libffi trampolines and stack traces? : was Overly brief stack traces for Java/linux ?


Bradley Schatz
 

Thanks for your help with this - and for the sketch of how to proceed with it.

I'm going to look into the feasibility of fixing the frame pointer in the trampoline before attempting what you have outlined.

Cheers!


On 18/4/21, 3:37 am, "Y Song" <ys114321@...> wrote:

On Fri, Apr 16, 2021 at 12:56 PM Yonghong Song via lists.iovisor.org
<ys114321=gmail.com@...> wrote:
>
> On Thu, Apr 15, 2021 at 11:09 PM Bradley Schatz
> <bradley@...> wrote:
> >
> > I still digging into this issue, and have hacked memleak/bcc to show addresses when they cant be resolved. In other places, I'm seeing jit complied java stack frames showing up alongside C ones, which is the expected behaviour when using Java with -XX:+PreserveFramePointer.
> >
> > I'm still seeing these allocation traces with only two frames per the below.
> >
> > 11534336 bytes in 11 allocations from stack
> > [7efeebb115d4 unknown] [jna7632521838566054573.tmp ]
> > [7efec27f7440 unknown] [perf-15445.map]
> > 14680064 bytes in 14 allocations from stack
> > [7efeebb115d4 unknown] [jna7632521838566054573.tmp ]
> > [7efec27f7380 unknown] [perf-15445.map]
> >
> > I suspect the lower frame above is a libffi generated trampoline. Are there any known issues with the eBPF stack tracing infrastructure and such trampolines? Workarounds?
>
> The bpf stack unwinder is using the kernel one which is the frame
> pointer based and it may have issues with generated trampoline code
> which may mess up frame pointer based stack chain. One possibility is
> to use perf call-graph "dwarf" mode to get the raw data to user space
> and use more powerful library like libunwind etc. to unwind the stack.
> But I haven't do this together with bpf program yet, and cannot
> describe whether and how to do "dwarf" call-graph with bpf program
> together.

I briefly looked at the perf and kernel code and experimented with
perf dwarf mode.
It is possible for bpf to copy user stack to user space. You could do
the following:
- when you do perf event open, the sample_type needs to be set
properly. The following
is what callchain dwarf mode had:
sample_type = PERF_SAMPLE_IP|PERF_SAMPLE_TID|PERF_SAMPLE_TIME|
PERF_SAMPLE_ADDR|PERF_SAMPLE_CALLCHAIN|PERF_SAMPLE_CPU|
PERF_SAMPLE_PERIOD|PERF_SAMPLE_REGS_USER|
PERF_SAMPLE_STACK_USER|PERF_SAMPLE_DATA_SRC

I do not think you need all of them, you definitely need PERF_SAMPLE_REGS_USER,
maybe you can PERF_SAMPLE_RAW and PERF_SAMPLE_REGS_USER
to see whether it works or not.

In kernel header file linux:include/linux/perf_event.h, we have
struct perf_sample_data {
...
struct perf_regs regs_user;
struct perf_regs regs_intr;
u64 stack_user_size;
}

If you have access to perf_sample_data, you can read regs_user and find
stack pointer and then you can copy some bytes (say 2KB) to user space
for analysis. Using perf_event_output may be too expensive. perf is using
user mmap memory. You can also use a map for this purpose.

How to access perf_sample_data?
Typically sampling bpf program has the signature like
int do_sample(struct bpf_perf_event_data *ctx)
During program run, the program really got the following ctx
struct bpf_perf_event_data_kern {
bpf_user_pt_regs_t *regs;
struct perf_sample_data *data;
struct perf_event *event;
};
So you can use bpf_probe_read_kernel() to get 'data' and then other
information inside perf_sample_data.

You can experiment this way if you really like to see whether it can
solve your problem or not. If it works, I guess we can add a bpf helper
to do the copy as such copy length tends to be big.

