| Message ID | cover.1725667397.git.scclevenger@os.amperecomputing.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | arm-cs-trace-disasm.py/perf must accommodate non-zero DSO text offset | expand |
Hello, Sorry for the long delay. But can you please explain your problem in more detail? Showing ELF program (or section) header would be helpful as well. Is the problem when the text mapping has non-zero pgoff only? Is the kernel symbols working correctly? Or is it just the Python script broken? Thanks, Namhyung On Mon, Sep 09, 2024 at 03:30:02PM -0600, Steve Clevenger wrote: > Changes in V8: > - in arm-cs-trace-disasm.py, ensure map_pgoff is not converted to > string. > - Remove map_pgoff integer conversion in dso not found print > message. > > Changes in V7: > - In arm-cs-trace-disasm.py, fix print message core dump resulting > from mixed type arithmetic. > - Modify CS_ETM_TRACE_ON filter to filter zero start_addr. The > CS_ETM_TRACE_ON message is changed to print only in verbose mode. > - Removed verbose mode only notification for start_addr/stop_addr > outside of dso address range. > > Changes in V6: > - In arm-cs-trace-disasm.py, zero map_pgoff for kernel files. Add > map_pgoff to start/end address for dso not found message. > - Added "Reviewed-by" trailer for patches 1-3 previously reviewed > by Leo Yan in V4 and V5. > > Changes in V5: > - In symbol-elf.c, branch to exit_close label if open file. > - In trace_event_python.c, correct indentation. set_sym_in_dict > call parameter "map_pgoff" renamed as "addr_map_pgoff" to > match local naming. > > Changes in V4: > - In trace-event-python.c, fixed perf-tools-next merge problem. > > Changes in V3: > - Rebased to linux-perf-tools branch. > - Squash symbol-elf.c and symbol.h into same commit. > - In map.c, merge dso__is_pie() call into existing if statement. > - In arm-cs-trace-disasm.py, remove debug artifacts. > > Changes in V2: > - In dso__is_pie() (symbol-elf.c), Decrease indentation, add null pointer > checks per Leo Yan review. > - Updated mailing list distribution. > > Steve Clevenger (4): > Add dso__is_pie call to identify ELF PIE > Force MAPPING_TYPE__IDENTIY for PIE > Add map pgoff to python dictionary based on MAPPING_TYPE > Adjust objdump start/end range per map pgoff parameter > > .../scripts/python/arm-cs-trace-disasm.py | 17 ++++-- > tools/perf/util/map.c | 4 +- > .../scripting-engines/trace-event-python.c | 13 +++- > tools/perf/util/symbol-elf.c | 61 +++++++++++++++++++ > tools/perf/util/symbol.h | 1 + > 5 files changed, 86 insertions(+), 10 deletions(-) > > -- > 2.44.0 >
Hello Namhyung, Sorry your question is so late. I don't include the ELF headers here, but the problem can be seen with a perf.data packet dump of user instruction trace capture. The problem is with the non-zero pgoff. The arm-cs-trace-disasm.py script was never passed pgoff information to adjust the start/end disassemble range passed to objdump. This patch distributes the fix between perf and the arm-cs-trace-disasm.py script. Here's a brief excerpt from an e-mail I sent to James Clark describing the patch before I submitted it. Regards, Steve C. ----------------------------------------------------------------- . . This e-mail is to document what I know and don’t know about the problem. The background is Fedora 38 introduced non-zero text offsets into common shared objects and executables (e.g. libc.so.6, etc.). If you were to ‘perf report --dump’ the perf.data of a user mode instruction trace on Fedora 37 and grep the PERF_RECORD_MMAP2 packets you’ll notice all zero values (@ 0) for shared binary and executable text offsets. Repeat the same for user trace collected on Fedora 38/39, and these text offsets