2 * mdadm - manage Linux "md" devices aka RAID arrays.
4 * Copyright (C) 2001-2013 Neil Brown <neilb@suse.de>
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License as published by
9 * the Free Software Foundation; either version 2 of the License, or
10 * (at your option) any later version.
12 * This program is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 * GNU General Public License for more details.
17 * You should have received a copy of the GNU General Public License
18 * along with this program; if not, write to the Free Software
19 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
22 * Email: <neilb@suse.de>
32 #if ! defined(__BIG_ENDIAN) && ! defined(__LITTLE_ENDIAN)
33 #error no endian defined
38 int restore_backup(struct supertype
*st
,
39 struct mdinfo
*content
,
49 int disk_count
= next_spare
+ working_disks
;
50 char *backup_file
= *backup_filep
;
52 dprintf("Called restore_backup()\n");
53 fdlist
= xmalloc(sizeof(int) * disk_count
);
55 enable_fds(next_spare
);
56 for (i
= 0; i
< next_spare
; i
++)
58 for (dev
= content
->devs
; dev
; dev
= dev
->next
) {
64 fd
= dev_open(buf
, O_RDWR
);
66 if (dev
->disk
.raid_disk
>= 0)
67 fdlist
[dev
->disk
.raid_disk
] = fd
;
69 fdlist
[next_spare
++] = fd
;
73 backup_file
= locate_backup(content
->sys_name
);
74 *backup_filep
= backup_file
;
77 if (st
->ss
->external
&& st
->ss
->recover_backup
)
78 err
= st
->ss
->recover_backup(st
, content
);
80 err
= Grow_restart(st
, content
, fdlist
, next_spare
,
81 backup_file
, verbose
> 0);
83 while (next_spare
> 0) {
85 if (fdlist
[next_spare
] >= 0)
86 close(fdlist
[next_spare
]);
90 pr_err("Failed to restore critical"
91 " section for reshape - sorry.\n");
93 pr_err("Possibly you need"
94 " to specify a --backup-file\n");
98 dprintf("restore_backup() returns status OK.\n");
102 int Grow_Add_device(char *devname
, int fd
, char *newdev
)
104 /* Add a device to an active array.
105 * Currently, just extend a linear array.
106 * This requires writing a new superblock on the
107 * new device, calling the kernel to add the device,
108 * and if that succeeds, update the superblock on
110 * This means that we need to *find* all other devices.
117 struct supertype
*st
= NULL
;
118 char *subarray
= NULL
;
120 if (ioctl(fd
, GET_ARRAY_INFO
, &info
.array
) < 0) {
121 pr_err("cannot get array info for %s\n", devname
);
125 if (info
.array
.level
!= -1) {
126 pr_err("can only add devices to linear arrays\n");
130 st
= super_by_fd(fd
, &subarray
);
132 pr_err("cannot handle arrays with superblock version %d\n",
133 info
.array
.major_version
);
138 pr_err("Cannot grow linear sub-arrays yet\n");
144 nfd
= open(newdev
, O_RDWR
|O_EXCL
|O_DIRECT
);
146 pr_err("cannot open %s\n", newdev
);
151 if ((stb
.st_mode
& S_IFMT
) != S_IFBLK
) {
152 pr_err("%s is not a block device!\n", newdev
);
157 /* now check out all the devices and make sure we can read the
159 for (d
=0 ; d
< info
.array
.raid_disks
; d
++) {
160 mdu_disk_info_t disk
;
163 st
->ss
->free_super(st
);
166 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0) {
167 pr_err("cannot get device detail for device %d\n",
173 dv
= map_dev(disk
.major
, disk
.minor
, 1);
175 pr_err("cannot find device file for device %d\n",
181 fd2
= dev_open(dv
, O_RDWR
);
183 pr_err("cannot open device file %s\n", dv
);
189 if (st
->ss
->load_super(st
, fd2
, NULL
)) {
190 pr_err("cannot find super block on %s\n", dv
);
198 /* Ok, looks good. Lets update the superblock and write it out to
202 info
.disk
.number
= d
;
203 info
.disk
.major
= major(stb
.st_rdev
);
204 info
.disk
.minor
= minor(stb
.st_rdev
);
205 info
.disk
.raid_disk
= d
;
206 info
.disk
.state
= (1 << MD_DISK_SYNC
) | (1 << MD_DISK_ACTIVE
);
207 st
->ss
->update_super(st
, &info
, "linear-grow-new", newdev
,
210 if (st
->ss
->store_super(st
, nfd
)) {
211 pr_err("Cannot store new superblock on %s\n",
218 if (ioctl(fd
, ADD_NEW_DISK
, &info
.disk
) != 0) {
219 pr_err("Cannot add new disk to this array\n");
222 /* Well, that seems to have worked.
223 * Now go through and update all superblocks
226 if (ioctl(fd
, GET_ARRAY_INFO
, &info
.array
) < 0) {
227 pr_err("cannot get array info for %s\n", devname
);
232 for (d
=0 ; d
< info
.array
.raid_disks
; d
++) {
233 mdu_disk_info_t disk
;
237 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0) {
238 pr_err("cannot get device detail for device %d\n",
242 dv
= map_dev(disk
.major
, disk
.minor
, 1);
244 pr_err("cannot find device file for device %d\n",
248 fd2
= dev_open(dv
, O_RDWR
);
250 pr_err("cannot open device file %s\n", dv
);
253 if (st
->ss
->load_super(st
, fd2
, NULL
)) {
254 pr_err("cannot find super block on %s\n", dv
);
258 info
.array
.raid_disks
= nd
+1;
259 info
.array
.nr_disks
= nd
+1;
260 info
.array
.active_disks
= nd
+1;
261 info
.array
.working_disks
= nd
+1;
263 st
->ss
->update_super(st
, &info
, "linear-grow-update", dv
,
266 if (st
->ss
->store_super(st
, fd2
)) {
267 pr_err("Cannot store new superblock on %s\n", dv
);
277 int Grow_addbitmap(char *devname
, int fd
, struct context
*c
, struct shape
*s
)
280 * First check that array doesn't have a bitmap
281 * Then create the bitmap
284 * For internal bitmaps, we need to check the version,
285 * find all the active devices, and write the bitmap block
288 mdu_bitmap_file_t bmf
;
289 mdu_array_info_t array
;
290 struct supertype
*st
;
291 char *subarray
= NULL
;
292 int major
= BITMAP_MAJOR_HI
;
293 int vers
= md_get_version(fd
);
294 unsigned long long bitmapsize
, array_size
;
297 major
= BITMAP_MAJOR_HOSTENDIAN
;
298 pr_err("Warning - bitmaps created on this kernel"
299 " are not portable\n"
300 " between different architectures. Consider upgrading"
301 " the Linux kernel.\n");
304 if (ioctl(fd
, GET_BITMAP_FILE
, &bmf
) != 0) {
306 pr_err("Memory allocation failure.\n");
308 pr_err("bitmaps not supported by this kernel.\n");
311 if (bmf
.pathname
[0]) {
312 if (strcmp(s
->bitmap_file
,"none")==0) {
313 if (ioctl(fd
, SET_BITMAP_FILE
, -1)!= 0) {
314 pr_err("failed to remove bitmap %s\n",
320 pr_err("%s already has a bitmap (%s)\n",
321 devname
, bmf
.pathname
);
324 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
325 pr_err("cannot get array status for %s\n", devname
);
328 if (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)) {
329 if (strcmp(s
->bitmap_file
, "none")==0) {
330 array
.state
&= ~(1<<MD_SB_BITMAP_PRESENT
);
331 if (ioctl(fd
, SET_ARRAY_INFO
, &array
)!= 0) {
332 pr_err("failed to remove internal bitmap.\n");
337 pr_err("Internal bitmap already present on %s\n",
342 if (strcmp(s
->bitmap_file
, "none") == 0) {
343 pr_err("no bitmap found on %s\n", devname
);
346 if (array
.level
<= 0) {
347 pr_err("Bitmaps not meaningful with level %s\n",
348 map_num(pers
, array
.level
)?:"of this array");
351 bitmapsize
= array
.size
;
353 if (get_dev_size(fd
, NULL
, &array_size
) &&
354 array_size
> (0x7fffffffULL
<<9)) {
355 /* Array is big enough that we cannot trust array.size
356 * try other approaches
358 bitmapsize
= get_component_size(fd
);
360 if (bitmapsize
== 0) {
361 pr_err("Cannot reliably determine size of array to create bitmap - sorry.\n");
365 if (array
.level
== 10) {
366 int ncopies
= (array
.layout
&255)*((array
.layout
>>8)&255);
367 bitmapsize
= bitmapsize
* array
.raid_disks
/ ncopies
;
370 st
= super_by_fd(fd
, &subarray
);
372 pr_err("Cannot understand version %d.%d\n",
373 array
.major_version
, array
.minor_version
);
377 pr_err("Cannot add bitmaps to sub-arrays yet\n");
382 if (strcmp(s
->bitmap_file
, "internal") == 0) {
385 int offset_setable
= 0;
387 if (st
->ss
->add_internal_bitmap
== NULL
) {
388 pr_err("Internal bitmaps not supported "
389 "with %s metadata\n", st
->ss
->name
);
392 mdi
= sysfs_read(fd
, NULL
, GET_BITMAP_LOCATION
);
395 for (d
=0; d
< st
->max_devs
; d
++) {
396 mdu_disk_info_t disk
;
399 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0)
401 if (disk
.major
== 0 &&
404 if ((disk
.state
& (1<<MD_DISK_SYNC
))==0)
406 dv
= map_dev(disk
.major
, disk
.minor
, 1);
408 int fd2
= dev_open(dv
, O_RDWR
);
411 if (st
->ss
->load_super(st
, fd2
, NULL
)==0) {
412 if (st
->ss
->add_internal_bitmap(
414 &s
->bitmap_chunk
, c
->delay
, s
->write_behind
,
415 bitmapsize
, offset_setable
,
418 st
->ss
->write_bitmap(st
, fd2
);
420 pr_err("failed to create internal bitmap"
421 " - chunksize problem.\n");
429 if (offset_setable
) {
430 st
->ss
->getinfo_super(st
, mdi
, NULL
);
431 sysfs_init(mdi
, fd
, NULL
);
432 rv
= sysfs_set_num_signed(mdi
, NULL
, "bitmap/location",
435 array
.state
|= (1<<MD_SB_BITMAP_PRESENT
);
436 rv
= ioctl(fd
, SET_ARRAY_INFO
, &array
);
440 pr_err("Cannot add bitmap while array is"
441 " resyncing or reshaping etc.\n");
442 pr_err("failed to set internal bitmap.\n");
449 int max_devs
= st
->max_devs
;
451 /* try to load a superblock */
452 for (d
= 0; d
< max_devs
; d
++) {
453 mdu_disk_info_t disk
;
457 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0)
459 if ((disk
.major
==0 && disk
.minor
==0) ||
460 (disk
.state
& (1<<MD_DISK_REMOVED
)))
462 dv
= map_dev(disk
.major
, disk
.minor
, 1);
465 fd2
= dev_open(dv
, O_RDONLY
);
467 if (st
->ss
->load_super(st
, fd2
, NULL
) == 0) {
469 st
->ss
->uuid_from_super(st
, uuid
);
476 pr_err("cannot find UUID for array!\n");
479 if (CreateBitmap(s
->bitmap_file
, c
->force
, (char*)uuid
, s
->bitmap_chunk
,
480 c
->delay
, s
->write_behind
, bitmapsize
, major
)) {
483 bitmap_fd
= open(s
->bitmap_file
, O_RDWR
);
485 pr_err("weird: %s cannot be opened\n",
489 if (ioctl(fd
, SET_BITMAP_FILE
, bitmap_fd
) < 0) {
492 pr_err("Cannot add bitmap while array is"
493 " resyncing or reshaping etc.\n");
494 pr_err("Cannot set bitmap file for %s: %s\n",
495 devname
, strerror(err
));
504 * When reshaping an array we might need to backup some data.
505 * This is written to all spares with a 'super_block' describing it.
506 * The superblock goes 4K from the end of the used space on the
508 * It if written after the backup is complete.
509 * It has the following structure.
512 static struct mdp_backup_super
{
513 char magic
[16]; /* md_backup_data-1 or -2 */
516 /* start/sizes in 512byte sectors */
517 __u64 devstart
; /* address on backup device/file of data */
520 __u32 sb_csum
; /* csum of preceeding bytes. */
522 __u64 devstart2
; /* offset in to data of second section */
525 __u32 sb_csum2
; /* csum of preceeding bytes. */
527 } __attribute__((aligned(512))) bsb
, bsb2
;
529 static __u32
bsb_csum(char *buf
, int len
)
533 for (i
= 0; i
< len
; i
++)
534 csum
= (csum
<<3) + buf
[0];
535 return __cpu_to_le32(csum
);
538 static int check_idle(struct supertype
*st
)
540 /* Check that all member arrays for this container, or the
541 * container of this array, are idle
543 char *container
= (st
->container_devnm
[0]
544 ? st
->container_devnm
: st
->devnm
);
545 struct mdstat_ent
*ent
, *e
;
548 ent
= mdstat_read(0, 0);
549 for (e
= ent
; e
; e
= e
->next
) {
550 if (!is_container_member(e
, container
))
552 if (e
->percent
>= 0) {
561 static int freeze_container(struct supertype
*st
)
563 char *container
= (st
->container_devnm
[0]
564 ? st
->container_devnm
: st
->devnm
);
569 if (block_monitor(container
, 1)) {
570 pr_err("failed to freeze container\n");
577 static void unfreeze_container(struct supertype
*st
)
579 char *container
= (st
->container_devnm
[0]
580 ? st
->container_devnm
: st
->devnm
);
582 unblock_monitor(container
, 1);
585 static int freeze(struct supertype
*st
)
587 /* Try to freeze resync/rebuild on this array/container.
588 * Return -1 if the array is busy,
589 * return -2 container cannot be frozen,
590 * return 0 if this kernel doesn't support 'frozen'
591 * return 1 if it worked.
593 if (st
->ss
->external
)
594 return freeze_container(st
);
596 struct mdinfo
*sra
= sysfs_read(-1, st
->devnm
, GET_VERSION
);
602 /* Need to clear any 'read-auto' status */
603 if (sysfs_get_str(sra
, NULL
, "array_state", buf
, 20) > 0 &&
604 strncmp(buf
, "read-auto", 9) == 0)
605 sysfs_set_str(sra
, NULL
, "array_state", "clean");
607 err
= sysfs_freeze_array(sra
);
613 static void unfreeze(struct supertype
*st
)
615 if (st
->ss
->external
)
616 return unfreeze_container(st
);
618 struct mdinfo
*sra
= sysfs_read(-1, st
->devnm
, GET_VERSION
);
622 sysfs_get_str(sra
, NULL
, "sync_action", buf
, 20) > 0
623 && strcmp(buf
, "frozen\n") == 0) {
624 printf("unfreeze\n");
625 sysfs_set_str(sra
, NULL
, "sync_action", "idle");
631 static void wait_reshape(struct mdinfo
*sra
)
633 int fd
= sysfs_get_fd(sra
, NULL
, "sync_action");
639 while (sysfs_fd_get_str(fd
, action
, 20) > 0 &&
640 strncmp(action
, "reshape", 7) == 0)
641 sysfs_wait(fd
, NULL
);
645 static int reshape_super(struct supertype
*st
, unsigned long long size
,
646 int level
, int layout
, int chunksize
, int raid_disks
,
647 int delta_disks
, char *backup_file
, char *dev
,
648 int direction
, int verbose
)
650 /* nothing extra to check in the native case */
651 if (!st
->ss
->external
)
653 if (!st
->ss
->reshape_super
||
654 !st
->ss
->manage_reshape
) {
655 pr_err("%s metadata does not support reshape\n",
660 return st
->ss
->reshape_super(st
, size
, level
, layout
, chunksize
,
661 raid_disks
, delta_disks
, backup_file
, dev
,
665 static void sync_metadata(struct supertype
*st
)
667 if (st
->ss
->external
) {
668 if (st
->update_tail
) {
669 flush_metadata_updates(st
);
670 st
->update_tail
= &st
->updates
;
672 st
->ss
->sync_metadata(st
);
676 static int subarray_set_num(char *container
, struct mdinfo
*sra
, char *name
, int n
)
678 /* when dealing with external metadata subarrays we need to be
679 * prepared to handle EAGAIN. The kernel may need to wait for
680 * mdmon to mark the array active so the kernel can handle
681 * allocations/writeback when preparing the reshape action
682 * (md_allow_write()). We temporarily disable safe_mode_delay
683 * to close a race with the array_state going clean before the
684 * next write to raid_disks / stripe_cache_size
689 /* only 'raid_disks' and 'stripe_cache_size' trigger md_allow_write */
691 (strcmp(name
, "raid_disks") != 0 &&
692 strcmp(name
, "stripe_cache_size") != 0))
693 return sysfs_set_num(sra
, NULL
, name
, n
);
695 rc
= sysfs_get_str(sra
, NULL
, "safe_mode_delay", safe
, sizeof(safe
));
698 sysfs_set_num(sra
, NULL
, "safe_mode_delay", 0);
699 rc
= sysfs_set_num(sra
, NULL
, name
, n
);
700 if (rc
< 0 && errno
== EAGAIN
) {
701 ping_monitor(container
);
702 /* if we get EAGAIN here then the monitor is not active
705 rc
= sysfs_set_num(sra
, NULL
, name
, n
);
707 sysfs_set_str(sra
, NULL
, "safe_mode_delay", safe
);
711 int start_reshape(struct mdinfo
*sra
, int already_running
,
712 int before_data_disks
, int data_disks
)
715 unsigned long long sync_max_to_set
;
717 sysfs_set_num(sra
, NULL
, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL
);
718 err
= sysfs_set_num(sra
, NULL
, "suspend_hi", sra
->reshape_progress
);
719 err
= err
?: sysfs_set_num(sra
, NULL
, "suspend_lo",
720 sra
->reshape_progress
);
721 if (before_data_disks
<= data_disks
)
722 sync_max_to_set
= sra
->reshape_progress
/ data_disks
;
724 sync_max_to_set
= (sra
->component_size
* data_disks
725 - sra
->reshape_progress
) / data_disks
;
726 if (!already_running
)
727 sysfs_set_num(sra
, NULL
, "sync_min", sync_max_to_set
);
728 err
= err
?: sysfs_set_num(sra
, NULL
, "sync_max", sync_max_to_set
);
729 if (!already_running
)
730 err
= err
?: sysfs_set_str(sra
, NULL
, "sync_action", "reshape");
735 void abort_reshape(struct mdinfo
*sra
)
737 sysfs_set_str(sra
, NULL
, "sync_action", "idle");
738 sysfs_set_num(sra
, NULL
, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL
);
739 sysfs_set_num(sra
, NULL
, "suspend_hi", 0);
740 sysfs_set_num(sra
, NULL
, "suspend_lo", 0);
741 sysfs_set_num(sra
, NULL
, "sync_min", 0);
742 // It isn't safe to reset sync_max as we aren't monitoring.
