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 section for reshape - sorry.\n");
92 pr_err("Possibly you need to specify a --backup-file\n");
96 dprintf("restore_backup() returns status OK.\n");
100 int Grow_Add_device(char *devname
, int fd
, char *newdev
)
102 /* Add a device to an active array.
103 * Currently, just extend a linear array.
104 * This requires writing a new superblock on the
105 * new device, calling the kernel to add the device,
106 * and if that succeeds, update the superblock on
108 * This means that we need to *find* all other devices.
115 struct supertype
*st
= NULL
;
116 char *subarray
= NULL
;
118 if (ioctl(fd
, GET_ARRAY_INFO
, &info
.array
) < 0) {
119 pr_err("cannot get array info for %s\n", devname
);
123 if (info
.array
.level
!= -1) {
124 pr_err("can only add devices to linear arrays\n");
128 st
= super_by_fd(fd
, &subarray
);
130 pr_err("cannot handle arrays with superblock version %d\n",
131 info
.array
.major_version
);
136 pr_err("Cannot grow linear sub-arrays yet\n");
142 nfd
= open(newdev
, O_RDWR
|O_EXCL
|O_DIRECT
);
144 pr_err("cannot open %s\n", newdev
);
149 if ((stb
.st_mode
& S_IFMT
) != S_IFBLK
) {
150 pr_err("%s is not a block device!\n", newdev
);
155 /* now check out all the devices and make sure we can read the
157 for (d
=0 ; d
< info
.array
.raid_disks
; d
++) {
158 mdu_disk_info_t disk
;
161 st
->ss
->free_super(st
);
164 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0) {
165 pr_err("cannot get device detail for device %d\n",
171 dv
= map_dev(disk
.major
, disk
.minor
, 1);
173 pr_err("cannot find device file for device %d\n",
179 fd2
= dev_open(dv
, O_RDWR
);
181 pr_err("cannot open device file %s\n", dv
);
187 if (st
->ss
->load_super(st
, fd2
, NULL
)) {
188 pr_err("cannot find super block on %s\n", dv
);
196 /* Ok, looks good. Lets update the superblock and write it out to
200 info
.disk
.number
= d
;
201 info
.disk
.major
= major(stb
.st_rdev
);
202 info
.disk
.minor
= minor(stb
.st_rdev
);
203 info
.disk
.raid_disk
= d
;
204 info
.disk
.state
= (1 << MD_DISK_SYNC
) | (1 << MD_DISK_ACTIVE
);
205 st
->ss
->update_super(st
, &info
, "linear-grow-new", newdev
,
208 if (st
->ss
->store_super(st
, nfd
)) {
209 pr_err("Cannot store new superblock on %s\n",
216 if (ioctl(fd
, ADD_NEW_DISK
, &info
.disk
) != 0) {
217 pr_err("Cannot add new disk to this array\n");
220 /* Well, that seems to have worked.
221 * Now go through and update all superblocks
224 if (ioctl(fd
, GET_ARRAY_INFO
, &info
.array
) < 0) {
225 pr_err("cannot get array info for %s\n", devname
);
230 for (d
=0 ; d
< info
.array
.raid_disks
; d
++) {
231 mdu_disk_info_t disk
;
235 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0) {
236 pr_err("cannot get device detail for device %d\n",
240 dv
= map_dev(disk
.major
, disk
.minor
, 1);
242 pr_err("cannot find device file for device %d\n",
246 fd2
= dev_open(dv
, O_RDWR
);
248 pr_err("cannot open device file %s\n", dv
);
251 if (st
->ss
->load_super(st
, fd2
, NULL
)) {
252 pr_err("cannot find super block on %s\n", dv
);
256 info
.array
.raid_disks
= nd
+1;
257 info
.array
.nr_disks
= nd
+1;
258 info
.array
.active_disks
= nd
+1;
259 info
.array
.working_disks
= nd
+1;
261 st
->ss
->update_super(st
, &info
, "linear-grow-update", dv
,
264 if (st
->ss
->store_super(st
, fd2
)) {
265 pr_err("Cannot store new superblock on %s\n", dv
);
275 int Grow_addbitmap(char *devname
, int fd
, struct context
*c
, struct shape
*s
)
278 * First check that array doesn't have a bitmap
279 * Then create the bitmap
282 * For internal bitmaps, we need to check the version,
283 * find all the active devices, and write the bitmap block
286 mdu_bitmap_file_t bmf
;
287 mdu_array_info_t array
;
288 struct supertype
*st
;
289 char *subarray
= NULL
;
290 int major
= BITMAP_MAJOR_HI
;
291 int vers
= md_get_version(fd
);
292 unsigned long long bitmapsize
, array_size
;
295 major
= BITMAP_MAJOR_HOSTENDIAN
;
296 pr_err("Warning - bitmaps created on this kernel are not portable\n"
297 " between different architectures. Consider upgrading the Linux kernel.\n");
300 if (ioctl(fd
, GET_BITMAP_FILE
, &bmf
) != 0) {
302 pr_err("Memory allocation failure.\n");
304 pr_err("bitmaps not supported by this kernel.\n");
307 if (bmf
.pathname
[0]) {
308 if (strcmp(s
->bitmap_file
,"none")==0) {
309 if (ioctl(fd
, SET_BITMAP_FILE
, -1)!= 0) {
310 pr_err("failed to remove bitmap %s\n",
316 pr_err("%s already has a bitmap (%s)\n",
317 devname
, bmf
.pathname
);
320 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
321 pr_err("cannot get array status for %s\n", devname
);
324 if (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)) {
325 if (strcmp(s
->bitmap_file
, "none")==0) {
326 array
.state
&= ~(1<<MD_SB_BITMAP_PRESENT
);
327 if (ioctl(fd
, SET_ARRAY_INFO
, &array
)!= 0) {
328 pr_err("failed to remove internal bitmap.\n");
333 pr_err("Internal bitmap already present on %s\n",
338 if (strcmp(s
->bitmap_file
, "none") == 0) {
339 pr_err("no bitmap found on %s\n", devname
);
342 if (array
.level
<= 0) {
343 pr_err("Bitmaps not meaningful with level %s\n",
344 map_num(pers
, array
.level
)?:"of this array");
347 bitmapsize
= array
.size
;
349 if (get_dev_size(fd
, NULL
, &array_size
) &&
350 array_size
> (0x7fffffffULL
<<9)) {
351 /* Array is big enough that we cannot trust array.size
352 * try other approaches
354 bitmapsize
= get_component_size(fd
);
356 if (bitmapsize
== 0) {
357 pr_err("Cannot reliably determine size of array to create bitmap - sorry.\n");
361 if (array
.level
== 10) {
362 int ncopies
= (array
.layout
&255)*((array
.layout
>>8)&255);
363 bitmapsize
= bitmapsize
* array
.raid_disks
/ ncopies
;
366 st
= super_by_fd(fd
, &subarray
);
368 pr_err("Cannot understand version %d.%d\n",
369 array
.major_version
, array
.minor_version
);
373 pr_err("Cannot add bitmaps to sub-arrays yet\n");
378 if (strcmp(s
->bitmap_file
, "internal") == 0) {
381 int offset_setable
= 0;
383 if (st
->ss
->add_internal_bitmap
== NULL
) {
384 pr_err("Internal bitmaps not supported with %s metadata\n", st
->ss
->name
);
387 mdi
= sysfs_read(fd
, NULL
, GET_BITMAP_LOCATION
);
390 for (d
=0; d
< st
->max_devs
; d
++) {
391 mdu_disk_info_t disk
;
394 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0)
396 if (disk
.major
== 0 &&
399 if ((disk
.state
& (1<<MD_DISK_SYNC
))==0)
401 dv
= map_dev(disk
.major
, disk
.minor
, 1);
403 int fd2
= dev_open(dv
, O_RDWR
);
406 if (st
->ss
->load_super(st
, fd2
, NULL
)==0) {
407 if (st
->ss
->add_internal_bitmap(
409 &s
->bitmap_chunk
, c
->delay
, s
->write_behind
,
410 bitmapsize
, offset_setable
,
413 st
->ss
->write_bitmap(st
, fd2
);
415 pr_err("failed to create internal bitmap - chunksize problem.\n");
423 if (offset_setable
) {
424 st
->ss
->getinfo_super(st
, mdi
, NULL
);
425 sysfs_init(mdi
, fd
, NULL
);
426 rv
= sysfs_set_num_signed(mdi
, NULL
, "bitmap/location",
429 array
.state
|= (1<<MD_SB_BITMAP_PRESENT
);
430 rv
= ioctl(fd
, SET_ARRAY_INFO
, &array
);
434 pr_err("Cannot add bitmap while array is resyncing or reshaping etc.\n");
435 pr_err("failed to set internal bitmap.\n");
442 int max_devs
= st
->max_devs
;
444 /* try to load a superblock */
445 for (d
= 0; d
< max_devs
; d
++) {
446 mdu_disk_info_t disk
;
450 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0)
452 if ((disk
.major
==0 && disk
.minor
==0) ||
453 (disk
.state
& (1<<MD_DISK_REMOVED
)))
455 dv
= map_dev(disk
.major
, disk
.minor
, 1);
458 fd2
= dev_open(dv
, O_RDONLY
);
460 if (st
->ss
->load_super(st
, fd2
, NULL
) == 0) {
462 st
->ss
->uuid_from_super(st
, uuid
);
469 pr_err("cannot find UUID for array!\n");
472 if (CreateBitmap(s
->bitmap_file
, c
->force
, (char*)uuid
, s
->bitmap_chunk
,
473 c
->delay
, s
->write_behind
, bitmapsize
, major
)) {
476 bitmap_fd
= open(s
->bitmap_file
, O_RDWR
);
478 pr_err("weird: %s cannot be opened\n",
482 if (ioctl(fd
, SET_BITMAP_FILE
, bitmap_fd
) < 0) {
485 pr_err("Cannot add bitmap while array is resyncing or reshaping etc.\n");
486 pr_err("Cannot set bitmap file for %s: %s\n",
487 devname
, strerror(err
));
496 * When reshaping an array we might need to backup some data.
497 * This is written to all spares with a 'super_block' describing it.
498 * The superblock goes 4K from the end of the used space on the
500 * It if written after the backup is complete.
501 * It has the following structure.
504 static struct mdp_backup_super
{
505 char magic
[16]; /* md_backup_data-1 or -2 */
508 /* start/sizes in 512byte sectors */
509 __u64 devstart
; /* address on backup device/file of data */
512 __u32 sb_csum
; /* csum of preceeding bytes. */
514 __u64 devstart2
; /* offset in to data of second section */
517 __u32 sb_csum2
; /* csum of preceeding bytes. */
519 } __attribute__((aligned(512))) bsb
, bsb2
;
521 static __u32
bsb_csum(char *buf
, int len
)
525 for (i
= 0; i
< len
; i
++)
526 csum
= (csum
<<3) + buf
[0];
527 return __cpu_to_le32(csum
);
530 static int check_idle(struct supertype
*st
)
532 /* Check that all member arrays for this container, or the
533 * container of this array, are idle
535 char *container
= (st
->container_devnm
[0]
536 ? st
->container_devnm
: st
->devnm
);
537 struct mdstat_ent
*ent
, *e
;
540 ent
= mdstat_read(0, 0);
541 for (e
= ent
; e
; e
= e
->next
) {
542 if (!is_container_member(e
, container
))
544 if (e
->percent
>= 0) {
553 static int freeze_container(struct supertype
*st
)
555 char *container
= (st
->container_devnm
[0]
556 ? st
->container_devnm
: st
->devnm
);
561 if (block_monitor(container
, 1)) {
562 pr_err("failed to freeze container\n");
569 static void unfreeze_container(struct supertype
*st
)
571 char *container
= (st
->container_devnm
[0]
572 ? st
->container_devnm
: st
->devnm
);
574 unblock_monitor(container
, 1);
577 static int freeze(struct supertype
*st
)
579 /* Try to freeze resync/rebuild on this array/container.
580 * Return -1 if the array is busy,
581 * return -2 container cannot be frozen,
582 * return 0 if this kernel doesn't support 'frozen'
583 * return 1 if it worked.
585 if (st
->ss
->external
)
586 return freeze_container(st
);
588 struct mdinfo
*sra
= sysfs_read(-1, st
->devnm
, GET_VERSION
);
594 /* Need to clear any 'read-auto' status */
595 if (sysfs_get_str(sra
, NULL
, "array_state", buf
, 20) > 0 &&
596 strncmp(buf
, "read-auto", 9) == 0)
597 sysfs_set_str(sra
, NULL
, "array_state", "clean");
599 err
= sysfs_freeze_array(sra
);
605 static void unfreeze(struct supertype
*st
)
607 if (st
->ss
->external
)
608 return unfreeze_container(st
);
610 struct mdinfo
*sra
= sysfs_read(-1, st
->devnm
, GET_VERSION
);
614 sysfs_get_str(sra
, NULL
, "sync_action", buf
, 20) > 0
615 && strcmp(buf
, "frozen\n") == 0) {
616 printf("unfreeze\n");
617 sysfs_set_str(sra
, NULL
, "sync_action", "idle");
623 static void wait_reshape(struct mdinfo
*sra
)
625 int fd
= sysfs_get_fd(sra
, NULL
, "sync_action");
631 while (sysfs_fd_get_str(fd
, action
, 20) > 0 &&
632 strncmp(action
, "reshape", 7) == 0)
633 sysfs_wait(fd
, NULL
);
637 static int reshape_super(struct supertype
*st
, unsigned long long size
,
638 int level
, int layout
, int chunksize
, int raid_disks
,
639 int delta_disks
, char *backup_file
, char *dev
,
640 int direction
, int verbose
)
642 /* nothing extra to check in the native case */
643 if (!st
->ss
->external
)
645 if (!st
->ss
->reshape_super
||
646 !st
->ss
->manage_reshape
) {
647 pr_err("%s metadata does not support reshape\n",
652 return st
->ss
->reshape_super(st
, size
, level
, layout
, chunksize
,
653 raid_disks
, delta_disks
, backup_file
, dev
,
657 static void sync_metadata(struct supertype
*st
)
659 if (st
->ss
->external
) {
660 if (st
->update_tail
) {
661 flush_metadata_updates(st
);
662 st
->update_tail
= &st
->updates
;
664 st
->ss
->sync_metadata(st
);
668 static int subarray_set_num(char *container
, struct mdinfo
*sra
, char *name
, int n
)
670 /* when dealing with external metadata subarrays we need to be
671 * prepared to handle EAGAIN. The kernel may need to wait for
672 * mdmon to mark the array active so the kernel can handle
673 * allocations/writeback when preparing the reshape action
674 * (md_allow_write()). We temporarily disable safe_mode_delay
675 * to close a race with the array_state going clean before the
676 * next write to raid_disks / stripe_cache_size
681 /* only 'raid_disks' and 'stripe_cache_size' trigger md_allow_write */
683 (strcmp(name
, "raid_disks") != 0 &&
684 strcmp(name
, "stripe_cache_size") != 0))
685 return sysfs_set_num(sra
, NULL
, name
, n
);
687 rc
= sysfs_get_str(sra
, NULL
, "safe_mode_delay", safe
, sizeof(safe
));
690 sysfs_set_num(sra
, NULL
, "safe_mode_delay", 0);
691 rc
= sysfs_set_num(sra
, NULL
, name
, n
);
692 if (rc
< 0 && errno
== EAGAIN
) {
693 ping_monitor(container
);
694 /* if we get EAGAIN here then the monitor is not active
697 rc
= sysfs_set_num(sra
, NULL
, name
, n
);
699 sysfs_set_str(sra
, NULL
, "safe_mode_delay", safe
);
703 int start_reshape(struct mdinfo
*sra
, int already_running
,
704 int before_data_disks
, int data_disks
)
707 unsigned long long sync_max_to_set
;
709 sysfs_set_num(sra
, NULL
, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL
);
710 err
= sysfs_set_num(sra
, NULL
, "suspend_hi", sra
->reshape_progress
);
711 err
= err
?: sysfs_set_num(sra
, NULL
, "suspend_lo",
712 sra
->reshape_progress
);
713 if (before_data_disks
<= data_disks
)
714 sync_max_to_set
= sra
->reshape_progress
/ data_disks
;
716 sync_max_to_set
= (sra
->component_size
* data_disks
717 - sra
->reshape_progress
) / data_disks
;
718 if (!already_running
)
719 sysfs_set_num(sra
, NULL
, "sync_min", sync_max_to_set
);
720 err
= err
?: sysfs_set_num(sra
, NULL
, "sync_max", sync_max_to_set
);
721 if (!already_running
)
722 err
= err
?: sysfs_set_str(sra
, NULL
, "sync_action", "reshape");
727 void abort_reshape(struct mdinfo
*sra
)
729 sysfs_set_str(sra
, NULL
, "sync_action", "idle");
730 sysfs_set_num(sra
, NULL
, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL
);
731 sysfs_set_num(sra
, NULL
, "suspend_hi", 0);
732 sysfs_set_num(sra
, NULL
, "suspend_lo", 0);
733 sysfs_set_num(sra
, NULL
, "sync_min", 0);
734 // It isn't safe to reset sync_max as we aren't monitoring.
735 // Array really should be stopped at this point.
