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 err
= err
?: sysfs_set_num(sra
, NULL
, "sync_min", 0);
637 err
= err
?: sysfs_set_num(sra
, NULL
, "sync_max", 0);
638 err
= err
?: sysfs_set_str(sra
, NULL
, "sync_action", "reshape");
643 void abort_reshape(struct mdinfo
*sra
)
645 sysfs_set_str(sra
, NULL
, "sync_action", "idle");
646 sysfs_set_num(sra
, NULL
, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL
);
647 sysfs_set_num(sra
, NULL
, "suspend_hi", 0);
648 sysfs_set_num(sra
, NULL
, "suspend_lo", 0);
649 sysfs_set_num(sra
, NULL
, "sync_min", 0);
650 sysfs_set_str(sra
, NULL
, "sync_max", "max");
653 int remove_disks_on_raid10_to_raid0_takeover(struct supertype
*st
,
658 struct mdinfo
*remaining
;
661 nr_of_copies
= layout
& 0xff;
663 remaining
= sra
->devs
;
665 /* for each 'copy', select one device and remove from the list. */
666 for (slot
= 0; slot
< sra
->array
.raid_disks
; slot
+= nr_of_copies
) {
667 struct mdinfo
**diskp
;
670 /* Find a working device to keep */
671 for (diskp
= &remaining
; *diskp
; diskp
= &(*diskp
)->next
) {
672 struct mdinfo
*disk
= *diskp
;
674 if (disk
->disk
.raid_disk
< slot
)
676 if (disk
->disk
.raid_disk
>= slot
+ nr_of_copies
)
678 if (disk
->disk
.state
& (1<<MD_DISK_REMOVED
))
680 if (disk
->disk
.state
& (1<<MD_DISK_FAULTY
))
682 if (!(disk
->disk
.state
& (1<<MD_DISK_SYNC
)))
685 /* We have found a good disk to use! */
687 disk
->next
= sra
->devs
;
696 if (slot
< sra
->array
.raid_disks
) {
697 /* didn't find all slots */
703 sra
->devs
= remaining
;
707 /* Remove all 'remaining' devices from the array */
709 struct mdinfo
*sd
= remaining
;
710 remaining
= sd
->next
;
712 sysfs_set_str(sra
, sd
, "state", "faulty");
713 sysfs_set_str(sra
, sd
, "slot", "none");
714 sysfs_set_str(sra
, sd
, "state", "remove");
715 sd
->disk
.state
|= (1<<MD_DISK_REMOVED
);
716 sd
->disk
.state
&= ~(1<<MD_DISK_SYNC
);
717 sd
->next
= sra
->devs
;
723 void reshape_free_fdlist(int *fdlist
,
724 unsigned long long *offsets
,
729 for (i
= 0; i
< size
; i
++)
737 int reshape_prepare_fdlist(char *devname
,
741 unsigned long blocks
,
744 unsigned long long *offsets
)
749 for (d
= 0; d
<= nrdisks
; d
++)
752 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
753 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
755 if (sd
->disk
.state
& (1<<MD_DISK_SYNC
)) {
756 char *dn
= map_dev(sd
->disk
.major
,
758 fdlist
[sd
->disk
.raid_disk
]
759 = dev_open(dn
, O_RDONLY
);
760 offsets
[sd
->disk
.raid_disk
] = sd
->data_offset
*512;
761 if (fdlist
[sd
->disk
.raid_disk
] < 0) {
763 Name
": %s: cannot open component %s\n",
764 devname
, dn
? dn
: "-unknown-");
768 } else if (backup_file
== NULL
) {
770 char *dn
= map_dev(sd
->disk
.major
,
772 fdlist
[d
] = dev_open(dn
, O_RDWR
);
773 offsets
[d
] = (sd
->data_offset
+ sra
->component_size
- blocks
- 8)*512;
775 fprintf(stderr
, Name
": %s: cannot open component %s\n",
776 devname
, dn
? dn
: "-unknown-");
787 int reshape_open_backup_file(char *backup_file
,
792 unsigned long long *offsets
)
794 /* Return 1 on success, 0 on any form of failure */
795 /* need to check backup file is large enough */
801 *fdlist
= open(backup_file
, O_RDWR
|O_CREAT
|O_EXCL
,
805 fprintf(stderr
, Name
": %s: cannot create backup file %s: %s\n",
806 devname
, backup_file
, strerror(errno
));
809 /* Guard against backup file being on array device.
810 * If array is partitioned or if LVM etc is in the
811 * way this will not notice, but it is better than
814 fstat(*fdlist
, &stb
);
817 if (stb
.st_rdev
== dev
) {
818 fprintf(stderr
, Name
": backup file must NOT be"
819 " on the array being reshaped.\n");
825 for (i
=0; i
< blocks
+ 1 ; i
++) {
826 if (write(*fdlist
, buf
, 512) != 512) {
827 fprintf(stderr
, Name
": %s: cannot create"
828 " backup file %s: %s\n",
829 devname
, backup_file
, strerror(errno
));
833 if (fsync(*fdlist
) != 0) {
834 fprintf(stderr
, Name
": %s: cannot create backup file %s: %s\n",
835 devname
, backup_file
, strerror(errno
));
842 unsigned long compute_backup_blocks(int nchunk
, int ochunk
,
843 unsigned int ndata
, unsigned int odata
)
845 unsigned long a
, b
, blocks
;
846 /* So how much do we need to backup.
847 * We need an amount of data which is both a whole number of
848 * old stripes and a whole number of new stripes.
849 * So LCM for (chunksize*datadisks).
851 a
= (ochunk
/512) * odata
;
852 b
= (nchunk
/512) * ndata
;
860 /* LCM == product / GCD */
861 blocks
= (ochunk
/512) * (nchunk
/512) * odata
* ndata
/ a
;
866 char *analyse_change(struct mdinfo
*info
, struct reshape
*re
)
868 /* Based on the current array state in info->array and
869 * the changes in info->new_* etc, determine:
870 * - whether the change is possible
871 * - Intermediate level/raid_disks/layout
872 * - whether a restriping reshape is needed
873 * - number of sectors in minimum change unit. This
874 * will cover a whole number of stripes in 'before' and
877 * Return message if the change should be rejected
878 * NULL if the change can be achieved
880 * This can be called as part of starting a reshape, or
881 * when assembling an array that is undergoing reshape.
885 /* If a new level not explicitly given, we assume no-change */
886 if (info
->new_level
== UnSet
)
887 info
->new_level
= info
->array
.level
;
890 switch (info
->new_level
) {
896 /* chunk size is meaningful, must divide component_size
899 if (info
->component_size
% (info
->new_chunk
/512))
900 return "New chunk size does not"
901 " divide component size";
904 return "chunk size not meaningful for this level";
907 info
->new_chunk
= info
->array
.chunk_size
;
909 switch (info
->array
.level
) {
911 /* RAID1 can convert to RAID1 with different disks, or
914 if (info
->new_level
== 1) {
915 if (info
->delta_disks
== UnSet
)
916 /* Don't know what to do */
917 return "no change requested for Growing RAID1";
919 re
->before
.data_disks
= (info
->array
.raid_disks
+
921 re
->before
.layout
= 0;
922 re
->backup_blocks
= 0;
926 if (info
->array
.raid_disks
== 2 &&
927 info
->array
.raid_disks
== 5) {
928 /* simple in-place conversion */
931 re
->before
.data_disks
= 1;
932 re
->before
.layout
= ALGORITHM_LEFT_SYMMETRIC
;
933 re
->backup_blocks
= 0;
936 /* Could do some multi-stage conversions, but leave that to
939 return "Impossibly level change request for RAID1";
942 /* RAID10 can only be converted from near mode to
943 * RAID0 by removing some devices
945 if ((info
->array
.layout
& ~0xff) != 0x100)
946 return "Cannot Grow RAID10 with far/offset layout";
947 /* number of devices must be multiple of number of copies */
948 if (info
->array
.raid_disks
% (info
->array
.layout
& 0xff))
949 return "RAID10 layout too complex for Grow operation";
951 if (info
->new_level
!= 0)
952 return "RAID10 can only be changed to RAID0";
953 new_disks
= (info
->array
.raid_disks
954 / (info
->array
.layout
& 0xff));
955 if (info
->delta_disks
!= UnSet
) {
956 info
->delta_disks
= (new_disks
957 - info
->array
.raid_disks
);
959 if (info
->delta_disks
!= new_disks
- info
->array
.raid_disks
)
960 return "New number of raid-devices impossible for RAID10";
961 if (info
->new_chunk
&&
962 info
->new_chunk
!= info
->array
.chunk_size
)
963 return "Cannot change chunk-size with RAID10 Grow";
968 re
->before
.data_disks
= new_disks
;
969 re
->before
.layout
= 0;
970 re
->backup_blocks
= 0;
974 /* RAID0 can be converted to RAID10, or to RAID456 */
975 if (info
->new_level
== 10) {
976 if (info
->new_layout
== UnSet
&& info
->delta_disks
== UnSet
) {
977 /* Assume near=2 layout */
978 info
->new_layout
= 0x102;
979 info
->delta_disks
= info
->array
.raid_disks
;
981 if (info
->new_layout
== UnSet
) {
982 int copies
= 1 + (info
->delta_disks
983 / info
->array
.raid_disks
);
984 if (info
->array
.raid_disks
* (copies
-1)
985 != info
->delta_disks
)
986 return "Impossible number of devices"
987 " for RAID0->RAID10";
988 info
->new_layout
= 0x100 + copies
;
990 if (info
->delta_disks
== UnSet
) {
991 int copies
= info
->new_layout
& 0xff;
992 if (info
->new_layout
!= 0x100 + copies
)
993 return "New layout impossible"
994 " for RAID0->RAID10";;
995 info
->delta_disks
= (copies
- 1) *
996 info
->array
.raid_disks
;
998 if (info
->new_chunk
&&
999 info
->new_chunk
!= info
->array
.chunk_size
)
1000 return "Cannot change chunk-size with RAID0->RAID10";
1004 re
->before
.data_disks
= (info
->array
.raid_disks
+
1006 re
->before
.layout
= info
->new_layout
;
1007 re
->backup_blocks
= 0;
1011 /* RAID0 can also covert to RAID0/4/5/6 by first converting to
1012 * a raid4 style layout of the final level.
