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
)
535 if (st
->ss
->external
)
536 return unfreeze_container(st
);
538 struct mdinfo
*sra
= sysfs_read(-1, st
->devnum
, GET_VERSION
);
541 sysfs_set_str(sra
, NULL
, "sync_action", "idle");
543 fprintf(stderr
, Name
": failed to unfreeze array\n");
548 static void wait_reshape(struct mdinfo
*sra
)
550 int fd
= sysfs_get_fd(sra
, NULL
, "sync_action");
556 while (sysfs_fd_get_str(fd
, action
, 20) > 0 &&
557 strncmp(action
, "reshape", 7) == 0) {
561 select(fd
+1, NULL
, NULL
, &rfds
, NULL
);
566 static int reshape_super(struct supertype
*st
, long long size
, int level
,
567 int layout
, int chunksize
, int raid_disks
,
568 char *backup_file
, char *dev
, int verbose
)
570 /* nothing extra to check in the native case */
571 if (!st
->ss
->external
)
573 if (!st
->ss
->reshape_super
||
574 !st
->ss
->manage_reshape
) {
575 fprintf(stderr
, Name
": %s metadata does not support reshape\n",
580 return st
->ss
->reshape_super(st
, size
, level
, layout
, chunksize
,
581 raid_disks
, backup_file
, dev
, verbose
);
584 static void sync_metadata(struct supertype
*st
)
586 if (st
->ss
->external
) {
587 if (st
->update_tail
) {
588 flush_metadata_updates(st
);
589 st
->update_tail
= &st
->updates
;
591 st
->ss
->sync_metadata(st
);
595 static int subarray_set_num(char *container
, struct mdinfo
*sra
, char *name
, int n
)
597 /* when dealing with external metadata subarrays we need to be
598 * prepared to handle EAGAIN. The kernel may need to wait for
599 * mdmon to mark the array active so the kernel can handle
600 * allocations/writeback when preparing the reshape action
601 * (md_allow_write()). We temporarily disable safe_mode_delay
602 * to close a race with the array_state going clean before the
603 * next write to raid_disks / stripe_cache_size
608 /* only 'raid_disks' and 'stripe_cache_size' trigger md_allow_write */
610 (strcmp(name
, "raid_disks") != 0 &&
611 strcmp(name
, "stripe_cache_size") != 0))
612 return sysfs_set_num(sra
, NULL
, name
, n
);
614 rc
= sysfs_get_str(sra
, NULL
, "safe_mode_delay", safe
, sizeof(safe
));
617 sysfs_set_num(sra
, NULL
, "safe_mode_delay", 0);
618 rc
= sysfs_set_num(sra
, NULL
, name
, n
);
619 if (rc
< 0 && errno
== EAGAIN
) {
620 ping_monitor(container
);
621 /* if we get EAGAIN here then the monitor is not active
624 rc
= sysfs_set_num(sra
, NULL
, name
, n
);
626 sysfs_set_str(sra
, NULL
, "safe_mode_delay", safe
);
630 int start_reshape(struct mdinfo
*sra
)
633 sysfs_set_num(sra
, NULL
, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL
);
634 err
= sysfs_set_num(sra
, NULL
, "suspend_hi", 0);
635 err
= err
?: sysfs_set_num(sra
, NULL
, "suspend_lo", 0);
636 /* Setting sync_min can fail if the recovery is already 'running',
637 * which can happen when restarting an array which is reshaping.
638 * So don't worry about errors here */
639 sysfs_set_num(sra
, NULL
, "sync_min", 0);
640 err
= err
?: sysfs_set_num(sra
, NULL
, "sync_max", 0);
641 err
= err
?: sysfs_set_str(sra
, NULL
, "sync_action", "reshape");
646 void abort_reshape(struct mdinfo
*sra
)
648 sysfs_set_str(sra
, NULL
, "sync_action", "idle");
649 sysfs_set_num(sra
, NULL
, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL
);
650 sysfs_set_num(sra
, NULL
, "suspend_hi", 0);
651 sysfs_set_num(sra
, NULL
, "suspend_lo", 0);
652 sysfs_set_num(sra
, NULL
, "sync_min", 0);
653 sysfs_set_str(sra
, NULL
, "sync_max", "max");
656 int remove_disks_on_raid10_to_raid0_takeover(struct supertype
*st
,
661 struct mdinfo
*remaining
;
664 nr_of_copies
= layout
& 0xff;
666 remaining
= sra
->devs
;
668 /* for each 'copy', select one device and remove from the list. */
669 for (slot
= 0; slot
< sra
->array
.raid_disks
; slot
+= nr_of_copies
) {
670 struct mdinfo
**diskp
;
673 /* Find a working device to keep */
674 for (diskp
= &remaining
; *diskp
; diskp
= &(*diskp
)->next
) {
675 struct mdinfo
*disk
= *diskp
;
677 if (disk
->disk
.raid_disk
< slot
)
679 if (disk
->disk
.raid_disk
>= slot
+ nr_of_copies
)
681 if (disk
->disk
.state
& (1<<MD_DISK_REMOVED
))
683 if (disk
->disk
.state
& (1<<MD_DISK_FAULTY
))
685 if (!(disk
->disk
.state
& (1<<MD_DISK_SYNC
)))
688 /* We have found a good disk to use! */
690 disk
->next
= sra
->devs
;
699 if (slot
< sra
->array
.raid_disks
) {
700 /* didn't find all slots */
706 sra
->devs
= remaining
;
710 /* Remove all 'remaining' devices from the array */
712 struct mdinfo
*sd
= remaining
;
713 remaining
= sd
->next
;
715 sysfs_set_str(sra
, sd
, "state", "faulty");
716 sysfs_set_str(sra
, sd
, "slot", "none");
717 sysfs_set_str(sra
, sd
, "state", "remove");
718 sd
->disk
.state
|= (1<<MD_DISK_REMOVED
);
719 sd
->disk
.state
&= ~(1<<MD_DISK_SYNC
);
720 sd
->next
= sra
->devs
;
726 void reshape_free_fdlist(int *fdlist
,
727 unsigned long long *offsets
,
732 for (i
= 0; i
< size
; i
++)
740 int reshape_prepare_fdlist(char *devname
,
744 unsigned long blocks
,
747 unsigned long long *offsets
)
752 for (d
= 0; d
<= nrdisks
; d
++)
755 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
756 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
758 if (sd
->disk
.state
& (1<<MD_DISK_SYNC
)) {
759 char *dn
= map_dev(sd
->disk
.major
,
761 fdlist
[sd
->disk
.raid_disk
]
762 = dev_open(dn
, O_RDONLY
);
763 offsets
[sd
->disk
.raid_disk
] = sd
->data_offset
*512;
764 if (fdlist
[sd
->disk
.raid_disk
] < 0) {
766 Name
": %s: cannot open component %s\n",
767 devname
, dn
? dn
: "-unknown-");
771 } else if (backup_file
== NULL
) {
773 char *dn
= map_dev(sd
->disk
.major
,
775 fdlist
[d
] = dev_open(dn
, O_RDWR
);
776 offsets
[d
] = (sd
->data_offset
+ sra
->component_size
- blocks
- 8)*512;
778 fprintf(stderr
, Name
": %s: cannot open component %s\n",
779 devname
, dn
? dn
: "-unknown-");
790 int reshape_open_backup_file(char *backup_file
,
795 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
|(restart
? O_TRUNC
: 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 char *analyse_change(struct mdinfo
*info
, struct reshape
*re
)
872 /* Based on the current array state in info->array and
873 * the changes in info->new_* etc, determine:
874 * - whether the change is possible
875 * - Intermediate level/raid_disks/layout
876 * - whether a restriping reshape is needed
877 * - number of sectors in minimum change unit. This
878 * will cover a whole number of stripes in 'before' and
881 * Return message if the change should be rejected
882 * NULL if the change can be achieved
884 * This can be called as part of starting a reshape, or
885 * when assembling an array that is undergoing reshape.
889 /* If a new level not explicitly given, we assume no-change */
890 if (info
->new_level
== UnSet
)
891 info
->new_level
= info
->array
.level
;
894 switch (info
->new_level
) {
900 /* chunk size is meaningful, must divide component_size
903 if (info
->component_size
% (info
->new_chunk
/512))
904 return "New chunk size does not"
905 " divide component size";
908 return "chunk size not meaningful for this level";
911 info
->new_chunk
= info
->array
.chunk_size
;
913 switch (info
->array
.level
) {
915 /* RAID1 can convert to RAID1 with different disks, or
918 if (info
->new_level
== 1) {
919 if (info
->delta_disks
== UnSet
)
920 /* Don't know what to do */
921 return "no change requested for Growing RAID1";
923 re
->before
.data_disks
= (info
->array
.raid_disks
+
925 re
->before
.layout
= 0;
926 re
->backup_blocks
= 0;
930 if (info
->array
.raid_disks
== 2 &&
931 info
->array
.raid_disks
== 5) {
932 /* simple in-place conversion */
935 re
->before
.data_disks
= 1;
936 re
->before
.layout
= ALGORITHM_LEFT_SYMMETRIC
;
937 re
->backup_blocks
= 0;
940 /* Could do some multi-stage conversions, but leave that to
943 return "Impossibly level change request for RAID1";
946 /* RAID10 can only be converted from near mode to
947 * RAID0 by removing some devices
949 if ((info
->array
.layout
& ~0xff) != 0x100)
950 return "Cannot Grow RAID10 with far/offset layout";
951 /* number of devices must be multiple of number of copies */
952 if (info
->array
.raid_disks
% (info
->array
.layout
& 0xff))
953 return "RAID10 layout too complex for Grow operation";
955 if (info
->new_level
!= 0)
956 return "RAID10 can only be changed to RAID0";
957 new_disks
= (info
->array
.raid_disks
958 / (info
->array
.layout
& 0xff));
959 if (info
->delta_disks
== UnSet
) {
960 info
->delta_disks
= (new_disks
961 - info
->array
.raid_disks
);
963 if (info
->delta_disks
!= new_disks
- info
->array
.raid_disks
)
964 return "New number of raid-devices impossible for RAID10";
965 if (info
->new_chunk
&&
966 info
->new_chunk
!= info
->array
.chunk_size
)
967 return "Cannot change chunk-size with RAID10 Grow";
972 re
->before
.data_disks
= new_disks
;
973 re
->before
.layout
= 0;
974 re
->backup_blocks
= 0;
978 /* RAID0 can be converted to RAID10, or to RAID456 */
979 if (info
->new_level
== 10) {
980 if (info
->new_layout
== UnSet
&& info
->delta_disks
== UnSet
) {
981 /* Assume near=2 layout */
982 info
->new_layout
= 0x102;
983 info
->delta_disks
= info
->array
.raid_disks
;
985 if (info
->new_layout
== UnSet
) {
986 int copies
= 1 + (info
->delta_disks
987 / info
->array
.raid_disks
);
988 if (info
->array
.raid_disks
* (copies
-1)
989 != info
->delta_disks
)
990 return "Impossible number of devices"
991 " for RAID0->RAID10";
992 info
->new_layout
= 0x100 + copies
;
994 if (info
->delta_disks
== UnSet
) {
995 int copies
= info
->new_layout
& 0xff;
996 if (info
->new_layout
!= 0x100 + copies
)
997 return "New layout impossible"
998 " for RAID0->RAID10";;
999 info
->delta_disks
= (copies
- 1) *
1000 info
->array
.raid_disks
;
1002 if (info
->new_chunk
&&
1003 info
->new_chunk
!= info
->array
.chunk_size
)
1004 return "Cannot change chunk-size with RAID0->RAID10";
1008 re
->before
.data_disks
= (info
->array
.raid_disks
+
1010 re
->before
.layout
= info
->new_layout
;
1011 re
->backup_blocks
= 0;
1015 /* RAID0 can also covert to RAID0/4/5/6 by first converting to
1016 * a raid4 style layout of the final level.