738 int remove_disks_for_takeover(struct supertype
*st
,
743 struct mdinfo
*remaining
;
746 if (sra
->array
.level
== 10)
747 nr_of_copies
= layout
& 0xff;
748 else if (sra
->array
.level
== 1)
749 nr_of_copies
= sra
->array
.raid_disks
;
753 remaining
= sra
->devs
;
755 /* for each 'copy', select one device and remove from the list. */
756 for (slot
= 0; slot
< sra
->array
.raid_disks
; slot
+= nr_of_copies
) {
757 struct mdinfo
**diskp
;
760 /* Find a working device to keep */
761 for (diskp
= &remaining
; *diskp
; diskp
= &(*diskp
)->next
) {
762 struct mdinfo
*disk
= *diskp
;
764 if (disk
->disk
.raid_disk
< slot
)
766 if (disk
->disk
.raid_disk
>= slot
+ nr_of_copies
)
768 if (disk
->disk
.state
& (1<<MD_DISK_REMOVED
))
770 if (disk
->disk
.state
& (1<<MD_DISK_FAULTY
))
772 if (!(disk
->disk
.state
& (1<<MD_DISK_SYNC
)))
775 /* We have found a good disk to use! */
777 disk
->next
= sra
->devs
;
786 if (slot
< sra
->array
.raid_disks
) {
787 /* didn't find all slots */
793 sra
->devs
= remaining
;
797 /* Remove all 'remaining' devices from the array */
799 struct mdinfo
*sd
= remaining
;
800 remaining
= sd
->next
;
802 sysfs_set_str(sra
, sd
, "state", "faulty");
803 sysfs_set_str(sra
, sd
, "slot", "none");
804 /* for external metadata disks should be removed in mdmon */
805 if (!st
->ss
->external
)
806 sysfs_set_str(sra
, sd
, "state", "remove");
807 sd
->disk
.state
|= (1<<MD_DISK_REMOVED
);
808 sd
->disk
.state
&= ~(1<<MD_DISK_SYNC
);
809 sd
->next
= sra
->devs
;
815 void reshape_free_fdlist(int *fdlist
,
816 unsigned long long *offsets
,
821 for (i
= 0; i
< size
; i
++)
829 int reshape_prepare_fdlist(char *devname
,
833 unsigned long blocks
,
836 unsigned long long *offsets
)
842 for (d
= 0; d
<= nrdisks
; d
++)
845 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
846 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
848 if (sd
->disk
.state
& (1<<MD_DISK_SYNC
)) {
849 char *dn
= map_dev(sd
->disk
.major
,
851 fdlist
[sd
->disk
.raid_disk
]
852 = dev_open(dn
, O_RDONLY
);
853 offsets
[sd
->disk
.raid_disk
] = sd
->data_offset
*512;
854 if (fdlist
[sd
->disk
.raid_disk
] < 0) {
855 pr_err("%s: cannot open component %s\n",
856 devname
, dn
? dn
: "-unknown-");
860 } else if (backup_file
== NULL
) {
862 char *dn
= map_dev(sd
->disk
.major
,
864 fdlist
[d
] = dev_open(dn
, O_RDWR
);
865 offsets
[d
] = (sd
->data_offset
+ sra
->component_size
- blocks
- 8)*512;
867 pr_err("%s: cannot open component %s\n",
868 devname
, dn
? dn
: "-unknown-");
879 int reshape_open_backup_file(char *backup_file
,
884 unsigned long long *offsets
,
888 /* Return 1 on success, 0 on any form of failure */
889 /* need to check backup file is large enough */
895 *fdlist
= open(backup_file
, O_RDWR
|O_CREAT
|(restart
? O_TRUNC
: O_EXCL
),
899 pr_err("%s: cannot create backup file %s: %s\n",
900 devname
, backup_file
, strerror(errno
));
903 /* Guard against backup file being on array device.
904 * If array is partitioned or if LVM etc is in the
905 * way this will not notice, but it is better than
908 fstat(*fdlist
, &stb
);
911 if (stb
.st_rdev
== dev
) {
912 pr_err("backup file must NOT be on the array being reshaped.\n");
918 for (i
=0; i
< blocks
+ 8 ; i
++) {
919 if (write(*fdlist
, buf
, 512) != 512) {
920 pr_err("%s: cannot create backup file %s: %s\n",
921 devname
, backup_file
, strerror(errno
));
925 if (fsync(*fdlist
) != 0) {
926 pr_err("%s: cannot create backup file %s: %s\n",
927 devname
, backup_file
, strerror(errno
));
931 if (!restart
&& strncmp(backup_file
, MAP_DIR
, strlen(MAP_DIR
)) != 0) {
932 char *bu
= make_backup(sys_name
);
933 if (symlink(backup_file
, bu
))
934 pr_err("Recording backup file in " MAP_DIR
" failed: %s\n",
942 unsigned long compute_backup_blocks(int nchunk
, int ochunk
,
943 unsigned int ndata
, unsigned int odata
)
945 unsigned long a
, b
, blocks
;
946 /* So how much do we need to backup.
947 * We need an amount of data which is both a whole number of
948 * old stripes and a whole number of new stripes.
949 * So LCM for (chunksize*datadisks).
951 a
= (ochunk
/512) * odata
;
952 b
= (nchunk
/512) * ndata
;
955 /* LCM == product / GCD */
956 blocks
= (ochunk
/512) * (nchunk
/512) * odata
* ndata
/ a
;
961 char *analyse_change(char *devname
, struct mdinfo
*info
, struct reshape
*re
)
963 /* Based on the current array state in info->array and
964 * the changes in info->new_* etc, determine:
965 * - whether the change is possible
966 * - Intermediate level/raid_disks/layout
967 * - whether a restriping reshape is needed
968 * - number of sectors in minimum change unit. This
969 * will cover a whole number of stripes in 'before' and
972 * Return message if the change should be rejected
973 * NULL if the change can be achieved
975 * This can be called as part of starting a reshape, or
976 * when assembling an array that is undergoing reshape.
978 int near
, far
, offset
, copies
;
980 int old_chunk
, new_chunk
;
981 /* delta_parity records change in number of devices
982 * caused by level change
984 int delta_parity
= 0;
986 memset(re
, 0, sizeof(*re
));
988 /* If a new level not explicitly given, we assume no-change */
989 if (info
->new_level
== UnSet
)
990 info
->new_level
= info
->array
.level
;
993 switch (info
->new_level
) {
999 /* chunk size is meaningful, must divide component_size
1002 if (info
->component_size
% (info
->new_chunk
/512)) {
1003 unsigned long long shrink
= info
->component_size
;
1004 shrink
&= ~(unsigned long long)(info
->new_chunk
/512-1);
1005 pr_err("New chunk size (%dK) does not evenly divide device size (%lluk)\n",
1006 info
->new_chunk
/1024, info
->component_size
/2);
1007 pr_err("After shrinking any filesystem, \"mdadm --grow %s --size %llu\"\n",
1009 pr_err("will shrink the array so the given chunk size would work.\n");
1014 return "chunk size not meaningful for this level";
1017 info
->new_chunk
= info
->array
.chunk_size
;
1019 switch (info
->array
.level
) {
1021 return "No reshape is possibly for this RAID level";
1023 if (info
->delta_disks
!= UnSet
)
1024 return "Only --add is supported for LINEAR, setting --raid-disks is not needed";
1026 return "Only --add is supported for LINEAR, other --grow options are not meaningful";
1028 /* RAID1 can convert to RAID1 with different disks, or
1029 * raid5 with 2 disks, or
1032 if (info
->new_level
> 1 &&
1033 (info
->component_size
& 7))
1034 return "Cannot convert RAID1 of this size - reduce size to multiple of 4K first.";
1035 if (info
->new_level
== 0) {
1036 if (info
->delta_disks
!= UnSet
&&
1037 info
->delta_disks
!= 0)
1038 return "Cannot change number of disks with RAID1->RAID0 conversion";
1040 re
->before
.data_disks
= 1;
1041 re
->after
.data_disks
= 1;
1044 if (info
->new_level
== 1) {
1045 if (info
->delta_disks
== UnSet
)
1046 /* Don't know what to do */
1047 return "no change requested for Growing RAID1";
1051 if (info
->array
.raid_disks
== 2 &&
1052 info
->new_level
== 5) {
1055 re
->before
.data_disks
= 1;
1056 if (info
->delta_disks
!= UnSet
&&
1057 info
->delta_disks
!= 0)
1058 re
->after
.data_disks
= 1 + info
->delta_disks
;
1060 re
->after
.data_disks
= 1;
1061 if (re
->after
.data_disks
< 1)
1062 return "Number of disks too small for RAID5";
1064 re
->before
.layout
= ALGORITHM_LEFT_SYMMETRIC
;
1065 info
->array
.chunk_size
= 65536;
1068 /* Could do some multi-stage conversions, but leave that to
1071 return "Impossibly level change request for RAID1";
1074 /* RAID10 can be converted from near mode to
1075 * RAID0 by removing some devices.
1076 * It can also be reshaped if the kernel supports
1079 switch (info
->new_level
) {
1081 if ((info
->array
.layout
& ~0xff) != 0x100)
1082 return "Cannot Grow RAID10 with far/offset layout";
1083 /* number of devices must be multiple of number of copies */
1084 if (info
->array
.raid_disks
% (info
->array
.layout
& 0xff))
1085 return "RAID10 layout too complex for Grow operation";
1087 new_disks
= (info
->array
.raid_disks
1088 / (info
->array
.layout
& 0xff));
1089 if (info
->delta_disks
== UnSet
)
1090 info
->delta_disks
= (new_disks
1091 - info
->array
.raid_disks
);
1093 if (info
->delta_disks
!= new_disks
- info
->array
.raid_disks
)
1094 return "New number of raid-devices impossible for RAID10";
1095 if (info
->new_chunk
&&
1096 info
->new_chunk
!= info
->array
.chunk_size
)
1097 return "Cannot change chunk-size with RAID10 Grow";
1101 re
->before
.data_disks
= new_disks
;
1102 re
->after
.data_disks
= re
->before
.data_disks
;
1106 near
= info
->array
.layout
& 0xff;
1107 far
= (info
->array
.layout
>> 8) & 0xff;
1108 offset
= info
->array
.layout
& 0x10000;
1109 if (far
> 1 && !offset
)
1110 return "Cannot reshape RAID10 in far-mode";
1111 copies
= near
* far
;
1113 old_chunk
= info
->array
.chunk_size
* far
;
1115 if (info
->new_layout
== UnSet
)
1116 info
->new_layout
= info
->array
.layout
;
1118 near
= info
->new_layout
& 0xff;
1119 far
= (info
->new_layout
>> 8) & 0xff;
1120 offset
= info
->new_layout
& 0x10000;
1121 if (far
> 1 && !offset
)
1122 return "Cannot reshape RAID10 to far-mode";
1123 if (near
* far
!= copies
)
1124 return "Cannot change number of copies when reshaping RAID10";
1126 if (info
->delta_disks
== UnSet
)
1127 info
->delta_disks
= 0;
1128 new_disks
= (info
->array
.raid_disks
+
1131 new_chunk
= info
->new_chunk
* far
;
1134 re
->before
.layout
= info
->array
.layout
;
1135 re
->before
.data_disks
= info
->array
.raid_disks
;
1136 re
->after
.layout
= info
->new_layout
;
1137 re
->after
.data_disks
= new_disks
;
1138 /* For RAID10 we don't do backup but do allow reshape,
1139 * so set backup_blocks to INVALID_SECTORS rather than
1141 * And there is no need to synchronise stripes on both
1142 * 'old' and 'new'. So the important
1143 * number is the minimum data_offset difference
1144 * which is the larger of (offset copies * chunk).
1146 re
->backup_blocks
= INVALID_SECTORS
;
1147 re
->min_offset_change
= max(old_chunk
, new_chunk
) / 512;
1148 if (new_disks
< re
->before
.data_disks
&&
1149 info
->space_after
< re
->min_offset_change
)
1150 /* Reduce component size by one chunk */
1151 re
->new_size
= (info
->component_size
-
1152 re
->min_offset_change
);
1154 re
->new_size
= info
->component_size
;
1155 re
->new_size
= re
->new_size
* new_disks
/ copies
;
1159 return "RAID10 can only be changed to RAID0";
1162 /* RAID0 can be converted to RAID10, or to RAID456 */
1163 if (info
->new_level
== 10) {
1164 if (info
->new_layout
== UnSet
&& info
->delta_disks
== UnSet
) {
1165 /* Assume near=2 layout */
1166 info
->new_layout
= 0x102;
1167 info
->delta_disks
= info
->array
.raid_disks
;
1169 if (info
->new_layout
== UnSet
) {
1170 int copies
= 1 + (info
->delta_disks
1171 / info
->array
.raid_disks
);
1172 if (info
->array
.raid_disks
* (copies
-1)
1173 != info
->delta_disks
)
1174 return "Impossible number of devices for RAID0->RAID10";
1175 info
->new_layout
= 0x100 + copies
;
1177 if (info
->delta_disks
== UnSet
) {
1178 int copies
= info
->new_layout
& 0xff;
1179 if (info
->new_layout
!= 0x100 + copies
)
1180 return "New layout impossible for RAID0->RAID10";;
1181 info
->delta_disks
= (copies
- 1) *
1182 info
->array
.raid_disks
;
1184 if (info
->new_chunk
&&
1185 info
->new_chunk
!= info
->array
.chunk_size
)
1186 return "Cannot change chunk-size with RAID0->RAID10";
1189 re
->before
.data_disks
= (info
->array
.raid_disks
+
1191 re
->after
.data_disks
= re
->before
.data_disks
;
1192 re
->before
.layout
= info
->new_layout
;
1196 /* RAID0 can also covert to RAID0/4/5/6 by first converting to
1197 * a raid4 style layout of the final level.
1199 switch (info
->new_level
) {
1204 re
->before
.layout
= 0;
1209 re
->before
.layout
= ALGORITHM_PARITY_N
;
1210 if (info
->new_layout
== UnSet
)
1211 info
->new_layout
= map_name(r5layout
, "default");
1216 re
->before
.layout
= ALGORITHM_PARITY_N
;
1217 if (info
->new_layout
== UnSet
)
1218 info
->new_layout
= map_name(r6layout
, "default");
1221 return "Impossible level change requested";
1223 re
->before
.data_disks
= info
->array
.raid_disks
;
1224 /* determining 'after' layout happens outside this 'switch' */
1228 info
->array
.layout
= ALGORITHM_PARITY_N
;
1230 switch (info
->new_level
) {
1234 re
->level
= info
->array
.level
;
1235 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1236 re
->before
.layout
= info
->array
.layout
;
1240 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1241 re
->before
.layout
= info
->array
.layout
;
1246 re
->before
.data_disks
= info
->array
.raid_disks
- 1;
1247 switch (info
->array
.layout
) {
1248 case ALGORITHM_LEFT_ASYMMETRIC
:
1249 re
->before
.layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
1251 case ALGORITHM_RIGHT_ASYMMETRIC
:
1252 re
->before
.layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
1254 case ALGORITHM_LEFT_SYMMETRIC
:
1255 re
->before
.layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
1257 case ALGORITHM_RIGHT_SYMMETRIC
:
1258 re
->before
.layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
1260 case ALGORITHM_PARITY_0
:
1261 re
->before
.layout
= ALGORITHM_PARITY_0_6
;
1263 case ALGORITHM_PARITY_N
:
1264 re
->before
.layout
= ALGORITHM_PARITY_N_6
;
1267 return "Cannot convert an array with this layout";
1271 if (info
->array
.raid_disks
!= 2)
1272 return "Can only convert a 2-device array to RAID1";
1273 if (info
->delta_disks
!= UnSet
&&
1274 info
->delta_disks
!= 0)
1275 return "Cannot set raid_disk when converting RAID5->RAID1";
1277 info
->new_chunk
= 0;
1280 return "Impossible level change requested";
1284 switch (info
->new_level
) {
1290 re
->before
.data_disks
= info
->array
.raid_disks
- 2;
1291 re
->before
.layout
= info
->array
.layout
;
1294 return "Impossible level change requested";
1299 /* If we reached here then it looks like a re-stripe is
1300 * happening. We have determined the intermediate level
1301 * and initial raid_disks/layout and stored these in 're'.
1303 * We need to deduce the final layout that can be atomically
1304 * converted to the end state.
1306 switch (info
->new_level
) {
1308 /* We can only get to RAID0 from RAID4 or RAID5
1309 * with appropriate layout and one extra device
1311 if (re
->level
!= 4 && re
->level
!= 5)
1312 return "Cannot covert to RAID0 from this level";
1314 switch (re
->level
) {
1316 re
->before
.layout
= 0;
1317 re
->after
.layout
= 0;
1320 re
->after
.layout
= ALGORITHM_PARITY_N
;
1326 /* We can only get to RAID4 from RAID5 */
1327 if (re
->level
!= 4 && re
->level
!= 5)
1328 return "Cannot convert to RAID4 from this level";
1330 switch (re
->level
) {
1332 re
->after
.layout
= 0;
1335 re
->after
.layout
= ALGORITHM_PARITY_N
;
1341 /* We get to RAID5 from RAID5 or RAID6 */
1342 if (re
->level
!= 5 && re
->level
!= 6)
1343 return "Cannot convert to RAID5 from this level";
1345 switch (re
->level
) {
1347 if (info
->new_layout
== UnSet
)
1348 re
->after
.layout
= re
->before
.layout
;
1350 re
->after
.layout
= info
->new_layout
;
1353 if (info
->new_layout
== UnSet
)
1354 info
->new_layout
= re
->before
.layout
;
1356 /* after.layout needs to be raid6 version of new_layout */
1357 if (info
->new_layout
== ALGORITHM_PARITY_N
)
1358 re
->after
.layout
= ALGORITHM_PARITY_N
;
1361 char *ls
= map_num(r5layout
, info
->new_layout
);
1364 /* Current RAID6 layout has a RAID5
1367 strcat(strcpy(layout
, ls
), "-6");
1368 l
= map_name(r6layout
, layout
);
1370 return "Cannot find RAID6 layout to convert to";
1372 /* Current RAID6 has no equivalent.
1373 * If it is already a '-6' layout we
1374 * can leave it unchanged, else we must
1377 ls
= map_num(r6layout
, info
->new_layout
);
1379 strcmp(ls
+strlen(ls
)-2, "-6") != 0)
1380 return "Please specify new layout";
1381 l
= info
->new_layout
;
1383 re
->after
.layout
= l
;
1389 /* We must already be at level 6 */
1391 return "Impossible level change";
1392 if (info
->new_layout
== UnSet
)
1393 re
->after
.layout
= info
->array
.layout
;
1395 re
->after
.layout
= info
->new_layout
;
1398 return "Impossible level change requested";
1400 if (info
->delta_disks
== UnSet
)
1401 info
->delta_disks
= delta_parity
;
1403 re
->after
.data_disks
= (re
->before
.data_disks
1406 switch (re
->level
) {
1407 case 6: re
->parity
= 2;
1410 case 5: re
->parity
= 1;
1412 default: re
->parity
= 0;
1415 /* So we have a restripe operation, we need to calculate the number
1416 * of blocks per reshape operation.