1014 switch (info
->new_level
) {
1018 re
->before
.layout
= 0;
1022 re
->before
.layout
= ALGORITHM_PARITY_N
;
1026 re
->before
.layout
= ALGORITHM_PARITY_N
;
1029 return "Impossible level change requested";
1031 re
->before
.data_disks
= info
->array
.raid_disks
;
1032 /* determining 'after' layout happens outside this 'switch' */
1036 info
->array
.layout
= ALGORITHM_PARITY_N
;
1038 switch (info
->new_level
) {
1040 re
->level
= info
->array
.level
;
1041 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1042 re
->before
.layout
= info
->array
.layout
;
1046 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1047 re
->before
.layout
= info
->array
.layout
;
1051 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1052 switch (info
->array
.layout
) {
1053 case ALGORITHM_LEFT_ASYMMETRIC
:
1054 re
->before
.layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
1056 case ALGORITHM_RIGHT_ASYMMETRIC
:
1057 re
->before
.layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
1059 case ALGORITHM_LEFT_SYMMETRIC
:
1060 re
->before
.layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
1062 case ALGORITHM_RIGHT_SYMMETRIC
:
1063 re
->before
.layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
1065 case ALGORITHM_PARITY_0
:
1066 re
->before
.layout
= ALGORITHM_PARITY_0_6
;
1068 case ALGORITHM_PARITY_N
:
1069 re
->before
.layout
= ALGORITHM_PARITY_N_6
;
1072 return "Cannot convert an array with this layout";
1076 if (info
->array
.raid_disks
!= 2)
1077 return "Can only convert a 2-device array to RAID1";
1079 re
->before
.data_disks
= 2;
1080 re
->before
.layout
= 0;
1083 return "Impossible level change requested";
1087 switch (info
->new_level
) {
1092 re
->before
.data_disks
= info
->array
.raid_disks
- 2;
1093 re
->before
.layout
= info
->array
.layout
;
1096 return "Impossible level change requested";
1101 /* If we reached here then it looks like a re-stripe is
1102 * happening. We have determined the intermediate level
1103 * and initial raid_disks/layout and stored these in 're'.
1105 * We need to deduce the final layout that can be atomically
1106 * converted to the end state.
1108 switch (info
->new_level
) {
1110 /* We can only get to RAID0 from RAID4 or RAID5
1111 * with appropriate layout and one extra device
1113 if (re
->level
!= 4 && re
->level
!= 5)
1114 return "Cannot covert to RAID0 from this level";
1115 if (info
->delta_disks
== UnSet
)
1116 re
->after
.data_disks
= re
->before
.data_disks
;
1118 re
->after
.data_disks
=
1119 info
->array
.raid_disks
+ info
->delta_disks
;
1120 switch (re
->level
) {
1122 re
->after
.layout
= 0 ; break;
1124 re
->after
.layout
= ALGORITHM_PARITY_N
; break;
1129 /* We can only get to RAID4 from RAID5 */
1130 if (re
->level
!= 4 && re
->level
!= 5)
1131 return "Cannot convert to RAID4 from this level";
1132 if (info
->delta_disks
== UnSet
)
1133 re
->after
.data_disks
= re
->before
.data_disks
;
1135 re
->after
.data_disks
=
1136 re
->before
.data_disks
+ info
->delta_disks
;
1137 switch (re
->level
) {
1139 re
->after
.layout
= 0 ; break;
1141 re
->after
.layout
= ALGORITHM_PARITY_N
; break;
1146 /* We get to RAID5 for RAID5 or RAID6 */
1147 if (re
->level
!= 5 && re
->level
!= 6)
1148 return "Cannot convert to RAID5 from this level";
1149 if (info
->delta_disks
== UnSet
)
1150 re
->after
.data_disks
= re
->before
.data_disks
;
1151 else if (re
->level
== 5)
1152 re
->after
.data_disks
=
1153 re
->before
.data_disks
+ info
->delta_disks
;
1155 re
->after
.data_disks
=
1156 info
->array
.raid_disks
+ info
->delta_disks
- 1;
1157 switch (re
->level
) {
1159 if (info
->new_layout
== UnSet
)
1160 re
->after
.layout
= re
->before
.layout
;
1162 re
->after
.layout
= info
->new_layout
;
1165 if (info
->new_layout
== UnSet
)
1166 info
->new_layout
= re
->before
.layout
;
1168 /* after.layout needs to be raid6 version of new_layout */
1169 if (info
->new_layout
== ALGORITHM_PARITY_N
)
1170 re
->after
.layout
= ALGORITHM_PARITY_N
;
1173 char *ls
= map_num(r5layout
, info
->new_layout
);
1175 strcat(strcpy(layout
, ls
), "-6");
1176 l
= map_name(r6layout
, layout
);
1178 return "Cannot find RAID6 layout"
1180 re
->after
.layout
= l
;
1186 /* We must already be at level 6 */
1188 return "Impossible level change";
1189 if (info
->delta_disks
== UnSet
)
1190 re
->after
.data_disks
= re
->before
.data_disks
;
1192 re
->after
.data_disks
= (info
->array
.raid_disks
+
1193 info
->delta_disks
) - 2;
1194 if (info
->new_layout
== UnSet
)
1195 re
->after
.layout
= info
->array
.layout
;
1197 re
->after
.layout
= info
->new_layout
;
1200 return "Impossible level change requested";
1202 switch (re
->level
) {
1203 case 6: re
->parity
= 2; break;
1205 case 5: re
->parity
= 1; break;
1206 default: re
->parity
= 0; break;
1208 /* So we have a restripe operation, we need to calculate the number
1209 * of blocks per reshape operation.
1211 if (info
->new_chunk
== 0)
1212 info
->new_chunk
= info
->array
.chunk_size
;
1213 if (re
->after
.data_disks
== re
->before
.data_disks
&&
1214 re
->after
.layout
== re
->before
.layout
&&
1215 info
->new_chunk
== info
->array
.chunk_size
) {
1216 /* Nothing to change */
1217 re
->backup_blocks
= 0;
1220 if (re
->after
.data_disks
== 1 && re
->before
.data_disks
== 1) {
1221 /* chunks can layout changes make no difference */
1222 re
->backup_blocks
= 0;
1226 if (re
->after
.data_disks
== re
->before
.data_disks
&&
1227 get_linux_version() < 2006032)
1228 return "in-place reshape is not safe before 2.6.32 - sorry.";
1230 if (re
->after
.data_disks
< re
->before
.data_disks
&&
1231 get_linux_version() < 2006030)
1232 return "reshape to fewer devices is not supported before 2.6.32 - sorry.";
1234 re
->backup_blocks
= compute_backup_blocks(
1235 info
->new_chunk
, info
->array
.chunk_size
,
1236 re
->after
.data_disks
,
1237 re
->before
.data_disks
);
1239 re
->new_size
= info
->component_size
* re
->after
.data_disks
;
1243 static int reshape_array(char *container
, int fd
, char *devname
,
1244 struct supertype
*st
, struct mdinfo
*info
,
1245 int force
, char *backup_file
, int quiet
, int forked
);
1246 static int reshape_container(char *container
, int cfd
, char *devname
,
1247 struct supertype
*st
,
1248 struct mdinfo
*info
,
1253 int Grow_reshape(char *devname
, int fd
, int quiet
, char *backup_file
,
1255 int level
, char *layout_str
, int chunksize
, int raid_disks
,
1258 /* Make some changes in the shape of an array.
1259 * The kernel must support the change.
1261 * There are three different changes. Each can trigger
1262 * a resync or recovery so we freeze that until we have
1263 * requested everything (if kernel supports freezing - 2.6.30).
1265 * - change size (i.e. component_size)
1267 * - change layout/chunksize/ndisks
1269 * The last can require a reshape. It is different on different
1270 * levels so we need to check the level before actioning it.
