[dm-crypt] dm-integrity standalone - lack of metadata

[dm-crypt] dm-integrity standalone - lack of metadata

[Harald Braumann]

Hi!

I'm using the dm-integrity target standalone with crc32c for component
devices of a RAID. I don't use journaling, because the perfomance hit
is too big and I don't really need it in that case.

I want to automatically activate these targets using udev
rules (see below). However, the superblock lacks some crucial
information:

- Integrity algorithm
The algorithm is not stored in the superblock and so I have to use
crc32 hard-coded in my udev rule. This is of course no general
solution. I don't quite understand why it is not stored in the
superblock. Is there a technical reason for that?

- Journal mode
I open the integrity devices without journaling. But again, that's
hard-coded and it would a be problem if I had other devices, where I
actually want journaling. It would be nice, if the superblock
contained a use_journal flag so the default could be configured on the
device itself.

- UUID/label support
Not strictly required for my use case, but would be nice for
persistent naming.

While playing around with that, I also noticed that integritysetup
format is extremly slow (this is on an HDD):

integritysetup format /dev/sdb2 -v --integrity crc32c --sector-size
4096
=> 39.6 MiB/s

Format with --no-wipe, open the target and use dd:
dd if=/dev/zero of=/dev/mapper/int bs=1M
=> 63 MiB/s

Same as above, but opening without journal (which really isn't
required for that step, even if the final target should be journaled):
=> 135 MiB/s

I have absolutely no problem with using the dd method, but it kind of
makes the built-in format method redundant. 

Best regards
harry

/etc/udev/rules.d/70-dm-integrity.rules:

ACTION!="add", GOTO="dm_integrity_end"
SUBSYSTEM!="block", GOTO="dm_integrity_end"
ENV{ID_FS_TYPE}!="DM_integrity", GOTO="dm_integrity_end"

RUN+="/sbin/integritysetup open $devnode $kernel-integrity --integrity crc32c --integrity-no-journal"

LABEL="dm_integrity_end"

Re: [dm-crypt] dm-integrity standalone - lack of metadata

[Milan Broz]

(added cc to dm-devel annd Mikulas)

Hi,

On 06/24/2018 11:53 AM, Harald Braumann wrote:
> Hi!
> 
> I'm using the dm-integrity target standalone with crc32c for component
> devices of a RAID. I don't use journaling, because the perfomance hit
> is too big and I don't really need it in that case.
> 
> I want to automatically activate these targets using udev
> rules (see below). However, the superblock lacks some crucial
> information:


yes, I agree. The reason is that dm-integrity was in the first
place designed for authenticated encryption (to store auth tags).
Then additional metadata ia stored in LUKS2 header.

For the rest I just I have not enough energy (and real-world use cases)
to convince Mikulas that we need attributes below in superblock (as the dm-integrity
code is written mostly by him).
At least for the algorithm I mention this idea several times ;-)

There are some patches for dm-integrity targeted to 4.19 floating around,
so let's add these ideas to the mix.

> - Integrity algorithm
> The algorithm is not stored in the superblock and so I have to use
> crc32 hard-coded in my udev rule. This is of course no general
> solution. I don't quite understand why it is not stored in the
> superblock. Is there a technical reason for that?
> 
> - Journal mode
> I open the integrity devices without journaling. But again, that's
> hard-coded and it would a be problem if I had other devices, where I
> actually want journaling. It would be nice, if the superblock
> contained a use_journal flag so the default could be configured on the
> device itself.
> 
> - UUID/label support
> Not strictly required for my use case, but would be nice for
> persistent naming.
> 
> While playing around with that, I also noticed that integritysetup
> format is extremly slow (this is on an HDD):
> 
> integritysetup format /dev/sdb2 -v --integrity crc32c --sector-size
> 4096
> => 39.6 MiB/s

This can be perhaps fine-tuned, but for standalone integrity I expect
we will use internal wipe, so this will be just fallback.
See https://www.redhat.com/archives/dm-devel/2018-June/msg00089.html

> Format with --no-wipe, open the target and use dd:
> dd if=/dev/zero of=/dev/mapper/int bs=1M
> => 63 MiB/s
> 
> Same as above, but opening without journal (which really isn't
> required for that step, even if the final target should be journaled):
> => 135 MiB/s

Internal wipe also uses activation without journal, it is slow probably
because we use direct-io (can you add oflag=direct to your dd command?
Is it the same speed as integritysetup wipe?)