1018 switch (info
->new_level
) {
1022 re
->before
.layout
= 0;
1026 re
->before
.layout
= ALGORITHM_PARITY_N
;
1030 re
->before
.layout
= ALGORITHM_PARITY_N
;
1033 return "Impossible level change requested";
1035 re
->before
.data_disks
= info
->array
.raid_disks
;
1036 /* determining 'after' layout happens outside this 'switch' */
1040 info
->array
.layout
= ALGORITHM_PARITY_N
;
1042 switch (info
->new_level
) {
1044 re
->level
= info
->array
.level
;
1045 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1046 re
->before
.layout
= info
->array
.layout
;
1050 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1051 re
->before
.layout
= info
->array
.layout
;
1055 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1056 switch (info
->array
.layout
) {
1057 case ALGORITHM_LEFT_ASYMMETRIC
:
1058 re
->before
.layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
1060 case ALGORITHM_RIGHT_ASYMMETRIC
:
1061 re
->before
.layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
1063 case ALGORITHM_LEFT_SYMMETRIC
:
1064 re
->before
.layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
1066 case ALGORITHM_RIGHT_SYMMETRIC
:
1067 re
->before
.layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
1069 case ALGORITHM_PARITY_0
:
1070 re
->before
.layout
= ALGORITHM_PARITY_0_6
;
1072 case ALGORITHM_PARITY_N
:
1073 re
->before
.layout
= ALGORITHM_PARITY_N_6
;
1076 return "Cannot convert an array with this layout";
1080 if (info
->array
.raid_disks
!= 2)
1081 return "Can only convert a 2-device array to RAID1";
1083 re
->before
.data_disks
= 2;
1084 re
->before
.layout
= 0;
1087 return "Impossible level change requested";
1091 switch (info
->new_level
) {
1096 re
->before
.data_disks
= info
->array
.raid_disks
- 2;
1097 re
->before
.layout
= info
->array
.layout
;
1100 return "Impossible level change requested";
1105 /* If we reached here then it looks like a re-stripe is
1106 * happening. We have determined the intermediate level
1107 * and initial raid_disks/layout and stored these in 're'.
1109 * We need to deduce the final layout that can be atomically
1110 * converted to the end state.
1112 switch (info
->new_level
) {
1114 /* We can only get to RAID0 from RAID4 or RAID5
1115 * with appropriate layout and one extra device
1117 if (re
->level
!= 4 && re
->level
!= 5)
1118 return "Cannot covert to RAID0 from this level";
1119 if (info
->delta_disks
== UnSet
)
1120 re
->after
.data_disks
= re
->before
.data_disks
;
1122 re
->after
.data_disks
=
1123 info
->array
.raid_disks
+ info
->delta_disks
;
1124 switch (re
->level
) {
1126 re
->after
.layout
= 0 ; break;
1128 re
->after
.layout
= ALGORITHM_PARITY_N
; break;
1133 /* We can only get to RAID4 from RAID5 */
1134 if (re
->level
!= 4 && re
->level
!= 5)
1135 return "Cannot convert to RAID4 from this level";
1136 if (info
->delta_disks
== UnSet
)
1137 re
->after
.data_disks
= re
->before
.data_disks
;
1139 re
->after
.data_disks
=
1140 re
->before
.data_disks
+ info
->delta_disks
;
1141 switch (re
->level
) {
1143 re
->after
.layout
= 0 ; break;
1145 re
->after
.layout
= ALGORITHM_PARITY_N
; break;
1150 /* We get to RAID5 for RAID5 or RAID6 */
1151 if (re
->level
!= 5 && re
->level
!= 6)
1152 return "Cannot convert to RAID5 from this level";
1153 if (info
->delta_disks
== UnSet
)
1154 re
->after
.data_disks
= re
->before
.data_disks
;
1155 else if (re
->level
== 5)
1156 re
->after
.data_disks
=
1157 re
->before
.data_disks
+ info
->delta_disks
;
1159 re
->after
.data_disks
=
1160 info
->array
.raid_disks
+ info
->delta_disks
- 1;
1161 switch (re
->level
) {
1163 if (info
->new_layout
== UnSet
)
1164 re
->after
.layout
= re
->before
.layout
;
1166 re
->after
.layout
= info
->new_layout
;
1169 if (info
->new_layout
== UnSet
)
1170 info
->new_layout
= re
->before
.layout
;
1172 /* after.layout needs to be raid6 version of new_layout */
1173 if (info
->new_layout
== ALGORITHM_PARITY_N
)
1174 re
->after
.layout
= ALGORITHM_PARITY_N
;
1177 char *ls
= map_num(r5layout
, info
->new_layout
);
1179 strcat(strcpy(layout
, ls
), "-6");
1180 l
= map_name(r6layout
, layout
);
1182 return "Cannot find RAID6 layout"
1184 re
->after
.layout
= l
;
1190 /* We must already be at level 6 */
1192 return "Impossible level change";
1193 if (info
->delta_disks
== UnSet
)
1194 re
->after
.data_disks
= re
->before
.data_disks
;
1196 re
->after
.data_disks
= (info
->array
.raid_disks
+
1197 info
->delta_disks
) - 2;
1198 if (info
->new_layout
== UnSet
)
1199 re
->after
.layout
= info
->array
.layout
;
1201 re
->after
.layout
= info
->new_layout
;
1204 return "Impossible level change requested";
1206 switch (re
->level
) {
1207 case 6: re
->parity
= 2; break;
1209 case 5: re
->parity
= 1; break;
1210 default: re
->parity
= 0; break;
1212 /* So we have a restripe operation, we need to calculate the number
1213 * of blocks per reshape operation.
1215 if (info
->new_chunk
== 0)
1216 info
->new_chunk
= info
->array
.chunk_size
;
1217 if (re
->after
.data_disks
== re
->before
.data_disks
&&
1218 re
->after
.layout
== re
->before
.layout
&&
1219 info
->new_chunk
== info
->array
.chunk_size
) {
1220 /* Nothing to change */
1221 re
->backup_blocks
= 0;
1224 if (re
->after
.data_disks
== 1 && re
->before
.data_disks
== 1) {
1225 /* chunks can layout changes make no difference */
1226 re
->backup_blocks
= 0;
1230 if (re
->after
.data_disks
== re
->before
.data_disks
&&
1231 get_linux_version() < 2006032)
1232 return "in-place reshape is not safe before 2.6.32 - sorry.";
1234 if (re
->after
.data_disks
< re
->before
.data_disks
&&
1235 get_linux_version() < 2006030)
1236 return "reshape to fewer devices is not supported before 2.6.32 - sorry.";
1238 re
->backup_blocks
= compute_backup_blocks(
1239 info
->new_chunk
, info
->array
.chunk_size
,
1240 re
->after
.data_disks
,
1241 re
->before
.data_disks
);
1243 re
->new_size
= info
->component_size
* re
->after
.data_disks
;
1247 static int reshape_array(char *container
, int fd
, char *devname
,
1248 struct supertype
*st
, struct mdinfo
*info
,
1249 int force
, char *backup_file
, int quiet
, int forked
,
1251 static int reshape_container(char *container
, int cfd
, char *devname
,
1252 struct supertype
*st
,
1253 struct mdinfo
*info
,
1258 int Grow_reshape(char *devname
, int fd
, int quiet
, char *backup_file
,
1260 int level
, char *layout_str
, int chunksize
, int raid_disks
,
1263 /* Make some changes in the shape of an array.
1264 * The kernel must support the change.
1266 * There are three different changes. Each can trigger
1267 * a resync or recovery so we freeze that until we have
1268 * requested everything (if kernel supports freezing - 2.6.30).
1270 * - change size (i.e. component_size)
1272 * - change layout/chunksize/ndisks
1274 * The last can require a reshape. It is different on different
1275 * levels so we need to check the level before actioning it.