1271 * Some times the level change needs to be requested after the
1272 * reshape (e.g. raid6->raid5, raid5->raid0)
1275 struct mdu_array_info_s array
;
1277 struct supertype
*st
;
1278 char *subarray
= NULL
;
1282 char *container
= NULL
;
1283 char container_buf
[20];
1289 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) < 0) {
1290 fprintf(stderr
, Name
": %s is not an active md array - aborting\n",
1296 (chunksize
|| level
!= UnSet
|| layout_str
|| raid_disks
)) {
1297 fprintf(stderr
, Name
": cannot change component size at the same time "
1298 "as other changes.\n"
1299 " Change size first, then check data is intact before "
1300 "making other changes.\n");
1304 if (raid_disks
&& raid_disks
< array
.raid_disks
&& array
.level
> 1 &&
1305 get_linux_version() < 2006032 &&
1306 !check_env("MDADM_FORCE_FEWER")) {
1307 fprintf(stderr
, Name
": reducing the number of devices is not safe before Linux 2.6.32\n"
1308 " Please use a newer kernel\n");
1312 st
= super_by_fd(fd
, &subarray
);
1314 fprintf(stderr
, Name
": Unable to determine metadata format for %s\n", devname
);
1317 if (raid_disks
> st
->max_devs
) {
1318 fprintf(stderr
, Name
": Cannot increase raid-disks on this array"
1319 " beyond %d\n", st
->max_devs
);
1323 /* in the external case we need to check that the requested reshape is
1324 * supported, and perform an initial check that the container holds the
1325 * pre-requisite spare devices (mdmon owns final validation)
1327 if (st
->ss
->external
) {
1332 container_dev
= st
->container_dev
;
1333 cfd
= open_dev_excl(st
->container_dev
);
1335 container_dev
= st
->devnum
;
1337 cfd
= open_dev_excl(st
->devnum
);
1341 fprintf(stderr
, Name
": Unable to open container for %s\n",
1347 fmt_devname(container_buf
, container_dev
);
1348 container
= container_buf
;
1350 rv
= st
->ss
->load_container(st
, cfd
, NULL
);
1353 fprintf(stderr
, Name
": Cannot read superblock for %s\n",
1359 if (mdmon_running(container_dev
))
1360 st
->update_tail
= &st
->updates
;
1363 if (raid_disks
> array
.raid_disks
&&
1364 array
.spare_disks
< (raid_disks
- array
.raid_disks
) &&
1367 Name
": Need %d spare%s to avoid degraded array,"
1368 " and only have %d.\n"
1369 " Use --force to over-ride this check.\n",
1370 raid_disks
- array
.raid_disks
,
1371 raid_disks
- array
.raid_disks
== 1 ? "" : "s",
1376 sra
= sysfs_read(fd
, 0, GET_LEVEL
| GET_DISKS
| GET_DEVS
1377 | GET_STATE
| GET_VERSION
);
1379 if (st
->ss
->external
&& subarray
== NULL
) {
1380 array
.level
= LEVEL_CONTAINER
;
1381 sra
->array
.level
= LEVEL_CONTAINER
;
1384 fprintf(stderr
, Name
": failed to read sysfs parameters for %s\n",
1388 frozen
= freeze(st
);
1390 /* freeze() already spewed the reason */
1392 } else if (frozen
< 0) {
1393 fprintf(stderr
, Name
": %s is performing resync/recovery and cannot"
1394 " be reshaped\n", devname
);
1398 /* ========= set size =============== */
1399 if (size
>= 0 && (size
== 0 || size
!= array
.size
)) {
1400 long long orig_size
= array
.size
;
1402 if (reshape_super(st
, size
, UnSet
, UnSet
, 0, 0, NULL
, devname
, !quiet
)) {
1408 if (array
.size
!= size
) {
1409 /* got truncated to 32bit, write to
1410 * component_size instead
1413 rv
= sysfs_set_num(sra
, NULL
,
1414 "component_size", size
);
1418 rv
= ioctl(fd
, SET_ARRAY_INFO
, &array
);
1422 /* restore metadata */
1423 if (reshape_super(st
, orig_size
, UnSet
, UnSet
, 0, 0,
1424 NULL
, devname
, !quiet
) == 0)
1426 fprintf(stderr
, Name
": Cannot set device size for %s: %s\n",
1427 devname
, strerror(err
));
1429 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
1430 fprintf(stderr
, " Bitmap must be removed before size can be changed\n");
1434 ioctl(fd
, GET_ARRAY_INFO
, &array
);
1435 size
= get_component_size(fd
)/2;
1439 fprintf(stderr
, Name
": component size of %s has been set to %lluK\n",
1442 } else if (array
.level
!= LEVEL_CONTAINER
) {
1443 size
= get_component_size(fd
)/2;
1448 /* ========= check for Raid10 -> Raid0 conversion ===============
1449 * current implementation assumes that following conditions must be met:
1451 * - near_copies == 2
1453 if (level
== 0 && array
.level
== 10 && sra
&&
1454 array
.layout
== ((1 << 8) + 2) && !(array
.raid_disks
& 1)) {
1456 err
= remove_disks_on_raid10_to_raid0_takeover(st
, sra
, array
.layout
);
1458 dprintf(Name
": Array cannot be reshaped\n");
1467 sysfs_init(&info
, fd
, NoMdDev
);
1468 strcpy(info
.text_version
, sra
->text_version
);
1469 info
.component_size
= size
*2;
1470 info
.new_level
= level
;
1471 info
.new_chunk
= chunksize
* 1024;
1473 info
.delta_disks
= raid_disks
- info
.array
.raid_disks
;
1475 info
.delta_disks
= UnSet
;
1476 if (layout_str
== NULL
) {
1477 info
.new_layout
= UnSet
;
1478 if (info
.array
.level
== 6 &&
1479 (info
.new_level
== 6 || info
.new_level
== UnSet
) &&
1480 info
.array
.layout
>= 16) {
1481 fprintf(stderr
, Name
1482 ": %s has a non-standard layout. If you"
1483 " wish to preserve this\n"
1484 " during the reshape, please specify"
1485 " --layout=preserve\n"
1486 " If you want to change it, specify a"
1487 " layout or use --layout=normalise\n",
1492 } else if (strcmp(layout_str
, "normalise") == 0 ||
1493 strcmp(layout_str
, "normalize") == 0) {
1494 /* If we have a -6 RAID6 layout, remove the '-6'. */
1495 info
.new_layout
= UnSet
;
1496 if (info
.array
.level
== 6 && info
.new_level
== UnSet
) {
1498 strcpy(l
, map_num(r6layout
, info
.array
.layout
));
1499 h
= strrchr(l
, '-');
1500 if (h
&& strcmp(h
, "-6") == 0) {
1502 info
.new_layout
= map_name(r6layout
, l
);
1505 } else if (strcmp(layout_str
, "preserve") == 0) {
1506 info
.new_layout
= UnSet
;
1508 int l
= info
.new_level
;
1510 l
= info
.array
.level
;
1513 info
.new_layout
= map_name(r5layout
, layout_str
);
1516 info
.new_layout
= map_name(r6layout
, layout_str
);
1519 info
.new_layout
= parse_layout_10(layout_str
);
1522 info
.new_layout
= parse_layout_faulty(layout_str
);
1525 fprintf(stderr
, Name
": layout not meaningful"
1526 " with this level\n");
1530 if (info
.new_layout
== UnSet
) {
1531 fprintf(stderr
, Name
": layout %s not understood"
1532 " for this level\n",
1539 if (array
.level
== LEVEL_CONTAINER
) {
1540 /* This change is to be applied to every array in the
1541 * container. This is only needed when the metadata imposes
1542 * restraints of the various arrays in the container.
1543 * Currently we only know that IMSM requires all arrays
1544 * to have the same number of devices so changing the
1545 * number of devices (On-Line Capacity Expansion) must be
1546 * performed at the level of the container
1548 rv
= reshape_container(container
, fd
, devname
, st
, &info
,
1549 force
, backup_file
, quiet
);
1552 /* Impose these changes on a single array. First
1553 * check that the metadata is OK with the change. */
1555 if (reshape_super(st
, info
.component_size
, info
.new_level
,
1556 info
.new_layout
, info
.new_chunk
,
1557 info
.array
.raid_disks
+ info
.delta_disks
,
1558 backup_file
, devname
, quiet
)) {
1563 rv
= reshape_array(container
, fd
, devname
, st
, &info
, force
,
1564 backup_file
, quiet
, 0);
1573 static int reshape_array(char *container
, int fd
, char *devname
,
1574 struct supertype
*st
, struct mdinfo
*info
,
1576 char *backup_file
, int quiet
, int forked
)
1578 struct reshape reshape
;
1581 int orig_level
= UnSet
;
1584 struct mdu_array_info_s array
;
1589 unsigned long long *offsets
;
1593 unsigned long blocks
;
1594 unsigned long cache
;
1595 unsigned long long array_size
;
1599 msg
= analyse_change(info
, &reshape
);
1601 fprintf(stderr
, Name
": %s\n", msg
);
1604 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
1605 dprintf("Canot get array information.\n");
1608 spares_needed
= max(reshape
.before
.data_disks
,
1609 reshape
.after
.data_disks
)
1610 + reshape
.parity
- array
.raid_disks
;
1613 info
->new_level
> 0 &&
1614 spares_needed
> info
->array
.spare_disks
) {
1616 Name
": Need %d spare%s to avoid degraded array,"
1617 " and only have %d.\n"
1618 " Use --force to over-ride this check.\n",
1620 spares_needed
== 1 ? "" : "s",
1621 info
->array
.spare_disks
);
1625 if (reshape
.level
!= info
->array
.level
) {
1626 char *c
= map_num(pers
, reshape
.level
);
1629 return 1; /* This should not be possible */
1631 err
= sysfs_set_str(info
, NULL
, "level", c
);
1634 fprintf(stderr
, Name
": %s: could not set level to %s\n",
1637 (info
->array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
1638 fprintf(stderr
, " Bitmap must be removed"
1639 " before level can be changed\n");
1643 fprintf(stderr
, Name
": level of %s changed to %s\n",
1645 orig_level
= info
->array
.level
;
1648 if (reshape
.level
> 0 && st
->ss
->external
&&
1649 !mdmon_running(st
->container_dev
)) {
1650 start_mdmon(st
->container_dev
);
1651 ping_monitor(container
);
1654 /* ->reshape_super might have chosen some spares from the
1655 * container that it wants to be part of the new array.
1656 * We can collect them with ->container_content and give
1657 * them to the kernel.
1659 if (st
->ss
->reshape_super
&& st
->ss
->container_content
) {
1660 char *subarray
= strchr(info
->text_version
+1, '/')+1;
1661 struct mdinfo
*info2
=
1662 st
->ss
->container_content(st
, subarray
);
1666 sysfs_init(info2
, fd
, st
->devnum
);
1667 for (d
= info2
->devs
; d
; d
= d
->next
) {
1668 if (d
->disk
.state
== 0 &&
1669 d
->disk
.raid_disk
>= 0) {
1670 /* This is a spare that wants to
1671 * be part of the array.