> I have absolutely no problem with using the dd method, but it kind of
> makes the built-in format method redundant. 
> 
> Best regards
> harry
> 
> /etc/udev/rules.d/70-dm-integrity.rules:
> 
> ACTION!="add", GOTO="dm_integrity_end"
> SUBSYSTEM!="block", GOTO="dm_integrity_end"
> ENV{ID_FS_TYPE}!="DM_integrity", GOTO="dm_integrity_end"
> 
> RUN+="/sbin/integritysetup open $devnode $kernel-integrity --integrity crc32c --integrity-no-journal"
> 
> LABEL="dm_integrity_end"

This is not the best way - integritysetup itself need udev synchronization, so it will
race here with udev.

I think we need similar system like /etc/crypttab (or extend crypttab of integrity targets).

Milan

Re: [dm-crypt] dm-integrity standalone - lack of metadata

[Mikulas Patocka]

On Sun, 24 Jun 2018, Milan Broz wrote:

> (added cc to dm-devel annd Mikulas)
> 
> Hi,
> 
> On 06/24/2018 11:53 AM, Harald Braumann wrote:
> > Hi!
> > 
> > I'm using the dm-integrity target standalone with crc32c for component
> > devices of a RAID. I don't use journaling, because the perfomance hit
> > is too big and I don't really need it in that case.
> > 
> > I want to automatically activate these targets using udev
> > rules (see below). However, the superblock lacks some crucial
> > information:
> 
> 
> yes, I agree. The reason is that dm-integrity was in the first
> place designed for authenticated encryption (to store auth tags).
> Then additional metadata ia stored in LUKS2 header.
> 
> For the rest I just I have not enough energy (and real-world use cases)
> to convince Mikulas that we need attributes below in superblock (as the dm-integrity
> code is written mostly by him).

You can write whatever you want to the end of the superblock. If 
kernelspace allocates superblock entries from the beginning and userspace 
allocates them from the end, they won't cross each other.

Mikulas

Re: [dm-crypt] dm-integrity standalone - lack of metadata

[Mikulas Patocka]

On Sun, 24 Jun 2018, Milan Broz wrote:

> (added cc to dm-devel annd Mikulas)
> 
> Hi,
> 
> On 06/24/2018 11:53 AM, Harald Braumann wrote:
> > Hi!
> > 
> > I'm using the dm-integrity target standalone with crc32c for component
> > devices of a RAID. I don't use journaling, because the perfomance hit
> > is too big and I don't really need it in that case.
> > 
> > I want to automatically activate these targets using udev
> > rules (see below). However, the superblock lacks some crucial
> > information:
> 
> 
> yes, I agree. The reason is that dm-integrity was in the first
> place designed for authenticated encryption (to store auth tags).
> Then additional metadata ia stored in LUKS2 header.
> 
> For the rest I just I have not enough energy (and real-world use cases)
> to convince Mikulas that we need attributes below in superblock (as the dm-integrity
> code is written mostly by him).

You can write whatever you want to the end of the superblock. If 
kernelspace allocates superblock entries from the beginning and userspace 
allocates them from the end, they won't cross each other.