show as non-zero. Fedora 37: 4294967295 18446744073709551615 0x8ac8 [0x78]: PERF_RECORD_MMAP2 2389577/2389577: [0xffff85124000(0x42000) @ 0 103:04 805306555 1941233998]: r-xp /usr/lib/ld-linux-aarch64.so.1 4294967295 18446744073709551615 0x8b40 [0x60]: PERF_RECORD_MMAP2 2389577/2389577: [0xffff85161000(0x1000) @ 0 00:00 0 0]: r-xp [vdso] 4294967295 18446744073709551615 0x8ba0 [0x68]: PERF_RECORD_MMAP2 2389577/2389577: [0xaaaab09b0000(0x10000) @ 0 103:04 1140851919 994218038]: r-xp /usr/bin/dd 4294967295 18446744073709551615 0x8c08 [0x70]: PERF_RECORD_MMAP2 2389577/2389577: [0xffff84f70000(0x1a9000) @ 0 103:04 1677721733 3891973508]: r-xp /usr/lib64/libc.so.6 Fedora 39: 4294967295 18446744073709551615 0x8ac8 [0x78]: PERF_RECORD_MMAP2 18229/18229: [0xffffa5512000(0x1d000) @ 0x10000 103:04 161 4093340249]: r-xp /usr/lib/ld-linux-aarch64.so.1 4294967295 18446744073709551615 0x8b40 [0x60]: PERF_RECORD_MMAP2 18229/18229: [0xffffa554f000(0x1000) @ 0 00:00 0 0]: r-xp [vdso] 4294967295 18446744073709551615 0x8ba0 [0x68]: PERF_RECORD_MMAP2 18229/18229: [0xaaaade7e0000(0xb000) @ 0x10000 103:04 536876423 421616320]: r-xp /usr/bin/dd 4294967295 18446744073709551615 0x8c08 [0x70]: PERF_RECORD_MMAP2 18229/18229: [0xffffa5360000(0x11f000) @ 0x30000 103:04 536873199 2415801053]: r-xp /usr/lib64/libc.so.6 The arm-cs-trace-disasm.py script never gets to see the text offset into the dso binaries, so has no opportunity to adjust the start/end address range passed to objdump. This wasn’t a problem on Fedora 37 and below since there’s no start/end adjustment for a zero text offset. On Fedora 38/39 distros, the instruction trace shows unconditional branch instructions which do not branch to the target address, the clearest indication of trouble. ----------------------------------------------------------------- On 10/7/2024 10:51 PM, Namhyung Kim wrote: > Hello, > > Sorry for the long delay. But can you please explain your problem in > more detail? Showing ELF program (or section) header would be helpful > as well. > > Is the problem when the text mapping has non-zero pgoff only? Is the > kernel symbols working correctly? Or is it just the Python script > broken? > > Thanks, > Namhyung > > On Mon, Sep 09, 2024 at 03:30:02PM -0600, Steve Clevenger wrote: >> Changes in V8: >> - in arm-cs-trace-disasm.py, ensure map_pgoff is not converted to >> string. >> - Remove map_pgoff integer conversion in dso not found print >> message. >> >> Changes in V7: >> - In arm-cs-trace-disasm.py, fix print message core dump resulting >> from mixed type arithmetic. >> - Modify CS_ETM_TRACE_ON filter to filter zero start_addr. The >> CS_ETM_TRACE_ON message is changed to print only in verbose mode. >> - Removed verbose mode only notification for start_addr/stop_addr >> outside of dso address range. >> >> Changes in V6: >> - In arm-cs-trace-disasm.py, zero map_pgoff for kernel files. Add >> map_pgoff to start/end address for dso not found message. >> - Added "Reviewed-by" trailer for patches 1-3 previously reviewed >> by Leo Yan in V4 and V5. >> >> Changes in V5: >> - In symbol-elf.c, branch to exit_close label if open file. >> - In trace_event_python.c, correct indentation. set_sym_in_dict >> call parameter "map_pgoff" renamed as "addr_map_pgoff" to >> match local naming. >> >> Changes in V4: >> - In trace-event-python.c, fixed perf-tools-next merge problem. >> >> Changes in V3: >> - Rebased to linux-perf-tools branch. >> - Squash symbol-elf.c and symbol.h into same commit. >> - In map.c, merge dso__is_pie() call into existing if statement. >> - In arm-cs-trace-disasm.py, remove debug artifacts. >> >> Changes in V2: >> - In dso__is_pie() (symbol-elf.c), Decrease indentation, add