1673 add_disk(fd
, st
, info2
, d
);
1680 if (reshape
.backup_blocks
== 0) {
1681 /* No restriping needed, but we might need to impose
1682 * some more changes: layout, raid_disks, chunk_size
1684 if (info
->new_layout
!= UnSet
&&
1685 info
->new_layout
!= info
->array
.layout
) {
1686 info
->array
.layout
= info
->new_layout
;
1687 if (ioctl(fd
, SET_ARRAY_INFO
, &info
->array
) != 0) {
1688 fprintf(stderr
, Name
": failed to set new layout\n");
1691 printf("layout for %s set to %d\n",
1692 devname
, info
->array
.layout
);
1694 if (info
->delta_disks
!= UnSet
&&
1695 info
->delta_disks
!= 0) {
1696 info
->array
.raid_disks
+= info
->delta_disks
;
1697 if (ioctl(fd
, SET_ARRAY_INFO
, &info
->array
) != 0) {
1698 fprintf(stderr
, Name
": failed to set raid disks\n");
1701 printf("raid_disks for %s set to %d\n",
1702 devname
, info
->array
.raid_disks
);
1704 if (info
->new_chunk
!= 0 &&
1705 info
->new_chunk
!= info
->array
.chunk_size
) {
1706 if (sysfs_set_num(info
, NULL
,
1707 "chunk_size", info
->new_chunk
) != 0) {
1708 fprintf(stderr
, Name
": failed to set chunk size\n");
1711 printf("chunk size for %s set to %d\n",
1712 devname
, info
->array
.chunk_size
);
1719 * There are three possibilities.
1720 * 1/ The array will shrink.
1721 * We need to ensure the reshape will pause before reaching
1722 * the 'critical section'. We also need to fork and wait for
1723 * that to happen. When it does we
1724 * suspend/backup/complete/unfreeze
1726 * 2/ The array will not change size.
1727 * This requires that we keep a backup of a sliding window
1728 * so that we can restore data after a crash. So we need
1729 * to fork and monitor progress.
1730 * In future we will allow the data_offset to change, so
1731 * a sliding backup becomes unnecessary.
1733 * 3/ The array will grow. This is relatively easy.
1734 * However the kernel's restripe routines will cheerfully
1735 * overwrite some early data before it is safe. So we
1736 * need to make a backup of the early parts of the array
1737 * and be ready to restore it if rebuild aborts very early.
1738 * For externally managed metadata, we still need a forked
1739 * child to monitor the reshape and suspend IO over the region
1740 * that is being reshaped.
1742 * We backup data by writing it to one spare, or to a
1743 * file which was given on command line.
1745 * In each case, we first make sure that storage is available
1746 * for the required backup.
1748 * - request the shape change.
1749 * - fork to handle backup etc.
1752 /* Check that we can hold all the data */
1753 get_dev_size(fd
, NULL
, &array_size
);
1754 if (reshape
.new_size
< (array_size
/512)) {
1756 Name
": this change will reduce the size of the array.\n"
1757 " use --grow --array-size first to truncate array.\n"
1758 " e.g. mdadm --grow %s --array-size %llu\n",
1759 devname
, reshape
.new_size
/2);
1764 sra
= sysfs_read(fd
, 0,
1765 GET_COMPONENT
|GET_DEVS
|GET_OFFSET
|GET_STATE
|GET_CHUNK
|
1769 fprintf(stderr
, Name
": %s: Cannot get array details from sysfs\n",
1775 /* Decide how many blocks (sectors) for a reshape
1776 * unit. The number we have so far is just a minimum
1778 blocks
= reshape
.backup_blocks
;
1779 if (reshape
.before
.data_disks
==
1780 reshape
.after
.data_disks
) {
1781 /* Make 'blocks' bigger for better throughput, but
1782 * not so big that we reject it below.
1783 * Try for 16 megabytes
1785 while (blocks
* 32 < sra
->component_size
&&
1789 fprintf(stderr
, Name
": Need to backup %luK of critical "
1790 "section..\n", blocks
/2);
1792 if (blocks
>= sra
->component_size
/2) {
1793 fprintf(stderr
, Name
": %s: Something wrong"
1794 " - reshape aborted\n",
1800 /* Now we need to open all these devices so we can read/write.
1802 nrdisks
= array
.raid_disks
+ sra
->array
.spare_disks
;
1803 fdlist
= malloc((1+nrdisks
) * sizeof(int));
1804 offsets
= malloc((1+nrdisks
) * sizeof(offsets
[0]));
1805 if (!fdlist
|| !offsets
) {
1806 fprintf(stderr
, Name
": malloc failed: grow aborted\n");
1811 d
= reshape_prepare_fdlist(devname
, sra
, array
.raid_disks
,
1812 nrdisks
, blocks
, backup_file
,
1818 if (backup_file
== NULL
) {
1819 if (reshape
.after
.data_disks
<= reshape
.before
.data_disks
) {
1821 Name
": %s: Cannot grow - need backup-file\n",
1825 } else if (sra
->array
.spare_disks
== 0) {
1826 fprintf(stderr
, Name
": %s: Cannot grow - need a spare or "
1827 "backup-file to backup critical section\n",
1833 if (!reshape_open_backup_file(backup_file
, fd
, devname
,
1835 fdlist
+d
, offsets
+d
)) {
1842 /* lastly, check that the internal stripe cache is
1843 * large enough, or it won't work.
1844 * It must hold at least 4 stripes of the larger
1847 cache
= max(info
->array
.chunk_size
, info
->new_chunk
);
1848 cache
*= 4; /* 4 stripes minimum */
1849 cache
/= 512; /* convert to sectors */
1850 disks
= min(reshape
.before
.data_disks
, reshape
.after
.data_disks
);
1851 /* make sure there is room for 'blocks' with a bit to spare */
1852 if (cache
< 16 + blocks
/ disks
)
1853 cache
= 16 + blocks
/ disks
;
1854 cache
/= (4096/512); /* Covert from sectors to pages */
1856 if (sra
->cache_size
< cache
)
1857 subarray_set_num(container
, sra
, "stripe_cache_size",
1860 /* Right, everything seems fine. Let's kick things off.
1861 * If only changing raid_disks, use ioctl, else use
1866 sra
->new_chunk
= info
->new_chunk
;
1868 if (info
->array
.chunk_size
== info
->new_chunk
&&
1869 reshape
.before
.layout
== reshape
.after
.layout
&&
1870 st
->ss
->external
== 0) {
1871 array
.raid_disks
= reshape
.after
.data_disks
+ reshape
.parity
;
1872 if (ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
1876 Name
": Cannot set device shape for %s: %s\n",
1877 devname
, strerror(errno
));
1880 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
1882 " Bitmap must be removed before"
1883 " shape can be changed\n");
1888 /* set them all just in case some old 'new_*' value
1889 * persists from some earlier problem
1891 int err
= err
; /* only used if rv==1, and always set if
1892 * rv==1, so initialisation not needed,
1893 * despite gcc warning
1895 if (sysfs_set_num(sra
, NULL
, "chunk_size", info
->new_chunk
) < 0)
1896 rv
= 1, err
= errno
;
1897 if (!rv
&& sysfs_set_num(sra
, NULL
, "layout",
1898 reshape
.after
.layout
) < 0)
1899 rv
= 1, err
= errno
;
1900 if (!rv
&& subarray_set_num(container
, sra
, "raid_disks",
1901 reshape
.after
.data_disks
+
1902 reshape
.parity
) < 0)
1903 rv
= 1, err
= errno
;
1905 fprintf(stderr
, Name
": Cannot set device shape for %s\n",
1909 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
1911 " Bitmap must be removed before"
1912 " shape can be changed\n");
1919 /* Now we just need to kick off the reshape and watch, while
1920 * handling backups of the data...
1921 * This is all done by a forked background process.
1923 switch(forked
? 0 : fork()) {
1925 fprintf(stderr
, Name
": Cannot run child to monitor reshape: %s\n",
1937 if (check_env("MDADM_GROW_VERIFY"))
1938 fd
= open(devname
, O_RDONLY
| O_DIRECT
);
1941 mlockall(MCL_FUTURE
);
1943 odisks
= reshape
.before
.data_disks
+ reshape
.parity
;
1945 if (st
->ss
->external
) {
1946 /* metadata handler takes it from here */
1947 done
= st
->ss
->manage_reshape(
1948 fd
, sra
, &reshape
, st
, blocks
,
1950 d
- odisks
, fdlist
+odisks
,
1953 done
= child_monitor(
1954 fd
, sra
, &reshape
, st
, blocks
,
1956 d
- odisks
, fdlist
+odisks
,
1959 if (backup_file
&& done
)
1960 unlink(backup_file
);
1966 if (!st
->ss
->external
&&
1967 !(reshape
.before
.data_disks
!= reshape
.after
.data_disks
1968 && info
->custom_array_size
) &&
1969 info
->new_level
== reshape
.level
&&
1971 /* no need to wait for the reshape to finish as
1972 * there is nothing more to do.
