[PATCH 2/6] sysroot: fix short release -r option handling in case of sysroot pass proper kernel modules

[Victor Kamensky]

On Wed, 14 Mar 2018, David Smith wrote:

> I tweaked this one. First, in commit 41efad047 I made sure that
> sysroot paths don't end in a '/'. That then simplifies this patch,
> since we don't have to special case sysroot == "/". I checked this in
> as commit a9ce89bcd.

Thanks David! Appreciate progress on accepting patches.

Patches 4, 5, 6 from the series are still in queue, right?

PATCH 4: https://sourceware.org/ml/systemtap/2018-q1/msg00071.html
PATCH 5: https://sourceware.org/ml/systemtap/2018-q1/msg00066.html
PATCH 6: https://sourceware.org/ml/systemtap/2018-q1/msg00070.html

Thanks,
Victor

> On Mon, Mar 5, 2018 at 11:37 AM, Victor Kamensky <kamensky@cisco.com> wrote:
>> In case of -r option getting short release (opposite to full path of
>> kernel build tree) stap passes short release name to
>> dwfl_linux_kernel_report_offline function and this function tries to
>> look up symbols in the host file system ignoring sysroot setting.
>>
>> Fix: in case of non empty, non "/" sysroot construct full path to
>> kernel modules directory and store it into elfutils_kernel_path variable.
>>
>> Here are steps how to reproduce the issue. Note <sysroot> is produced by
>> yocto poky build.
>>
>> [kamensky@coreos-lnx2 tests]$ cat meminfo.stp
>> probe kernel.function("meminfo_proc_show") {
>>         println("meminfo_proc_show called")
>> }
>> [kamensky@coreos-lnx2 tests]$ /home/wd8/systemtap/20180208_2/packages/bin/stap --sysroot=/home/wd8/yocto/20180128/build_x86_64/tmp/work/intel_corei7_64-poky-linux/kdevel-console-devel-image/1.0-r0/rootfs -a x86_64 -r 4.9.78-yocto-standard -B CROSS_COMPILE=x86_64-poky-linux- --sysenv=PATH=/usr/bin:/bin:/usr/sbin:/sbin --sysenv=LD_LIBRARY_PATH=/lib:/usr/lib -p4 -v -m meminfo meminfo.stp
>> Pass 1: parsed user script and 480 library scripts using 230196virt/93452res/5364shr/88624data kb, in 340usr/30sys/375real ms.
>> semantic error: while resolving probe point: identifier 'kernel' at meminfo.stp:1:7
>>         source: probe kernel.function("meminfo_proc_show") {
>>                       ^
>>
>> Running command under strace shows that stap is trying to open linux symbol file
>> on host file system, instead of looking at sysroot location:
>>
>> 19517 openat(AT_FDCWD, "/boot/vmlinux-4.9.78-yocto-standard.debug", O_RDONLY) = -1 ENOENT (No such file or directory)
>> 19517 openat(AT_FDCWD, "/boot/vmlinux-4.9.78-yocto-standard", O_RDONLY) = -1 ENOENT (No such file or directory)
>>
>> Signed-off-by: Victor Kamensky <kamensky@cisco.com>
>> ---
>>  setupdwfl.cxx | 16 +++++++++++++---
>>  1 file changed, 13 insertions(+), 3 deletions(-)
>>
>> diff --git a/setupdwfl.cxx b/setupdwfl.cxx
>> index 11e0bb2af..f00cf755b 100644
>> --- a/setupdwfl.cxx
>> +++ b/setupdwfl.cxx
>> @@ -359,9 +359,19 @@ setup_dwfl_kernel (unsigned *modules_found, systemtap_session &s)
>>    // passs the plain kernel_release here.  So instead we have to
>>    // hard-code this magic here.
>>     string lib_path = "/lib/modules/" + s.kernel_release + "/build";
>> -   if (s.kernel_build_tree == string(s.sysroot + lib_path) ||
>> -       (s.kernel_build_tree == lib_path
>> -       && s.sysroot == "/"))
>> +   if (s.kernel_build_tree == string(s.sysroot + lib_path))
>> +     {
>> +       // If we have sysroot set does not make sense to pass
>> +       // short release to dwfl, it won't take a sysroot into
>> +       // account. Let's construct full path in such case.
>> +       if (s.sysroot != "" && s.sysroot != "/")
>> +        elfutils_kernel_path = string(s.sysroot + "/lib/modules/" + s.kernel_release);
>> +       else
>> +        elfutils_kernel_path = s.kernel_release;
>> +     }
>> +   else
>> +     if (s.kernel_build_tree == lib_path
>> +        && s.sysroot == "/")
>>        elfutils_kernel_path = s.kernel_release;
>>     else
>>        elfutils_kernel_path = s.kernel_build_tree;
>> --
>> 2.14.3
>>
>
>
>
> -- 
> David Smith
> Associate Manager
> Red Hat
>