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
->after
.data_disks
= re
->before
.data_disks
;
974 re
->before
.layout
= 0;
975 re
->backup_blocks
= 0;
979 /* RAID0 can be converted to RAID10, or to RAID456 */
980 if (info
->new_level
== 10) {
981 if (info
->new_layout
== UnSet
&& info
->delta_disks
== UnSet
) {
982 /* Assume near=2 layout */
983 info
->new_layout
= 0x102;
984 info
->delta_disks
= info
->array
.raid_disks
;
986 if (info
->new_layout
== UnSet
) {
987 int copies
= 1 + (info
->delta_disks
988 / info
->array
.raid_disks
);
989 if (info
->array
.raid_disks
* (copies
-1)
990 != info
->delta_disks
)
991 return "Impossible number of devices"
992 " for RAID0->RAID10";
993 info
->new_layout
= 0x100 + copies
;
995 if (info
->delta_disks
== UnSet
) {
996 int copies
= info
->new_layout
& 0xff;
997 if (info
->new_layout
!= 0x100 + copies
)
998 return "New layout impossible"
999 " for RAID0->RAID10";;
1000 info
->delta_disks
= (copies
- 1) *
1001 info
->array
.raid_disks
;
1003 if (info
->new_chunk
&&
1004 info
->new_chunk
!= info
->array
.chunk_size
)
1005 return "Cannot change chunk-size with RAID0->RAID10";
1009 re
->before
.data_disks
= (info
->array
.raid_disks
+
1011 re
->after
.data_disks
= re
->before
.data_disks
;
1012 re
->before
.layout
= info
->new_layout
;
1013 re
->backup_blocks
= 0;
1017 /* RAID0 can also covert to RAID0/4/5/6 by first converting to
1018 * a raid4 style layout of the final level.
1020 switch (info
->new_level
) {
1024 re
->before
.layout
= 0;
1028 re
->before
.layout
= ALGORITHM_PARITY_N
;
1032 re
->before
.layout
= ALGORITHM_PARITY_N
;
1035 return "Impossible level change requested";
1037 re
->before
.data_disks
= info
->array
.raid_disks
;
1038 /* determining 'after' layout happens outside this 'switch' */
1042 info
->array
.layout
= ALGORITHM_PARITY_N
;
1044 switch (info
->new_level
) {
1046 re
->level
= info
->array
.level
;
1047 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1048 re
->before
.layout
= info
->array
.layout
;
1052 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1053 re
->before
.layout
= info
->array
.layout
;
1057 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1058 switch (info
->array
.layout
) {
1059 case ALGORITHM_LEFT_ASYMMETRIC
:
1060 re
->before
.layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
1062 case ALGORITHM_RIGHT_ASYMMETRIC
:
1063 re
->before
.layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
1065 case ALGORITHM_LEFT_SYMMETRIC
:
1066 re
->before
.layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
1068 case ALGORITHM_RIGHT_SYMMETRIC
:
1069 re
->before
.layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
1071 case ALGORITHM_PARITY_0
:
1072 re
->before
.layout
= ALGORITHM_PARITY_0_6
;
1074 case ALGORITHM_PARITY_N
:
1075 re
->before
.layout
= ALGORITHM_PARITY_N_6
;
1078 return "Cannot convert an array with this layout";
1082 if (info
->array
.raid_disks
!= 2)
1083 return "Can only convert a 2-device array to RAID1";
1085 re
->before
.data_disks
= 2;
1086 re
->before
.layout
= 0;
1089 return "Impossible level change requested";
1093 switch (info
->new_level
) {
1098 re
->before
.data_disks
= info
->array
.raid_disks
- 2;
1099 re
->before
.layout
= info
->array
.layout
;
1102 return "Impossible level change requested";
1107 /* If we reached here then it looks like a re-stripe is
1108 * happening. We have determined the intermediate level
1109 * and initial raid_disks/layout and stored these in 're'.
1111 * We need to deduce the final layout that can be atomically
1112 * converted to the end state.
1114 switch (info
->new_level
) {
1116 /* We can only get to RAID0 from RAID4 or RAID5
1117 * with appropriate layout and one extra device
1119 if (re
->level
!= 4 && re
->level
!= 5)
1120 return "Cannot covert to RAID0 from this level";
1121 if (info
->delta_disks
== UnSet
)
1122 re
->after
.data_disks
= re
->before
.data_disks
;
1124 re
->after
.data_disks
=
1125 info
->array
.raid_disks
+ info
->delta_disks
;
1126 switch (re
->level
) {
1128 re
->after
.layout
= 0 ; break;
1130 re
->after
.layout
= ALGORITHM_PARITY_N
; break;
1135 /* We can only get to RAID4 from RAID5 */
1136 if (re
->level
!= 4 && re
->level
!= 5)
1137 return "Cannot convert to RAID4 from this level";
1138 if (info
->delta_disks
== UnSet
)
1139 re
->after
.data_disks
= re
->before
.data_disks
;
1141 re
->after
.data_disks
=
1142 re
->before
.data_disks
+ info
->delta_disks
;
1143 switch (re
->level
) {
1145 re
->after
.layout
= 0 ; break;
1147 re
->after
.layout
= ALGORITHM_PARITY_N
; break;
1152 /* We get to RAID5 for RAID5 or RAID6 */
1153 if (re
->level
!= 5 && re
->level
!= 6)
1154 return "Cannot convert to RAID5 from this level";
1155 if (info
->delta_disks
== UnSet
)
1156 re
->after
.data_disks
= re
->before
.data_disks
;
1157 else if (re
->level
== 5)
1158 re
->after
.data_disks
=
1159 re
->before
.data_disks
+ info
->delta_disks
;
1161 re
->after
.data_disks
=
1162 info
->array
.raid_disks
+ info
->delta_disks
- 1;
1163 switch (re
->level
) {
1165 if (info
->new_layout
== UnSet
)
1166 re
->after
.layout
= re
->before
.layout
;
1168 re
->after
.layout
= info
->new_layout
;
1171 if (info
->new_layout
== UnSet
)
1172 info
->new_layout
= re
->before
.layout
;
1174 /* after.layout needs to be raid6 version of new_layout */
1175 if (info
->new_layout
== ALGORITHM_PARITY_N
)
1176 re
->after
.layout
= ALGORITHM_PARITY_N
;
1179 char *ls
= map_num(r5layout
, info
->new_layout
);
1181 strcat(strcpy(layout
, ls
), "-6");
1182 l
= map_name(r6layout
, layout
);
1184 return "Cannot find RAID6 layout"
1186 re
->after
.layout
= l
;
1192 /* We must already be at level 6 */
1194 return "Impossible level change";
1195 if (info
->delta_disks
== UnSet
)
1196 re
->after
.data_disks
= re
->before
.data_disks
;
1198 re
->after
.data_disks
= (info
->array
.raid_disks
+
1199 info
->delta_disks
) - 2;
1200 if (info
->new_layout
== UnSet
)
1201 re
->after
.layout
= info
->array
.layout
;
1203 re
->after
.layout
= info
->new_layout
;
1206 return "Impossible level change requested";
1208 switch (re
->level
) {
1209 case 6: re
->parity
= 2; break;
1211 case 5: re
->parity
= 1; break;
1212 default: re
->parity
= 0; break;
1214 /* So we have a restripe operation, we need to calculate the number
1215 * of blocks per reshape operation.
1217 if (info
->new_chunk
== 0)
1218 info
->new_chunk
= info
->array
.chunk_size
;
1219 if (re
->after
.data_disks
== re
->before
.data_disks
&&
1220 re
->after
.layout
== re
->before
.layout
&&
1221 info
->new_chunk
== info
->array
.chunk_size
) {
1222 /* Nothing to change */
1223 re
->backup_blocks
= 0;
1226 if (re
->after
.data_disks
== 1 && re
->before
.data_disks
== 1) {
1227 /* chunks can layout changes make no difference */
1228 re
->backup_blocks
= 0;
1232 if (re
->after
.data_disks
== re
->before
.data_disks
&&
1233 get_linux_version() < 2006032)
1234 return "in-place reshape is not safe before 2.6.32 - sorry.";
1236 if (re
->after
.data_disks
< re
->before
.data_disks
&&
1237 get_linux_version() < 2006030)
1238 return "reshape to fewer devices is not supported before 2.6.32 - sorry.";
1240 re
->backup_blocks
= compute_backup_blocks(
1241 info
->new_chunk
, info
->array
.chunk_size
,
1242 re
->after
.data_disks
,
1243 re
->before
.data_disks
);
1245 re
->new_size
= info
->component_size
* re
->after
.data_disks
;
1249 static int reshape_array(char *container
, int fd
, char *devname
,
1250 struct supertype
*st
, struct mdinfo
*info
,
1251 int force
, char *backup_file
, int quiet
, int forked
,
1253 static int reshape_container(char *container
, int cfd
, char *devname
,
1254 struct supertype
*st
,
1255 struct mdinfo
*info
,
1260 int Grow_reshape(char *devname
, int fd
, int quiet
, char *backup_file
,
1262 int level
, char *layout_str
, int chunksize
, int raid_disks
,
1265 /* Make some changes in the shape of an array.
