Discard is not working

Hello,

after reading the "invitation" for some one-off questions without subscribing excuse my disturbing, please.

At first my setup of host:

* in general Debian Wheezy
* Kernel: 3.11-0.bpo.2-amd64
(http://packages.debian.org/wheezy-backports/linux-image-3.11-0.bpo.2-amd64)
* LVM 2.02.98 (http://packages.debian.org/jessie/lvm2)
* thin-provisioning-tools 0.2.8-1
(http://packages.debian.org/jessie/thin-provisioning-tools)
* Qemu-KVM: 1.7.0 (http://packages.debian.org/jessie/qemu-kvm)

The new kernel and dirty things from Jessie are for working thin volumes with discard on my ssd. Everything's fine up to this point.

Afterwards I start a guest (also Debian Wheezy) over the following command (over libvirt) as root (for test only):

kvm [...] -drive file=/dev/ssd0/sarabi,if=none,id=drive-scsi0-0-0-0,format=raw,discard=unmap -device scsi-hd,bus=scsi0.0,channel=0,scsi-id=0,lun=0,drive=drive-scsi0-0-0-0,id=scsi0-0-0-0,bootindex=2 [...]

Neither lsblk -D nor the try of fstrim ends positive. Both say that discard is not supported.
I read that guests need at least kernel 3.4 to support discard (and virtio-scsi). But the same backports kernel 3.11 did not changed anything.

Did I understand something wrong? Is the command wrong? Are any requirements not met?

Bye Markus

Hello,

I did not get response to my private mail address. So I'm subscribed now.

Bye Markus

-------- Original-Nachricht --------
Betreff: Discard is not working
Datum: Wed, 05 Feb 2014 23:53:56 +0100
Von: [email protected]
An: [email protected]

Hello,

after reading the "invitation" for some one-off questions without subscribing excuse my disturbing, please.

At first my setup of host:

* in general Debian Wheezy
* Kernel: 3.11-0.bpo.2-amd64 (http://packages.debian.org/wheezy-backports/linux-image-3.11-0.bpo.2-amd64)
* LVM 2.02.98 (http://packages.debian.org/jessie/lvm2)
* thin-provisioning-tools 0.2.8-1 (http://packages.debian.org/jessie/thin-provisioning-tools)
* Qemu-KVM: 1.7.0 (http://packages.debian.org/jessie/qemu-kvm)

The new kernel and dirty things from Jessie are for working thin volumes with discard on my ssd. Everything's fine up to this point.

Afterwards I start a guest (also Debian Wheezy) over the following command (over libvirt) as root (for test only):

kvm [...] -drive file=/dev/ssd0/sarabi,if=none,id=drive-scsi0-0-0-0,format=raw,discard=unmap -device scsi-hd,bus=scsi0.0,channel=0,scsi-id=0,lun=0,drive=drive-scsi0-0-0-0,id=scsi0-0-0-0,bootindex=2 [...]

Neither lsblk -D nor the try of fstrim ends positive. Both say that discard is not supported.
I read that guests need at least kernel 3.4 to support discard (and virtio-scsi). But the same backports kernel 3.11 did not changed anything.

Did I understand something wrong? Is the command wrong? Are any requirements not met?

Bye Markus

This is the first time I tried to deal with discard. I've no discard-capable hardware, so never bothered to even look how it works (or not).

I fired up my sample linux guest here, using 3.10 kernel on host and 3.2 kernel on guest, using something similar to what your command looks like:

-device megasas,id=scsi0 \
-drive file=test.raw,if=none,id=sd0,discard=unmap \
-device scsi-hd,bus=scsi0.0,channel=0,scsi-id=0,lun=0,drive=sd0

Note: this is a plain file (test.raw) on an ext4 filesystem on a device without discard support (regular rotational hard drive).

In the guest, I can see discard granularity (4K) for that drive in lsblk -D output.

fstrim does not error out, but I don't know if it actually does something or not.

The same result can be observed with 3.10 kernel in guest, and with other scsi devices/drivers (eg lsi53c895a).

However, when I export the same test.raw file using virtio, like this:

-drive file=test.raw,if=virtio,discard=unmap

it doesn't work: guest does not see the virtio drive as discardable.

So I've no idea what's going on here... ;)

Thanks,

/mjt

Hello,

I played around with raw images now.

Megasas is not supported by libvirt. The default controller crashs during installation while accessing to disk.

qemu-system-x86_64:
/build/qemu-j0uSxv/qemu-1.7.0+dfsg/hw/scsi/lsi53c895a.c:541:
lsi_do_dma: Assertion `s->current' failed.

So I took Debian Testing (Jessie with kernel 3.12) to use virtio-scsi over following options.

   -device virtio-scsi-pci,id=scsi0,bus=pci.0,addr=0x4 \
   -drive
   file=/home/vm/trim.raw,if=none,id=drive-scsi0-0-0-0,format=raw,discard=unmap
   \
   -device
   scsi-hd,bus=scsi0.0,channel=0,scsi-id=0,lun=0,drive=drive-scsi0-0-0-0,id=scsi0-0-0-0,bootindex=2

Result of this: fstrim works.
Take a look into lsblk -D: DISC-ZERO=0? I read DISC-ZERO has to be "1" to propagate discard.

Switch back to my thin volume vm. DISC-ZERO=0 and fstrim works! Booting from old kernel 3.2 - works too. LVM recognize the increasing of thin volume after discard.

Conclusion of my issues:

Yes!
Discard granularity (DISC-GRAN) of lsblk -D is the important number. It must not be "0B".

I did so much the last weeks. Maybe the error of fstrim occurs before getting thin volumes work. Thanks a lot for your help (and hint to try it with an image)!

Bye Markus

Am 07.02.2014 17:36, schrieb Michael Tokarev:

Il 07/02/2014 17:36, Michael Tokarev ha scritto:

Thanks Michael for the experiments! I missed this email.

virtio-blk does not support discard, so this is expected.

Regarding lsblk -D, DISC-ZERO is never true in VMs (not yet at least!), but the important one is DISC-GRAN.

Paolo