1418 re
->new_size
= info
->component_size
* re
->before
.data_disks
;
1419 if (info
->new_chunk
== 0)
1420 info
->new_chunk
= info
->array
.chunk_size
;
1421 if (re
->after
.data_disks
== re
->before
.data_disks
&&
1422 re
->after
.layout
== re
->before
.layout
&&
1423 info
->new_chunk
== info
->array
.chunk_size
) {
1424 /* Nothing to change, can change level immediately. */
1425 re
->level
= info
->new_level
;
1426 re
->backup_blocks
= 0;
1429 if (re
->after
.data_disks
== 1 && re
->before
.data_disks
== 1) {
1430 /* chunk and layout changes make no difference */
1431 re
->level
= info
->new_level
;
1432 re
->backup_blocks
= 0;
1436 if (re
->after
.data_disks
== re
->before
.data_disks
&&
1437 get_linux_version() < 2006032)
1438 return "in-place reshape is not safe before 2.6.32 - sorry.";
1440 if (re
->after
.data_disks
< re
->before
.data_disks
&&
1441 get_linux_version() < 2006030)
1442 return "reshape to fewer devices is not supported before 2.6.30 - sorry.";
1444 re
->backup_blocks
= compute_backup_blocks(
1445 info
->new_chunk
, info
->array
.chunk_size
,
1446 re
->after
.data_disks
,
1447 re
->before
.data_disks
);
1448 re
->min_offset_change
= re
->backup_blocks
/ re
->before
.data_disks
;
1450 re
->new_size
= info
->component_size
* re
->after
.data_disks
;
1454 static int set_array_size(struct supertype
*st
, struct mdinfo
*sra
,
1457 struct mdinfo
*info
;
1461 if ((st
== NULL
) || (sra
== NULL
))
1464 if (text_version
== NULL
)
1465 text_version
= sra
->text_version
;
1466 subarray
= strchr(text_version
+1, '/')+1;
1467 info
= st
->ss
->container_content(st
, subarray
);
1469 unsigned long long current_size
= 0;
1470 unsigned long long new_size
=
1471 info
->custom_array_size
/2;
1473 if (sysfs_get_ll(sra
, NULL
, "array_size", ¤t_size
) == 0 &&
1474 new_size
> current_size
) {
1475 if (sysfs_set_num(sra
, NULL
, "array_size", new_size
)
1477 dprintf("Error: Cannot set array size");
1480 dprintf("Array size changed");
1482 dprintf_cont(" from %llu to %llu.\n",
1483 current_size
, new_size
);
1487 dprintf("Error: set_array_size(): info pointer in NULL\n");
1492 static int reshape_array(char *container
, int fd
, char *devname
,
1493 struct supertype
*st
, struct mdinfo
*info
,
1494 int force
, struct mddev_dev
*devlist
,
1495 unsigned long long data_offset
,
1496 char *backup_file
, int verbose
, int forked
,
1497 int restart
, int freeze_reshape
);
1498 static int reshape_container(char *container
, char *devname
,
1500 struct supertype
*st
,
1501 struct mdinfo
*info
,
1503 char *backup_file
, int verbose
,
1504 int forked
, int restart
, int freeze_reshape
);
1506 int Grow_reshape(char *devname
, int fd
,
1507 struct mddev_dev
*devlist
,
1508 unsigned long long data_offset
,
1509 struct context
*c
, struct shape
*s
)
1511 /* Make some changes in the shape of an array.
1512 * The kernel must support the change.
1514 * There are three different changes. Each can trigger
1515 * a resync or recovery so we freeze that until we have
1516 * requested everything (if kernel supports freezing - 2.6.30).
1518 * - change size (i.e. component_size)
1520 * - change layout/chunksize/ndisks
1522 * The last can require a reshape. It is different on different
1523 * levels so we need to check the level before actioning it.
1524 * Some times the level change needs to be requested after the
1525 * reshape (e.g. raid6->raid5, raid5->raid0)
1528 struct mdu_array_info_s array
;
1530 struct supertype
*st
;
1531 char *subarray
= NULL
;
1535 char *container
= NULL
;
1538 struct mddev_dev
*dv
;
1544 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) < 0) {
1545 pr_err("%s is not an active md array - aborting\n",
1549 if (data_offset
!= INVALID_SECTORS
&& array
.level
!= 10
1550 && (array
.level
< 4 || array
.level
> 6)) {
1551 pr_err("--grow --data-offset not yet supported\n");
1556 (s
->chunk
|| s
->level
!= UnSet
|| s
->layout_str
|| s
->raiddisks
)) {
1557 pr_err("cannot change component size at the same time as other changes.\n"
1558 " Change size first, then check data is intact before making other changes.\n");
1562 if (s
->raiddisks
&& s
->raiddisks
< array
.raid_disks
&& array
.level
> 1 &&
1563 get_linux_version() < 2006032 &&
1564 !check_env("MDADM_FORCE_FEWER")) {
1565 pr_err("reducing the number of devices is not safe before Linux 2.6.32\n"
1566 " Please use a newer kernel\n");
1570 st
= super_by_fd(fd
, &subarray
);
1572 pr_err("Unable to determine metadata format for %s\n", devname
);
1575 if (s
->raiddisks
> st
->max_devs
) {
1576 pr_err("Cannot increase raid-disks on this array beyond %d\n", st
->max_devs
);
1580 /* in the external case we need to check that the requested reshape is
1581 * supported, and perform an initial check that the container holds the
1582 * pre-requisite spare devices (mdmon owns final validation)
1584 if (st
->ss
->external
) {
1588 container
= st
->container_devnm
;
1589 cfd
= open_dev_excl(st
->container_devnm
);
1591 container
= st
->devnm
;
1593 cfd
= open_dev_excl(st
->devnm
);
1597 pr_err("Unable to open container for %s\n",
1603 rv
= st
->ss
->load_container(st
, cfd
, NULL
);
1606 pr_err("Cannot read superblock for %s\n",
1612 /* check if operation is supported for metadata handler */
1613 if (st
->ss
->container_content
) {
1614 struct mdinfo
*cc
= NULL
;
1615 struct mdinfo
*content
= NULL
;
1617 cc
= st
->ss
->container_content(st
, subarray
);
1618 for (content
= cc
; content
; content
= content
->next
) {
1619 int allow_reshape
= 1;
1621 /* check if reshape is allowed based on metadata
1622 * indications stored in content.array.status
1624 if (content
->array
.state
& (1<<MD_SB_BLOCK_VOLUME
))
1626 if (content
->array
.state
1627 & (1<<MD_SB_BLOCK_CONTAINER_RESHAPE
))
1629 if (!allow_reshape
) {
1630 pr_err("cannot reshape arrays in container with unsupported metadata: %s(%s)\n",
1631 devname
, container
);
1639 if (mdmon_running(container
))
1640 st
->update_tail
= &st
->updates
;
1644 for (dv
= devlist
; dv
; dv
= dv
->next
)
1646 if (s
->raiddisks
> array
.raid_disks
&&
1647 array
.spare_disks
+added_disks
< (s
->raiddisks
- array
.raid_disks
) &&
1649 pr_err("Need %d spare%s to avoid degraded array, and only have %d.\n"
1650 " Use --force to over-ride this check.\n",
1651 s
->raiddisks
- array
.raid_disks
,
1652 s
->raiddisks
- array
.raid_disks
== 1 ? "" : "s",
1653 array
.spare_disks
+ added_disks
);
1657 sra
= sysfs_read(fd
, NULL
, GET_LEVEL
| GET_DISKS
| GET_DEVS
1658 | GET_STATE
| GET_VERSION
);
1660 if (st
->ss
->external
&& subarray
== NULL
) {
1661 array
.level
= LEVEL_CONTAINER
;
1662 sra
->array
.level
= LEVEL_CONTAINER
;
1665 pr_err("failed to read sysfs parameters for %s\n",
1669 frozen
= freeze(st
);
1671 /* freeze() already spewed the reason */
1674 } else if (frozen
< 0) {
1675 pr_err("%s is performing resync/recovery and cannot be reshaped\n", devname
);
1680 /* ========= set size =============== */
1681 if (s
->size
> 0 && (s
->size
== MAX_SIZE
|| s
->size
!= (unsigned)array
.size
)) {
1682 unsigned long long orig_size
= get_component_size(fd
)/2;
1683 unsigned long long min_csize
;
1685 int raid0_takeover
= 0;
1688 orig_size
= (unsigned) array
.size
;
1690 if (orig_size
== 0) {
1691 pr_err("Cannot set device size in this type of array.\n");
1696 if (reshape_super(st
, s
->size
, UnSet
, UnSet
, 0, 0, UnSet
, NULL
,
1697 devname
, APPLY_METADATA_CHANGES
, c
->verbose
> 0)) {
1702 if (st
->ss
->external
) {
1703 /* metadata can have size limitation
1704 * update size value according to metadata information
1706 struct mdinfo
*sizeinfo
=
1707 st
->ss
->container_content(st
, subarray
);
1709 unsigned long long new_size
=
1710 sizeinfo
->custom_array_size
/2;
1711 int data_disks
= get_data_disks(
1712 sizeinfo
->array
.level
,
1713 sizeinfo
->array
.layout
,
1714 sizeinfo
->array
.raid_disks
);
1715 new_size
/= data_disks
;
1716 dprintf("Metadata size correction from %llu to %llu (%llu)\n", orig_size
, new_size
,
1717 new_size
* data_disks
);
1719 sysfs_free(sizeinfo
);
1723 /* Update the size of each member device in case
1724 * they have been resized. This will never reduce
1725 * below the current used-size. The "size" attribute
1726 * understands '0' to mean 'max'.
1730 for (mdi
= sra
->devs
; mdi
; mdi
= mdi
->next
) {
1731 if (sysfs_set_num(sra
, mdi
, "size",
1732 s
->size
== MAX_SIZE
? 0 : s
->size
) < 0) {
1733 /* Probably kernel refusing to let us
1734 * reduce the size - not an error.
1738 if (array
.not_persistent
== 0 &&
1739 array
.major_version
== 0 &&
1740 get_linux_version() < 3001000) {
1741 /* Dangerous to allow size to exceed 2TB */
1742 unsigned long long csize
;
1743 if (sysfs_get_ll(sra
, mdi
, "size", &csize
) == 0) {
1744 if (csize
>= 2ULL*1024*1024*1024)
1745 csize
= 2ULL*1024*1024*1024;
1746 if ((min_csize
== 0 || (min_csize
1753 pr_err("Cannot set size on array members.\n");
1754 goto size_change_error
;
1756 if (min_csize
&& s
->size
> min_csize
) {
1757 pr_err("Cannot safely make this array use more than 2TB per device on this kernel.\n");
1759 goto size_change_error
;
1761 if (min_csize
&& s
->size
== MAX_SIZE
) {
1762 /* Don't let the kernel choose a size - it will get
1765 pr_err("Limited v0.90 array to 2TB per device\n");
1766 s
->size
= min_csize
;
1768 if (st
->ss
->external
) {
1769 if (sra
->array
.level
== 0) {
1770 rv
= sysfs_set_str(sra
, NULL
, "level",
1774 /* get array parametes after takeover
1775 * to chane one parameter at time only
1777 rv
= ioctl(fd
, GET_ARRAY_INFO
, &array
);
1780 /* make sure mdmon is
1781 * aware of the new level */
1782 if (!mdmon_running(st
->container_devnm
))
1783 start_mdmon(st
->container_devnm
);
1784 ping_monitor(container
);
1785 if (mdmon_running(st
->container_devnm
) &&
1786 st
->update_tail
== NULL
)
1787 st
->update_tail
= &st
->updates
;
1790 if (s
->size
== MAX_SIZE
)
1792 array
.size
= s
->size
;
1793 if (s
->size
& ~INT32_MAX
) {
1794 /* got truncated to 32bit, write to
1795 * component_size instead
1798 rv
= sysfs_set_num(sra
, NULL
,
1799 "component_size", s
->size
);
1803 rv
= ioctl(fd
, SET_ARRAY_INFO
, &array
);
1805 /* manage array size when it is managed externally
1807 if ((rv
== 0) && st
->ss
->external
)
1808 rv
= set_array_size(st
, sra
, sra
->text_version
);
1811 if (raid0_takeover
) {
1812 /* do not recync non-existing parity,
1813 * we will drop it anyway
1815 sysfs_set_str(sra
, NULL
, "sync_action", "frozen");
1816 /* go back to raid0, drop parity disk
1818 sysfs_set_str(sra
, NULL
, "level", "raid0");
1819 ioctl(fd
, GET_ARRAY_INFO
, &array
);
1826 /* restore metadata */
1827 if (reshape_super(st
, orig_size
, UnSet
, UnSet
, 0, 0,
1828 UnSet
, NULL
, devname
,
1829 ROLLBACK_METADATA_CHANGES
,
1832 pr_err("Cannot set device size for %s: %s\n",
1833 devname
, strerror(err
));
1835 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
1836 cont_err("Bitmap must be removed before size can be changed\n");
1840 if (s
->assume_clean
) {
1841 /* This will fail on kernels older than 3.0 unless
1842 * a backport has been arranged.
1845 sysfs_set_str(sra
, NULL
, "resync_start", "none") < 0)
1846 pr_err("--assume-clean not supported with --grow on this kernel\n");
1848 ioctl(fd
, GET_ARRAY_INFO
, &array
);
1849 s
->size
= get_component_size(fd
)/2;
1851 s
->size
= array
.size
;
1852 if (c
->verbose
>= 0) {
1853 if (s
->size
== orig_size
)
1854 pr_err("component size of %s unchanged at %lluK\n",
1857 pr_err("component size of %s has been set to %lluK\n",
1861 } else if (array
.level
!= LEVEL_CONTAINER
) {
1862 s
->size
= get_component_size(fd
)/2;
1864 s
->size
= array
.size
;
1867 /* See if there is anything else to do */
1868 if ((s
->level
== UnSet
|| s
->level
== array
.level
) &&
1869 (s
->layout_str
== NULL
) &&
1870 (s
->chunk
== 0 || s
->chunk
== array
.chunk_size
) &&
1871 data_offset
== INVALID_SECTORS
&&
1872 (s
->raiddisks
== 0 || s
->raiddisks
== array
.raid_disks
)) {
1873 /* Nothing more to do */
1874 if (!changed
&& c
->verbose
>= 0)
1875 pr_err("%s: no change requested\n",
1880 /* ========= check for Raid10/Raid1 -> Raid0 conversion ===============
1881 * current implementation assumes that following conditions must be met:
1884 * - near_copies == 2
1886 if ((s
->level
== 0 && array
.level
== 10 && sra
&&
1887 array
.layout
== ((1 << 8) + 2) && !(array
.raid_disks
& 1)) ||
1888 (s
->level
== 0 && array
.level
== 1 && sra
)) {
1890 err
= remove_disks_for_takeover(st
, sra
, array
.layout
);
1892 dprintf("Array cannot be reshaped\n");
1898 /* Make sure mdmon has seen the device removal
1899 * and updated metadata before we continue with
1903 ping_monitor(container
);
1906 memset(&info
, 0, sizeof(info
));
1908 sysfs_init(&info
, fd
, NULL
);
1909 strcpy(info
.text_version
, sra
->text_version
);
1910 info
.component_size
= s
->size
*2;
1911 info
.new_level
= s
->level
;
1912 info
.new_chunk
= s
->chunk
* 1024;
1913 if (info
.array
.level
== LEVEL_CONTAINER
) {
1914 info
.delta_disks
= UnSet
;
1915 info
.array
.raid_disks
= s
->raiddisks
;
1916 } else if (s
->raiddisks
)
1917 info
.delta_disks
= s
->raiddisks
- info
.array
.raid_disks
;
1919 info
.delta_disks
= UnSet
;
1920 if (s
->layout_str
== NULL
) {
1921 info
.new_layout
= UnSet
;
1922 if (info
.array
.level
== 6 &&
1923 (info
.new_level
== 6 || info
.new_level
== UnSet
) &&
1924 info
.array
.layout
>= 16) {
1925 pr_err("%s has a non-standard layout. If you wish to preserve this\n", devname
);
1926 cont_err("during the reshape, please specify --layout=preserve\n");
1927 cont_err("If you want to change it, specify a layout or use --layout=normalise\n");
1931 } else if (strcmp(s
->layout_str
, "normalise") == 0 ||
1932 strcmp(s
->layout_str
, "normalize") == 0) {
1933 /* If we have a -6 RAID6 layout, remove the '-6'. */
1934 info
.new_layout
= UnSet
;
1935 if (info
.array
.level
== 6 && info
.new_level
== UnSet
) {
1937 strcpy(l
, map_num(r6layout
, info
.array
.layout
));
1938 h
= strrchr(l
, '-');
1939 if (h
&& strcmp(h
, "-6") == 0) {
1941 info
.new_layout
= map_name(r6layout
, l
);
1944 pr_err("%s is only meaningful when reshaping a RAID6 array.\n", s
->layout_str
);
1948 } else if (strcmp(s
->layout_str
, "preserve") == 0) {
1949 /* This means that a non-standard RAID6 layout
1952 * - When reshape a RAID6 (e.g. adding a device)
1953 * which is in a non-standard layout, it is OK
1954 * to preserve that layout.
1955 * - When converting a RAID5 to RAID6, leave it in
1956 * the XXX-6 layout, don't re-layout.
