[Crash-utility] [PATCH] add arm support for libgcore

[HATAYAMA Daisuke]

From: Lei Wen <adrian.wenl at gmail.com>
Subject: Re: [Crash-utility] [PATCH] add arm support for libgcore
Date: Mon, 13 Feb 2012 17:03:02 +0800

> Hi Hatayama,
> 
> On Mon, Feb 13, 2012 at 4:33 PM, HATAYAMA Daisuke
> <d.hatayama at jp.fujitsu.com> wrote:
>> Hello Lei,
>>
>> Thanks for making patch. I'll check your patch this week, but I have
>> two things to ask you.
>>
>> 1. I don't know arm architecture at all and I don't have arm
>> machine. What I can do is only testing common part and regression test
>> on x86 architecture. Please maintain arm part yourself.
> 
> Sure, it is my pleasure. :)
> 
>>
>> 2. Could you tell me specific kernel versions you have tested this
>> patch in? I myself have yet to do this, but now I think it necessary
>> to make such a list. I imagine just like makedumpfile's SUPPORTED
>> KERNELS described in its README. I'll put them in gcore's README and
>> then ask Dave to add them into description in distribution page.
> 
> I am current testing with kernel 2.6.35.7 and 3.0.8, and they are both ok.

I see.

> But I see below warnings during extracting, while the extracted core
> dump image is
> OK for gdb, I don't know whether it there is still some missing in
> original implementation,
> or those pages just don't existed in memory?
> 
> gcore: PT_LOAD[165]: af900000 - af90e000
> gcore: WARNING: page fault at af909000
> gcore: WARNING: page fault at af90d000

These are verbose messages, which you can specify which to display via
-v option. Please see help message in detail. However, important is
"page fault" message below.

> gcore: WARNING: page fault at af909000
> gcore: WARNING: page fault at af90d000

Most of crash dump mechanism doesn't collect swap space, such as
kdump, diskdump, so in essence, crash gcore doesn't try to collect
such paged-out user-space memory.

crash gcore instead fills the paged-out memory with zero; this
is easier implementation than reconstracting program headers.

The reason why I've included the ``page fault'' in the default warning
message is to avoid the situation where users get confused they have
successfully got complete user-space coredump.

GDB tends to work well because part of user stack necessary for
backtrace is not paged out most of time; of course, the backtrace
would fail if paged-out.

Thanks.
HATAYAMA, Daisuke