_version
+1, '/')+1;
1691 struct mdinfo
*info2
=
1692 st
->ss
->container_content(st
, subarray
);
1696 for (d
= info2
->devs
; d
; d
= d
->next
) {
1697 if (d
->disk
.state
== 0 &&
1698 d
->disk
.raid_disk
>= 0) {
1699 /* This is a spare that wants to
1700 * be part of the array.
1702 add_disk(fd
, st
, info2
, d
);
1708 if (reshape
.blocks
== 0) {
1709 /* No restriping needed, but we might need to impose
1710 * some more changes: layout, raid_disks, chunk_size
1712 if (info
->new_layout
!= UnSet
&&
1713 info
->new_layout
!= info
->array
.layout
) {
1714 info
->array
.layout
= info
->new_layout
;
1715 if (ioctl(fd
, SET_ARRAY_INFO
, &info
->array
) != 0) {
1716 fprintf(stderr
, Name
": failed to set new layout\n");
1719 printf("layout for %s set to %d\n",
1720 devname
, info
->array
.layout
);
1722 if (info
->delta_disks
!= UnSet
&&
1723 info
->delta_disks
!= 0) {
1724 info
->array
.raid_disks
+= info
->delta_disks
;
1725 if (ioctl(fd
, SET_ARRAY_INFO
, &info
->array
) != 0) {
1726 fprintf(stderr
, Name
": failed to set raid disks\n");
1729 printf("raid_disks for %s set to %d\n",
1730 devname
, info
->array
.raid_disks
);
1732 if (info
->new_chunk
!= 0 &&
1733 info
->new_chunk
!= info
->array
.chunk_size
) {
1734 if (sysfs_set_num(info
, NULL
,
1735 "chunk_size", info
->new_chunk
) != 0) {
1736 fprintf(stderr
, Name
": failed to set chunk size\n");
1739 printf("chunk size for %s set to %d\n",
1740 devname
, info
->array
.chunk_size
);
1747 * There are three possibilities.
1748 * 1/ The array will shrink.
1749 * We need to ensure the reshape will pause before reaching
1750 * the 'critical section'. We also need to fork and wait for
1751 * that to happen. When it does we
1752 * suspend/backup/complete/unfreeze
1754 * 2/ The array will not change size.
1755 * This requires that we keep a backup of a sliding window
1756 * so that we can restore data after a crash. So we need
1757 * to fork and monitor progress.
1758 * In future we will allow the data_offset to change, so
1759 * a sliding backup becomes unnecessary.
1761 * 3/ The array will grow. This is relatively easy.
1762 * However the kernel's restripe routines will cheerfully
1763 * overwrite some early data before it is safe. So we
1764 * need to make a backup of the early parts of the array
1765 * and be ready to restore it if rebuild aborts very early.
1766 * For externally managed metadata, we still need a forked
1767 * child to monitor the reshape and suspend IO over the region
1768 * that is being reshaped.
1770 * We backup data by writing it to one spare, or to a
1771 * file which was given on command line.
1773 * In each case, we first make sure that storage is available
1774 * for the required backup.
1776 * - request the shape change.
1777 * - fork to handle backup etc.
1780 /* Check that we can hold all the data */
1781 get_dev_size(fd
, NULL
, &array_size
);
1782 if (reshape
.new_size
< (array_size
/512)) {
1784 Name
": this change will reduce the size of the array.\n"
1785 " use --grow --array-size first to truncate array.\n"
1786 " e.g. mdadm --grow %s --array-size %llu\n",
1787 devname
, reshape
.new_size
/2);
1792 sra
= sysfs_read(fd
, 0,
1793 GET_COMPONENT
|GET_DEVS
|GET_OFFSET
|GET_STATE
|GET_CHUNK
|
1797 fprintf(stderr
, Name
": %s: Cannot get array details from sysfs\n",
1803 /* Decide how many blocks (sectors) for a reshape
1804 * unit. The number we have so far is just a minimum
1806 blocks
= reshape
.blocks
;
1807 if (reshape
.before
.data_disks
==
1808 reshape
.after
.data_disks
) {
1809 /* Make 'blocks' bigger for better throughput, but
1810 * not so big that we reject it below.
1811 * Try for 16 megabytes
1813 while (blocks
* 32 < sra
->component_size
&&
1817 fprintf(stderr
, Name
": Need to backup %luK of critical "
1818 "section..\n", blocks
/2);
1820 if (blocks
>= sra
->component_size
/2) {
1821 fprintf(stderr
, Name
": %s: Something wrong"
1822 " - reshape aborted\n",
1828 /* Now we need to open all these devices so we can read/write.
1830 nrdisks
= array
.raid_disks
+ sra
->array
.spare_disks
;
1831 fdlist
= malloc((1+nrdisks
) * sizeof(int));
1832 offsets
= malloc((1+nrdisks
) * sizeof(offsets
[0]));
1833 if (!fdlist
|| !offsets
) {
1834 fprintf(stderr
, Name
": malloc failed: grow aborted\n");
1839 d
= reshape_prepare_fdlist(devname
, sra
, array
.raid_disks
,
1840 nrdisks
, blocks
, backup_file
,
1846 if (backup_file
== NULL
) {
1847 if (reshape
.after
.data_disks
<= reshape
.before
.data_disks
) {
1849 Name
": %s: Cannot grow - need backup-file\n",
1853 } else if (sra
->array
.spare_disks
== 0) {
1854 fprintf(stderr
, Name
": %s: Cannot grow - need a spare or "
1855 "backup-file to backup critical section\n",
1861 if (!reshape_open_backup_file(backup_file
, fd
, devname
,
1863 fdlist
+d
, offsets
+d
)) {
1870 /* lastly, check that the internal stripe cache is
1871 * large enough, or it won't work.