null pointer >> checks per Leo Yan review. >> - Updated mailing list distribution. >> >> Steve Clevenger (4): >> Add dso__is_pie call to identify ELF PIE >> Force MAPPING_TYPE__IDENTIY for PIE >> Add map pgoff to python dictionary based on MAPPING_TYPE >> Adjust objdump start/end range per map pgoff parameter >> >> .../scripts/python/arm-cs-trace-disasm.py | 17 ++++-- >> tools/perf/util/map.c | 4 +- >> .../scripting-engines/trace-event-python.c | 13 +++- >> tools/perf/util/symbol-elf.c | 61 +++++++++++++++++++ >> tools/perf/util/symbol.h | 1 + >> 5 files changed, 86 insertions(+), 10 deletions(-) >> >> -- >> 2.44.0 >>
On Tue, Oct 08, 2024 at 11:24:45AM -0700, Steve Clevenger wrote: > > Hello Namhyung, > > Sorry your question is so late. I don't include the ELF headers here, > but the problem can be seen with a perf.data packet dump of user > instruction trace capture. The problem is with the non-zero pgoff. The > arm-cs-trace-disasm.py script was never passed pgoff information to > adjust the start/end disassemble range passed to objdump. This patch > distributes the fix between perf and the arm-cs-trace-disasm.py script. > > Here's a brief excerpt from an e-mail I sent to James Clark describing > the patch before I submitted it. Oh, is it applied to the coresight tree already? I don't think I got any notification for that. Please make sure to notify perf maintainers (and hopefullly to get Ack's) when you touch non-coresight part of the code. Also I think we need to clarify how to route coresight tool patches. I thought they are going through the perf-tools tree. But it doesn't seem to be the case? Anyway, sorry about being late on this. > > ----------------------------------------------------------------- > . > . > This e-mail is to document what I know and don’t know about the problem. > The background is Fedora 38 introduced non-zero text offsets into common > shared objects and executables (e.g. libc.so.6, etc.). > > If you were to ‘perf report --dump’ the perf.data of a user mode > instruction trace on Fedora 37 and grep the PERF_RECORD_MMAP2 packets > you’ll notice all zero values (@ 0) for shared binary and executable > text offsets. Repeat the same for user trace collected on Fedora 38/39, > and these text offsets show as non-zero. > > Fedora 37: > > 4294967295 18446744073709551615 0x8ac8 [0x78]: PERF_RECORD_MMAP2 > 2389577/2389577: [0xffff85124000(0x42000) @ 0 103:04 805306555 > 1941233998]: r-xp /usr/lib/ld-linux-aarch64.so.1 > 4294967295 18446744073709551615 0x8b40 [0x60]: PERF_RECORD_MMAP2 > 2389577/2389577: [0xffff85161000(0x1000) @ 0 00:00 0 0]: r-xp [vdso] > 4294967295 18446744073709551615 0x8ba0 [0x68]: PERF_RECORD_MMAP2 > 2389577/2389577: [0xaaaab09b0000(0x10000) @ 0 103:04 1140851919 > 994218038]: r-xp /usr/bin/dd > 4294967295 18446744073709551615 0x8c08 [0x70]: PERF_RECORD_MMAP2 > 2389577/2389577: [0xffff84f70000(0x1a9000) @ 0 103:04 1677721733 > 3891973508]: r-xp /usr/lib64/libc.so.6 > > Fedora 39: > > 4294967295 18446744073709551615 0x8ac8 [0x78]: PERF_RECORD_MMAP2 > 18229/18229: [0xffffa5512000(0x1d000) @ 0x10000 103:04 161 4093340249]: > r-xp /usr/lib/ld-linux-aarch64.so.1 > 4294967295 18446744073709551615 0x8b40 [0x60]: PERF_RECORD_MMAP2 > 18229/18229: [0xffffa554f000(0x1000) @ 0 00:00 0 0]: r-xp [vdso] > 4294967295 18446744073709551615 0x8ba0 [0x68]: PERF_RECORD_MMAP2 > 18229/18229: [0xaaaade7e0000(0xb000) @ 0x10000 103:04 536876423 > 421616320]: r-xp /usr/bin/dd > 4294967295 18446744073709551615 0x8c08 [0x70]: PERF_RECORD_MMAP2 > 18229/18229: [0xffffa5360000(0x11f000) @ 0x30000 103:04 536873199 > 2415801053]: r-xp /usr/lib64/libc.so.6 > > The arm-cs-trace-disasm.py script never gets to see the text offset into > the dso binaries, so has no opportunity to adjust the start/end address > range passed to objdump. This wasn’t a problem on Fedora 37 and below > since there’s no start/end adjustment for a zero text offset. On Fedora > 38/39 distros, the instruction trace shows unconditional branch > instructions which do not branch to the target address, the clearest > indication of trouble. Ok, thanks for sharing this. I'm ok with exposing pgoff to the python script itself. But I'd like to make sure if it works correctly with PIEs that have non-zero page offsets in other places (probably ok). Thanks, Namhyung > ----------------------------------------------------------------- > > On 10/7/2024 10:51 PM, Namhyung Kim wrote: > > Hello, > > > > Sorry for the long delay. But can you please explain your problem in > > more detail? Showing ELF program (or section) header would be helpful > > as well. > > > > Is the problem when the text mapping has non-zero pgoff only? Is the > > kernel symbols working correctly? Or is it just the Python script > > broken? > > > > Thanks, > > Namhyung > > > > On Mon, Sep 09, 2024 at 03:30:02PM -0600, Steve Clevenger wrote: > >> Changes in V8: > >> - in arm-cs-trace-disasm.py, ensure map_pgoff is not converted to > >> string. > >> - Remove map_pgoff integer conversion in dso not found print > >> message. > >> > >> Changes in V7: > >> - In arm-cs-trace-disasm.py, fix print message core dump resulting > >> from mixed type arithmetic. > >> - Modify CS_ETM_TRACE_ON filter to filter zero start_addr. The > >> CS_ETM_TRACE_ON message is changed to print only in verbose mode. > >> - Removed verbose mode only notification for start_addr/stop_addr > >> outside of dso address range. > >> > >> Changes in V6: > >> - In arm-cs-trace-disasm.py, zero map_pgoff for kernel files. Add > >> map_pgoff to start/end address for dso not found message. > >> - Added "Reviewed-by" trailer for patches 1-3 previously reviewed > >> by Leo Yan in V4 and V5. > >> > >> Changes in V5: > >> - In symbol-elf.c, branch to exit_close label if open file. > >> - In trace_event_python.c, correct indentation. set_sym_in_dict > >> call parameter "map_pgoff" renamed as "addr_map_pgoff" to > >> match local naming. > >> > >> Changes in V4: > >> - In trace-event-python.c, fixed perf-tools-next merge problem. > >> > >> Changes in V3: > >> - Rebased to linux-perf-tools branch. > >> - Squash symbol-elf.c and symbol.h into same commit. > >> - In map.c, merge dso__is_pie() call into existing if statement. > >> - In arm-cs-trace-disasm.py, remove debug artifacts. > >> > >> Changes in V2: > >> - In dso__is_pie() (symbol-elf.c), Decrease indentation, add null pointer > >> checks per Leo Yan review. > >> - Updated mailing list distribution. > >> > >> Steve Clevenger (4): > >> Add dso__is_pie call to identify ELF PIE > >> Force MAPPING_TYPE__IDENTIY for PIE > >> Add map pgoff to python dictionary based on MAPPING_TYPE > >> Adjust objdump start/end range per map pgoff parameter > >> > >> .../scripts/python/arm-cs-trace-disasm.py | 17 ++++-- > >> tools/perf/util/map.c | 4 +- > >> .../scripting-engines/trace-event-python.c | 13 +++- > >> tools/perf/util/symbol-elf.c | 61 +++++++++++++++++++ > >> tools/perf/util/symbol.h | 1 + > >> 5 files changed, 86 insertions(+), 10 deletions(-) > >> > >> -- > >> 2.44.0 > >> >