1978 if (st
->ss
->external
) {
1979 /* Re-load the metadata as much could have changed */
1980 int cfd
= open_dev(st
->container_dev
);
1982 ping_monitor(container
);
1983 st
->ss
->free_super(st
);
1984 st
->ss
->load_container(st
, cfd
, container
);
1989 /* set new array size if required customer_array_size is used
1992 if (reshape
.before
.data_disks
!=
1993 reshape
.after
.data_disks
&&
1994 info
->custom_array_size
) {
1995 struct mdinfo
*info2
;
1996 char *subarray
= strchr(info
->text_version
+1, '/')+1;
1998 info2
= st
->ss
->container_content(st
, subarray
);
2000 unsigned long long current_size
= 0;
2001 unsigned long long new_size
=
2002 info2
->custom_array_size
/2;
2004 if (sysfs_get_ll(sra
,
2007 ¤t_size
) == 0 &&
2008 new_size
> current_size
) {
2009 if (sysfs_set_num(sra
, NULL
,
2010 "array_size", new_size
)
2012 dprintf("Error: Cannot"
2015 dprintf("Array size "
2017 dprintf(" from %llu to %llu.\n",
2018 current_size
, new_size
);
2024 if (info
->new_level
!= reshape
.level
) {
2026 c
= map_num(pers
, info
->new_level
);
2028 err
= sysfs_set_str(sra
, NULL
, "level", c
);
2030 fprintf(stderr
, Name\
2031 ": %s: could not set level "
2032 "to %s\n", devname
, c
);
2041 if (orig_level
!= UnSet
&& sra
) {
2042 c
= map_num(pers
, orig_level
);
2043 if (c
&& sysfs_set_str(sra
, NULL
, "level", c
) == 0)
2044 fprintf(stderr
, Name
": aborting level change\n");
2051 int reshape_container(char *container
, int cfd
, char *devname
,
2052 struct supertype
*st
,
2053 struct mdinfo
*info
,
2058 struct mdinfo
*cc
= NULL
;
2060 /* component_size is not meaningful for a container,
2061 * so pass '-1' meaning 'no change'
2063 if (reshape_super(st
, -1, info
->new_level
,
2064 info
->new_layout
, info
->new_chunk
,
2065 info
->array
.raid_disks
+ info
->delta_disks
,
2066 backup_file
, devname
, quiet
))
2071 /* ping monitor to be sure that update is on disk
2073 ping_monitor(container
);
2076 case -1: /* error */
2077 perror("Cannot fork to complete reshape\n");
2079 default: /* parent */
2080 printf(Name
": multi-array reshape continues in background\n");
2087 /* For each member array with reshape_active,
2088 * we need to perform the reshape.
2089 * We pick the first array that needs reshaping and
2090 * reshape it. reshape_array() will re-read the metadata
2091 * so the next time through a different array should be
2092 * ready for reshape.
2094 struct mdinfo
*content
;
2097 struct mdstat_ent
*mdstat
;
2102 cc
= st
->ss
->container_content(st
, NULL
);
2104 for (content
= cc
; content
; content
= content
->next
) {
2106 if (!content
->reshape_active
)
2109 subarray
= strchr(content
->text_version
+1, '/')+1;
2110 mdstat
= mdstat_by_subdev(subarray
,
2111 devname2devnum(container
));
2119 fd
= open_dev_excl(mdstat
->devnum
);
2122 adev
= map_dev(dev2major(mdstat
->devnum
),
2123 dev2minor(mdstat
->devnum
),
2126 adev
= content
->text_version
;
2128 sysfs_init(content
, fd
, mdstat
->devnum
);
2130 rv
= reshape_array(container
, fd
, adev
, st
,
2132 backup_file
, quiet
, 1);
2143 * We run a child process in the background which performs the following
2145 * - wait for resync to reach a certain point
2146 * - suspend io to the following section
2147 * - backup that section
2148 * - allow resync to proceed further
2150 * - discard the backup.
2152 * When are combined in slightly different ways in the three cases.
2154 * - suspend/backup/allow/wait/resume/discard
2156 * - allow/wait/suspend/backup/allow/wait/resume/discard
2158 * - wait/resume/discard/suspend/backup/allow
2160 * suspend/backup/allow always come together
2161 * wait/resume/discard do too.
2162 * For the same-size case we have two backups to improve flow.
2166 int progress_reshape(struct mdinfo
*info
, struct reshape
*reshape
,
2167 unsigned long long backup_point
,
2168 unsigned long long wait_point
,
2169 unsigned long long *suspend_point
,
2170 unsigned long long *reshape_completed
)
2172 /* This function is called repeatedly by the reshape manager.
2173 * It determines how much progress can safely be made and allows
2175 * - 'info' identifies the array and particularly records in
2176 * ->reshape_progress the metadata's knowledge of progress
2177 * This is a sector offset from the start of the array
2178 * of the next array block to be relocated. This number
2179 * may increase from 0 or decrease from array_size, depending
2180 * on the type of reshape that is happening.
2181 * Note that in contrast, 'sync_completed' is a block count of the
2182 * reshape so far. It gives the distance between the start point
2183 * (head or tail of device) and the next place that data will be
2184 * written. It always increases.
2185 * - 'reshape' is the structure created by analyse_change
2186 * - 'backup_point' shows how much the metadata manager has backed-up
2187 * data. For reshapes with increasing progress, it is the next address
2188 * to be backed up, previous addresses have been backed-up. For
2189 * decreasing progress, it is the earliest address that has been
2190 * backed up - later address are also backed up.
2191 * So addresses between reshape_progress and backup_point are
2192 * backed up providing those are in the 'correct' order.
2193 * - 'wait_point' is an array address. When reshape_completed
2194 * passes this point, progress_reshape should return. It might
2195 * return earlier if it determines that ->reshape_progress needs
2196 * to be updated or further backup is needed.
2197 * - suspend_point is maintained by progress_reshape and the caller
2198 * should not touch it except to initialise to zero.
2199 * It is an array address and it only increases in 2.6.37 and earlier.
2200 * This makes it difficult to handle reducing reshapes with
2201 * external metadata.
2202 * However: it is similar to backup_point in that it records the
2203 * other end of a suspended region from reshape_progress.
2204 * it is moved to extend the region that is safe to backup and/or
2206 * - reshape_completed is read from sysfs and returned. The caller
2207 * should copy this into ->reshape_progress when it has reason to
2208 * believe that the metadata knows this, and any backup outside this
2212 * 1 if more data from backup_point - but only as far as suspend_point,
2213 * should be backed up
2214 * 0 if things are progressing smoothly
2215 * -1 if the reshape is finished, either because it is all done,
2216 * or due to an error.
2219 int advancing
= (reshape
->after
.data_disks
2220 >= reshape
->before
.data_disks
);
2221 unsigned long long need_backup
; /* need to eventually backup all the way
2224 unsigned long long read_offset
, write_offset
;
2225 unsigned long long write_range
;
2226 unsigned long long max_progress
, target
, completed
;
2227 unsigned long long array_size
= (info
->component_size
2228 * reshape
->before
.data_disks
);
2231 /* First, we unsuspend any region that is now known to be safe.
2232 * If suspend_point is on the 'wrong' side of reshape_progress, then
2233 * we don't have or need suspension at the moment. This is true for
2234 * native metadata when we don't need to back-up.
2237 if (info
->reshape_progress
<= *suspend_point
)
2238 sysfs_set_num(info
, NULL
, "suspend_lo",
2239 info
->reshape_progress
);
2241 /* Note: this won't work in 2.6.37 and before.
2242 * Something somewhere should make sure we don't need it!
2244 if (info
->reshape_progress
>= *suspend_point
)
2245 sysfs_set_num(info
, NULL
, "suspend_hi",
2246 info
->reshape_progress
);
2249 /* Now work out how far it is safe to progress.
2250 * If the read_offset for ->reshape_progress is less than
2251 * 'blocks' beyond the write_offset, we can only progress as far
2253 * Otherwise we can progress until the write_offset for the new location
2254 * reaches (within 'blocks' of) the read_offset at the current location.
2255 * However that region must be suspended unless we are using native
2257 * If we need to suspend more, we limit it to 128M per device, which is
2258 * rather arbitrary and should be some time-based calculation.
2260 read_offset
= info
->reshape_progress
/ reshape
->before
.data_disks
;
2261 write_offset
= info
->reshape_progress
/ reshape
->after
.data_disks
;
2262 write_range
= info
->new_chunk
/512;
2265 if (read_offset
< write_offset
+ write_range
) {
2266 max_progress
= backup_point
;
2267 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
)
2268 need_backup
= array_size
;
2270 need_backup
= reshape
->backup_blocks
;
2274 reshape
->after
.data_disks
;
2277 need_backup
= array_size
;
2278 if (read_offset
> write_offset
- write_range
) {
2279 max_progress
= backup_point
;
2280 if (max_progress
>= info
->reshape_progress
)
2285 reshape
->after
.data_disks
;
2286 /* If we are using internal metadata, then we can
2287 * progress all the way to the suspend_point without
2288 * worrying about backing-up/suspending along the
2291 if (max_progress
< *suspend_point
&&
2292 info
->array
.major_version
>= 0)
2293 max_progress
= *suspend_point
;
2297 /* We know it is safe to progress to 'max_progress' providing
2298 * it is suspended or we are using native metadata.
2299 * Consider extending suspend_point 128M per device if it
2300 * is less than 64M per device beyond reshape_progress.
2301 * But always do a multiple of 'blocks'
2302 * FIXME this is too big - it takes to long to complete
2305 target
= 64*1024*2 * min(reshape
->before
.data_disks
,
2306 reshape
->after
.data_disks
);
2307 target
/= reshape
->backup_blocks
;
2310 target
*= reshape
->backup_blocks
;
2312 /* For externally managed metadata we always need to suspend IO to
2313 * the area being reshaped so we regularly push suspend_point forward.
