How do you read a segfault kernel log message

Question

This can be a very simple question, I'm am attempting to debug an application which generates the following segfault error in the kern.log

kernel: myapp[15514]: segfault at 794ef0 ip 080513b sp 794ef0 error 6 in myapp[8048000+24000]

Here are my questions:

1. Is there any documentation as to what are the diff error numbers on segfault, in this instance it is error 6, but i've seen error 4, 5

2. What is the meaning of the information at bf794ef0 ip 0805130b sp bf794ef0 and myapp[8048000+24000]?

So far i was able to compile with symbols, and when i do a x 0x8048000+24000 it returns a symbol, is that the correct way of doing it? My assumptions thus far are the following:

• sp = stack pointer?
• ip = instruction pointer
• at = ????
• myapp[8048000+24000] = address of symbol?

Solution 1

When the report points to a program, not a shared library

Run addr2line -e myapp 080513b (and repeat for the other instruction pointer values given) to see where the error is happening. Better, get a debug-instrumented build, and reproduce the problem under a debugger such as gdb.

If it's a shared library

In the libfoo.so[NNNNNN+YYYY] part, the NNNNNN is where the library was loaded. Subtract this from the instruction pointer (ip) and you'll get the offset into the .so of the offending instruction. Then you can use objdump -DCgl libfoo.so and search for the instruction at that offset. You should easily be able to figure out which function it is from the asm labels. If the .so doesn't have optimizations you can also try using addr2line -e libfoo.so <offset>.

What the error means

Here's the breakdown of the fields:

• address - the location in memory the code is trying to access (it's likely that 10 and 11 are offsets from a pointer we expect to be set to a valid value but which is instead pointing to 0)
• ip - instruction pointer, ie. where the code which is trying to do this lives
• sp - stack pointer
• error - Architecture-specific flags; see arch/*/mm/fault.c for your platform.

Solution 2

Based on my limited knowledge, your assumptions are correct.

• sp = stack pointer
• ip = instruction pointer
• myapp[8048000+24000] = address

If I were debugging the problem I would modify the code to produce a core dump or log a stack backtrace on the crash. You might also run the program under (or attach) GDB.

The error code is just the architectural error code for page faults and seems to be architecture specific. They are often documented in arch/*/mm/fault.c in the kernel source. My copy of Linux/arch/i386/mm/fault.c has the following definition for error_code:

• bit 0 == 0 means no page found, 1 means protection fault
• bit 1 == 0 means read, 1 means write
• bit 2 == 0 means kernel, 1 means user-mode

My copy of Linux/arch/x86_64/mm/fault.c adds the following:

• bit 3 == 1 means fault was an instruction fetch

Solution 3

If it's a shared library

You're hosed, unfortunately; it's not possible to know where the libraries were placed in memory by the dynamic linker after-the-fact.

Well, there is still a possibility to retrieve the information, not from the binary, but from the object. But you need the base address of the object. And this information still is within the coredump, in the link_map structure.

So first you want to import the struct link_map into GDB. So lets compile a program with it with debug symbol and add it to the GDB.

link.c

#include <link.h>
toto(){struct link_map * s = 0x400;}

get_baseaddr_from_coredump.sh

#!/bin/bash

BINARY=$(which myapplication)

IsBinPIE ()
{
    readelf -h $1|grep 'Type' |grep "EXEC">/dev/null || return 0
    return 1
}

Hex2Decimal ()
{
    export number="`echo "$1" | sed -e 's:^0[xX]::' | tr '[a-f]' '[A-F]'`"
    export number=`echo "ibase=16; $number" | bc`
}

GetBinaryLength ()
{
    if [ $# != 1 ]; then
    echo "Error, no argument provided"
    fi
    IsBinPIE $1 || (echo "ET_EXEC file, need a base_address"; exit 0)
    export totalsize=0
    # Get PT_LOAD's size segment out of Program Header Table (ELF format)
    export sizes="$(readelf -l $1 |grep LOAD |awk '{print $6}'|tr '\n' ' ')"
    for size in $sizes
    do Hex2Decimal "$size"; export totalsize=$(expr $number + $totalsize); export totalsize=$(expr $number + $totalsize)
    done
    return $totalsize
}

if [ $# = 1 ]; then
    echo "Using binary $1"
    IsBinPIE $1 && (echo "NOT ET_EXEC, need a base_address..."; exit 0)
    BINARY=$1
fi

gcc -g3 -fPIC -shared link.c -o link.so

GOTADDR=$(readelf -S $BINARY|grep -E '\.got.plt[ \t]'|awk '{print $4}')

echo "First do the following command :"
echo file $BINARY
echo add-symbol-file ./link.so 0x0
read
echo "Now copy/paste the following into your gdb session with attached coredump"
cat <<EOF
set \$linkmapaddr = *(0x$GOTADDR + 4)
set \$mylinkmap = (struct link_map *) \$linkmapaddr
while (\$mylinkmap != 0)
if (\$mylinkmap->l_addr)
printf "add-symbol-file .%s %#.08x\n", \$mylinkmap->l_name, \$mylinkmap->l_addr
end
set \$mylinkmap = \$mylinkmap->l_next
end

it will print you the whole link_map content, within a set of GDB command.

It itself it might seems unnesseray but with the base_addr of the shared object we are about, you might get some more information out of an address by debuging directly the involved shared object in another GDB instance. Keep the first gdb to have an idee of the symbol.

NOTE : the script is rather incomplete i suspect you may add to the second parameter of add-symbol-file printed the sum with this value :

readelf -S $SO_PATH|grep -E '\.text[ \t]'|awk '{print $5}'

where $SO_PATH is the first argument of the add-symbol-file

Hope it helps

Comments

[Sullenx]

The issue i have is that: 1) The application is segfaulting in a production environment, where symbols are stripped, all i have is just the logs 2) I'm trying to find that memory location in the development env, so at least i can see where it is crashing.

[jschmier]

If you have the pre-stripped binary, try running it through nm or objdump.

[Sullenx]

nm is pretty helpful, at least I have an idea where the crash happened. One last thing, what is an error 6? ... is there any table out there?

[Nikolai Fetissov]

segfault at 794ef0 ... sp 794ef0 - stack is obviously corrupted.

[Sullenx]

Thank you, this is very helpful

[oliver]

Event for a shared lib, the "[8048000+24000]" part should give a hint where the crashing segment of the lib was mapped in memory. "readelf --segments mylib.so" lists these segments, and then you can calculate the EIP offset into the crashing segment and feed that to addr2line (or view it in "objdump -dgS").

[Tectrendz]

How do you compile link.c ?

[Nemo]

I believe 0x8048000 is (probably) the address where the text segment was mapped, so you will want to pass -j .text to the objdump command. (At least, that is what I needed when diagnosing one of these just now.)

[Baroudi Safwen]

@Charles Duffy If I ever see you I will hug like I never hugged a living soul.

[emkey08]

On a 64 bit Debian system, I have trap divide error ip:55fc4735e6d8 sp:7fa4d77fcd80 error:0 in myapp[55fc472d4000+ed000], but addr2line -e myapp 55fc4735e6d8 just returns ??:0. The app is a debug build with no optimizations enabled. Any advice?

[emkey08]

Note: I also tried addr2line -e myapp 8a6d8 (since 55fc4735e6d8 - 55fc472d4000 = 8a6d8), which actually does find a source line, but obviously an incorrect one, since no division is happening there at all. (I'm having a trap divide error.)

[pktiuk]

In my case, using addr2line with executable returns only ??:0 (even after recompiling executable with debug info) Could you clarify what is needed for proper output?