The 'perf' tool has some built-in test cases and one of them checks to
see if the symbols in vmlinux match those in /proc/kallsyms.
This test is failing on Power for several reasons. I fixed a couple of
them (described briefly at the end of the mail) and these fixes take
the test further.
One pending failure is because maps in vmlinux differ from kallsyms.
Here is some output:
$ ./perf test -v 1
---
Using /lib/modules/3.8.0-rc4-perf-core+/build/vmlinux for symbols
0xc000000000001800: diff name v: machine_check_fwnmi k: machine_check_pSeries
0xc000000000005780: diff name v: __end_interrupts k: system_call_entry_direct
0xc000000000006800: diff name v: __end_handlers k: .do_hash_page
0xc000000000009300: diff name v: __secondary_start k: copy_to_here
0xc00000000063d7a0: diff name v: __kprobes_text_end k: .vdso_getcpu_init
Maps only in vmlinux:
c00000000063d804-c000000000642077 0 [kernel].cpuinit.text
c000000000642078-c00000000064225f 0 [kernel].text.unlikely
c000000000642260-c0000000007cffff 0 [kernel].meminit.text
c0000000007d0000-c000000000812043 0 [kernel].init.text
c000000000812044-d000000000f8ffff 0 [kernel].exit.text
Maps in vmlinux with a different name in kallsyms:
---
Looking a the first pair of differing symbols in kallsyms and vmlinux:
$ grep c000000000001800 /proc/kallsyms
c000000000001800 T machine_check_fwnmi
c000000000001800 t machine_check_pSeries
$ nm /boot/vmlinux-3.8.0-rc4-perf-core |grep c000000000001800
c000000000001800 T machine_check_fwnmi
c000000000001800 t machine_check_pSeries
Now perf attempts to "fixup" the duplicates (see choose_best_symbol() in
tools/perf/util/symbol.c) and use heuristics to choose one from the pair
and discard the other.
When processing the pair of 'machine_check_fwnmi' and 'machine_check_pSeries'
from kallsyms, it finds that they both have global binding and it chooses
'machine_check_pSeries' based on the string length of the symbol name.
When processing the same pair from vmlinux, it finds that 'machine_check_fwnmi'
has the global binding and chooses that.
Finally, when comparing the vmlinux symbols with kallsyms, this mismatch
is shows up as a failure.
Some questions:
- Is perf trying to choose one from the pair, so we have only one
symbol to use when building a profile or call-graph ?
- Why do some symbols like 'machine_check_fwnmi' above, appear with
global binding in /proc/kallsyms and local binding in vmlinux ?
----
Here are couple of other failures in 'perf test' on Power. I applied some
quick fixes for these to expose the above failures.
1. Look for .__start or _stext.
On powerpc, multiple symbols match the address of '_stext':
$ grep c000000000000000 /proc/kallsyms
c000000000000000 T .__start
c000000000000000 T _stext
c000000000000000 T _text
choose_best_symbol() discards all but .__start() because, ironically, it
has the least preceding underscores. The vmlinux test case later looks for
_stext and fails.
Quick fix, have the test case look for either _stext or .__start.
2. In vmlinux, system call names have '.sys' or '_sys_'.
In kallsyms, the names have '.SyS' or '_SyS_'
0xc000000000067660: diff name v: .sys_sched_get_priority_max k:
.SyS_sched_get_priority_max
0xc000000000067690: diff name v: .sys_sched_get_priority_min k:
.SyS_sched_get_priority_min
0xc0000000000f2650: diff name v: .compat_sys_waitid k: .compat_SyS_waitid
0xc0000000000f2780: diff name v: .compat_sys_wait4 k: .compat_SyS_wait4
Quick fix: treat names as identical if they only differ in upper/lower
case of the substring 'sys'.
Curious though why kallsyms and vmlinux differ this way.
_______________________________________________
Linuxppc-dev mailing list
Linuxppc-dev@lists.ozlabs.org
https://lists.ozlabs.org/listinfo/linuxppc-dev