On 10/8/2024 2:00 PM, Namhyung Kim wrote: > On Tue, Oct 08, 2024 at 11:24:45AM -0700, Steve Clevenger wrote: >> >> Hello Namhyung, >> >> Sorry your question is so late. I don't include the ELF headers here, >> but the problem can be seen with a perf.data packet dump of user >> instruction trace capture. The problem is with the non-zero pgoff. The >> arm-cs-trace-disasm.py script was never passed pgoff information to >> adjust the start/end disassemble range passed to objdump. This patch >> distributes the fix between perf and the arm-cs-trace-disasm.py script. >> >> Here's a brief excerpt from an e-mail I sent to James Clark describing >> the patch before I submitted it. > > Oh, is it applied to the coresight tree already? I don't think I got > any notification for that. Please make sure to notify perf maintainers > (and hopefullly to get Ack's) when you touch non-coresight part of the > code. > > Also I think we need to clarify how to route coresight tool patches. I > thought they are going through the perf-tools tree. But it doesn't seem > to be the case? > > Anyway, sorry about being late on this. FYI. These patch(es) were based on perf-tools-next at revision 6.11.0-rc3. All ack'd by Leo Yan. If they could get merged to perf-tools, that would be great! > >> >> ----------------------------------------------------------------- >> . >> . >> This e-mail is to document what I know and don’t know about the problem. >> The background is Fedora 38 introduced non-zero text offsets into common >> shared objects and executables (e.g. libc.so.6, etc.). >> >> If you were to ‘perf report --dump’ the perf.data of a user mode >> instruction trace on Fedora 37 and grep the PERF_RECORD_MMAP2 packets >> you’ll notice all zero values (@ 0) for shared binary and executable >> text offsets. Repeat the same for user trace collected on Fedora 38/39, >> and these text offsets show as non-zero. >> >> Fedora 37: >> >> 4294967295 18446744073709551615 0x8ac8 [0x78]: PERF_RECORD_MMAP2 >> 2389577/2389577: [0xffff85124000(0x42000) @ 0 103:04 805306555 >> 1941233998]: r-xp /usr/lib/ld-linux-aarch64.so.1 >> 4294967295 18446744073709551615 0x8b40 [0x60]: PERF_RECORD_MMAP2 >> 2389577/2389577: [0xffff85161000(0x1000) @ 0 00:00 0 0]: r-xp [vdso] >> 4294967295 18446744073709551615 0x8ba0 [0x68]: PERF_RECORD_MMAP2 >> 2389577/2389577: [0xaaaab09b0000(0x10000) @ 0 103:04 1140851919 >> 994218038]: r-xp /usr/bin/dd >> 4294967295 18446744073709551615 0x8c08 [0x70]: PERF_RECORD_MMAP2 >> 2389577/2389577: [0xffff84f70000(0x1a9000) @ 0 103:04 1677721733 >> 3891973508]: r-xp /usr/lib64/libc.so.6 >> >> Fedora 39: >> >> 4294967295 18446744073709551615 0x8ac8 [0x78]: PERF_RECORD_MMAP2 >> 18229/18229: [0xffffa5512000(0x1d000) @ 0x10000 103:04 161 4093340249]: >> r-xp /usr/lib/ld-linux-aarch64.so.1 >> 4294967295 18446744073709551615 0x8b40 [0x60]: PERF_RECORD_MMAP2 >> 18229/18229: [0xffffa554f000(0x1000) @ 0 00:00 0 0]: r-xp [vdso] >> 4294967295 18446744073709551615 0x8ba0 [0x68]: PERF_RECORD_MMAP2 >> 18229/18229: [0xaaaade7e0000(0xb000) @ 0x10000 103:04 536876423 >> 421616320]: r-xp /usr/bin/dd >> 4294967295 18446744073709551615 0x8c08 [0x70]: PERF_RECORD_MMAP2 >> 18229/18229: [0xffffa5360000(0x11f000) @ 0x30000 103:04 536873199 >> 2415801053]: r-xp /usr/lib64/libc.so.6 >> >> The arm-cs-trace-disasm.py script never gets to see the text offset into >> the dso binaries, so has no opportunity to adjust the start/end address >> range passed to objdump. This wasn’t a problem on Fedora 37 and below >> since there’s no start/end adjustment for a zero text offset. On Fedora >> 38/39 distros, the instruction trace shows unconditional branch >> instructions which do not branch to the target address, the clearest >> indication of trouble. > > Ok, thanks for sharing this. I'm ok with exposing pgoff to the python > script itself. But I'd like to make sure if it works correctly with > PIEs that have non-zero page offsets in other places (probably ok). > > Thanks, > Namhyung > > >> ----------------------------------------------------------------- >> >> On 10/7/2024 10:51 