1266 * The kernel must support the change.
1268 * There are three different changes. Each can trigger
1269 * a resync or recovery so we freeze that until we have
1270 * requested everything (if kernel supports freezing - 2.6.30).
1272 * - change size (i.e. component_size)
1274 * - change layout/chunksize/ndisks
1276 * The last can require a reshape. It is different on different
1277 * levels so we need to check the level before actioning it.
1278 * Some times the level change needs to be requested after the
1279 * reshape (e.g. raid6->raid5, raid5->raid0)
1282 struct mdu_array_info_s array
;
1284 struct supertype
*st
;
1285 char *subarray
= NULL
;
1289 char *container
= NULL
;
1290 char container_buf
[20];
1296 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) < 0) {
1297 fprintf(stderr
, Name
": %s is not an active md array - aborting\n",
1303 (chunksize
|| level
!= UnSet
|| layout_str
|| raid_disks
)) {
1304 fprintf(stderr
, Name
": cannot change component size at the same time "
1305 "as other changes.\n"
1306 " Change size first, then check data is intact before "
1307 "making other changes.\n");
1311 if (raid_disks
&& raid_disks
< array
.raid_disks
&& array
.level
> 1 &&
1312 get_linux_version() < 2006032 &&
1313 !check_env("MDADM_FORCE_FEWER")) {
1314 fprintf(stderr
, Name
": reducing the number of devices is not safe before Linux 2.6.32\n"
1315 " Please use a newer kernel\n");
1319 st
= super_by_fd(fd
, &subarray
);
1321 fprintf(stderr
, Name
": Unable to determine metadata format for %s\n", devname
);
1324 if (raid_disks
> st
->max_devs
) {
1325 fprintf(stderr
, Name
": Cannot increase raid-disks on this array"
1326 " beyond %d\n", st
->max_devs
);
1330 /* in the external case we need to check that the requested reshape is
1331 * supported, and perform an initial check that the container holds the
1332 * pre-requisite spare devices (mdmon owns final validation)
1334 if (st
->ss
->external
) {
1339 container_dev
= st
->container_dev
;
1340 cfd
= open_dev_excl(st
->container_dev
);
1342 container_dev
= st
->devnum
;
1344 cfd
= open_dev_excl(st
->devnum
);
1348 fprintf(stderr
, Name
": Unable to open container for %s\n",
1354 fmt_devname(container_buf
, container_dev
);
1355 container
= container_buf
;
1357 rv
= st
->ss
->load_container(st
, cfd
, NULL
);
1360 fprintf(stderr
, Name
": Cannot read superblock for %s\n",
1366 if (mdmon_running(container_dev
))
1367 st
->update_tail
= &st
->updates
;
1370 if (raid_disks
> array
.raid_disks
&&
1371 array
.spare_disks
< (raid_disks
- array
.raid_disks
) &&
1374 Name
": Need %d spare%s to avoid degraded array,"
1375 " and only have %d.\n"
1376 " Use --force to over-ride this check.\n",
1377 raid_disks
- array
.raid_disks
,
1378 raid_disks
- array
.raid_disks
== 1 ? "" : "s",
1383 sra
= sysfs_read(fd
, 0, GET_LEVEL
| GET_DISKS
| GET_DEVS
1384 | GET_STATE
| GET_VERSION
);
1386 if (st
->ss
->external
&& subarray
== NULL
) {
1387 array
.level
= LEVEL_CONTAINER
;
1388 sra
->array
.level
= LEVEL_CONTAINER
;
1391 fprintf(stderr
, Name
": failed to read sysfs parameters for %s\n",
1395 frozen
= freeze(st
);
1397 /* freeze() already spewed the reason */
1399 } else if (frozen
< 0) {
1400 fprintf(stderr
, Name
": %s is performing resync/recovery and cannot"
1401 " be reshaped\n", devname
);
1405 /* ========= set size =============== */
1406 if (size
>= 0 && (size
== 0 || size
!= array
.size
)) {
1407 long long orig_size
= array
.size
;
1409 if (reshape_super(st
, size
, UnSet
, UnSet
, 0, 0, NULL
, devname
, !quiet
)) {
1415 if (array
.size
!= size
) {
1416 /* got truncated to 32bit, write to
1417 * component_size instead
1420 rv
= sysfs_set_num(sra
, NULL
,
1421 "component_size", size
);
1425 rv
= ioctl(fd
, SET_ARRAY_INFO
, &array
);
1429 /* restore metadata */
1430 if (reshape_super(st
, orig_size
, UnSet
, UnSet
, 0, 0,
1431 NULL
, devname
, !quiet
) == 0)
1433 fprintf(stderr
, Name
": Cannot set device size for %s: %s\n",
1434 devname
, strerror(err
));
1436 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
1437 fprintf(stderr
, " Bitmap must be removed before size can be changed\n");
1441 ioctl(fd
, GET_ARRAY_INFO
, &array
);
1442 size
= get_component_size(fd
)/2;
1446 fprintf(stderr
, Name
": component size of %s has been set to %lluK\n",
1449 } else if (array
.level
!= LEVEL_CONTAINER
) {
1450 size
= get_component_size(fd
)/2;
1455 /* ========= check for Raid10 -> Raid0 conversion ===============
1456 * current implementation assumes that following conditions must be met:
1458 * - near_copies == 2
1460 if (level
== 0 && array
.level
== 10 && sra
&&
1461 array
.layout
== ((1 << 8) + 2) && !(array
.raid_disks
& 1)) {
1463 err
= remove_disks_on_raid10_to_raid0_takeover(st
, sra
, array
.layout
);
1465 dprintf(Name
": Array cannot be reshaped\n");
1474 sysfs_init(&info
, fd
, NoMdDev
);
1475 strcpy(info
.text_version
, sra
->text_version
);
1476 info
.component_size
= size
*2;
1477 info
.new_level
= level
;
1478 info
.new_chunk
= chunksize
* 1024;
1480 info
.delta_disks
= raid_disks
- info
.array
.raid_disks
;
1482 info
.delta_disks
= UnSet
;
1483 if (layout_str
== NULL
) {
1484 info
.new_layout
= UnSet
;
1485 if (info
.array
.level
== 6 &&
1486 (info
.new_level
== 6 || info
.new_level
== UnSet
) &&
1487 info
.array
.layout
>= 16) {
1488 fprintf(stderr
, Name
1489 ": %s has a non-standard layout. If you"
1490 " wish to preserve this\n"
1491 " during the reshape, please specify"
1492 " --layout=preserve\n"
1493 " If you want to change it, specify a"
1494 " layout or use --layout=normalise\n",
1499 } else if (strcmp(layout_str
, "normalise") == 0 ||
1500 strcmp(layout_str
, "normalize") == 0) {
1501 /* If we have a -6 RAID6 layout, remove the '-6'. */
1502 info
.new_layout
= UnSet
;
1503 if (info
.array
.level
== 6 && info
.new_level
== UnSet
) {
1505 strcpy(l
, map_num(r6layout
, info
.array
.layout
));
1506 h
= strrchr(l
, '-');
1507 if (h
&& strcmp(h
, "-6") == 0) {
1509 info
.new_layout
= map_name(r6layout
, l
);
1512 } else if (strcmp(layout_str
, "preserve") == 0) {
1513 info
.new_layout
= UnSet
;
1515 int l
= info
.new_level
;
1517 l
= info
.array
.level
;
1520 info
.new_layout
= map_name(r5layout
, layout_str
);
1523 info
.new_layout
= map_name(r6layout
, layout_str
);
1526 info
.new_layout
= parse_layout_10(layout_str
);
1529 info
.new_layout
= parse_layout_faulty(layout_str
);
1532 fprintf(stderr
, Name
": layout not meaningful"
1533 " with this level\n");
1537 if (info
.new_layout
== UnSet
) {
1538 fprintf(stderr
, Name
": layout %s not understood"
1539 " for this level\n",
1546 if (array
.level
== LEVEL_CONTAINER
) {
1547 /* This change is to be applied to every array in the
1548 * container. This is only needed when the metadata imposes
1549 * restraints of the various arrays in the container.
