94

Sometimes, when resizing or otherwise mucking about with partitions on a disk, cfdisk will say:

Wrote partition table, but re-read table failed. Reboot to update table.

(This also happens with other partitioning tools, so I'm thinking this is a Linux issue rather than a cfdisk issue.) Why is this, and why does it only happens sometimes, and what can I do to avoid it?

Note: Please assume that none of the partitions I am actually editing are opened, mounted or otherwise in use.


Update:

cfdisk uses ioctl(fd, BLKRRPART, NULL) to tell Linux to reread the partition table. Two of the other tools recommended so far (hdparm -z DEVICE, sfdisk -R DEVICE) does exactly the same thing. The partprobe DEVICE command, on the other hand, seems to use a new ioctl called BLKPG, which might be better; I don't know. (It also falls back on BLKRRPART if BLKPG fails.)

BLKPG seems to be a "this partition has changed; here is the new size" operation, and it looked like partprobe called it individually on all the partitions on the device passed, so it should work if the individual partitions are unused. However, I have not had the opportunity to try it.

2
  • 1
    man sfdisk says: Since version 2.26 sfdisk no longer provides the -R or --re-read option to force the kernel to reread the partition table. Use blockdev --rereadpt instead.
    – Tom Hale
    Commented Jan 18, 2017 at 2:26
  • Check for a running software raid! Maybe mdadm --stop /dev/mdXX is needed first?
    – guettli
    Commented Mar 8 at 11:16

18 Answers 18

90

IMHO the most reliable/best answer is

partprobe /dev/sdX
7
  • 1
    I just expanded a dev under Ubuntu Server, this dev is a hardware raid. After expanding the underlying raid using the raidcontroller, I dismounted the filesystem and tried "partprobe /dev/sda" - this did not work. "fdisk -l" still showed the old size. I then ran "hdparm -z /dev/sda" and this did the trick. I could then mount and resize my filesystem with no reboot. I know I'm not adding anything really other than YMMV.
    – Mwuanno
    Commented May 28, 2014 at 13:30
  • i'm on centos 6.5 ; kernel 2.6.32 . all the following commands did not make kernel reread partition : - partprobe /dev/sda (warnikg : kernel failed to reread)
    – Max
    Commented May 29, 2015 at 7:56
  • @Max, I've also noticed that sometimes even partprobe prints out an error that it didn't work. Sometimes a reboot is the only option to be certain. Many times it does seem to work for me though.
    – hookenz
    Commented Oct 1, 2015 at 3:38
  • This did not work for me because there were some directories still mounted with --bind. The partition itself was already unmounted but the bind-mounts pointing to that partition were still there. Strange that umount worked and partprobe not, but after umounting the bind-mounts too, I could also partprobe the disk. Commented Nov 23, 2015 at 19:32
  • This worked for me on CentOS 6 after flagellating around with kpartx and udevadm trigger for 10 minutes. Thank you! Commented Apr 15, 2016 at 14:14
24

Rereading partition table information doesn't always work, but try

hdparm -z /dev/sda

or

sfdisk -R /dev/sda

If it works the values in /proc/partitions will change.

6
  • hdparm worked for me. Commented Jun 9, 2011 at 22:03
  • 5
    the sfdisk -R option doesn't exist.
    – hookenz
    Commented May 19, 2015 at 0:09
  • It should be noted that hdparm command will only work if partitions are not mounted.
    – user66336
    Commented Nov 16, 2015 at 23:45
  • 1
    ...indeed, it looks like sfdisk -R was removed somewhere between util-linux 2.24.2 and 2.26.1 Commented Jan 25, 2016 at 16:21
  • 6
    man sfdisk says: Since version 2.26 sfdisk no longer provides the -R or --re-read option to force the kernel to reread the partition table. Use blockdev --rereadpt instead.
    – Tom Hale
    Commented Jan 18, 2017 at 2:26
19

I (the original questioner) had a situation a few days ago when none of the other answers (including partprobe /dev/sdX, currently the accepted and highest-voted answer) worked. What did work, however, was this:

blockdev --rereadpt /dev/sdX

(I don't know why this worked and the others didn't, but I'm happy it did work, as it saved me a reboot on a busy server.)

