enhance makedumpfile to dump process list from vmcore

enhance makedumpfile to dump process list from vmcore

[Prasad Koya]

Hi

We use makedumpfile to read the dmesg buffer from the vmcore. We are
working on adding a feature to makedumpfile to extract the "process
list" from /proc/vmcore. Our main use case is to save the processes
that were running at the time of kernel crash to a file on the
permanent storage from the kdump kernel. We intend to read a handful
of useful data for each running task: pid, parent pid, real parent
pid, priority, nice value, RSS, VM size, command, utime, stime, OOM
score.

Process list from vmcore can probably be read using crash utility, but
in our embedded system using makedumpfile to do that job makes sense
for us.

If such an option is useful for the general users of makedumpfile,
we'd like to get more inputs and contribute to this feature.

Thank you
Prasad

_______________________________________________
kexec mailing list
kexec@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/kexec

Re: enhance makedumpfile to dump process list from vmcore

[HAGIO KAZUHITO(萩尾 一仁)]

On 2024/06/06 3:10, Prasad Koya wrote:
> Hi
> 
> We use makedumpfile to read the dmesg buffer from the vmcore. We are
> working on adding a feature to makedumpfile to extract the "process
> list" from /proc/vmcore. Our main use case is to save the processes
> that were running at the time of kernel crash to a file on the
> permanent storage from the kdump kernel. We intend to read a handful
> of useful data for each running task: pid, parent pid, real parent
> pid, priority, nice value, RSS, VM size, command, utime, stime, OOM
> score.
> 
> Process list from vmcore can probably be read using crash utility, but
> in our embedded system using makedumpfile to do that job makes sense
> for us.
> 
> If such an option is useful for the general users of makedumpfile,
> we'd like to get more inputs and contribute to this feature.

Hi Prasad,

thank you for sharing the idea.  But sorry, currently we are not 
interested in implementing such a big function in makedumpfile.

The dmesg buffer is the most important information to be able to 
determine the cause of a crash only with it, so the --dump-dmesg option 
was implemented in makedumpfile first once upon a time.  But now we have 
handier vmcore-dmesg command in kexec-tools, so we don't intend to add 
other functions to makedumpfile than making a dumpfile.

I'm not sure whether these are doable, but I would suggest a few ideas 
instead:
- make e.g. vmcore-process-list command like vmcore-dmesg.
- use panic notifier to dump process list to the dmesg buffer before 
kdump, and use vmcore-dmesg to get it.

Thanks,
Kazu
_______________________________________________
kexec mailing list
kexec@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/kexec

Re: enhance makedumpfile to dump process list from vmcore

[Dave Young]

Hi Kazu,

On Wed, 19 Jun 2024 at 13:45, HAGIO KAZUHITO(萩尾　一仁) <k-hagio-ab@nec.com> wrote:
>
> On 2024/06/06 3:10, Prasad Koya wrote:
> > Hi
> >
> > We use makedumpfile to read the dmesg buffer from the vmcore. We are
> > working on adding a feature to makedumpfile to extract the "process
> > list" from /proc/vmcore. Our main use case is to save the processes
> > that were running at the time of kernel crash to a file on the
> > permanent storage from the kdump kernel. We intend to read a handful
> > of useful data for each running task: pid, parent pid, real parent
> > pid, priority, nice value, RSS, VM size, command, utime, stime, OOM
> > score.
> >
> > Process list from vmcore can probably be read using crash utility, but
> > in our embedded system using makedumpfile to do that job makes sense
> > for us.
> >
> > If such an option is useful for the general users of makedumpfile,
> > we'd like to get more inputs and contribute to this feature.
>
> Hi Prasad,
>
> thank you for sharing the idea.  But sorry, currently we are not
> interested in implementing such a big function in makedumpfile.
>
> The dmesg buffer is the most important information to be able to
> determine the cause of a crash only with it, so the --dump-dmesg option
> was implemented in makedumpfile first once upon a time.  But now we have
> handier vmcore-dmesg command in kexec-tools, so we don't intend to add
> other functions to makedumpfile than making a dumpfile.
>
> I'm not sure whether these are doable, but I would suggest a few ideas
> instead:
> - make e.g. vmcore-process-list command like vmcore-dmesg.

Not sure if this is worth either.    I guess it can be done with some
crash tool scripts or a crash plugin?

> - use panic notifier to dump process list to the dmesg buffer before
> kdump, and use vmcore-dmesg to get it.

It is not recommended to add these before kdump if people still want a
reliable kdump.

>
> Thanks,
> Kazu
> _______________________________________________
> kexec mailing list


Thanks
Dave
> kexec@lists.infradead.org
> http://lists.infradead.org/mailman/listinfo/kexec
>


_______________________________________________
kexec mailing list
kexec@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/kexec

Re: enhance makedumpfile to dump process list from vmcore

[HAGIO KAZUHITO(萩尾 一仁)]

Hi Dave,

thank you for your comments.

On 2024/06/24 10:52, Dave Young wrote:
> Hi Kazu,
> 
> On Wed, 19 Jun 2024 at 13:45, HAGIO KAZUHITO(萩尾　一仁) <k-hagio-ab@nec.com> wrote:
>>
>> On 2024/06/06 3:10, Prasad Koya wrote:
>>> Hi
>>>
>>> We use makedumpfile to read the dmesg buffer from the vmcore. We are
>>> working on adding a feature to makedumpfile to extract the "process
>>> list" from /proc/vmcore. Our main use case is to save the processes
>>> that were running at the time of kernel crash to a file on the
>>> permanent storage from the kdump kernel. We intend to read a handful
>>> of useful data for each running task: pid, parent pid, real parent
>>> pid, priority, nice value, RSS, VM size, command, utime, stime, OOM
>>> score.
>>>
>>> Process list from vmcore can probably be read using crash utility, but
>>> in our embedded system using makedumpfile to do that job makes sense
>>> for us.
>>>
>>> If such an option is useful for the general users of makedumpfile,
>>> we'd like to get more inputs and contribute to this feature.
>>
>> Hi Prasad,
>>
>> thank you for sharing the idea.  But sorry, currently we are not
>> interested in implementing such a big function in makedumpfile.
>>
>> The dmesg buffer is the most important information to be able to
>> determine the cause of a crash only with it, so the --dump-dmesg option
>> was implemented in makedumpfile first once upon a time.  But now we have
>> handier vmcore-dmesg command in kexec-tools, so we don't intend to add
>> other functions to makedumpfile than making a dumpfile.
>>
>> I'm not sure whether these are doable, but I would suggest a few ideas
>> instead:
>> - make e.g. vmcore-process-list command like vmcore-dmesg.
> 
> Not sure if this is worth either.

I also don't think that this (dumping a process list in 2nd kernel) is 
worth such a (probably) big effort, because we can make a dumpfile..

I assumed that they cannot create a dumpfile, because they know that the 
crash-utility can read it from a vmcore.  I'm not sure why makedumpfile 
doing it makes sense to them..

>   I guess it can be done with some
> crash tool scripts or a crash plugin?

Yes, agree.

> 
>> - use panic notifier to dump process list to the dmesg buffer before
>> kdump, and use vmcore-dmesg to get it.
> 
> It is not recommended to add these before kdump if people still want a
> reliable kdump.

Agree, a panic situation is unstable, I don't intend to recommend doing 
this generally.  I just thought that if they cannot create a dumpfile, 
maybe they can make a module that uses the notifier, as a last resort.

Thanks,
Kazu
_______________________________________________
kexec mailing list
kexec@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/kexec