1958 if (info
.array
.level
== 6 && info
.new_level
== UnSet
)
1959 info
.new_layout
= info
.array
.layout
;
1960 else if (info
.array
.level
== 5 && info
.new_level
== 6) {
1962 strcpy(l
, map_num(r5layout
, info
.array
.layout
));
1964 info
.new_layout
= map_name(r6layout
, l
);
1966 pr_err("%s in only meaningful when reshaping to RAID6\n", s
->layout_str
);
1971 int l
= info
.new_level
;
1973 l
= info
.array
.level
;
1976 info
.new_layout
= map_name(r5layout
, s
->layout_str
);
1979 info
.new_layout
= map_name(r6layout
, s
->layout_str
);
1982 info
.new_layout
= parse_layout_10(s
->layout_str
);
1985 info
.new_layout
= parse_layout_faulty(s
->layout_str
);
1988 pr_err("layout not meaningful with this level\n");
1992 if (info
.new_layout
== UnSet
) {
1993 pr_err("layout %s not understood for this level\n",
2000 if (array
.level
== LEVEL_FAULTY
) {
2001 if (s
->level
!= UnSet
&& s
->level
!= array
.level
) {
2002 pr_err("cannot change level of Faulty device\n");
2006 pr_err("cannot set chunksize of Faulty device\n");
2009 if (s
->raiddisks
&& s
->raiddisks
!= 1) {
2010 pr_err("cannot set raid_disks of Faulty device\n");
2013 if (s
->layout_str
) {
2014 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
2015 dprintf("Cannot get array information.\n");
2018 array
.layout
= info
.new_layout
;
2019 if (ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
2020 pr_err("failed to set new layout\n");
2022 } else if (c
->verbose
>= 0)
2023 printf("layout for %s set to %d\n",
2024 devname
, array
.layout
);
2026 } else if (array
.level
== LEVEL_CONTAINER
) {
2027 /* This change is to be applied to every array in the
2028 * container. This is only needed when the metadata imposes
2029 * restraints of the various arrays in the container.
2030 * Currently we only know that IMSM requires all arrays
2031 * to have the same number of devices so changing the
2032 * number of devices (On-Line Capacity Expansion) must be
2033 * performed at the level of the container
2035 rv
= reshape_container(container
, devname
, -1, st
, &info
,
2036 c
->force
, c
->backup_file
, c
->verbose
, 0, 0, 0);
2039 /* get spare devices from external metadata
2041 if (st
->ss
->external
) {
2042 struct mdinfo
*info2
;
2044 info2
= st
->ss
->container_content(st
, subarray
);
2046 info
.array
.spare_disks
=
2047 info2
->array
.spare_disks
;
2052 /* Impose these changes on a single array. First
2053 * check that the metadata is OK with the change. */
2055 if (reshape_super(st
, 0, info
.new_level
,
2056 info
.new_layout
, info
.new_chunk
,
2057 info
.array
.raid_disks
, info
.delta_disks
,
2058 c
->backup_file
, devname
, APPLY_METADATA_CHANGES
,
2064 rv
= reshape_array(container
, fd
, devname
, st
, &info
, c
->force
,
2065 devlist
, data_offset
, c
->backup_file
, c
->verbose
,
2076 /* verify_reshape_position()
2077 * Function checks if reshape position in metadata is not farther
2078 * than position in md.
2080 * 0 : not valid sysfs entry
2081 * it can be caused by not started reshape, it should be started
2082 * by reshape array or raid0 array is before takeover
2083 * -1 : error, reshape position is obviously wrong
2084 * 1 : success, reshape progress correct or updated
2086 static int verify_reshape_position(struct mdinfo
*info
, int level
)
2092 /* read sync_max, failure can mean raid0 array */
2093 rv
= sysfs_get_str(info
, NULL
, "sync_max", buf
, 40);
2097 unsigned long long position
= strtoull(buf
, &ep
, 0);
2099 dprintf("Read sync_max sysfs entry is: %s\n", buf
);
2100 if (!(ep
== buf
|| (*ep
!= 0 && *ep
!= '\n' && *ep
!= ' '))) {
2101 position
*= get_data_disks(level
,
2103 info
->array
.raid_disks
);
2104 if (info
->reshape_progress
< position
) {
2105 dprintf("Corrected reshape progress (%llu) to md position (%llu)\n",
2106 info
->reshape_progress
, position
);
2107 info
->reshape_progress
= position
;
2109 } else if (info
->reshape_progress
> position
) {
2110 pr_err("Fatal error: array reshape was not properly frozen (expected reshape position is %llu, but reshape progress is %llu.\n",
2111 position
, info
->reshape_progress
);
2114 dprintf("Reshape position in md and metadata are the same;");
2118 } else if (rv
== 0) {
2119 /* for valid sysfs entry, 0-length content
2120 * should be indicated as error
2128 static unsigned long long choose_offset(unsigned long long lo
,
2129 unsigned long long hi
,
2130 unsigned long long min
,
2131 unsigned long long max
)
2133 /* Choose a new offset between hi and lo.
2134 * It must be between min and max, but
2135 * we would prefer something near the middle of hi/lo, and also
2136 * prefer to be aligned to a big power of 2.
2138 * So we start with the middle, then for each bit,
2139 * starting at '1' and increasing, if it is set, we either
2140 * add it or subtract it if possible, preferring the option
2141 * which is furthest from the boundary.
2143 * We stop once we get a 1MB alignment. As units are in sectors,
2144 * 1MB = 2*1024 sectors.
2146 unsigned long long choice
= (lo
+ hi
) / 2;
2147 unsigned long long bit
= 1;
2149 for (bit
= 1; bit
< 2*1024; bit
= bit
<< 1) {
2150 unsigned long long bigger
, smaller
;
2151 if (! (bit
& choice
))
2153 bigger
= choice
+ bit
;
2154 smaller
= choice
- bit
;
2155 if (bigger
> max
&& smaller
< min
)
2159 else if (smaller
< min
)
2161 else if (hi
- bigger
> smaller
- lo
)
2169 static int set_new_data_offset(struct mdinfo
*sra
, struct supertype
*st
,
2170 char *devname
, int delta_disks
,
2171 unsigned long long data_offset
,
2172 unsigned long long min
,
2178 unsigned long long before
, after
;
2180 /* Need to find min space before and after so same is used
2183 before
= UINT64_MAX
;
2185 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
2189 struct supertype
*st2
;
2190 struct mdinfo info2
;
2192 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
2194 dn
= map_dev(sd
->disk
.major
, sd
->disk
.minor
, 0);
2195 dfd
= dev_open(dn
, O_RDONLY
);
2197 pr_err("%s: cannot open component %s\n",
2198 devname
, dn
? dn
: "-unknown-");
2201 st2
= dup_super(st
);
2202 rv
= st2
->ss
->load_super(st2
,dfd
, NULL
);
2206 pr_err("%s: cannot get superblock from %s\n",
2210 st2
->ss
->getinfo_super(st2
, &info2
, NULL
);
2211 st2
->ss
->free_super(st2
);
2213 if (info2
.space_before
== 0 &&
2214 info2
.space_after
== 0) {
2215 /* Metadata doesn't support data_offset changes */
2217 pr_err("%s: Metadata version doesn't support data_offset changes\n",
2221 if (before
> info2
.space_before
)
2222 before
= info2
.space_before
;
2223 if (after
> info2
.space_after
)
2224 after
= info2
.space_after
;
2226 if (data_offset
!= INVALID_SECTORS
) {
2228 if (info2
.data_offset
== data_offset
) {
2229 pr_err("%s: already has that data_offset\n",
2233 if (data_offset
< info2
.data_offset
)
2237 } else if ((data_offset
<= info2
.data_offset
&& dir
== 1) ||
2238 (data_offset
>= info2
.data_offset
&& dir
== -1)) {
2239 pr_err("%s: differing data offsets on devices make this --data-offset setting impossible\n",
2245 if (before
== UINT64_MAX
)
2246 /* impossible really, there must be no devices */
2249 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
2250 char *dn
= map_dev(sd
->disk
.major
, sd
->disk
.minor
, 0);
2251 unsigned long long new_data_offset
;
2253 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
2255 if (delta_disks
< 0) {
2256 /* Don't need any space as array is shrinking
2257 * just move data_offset up by min
2259 if (data_offset
== INVALID_SECTORS
)
2260 new_data_offset
= sd
->data_offset
+ min
;
2262 if (data_offset
< sd
->data_offset
+ min
) {
2263 pr_err("--data-offset too small for %s\n",
2267 new_data_offset
= data_offset
;
2269 } else if (delta_disks
> 0) {
2270 /* need space before */
2274 pr_err("Insufficient head-space for reshape on %s\n",
2278 if (data_offset
== INVALID_SECTORS
)
2279 new_data_offset
= sd
->data_offset
- min
;
2281 if (data_offset
> sd
->data_offset
- min
) {
2282 pr_err("--data-offset too large for %s\n",
2286 new_data_offset
= data_offset
;
2290 /* can move up or down. If 'data_offset'
2291 * was set we would have already decided,
2292 * so just choose direction with most space.
2299 sysfs_set_str(sra
, NULL
, "reshape_direction",
2300 dir
== 1 ? "backwards" : "forwards");
2302 /* Increase data offset */
2306 pr_err("Insufficient tail-space for reshape on %s\n",
2310 if (data_offset
!= INVALID_SECTORS
&&
2311 data_offset
< sd
->data_offset
+ min
) {
2312 pr_err("--data-offset too small on %s\n",
2316 if (data_offset
!= INVALID_SECTORS
)
2317 new_data_offset
= data_offset
;
2319 new_data_offset
= choose_offset(sd
->data_offset
,
2320 sd
->data_offset
+ after
,
2321 sd
->data_offset
+ min
,
2322 sd
->data_offset
+ after
);
2324 /* Decrease data offset */
2328 pr_err("insufficient head-room on %s\n",
2332 if (data_offset
!= INVALID_SECTORS
&&
2333 data_offset
< sd
->data_offset
- min
) {
2334 pr_err("--data-offset too small on %s\n",
2338 if (data_offset
!= INVALID_SECTORS
)
2339 new_data_offset
= data_offset
;
2341 new_data_offset
= choose_offset(sd
->data_offset
- before
,
2343 sd
->data_offset
- before
,
2344 sd
->data_offset
- min
);
2347 err
= sysfs_set_num(sra
, sd
, "new_offset", new_data_offset
);
2348 if (err
< 0 && errno
== E2BIG
) {
2349 /* try again after increasing data size to max */
2350 err
= sysfs_set_num(sra
, sd
, "size", 0);
2351 if (err
< 0 && errno
== EINVAL
&&
2352 !(sd
->disk
.state
& (1<<MD_DISK_SYNC
))) {
2353 /* some kernels have a bug where you cannot
2354 * use '0' on spare devices. */
2355 sysfs_set_num(sra
, sd
, "size",
2356 (sra
->component_size
+ after
)/2);
2358 err
= sysfs_set_num(sra
, sd
, "new_offset",
2362 if (errno
== E2BIG
&& data_offset
!= INVALID_SECTORS
) {
2363 pr_err("data-offset is too big for %s\n",
2367 if (sd
== sra
->devs
&&
2368 (errno
== ENOENT
|| errno
== E2BIG
))
2369 /* Early kernel, no 'new_offset' file,
2370 * or kernel doesn't like us.
2371 * For RAID5/6 this is not fatal
2374 pr_err("Cannot set new_offset for %s\n",
2383 /* Just use a backup file */
2387 static int raid10_reshape(char *container
, int fd
, char *devname
,
2388 struct supertype
*st
, struct mdinfo
*info
,
2389 struct reshape
*reshape
,
2390 unsigned long long data_offset
,
2391 int force
, int verbose
)
2393 /* Changing raid_disks, layout, chunksize or possibly
2394 * just data_offset for a RAID10.
2395 * We must always change data_offset. We change by at least
2396 * ->min_offset_change which is the largest of the old and new
2398 * If raid_disks is increasing, then data_offset must decrease
2399 * by at least this copy size.
2400 * If raid_disks is unchanged, data_offset must increase or
2401 * decrease by at least min_offset_change but preferably by much more.
2402 * We choose half of the available space.
2403 * If raid_disks is decreasing, data_offset must increase by
2404 * at least min_offset_change. To allow of this, component_size
2405 * must be decreased by the same amount.
2407 * So we calculate the required minimum and direction, possibly
2408 * reduce the component_size, then iterate through the devices
2409 * and set the new_data_offset.
2410 * If that all works, we set chunk_size, layout, raid_disks, and start
2414 unsigned long long min
;
2417 sra
= sysfs_read(fd
, NULL
,
2418 GET_COMPONENT
|GET_DEVS
|GET_OFFSET
|GET_STATE
|GET_CHUNK
2421 pr_err("%s: Cannot get array details from sysfs\n",
2425 min
= reshape
->min_offset_change
;
2427 if (info
->delta_disks
)
2428 sysfs_set_str(sra
, NULL
, "reshape_direction",
2429 info
->delta_disks
< 0 ? "backwards" : "forwards");
2430 if (info
->delta_disks
< 0 &&
2431 info
->space_after
< min
) {
2432 int rv
= sysfs_set_num(sra
, NULL
, "component_size",
2433 (sra
->component_size
-
2436 pr_err("cannot reduce component size\n");
2440 err
= set_new_data_offset(sra
, st
, devname
, info
->delta_disks
, data_offset
,
2443 pr_err("Cannot set new_data_offset: RAID10 reshape not\n");
2444 cont_err("supported on this kernel\n");
2450 if (!err
&& sysfs_set_num(sra
, NULL
, "chunk_size", info
->new_chunk
) < 0)
2452 if (!err
&& sysfs_set_num(sra
, NULL
, "layout", reshape
->after
.layout
) < 0)
2454 if (!err
&& sysfs_set_num(sra
, NULL
, "raid_disks",
2455 info
->array
.raid_disks
+ info
->delta_disks
) < 0)
2457 if (!err
&& sysfs_set_str(sra
, NULL
, "sync_action", "reshape") < 0)
2460 pr_err("Cannot set array shape for %s\n",
2463 (info
->array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
2464 cont_err(" Bitmap must be removed before shape can be changed\n");
2474 static void get_space_after(int fd
, struct supertype
*st
, struct mdinfo
*info
)
2476 struct mdinfo
*sra
, *sd
;
2477 /* Initialisation to silence compiler warning */
2478 unsigned long long min_space_before
= 0, min_space_after
= 0;
2481 sra
= sysfs_read(fd
, NULL
, GET_DEVS
);
2484 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
2487 struct supertype
*st2
;
2488 struct mdinfo info2
;
2490 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
2492 dn
= map_dev(sd
->disk
.major
, sd
->disk
.minor
, 0);
2493 dfd
= dev_open(dn
, O_RDONLY
);
2496 st2
= dup_super(st
);
2497 if (st2
->ss
->load_super(st2
,dfd
, NULL
)) {
2503 st2
->ss
->getinfo_super(st2
, &info2
, NULL
);
2504 st2
->ss
->free_super(st2
);
2507 min_space_before
> info2
.space_before
)
2508 min_space_before
= info2
.space_before
;
2510 min_space_after
> info2
.space_after
)
2511 min_space_after
= info2
.space_after
;
2514 if (sd
== NULL
&& !first
) {
2515 info
->space_after
= min_space_after
;
2516 info
->space_before
= min_space_before
;
2521 static void update_cache_size(char *container
, struct mdinfo
*sra
,
2522 struct mdinfo
*info
,
2523 int disks
, unsigned long long blocks
)
2525 /* Check that the internal stripe cache is
2526 * large enough, or it won't work.
2527 * It must hold at least 4 stripes of the larger
2530 unsigned long cache
;
2531 cache
= max(info
->array
.chunk_size
, info
->new_chunk
);
2532 cache
*= 4; /* 4 stripes minimum */
2533 cache
/= 512; /* convert to sectors */
2534 /* make sure there is room for 'blocks' with a bit to spare */
2535 if (cache
< 16 + blocks
/ disks
)
2536 cache
= 16 + blocks
/ disks
;
2537 cache
/= (4096/512); /* Covert from sectors to pages */
2539 if (sra
->cache_size
< cache
)
2540 subarray_set_num(container
, sra
, "stripe_cache_size",
2544 static int impose_reshape(struct mdinfo
*sra
,
2545 struct mdinfo
*info
,
2546 struct supertype
*st
,
2549 char *devname
, char *container
,
2550 struct reshape
*reshape
)
2552 struct mdu_array_info_s array
;
2554 sra
->new_chunk
= info
->new_chunk
;
2557 /* for external metadata checkpoint saved by mdmon can be lost
2558 * or missed /due to e.g. crash/. Check if md is not during
2559 * restart farther than metadata points to.
2560 * If so, this means metadata information is obsolete.
2562 if (st
->ss
->external
)
2563 verify_reshape_position(info
, reshape
->level
);
2564 sra
->reshape_progress
= info
->reshape_progress
;
2566 sra
->reshape_progress
= 0;
2567 if (reshape
->after
.data_disks
< reshape
->before
.data_disks
)
2568 /* start from the end of the new array */
2569 sra
->reshape_progress
= (sra
->component_size
2570 * reshape
->after
.data_disks
);
2573 ioctl(fd
, GET_ARRAY_INFO
, &array
);
2574 if (info
->array
.chunk_size
== info
->new_chunk
&&
2575 reshape
->before
.layout
== reshape
->after
.layout
&&
2576 st
->ss
->external
== 0) {
2577 /* use SET_ARRAY_INFO but only if reshape hasn't started */
2578 array
.raid_disks
= reshape
->after
.data_disks
+ reshape
->parity
;
2580 ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
2583 pr_err("Cannot set device shape for %s: %s\n",
2584 devname
, strerror(errno
));
2587 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
2588 cont_err("Bitmap must be removed before shape can be changed\n");
2592 } else if (!restart
) {
2593 /* set them all just in case some old 'new_*' value
2594 * persists from some earlier problem.