1550 * Currently we only know that IMSM requires all arrays
1551 * to have the same number of devices so changing the
1552 * number of devices (On-Line Capacity Expansion) must be
1553 * performed at the level of the container
1555 rv
= reshape_container(container
, fd
, devname
, st
, &info
,
1556 force
, backup_file
, quiet
);
1559 /* Impose these changes on a single array. First
1560 * check that the metadata is OK with the change. */
1562 if (reshape_super(st
, info
.component_size
, info
.new_level
,
1563 info
.new_layout
, info
.new_chunk
,
1564 info
.array
.raid_disks
+ info
.delta_disks
,
1565 backup_file
, devname
, quiet
)) {
1570 rv
= reshape_array(container
, fd
, devname
, st
, &info
, force
,
1571 backup_file
, quiet
, 0, 0);
1580 static int reshape_array(char *container
, int fd
, char *devname
,
1581 struct supertype
*st
, struct mdinfo
*info
,
1583 char *backup_file
, int quiet
, int forked
,
1586 struct reshape reshape
;
1589 int orig_level
= UnSet
;
1592 struct mdu_array_info_s array
;
1596 unsigned long long *offsets
;
1600 unsigned long blocks
;
1601 unsigned long cache
;
1602 unsigned long long array_size
;
1604 struct mdinfo
*sra
= NULL
;
1606 msg
= analyse_change(info
, &reshape
);
1608 fprintf(stderr
, Name
": %s\n", msg
);
1611 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
1612 dprintf("Canot get array information.\n");
1617 /* reshape already started. just skip to monitoring the reshape */
1618 if (reshape
.backup_blocks
== 0)
1622 spares_needed
= max(reshape
.before
.data_disks
,
1623 reshape
.after
.data_disks
)
1624 + reshape
.parity
- array
.raid_disks
;
1627 info
->new_level
> 0 &&
1628 spares_needed
> info
->array
.spare_disks
) {
1630 Name
": Need %d spare%s to avoid degraded array,"
1631 " and only have %d.\n"
1632 " Use --force to over-ride this check.\n",
1634 spares_needed
== 1 ? "" : "s",
1635 info
->array
.spare_disks
);
1639 if (reshape
.level
!= info
->array
.level
) {
1640 char *c
= map_num(pers
, reshape
.level
);
1645 err
= sysfs_set_str(info
, NULL
, "level", c
);
1648 fprintf(stderr
, Name
": %s: could not set level to %s\n",
1651 (info
->array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
1652 fprintf(stderr
, " Bitmap must be removed"
1653 " before level can be changed\n");
1657 fprintf(stderr
, Name
": level of %s changed to %s\n",
1659 orig_level
= info
->array
.level
;
1662 if (reshape
.level
> 0 && st
->ss
->external
&&
1663 !mdmon_running(st
->container_dev
)) {
1664 start_mdmon(st
->container_dev
);
1665 ping_monitor(container
);
1668 /* ->reshape_super might have chosen some spares from the
1669 * container that it wants to be part of the new array.
1670 * We can collect them with ->container_content and give
1671 * them to the kernel.
1673 if (st
->ss
->reshape_super
&& st
->ss
->container_content
) {
1674 char *subarray
= strchr(info
->text_version
+1, '/')+1;
1675 struct mdinfo
*info2
=
1676 st
->ss
->container_content(st
, subarray
);
1680 sysfs_init(info2
, fd
, st
->devnum
);
1681 for (d
= info2
->devs
; d
; d
= d
->next
) {
1682 if (d
->disk
.state
== 0 &&
1683 d
->disk
.raid_disk
>= 0) {
1684 /* This is a spare that wants to
1685 * be part of the array.
1687 add_disk(fd
, st
, info2
, d
);
1694 if (reshape
.backup_blocks
== 0) {
1695 /* No restriping needed, but we might need to impose
1696 * some more changes: layout, raid_disks, chunk_size
1698 if (info
->new_layout
!= UnSet
&&
1699 info
->new_layout
!= info
->array
.layout
) {
1700 info
->array
.layout
= info
->new_layout
;
1701 if (ioctl(fd
, SET_ARRAY_INFO
, &info
->array
) != 0) {
1702 fprintf(stderr
, Name
": failed to set new layout\n");
1705 printf("layout for %s set to %d\n",
1706 devname
, info
->array
.layout
);
1708 if (info
->delta_disks
!= UnSet
&&
1709 info
->delta_disks
!= 0) {
1710 info
->array
.raid_disks
+= info
->delta_disks
;
1711 if (ioctl(fd
, SET_ARRAY_INFO
, &info
->array
) != 0) {
1712 fprintf(stderr
, Name
": failed to set raid disks\n");
1715 printf("raid_disks for %s set to %d\n",
1716 devname
, info
->array
.raid_disks
);
1718 if (info
->new_chunk
!= 0 &&
1719 info
->new_chunk
!= info
->array
.chunk_size
) {
1720 if (sysfs_set_num(info
, NULL
,
1721 "chunk_size", info
->new_chunk
) != 0) {
1722 fprintf(stderr
, Name
": failed to set chunk size\n");
1725 printf("chunk size for %s set to %d\n",
1726 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 sra
->reshape_progress
= 0;
1874 if (reshape
.after
.data_disks
< reshape
.before
.data_disks
)
1875 /* start from the end of the new array */
1876 sra
->reshape_progress
= (sra
->component_size
1877 * reshape
.after
.data_disks
);
1879 if (info
->reshape_active
)
1880 sra
->reshape_progress
= info
->reshape_progress
;
1881 else if (info
->array
.chunk_size
== info
->new_chunk
&&
1882 reshape
.before
.layout
== reshape
.after
.layout
&&
1883 st
->ss
->external
== 0) {
1884 array
.raid_disks
= reshape
.after
.data_disks
+ reshape
.parity
;
1885 if (ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
1889 Name
": Cannot set device shape for %s: %s\n",
1890 devname
, strerror(errno
));
1893 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
1895 " Bitmap must be removed before"
1896 " shape can be changed\n");
1901 /* set them all just in case some old 'new_*' value
1902 * persists from some earlier problem
1905 if (sysfs_set_num(sra
, NULL
, "chunk_size", info
->new_chunk
) < 0)
1907 if (!err
&& sysfs_set_num(sra
, NULL
, "layout",
1908 reshape
.after
.layout
) < 0)
1910 if (!err
&& subarray_set_num(container
, sra
, "raid_disks",
1911 reshape
.after
.data_disks
+
1912 reshape
.parity
) < 0)
1915 fprintf(stderr
, Name
": Cannot set device shape for %s\n",
1919 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
1921 " Bitmap must be removed before"
1922 " shape can be changed\n");
1929 sysfs_set_str(sra
, NULL
, "array_state", "active");
1931 /* Now we just need to kick off the reshape and watch, while
1932 * handling backups of the data...
1933 * This is all done by a forked background process.
1935 switch(forked
? 0 : fork()) {
1937 fprintf(stderr
, Name
": Cannot run child to monitor reshape: %s\n",
1948 if (check_env("MDADM_GROW_VERIFY"))
1949 fd
= open(devname
, O_RDONLY
| O_DIRECT
);
1952 mlockall(MCL_FUTURE
);
1954 odisks
= reshape
.before
.data_disks
+ reshape
.parity
;
1956 if (st
->ss
->external
) {
1957 /* metadata handler takes it from here */
1958 done
= st
->ss
->manage_reshape(
1959 fd
, sra
, &reshape
, st
, blocks
,
1961 d
- odisks
, fdlist
+odisks
,
1964 done
= child_monitor(
1965 fd
, sra
, &reshape
, st
, blocks
,
1967 d
- odisks
, fdlist
+odisks
,
1970 if (backup_file
&& done
)
1971 unlink(backup_file
);
1977 if (!st
->ss
->external
&&
1978 !(reshape
.before
.data_disks
!= reshape
.after
.data_disks
1979 && info
->custom_array_size
) &&
1980 info
->new_level
== reshape
.level
&&
1982 /* no need to wait for the reshape to finish as
1983 * there is nothing more to do.