743 // Array really should be stopped at this point.
746 int remove_disks_for_takeover(struct supertype
*st
,
751 struct mdinfo
*remaining
;
754 if (sra
->array
.level
== 10)
755 nr_of_copies
= layout
& 0xff;
756 else if (sra
->array
.level
== 1)
757 nr_of_copies
= sra
->array
.raid_disks
;
761 remaining
= sra
->devs
;
763 /* for each 'copy', select one device and remove from the list. */
764 for (slot
= 0; slot
< sra
->array
.raid_disks
; slot
+= nr_of_copies
) {
765 struct mdinfo
**diskp
;
768 /* Find a working device to keep */
769 for (diskp
= &remaining
; *diskp
; diskp
= &(*diskp
)->next
) {
770 struct mdinfo
*disk
= *diskp
;
772 if (disk
->disk
.raid_disk
< slot
)
774 if (disk
->disk
.raid_disk
>= slot
+ nr_of_copies
)
776 if (disk
->disk
.state
& (1<<MD_DISK_REMOVED
))
778 if (disk
->disk
.state
& (1<<MD_DISK_FAULTY
))
780 if (!(disk
->disk
.state
& (1<<MD_DISK_SYNC
)))
783 /* We have found a good disk to use! */
785 disk
->next
= sra
->devs
;
794 if (slot
< sra
->array
.raid_disks
) {
795 /* didn't find all slots */
801 sra
->devs
= remaining
;
805 /* Remove all 'remaining' devices from the array */
807 struct mdinfo
*sd
= remaining
;
808 remaining
= sd
->next
;
810 sysfs_set_str(sra
, sd
, "state", "faulty");
811 sysfs_set_str(sra
, sd
, "slot", "none");
812 /* for external metadata disks should be removed in mdmon */
813 if (!st
->ss
->external
)
814 sysfs_set_str(sra
, sd
, "state", "remove");
815 sd
->disk
.state
|= (1<<MD_DISK_REMOVED
);
816 sd
->disk
.state
&= ~(1<<MD_DISK_SYNC
);
817 sd
->next
= sra
->devs
;
823 void reshape_free_fdlist(int *fdlist
,
824 unsigned long long *offsets
,
829 for (i
= 0; i
< size
; i
++)
837 int reshape_prepare_fdlist(char *devname
,
841 unsigned long blocks
,
844 unsigned long long *offsets
)
850 for (d
= 0; d
<= nrdisks
; d
++)
853 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
854 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
856 if (sd
->disk
.state
& (1<<MD_DISK_SYNC
)) {
857 char *dn
= map_dev(sd
->disk
.major
,
859 fdlist
[sd
->disk
.raid_disk
]
860 = dev_open(dn
, O_RDONLY
);
861 offsets
[sd
->disk
.raid_disk
] = sd
->data_offset
*512;
862 if (fdlist
[sd
->disk
.raid_disk
] < 0) {
863 pr_err("%s: cannot open component %s\n",
864 devname
, dn
? dn
: "-unknown-");
868 } else if (backup_file
== NULL
) {
870 char *dn
= map_dev(sd
->disk
.major
,
872 fdlist
[d
] = dev_open(dn
, O_RDWR
);
873 offsets
[d
] = (sd
->data_offset
+ sra
->component_size
- blocks
- 8)*512;
875 pr_err("%s: cannot open component %s\n",
876 devname
, dn
? dn
: "-unknown-");
887 int reshape_open_backup_file(char *backup_file
,
892 unsigned long long *offsets
,
896 /* Return 1 on success, 0 on any form of failure */
897 /* need to check backup file is large enough */
903 *fdlist
= open(backup_file
, O_RDWR
|O_CREAT
|(restart
? O_TRUNC
: O_EXCL
),
907 pr_err("%s: cannot create backup file %s: %s\n",
908 devname
, backup_file
, strerror(errno
));
911 /* Guard against backup file being on array device.
912 * If array is partitioned or if LVM etc is in the
913 * way this will not notice, but it is better than
916 fstat(*fdlist
, &stb
);
919 if (stb
.st_rdev
== dev
) {
920 pr_err("backup file must NOT be"
921 " on the array being reshaped.\n");
927 for (i
=0; i
< blocks
+ 8 ; i
++) {
928 if (write(*fdlist
, buf
, 512) != 512) {
929 pr_err("%s: cannot create"
930 " backup file %s: %s\n",
931 devname
, backup_file
, strerror(errno
));
935 if (fsync(*fdlist
) != 0) {
936 pr_err("%s: cannot create backup file %s: %s\n",
937 devname
, backup_file
, strerror(errno
));
941 if (!restart
&& strncmp(backup_file
, MAP_DIR
, strlen(MAP_DIR
)) != 0) {
942 char *bu
= make_backup(sys_name
);
943 symlink(backup_file
, bu
);
950 unsigned long compute_backup_blocks(int nchunk
, int ochunk
,
951 unsigned int ndata
, unsigned int odata
)
953 unsigned long a
, b
, blocks
;
954 /* So how much do we need to backup.
955 * We need an amount of data which is both a whole number of
956 * old stripes and a whole number of new stripes.
957 * So LCM for (chunksize*datadisks).
959 a
= (ochunk
/512) * odata
;
960 b
= (nchunk
/512) * ndata
;
963 /* LCM == product / GCD */
964 blocks
= (ochunk
/512) * (nchunk
/512) * odata
* ndata
/ a
;
969 char *analyse_change(char *devname
, struct mdinfo
*info
, struct reshape
*re
)
971 /* Based on the current array state in info->array and
972 * the changes in info->new_* etc, determine:
973 * - whether the change is possible
974 * - Intermediate level/raid_disks/layout
975 * - whether a restriping reshape is needed
976 * - number of sectors in minimum change unit. This
977 * will cover a whole number of stripes in 'before' and
980 * Return message if the change should be rejected
981 * NULL if the change can be achieved
983 * This can be called as part of starting a reshape, or
984 * when assembling an array that is undergoing reshape.
986 int near
, far
, offset
, copies
;
988 int old_chunk
, new_chunk
;
989 /* delta_parity records change in number of devices
990 * caused by level change
992 int delta_parity
= 0;
994 memset(re
, 0, sizeof(*re
));
996 /* If a new level not explicitly given, we assume no-change */
997 if (info
->new_level
== UnSet
)
998 info
->new_level
= info
->array
.level
;
1000 if (info
->new_chunk
)
1001 switch (info
->new_level
) {
1007 /* chunk size is meaningful, must divide component_size
1010 if (info
->component_size
% (info
->new_chunk
/512)) {
1011 unsigned long long shrink
= info
->component_size
;
1012 shrink
&= ~(unsigned long long)(info
->new_chunk
/512-1);
1013 pr_err("New chunk size (%dK) does not evenly divide device size (%lluk)\n",
1014 info
->new_chunk
/1024, info
->component_size
/2);
1015 pr_err("After shrinking any filesystem, \"mdadm --grow %s --size %llu\"\n",
1017 pr_err("will shrink the array so the given chunk size would work.\n");
1022 return "chunk size not meaningful for this level";
1025 info
->new_chunk
= info
->array
.chunk_size
;
1027 switch (info
->array
.level
) {
1029 return "Cannot understand this RAID level";
1031 /* RAID1 can convert to RAID1 with different disks, or
1032 * raid5 with 2 disks, or
1035 if (info
->new_level
> 1 &&
1036 (info
->component_size
& 7))
1037 return "Cannot convert RAID1 of this size - "
1038 "reduce size to multiple of 4K first.";
1039 if (info
->new_level
== 0) {
1040 if (info
->delta_disks
!= UnSet
&&
1041 info
->delta_disks
!= 0)
1042 return "Cannot change number of disks "
1043 "with RAID1->RAID0 conversion";
1045 re
->before
.data_disks
= 1;
1046 re
->after
.data_disks
= 1;
1049 if (info
->new_level
== 1) {
1050 if (info
->delta_disks
== UnSet
)
1051 /* Don't know what to do */
1052 return "no change requested for Growing RAID1";
1056 if (info
->array
.raid_disks
== 2 &&
1057 info
->new_level
== 5) {
1060 re
->before
.data_disks
= 1;
1061 if (info
->delta_disks
!= UnSet
&&
1062 info
->delta_disks
!= 0)
1063 re
->after
.data_disks
= 1 + info
->delta_disks
;
1065 re
->after
.data_disks
= 1;
1066 if (re
->after
.data_disks
< 1)
1067 return "Number of disks too small for RAID5";
1069 re
->before
.layout
= ALGORITHM_LEFT_SYMMETRIC
;
1070 info
->array
.chunk_size
= 65536;
1073 /* Could do some multi-stage conversions, but leave that to
1076 return "Impossibly level change request for RAID1";
1079 /* RAID10 can be converted from near mode to
1080 * RAID0 by removing some devices.
1081 * It can also be reshaped if the kernel supports
1084 switch (info
->new_level
) {
1086 if ((info
->array
.layout
& ~0xff) != 0x100)
1087 return "Cannot Grow RAID10 with far/offset layout";
1088 /* number of devices must be multiple of number of copies */
1089 if (info
->array
.raid_disks
% (info
->array
.layout
& 0xff))
1090 return "RAID10 layout too complex for Grow operation";
1092 new_disks
= (info
->array
.raid_disks
1093 / (info
->array
.layout
& 0xff));
1094 if (info
->delta_disks
== UnSet
)
1095 info
->delta_disks
= (new_disks
1096 - info
->array
.raid_disks
);
1098 if (info
->delta_disks
!= new_disks
- info
->array
.raid_disks
)
1099 return "New number of raid-devices impossible for RAID10";
1100 if (info
->new_chunk
&&
1101 info
->new_chunk
!= info
->array
.chunk_size
)
1102 return "Cannot change chunk-size with RAID10 Grow";
1106 re
->before
.data_disks
= new_disks
;
1107 re
->after
.data_disks
= re
->before
.data_disks
;
1111 near
= info
->array
.layout
& 0xff;
1112 far
= (info
->array
.layout
>> 8) & 0xff;
1113 offset
= info
->array
.layout
& 0x10000;
1114 if (far
> 1 && !offset
)
1115 return "Cannot reshape RAID10 in far-mode";
1116 copies
= near
* far
;
1118 old_chunk
= info
->array
.chunk_size
* far
;
1120 if (info
->new_layout
== UnSet
)
1121 info
->new_layout
= info
->array
.layout
;
1123 near
= info
->new_layout
& 0xff;
1124 far
= (info
->new_layout
>> 8) & 0xff;
1125 offset
= info
->new_layout
& 0x10000;
1126 if (far
> 1 && !offset
)
1127 return "Cannot reshape RAID10 to far-mode";
1128 if (near
* far
!= copies
)
1129 return "Cannot change number of copies"
1130 " when reshaping RAID10";
1132 if (info
->delta_disks
== UnSet
)
1133 info
->delta_disks
= 0;
1134 new_disks
= (info
->array
.raid_disks
+
1137 new_chunk
= info
->new_chunk
* far
;
1140 re
->before
.layout
= info
->array
.layout
;
1141 re
->before
.data_disks
= info
->array
.raid_disks
;
1142 re
->after
.layout
= info
->new_layout
;
1143 re
->after
.data_disks
= new_disks
;
1144 /* For RAID10 we don't do backup but do allow reshape,
1145 * so set backup_blocks to INVALID_SECTORS rather than
1147 * And there is no need to synchronise stripes on both
1148 * 'old' and 'new'. So the important
1149 * number is the minimum data_offset difference
1150 * which is the larger of (offset copies * chunk).
1152 re
->backup_blocks
= INVALID_SECTORS
;
1153 re
->min_offset_change
= max(old_chunk
, new_chunk
) / 512;
1154 if (new_disks
< re
->before
.data_disks
&&
1155 info
->space_after
< re
->min_offset_change
)
1156 /* Reduce component size by one chunk */
1157 re
->new_size
= (info
->component_size
-
1158 re
->min_offset_change
);
1160 re
->new_size
= info
->component_size
;
1161 re
->new_size
= re
->new_size
* new_disks
/ copies
;
1165 return "RAID10 can only be changed to RAID0";
1168 /* RAID0 can be converted to RAID10, or to RAID456 */
1169 if (info
->new_level
== 10) {
1170 if (info
->new_layout
== UnSet
&& info
->delta_disks
== UnSet
) {
1171 /* Assume near=2 layout */
1172 info
->new_layout
= 0x102;
1173 info
->delta_disks
= info
->array
.raid_disks
;
1175 if (info
->new_layout
== UnSet
) {
1176 int copies
= 1 + (info
->delta_disks
1177 / info
->array
.raid_disks
);
1178 if (info
->array
.raid_disks
* (copies
-1)
1179 != info
->delta_disks
)
1180 return "Impossible number of devices"
1181 " for RAID0->RAID10";
1182 info
->new_layout
= 0x100 + copies
;
1184 if (info
->delta_disks
== UnSet
) {
1185 int copies
= info
->new_layout
& 0xff;
1186 if (info
->new_layout
!= 0x100 + copies
)
1187 return "New layout impossible"
1188 " for RAID0->RAID10";;
1189 info
->delta_disks
= (copies
- 1) *
1190 info
->array
.raid_disks
;
1192 if (info
->new_chunk
&&
1193 info
->new_chunk
!= info
->array
.chunk_size
)
1194 return "Cannot change chunk-size with RAID0->RAID10";
1197 re
->before
.data_disks
= (info
->array
.raid_disks
+
1199 re
->after
.data_disks
= re
->before
.data_disks
;
1200 re
->before
.layout
= info
->new_layout
;
1204 /* RAID0 can also covert to RAID0/4/5/6 by first converting to
1205 * a raid4 style layout of the final level.
1207 switch (info
->new_level
) {
1212 re
->before
.layout
= 0;
1217 re
->before
.layout
= ALGORITHM_PARITY_N
;
1218 if (info
->new_layout
== UnSet
)
1219 info
->new_layout
= map_name(r5layout
, "default");
1224 re
->before
.layout
= ALGORITHM_PARITY_N
;
1225 if (info
->new_layout
== UnSet
)
1226 info
->new_layout
= map_name(r6layout
, "default");
1229 return "Impossible level change requested";
1231 re
->before
.data_disks
= info
->array
.raid_disks
;
1232 /* determining 'after' layout happens outside this 'switch' */
1236 info
->array
.layout
= ALGORITHM_PARITY_N
;
1238 switch (info
->new_level
) {
1242 re
->level
= info
->array
.level
;
1243 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1244 re
->before
.layout
= info
->array
.layout
;
1248 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1249 re
->before
.layout
= info
->array
.layout
;
1254 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1255 switch (info
->array
.layout
) {
1256 case ALGORITHM_LEFT_ASYMMETRIC
:
1257 re
->before
.layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
1259 case ALGORITHM_RIGHT_ASYMMETRIC
:
1260 re
->before
.layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
1262 case ALGORITHM_LEFT_SYMMETRIC
:
1263 re
->before
.layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
1265 case ALGORITHM_RIGHT_SYMMETRIC
:
1266 re
->before
.layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
1268 case ALGORITHM_PARITY_0
:
1269 re
->before
.layout
= ALGORITHM_PARITY_0_6
;
1271 case ALGORITHM_PARITY_N
:
1272 re
->before
.layout
= ALGORITHM_PARITY_N_6
;
1275 return "Cannot convert an array with this layout";
1279 if (info
->array
.raid_disks
!= 2)
1280 return "Can only convert a 2-device array to RAID1";
1281 if (info
->delta_disks
!= UnSet
&&
1282 info
->delta_disks
!= 0)
1283 return "Cannot set raid_disk when "
1284 "converting RAID5->RAID1";
1286 info
->new_chunk
= 0;
1289 return "Impossible level change requested";
1293 switch (info
->new_level
) {
1299 re
->before
.data_disks
= info
->array
.raid_disks
- 2;
1300 re
->before
.layout
= info
->array
.layout
;
1303 return "Impossible level change requested";
1308 /* If we reached here then it looks like a re-stripe is
1309 * happening. We have determined the intermediate level
1310 * and initial raid_disks/layout and stored these in 're'.
1312 * We need to deduce the final layout that can be atomically
1313 * converted to the end state.
1315 switch (info
->new_level
) {
1317 /* We can only get to RAID0 from RAID4 or RAID5
1318 * with appropriate layout and one extra device
1320 if (re
->level
!= 4 && re
->level
!= 5)
1321 return "Cannot covert to RAID0 from this level";
1323 switch (re
->level
) {
1325 re
->before
.layout
= 0;
1326 re
->after
.layout
= 0;
1329 re
->after
.layout
= ALGORITHM_PARITY_N
;
1335 /* We can only get to RAID4 from RAID5 */
1336 if (re
->level
!= 4 && re
->level
!= 5)
1337 return "Cannot convert to RAID4 from this level";
1339 switch (re
->level
) {
1341 re
->before
.layout
= 0;
1342 re
->after
.layout
= 0;
1345 re
->after
.layout
= ALGORITHM_PARITY_N
;
1351 /* We get to RAID5 from RAID5 or RAID6 */
1352 if (re
->level
!= 5 && re
->level
!= 6)
1353 return "Cannot convert to RAID5 from this level";
1355 switch (re
->level
) {
1357 if (info
->new_layout
== UnSet
)
1358 re
->after
.layout
= re
->before
.layout
;
1360 re
->after
.layout
= info
->new_layout
;
1363 if (info
->new_layout
== UnSet
)
1364 info
->new_layout
= re
->before
.layout
;
1366 /* after.layout needs to be raid6 version of new_layout */
1367 if (info
->new_layout
== ALGORITHM_PARITY_N
)
1368 re
->after
.layout
= ALGORITHM_PARITY_N
;
1371 char *ls
= map_num(r5layout
, info
->new_layout
);
1374 /* Current RAID6 layout has a RAID5
1377 strcat(strcpy(layout
, ls
), "-6");
1378 l
= map_name(r6layout
, layout
);
1380 return "Cannot find RAID6 layout"
1383 /* Current RAID6 has no equivalent.