2314 * For native metadata we only need the suspend if we are going to do
2318 if ((need_backup
> info
->reshape_progress
2319 || info
->array
.major_version
< 0) &&
2320 *suspend_point
< info
->reshape_progress
+ target
) {
2321 if (need_backup
< *suspend_point
+ 2 * target
)
2322 *suspend_point
= need_backup
;
2323 else if (*suspend_point
+ 2 * target
< array_size
)
2324 *suspend_point
+= 2 * target
;
2326 *suspend_point
= array_size
;
2327 sysfs_set_num(info
, NULL
, "suspend_hi", *suspend_point
);
2328 if (max_progress
> *suspend_point
)
2329 max_progress
= *suspend_point
;
2332 if ((need_backup
< info
->reshape_progress
2333 || info
->array
.major_version
< 0) &&
2334 *suspend_point
> info
->reshape_progress
- target
) {
2335 if (need_backup
> *suspend_point
- 2 * target
)
2336 *suspend_point
= need_backup
;
2337 else if (*suspend_point
>= 2 * target
)
2338 *suspend_point
-= 2 * target
;
2341 sysfs_set_num(info
, NULL
, "suspend_lo", *suspend_point
);
2342 if (max_progress
< *suspend_point
)
2343 max_progress
= *suspend_point
;
2347 /* now set sync_max to allow that progress. sync_max, like
2348 * sync_completed is a count of sectors written per device, so
2349 * we find the difference between max_progress and the start point,
2350 * and divide that by after.data_disks to get a sync_max
2352 * At the same time we convert wait_point to a similar number
2353 * for comparing against sync_completed.
2355 /* scale down max_progress to per_disk */
2356 max_progress
/= reshape
->after
.data_disks
;
2357 /* Round to chunk size as some kernels give an erroneously high number */
2358 max_progress
/= info
->new_chunk
/512;
2359 max_progress
*= info
->new_chunk
/512;
2360 /* Limit progress to the whole device */
2361 if (max_progress
> info
->component_size
)
2362 max_progress
= info
->component_size
;
2363 wait_point
/= reshape
->after
.data_disks
;
2365 /* switch from 'device offset' to 'processed block count' */
2366 max_progress
= info
->component_size
- max_progress
;
2367 wait_point
= info
->component_size
- wait_point
;
2370 sysfs_set_num(info
, NULL
, "sync_max", max_progress
);
2372 /* Now wait. If we have already reached the point that we were
2373 * asked to wait to, don't wait at all, else wait for any change.
2374 * We need to select on 'sync_completed' as that is the place that
2375 * notifications happen, but we are really interested in
2376 * 'reshape_position'
2378 fd
= sysfs_get_fd(info
, NULL
, "sync_completed");
2382 if (sysfs_fd_get_ll(fd
, &completed
) < 0) {
2386 while (completed
< max_progress
&& completed
< wait_point
) {
2387 /* Check that sync_action is still 'reshape' to avoid
2388 * waiting forever on a dead array
2392 if (sysfs_get_str(info
, NULL
, "sync_action",
2394 strncmp(action
, "reshape", 7) != 0)
2398 select(fd
+1, NULL
, NULL
, &rfds
, NULL
);
2399 if (sysfs_fd_get_ll(fd
, &completed
) < 0) {
2404 /* some kernels can give an incorrectly high 'completed' number */
2405 completed
/= (info
->new_chunk
/512);
2406 completed
*= (info
->new_chunk
/512);
2407 /* Convert 'completed' back in to a 'progress' number */
2408 completed
*= reshape
->after
.data_disks
;
2410 completed
= info
->component_size
* reshape
->after
.data_disks
2413 *reshape_completed
= completed
;
2417 /* We return the need_backup flag. Caller will decide
2418 * how much - a multiple of ->backup_blocks up to *suspend_point
2421 ? (need_backup
> info
->reshape_progress
)
2422 : (need_backup
< info
->reshape_progress
);
2426 /* FIXME return status is never checked */
2427 static int grow_backup(struct mdinfo
*sra
,
2428 unsigned long long offset
, /* per device */
2429 unsigned long stripes
, /* per device, in old chunks */
2430 int *sources
, unsigned long long *offsets
,
2431 int disks
, int chunk
, int level
, int layout
,
2432 int dests
, int *destfd
, unsigned long long *destoffsets
,
2433 int part
, int *degraded
,
2436 /* Backup 'blocks' sectors at 'offset' on each device of the array,
2437 * to storage 'destfd' (offset 'destoffsets'), after first
2438 * suspending IO. Then allow resync to continue
2439 * over the suspended section.
2440 * Use part 'part' of the backup-super-block.
2445 unsigned long long ll
;
2447 //printf("offset %llu\n", offset);
2453 /* Check that array hasn't become degraded, else we might backup the wrong data */
2454 sysfs_get_ll(sra
, NULL
, "degraded", &ll
);
2455 new_degraded
= (int)ll
;
2456 if (new_degraded
!= *degraded
) {
2457 /* check each device to ensure it is still working */
2459 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
2460 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
2462 if (sd
->disk
.state
& (1<<MD_DISK_SYNC
)) {
2464 if (sysfs_get_str(sra
, sd
, "state", sbuf
, 20) < 0 ||
2465 strstr(sbuf
, "faulty") ||
2466 strstr(sbuf
, "in_sync") == NULL
) {
2467 /* this device is dead */
2468 sd
->disk
.state
= (1<<MD_DISK_FAULTY
);
2469 if (sd
->disk
.raid_disk
>= 0 &&
2470 sources
[sd
->disk
.raid_disk
] >= 0) {
2471 close(sources
[sd
->disk
.raid_disk
]);
2472 sources
[sd
->disk
.raid_disk
] = -1;
2477 *degraded
= new_degraded
;
2480 bsb
.arraystart2
= __cpu_to_le64(offset
* odata
);
2481 bsb
.length2
= __cpu_to_le64(stripes
* (chunk
/512) * odata
);
2483 bsb
.arraystart
= __cpu_to_le64(offset
* odata
);
2484 bsb
.length
= __cpu_to_le64(stripes
* (chunk
/512) * odata
);
2487 bsb
.magic
[15] = '2';
2488 for (i
= 0; i
< dests
; i
++)
2490 lseek64(destfd
[i
], destoffsets
[i
] + __le64_to_cpu(bsb
.devstart2
)*512, 0);
2492 lseek64(destfd
[i
], destoffsets
[i
], 0);
2494 rv
= save_stripes(sources
, offsets
,
2495 disks
, chunk
, level
, layout
,
2497 offset
*512*odata
, stripes
* chunk
* odata
,
2502 bsb
.mtime
= __cpu_to_le64(time(0));
2503 for (i
= 0; i
< dests
; i
++) {
2504 bsb
.devstart
= __cpu_to_le64(destoffsets
[i
]/512);
2506 bsb
.sb_csum
= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
));
2507 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0)
2508 bsb
.sb_csum2
= bsb_csum((char*)&bsb
,
2509 ((char*)&bsb
.sb_csum2
)-((char*)&bsb
));
2512 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
] - 4096, 0)
2513 != destoffsets
[i
] - 4096)
2515 if (write(destfd
[i
], &bsb
, 512) != 512)
2517 if (destoffsets
[i
] > 4096) {
2518 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
]+stripes
*chunk
*odata
, 0) !=
2519 destoffsets
[i
]+stripes
*chunk
*odata
)
2521 if (write(destfd
[i
], &bsb
, 512) != 512)
2531 /* in 2.6.30, the value reported by sync_completed can be
2532 * less that it should be by one stripe.
2533 * This only happens when reshape hits sync_max and pauses.
2534 * So allow wait_backup to either extent sync_max further
2535 * than strictly necessary, or return before the
2536 * sync has got quite as far as we would really like.
2537 * This is what 'blocks2' is for.
2538 * The various caller give appropriate values so that
2541 /* FIXME return value is often ignored */
2542 static int forget_backup(
2543 int dests
, int *destfd
, unsigned long long *destoffsets
,
2547 * Erase backup 'part' (which is 0 or 1)
2553 bsb
.arraystart2
= __cpu_to_le64(0);
2554 bsb
.length2
= __cpu_to_le64(0);
2556 bsb
.arraystart
= __cpu_to_le64(0);
2557 bsb
.length
= __cpu_to_le64(0);
2559 bsb
.mtime
= __cpu_to_le64(time(0));
2561 for (i
= 0; i
< dests
; i
++) {
2562 bsb
.devstart
= __cpu_to_le64(destoffsets
[i
]/512);
2563 bsb
.sb_csum
= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
));
2564 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0)
2565 bsb
.sb_csum2
= bsb_csum((char*)&bsb
,
2566 ((char*)&bsb
.sb_csum2
)-((char*)&bsb
));
2567 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
]-4096, 0) !=
2568 destoffsets
[i
]-4096)
2571 write(destfd
[i
], &bsb
, 512) != 512)
2578 static void fail(char *msg
)
2581 rv
= (write(2, msg
, strlen(msg
)) != (int)strlen(msg
));
2582 rv
|= (write(2, "\n", 1) != 1);
2586 static char *abuf
, *bbuf
;
2587 static unsigned long long abuflen
;
2588 static void validate(int afd
, int bfd
, unsigned long long offset
)
2590 /* check that the data in the backup against the array.
