What is the deal with the partition separator?

What is the deal with the partition separator?

[Phillip Susi]

It used to be that partitions device names just had a digit added on to 
the base disk device name.  It seems that this became problematic at 
some point with device mapper and oddly named disks, and there have been 
several responses to it:

1)  dmraid and (lib)parted now always add a 'p' between the base name 
and the partition number

2)  kpartx from multipath-tools adds the 'p' only if the base name ends 
in a digit

3)  Debian and Ubuntu's udev and init scripts tell kpartx to use 'part' 
instead of just 'p'.

4)  gparted now explicitly tells dmraid to not use any character so that 
it behaves like older versions and is therefore compatible with the 
kpartx method that gparted has adopted, at least when the base name does 
not end with a digit.

Each of these components needs to agree on what the correct name is or 
chaos ensues.  I would like to discuss the merits of each and try to 
decide on a standard.

Having thought about it for a moment, it seems to me that deciding on 
always adding the 'p' is the way to go, since the 'art' just makes 
things longer for no good reason, and if you only sometimes add the 'p' 
then you can't tell if a device name that ends in a digit that does not 
follow a 'p' is a whole disk, or a partition.

Re: What is the deal with the partition separator?

[Hannes Reinecke]

On 02/16/2011 04:21 AM, Phillip Susi wrote:
> It used to be that partitions device names just had a digit added on to
> the base disk device name.  It seems that this became problematic at
> some point with device mapper and oddly named disks, and there have been
> several responses to it:
> 
> 1)  dmraid and (lib)parted now always add a 'p' between the base name
> and the partition number
> 
> 2)  kpartx from multipath-tools adds the 'p' only if the base name ends
> in a digit
> 
> 3)  Debian and Ubuntu's udev and init scripts tell kpartx to use 'part'
> instead of just 'p'.
> 
> 4)  gparted now explicitly tells dmraid to not use any character so that
> it behaves like older versions and is therefore compatible with the
> kpartx method that gparted has adopted, at least when the base name does
> not end with a digit.
> 
> Each of these components needs to agree on what the correct name is or
> chaos ensues.  I would like to discuss the merits of each and try to
> decide on a standard.
> 
> Having thought about it for a moment, it seems to me that deciding on
> always adding the 'p' is the way to go, since the 'art' just makes
> things longer for no good reason, and if you only sometimes add the 'p'
> then you can't tell if a device name that ends in a digit that does not
> follow a 'p' is a whole disk, or a partition.
> 
No.

The linux scheme since the dawn of time is to
a) Add the partition number to the device node name
b) If last letter of the device node name is a number, insert a 'p'
between device node name and partition number

with the advent of persistent device names (via udev) the partition
separator (for persistent links only!) is '-part'.

So you have

/dev/sda
/dev/sda1
/dev/disk/by-id/scsi-0WhatAStupidName
/dev/disk/by-id/scsi-0WhatAStupidName-part1

persistent device names are longish anyway, so we can as well use
something readable for partitions.

Cheers,

Hannes
-- 
Dr. Hannes Reinecke		      zSeries & Storage
hare@suse.de			      +49 911 74053 688
SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg
GF: Markus Rex, HRB 16746 (AG Nürnberg)

Re: [dm-devel] What is the deal with the partition separator?

[Curtis Gedak]

On 11-02-16 12:38 AM, Hannes Reinecke wrote:
> On 02/16/2011 04:21 AM, Phillip Susi wrote:
>> It used to be that partitions device names just had a digit added on to
>> the base disk device name.  It seems that this became problematic at
>> some point with device mapper and oddly named disks, and there have been
>> several responses to it:
>>
>> 1)  dmraid and (lib)parted now always add a 'p' between the base name
>> and the partition number

Only newer versions of (lib)parted which link with the dmraid library 
work this way with dmraid devices.  In this situation it is the dmraid 
library that is coming up with the device names.  For all other devices, 
(lib)parted uses the "linux scheme since the dawn of time" mentioned 
below.  Older versions of (lib)parted follow the "linux scheme since the 
dawn of time".

>> 2)  kpartx from multipath-tools adds the 'p' only if the base name ends
>> in a digit

kpartx follows the "linux scheme since the dawn of time" mentioned below.