2597 if (sysfs_set_num(sra
, NULL
, "chunk_size", info
->new_chunk
) < 0)
2599 if (!err
&& sysfs_set_num(sra
, NULL
, "layout",
2600 reshape
->after
.layout
) < 0)
2602 if (!err
&& subarray_set_num(container
, sra
, "raid_disks",
2603 reshape
->after
.data_disks
+
2604 reshape
->parity
) < 0)
2607 pr_err("Cannot set device shape for %s\n",
2611 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
2612 cont_err("Bitmap must be removed before shape can be changed\n");
2621 static int impose_level(int fd
, int level
, char *devname
, int verbose
)
2624 struct mdu_array_info_s array
;
2626 sysfs_init(&info
, fd
, NULL
);
2628 ioctl(fd
, GET_ARRAY_INFO
, &array
);
2630 (array
.level
>= 4 && array
.level
<= 6)) {
2631 /* To convert to RAID0 we need to fail and
2632 * remove any non-data devices. */
2635 int data_disks
= array
.raid_disks
- 1;
2636 if (array
.level
== 6)
2638 if (array
.level
== 5 &&
2639 array
.layout
!= ALGORITHM_PARITY_N
)
2641 if (array
.level
== 6 &&
2642 array
.layout
!= ALGORITHM_PARITY_N_6
)
2644 sysfs_set_str(&info
, NULL
,"sync_action", "idle");
2645 /* First remove any spares so no recovery starts */
2646 for (d
= 0, found
= 0;
2647 d
< MAX_DISKS
&& found
< array
.nr_disks
;
2649 mdu_disk_info_t disk
;
2651 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0)
2653 if (disk
.major
== 0 && disk
.minor
== 0)
2656 if ((disk
.state
& (1 << MD_DISK_ACTIVE
))
2657 && disk
.raid_disk
< data_disks
)
2660 ioctl(fd
, HOT_REMOVE_DISK
,
2661 makedev(disk
.major
, disk
.minor
));
2663 /* Now fail anything left */
2664 ioctl(fd
, GET_ARRAY_INFO
, &array
);
2665 for (d
= 0, found
= 0;
2666 d
< MAX_DISKS
&& found
< array
.nr_disks
;
2669 mdu_disk_info_t disk
;
2671 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0)
2673 if (disk
.major
== 0 && disk
.minor
== 0)
2676 if ((disk
.state
& (1 << MD_DISK_ACTIVE
))
2677 && disk
.raid_disk
< data_disks
)
2680 ioctl(fd
, SET_DISK_FAULTY
,
2681 makedev(disk
.major
, disk
.minor
));
2683 while (ioctl(fd
, HOT_REMOVE_DISK
,
2684 makedev(disk
.major
, disk
.minor
)) < 0
2691 c
= map_num(pers
, level
);
2693 int err
= sysfs_set_str(&info
, NULL
, "level", c
);
2696 pr_err("%s: could not set level to %s\n",
2699 (array
.state
& (1<<MD_SB_BITMAP_PRESENT
)))
2700 cont_err("Bitmap must be removed before level can be changed\n");
2704 pr_err("level of %s changed to %s\n",
2711 static void catch_term(int sig
)
2716 static int continue_via_systemd(char *devnm
)
2718 int skipped
, i
, pid
, status
;
2720 /* In a systemd/udev world, it is best to get systemd to
2721 * run "mdadm --grow --continue" rather than running in the
2726 /* FIXME yuk. CLOSE_EXEC?? */
2728 for (i
= 3; skipped
< 20; i
++)
2734 /* Don't want to see error messages from
2735 * systemctl. If the service doesn't exist,
2736 * we fork ourselves.
2739 open("/dev/null", O_WRONLY
);
2740 snprintf(pathbuf
, sizeof(pathbuf
), "mdadm-grow-continue@%s.service",
2742 status
= execl("/usr/bin/systemctl", "systemctl",
2745 status
= execl("/bin/systemctl", "systemctl", "start",
2748 case -1: /* Just do it ourselves. */
2750 default: /* parent - good */
2751 pid
= wait(&status
);
2752 if (pid
>= 0 && status
== 0)
2758 static int reshape_array(char *container
, int fd
, char *devname
,
2759 struct supertype
*st
, struct mdinfo
*info
,
2760 int force
, struct mddev_dev
*devlist
,
2761 unsigned long long data_offset
,
2762 char *backup_file
, int verbose
, int forked
,
2763 int restart
, int freeze_reshape
)
2765 struct reshape reshape
;
2768 int orig_level
= UnSet
;
2772 struct mdu_array_info_s array
;
2775 struct mddev_dev
*dv
;
2779 unsigned long long *offsets
= NULL
;
2783 unsigned long blocks
;
2784 unsigned long long array_size
;
2786 struct mdinfo
*sra
= NULL
;
2789 /* when reshaping a RAID0, the component_size might be zero.
2790 * So try to fix that up.
2792 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
2793 dprintf("Cannot get array information.\n");
2796 if (array
.level
== 0 && info
->component_size
== 0) {
2797 get_dev_size(fd
, NULL
, &array_size
);
2798 info
->component_size
= array_size
/ array
.raid_disks
;
2801 if (array
.level
== 10)
2802 /* Need space_after info */
2803 get_space_after(fd
, st
, info
);
2805 if (info
->reshape_active
) {
2806 int new_level
= info
->new_level
;
2807 info
->new_level
= UnSet
;
2808 if (info
->delta_disks
> 0)
2809 info
->array
.raid_disks
-= info
->delta_disks
;
2810 msg
= analyse_change(devname
, info
, &reshape
);
2811 info
->new_level
= new_level
;
2812 if (info
->delta_disks
> 0)
2813 info
->array
.raid_disks
+= info
->delta_disks
;
2815 /* Make sure the array isn't read-only */
2816 ioctl(fd
, RESTART_ARRAY_RW
, 0);
2818 msg
= analyse_change(devname
, info
, &reshape
);
2820 /* if msg == "", error has already been printed */
2822 pr_err("%s\n", msg
);
2826 (reshape
.level
!= info
->array
.level
||
2827 reshape
.before
.layout
!= info
->array
.layout
||
2828 reshape
.before
.data_disks
+ reshape
.parity
2829 != info
->array
.raid_disks
- max(0, info
->delta_disks
))) {
2830 pr_err("reshape info is not in native format - cannot continue.\n");
2834 if (st
->ss
->external
&& restart
&& (info
->reshape_progress
== 0) &&
2835 !((sysfs_get_str(info
, NULL
, "sync_action", buf
, sizeof(buf
)) > 0) &&
2836 (strncmp(buf
, "reshape", 7) == 0))) {
2837 /* When reshape is restarted from '0', very begin of array
2838 * it is possible that for external metadata reshape and array
2839 * configuration doesn't happen.
2840 * Check if md has the same opinion, and reshape is restarted
2841 * from 0. If so, this is regular reshape start after reshape
2842 * switch in metadata to next array only.
2844 if ((verify_reshape_position(info
, reshape
.level
) >= 0) &&
2845 (info
->reshape_progress
== 0))
2849 /* reshape already started. just skip to monitoring the reshape */
2850 if (reshape
.backup_blocks
== 0)
2852 if (restart
& RESHAPE_NO_BACKUP
)
2855 /* Need 'sra' down at 'started:' */
2856 sra
= sysfs_read(fd
, NULL
,
2857 GET_COMPONENT
|GET_DEVS
|GET_OFFSET
|GET_STATE
|GET_CHUNK
|
2860 pr_err("%s: Cannot get array details from sysfs\n",
2866 backup_file
= locate_backup(sra
->sys_name
);
2870 /* The container is frozen but the array may not be.
2871 * So freeze the array so spares don't get put to the wrong use
2872 * FIXME there should probably be a cleaner separation between
2873 * freeze_array and freeze_container.
2875 sysfs_freeze_array(info
);
2876 /* Check we have enough spares to not be degraded */
2878 for (dv
= devlist
; dv
; dv
=dv
->next
)
2880 spares_needed
= max(reshape
.before
.data_disks
,
2881 reshape
.after
.data_disks
)
2882 + reshape
.parity
- array
.raid_disks
;
2885 info
->new_level
> 1 && info
->array
.level
> 1 &&
2886 spares_needed
> info
->array
.spare_disks
+ added_disks
) {
2887 pr_err("Need %d spare%s to avoid degraded array, and only have %d.\n"
2888 " Use --force to over-ride this check.\n",
2890 spares_needed
== 1 ? "" : "s",
2891 info
->array
.spare_disks
+ added_disks
);
2894 /* Check we have enough spares to not fail */
2895 spares_needed
= max(reshape
.before
.data_disks
,
2896 reshape
.after
.data_disks
)
2898 if ((info
->new_level
> 1 || info
->new_level
== 0) &&
2899 spares_needed
> info
->array
.spare_disks
+added_disks
) {
2900 pr_err("Need %d spare%s to create working array, and only have %d.\n",
2902 spares_needed
== 1 ? "" : "s",
2903 info
->array
.spare_disks
+ added_disks
);
2907 if (reshape
.level
!= array
.level
) {
2908 int err
= impose_level(fd
, reshape
.level
, devname
, verbose
);
2911 info
->new_layout
= UnSet
; /* after level change,
2912 * layout is meaningless */
2913 orig_level
= array
.level
;
2914 sysfs_freeze_array(info
);
2916 if (reshape
.level
> 0 && st
->ss
->external
) {
2917 /* make sure mdmon is aware of the new level */
2918 if (mdmon_running(container
))
2919 flush_mdmon(container
);
2921 if (!mdmon_running(container
))
2922 start_mdmon(container
);
2923 ping_monitor(container
);
2924 if (mdmon_running(container
) &&
2925 st
->update_tail
== NULL
)
2926 st
->update_tail
= &st
->updates
;
2929 /* ->reshape_super might have chosen some spares from the
2930 * container that it wants to be part of the new array.
2931 * We can collect them with ->container_content and give
2932 * them to the kernel.
2934 if (st
->ss
->reshape_super
&& st
->ss
->container_content
) {
2935 char *subarray
= strchr(info
->text_version
+1, '/')+1;
2936 struct mdinfo
*info2
=
2937 st
->ss
->container_content(st
, subarray
);
2941 sysfs_init(info2
, fd
, st
->devnm
);
2942 /* When increasing number of devices, we need to set
2943 * new raid_disks before adding these, or they might
2946 if (reshape
.backup_blocks
&&
2947 reshape
.after
.data_disks
> reshape
.before
.data_disks
)
2948 subarray_set_num(container
, info2
, "raid_disks",
2949 reshape
.after
.data_disks
+
2951 for (d
= info2
->devs
; d
; d
= d
->next
) {
2952 if (d
->disk
.state
== 0 &&
2953 d
->disk
.raid_disk
>= 0) {
2954 /* This is a spare that wants to
2955 * be part of the array.
2957 add_disk(fd
, st
, info2
, d
);
2963 /* We might have been given some devices to add to the
2964 * array. Now that the array has been changed to the right
2965 * level and frozen, we can safely add them.
2968 Manage_subdevs(devname
, fd
, devlist
, verbose
,
2971 if (reshape
.backup_blocks
== 0 && data_offset
!= INVALID_SECTORS
)
2972 reshape
.backup_blocks
= reshape
.before
.data_disks
* info
->array
.chunk_size
/512;
2973 if (reshape
.backup_blocks
== 0) {
2974 /* No restriping needed, but we might need to impose
2975 * some more changes: layout, raid_disks, chunk_size
2977 /* read current array info */
2978 if (ioctl(fd
, GET_ARRAY_INFO
, &array
) != 0) {
2979 dprintf("Cannot get array information.\n");
2982 /* compare current array info with new values and if
2983 * it is different update them to new */
2984 if (info
->new_layout
!= UnSet
&&
2985 info
->new_layout
!= array
.layout
) {
2986 array
.layout
= info
->new_layout
;
2987 if (ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
2988 pr_err("failed to set new layout\n");
2990 } else if (verbose
>= 0)
2991 printf("layout for %s set to %d\n",
2992 devname
, array
.layout
);
2994 if (info
->delta_disks
!= UnSet
&&
2995 info
->delta_disks
!= 0 &&
2996 array
.raid_disks
!= (info
->array
.raid_disks
+ info
->delta_disks
)) {
2997 array
.raid_disks
+= info
->delta_disks
;
2998 if (ioctl(fd
, SET_ARRAY_INFO
, &array
) != 0) {
2999 pr_err("failed to set raid disks\n");
3001 } else if (verbose
>= 0) {
3002 printf("raid_disks for %s set to %d\n",
3003 devname
, array
.raid_disks
);
3006 if (info
->new_chunk
!= 0 &&
3007 info
->new_chunk
!= array
.chunk_size
) {
3008 if (sysfs_set_num(info
, NULL
,
3009 "chunk_size", info
->new_chunk
) != 0) {
3010 pr_err("failed to set chunk size\n");
3012 } else if (verbose
>= 0)
3013 printf("chunk size for %s set to %d\n",
3014 devname
, array
.chunk_size
);
3021 * There are three possibilities.
3022 * 1/ The array will shrink.
3023 * We need to ensure the reshape will pause before reaching
3024 * the 'critical section'. We also need to fork and wait for
3025 * that to happen. When it does we
3026 * suspend/backup/complete/unfreeze
3028 * 2/ The array will not change size.
3029 * This requires that we keep a backup of a sliding window
3030 * so that we can restore data after a crash. So we need
3031 * to fork and monitor progress.
3032 * In future we will allow the data_offset to change, so
3033 * a sliding backup becomes unnecessary.
3035 * 3/ The array will grow. This is relatively easy.
3036 * However the kernel's restripe routines will cheerfully
3037 * overwrite some early data before it is safe. So we
3038 * need to make a backup of the early parts of the array
3039 * and be ready to restore it if rebuild aborts very early.
3040 * For externally managed metadata, we still need a forked
3041 * child to monitor the reshape and suspend IO over the region
3042 * that is being reshaped.
3044 * We backup data by writing it to one spare, or to a
3045 * file which was given on command line.
3047 * In each case, we first make sure that storage is available
3048 * for the required backup.
3050 * - request the shape change.
3051 * - fork to handle backup etc.
3053 /* Check that we can hold all the data */
3054 get_dev_size(fd
, NULL
, &array_size
);
3055 if (reshape
.new_size
< (array_size
/512)) {
3056 pr_err("this change will reduce the size of the array.\n"
3057 " use --grow --array-size first to truncate array.\n"
3058 " e.g. mdadm --grow %s --array-size %llu\n",
3059 devname
, reshape
.new_size
/2);
3063 if (array
.level
== 10) {
3064 /* Reshaping RAID10 does not require any data backup by
3065 * user-space. Instead it requires that the data_offset
3066 * is changed to avoid the need for backup.
3067 * So this is handled very separately
3070 /* Nothing to do. */
3072 return raid10_reshape(container
, fd
, devname
, st
, info
,
3073 &reshape
, data_offset
,
3076 sra
= sysfs_read(fd
, NULL
,
3077 GET_COMPONENT
|GET_DEVS
|GET_OFFSET
|GET_STATE
|GET_CHUNK
|
3080 pr_err("%s: Cannot get array details from sysfs\n",
3086 switch(set_new_data_offset(sra
, st
, devname
,
3087 reshape
.after
.data_disks
- reshape
.before
.data_disks
,
3089 reshape
.min_offset_change
, 1)) {
3093 /* Updated data_offset, so it's easy now */
3094 update_cache_size(container
, sra
, info
,
3095 min(reshape
.before
.data_disks
,
3096 reshape
.after
.data_disks
),
3097 reshape
.backup_blocks
);
3099 /* Right, everything seems fine. Let's kick things off.
3103 if (impose_reshape(sra
, info
, st
, fd
, restart
,
3104 devname
, container
, &reshape
) < 0)
3106 if (sysfs_set_str(sra
, NULL
, "sync_action", "reshape") < 0) {
3108 if (errno
!= EINVAL
) {
3109 pr_err("Failed to initiate reshape!\n");
3112 /* revert data_offset and try the old way */
3113 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
3114 sysfs_set_num(sra
, sd
, "new_offset",
3116 sysfs_set_str(sra
, NULL
, "reshape_direction",
3121 if (info
->new_level
== reshape
.level
)
3123 /* need to adjust level when reshape completes */
3125 case -1: /* ignore error, but don't wait */
3127 default: /* parent */
3135 fd
= open_dev(sra
->sys_name
);
3137 impose_level(fd
, info
->new_level
, devname
, verbose
);
3139 case 1: /* Couldn't set data_offset, try the old way */
3140 if (data_offset
!= INVALID_SECTORS
) {
3141 pr_err("Cannot update data_offset on this array\n");
3148 /* Decide how many blocks (sectors) for a reshape
3149 * unit. The number we have so far is just a minimum
3151 blocks
= reshape
.backup_blocks
;
3152 if (reshape
.before
.data_disks
==
3153 reshape
.after
.data_disks
) {
3154 /* Make 'blocks' bigger for better throughput, but
3155 * not so big that we reject it below.
3156 * Try for 16 megabytes
3158 while (blocks
* 32 < sra
->component_size
&&
3162 pr_err("Need to backup %luK of critical section..\n", blocks
/2);
3164 if (blocks
>= sra
->component_size
/2) {
3165 pr_err("%s: Something wrong - reshape aborted\n",
3170 /* Now we need to open all these devices so we can read/write.
3172 nrdisks
= max(reshape
.before
.data_disks
,
3173 reshape
.after
.data_disks
) + reshape
.parity
3174 + sra
->array
.spare_disks
;
3175 fdlist
= xcalloc((1+nrdisks
), sizeof(int));
3176 offsets
= xcalloc((1+nrdisks
), sizeof(offsets
[0]));
3178 odisks
= reshape
.before
.data_disks
+ reshape
.parity
;
3179 d
= reshape_prepare_fdlist(devname
, sra
, odisks
,
3180 nrdisks
, blocks
, backup_file
,
3185 if ((st
->ss
->manage_reshape
== NULL
) ||
3186 (st
->ss
->recover_backup
== NULL
)) {
3187 if (backup_file
== NULL
) {
3188 if (reshape
.after
.data_disks
<=
3189 reshape
.before
.data_disks
) {
3190 pr_err("%s: Cannot grow - need backup-file\n",
3192 pr_err(" Please provide one with \"--backup=...\"\n");
3194 } else if (sra
->array
.spare_disks
== 0) {
3195 pr_err("%s: Cannot grow - need a spare or backup-file to backup critical section\n", devname
);
3199 if (!reshape_open_backup_file(backup_file
, fd
, devname
,
3201 fdlist
+d
, offsets
+d
,
3210 update_cache_size(container
, sra
, info
,
3211 min(reshape
.before
.data_disks
, reshape
.after
.data_disks
),
3214 /* Right, everything seems fine. Let's kick things off.