1872 * It must hold at least 4 stripes of the larger
1875 cache
= max(info
->array
.chunk_size
, info
->new_chunk
);
1876 cache
*= 4; /* 4 stripes minimum */
1877 cache
/= 512; /* convert to sectors */
1878 disks
= min(reshape
.before
.data_disks
, reshape
.after
.data_disks
);
1879 /* make sure there is room for 'blocks' with a bit to spare */
1880 if (cache
< 16 + blocks
/ disks
)
1881 cache
= 16 + blocks
/ disks
;
1882 cache
/= (4096/512); /* Covert from sectors to pages */
1884 if (sra
->cache_size
< cache
)
1885 subarray_set_num(container
, sra
, "stripe_cache_size",
1888 /* Right, everything seems fine. Let's kick things off.
1889 * If only changing raid_disks, use ioctl, else use
1894 if (info
->array
.chunk_size
== info
->new_chunk
&&
1895 reshape
.before
.layout
== reshape
.after
.layout
&&
1896 st
->ss
->external
== 0) {
1897 array
.raid_disks
= reshape
.after
.data_disks
+ reshape
.parity
;
1898 if (ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
1902 Name
": Cannot set device shape for %s: %s\n",
1903 devname
, strerror(errno
));
1906 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
1908 " Bitmap must be removed before"
1909 " shape can be changed\n");
1914 /* set them all just in case some old 'new_*' value
1915 * persists from some earlier problem
1917 int err
= err
; /* only used if rv==1, and always set if
1918 * rv==1, so initialisation not needed,
1919 * despite gcc warning
1921 if (sysfs_set_num(sra
, NULL
, "chunk_size", info
->new_chunk
) < 0)
1922 rv
= 1, err
= errno
;
1923 if (!rv
&& sysfs_set_num(sra
, NULL
, "layout",
1924 reshape
.after
.layout
) < 0)
1925 rv
= 1, err
= errno
;
1926 if (!rv
&& subarray_set_num(container
, sra
, "raid_disks",
1927 reshape
.after
.data_disks
+
1928 reshape
.parity
) < 0)
1929 rv
= 1, err
= errno
;
1931 fprintf(stderr
, Name
": Cannot set device shape for %s\n",
1935 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
1937 " Bitmap must be removed before"
1938 " shape can be changed\n");
1944 if (st
->ss
->external
) {
1945 /* metadata handler takes it from here */
1946 ping_manager(container
);
1947 st
->ss
->manage_reshape(st
, backup_file
);
1952 /* set up the backup-super-block. This requires the
1953 * uuid from the array.
1955 /* Find a superblock */
1956 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
1960 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
1962 dn
= map_dev(sd
->disk
.major
, sd
->disk
.minor
, 1);
1963 devfd
= dev_open(dn
, O_RDONLY
);
1966 ok
= st
->ss
->load_super(st
, devfd
, NULL
);
1972 fprintf(stderr
, Name
": %s: Cannot find a superblock\n",
1979 memset(&bsb
, 0, 512);
1980 memcpy(bsb
.magic
, "md_backup_data-1", 16);
1981 st
->ss
->uuid_from_super(st
, (int*)&bsb
.set_uuid
);
1982 bsb
.mtime
= __cpu_to_le64(time(0));
1983 bsb
.devstart2
= blocks
;
1985 stripes
= reshape
.blocks
/ (info
->array
.chunk_size
/512) /
1986 reshape
.before
.data_disks
;
1988 /* Now we just need to kick off the reshape and watch, while
1989 * handling backups of the data...
1990 * This is all done by a forked background process.
1992 switch(forked
? 0 : fork()) {
1995 if (check_env("MDADM_GROW_VERIFY"))
1996 fd
= open(devname
, O_RDONLY
| O_DIRECT
);
1999 mlockall(MCL_FUTURE
);
2001 odisks
= reshape
.before
.data_disks
+ reshape
.parity
;
2003 done
= child_monitor(fd
, sra
, &reshape
, stripes
,
2005 d
- odisks
, fdlist
+odisks
, offsets
+odisks
);
2007 if (backup_file
&& done
)
2008 unlink(backup_file
);
2013 /* set new array size if required customer_array_size is used
2016 if (reshape
.before
.data_disks
!=
2017 reshape
.after
.data_disks
&&
2018 info
->custom_array_size
) {
2019 struct mdinfo
*info2
;
2020 char *subarray
= strchr(info
->text_version
+1, '/')+1;
2023 ping_monitor(container
);
2025 info2
= st
->ss
->container_content(st
, subarray
);
2027 unsigned long long current_size
= 0;
2028 unsigned long long new_size
=
2029 info2
->custom_array_size
/2;
2031 if (sysfs_get_ll(sra
,
2034 ¤t_size
) == 0 &&
2035 new_size
> current_size
) {
2036 if (sysfs_set_num(sra
, NULL
,
2037 "array_size", new_size
)
2039 dprintf("Error: Cannot"
2042 dprintf("Array size "
2044 dprintf(" from %llu to %llu.\n",
2045 current_size
, new_size
);
2051 if (info
->new_level
!= reshape
.level
) {
2052 /* We need to wait for the reshape to finish
2053 * (which will have happened unless
2054 * odata < ndata) and then set the level
2057 if (reshape
.before
.data_disks
<
2058 reshape
.after
.data_disks
)
2061 c
= map_num(pers
, info
->new_level
);
2065 exit(0);/* not possible */
2068 err
= sysfs_set_str(sra
, NULL
, "level", c
);
2070 fprintf(stderr
, Name\
2071 ": %s: could not set level"
2072 "to %s\n", devname
, c
);
2079 fprintf(stderr
, Name
": Cannot run child to monitor reshape: %s\n",
2085 /* The child will take care of unfreezing the array */
2093 unfreeze(st
, frozen
);
2097 ping_monitor(container
);
2098 if (st
->ss
->external
) {
2099 /* Re-load the metadata as much could have changed */
2100 int cfd
= open_dev(st
->container_dev
);
2102 st
->ss
->free_super(st
);
2103 st
->ss
->load_container(st
, cfd
, container
);
2107 if (rv
&& orig_level
!= UnSet
&& sra
) {
2108 c
= map_num(pers
, orig_level
);
2109 if (c
&& sysfs_set_str(sra
, NULL
, "level", c
) == 0)
2110 fprintf(stderr
, Name
": aborting level change\n");
2112 unfreeze(st
, frozen
);
2116 int reshape_container(char *container
, int cfd
, char *devname
,
2117 struct supertype
*st
,
2118 struct mdinfo
*info
,
2123 struct mdinfo
*cc
= NULL
;
2125 if (reshape_super(st
, info
->component_size
, info
->new_level
,
2126 info
->new_layout
, info
->new_chunk
,
2127 info
->array
.raid_disks
+ info
->delta_disks
,
2128 backup_file
, devname
, quiet
))
2133 /* ping monitor to be sure that update is on disk
2135 ping_monitor(container
);
2138 case -1: /* error */
2139 perror("Cannot fork to complete reshape\n");
2141 default: /* parent */
2142 printf(Name
": multi-array reshape continues in background\n");
2149 /* For each member array with reshape_active,
2150 * we need to perform the reshape.
