Re: Review of the systemtap examples testsuite results x86 Fedora rawhide

[William Cohen <wcohen@redhat.com>]

On 2/6/19 4:41 PM, William Cohen wrote:
> Now that systemtap is working with 5.0.0-rcx kernel took the time to run the
> systemtap examples to see what failed.
> 
> The following two tests are failing because "vitural memory exhausted"
> Both are using --all-modules. There are some other test other PASS:
> but it looks like they result in smaller stap-symbols.h files.  These
> particular tests look to be because stap-symbols.h files are huge,
> over 70MB.  This fail might be more of a result of the guest VM having
> 2GB of RAM. However, not sure why there is such variations in the
> in the size of the stap-symbols.h files
> 
> FAIL: systemtap.examples/profiling/fileline-profile run
> FAIL: systemtap.examples/profiling/periodic build

When running the fileline-profile see cc1 get way up there in memory use on successful run on a machine with ample resources:

  PID USER      PR  NI    VIRT    RES    SHR S  %CPU  %MEM     TIME+ COMMAND                                            
17936 root      20   0 3013476   2.4g  12484 R 100.0   7.6   0:23.33 cc1          

similarly for periodic.stp see:

  PID USER      PR  NI    VIRT    RES    SHR S  %CPU  %MEM     TIME+ COMMAND                                                
21631 root      20   0 3656892   2.7g  21588 R  99.7   8.8   0:29.40 cc1  

> 
>> Kernel commit 23c9deeb328 eliminates all the FAN_ALL_* defines causing
> FAN_ALL_CLASS_BITS to be undefined. This commit is also in the 4.20
> kernels, so the errsnoop and strace examples will also break on Fedora
> 29/28.  _fanotify_init_flags_str function in
> tapset/linux/aux_syscalls.stp will need to be fixed up.
> 
> FAIL: systemtap.examples/process/errsnoop build
> FAIL: systemtap.examples/process/strace build

Both errsnoop and strace are working now with systemtap git commits:
4e76869512d2d05fc3347861bdeed92395e34d63
1ac5a4499ecc17f612a94bf4dff39ba90d4cd532

However, noticed that there appears to be some nd_syscall.*.retrun probes missing retstr given the following warnings when running strace.stp:

WARNING: never-assigned local variable 'retstr' (similar: argstr, status, name, _target_set, thread_argstr): identifier 'retstr' at strace.stp:49:38
 source:     report(name,thread_argstr[tid()],retstr)
                                              ^
WARNING: never-assigned local variable 'retstr' (similar: argstr, status, name, _target_set, thread_argstr): identifier 'retstr' at :49:38
 source:     report(name,thread_argstr[tid()],retstr)
                                              ^
WARNING: never-assigned local variable 'retstr' (similar: argstr, status, name, _target_set, thread_argstr): identifier 'retstr' at :49:38
 source:     report(name,thread_argstr[tid()],retstr)
                                              ^
WARNING: never-assigned local variable 'retstr' (similar: argstr, status, name, _target_set, thread_argstr): identifier 'retstr' at :49:38
 source:     report(name,thread_argstr[tid()],retstr)
   
> 
> 
> The following two tests seem to be having issues with the
> kernel.statement() on do_sys_open being used for them.  Both get
> "inconsistent relocation address".
> 
> FAIL: systemtap.examples/general/varwatch build
> FAIL: systemtap.examples/general/whythefail build
> 
> 
> The following test appears not to go down some other function other
> than the vfs_* functions being currently monitored.  This does work on
> RHEL7.  Looking through the list of EXPORT_SYMBOLS(vfs_*) in
> linux/fs/namei.c it looks like vfs_tmpfile or vfs_mkobj are likely
> missing probes:
> 
> FAIL: systemtap.examples/general/badname run

Found that badname.stp last worked in linux 4.6 and was broken in 4.7.  Did a kernel bisect today and found that the following kernel commit caused this example to stop working:

commit 6ac087099edf09ca357e2f765e3e24677543897c
Author: Al Viro <viro@zeniv.linux.org.uk>
Date:   Tue Apr 26 00:02:50 2016 -0400

    path_openat(): take O_PATH handling out of do_last()
    
    do_last() and lookup_open() simpler that way and so does O_PATH
    itself.  As it bloody well should: we find what the pathname
    resolves to, same way as in stat() et.al. and associate it with
    FMODE_PATH struct file.
    
    Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>

> 
> 
> The syscall.ptrace probe point is now using the nodwarf version, so
> $request target variable isn't available
> 
> FAIL: systemtap.examples/process/noptrace build
> 
> 
> Lots of kernel internal ABI changes making the following fail to build:
> 
> FAIL: systemtap.examples/process/pfiles build
> 
> -Will
>