1384 * If it is already a '-6' layout we
1385 * can leave it unchanged, else we must
1388 ls
= map_num(r6layout
, info
->new_layout
);
1390 strcmp(ls
+strlen(ls
)-2, "-6") != 0)
1391 return "Please specify new layout";
1392 l
= info
->new_layout
;
1394 re
->after
.layout
= l
;
1400 /* We must already be at level 6 */
1402 return "Impossible level change";
1403 if (info
->new_layout
== UnSet
)
1404 re
->after
.layout
= info
->array
.layout
;
1406 re
->after
.layout
= info
->new_layout
;
1409 return "Impossible level change requested";
1411 if (info
->delta_disks
== UnSet
)
1412 info
->delta_disks
= delta_parity
;
1414 re
->after
.data_disks
= (re
->before
.data_disks
1417 switch (re
->level
) {
1418 case 6: re
->parity
= 2;
1421 case 5: re
->parity
= 1;
1423 default: re
->parity
= 0;
1426 /* So we have a restripe operation, we need to calculate the number
1427 * of blocks per reshape operation.
1429 re
->new_size
= info
->component_size
* re
->before
.data_disks
;
1430 if (info
->new_chunk
== 0)
1431 info
->new_chunk
= info
->array
.chunk_size
;
1432 if (re
->after
.data_disks
== re
->before
.data_disks
&&
1433 re
->after
.layout
== re
->before
.layout
&&
1434 info
->new_chunk
== info
->array
.chunk_size
) {
1435 /* Nothing to change, can change level immediately. */
1436 re
->level
= info
->new_level
;
1437 re
->backup_blocks
= 0;
1440 if (re
->after
.data_disks
== 1 && re
->before
.data_disks
== 1) {
1441 /* chunk and layout changes make no difference */
1442 re
->level
= info
->new_level
;
1443 re
->backup_blocks
= 0;
1447 if (re
->after
.data_disks
== re
->before
.data_disks
&&
1448 get_linux_version() < 2006032)
1449 return "in-place reshape is not safe before 2.6.32 - sorry.";
1451 if (re
->after
.data_disks
< re
->before
.data_disks
&&
1452 get_linux_version() < 2006030)
1453 return "reshape to fewer devices is not supported before 2.6.30 - sorry.";
1455 re
->backup_blocks
= compute_backup_blocks(
1456 info
->new_chunk
, info
->array
.chunk_size
,
1457 re
->after
.data_disks
,
1458 re
->before
.data_disks
);
1459 re
->min_offset_change
= re
->backup_blocks
/ re
->before
.data_disks
;
1461 re
->new_size
= info
->component_size
* re
->after
.data_disks
;
1465 static int set_array_size(struct supertype
*st
, struct mdinfo
*sra
,
1468 struct mdinfo
*info
;
1472 if ((st
== NULL
) || (sra
== NULL
))
1475 if (text_version
== NULL
)
1476 text_version
= sra
->text_version
;
1477 subarray
= strchr(text_version
+1, '/')+1;
1478 info
= st
->ss
->container_content(st
, subarray
);
1480 unsigned long long current_size
= 0;
1481 unsigned long long new_size
=
1482 info
->custom_array_size
/2;
1484 if (sysfs_get_ll(sra
, NULL
, "array_size", ¤t_size
) == 0 &&
1485 new_size
> current_size
) {
1486 if (sysfs_set_num(sra
, NULL
, "array_size", new_size
)
1488 dprintf("Error: Cannot set array size");
1491 dprintf("Array size changed");
1493 dprintf(" from %llu to %llu.\n",
1494 current_size
, new_size
);
1498 dprintf("Error: set_array_size(): info pointer in NULL\n");
1503 static int reshape_array(char *container
, int fd
, char *devname
,
1504 struct supertype
*st
, struct mdinfo
*info
,
1505 int force
, struct mddev_dev
*devlist
,
1506 unsigned long long data_offset
,
1507 char *backup_file
, int verbose
, int forked
,
1508 int restart
, int freeze_reshape
);
1509 static int reshape_container(char *container
, char *devname
,
1511 struct supertype
*st
,
1512 struct mdinfo
*info
,
1514 char *backup_file
, int verbose
,
1515 int forked
, int restart
, int freeze_reshape
);
1517 int Grow_reshape(char *devname
, int fd
,
1518 struct mddev_dev
*devlist
,
1519 unsigned long long data_offset
,
1520 struct context
*c
, struct shape
*s
)
1522 /* Make some changes in the shape of an array.
1523 * The kernel must support the change.
1525 * There are three different changes. Each can trigger
1526 * a resync or recovery so we freeze that until we have
1527 * requested everything (if kernel supports freezing - 2.6.30).
1529 * - change size (i.e. component_size)
1531 * - change layout/chunksize/ndisks
1533 * The last can require a reshape. It is different on different
1534 * levels so we need to check the level before actioning it.
1535 * Some times the level change needs to be requested after the
1536 * reshape (e.g. raid6->raid5, raid5->raid0)
1539 struct mdu_array_info_s array
;
1541 struct supertype
*st
;
1542 char *subarray
= NULL
;
1546 char *container
= NULL
;
1549 struct mddev_dev
*dv
;
1555 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) < 0) {
1556 pr_err("%s is not an active md array - aborting\n",
1560 if (data_offset
!= INVALID_SECTORS
&& array
.level
!= 10
1561 && (array
.level
< 4 || array
.level
> 6)) {
1562 pr_err("--grow --data-offset not yet supported\n");
1567 (s
->chunk
|| s
->level
!= UnSet
|| s
->layout_str
|| s
->raiddisks
)) {
1568 pr_err("cannot change component size at the same time "
1569 "as other changes.\n"
1570 " Change size first, then check data is intact before "
1571 "making other changes.\n");
1575 if (s
->raiddisks
&& s
->raiddisks
< array
.raid_disks
&& array
.level
> 1 &&
1576 get_linux_version() < 2006032 &&
1577 !check_env("MDADM_FORCE_FEWER")) {
1578 pr_err("reducing the number of devices is not safe before Linux 2.6.32\n"
1579 " Please use a newer kernel\n");
1583 st
= super_by_fd(fd
, &subarray
);
1585 pr_err("Unable to determine metadata format for %s\n", devname
);
1588 if (s
->raiddisks
> st
->max_devs
) {
1589 pr_err("Cannot increase raid-disks on this array"
1590 " beyond %d\n", st
->max_devs
);
1594 /* in the external case we need to check that the requested reshape is
1595 * supported, and perform an initial check that the container holds the
1596 * pre-requisite spare devices (mdmon owns final validation)
1598 if (st
->ss
->external
) {
1602 container
= st
->container_devnm
;
1603 cfd
= open_dev_excl(st
->container_devnm
);
1605 container
= st
->devnm
;
1607 cfd
= open_dev_excl(st
->devnm
);
1611 pr_err("Unable to open container for %s\n",
1617 rv
= st
->ss
->load_container(st
, cfd
, NULL
);
1620 pr_err("Cannot read superblock for %s\n",
1626 /* check if operation is supported for metadata handler */
1627 if (st
->ss
->container_content
) {
1628 struct mdinfo
*cc
= NULL
;
1629 struct mdinfo
*content
= NULL
;
1631 cc
= st
->ss
->container_content(st
, subarray
);
1632 for (content
= cc
; content
; content
= content
->next
) {
1633 int allow_reshape
= 1;
1635 /* check if reshape is allowed based on metadata
1636 * indications stored in content.array.status
1638 if (content
->array
.state
& (1<<MD_SB_BLOCK_VOLUME
))
1640 if (content
->array
.state
1641 & (1<<MD_SB_BLOCK_CONTAINER_RESHAPE
))
1643 if (!allow_reshape
) {
1644 pr_err("cannot reshape arrays in"
1645 " container with unsupported"
1646 " metadata: %s(%s)\n",
1647 devname
, container
);
1655 if (mdmon_running(container
))
1656 st
->update_tail
= &st
->updates
;
1660 for (dv
= devlist
; dv
; dv
= dv
->next
)
1662 if (s
->raiddisks
> array
.raid_disks
&&
1663 array
.spare_disks
+added_disks
< (s
->raiddisks
- array
.raid_disks
) &&
1665 pr_err("Need %d spare%s to avoid degraded array,"
1666 " and only have %d.\n"
1667 " Use --force to over-ride this check.\n",
1668 s
->raiddisks
- array
.raid_disks
,
1669 s
->raiddisks
- array
.raid_disks
== 1 ? "" : "s",
1670 array
.spare_disks
+ added_disks
);
1674 sra
= sysfs_read(fd
, NULL
, GET_LEVEL
| GET_DISKS
| GET_DEVS
1675 | GET_STATE
| GET_VERSION
);
1677 if (st
->ss
->external
&& subarray
== NULL
) {
1678 array
.level
= LEVEL_CONTAINER
;
1679 sra
->array
.level
= LEVEL_CONTAINER
;
1682 pr_err("failed to read sysfs parameters for %s\n",
1686 frozen
= freeze(st
);
1688 /* freeze() already spewed the reason */
1691 } else if (frozen
< 0) {
1692 pr_err("%s is performing resync/recovery and cannot"
1693 " be reshaped\n", devname
);
1698 /* ========= set size =============== */
1699 if (s
->size
> 0 && (s
->size
== MAX_SIZE
|| s
->size
!= (unsigned)array
.size
)) {
1700 unsigned long long orig_size
= get_component_size(fd
)/2;
1701 unsigned long long min_csize
;
1703 int raid0_takeover
= 0;
1706 orig_size
= (unsigned) array
.size
;
1708 if (orig_size
== 0) {
1709 pr_err("Cannot set device size in this type of array.\n");
1714 if (reshape_super(st
, s
->size
, UnSet
, UnSet
, 0, 0, UnSet
, NULL
,
1715 devname
, APPLY_METADATA_CHANGES
, c
->verbose
> 0)) {
1720 if (st
->ss
->external
) {
1721 /* metadata can have size limitation
1722 * update size value according to metadata information
1724 struct mdinfo
*sizeinfo
=
1725 st
->ss
->container_content(st
, subarray
);
1727 unsigned long long new_size
=
1728 sizeinfo
->custom_array_size
/2;
1729 int data_disks
= get_data_disks(
1730 sizeinfo
->array
.level
,
1731 sizeinfo
->array
.layout
,
1732 sizeinfo
->array
.raid_disks
);
1733 new_size
/= data_disks
;
1734 dprintf("Metadata size correction from %llu to "
1735 "%llu (%llu)\n", orig_size
, new_size
,
1736 new_size
* data_disks
);
1738 sysfs_free(sizeinfo
);
1742 /* Update the size of each member device in case
1743 * they have been resized. This will never reduce
1744 * below the current used-size. The "size" attribute
1745 * understands '0' to mean 'max'.
1749 for (mdi
= sra
->devs
; mdi
; mdi
= mdi
->next
) {
1750 if (sysfs_set_num(sra
, mdi
, "size",
1751 s
->size
== MAX_SIZE
? 0 : s
->size
) < 0) {
1752 /* Probably kernel refusing to let us
1753 * reduce the size - not an error.
1757 if (array
.not_persistent
== 0 &&
1758 array
.major_version
== 0 &&
1759 get_linux_version() < 3001000) {
1760 /* Dangerous to allow size to exceed 2TB */
1761 unsigned long long csize
;
1762 if (sysfs_get_ll(sra
, mdi
, "size", &csize
) == 0) {
1763 if (csize
>= 2ULL*1024*1024*1024)
1764 csize
= 2ULL*1024*1024*1024;
1765 if ((min_csize
== 0 || (min_csize
1772 pr_err("Cannot set size on "
1773 "array members.\n");
1774 goto size_change_error
;
1776 if (min_csize
&& s
->size
> min_csize
) {
1777 pr_err("Cannot safely make this array "
1778 "use more than 2TB per device on this kernel.\n");
1780 goto size_change_error
;
1782 if (min_csize
&& s
->size
== MAX_SIZE
) {
1783 /* Don't let the kernel choose a size - it will get
1786 pr_err("Limited v0.90 array to "
1787 "2TB per device\n");
1788 s
->size
= min_csize
;
1790 if (st
->ss
->external
) {
1791 if (sra
->array
.level
== 0) {
1792 rv
= sysfs_set_str(sra
, NULL
, "level",
1796 /* get array parametes after takeover
1797 * to chane one parameter at time only
1799 rv
= ioctl(fd
, GET_ARRAY_INFO
, &array
);
1802 /* make sure mdmon is
1803 * aware of the new level */
1804 if (!mdmon_running(st
->container_devnm
))
1805 start_mdmon(st
->container_devnm
);
1806 ping_monitor(container
);
1807 if (mdmon_running(st
->container_devnm
) &&
1808 st
->update_tail
== NULL
)
1809 st
->update_tail
= &st
->updates
;
1812 if (s
->size
== MAX_SIZE
)
1814 array
.size
= s
->size
;
1815 if ((unsigned)array
.size
!= s
->size
) {
1816 /* got truncated to 32bit, write to
1817 * component_size instead
1820 rv
= sysfs_set_num(sra
, NULL
,
1821 "component_size", s
->size
);
1825 rv
= ioctl(fd
, SET_ARRAY_INFO
, &array
);
1827 /* manage array size when it is managed externally
1829 if ((rv
== 0) && st
->ss
->external
)
1830 rv
= set_array_size(st
, sra
, sra
->text_version
);
1833 if (raid0_takeover
) {
1834 /* do not recync non-existing parity,
1835 * we will drop it anyway
1837 sysfs_set_str(sra
, NULL
, "sync_action", "frozen");
1838 /* go back to raid0, drop parity disk
1840 sysfs_set_str(sra
, NULL
, "level", "raid0");
1841 ioctl(fd
, GET_ARRAY_INFO
, &array
);
1848 /* restore metadata */
1849 if (reshape_super(st
, orig_size
, UnSet
, UnSet
, 0, 0,
1850 UnSet
, NULL
, devname
,
1851 ROLLBACK_METADATA_CHANGES
,
1854 pr_err("Cannot set device size for %s: %s\n",
1855 devname
, strerror(err
));
1857 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
1858 cont_err("Bitmap must be removed before size can be changed\n");
1862 if (s
->assume_clean
) {
1863 /* This will fail on kernels older than 3.0 unless
1864 * a backport has been arranged.
1867 sysfs_set_str(sra
, NULL
, "resync_start", "none") < 0)
1868 pr_err("--assume-clean not supported with --grow on this kernel\n");
1870 ioctl(fd
, GET_ARRAY_INFO
, &array
);
1871 s
->size
= get_component_size(fd
)/2;
1873 s
->size
= array
.size
;
1874 if (c
->verbose
>= 0) {
1875 if (s
->size
== orig_size
)
1876 pr_err("component size of %s "
1877 "unchanged at %lluK\n",
1880 pr_err("component size of %s "
1881 "has been set to %lluK\n",
1885 } else if (array
.level
!= LEVEL_CONTAINER
) {
1886 s
->size
= get_component_size(fd
)/2;
1888 s
->size
= array
.size
;
1891 /* See if there is anything else to do */
1892 if ((s
->level
== UnSet
|| s
->level
== array
.level
) &&
1893 (s
->layout_str
== NULL
) &&
1894 (s
->chunk
== 0 || s
->chunk
== array
.chunk_size
) &&
1895 data_offset
== INVALID_SECTORS
&&
1896 (s
->raiddisks
== 0 || s
->raiddisks
== array
.raid_disks
)) {
1897 /* Nothing more to do */
1898 if (!changed
&& c
->verbose
>= 0)
1899 pr_err("%s: no change requested\n",
1904 /* ========= check for Raid10/Raid1 -> Raid0 conversion ===============
1905 * current implementation assumes that following conditions must be met:
1908 * - near_copies == 2
1910 if ((s
->level
== 0 && array
.level
== 10 && sra
&&
1911 array
.layout
== ((1 << 8) + 2) && !(array
.raid_disks
& 1)) ||
1912 (s
->level
== 0 && array
.level
== 1 && sra
)) {
1914 err
= remove_disks_for_takeover(st
, sra
, array
.layout
);
1916 dprintf(Name
": Array cannot be reshaped\n");
1922 /* Make sure mdmon has seen the device removal
1923 * and updated metadata before we continue with
1927 ping_monitor(container
);
1930 memset(&info
, 0, sizeof(info
));
1932 sysfs_init(&info
, fd
, NULL
);
1933 strcpy(info
.text_version
, sra
->text_version
);
1934 info
.component_size
= s
->size
*2;
1935 info
.new_level
= s
->level
;
1936 info
.new_chunk
= s
->chunk
* 1024;
1937 if (info
.array
.level
== LEVEL_CONTAINER
) {
1938 info
.delta_disks
= UnSet
;
1939 info
.array
.raid_disks
= s
->raiddisks
;
1940 } else if (s
->raiddisks
)
1941 info
.delta_disks
= s
->raiddisks
- info
.array
.raid_disks
;
1943 info
.delta_disks
= UnSet
;
1944 if (s
->layout_str
== NULL
) {
1945 info
.new_layout
= UnSet
;
1946 if (info
.array
.level
== 6 &&
1947 (info
.new_level
== 6 || info
.new_level
== UnSet
) &&
1948 info
.array
.layout
>= 16) {
1949 pr_err("%s has a non-standard layout. If you"
1950 " wish to preserve this\n", devname
);
1951 cont_err("during the reshape, please specify"
1952 " --layout=preserve\n");
1953 cont_err("If you want to change it, specify a"
1954 " layout or use --layout=normalise\n");
1958 } else if (strcmp(s
->layout_str
, "normalise") == 0 ||
1959 strcmp(s
->layout_str
, "normalize") == 0) {
1960 /* If we have a -6 RAID6 layout, remove the '-6'. */
1961 info
.new_layout
= UnSet
;
1962 if (info
.array
.level
== 6 && info
.new_level
== UnSet
) {
1964 strcpy(l
, map_num(r6layout
, info
.array
.layout
));
1965 h
= strrchr(l
, '-');
1966 if (h
&& strcmp(h
, "-6") == 0) {
1968 info
.new_layout
= map_name(r6layout
, l
);
1971 pr_err("%s is only meaningful when reshaping"
1972 " a RAID6 array.\n", s
->layout_str
);
1976 } else if (strcmp(s
->layout_str
, "preserve") == 0) {
1977 /* This means that a non-standard RAID6 layout
1980 * - When reshape a RAID6 (e.g. adding a device)
1981 * which is in a non-standard layout, it is OK
1982 * to preserve that layout.