1276 * Some times the level change needs to be requested after the
1277 * reshape (e.g. raid6->raid5, raid5->raid0)
1280 struct mdu_array_info_s array
;
1282 struct supertype
*st
;
1283 char *subarray
= NULL
;
1287 char *container
= NULL
;
1288 char container_buf
[20];
1294 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) < 0) {
1295 fprintf(stderr
, Name
": %s is not an active md array - aborting\n",
1301 (chunksize
|| level
!= UnSet
|| layout_str
|| raid_disks
)) {
1302 fprintf(stderr
, Name
": cannot change component size at the same time "
1303 "as other changes.\n"
1304 " Change size first, then check data is intact before "
1305 "making other changes.\n");
1309 if (raid_disks
&& raid_disks
< array
.raid_disks
&& array
.level
> 1 &&
1310 get_linux_version() < 2006032 &&
1311 !check_env("MDADM_FORCE_FEWER")) {
1312 fprintf(stderr
, Name
": reducing the number of devices is not safe before Linux 2.6.32\n"
1313 " Please use a newer kernel\n");
1317 st
= super_by_fd(fd
, &subarray
);
1319 fprintf(stderr
, Name
": Unable to determine metadata format for %s\n", devname
);
1322 if (raid_disks
> st
->max_devs
) {
1323 fprintf(stderr
, Name
": Cannot increase raid-disks on this array"
1324 " beyond %d\n", st
->max_devs
);
1328 /* in the external case we need to check that the requested reshape is
1329 * supported, and perform an initial check that the container holds the
1330 * pre-requisite spare devices (mdmon owns final validation)
1332 if (st
->ss
->external
) {
1337 container_dev
= st
->container_dev
;
1338 cfd
= open_dev_excl(st
->container_dev
);
1340 container_dev
= st
->devnum
;
1342 cfd
= open_dev_excl(st
->devnum
);
1346 fprintf(stderr
, Name
": Unable to open container for %s\n",
1352 fmt_devname(container_buf
, container_dev
);
1353 container
= container_buf
;
1355 rv
= st
->ss
->load_container(st
, cfd
, NULL
);
1358 fprintf(stderr
, Name
": Cannot read superblock for %s\n",
1364 if (mdmon_running(container_dev
))
1365 st
->update_tail
= &st
->updates
;
1368 if (raid_disks
> array
.raid_disks
&&
1369 array
.spare_disks
< (raid_disks
- array
.raid_disks
) &&
1372 Name
": Need %d spare%s to avoid degraded array,"
1373 " and only have %d.\n"
1374 " Use --force to over-ride this check.\n",
1375 raid_disks
- array
.raid_disks
,
1376 raid_disks
- array
.raid_disks
== 1 ? "" : "s",
1381 sra
= sysfs_read(fd
, 0, GET_LEVEL
| GET_DISKS
| GET_DEVS
1382 | GET_STATE
| GET_VERSION
);
1384 if (st
->ss
->external
&& subarray
== NULL
) {
1385 array
.level
= LEVEL_CONTAINER
;
1386 sra
->array
.level
= LEVEL_CONTAINER
;
1389 fprintf(stderr
, Name
": failed to read sysfs parameters for %s\n",
1393 frozen
= freeze(st
);
1395 /* freeze() already spewed the reason */
1397 } else if (frozen
< 0) {
1398 fprintf(stderr
, Name
": %s is performing resync/recovery and cannot"
1399 " be reshaped\n", devname
);
1403 /* ========= set size =============== */
1404 if (size
>= 0 && (size
== 0 || size
!= array
.size
)) {
1405 long long orig_size
= array
.size
;
1407 if (reshape_super(st
, size
, UnSet
, UnSet
, 0, 0, NULL
, devname
, !quiet
)) {
1413 if (array
.size
!= size
) {
1414 /* got truncated to 32bit, write to
1415 * component_size instead
1418 rv
= sysfs_set_num(sra
, NULL
,
1419 "component_size", size
);
1423 rv
= ioctl(fd
, SET_ARRAY_INFO
, &array
);
1427 /* restore metadata */
1428 if (reshape_super(st
, orig_size
, UnSet
, UnSet
, 0, 0,
1429 NULL
, devname
, !quiet
) == 0)
1431 fprintf(stderr
, Name
": Cannot set device size for %s: %s\n",
1432 devname
, strerror(err
));
1434 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
1435 fprintf(stderr
, " Bitmap must be removed before size can be changed\n");
1439 ioctl(fd
, GET_ARRAY_INFO
, &array
);
1440 size
= get_component_size(fd
)/2;
1444 fprintf(stderr
, Name
": component size of %s has been set to %lluK\n",
1447 } else if (array
.level
!= LEVEL_CONTAINER
) {
1448 size
= get_component_size(fd
)/2;
1453 /* ========= check for Raid10 -> Raid0 conversion ===============
1454 * current implementation assumes that following conditions must be met:
1456 * - near_copies == 2
1458 if (level
== 0 && array
.level
== 10 && sra
&&
1459 array
.layout
== ((1 << 8) + 2) && !(array
.raid_disks
& 1)) {
1461 err
= remove_disks_on_raid10_to_raid0_takeover(st
, sra
, array
.layout
);
1463 dprintf(Name
": Array cannot be reshaped\n");
1472 sysfs_init(&info
, fd
, NoMdDev
);
1473 strcpy(info
.text_version
, sra
->text_version
);
1474 info
.component_size
= size
*2;
1475 info
.new_level
= level
;
1476 info
.new_chunk
= chunksize
* 1024;
1478 info
.delta_disks
= raid_disks
- info
.array
.raid_disks
;
1480 info
.delta_disks
= UnSet
;
1481 if (layout_str
== NULL
) {
1482 info
.new_layout
= UnSet
;
1483 if (info
.array
.level
== 6 &&
1484 (info
.new_level
== 6 || info
.new_level
== UnSet
) &&
1485 info
.array
.layout
>= 16) {
1486 fprintf(stderr
, Name
1487 ": %s has a non-standard layout. If you"
1488 " wish to preserve this\n"
1489 " during the reshape, please specify"
1490 " --layout=preserve\n"
1491 " If you want to change it, specify a"
1492 " layout or use --layout=normalise\n",
1497 } else if (strcmp(layout_str
, "normalise") == 0 ||
1498 strcmp(layout_str
, "normalize") == 0) {
1499 /* If we have a -6 RAID6 layout, remove the '-6'. */
1500 info
.new_layout
= UnSet
;
1501 if (info
.array
.level
== 6 && info
.new_level
== UnSet
) {
1503 strcpy(l
, map_num(r6layout
, info
.array
.layout
));
1504 h
= strrchr(l
, '-');
1505 if (h
&& strcmp(h
, "-6") == 0) {
1507 info
.new_layout
= map_name(r6layout
, l
);
1510 } else if (strcmp(layout_str
, "preserve") == 0) {
1511 info
.new_layout
= UnSet
;
1513 int l
= info
.new_level
;
1515 l
= info
.array
.level
;
1518 info
.new_layout
= map_name(r5layout
, layout_str
);
1521 info
.new_layout
= map_name(r6layout
, layout_str
);
1524 info
.new_layout
= parse_layout_10(layout_str
);
1527 info
.new_layout
= parse_layout_faulty(layout_str
);
1530 fprintf(stderr
, Name
": layout not meaningful"
1531 " with this level\n");
1535 if (info
.new_layout
== UnSet
) {
1536 fprintf(stderr
, Name
": layout %s not understood"
1537 " for this level\n",
1544 if (array
.level
== LEVEL_CONTAINER
) {
1545 /* This change is to be applied to every array in the
1546 * container. This is only needed when the metadata imposes
1547 * restraints of the various arrays in the container.
1548 * Currently we only know that IMSM requires all arrays
1549 * to have the same number of devices so changing the
1550 * number of devices (On-Line Capacity Expansion) must be
1551 * performed at the level of the container
1553 rv
= reshape_container(container
, fd
, devname
, st
, &info
,
1554 force
, backup_file
, quiet
);
1557 /* Impose these changes on a single array. First
1558 * check that the metadata is OK with the change. */
1560 if (reshape_super(st
, info
.component_size
, info
.new_level
,
1561 info
.new_layout
, info
.new_chunk
,
1562 info
.array
.raid_disks
+ info
.delta_disks
,
1563 backup_file
, devname
, quiet
)) {
1568 rv
= reshape_array(container
, fd
, devname
, st
, &info
, force
,
1569 backup_file
, quiet
, 0, 0);
1578 static int reshape_array(char *container
, int fd
, char *devname
,
1579 struct supertype
*st
, struct mdinfo
*info
,
1581 char *backup_file
, int quiet
, int forked
,
1584 struct reshape reshape
;
1587 int orig_level
= UnSet
;
1590 struct mdu_array_info_s array
;
1594 unsigned long long *offsets
;
1598 unsigned long blocks
;
1599 unsigned long cache
;
1600 unsigned long long array_size
;
1602 struct mdinfo
*sra
= NULL
;
1604 msg
= analyse_change(info
, &reshape
);
1606 fprintf(stderr
, Name
": %s\n", msg
);
1609 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
1610 dprintf("Canot get array information.\n");
1615 /* reshape already started. just skip to monitoring the reshape */
1616 if (reshape
.backup_blocks
== 0)
1620 spares_needed
= max(reshape
.before
.data_disks
,
1621 reshape
.after
.data_disks
)
1622 + reshape
.parity
- array
.raid_disks
;
1625 info
->new_level
> 0 &&
1626 spares_needed
> info
->array
.spare_disks
) {
1628 Name
": Need %d spare%s to avoid degraded array,"
1629 " and only have %d.\n"
1630 " Use --force to over-ride this check.\n",
1632 spares_needed
== 1 ? "" : "s",
1633 info
->array
.spare_disks
);
1637 if (reshape
.level
!= info
->array
.level
) {
1638 char *c
= map_num(pers
, reshape
.level
);
1643 err
= sysfs_set_str(info
, NULL
, "level", c
);
1646 fprintf(stderr
, Name
": %s: could not set level to %s\n",
1649 (info
->array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
1650 fprintf(stderr
, " Bitmap must be removed"
1651 " before level can be changed\n");
1655 fprintf(stderr
, Name
": level of %s changed to %s\n",
1657 orig_level
= info
->array
.level
;
1660 if (reshape
.level
> 0 && st
->ss
->external
&&
1661 !mdmon_running(st
->container_dev
)) {
1662 start_mdmon(st
->container_dev
);
1663 ping_monitor(container
);
1666 /* ->reshape_super might have chosen some spares from the
1667 * container that it wants to be part of the new array.