1989 if (st
->ss
->external
) {
1990 /* Re-load the metadata as much could have changed */
1991 int cfd
= open_dev(st
->container_dev
);
1993 ping_monitor(container
);
1994 st
->ss
->free_super(st
);
1995 st
->ss
->load_container(st
, cfd
, container
);
2000 /* set new array size if required customer_array_size is used
2003 if (reshape
.before
.data_disks
!=
2004 reshape
.after
.data_disks
&&
2005 info
->custom_array_size
) {
2006 struct mdinfo
*info2
;
2007 char *subarray
= strchr(info
->text_version
+1, '/')+1;
2009 info2
= st
->ss
->container_content(st
, subarray
);
2011 unsigned long long current_size
= 0;
2012 unsigned long long new_size
=
2013 info2
->custom_array_size
/2;
2015 if (sysfs_get_ll(sra
,
2018 ¤t_size
) == 0 &&
2019 new_size
> current_size
) {
2020 if (sysfs_set_num(sra
, NULL
,
2021 "array_size", new_size
)
2023 dprintf("Error: Cannot"
2026 dprintf("Array size "
2028 dprintf(" from %llu to %llu.\n",
2029 current_size
, new_size
);
2035 if (info
->new_level
!= reshape
.level
) {
2037 c
= map_num(pers
, info
->new_level
);
2039 err
= sysfs_set_str(sra
, NULL
, "level", c
);
2041 fprintf(stderr
, Name\
2042 ": %s: could not set level "
2043 "to %s\n", devname
, c
);
2052 if (orig_level
!= UnSet
&& sra
) {
2053 c
= map_num(pers
, orig_level
);
2054 if (c
&& sysfs_set_str(sra
, NULL
, "level", c
) == 0)
2055 fprintf(stderr
, Name
": aborting level change\n");
2062 int reshape_container(char *container
, int cfd
, char *devname
,
2063 struct supertype
*st
,
2064 struct mdinfo
*info
,
2069 struct mdinfo
*cc
= NULL
;
2071 /* component_size is not meaningful for a container,
2072 * so pass '-1' meaning 'no change'
2074 if (reshape_super(st
, -1, info
->new_level
,
2075 info
->new_layout
, info
->new_chunk
,
2076 info
->array
.raid_disks
+ info
->delta_disks
,
2077 backup_file
, devname
, quiet
))
2082 /* ping monitor to be sure that update is on disk
2084 ping_monitor(container
);
2087 case -1: /* error */
2088 perror("Cannot fork to complete reshape\n");
2090 default: /* parent */
2091 printf(Name
": multi-array reshape continues in background\n");
2098 /* For each member array with reshape_active,
2099 * we need to perform the reshape.
2100 * We pick the first array that needs reshaping and
2101 * reshape it. reshape_array() will re-read the metadata
2102 * so the next time through a different array should be
2103 * ready for reshape.
2105 struct mdinfo
*content
;
2108 struct mdstat_ent
*mdstat
;
2113 cc
= st
->ss
->container_content(st
, NULL
);
2115 for (content
= cc
; content
; content
= content
->next
) {
2117 if (!content
->reshape_active
)
2120 subarray
= strchr(content
->text_version
+1, '/')+1;
2121 mdstat
= mdstat_by_subdev(subarray
,
2122 devname2devnum(container
));
2130 fd
= open_dev(mdstat
->devnum
);
2133 adev
= map_dev(dev2major(mdstat
->devnum
),
2134 dev2minor(mdstat
->devnum
),
2137 adev
= content
->text_version
;
2139 sysfs_init(content
, fd
, mdstat
->devnum
);
2141 rv
= reshape_array(container
, fd
, adev
, st
,
2143 backup_file
, quiet
, 1, 0);
2154 * We run a child process in the background which performs the following
2156 * - wait for resync to reach a certain point
2157 * - suspend io to the following section
2158 * - backup that section
2159 * - allow resync to proceed further
2161 * - discard the backup.
2163 * When are combined in slightly different ways in the three cases.
2165 * - suspend/backup/allow/wait/resume/discard
2167 * - allow/wait/suspend/backup/allow/wait/resume/discard
2169 * - wait/resume/discard/suspend/backup/allow
2171 * suspend/backup/allow always come together
2172 * wait/resume/discard do too.
2173 * For the same-size case we have two backups to improve flow.
2177 int progress_reshape(struct mdinfo
*info
, struct reshape
*reshape
,
2178 unsigned long long backup_point
,
2179 unsigned long long wait_point
,
2180 unsigned long long *suspend_point
,
2181 unsigned long long *reshape_completed
)
2183 /* This function is called repeatedly by the reshape manager.
2184 * It determines how much progress can safely be made and allows
2186 * - 'info' identifies the array and particularly records in
2187 * ->reshape_progress the metadata's knowledge of progress
2188 * This is a sector offset from the start of the array
2189 * of the next array block to be relocated. This number
2190 * may increase from 0 or decrease from array_size, depending
2191 * on the type of reshape that is happening.
2192 * Note that in contrast, 'sync_completed' is a block count of the
2193 * reshape so far. It gives the distance between the start point
2194 * (head or tail of device) and the next place that data will be
2195 * written. It always increases.
2196 * - 'reshape' is the structure created by analyse_change
2197 * - 'backup_point' shows how much the metadata manager has backed-up
2198 * data. For reshapes with increasing progress, it is the next address
2199 * to be backed up, previous addresses have been backed-up. For
2200 * decreasing progress, it is the earliest address that has been
2201 * backed up - later address are also backed up.
2202 * So addresses between reshape_progress and backup_point are
2203 * backed up providing those are in the 'correct' order.
2204 * - 'wait_point' is an array address. When reshape_completed
2205 * passes this point, progress_reshape should return. It might
2206 * return earlier if it determines that ->reshape_progress needs
2207 * to be updated or further backup is needed.
2208 * - suspend_point is maintained by progress_reshape and the caller
2209 * should not touch it except to initialise to zero.
2210 * It is an array address and it only increases in 2.6.37 and earlier.
2211 * This makes it difficult to handle reducing reshapes with
2212 * external metadata.
2213 * However: it is similar to backup_point in that it records the
2214 * other end of a suspended region from reshape_progress.
2215 * it is moved to extend the region that is safe to backup and/or
2217 * - reshape_completed is read from sysfs and returned. The caller
2218 * should copy this into ->reshape_progress when it has reason to
2219 * believe that the metadata knows this, and any backup outside this
2223 * 1 if more data from backup_point - but only as far as suspend_point,
2224 * should be backed up
2225 * 0 if things are progressing smoothly
2226 * -1 if the reshape is finished, either because it is all done,
2227 * or due to an error.
2230 int advancing
= (reshape
->after
.data_disks
2231 >= reshape
->before
.data_disks
);
2232 unsigned long long need_backup
; /* All data between start of array and
2233 * here will at some point need to
2236 unsigned long long read_offset
, write_offset
;
2237 unsigned long long write_range
;
2238 unsigned long long max_progress
, target
, completed
;
2239 unsigned long long array_size
= (info
->component_size
2240 * reshape
->before
.data_disks
);
2244 /* First, we unsuspend any region that is now known to be safe.
2245 * If suspend_point is on the 'wrong' side of reshape_progress, then
2246 * we don't have or need suspension at the moment. This is true for
2247 * native metadata when we don't need to back-up.
2250 if (info
->reshape_progress
<= *suspend_point
)
2251 sysfs_set_num(info
, NULL
, "suspend_lo",
2252 info
->reshape_progress
);
2254 /* Note: this won't work in 2.6.37 and before.
2255 * Something somewhere should make sure we don't need it!
2257 if (info
->reshape_progress
>= *suspend_point
)
2258 sysfs_set_num(info
, NULL
, "suspend_hi",
2259 info
->reshape_progress
);
2262 /* Now work out how far it is safe to progress.