>
> >
> > Thanks,
> > Bradley
> >
> >
> >
> > On 7/4/21, 9:04 pm, "Bradley Schatz" <bradley@...> wrote:
> >
> > > What does '[unknown] [perf-18047.map]' mean? Does this mean
> > > perf-18047.map is not found? If the perf-<pid>.map file cannot be found,
> > > symbolization won't be possible. Maybe you want to double check this?
> >
> > The file perf-18047.map is there and from other parts of the stack trace I can see it being used to successfully resolve symbols.
> >
> > Thanks!
> >
> >
> >
> > On 7/4/21, 4:44 am, "Y Song" <ys114321@...> wrote:
> >
> > On Mon, Apr 5, 2021 at 10:08 PM Bradley Schatz
> > <bradley@...> wrote:
> > >
> > > Thanks for the suggestion. I found a tunable to keep the JNI shared library in memory after loading. As you can see below, it is no longer showing as deleted.
> > >
> > > 13238272 bytes in 404 allocations from stack
> > > [unknown] [jna2576903844543447777.tmp]
> > > [unknown] [perf-18047.map]
> >
> > I have no experience with perf-map-agent, but the following is what I guess:
> > [perf-18047.map] is used to find the mapping between address and symbol.
> > What does '[unknown] [perf-18047.map]' mean? Does this mean
> > perf-18047.map is not found? If the perf-<pid>.map file cannot be found,
> > symbolization won't be possible. Maybe you want to double check this?
> >
> > >
> > > No improvement in granularity though.
> > >
> > > In the VM I'm using -XX:+PreserveFramePointer -XX:+UnlockDiagnosticVMOptions -XX:+DebugNonSafepoints. In perf_maps_agent, I'm using "unfoldall"
> > >
> > > Any other suggestions?
> > >
> > > Thanks!
> > >
> > >
> > >
> > >
> > > On 3/4/21, 2:42 am, "Y Song" <ys114321@...> wrote:
> > >
> > > On Wed, Mar 31, 2021 at 11:25 PM Bradley Schatz
> > > <bradley@...> wrote:
> > > >
> > > > Hi,
> > > >
> > > >
> > > >
> > > > I’m just starting to come to grips with bcc & perf-map-agent for introspecting java on linux, with the goal of identifying what appears to be an off-heap memory leak (using memleak).
> > > >
> > > >
> > > >
> > > > I appear to be getting reliable stack decoding for jvm library code and for jit’ed java methods (see below for an example of the former). However I am seeing some very short stack traces which don’t seem to decode (the latter three stacks of below).
> > > >
> > > >
> > > >
> > > > It’s looking to me like the frame starting with “jna…” is likely the native JNI shared library for the FFI library “JNA”.
> > > >
> > > >
> > > >
> > > > Any suggestions as to why these latter three are so brief and/or how I can increase the resolution?
> > >
> > > I can see the file has been marked as deleted.
> > >
> > > 34603008 bytes in 33 allocations from stack
> > >
> > > [unknown] [jna9005484735610534564.tmp (deleted)]
> > >
> > > [unknown] [perf-31566.map]
> > >
> > > 96468992 bytes in 92 allocations from stack
> > >
> > > [unknown] [jna9005484735610534564.tmp (deleted)]
> > >
> > > [unknown] [perf-31566.map]
> > >
> > > So the file has been removed in userspace and current bcc won't be
> > > pass to parse it since it takes the file name as
> > > "jna9005484735610534564.tmp (deleted)"
> > > The file name is actually taken from /proc/<pid>/maps.
> > >
> > > I am not sure whether you can hack to parse "jna9005484735610534564.tmp" or not.
> > > But I would consider it is unsafe to do that as the original file
> > > related info may just
> > > exist in kernel and there is a reference to it. For user space, it is
> > > either gone or
> > > could be replaced by something else. So the safest way is to find a place to
> > > do symbolization before file is gone or keep tmp file a little bit longer.
> > > >
> > > >
> > > >
> > > > Apologies if this is the wrong place for such a question. Thank you for your help.
> > > >
> > > >
> > > >
> > > > Kind regards,
> > > >
> > > > Bradley
> > > >
> > > >
> > > >
> > > >
> > > >
> > > >
> > > >
> > > >
> > > >
> > > > 119408 bytes in 71 allocations from stack
> > > >
> > > > os::malloc(unsigned long, MemoryType, NativeCallStack const&)+0xb5 [libjvm.so]
> > > >
> > > > CodeBlob::set_oop_maps(OopMapSet*) [clone .part.5]+0x75 [libjvm.so]
> > > >
> > > > CodeBlob::CodeBlob(char const*, CodeBuffer*, int, int, int, int, OopMapSet*)+0xe3 [libjvm.so]
> > > >
> > > > nmethod::nmethod(Method*, int, int, int, CodeOffsets*, int, DebugInformationRecorder*, Dependencies*, CodeBuffer*, int, OopMapSet*, ExceptionHandlerTable*, ImplicitExceptionTable*, AbstractCompiler*, int)+0x4d [libjvm.so]
> > > >
> > > > nmethod::new_nmethod(methodHandle, int, int, CodeOffsets*, int, DebugInformationRecorder*, Dependencies*, CodeBuffer*, int, OopMapSet*, ExceptionHandlerTable*, ImplicitExceptionTable*, AbstractCompiler*, int)+0x219 [libjvm.so]
> > > >
> > > > ciEnv::register_method(ciMethod*, int, CodeOffsets*, int, CodeBuffer*, int, OopMapSet*, ExceptionHandlerTable*, ImplicitExceptionTable*, AbstractCompiler*, int, bool, bool, RTMState)+0x1b1 [libjvm.so]
> > > >
> > > > Compile::Compile(ciEnv*, C2Compiler*, ciMethod*, int, bool, bool, bool)+0xe60 [libjvm.so]
> > > >
> > > > C2Compiler::compile_method(ciEnv*, ciMethod*, int)+0xa3 [libjvm.so]
> > > >
> > > > CompileBroker::invoke_compiler_on_method(CompileTask*)+0x808 [libjvm.so]
> > > >
> > > > CompileBroker::compiler_thread_loop()+0x6d8 [libjvm.so]
> > > >
> > > > JavaThread::thread_main_inner()+0x1c7 [libjvm.so]
> > > >
> > > > JavaThread::run()+0x2fa [libjvm.so]
> > > >
> > > > java_start(Thread*)+0x102 [libjvm.so]
> > > >
> > > > start_thread+0xf3 [libpthread-2.28.so]
> > > >
> > > > 34603008 bytes in 33 allocations from stack
> > > >
> > > > [unknown] [jna9005484735610534564.tmp (deleted)]
> > > >
> > > > [unknown] [perf-31566.map]
> > > >
> > > > 96468992 bytes in 92 allocations from stack
> > > >
> > > > [unknown] [jna9005484735610534564.tmp (deleted)]
> > > >
> > > > [unknown] [perf-31566.map]
> > > >
> > > > 295698432 bytes in 282 allocations from stack
> > > >
> > > > [unknown] [jna9005484735610534564.tmp (deleted)]
> > > >
> > > > [unknown] [perf-31566.map]
> > > >
> > > >
> > > >
> > > >
> > >
> >
> >
>
>
>
>
>