1
  • I like this most because blockdev is part of the util-linux package which is independent of certain partitioning programs being installed.
    – PoC
    Commented Apr 12 at 22:58
15

On Centos7 :

According to https://access.redhat.com/solutions/199573

You should try :

partx -u <partition>

It worked for me.

1
  • 2
    That was the only one that work for me. Thank you so much for sharing!! Top of the day to you, sir!
    – NotGaeL
    Commented Apr 8, 2016 at 18:36
8

Note: Please assume that none of the partitions I am actually editing are opened, mounted or otherwise in use.

Given that assumption, the partition table can be successfully rescanned, and the issue won't arise. If you're getting that error, it's because the partition table is currently in use, and hence can't be re-scanned without creating inconsistencies.

2
  • Some partitions might be in use, but none of them are the ones that I am actually changing, though they might be in the same partition table.
    – Teddy
    Commented Jul 11, 2009 at 1:22
  • 9
    The kernel isn't that smart. If any partition in the table is in use, the kernel doesn't rescan. Getting that wrong in the other direction could be catastrophic, so it's being safe. If you want to stuff around with partitions at will, use LVM.
    – womble
    Commented Jul 11, 2009 at 2:05
7

When a command like blockdev --rereadpt /dev/sdX fails with

blockdev: ioctl error on BLKRRPART: Device or resource busy

this usually means that some (old) partition is indeed still somehow used by the kernel.

Possible causes/fixes:

  1. an sdX partition - say sdX1 - is still mounted - check with mount and umount it
  2. /dev/sdX1 is part of a software raid - check cat /proc/mdstat and possibly stop the relevant arrays, e.g. mdadm --stop /dev/md126
  3. /dev/sdX1 is part of an LVM physical volume - check with pvdisplay/vgdisplay and possibly deactivate with vgchange
  4. /dev/sdX1 is part of some device mapping - e.g. via cryptsetup - check /dev/mapper and lsblk and possibly remove the mapping (e.g. cryptsetup luksClose)
  5. Race condition with some udev probing - check running processes with ps and possibly kill one

If one tool - say blockdev --rereadpt fails usually similar ones like (partx -uv, kpartx, partprobe, kpartprobe) fail in a similar way until the root cause is eliminated.

6

It is not based on partition that you are editing.

Suppose you have only one harddisk (/dev/sda) and two partitions (/dev/sda1, /dev/sda2) and you have mounted only one partition (/dev/sda1). If you delete or change anything about other partition which is not even mounted (/dev/sda2) you will get the error that re-reading of partition table failed and kernel will use old table.

But if you have two harddisks (/dev/sda, /dev/sdb) and none of the partitions of (/dev/sdb) are in use. Then you can add / delete / resize /edit partitions of /dev/sdb and they will be re-read without any problem. But even if one partition of /dev/sdb was mounted during change. Then kernel will keep using old table.

5

With all mount points unmounted, running Yocto 2.4:

partprobe /dev/sda 

Still failed to re-load the partition table after partitions had been deleted on the device. Also tried -- and failed were:

udevadm trigger --subsystem-match=block; udevadm settle
hdparm -z /dev/sda
blockdev --rereadpt /dev/sda

All reported similar "BLKRRPART failed : device or resource busy..." errors instructing me to reboot. Is this failure of previously working methods possibly due to the fact that udev is now under systemd control? Thinking along those lines I tried:

systemctl restart systemd-udevd.service

And suddenly my disk is available again, without a reboot!

2
  • 1
    The top accepted answer is incomplete: in the modern systemd world, THIS is the correct answer. Please note you also need to restart one of those (or both) systemd-udev-settle and systemd-udev-trigger. Restarting just systemd-udevd like Camp said was not enough for me. But restarting also the other two did the trick! Commented Feb 5, 2019 at 18:03
  • systemctl restart systemd-udev* worked for me (although it printed a lot of warnings). Commented Nov 8, 2021 at 18:36
4

i'm on centos 6.5 x64 ; kernel 2.6.32 . and i'm testing the fdisk trick to resize.