3215 * If only changing raid_disks, use ioctl, else use
3220 if (impose_reshape(sra
, info
, st
, fd
, restart
,
3221 devname
, container
, &reshape
) < 0)
3224 err
= start_reshape(sra
, restart
, reshape
.before
.data_disks
,
3225 reshape
.after
.data_disks
);
3227 pr_err("Cannot %s reshape for %s\n",
3228 restart
? "continue" : "start",
3233 sysfs_set_str(sra
, NULL
, "array_state", "active");
3234 if (freeze_reshape
) {
3238 pr_err("Reshape has to be continued from location %llu when root filesystem has been mounted.\n",
3239 sra
->reshape_progress
);
3243 if (!forked
&& !check_env("MDADM_NO_SYSTEMCTL"))
3244 if (continue_via_systemd(container
?: sra
->sys_name
)) {
3251 /* Now we just need to kick off the reshape and watch, while
3252 * handling backups of the data...
3253 * This is all done by a forked background process.
3255 switch(forked
? 0 : fork()) {
3257 pr_err("Cannot run child to monitor reshape: %s\n",
3271 /* If another array on the same devices is busy, the
3272 * reshape will wait for them. This would mean that
3273 * the first section that we suspend will stay suspended
3274 * for a long time. So check on that possibility
3275 * by looking for "DELAYED" in /proc/mdstat, and if found,
3279 struct mdstat_ent
*mds
, *m
;
3281 mds
= mdstat_read(1, 0);
3282 for (m
= mds
; m
; m
= m
->next
)
3283 if (strcmp(m
->devnm
, sra
->sys_name
) == 0) {
3285 m
->percent
== RESYNC_DELAYED
)
3288 /* Haven't started the reshape thread
3295 if (delayed
== 1 && get_linux_version() < 3007000) {
3296 pr_err("Reshape is delayed, but cannot wait carefully with this kernel.\n"
3297 " You might experience problems until other reshapes complete.\n");
3301 mdstat_wait(30 - (delayed
-1) * 25);
3305 if (check_env("MDADM_GROW_VERIFY"))
3306 fd
= open(devname
, O_RDONLY
| O_DIRECT
);
3309 mlockall(MCL_FUTURE
);
3311 signal(SIGTERM
, catch_term
);
3313 if (st
->ss
->external
) {
3314 /* metadata handler takes it from here */
3315 done
= st
->ss
->manage_reshape(
3316 fd
, sra
, &reshape
, st
, blocks
,
3318 d
- odisks
, fdlist
+odisks
,
3321 done
= child_monitor(
3322 fd
, sra
, &reshape
, st
, blocks
,
3324 d
- odisks
, fdlist
+odisks
,
3330 if (backup_file
&& done
) {
3332 bul
= make_backup(sra
->sys_name
);
3335 int l
= readlink(bul
, buf
, sizeof(buf
) - 1);
3343 unlink(backup_file
);
3350 if (!st
->ss
->external
&&
3351 !(reshape
.before
.data_disks
!= reshape
.after
.data_disks
3352 && info
->custom_array_size
) &&
3353 info
->new_level
== reshape
.level
&&
3355 /* no need to wait for the reshape to finish as
3356 * there is nothing more to do.
3363 if (st
->ss
->external
) {
3364 /* Re-load the metadata as much could have changed */
3365 int cfd
= open_dev(st
->container_devnm
);
3367 flush_mdmon(container
);
3368 st
->ss
->free_super(st
);
3369 st
->ss
->load_container(st
, cfd
, container
);
3374 /* set new array size if required customer_array_size is used
3377 if (reshape
.before
.data_disks
!=
3378 reshape
.after
.data_disks
&&
3379 info
->custom_array_size
)
3380 set_array_size(st
, info
, info
->text_version
);
3382 if (info
->new_level
!= reshape
.level
) {
3384 fd
= open(devname
, O_RDONLY
);
3385 impose_level(fd
, info
->new_level
, devname
, verbose
);
3387 if (info
->new_level
== 0)
3388 st
->update_tail
= NULL
;
3400 if (orig_level
!= UnSet
&& sra
) {
3401 c
= map_num(pers
, orig_level
);
3402 if (c
&& sysfs_set_str(sra
, NULL
, "level", c
) == 0)
3403 pr_err("aborting level change\n");
3411 /* mdfd handle is passed to be closed in child process (after fork).
3413 int reshape_container(char *container
, char *devname
,
3415 struct supertype
*st
,
3416 struct mdinfo
*info
,
3418 char *backup_file
, int verbose
,
3419 int forked
, int restart
, int freeze_reshape
)
3421 struct mdinfo
*cc
= NULL
;
3423 char last_devnm
[32] = "";
3425 /* component_size is not meaningful for a container,
3426 * so pass '0' meaning 'no change'
3429 reshape_super(st
, 0, info
->new_level
,
3430 info
->new_layout
, info
->new_chunk
,
3431 info
->array
.raid_disks
, info
->delta_disks
,
3432 backup_file
, devname
, APPLY_METADATA_CHANGES
,
3440 /* ping monitor to be sure that update is on disk
3442 ping_monitor(container
);
3444 if (!forked
&& !freeze_reshape
&& !check_env("MDADM_NO_SYSTEMCTL"))
3445 if (continue_via_systemd(container
))
3448 switch (forked
? 0 : fork()) {
3449 case -1: /* error */
3450 perror("Cannot fork to complete reshape\n");
3453 default: /* parent */
3454 if (!freeze_reshape
)
3455 printf("%s: multi-array reshape continues in background\n", Name
);
3462 /* close unused handle in child process
3468 /* For each member array with reshape_active,
3469 * we need to perform the reshape.
3470 * We pick the first array that needs reshaping and
3471 * reshape it. reshape_array() will re-read the metadata
3472 * so the next time through a different array should be
3473 * ready for reshape.
3474 * It is possible that the 'different' array will not
3475 * be assembled yet. In that case we simple exit.
3476 * When it is assembled, the mdadm which assembles it
3477 * will take over the reshape.
3479 struct mdinfo
*content
;
3481 struct mdstat_ent
*mdstat
;
3487 cc
= st
->ss
->container_content(st
, NULL
);
3489 for (content
= cc
; content
; content
= content
->next
) {
3491 if (!content
->reshape_active
)
3494 subarray
= strchr(content
->text_version
+1, '/')+1;
3495 mdstat
= mdstat_by_subdev(subarray
, container
);
3498 if (mdstat
->active
== 0) {
3499 pr_err("Skipping inactive array %s.\n",
3501 free_mdstat(mdstat
);
3510 devid
= devnm2devid(mdstat
->devnm
);
3511 adev
= map_dev(major(devid
), minor(devid
), 0);
3513 adev
= content
->text_version
;
3515 fd
= open_dev(mdstat
->devnm
);
3517 pr_err("Device %s cannot be opened for reshape.\n", adev
);
3521 if (strcmp(last_devnm
, mdstat
->devnm
) == 0) {
3522 /* Do not allow for multiple reshape_array() calls for
3524 * It can happen when reshape_array() returns without
3525 * error, when reshape is not finished (wrong reshape
3526 * starting/continuation conditions). Mdmon doesn't
3527 * switch to next array in container and reentry
3528 * conditions for the same array occur.
3529 * This is possibly interim until the behaviour of
3530 * reshape_array is resolved().
3532 printf("%s: Multiple reshape execution detected for device %s.\n", Name
, adev
);
3536 strcpy(last_devnm
, mdstat
->devnm
);
3538 sysfs_init(content
, fd
, mdstat
->devnm
);
3540 if (mdmon_running(container
))
3541 flush_mdmon(container
);
3543 rv
= reshape_array(container
, fd
, adev
, st
,
3544 content
, force
, NULL
, INVALID_SECTORS
,
3545 backup_file
, verbose
, 1, restart
,
3549 if (freeze_reshape
) {
3558 if (mdmon_running(container
))
3559 flush_mdmon(container
);
3568 * We run a child process in the background which performs the following
3570 * - wait for resync to reach a certain point
3571 * - suspend io to the following section
3572 * - backup that section
3573 * - allow resync to proceed further
3575 * - discard the backup.
3577 * When are combined in slightly different ways in the three cases.
3579 * - suspend/backup/allow/wait/resume/discard
3581 * - allow/wait/suspend/backup/allow/wait/resume/discard
3583 * - wait/resume/discard/suspend/backup/allow
3585 * suspend/backup/allow always come together
3586 * wait/resume/discard do too.
3587 * For the same-size case we have two backups to improve flow.
3591 int progress_reshape(struct mdinfo
*info
, struct reshape
*reshape
,
3592 unsigned long long backup_point
,
3593 unsigned long long wait_point
,
3594 unsigned long long *suspend_point
,
3595 unsigned long long *reshape_completed
, int *frozen
)
3597 /* This function is called repeatedly by the reshape manager.
3598 * It determines how much progress can safely be made and allows
3600 * - 'info' identifies the array and particularly records in
3601 * ->reshape_progress the metadata's knowledge of progress
3602 * This is a sector offset from the start of the array
3603 * of the next array block to be relocated. This number
3604 * may increase from 0 or decrease from array_size, depending
3605 * on the type of reshape that is happening.
3606 * Note that in contrast, 'sync_completed' is a block count of the
3607 * reshape so far. It gives the distance between the start point
3608 * (head or tail of device) and the next place that data will be
3609 * written. It always increases.
3610 * - 'reshape' is the structure created by analyse_change
3611 * - 'backup_point' shows how much the metadata manager has backed-up
3612 * data. For reshapes with increasing progress, it is the next address
3613 * to be backed up, previous addresses have been backed-up. For
3614 * decreasing progress, it is the earliest address that has been
3615 * backed up - later address are also backed up.
3616 * So addresses between reshape_progress and backup_point are
3617 * backed up providing those are in the 'correct' order.
3618 * - 'wait_point' is an array address. When reshape_completed
3619 * passes this point, progress_reshape should return. It might
3620 * return earlier if it determines that ->reshape_progress needs
3621 * to be updated or further backup is needed.
3622 * - suspend_point is maintained by progress_reshape and the caller
3623 * should not touch it except to initialise to zero.
3624 * It is an array address and it only increases in 2.6.37 and earlier.
3625 * This makes it difficult to handle reducing reshapes with
3626 * external metadata.
3627 * However: it is similar to backup_point in that it records the
3628 * other end of a suspended region from reshape_progress.
3629 * it is moved to extend the region that is safe to backup and/or
3631 * - reshape_completed is read from sysfs and returned. The caller
3632 * should copy this into ->reshape_progress when it has reason to
3633 * believe that the metadata knows this, and any backup outside this
3637 * 1 if more data from backup_point - but only as far as suspend_point,
3638 * should be backed up
3639 * 0 if things are progressing smoothly
3640 * -1 if the reshape is finished because it is all done,
3641 * -2 if the reshape is finished due to an error.
3644 int advancing
= (reshape
->after
.data_disks
3645 >= reshape
->before
.data_disks
);
3646 unsigned long long need_backup
; /* All data between start of array and
3647 * here will at some point need to
3650 unsigned long long read_offset
, write_offset
;
3651 unsigned long long write_range
;
3652 unsigned long long max_progress
, target
, completed
;
3653 unsigned long long array_size
= (info
->component_size
3654 * reshape
->before
.data_disks
);
3658 /* First, we unsuspend any region that is now known to be safe.
3659 * If suspend_point is on the 'wrong' side of reshape_progress, then
3660 * we don't have or need suspension at the moment. This is true for
3661 * native metadata when we don't need to back-up.
3664 if (info
->reshape_progress
<= *suspend_point
)
3665 sysfs_set_num(info
, NULL
, "suspend_lo",
3666 info
->reshape_progress
);
3668 /* Note: this won't work in 2.6.37 and before.
3669 * Something somewhere should make sure we don't need it!
3671 if (info
->reshape_progress
>= *suspend_point
)
3672 sysfs_set_num(info
, NULL
, "suspend_hi",
3673 info
->reshape_progress
);
3676 /* Now work out how far it is safe to progress.
3677 * If the read_offset for ->reshape_progress is less than
3678 * 'blocks' beyond the write_offset, we can only progress as far
3680 * Otherwise we can progress until the write_offset for the new location
3681 * reaches (within 'blocks' of) the read_offset at the current location.
3682 * However that region must be suspended unless we are using native
3684 * If we need to suspend more, we limit it to 128M per device, which is
3685 * rather arbitrary and should be some time-based calculation.
3687 read_offset
= info
->reshape_progress
/ reshape
->before
.data_disks
;
3688 write_offset
= info
->reshape_progress
/ reshape
->after
.data_disks
;
3689 write_range
= info
->new_chunk
/512;
3690 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
)
3691 need_backup
= array_size
;
3693 need_backup
= reshape
->backup_blocks
;
3695 if (read_offset
< write_offset
+ write_range
)
3696 max_progress
= backup_point
;
3700 reshape
->after
.data_disks
;
3702 if (read_offset
> write_offset
- write_range
)
3703 /* Can only progress as far as has been backed up,
3704 * which must be suspended */
3705 max_progress
= backup_point
;
3706 else if (info
->reshape_progress
<= need_backup
)
3707 max_progress
= backup_point
;
3709 if (info
->array
.major_version
>= 0)
3710 /* Can progress until backup is needed */
3711 max_progress
= need_backup
;
3713 /* Can progress until metadata update is required */
3716 reshape
->after
.data_disks
;
3717 /* but data must be suspended */
3718 if (max_progress
< *suspend_point
)
3719 max_progress
= *suspend_point
;
3724 /* We know it is safe to progress to 'max_progress' providing
3725 * it is suspended or we are using native metadata.
3726 * Consider extending suspend_point 128M per device if it
3727 * is less than 64M per device beyond reshape_progress.
3728 * But always do a multiple of 'blocks'
3729 * FIXME this is too big - it takes to long to complete
3732 target
= 64*1024*2 * min(reshape
->before
.data_disks
,
3733 reshape
->after
.data_disks
);
3734 target
/= reshape
->backup_blocks
;
3737 target
*= reshape
->backup_blocks
;
3739 /* For externally managed metadata we always need to suspend IO to
3740 * the area being reshaped so we regularly push suspend_point forward.
3741 * For native metadata we only need the suspend if we are going to do
3745 if ((need_backup
> info
->reshape_progress
3746 || info
->array
.major_version
< 0) &&
3747 *suspend_point
< info
->reshape_progress
+ target
) {
3748 if (need_backup
< *suspend_point
+ 2 * target
)
3749 *suspend_point
= need_backup
;
3750 else if (*suspend_point
+ 2 * target
< array_size
)
3751 *suspend_point
+= 2 * target
;
3753 *suspend_point
= array_size
;
3754 sysfs_set_num(info
, NULL
, "suspend_hi", *suspend_point
);
3755 if (max_progress
> *suspend_point
)
3756 max_progress
= *suspend_point
;
3759 if (info
->array
.major_version
>= 0) {
3760 /* Only need to suspend when about to backup */
3761 if (info
->reshape_progress
< need_backup
* 2 &&
3762 *suspend_point
> 0) {
3764 sysfs_set_num(info
, NULL
, "suspend_lo", 0);
3765 sysfs_set_num(info
, NULL
, "suspend_hi", need_backup
);
3768 /* Need to suspend continually */
3769 if (info
->reshape_progress
< *suspend_point
)
3770 *suspend_point
= info
->reshape_progress
;
3771 if (*suspend_point
+ target
< info
->reshape_progress
)
3772 /* No need to move suspend region yet */;
3774 if (*suspend_point
>= 2 * target
)
3775 *suspend_point
-= 2 * target
;
3778 sysfs_set_num(info
, NULL
, "suspend_lo",
3781 if (max_progress
< *suspend_point
)
3782 max_progress
= *suspend_point
;
3786 /* now set sync_max to allow that progress. sync_max, like
3787 * sync_completed is a count of sectors written per device, so
3788 * we find the difference between max_progress and the start point,
3789 * and divide that by after.data_disks to get a sync_max
3791 * At the same time we convert wait_point to a similar number
3792 * for comparing against sync_completed.
3794 /* scale down max_progress to per_disk */
3795 max_progress
/= reshape
->after
.data_disks
;
3796 /* Round to chunk size as some kernels give an erroneously high number */
3797 max_progress
/= info
->new_chunk
/512;
3798 max_progress
*= info
->new_chunk
/512;
3799 /* And round to old chunk size as the kernel wants that */
3800 max_progress
/= info
->array
.chunk_size
/512;
3801 max_progress
*= info
->array
.chunk_size
/512;
3802 /* Limit progress to the whole device */
3803 if (max_progress
> info
->component_size
)
3804 max_progress
= info
->component_size
;
3805 wait_point
/= reshape
->after
.data_disks
;
3807 /* switch from 'device offset' to 'processed block count' */
3808 max_progress
= info
->component_size
- max_progress
;
3809 wait_point
= info
->component_size
- wait_point
;
3813 sysfs_set_num(info
, NULL
, "sync_max", max_progress
);
3815 /* Now wait. If we have already reached the point that we were
3816 * asked to wait to, don't wait at all, else wait for any change.
3817 * We need to select on 'sync_completed' as that is the place that
3818 * notifications happen, but we are really interested in
3819 * 'reshape_position'
3821 fd
= sysfs_get_fd(info
, NULL
, "sync_completed");
3823 goto check_progress
;
3825 if (sysfs_fd_get_ll(fd
, &completed
) < 0)
3826 goto check_progress
;
3828 while (completed
< max_progress
&& completed
< wait_point
) {
3829 /* Check that sync_action is still 'reshape' to avoid
3830 * waiting forever on a dead array
3833 if (sysfs_get_str(info
, NULL
, "sync_action",
3835 strncmp(action
, "reshape", 7) != 0)
3837 /* Some kernels reset 'sync_completed' to zero
3838 * before setting 'sync_action' to 'idle'.