>> 4)  gparted now explicitly tells dmraid to not use any character so that
>> it behaves like older versions and is therefore compatible with the
>> kpartx method that gparted has adopted, at least when the base name does
>> not end with a digit.
>>

GParted uses (lib)parted to come up with device names, except in the 
case of dmraid devices.  Because dmraid does not follow the "linux 
scheme since the dawn of time", gparted calls dmraid directly to create 
names following the old dmraid standard (only append partition number to 
device name).  If kpartx is available, then GParted will call kpartx to 
ensure that dmraid device names follow the "linux scheme since the dawn 
of time".  This behaviour enables GParted to work with all versions of 
dmraid, and maintain compatibility with (lib)parted versions 1.7.1 and 
higher.

>> Each of these components needs to agree on what the correct name is or
>> chaos ensues.  I would like to discuss the merits of each and try to
>> decide on a standard.
>>
>> Having thought about it for a moment, it seems to me that deciding on
>> always adding the 'p' is the way to go, since the 'art' just makes
>> things longer for no good reason, and if you only sometimes add the 'p'
>> then you can't tell if a device name that ends in a digit that does not
>> follow a 'p' is a whole disk, or a partition.
>>
> No.
>
> The linux scheme since the dawn of time is to
> a) Add the partition number to the device node name
> b) If last letter of the device node name is a number, insert a 'p'
> between device node name and partition number
>
> with the advent of persistent device names (via udev) the partition
> separator (for persistent links only!) is '-part'.
>
> So you have
>
> /dev/sda
> /dev/sda1
> /dev/disk/by-id/scsi-0WhatAStupidName
> /dev/disk/by-id/scsi-0WhatAStupidName-part1
>
> persistent device names are longish anyway, so we can as well use
> something readable for partitions.
>
>

My thoughts are that if possible, tools should try to create partition 
names using the "linux scheme since the dawn of time", unless there are 
good reasons to not do this.

Regards,
Curtis Gedak
(Maintainer of GParted)

Re: [dm-devel] What is the deal with the partition separator?

[Phillip Susi]

On 2/16/2011 12:19 PM, Curtis Gedak wrote:
> Only newer versions of (lib)parted which link with the dmraid library
> work this way with dmraid devices.  In this situation it is the dmraid
> library that is coming up with the device names.  For all other devices,
> (lib)parted uses the "linux scheme since the dawn of time" mentioned
> below.  Older versions of (lib)parted follow the "linux scheme since the
> dawn of time".

(lib)parted does not link to dmraid; it contains its own code to
generate the name by always adding the 'p' in libparted/arch/linux.c.

> GParted uses (lib)parted to come up with device names, except in the
> case of dmraid devices.  Because dmraid does not follow the "linux
> scheme since the dawn of time", gparted calls dmraid directly to create
> names following the old dmraid standard (only append partition number to
> device name).  If kpartx is available, then GParted will call kpartx to
> ensure that dmraid device names follow the "linux scheme since the dawn
> of time".  This behaviour enables GParted to work with all versions of
> dmraid, and maintain compatibility with (lib)parted versions 1.7.1 and
> higher.

It shouldn't be using both dmraid and kpartx to create the partition
devices.  If dmraid does it, then calling kpartx at best does nothing,
and at worst creates duplicate devices.

Re: [dm-devel] What is the deal with the partition separator?

[Phillip Susi]

On 2/16/2011 2:38 AM, Hannes Reinecke wrote:
> The linux scheme since the dawn of time is to
> a) Add the partition number to the device node name
> b) If last letter of the device node name is a number, insert a 'p'
> between device node name and partition number

Since the dawn of time?  Before device mapper it wasn't possible to have
a disk device name ending in a digit.

> with the advent of persistent device names (via udev) the partition
> separator (for persistent links only!) is '-part'.
> 
> So you have
> 
> /dev/sda
> /dev/sda1
> /dev/disk/by-id/scsi-0WhatAStupidName
> /dev/disk/by-id/scsi-0WhatAStupidName-part1
> 
> persistent device names are longish anyway, so we can as well use
> something readable for partitions.

That's fine for the symlinks in by-id, but what about the names in
/dev/mapper and /dev/VG?

Re: [dm-devel] What is the deal with the partition separator?

[Curtis Gedak]

On 11-02-16 12:33 PM, Phillip Susi wrote:
> On 2/16/2011 12:19 PM, Curtis Gedak wrote:
>> Only newer versions of (lib)parted which link with the dmraid library
>> work this way with dmraid devices.  In this situation it is the dmraid
>> library that is coming up with the device names.  For all other devices,
>> (lib)parted uses the "linux scheme since the dawn of time" mentioned
>> below.  Older versions of (lib)parted follow the "linux scheme since the
>> dawn of time".
> (lib)parted does not link to dmraid; it contains its own code to
> generate the name by always adding the 'p' in libparted/arch/linux.c.