1668 * We can collect them with ->container_content and give
1669 * them to the kernel.
1671 if (st
->ss
->reshape_super
&& st
->ss
->container_content
) {
1672 char *subarray
= strchr(info
->text_version
+1, '/')+1;
1673 struct mdinfo
*info2
=
1674 st
->ss
->container_content(st
, subarray
);
1678 sysfs_init(info2
, fd
, st
->devnum
);
1679 for (d
= info2
->devs
; d
; d
= d
->next
) {
1680 if (d
->disk
.state
== 0 &&
1681 d
->disk
.raid_disk
>= 0) {
1682 /* This is a spare that wants to
1683 * be part of the array.
1685 add_disk(fd
, st
, info2
, d
);
1692 if (reshape
.backup_blocks
== 0) {
1693 /* No restriping needed, but we might need to impose
1694 * some more changes: layout, raid_disks, chunk_size
1696 if (info
->new_layout
!= UnSet
&&
1697 info
->new_layout
!= info
->array
.layout
) {
1698 info
->array
.layout
= info
->new_layout
;
1699 if (ioctl(fd
, SET_ARRAY_INFO
, &info
->array
) != 0) {
1700 fprintf(stderr
, Name
": failed to set new layout\n");
1703 printf("layout for %s set to %d\n",
1704 devname
, info
->array
.layout
);
1706 if (info
->delta_disks
!= UnSet
&&
1707 info
->delta_disks
!= 0) {
1708 info
->array
.raid_disks
+= info
->delta_disks
;
1709 if (ioctl(fd
, SET_ARRAY_INFO
, &info
->array
) != 0) {
1710 fprintf(stderr
, Name
": failed to set raid disks\n");
1713 printf("raid_disks for %s set to %d\n",
1714 devname
, info
->array
.raid_disks
);
1716 if (info
->new_chunk
!= 0 &&
1717 info
->new_chunk
!= info
->array
.chunk_size
) {
1718 if (sysfs_set_num(info
, NULL
,
1719 "chunk_size", info
->new_chunk
) != 0) {
1720 fprintf(stderr
, Name
": failed to set chunk size\n");
1723 printf("chunk size for %s set to %d\n",
1724 devname
, info
->array
.chunk_size
);
1733 * There are three possibilities.
1734 * 1/ The array will shrink.
1735 * We need to ensure the reshape will pause before reaching
1736 * the 'critical section'. We also need to fork and wait for
1737 * that to happen. When it does we
1738 * suspend/backup/complete/unfreeze
1740 * 2/ The array will not change size.
1741 * This requires that we keep a backup of a sliding window
1742 * so that we can restore data after a crash. So we need
1743 * to fork and monitor progress.
1744 * In future we will allow the data_offset to change, so
1745 * a sliding backup becomes unnecessary.
1747 * 3/ The array will grow. This is relatively easy.
1748 * However the kernel's restripe routines will cheerfully
1749 * overwrite some early data before it is safe. So we
1750 * need to make a backup of the early parts of the array
1751 * and be ready to restore it if rebuild aborts very early.
1752 * For externally managed metadata, we still need a forked
1753 * child to monitor the reshape and suspend IO over the region
1754 * that is being reshaped.
1756 * We backup data by writing it to one spare, or to a
1757 * file which was given on command line.
1759 * In each case, we first make sure that storage is available
1760 * for the required backup.
1762 * - request the shape change.
1763 * - fork to handle backup etc.
1766 /* Check that we can hold all the data */
1767 get_dev_size(fd
, NULL
, &array_size
);
1768 if (reshape
.new_size
< (array_size
/512)) {
1770 Name
": this change will reduce the size of the array.\n"
1771 " use --grow --array-size first to truncate array.\n"
1772 " e.g. mdadm --grow %s --array-size %llu\n",
1773 devname
, reshape
.new_size
/2);
1777 sra
= sysfs_read(fd
, 0,
1778 GET_COMPONENT
|GET_DEVS
|GET_OFFSET
|GET_STATE
|GET_CHUNK
|
1781 fprintf(stderr
, Name
": %s: Cannot get array details from sysfs\n",
1786 /* Decide how many blocks (sectors) for a reshape
1787 * unit. The number we have so far is just a minimum
1789 blocks
= reshape
.backup_blocks
;
1790 if (reshape
.before
.data_disks
==
1791 reshape
.after
.data_disks
) {
1792 /* Make 'blocks' bigger for better throughput, but
1793 * not so big that we reject it below.
1794 * Try for 16 megabytes
1796 while (blocks
* 32 < sra
->component_size
&&
1800 fprintf(stderr
, Name
": Need to backup %luK of critical "
1801 "section..\n", blocks
/2);
1803 if (blocks
>= sra
->component_size
/2) {
1804 fprintf(stderr
, Name
": %s: Something wrong"
1805 " - reshape aborted\n",
1810 /* Now we need to open all these devices so we can read/write.
1812 nrdisks
= array
.raid_disks
+ sra
->array
.spare_disks
;
1813 fdlist
= malloc((1+nrdisks
) * sizeof(int));
1814 offsets
= malloc((1+nrdisks
) * sizeof(offsets
[0]));
1815 if (!fdlist
|| !offsets
) {
1816 fprintf(stderr
, Name
": malloc failed: grow aborted\n");
1820 d
= reshape_prepare_fdlist(devname
, sra
, array
.raid_disks
,
1821 nrdisks
, blocks
, backup_file
,
1826 if (backup_file
== NULL
) {
1827 if (reshape
.after
.data_disks
<= reshape
.before
.data_disks
) {
1829 Name
": %s: Cannot grow - need backup-file\n",
1832 } else if (sra
->array
.spare_disks
== 0) {
1833 fprintf(stderr
, Name
": %s: Cannot grow - need a spare or "
1834 "backup-file to backup critical section\n",
1839 if (!reshape_open_backup_file(backup_file
, fd
, devname
,
1841 fdlist
+d
, offsets
+d
, restart
)) {
1847 /* lastly, check that the internal stripe cache is
1848 * large enough, or it won't work.
1849 * It must hold at least 4 stripes of the larger
1852 cache
= max(info
->array
.chunk_size
, info
->new_chunk
);
1853 cache
*= 4; /* 4 stripes minimum */
1854 cache
/= 512; /* convert to sectors */
1855 disks
= min(reshape
.before
.data_disks
, reshape
.after
.data_disks
);
1856 /* make sure there is room for 'blocks' with a bit to spare */
1857 if (cache
< 16 + blocks
/ disks
)
1858 cache
= 16 + blocks
/ disks
;
1859 cache
/= (4096/512); /* Covert from sectors to pages */
1861 if (sra
->cache_size
< cache
)
1862 subarray_set_num(container
, sra
, "stripe_cache_size",
1865 /* Right, everything seems fine. Let's kick things off.
1866 * If only changing raid_disks, use ioctl, else use
1871 sra
->new_chunk
= info
->new_chunk
;
1873 if (info
->reshape_active
)
1874 /* nothing needed here */;
1875 else if (info
->array
.chunk_size
== info
->new_chunk
&&
1876 reshape
.before
.layout
== reshape
.after
.layout
&&
1877 st
->ss
->external
== 0) {
1878 array
.raid_disks
= reshape
.after
.data_disks
+ reshape
.parity
;
1879 if (ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
1883 Name
": Cannot set device shape for %s: %s\n",
1884 devname
, strerror(errno
));
1887 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
1889 " Bitmap must be removed before"
1890 " shape can be changed\n");
1895 /* set them all just in case some old 'new_*' value
1896 * persists from some earlier problem
1899 if (sysfs_set_num(sra
, NULL
, "chunk_size", info
->new_chunk
) < 0)
1901 if (!err
&& sysfs_set_num(sra
, NULL
, "layout",
1902 reshape
.after
.layout
) < 0)
1904 if (!err
&& subarray_set_num(container
, sra
, "raid_disks",
1905 reshape
.after
.data_disks
+
1906 reshape
.parity
) < 0)
1909 fprintf(stderr
, Name
": Cannot set device shape for %s\n",
1913 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
1915 " Bitmap must be removed before"
1916 " shape can be changed\n");
1923 sysfs_set_str(sra
, NULL
, "array_state", "active");
1925 /* Now we just need to kick off the reshape and watch, while
1926 * handling backups of the data...
1927 * This is all done by a forked background process.
1929 switch(forked
? 0 : fork()) {
1931 fprintf(stderr
, Name
": Cannot run child to monitor reshape: %s\n",
1942 if (check_env("MDADM_GROW_VERIFY"))
1943 fd
= open(devname
, O_RDONLY
| O_DIRECT
);
1946 mlockall(MCL_FUTURE
);
1948 odisks
= reshape
.before
.data_disks
+ reshape
.parity
;
1950 if (st
->ss
->external
) {
1951 /* metadata handler takes it from here */
1952 done
= st
->ss
->manage_reshape(
1953 fd
, sra
, &reshape
, st
, blocks
,
1955 d
- odisks
, fdlist
+odisks
,
1958 done
= child_monitor(
1959 fd
, sra
, &reshape
, st
, blocks
,
1961 d
- odisks
, fdlist
+odisks
,
1964 if (backup_file
&& done
)
1965 unlink(backup_file
);
1971 if (!st
->ss
->external
&&
1972 !(reshape
.before
.data_disks
!= reshape
.after
.data_disks
1973 && info
->custom_array_size
) &&
1974 info
->new_level
== reshape
.level
&&
1976 /* no need to wait for the reshape to finish as
1977 * there is nothing more to do.