1983 * - When converting a RAID5 to RAID6, leave it in
1984 * the XXX-6 layout, don't re-layout.
1986 if (info
.array
.level
== 6 && info
.new_level
== UnSet
)
1987 info
.new_layout
= info
.array
.layout
;
1988 else if (info
.array
.level
== 5 && info
.new_level
== 6) {
1990 strcpy(l
, map_num(r5layout
, info
.array
.layout
));
1992 info
.new_layout
= map_name(r6layout
, l
);
1994 pr_err("%s in only meaningful when reshaping"
1995 " to RAID6\n", s
->layout_str
);
2000 int l
= info
.new_level
;
2002 l
= info
.array
.level
;
2005 info
.new_layout
= map_name(r5layout
, s
->layout_str
);
2008 info
.new_layout
= map_name(r6layout
, s
->layout_str
);
2011 info
.new_layout
= parse_layout_10(s
->layout_str
);
2014 info
.new_layout
= parse_layout_faulty(s
->layout_str
);
2017 pr_err("layout not meaningful"
2018 " with this level\n");
2022 if (info
.new_layout
== UnSet
) {
2023 pr_err("layout %s not understood"
2024 " for this level\n",
2031 if (array
.level
== LEVEL_FAULTY
) {
2032 if (s
->level
!= UnSet
&& s
->level
!= array
.level
) {
2033 pr_err("cannot change level of Faulty device\n");
2037 pr_err("cannot set chunksize of Faulty device\n");
2040 if (s
->raiddisks
&& s
->raiddisks
!= 1) {
2041 pr_err("cannot set raid_disks of Faulty device\n");
2044 if (s
->layout_str
) {
2045 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
2046 dprintf("Cannot get array information.\n");
2049 array
.layout
= info
.new_layout
;
2050 if (ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
2051 pr_err("failed to set new layout\n");
2053 } else if (c
->verbose
>= 0)
2054 printf("layout for %s set to %d\n",
2055 devname
, array
.layout
);
2057 } else if (array
.level
== LEVEL_CONTAINER
) {
2058 /* This change is to be applied to every array in the
2059 * container. This is only needed when the metadata imposes
2060 * restraints of the various arrays in the container.
2061 * Currently we only know that IMSM requires all arrays
2062 * to have the same number of devices so changing the
2063 * number of devices (On-Line Capacity Expansion) must be
2064 * performed at the level of the container
2066 rv
= reshape_container(container
, devname
, -1, st
, &info
,
2067 c
->force
, c
->backup_file
, c
->verbose
, 0, 0, 0);
2070 /* get spare devices from external metadata
2072 if (st
->ss
->external
) {
2073 struct mdinfo
*info2
;
2075 info2
= st
->ss
->container_content(st
, subarray
);
2077 info
.array
.spare_disks
=
2078 info2
->array
.spare_disks
;
2083 /* Impose these changes on a single array. First
2084 * check that the metadata is OK with the change. */
2086 if (reshape_super(st
, 0, info
.new_level
,
2087 info
.new_layout
, info
.new_chunk
,
2088 info
.array
.raid_disks
, info
.delta_disks
,
2089 c
->backup_file
, devname
, APPLY_METADATA_CHANGES
,
2095 rv
= reshape_array(container
, fd
, devname
, st
, &info
, c
->force
,
2096 devlist
, data_offset
, c
->backup_file
, c
->verbose
,
2107 /* verify_reshape_position()
2108 * Function checks if reshape position in metadata is not farther
2109 * than position in md.
2111 * 0 : not valid sysfs entry
2112 * it can be caused by not started reshape, it should be started
2113 * by reshape array or raid0 array is before takeover
2114 * -1 : error, reshape position is obviously wrong
2115 * 1 : success, reshape progress correct or updated
2117 static int verify_reshape_position(struct mdinfo
*info
, int level
)
2123 /* read sync_max, failure can mean raid0 array */
2124 rv
= sysfs_get_str(info
, NULL
, "sync_max", buf
, 40);
2128 unsigned long long position
= strtoull(buf
, &ep
, 0);
2130 dprintf(Name
": Read sync_max sysfs entry is: %s\n", buf
);
2131 if (!(ep
== buf
|| (*ep
!= 0 && *ep
!= '\n' && *ep
!= ' '))) {
2132 position
*= get_data_disks(level
,
2134 info
->array
.raid_disks
);
2135 if (info
->reshape_progress
< position
) {
2136 dprintf("Corrected reshape progress (%llu) to "
2137 "md position (%llu)\n",
2138 info
->reshape_progress
, position
);
2139 info
->reshape_progress
= position
;
2141 } else if (info
->reshape_progress
> position
) {
2142 pr_err("Fatal error: array "
2143 "reshape was not properly frozen "
2144 "(expected reshape position is %llu, "
2145 "but reshape progress is %llu.\n",
2146 position
, info
->reshape_progress
);
2149 dprintf("Reshape position in md and metadata "
2154 } else if (rv
== 0) {
2155 /* for valid sysfs entry, 0-length content
2156 * should be indicated as error
2164 static unsigned long long choose_offset(unsigned long long lo
,
2165 unsigned long long hi
,
2166 unsigned long long min
,
2167 unsigned long long max
)
2169 /* Choose a new offset between hi and lo.
2170 * It must be between min and max, but
2171 * we would prefer something near the middle of hi/lo, and also
2172 * prefer to be aligned to a big power of 2.
2174 * So we start with the middle, then for each bit,
2175 * starting at '1' and increasing, if it is set, we either
2176 * add it or subtract it if possible, preferring the option
2177 * which is furthest from the boundary.
2179 * We stop once we get a 1MB alignment. As units are in sectors,
2180 * 1MB = 2*1024 sectors.
2182 unsigned long long choice
= (lo
+ hi
) / 2;
2183 unsigned long long bit
= 1;
2185 for (bit
= 1; bit
< 2*1024; bit
= bit
<< 1) {
2186 unsigned long long bigger
, smaller
;
2187 if (! (bit
& choice
))
2189 bigger
= choice
+ bit
;
2190 smaller
= choice
- bit
;
2191 if (bigger
> max
&& smaller
< min
)
2195 else if (smaller
< min
)
2197 else if (hi
- bigger
> smaller
- lo
)
2205 static int set_new_data_offset(struct mdinfo
*sra
, struct supertype
*st
,
2206 char *devname
, int delta_disks
,
2207 unsigned long long data_offset
,
2208 unsigned long long min
,
2214 unsigned long long before
, after
;
2216 /* Need to find min space before and after so same is used
2219 before
= UINT64_MAX
;
2221 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
2225 struct supertype
*st2
;
2226 struct mdinfo info2
;
2228 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
2230 dn
= map_dev(sd
->disk
.major
, sd
->disk
.minor
, 0);
2231 dfd
= dev_open(dn
, O_RDONLY
);
2233 pr_err("%s: cannot open component %s\n",
2234 devname
, dn
? dn
: "-unknown-");
2237 st2
= dup_super(st
);
2238 rv
= st2
->ss
->load_super(st2
,dfd
, NULL
);
2242 pr_err("%s: cannot get superblock from %s\n",
2246 st2
->ss
->getinfo_super(st2
, &info2
, NULL
);
2247 st2
->ss
->free_super(st2
);
2249 if (info2
.space_before
== 0 &&
2250 info2
.space_after
== 0) {
2251 /* Metadata doesn't support data_offset changes */
2254 if (before
> info2
.space_before
)
2255 before
= info2
.space_before
;
2256 if (after
> info2
.space_after
)
2257 after
= info2
.space_after
;
2259 if (data_offset
!= INVALID_SECTORS
) {
2261 if (info2
.data_offset
== data_offset
) {
2262 pr_err("%s: already has that data_offset\n",
2266 if (data_offset
< info2
.data_offset
)
2270 } else if ((data_offset
<= info2
.data_offset
&& dir
== 1) ||
2271 (data_offset
>= info2
.data_offset
&& dir
== -1)) {
2272 pr_err("%s: differing data offsets on devices make this --data-offset setting impossible\n",
2278 if (before
== UINT64_MAX
)
2279 /* impossible really, there must be no devices */
2282 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
2283 char *dn
= map_dev(sd
->disk
.major
, sd
->disk
.minor
, 0);
2284 unsigned long long new_data_offset
;
2286 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
2288 if (delta_disks
< 0) {
2289 /* Don't need any space as array is shrinking
2290 * just move data_offset up by min
2292 if (data_offset
== INVALID_SECTORS
)
2293 new_data_offset
= sd
->data_offset
+ min
;
2295 if (data_offset
< sd
->data_offset
+ min
) {
2296 pr_err("--data-offset too small for %s\n",
2300 new_data_offset
= data_offset
;
2302 } else if (delta_disks
> 0) {
2303 /* need space before */
2307 pr_err("Insufficient head-space for reshape on %s\n",
2311 if (data_offset
== INVALID_SECTORS
)
2312 new_data_offset
= sd
->data_offset
- min
;
2314 if (data_offset
> sd
->data_offset
- min
) {
2315 pr_err("--data-offset too large for %s\n",
2319 new_data_offset
= data_offset
;
2323 /* can move up or down. If 'data_offset'
2324 * was set we would have already decided,
2325 * so just choose direction with most space.
2332 sysfs_set_str(sra
, NULL
, "reshape_direction",
2333 dir
== 1 ? "backwards" : "forwards");
2335 /* Increase data offset */
2339 pr_err("Insufficient tail-space for reshape on %s\n",
2343 if (data_offset
!= INVALID_SECTORS
&&
2344 data_offset
< sd
->data_offset
+ min
) {
2345 pr_err("--data-offset too small on %s\n",
2349 if (data_offset
!= INVALID_SECTORS
)
2350 new_data_offset
= data_offset
;
2352 new_data_offset
= choose_offset(sd
->data_offset
,
2353 sd
->data_offset
+ after
,
2354 sd
->data_offset
+ min
,
2355 sd
->data_offset
+ after
);
2357 /* Decrease data offset */
2361 pr_err("insufficient head-room on %s\n",
2365 if (data_offset
!= INVALID_SECTORS
&&
2366 data_offset
< sd
->data_offset
- min
) {
2367 pr_err("--data-offset too small on %s\n",
2371 if (data_offset
!= INVALID_SECTORS
)
2372 new_data_offset
= data_offset
;
2374 new_data_offset
= choose_offset(sd
->data_offset
- before
,
2376 sd
->data_offset
- before
,
2377 sd
->data_offset
- min
);
2380 err
= sysfs_set_num(sra
, sd
, "new_offset", new_data_offset
);
2381 if (err
< 0 && errno
== E2BIG
) {
2382 /* try again after increasing data size to max */
2383 err
= sysfs_set_num(sra
, sd
, "size", 0);
2384 if (err
< 0 && errno
== EINVAL
&&
2385 !(sd
->disk
.state
& (1<<MD_DISK_SYNC
))) {
2386 /* some kernels have a bug where you cannot
2387 * use '0' on spare devices. */
2388 sysfs_set_num(sra
, sd
, "size",
2389 (sra
->component_size
+ after
)/2);
2391 err
= sysfs_set_num(sra
, sd
, "new_offset",
2395 if (errno
== E2BIG
&& data_offset
!= INVALID_SECTORS
) {
2396 pr_err("data-offset is too big for %s\n",
2400 if (sd
== sra
->devs
&&
2401 (errno
== ENOENT
|| errno
== E2BIG
))
2402 /* Early kernel, no 'new_offset' file,
2403 * or kernel doesn't like us.
2404 * For RAID5/6 this is not fatal
2407 pr_err("Cannot set new_offset for %s\n",
2416 /* Just use a backup file */
2420 static int raid10_reshape(char *container
, int fd
, char *devname
,
2421 struct supertype
*st
, struct mdinfo
*info
,
2422 struct reshape
*reshape
,
2423 unsigned long long data_offset
,
2424 int force
, int verbose
)
2426 /* Changing raid_disks, layout, chunksize or possibly
2427 * just data_offset for a RAID10.
2428 * We must always change data_offset. We change by at least
2429 * ->min_offset_change which is the largest of the old and new
2431 * If raid_disks is increasing, then data_offset must decrease
2432 * by at least this copy size.
2433 * If raid_disks is unchanged, data_offset must increase or
2434 * decrease by at least min_offset_change but preferably by much more.
2435 * We choose half of the available space.
2436 * If raid_disks is decreasing, data_offset must increase by
2437 * at least min_offset_change. To allow of this, component_size
2438 * must be decreased by the same amount.
2440 * So we calculate the required minimum and direction, possibly
2441 * reduce the component_size, then iterate through the devices
2442 * and set the new_data_offset.
2443 * If that all works, we set chunk_size, layout, raid_disks, and start
2447 unsigned long long min
;
2450 sra
= sysfs_read(fd
, NULL
,
2451 GET_COMPONENT
|GET_DEVS
|GET_OFFSET
|GET_STATE
|GET_CHUNK
2454 pr_err("%s: Cannot get array details from sysfs\n",
2458 min
= reshape
->min_offset_change
;
2460 if (info
->delta_disks
)
2461 sysfs_set_str(sra
, NULL
, "reshape_direction",
2462 info
->delta_disks
< 0 ? "backwards" : "forwards");
2463 if (info
->delta_disks
< 0 &&
2464 info
->space_after
< min
) {
2465 int rv
= sysfs_set_num(sra
, NULL
, "component_size",
2466 (sra
->component_size
-
2469 pr_err("cannot reduce component size\n");
2473 err
= set_new_data_offset(sra
, st
, devname
, info
->delta_disks
, data_offset
,
2476 pr_err("Cannot set new_data_offset: RAID10 reshape not\n");
2477 cont_err("supported on this kernel\n");
2483 if (!err
&& sysfs_set_num(sra
, NULL
, "chunk_size", info
->new_chunk
) < 0)
2485 if (!err
&& sysfs_set_num(sra
, NULL
, "layout", reshape
->after
.layout
) < 0)
2487 if (!err
&& sysfs_set_num(sra
, NULL
, "raid_disks",
2488 info
->array
.raid_disks
+ info
->delta_disks
) < 0)
2490 if (!err
&& sysfs_set_str(sra
, NULL
, "sync_action", "reshape") < 0)
2493 pr_err("Cannot set array shape for %s\n",
2496 (info
->array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
2497 cont_err(" Bitmap must be removed before"
2498 " shape can be changed\n");
2508 static void get_space_after(int fd
, struct supertype
*st
, struct mdinfo
*info
)
2510 struct mdinfo
*sra
, *sd
;
2511 /* Initialisation to silence compiler warning */
2512 unsigned long long min_space_before
= 0, min_space_after
= 0;
2515 sra
= sysfs_read(fd
, NULL
, GET_DEVS
);
2518 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
2521 struct supertype
*st2
;
2522 struct mdinfo info2
;
2524 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
2526 dn
= map_dev(sd
->disk
.major
, sd
->disk
.minor
, 0);
2527 dfd
= dev_open(dn
, O_RDONLY
);
2530 st2
= dup_super(st
);
2531 if (st2
->ss
->load_super(st2
,dfd
, NULL
)) {
2537 st2
->ss
->getinfo_super(st2
, &info2
, NULL
);
2538 st2
->ss
->free_super(st2
);
2541 min_space_before
> info2
.space_before
)
2542 min_space_before
= info2
.space_before
;
2544 min_space_after
> info2
.space_after
)
2545 min_space_after
= info2
.space_after
;
2548 if (sd
== NULL
&& !first
) {
2549 info
->space_after
= min_space_after
;
2550 info
->space_before
= min_space_before
;
2555 static void update_cache_size(char *container
, struct mdinfo
*sra
,
2556 struct mdinfo
*info
,
2557 int disks
, unsigned long long blocks
)
2559 /* Check that the internal stripe cache is
2560 * large enough, or it won't work.
2561 * It must hold at least 4 stripes of the larger
2564 unsigned long cache
;
2565 cache
= max(info
->array
.chunk_size
, info
->new_chunk
);
2566 cache
*= 4; /* 4 stripes minimum */
2567 cache
/= 512; /* convert to sectors */
2568 /* make sure there is room for 'blocks' with a bit to spare */
2569 if (cache
< 16 + blocks
/ disks
)
2570 cache
= 16 + blocks
/ disks
;
2571 cache
/= (4096/512); /* Covert from sectors to pages */
2573 if (sra
->cache_size
< cache
)
2574 subarray_set_num(container
, sra
, "stripe_cache_size",
2578 static int impose_reshape(struct mdinfo
*sra
,
2579 struct mdinfo
*info
,
2580 struct supertype
*st
,
2583 char *devname
, char *container
,
2584 struct reshape
*reshape
)
2586 struct mdu_array_info_s array
;
2588 sra
->new_chunk
= info
->new_chunk
;
2591 /* for external metadata checkpoint saved by mdmon can be lost
2592 * or missed /due to e.g. crash/. Check if md is not during
2593 * restart farther than metadata points to.
2594 * If so, this means metadata information is obsolete.