My bad.  (lib)parted does not link with dmraid.  It has been a while 
since I looked deeply into the libparted code.

I do know that the current version of parted in git has a 
"--enable-device-mapper" option for configure that is enabled by default.
http://git.debian.org/?p=parted/parted.git;a=blob;f=configure.ac;h=25960a74abe909f7b11580aa053c0689a966820d;hb=HEAD#l81

The function _device_get_part_path in libparted/arc/linux.c, does appear 
to always insert a 'p'  for devices such as PED_DEVICE_DM and 
PED_DEVICE_ATARAID.
http://git.debian.org/?p=parted/parted.git;a=blob;f=libparted/arch/linux.c;h=0288a150769298b92f06d96f8515f96bc27066ea;hb=HEAD#l2198


Regarding the GParted code, it has been written to will work with dmraid 
devices regardless of whether the "--enable-device-mapper" configure 
option is enabled or disabled.

>> GParted uses (lib)parted to come up with device names, except in the
>> case of dmraid devices.  Because dmraid does not follow the "linux
>> scheme since the dawn of time", gparted calls dmraid directly to create
>> names following the old dmraid standard (only append partition number to
>> device name).  If kpartx is available, then GParted will call kpartx to
>> ensure that dmraid device names follow the "linux scheme since the dawn
>> of time".  This behaviour enables GParted to work with all versions of
>> dmraid, and maintain compatibility with (lib)parted versions 1.7.1 and
>> higher.
> It shouldn't be using both dmraid and kpartx to create the partition
> devices.  If dmraid does it, then calling kpartx at best does nothing,
> and at worst creates duplicate devices.

It would be easy in GParted to not call kpartx at all.  This is in the 
code to try to follow the normal partition naming standards for GNU/Linux.

Re: [dm-devel] What is the deal with the partition separator?

[Hannes Reinecke]

On 02/16/2011 08:37 PM, Phillip Susi wrote:
> On 2/16/2011 2:38 AM, Hannes Reinecke wrote:
>> The linux scheme since the dawn of time is to
>> a) Add the partition number to the device node name
>> b) If last letter of the device node name is a number, insert a 'p'
>> between device node name and partition number
> 
> Since the dawn of time?  Before device mapper it wasn't possible to have
> a disk device name ending in a digit.
> 
Unless you happen to have a raid HBA, which has names like

cciss/c0d0

or

rd/c0d0

and, of course, you can have partitions on md devices, too, which
will also give you names like
md0p1

>> with the advent of persistent device names (via udev) the partition
>> separator (for persistent links only!) is '-part'.
>>
>> So you have
>>
>> /dev/sda
>> /dev/sda1
>> /dev/disk/by-id/scsi-0WhatAStupidName
>> /dev/disk/by-id/scsi-0WhatAStupidName-part1
>>
>> persistent device names are longish anyway, so we can as well use
>> something readable for partitions.
> 
> That's fine for the symlinks in by-id, but what about the names in
> /dev/mapper and /dev/VG?
> 
/dev/mapper contains the names of the device-mapper devices
themselves; for those I don't have any preference.
kpartx as called from udev is using the '-part' suffix here, too.

Cheers,

Hannes
-- 
Dr. Hannes Reinecke		      zSeries & Storage
hare-l3A5Bk7waGM@public.gmane.org			      +49 911 74053 688
SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg
GF: Markus Rex, HRB 16746 (AG Nürnberg)

Re: [dm-devel] What is the deal with the partition separator?

[Phillip Susi]

On 2/17/2011 2:29 AM, Hannes Reinecke wrote:
> /dev/mapper contains the names of the device-mapper devices
> themselves; for those I don't have any preference.
> kpartx as called from udev is using the '-part' suffix here, too.

Right; it needs to not do that.  Everyone needs to use the same naming
scheme.  When you change naming schemes, you get duplicate devices.  If
dmraid creates them with 'p' and then udev runs kpartx and tells it to
use '-part', then you get two partition devices, which will cause all
kinds of hell.

Is there any harm done by adding the 'p' even when the base name does
not end in a digit?  Is there any good reason to?

Re: [dm-devel] What is the deal with the partition separator?

[Phillip Susi]

On 2/17/2011 10:13 AM, Hannes Reinecke wrote:
> Which is why you should call 'dmraid' with '-p' to avoid having it
> creating partitions.
> We'll be getting another event via udev, which then trigger kpartx
> to create the partitons there.