1983 if (st
->ss
->external
) {
1984 /* Re-load the metadata as much could have changed */
1985 int cfd
= open_dev(st
->container_dev
);
1987 ping_monitor(container
);
1988 st
->ss
->free_super(st
);
1989 st
->ss
->load_container(st
, cfd
, container
);
1994 /* set new array size if required customer_array_size is used
1997 if (reshape
.before
.data_disks
!=
1998 reshape
.after
.data_disks
&&
1999 info
->custom_array_size
) {
2000 struct mdinfo
*info2
;
2001 char *subarray
= strchr(info
->text_version
+1, '/')+1;
2003 info2
= st
->ss
->container_content(st
, subarray
);
2005 unsigned long long current_size
= 0;
2006 unsigned long long new_size
=
2007 info2
->custom_array_size
/2;
2009 if (sysfs_get_ll(sra
,
2012 ¤t_size
) == 0 &&
2013 new_size
> current_size
) {
2014 if (sysfs_set_num(sra
, NULL
,
2015 "array_size", new_size
)
2017 dprintf("Error: Cannot"
2020 dprintf("Array size "
2022 dprintf(" from %llu to %llu.\n",
2023 current_size
, new_size
);
2029 if (info
->new_level
!= reshape
.level
) {
2031 c
= map_num(pers
, info
->new_level
);
2033 err
= sysfs_set_str(sra
, NULL
, "level", c
);
2035 fprintf(stderr
, Name\
2036 ": %s: could not set level "
2037 "to %s\n", devname
, c
);
2046 if (orig_level
!= UnSet
&& sra
) {
2047 c
= map_num(pers
, orig_level
);
2048 if (c
&& sysfs_set_str(sra
, NULL
, "level", c
) == 0)
2049 fprintf(stderr
, Name
": aborting level change\n");
2056 int reshape_container(char *container
, int cfd
, char *devname
,
2057 struct supertype
*st
,
2058 struct mdinfo
*info
,
2063 struct mdinfo
*cc
= NULL
;
2065 /* component_size is not meaningful for a container,
2066 * so pass '-1' meaning 'no change'
2068 if (reshape_super(st
, -1, info
->new_level
,
2069 info
->new_layout
, info
->new_chunk
,
2070 info
->array
.raid_disks
+ info
->delta_disks
,
2071 backup_file
, devname
, quiet
))
2076 /* ping monitor to be sure that update is on disk
2078 ping_monitor(container
);
2081 case -1: /* error */
2082 perror("Cannot fork to complete reshape\n");
2084 default: /* parent */
2085 printf(Name
": multi-array reshape continues in background\n");
2092 /* For each member array with reshape_active,
2093 * we need to perform the reshape.
2094 * We pick the first array that needs reshaping and
2095 * reshape it. reshape_array() will re-read the metadata
2096 * so the next time through a different array should be
2097 * ready for reshape.
2099 struct mdinfo
*content
;
2102 struct mdstat_ent
*mdstat
;
2107 cc
= st
->ss
->container_content(st
, NULL
);
2109 for (content
= cc
; content
; content
= content
->next
) {
2111 if (!content
->reshape_active
)
2114 subarray
= strchr(content
->text_version
+1, '/')+1;
2115 mdstat
= mdstat_by_subdev(subarray
,
2116 devname2devnum(container
));
2124 fd
= open_dev_excl(mdstat
->devnum
);
2127 adev
= map_dev(dev2major(mdstat
->devnum
),
2128 dev2minor(mdstat
->devnum
),
2131 adev
= content
->text_version
;
2133 sysfs_init(content
, fd
, mdstat
->devnum
);
2135 rv
= reshape_array(container
, fd
, adev
, st
,
2137 backup_file
, quiet
, 1, 0);
2148 * We run a child process in the background which performs the following
2150 * - wait for resync to reach a certain point
2151 * - suspend io to the following section
2152 * - backup that section
2153 * - allow resync to proceed further
2155 * - discard the backup.
2157 * When are combined in slightly different ways in the three cases.
2159 * - suspend/backup/allow/wait/resume/discard
2161 * - allow/wait/suspend/backup/allow/wait/resume/discard
2163 * - wait/resume/discard/suspend/backup/allow
2165 * suspend/backup/allow always come together
2166 * wait/resume/discard do too.
2167 * For the same-size case we have two backups to improve flow.
2171 int progress_reshape(struct mdinfo
*info
, struct reshape
*reshape
,
2172 unsigned long long backup_point
,
2173 unsigned long long wait_point
,
2174 unsigned long long *suspend_point
,
2175 unsigned long long *reshape_completed
)
2177 /* This function is called repeatedly by the reshape manager.
2178 * It determines how much progress can safely be made and allows
2180 * - 'info' identifies the array and particularly records in
2181 * ->reshape_progress the metadata's knowledge of progress
2182 * This is a sector offset from the start of the array
2183 * of the next array block to be relocated. This number
2184 * may increase from 0 or decrease from array_size, depending
2185 * on the type of reshape that is happening.
2186 * Note that in contrast, 'sync_completed' is a block count of the
2187 * reshape so far. It gives the distance between the start point
2188 * (head or tail of device) and the next place that data will be
2189 * written. It always increases.
2190 * - 'reshape' is the structure created by analyse_change
2191 * - 'backup_point' shows how much the metadata manager has backed-up
2192 * data. For reshapes with increasing progress, it is the next address
2193 * to be backed up, previous addresses have been backed-up. For
2194 * decreasing progress, it is the earliest address that has been
2195 * backed up - later address are also backed up.
2196 * So addresses between reshape_progress and backup_point are
2197 * backed up providing those are in the 'correct' order.
2198 * - 'wait_point' is an array address. When reshape_completed
2199 * passes this point, progress_reshape should return. It might
2200 * return earlier if it determines that ->reshape_progress needs
2201 * to be updated or further backup is needed.
2202 * - suspend_point is maintained by progress_reshape and the caller
2203 * should not touch it except to initialise to zero.
2204 * It is an array address and it only increases in 2.6.37 and earlier.
2205 * This makes it difficult to handle reducing reshapes with
2206 * external metadata.
2207 * However: it is similar to backup_point in that it records the
2208 * other end of a suspended region from reshape_progress.
2209 * it is moved to extend the region that is safe to backup and/or
2211 * - reshape_completed is read from sysfs and returned. The caller
2212 * should copy this into ->reshape_progress when it has reason to
2213 * believe that the metadata knows this, and any backup outside this
2217 * 1 if more data from backup_point - but only as far as suspend_point,
2218 * should be backed up
2219 * 0 if things are progressing smoothly
2220 * -1 if the reshape is finished, either because it is all done,
2221 * or due to an error.
2224 int advancing
= (reshape
->after
.data_disks
2225 >= reshape
->before
.data_disks
);
2226 unsigned long long need_backup
; /* need to eventually backup all the way
2229 unsigned long long read_offset
, write_offset
;
2230 unsigned long long write_range
;
2231 unsigned long long max_progress
, target
, completed
;
2232 unsigned long long array_size
= (info
->component_size
2233 * reshape
->before
.data_disks
);
2237 /* First, we unsuspend any region that is now known to be safe.
2238 * If suspend_point is on the 'wrong' side of reshape_progress, then
2239 * we don't have or need suspension at the moment. This is true for
2240 * native metadata when we don't need to back-up.
2243 if (info
->reshape_progress
<= *suspend_point
)
2244 sysfs_set_num(info
, NULL
, "suspend_lo",
2245 info
->reshape_progress
);
2247 /* Note: this won't work in 2.6.37 and before.
2248 * Something somewhere should make sure we don't need it!
2250 if (info
->reshape_progress
>= *suspend_point
)
2251 sysfs_set_num(info
, NULL
, "suspend_hi",
2252 info
->reshape_progress
);
2255 /* Now work out how far it is safe to progress.
2256 * If the read_offset for ->reshape_progress is less than
2257 * 'blocks' beyond the write_offset, we can only progress as far
2259 * Otherwise we can progress until the write_offset for the new location
2260 * reaches (within 'blocks' of) the read_offset at the current location.
2261 * However that region must be suspended unless we are using native
2263 * If we need to suspend more, we limit it to 128M per device, which is
2264 * rather arbitrary and should be some time-based calculation.
2266 read_offset
= info
->reshape_progress
/ reshape
->before
.data_disks
;
2267 write_offset
= info
->reshape_progress
/ reshape
->after
.data_disks
;
2268 write_range
= info
->new_chunk
/512;
2271 if (read_offset
< write_offset
+ write_range
) {
2272 max_progress
= backup_point
;
2273 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
)
2274 need_backup
= array_size
;
2276 need_backup
= reshape
->backup_blocks
;
2280 reshape
->after
.data_disks
;
2283 need_backup
= array_size
;
2284 if (read_offset
> write_offset
- write_range
) {
2285 max_progress
= backup_point
;
2286 if (max_progress
>= info
->reshape_progress
)
2291 reshape
->after
.data_disks
;
2292 /* If we are using internal metadata, then we can
2293 * progress all the way to the suspend_point without
2294 * worrying about backing-up/suspending along the
2297 if (max_progress
< *suspend_point
&&
2298 info
->array
.major_version
>= 0)
2299 max_progress
= *suspend_point
;
2303 /* We know it is safe to progress to 'max_progress' providing
2304 * it is suspended or we are using native metadata.
2305 * Consider extending suspend_point 128M per device if it
2306 * is less than 64M per device beyond reshape_progress.
2307 * But always do a multiple of 'blocks'
2308 * FIXME this is too big - it takes to long to complete
2311 target
= 64*1024*2 * min(reshape
->before
.data_disks
,
2312 reshape
->after
.data_disks
);
2313 target
/= reshape
->backup_blocks
;
2316 target
*= reshape
->backup_blocks
;
2318 /* For externally managed metadata we always need to suspend IO to
2319 * the area being reshaped so we regularly push suspend_point forward.