2596 if (st
->ss
->external
)
2597 verify_reshape_position(info
, reshape
->level
);
2598 sra
->reshape_progress
= info
->reshape_progress
;
2600 sra
->reshape_progress
= 0;
2601 if (reshape
->after
.data_disks
< reshape
->before
.data_disks
)
2602 /* start from the end of the new array */
2603 sra
->reshape_progress
= (sra
->component_size
2604 * reshape
->after
.data_disks
);
2607 ioctl(fd
, GET_ARRAY_INFO
, &array
);
2608 if (info
->array
.chunk_size
== info
->new_chunk
&&
2609 reshape
->before
.layout
== reshape
->after
.layout
&&
2610 st
->ss
->external
== 0) {
2611 /* use SET_ARRAY_INFO but only if reshape hasn't started */
2612 array
.raid_disks
= reshape
->after
.data_disks
+ reshape
->parity
;
2614 ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
2617 pr_err("Cannot set device shape for %s: %s\n",
2618 devname
, strerror(errno
));
2621 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
2622 cont_err("Bitmap must be removed before"
2623 " shape can be changed\n");
2627 } else if (!restart
) {
2628 /* set them all just in case some old 'new_*' value
2629 * persists from some earlier problem.
2632 if (sysfs_set_num(sra
, NULL
, "chunk_size", info
->new_chunk
) < 0)
2634 if (!err
&& sysfs_set_num(sra
, NULL
, "layout",
2635 reshape
->after
.layout
) < 0)
2637 if (!err
&& subarray_set_num(container
, sra
, "raid_disks",
2638 reshape
->after
.data_disks
+
2639 reshape
->parity
) < 0)
2642 pr_err("Cannot set device shape for %s\n",
2646 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
2647 cont_err("Bitmap must be removed before"
2648 " shape can be changed\n");
2657 static int impose_level(int fd
, int level
, char *devname
, int verbose
)
2660 struct mdu_array_info_s array
;
2662 sysfs_init(&info
, fd
, NULL
);
2664 ioctl(fd
, GET_ARRAY_INFO
, &array
);
2666 (array
.level
>= 4 && array
.level
<= 6)) {
2667 /* To convert to RAID0 we need to fail and
2668 * remove any non-data devices. */
2671 int data_disks
= array
.raid_disks
- 1;
2672 if (array
.level
== 6)
2674 if (array
.level
== 5 &&
2675 array
.layout
!= ALGORITHM_PARITY_N
)
2677 if (array
.level
== 6 &&
2678 array
.layout
!= ALGORITHM_PARITY_N_6
)
2680 sysfs_set_str(&info
, NULL
,"sync_action", "idle");
2681 /* First remove any spares so no recovery starts */
2682 for (d
= 0, found
= 0;
2683 d
< MAX_DISKS
&& found
< array
.nr_disks
;
2685 mdu_disk_info_t disk
;
2687 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0)
2689 if (disk
.major
== 0 && disk
.minor
== 0)
2692 if ((disk
.state
& (1 << MD_DISK_ACTIVE
))
2693 && disk
.raid_disk
< data_disks
)
2696 ioctl(fd
, HOT_REMOVE_DISK
,
2697 makedev(disk
.major
, disk
.minor
));
2699 /* Now fail anything left */
2700 ioctl(fd
, GET_ARRAY_INFO
, &array
);
2701 for (d
= 0, found
= 0;
2702 d
< MAX_DISKS
&& found
< array
.nr_disks
;
2705 mdu_disk_info_t disk
;
2707 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0)
2709 if (disk
.major
== 0 && disk
.minor
== 0)
2712 if ((disk
.state
& (1 << MD_DISK_ACTIVE
))
2713 && disk
.raid_disk
< data_disks
)
2716 ioctl(fd
, SET_DISK_FAULTY
,
2717 makedev(disk
.major
, disk
.minor
));
2719 while (ioctl(fd
, HOT_REMOVE_DISK
,
2720 makedev(disk
.major
, disk
.minor
)) < 0
2727 c
= map_num(pers
, level
);
2729 int err
= sysfs_set_str(&info
, NULL
, "level", c
);
2732 pr_err("%s: could not set level to %s\n",
2735 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
2736 cont_err("Bitmap must be removed"
2737 " before level can be changed\n");
2741 pr_err("level of %s changed to %s\n",
2748 static void catch_term(int sig
)
2753 static int continue_via_systemd(char *devnm
)
2755 int skipped
, i
, pid
, status
;
2757 /* In a systemd/udev world, it is best to get systemd to
2758 * run "mdadm --grow --continue" rather than running in the
2763 /* FIXME yuk. CLOSE_EXEC?? */
2765 for (i
= 3; skipped
< 20; i
++)
2771 /* Don't want to see error messages from
2772 * systemctl. If the service doesn't exist,
2773 * we fork ourselves.
2776 open("/dev/null", O_WRONLY
);
2777 snprintf(pathbuf
, sizeof(pathbuf
), "mdadm-grow-continue@%s.service",
2779 status
= execl("/usr/bin/systemctl", "systemctl",
2782 status
= execl("/bin/systemctl", "systemctl", "start",
2785 case -1: /* Just do it ourselves. */
2787 default: /* parent - good */
2788 pid
= wait(&status
);
2789 if (pid
>= 0 && status
== 0)
2795 static int reshape_array(char *container
, int fd
, char *devname
,
2796 struct supertype
*st
, struct mdinfo
*info
,
2797 int force
, struct mddev_dev
*devlist
,
2798 unsigned long long data_offset
,
2799 char *backup_file
, int verbose
, int forked
,
2800 int restart
, int freeze_reshape
)
2802 struct reshape reshape
;
2805 int orig_level
= UnSet
;
2809 struct mdu_array_info_s array
;
2812 struct mddev_dev
*dv
;
2816 unsigned long long *offsets
= NULL
;
2820 unsigned long blocks
;
2821 unsigned long long array_size
;
2823 struct mdinfo
*sra
= NULL
;
2825 /* when reshaping a RAID0, the component_size might be zero.
2826 * So try to fix that up.
2828 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
2829 dprintf("Cannot get array information.\n");
2832 if (array
.level
== 0 && info
->component_size
== 0) {
2833 get_dev_size(fd
, NULL
, &array_size
);
2834 info
->component_size
= array_size
/ array
.raid_disks
;
2837 if (array
.level
== 10)
2838 /* Need space_after info */
2839 get_space_after(fd
, st
, info
);
2841 if (info
->reshape_active
) {
2842 int new_level
= info
->new_level
;
2843 info
->new_level
= UnSet
;
2844 if (info
->delta_disks
> 0)
2845 info
->array
.raid_disks
-= info
->delta_disks
;
2846 msg
= analyse_change(devname
, info
, &reshape
);
2847 info
->new_level
= new_level
;
2848 if (info
->delta_disks
> 0)
2849 info
->array
.raid_disks
+= info
->delta_disks
;
2851 /* Make sure the array isn't read-only */
2852 ioctl(fd
, RESTART_ARRAY_RW
, 0);
2854 msg
= analyse_change(devname
, info
, &reshape
);
2856 /* if msg == "", error has already been printed */
2858 pr_err("%s\n", msg
);
2862 (reshape
.level
!= info
->array
.level
||
2863 reshape
.before
.layout
!= info
->array
.layout
||
2864 reshape
.before
.data_disks
+ reshape
.parity
2865 != info
->array
.raid_disks
- max(0, info
->delta_disks
))) {
2866 pr_err("reshape info is not in native format -"
2867 " cannot continue.\n");
2871 if (st
->ss
->external
&& restart
&& (info
->reshape_progress
== 0)) {
2872 /* When reshape is restarted from '0', very begin of array
2873 * it is possible that for external metadata reshape and array
2874 * configuration doesn't happen.
2875 * Check if md has the same opinion, and reshape is restarted
2876 * from 0. If so, this is regular reshape start after reshape
2877 * switch in metadata to next array only.
2879 if ((verify_reshape_position(info
, reshape
.level
) >= 0) &&
2880 (info
->reshape_progress
== 0))
2884 /* reshape already started. just skip to monitoring the reshape */
2885 if (reshape
.backup_blocks
== 0)
2887 if (restart
& RESHAPE_NO_BACKUP
)
2890 /* Need 'sra' down at 'started:' */
2891 sra
= sysfs_read(fd
, NULL
,
2892 GET_COMPONENT
|GET_DEVS
|GET_OFFSET
|GET_STATE
|GET_CHUNK
|
2895 pr_err("%s: Cannot get array details from sysfs\n",
2901 backup_file
= locate_backup(sra
->sys_name
);
2905 /* The container is frozen but the array may not be.
2906 * So freeze the array so spares don't get put to the wrong use
2907 * FIXME there should probably be a cleaner separation between
2908 * freeze_array and freeze_container.
2910 sysfs_freeze_array(info
);
2911 /* Check we have enough spares to not be degraded */
2913 for (dv
= devlist
; dv
; dv
=dv
->next
)
2915 spares_needed
= max(reshape
.before
.data_disks
,
2916 reshape
.after
.data_disks
)
2917 + reshape
.parity
- array
.raid_disks
;
2920 info
->new_level
> 1 && info
->array
.level
> 1 &&
2921 spares_needed
> info
->array
.spare_disks
+ added_disks
) {
2922 pr_err("Need %d spare%s to avoid degraded array,"
2923 " and only have %d.\n"
2924 " Use --force to over-ride this check.\n",
2926 spares_needed
== 1 ? "" : "s",
2927 info
->array
.spare_disks
+ added_disks
);
2930 /* Check we have enough spares to not fail */
2931 spares_needed
= max(reshape
.before
.data_disks
,
2932 reshape
.after
.data_disks
)
2934 if ((info
->new_level
> 1 || info
->new_level
== 0) &&
2935 spares_needed
> info
->array
.spare_disks
+added_disks
) {
2936 pr_err("Need %d spare%s to create working array,"
2937 " and only have %d.\n",
2939 spares_needed
== 1 ? "" : "s",
2940 info
->array
.spare_disks
+ added_disks
);
2944 if (reshape
.level
!= array
.level
) {
2945 int err
= impose_level(fd
, reshape
.level
, devname
, verbose
);
2948 info
->new_layout
= UnSet
; /* after level change,
2949 * layout is meaningless */
2950 orig_level
= array
.level
;
2951 sysfs_freeze_array(info
);
2953 if (reshape
.level
> 0 && st
->ss
->external
) {
2954 /* make sure mdmon is aware of the new level */
2955 if (mdmon_running(container
))
2956 flush_mdmon(container
);
2958 if (!mdmon_running(container
))
2959 start_mdmon(container
);
2960 ping_monitor(container
);
2961 if (mdmon_running(container
) &&
2962 st
->update_tail
== NULL
)
2963 st
->update_tail
= &st
->updates
;
2966 /* ->reshape_super might have chosen some spares from the
2967 * container that it wants to be part of the new array.
2968 * We can collect them with ->container_content and give
2969 * them to the kernel.
2971 if (st
->ss
->reshape_super
&& st
->ss
->container_content
) {
2972 char *subarray
= strchr(info
->text_version
+1, '/')+1;
2973 struct mdinfo
*info2
=
2974 st
->ss
->container_content(st
, subarray
);
2978 sysfs_init(info2
, fd
, st
->devnm
);
2979 /* When increasing number of devices, we need to set
2980 * new raid_disks before adding these, or they might
2983 if (reshape
.backup_blocks
&&
2984 reshape
.after
.data_disks
> reshape
.before
.data_disks
)
2985 subarray_set_num(container
, info2
, "raid_disks",
2986 reshape
.after
.data_disks
+
2988 for (d
= info2
->devs
; d
; d
= d
->next
) {
2989 if (d
->disk
.state
== 0 &&
2990 d
->disk
.raid_disk
>= 0) {
2991 /* This is a spare that wants to
2992 * be part of the array.
2994 add_disk(fd
, st
, info2
, d
);
3000 /* We might have been given some devices to add to the
3001 * array. Now that the array has been changed to the right
3002 * level and frozen, we can safely add them.
3005 Manage_subdevs(devname
, fd
, devlist
, verbose
,
3008 if (reshape
.backup_blocks
== 0 && data_offset
!= INVALID_SECTORS
)
3009 reshape
.backup_blocks
= reshape
.before
.data_disks
* info
->array
.chunk_size
/512;
3010 if (reshape
.backup_blocks
== 0) {
3011 /* No restriping needed, but we might need to impose
3012 * some more changes: layout, raid_disks, chunk_size
3014 /* read current array info */
3015 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
3016 dprintf("Cannot get array information.\n");
3019 /* compare current array info with new values and if
3020 * it is different update them to new */
3021 if (info
->new_layout
!= UnSet
&&
3022 info
->new_layout
!= array
.layout
) {
3023 array
.layout
= info
->new_layout
;
3024 if (ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
3025 pr_err("failed to set new layout\n");
3027 } else if (verbose
>= 0)
3028 printf("layout for %s set to %d\n",
3029 devname
, array
.layout
);
3031 if (info
->delta_disks
!= UnSet
&&
3032 info
->delta_disks
!= 0 &&
3033 array
.raid_disks
!= (info
->array
.raid_disks
+ info
->delta_disks
)) {
3034 array
.raid_disks
+= info
->delta_disks
;
3035 if (ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
3036 pr_err("failed to set raid disks\n");
3038 } else if (verbose
>= 0) {
3039 printf("raid_disks for %s set to %d\n",
3040 devname
, array
.raid_disks
);
3043 if (info
->new_chunk
!= 0 &&
3044 info
->new_chunk
!= array
.chunk_size
) {
3045 if (sysfs_set_num(info
, NULL
,
3046 "chunk_size", info
->new_chunk
) != 0) {
3047 pr_err("failed to set chunk size\n");
3049 } else if (verbose
>= 0)
3050 printf("chunk size for %s set to %d\n",
3051 devname
, array
.chunk_size
);
3058 * There are three possibilities.
3059 * 1/ The array will shrink.
3060 * We need to ensure the reshape will pause before reaching
3061 * the 'critical section'. We also need to fork and wait for
3062 * that to happen. When it does we
3063 * suspend/backup/complete/unfreeze
3065 * 2/ The array will not change size.
3066 * This requires that we keep a backup of a sliding window
3067 * so that we can restore data after a crash. So we need
3068 * to fork and monitor progress.
3069 * In future we will allow the data_offset to change, so
3070 * a sliding backup becomes unnecessary.
3072 * 3/ The array will grow. This is relatively easy.
3073 * However the kernel's restripe routines will cheerfully
3074 * overwrite some early data before it is safe. So we
3075 * need to make a backup of the early parts of the array
3076 * and be ready to restore it if rebuild aborts very early.
3077 * For externally managed metadata, we still need a forked
3078 * child to monitor the reshape and suspend IO over the region
3079 * that is being reshaped.
3081 * We backup data by writing it to one spare, or to a
3082 * file which was given on command line.
3084 * In each case, we first make sure that storage is available
3085 * for the required backup.
3087 * - request the shape change.
3088 * - fork to handle backup etc.
3090 /* Check that we can hold all the data */
3091 get_dev_size(fd
, NULL
, &array_size
);
3092 if (reshape
.new_size
< (array_size
/512)) {
3093 pr_err("this change will reduce the size of the array.\n"
3094 " use --grow --array-size first to truncate array.\n"
3095 " e.g. mdadm --grow %s --array-size %llu\n",
3096 devname
, reshape
.new_size
/2);
3100 if (array
.level
== 10) {
3101 /* Reshaping RAID10 does not require any data backup by
3102 * user-space. Instead it requires that the data_offset
3103 * is changed to avoid the need for backup.
3104 * So this is handled very separately
3107 /* Nothing to do. */
3109 return raid10_reshape(container
, fd
, devname
, st
, info
,
3110 &reshape
, data_offset
,
3113 sra
= sysfs_read(fd
, NULL
,
3114 GET_COMPONENT
|GET_DEVS
|GET_OFFSET
|GET_STATE
|GET_CHUNK
|
3117 pr_err("%s: Cannot get array details from sysfs\n",
3123 switch(set_new_data_offset(sra
, st
, devname
,
3124 reshape
.after
.data_disks
- reshape
.before
.data_disks
,
3126 reshape
.min_offset_change
, 1)) {
3130 /* Updated data_offset, so it's easy now */
3131 update_cache_size(container
, sra
, info
,
3132 min(reshape
.before
.data_disks
,
3133 reshape
.after
.data_disks
),
3134 reshape
.backup_blocks
);
3136 /* Right, everything seems fine. Let's kick things off.
3140 if (impose_reshape(sra
, info
, st
, fd
, restart
,
3141 devname
, container
, &reshape
) < 0)
3143 if (sysfs_set_str(sra
, NULL
, "sync_action", "reshape") < 0) {
3144 pr_err("Failed to initiate reshape!\n");
3147 if (info
->new_level
== reshape
.level
)
3149 /* need to adjust level when reshape completes */
3151 case -1: /* ignore error, but don't wait */
3153 default: /* parent */
3161 fd
= open_dev(sra
->sys_name
);
3163 impose_level(fd
, info
->new_level
, devname
, verbose
);
3165 case 1: /* Couldn't set data_offset, try the old way */
3166 if (data_offset
!= INVALID_SECTORS
) {
3167 pr_err("Cannot update data_offset on this array\n");
3174 /* Decide how many blocks (sectors) for a reshape
3175 * unit. The number we have so far is just a minimum
3177 blocks
= reshape
.backup_blocks
;
3178 if (reshape
.before
.data_disks
==
3179 reshape
.after
.data_disks
) {
3180 /* Make 'blocks' bigger for better throughput, but
3181 * not so big that we reject it below.
3182 * Try for 16 megabytes
3184 while (blocks
* 32 < sra
->component_size
&&
3188 pr_err("Need to backup %luK of critical "
3189 "section..\n", blocks
/2);
3191 if (blocks
>= sra
->component_size
/2) {
3192 pr_err("%s: Something wrong"
3193 " - reshape aborted\n",
3198 /* Now we need to open all these devices so we can read/write.