2263 * If the read_offset for ->reshape_progress is less than
2264 * 'blocks' beyond the write_offset, we can only progress as far
2266 * Otherwise we can progress until the write_offset for the new location
2267 * reaches (within 'blocks' of) the read_offset at the current location.
2268 * However that region must be suspended unless we are using native
2270 * If we need to suspend more, we limit it to 128M per device, which is
2271 * rather arbitrary and should be some time-based calculation.
2273 read_offset
= info
->reshape_progress
/ reshape
->before
.data_disks
;
2274 write_offset
= info
->reshape_progress
/ reshape
->after
.data_disks
;
2275 write_range
= info
->new_chunk
/512;
2276 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
)
2277 need_backup
= array_size
;
2279 need_backup
= reshape
->backup_blocks
;
2281 if (read_offset
< write_offset
+ write_range
)
2282 max_progress
= backup_point
;
2286 reshape
->after
.data_disks
;
2288 if (read_offset
> write_offset
- write_range
)
2289 /* Can only progress as far as has been backed up,
2290 * which must be suspended */
2291 max_progress
= backup_point
;
2292 else if (info
->reshape_progress
<= need_backup
)
2293 max_progress
= backup_point
;
2295 if (info
->array
.major_version
>= 0)
2296 /* Can progress until backup is needed */
2297 max_progress
= need_backup
;
2299 /* Can progress until metadata update is required */
2302 reshape
->after
.data_disks
;
2303 /* but data must be suspended */
2304 if (max_progress
< *suspend_point
)
2305 max_progress
= *suspend_point
;
2310 /* We know it is safe to progress to 'max_progress' providing
2311 * it is suspended or we are using native metadata.
2312 * Consider extending suspend_point 128M per device if it
2313 * is less than 64M per device beyond reshape_progress.
2314 * But always do a multiple of 'blocks'
2315 * FIXME this is too big - it takes to long to complete
2318 target
= 64*1024*2 * min(reshape
->before
.data_disks
,
2319 reshape
->after
.data_disks
);
2320 target
/= reshape
->backup_blocks
;
2323 target
*= reshape
->backup_blocks
;
2325 /* For externally managed metadata we always need to suspend IO to
2326 * the area being reshaped so we regularly push suspend_point forward.
2327 * For native metadata we only need the suspend if we are going to do
2331 if ((need_backup
> info
->reshape_progress
2332 || info
->array
.major_version
< 0) &&
2333 *suspend_point
< info
->reshape_progress
+ target
) {
2334 if (need_backup
< *suspend_point
+ 2 * target
)
2335 *suspend_point
= need_backup
;
2336 else if (*suspend_point
+ 2 * target
< array_size
)
2337 *suspend_point
+= 2 * target
;
2339 *suspend_point
= array_size
;
2340 sysfs_set_num(info
, NULL
, "suspend_hi", *suspend_point
);
2341 if (max_progress
> *suspend_point
)
2342 max_progress
= *suspend_point
;
2345 if (info
->array
.major_version
>= 0) {
2346 /* Only need to suspend when about to backup */
2347 if (info
->reshape_progress
< need_backup
* 2 &&
2348 *suspend_point
> 0) {
2350 sysfs_set_num(info
, NULL
, "suspend_lo", 0);
2351 sysfs_set_num(info
, NULL
, "suspend_hi", need_backup
);
2354 /* Need to suspend continually */
2355 if (info
->reshape_progress
< *suspend_point
)
2356 *suspend_point
= info
->reshape_progress
;
2357 if (*suspend_point
+ target
< info
->reshape_progress
)
2358 /* No need to move suspend region yet */;
2360 if (*suspend_point
>= 2 * target
)
2361 *suspend_point
-= 2 * target
;
2364 sysfs_set_num(info
, NULL
, "suspend_lo",
2367 if (max_progress
< *suspend_point
)
2368 max_progress
= *suspend_point
;
2372 /* now set sync_max to allow that progress. sync_max, like
2373 * sync_completed is a count of sectors written per device, so
2374 * we find the difference between max_progress and the start point,
2375 * and divide that by after.data_disks to get a sync_max
2377 * At the same time we convert wait_point to a similar number
2378 * for comparing against sync_completed.
2380 /* scale down max_progress to per_disk */
2381 max_progress
/= reshape
->after
.data_disks
;
2382 /* Round to chunk size as some kernels give an erroneously high number */
2383 max_progress
/= info
->new_chunk
/512;
2384 max_progress
*= info
->new_chunk
/512;
2385 /* Limit progress to the whole device */
2386 if (max_progress
> info
->component_size
)
2387 max_progress
= info
->component_size
;
2388 wait_point
/= reshape
->after
.data_disks
;
2390 /* switch from 'device offset' to 'processed block count' */
2391 max_progress
= info
->component_size
- max_progress
;
2392 wait_point
= info
->component_size
- wait_point
;
2395 sysfs_set_num(info
, NULL
, "sync_max", max_progress
);
2397 /* Now wait. If we have already reached the point that we were
2398 * asked to wait to, don't wait at all, else wait for any change.
2399 * We need to select on 'sync_completed' as that is the place that
2400 * notifications happen, but we are really interested in
2401 * 'reshape_position'
2403 fd
= sysfs_get_fd(info
, NULL
, "sync_completed");
2405 goto check_progress
;
2407 if (sysfs_fd_get_ll(fd
, &completed
) < 0) {
2409 goto check_progress
;
2411 while (completed
< max_progress
&& completed
< wait_point
) {
2412 /* Check that sync_action is still 'reshape' to avoid
2413 * waiting forever on a dead array
2417 if (sysfs_get_str(info
, NULL
, "sync_action",
2419 strncmp(action
, "reshape", 7) != 0)
2423 select(fd
+1, NULL
, NULL
, &rfds
, NULL
);
2424 if (sysfs_fd_get_ll(fd
, &completed
) < 0) {
2426 goto check_progress
;
2429 /* some kernels can give an incorrectly high 'completed' number */
2430 completed
/= (info
->new_chunk
/512);
2431 completed
*= (info
->new_chunk
/512);
2432 /* Convert 'completed' back in to a 'progress' number */
2433 completed
*= reshape
->after
.data_disks
;
2435 completed
= info
->component_size
* reshape
->after
.data_disks
2438 *reshape_completed
= completed
;
2442 /* We return the need_backup flag. Caller will decide
2443 * how much - a multiple of ->backup_blocks up to *suspend_point
2446 return need_backup
> info
->reshape_progress
;
2448 return need_backup
>= info
->reshape_progress
;
2451 /* if we couldn't read a number from sync_completed, then
2452 * either the reshape did complete, or it aborted.
2453 * We can tell which by checking for 'none' in reshape_position.
2456 if (sysfs_get_str(info
, NULL
, "reshape_position", buf
, sizeof(buf
)) < 0
2457 || strncmp(buf
, "none", 4) != 0)
2458 return -2; /* abort */
2460 /* Maybe racing with array shutdown - check state */
2461 if (sysfs_get_str(info
, NULL
, "array_state", buf
, sizeof(buf
)) < 0
2462 || strncmp(buf
, "inactive", 8) == 0
2463 || strncmp(buf
, "clear",5) == 0)
2464 return -2; /* abort */
2465 return -1; /* complete */
2470 /* FIXME return status is never checked */
2471 static int grow_backup(struct mdinfo
*sra
,
2472 unsigned long long offset
, /* per device */
2473 unsigned long stripes
, /* per device, in old chunks */
2474 int *sources
, unsigned long long *offsets
,
2475 int disks
, int chunk
, int level
, int layout
,
2476 int dests
, int *destfd
, unsigned long long *destoffsets
,
2477 int part
, int *degraded
,
2480 /* Backup 'blocks' sectors at 'offset' on each device of the array,
2481 * to storage 'destfd' (offset 'destoffsets'), after first
2482 * suspending IO. Then allow resync to continue
2483 * over the suspended section.
2484 * Use part 'part' of the backup-super-block.