3839 * So we need these extra tests.
3841 if (completed
== 0 && advancing
3842 && info
->reshape_progress
> 0)
3844 if (completed
== 0 && !advancing
3845 && info
->reshape_progress
< (info
->component_size
3846 * reshape
->after
.data_disks
))
3848 sysfs_wait(fd
, NULL
);
3849 if (sysfs_fd_get_ll(fd
, &completed
) < 0)
3850 goto check_progress
;
3852 /* Some kernels reset 'sync_completed' to zero,
3853 * we need to have real point we are in md
3856 completed
= max_progress
;
3858 /* some kernels can give an incorrectly high 'completed' number */
3859 completed
/= (info
->new_chunk
/512);
3860 completed
*= (info
->new_chunk
/512);
3861 /* Convert 'completed' back in to a 'progress' number */
3862 completed
*= reshape
->after
.data_disks
;
3864 completed
= info
->component_size
* reshape
->after
.data_disks
3867 *reshape_completed
= completed
;
3871 /* We return the need_backup flag. Caller will decide
3872 * how much - a multiple of ->backup_blocks up to *suspend_point
3875 return need_backup
> info
->reshape_progress
;
3877 return need_backup
>= info
->reshape_progress
;
3880 /* if we couldn't read a number from sync_completed, then
3881 * either the reshape did complete, or it aborted.
3882 * We can tell which by checking for 'none' in reshape_position.
3883 * If it did abort, then it might immediately restart if it
3884 * it was just a device failure that leaves us degraded but
3888 if (sysfs_get_str(info
, NULL
, "reshape_position", buf
, sizeof(buf
)) < 0
3889 || strncmp(buf
, "none", 4) != 0) {
3890 /* The abort might only be temporary. Wait up to 10
3891 * seconds for fd to contain a valid number again.
3895 unsigned long long new_sync_max
;
3896 while (fd
>= 0 && rv
< 0 && wait
> 0) {
3897 if (sysfs_wait(fd
, &wait
) != 1)
3899 switch (sysfs_fd_get_ll(fd
, &completed
)) {
3901 /* all good again */
3903 /* If "sync_max" is no longer max_progress
3904 * we need to freeze things
3906 sysfs_get_ll(info
, NULL
, "sync_max", &new_sync_max
);
3907 *frozen
= (new_sync_max
!= max_progress
);
3909 case -2: /* read error - abort */
3916 return rv
; /* abort */
3918 /* Maybe racing with array shutdown - check state */
3921 if (sysfs_get_str(info
, NULL
, "array_state", buf
, sizeof(buf
)) < 0
3922 || strncmp(buf
, "inactive", 8) == 0
3923 || strncmp(buf
, "clear",5) == 0)
3924 return -2; /* abort */
3925 return -1; /* complete */
3929 /* FIXME return status is never checked */
3930 static int grow_backup(struct mdinfo
*sra
,
3931 unsigned long long offset
, /* per device */
3932 unsigned long stripes
, /* per device, in old chunks */
3933 int *sources
, unsigned long long *offsets
,
3934 int disks
, int chunk
, int level
, int layout
,
3935 int dests
, int *destfd
, unsigned long long *destoffsets
,
3936 int part
, int *degraded
,
3939 /* Backup 'blocks' sectors at 'offset' on each device of the array,
3940 * to storage 'destfd' (offset 'destoffsets'), after first
3941 * suspending IO. Then allow resync to continue
3942 * over the suspended section.
3943 * Use part 'part' of the backup-super-block.
3948 unsigned long long ll
;
3950 //printf("offset %llu\n", offset);
3956 /* Check that array hasn't become degraded, else we might backup the wrong data */
3957 if (sysfs_get_ll(sra
, NULL
, "degraded", &ll
) < 0)
3958 return -1; /* FIXME this error is ignored */
3959 new_degraded
= (int)ll
;
3960 if (new_degraded
!= *degraded
) {
3961 /* check each device to ensure it is still working */
3963 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
3964 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
3966 if (sd
->disk
.state
& (1<<MD_DISK_SYNC
)) {
3968 if (sysfs_get_str(sra
, sd
, "state", sbuf
, 20) < 0 ||
3969 strstr(sbuf
, "faulty") ||
3970 strstr(sbuf
, "in_sync") == NULL
) {
3971 /* this device is dead */
3972 sd
->disk
.state
= (1<<MD_DISK_FAULTY
);
3973 if (sd
->disk
.raid_disk
>= 0 &&
3974 sources
[sd
->disk
.raid_disk
] >= 0) {
3975 close(sources
[sd
->disk
.raid_disk
]);
3976 sources
[sd
->disk
.raid_disk
] = -1;
3981 *degraded
= new_degraded
;
3984 bsb
.arraystart2
= __cpu_to_le64(offset
* odata
);
3985 bsb
.length2
= __cpu_to_le64(stripes
* (chunk
/512) * odata
);
3987 bsb
.arraystart
= __cpu_to_le64(offset
* odata
);
3988 bsb
.length
= __cpu_to_le64(stripes
* (chunk
/512) * odata
);
3991 bsb
.magic
[15] = '2';
3992 for (i
= 0; i
< dests
; i
++)
3994 lseek64(destfd
[i
], destoffsets
[i
] + __le64_to_cpu(bsb
.devstart2
)*512, 0);
3996 lseek64(destfd
[i
], destoffsets
[i
], 0);
3998 rv
= save_stripes(sources
, offsets
,
3999 disks
, chunk
, level
, layout
,
4001 offset
*512*odata
, stripes
* chunk
* odata
,
4006 bsb
.mtime
= __cpu_to_le64(time(0));
4007 for (i
= 0; i
< dests
; i
++) {
4008 bsb
.devstart
= __cpu_to_le64(destoffsets
[i
]/512);
4010 bsb
.sb_csum
= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
));
4011 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0)
4012 bsb
.sb_csum2
= bsb_csum((char*)&bsb
,
4013 ((char*)&bsb
.sb_csum2
)-((char*)&bsb
));
4016 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
] - 4096, 0)
4017 != destoffsets
[i
] - 4096)
4019 if (write(destfd
[i
], &bsb
, 512) != 512)
4021 if (destoffsets
[i
] > 4096) {
4022 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
]+stripes
*chunk
*odata
, 0) !=
4023 destoffsets
[i
]+stripes
*chunk
*odata
)
4025 if (write(destfd
[i
], &bsb
, 512) != 512)
4035 /* in 2.6.30, the value reported by sync_completed can be
4036 * less that it should be by one stripe.
4037 * This only happens when reshape hits sync_max and pauses.
4038 * So allow wait_backup to either extent sync_max further
4039 * than strictly necessary, or return before the
4040 * sync has got quite as far as we would really like.
4041 * This is what 'blocks2' is for.
4042 * The various caller give appropriate values so that
4045 /* FIXME return value is often ignored */
4046 static int forget_backup(int dests
, int *destfd
,
4047 unsigned long long *destoffsets
,
4051 * Erase backup 'part' (which is 0 or 1)
4057 bsb
.arraystart2
= __cpu_to_le64(0);
4058 bsb
.length2
= __cpu_to_le64(0);
4060 bsb
.arraystart
= __cpu_to_le64(0);
4061 bsb
.length
= __cpu_to_le64(0);
4063 bsb
.mtime
= __cpu_to_le64(time(0));
4065 for (i
= 0; i
< dests
; i
++) {
4066 bsb
.devstart
= __cpu_to_le64(destoffsets
[i
]/512);
4067 bsb
.sb_csum
= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
));
4068 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0)
4069 bsb
.sb_csum2
= bsb_csum((char*)&bsb
,
4070 ((char*)&bsb
.sb_csum2
)-((char*)&bsb
));
4071 if ((unsigned long long)lseek64(destfd
[i
], destoffsets
[i
]-4096, 0) !=
4072 destoffsets
[i
]-4096)
4075 write(destfd
[i
], &bsb
, 512) != 512)
4082 static void fail(char *msg
)
4085 rv
= (write(2, msg
, strlen(msg
)) != (int)strlen(msg
));
4086 rv
|= (write(2, "\n", 1) != 1);
4090 static char *abuf
, *bbuf
;
4091 static unsigned long long abuflen
;
4092 static void validate(int afd
, int bfd
, unsigned long long offset
)
4094 /* check that the data in the backup against the array.
4095 * This is only used for regression testing and should not
4096 * be used while the array is active
4100 lseek64(bfd
, offset
- 4096, 0);
4101 if (read(bfd
, &bsb2
, 512) != 512)
4102 fail("cannot read bsb");
4103 if (bsb2
.sb_csum
!= bsb_csum((char*)&bsb2
,
4104 ((char*)&bsb2
.sb_csum
)-((char*)&bsb2
)))
4105 fail("first csum bad");
4106 if (memcmp(bsb2
.magic
, "md_backup_data", 14) != 0)
4107 fail("magic is bad");
4108 if (memcmp(bsb2
.magic
, "md_backup_data-2", 16) == 0 &&
4109 bsb2
.sb_csum2
!= bsb_csum((char*)&bsb2
,
4110 ((char*)&bsb2
.sb_csum2
)-((char*)&bsb2
)))
4111 fail("second csum bad");
4113 if (__le64_to_cpu(bsb2
.devstart
)*512 != offset
)
4114 fail("devstart is wrong");
4117 unsigned long long len
= __le64_to_cpu(bsb2
.length
)*512;
4119 if (abuflen
< len
) {
4123 if (posix_memalign((void**)&abuf
, 4096, abuflen
) ||
4124 posix_memalign((void**)&bbuf
, 4096, abuflen
)) {
4126 /* just stop validating on mem-alloc failure */
4131 lseek64(bfd
, offset
, 0);
4132 if ((unsigned long long)read(bfd
, bbuf
, len
) != len
) {
4133 //printf("len %llu\n", len);
4134 fail("read first backup failed");
4136 lseek64(afd
, __le64_to_cpu(bsb2
.arraystart
)*512, 0);
4137 if ((unsigned long long)read(afd
, abuf
, len
) != len
)
4138 fail("read first from array failed");
4139 if (memcmp(bbuf
, abuf
, len
) != 0) {
4142 printf("offset=%llu len=%llu\n",
4143 (unsigned long long)__le64_to_cpu(bsb2
.arraystart
)*512, len
);
4144 for (i
=0; i
<len
; i
++)
4145 if (bbuf
[i
] != abuf
[i
]) {
4146 printf("first diff byte %d\n", i
);
4150 fail("data1 compare failed");
4154 unsigned long long len
= __le64_to_cpu(bsb2
.length2
)*512;
4156 if (abuflen
< len
) {
4160 abuf
= xmalloc(abuflen
);
4161 bbuf
= xmalloc(abuflen
);
4164 lseek64(bfd
, offset
+__le64_to_cpu(bsb2
.devstart2
)*512, 0);
4165 if ((unsigned long long)read(bfd
, bbuf
, len
) != len
)
4166 fail("read second backup failed");
4167 lseek64(afd
, __le64_to_cpu(bsb2
.arraystart2
)*512, 0);
4168 if ((unsigned long long)read(afd
, abuf
, len
) != len
)
4169 fail("read second from array failed");
4170 if (memcmp(bbuf
, abuf
, len
) != 0)
4171 fail("data2 compare failed");
4175 int child_monitor(int afd
, struct mdinfo
*sra
, struct reshape
*reshape
,
4176 struct supertype
*st
, unsigned long blocks
,
4177 int *fds
, unsigned long long *offsets
,
4178 int dests
, int *destfd
, unsigned long long *destoffsets
)
4180 /* Monitor a reshape where backup is being performed using
4181 * 'native' mechanism - either to a backup file, or
4182 * to some space in a spare.
4186 unsigned long long speed
;
4187 unsigned long long suspend_point
, array_size
;
4188 unsigned long long backup_point
, wait_point
;
4189 unsigned long long reshape_completed
;
4191 int increasing
= reshape
->after
.data_disks
>= reshape
->before
.data_disks
;
4192 int part
= 0; /* The next part of the backup area to fill. It may already
4193 * be full, so we need to check */
4194 int level
= reshape
->level
;
4195 int layout
= reshape
->before
.layout
;
4196 int data
= reshape
->before
.data_disks
;
4197 int disks
= reshape
->before
.data_disks
+ reshape
->parity
;
4198 int chunk
= sra
->array
.chunk_size
;
4200 unsigned long stripes
;
4204 /* set up the backup-super-block. This requires the
4205 * uuid from the array.
4207 /* Find a superblock */
4208 for (sd
= sra
->devs
; sd
; sd
= sd
->next
) {
4212 if (sd
->disk
.state
& (1<<MD_DISK_FAULTY
))
4214 dn
= map_dev(sd
->disk
.major
, sd
->disk
.minor
, 1);
4215 devfd
= dev_open(dn
, O_RDONLY
);
4218 ok
= st
->ss
->load_super(st
, devfd
, NULL
);
4224 pr_err("Cannot find a superblock\n");
4228 memset(&bsb
, 0, 512);
4229 memcpy(bsb
.magic
, "md_backup_data-1", 16);
4230 st
->ss
->uuid_from_super(st
, uuid
);
4231 memcpy(bsb
.set_uuid
, uuid
, 16);
4232 bsb
.mtime
= __cpu_to_le64(time(0));
4233 bsb
.devstart2
= blocks
;
4235 stripes
= blocks
/ (sra
->array
.chunk_size
/512) /
4236 reshape
->before
.data_disks
;
4238 if (posix_memalign((void**)&buf
, 4096, disks
* chunk
))
4239 /* Don't start the 'reshape' */
4241 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
) {
4242 sysfs_get_ll(sra
, NULL
, "sync_speed_min", &speed
);
4243 sysfs_set_num(sra
, NULL
, "sync_speed_min", 200000);
4247 array_size
= sra
->component_size
* reshape
->after
.data_disks
;
4248 backup_point
= sra
->reshape_progress
;
4251 array_size
= sra
->component_size
* reshape
->before
.data_disks
;
4252 backup_point
= reshape
->backup_blocks
;
4253 suspend_point
= array_size
;
4259 /* Want to return as soon the oldest backup slot can
4260 * be released as that allows us to start backing up
4261 * some more, providing suspend_point has been
4262 * advanced, which it should have.
4265 wait_point
= array_size
;
4266 if (part
== 0 && __le64_to_cpu(bsb
.length
) > 0)
4267 wait_point
= (__le64_to_cpu(bsb
.arraystart
) +
4268 __le64_to_cpu(bsb
.length
));
4269 if (part
== 1 && __le64_to_cpu(bsb
.length2
) > 0)
4270 wait_point
= (__le64_to_cpu(bsb
.arraystart2
) +
4271 __le64_to_cpu(bsb
.length2
));
4274 if (part
== 0 && __le64_to_cpu(bsb
.length
) > 0)
4275 wait_point
= __le64_to_cpu(bsb
.arraystart
);
4276 if (part
== 1 && __le64_to_cpu(bsb
.length2
) > 0)
4277 wait_point
= __le64_to_cpu(bsb
.arraystart2
);
4280 reshape_completed
= sra
->reshape_progress
;
4281 rv
= progress_reshape(sra
, reshape
,
4282 backup_point
, wait_point
,
4283 &suspend_point
, &reshape_completed
,
4285 /* external metadata would need to ping_monitor here */
4286 sra
->reshape_progress
= reshape_completed
;
4288 /* Clear any backup region that is before 'here' */
4290 if (__le64_to_cpu(bsb
.length
) > 0 &&
4291 reshape_completed
>= (__le64_to_cpu(bsb
.arraystart
) +
4292 __le64_to_cpu(bsb
.length
)))
4293 forget_backup(dests
, destfd
,
4295 if (__le64_to_cpu(bsb
.length2
) > 0 &&
4296 reshape_completed
>= (__le64_to_cpu(bsb
.arraystart2
) +
4297 __le64_to_cpu(bsb
.length2
)))
4298 forget_backup(dests
, destfd
,
4301 if (__le64_to_cpu(bsb
.length
) > 0 &&
4302 reshape_completed
<= (__le64_to_cpu(bsb
.arraystart
)))
4303 forget_backup(dests
, destfd
,
4305 if (__le64_to_cpu(bsb
.length2
) > 0 &&
4306 reshape_completed
<= (__le64_to_cpu(bsb
.arraystart2
)))
4307 forget_backup(dests
, destfd
,
4317 if (rv
== 0 && increasing
&& !st
->ss
->external
) {
4318 /* No longer need to monitor this reshape */
4319 sysfs_set_str(sra
, NULL
, "sync_max", "max");
4325 unsigned long long offset
;
4326 unsigned long actual_stripes
;
4327 /* Need to backup some data.
4328 * If 'part' is not used and the desired
4329 * backup size is suspended, do a backup,
4330 * then consider the next part.