2320 * For native metadata we only need the suspend if we are going to do
2324 if ((need_backup
> info
->reshape_progress
2325 || info
->array
.major_version
< 0) &&
2326 *suspend_point
< info
->reshape_progress
+ target
) {
2327 if (need_backup
< *suspend_point
+ 2 * target
)
2328 *suspend_point
= need_backup
;
2329 else if (*suspend_point
+ 2 * target
< array_size
)
2330 *suspend_point
+= 2 * target
;
2332 *suspend_point
= array_size
;
2333 sysfs_set_num(info
, NULL
, "suspend_hi", *suspend_point
);
2334 if (max_progress
> *suspend_point
)
2335 max_progress
= *suspend_point
;
2338 if ((need_backup
< info
->reshape_progress
2339 || info
->array
.major_version
< 0) &&
2340 *suspend_point
> info
->reshape_progress
- target
) {
2341 if (need_backup
> *suspend_point
- 2 * target
)
2342 *suspend_point
= need_backup
;
2343 else if (*suspend_point
>= 2 * target
)
2344 *suspend_point
-= 2 * target
;
2347 sysfs_set_num(info
, NULL
, "suspend_lo", *suspend_point
);
2348 if (max_progress
< *suspend_point
)
2349 max_progress
= *suspend_point
;
2353 /* now set sync_max to allow that progress. sync_max, like
2354 * sync_completed is a count of sectors written per device, so
2355 * we find the difference between max_progress and the start point,
2356 * and divide that by after.data_disks to get a sync_max
2358 * At the same time we convert wait_point to a similar number
2359 * for comparing against sync_completed.
2361 /* scale down max_progress to per_disk */
2362 max_progress
/= reshape
->after
.data_disks
;
2363 /* Round to chunk size as some kernels give an erroneously high number */
2364 max_progress
/= info
->new_chunk
/512;
2365 max_progress
*= info
->new_chunk
/512;
2366 /* Limit progress to the whole device */
2367 if (max_progress
> info
->component_size
)
2368 max_progress
= info
->component_size
;
2369 wait_point
/= reshape
->after
.data_disks
;
2371 /* switch from 'device offset' to 'processed block count' */
2372 max_progress
= info
->component_size
- max_progress
;
2373 wait_point
= info
->component_size
- wait_point
;
2376 sysfs_set_num(info
, NULL
, "sync_max", max_progress
);
2378 /* Now wait. If we have already reached the point that we were
2379 * asked to wait to, don't wait at all, else wait for any change.
2380 * We need to select on 'sync_completed' as that is the place that
2381 * notifications happen, but we are really interested in
2382 * 'reshape_position'
2384 fd
= sysfs_get_fd(info
, NULL
, "sync_completed");
2386 goto check_progress
;
2388 if (sysfs_fd_get_ll(fd
, &completed
) < 0) {
2390 goto check_progress
;
2392 while (completed
< max_progress
&& completed
< wait_point
) {
2393 /* Check that sync_action is still 'reshape' to avoid
2394 * waiting forever on a dead array
2398 if (sysfs_get_str(info
, NULL
, "sync_action",
2400 strncmp(action
, "reshape", 7) != 0)
2404 select(fd
+1, NULL
, NULL
, &rfds
, NULL
);
2405 if (sysfs_fd_get_ll(fd
, &completed
) < 0) {
2407 goto check_progress
;
2410 /* some kernels can give an incorrectly high 'completed' number */
2411 completed
/= (info
->new_chunk
/512);
2412 completed
*= (info
->new_chunk
/512);
2413 /* Convert 'completed' back in to a 'progress' number */
2414 completed
*= reshape
->after
.data_disks
;
2416 completed
= info
->component_size
* reshape
->after
.data_disks
2419 *reshape_completed
= completed
;
2423 /* We return the need_backup flag. Caller will decide
2424 * how much - a multiple of ->backup_blocks up to *suspend_point
2427 ? (need_backup
> info
->reshape_progress
)
2428 : (need_backup
< info
->reshape_progress
);
2431 /* if we couldn't read a number from sync_completed, then
2432 * either the reshape did complete, or it aborted.
2433 * We can tell which by checking for 'none' in reshape_position.
2436 if (sysfs_get_str(info
, NULL
, "reshape_position", buf
, sizeof(buf
)) < 0
2437 || strncmp(buf
, "none", 4) != 0)
2438 return -2; /* abort */
2440 /* Maybe racing with array shutdown - check state */
2441 if (sysfs_get_str(info
, NULL
, "array_state", buf
, sizeof(buf
)) < 0
2442 || strncmp(buf
, "inactive", 8) == 0
2443 || strncmp(buf
, "clear",5) == 0)
2444 return -2; /* abort */
2445 return -1; /* complete */
2450 /* FIXME return status is never checked */
2451 static int grow_backup(struct mdinfo
*sra
,
2452 unsigned long long offset
, /* per device */
2453 unsigned long stripes
, /* per device, in old chunks */
2454 int *sources
, unsigned long long *offsets
,
2455 int disks
, int chunk
, int level
, int layout
,
2456 int dests
, int *destfd
, unsigned long long *destoffsets
,
2457 int part
, int *degraded
,
2460 /* Backup 'blocks' sectors at 'offset' on each device of the array,
2461 * to storage 'destfd' (offset 'destoffsets'), after first
2462 * suspending IO. Then allow resync to continue
2463 * over the suspended section.
2464 * Use part 'part' of the backup-super-block.
2469 unsigned long long ll
;
2471 //printf("offset %llu\n", offset);
2477 /* Check that array hasn't become degraded, else we might backup the wrong data */
2478 sysfs_get_ll(sra
, NULL
, "degraded", &ll
);
2479 new_degraded
= (int)ll
;
2480 if (new_degraded
!= *degraded
) {
2481 /* check each device to ensure it is still working */
2483 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
2484 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
2486 if (sd
->disk
.state
& (1<<MD_DISK_SYNC
)) {
2488 if (sysfs_get_str(sra
, sd
, "state", sbuf
, 20) < 0 ||
2489 strstr(sbuf
, "faulty") ||
2490 strstr(sbuf
, "in_sync") == NULL
) {
2491 /* this device is dead */
2492 sd
->disk
.state
= (1<<MD_DISK_FAULTY
);
2493 if (sd
->disk
.raid_disk
>= 0 &&
2494 sources
[sd
->disk
.raid_disk
] >= 0) {
2495 close(sources
[sd
->disk
.raid_disk
]);
2496 sources
[sd
->disk
.raid_disk
] = -1;
2501 *degraded
= new_degraded
;
2504 bsb
.arraystart2
= __cpu_to_le64(offset
* odata
);
2505 bsb
.length2
= __cpu_to_le64(stripes
* (chunk
/512) * odata
);
2507 bsb
.arraystart
= __cpu_to_le64(offset
* odata
);
2508 bsb
.length
= __cpu_to_le64(stripes
* (chunk
/512) * odata
);
2511 bsb
.magic
[15] = '2';
2512 for (i
= 0; i
< dests
; i
++)
2514 lseek64(destfd
[i
], destoffsets
[i
] + __le64_to_cpu(bsb
.devstart2
)*512, 0);
2516 lseek64(destfd
[i
], destoffsets
[i
], 0);
2518 rv
= save_stripes(sources
, offsets
,
2519 disks
, chunk
, level
, layout
,
2521 offset
*512*odata
, stripes
* chunk
* odata
,
2526 bsb
.mtime
= __cpu_to_le64(time(0));
2527 for (i
= 0; i
< dests
; i
++) {
2528 bsb
.devstart
= __cpu_to_le64(destoffsets
[i
]/512);
2530 bsb
.sb_csum
= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
));
2531 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0)
2532 bsb
.sb_csum2
= bsb_csum((char*)&bsb
,
2533 ((char*)&bsb
.sb_csum2
)-((char*)&bsb
));
2536 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
] - 4096, 0)
2537 != destoffsets
[i
] - 4096)
2539 if (write(destfd
[i
], &bsb
, 512) != 512)
2541 if (destoffsets
[i
] > 4096) {
2542 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
]+stripes
*chunk
*odata
, 0) !=
2543 destoffsets
[i
]+stripes
*chunk
*odata
)
2545 if (write(destfd
[i
], &bsb
, 512) != 512)
2555 /* in 2.6.30, the value reported by sync_completed can be
2556 * less that it should be by one stripe.
2557 * This only happens when reshape hits sync_max and pauses.
2558 * So allow wait_backup to either extent sync_max further
2559 * than strictly necessary, or return before the
2560 * sync has got quite as far as we would really like.
2561 * This is what 'blocks2' is for.
2562 * The various caller give appropriate values so that
2565 /* FIXME return value is often ignored */
2566 static int forget_backup(
2567 int dests
, int *destfd
, unsigned long long *destoffsets
,
2571 * Erase backup 'part' (which is 0 or 1)
2577 bsb
.arraystart2
= __cpu_to_le64(0);
2578 bsb
.length2
= __cpu_to_le64(0);
2580 bsb
.arraystart
= __cpu_to_le64(0);
2581 bsb
.length
= __cpu_to_le64(0);
2583 bsb
.mtime
= __cpu_to_le64(time(0));
2585 for (i
= 0; i
< dests
; i
++) {
2586 bsb
.devstart
= __cpu_to_le64(destoffsets
[i
]/512);
2587 bsb
.sb_csum
= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
));
2588 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0)
2589 bsb
.sb_csum2
= bsb_csum((char*)&bsb
,
2590 ((char*)&bsb
.sb_csum2
)-((char*)&bsb
));
2591 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
]-4096, 0) !=
2592 destoffsets
[i
]-4096)
2595 write(destfd
[i
], &bsb
, 512) != 512)
2602 static void fail(char *msg
)
2605 rv
= (write(2, msg
, strlen(msg
)) != (int)strlen(msg
));
2606 rv
|= (write(2, "\n", 1) != 1);
2610 static char *abuf
, *bbuf
;
2611 static unsigned long long abuflen
;
2612 static void validate(int afd
, int bfd
, unsigned long long offset
)
2614 /* check that the data in the backup against the array.