/dev/sda1 /boot
/dev/sda2 /

All the following commands did not make kernel reread partition :

  • partprobe /dev/sda (warning : kernel failed to reread ....)
  • hdparm -z /dev/sda (BLKRRPART failed : device or resource busy)
  • blockdev -rereadpt /dev/sda (BLKRRPART failed : device or resource busy)
  • sfdisk -R /dev/sda (BLKRRPART failed : device or resource busy)

i still need a reboot to make it work

4
  • none of that worked for me either (proxmox VM, centos 7, xfs partition, no lvm). @uus answer did work, though: serverfault.com/a/722386/102252
    – NotGaeL
    Commented Apr 8, 2016 at 18:37
  • All the above commands did not work for me either.
    – Fadi Asbih
    Commented Mar 15, 2017 at 10:59
  • I think kernel 2.6.32 has issue, I used these before on other machines, it worked just fine, even when adding partitions with higher numbers between older partitions. i.e. sdb1 sdb2 sdb3 - delete sdb2 , then sdb1 sdb4 sdb5 sdb3. In addition to above, partx, kpartx, blockdev didn't work as well.
    – user235422
    Commented Jun 30, 2017 at 13:45
  • I don't think that it's unusual that if one command fails rereading the partitions, all fail - see also my answer for how to eliminate some causes for this. Commented Nov 17, 2018 at 20:58
3

You can also try:

echo 1 > /sys/block/sdX/device/rescan

(But might not work, see the comment below)

3
  • 4
    This does not reread the partition table. It just updates the geometry information, cache mode, etc. SCSI IDENTIFY_DRIVE is being issued. Commented Jun 9, 2012 at 8:27
  • 2
    This is the only solution, that helped me on vmware vm. Commented Feb 11, 2020 at 11:57
  • @user3132194, I assume you'd just resized the drive? Then you need to rescan the device, not reread the partition table. They're two completely different operations, and only the partition table reread is the one that's on-topic for this question. Commented Nov 7, 2020 at 1:12
2

in Alpine-Linux netboot mode , new partitions were shown in dmesg, main disks /dev/sda were present but no /dev/sdaX partition device files , tried all tricks above , finally


mknod -m 660 ${chroot_dir}/dev/sda1 b 8 1
mknod -m 660 ${chroot_dir}/dev/sda2 b 8 2

worked

https://wiki.alpinelinux.org/wiki/Alpine_Linux_in_a_chroot

2
  • 1
    666 on a device node is.. not particularly wise. You don't want non-administrative users going straight to the underlying block device to bypass file permissions. Commented Nov 7, 2020 at 1:10
  • fixed... 660 now ;) Commented Mar 16, 2021 at 3:17
1

For me neither partprobe or blockdev solution worked. Although, this one works:

udevadm settle --exit-if-exists=/dev/sdb1
1

I was having a very similar issue in Alpine Linux (not chrooted) where partprobe did nothing. It was using mdev to manage the /dev and /sys mounts, and so mdev -s to trigger a re-scan after imaging the mmcblk0 device worked.

0

kpartx -a <partition> can be run two times on newly created partition .... instead of rebooting the system.

2
  • 2
    Two times? Do you also run “sync; sync; sync”? ☺ I smell superstition…
    – Teddy
    Commented Mar 12, 2015 at 13:04
  • 2
    I think this superstition came from the fact that you check if your sync, by doing a second sync. Except that the second one is only valuable for the operator, to confirm it returns to prompt immediately, thus indicating the first sync completed as expected. a few blogs and tutorials later, and .... Commented Oct 5, 2015 at 13:59
0

Remember to check udev service is running. This is especially useful when partprobe, hdparm, blockdev, and various other commands do not seem to make any difference what device files are available in /dev/ directory.

0
fdisk /dev/sdX

and exit by "w"

0

In case if you want to ask kernel to "reread partition table", or update the size of specific partition to its actual size on Oracle Linux 6, these instructions could be helpful.