2151 * We pick the first array that needs reshaping and
2152 * reshape it. reshape_array() will re-read the metadata
2153 * so the next time through a different array should be
2154 * ready for reshape.
2156 struct mdinfo
*content
;
2159 struct mdstat_ent
*mdstat
;
2164 cc
= st
->ss
->container_content(st
, NULL
);
2166 for (content
= cc
; content
; content
= content
->next
) {
2168 if (!content
->reshape_active
)
2171 subarray
= strchr(content
->text_version
+1, '/')+1;
2172 mdstat
= mdstat_by_subdev(subarray
,
2173 devname2devnum(container
));
2181 fd
= open_dev_excl(mdstat
->devnum
);
2184 adev
= map_dev(dev2major(mdstat
->devnum
),
2185 dev2minor(mdstat
->devnum
),
2188 adev
= content
->text_version
;
2190 sysfs_init(content
, fd
, mdstat
->devnum
);
2192 rv
= reshape_array(container
, fd
, adev
, st
,
2194 backup_file
, quiet
, 1);
2204 * We run a child process in the background which performs the following
2206 * - wait for resync to reach a certain point
2207 * - suspend io to the following section
2208 * - backup that section
2209 * - allow resync to proceed further
2211 * - discard the backup.
2213 * When are combined in slightly different ways in the three cases.
2215 * - suspend/backup/allow/wait/resume/discard
2217 * - allow/wait/suspend/backup/allow/wait/resume/discard
2219 * - wait/resume/discard/suspend/backup/allow
2221 * suspend/backup/allow always come together
2222 * wait/resume/discard do too.
2223 * For the same-size case we have two backups to improve flow.
2227 int progress_reshape(struct mdinfo
*info
, struct reshape
*reshape
,
2228 unsigned long long backup_point
,
2229 unsigned long long wait_point
,
2230 unsigned long long *suspend_point
,
2231 unsigned long long *reshape_completed
)
2233 /* This function is called repeatedly by the reshape manager.
2234 * It determines how much progress can safely be made and allows
2236 * - 'info' identifies the array and particularly records in
2237 * ->reshape_progress the metadata's knowledge of progress
2238 * This is a sector offset from the start of the array
2239 * of the next array block to be relocated. This number
2240 * may increase from 0 or decrease from array_size, depending
2241 * on the type of reshape that is happening.
2242 * Note that in contrast, 'sync_completed' is a block count of the
2243 * reshape so far. It gives the distance between the start point
2244 * (head or tail of device) and the next place that data will be
2245 * written. It always increases.
2246 * - 'reshape' is the structure created by analyse_change
2247 * - 'backup_point' shows how much the metadata manager has backed-up
2248 * data. For reshapes with increasing progress, it is the next address
2249 * to be backed up, previous addresses have been backed-up. For
2250 * decreasing progress, it is the earliest address that has been
2251 * backed up - later address are also backed up.
2252 * So addresses between reshape_progress and backup_point are
2253 * backed up providing those are in the 'correct' order.
2254 * - 'wait_point' is an array address. When reshape_completed
2255 * passes this point, progress_reshape should return. It might
2256 * return earlier if it determines that ->reshape_progress needs
2257 * to be updated or further backup is needed.
2258 * - suspend_point is maintained by progress_reshape and the caller
2259 * should not touch it except to initialise to zero.
2260 * It is an array address and it only increases in 2.6.37 and earlier.
2261 * This makes it difficulty to handle reducing reshapes with
2262 * external metadata.
2263 * However: it is similar to backup_point in that it records the
2264 * other end of a suspended region from reshape_progress.
2265 * it is moved to extend the region that is safe to backup and/or
2267 * - reshape_completed is read from sysfs and returned. The caller
2268 * should copy this into ->reshape_progress when it has reason to
2269 * believe that the metadata knows this, and any backup outside this
2273 * 1 if more data from backup_point - but only as far as suspend_point,
2274 * should be backed up
2275 * 0 if things are progressing smoothly
2276 * -1 if the reshape is finished, either because it is all done,
2277 * or due to an error.
2280 int advancing
= (reshape
->after
.data_disks
2281 >= reshape
->before
.data_disks
);
2282 int need_backup
= (reshape
->after
.data_disks
2283 == reshape
->before
.data_disks
);
2284 unsigned long long read_offset
, write_offset
;
2285 unsigned long long read_range
, write_range
;
2286 unsigned long long max_progress
, target
, completed
;
2289 /* First, we unsuspend any region that is now known to be safe.
2290 * If suspend_point is on the 'wrong' side of reshape_progress, then
2291 * we don't have or need suspension at the moment. This is true for
2292 * native metadata when we don't need to back-up.
2295 if (info
->reshape_progress
< *suspend_point
)
2296 sysfs_set_num(info
, NULL
, "suspend_lo",
2297 info
->reshape_progress
);
2299 /* Note: this won't work in 2.6.37 and before.