2615 * This is only used for regression testing and should not
2616 * be used while the array is active
2620 lseek64(bfd
, offset
- 4096, 0);
2621 if (read(bfd
, &bsb2
, 512) != 512)
2622 fail("cannot read bsb");
2623 if (bsb2
.sb_csum
!= bsb_csum((char*)&bsb2
,
2624 ((char*)&bsb2
.sb_csum
)-((char*)&bsb2
)))
2625 fail("first csum bad");
2626 if (memcmp(bsb2
.magic
, "md_backup_data", 14) != 0)
2627 fail("magic is bad");
2628 if (memcmp(bsb2
.magic
, "md_backup_data-2", 16) == 0 &&
2629 bsb2
.sb_csum2
!= bsb_csum((char*)&bsb2
,
2630 ((char*)&bsb2
.sb_csum2
)-((char*)&bsb2
)))
2631 fail("second csum bad");
2633 if (__le64_to_cpu(bsb2
.devstart
)*512 != offset
)
2634 fail("devstart is wrong");
2637 unsigned long long len
= __le64_to_cpu(bsb2
.length
)*512;
2639 if (abuflen
< len
) {
2643 if (posix_memalign((void**)&abuf
, 4096, abuflen
) ||
2644 posix_memalign((void**)&bbuf
, 4096, abuflen
)) {
2646 /* just stop validating on mem-alloc failure */
2651 lseek64(bfd
, offset
, 0);
2652 if ((unsigned long long)read(bfd
, bbuf
, len
) != len
) {
2653 //printf("len %llu\n", len);
2654 fail("read first backup failed");
2656 lseek64(afd
, __le64_to_cpu(bsb2
.arraystart
)*512, 0);
2657 if ((unsigned long long)read(afd
, abuf
, len
) != len
)
2658 fail("read first from array failed");
2659 if (memcmp(bbuf
, abuf
, len
) != 0) {
2662 printf("offset=%llu len=%llu\n",
2663 (unsigned long long)__le64_to_cpu(bsb2
.arraystart
)*512, len
);
2664 for (i
=0; i
<len
; i
++)
2665 if (bbuf
[i
] != abuf
[i
]) {
2666 printf("first diff byte %d\n", i
);
2670 fail("data1 compare failed");
2674 unsigned long long len
= __le64_to_cpu(bsb2
.length2
)*512;
2676 if (abuflen
< len
) {
2680 abuf
= malloc(abuflen
);
2681 bbuf
= malloc(abuflen
);
2684 lseek64(bfd
, offset
+__le64_to_cpu(bsb2
.devstart2
)*512, 0);
2685 if ((unsigned long long)read(bfd
, bbuf
, len
) != len
)
2686 fail("read second backup failed");
2687 lseek64(afd
, __le64_to_cpu(bsb2
.arraystart2
)*512, 0);
2688 if ((unsigned long long)read(afd
, abuf
, len
) != len
)
2689 fail("read second from array failed");
2690 if (memcmp(bbuf
, abuf
, len
) != 0)
2691 fail("data2 compare failed");
2695 int child_monitor(int afd
, struct mdinfo
*sra
, struct reshape
*reshape
,
2696 struct supertype
*st
, unsigned long blocks
,
2697 int *fds
, unsigned long long *offsets
,
2698 int dests
, int *destfd
, unsigned long long *destoffsets
)
2700 /* Monitor a reshape where backup is being performed using
2701 * 'native' mechanism - either to a backup file, or
2702 * to some space in a spare.
2706 unsigned long long speed
;
2707 unsigned long long suspend_point
, array_size
;
2708 unsigned long long backup_point
, wait_point
;
2709 unsigned long long reshape_completed
;
2711 int increasing
= reshape
->after
.data_disks
>= reshape
->before
.data_disks
;
2712 int part
= 0; /* The next part of the backup area to fill. It may already
2713 * be full, so we need to check */
2714 int level
= reshape
->level
;
2715 int layout
= reshape
->before
.layout
;
2716 int data
= reshape
->before
.data_disks
;
2717 int disks
= reshape
->before
.data_disks
+ reshape
->parity
;
2718 int chunk
= sra
->array
.chunk_size
;
2720 unsigned long stripes
;
2722 /* set up the backup-super-block. This requires the
2723 * uuid from the array.
2725 /* Find a superblock */
2726 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
2730 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
2732 dn
= map_dev(sd
->disk
.major
, sd
->disk
.minor
, 1);
2733 devfd
= dev_open(dn
, O_RDONLY
);
2736 ok
= st
->ss
->load_super(st
, devfd
, NULL
);
2742 fprintf(stderr
, Name
": Cannot find a superblock\n");
2746 memset(&bsb
, 0, 512);
2747 memcpy(bsb
.magic
, "md_backup_data-1", 16);
2748 st
->ss
->uuid_from_super(st
, (int*)&bsb
.set_uuid
);
2749 bsb
.mtime
= __cpu_to_le64(time(0));
2750 bsb
.devstart2
= blocks
;
2752 stripes
= blocks
/ (sra
->array
.chunk_size
/512) /
2753 reshape
->before
.data_disks
;
2755 if (posix_memalign((void**)&buf
, 4096, disks
* chunk
))
2756 /* Don't start the 'reshape' */
2758 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
) {
2759 sysfs_get_ll(sra
, NULL
, "sync_speed_min", &speed
);
2760 sysfs_set_num(sra
, NULL
, "sync_speed_min", 200000);
2764 array_size
= sra
->component_size
* reshape
->after
.data_disks
;
2765 backup_point
= sra
->reshape_progress
;
2768 array_size
= sra
->component_size
* reshape
->before
.data_disks
;
2769 backup_point
= array_size
;
2770 suspend_point
= array_size
;
2776 /* Want to return as soon the oldest backup slot can
2777 * be released as that allows us to start backing up
2778 * some more, providing suspend_point has been
2779 * advanced, which it should have.
2782 wait_point
= array_size
;
2783 if (part
== 0 && __le64_to_cpu(bsb
.length
) > 0)
2784 wait_point
= (__le64_to_cpu(bsb
.arraystart
) +
2785 __le64_to_cpu(bsb
.length
));
2786 if (part
== 1 && __le64_to_cpu(bsb
.length2
) > 0)
2787 wait_point
= (__le64_to_cpu(bsb
.arraystart2
) +
2788 __le64_to_cpu(bsb
.length2
));
2791 if (part
== 0 && __le64_to_cpu(bsb
.length
) > 0)
2792 wait_point
= __le64_to_cpu(bsb
.arraystart
);
2793 if (part
== 1 && __le64_to_cpu(bsb
.length2
) > 0)
2794 wait_point
= __le64_to_cpu(bsb
.arraystart2
);
2797 rv
= progress_reshape(sra
, reshape
,
2798 backup_point
, wait_point
,
2799 &suspend_point
, &reshape_completed
);
2800 /* external metadata would need to ping_monitor here */
2801 sra
->reshape_progress
= reshape_completed
;
2803 /* Clear any backup region that is before 'here' */
2805 if (reshape_completed
>= (__le64_to_cpu(bsb
.arraystart
) +
2806 __le64_to_cpu(bsb
.length
)))
2807 forget_backup(dests
, destfd
,
2809 if (reshape_completed
>= (__le64_to_cpu(bsb
.arraystart2
) +
2810 __le64_to_cpu(bsb
.length2
)))
2811 forget_backup(dests
, destfd
,
2814 if (reshape_completed
<= (__le64_to_cpu(bsb
.arraystart
)))
2815 forget_backup(dests
, destfd
,
2817 if (reshape_completed
<= (__le64_to_cpu(bsb
.arraystart2
)))
2818 forget_backup(dests
, destfd
,
2829 unsigned long long offset
;
2830 unsigned long actual_stripes
;
2831 /* Need to backup some data.
2832 * If 'part' is not used and the desired
2833 * backup size is suspended, do a backup,
2834 * then consider the next part.
2836 /* Check that 'part' is unused */
2837 if (part
== 0 && __le64_to_cpu(bsb
.length
) != 0)
2839 if (part
== 1 && __le64_to_cpu(bsb
.length2
) != 0)
2842 offset
= backup_point
/ data
;
2843 actual_stripes
= stripes
;
2845 if (offset
+ actual_stripes
* (chunk
/512) >
2846 sra
->component_size
)
2847 actual_stripes
= ((sra
->component_size
- offset
)
2849 if (offset
+ actual_stripes
* (chunk
/512) >
2853 if (offset
< actual_stripes
* (chunk
/512))
2854 actual_stripes
= offset
/ (chunk
/512);
2855 offset
-= actual_stripes
* (chunk
/512);
2856 if (offset
< suspend_point
/data
)
2859 grow_backup(sra
, offset
, actual_stripes
,
2861 disks
, chunk
, level
, layout
,
2862 dests
, destfd
, destoffsets
,
2863 part
, °raded
, buf
);
2864 validate(afd
, destfd
[0], destoffsets
[0]);
2865 /* record where 'part' is up to */
2868 backup_point
+= actual_stripes
* (chunk
/512) * data
;
2870 backup_point
-= actual_stripes
* (chunk
/512) * data
;
2874 /* FIXME maybe call progress_reshape one more time instead */
2875 abort_reshape(sra
); /* remove any remaining suspension */
2876 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
)
2877 sysfs_set_num(sra
, NULL
, "sync_speed_min", speed
);
2883 * If any spare contains md_back_data-1 which is recent wrt mtime,
2884 * write that data into the array and update the super blocks with
2885 * the new reshape_progress
2887 int Grow_restart(struct supertype
*st
, struct mdinfo
*info
, int *fdlist
, int cnt
,
2888 char *backup_file
, int verbose
)
2892 unsigned long long *offsets
;
2893 unsigned long long nstripe
, ostripe
;
2896 if (info
->new_level
!= info
->array
.level
)
2897 return 1; /* Cannot handle level changes (they are instantaneous) */
2899 odata
= info
->array
.raid_disks
- info
->delta_disks
- 1;
2900 if (info
->array
.level
== 6) odata
--; /* number of data disks */
2901 ndata
= info
->array
.raid_disks
- 1;
2902 if (info
->new_level
== 6) ndata
--;
2904 old_disks
= info
->array
.raid_disks
- info
->delta_disks
;
2906 if (info
->delta_disks
<= 0)
2907 /* Didn't grow, so the backup file must have
2911 for (i
=old_disks
-(backup_file
?1:0); i
<cnt
; i
++) {
2912 struct mdinfo dinfo
;
2915 char *devname
, namebuf
[20];
2917 /* This was a spare and may have some saved data on it.