Mikulas

Re: [dm-crypt] dm-integrity standalone - lack of metadata

[Milan Broz]

On 06/25/2018 08:05 PM, Mikulas Patocka wrote:
> On Sun, 24 Jun 2018, Milan Broz wrote:
> 
>> (added cc to dm-devel annd Mikulas)
>>
>> Hi,
>>
>> On 06/24/2018 11:53 AM, Harald Braumann wrote:
>>> Hi!
>>>
>>> I'm using the dm-integrity target standalone with crc32c for component
>>> devices of a RAID. I don't use journaling, because the perfomance hit
>>> is too big and I don't really need it in that case.
>>>
>>> I want to automatically activate these targets using udev
>>> rules (see below). However, the superblock lacks some crucial
>>> information:
>>
>>
>> yes, I agree. The reason is that dm-integrity was in the first
>> place designed for authenticated encryption (to store auth tags).
>> Then additional metadata ia stored in LUKS2 header.
>>
>> For the rest I just I have not enough energy (and real-world use cases)
>> to convince Mikulas that we need attributes below in superblock (as the dm-integrity
>> code is written mostly by him).
> 
> You can write whatever you want to the end of the superblock. If 
> kernelspace allocates superblock entries from the beginning and userspace 
> allocates them from the end, they won't cross each other.

This is really hackish solution :)

I think the superblock should be either maintained inside kernel or in userspace.

Or at least kernel superblock should define area reserved for userspace
(IOW the area that kernel guarantees to be never overwritten by kernel driver).

I really do not want to define any userspace-only attributes until there
is some plan how (and if) to integrate plain dm-integrity devices.

Milan

Re: [dm-crypt] dm-integrity standalone - lack of metadata

[Milan Broz]

On 06/25/2018 08:05 PM, Mikulas Patocka wrote:
> On Sun, 24 Jun 2018, Milan Broz wrote:
> 
>> (added cc to dm-devel annd Mikulas)
>>
>> Hi,
>>
>> On 06/24/2018 11:53 AM, Harald Braumann wrote:
>>> Hi!
>>>
>>> I'm using the dm-integrity target standalone with crc32c for component
>>> devices of a RAID. I don't use journaling, because the perfomance hit
>>> is too big and I don't really need it in that case.
>>>
>>> I want to automatically activate these targets using udev
>>> rules (see below). However, the superblock lacks some crucial
>>> information:
>>
>>
>> yes, I agree. The reason is that dm-integrity was in the first
>> place designed for authenticated encryption (to store auth tags).
>> Then additional metadata ia stored in LUKS2 header.
>>
>> For the rest I just I have not enough energy (and real-world use cases)
>> to convince Mikulas that we need attributes below in superblock (as the dm-integrity
>> code is written mostly by him).
> 
> You can write whatever you want to the end of the superblock. If 
> kernelspace allocates superblock entries from the beginning and userspace 
> allocates them from the end, they won't cross each other.

This is really hackish solution :)

I think the superblock should be either maintained inside kernel or in userspace.

Or at least kernel superblock should define area reserved for userspace
(IOW the area that kernel guarantees to be never overwritten by kernel driver).

I really do not want to define any userspace-only attributes until there
is some plan how (and if) to integrate plain dm-integrity devices.

Milan

Re: [dm-crypt] dm-integrity standalone - lack of metadata

[Mikulas Patocka]

On Tue, 3 Jul 2018, Milan Broz wrote:

> On 06/25/2018 08:05 PM, Mikulas Patocka wrote:
> > On Sun, 24 Jun 2018, Milan Broz wrote:
> > 
> >> (added cc to dm-devel annd Mikulas)
> >>
> >> Hi,
> >>
> >> On 06/24/2018 11:53 AM, Harald Braumann wrote:
> >>> Hi!
> >>>
> >>> I'm using the dm-integrity target standalone with crc32c for component
> >>> devices of a RAID. I don't use journaling, because the perfomance hit
> >>> is too big and I don't really need it in that case.
> >>>
> >>> I want to automatically activate these targets using udev
> >>> rules (see below). However, the superblock lacks some crucial
> >>> information:
> >>
> >>
> >> yes, I agree. The reason is that dm-integrity was in the first
> >> place designed for authenticated encryption (to store auth tags).
> >> Then additional metadata ia stored in LUKS2 header.
> >>
> >> For the rest I just I have not enough energy (and real-world use cases)
> >> to convince Mikulas that we need attributes below in superblock (as the dm-integrity
> >> code is written mostly by him).
> > 
> > You can write whatever you want to the end of the superblock. If 
> > kernelspace allocates superblock entries from the beginning and userspace 
> > allocates them from the end, they won't cross each other.
> 
> This is really hackish solution :)

Then - there's an "offset" argument - you can use it and create your own 
superblock before the offset.