2591 * This is only used for regression testing and should not
2592 * be used while the array is active
2596 lseek64(bfd
, offset
- 4096, 0);
2597 if (read(bfd
, &bsb2
, 512) != 512)
2598 fail("cannot read bsb");
2599 if (bsb2
.sb_csum
!= bsb_csum((char*)&bsb2
,
2600 ((char*)&bsb2
.sb_csum
)-((char*)&bsb2
)))
2601 fail("first csum bad");
2602 if (memcmp(bsb2
.magic
, "md_backup_data", 14) != 0)
2603 fail("magic is bad");
2604 if (memcmp(bsb2
.magic
, "md_backup_data-2", 16) == 0 &&
2605 bsb2
.sb_csum2
!= bsb_csum((char*)&bsb2
,
2606 ((char*)&bsb2
.sb_csum2
)-((char*)&bsb2
)))
2607 fail("second csum bad");
2609 if (__le64_to_cpu(bsb2
.devstart
)*512 != offset
)
2610 fail("devstart is wrong");
2613 unsigned long long len
= __le64_to_cpu(bsb2
.length
)*512;
2615 if (abuflen
< len
) {
2619 if (posix_memalign((void**)&abuf
, 4096, abuflen
) ||
2620 posix_memalign((void**)&bbuf
, 4096, abuflen
)) {
2622 /* just stop validating on mem-alloc failure */
2627 lseek64(bfd
, offset
, 0);
2628 if ((unsigned long long)read(bfd
, bbuf
, len
) != len
) {
2629 //printf("len %llu\n", len);
2630 fail("read first backup failed");
2632 lseek64(afd
, __le64_to_cpu(bsb2
.arraystart
)*512, 0);
2633 if ((unsigned long long)read(afd
, abuf
, len
) != len
)
2634 fail("read first from array failed");
2635 if (memcmp(bbuf
, abuf
, len
) != 0) {
2638 printf("offset=%llu len=%llu\n",
2639 (unsigned long long)__le64_to_cpu(bsb2
.arraystart
)*512, len
);
2640 for (i
=0; i
<len
; i
++)
2641 if (bbuf
[i
] != abuf
[i
]) {
2642 printf("first diff byte %d\n", i
);
2646 fail("data1 compare failed");
2650 unsigned long long len
= __le64_to_cpu(bsb2
.length2
)*512;
2652 if (abuflen
< len
) {
2656 abuf
= malloc(abuflen
);
2657 bbuf
= malloc(abuflen
);
2660 lseek64(bfd
, offset
+__le64_to_cpu(bsb2
.devstart2
)*512, 0);
2661 if ((unsigned long long)read(bfd
, bbuf
, len
) != len
)
2662 fail("read second backup failed");
2663 lseek64(afd
, __le64_to_cpu(bsb2
.arraystart2
)*512, 0);
2664 if ((unsigned long long)read(afd
, abuf
, len
) != len
)
2665 fail("read second from array failed");
2666 if (memcmp(bbuf
, abuf
, len
) != 0)
2667 fail("data2 compare failed");
2671 int child_monitor(int afd
, struct mdinfo
*sra
, struct reshape
*reshape
,
2672 struct supertype
*st
, unsigned long blocks
,
2673 int *fds
, unsigned long long *offsets
,
2674 int dests
, int *destfd
, unsigned long long *destoffsets
)
2676 /* Monitor a reshape where backup is being performed using
2677 * 'native' mechanism - either to a backup file, or
2678 * to some space in a spare.
2682 unsigned long long speed
;
2683 unsigned long long suspend_point
, array_size
;
2684 unsigned long long backup_point
, wait_point
;
2685 unsigned long long reshape_completed
;
2687 int increasing
= reshape
->after
.data_disks
>= reshape
->before
.data_disks
;
2688 int part
= 0; /* The next part of the backup area to fill. It may already
2689 * be full, so we need to check */
2690 int level
= reshape
->level
;
2691 int layout
= reshape
->before
.layout
;
2692 int data
= reshape
->before
.data_disks
;
2693 int disks
= reshape
->before
.data_disks
+ reshape
->parity
;
2694 int chunk
= sra
->array
.chunk_size
;
2696 unsigned long stripes
;
2698 /* set up the backup-super-block. This requires the
2699 * uuid from the array.
2701 /* Find a superblock */
2702 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
2706 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
2708 dn
= map_dev(sd
->disk
.major
, sd
->disk
.minor
, 1);
2709 devfd
= dev_open(dn
, O_RDONLY
);
2712 ok
= st
->ss
->load_super(st
, devfd
, NULL
);
2718 fprintf(stderr
, Name
": Cannot find a superblock\n");
2722 memset(&bsb
, 0, 512);
2723 memcpy(bsb
.magic
, "md_backup_data-1", 16);
2724 st
->ss
->uuid_from_super(st
, (int*)&bsb
.set_uuid
);
2725 bsb
.mtime
= __cpu_to_le64(time(0));
2726 bsb
.devstart2
= blocks
;
2728 stripes
= blocks
/ (sra
->array
.chunk_size
/512) /
2729 reshape
->before
.data_disks
;
2731 if (posix_memalign((void**)&buf
, 4096, disks
* chunk
))
2732 /* Don't start the 'reshape' */
2734 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
) {
2735 sysfs_get_ll(sra
, NULL
, "sync_speed_min", &speed
);
2736 sysfs_set_num(sra
, NULL
, "sync_speed_min", 200000);
2740 array_size
= sra
->component_size
* reshape
->after
.data_disks
;
2741 backup_point
= sra
->reshape_progress
;
2744 array_size
= sra
->component_size
* reshape
->before
.data_disks
;
2745 backup_point
= array_size
;
2746 suspend_point
= array_size
;
2752 /* Want to return as soon the oldest backup slot can
2753 * be released as that allows us to start backing up
2754 * some more, providing suspend_point has been
2755 * advanced, which it should have.
2758 wait_point
= array_size
;
2759 if (part
== 0 && __le64_to_cpu(bsb
.length
) > 0)
2760 wait_point
= (__le64_to_cpu(bsb
.arraystart
) +
2761 __le64_to_cpu(bsb
.length
));
2762 if (part
== 1 && __le64_to_cpu(bsb
.length2
) > 0)
2763 wait_point
= (__le64_to_cpu(bsb
.arraystart2
) +
2764 __le64_to_cpu(bsb
.length2
));
2767 if (part
== 0 && __le64_to_cpu(bsb
.length
) > 0)
2768 wait_point
= __le64_to_cpu(bsb
.arraystart
);
2769 if (part
== 1 && __le64_to_cpu(bsb
.length2
) > 0)
2770 wait_point
= __le64_to_cpu(bsb
.arraystart2
);
2773 rv
= progress_reshape(sra
, reshape
,
2774 backup_point
, wait_point
,
2775 &suspend_point
, &reshape_completed
);
2776 /* external metadata would need to ping_monitor here */
2777 sra
->reshape_progress
= reshape_completed
;
2779 /* Clear any backup region that is before 'here' */
2781 if (reshape_completed
>= (__le64_to_cpu(bsb
.arraystart
) +
2782 __le64_to_cpu(bsb
.length
)))
2783 forget_backup(dests
, destfd
,
2785 if (reshape_completed
>= (__le64_to_cpu(bsb
.arraystart2
) +
2786 __le64_to_cpu(bsb
.length2
)))
2787 forget_backup(dests
, destfd
,
2790 if (reshape_completed
<= (__le64_to_cpu(bsb
.arraystart
)))
2791 forget_backup(dests
, destfd
,
2793 if (reshape_completed
<= (__le64_to_cpu(bsb
.arraystart2
)))
2794 forget_backup(dests
, destfd
,
2804 unsigned long long offset
;
2805 unsigned long actual_stripes
;
2806 /* Need to backup some data.
2807 * If 'part' is not used and the desired
2808 * backup size is suspended, do a backup,
2809 * then consider the next part.
2811 /* Check that 'part' is unused */
2812 if (part
== 0 && __le64_to_cpu(bsb
.length
) != 0)
2814 if (part
== 1 && __le64_to_cpu(bsb
.length2
) != 0)
2817 offset
= backup_point
/ data
;
2818 actual_stripes
= stripes
;
2820 if (offset
+ actual_stripes
* (chunk
/512) >
2821 sra
->component_size
)
2822 actual_stripes
= ((sra
->component_size
- offset
)
2824 if (offset
+ actual_stripes
* (chunk
/512) >
2828 if (offset
< actual_stripes
* (chunk
/512))
2829 actual_stripes
= offset
/ (chunk
/512);
2830 offset
-= actual_stripes
* (chunk
/512);
2831 if (offset
< suspend_point
/data
)
2834 grow_backup(sra
, offset
, actual_stripes
,
2836 disks
, chunk
, level
, layout
,
2837 dests
, destfd
, destoffsets
,
2838 part
, °raded
, buf
);
2839 validate(afd
, destfd
[0], destoffsets
[0]);
2840 /* record where 'part' is up to */
2843 backup_point
+= actual_stripes
* (chunk
/512) * data
;
2845 backup_point
-= actual_stripes
* (chunk
/512) * data
;
2849 /* FIXME maybe call progress_reshape one more time instead */
2850 abort_reshape(sra
); /* remove any remaining suspension */
2851 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
)
2852 sysfs_set_num(sra
, NULL
, "sync_speed_min", speed
);
2854 return 1; /* FIXME what does this mean? */
2858 * If any spare contains md_back_data-1 which is recent wrt mtime,
2859 * write that data into the array and update the super blocks with
2860 * the new reshape_progress
2862 int Grow_restart(struct supertype
*st
, struct mdinfo
*info
, int *fdlist
, int cnt
,
2863 char *backup_file
, int verbose
)
2867 unsigned long long *offsets
;
2868 unsigned long long nstripe
, ostripe
;
2871 if (info
->new_level
!= info
->array
.level
)
2872 return 1; /* Cannot handle level changes (they are instantaneous) */
2874 odata
= info
->array
.raid_disks
- info
->delta_disks
- 1;
2875 if (info
->array
.level
== 6) odata
--; /* number of data disks */
2876 ndata
= info
->array
.raid_disks
- 1;
2877 if (info
->new_level
== 6) ndata
--;
2879 old_disks
= info
->array
.raid_disks
- info
->delta_disks
;
2881 if (info
->delta_disks
<= 0)
2882 /* Didn't grow, so the backup file must have
2886 for (i
=old_disks
-(backup_file
?1:0); i
<cnt
; i
++) {
2887 struct mdinfo dinfo
;
2890 char *devname
, namebuf
[20];
2892 /* This was a spare and may have some saved data on it.
2893 * Load the superblock, find and load the
2894 * backup_super_block.
