Skip to contentSkip to conversation

remote GDB debugging and the __init macro of init.h

DISCUSSION Steve Kranz 5 participants 6 messages Oct 27, 2000 – Oct 31, 2000 MIPS

lore
Share & tools
lore
#0 lore raw

Note:

I had to make a change to allow remote MIPS kernel debugging (GDB). The change I found necessary was in the file:

    include/linux/init.h     (2.4.0-test9)

As you can see from the snippet below the change involves conditionally defining the "__init" macro as a function of whether remote debugging is enabled or not. Am I missing something, or does this seem like a reasonable change?

===========
Was this...
===========
/*
* Mark functions and data as being only used at initialization
* or exit time.
*/
#define __init __attribute__ ((__section__ (".text.init")))

==================================
I changed my local copy to this...
==================================
/*
* Mark functions and data as being only used at initialization
* or exit time.
*/
#ifdef CONFIG_REMOTE_DEBUG
// Note: While running the mips-linux-elf-gdb (GNU gdb 5.0), RidgeRun Inc
// noticed that gdb could not correctly derive the true address of any symbol
// declared with the __init pragma. This prevented being able to correctly
// set breakpoints on any of those functions. So, if we are building // with the GDB remote debugger in mind, then null out the __init // definition making those functions look like a normal functions // since this seems to satisfy things for remote kernel debugging. // Incidentally, for reference, the GDB being used at the time of this writing
// was configured as "--host=i686-pc-linux-gnu --target=mips-linux-elf".

// and the mips-linux-gcc crosscompiler being used is egcs-2.90.29 980515
// (egcs-1.0.3 release) with binutils version 2.8.1. (These tools running on
// a x86 host producing code for target CONFIG_CPU_R5000). #define __init
#else
#define __init __attribute__ ((__section__ (".text.init"))) #endif

Steve Kranz
[email protected]
Senior Kernel Developer
RidgeRun Inc.

#1 lore raw

What you've done should solve the problem, but note that it has the side effect of preventing the text and data sections in question from getting freed up at the end of initialization. I probably should have done so myself last year when I was struggling with debugging some init code using kgdb, but instead I simply got used to finding the address in the symbol table and setting the breakpoints by hex address instead of by symbol.

The real fix would be to teach gdb to treat symbols in the init section as valid targets.

Kevin K.

----- Original Message -----
From: "Steve Kranz" <[email protected]>
To: <[email protected]>; <[email protected]>
Sent: Friday, October 27, 2000 5:24 PM
Subject: remote GDB debugging and the __init macro of init.h

Quoted 61 lines

Note:

I had to make a change to allow remote MIPS kernel debugging (GDB). The change I found necessary was in the file:

    include/linux/init.h     (2.4.0-test9)

As you can see from the snippet below the change involves conditionally defining the "__init" macro as a function of whether remote debugging is enabled or not. Am I missing something, or does this seem like a reasonable change?

===========
Was this...
===========
/*
* Mark functions and data as being only used at initialization
* or exit time.
*/
#define __init __attribute__ ((__section__ (".text.init")))

==================================
I changed my local copy to this...
==================================
/*
* Mark functions and data as being only used at initialization
* or exit time.
*/
#ifdef CONFIG_REMOTE_DEBUG
// Note: While running the mips-linux-elf-gdb (GNU gdb 5.0), RidgeRun Inc
// noticed that gdb could not correctly derive the true address of any symbol
// declared with the __init pragma. This prevented being able to correctly
// set breakpoints on any of those functions. So, if we are building // with the GDB remote debugger in mind, then null out the __init // definition making those functions look like a normal functions // since this seems to satisfy things for remote kernel debugging. // Incidentally, for reference, the GDB being used at the time of this writing
// was configured as "--host=i686-pc-linux-gnu --target=mips-linux-elf".

// and the mips-linux-gcc crosscompiler being used is egcs-2.90.29 980515
// (egcs-1.0.3 release) with binutils version 2.8.1. (These tools running on
// a x86 host producing code for target CONFIG_CPU_R5000). #define __init
#else
#define __init __attribute__ ((__section__ (".text.init"))) #endif

Steve Kranz
[email protected]
Senior Kernel Developer
RidgeRun Inc.

#2 lore raw
Quoted 11 lines · Kevin D. Kissell

What you've done should solve the problem, but note that it has the side effect of preventing the text and data sections in question from getting freed up at the end of initialization. I probably should have done so myself last year when I was struggling with debugging some init code using kgdb, but instead I simply got used to finding the address in the symbol table and setting the breakpoints by hex address instead of by symbol.

The real fix would be to teach gdb to treat symbols in the init section as valid targets.

Did somebody look into porting KDB?

Ralf

#3 lore raw
In reply to Kevin D. Kissell
Quoted 10 lines · Kevin D. Kissell

What you've done should solve the problem, but note that it has the side effect of preventing the text and data sections in question from getting freed up at the end of initialization. I probably should have done so myself last year when I was struggling with debugging some init code using kgdb, but instead I simply got used to finding the address in the symbol table and setting the breakpoints by hex address instead of by symbol.

Kevin,

A dumb question - how do you set breakpoint at specified address? I was trying to do that with "b 0xabcdabcd" or "b @0xabcdabcd", none of them worked.

Jun

#4 lore raw
Quoted 9 lines · Jun Sun
Quoted 8 lines

What you've done should solve the problem, but note that it has the side effect of preventing the text and data sections in question from getting freed up at the end of initialization. I probably should have done so myself last year when I was struggling with debugging some init code using kgdb, but instead I simply got used to finding the address in the symbol table and setting the breakpoints by hex address instead of by symbol.

Quoted 3 lines

A dumb question - how do you set breakpoint at specified address? I was trying to do that with "b 0xabcdabcd" or "b @0xabcdabcd", none of them worked.

b *0xabcdabcd

Try ``help break'' :-)

Ralf

#5 lore raw
In reply to Steve Kranz
Quoted 11 lines · Steve Kranz

I had to make a change to allow remote MIPS kernel debugging (GDB). The change I found necessary was in the file:

   include/linux/init.h     (2.4.0-test9)

As you can see from the snippet below the change involves conditionally defining the "__init" macro as a function of whether remote debugging is enabled or not. Am I missing something, or does this seem like a reasonable change?

It would be better to teach kgdb that the kernel has symbols in sections other than .text. Even with your patch, you do not get all the symbols, there is also code in sections .setup.init, .initcall.init and .exitcall.exit, with similar sections for data. You cannot remove .initcall.init without destroying the kernel initialization procedure.

Recent versions of kgdb for ix86 use a modified version of gdb that knows about multiple sections for modules. I do not know if they have the same functionality for the kernel.