3200 nrdisks
= max(reshape
.before
.data_disks
,
3201 reshape
.after
.data_disks
) + reshape
.parity
3202 + sra
->array
.spare_disks
;
3203 fdlist
= xcalloc((1+nrdisks
), sizeof(int));
3204 offsets
= xcalloc((1+nrdisks
), sizeof(offsets
[0]));
3206 odisks
= reshape
.before
.data_disks
+ reshape
.parity
;
3207 d
= reshape_prepare_fdlist(devname
, sra
, odisks
,
3208 nrdisks
, blocks
, backup_file
,
3213 if ((st
->ss
->manage_reshape
== NULL
) ||
3214 (st
->ss
->recover_backup
== NULL
)) {
3215 if (backup_file
== NULL
) {
3216 if (reshape
.after
.data_disks
<=
3217 reshape
.before
.data_disks
) {
3218 pr_err("%s: Cannot grow - need backup-file\n",
3220 pr_err(" Please provide one with \"--backup=...\"\n");
3222 } else if (sra
->array
.spare_disks
== 0) {
3223 pr_err("%s: Cannot grow - "
3224 "need a spare or backup-file to backup "
3225 "critical section\n", devname
);
3229 if (!reshape_open_backup_file(backup_file
, fd
, devname
,
3231 fdlist
+d
, offsets
+d
,
3240 update_cache_size(container
, sra
, info
,
3241 min(reshape
.before
.data_disks
, reshape
.after
.data_disks
),
3244 /* Right, everything seems fine. Let's kick things off.
3245 * If only changing raid_disks, use ioctl, else use
3250 if (impose_reshape(sra
, info
, st
, fd
, restart
,
3251 devname
, container
, &reshape
) < 0)
3254 err
= start_reshape(sra
, restart
, reshape
.before
.data_disks
,
3255 reshape
.after
.data_disks
);
3257 pr_err("Cannot %s reshape for %s\n",
3258 restart
? "continue" : "start",
3263 sysfs_set_str(sra
, NULL
, "array_state", "active");
3264 if (freeze_reshape
) {
3268 pr_err("Reshape has to be continued from"
3269 " location %llu when root filesystem has been mounted.\n",
3270 sra
->reshape_progress
);
3274 if (!forked
&& !check_env("MDADM_NO_SYSTEMCTL"))
3275 if (continue_via_systemd(container
?: sra
->sys_name
)) {
3282 /* Now we just need to kick off the reshape and watch, while
3283 * handling backups of the data...
3284 * This is all done by a forked background process.
3286 switch(forked
? 0 : fork()) {
3288 pr_err("Cannot run child to monitor reshape: %s\n",
3302 /* If another array on the same devices is busy, the
3303 * reshape will wait for them. This would mean that
3304 * the first section that we suspend will stay suspended
3305 * for a long time. So check on that possibility
3306 * by looking for "DELAYED" in /proc/mdstat, and if found,
3310 struct mdstat_ent
*mds
, *m
;
3312 mds
= mdstat_read(1, 0);
3313 for (m
= mds
; m
; m
= m
->next
)
3314 if (strcmp(m
->devnm
, sra
->sys_name
) == 0) {
3316 m
->percent
== RESYNC_DELAYED
)
3319 /* Haven't started the reshape thread
3326 if (delayed
== 1 && get_linux_version() < 3007000) {
3327 pr_err("Reshape is delayed, but cannot wait carefully with this kernel.\n"
3328 " You might experience problems until other reshapes complete.\n");
3332 mdstat_wait(30 - (delayed
-1) * 25);
3336 if (check_env("MDADM_GROW_VERIFY"))
3337 fd
= open(devname
, O_RDONLY
| O_DIRECT
);
3340 mlockall(MCL_FUTURE
);
3342 signal(SIGTERM
, catch_term
);
3344 if (st
->ss
->external
) {
3345 /* metadata handler takes it from here */
3346 done
= st
->ss
->manage_reshape(
3347 fd
, sra
, &reshape
, st
, blocks
,
3349 d
- odisks
, fdlist
+odisks
,
3352 done
= child_monitor(
3353 fd
, sra
, &reshape
, st
, blocks
,
3355 d
- odisks
, fdlist
+odisks
,
3361 if (backup_file
&& done
) {
3363 bul
= make_backup(sra
->sys_name
);
3366 int l
= readlink(bul
, buf
, sizeof(buf
));
3374 unlink(backup_file
);
3381 if (!st
->ss
->external
&&
3382 !(reshape
.before
.data_disks
!= reshape
.after
.data_disks
3383 && info
->custom_array_size
) &&
3384 info
->new_level
== reshape
.level
&&
3386 /* no need to wait for the reshape to finish as
3387 * there is nothing more to do.
3394 if (st
->ss
->external
) {
3395 /* Re-load the metadata as much could have changed */
3396 int cfd
= open_dev(st
->container_devnm
);
3398 flush_mdmon(container
);
3399 st
->ss
->free_super(st
);
3400 st
->ss
->load_container(st
, cfd
, container
);
3405 /* set new array size if required customer_array_size is used
3408 if (reshape
.before
.data_disks
!=
3409 reshape
.after
.data_disks
&&
3410 info
->custom_array_size
)
3411 set_array_size(st
, info
, info
->text_version
);
3413 if (info
->new_level
!= reshape
.level
) {
3415 fd
= open(devname
, O_RDONLY
);
3416 impose_level(fd
, info
->new_level
, devname
, verbose
);
3418 if (info
->new_level
== 0)
3419 st
->update_tail
= NULL
;
3431 if (orig_level
!= UnSet
&& sra
) {
3432 c
= map_num(pers
, orig_level
);
3433 if (c
&& sysfs_set_str(sra
, NULL
, "level", c
) == 0)
3434 pr_err("aborting level change\n");
3442 /* mdfd handle is passed to be closed in child process (after fork).
3444 int reshape_container(char *container
, char *devname
,
3446 struct supertype
*st
,
3447 struct mdinfo
*info
,
3449 char *backup_file
, int verbose
,
3450 int forked
, int restart
, int freeze_reshape
)
3452 struct mdinfo
*cc
= NULL
;
3454 char last_devnm
[32] = "";
3456 /* component_size is not meaningful for a container,
3457 * so pass '0' meaning 'no change'
3460 reshape_super(st
, 0, info
->new_level
,
3461 info
->new_layout
, info
->new_chunk
,
3462 info
->array
.raid_disks
, info
->delta_disks
,
3463 backup_file
, devname
, APPLY_METADATA_CHANGES
,
3471 /* ping monitor to be sure that update is on disk
3473 ping_monitor(container
);
3475 if (!forked
&& !check_env("MDADM_NO_SYSTEMCTL"))
3476 if (continue_via_systemd(container
))
3479 switch (forked
? 0 : fork()) {
3480 case -1: /* error */
3481 perror("Cannot fork to complete reshape\n");
3484 default: /* parent */
3485 if (!freeze_reshape
)
3486 printf(Name
": multi-array reshape continues"
3487 " in background\n");
3494 /* close unused handle in child process
3500 /* For each member array with reshape_active,
3501 * we need to perform the reshape.
3502 * We pick the first array that needs reshaping and
3503 * reshape it. reshape_array() will re-read the metadata
3504 * so the next time through a different array should be
3505 * ready for reshape.
3506 * It is possible that the 'different' array will not
3507 * be assembled yet. In that case we simple exit.
3508 * When it is assembled, the mdadm which assembles it
3509 * will take over the reshape.
3511 struct mdinfo
*content
;
3513 struct mdstat_ent
*mdstat
;
3519 cc
= st
->ss
->container_content(st
, NULL
);
3521 for (content
= cc
; content
; content
= content
->next
) {
3523 if (!content
->reshape_active
)
3526 subarray
= strchr(content
->text_version
+1, '/')+1;
3527 mdstat
= mdstat_by_subdev(subarray
, container
);
3530 if (mdstat
->active
== 0) {
3531 pr_err("Skipping inactive array %s.\n",
3533 free_mdstat(mdstat
);
3542 devid
= devnm2devid(mdstat
->devnm
);
3543 adev
= map_dev(major(devid
), minor(devid
), 0);
3545 adev
= content
->text_version
;
3547 fd
= open_dev(mdstat
->devnm
);
3549 printf(Name
": Device %s cannot be opened for reshape.",
3554 if (strcmp(last_devnm
, mdstat
->devnm
) == 0) {
3555 /* Do not allow for multiple reshape_array() calls for
3557 * It can happen when reshape_array() returns without
3558 * error, when reshape is not finished (wrong reshape
3559 * starting/continuation conditions). Mdmon doesn't
3560 * switch to next array in container and reentry
3561 * conditions for the same array occur.
3562 * This is possibly interim until the behaviour of
3563 * reshape_array is resolved().
3565 printf(Name
": Multiple reshape execution detected for "
3566 "device %s.", adev
);
3570 strcpy(last_devnm
, mdstat
->devnm
);
3572 sysfs_init(content
, fd
, mdstat
->devnm
);
3574 if (mdmon_running(container
))
3575 flush_mdmon(container
);
3577 rv
= reshape_array(container
, fd
, adev
, st
,
3578 content
, force
, NULL
, INVALID_SECTORS
,
3579 backup_file
, verbose
, 1, restart
,
3583 if (freeze_reshape
) {
3592 if (mdmon_running(container
))
3593 flush_mdmon(container
);
3602 * We run a child process in the background which performs the following
3604 * - wait for resync to reach a certain point
3605 * - suspend io to the following section
3606 * - backup that section
3607 * - allow resync to proceed further
3609 * - discard the backup.
3611 * When are combined in slightly different ways in the three cases.
3613 * - suspend/backup/allow/wait/resume/discard
3615 * - allow/wait/suspend/backup/allow/wait/resume/discard
3617 * - wait/resume/discard/suspend/backup/allow
3619 * suspend/backup/allow always come together
3620 * wait/resume/discard do too.
3621 * For the same-size case we have two backups to improve flow.
3625 int progress_reshape(struct mdinfo
*info
, struct reshape
*reshape
,
3626 unsigned long long backup_point
,
3627 unsigned long long wait_point
,
3628 unsigned long long *suspend_point
,
3629 unsigned long long *reshape_completed
, int *frozen
)
3631 /* This function is called repeatedly by the reshape manager.
3632 * It determines how much progress can safely be made and allows
3634 * - 'info' identifies the array and particularly records in
3635 * ->reshape_progress the metadata's knowledge of progress
3636 * This is a sector offset from the start of the array
3637 * of the next array block to be relocated. This number
3638 * may increase from 0 or decrease from array_size, depending
3639 * on the type of reshape that is happening.
3640 * Note that in contrast, 'sync_completed' is a block count of the
3641 * reshape so far. It gives the distance between the start point
3642 * (head or tail of device) and the next place that data will be
3643 * written. It always increases.
3644 * - 'reshape' is the structure created by analyse_change
3645 * - 'backup_point' shows how much the metadata manager has backed-up
3646 * data. For reshapes with increasing progress, it is the next address
3647 * to be backed up, previous addresses have been backed-up. For
3648 * decreasing progress, it is the earliest address that has been
3649 * backed up - later address are also backed up.
3650 * So addresses between reshape_progress and backup_point are
3651 * backed up providing those are in the 'correct' order.
3652 * - 'wait_point' is an array address. When reshape_completed
3653 * passes this point, progress_reshape should return. It might
3654 * return earlier if it determines that ->reshape_progress needs
3655 * to be updated or further backup is needed.
3656 * - suspend_point is maintained by progress_reshape and the caller
3657 * should not touch it except to initialise to zero.
3658 * It is an array address and it only increases in 2.6.37 and earlier.
3659 * This makes it difficult to handle reducing reshapes with
3660 * external metadata.
3661 * However: it is similar to backup_point in that it records the
3662 * other end of a suspended region from reshape_progress.
3663 * it is moved to extend the region that is safe to backup and/or
3665 * - reshape_completed is read from sysfs and returned. The caller
3666 * should copy this into ->reshape_progress when it has reason to
3667 * believe that the metadata knows this, and any backup outside this
3671 * 1 if more data from backup_point - but only as far as suspend_point,
3672 * should be backed up
3673 * 0 if things are progressing smoothly
3674 * -1 if the reshape is finished because it is all done,
3675 * -2 if the reshape is finished due to an error.
3678 int advancing
= (reshape
->after
.data_disks
3679 >= reshape
->before
.data_disks
);
3680 unsigned long long need_backup
; /* All data between start of array and
3681 * here will at some point need to
3684 unsigned long long read_offset
, write_offset
;
3685 unsigned long long write_range
;
3686 unsigned long long max_progress
, target
, completed
;
3687 unsigned long long array_size
= (info
->component_size
3688 * reshape
->before
.data_disks
);
3692 /* First, we unsuspend any region that is now known to be safe.
3693 * If suspend_point is on the 'wrong' side of reshape_progress, then
3694 * we don't have or need suspension at the moment. This is true for
3695 * native metadata when we don't need to back-up.
3698 if (info
->reshape_progress
<= *suspend_point
)
3699 sysfs_set_num(info
, NULL
, "suspend_lo",
3700 info
->reshape_progress
);
3702 /* Note: this won't work in 2.6.37 and before.
3703 * Something somewhere should make sure we don't need it!
3705 if (info
->reshape_progress
>= *suspend_point
)
3706 sysfs_set_num(info
, NULL
, "suspend_hi",
3707 info
->reshape_progress
);
3710 /* Now work out how far it is safe to progress.
3711 * If the read_offset for ->reshape_progress is less than
3712 * 'blocks' beyond the write_offset, we can only progress as far
3714 * Otherwise we can progress until the write_offset for the new location
3715 * reaches (within 'blocks' of) the read_offset at the current location.
3716 * However that region must be suspended unless we are using native
3718 * If we need to suspend more, we limit it to 128M per device, which is
3719 * rather arbitrary and should be some time-based calculation.
3721 read_offset
= info
->reshape_progress
/ reshape
->before
.data_disks
;
3722 write_offset
= info
->reshape_progress
/ reshape
->after
.data_disks
;
3723 write_range
= info
->new_chunk
/512;
3724 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
)
3725 need_backup
= array_size
;
3727 need_backup
= reshape
->backup_blocks
;
3729 if (read_offset
< write_offset
+ write_range
)
3730 max_progress
= backup_point
;
3734 reshape
->after
.data_disks
;
3736 if (read_offset
> write_offset
- write_range
)
3737 /* Can only progress as far as has been backed up,
3738 * which must be suspended */
3739 max_progress
= backup_point
;
3740 else if (info
->reshape_progress
<= need_backup
)
3741 max_progress
= backup_point
;
3743 if (info
->array
.major_version
>= 0)
3744 /* Can progress until backup is needed */
3745 max_progress
= need_backup
;
3747 /* Can progress until metadata update is required */
3750 reshape
->after
.data_disks
;
3751 /* but data must be suspended */
3752 if (max_progress
< *suspend_point
)
3753 max_progress
= *suspend_point
;
3758 /* We know it is safe to progress to 'max_progress' providing
3759 * it is suspended or we are using native metadata.
3760 * Consider extending suspend_point 128M per device if it
3761 * is less than 64M per device beyond reshape_progress.
3762 * But always do a multiple of 'blocks'
3763 * FIXME this is too big - it takes to long to complete
3766 target
= 64*1024*2 * min(reshape
->before
.data_disks
,
3767 reshape
->after
.data_disks
);
3768 target
/= reshape
->backup_blocks
;
3771 target
*= reshape
->backup_blocks
;
3773 /* For externally managed metadata we always need to suspend IO to
3774 * the area being reshaped so we regularly push suspend_point forward.
3775 * For native metadata we only need the suspend if we are going to do
3779 if ((need_backup
> info
->reshape_progress
3780 || info
->array
.major_version
< 0) &&
3781 *suspend_point
< info
->reshape_progress
+ target
) {
3782 if (need_backup
< *suspend_point
+ 2 * target
)
3783 *suspend_point
= need_backup
;
3784 else if (*suspend_point
+ 2 * target
< array_size
)
3785 *suspend_point
+= 2 * target
;
3787 *suspend_point
= array_size
;
3788 sysfs_set_num(info
, NULL
, "suspend_hi", *suspend_point
);
3789 if (max_progress
> *suspend_point
)
3790 max_progress
= *suspend_point
;
3793 if (info
->array
.major_version
>= 0) {
3794 /* Only need to suspend when about to backup */
3795 if (info
->reshape_progress
< need_backup
* 2 &&
3796 *suspend_point
> 0) {
3798 sysfs_set_num(info
, NULL
, "suspend_lo", 0);
3799 sysfs_set_num(info
, NULL
, "suspend_hi", need_backup
);
3802 /* Need to suspend continually */
3803 if (info
->reshape_progress
< *suspend_point
)
3804 *suspend_point
= info
->reshape_progress
;
3805 if (*suspend_point
+ target
< info
->reshape_progress
)
3806 /* No need to move suspend region yet */;
3808 if (*suspend_point
>= 2 * target
)
3809 *suspend_point
-= 2 * target
;
3812 sysfs_set_num(info
, NULL
, "suspend_lo",
3815 if (max_progress
< *suspend_point
)
3816 max_progress
= *suspend_point
;
3820 /* now set sync_max to allow that progress. sync_max, like
3821 * sync_completed is a count of sectors written per device, so
3822 * we find the difference between max_progress and the start point,
3823 * and divide that by after.data_disks to get a sync_max
3825 * At the same time we convert wait_point to a similar number
3826 * for comparing against sync_completed.
3828 /* scale down max_progress to per_disk */
3829 max_progress
/= reshape
->after
.data_disks
;
3830 /* Round to chunk size as some kernels give an erroneously high number */
3831 max_progress
/= info
->new_chunk
/512;
3832 max_progress
*= info
->new_chunk
/512;
3833 /* And round to old chunk size as the kernel wants that */
3834 max_progress
/= info
->array
.chunk_size
/512;
3835 max_progress
*= info
->array
.chunk_size
/512;
3836 /* Limit progress to the whole device */
3837 if (max_progress
> info
->component_size
)
3838 max_progress
= info
->component_size
;
3839 wait_point
/= reshape
->after
.data_disks
;
3841 /* switch from 'device offset' to 'processed block count' */
3842 max_progress
= info
->component_size
- max_progress
;
3843 wait_point
= info
->component_size
- wait_point
;
3847 sysfs_set_num(info
, NULL
, "sync_max", max_progress
);
3849 /* Now wait. If we have already reached the point that we were
3850 * asked to wait to, don't wait at all, else wait for any change.