Except that won't work if you don't have kpartx installed, and also
users tend to run dmraid directly and don't use -p to change its
behavior.  The same goes with kpartx.  Everyone needs to get on the same
page here and do away with -p switches to deviate from the standard
behavior, or you run into possibly data corrupting problems.

> Which was the reason why I chose the '-part' naming scheme; this way
> it's pretty obvious that a new naming scheme is used. So any
> objections for it not being compliant to the linux naming scheme are
> immediately voided.

Being VERY non compliant does not void objections that you aren't compliant.

Re: [dm-devel] What is the deal with the partition separator?

[Hannes Reinecke]

On 02/17/2011 03:57 PM, Phillip Susi wrote:
> On 2/17/2011 2:29 AM, Hannes Reinecke wrote:
>> /dev/mapper contains the names of the device-mapper devices
>> themselves; for those I don't have any preference.
>> kpartx as called from udev is using the '-part' suffix here, too.
> 
> Right; it needs to not do that.  Everyone needs to use the same naming
> scheme.  When you change naming schemes, you get duplicate devices.  If
> dmraid creates them with 'p' and then udev runs kpartx and tells it to
> use '-part', then you get two partition devices, which will cause all
> kinds of hell.
> 
Which is why you should call 'dmraid' with '-p' to avoid having it
creating partitions.
We'll be getting another event via udev, which then trigger kpartx
to create the partitons there.

> Is there any harm done by adding the 'p' even when the base name does
> not end in a digit?  Is there any good reason to?

The main problem here is that each and everyone has their own
preferred way of naming.
Most tend to stick to the linux model (occasionally inserting a
'p'), some tend to use persistent device names, some tend to use
device-mapper / LVM and do away with all partitions etc.

So whichever way you proceed, you should avoid using a naming scheme
which resembles the linux one.
Otherwise inevitably someone doesn't get it and starts complaining,
one way of the other.

Which was the reason why I chose the '-part' naming scheme; this way
it's pretty obvious that a new naming scheme is used. So any
objections for it not being compliant to the linux naming scheme are
immediately voided.

Cheers,

Hannes
-- 
Dr. Hannes Reinecke		      zSeries & Storage
hare-l3A5Bk7waGM@public.gmane.org			      +49 911 74053 688
SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg
GF: Markus Rex, HRB 16746 (AG Nürnberg)

Re: What is the deal with the partition separator?

[Alasdair G Kergon]

Partitions should be created only once, in one place.

If a second instance of anything tries to create them too, it should,
by default, detect that they are already there and not proceed.

You should not be able to run kpartx twice on the same underlying
disk with different delimiters without some sort of 'force' option.

(Check by querying /sys/.../holders ?)

Alasdair

Re: [dm-devel] What is the deal with the partition separator?

[Curtis Gedak]

On 11-02-17 08:13 AM, Hannes Reinecke wrote:
> So whichever way you proceed, you should avoid using a naming scheme
> which resembles the linux one.

 From my perspective as a maintainer of partitioning software it would 
be nice if given the device name and partition number I could easily 
determine the partition name.

Unfortunately this is not easily done when there are deviations from the 
linux model ("insert a 'p' between the device name and the partition 
number when the device name ends in a number").

Each deviation from the linux partition naming model results in more 
code to handle the deviation, and a corresponding increased number of 
test cases to avoid errors.

Regards,
Curtis Gedak
(Maintainer of GParted)

Re: What is the deal with the partition separator?

[Alasdair G Kergon]

On Thu, Feb 17, 2011 at 04:53:32PM +0000, Alasdair G Kergon wrote:
> (Check by querying /sys/.../holders ?)
 
Actually, would it be sufficient for kpartx to attempt an exclusive
open?

Alasdair

Re: [dm-devel] What is the deal with the partition separator?

[Phillip Susi]

Oops, didn't see this until today because this darn mailing list mangles 
the reply-to header. *shakes fist*

Currently all of the tools check by trying to open the expected device 
name.  If they have different ideas of what that is, the check fails to 
find the existing instance.  It is looking to me like the way to fix 
this is to patch dmraid and libparted so that they conform with the 
kpartx and "linux since the dawn of time" method of only adding the 'p' 
if the previous character is a digit.

On 02/17/2011 11:53 AM, Alasdair G Kergon wrote:
> Partitions should be created only once, in one place.
>
> If a second instance of anything tries to create them too, it should,
> by default, detect that they are already there and not proceed.
>
> You should not be able to run kpartx twice on the same underlying
> disk with different delimiters without some sort of 'force' option.
>
> (Check by querying /sys/.../holders ?)