2489 unsigned long long ll
;
2491 //printf("offset %llu\n", offset);
2497 /* Check that array hasn't become degraded, else we might backup the wrong data */
2498 sysfs_get_ll(sra
, NULL
, "degraded", &ll
);
2499 new_degraded
= (int)ll
;
2500 if (new_degraded
!= *degraded
) {
2501 /* check each device to ensure it is still working */
2503 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
2504 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
2506 if (sd
->disk
.state
& (1<<MD_DISK_SYNC
)) {
2508 if (sysfs_get_str(sra
, sd
, "state", sbuf
, 20) < 0 ||
2509 strstr(sbuf
, "faulty") ||
2510 strstr(sbuf
, "in_sync") == NULL
) {
2511 /* this device is dead */
2512 sd
->disk
.state
= (1<<MD_DISK_FAULTY
);
2513 if (sd
->disk
.raid_disk
>= 0 &&
2514 sources
[sd
->disk
.raid_disk
] >= 0) {
2515 close(sources
[sd
->disk
.raid_disk
]);
2516 sources
[sd
->disk
.raid_disk
] = -1;
2521 *degraded
= new_degraded
;
2524 bsb
.arraystart2
= __cpu_to_le64(offset
* odata
);
2525 bsb
.length2
= __cpu_to_le64(stripes
* (chunk
/512) * odata
);
2527 bsb
.arraystart
= __cpu_to_le64(offset
* odata
);
2528 bsb
.length
= __cpu_to_le64(stripes
* (chunk
/512) * odata
);
2531 bsb
.magic
[15] = '2';
2532 for (i
= 0; i
< dests
; i
++)
2534 lseek64(destfd
[i
], destoffsets
[i
] + __le64_to_cpu(bsb
.devstart2
)*512, 0);
2536 lseek64(destfd
[i
], destoffsets
[i
], 0);
2538 rv
= save_stripes(sources
, offsets
,
2539 disks
, chunk
, level
, layout
,
2541 offset
*512*odata
, stripes
* chunk
* odata
,
2546 bsb
.mtime
= __cpu_to_le64(time(0));
2547 for (i
= 0; i
< dests
; i
++) {
2548 bsb
.devstart
= __cpu_to_le64(destoffsets
[i
]/512);
2550 bsb
.sb_csum
= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
));
2551 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0)
2552 bsb
.sb_csum2
= bsb_csum((char*)&bsb
,
2553 ((char*)&bsb
.sb_csum2
)-((char*)&bsb
));
2556 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
] - 4096, 0)
2557 != destoffsets
[i
] - 4096)
2559 if (write(destfd
[i
], &bsb
, 512) != 512)
2561 if (destoffsets
[i
] > 4096) {
2562 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
]+stripes
*chunk
*odata
, 0) !=
2563 destoffsets
[i
]+stripes
*chunk
*odata
)
2565 if (write(destfd
[i
], &bsb
, 512) != 512)
2575 /* in 2.6.30, the value reported by sync_completed can be
2576 * less that it should be by one stripe.
2577 * This only happens when reshape hits sync_max and pauses.
2578 * So allow wait_backup to either extent sync_max further
2579 * than strictly necessary, or return before the
2580 * sync has got quite as far as we would really like.
2581 * This is what 'blocks2' is for.
2582 * The various caller give appropriate values so that
2585 /* FIXME return value is often ignored */
2586 static int forget_backup(
2587 int dests
, int *destfd
, unsigned long long *destoffsets
,
2591 * Erase backup 'part' (which is 0 or 1)
2597 bsb
.arraystart2
= __cpu_to_le64(0);
2598 bsb
.length2
= __cpu_to_le64(0);
2600 bsb
.arraystart
= __cpu_to_le64(0);
2601 bsb
.length
= __cpu_to_le64(0);
2603 bsb
.mtime
= __cpu_to_le64(time(0));
2605 for (i
= 0; i
< dests
; i
++) {
2606 bsb
.devstart
= __cpu_to_le64(destoffsets
[i
]/512);
2607 bsb
.sb_csum
= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
));
2608 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0)
2609 bsb
.sb_csum2
= bsb_csum((char*)&bsb
,
2610 ((char*)&bsb
.sb_csum2
)-((char*)&bsb
));
2611 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
]-4096, 0) !=
2612 destoffsets
[i
]-4096)
2615 write(destfd
[i
], &bsb
, 512) != 512)
2622 static void fail(char *msg
)
2625 rv
= (write(2, msg
, strlen(msg
)) != (int)strlen(msg
));
2626 rv
|= (write(2, "\n", 1) != 1);
2630 static char *abuf
, *bbuf
;
2631 static unsigned long long abuflen
;
2632 static void validate(int afd
, int bfd
, unsigned long long offset
)
2634 /* check that the data in the backup against the array.
2635 * This is only used for regression testing and should not
2636 * be used while the array is active
2640 lseek64(bfd
, offset
- 4096, 0);
2641 if (read(bfd
, &bsb2
, 512) != 512)
2642 fail("cannot read bsb");
2643 if (bsb2
.sb_csum
!= bsb_csum((char*)&bsb2
,
2644 ((char*)&bsb2
.sb_csum
)-((char*)&bsb2
)))
2645 fail("first csum bad");
2646 if (memcmp(bsb2
.magic
, "md_backup_data", 14) != 0)
2647 fail("magic is bad");
2648 if (memcmp(bsb2
.magic
, "md_backup_data-2", 16) == 0 &&
2649 bsb2
.sb_csum2
!= bsb_csum((char*)&bsb2
,
2650 ((char*)&bsb2
.sb_csum2
)-((char*)&bsb2
)))
2651 fail("second csum bad");
2653 if (__le64_to_cpu(bsb2
.devstart
)*512 != offset
)
2654 fail("devstart is wrong");
2657 unsigned long long len
= __le64_to_cpu(bsb2
.length
)*512;
2659 if (abuflen
< len
) {
2663 if (posix_memalign((void**)&abuf
, 4096, abuflen
) ||
2664 posix_memalign((void**)&bbuf
, 4096, abuflen
)) {
2666 /* just stop validating on mem-alloc failure */
2671 lseek64(bfd
, offset
, 0);
2672 if ((unsigned long long)read(bfd
, bbuf
, len
) != len
) {
2673 //printf("len %llu\n", len);
2674 fail("read first backup failed");
2676 lseek64(afd
, __le64_to_cpu(bsb2
.arraystart
)*512, 0);
2677 if ((unsigned long long)read(afd
, abuf
, len
) != len
)
2678 fail("read first from array failed");
2679 if (memcmp(bbuf
, abuf
, len
) != 0) {
2682 printf("offset=%llu len=%llu\n",
2683 (unsigned long long)__le64_to_cpu(bsb2
.arraystart
)*512, len
);
2684 for (i
=0; i
<len
; i
++)
2685 if (bbuf
[i
] != abuf
[i
]) {
2686 printf("first diff byte %d\n", i
);
2690 fail("data1 compare failed");
2694 unsigned long long len
= __le64_to_cpu(bsb2
.length2
)*512;
2696 if (abuflen
< len
) {
2700 abuf
= malloc(abuflen
);
2701 bbuf
= malloc(abuflen
);
2704 lseek64(bfd
, offset
+__le64_to_cpu(bsb2
.devstart2
)*512, 0);
2705 if ((unsigned long long)read(bfd
, bbuf
, len
) != len
)
2706 fail("read second backup failed");
2707 lseek64(afd
, __le64_to_cpu(bsb2
.arraystart2
)*512, 0);
2708 if ((unsigned long long)read(afd
, abuf
, len
) != len
)
2709 fail("read second from array failed");
2710 if (memcmp(bbuf
, abuf
, len
) != 0)
2711 fail("data2 compare failed");
2715 int child_monitor(int afd
, struct mdinfo
*sra
, struct reshape
*reshape
,
2716 struct supertype
*st
, unsigned long blocks
,
2717 int *fds
, unsigned long long *offsets
,
2718 int dests
, int *destfd
, unsigned long long *destoffsets
)
2720 /* Monitor a reshape where backup is being performed using
2721 * 'native' mechanism - either to a backup file, or
2722 * to some space in a spare.
2726 unsigned long long speed
;
2727 unsigned long long suspend_point
, array_size
;
2728 unsigned long long backup_point
, wait_point
;
2729 unsigned long long reshape_completed
;
2731 int increasing
= reshape
->after
.data_disks
>= reshape
->before
.data_disks
;
2732 int part
= 0; /* The next part of the backup area to fill. It may already
2733 * be full, so we need to check */
2734 int level
= reshape
->level
;
2735 int layout
= reshape
->before
.layout
;
2736 int data
= reshape
->before
.data_disks
;
2737 int disks
= reshape
->before
.data_disks
+ reshape
->parity
;
2738 int chunk
= sra
->array
.chunk_size
;
2740 unsigned long stripes
;
2742 /* set up the backup-super-block. This requires the
2743 * uuid from the array.