PM, Namhyung Kim wrote: >>> Hello, >>> >>> Sorry for the long delay. But can you please explain your problem in >>> more detail? Showing ELF program (or section) header would be helpful >>> as well. >>> >>> Is the problem when the text mapping has non-zero pgoff only? Is the >>> kernel symbols working correctly? Or is it just the Python script >>> broken? >>> >>> Thanks, >>> Namhyung >>> >>> On Mon, Sep 09, 2024 at 03:30:02PM -0600, Steve Clevenger wrote: >>>> Changes in V8: >>>> - in arm-cs-trace-disasm.py, ensure map_pgoff is not converted to >>>> string. >>>> - Remove map_pgoff integer conversion in dso not found print >>>> message. >>>> >>>> Changes in V7: >>>> - In arm-cs-trace-disasm.py, fix print message core dump resulting >>>> from mixed type arithmetic. >>>> - Modify CS_ETM_TRACE_ON filter to filter zero start_addr. The >>>> CS_ETM_TRACE_ON message is changed to print only in verbose mode. >>>> - Removed verbose mode only notification for start_addr/stop_addr >>>> outside of dso address range. >>>> >>>> Changes in V6: >>>> - In arm-cs-trace-disasm.py, zero map_pgoff for kernel files. Add >>>> map_pgoff to start/end address for dso not found message. >>>> - Added "Reviewed-by" trailer for patches 1-3 previously reviewed >>>> by Leo Yan in V4 and V5. >>>> >>>> Changes in V5: >>>> - In symbol-elf.c, branch to exit_close label if open file. >>>> - In trace_event_python.c, correct indentation. set_sym_in_dict >>>> call parameter "map_pgoff" renamed as "addr_map_pgoff" to >>>> match local naming. >>>> >>>> Changes in V4: >>>> - In trace-event-python.c, fixed perf-tools-next merge problem. >>>> >>>> Changes in V3: >>>> - Rebased to linux-perf-tools branch. >>>> - Squash symbol-elf.c and symbol.h into same commit. >>>> - In map.c, merge dso__is_pie() call into existing if statement. >>>> - In arm-cs-trace-disasm.py, remove debug artifacts. >>>> >>>> Changes in V2: >>>> - In dso__is_pie() (symbol-elf.c), Decrease indentation, add null pointer >>>> checks per Leo Yan review. >>>> - Updated mailing list distribution. >>>> >>>> Steve Clevenger (4): >>>> Add dso__is_pie call to identify ELF PIE >>>> Force MAPPING_TYPE__IDENTIY for PIE >>>> Add map pgoff to python dictionary based on MAPPING_TYPE >>>> Adjust objdump start/end range per map pgoff parameter >>>> >>>> .../scripts/python/arm-cs-trace-disasm.py | 17 ++++-- >>>> tools/perf/util/map.c | 4 +- >>>> .../scripting-engines/trace-event-python.c | 13 +++- >>>> tools/perf/util/symbol-elf.c | 61 +++++++++++++++++++ >>>> tools/perf/util/symbol.h | 1 + >>>> 5 files changed, 86 insertions(+), 10 deletions(-) >>>> >>>> -- >>>> 2.44.0 >>>> >>
Hi Namhyung, On 10/8/2024 10:00 PM, Namhyung Kim wrote: [...] > On Tue, Oct 08, 2024 at 11:24:45AM -0700, Steve Clevenger wrote: >> >> Hello Namhyung, >> >> Sorry your question is so late. I don't include the ELF headers here, >> but the problem can be seen with a perf.data packet dump of user >> instruction trace capture. The problem is with the non-zero pgoff. The >> arm-cs-trace-disasm.py script was never passed pgoff information to >> adjust the start/end disassemble range passed to objdump. This patch >> distributes the fix between perf and the arm-cs-trace-disasm.py script. >> >> Here's a brief excerpt from an e-mail I sent to James Clark describing >> the patch before I submitted it. > > Oh, is it applied to the coresight tree already? No. > I don't think I got> any notification for that. Please make sure to notify perf maintainers > (and hopefullly to get Ack's) when you touch non-coresight part of the > code. > > Also I think we need to clarify how to route coresight tool patches. I > thought they are going through the perf-tools tree. But it doesn't