forked from luck/tmp_suning_uos_patched
3ec42508a6
[ Upstream commit 2d86612aacb7805f72873691a2644d7279ed0630 ]
When using 'perf mem' and 'perf c2c', an issue is observed that tool
reports the wrong offset for global data symbols. This is a common
issue on both x86 and Arm64 platforms.
Let's see an example, for a test program, below is the disassembly for
its .bss section which is dumped with objdump:
...
Disassembly of section .bss:
0000000000004040 <completed.0>:
...
0000000000004080 <buf1>:
...
00000000000040c0 <buf2>:
...
0000000000004100 <thread>:
...
First we used 'perf mem record' to run the test program and then used
'perf --debug verbose=4 mem report' to observe what's the symbol info
for 'buf1' and 'buf2' structures.
# ./perf mem record -e ldlat-loads,ldlat-stores -- false_sharing.exe 8
# ./perf --debug verbose=4 mem report
...
dso__load_sym_internal: adjusting symbol: st_value: 0x40c0 sh_addr: 0x4040 sh_offset: 0x3028
symbol__new: buf2 0x30a8-0x30e8
...
dso__load_sym_internal: adjusting symbol: st_value: 0x4080 sh_addr: 0x4040 sh_offset: 0x3028
symbol__new: buf1 0x3068-0x30a8
...
The perf tool relies on libelf to parse symbols, in executable and
shared object files, 'st_value' holds a virtual address; 'sh_addr' is
the address at which section's first byte should reside in memory, and
'sh_offset' is the byte offset from the beginning of the file to the
first byte in the section. The perf tool uses below formula to convert
a symbol's memory address to a file address:
file_address = st_value - sh_addr + sh_offset
^
` Memory address
We can see the final adjusted address ranges for buf1 and buf2 are
[0x30a8-0x30e8) and [0x3068-0x30a8) respectively, apparently this is
incorrect, in the code, the structure for 'buf1' and 'buf2' specifies
compiler attribute with 64-byte alignment.
The problem happens for 'sh_offset', libelf returns it as 0x3028 which
is not 64-byte aligned, combining with disassembly, it's likely libelf
doesn't respect the alignment for .bss section, therefore, it doesn't
return the aligned value for 'sh_offset'.
Suggested by Fangrui Song, ELF file contains program header which
contains PT_LOAD segments, the fields p_vaddr and p_offset in PT_LOAD
segments contain the execution info. A better choice for converting
memory address to file address is using the formula:
file_address = st_value - p_vaddr + p_offset
This patch introduces elf_read_program_header() which returns the
program header based on the passed 'st_value', then it uses the formula
above to calculate the symbol file address; and the debugging log is
updated respectively.
After applying the change:
# ./perf --debug verbose=4 mem report
...
dso__load_sym_internal: adjusting symbol: st_value: 0x40c0 p_vaddr: 0x3d28 p_offset: 0x2d28
symbol__new: buf2 0x30c0-0x3100
...
dso__load_sym_internal: adjusting symbol: st_value: 0x4080 p_vaddr: 0x3d28 p_offset: 0x2d28
symbol__new: buf1 0x3080-0x30c0
...
Fixes:
|
||
---|---|---|
.. | ||
arch | ||
bench | ||
Documentation | ||
examples/bpf | ||
include/bpf | ||
jvmti | ||
pmu-events | ||
python | ||
scripts | ||
tests | ||
trace | ||
ui | ||
util | ||
.gitignore | ||
Build | ||
builtin-annotate.c | ||
builtin-bench.c | ||
builtin-buildid-cache.c | ||
builtin-buildid-list.c | ||
builtin-c2c.c | ||
builtin-config.c | ||
builtin-data.c | ||
builtin-diff.c | ||
builtin-evlist.c | ||
builtin-ftrace.c | ||
builtin-help.c | ||
builtin-inject.c | ||
builtin-kallsyms.c | ||
builtin-kmem.c | ||
builtin-kvm.c | ||
builtin-list.c | ||
builtin-lock.c | ||
builtin-mem.c | ||
builtin-probe.c | ||
builtin-record.c | ||
builtin-report.c | ||
builtin-sched.c | ||
builtin-script.c | ||
builtin-stat.c | ||
builtin-timechart.c | ||
builtin-top.c | ||
builtin-trace.c | ||
builtin-version.c | ||
builtin.h | ||
check-headers.sh | ||
command-list.txt | ||
CREDITS | ||
design.txt | ||
Makefile | ||
Makefile.config | ||
Makefile.perf | ||
MANIFEST | ||
perf-archive.sh | ||
perf-completion.sh | ||
perf-read-vdso.c | ||
perf-sys.h | ||
perf-with-kcore.sh | ||
perf.c | ||
perf.h |