4332 /* Check that 'part' is unused */
4333 if (part
== 0 && __le64_to_cpu(bsb
.length
) != 0)
4335 if (part
== 1 && __le64_to_cpu(bsb
.length2
) != 0)
4338 offset
= backup_point
/ data
;
4339 actual_stripes
= stripes
;
4341 if (offset
+ actual_stripes
* (chunk
/512) >
4342 sra
->component_size
)
4343 actual_stripes
= ((sra
->component_size
- offset
)
4345 if (offset
+ actual_stripes
* (chunk
/512) >
4349 if (offset
< actual_stripes
* (chunk
/512))
4350 actual_stripes
= offset
/ (chunk
/512);
4351 offset
-= actual_stripes
* (chunk
/512);
4352 if (offset
< suspend_point
/data
)
4355 if (actual_stripes
== 0)
4357 grow_backup(sra
, offset
, actual_stripes
,
4359 disks
, chunk
, level
, layout
,
4360 dests
, destfd
, destoffsets
,
4361 part
, °raded
, buf
);
4362 validate(afd
, destfd
[0], destoffsets
[0]);
4363 /* record where 'part' is up to */
4366 backup_point
+= actual_stripes
* (chunk
/512) * data
;
4368 backup_point
-= actual_stripes
* (chunk
/512) * data
;
4372 /* FIXME maybe call progress_reshape one more time instead */
4373 /* remove any remaining suspension */
4374 sysfs_set_num(sra
, NULL
, "suspend_lo", 0x7FFFFFFFFFFFFFFFULL
);
4375 sysfs_set_num(sra
, NULL
, "suspend_hi", 0);
4376 sysfs_set_num(sra
, NULL
, "suspend_lo", 0);
4377 sysfs_set_num(sra
, NULL
, "sync_min", 0);
4379 if (reshape
->before
.data_disks
== reshape
->after
.data_disks
)
4380 sysfs_set_num(sra
, NULL
, "sync_speed_min", speed
);
4386 * If any spare contains md_back_data-1 which is recent wrt mtime,
4387 * write that data into the array and update the super blocks with
4388 * the new reshape_progress
4390 int Grow_restart(struct supertype
*st
, struct mdinfo
*info
, int *fdlist
, int cnt
,
4391 char *backup_file
, int verbose
)
4395 unsigned long long *offsets
;
4396 unsigned long long nstripe
, ostripe
;
4399 odata
= info
->array
.raid_disks
- info
->delta_disks
- 1;
4400 if (info
->array
.level
== 6) odata
--; /* number of data disks */
4401 ndata
= info
->array
.raid_disks
- 1;
4402 if (info
->new_level
== 6) ndata
--;
4404 old_disks
= info
->array
.raid_disks
- info
->delta_disks
;
4406 if (info
->delta_disks
<= 0)
4407 /* Didn't grow, so the backup file must have
4411 for (i
=old_disks
-(backup_file
?1:0); i
<cnt
; i
++) {
4412 struct mdinfo dinfo
;
4415 char *devname
, namebuf
[20];
4416 unsigned long long lo
, hi
;
4418 /* This was a spare and may have some saved data on it.
4419 * Load the superblock, find and load the
4420 * backup_super_block.
4421 * If either fail, go on to next device.
4422 * If the backup contains no new info, just return
4423 * else restore data and update all superblocks
4425 if (i
== old_disks
-1) {
4426 fd
= open(backup_file
, O_RDONLY
);
4428 pr_err("backup file %s inaccessible: %s\n",
4429 backup_file
, strerror(errno
));
4432 devname
= backup_file
;
4437 if (st
->ss
->load_super(st
, fd
, NULL
))
4440 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
4441 st
->ss
->free_super(st
);
4444 (dinfo
.data_offset
+ dinfo
.component_size
- 8) <<9,
4446 pr_err("Cannot seek on device %d\n", i
);
4447 continue; /* Cannot seek */
4449 sprintf(namebuf
, "device-%d", i
);
4452 if (read(fd
, &bsb
, sizeof(bsb
)) != sizeof(bsb
)) {
4454 pr_err("Cannot read from %s\n", devname
);
4455 continue; /* Cannot read */
4457 if (memcmp(bsb
.magic
, "md_backup_data-1", 16) != 0 &&
4458 memcmp(bsb
.magic
, "md_backup_data-2", 16) != 0) {
4460 pr_err("No backup metadata on %s\n", devname
);
4463 if (bsb
.sb_csum
!= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum
)-((char*)&bsb
))) {
4465 pr_err("Bad backup-metadata checksum on %s\n", devname
);
4466 continue; /* bad checksum */
4468 if (memcmp(bsb
.magic
, "md_backup_data-2", 16) == 0 &&
4469 bsb
.sb_csum2
!= bsb_csum((char*)&bsb
, ((char*)&bsb
.sb_csum2
)-((char*)&bsb
))) {
4471 pr_err("Bad backup-metadata checksum2 on %s\n", devname
);
4472 continue; /* Bad second checksum */
4474 if (memcmp(bsb
.set_uuid
,info
->uuid
, 16) != 0) {
4476 pr_err("Wrong uuid on backup-metadata on %s\n", devname
);
4477 continue; /* Wrong uuid */
4480 /* array utime and backup-mtime should be updated at much the same time, but it seems that
4481 * sometimes they aren't... So allow considerable flexability in matching, and allow
4482 * this test to be overridden by an environment variable.
4484 if (info
->array
.utime
> (int)__le64_to_cpu(bsb
.mtime
) + 2*60*60 ||
4485 info
->array
.utime
< (int)__le64_to_cpu(bsb
.mtime
) - 10*60) {
4486 if (check_env("MDADM_GROW_ALLOW_OLD")) {
4487 pr_err("accepting backup with timestamp %lu for array with timestamp %lu\n",
4488 (unsigned long)__le64_to_cpu(bsb
.mtime
),
4489 (unsigned long)info
->array
.utime
);
4491 pr_err("too-old timestamp on backup-metadata on %s\n", devname
);
4492 pr_err("If you think it is should be safe, try 'export MDADM_GROW_ALLOW_OLD=1'\n");
4493 continue; /* time stamp is too bad */
4497 if (bsb
.magic
[15] == '1') {
4498 if (bsb
.length
== 0)
4500 if (info
->delta_disks
>= 0) {
4501 /* reshape_progress is increasing */
4502 if (__le64_to_cpu(bsb
.arraystart
)
4503 + __le64_to_cpu(bsb
.length
)
4504 < info
->reshape_progress
) {
4507 pr_err("backup-metadata found on %s but is not needed\n", devname
);
4508 continue; /* No new data here */
4511 /* reshape_progress is decreasing */
4512 if (__le64_to_cpu(bsb
.arraystart
) >=
4513 info
->reshape_progress
)
4514 goto nonew
; /* No new data here */
4517 if (bsb
.length
== 0 && bsb
.length2
== 0)
4519 if (info
->delta_disks
>= 0) {
4520 /* reshape_progress is increasing */
4521 if ((__le64_to_cpu(bsb
.arraystart
)
4522 + __le64_to_cpu(bsb
.length
)
4523 < info
->reshape_progress
)
4525 (__le64_to_cpu(bsb
.arraystart2
)
4526 + __le64_to_cpu(bsb
.length2
)
4527 < info
->reshape_progress
))
4528 goto nonew
; /* No new data here */
4530 /* reshape_progress is decreasing */
4531 if (__le64_to_cpu(bsb
.arraystart
) >=
4532 info
->reshape_progress
&&
4533 __le64_to_cpu(bsb
.arraystart2
) >=
4534 info
->reshape_progress
)
4535 goto nonew
; /* No new data here */
4538 if (lseek64(fd
, __le64_to_cpu(bsb
.devstart
)*512, 0)< 0) {
4541 pr_err("Failed to verify secondary backup-metadata block on %s\n",
4543 continue; /* Cannot seek */
4545 /* There should be a duplicate backup superblock 4k before here */
4546 if (lseek64(fd
, -4096, 1) < 0 ||
4547 read(fd
, &bsb2
, sizeof(bsb2
)) != sizeof(bsb2
))
4548 goto second_fail
; /* Cannot find leading superblock */
4549 if (bsb
.magic
[15] == '1')
4550 bsbsize
= offsetof(struct mdp_backup_super
, pad1
);
4552 bsbsize
= offsetof(struct mdp_backup_super
, pad
);
4553 if (memcmp(&bsb2
, &bsb
, bsbsize
) != 0)
4554 goto second_fail
; /* Cannot find leading superblock */
4556 /* Now need the data offsets for all devices. */
4557 offsets
= xmalloc(sizeof(*offsets
)*info
->array
.raid_disks
);
4558 for(j
=0; j
<info
->array
.raid_disks
; j
++) {
4561 if (st
->ss
->load_super(st
, fdlist
[j
], NULL
))
4562 /* FIXME should be this be an error */
4564 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
4565 st
->ss
->free_super(st
);
4566 offsets
[j
] = dinfo
.data_offset
* 512;
4568 printf("%s: restoring critical section\n", Name
);
4570 if (restore_stripes(fdlist
, offsets
,
4571 info
->array
.raid_disks
,
4575 fd
, __le64_to_cpu(bsb
.devstart
)*512,
4576 __le64_to_cpu(bsb
.arraystart
)*512,
4577 __le64_to_cpu(bsb
.length
)*512, NULL
)) {
4578 /* didn't succeed, so giveup */
4580 pr_err("Error restoring backup from %s\n",
4586 if (bsb
.magic
[15] == '2' &&
4587 restore_stripes(fdlist
, offsets
,
4588 info
->array
.raid_disks
,
4592 fd
, __le64_to_cpu(bsb
.devstart
)*512 +
4593 __le64_to_cpu(bsb
.devstart2
)*512,
4594 __le64_to_cpu(bsb
.arraystart2
)*512,
4595 __le64_to_cpu(bsb
.length2
)*512, NULL
)) {
4596 /* didn't succeed, so giveup */
4598 pr_err("Error restoring second backup from %s\n",
4606 /* Ok, so the data is restored. Let's update those superblocks. */
4610 lo
= __le64_to_cpu(bsb
.arraystart
);
4611 hi
= lo
+ __le64_to_cpu(bsb
.length
);
4613 if (bsb
.magic
[15] == '2' && bsb
.length2
) {
4614 unsigned long long lo1
, hi1
;
4615 lo1
= __le64_to_cpu(bsb
.arraystart2
);
4616 hi1
= lo1
+ __le64_to_cpu(bsb
.length2
);
4620 } else if (lo
< lo1
)
4626 (info
->reshape_progress
< lo
||
4627 info
->reshape_progress
> hi
))
4628 /* backup does not affect reshape_progress*/ ;
4629 else if (info
->delta_disks
>= 0) {
4630 info
->reshape_progress
= __le64_to_cpu(bsb
.arraystart
) +
4631 __le64_to_cpu(bsb
.length
);
4632 if (bsb
.magic
[15] == '2') {
4633 unsigned long long p2
= __le64_to_cpu(bsb
.arraystart2
) +
4634 __le64_to_cpu(bsb
.length2
);
4635 if (p2
> info
->reshape_progress
)
4636 info
->reshape_progress
= p2
;
4639 info
->reshape_progress
= __le64_to_cpu(bsb
.arraystart
);
4640 if (bsb
.magic
[15] == '2') {
4641 unsigned long long p2
= __le64_to_cpu(bsb
.arraystart2
);
4642 if (p2
< info
->reshape_progress
)
4643 info
->reshape_progress
= p2
;
4646 for (j
=0; j
<info
->array
.raid_disks
; j
++) {
4649 if (st
->ss
->load_super(st
, fdlist
[j
], NULL
))
4651 st
->ss
->getinfo_super(st
, &dinfo
, NULL
);
4652 dinfo
.reshape_progress
= info
->reshape_progress
;
4653 st
->ss
->update_super(st
, &dinfo
,
4654 "_reshape_progress",
4656 st
->ss
->store_super(st
, fdlist
[j
]);
4657 st
->ss
->free_super(st
);
4661 /* Didn't find any backup data, try to see if any
4664 if (info
->delta_disks
< 0) {
4665 /* When shrinking, the critical section is at the end.
4666 * So see if we are before the critical section.
4668 unsigned long long first_block
;
4669 nstripe
= ostripe
= 0;
4671 while (ostripe
>= nstripe
) {
4672 ostripe
+= info
->array
.chunk_size
/ 512;
4673 first_block
= ostripe
* odata
;
4674 nstripe
= first_block
/ ndata
/ (info
->new_chunk
/512) *
4675 (info
->new_chunk
/512);
4678 if (info
->reshape_progress
>= first_block
)
4681 if (info
->delta_disks
> 0) {
4682 /* See if we are beyond the critical section. */
4683 unsigned long long last_block
;
4684 nstripe
= ostripe
= 0;
4686 while (nstripe
>= ostripe
) {
4687 nstripe
+= info
->new_chunk
/ 512;
4688 last_block
= nstripe
* ndata
;
4689 ostripe
= last_block
/ odata
/ (info
->array
.chunk_size
/512) *
4690 (info
->array
.chunk_size
/512);
4693 if (info
->reshape_progress
>= last_block
)
4696 /* needed to recover critical section! */
4698 pr_err("Failed to find backup of critical section\n");
4702 int Grow_continue_command(char *devname
, int fd
,
4703 char *backup_file
, int verbose
)
4706 struct supertype
*st
= NULL
;
4707 struct mdinfo
*content
= NULL
;
4708 struct mdinfo array
;
4709 char *subarray
= NULL
;
4710 struct mdinfo
*cc
= NULL
;
4711 struct mdstat_ent
*mdstat
= NULL
;
4715 dprintf("Grow continue from command line called for %s\n",
4718 st
= super_by_fd(fd
, &subarray
);
4719 if (!st
|| !st
->ss
) {
4720 pr_err("Unable to determine metadata format for %s\n",
4724 dprintf("Grow continue is run for ");
4725 if (st
->ss
->external
== 0) {
4727 dprintf_cont("native array (%s)\n", devname
);
4728 if (ioctl(fd
, GET_ARRAY_INFO
, &array
.array
) < 0) {
4729 pr_err("%s is not an active md array - aborting\n", devname
);
4731 goto Grow_continue_command_exit
;
4734 /* Need to load a superblock.
4735 * FIXME we should really get what we need from
4738 for (d
= 0; d
< MAX_DISKS
; d
++) {
4739 mdu_disk_info_t disk
;
4743 if (ioctl(fd
, GET_DISK_INFO
, &disk
) < 0)
4745 if (disk
.major
== 0 && disk
.minor
== 0)
4747 if ((disk
.state
& (1 << MD_DISK_ACTIVE
)) == 0)
4749 dv
= map_dev(disk
.major
, disk
.minor
, 1);
4752 fd2
= dev_open(dv
, O_RDONLY
);
4755 err
= st
->ss
->load_super(st
, fd2
, NULL
);
4757 /* invalidate fd2 to avoid possible double close() */
4763 if (d
== MAX_DISKS
) {
4764 pr_err("Unable to load metadata for %s\n",
4767 goto Grow_continue_command_exit
;
4769 st
->ss
->getinfo_super(st
, content
, NULL
);
4774 dprintf_cont("subarray (%s)\n", subarray
);
4775 container
= st
->container_devnm
;
4776 cfd
= open_dev_excl(st
->container_devnm
);
4778 container
= st
->devnm
;
4780 cfd
= open_dev_excl(st
->devnm
);
4781 dprintf_cont("container (%s)\n", container
);
4785 pr_err("Unable to open container for %s\n", devname
);
4787 goto Grow_continue_command_exit
;
4790 /* find in container array under reshape
4792 ret_val
= st
->ss
->load_container(st
, cfd
, NULL
);
4794 pr_err("Cannot read superblock for %s\n",
4797 goto Grow_continue_command_exit
;
4800 cc
= st
->ss
->container_content(st
, subarray
);
4801 for (content
= cc
; content
; content
= content
->next
) {
4803 int allow_reshape
= 1;
4805 if (content
->reshape_active
== 0)
4807 /* The decision about array or container wide
4808 * reshape is taken in Grow_continue based
4809 * content->reshape_active state, therefore we
4810 * need to check_reshape based on
4811 * reshape_active and subarray name
4813 if (content
->array
.state
& (1<<MD_SB_BLOCK_VOLUME
))
4815 if (content
->reshape_active
== CONTAINER_RESHAPE
&&
4816 (content
->array
.state
4817 & (1<<MD_SB_BLOCK_CONTAINER_RESHAPE
)))
4820 if (!allow_reshape
) {
4821 pr_err("cannot continue reshape of an array in container with unsupported metadata: %s(%s)\n",
4822 devname
, container
);
4824 goto Grow_continue_command_exit
;
4827 array
= strchr(content
->text_version
+1, '/')+1;
4828 mdstat
= mdstat_by_subdev(array
, container
);
4831 if (mdstat
->active
== 0) {
4832 pr_err("Skipping inactive array %s.\n",
4834 free_mdstat(mdstat
);
4841 pr_err("Unable to determine reshaped array for %s\n", devname
);
4843 goto Grow_continue_command_exit
;
4845 fd2
= open_dev(mdstat
->devnm
);
4847 pr_err("cannot open (%s)\n", mdstat
->devnm
);
4849 goto Grow_continue_command_exit
;
4852 sysfs_init(content
, fd2
, mdstat
->devnm
);
4854 /* start mdmon in case it is not running
4856 if (!mdmon_running(container
))
4857 start_mdmon(container
);
4858 ping_monitor(container
);
4860 if (mdmon_running(container
))
4861 st
->update_tail
= &st
->updates
;
4863 pr_err("No mdmon found. Grow cannot continue.\n");
4865 goto Grow_continue_command_exit
;
4869 /* verify that array under reshape is started from
4872 if (verify_reshape_position(content
, content
->array
.level
) < 0) {
4874 goto Grow_continue_command_exit
;
4879 ret_val
= Grow_continue(fd
, st
, content
, backup_file
, 1, 0);
4881 Grow_continue_command_exit
:
4886 st
->ss
->free_super(st
);
4887 free_mdstat(mdstat
);
4894 int Grow_continue(int mdfd
, struct supertype
*st
, struct mdinfo
*info
,
4895 char *backup_file
, int forked
, int freeze_reshape
)
4899 if (!info
->reshape_active
)
4902 if (st
->ss
->external
) {
4903 int cfd
= open_dev(st
->container_devnm
);
4908 st
->ss
->load_container(st
, cfd
, st
->container_devnm
);
4910 ret_val
= reshape_container(st
->container_devnm
, NULL
, mdfd
,
4911 st
, info
, 0, backup_file
,
4913 1 | info
->reshape_active
,
4916 ret_val
= reshape_array(NULL
, mdfd
, "array", st
, info
, 1,
4917 NULL
, INVALID_SECTORS
,
4918 backup_file
, 0, forked
,
4919 1 | info
->reshape_active
,
4925 char *make_backup(char *name
)
4927 char *base
= "backup_file-";
4931 len
= strlen(MAP_DIR
) + 1 + strlen(base
) + strlen(name
)+1;
4932 fname
= xmalloc(len
);
4933 sprintf(fname
, "%s/%s%s", MAP_DIR
, base
, name
);
4937 char *locate_backup(char *name
)
4939 char *fl
= make_backup(name
);
4942 if (stat(fl
, &stb
) == 0 &&
4943 S_ISREG(stb
.st_mode
))