> I think the superblock should be either maintained inside kernel or in userspace.
> 
> Or at least kernel superblock should define area reserved for userspace
> (IOW the area that kernel guarantees to be never overwritten by kernel driver).
> 
> I really do not want to define any userspace-only attributes until there
> is some plan how (and if) to integrate plain dm-integrity devices.
> 
> Milan

Mikulas

Re: [dm-crypt] dm-integrity standalone - lack of metadata

[Mikulas Patocka]

On Tue, 3 Jul 2018, Milan Broz wrote:

> On 06/25/2018 08:05 PM, Mikulas Patocka wrote:
> > On Sun, 24 Jun 2018, Milan Broz wrote:
> > 
> >> (added cc to dm-devel annd Mikulas)
> >>
> >> Hi,
> >>
> >> On 06/24/2018 11:53 AM, Harald Braumann wrote:
> >>> Hi!
> >>>
> >>> I'm using the dm-integrity target standalone with crc32c for component
> >>> devices of a RAID. I don't use journaling, because the perfomance hit
> >>> is too big and I don't really need it in that case.
> >>>
> >>> I want to automatically activate these targets using udev
> >>> rules (see below). However, the superblock lacks some crucial
> >>> information:
> >>
> >>
> >> yes, I agree. The reason is that dm-integrity was in the first
> >> place designed for authenticated encryption (to store auth tags).
> >> Then additional metadata ia stored in LUKS2 header.
> >>
> >> For the rest I just I have not enough energy (and real-world use cases)
> >> to convince Mikulas that we need attributes below in superblock (as the dm-integrity
> >> code is written mostly by him).
> > 
> > You can write whatever you want to the end of the superblock. If 
> > kernelspace allocates superblock entries from the beginning and userspace 
> > allocates them from the end, they won't cross each other.
> 
> This is really hackish solution :)

Then - there's an "offset" argument - you can use it and create your own 
superblock before the offset.

> I think the superblock should be either maintained inside kernel or in userspace.
> 
> Or at least kernel superblock should define area reserved for userspace
> (IOW the area that kernel guarantees to be never overwritten by kernel driver).
> 
> I really do not want to define any userspace-only attributes until there
> is some plan how (and if) to integrate plain dm-integrity devices.
> 
> Milan

Mikulas

Re: [dm-crypt] dm-integrity standalone - lack of metadata

[Harald Braumann]

On 24 June 2018 13:01:06 CEST, Milan Broz <gmazyland@gmail.com> wrote:
>(added cc to dm-devel annd Mikulas)

>On 06/24/2018 11:53 AM, Harald Braumann wrote:
>> However, the superblock lacks some crucial information:

>yes, I agree. The reason is that dm-integrity was in the first
>place designed for authenticated encryption (to store auth tags).
>Then additional metadata ia stored in LUKS2 header.
I understand.

>For the rest I just I have not enough energy (and real-world use cases)
>to convince Mikulas that we need attributes below in superblock (as the
>dm-integrity
>code is written mostly by him).
>At least for the algorithm I mention this idea several times ;-)
OK, good. I was just afraid that there might be some technical reason and this was out of the question.