seem > to be the case? AFAIK, in most cases, Arm patches for the perf will be picked up via perf-tools tree, including this patch series. [...] >> Fedora 39: >> >> 4294967295 18446744073709551615 0x8ac8 [0x78]: PERF_RECORD_MMAP2 >> 18229/18229: [0xffffa5512000(0x1d000) @ 0x10000 103:04 161 4093340249]: >> r-xp /usr/lib/ld-linux-aarch64.so.1 >> >> 4294967295 18446744073709551615 0x8b40 [0x60]: PERF_RECORD_MMAP2 >> 18229/18229: [0xffffa554f000(0x1000) @ 0 00:00 0 0]: r-xp [vdso] >> 4294967295 18446744073709551615 0x8ba0 [0x68]: PERF_RECORD_MMAP2 >> 18229/18229: [0xaaaade7e0000(0xb000) @ 0x10000 103:04 536876423 >> 421616320]: r-xp /usr/bin/dd >> 4294967295 18446744073709551615 0x8c08 [0x70]: PERF_RECORD_MMAP2 >> 18229/18229: [0xffffa5360000(0x11f000) @ 0x30000 103:04 536873199 >> 2415801053]: r-xp /usr/lib64/libc.so.6 >> >> The arm-cs-trace-disasm.py script never gets to see the text offset into >> the dso binaries, so has no opportunity to adjust the start/end address >> range passed to objdump. This wasn’t a problem on Fedora 37 and below >> since there’s no start/end adjustment for a zero text offset. On Fedora >> 38/39 distros, the instruction trace shows unconditional branch >> instructions which do not branch to the target address, the clearest >> indication of trouble. > > Ok, thanks for sharing this. I'm ok with exposing pgoff to the python > script itself. But I'd like to make sure if it works correctly with > PIEs that have non-zero page offsets in other places (probably ok). Fair point. The 'pgoff' is exposed from the kernel's VMA info, it is a runtime value after the DSO loading. After set the MAPPING_TYPE__IDENTITY type for PIE DSO, the 'pgoff' will be ignored in both the perf tool (see map__map_ip()) and the script. Sorry that I missed this part when reviewed the change. I understood Steve have verified the script result. Seems to me, the critical question is kernel has an offset for PIE DSO but why ignore it in the tool. @Steve, if you have more info, could you explain a bit what is the reason for ignoring pgoff for PIE DSO? Thanks, Leo
On Wed, Oct 09, 2024 at 04:24:29PM +0100, Leo Yan wrote: > Hi Namhyung, > > On 10/8/2024 10:00 PM, Namhyung Kim wrote: > > [...] > > > On Tue, Oct 08, 2024 at 11:24:45AM -0700, Steve Clevenger wrote: > >> > >> Hello Namhyung, > >> > >> Sorry your question is so late. I don't include the ELF headers here, > >> but the problem can be seen with a perf.data packet dump of user > >> instruction trace capture. The problem is with the non-zero pgoff. The > >> arm-cs-trace-disasm.py script was never passed pgoff information to > >> adjust the start/end disassemble range passed to objdump. This patch > >> distributes the fix between perf and the arm-cs-trace-disasm.py script. > >> > >> Here's a brief excerpt from an e-mail I sent to James Clark describing > >> the patch before I submitted it. > > > > Oh, is it applied to the coresight tree already? > > No. > > > I don't think I got> any notification for that. Please make sure to notify > perf maintainers > > (and hopefullly to get Ack's) when you touch non-coresight part of the > > code. > > > > Also I think we need to clarify how to route coresight tool patches. I > > thought they are going through the perf-tools tree. But it doesn't seem > > to be the case? > > AFAIK, in most cases, Arm patches for the perf will be picked up via > perf-tools tree, including this patch series. Ok, I was confused and thought it was applied there. Thanks, Namhyung > > [...] > >> Fedora 39: > >> > >> 4294967295 18446744073709551615 0x8ac8 [0x78]: PERF_RECORD_MMAP2 > >> 18229/18229: [0xffffa5512000(0x1d000) @ 0x10000 103:04 161 4093340249]: > >> r-xp /usr/lib/ld-linux-aarch64.so.1 > >> > >> 4294967295 18446744073709551615 0x8b40 [0x60]: PERF_RECORD_MMAP2 > >> 18229/18229: [0xffffa554f000(0x1000) @ 0 00:00 0 0]: r-xp [vdso] > >> 4294967295 18446744073709551615 0x8ba0 [0x68]: PERF_RECORD_MMAP2 > >> 18229/18229: [0xaaaade7e0000(0xb000) @ 0x10000 103:04 536876423 > >> 421616320]: r-xp /usr/bin/dd > >> 4294967295 18446744073709551615 0x8c08 [0x70]: PERF_RECORD_MMAP2 > >> 18229/18229: [0xffffa5360000(0x11f000) @ 0x30000 103:04 536873199 > >> 2415801053]: r-xp /usr/lib64/libc.so.6 > >> > >> The arm-cs-trace-disasm.py script never gets to see the text offset into > >> the dso binaries, so has no opportunity to adjust the start/end address > >> range passed to objdump. This wasn’t a problem on Fedora 37 and below > >> since there’s no start/end adjustment for a zero text offset. On Fedora > >> 38/39 distros, the instruction trace shows unconditional branch > >> instructions which do not branch to the target address, the clearest > >> indication of trouble. > > > > Ok, thanks for sharing this. I'm ok with exposing pgoff to the python > > script itself. But I'd like to make sure if it works correctly with > > PIEs that have non-zero page offsets in other places (probably ok). > > Fair point. The 'pgoff' is exposed from the kernel's VMA info, it is a runtime > value after the DSO loading. After set the MAPPING_TYPE__IDENTITY type for PIE > DSO, the 'pgoff' will be ignored in both the perf tool (see map__map_ip()) and > the script. > > Sorry that I missed this part when reviewed the change. I understood Steve > have verified the script result. Seems to me, the critical question is kernel > has an offset for PIE DSO but why ignore it in the tool. > > @Steve, if you have more info, could you explain a bit what is the reason for > ignoring pgoff for PIE DSO? > > Thanks, > Leo
Changes in V8: - in arm-cs-trace-disasm.py, ensure map_pgoff is not converted to string. - Remove map_pgoff integer conversion in dso not found print message. Changes in V7: - In arm-cs-trace-disasm.py, fix print message core dump resulting from mixed type arithmetic. - Modify CS_ETM_TRACE_ON filter to filter zero start_addr. The CS_ETM_TRACE_ON message is changed to print only in verbose mode. - Removed verbose mode only notification for start_addr/stop_addr outside of dso address range. Changes in V6: - In arm-cs-trace-disasm.py, zero map_pgoff for kernel files. Add map_pgoff to start/end address for dso not found message. - Added "Reviewed-by" trailer for patches 1-3 previously reviewed by Leo Yan in V4 and V5. Changes in V5: - In symbol-elf.c, branch to exit_close label if open file. - In trace_event_python.c, correct indentation. set_sym_in_dict call parameter "map_pgoff" renamed as "addr_map_pgoff" to match local naming. Changes in V4: - In trace-event-python.c, fixed perf-tools-next merge problem. Changes in V3: - Rebased to linux-perf-tools branch. - Squash symbol-elf.c and symbol.h into same commit. - In map.c, merge dso__is_pie() call into existing if statement. - In arm-cs-trace-disasm.py, remove debug artifacts. Changes in V2: - In dso__is_pie() (symbol-elf.c), Decrease indentation, add null pointer checks per Leo Yan review. - Updated mailing list distribution. Steve Clevenger (4): Add dso__is_pie call to identify ELF PIE Force MAPPING_TYPE__IDENTIY for PIE Add map pgoff to python dictionary based on MAPPING_TYPE Adjust objdump start/end range per map pgoff parameter .../scripts/python/arm-cs-trace-disasm.py | 17 ++++-- tools/perf/util/map.c | 4 +- .../scripting-engines/trace-event-python.c | 13 +++- tools/perf/util/symbol-elf.c | 61 +++++++++++++++++++ tools/perf/util/symbol.h | 1 + 5 files changed, 86 insertions(+), 10 deletions(-)