device syntax again

device syntax again

[Hollis Blanchard]

I'd like to discuss device syntax again. Let me start by giving an 
example. When a client is booted on an OF-based machine, it can find 
out what the booted device and path was. This is the /chosen/bootpath 
property. Here is the bootpath property from a briQ I just booted:
	ok > boot net:,grubof.chrp
	bootpath = /pci@FF500000/ethernet@7
I would like to load a config file from the same location, so I should 
set the GRUB environment variable "prefix".

(Remember: the original problem I had regarding device syntax was use 
of the comma in device paths. OF paths sometime contain commas, but 
that confuses GRUB's parser, which was written when all it had to worry 
about was the devices on a PC.)

What should I set prefix to if the bootpath property has a comma in it? 
It is not possible to ask the user to reboot and create a devalias for 
this device. Aside from the obvious usability issues, the devalias is 
actually expanded to its true path for the bootpath property. You'll 
notice above I used the alias "net", but bootpath was the full 
"/pci@FF500000/ethernet@7".

-Hollis

Re: device syntax again

[Hollis Blanchard]

On Jan 18, 2005, at 5:04 AM, Yoshinori K. Okuji wrote:

> On Tuesday 18 January 2005 06:25, Hollis Blanchard wrote:
>> Nobody has replied to my previous mail about device syntax, so I will
>> assume that means you all agree that I'm right about requiring
>> different device syntaxes for different architectures. ;)
>
> You assume that aliases are made by a firmware, but I actually meant
> that GRUB could make arbitrary aliases in itself, since GRUB has its
> own device drivers. For example:
>
> alias hd0 /path/to/a/device
>
> This does not have to call Open Firmware, because GRUB itself can map
> "hd0" to "/path/to/a/device" transparently.

(Your mail above is actually regarding device syntax, so I will reply 
to this thread instead.)

For now let's not talk about where aliases are created. What about my 
original question: grub has just booted, and when it asks what device 
it was booted from (the /chosen/bootpath property) it gets this:
	/pci@1f,0/pci@1,1/ide@3/disk@0,0
(that is the real "disk" device on Vincent's UltraSparc). The next 
thing we want to do is load a config file from the same device. What 
value should we put into "prefix"?

Is the only reason you're suggesting this "alias" scheme is to keep the 
PC-style device syntax? Please remember that PC BIOS has no native 
syntax, so grub had to create its own. That is fine, but other systems 
(like OpenBoot/Open Firmware and EFI) *do* have a native syntax, and 
one that does not easily fit into the manufactured syntax. Please do 
not try to unnaturally conform the native syntaxes to a primitive PC 
syntax. That will cause
1) complexity and bugs in code to translate between the two,
2) frustrating UI requirements ("no no, you have to run 'alias' 
first"), and
3) require users to learn yet another syntax.

If you don't agree with any of those three points, please let me know 
and I will provide examples.

-Hollis

Re: device syntax again

[Yoshinori K. Okuji]

On Tuesday 18 January 2005 15:52, Hollis Blanchard wrote:
> For now let's not talk about where aliases are created. What about my
> original question: grub has just booted, and when it asks what device
> it was booted from (the /chosen/bootpath property) it gets this:
> 	/pci@1f,0/pci@1,1/ide@3/disk@0,0
> (that is the real "disk" device on Vincent's UltraSparc). The next
> thing we want to do is load a config file from the same device. What
> value should we put into "prefix"?

This has already been discussed. Marco suggested to have a "boot drive" 
for this.

> Is the only reason you're suggesting this "alias" scheme is to keep
> the PC-style device syntax?

Don't call it "PC-style". This is not accurate. It is GRUB-style as _we_ 
define the syntax.

> 1) complexity and bugs in code to translate between the two,

Why? You can simply register a new device with a device path. What is 
complex?

> 2) frustrating UI requirements ("no no, you have to run 'alias'
> first"), and

GRUB can automatically create aliases.

> 3) require users to learn yet another syntax.

We have already discussed this, and I got no objection to my opinion at 
that time. I'm surprised that you refrain this.

Don't assume that ordinary users know Open Firmware or EFI. Basically, 
they don't know anything quite technical. So GRUB should provide an 
unique syntax rule so that users do not have to learn new things when 
they get new architectures.

And, have you ever asked any ordinary person to 
type /pci@1f,0/pci@1,1/ide@3/disk@0,0 rather than hd0? I'm sure that 
she will be scared. In addition, this would make remote assistance very 
hard. It should be easy to imagine how difficult it is to pronounce the 
long and mysterios string by phone.

