This is the mail archive of the
systemtap@sourceware.org
mailing list for the systemtap project.
Better module/symbol tracking for user space
- From: Mark Wielaard <mjw at redhat dot com>
- To: systemtap at sourceware dot org
- Date: Mon, 21 Dec 2009 13:36:31 +0100
- Subject: Better module/symbol tracking for user space
Hi,
I added some code to solve PR11015 Support shared library reloading (in
different processes), which should resolve the "ERROR Reloaded module
'XXX' at 0xYYY, was 0xZZZ" that people were seeing recently. This isn't
really new, we used to just silently ignore this issue in the past.
Which meant we would not track modules/symbols correctly for user space
shared libraries which were loaded at different addresses in different
processes. This patch fixes that:
commit ea549ffc2915aa58861637472b12196222673fa2
Author: Mark Wielaard <mjw@redhat.com>
Date: Mon Dec 21 13:02:19 2009 +0100
PR11015 Support shared library reloading (in different processes)
* runtime/task_finder_vma.c (stap_remove_vma_map_info): Return negative
status on failure.
(stap_find_vma_map_info): Likewise.
(stap_find_vma_map_info_user): New function.
(stap_drop_vma_maps): New function.
* runtime/sym.h (addr): Renamed to static_addr, to store addresses for
sections which are always mapped at the same address.
(_stp_module_relocate): Add extra struct task_struct *tsk argument.
* runtime/sym.c (_stp_tf_exec_cb): New callback, calls stap_drop_vma_maps.
(_stp_tf_mmap_cb): Don't store address in module.section, but call
stap_add_vma_map_info() per tsk->group_leader for matched module.
Don't register empty/null modules.
(_stp_module_relocate): Take extra struct task_struct *tsk argument,
cache last tsk used. Only use section->static_addr for none dynamic
modules. Use stap_find_vma_map_info_user() to locate dynamic modules.
(_stp_mod_sec_lookup): Add extra argument unsigned long *rel_addr to
optionally store relative address when module/section found.
(_stp_kallsyms_lookup): Use _stp_mod_sec_lookup to find relative address.
(_stp_sym_init): Register _stp_tf_exec_cb in stap_task_finder_target.
Add error check to see if task finder could be initialized.
* dwflpp.cxx (emit_address): Pass NULL for kernel/modules and current for
user tasks to _stp_module_relocate.
* runtime/transport/symbols.c (_stp_do_relocation): Set new static_addr
_stp_section field.
* runtime/unwind.c (adjustStartLoc): Take new struct task_struct *tsk
argument and pass to stap_find_vma_map_info_user and _stp_module_relocate
to find adjusted addr.
(_stp_search_unwind_hdr): Pass through struct task_struct *tsk.
(unwind_frame): Likewise.
* tapset/context-symbols.stp (probemod): Add NULL to _stp_mod_sec_lookup
call to indicate we aren't interested in relative address.
* tapsets.cxx (dwarf_derived_probe_group::emit_module_init): Pass NULL to
_stp_module_relocate to indicate kernel/module address.
I didn't add new test cases yet, so would appreciate if people who saw
the above ERROR could double check things work as expected now. I also
added some XXX comments in runtime/sym.c for things I think we are still
doing which are not 100% correct/optimal yet. Also the unwind.c
(adjustStartLoc) function should really, really be rewritten. I only
fixed it up enough to fit the new structure, but it really shouldn't be
doing all the work it is now doing directly anymore.
Cheers,
Mark