3851 * We need to select on 'sync_completed' as that is the place that
3852 * notifications happen, but we are really interested in
3853 * 'reshape_position'
3855 fd
= sysfs_get_fd(info
, NULL
, "sync_completed");
3857 goto check_progress
;
3859 if (sysfs_fd_get_ll(fd
, &completed
) < 0)
3860 goto check_progress
;
3862 while (completed
< max_progress
&& completed
< wait_point
) {
3863 /* Check that sync_action is still 'reshape' to avoid
3864 * waiting forever on a dead array
3867 if (sysfs_get_str(info
, NULL
, "sync_action",
3869 strncmp(action
, "reshape", 7) != 0)
3871 /* Some kernels reset 'sync_completed' to zero
3872 * before setting 'sync_action' to 'idle'.
3873 * So we need these extra tests.
3875 if (completed
== 0 && advancing
3876 && info
->reshape_progress
> 0)
3878 if (completed
== 0 && !advancing
3879 && info
->reshape_progress
< (info
->component_size
3880 * reshape
->after
.data_disks
))
3882 sysfs_wait(fd
, NULL
);
3883 if (sysfs_fd_get_ll(fd
, &completed
) < 0)
3884 goto check_progress
;
3886 /* Some kernels reset 'sync_completed' to zero,
3887 * we need to have real point we are in md
3890 completed
= max_progress
;
3892 /* some kernels can give an incorrectly high 'completed' number */
3893 completed
/= (info
->new_chunk
/512);
3894 completed
*= (info
->new_chunk
/512);
3895 /* Convert 'completed' back in to a 'progress' number */
3896 completed
*= reshape
->after
.data_disks
;
3898 completed
= info
->component_size
* reshape
->after
.data_disks
3901 *reshape_completed
= completed
;
3905 /* We return the need_backup flag. Caller will decide
3906 * how much - a multiple of ->backup_blocks up to *suspend_point
3909 return need_backup
> info
->reshape_progress
;
3911 return need_backup
>= info
->reshape_progress
;
3914 /* if we couldn't read a number from sync_completed, then
3915 * either the reshape did complete, or it aborted.
3916 * We can tell which by checking for 'none' in reshape_position.
3917 * If it did abort, then it might immediately restart if it
3918 * it was just a device failure that leaves us degraded but
3922 if (sysfs_get_str(info
, NULL
, "reshape_position", buf
, sizeof(buf
)) < 0
3923 || strncmp(buf
, "none", 4) != 0) {
3924 /* The abort might only be temporary. Wait up to 10
3925 * seconds for fd to contain a valid number again.
3929 unsigned long long new_sync_max
;
3930 while (fd
>= 0 && rv
< 0 && wait
> 0) {
3931 if (sysfs_wait(fd
, &wait
) != 1)
3933 switch (sysfs_fd_get_ll(fd
, &completed
)) {
3935 /* all good again */
3937 /* If "sync_max" is no longer max_progress
3938 * we need to freeze things
3940 sysfs_get_ll(info
, NULL
, "sync_max", &new_sync_max
);
3941 *frozen
= (new_sync_max
!= max_progress
);
3943 case -2: /* read error - abort */
3950 return rv
; /* abort */
3952 /* Maybe racing with array shutdown - check state */
3955 if (sysfs_get_str(info
, NULL
, "array_state", buf
, sizeof(buf
)) < 0
3956 || strncmp(buf
, "inactive", 8) == 0
3957 || strncmp(buf
, "clear",5) == 0)
3958 return -2; /* abort */
3959 return -1; /* complete */
3963 /* FIXME return status is never checked */
3964 static int grow_backup(struct mdinfo
*sra
,
3965 unsigned long long offset
, /* per device */
3966 unsigned long stripes
, /* per device, in old chunks */
3967 int *sources
, unsigned long long *offsets
,
3968 int disks
, int chunk
, int level
, int layout
,
3969 int dests
, int *destfd
, unsigned long long *destoffsets
,
3970 int part
, int *degraded
,
3973 /* Backup 'blocks' sectors at 'offset' on each device of the array,
3974 * to storage 'destfd' (offset 'destoffsets'), after first
3975 * suspending IO. Then allow resync to continue
3976 * over the suspended section.
3977 * Use part 'part' of the backup-super-block.
3982 unsigned long long ll
;
3984 //printf("offset %llu\n", offset);
3990 /* Check that array hasn't become degraded, else we might backup the wrong data */
3991 if (sysfs_get_ll(sra
, NULL
, "degraded", &ll
) < 0)
3992 return -1; /* FIXME this error is ignored */
3993 new_degraded
= (int)ll
;
3994 if (new_degraded
!= *degraded
) {
3995 /* check each device to ensure it is still working */
3997 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
3998 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
4000 if (sd
->disk
.state
& (1<<MD_DISK_SYNC
)) {
4002 if (sysfs_get_str(sra
, sd
, "state", sbuf
, 20) < 0 ||
4003 strstr(sbuf
, "faulty") ||
4004 strstr(sbuf
, "in_sync") == NULL
) {
4005 /* this device is dead */
4006 sd
->disk
.state
= (1<<MD_DISK_FAULTY
);
4007 if (sd
->disk
.raid_disk
>= 0 &&
4008 sources
[sd
->disk
.raid_disk
] >= 0) {
4009 close(sources
[sd
->disk
.raid_disk
]);
4010 sources
[sd
->disk
.raid_disk
] = -1;
4015 *degraded
= new_degraded
;
4018 bsb
.arraystart2
= __cpu_to_le64(offset
* odata
);
4019 bsb
.length2
= __cpu_to_le64(stripes
* (chunk
/512) * odata
);
4021 bsb
.arraystart
= __cpu_to_le64(offset
* odata
);
4022 bsb
.length
= __cpu_to_le64(stripes
* (chunk
/512) * odata
);
4025 bsb
.magic
[15] = '2';
4026 for (i
= 0; i
< dests
; i
++)
4028 lseek64(destfd
[i
], destoffsets
[i
] + __le64_to_cpu(bsb
.devstart2
)*512, 0);
4030 lseek64(destfd
[i
], destoffsets
[i
], 0);
4032 rv
= save_stripes(sources
, offsets
,
4033 disks
, chunk
, level
, layout
,
4035 offset
*512*odata
, stripes
* chunk
* odata
,
4040 bsb
.mtime
= __cpu_to_le64(time(0));
4041 for (i
= 0; i
< dests
; i
++) {
4042 bsb
.devstart
= __cpu_to_le64(destoffsets
[i
]/512);
4044 bsb
.sb_csum
= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
));
4045 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0)
4046 bsb
.sb_csum2
= bsb_csum((char*)&bsb
,
4047 ((char*)&bsb
.sb_csum2
)-((char*)&bsb
));
4050 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
] - 4096, 0)
4051 != destoffsets
[i
] - 4096)
4053 if (write(destfd
[i
], &bsb
, 512) != 512)
4055 if (destoffsets
[i
] > 4096) {
4056 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
]+stripes
*chunk
*odata
, 0) !=
4057 destoffsets
[i
]+stripes
*chunk
*odata
)
4059 if (write(destfd
[i
], &bsb
, 512) != 512)
4069 /* in 2.6.30, the value reported by sync_completed can be
4070 * less that it should be by one stripe.
4071 * This only happens when reshape hits sync_max and pauses.
4072 * So allow wait_backup to either extent sync_max further
4073 * than strictly necessary, or return before the
4074 * sync has got quite as far as we would really like.
4075 * This is what 'blocks2' is for.
4076 * The various caller give appropriate values so that
4079 /* FIXME return value is often ignored */
4080 static int forget_backup(int dests
, int *destfd
,
4081 unsigned long long *destoffsets
,
4085 * Erase backup 'part' (which is 0 or 1)
4091 bsb
.arraystart2
= __cpu_to_le64(0);
4092 bsb
.length2
= __cpu_to_le64(0);
4094 bsb
.arraystart
= __cpu_to_le64(0);
4095 bsb
.length
= __cpu_to_le64(0);
4097 bsb
.mtime
= __cpu_to_le64(time(0));
4099 for (i
= 0; i
< dests
; i
++) {
4100 bsb
.devstart
= __cpu_to_le64(destoffsets
[i
]/512);
4101 bsb
.sb_csum
= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
));
4102 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0)
4103 bsb
.sb_csum2
= bsb_csum((char*)&bsb
,
4104 ((char*)&bsb
.sb_csum2
)-((char*)&bsb
));
4105 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
]-4096, 0) !=
4106 destoffsets
[i
]-4096)
4109 write(destfd
[i
], &bsb
, 512) != 512)
4116 static void fail(char *msg
)
4119 rv
= (write(2, msg
, strlen(msg
)) != (int)strlen(msg
));
4120 rv
|= (write(2, "\n", 1) != 1);
4124 static char *abuf
, *bbuf
;
4125 static unsigned long long abuflen
;
4126 static void validate(int afd
, int bfd
, unsigned long long offset
)
4128 /* check that the data in the backup against the array.
4129 * This is only used for regression testing and should not
4130 * be used while the array is active
4134 lseek64(bfd
, offset
- 4096, 0);
4135 if (read(bfd
, &bsb2
, 512) != 512)
4136 fail("cannot read bsb");
4137 if (bsb2
.sb_csum
!= bsb_csum((char*)&bsb2
,
4138 ((char*)&bsb2
.sb_csum
)-((char*)&bsb2
)))
4139 fail("first csum bad");
4140 if (memcmp(bsb2
.magic
, "md_backup_data", 14) != 0)
4141 fail("magic is bad");
4142 if (memcmp(bsb2
.magic
, "md_backup_data-2", 16) == 0 &&
4143 bsb2
.sb_csum2
!= bsb_csum((char*)&bsb2
,
4144 ((char*)&bsb2
.sb_csum2
)-((char*)&bsb2
)))
4145 fail("second csum bad");
4147 if (__le64_to_cpu(bsb2
.devstart
)*512 != offset
)
4148 fail("devstart is wrong");
4151 unsigned long long len
= __le64_to_cpu(bsb2
.length
)*512;
4153 if (abuflen
< len
) {
4157 if (posix_memalign((void**)&abuf
, 4096, abuflen
) ||
4158 posix_memalign((void**)&bbuf
, 4096, abuflen
)) {
4160 /* just stop validating on mem-alloc failure */
4165 lseek64(bfd
, offset
, 0);
4166 if ((unsigned long long)read(bfd
, bbuf
, len
) != len
) {
4167 //printf("len %llu\n", len);
4168 fail("read first backup failed");
4170 lseek64(afd
, __le64_to_cpu(bsb2
.arraystart
)*512, 0);
4171 if ((unsigned long long)read(afd
, abuf
, len
) != len
)
4172 fail("read first from array failed");
4173 if (memcmp(bbuf
, abuf
, len
) != 0) {
4176 printf("offset=%llu len=%llu\n",
4177 (unsigned long long)__le64_to_cpu(bsb2
.arraystart
)*512, len
);
4178 for (i
=0; i
<len
; i
++)
4179 if (bbuf
[i
] != abuf
[i
]) {
4180 printf("first diff byte %d\n", i
);
4184 fail("data1 compare failed");
4188 unsigned long long len
= __le64_to_cpu(bsb2
.length2
)*512;
4190 if (abuflen
< len
) {
4194 abuf
= xmalloc(abuflen
);
4195 bbuf
= xmalloc(abuflen
);
4198 lseek64(bfd
, offset
+__le64_to_cpu(bsb2
.devstart2
)*512, 0);
4199 if ((unsigned long long)read(bfd
, bbuf
, len
) != len
)
4200 fail("read second backup failed");
4201 lseek64(afd
, __le64_to_cpu(bsb2
.arraystart2
)*512, 0);
4202 if ((unsigned long long)read(afd
, abuf
, len
) != len
)
4203 fail("read second from array failed");
4204 if (memcmp(bbuf
, abuf
, len
) != 0)
4205 fail("data2 compare failed");
4209 int child_monitor(int afd
, struct mdinfo
*sra
, struct reshape
*reshape
,
4210 struct supertype
*st
, unsigned long blocks
,
4211 int *fds
, unsigned long long *offsets
,
4212 int dests
, int *destfd
, unsigned long long *destoffsets
)
4214 /* Monitor a reshape where backup is being performed using
4215 * 'native' mechanism - either to a backup file, or
4216 * to some space in a spare.
4220 unsigned long long speed
;
4221 unsigned long long suspend_point
, array_size
;
4222 unsigned long long backup_point
, wait_point
;
4223 unsigned long long reshape_completed
;
4225 int increasing
= reshape
->after
.data_disks
>= reshape
->before
.data_disks
;
4226 int part
= 0; /* The next part of the backup area to fill. It may already
4227 * be full, so we need to check */
4228 int level
= reshape
->level
;
4229 int layout
= reshape
->before
.layout
;
4230 int data
= reshape
->before
.data_disks
;
4231 int disks
= reshape
->before
.data_disks
+ reshape
->parity
;
4232 int chunk
= sra
->array
.chunk_size
;
4234 unsigned long stripes
;
4238 /* set up the backup-super-block. This requires the
4239 * uuid from the array.
4241 /* Find a superblock */
4242 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
4246 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
4248 dn
= map_dev(sd
->disk
.major
, sd
->disk
.minor
, 1);
4249 devfd
= dev_open(dn
, O_RDONLY
);
4252 ok
= st
->ss
->load_super(st
, devfd
, NULL
);
4258 pr_err("Cannot find a superblock\n");
4262 memset(&bsb
, 0, 512);
4263 memcpy(bsb
.magic
, "md_backup_data-1", 16);
4264 st
->ss
->uuid_from_super(st
, uuid
);
4265 memcpy(bsb
.set_uuid
, uuid
, 16);
4266 bsb
.mtime
= __cpu_to_le64(time(0));
4267 bsb
.devstart2
= blocks
;
4269 stripes
= blocks
/ (sra
->array
.chunk_size
/512) /
4270 reshape
->before
.data_disks
;
4272 if (posix_memalign((void**)&buf
, 4096, disks
* chunk
))
4273 /* Don't start the 'reshape' */
4275 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
) {
4276 sysfs_get_ll(sra
, NULL
, "sync_speed_min", &speed
);
4277 sysfs_set_num(sra
, NULL
, "sync_speed_min", 200000);
4281 array_size
= sra
->component_size
* reshape
->after
.data_disks
;
4282 backup_point
= sra
->reshape_progress
;
4285 array_size
= sra
->component_size
* reshape
->before
.data_disks
;
4286 backup_point
= reshape
->backup_blocks
;
4287 suspend_point
= array_size
;
4293 /* Want to return as soon the oldest backup slot can
4294 * be released as that allows us to start backing up
4295 * some more, providing suspend_point has been
4296 * advanced, which it should have.
4299 wait_point
= array_size
;
4300 if (part
== 0 && __le64_to_cpu(bsb
.length
) > 0)
4301 wait_point
= (__le64_to_cpu(bsb
.arraystart
) +
4302 __le64_to_cpu(bsb
.length
));
4303 if (part
== 1 && __le64_to_cpu(bsb
.length2
) > 0)
4304 wait_point
= (__le64_to_cpu(bsb
.arraystart2
) +
4305 __le64_to_cpu(bsb
.length2
));
4308 if (part
== 0 && __le64_to_cpu(bsb
.length
) > 0)
4309 wait_point
= __le64_to_cpu(bsb
.arraystart
);
4310 if (part
== 1 && __le64_to_cpu(bsb
.length2
) > 0)
4311 wait_point
= __le64_to_cpu(bsb
.arraystart2
);
4314 reshape_completed
= sra
->reshape_progress
;
4315 rv
= progress_reshape(sra
, reshape
,
4316 backup_point
, wait_point
,
4317 &suspend_point
, &reshape_completed
,
4319 /* external metadata would need to ping_monitor here */
4320 sra
->reshape_progress
= reshape_completed
;
4322 /* Clear any backup region that is before 'here' */
4324 if (__le64_to_cpu(bsb
.length
) > 0 &&
4325 reshape_completed
>= (__le64_to_cpu(bsb
.arraystart
) +
4326 __le64_to_cpu(bsb
.length
)))
4327 forget_backup(dests
, destfd
,
4329 if (__le64_to_cpu(bsb
.length2
) > 0 &&
4330 reshape_completed
>= (__le64_to_cpu(bsb
.arraystart2
) +
4331 __le64_to_cpu(bsb
.length2
)))
4332 forget_backup(dests
, destfd
,
4335 if (__le64_to_cpu(bsb
.length
) > 0 &&
4336 reshape_completed
<= (__le64_to_cpu(bsb
.arraystart
)))
4337 forget_backup(dests
, destfd
,
4339 if (__le64_to_cpu(bsb
.length2
) > 0 &&
4340 reshape_completed
<= (__le64_to_cpu(bsb
.arraystart2
)))
4341 forget_backup(dests
, destfd
,
4351 if (rv
== 0 && increasing
&& !st
->ss
->external
) {
4352 /* No longer need to monitor this reshape */
4353 sysfs_set_str(sra
, NULL
, "sync_max", "max");
4359 unsigned long long offset
;
4360 unsigned long actual_stripes
;
4361 /* Need to backup some data.
4362 * If 'part' is not used and the desired
4363 * backup size is suspended, do a backup,
4364 * then consider the next part.