2895 * If either fail, go on to next device.
2896 * If the backup contains no new info, just return
2897 * else restore data and update all superblocks
2899 if (i
== old_disks
-1) {
2900 fd
= open(backup_file
, O_RDONLY
);
2902 fprintf(stderr
, Name
": backup file %s inaccessible: %s\n",
2903 backup_file
, strerror(errno
));
2906 devname
= backup_file
;
2911 if (st
->ss
->load_super(st
, fd
, NULL
))
2914 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
2915 st
->ss
->free_super(st
);
2918 (dinfo
.data_offset
+ dinfo
.component_size
- 8) <<9,
2920 fprintf(stderr
, Name
": Cannot seek on device %d\n", i
);
2921 continue; /* Cannot seek */
2923 sprintf(namebuf
, "device-%d", i
);
2926 if (read(fd
, &bsb
, sizeof(bsb
)) != sizeof(bsb
)) {
2928 fprintf(stderr
, Name
": Cannot read from %s\n", devname
);
2929 continue; /* Cannot read */
2931 if (memcmp(bsb
.magic
, "md_backup_data-1", 16) != 0 &&
2932 memcmp(bsb
.magic
, "md_backup_data-2", 16) != 0) {
2934 fprintf(stderr
, Name
": No backup metadata on %s\n", devname
);
2937 if (bsb
.sb_csum
!= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
))) {
2939 fprintf(stderr
, Name
": Bad backup-metadata checksum on %s\n", devname
);
2940 continue; /* bad checksum */
2942 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0 &&
2943 bsb
.sb_csum2
!= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum2
)-((char*)&bsb
))) {
2945 fprintf(stderr
, Name
": Bad backup-metadata checksum2 on %s\n", devname
);
2946 continue; /* Bad second checksum */
2948 if (memcmp(bsb
.set_uuid
,info
->uuid
, 16) != 0) {
2950 fprintf(stderr
, Name
": Wrong uuid on backup-metadata on %s\n", devname
);
2951 continue; /* Wrong uuid */
2954 /* array utime and backup-mtime should be updated at much the same time, but it seems that
2955 * sometimes they aren't... So allow considerable flexability in matching, and allow
2956 * this test to be overridden by an environment variable.
2958 if (info
->array
.utime
> (int)__le64_to_cpu(bsb
.mtime
) + 2*60*60 ||
2959 info
->array
.utime
< (int)__le64_to_cpu(bsb
.mtime
) - 10*60) {
2960 if (check_env("MDADM_GROW_ALLOW_OLD")) {
2961 fprintf(stderr
, Name
": accepting backup with timestamp %lu "
2962 "for array with timestamp %lu\n",
2963 (unsigned long)__le64_to_cpu(bsb
.mtime
),
2964 (unsigned long)info
->array
.utime
);
2967 fprintf(stderr
, Name
": too-old timestamp on "
2968 "backup-metadata on %s\n", devname
);
2969 continue; /* time stamp is too bad */
2973 if (bsb
.magic
[15] == '1') {
2974 if (info
->delta_disks
>= 0) {
2975 /* reshape_progress is increasing */
2976 if (__le64_to_cpu(bsb
.arraystart
) + __le64_to_cpu(bsb
.length
) <
2977 info
->reshape_progress
) {
2980 fprintf(stderr
, Name
": backup-metadata found on %s but is not needed\n", devname
);
2981 continue; /* No new data here */
2984 /* reshape_progress is decreasing */
2985 if (__le64_to_cpu(bsb
.arraystart
) >=
2986 info
->reshape_progress
)
2987 goto nonew
; /* No new data here */
2990 if (info
->delta_disks
>= 0) {
2991 /* reshape_progress is increasing */
2992 if (__le64_to_cpu(bsb
.arraystart
) + __le64_to_cpu(bsb
.length
) <
2993 info
->reshape_progress
&&
2994 __le64_to_cpu(bsb
.arraystart2
) + __le64_to_cpu(bsb
.length2
) <
2995 info
->reshape_progress
)
2996 goto nonew
; /* No new data here */
2998 /* reshape_progress is decreasing */
2999 if (__le64_to_cpu(bsb
.arraystart
) >=
3000 info
->reshape_progress
&&
3001 __le64_to_cpu(bsb
.arraystart2
) >=
3002 info
->reshape_progress
)
3003 goto nonew
; /* No new data here */
3006 if (lseek64(fd
, __le64_to_cpu(bsb
.devstart
)*512, 0)< 0) {
3009 fprintf(stderr
, Name
": Failed to verify secondary backup-metadata block on %s\n",
3011 continue; /* Cannot seek */
3013 /* There should be a duplicate backup superblock 4k before here */
3014 if (lseek64(fd
, -4096, 1) < 0 ||
3015 read(fd
, &bsb2
, sizeof(bsb2
)) != sizeof(bsb2
))
3016 goto second_fail
; /* Cannot find leading superblock */
3017 if (bsb
.magic
[15] == '1')
3018 bsbsize
= offsetof(struct mdp_backup_super
, pad1
);
3020 bsbsize
= offsetof(struct mdp_backup_super
, pad
);
3021 if (memcmp(&bsb2
, &bsb
, bsbsize
) != 0)
3022 goto second_fail
; /* Cannot find leading superblock */
3024 /* Now need the data offsets for all devices. */
3025 offsets
= malloc(sizeof(*offsets
)*info
->array
.raid_disks
);
3026 for(j
=0; j
<info
->array
.raid_disks
; j
++) {
3029 if (st
->ss
->load_super(st
, fdlist
[j
], NULL
))
3030 /* FIXME should be this be an error */
3032 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
3033 st
->ss
->free_super(st
);
3034 offsets
[j
] = dinfo
.data_offset
* 512;
3036 printf(Name
": restoring critical section\n");
3038 if (restore_stripes(fdlist
, offsets
,
3039 info
->array
.raid_disks
,
3043 fd
, __le64_to_cpu(bsb
.devstart
)*512,
3044 __le64_to_cpu(bsb
.arraystart
)*512,
3045 __le64_to_cpu(bsb
.length
)*512)) {
3046 /* didn't succeed, so giveup */
3048 fprintf(stderr
, Name
": Error restoring backup from %s\n",
3053 if (bsb
.magic
[15] == '2' &&
3054 restore_stripes(fdlist
, offsets
,
3055 info
->array
.raid_disks
,
3059 fd
, __le64_to_cpu(bsb
.devstart
)*512 +
3060 __le64_to_cpu(bsb
.devstart2
)*512,
3061 __le64_to_cpu(bsb
.arraystart2
)*512,
3062 __le64_to_cpu(bsb
.length2
)*512)) {
3063 /* didn't succeed, so giveup */
3065 fprintf(stderr
, Name
": Error restoring second backup from %s\n",
3071 /* Ok, so the data is restored. Let's update those superblocks. */
3073 if (info
->delta_disks
>= 0) {
3074 info
->reshape_progress
= __le64_to_cpu(bsb
.arraystart
) +
3075 __le64_to_cpu(bsb
.length
);
3076 if (bsb
.magic
[15] == '2') {
3077 unsigned long long p2
= __le64_to_cpu(bsb
.arraystart2
) +
3078 __le64_to_cpu(bsb
.length2
);
3079 if (p2
> info
->reshape_progress
)
3080 info
->reshape_progress
= p2
;
3083 info
->reshape_progress
= __le64_to_cpu(bsb
.arraystart
);
3084 if (bsb
.magic
[15] == '2') {
3085 unsigned long long p2
= __le64_to_cpu(bsb
.arraystart2
);
3086 if (p2
< info
->reshape_progress
)
3087 info
->reshape_progress
= p2
;
3090 for (j
=0; j
<info
->array
.raid_disks
; j
++) {
3091 if (fdlist
[j
] < 0) continue;
3092 if (st
->ss
->load_super(st
, fdlist
[j
], NULL
))
3094 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
3095 dinfo
.reshape_progress
= info
->reshape_progress
;
3096 st
->ss
->update_super(st
, &dinfo
,
3097 "_reshape_progress",
3099 st
->ss
->store_super(st
, fdlist
[j
]);
3100 st
->ss
->free_super(st
);
3104 /* Didn't find any backup data, try to see if any
3107 if (info
->delta_disks
< 0) {
3108 /* When shrinking, the critical section is at the end.
3109 * So see if we are before the critical section.
3111 unsigned long long first_block
;
3112 nstripe
= ostripe
= 0;
3114 while (ostripe
>= nstripe
) {
3115 ostripe
+= info
->array
.chunk_size
/ 512;
3116 first_block
= ostripe
* odata
;
3117 nstripe
= first_block
/ ndata
/ (info
->new_chunk
/512) *
3118 (info
->new_chunk
/512);
3121 if (info
->reshape_progress
>= first_block
)
3124 if (info
->delta_disks
> 0) {
3125 /* See if we are beyond the critical section. */
3126 unsigned long long last_block
;
3127 nstripe
= ostripe
= 0;
3129 while (nstripe
>= ostripe
) {
3130 nstripe
+= info
->new_chunk
/ 512;
3131 last_block
= nstripe
* ndata
;
3132 ostripe
= last_block
/ odata
/ (info
->array
.chunk_size
/512) *
3133 (info
->array
.chunk_size
/512);
3136 if (info
->reshape_progress
>= last_block
)
3139 /* needed to recover critical section! */
3141 fprintf(stderr
, Name
": Failed to find backup of critical section\n");
3145 int Grow_continue(int mdfd
, struct supertype
*st
, struct mdinfo
*info
,
3148 int err
= sysfs_set_str(info
, NULL
, "array_state", "readonly");
3151 return reshape_array(NULL
, mdfd
, "array", st
, info
, 1, backup_file
, 0, 0);