>> integritysetup format /dev/sdb2 -v --integrity crc32c --sector-size
>> 4096
>> => 39.6 MiB/s
>>
>> Format with --no-wipe, open the target and use dd:
>> dd if=/dev/zero of=/dev/mapper/int bs=1M
>> => 63 MiB/s
>> 
>> Same as above, but opening without journal (which really isn't
>> required for that step, even if the final target should be journaled):
>> => 135 MiB/s

>Internal wipe also uses activation without journal, it is slow probably
>because we use direct-io (can you add oflag=direct to your dd command?
>Is it the same speed as integritysetup wipe?)
Sorry, I haven't gotten around to it, yet. 

>> /etc/udev/rules.d/70-dm-integrity.rules:
>> 
>> ACTION!="add", GOTO="dm_integrity_end"
>> SUBSYSTEM!="block", GOTO="dm_integrity_end"
>> ENV{ID_FS_TYPE}!="DM_integrity", GOTO="dm_integrity_end"
>> 
>> RUN+="/sbin/integritysetup open $devnode $kernel-integrity
>--integrity crc32c --integrity-no-journal"
>> 
>> LABEL="dm_integrity_end"

>This is not the best way - integritysetup itself need udev
>synchronization, so it will
>race here with udev.
Could you elaborate on that? How does it need udev and could I add anything to the rules to ensure proper synchronization?

>I think we need similar system like /etc/crypttab (or extend crypttab
>of integrity targets).
Thats what I wanted to avoid. I would like it to work like LVM where volumes can be set up automatically with just metadata stored on the device itself.

Best regards
harry

Re: dm-integrity standalone - lack of metadata

[Harald Braumann]

On 24 June 2018 13:01:06 CEST, Milan Broz <gmazyland-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org> wrote:
>(added cc to dm-devel annd Mikulas)

>On 06/24/2018 11:53 AM, Harald Braumann wrote:
>> However, the superblock lacks some crucial information:

>yes, I agree. The reason is that dm-integrity was in the first
>place designed for authenticated encryption (to store auth tags).
>Then additional metadata ia stored in LUKS2 header.
I understand.

>For the rest I just I have not enough energy (and real-world use cases)
>to convince Mikulas that we need attributes below in superblock (as the
>dm-integrity
>code is written mostly by him).
>At least for the algorithm I mention this idea several times ;-)
OK, good. I was just afraid that there might be some technical reason and this was out of the question.

>> integritysetup format /dev/sdb2 -v --integrity crc32c --sector-size
>> 4096
>> => 39.6 MiB/s
>>
>> Format with --no-wipe, open the target and use dd:
>> dd if=/dev/zero of=/dev/mapper/int bs=1M
>> => 63 MiB/s
>> 
>> Same as above, but opening without journal (which really isn't
>> required for that step, even if the final target should be journaled):
>> => 135 MiB/s

>Internal wipe also uses activation without journal, it is slow probably
>because we use direct-io (can you add oflag=direct to your dd command?
>Is it the same speed as integritysetup wipe?)
Sorry, I haven't gotten around to it, yet. 

>> /etc/udev/rules.d/70-dm-integrity.rules:
>> 
>> ACTION!="add", GOTO="dm_integrity_end"
>> SUBSYSTEM!="block", GOTO="dm_integrity_end"
>> ENV{ID_FS_TYPE}!="DM_integrity", GOTO="dm_integrity_end"
>> 
>> RUN+="/sbin/integritysetup open $devnode $kernel-integrity
>--integrity crc32c --integrity-no-journal"
>> 
>> LABEL="dm_integrity_end"

>This is not the best way - integritysetup itself need udev
>synchronization, so it will
>race here with udev.
Could you elaborate on that? How does it need udev and could I add anything to the rules to ensure proper synchronization?

>I think we need similar system like /etc/crypttab (or extend crypttab
>of integrity targets).
Thats what I wanted to avoid. I would like it to work like LVM where volumes can be set up automatically with just metadata stored on the device itself.

Best regards
harry
_______________________________________________
dm-crypt mailing list
dm-crypt-4q3lyFh4P1g@public.gmane.org
https://www.saout.de/mailman/listinfo/dm-crypt