4366 /* Check that 'part' is unused */
4367 if (part
== 0 && __le64_to_cpu(bsb
.length
) != 0)
4369 if (part
== 1 && __le64_to_cpu(bsb
.length2
) != 0)
4372 offset
= backup_point
/ data
;
4373 actual_stripes
= stripes
;
4375 if (offset
+ actual_stripes
* (chunk
/512) >
4376 sra
->component_size
)
4377 actual_stripes
= ((sra
->component_size
- offset
)
4379 if (offset
+ actual_stripes
* (chunk
/512) >
4383 if (offset
< actual_stripes
* (chunk
/512))
4384 actual_stripes
= offset
/ (chunk
/512);
4385 offset
-= actual_stripes
* (chunk
/512);
4386 if (offset
< suspend_point
/data
)
4389 if (actual_stripes
== 0)
4391 grow_backup(sra
, offset
, actual_stripes
,
4393 disks
, chunk
, level
, layout
,
4394 dests
, destfd
, destoffsets
,
4395 part
, °raded
, buf
);
4396 validate(afd
, destfd
[0], destoffsets
[0]);
4397 /* record where 'part' is up to */
4400 backup_point
+= actual_stripes
* (chunk
/512) * data
;
4402 backup_point
-= actual_stripes
* (chunk
/512) * data
;
4406 /* FIXME maybe call progress_reshape one more time instead */
4407 /* remove any remaining suspension */
4408 sysfs_set_num(sra
, NULL
, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL
);
4409 sysfs_set_num(sra
, NULL
, "suspend_hi", 0);
4410 sysfs_set_num(sra
, NULL
, "suspend_lo", 0);
4411 sysfs_set_num(sra
, NULL
, "sync_min", 0);
4413 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
)
4414 sysfs_set_num(sra
, NULL
, "sync_speed_min", speed
);
4420 * If any spare contains md_back_data-1 which is recent wrt mtime,
4421 * write that data into the array and update the super blocks with
4422 * the new reshape_progress
4424 int Grow_restart(struct supertype
*st
, struct mdinfo
*info
, int *fdlist
, int cnt
,
4425 char *backup_file
, int verbose
)
4429 unsigned long long *offsets
;
4430 unsigned long long nstripe
, ostripe
;
4433 odata
= info
->array
.raid_disks
- info
->delta_disks
- 1;
4434 if (info
->array
.level
== 6) odata
--; /* number of data disks */
4435 ndata
= info
->array
.raid_disks
- 1;
4436 if (info
->new_level
== 6) ndata
--;
4438 old_disks
= info
->array
.raid_disks
- info
->delta_disks
;
4440 if (info
->delta_disks
<= 0)
4441 /* Didn't grow, so the backup file must have
4445 for (i
=old_disks
-(backup_file
?1:0); i
<cnt
; i
++) {
4446 struct mdinfo dinfo
;
4449 char *devname
, namebuf
[20];
4450 unsigned long long lo
, hi
;
4452 /* This was a spare and may have some saved data on it.
4453 * Load the superblock, find and load the
4454 * backup_super_block.
4455 * If either fail, go on to next device.
4456 * If the backup contains no new info, just return
4457 * else restore data and update all superblocks
4459 if (i
== old_disks
-1) {
4460 fd
= open(backup_file
, O_RDONLY
);
4462 pr_err("backup file %s inaccessible: %s\n",
4463 backup_file
, strerror(errno
));
4466 devname
= backup_file
;
4471 if (st
->ss
->load_super(st
, fd
, NULL
))
4474 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
4475 st
->ss
->free_super(st
);
4478 (dinfo
.data_offset
+ dinfo
.component_size
- 8) <<9,
4480 pr_err("Cannot seek on device %d\n", i
);
4481 continue; /* Cannot seek */
4483 sprintf(namebuf
, "device-%d", i
);
4486 if (read(fd
, &bsb
, sizeof(bsb
)) != sizeof(bsb
)) {
4488 pr_err("Cannot read from %s\n", devname
);
4489 continue; /* Cannot read */
4491 if (memcmp(bsb
.magic
, "md_backup_data-1", 16) != 0 &&
4492 memcmp(bsb
.magic
, "md_backup_data-2", 16) != 0) {
4494 pr_err("No backup metadata on %s\n", devname
);
4497 if (bsb
.sb_csum
!= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
))) {
4499 pr_err("Bad backup-metadata checksum on %s\n", devname
);
4500 continue; /* bad checksum */
4502 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0 &&
4503 bsb
.sb_csum2
!= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum2
)-((char*)&bsb
))) {
4505 pr_err("Bad backup-metadata checksum2 on %s\n", devname
);
4506 continue; /* Bad second checksum */
4508 if (memcmp(bsb
.set_uuid
,info
->uuid
, 16) != 0) {
4510 pr_err("Wrong uuid on backup-metadata on %s\n", devname
);
4511 continue; /* Wrong uuid */
4514 /* array utime and backup-mtime should be updated at much the same time, but it seems that
4515 * sometimes they aren't... So allow considerable flexability in matching, and allow
4516 * this test to be overridden by an environment variable.
4518 if (info
->array
.utime
> (int)__le64_to_cpu(bsb
.mtime
) + 2*60*60 ||
4519 info
->array
.utime
< (int)__le64_to_cpu(bsb
.mtime
) - 10*60) {
4520 if (check_env("MDADM_GROW_ALLOW_OLD")) {
4521 pr_err("accepting backup with timestamp %lu "
4522 "for array with timestamp %lu\n",
4523 (unsigned long)__le64_to_cpu(bsb
.mtime
),
4524 (unsigned long)info
->array
.utime
);
4526 pr_err("too-old timestamp on backup-metadata on %s\n", devname
);
4527 pr_err("If you think it is should be safe, try 'export MDADM_GROW_ALLOW_OLD=1'\n");
4528 continue; /* time stamp is too bad */
4532 if (bsb
.magic
[15] == '1') {
4533 if (bsb
.length
== 0)
4535 if (info
->delta_disks
>= 0) {
4536 /* reshape_progress is increasing */
4537 if (__le64_to_cpu(bsb
.arraystart
)
4538 + __le64_to_cpu(bsb
.length
)
4539 < info
->reshape_progress
) {
4542 pr_err("backup-metadata found on %s but is not needed\n", devname
);
4543 continue; /* No new data here */
4546 /* reshape_progress is decreasing */
4547 if (__le64_to_cpu(bsb
.arraystart
) >=
4548 info
->reshape_progress
)
4549 goto nonew
; /* No new data here */
4552 if (bsb
.length
== 0 && bsb
.length2
== 0)
4554 if (info
->delta_disks
>= 0) {
4555 /* reshape_progress is increasing */
4556 if ((__le64_to_cpu(bsb
.arraystart
)
4557 + __le64_to_cpu(bsb
.length
)
4558 < info
->reshape_progress
)
4560 (__le64_to_cpu(bsb
.arraystart2
)
4561 + __le64_to_cpu(bsb
.length2
)
4562 < info
->reshape_progress
))
4563 goto nonew
; /* No new data here */
4565 /* reshape_progress is decreasing */
4566 if (__le64_to_cpu(bsb
.arraystart
) >=
4567 info
->reshape_progress
&&
4568 __le64_to_cpu(bsb
.arraystart2
) >=
4569 info
->reshape_progress
)
4570 goto nonew
; /* No new data here */
4573 if (lseek64(fd
, __le64_to_cpu(bsb
.devstart
)*512, 0)< 0) {
4576 pr_err("Failed to verify secondary backup-metadata block on %s\n",
4578 continue; /* Cannot seek */
4580 /* There should be a duplicate backup superblock 4k before here */
4581 if (lseek64(fd
, -4096, 1) < 0 ||
4582 read(fd
, &bsb2
, sizeof(bsb2
)) != sizeof(bsb2
))
4583 goto second_fail
; /* Cannot find leading superblock */
4584 if (bsb
.magic
[15] == '1')
4585 bsbsize
= offsetof(struct mdp_backup_super
, pad1
);
4587 bsbsize
= offsetof(struct mdp_backup_super
, pad
);
4588 if (memcmp(&bsb2
, &bsb
, bsbsize
) != 0)
4589 goto second_fail
; /* Cannot find leading superblock */
4591 /* Now need the data offsets for all devices. */
4592 offsets
= xmalloc(sizeof(*offsets
)*info
->array
.raid_disks
);
4593 for(j
=0; j
<info
->array
.raid_disks
; j
++) {
4596 if (st
->ss
->load_super(st
, fdlist
[j
], NULL
))
4597 /* FIXME should be this be an error */
4599 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
4600 st
->ss
->free_super(st
);
4601 offsets
[j
] = dinfo
.data_offset
* 512;
4603 printf(Name
": restoring critical section\n");
4605 if (restore_stripes(fdlist
, offsets
,
4606 info
->array
.raid_disks
,
4610 fd
, __le64_to_cpu(bsb
.devstart
)*512,
4611 __le64_to_cpu(bsb
.arraystart
)*512,
4612 __le64_to_cpu(bsb
.length
)*512, NULL
)) {
4613 /* didn't succeed, so giveup */
4615 pr_err("Error restoring backup from %s\n",
4621 if (bsb
.magic
[15] == '2' &&
4622 restore_stripes(fdlist
, offsets
,
4623 info
->array
.raid_disks
,
4627 fd
, __le64_to_cpu(bsb
.devstart
)*512 +
4628 __le64_to_cpu(bsb
.devstart2
)*512,
4629 __le64_to_cpu(bsb
.arraystart2
)*512,
4630 __le64_to_cpu(bsb
.length2
)*512, NULL
)) {
4631 /* didn't succeed, so giveup */
4633 pr_err("Error restoring second backup from %s\n",
4641 /* Ok, so the data is restored. Let's update those superblocks. */
4645 lo
= __le64_to_cpu(bsb
.arraystart
);
4646 hi
= lo
+ __le64_to_cpu(bsb
.length
);
4648 if (bsb
.magic
[15] == '2' && bsb
.length2
) {
4649 unsigned long long lo1
, hi1
;
4650 lo1
= __le64_to_cpu(bsb
.arraystart2
);
4651 hi1
= lo1
+ __le64_to_cpu(bsb
.length2
);
4655 } else if (lo
< lo1
)
4661 (info
->reshape_progress
< lo
||
4662 info
->reshape_progress
> hi
))
4663 /* backup does not affect reshape_progress*/ ;
4664 else if (info
->delta_disks
>= 0) {
4665 info
->reshape_progress
= __le64_to_cpu(bsb
.arraystart
) +
4666 __le64_to_cpu(bsb
.length
);
4667 if (bsb
.magic
[15] == '2') {
4668 unsigned long long p2
= __le64_to_cpu(bsb
.arraystart2
) +
4669 __le64_to_cpu(bsb
.length2
);
4670 if (p2
> info
->reshape_progress
)
4671 info
->reshape_progress
= p2
;
4674 info
->reshape_progress
= __le64_to_cpu(bsb
.arraystart
);
4675 if (bsb
.magic
[15] == '2') {
4676 unsigned long long p2
= __le64_to_cpu(bsb
.arraystart2
);
4677 if (p2
< info
->reshape_progress
)
4678 info
->reshape_progress
= p2
;
4681 for (j
=0; j
<info
->array
.raid_disks
; j
++) {
4684 if (st
->ss
->load_super(st
, fdlist
[j
], NULL
))
4686 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
4687 dinfo
.reshape_progress
= info
->reshape_progress
;
4688 st
->ss
->update_super(st
, &dinfo
,
4689 "_reshape_progress",
4691 st
->ss
->store_super(st
, fdlist
[j
]);
4692 st
->ss
->free_super(st
);
4696 /* Didn't find any backup data, try to see if any
4699 if (info
->delta_disks
< 0) {
4700 /* When shrinking, the critical section is at the end.
4701 * So see if we are before the critical section.
4703 unsigned long long first_block
;
4704 nstripe
= ostripe
= 0;
4706 while (ostripe
>= nstripe
) {
4707 ostripe
+= info
->array
.chunk_size
/ 512;
4708 first_block
= ostripe
* odata
;
4709 nstripe
= first_block
/ ndata
/ (info
->new_chunk
/512) *
4710 (info
->new_chunk
/512);
4713 if (info
->reshape_progress
>= first_block
)
4716 if (info
->delta_disks
> 0) {
4717 /* See if we are beyond the critical section. */
4718 unsigned long long last_block
;
4719 nstripe
= ostripe
= 0;
4721 while (nstripe
>= ostripe
) {
4722 nstripe
+= info
->new_chunk
/ 512;
4723 last_block
= nstripe
* ndata
;
4724 ostripe
= last_block
/ odata
/ (info
->array
.chunk_size
/512) *
4725 (info
->array
.chunk_size
/512);
4728 if (info
->reshape_progress
>= last_block
)
4731 /* needed to recover critical section! */
4733 pr_err("Failed to find backup of critical section\n");
4737 int Grow_continue_command(char *devname
, int fd
,
4738 char *backup_file
, int verbose
)
4741 struct supertype
*st
= NULL
;
4742 struct mdinfo
*content
= NULL
;
4743 struct mdinfo array
;
4744 char *subarray
= NULL
;
4745 struct mdinfo
*cc
= NULL
;
4746 struct mdstat_ent
*mdstat
= NULL
;
4750 dprintf("Grow continue from command line called for %s\n",
4753 st
= super_by_fd(fd
, &subarray
);
4754 if (!st
|| !st
->ss
) {
4755 pr_err("Unable to determine metadata format for %s\n",
4759 dprintf("Grow continue is run for ");
4760 if (st
->ss
->external
== 0) {
4762 dprintf("native array (%s)\n", devname
);
4763 if (ioctl(fd
, GET_ARRAY_INFO
, &array
.array
) < 0) {
4764 pr_err("%s is not an active md array -"
4765 " aborting\n", devname
);
4767 goto Grow_continue_command_exit
;
4770 /* Need to load a superblock.
4771 * FIXME we should really get what we need from
4774 for (d
= 0; d
< MAX_DISKS
; d
++) {
4775 mdu_disk_info_t disk
;
4779 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0)
4781 if (disk
.major
== 0 && disk
.minor
== 0)
4783 if ((disk
.state
& (1 << MD_DISK_ACTIVE
)) == 0)
4785 dv
= map_dev(disk
.major
, disk
.minor
, 1);
4788 fd2
= dev_open(dv
, O_RDONLY
);
4791 err
= st
->ss
->load_super(st
, fd2
, NULL
);
4793 /* invalidate fd2 to avoid possible double close() */
4799 if (d
== MAX_DISKS
) {
4800 pr_err("Unable to load metadata for %s\n",
4803 goto Grow_continue_command_exit
;
4805 st
->ss
->getinfo_super(st
, content
, NULL
);
4810 dprintf("subarray (%s)\n", subarray
);
4811 container
= st
->container_devnm
;
4812 cfd
= open_dev_excl(st
->container_devnm
);
4814 container
= st
->devnm
;
4816 cfd
= open_dev_excl(st
->devnm
);
4817 dprintf("container (%s)\n", container
);
4821 pr_err("Unable to open container "
4822 "for %s\n", devname
);
4824 goto Grow_continue_command_exit
;
4827 /* find in container array under reshape
4829 ret_val
= st
->ss
->load_container(st
, cfd
, NULL
);
4831 pr_err("Cannot read superblock for %s\n",
4834 goto Grow_continue_command_exit
;
4837 cc
= st
->ss
->container_content(st
, subarray
);
4838 for (content
= cc
; content
; content
= content
->next
) {
4840 int allow_reshape
= 1;
4842 if (content
->reshape_active
== 0)
4844 /* The decision about array or container wide
4845 * reshape is taken in Grow_continue based
4846 * content->reshape_active state, therefore we
4847 * need to check_reshape based on
4848 * reshape_active and subarray name
4850 if (content
->array
.state
& (1<<MD_SB_BLOCK_VOLUME
))
4852 if (content
->reshape_active
== CONTAINER_RESHAPE
&&
4853 (content
->array
.state
4854 & (1<<MD_SB_BLOCK_CONTAINER_RESHAPE
)))
4857 if (!allow_reshape
) {
4858 pr_err("cannot continue reshape of an array"
4859 " in container with unsupported"
4860 " metadata: %s(%s)\n",
4861 devname
, container
);
4863 goto Grow_continue_command_exit
;
4866 array
= strchr(content
->text_version
+1, '/')+1;
4867 mdstat
= mdstat_by_subdev(array
, container
);
4870 if (mdstat
->active
== 0) {
4871 pr_err("Skipping inactive array %s.\n",
4873 free_mdstat(mdstat
);
4880 pr_err("Unable to determine reshaped "
4881 "array for %s\n", devname
);
4883 goto Grow_continue_command_exit
;
4885 fd2
= open_dev(mdstat
->devnm
);
4887 pr_err("cannot open (%s)\n", mdstat
->devnm
);
4889 goto Grow_continue_command_exit
;
4892 sysfs_init(content
, fd2
, mdstat
->devnm
);
4894 /* start mdmon in case it is not running
4896 if (!mdmon_running(container
))
4897 start_mdmon(container
);
4898 ping_monitor(container
);
4900 if (mdmon_running(container
))
4901 st
->update_tail
= &st
->updates
;
4903 pr_err("No mdmon found. "
4904 "Grow cannot continue.\n");
4906 goto Grow_continue_command_exit
;
4910 /* verify that array under reshape is started from
4913 if (verify_reshape_position(content
, content
->array
.level
) < 0) {
4915 goto Grow_continue_command_exit
;
4920 ret_val
= Grow_continue(fd
, st
, content
, backup_file
, 1, 0);
4922 Grow_continue_command_exit
:
4927 st
->ss
->free_super(st
);
4928 free_mdstat(mdstat
);
4935 int Grow_continue(int mdfd
, struct supertype
*st
, struct mdinfo
*info
,
4936 char *backup_file
, int forked
, int freeze_reshape
)
4940 if (!info
->reshape_active
)
4943 if (st
->ss
->external
) {
4944 int cfd
= open_dev(st
->container_devnm
);
4949 st
->ss
->load_container(st
, cfd
, st
->container_devnm
);
4951 ret_val
= reshape_container(st
->container_devnm
, NULL
, mdfd
,
4952 st
, info
, 0, backup_file
,
4954 1 | info
->reshape_active
,
4957 ret_val
= reshape_array(NULL
, mdfd
, "array", st
, info
, 1,
4958 NULL
, INVALID_SECTORS
,
4959 backup_file
, 0, forked
,
4960 1 | info
->reshape_active
,
4966 char *make_backup(char *name
)
4968 char *base
= "backup_file-";
4972 len
= strlen(MAP_DIR
) + 1 + strlen(base
) + strlen(name
)+1;
4973 fname
= xmalloc(len
);
4974 sprintf(fname
, "%s/%s%s", MAP_DIR
, base
, name
);
4978 char *locate_backup(char *name
)
4980 char *fl
= make_backup(name
);
4983 if (stat(fl
, &stb
) == 0 &&
4984 S_ISREG(stb
.st_mode
))