2745 /* Find a superblock */
2746 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
2750 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
2752 dn
= map_dev(sd
->disk
.major
, sd
->disk
.minor
, 1);
2753 devfd
= dev_open(dn
, O_RDONLY
);
2756 ok
= st
->ss
->load_super(st
, devfd
, NULL
);
2762 fprintf(stderr
, Name
": Cannot find a superblock\n");
2766 memset(&bsb
, 0, 512);
2767 memcpy(bsb
.magic
, "md_backup_data-1", 16);
2768 st
->ss
->uuid_from_super(st
, (int*)&bsb
.set_uuid
);
2769 bsb
.mtime
= __cpu_to_le64(time(0));
2770 bsb
.devstart2
= blocks
;
2772 stripes
= blocks
/ (sra
->array
.chunk_size
/512) /
2773 reshape
->before
.data_disks
;
2775 if (posix_memalign((void**)&buf
, 4096, disks
* chunk
))
2776 /* Don't start the 'reshape' */
2778 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
) {
2779 sysfs_get_ll(sra
, NULL
, "sync_speed_min", &speed
);
2780 sysfs_set_num(sra
, NULL
, "sync_speed_min", 200000);
2784 array_size
= sra
->component_size
* reshape
->after
.data_disks
;
2785 backup_point
= sra
->reshape_progress
;
2788 array_size
= sra
->component_size
* reshape
->before
.data_disks
;
2789 backup_point
= array_size
;
2790 suspend_point
= array_size
;
2796 /* Want to return as soon the oldest backup slot can
2797 * be released as that allows us to start backing up
2798 * some more, providing suspend_point has been
2799 * advanced, which it should have.
2802 wait_point
= array_size
;
2803 if (part
== 0 && __le64_to_cpu(bsb
.length
) > 0)
2804 wait_point
= (__le64_to_cpu(bsb
.arraystart
) +
2805 __le64_to_cpu(bsb
.length
));
2806 if (part
== 1 && __le64_to_cpu(bsb
.length2
) > 0)
2807 wait_point
= (__le64_to_cpu(bsb
.arraystart2
) +
2808 __le64_to_cpu(bsb
.length2
));
2811 if (part
== 0 && __le64_to_cpu(bsb
.length
) > 0)
2812 wait_point
= __le64_to_cpu(bsb
.arraystart
);
2813 if (part
== 1 && __le64_to_cpu(bsb
.length2
) > 0)
2814 wait_point
= __le64_to_cpu(bsb
.arraystart2
);
2817 rv
= progress_reshape(sra
, reshape
,
2818 backup_point
, wait_point
,
2819 &suspend_point
, &reshape_completed
);
2820 /* external metadata would need to ping_monitor here */
2821 sra
->reshape_progress
= reshape_completed
;
2823 /* Clear any backup region that is before 'here' */
2825 if (reshape_completed
>= (__le64_to_cpu(bsb
.arraystart
) +
2826 __le64_to_cpu(bsb
.length
)))
2827 forget_backup(dests
, destfd
,
2829 if (reshape_completed
>= (__le64_to_cpu(bsb
.arraystart2
) +
2830 __le64_to_cpu(bsb
.length2
)))
2831 forget_backup(dests
, destfd
,
2834 if (reshape_completed
<= (__le64_to_cpu(bsb
.arraystart
)))
2835 forget_backup(dests
, destfd
,
2837 if (reshape_completed
<= (__le64_to_cpu(bsb
.arraystart2
)))
2838 forget_backup(dests
, destfd
,
2849 unsigned long long offset
;
2850 unsigned long actual_stripes
;
2851 /* Need to backup some data.
2852 * If 'part' is not used and the desired
2853 * backup size is suspended, do a backup,
2854 * then consider the next part.
2856 /* Check that 'part' is unused */
2857 if (part
== 0 && __le64_to_cpu(bsb
.length
) != 0)
2859 if (part
== 1 && __le64_to_cpu(bsb
.length2
) != 0)
2862 offset
= backup_point
/ data
;
2863 actual_stripes
= stripes
;
2865 if (offset
+ actual_stripes
* (chunk
/512) >
2866 sra
->component_size
)
2867 actual_stripes
= ((sra
->component_size
- offset
)
2869 if (offset
+ actual_stripes
* (chunk
/512) >
2873 if (offset
< actual_stripes
* (chunk
/512))
2874 actual_stripes
= offset
/ (chunk
/512);
2875 offset
-= actual_stripes
* (chunk
/512);
2876 if (offset
< suspend_point
/data
)
2879 grow_backup(sra
, offset
, actual_stripes
,
2881 disks
, chunk
, level
, layout
,
2882 dests
, destfd
, destoffsets
,
2883 part
, °raded
, buf
);
2884 validate(afd
, destfd
[0], destoffsets
[0]);
2885 /* record where 'part' is up to */
2888 backup_point
+= actual_stripes
* (chunk
/512) * data
;
2890 backup_point
-= actual_stripes
* (chunk
/512) * data
;
2894 /* FIXME maybe call progress_reshape one more time instead */
2895 abort_reshape(sra
); /* remove any remaining suspension */
2896 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
)
2897 sysfs_set_num(sra
, NULL
, "sync_speed_min", speed
);
2903 * If any spare contains md_back_data-1 which is recent wrt mtime,
2904 * write that data into the array and update the super blocks with
2905 * the new reshape_progress
2907 int Grow_restart(struct supertype
*st
, struct mdinfo
*info
, int *fdlist
, int cnt
,
2908 char *backup_file
, int verbose
)
2912 unsigned long long *offsets
;
2913 unsigned long long nstripe
, ostripe
;
2916 if (info
->new_level
!= info
->array
.level
)
2917 return 1; /* Cannot handle level changes (they are instantaneous) */
2919 odata
= info
->array
.raid_disks
- info
->delta_disks
- 1;
2920 if (info
->array
.level
== 6) odata
--; /* number of data disks */
2921 ndata
= info
->array
.raid_disks
- 1;
2922 if (info
->new_level
== 6) ndata
--;
2924 old_disks
= info
->array
.raid_disks
- info
->delta_disks
;
2926 if (info
->delta_disks
<= 0)
2927 /* Didn't grow, so the backup file must have
2931 for (i
=old_disks
-(backup_file
?1:0); i
<cnt
; i
++) {
2932 struct mdinfo dinfo
;
2935 char *devname
, namebuf
[20];
2937 /* This was a spare and may have some saved data on it.
2938 * Load the superblock, find and load the
2939 * backup_super_block.
2940 * If either fail, go on to next device.
2941 * If the backup contains no new info, just return
2942 * else restore data and update all superblocks
2944 if (i
== old_disks
-1) {
2945 fd
= open(backup_file
, O_RDONLY
);
2947 fprintf(stderr
, Name
": backup file %s inaccessible: %s\n",
2948 backup_file
, strerror(errno
));
2951 devname
= backup_file
;
2956 if (st
->ss
->load_super(st
, fd
, NULL
))
2959 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
2960 st
->ss
->free_super(st
);
2963 (dinfo
.data_offset
+ dinfo
.component_size
- 8) <<9,
2965 fprintf(stderr
, Name
": Cannot seek on device %d\n", i
);
2966 continue; /* Cannot seek */
2968 sprintf(namebuf
, "device-%d", i
);
2971 if (read(fd
, &bsb
, sizeof(bsb
)) != sizeof(bsb
)) {
2973 fprintf(stderr
, Name
": Cannot read from %s\n", devname
);
2974 continue; /* Cannot read */
2976 if (memcmp(bsb
.magic
, "md_backup_data-1", 16) != 0 &&
2977 memcmp(bsb
.magic
, "md_backup_data-2", 16) != 0) {
2979 fprintf(stderr
, Name
": No backup metadata on %s\n", devname
);
2982 if (bsb
.sb_csum
!= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
))) {
2984 fprintf(stderr
, Name
": Bad backup-metadata checksum on %s\n", devname
);
2985 continue; /* bad checksum */
2987 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0 &&
2988 bsb
.sb_csum2
!= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum2
)-((char*)&bsb
))) {
2990 fprintf(stderr
, Name
": Bad backup-metadata checksum2 on %s\n", devname
);
2991 continue; /* Bad second checksum */
2993 if (memcmp(bsb
.set_uuid
,info
->uuid
, 16) != 0) {
2995 fprintf(stderr
, Name
": Wrong uuid on backup-metadata on %s\n", devname
);
2996 continue; /* Wrong uuid */
2999 /* array utime and backup-mtime should be updated at much the same time, but it seems that
3000 * sometimes they aren't... So allow considerable flexability in matching, and allow
3001 * this test to be overridden by an environment variable.