As far as I know, most people use /dev/hda instead 
of /dev/ide/host0/bus0/target0/lun0/disc on Linux, whenever possible. 
This is simply because shorter is easier. Ordinary people think "the 
first IDE disk" instead of "the hard disk attached to the master of the 
first IDE bus of the main controller". Don't you agree?

Okuji

Re: device syntax again

[Hollis Blanchard]

On Jan 18, 2005, at 9:36 AM, Yoshinori K. Okuji wrote:

> On Tuesday 18 January 2005 15:52, Hollis Blanchard wrote:
>> For now let's not talk about where aliases are created. What about my
>> original question: grub has just booted, and when it asks what device
>> it was booted from (the /chosen/bootpath property) it gets this:
>> 	/pci@1f,0/pci@1,1/ide@3/disk@0,0
>> (that is the real "disk" device on Vincent's UltraSparc). The next
>> thing we want to do is load a config file from the same device. What
>> value should we put into "prefix"?
>
> This has already been discussed. Marco suggested to have a "boot drive"
> for this.

He mentioned that to me on IRC today, but I had not heard the idea 
before. I believe that will work fine.

>> Is the only reason you're suggesting this "alias" scheme is to keep
>> the PC-style device syntax?
>
> Don't call it "PC-style". This is not accurate. It is GRUB-style as 
> _we_
> define the syntax.

I assumed it was written solely with PCs in mind, as that is where GRUB 
Legacy is used. Obviously, this syntax works perfectly well for PCs, 
but I'm trying to explain that it will not work very well for larger 
systems. That is fine too! I don't think GRUB was originally designed 
with such systems in mind, so of course some changes may be needed if 
we wish to support them.

>> 2) frustrating UI requirements ("no no, you have to run 'alias'
>> first"), and
>
> GRUB can automatically create aliases.

Here is what I see happening:
	> boot /pci@1f,0/<tab><tab><tab>
	>
	> boot /pci@1f,0/pci@1,1/ide@3/disk@0,0
	No such device
	> alias thatdisk /pci@1f,0/pci@1,1/ide@3/disk@0,0
	> boot thatdisk

I don't think GRUB can automatically create aliases in this case.

>> 3) require users to learn yet another syntax.
>
> We have already discussed this, and I got no objection to my opinion at
> that time. I'm surprised that you refrain this.

I agreed at the time, but the more I thought about it I changed my mind.

> Don't assume that ordinary users know Open Firmware or EFI. Basically,
> they don't know anything quite technical.

It is true that many home users do not understand technical things. It 
is also true that advanced users (e.g. users with complex hardware) 
must be allowed to do advanced things without being forced into a 
too-simple model...

> So GRUB should provide an unique syntax rule so that users do not have 
> to learn
> new things when they get new architectures.

Users must be expected to learn how to boot their new architectures... 
In the case of a small Open Firmware system, this need not be any more 
complicated than an x86 PC. In the case of a large Open Firmware 
system, though, users will need to become at least slightly familiar 
with their firmware. I speak from experience here. :)

> And, have you ever asked any ordinary person to
> type /pci@1f,0/pci@1,1/ide@3/disk@0,0 rather than hd0? I'm sure that
> she will be scared. In addition, this would make remote assistance very
> hard. It should be easy to imagine how difficult it is to pronounce the
> long and mysterios string by phone.

I agree, phone support would be difficult, but also very unlikely. :)

An Open Firmware devalias (e.g. "disk" = 
"/pci@1f,0/pci@1,1/ide@3/disk@0,0") is a convenience. It is not 
intended to entirely replace the device tree. Macintosh users will 
likely find that they already have devaliases for the internal disks, 
so you will not have to tell your aunt about the device tree over the 
phone. However, pSeries and Sparc systems may have many more disks than 
devaliases, and these users need to be able to boot easily as well.

> As far as I know, most people use /dev/hda instead
> of /dev/ide/host0/bus0/target0/lun0/disc on Linux, whenever possible.
> This is simply because shorter is easier. Ordinary people think "the
> first IDE disk" instead of "the hard disk attached to the master of the
> first IDE bus of the main controller". Don't you agree?

This is a key point.

Even on small systems, the difficulty in naming is easy to see. I have 
one SCSI disk and one ATA disk. Which is "the first"? On a PC I guess 
you would say "it depends on the BIOS and SCSI BIOS, but the ATA disk 
is probably first". With Open Firmware, the question is unambiguously 
answered.

The systems I use may have dozens of PCI buses. Each bus may contain 
many PCI SCSI adapters, and each adapter may have many disks attached. 
How would you number these devices? (The bus connections do not match 
the physical positions, so you cannot say "the top left disk is #0, the 
next to the right, is #1, etc".)

Open Firmware answers this question too. There are tools that can map a 
device path to a physical location in the chassis, or blink a disk 
light from just a device path. In Linux (and AIX) we already have a 
namespace problem: "I have just installed my system, and now I want to 
boot from /dev/sda. What device name should I use in Open Firmware?" 
Would you add yet a third namespace? In Open Firmware use "disk"; in 
GRUB use "hd0"; in Linux use "/dev/sda"?




Most of us have only had to deal with small computers, and I agree that 
things should be optimized to make life as easy as possible for these 
users. However, please also keep large systems in mind. Try to imagine 
yourself in front of a small rackmount system with only 12 hotplug 
disks wondering which one is /dev/sda under Linux. Then imagine you 
need to hotplug the correct disk into another system, and after telling 
firmware to boot from the new disk, GRUB and Linux "just work". It is 
hard to imagine, but I hope we can agree that such users have different 
needs from our aunts sitting at their PCs at home. Thanks!

-Hollis

Re: device syntax again

[Vincent Pelletier]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Hello.

| Would you add
| yet a third namespace? In Open Firmware use "disk"; in GRUB use "hd0";
| in Linux use "/dev/sda"?

Isn't it possible to support both syntax ?
(hd0,0)/path/file
and
~ (/pci@1f,0/pci@1,1/ide@3/disk@0,0)/path/file
~ or
~ (disk)/path/file

That would keep backward compatibility, offer "low end" users the easy
although ambiguous syntax and "high end" users the full advantage of
their usual and complex Open Firmware syntax.

Maybe even could it be separated in another module (like "boot" and
"boot_of"). It would solve any problem detecting the type of syntax
used, but may be confusing ("I'm using an OF based box, can't I use the
boot module with simple syntax ?").

Vincent Pelletier
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.5 (GNU/Linux)
Comment: Using GnuPG with Thunderbird - http://enigmail.mozdev.org

iD8DBQFB7hMKFEQoKRQyjtURAv7/AJ9l4iSyY/WkziVc9IB3qFFeih+BowCgmQs1
tm8pbKinLhfmYUkNtgcRPbk=
=Yr/2
-----END PGP SIGNATURE-----

Re: device syntax again

[Hollis Blanchard]

On Jan 19, 2005, at 1:58 AM, Vincent Pelletier wrote:
>
> | Would you add
> | yet a third namespace? In Open Firmware use "disk"; in GRUB use 
> "hd0";
> | in Linux use "/dev/sda"?
>
> Isn't it possible to support both syntax ?
> (hd0,0)/path/file
> and
> ~ (/pci@1f,0/pci@1,1/ide@3/disk@0,0)/path/file
> ~ or
> ~ (disk)/path/file

That is exactly what I'd like to see.

Unfortunately, various code throughout GRUB looks for the first comma 
as separating the device name and partition number. So in the first 
example, code would decide that the device name is "/pci@1f" and the 
partition number is "0/pci@1,1/ide@3/disk@0,0", and cause an error.

Okuji and Marco feel it would be too confusing to even allow the first 
example, despite the fact that the second example would continue to 
work.

-Hollis

Re: device syntax again

[Yoshinori K. Okuji]

On Wednesday 19 January 2005 02:59, Hollis Blanchard wrote:
> Here is what I see happening:
> 	> boot /pci@1f,0/<tab><tab><tab>
> 	>
> 	> boot /pci@1f,0/pci@1,1/ide@3/disk@0,0
>
> 	No such device
>
> 	> alias thatdisk /pci@1f,0/pci@1,1/ide@3/disk@0,0
> 	> boot thatdisk
>
> I don't think GRUB can automatically create aliases in this case.

I don't understand this. Could you elaborate on this in details?

> I agreed at the time, but the more I thought about it I changed my
> mind.

:(

> It is true that many home users do not understand technical things.
> It is also true that advanced users (e.g. users with complex
> hardware) must be allowed to do advanced things without being forced
> into a too-simple model...

It is not too simple. You are allowed to make an arbitrary alias on 
GRUB, so you just need to do one more step to access a drive.

alias my_drive /some/path
root (my_drive)

This should be acceptable for experts. Do I miss anything?

> Users must be expected to learn how to boot their new
> architectures... In the case of a small Open Firmware system, this
> need not be any more complicated than an x86 PC. In the case of a
> large Open Firmware system, though, users will need to become at
> least slightly familiar with their firmware. I speak from experience
> here. :)

I don't think so, if the software is good.

For instance, when I installed Debian GNU/Linux into my iBook, all I had 
to learn was how to boot a CD-ROM from Open Firmware (pressing 'c' or 
something like that). The other dirty jobs were done by yaboot and 
ybin. So I didn't need to learn anything about the device syntax.

The reason why I know Open Firmware somehow is simply because I studied 
it intentionally when Gordon and I were thinking of porting GRUB to 
other architectures (and I played Open Firmware with SparcStation and 
PowerBook). For most people, there is no reason why they must learn 
Open Firmware.

> I agree, phone support would be difficult, but also very unlikely. :)

It happens frequently in the support section of each vendor working with 
GNU/Linux. For example, I know Red Hat does this.

> Even on small systems, the difficulty in naming is easy to see. I
> have one SCSI disk and one ATA disk. Which is "the first"? On a PC I
> guess you would say "it depends on the BIOS and SCSI BIOS, but the
> ATA disk is probably first". With Open Firmware, the question is
> unambiguously answered.

I have never said that you should not distinguish SCSI from ATA.

> The systems I use may have dozens of PCI buses. Each bus may contain
> many PCI SCSI adapters, and each adapter may have many disks
> attached. How would you number these devices? (The bus connections do
> not match the physical positions, so you cannot say "the top left
> disk is #0, the next to the right, is #1, etc".)

In this kind of configuration, you should specify a drive logically, 
otherwise it is unmaintainable.

I have a more confusing example in our company. The machine uses an 
internal ATA disk and many Firewire disks. Unlike SCSI, Firewire disks 
do not have IDs, so there is no way to expect how to count (name) the 
disks. And, we sometimes want to connect a new disk and disconnect a 
disk, which must not affect the booting process of the machine.

Currently I'm using disk labels to address this issue. Although GRUB 
does not support disk labels (at the moment), this is not a problem 
only because I use the internal ATA disk to boot the machine.

Suppose that you have a machine only with Firewire disks and you want to 
boot it from one of the disks... Probably any kind of physical way does 
not help you. It is necessary to let GRUB find a desired disk in a 
logical way. You could use a disk label, an UUID or a filename for 
this.

> Most of us have only had to deal with small computers, and I agree
> that things should be optimized to make life as easy as possible for
> these users. However, please also keep large systems in mind. Try to
> imagine yourself in front of a small rackmount system with only 12
> hotplug disks wondering which one is /dev/sda under Linux. Then
> imagine you need to hotplug the correct disk into another system, and
> after telling firmware to boot from the new disk, GRUB and Linux
> "just work". It is hard to imagine, but I hope we can agree that such
> users have different needs from our aunts sitting at their PCs at
> home. Thanks!

Don't worry. I have experience with large systems, and I take it into 
account. I think using disk labels is a good answer for your question.

Okuji

Re: device syntax again

[Hollis Blanchard]

On Jan 19, 2005, at 4:06 AM, Yoshinori K. Okuji wrote:
>
> For instance, when I installed Debian GNU/Linux into my iBook, all I 
> had
> to learn was how to boot a CD-ROM from Open Firmware (pressing 'c' or
> something like that). The other dirty jobs were done by yaboot and
> ybin. So I didn't need to learn anything about the device syntax.

In my proposal, _you can still do that_.

> The reason why I know Open Firmware somehow is simply because I studied
> it intentionally when Gordon and I were thinking of porting GRUB to
> other architectures (and I played Open Firmware with SparcStation and
> PowerBook). For most people, there is no reason why they must learn
> Open Firmware.

I agree, there is no reason. But people who want to can, and will enjoy 
the benefits.

> Don't worry. I have experience with large systems, and I take it into
> account. I think using disk labels is a good answer for your question.

Glad to hear it.

Please understand: I am not requiring users to learn the device tree. I 
am asking that users who already know the device tree can use it more 
easily. You can still hold down the 'c' key and never worry about it

The 'alias' suggestion is not bad, but it is one extra annoyance. "You 
can use Open Firmware device names in GRUB, unless they contain a 
comma. If they contain a comma, you must create an alias first." I see 
that as an implementation detail that users shouldn't have to think 
about.

-Hollis

Re: device syntax again

[Marco Gerards]

Hollis Blanchard <hollis@penguinppc.org> writes:

> The 'alias' suggestion is not bad, but it is one extra annoyance. "You
> can use Open Firmware device names in GRUB, unless they contain a
> comma. If they contain a comma, you must create an alias first." I see
> that as an implementation detail that users shouldn't have to think
> about.

Right.  But if you allow a comma, you'll create an ugly exception and
I think the code shared between the archs get really ugly because of
this.

--
Marco

Re: device syntax again

[Hollis Blanchard]

On Jan 21, 2005, at 6:22 AM, Marco Gerards wrote:

> Hollis Blanchard <hollis@penguinppc.org> writes:
>
>> The 'alias' suggestion is not bad, but it is one extra annoyance. "You
>> can use Open Firmware device names in GRUB, unless they contain a
>> comma. If they contain a comma, you must create an alias first." I see
>> that as an implementation detail that users shouldn't have to think
>> about.
>
> Right.  But if you allow a comma, you'll create an ugly exception and
> I think the code shared between the archs get really ugly because of
> this.

Yes, we're impacting users because doing the right thing would be 
difficult.

I agree that a non-trivial amount of code would need to change, but 
there is perhaps a simpler alternative: choosing a separator character 
for each architecture that is guaranteed by spec to be ok. For example:
	#define GRUB_DEVICE_SEPARATOR ',' // x86
	#define GRUB_DEVICE_SEPARATOR '|' // Open Firmware

-Hollis

Re: device syntax again

[Yoshinori K. Okuji]

On Friday 21 January 2005 15:38, Hollis Blanchard wrote:
> Yes, we're impacting users because doing the right thing would be
> difficult.

I disagree strongly. How do you define what is the right thing? For me, 
the right thing is consistency in GRUB. I think you stick to your 
firmware too much.

You must consider the balance. How much does it impact users in reality? 
If you cannot do something because of this, I would agree with you. 
But, in this case, you just need one more step and it must be rarely 
required.

> I agree that a non-trivial amount of code would need to change, but
> there is perhaps a simpler alternative: choosing a separator
> character for each architecture that is guaranteed by spec to be ok.
> For example: #define GRUB_DEVICE_SEPARATOR ',' // x86
> 	#define GRUB_DEVICE_SEPARATOR '|' // Open Firmware

Don't do it, please.

Hollis, I feel that your way of thinking is "GRUB should be a convenient 
way to wrap Open Firmware". However, I have heard that many people want 
to see GRUB in other architectures and their way of thinking is "I want 
the same features and the same interface on my architecture". 
Basically, they are interested in the user-visible part of GRUB. If 
GRUB provides a different interface for each architecture, what would 
they feel? I'm sure that they would be surprised and complain.

Okuji

Re: device syntax again

[Marco Gerards]

Hollis Blanchard <hollis@penguinppc.org> writes:

> On Jan 18, 2005, at 9:36 AM, Yoshinori K. Okuji wrote:
>
>> On Tuesday 18 January 2005 15:52, Hollis Blanchard wrote:
>>> For now let's not talk about where aliases are created. What about my
>>> original question: grub has just booted, and when it asks what device
>>> it was booted from (the /chosen/bootpath property) it gets this:
>>> 	/pci@1f,0/pci@1,1/ide@3/disk@0,0
>>> (that is the real "disk" device on Vincent's UltraSparc). The next
>>> thing we want to do is load a config file from the same device. What
>>> value should we put into "prefix"?
>>
>> This has already been discussed. Marco suggested to have a "boot drive"
>> for this.
>
> He mentioned that to me on IRC today, but I had not heard the idea
> before. I believe that will work fine.

Huh?  I even explained it to you.  But I guess I misunderstood the
question you asked me, sorry about that.

The boot drive Okuji mentioned is the alias that is always created for
the device booted from.  This should be implemented on the PC as well
because it will solve some problems.

So (boot) will in all cases point to the string in /chosen/bootpath.
After we have this we can set the prefix.  If it is possible to
determine the partition number from bootpath it should be set in
prefix as well.

If (boot) is not the right drive for prefix we should considering
storing prefix in the ELF, this is how it is done on the PC as well.
This might be required when the modules are not stored on the drive
the ELF was loaded from.

Thanks,
Marco