2918 * Load the superblock, find and load the
2919 * backup_super_block.
2920 * If either fail, go on to next device.
2921 * If the backup contains no new info, just return
2922 * else restore data and update all superblocks
2924 if (i
== old_disks
-1) {
2925 fd
= open(backup_file
, O_RDONLY
);
2927 fprintf(stderr
, Name
": backup file %s inaccessible: %s\n",
2928 backup_file
, strerror(errno
));
2931 devname
= backup_file
;
2936 if (st
->ss
->load_super(st
, fd
, NULL
))
2939 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
2940 st
->ss
->free_super(st
);
2943 (dinfo
.data_offset
+ dinfo
.component_size
- 8) <<9,
2945 fprintf(stderr
, Name
": Cannot seek on device %d\n", i
);
2946 continue; /* Cannot seek */
2948 sprintf(namebuf
, "device-%d", i
);
2951 if (read(fd
, &bsb
, sizeof(bsb
)) != sizeof(bsb
)) {
2953 fprintf(stderr
, Name
": Cannot read from %s\n", devname
);
2954 continue; /* Cannot read */
2956 if (memcmp(bsb
.magic
, "md_backup_data-1", 16) != 0 &&
2957 memcmp(bsb
.magic
, "md_backup_data-2", 16) != 0) {
2959 fprintf(stderr
, Name
": No backup metadata on %s\n", devname
);
2962 if (bsb
.sb_csum
!= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
))) {
2964 fprintf(stderr
, Name
": Bad backup-metadata checksum on %s\n", devname
);
2965 continue; /* bad checksum */
2967 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0 &&
2968 bsb
.sb_csum2
!= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum2
)-((char*)&bsb
))) {
2970 fprintf(stderr
, Name
": Bad backup-metadata checksum2 on %s\n", devname
);
2971 continue; /* Bad second checksum */
2973 if (memcmp(bsb
.set_uuid
,info
->uuid
, 16) != 0) {
2975 fprintf(stderr
, Name
": Wrong uuid on backup-metadata on %s\n", devname
);
2976 continue; /* Wrong uuid */
2979 /* array utime and backup-mtime should be updated at much the same time, but it seems that
2980 * sometimes they aren't... So allow considerable flexability in matching, and allow
2981 * this test to be overridden by an environment variable.
2983 if (info
->array
.utime
> (int)__le64_to_cpu(bsb
.mtime
) + 2*60*60 ||
2984 info
->array
.utime
< (int)__le64_to_cpu(bsb
.mtime
) - 10*60) {
2985 if (check_env("MDADM_GROW_ALLOW_OLD")) {
2986 fprintf(stderr
, Name
": accepting backup with timestamp %lu "
2987 "for array with timestamp %lu\n",
2988 (unsigned long)__le64_to_cpu(bsb
.mtime
),
2989 (unsigned long)info
->array
.utime
);
2992 fprintf(stderr
, Name
": too-old timestamp on "
2993 "backup-metadata on %s\n", devname
);
2994 continue; /* time stamp is too bad */
2998 if (bsb
.magic
[15] == '1') {
2999 if (info
->delta_disks
>= 0) {
3000 /* reshape_progress is increasing */
3001 if (__le64_to_cpu(bsb
.arraystart
) + __le64_to_cpu(bsb
.length
) <
3002 info
->reshape_progress
) {
3005 fprintf(stderr
, Name
": backup-metadata found on %s but is not needed\n", devname
);
3006 continue; /* No new data here */
3009 /* reshape_progress is decreasing */
3010 if (__le64_to_cpu(bsb
.arraystart
) >=
3011 info
->reshape_progress
)
3012 goto nonew
; /* No new data here */
3015 if (info
->delta_disks
>= 0) {
3016 /* reshape_progress is increasing */
3017 if (__le64_to_cpu(bsb
.arraystart
) + __le64_to_cpu(bsb
.length
) <
3018 info
->reshape_progress
&&
3019 __le64_to_cpu(bsb
.arraystart2
) + __le64_to_cpu(bsb
.length2
) <
3020 info
->reshape_progress
)
3021 goto nonew
; /* No new data here */
3023 /* reshape_progress is decreasing */
3024 if (__le64_to_cpu(bsb
.arraystart
) >=
3025 info
->reshape_progress
&&
3026 __le64_to_cpu(bsb
.arraystart2
) >=
3027 info
->reshape_progress
)
3028 goto nonew
; /* No new data here */
3031 if (lseek64(fd
, __le64_to_cpu(bsb
.devstart
)*512, 0)< 0) {
3034 fprintf(stderr
, Name
": Failed to verify secondary backup-metadata block on %s\n",
3036 continue; /* Cannot seek */
3038 /* There should be a duplicate backup superblock 4k before here */
3039 if (lseek64(fd
, -4096, 1) < 0 ||
3040 read(fd
, &bsb2
, sizeof(bsb2
)) != sizeof(bsb2
))
3041 goto second_fail
; /* Cannot find leading superblock */
3042 if (bsb
.magic
[15] == '1')
3043 bsbsize
= offsetof(struct mdp_backup_super
, pad1
);
3045 bsbsize
= offsetof(struct mdp_backup_super
, pad
);
3046 if (memcmp(&bsb2
, &bsb
, bsbsize
) != 0)
3047 goto second_fail
; /* Cannot find leading superblock */
3049 /* Now need the data offsets for all devices. */
3050 offsets
= malloc(sizeof(*offsets
)*info
->array
.raid_disks
);
3051 for(j
=0; j
<info
->array
.raid_disks
; j
++) {
3054 if (st
->ss
->load_super(st
, fdlist
[j
], NULL
))
3055 /* FIXME should be this be an error */
3057 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
3058 st
->ss
->free_super(st
);
3059 offsets
[j
] = dinfo
.data_offset
* 512;
3061 printf(Name
": restoring critical section\n");
3063 if (restore_stripes(fdlist
, offsets
,
3064 info
->array
.raid_disks
,
3068 fd
, __le64_to_cpu(bsb
.devstart
)*512,
3069 __le64_to_cpu(bsb
.arraystart
)*512,
3070 __le64_to_cpu(bsb
.length
)*512)) {
3071 /* didn't succeed, so giveup */
3073 fprintf(stderr
, Name
": Error restoring backup from %s\n",
3078 if (bsb
.magic
[15] == '2' &&
3079 restore_stripes(fdlist
, offsets
,
3080 info
->array
.raid_disks
,
3084 fd
, __le64_to_cpu(bsb
.devstart
)*512 +
3085 __le64_to_cpu(bsb
.devstart2
)*512,
3086 __le64_to_cpu(bsb
.arraystart2
)*512,
3087 __le64_to_cpu(bsb
.length2
)*512)) {
3088 /* didn't succeed, so giveup */
3090 fprintf(stderr
, Name
": Error restoring second backup from %s\n",
3096 /* Ok, so the data is restored. Let's update those superblocks. */
3098 if (info
->delta_disks
>= 0) {
3099 info
->reshape_progress
= __le64_to_cpu(bsb
.arraystart
) +
3100 __le64_to_cpu(bsb
.length
);
3101 if (bsb
.magic
[15] == '2') {
3102 unsigned long long p2
= __le64_to_cpu(bsb
.arraystart2
) +
3103 __le64_to_cpu(bsb
.length2
);
3104 if (p2
> info
->reshape_progress
)
3105 info
->reshape_progress
= p2
;
3108 info
->reshape_progress
= __le64_to_cpu(bsb
.arraystart
);
3109 if (bsb
.magic
[15] == '2') {
3110 unsigned long long p2
= __le64_to_cpu(bsb
.arraystart2
);
3111 if (p2
< info
->reshape_progress
)
3112 info
->reshape_progress
= p2
;
3115 for (j
=0; j
<info
->array
.raid_disks
; j
++) {
3116 if (fdlist
[j
] < 0) continue;
3117 if (st
->ss
->load_super(st
, fdlist
[j
], NULL
))
3119 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
3120 dinfo
.reshape_progress
= info
->reshape_progress
;
3121 st
->ss
->update_super(st
, &dinfo
,
3122 "_reshape_progress",
3124 st
->ss
->store_super(st
, fdlist
[j
]);
3125 st
->ss
->free_super(st
);
3129 /* Didn't find any backup data, try to see if any
3132 if (info
->delta_disks
< 0) {
3133 /* When shrinking, the critical section is at the end.
3134 * So see if we are before the critical section.
3136 unsigned long long first_block
;
3137 nstripe
= ostripe
= 0;
3139 while (ostripe
>= nstripe
) {
3140 ostripe
+= info
->array
.chunk_size
/ 512;
3141 first_block
= ostripe
* odata
;
3142 nstripe
= first_block
/ ndata
/ (info
->new_chunk
/512) *
3143 (info
->new_chunk
/512);
3146 if (info
->reshape_progress
>= first_block
)
3149 if (info
->delta_disks
> 0) {
3150 /* See if we are beyond the critical section. */
3151 unsigned long long last_block
;
3152 nstripe
= ostripe
= 0;
3154 while (nstripe
>= ostripe
) {
3155 nstripe
+= info
->new_chunk
/ 512;
3156 last_block
= nstripe
* ndata
;
3157 ostripe
= last_block
/ odata
/ (info
->array
.chunk_size
/512) *
3158 (info
->array
.chunk_size
/512);
3161 if (info
->reshape_progress
>= last_block
)
3164 /* needed to recover critical section! */
3166 fprintf(stderr
, Name
": Failed to find backup of critical section\n");
3170 int Grow_continue(int mdfd
, struct supertype
*st
, struct mdinfo
*info
,
3173 int err
= sysfs_set_str(info
, NULL
, "array_state", "readonly");
3176 return reshape_array(NULL
, mdfd
, "array", st
, info
, 1, backup_file
, 0, 0, 1);