Let's say you have a machine with Oracle Linux 6, relatively new kernel 4.1.12-124.28.3.el6uek.x86_64, you just changed the size of your virtual disk, and now you're in situation that your /dev/sdc is 2.1 TiB but its own only MOUNTED partition sdc1 is 2 TiB, and you want to resize it to 2.1 TiB without unmounting it or rebooting. lsblk will show you something like that:

sdc                              8:32   0  2.1T  0 disk
`-sdc1                           8:33   0    2T  0 part

What I've found is that there is no ready-to-use application that you can install from package repository to help you with the issue: all programs in repository are too old and in the best case they will rewrite your partition table to disk (it's still required), but will use BLKRRPART to notify kernel about changes that won't work with mounted/"busy" partitions.

What you can do instead is to build your own simple C program that will run newer BLKPG ioctl to do the trick. Fortunately, you don't need to write it from scratch: just download this C file, modify it, build and run:

wget https://raw.githubusercontent.com/gportay/blkpg-part/master/blkpg-part.c
sed -i 's/BLKPG_RESIZE_PARTITION/3/' blkpg-part.c
make blkpg-part

Why you need to replace BLKPG_RESIZE_PARTITION with hardcoded value 3? It's because Oracle Linux 6 ships with the old conservative UAPI that is not synchronized with actual Linux kernel API, and therefore BLKPG_RESIZE_PARTITION constant from include/uapi/linux/blkkg.h is not available, even though the feature is present in kernel (you can check it by grepping the Linux kernel source code with BLKPG_RESIZE_PARTITION).

Okay, so you have your blkpg-part built, now you can check that it's working. Let's start with running it "dry-run": just ask it to resize the partition /dev/sdc1 with its own current size (reminder: the size of partition from kernel's point of view is still the "old" one, therefore /sys/class/block/sdc1/size will return the "old" size, not the one that is written in the new partition table):

./blkpg-part resize /dev/sdc 1 $(( $(cat /sys/class/block/sdc1/start) * 512 )) $(( $(cat /sys/class/block/sdc1/size) * 512 ))

If it runs without errors, it means that this userspace application succeeded with running BLKPG_RESIZE_PARTITION and interacting with kernel; so, we are close to actually ask kernel to change the size of the partition.

Now you need to run blkpg-part with the new partition size in bytes that is written in the new partition table. parted /dev/sdc unit B print can be used to identify the actual size of the partition, you'll get something like that:

Number  Start   End             Size            File system  Name     Flags
 1      20480B  2254857813503B  2254857793024B               primary  lvm

So, the new size of partition is 2254857793024, and you can finally run this command below:

./blkpkg-part3 resize /dev/sdc 1 $(( $(cat /sys/class/block/sdc1/start) * 512 )) 2254857793024

Immediately after that kernel will update the state of the partition, and in lsblk you'll see:

sdc                              8:32   0  2.1T  0 disk
`-sdc1                           8:33   0  2.1T  0 part

Useful links:

-5

If you read the manpage for 'man oracleasm-scandisks' you will note the text below. oracleasm is using /proc/partitions as the source of all scanning it performs. You must get your raw devices listed in /proc/partitions before you can do a scandisk. The Scanorder and Scanexclude parameters you place in /etc/sysconfig/oracleasm relate to the names found in /proc/partitions (!!!!).

---------- man oracleasm-scandisks ------ ...

HOW SCANNING HAPPENS The scan proceeds in four basic stages.

   First, the list of disks to scan is created. If disks were specified on the command line, this is the list.
   If not, /proc/partitions is read, and each block device is added, subject to the -o and -x options.

   Second, the partition tables of each disk in the scan are reloaded unless the -s option was specified. Any
   disks that no longer exist are dropped.

   Third, the list of disks is recreated based on the new partition tables.

   Finally, each disk in the list is checked to see if it is marked for ASM use. Disks that are marked are
   instantiated.
2
  • 2
    ...he's mentioned nothing about using oracleasm-scandisks
    – voretaq7
    Commented Apr 26, 2013 at 17:45
  • This is a separate piece of (Oracle-provided) software that's reading /proc/partitions. The goal of this question is how to tell the kernel itself to update the data it uses to generate /proc/partitions. Commented Nov 7, 2020 at 1:13

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .