Re: [PATCH]Diskdump - yet another crash dump function

Re: [PATCH]Diskdump - yet another crash dump function

[Andi Kleen]

Takao Indoh <indou.takao@soft.fujitsu.com> writes:
>
> Diskdump can be downloaded from here.
>    http://sourceforge.net/projects/lkdump
> Please see readme.txt which can be downloaded from this site.
>
> Any comments?

I like the concept - it's basically netconsole for SCSI drivers
and the driver changes are surprisingly simple and clean.

But it would be better if it coexisted with LKCD so that netdump
would also work. Can't you make it a low level driver for LKCD? 

-Andi

Re: [PATCH]Diskdump - yet another crash dump function

[Christoph Hellwig]

On Thu, May 27, 2004 at 03:19:44PM +0200, Andi Kleen wrote:
> Takao Indoh <indou.takao@soft.fujitsu.com> writes:
> >
> > Diskdump can be downloaded from here.
> >    http://sourceforge.net/projects/lkdump
> > Please see readme.txt which can be downloaded from this site.
> >
> > Any comments?
> 
> I like the concept - it's basically netconsole for SCSI drivers
> and the driver changes are surprisingly simple and clean.
> 
> But it would be better if it coexisted with LKCD so that netdump
> would also work. Can't you make it a low level driver for LKCD? 

LKCD is horribly bloated.  IMHO it would be better to get a lean
variant of netdump + possible this diskdump in instead of integrating
it into the monster LKCD patch.

Re: [PATCH]Diskdump - yet another crash dump function

[Takao Indoh]

Hi,

On Thu, 27 May 2004 15:19:44 +0200, Andi Kleen wrote:

>I like the concept - it's basically netconsole for SCSI drivers
>and the driver changes are surprisingly simple and clean.
>
>But it would be better if it coexisted with LKCD so that netdump
>would also work. Can't you make it a low level driver for LKCD? 

One of the policy of diskdump is "simple structure".
LKCD has many funcitons (normal dump to disk, network dump, memdump,
 ...), so its structure is somewhat complex.

Re: [PATCH]Diskdump - yet another crash dump function

[Nobuhiro Tachino]

Andi Kleen wrote:
> I like the concept - it's basically netconsole for SCSI drivers
> and the driver changes are surprisingly simple and clean.
> 
> But it would be better if it coexisted with LKCD so that netdump
> would also work. Can't you make it a low level driver for LKCD? 

Diskdump can exist with netdump without LKCD. Actually in our
system, diskdump and netdump both exist and can be selected by
module loading. When both modules are loaded, diskdump runs first
and if it detects I/O error when writing dump, it falls backs
to netdump.

---
Nobuhiro Tachino
FUJITSU LIMITED

Re: [3/4] [PATCH]Diskdump - yet another crash dump function

[Andi Kleen]

Ingo Molnar <mingo@elte.hu> writes:

> yeah, this is arguably the biggest (and i think only) conceptual item
> that needs to be solved before this can be integrated.

I would think netdump is more important than this though
(so if anything should be integrated i would start with netdump) 

> it would also be easier to enable diskdump in a driver if this was
> handled in add_timer()/del_timer()/mod_timer()/tasklet_schedule().

I don't think it's a good idea to add this to these fast paths.
Timers are critical for lots of things, tasklet_schedule too.

How about a standard wrapper that does the check and everybody
who may need that uses the wrappers ?

-Andi

Re: [3/4] [PATCH]Diskdump - yet another crash dump function

[Ingo Molnar]

* Andi Kleen <ak@muc.de> wrote:

> > yeah, this is arguably the biggest (and i think only) conceptual item
> > that needs to be solved before this can be integrated.
> 
> I would think netdump is more important than this though (so if
> anything should be integrated i would start with netdump)

i think both are important. Dumping to disks is a frequent customer
request, and not all setups can dump to a network, for performance,
security or reliability reasons. The diskdump format is the same as the
netdump format, and can be handled/analyzed by the same userspace tools,
so in this sense it's an extension of netdump to another class of IO
devices.

> > it would also be easier to enable diskdump in a driver if this was
> > handled in add_timer()/del_timer()/mod_timer()/tasklet_schedule().
> 
> I don't think it's a good idea to add this to these fast paths. Timers
> are critical for lots of things, tasklet_schedule too.
> 
> How about a standard wrapper that does the check and everybody who may
> need that uses the wrappers ?

that's what the patch does currently. It needs to be decided one way or
another. But it doesnt make alot of sense to introduce another vector of
APIs for this purpose - it would need possibly wide driver changes for
no good reason.

	Ingo

[PATCH]Diskdump - yet another crash dump function

[Takao Indoh]

Although I know about LKCD and netdump, I'm developing yet another crash
dump, which is a polling-based disk dump as netdump do. Because it
disables any interrupts during doing dump, it can avoid lots of problems
LKCD has.

Main Feature
- Reliability
   Diskdump disables interrupts, stops other cpus and writes to the 
   disk with polling mode. Therefore, unnecessary functions(like
   interrupt handler) don't disturb dumping.
- Safety
   Before writing to the disk, diskdump checks its disk partition to 
   confirm whether it is a really dump device. Therefore, diskdump
   doesn't overwrite filesystems by mistake.

Currently, diskdump supports only scsi disk(aic7xxx/aic79xx), but, I'm
planning to support IDE disk in the near future.

Diskdump can be downloaded from here.
   http://sourceforge.net/projects/lkdump
Please see readme.txt which can be downloaded from this site.

Any comments?

Best Regards,
Takao Indoh

Re: [PATCH]Diskdump - yet another crash dump function

[Pavel Machek]

Hi!

> Although I know about LKCD and netdump, I'm developing yet another crash
> dump, which is a polling-based disk dump as netdump do. Because it
> disables any interrupts during doing dump, it can avoid lots of problems
> LKCD has.
> 
> Main Feature
> - Reliability
>    Diskdump disables interrupts, stops other cpus and writes to the 
>    disk with polling mode. Therefore, unnecessary functions(like
>    interrupt handler) don't disturb dumping.

Hmm... with this, better design of swsusp mifht be feasible.

				Pavel
-- 
64 bytes from 195.113.31.123: icmp_seq=28 ttl=51 time=448769.1 ms         

Re: [PATCH]Diskdump - yet another crash dump function

[Takao Indoh]

Hello, 

On Thu, 03 Jun 2004 15:10:07 +0200, Pavel Machek wrote:

>> Although I know about LKCD and netdump, I'm developing yet another crash
>> dump, which is a polling-based disk dump as netdump do. Because it
>> disables any interrupts during doing dump, it can avoid lots of problems
>> LKCD has.
>> 
>> Main Feature
>> - Reliability
>>    Diskdump disables interrupts, stops other cpus and writes to the 
>>    disk with polling mode. Therefore, unnecessary functions(like
>>    interrupt handler) don't disturb dumping.
>
>Hmm... with this, better design of swsusp mifht be feasible.

I don't know mechanism of swsusp well.
Do you mean diskdump is useful when saving system memory to the disk?

Regards,
Takao Indoh

Re: [PATCH]Diskdump - yet another crash dump function

[Pavel Machek]

Hi!

> >> Although I know about LKCD and netdump, I'm developing yet another crash
> >> dump, which is a polling-based disk dump as netdump do. Because it
> >> disables any interrupts during doing dump, it can avoid lots of problems
> >> LKCD has.
> >> 
> >> Main Feature
> >> - Reliability
> >>    Diskdump disables interrupts, stops other cpus and writes to the 
> >>    disk with polling mode. Therefore, unnecessary functions(like
> >>    interrupt handler) don't disturb dumping.
> >
> >Hmm... with this, better design of swsusp mifht be feasible.
> 
> I don't know mechanism of swsusp well.
> Do you mean diskdump is useful when saving system memory to the disk?

Yes. Self-contained code accessing disk would be usefull both when
saving system memory and when reading it back. (This is probably 2.8
material, through).
							Pavel

-- 
934a471f20d6580d5aad759bf0d97ddc