In linux, how can I determine what processes are using a block device?

Question

I have a disk in a server that I'm migrating to a LVM volume group. Previously, it was using traditional DOS-disk partitioning, hdb[1-5].

I've unmounted every filesystem from hdb, shut off swap using hdb, removed a smaller VG on the device already, and went to repartition it using fdisk, deleted existing partitions, and created 2 partitions, but upon writing it out, linux refused to re-read the partition table. Trying again using hdparm -z reports: BLKRRPART failed: Device or resource busy.

I've checked the following places to ensure the device and it's partitions arn't listed anywhere:

• /proc/swaps
• /proc/mdadm
• output from 'pvs' command
• output from 'mount' command
• /etc/mtab
• lsof | grep hdb

But cat /proc/partitions still lists the partitioning, and hdparm -z /dev/hdb still gives me device busy.

Is there a something I'm missing, or a secret place I don't yet know about to find what's still holding on to my block device? and more importantly, How can I release it's hold so I can reload the partition table?

FWIW, on this specific case, I can simply reboot the server w/o much worry, but this has plagued me before, and I'm curious if there is a better way.

(Edit: added more precise wording) (Edit: details re repartitioning)

Update: I used partprobe /dev/hdb, and it did change things: in /dev /hdb1, /deb/hdb[3-5] are now gone, and partprobe is reporting Error: Error informing the kernel about modifications to partition /dev/hdb1 -- Device or resource busy. <-- specifically about hdb1. hdb1 formerly was a Physical Volume (PV) in a LVM Volume Group (VG), abut i vgremove + pvremoved em before I repartitioned......

Update 2: FWIW, I still haven't corrected this problem, fortunatly it's not urgent. I've learned that partprobe is using a newer API call which is why it did seem to do something earlier. I still haven't found an simple and effective way to, given a device, and it's major/minor numbers, figure out which resources (kernel or userspace) are using it. Any ideas?

Solution 1

Try using fuser

fuser -vam /dev/hdb1

Solution 2

Eddy's fuser -vam /dev/hdb1 example was essentially correct, but it lacked some completeness. In my case, I ran into a similar problem while recovering files from someone off the last drive of a raid1 array where the partition holding the data was in LVM.

In this case, I had started photorec to examine the drive, seen that there was a volume group, and then closed the terminal running photorec. Unbeknownst to me, photorec was still holding on to /dev/mapper/vg0-lv0. So, in the future, try using fuser, but on the contents of /dev/mapper/

fuser -vam /dev/mapper/*

This still probably not the best answer, but remember to try to check against any files under /dev/ that might also map in some way to the block device you're trying to use.

Comments

[Sunny]

Check again with pvdisplay and vgdisplay if this partition is still assigned to any LVM

[3dinfluence]

disregard my answer...I see that you did lsof at the bottom of your list....missed that when I answered the question.

[Jason]

mount | grep hdb doesn't list anything; mount | grep /oldmountpoint doesn't doesn't list anything under the old mount point, and only shows the new device @ the old mount point, that I had moved stuff to. I was able to completely unmount it before the repartition, which indicates (to me at least) that there was nothing binding under it, and it wasn't bound elsewhere.

[Jason]

I tried fuser on /dev/hdb and /dev/hdb[1-5], and it's still reporting nothing using those devices.

[Jason]

FWIW, I never fully resolved this, but I've since restarted the machine, and that corrected it. Thanks for your help!