2300 * Something somewhere should make sure we don't need it!
2302 if (info
->reshape_progress
> *suspend_point
)
2303 sysfs_set_num(info
, NULL
, "suspend_hi",
2304 info
->reshape_progress
);
2307 /* Now work out how far it is safe to progress.
2308 * If the read_offset for ->reshape_progress is less than
2309 * 'blocks' beyond the write_offset, we can only progress as far
2311 * Otherwise we can progress until the write_offset for the new location
2312 * reaches (within 'blocks' of) the read_offset at the current location.
2313 * However that region must be suspended unless we are using native
2315 * If we need to suspend more, we limit it to 128M per device, which is
2316 * rather arbitrary and should be some time-based calculation.
2318 write_offset
= info
->reshape_progress
/ reshape
->before
.data_disks
;
2319 read_offset
= info
->reshape_progress
/ reshape
->after
.data_disks
;
2320 write_range
= reshape
->blocks
/ reshape
->before
.data_disks
;
2321 read_range
= reshape
->blocks
/ reshape
->after
.data_disks
;
2323 if (read_offset
< write_offset
+ write_range
) {
2324 max_progress
= backup_point
;
2325 if (max_progress
<= info
->reshape_progress
)
2329 (read_offset
- write_range
) *
2330 reshape
->before
.data_disks
;
2333 if (read_offset
> write_offset
- write_range
) {
2334 max_progress
= backup_point
;
2335 if (max_progress
>= info
->reshape_progress
)
2339 (read_offset
+ write_range
) *
2340 reshape
->before
.data_disks
;
2341 /* If we are using internal metadata, then we can
2342 * progress all the way to the suspend_point without
2343 * worrying about backing-up/suspending along the
2346 if (max_progress
< *suspend_point
&&
2347 info
->array
.major_version
>= 0)
2348 max_progress
= *suspend_point
;
2352 /* We know it is safe to progress to 'max_progress' providing
2353 * it is suspended or we are using native metadata.
2354 * Consider extending suspend_point 128M per device if it
2355 * is less than 64M per device beyond reshape_progress.
2356 * But always do a multiple of 'blocks'
2358 target
= 64*1024*2 * min(reshape
->before
.data_disks
,
2359 reshape
->after
.data_disks
);
2360 target
/= reshape
->blocks
;
2363 target
*= reshape
->blocks
;
2365 /* For externally managed metadata we always need to suspend IO to
2366 * the area being reshaped so we regularly push suspend_point forward.
2367 * For native metadata we only need the suspend if we are going to do
2371 if ((need_backup
|| info
->array
.major_version
< 0) &&
2372 *suspend_point
< info
->reshape_progress
+ target
) {
2373 if (max_progress
< *suspend_point
+ 2 * target
)
2374 *suspend_point
= max_progress
;
2376 *suspend_point
+= 2 * target
;
2377 sysfs_set_num(info
, NULL
, "suspend_hi", *suspend_point
);
2378 max_progress
= *suspend_point
;
2381 if ((need_backup
|| info
->array
.major_version
< 0) &&
2382 *suspend_point
> info
->reshape_progress
- target
) {
2383 if (max_progress
> *suspend_point
- 2 * target
)
2384 *suspend_point
= max_progress
;
2386 *suspend_point
-= 2 * target
;
2387 sysfs_set_num(info
, NULL
, "suspend_lo", *suspend_point
);
2388 max_progress
= *suspend_point
;
2392 /* now set sync_max to allow that progress. sync_max, like
2393 * sync_completed is a count of sectors written per device, so
2394 * we find the difference between max_progress and the start point,
2395 * and divide that by after.data_disks to get a sync_max
2397 * At the same time we convert wait_point to a similar number
2398 * for comparing against sync_completed.
2401 max_progress
= info
->component_size
* reshape
->after
.data_disks
2403 wait_point
= info
->component_size
* reshape
->after
.data_disks
2406 max_progress
/= reshape
->after
.data_disks
;
2407 wait_point
/= reshape
->after
.data_disks
;
2409 sysfs_set_num(info
, NULL
, "sync_max", max_progress
);
2411 /* Now wait. If we have already reached the point that we were
2412 * asked to wait to, don't wait at all, else wait for any change.
2413 * We need to select on 'sync_completed' as that is the place that
2414 * notifications happen, but we are really interested in
2415 * 'reshape_position'
2417 fd
= sysfs_get_fd(info
, NULL
, "sync_completed");
2421 if (sysfs_fd_get_ll(fd
, &completed
) < 0) {
2425 while (completed
< max_progress
&& completed
< wait_point
) {
2426 /* Check that sync_action is still 'reshape' to avoid
2427 * waiting forever on a dead array
2431 if (sysfs_get_str(info
, NULL
, "sync_action",
2433 strncmp(action
, "reshape", 7) != 0)
2437 select(fd
+1, NULL
, NULL
, &rfds
, NULL
);
2438 if (sysfs_fd_get_ll(fd
, &completed
) < 0) {
2443 /* Convert 'completed' back in to a 'progress' number */
2444 completed
*= reshape
->after
.data_disks
;
2446 completed
= info
->component_size
* reshape
->after
.data_disks
2449 *reshape_completed
= completed
;
2453 /* We return the need_backup flag. Caller will decide
2454 * how much (a multiple of ->blocks) and will adjust
2455 * suspend_{lo,hi} and suspend_point.
2461 /* FIXME return status is never checked */
2462 static int grow_backup(struct mdinfo
*sra
,
2463 unsigned long long offset
, /* per device */
2464 unsigned long stripes
, /* per device, in old chunks */
2465 int *sources
, unsigned long long *offsets
,
2466 int disks
, int chunk
, int level
, int layout
,
2467 int dests
, int *destfd
, unsigned long long *destoffsets
,
2468 int part
, int *degraded
,
2471 /* Backup 'blocks' sectors at 'offset' on each device of the array,
2472 * to storage 'destfd' (offset 'destoffsets'), after first
2473 * suspending IO. Then allow resync to continue
2474 * over the suspended section.
2475 * Use part 'part' of the backup-super-block.
2480 unsigned long long ll
;
2482 //printf("offset %llu\n", offset);
2488 /* Check that array hasn't become degraded, else we might backup the wrong data */
2489 sysfs_get_ll(sra
, NULL
, "degraded", &ll
);
2490 new_degraded
= (int)ll
;
2491 if (new_degraded
!= *degraded
) {
2492 /* check each device to ensure it is still working */
2494 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
2495 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
2497 if (sd
->disk
.state
& (1<<MD_DISK_SYNC
)) {
2499 if (sysfs_get_str(sra
, sd
, "state", sbuf
, 20) < 0 ||
2500 strstr(sbuf
, "faulty") ||
2501 strstr(sbuf
, "in_sync") == NULL
) {
2502 /* this device is dead */
2503 sd
->disk
.state
= (1<<MD_DISK_FAULTY
);
2504 if (sd
->disk
.raid_disk
>= 0 &&
2505 sources
[sd
->disk
.raid_disk
] >= 0) {
2506 close(sources
[sd
->disk
.raid_disk
]);
2507 sources
[sd
->disk
.raid_disk
] = -1;
2512 *degraded
= new_degraded
;
2515 bsb
.arraystart2
= __cpu_to_le64(offset
* odata
);
2516 bsb
.length2
= __cpu_to_le64(stripes
* (chunk
/512) * odata
);
2518 bsb
.arraystart
= __cpu_to_le64(offset
* odata
);
2519 bsb
.length
= __cpu_to_le64(stripes
* (chunk
/512) * odata
);
2522 bsb
.magic
[15] = '2';
2523 for (i
= 0; i
< dests
; i
++)
2525 lseek64(destfd
[i
], destoffsets
[i
] + __le64_to_cpu(bsb
.devstart2
)*512, 0);
2527 lseek64(destfd
[i
], destoffsets
[i
], 0);
2529 rv
= save_stripes(sources
, offsets
,
2530 disks
, chunk
, level
, layout
,
2532 offset
*512*odata
, stripes
* chunk
* odata
,
2537 bsb
.mtime
= __cpu_to_le64(time(0));
2538 for (i
= 0; i
< dests
; i
++) {
2539 bsb
.devstart
= __cpu_to_le64(destoffsets
[i
]/512);
2541 bsb
.sb_csum
= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
));
2542 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0)
2543 bsb
.sb_csum2
= bsb_csum((char*)&bsb
,
2544 ((char*)&bsb
.sb_csum2
)-((char*)&bsb
));
2547 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
] - 4096, 0)
2548 != destoffsets
[i
] - 4096)
2550 if (write(destfd
[i
], &bsb
, 512) != 512)
2552 if (destoffsets
[i
] > 4096) {
2553 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
]+stripes
*chunk
*odata
, 0) !=
2554 destoffsets
[i
]+stripes
*chunk
*odata
)
2556 if (write(destfd
[i
], &bsb
, 512) != 512)
2566 /* in 2.6.30, the value reported by sync_completed can be
2567 * less that it should be by one stripe.
2568 * This only happens when reshape hits sync_max and pauses.
2569 * So allow wait_backup to either extent sync_max further
2570 * than strictly necessary, or return before the
2571 * sync has got quite as far as we would really like.
2572 * This is what 'blocks2' is for.
2573 * The various caller give appropriate values so that
2576 /* FIXME return value is often ignored */
2577 static int forget_backup(
2578 int dests
, int *destfd
, unsigned long long *destoffsets
,
2582 * Erase backup 'part' (which is 0 or 1)
2588 bsb
.arraystart2
= __cpu_to_le64(0);
2589 bsb
.length2
= __cpu_to_le64(0);
2591 bsb
.arraystart
= __cpu_to_le64(0);
2592 bsb
.length
= __cpu_to_le64(0);
2594 bsb
.mtime
= __cpu_to_le64(time(0));
2596 for (i
= 0; i
< dests
; i
++) {
2597 bsb
.devstart
= __cpu_to_le64(destoffsets
[i
]/512);
2598 bsb
.sb_csum
= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
));
2599 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0)
2600 bsb
.sb_csum2
= bsb_csum((char*)&bsb
,
2601 ((char*)&bsb
.sb_csum2
)-((char*)&bsb
));
2602 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
]-4096, 0) !=
2603 destoffsets
[i
]-4096)
2606 write(destfd
[i
], &bsb
, 512) != 512)
2613 static void fail(char *msg
)
2616 rv
= (write(2, msg
, strlen(msg
)) != (int)strlen(msg
));
2617 rv
|= (write(2, "\n", 1) != 1);
2621 static char *abuf
, *bbuf
;
2622 static unsigned long long abuflen
;
2623 static void validate(int afd
, int bfd
, unsigned long long offset
)
2625 /* check that the data in the backup against the array.
2626 * This is only used for regression testing and should not
2627 * be used while the array is active
2631 lseek64(bfd
, offset
- 4096, 0);
2632 if (read(bfd
, &bsb2
, 512) != 512)
2633 fail("cannot read bsb");
2634 if (bsb2
.sb_csum
!= bsb_csum((char*)&bsb2
,
2635 ((char*)&bsb2
.sb_csum
)-((char*)&bsb2
)))
2636 fail("first csum bad");
2637 if (memcmp(bsb2
.magic
, "md_backup_data", 14) != 0)
2638 fail("magic is bad");
2639 if (memcmp(bsb2
.magic
, "md_backup_data-2", 16) == 0 &&
2640 bsb2
.sb_csum2
!= bsb_csum((char*)&bsb2
,
2641 ((char*)&bsb2
.sb_csum2
)-((char*)&bsb2
)))
2642 fail("second csum bad");
2644 if (__le64_to_cpu(bsb2
.devstart
)*512 != offset
)
2645 fail("devstart is wrong");
2648 unsigned long long len
= __le64_to_cpu(bsb2
.length
)*512;
2650 if (abuflen
< len
) {
2654 if (posix_memalign((void**)&abuf
, 4096, abuflen
) ||
2655 posix_memalign((void**)&bbuf
, 4096, abuflen
)) {
2657 /* just stop validating on mem-alloc failure */
2662 lseek64(bfd
, offset
, 0);
2663 if ((unsigned long long)read(bfd
, bbuf
, len
) != len
) {
2664 //printf("len %llu\n", len);
2665 fail("read first backup failed");
2667 lseek64(afd
, __le64_to_cpu(bsb2
.arraystart
)*512, 0);
2668 if ((unsigned long long)read(afd
, abuf
, len
) != len
)
2669 fail("read first from array failed");
2670 if (memcmp(bbuf
, abuf
, len
) != 0) {
2673 printf("offset=%llu len=%llu\n",
2674 (unsigned long long)__le64_to_cpu(bsb2
.arraystart
)*512, len
);
2675 for (i
=0; i
<len
; i
++)
2676 if (bbuf
[i
] != abuf
[i
]) {
2677 printf("first diff byte %d\n", i
);
2681 fail("data1 compare failed");
2685 unsigned long long len
= __le64_to_cpu(bsb2
.length2
)*512;
2687 if (abuflen
< len
) {
2691 abuf
= malloc(abuflen
);
2692 bbuf
= malloc(abuflen
);
2695 lseek64(bfd
, offset
+__le64_to_cpu(bsb2
.devstart2
)*512, 0);
2696 if ((unsigned long long)read(bfd
, bbuf
, len
) != len
)
2697 fail("read second backup failed");
2698 lseek64(afd
, __le64_to_cpu(bsb2
.arraystart2
)*512, 0);
2699 if ((unsigned long long)read(afd
, abuf
, len
) != len
)
2700 fail("read second from array failed");
2701 if (memcmp(bbuf
, abuf
, len
) != 0)
2702 fail("data2 compare failed");
2706 static int child_monitor(int afd
, struct mdinfo
*sra
, struct reshape
*reshape
,
2707 unsigned long stripes
,
2708 int *fds
, unsigned long long *offsets
,
2709 int dests
, int *destfd
, unsigned long long *destoffsets
)
2711 /* Monitor a reshape where backup is being performed using
2712 * 'native' mechanism - either to a backup file, or
2713 * to some space in a spare.
2717 unsigned long long speed
;
2718 unsigned long long suspend_point
, array_size
;
2719 unsigned long long backup_point
, wait_point
;
2720 unsigned long long reshape_completed
;
2722 int increasing
= reshape
->after
.data_disks
>= reshape
->before
.data_disks
;
2723 int part
= 0; /* The next part of the backup area to fill. It may already
2724 * be full, so we need to check */
2725 int level
= reshape
->level
;
2726 int layout
= reshape
->before
.layout
;
2727 int data
= reshape
->before
.data_disks
;
2728 int disks
= reshape
->before
.data_disks
+ reshape
->parity
;
2729 int chunk
= sra
->array
.chunk_size
;
2731 if (posix_memalign((void**)&buf
, 4096, disks
* chunk
))
2732 /* Don't start the 'reshape' */
2734 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
) {
2735 sysfs_get_ll(sra
, NULL
, "sync_speed_min", &speed
);
2736 sysfs_set_num(sra
, NULL
, "sync_speed_min", 200000);
2739 array_size
= sra
->component_size
* data
;
2741 backup_point
= sra
->reshape_progress
;
2744 backup_point
= array_size
;
2745 suspend_point
= array_size
;
2751 /* Want to return as soon the oldest backup slot can
2752 * be released as that allows us to start backing up
2753 * some more, providing suspend_point has been
2754 * advanced, which it should have
2757 wait_point
= array_size
;
2758 if (part
== 0 && __le64_to_cpu(bsb
.length
) > 0)
2759 wait_point
= (__le64_to_cpu(bsb
.arraystart
) +
2760 __le64_to_cpu(bsb
.length
));
2761 if (part
== 1 && __le64_to_cpu(bsb
.length2
) > 0)
2762 wait_point
= (__le64_to_cpu(bsb
.arraystart2
) +
2763 __le64_to_cpu(bsb
.length2
));
2766 if (part
== 0 && __le64_to_cpu(bsb
.length
) > 0)
2767 wait_point
= __le64_to_cpu(bsb
.arraystart
);
2768 if (part
== 1 && __le64_to_cpu(bsb
.length2
) > 0)
2769 wait_point
= __le64_to_cpu(bsb
.arraystart2
);
2772 rv
= progress_reshape(sra
, reshape
,
2773 backup_point
, wait_point
,
2774 &suspend_point
, &reshape_completed
);
2780 /* external metadata would need to ping_monitor here */
2781 sra
->reshape_progress
= reshape_completed
;
2783 /* Clear any backup region that is before 'here' */
2785 if (reshape_completed
>= (__le64_to_cpu(bsb
.arraystart
) +
2786 __le64_to_cpu(bsb
.length
)))
2787 forget_backup(dests
, destfd
,
2789 if (reshape_completed
>= (__le64_to_cpu(bsb
.arraystart2
) +
2790 __le64_to_cpu(bsb
.length2
)))
2791 forget_backup(dests
, destfd
,
2794 if (reshape_completed
<= (__le64_to_cpu(bsb
.arraystart
)))
2795 forget_backup(dests
, destfd
,
2797 if (reshape_completed
<= (__le64_to_cpu(bsb
.arraystart2
)))
2798 forget_backup(dests
, destfd
,
2803 unsigned long long offset
;
2804 /* need to backup some space... */
2805 /* Check that 'part' is unused */
2806 if (part
== 0 && __le64_to_cpu(bsb
.length
) != 0)
2807 abort(); /* BUG here */
2808 if (part
== 1 && __le64_to_cpu(bsb
.length2
) != 0)
2811 offset
= backup_point
/ data
;
2813 offset
-= stripes
* (chunk
/512);
2814 grow_backup(sra
, offset
, stripes
,
2816 disks
, chunk
, level
, layout
,
2817 dests
, destfd
, destoffsets
,
2818 part
, °raded
, buf
);
2819 validate(afd
, destfd
[0], destoffsets
[0]);
2820 /* record where 'part' is up to */
2823 backup_point
+= stripes
* (chunk
/512) * data
;
2825 backup_point
-= stripes
* (chunk
/512) * data
;
2829 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
)
2830 sysfs_set_num(sra
, NULL
, "sync_speed_min", speed
);
2832 return 1; /* FIXME what does this mean? */
2836 * If any spare contains md_back_data-1 which is recent wrt mtime,
2837 * write that data into the array and update the super blocks with
2838 * the new reshape_progress
2840 int Grow_restart(struct supertype
*st
, struct mdinfo
*info
, int *fdlist
, int cnt
,
2841 char *backup_file
, int verbose
)
2845 unsigned long long *offsets
;
2846 unsigned long long nstripe
, ostripe
;
2849 if (info
->new_level
!= info
->array
.level
)
2850 return 1; /* Cannot handle level changes (they are instantaneous) */
2852 odata
= info
->array
.raid_disks
- info
->delta_disks
- 1;
2853 if (info
->array
.level
== 6) odata
--; /* number of data disks */
2854 ndata
= info
->array
.raid_disks
- 1;
2855 if (info
->new_level
== 6) ndata
--;
2857 old_disks
= info
->array
.raid_disks
- info
->delta_disks
;
2859 if (info
->delta_disks
<= 0)
2860 /* Didn't grow, so the backup file must have
2864 for (i
=old_disks
-(backup_file
?1:0); i
<cnt
; i
++) {
2865 struct mdinfo dinfo
;
2868 char *devname
, namebuf
[20];
2870 /* This was a spare and may have some saved data on it.
2871 * Load the superblock, find and load the
2872 * backup_super_block.
2873 * If either fail, go on to next device.
2874 * If the backup contains no new info, just return
2875 * else restore data and update all superblocks
2877 if (i
== old_disks
-1) {
2878 fd
= open(backup_file
, O_RDONLY
);
2880 fprintf(stderr
, Name
": backup file %s inaccessible: %s\n",
2881 backup_file
, strerror(errno
));
2884 devname
= backup_file
;
2889 if (st
->ss
->load_super(st
, fd
, NULL
))
2892 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
2893 st
->ss
->free_super(st
);
2896 (dinfo
.data_offset
+ dinfo
.component_size
- 8) <<9,
2898 fprintf(stderr
, Name
": Cannot seek on device %d\n", i
);
2899 continue; /* Cannot seek */
2901 sprintf(namebuf
, "device-%d", i
);
2904 if (read(fd
, &bsb
, sizeof(bsb
)) != sizeof(bsb
)) {
2906 fprintf(stderr
, Name
": Cannot read from %s\n", devname
);
2907 continue; /* Cannot read */
2909 if (memcmp(bsb
.magic
, "md_backup_data-1", 16) != 0 &&
2910 memcmp(bsb
.magic
, "md_backup_data-2", 16) != 0) {
2912 fprintf(stderr
, Name
": No backup metadata on %s\n", devname
);
2915 if (bsb
.sb_csum
!= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
))) {
2917 fprintf(stderr
, Name
": Bad backup-metadata checksum on %s\n", devname
);
2918 continue; /* bad checksum */
2920 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0 &&
2921 bsb
.sb_csum2
!= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum2
)-((char*)&bsb
))) {
2923 fprintf(stderr
, Name
": Bad backup-metadata checksum2 on %s\n", devname
);
2924 continue; /* Bad second checksum */
2926 if (memcmp(bsb
.set_uuid
,info
->uuid
, 16) != 0) {
2928 fprintf(stderr
, Name
": Wrong uuid on backup-metadata on %s\n", devname
);
2929 continue; /* Wrong uuid */
2932 /* array utime and backup-mtime should be updated at much the same time, but it seems that
2933 * sometimes they aren't... So allow considerable flexability in matching, and allow
2934 * this test to be overridden by an environment variable.
2936 if (info
->array
.utime
> (int)__le64_to_cpu(bsb
.mtime
) + 2*60*60 ||
2937 info
->array
.utime
< (int)__le64_to_cpu(bsb
.mtime
) - 10*60) {
2938 if (check_env("MDADM_GROW_ALLOW_OLD")) {
2939 fprintf(stderr
, Name
": accepting backup with timestamp %lu "
2940 "for array with timestamp %lu\n",
2941 (unsigned long)__le64_to_cpu(bsb
.mtime
),
2942 (unsigned long)info
->array
.utime
);
2945 fprintf(stderr
, Name
": too-old timestamp on "
2946 "backup-metadata on %s\n", devname
);
2947 continue; /* time stamp is too bad */
2951 if (bsb
.magic
[15] == '1') {
2952 if (info
->delta_disks
>= 0) {
2953 /* reshape_progress is increasing */
2954 if (__le64_to_cpu(bsb
.arraystart
) + __le64_to_cpu(bsb
.length
) <
2955 info
->reshape_progress
) {
2958 fprintf(stderr
, Name
": backup-metadata found on %s but is not needed\n", devname
);
2959 continue; /* No new data here */
2962 /* reshape_progress is decreasing */
2963 if (__le64_to_cpu(bsb
.arraystart
) >=
2964 info
->reshape_progress
)
2965 goto nonew
; /* No new data here */
2968 if (info
->delta_disks
>= 0) {
2969 /* reshape_progress is increasing */
2970 if (__le64_to_cpu(bsb
.arraystart
) + __le64_to_cpu(bsb
.length
) <
2971 info
->reshape_progress
&&
2972 __le64_to_cpu(bsb
.arraystart2
) + __le64_to_cpu(bsb
.length2
) <
2973 info
->reshape_progress
)
2974 goto nonew
; /* No new data here */
2976 /* reshape_progress is decreasing */
2977 if (__le64_to_cpu(bsb
.arraystart
) >=
2978 info
->reshape_progress
&&
2979 __le64_to_cpu(bsb
.arraystart2
) >=
2980 info
->reshape_progress
)
2981 goto nonew
; /* No new data here */
2984 if (lseek64(fd
, __le64_to_cpu(bsb
.devstart
)*512, 0)< 0) {
2987 fprintf(stderr
, Name
": Failed to verify secondary backup-metadata block on %s\n",
2989 continue; /* Cannot seek */
2991 /* There should be a duplicate backup superblock 4k before here */
2992 if (lseek64(fd
, -4096, 1) < 0 ||
2993 read(fd
, &bsb2
, sizeof(bsb2
)) != sizeof(bsb2
))
2994 goto second_fail
; /* Cannot find leading superblock */
2995 if (bsb
.magic
[15] == '1')
2996 bsbsize
= offsetof(struct mdp_backup_super
, pad1
);
2998 bsbsize
= offsetof(struct mdp_backup_super
, pad
);
2999 if (memcmp(&bsb2
, &bsb
, bsbsize
) != 0)
3000 goto second_fail
; /* Cannot find leading superblock */
3002 /* Now need the data offsets for all devices. */
3003 offsets
= malloc(sizeof(*offsets
)*info
->array
.raid_disks
);
3004 for(j
=0; j
<info
->array
.raid_disks
; j
++) {
3007 if (st
->ss
->load_super(st
, fdlist
[j
], NULL
))
3008 /* FIXME should be this be an error */
3010 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
3011 st
->ss
->free_super(st
);
3012 offsets
[j
] = dinfo
.data_offset
* 512;
3014 printf(Name
": restoring critical section\n");
3016 if (restore_stripes(fdlist
, offsets
,
3017 info
->array
.raid_disks
,
3021 fd
, __le64_to_cpu(bsb
.devstart
)*512,
3022 __le64_to_cpu(bsb
.arraystart
)*512,
3023 __le64_to_cpu(bsb
.length
)*512)) {
3024 /* didn't succeed, so giveup */
3026 fprintf(stderr
, Name
": Error restoring backup from %s\n",
3031 if (bsb
.magic
[15] == '2' &&
3032 restore_stripes(fdlist
, offsets
,
3033 info
->array
.raid_disks
,
3037 fd
, __le64_to_cpu(bsb
.devstart
)*512 +
3038 __le64_to_cpu(bsb
.devstart2
)*512,
3039 __le64_to_cpu(bsb
.arraystart2
)*512,
3040 __le64_to_cpu(bsb
.length2
)*512)) {
3041 /* didn't succeed, so giveup */
3043 fprintf(stderr
, Name
": Error restoring second backup from %s\n",
3049 /* Ok, so the data is restored. Let's update those superblocks. */
3051 if (info
->delta_disks
>= 0) {
3052 info
->reshape_progress
= __le64_to_cpu(bsb
.arraystart
) +
3053 __le64_to_cpu(bsb
.length
);
3054 if (bsb
.magic
[15] == '2') {
3055 unsigned long long p2
= __le64_to_cpu(bsb
.arraystart2
) +
3056 __le64_to_cpu(bsb
.length2
);
3057 if (p2
> info
->reshape_progress
)
3058 info
->reshape_progress
= p2
;
3061 info
->reshape_progress
= __le64_to_cpu(bsb
.arraystart
);
3062 if (bsb
.magic
[15] == '2') {
3063 unsigned long long p2
= __le64_to_cpu(bsb
.arraystart2
);
3064 if (p2
< info
->reshape_progress
)
3065 info
->reshape_progress
= p2
;
3068 for (j
=0; j
<info
->array
.raid_disks
; j
++) {
3069 if (fdlist
[j
] < 0) continue;
3070 if (st
->ss
->load_super(st
, fdlist
[j
], NULL
))
3072 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
3073 dinfo
.reshape_progress
= info
->reshape_progress
;
3074 st
->ss
->update_super(st
, &dinfo
,
3075 "_reshape_progress",
3077 st
->ss
->store_super(st
, fdlist
[j
]);
3078 st
->ss
->free_super(st
);
3082 /* Didn't find any backup data, try to see if any
3085 if (info
->delta_disks
< 0) {
3086 /* When shrinking, the critical section is at the end.
3087 * So see if we are before the critical section.
3089 unsigned long long first_block
;
3090 nstripe
= ostripe
= 0;
3092 while (ostripe
>= nstripe
) {
3093 ostripe
+= info
->array
.chunk_size
/ 512;
3094 first_block
= ostripe
* odata
;
3095 nstripe
= first_block
/ ndata
/ (info
->new_chunk
/512) *
3096 (info
->new_chunk
/512);
3099 if (info
->reshape_progress
>= first_block
)
3102 if (info
->delta_disks
> 0) {
3103 /* See if we are beyond the critical section. */
3104 unsigned long long last_block
;
3105 nstripe
= ostripe
= 0;
3107 while (nstripe
>= ostripe
) {
3108 nstripe
+= info
->new_chunk
/ 512;
3109 last_block
= nstripe
* ndata
;
3110 ostripe
= last_block
/ odata
/ (info
->array
.chunk_size
/512) *
3111 (info
->array
.chunk_size
/512);
3114 if (info
->reshape_progress
>= last_block
)
3117 /* needed to recover critical section! */
3119 fprintf(stderr
, Name
": Failed to find backup of critical section\n");
3123 int Grow_continue(int mdfd
, struct supertype
*st
, struct mdinfo
*info
,
3126 int err
= sysfs_set_str(info
, NULL
, "array_state", "readonly");
3129 return reshape_array(NULL
, mdfd
, "array", st
, info
, 1, backup_file
, 0, 0);