3003 if (info
->array
.utime
> (int)__le64_to_cpu(bsb
.mtime
) + 2*60*60 ||
3004 info
->array
.utime
< (int)__le64_to_cpu(bsb
.mtime
) - 10*60) {
3005 if (check_env("MDADM_GROW_ALLOW_OLD")) {
3006 fprintf(stderr
, Name
": accepting backup with timestamp %lu "
3007 "for array with timestamp %lu\n",
3008 (unsigned long)__le64_to_cpu(bsb
.mtime
),
3009 (unsigned long)info
->array
.utime
);
3012 fprintf(stderr
, Name
": too-old timestamp on "
3013 "backup-metadata on %s\n", devname
);
3014 continue; /* time stamp is too bad */
3018 if (bsb
.magic
[15] == '1') {
3019 if (info
->delta_disks
>= 0) {
3020 /* reshape_progress is increasing */
3021 if (__le64_to_cpu(bsb
.arraystart
) + __le64_to_cpu(bsb
.length
) <
3022 info
->reshape_progress
) {
3025 fprintf(stderr
, Name
": backup-metadata found on %s but is not needed\n", devname
);
3026 continue; /* No new data here */
3029 /* reshape_progress is decreasing */
3030 if (__le64_to_cpu(bsb
.arraystart
) >=
3031 info
->reshape_progress
)
3032 goto nonew
; /* No new data here */
3035 if (info
->delta_disks
>= 0) {
3036 /* reshape_progress is increasing */
3037 if (__le64_to_cpu(bsb
.arraystart
) + __le64_to_cpu(bsb
.length
) <
3038 info
->reshape_progress
&&
3039 __le64_to_cpu(bsb
.arraystart2
) + __le64_to_cpu(bsb
.length2
) <
3040 info
->reshape_progress
)
3041 goto nonew
; /* No new data here */
3043 /* reshape_progress is decreasing */
3044 if (__le64_to_cpu(bsb
.arraystart
) >=
3045 info
->reshape_progress
&&
3046 __le64_to_cpu(bsb
.arraystart2
) >=
3047 info
->reshape_progress
)
3048 goto nonew
; /* No new data here */
3051 if (lseek64(fd
, __le64_to_cpu(bsb
.devstart
)*512, 0)< 0) {
3054 fprintf(stderr
, Name
": Failed to verify secondary backup-metadata block on %s\n",
3056 continue; /* Cannot seek */
3058 /* There should be a duplicate backup superblock 4k before here */
3059 if (lseek64(fd
, -4096, 1) < 0 ||
3060 read(fd
, &bsb2
, sizeof(bsb2
)) != sizeof(bsb2
))
3061 goto second_fail
; /* Cannot find leading superblock */
3062 if (bsb
.magic
[15] == '1')
3063 bsbsize
= offsetof(struct mdp_backup_super
, pad1
);
3065 bsbsize
= offsetof(struct mdp_backup_super
, pad
);
3066 if (memcmp(&bsb2
, &bsb
, bsbsize
) != 0)
3067 goto second_fail
; /* Cannot find leading superblock */
3069 /* Now need the data offsets for all devices. */
3070 offsets
= malloc(sizeof(*offsets
)*info
->array
.raid_disks
);
3071 for(j
=0; j
<info
->array
.raid_disks
; j
++) {
3074 if (st
->ss
->load_super(st
, fdlist
[j
], NULL
))
3075 /* FIXME should be this be an error */
3077 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
3078 st
->ss
->free_super(st
);
3079 offsets
[j
] = dinfo
.data_offset
* 512;
3081 printf(Name
": restoring critical section\n");
3083 if (restore_stripes(fdlist
, offsets
,
3084 info
->array
.raid_disks
,
3088 fd
, __le64_to_cpu(bsb
.devstart
)*512,
3089 __le64_to_cpu(bsb
.arraystart
)*512,
3090 __le64_to_cpu(bsb
.length
)*512)) {
3091 /* didn't succeed, so giveup */
3093 fprintf(stderr
, Name
": Error restoring backup from %s\n",
3098 if (bsb
.magic
[15] == '2' &&
3099 restore_stripes(fdlist
, offsets
,
3100 info
->array
.raid_disks
,
3104 fd
, __le64_to_cpu(bsb
.devstart
)*512 +
3105 __le64_to_cpu(bsb
.devstart2
)*512,
3106 __le64_to_cpu(bsb
.arraystart2
)*512,
3107 __le64_to_cpu(bsb
.length2
)*512)) {
3108 /* didn't succeed, so giveup */
3110 fprintf(stderr
, Name
": Error restoring second backup from %s\n",
3116 /* Ok, so the data is restored. Let's update those superblocks. */
3118 if (info
->delta_disks
>= 0) {
3119 info
->reshape_progress
= __le64_to_cpu(bsb
.arraystart
) +
3120 __le64_to_cpu(bsb
.length
);
3121 if (bsb
.magic
[15] == '2') {
3122 unsigned long long p2
= __le64_to_cpu(bsb
.arraystart2
) +
3123 __le64_to_cpu(bsb
.length2
);
3124 if (p2
> info
->reshape_progress
)
3125 info
->reshape_progress
= p2
;
3128 info
->reshape_progress
= __le64_to_cpu(bsb
.arraystart
);
3129 if (bsb
.magic
[15] == '2') {
3130 unsigned long long p2
= __le64_to_cpu(bsb
.arraystart2
);
3131 if (p2
< info
->reshape_progress
)
3132 info
->reshape_progress
= p2
;
3135 for (j
=0; j
<info
->array
.raid_disks
; j
++) {
3136 if (fdlist
[j
] < 0) continue;
3137 if (st
->ss
->load_super(st
, fdlist
[j
], NULL
))
3139 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
3140 dinfo
.reshape_progress
= info
->reshape_progress
;
3141 st
->ss
->update_super(st
, &dinfo
,
3142 "_reshape_progress",
3144 st
->ss
->store_super(st
, fdlist
[j
]);
3145 st
->ss
->free_super(st
);
3149 /* Didn't find any backup data, try to see if any
3152 if (info
->delta_disks
< 0) {
3153 /* When shrinking, the critical section is at the end.
3154 * So see if we are before the critical section.
3156 unsigned long long first_block
;
3157 nstripe
= ostripe
= 0;
3159 while (ostripe
>= nstripe
) {
3160 ostripe
+= info
->array
.chunk_size
/ 512;
3161 first_block
= ostripe
* odata
;
3162 nstripe
= first_block
/ ndata
/ (info
->new_chunk
/512) *
3163 (info
->new_chunk
/512);
3166 if (info
->reshape_progress
>= first_block
)
3169 if (info
->delta_disks
> 0) {
3170 /* See if we are beyond the critical section. */
3171 unsigned long long last_block
;
3172 nstripe
= ostripe
= 0;
3174 while (nstripe
>= ostripe
) {
3175 nstripe
+= info
->new_chunk
/ 512;
3176 last_block
= nstripe
* ndata
;
3177 ostripe
= last_block
/ odata
/ (info
->array
.chunk_size
/512) *
3178 (info
->array
.chunk_size
/512);
3181 if (info
->reshape_progress
>= last_block
)
3184 /* needed to recover critical section! */
3186 fprintf(stderr
, Name
": Failed to find backup of critical section\n");
3190 int Grow_continue(int mdfd
, struct supertype
*st
, struct mdinfo
*info
,
3193 int err
= sysfs_set_str(info
, NULL
, "array_state", "readonly");
3196 return reshape_array(NULL
, mdfd
, "array", st
, info
, 1, backup_file
, 0, 0, 1);