Yonghong Song
 

On Fri, Apr 16, 2021 at 12:56 PM Yonghong Song via lists.iovisor.org
<ys114321=gmail.com@...> wrote:

On Thu, Apr 15, 2021 at 11:09 PM Bradley Schatz
<bradley@...> wrote:

I still digging into this issue, and have hacked memleak/bcc to show addresses when they cant be resolved. In other places, I'm seeing jit complied java stack frames showing up alongside C ones, which is the expected behaviour when using Java with -XX:+PreserveFramePointer.

I'm still seeing these allocation traces with only two frames per the below.

11534336 bytes in 11 allocations from stack
[7efeebb115d4 unknown] [jna7632521838566054573.tmp ]
[7efec27f7440 unknown] [perf-15445.map]
14680064 bytes in 14 allocations from stack
[7efeebb115d4 unknown] [jna7632521838566054573.tmp ]
[7efec27f7380 unknown] [perf-15445.map]

I suspect the lower frame above is a libffi generated trampoline. Are there any known issues with the eBPF stack tracing infrastructure and such trampolines? Workarounds?
The bpf stack unwinder is using the kernel one which is the frame
pointer based and it may have issues with generated trampoline code
which may mess up frame pointer based stack chain. One possibility is
to use perf call-graph "dwarf" mode to get the raw data to user space
and use more powerful library like libunwind etc. to unwind the stack.
But I haven't do this together with bpf program yet, and cannot
describe whether and how to do "dwarf" call-graph with bpf program
together.
I briefly looked at the perf and kernel code and experimented with
perf dwarf mode.
It is possible for bpf to copy user stack to user space. You could do
the following:
- when you do perf event open, the sample_type needs to be set
properly. The following
is what callchain dwarf mode had:
sample_type = PERF_SAMPLE_IP|PERF_SAMPLE_TID|PERF_SAMPLE_TIME|
PERF_SAMPLE_ADDR|PERF_SAMPLE_CALLCHAIN|PERF_SAMPLE_CPU|
PERF_SAMPLE_PERIOD|PERF_SAMPLE_REGS_USER|
PERF_SAMPLE_STACK_USER|PERF_SAMPLE_DATA_SRC

I do not think you need all of them, you definitely need PERF_SAMPLE_REGS_USER,
maybe you can PERF_SAMPLE_RAW and PERF_SAMPLE_REGS_USER
to see whether it works or not.

In kernel header file linux:include/linux/perf_event.h, we have
struct perf_sample_data {
...
struct perf_regs regs_user;
struct perf_regs regs_intr;
u64 stack_user_size;
}

If you have access to perf_sample_data, you can read regs_user and find
stack pointer and then you can copy some bytes (say 2KB) to user space
for analysis. Using perf_event_output may be too expensive. perf is using
user mmap memory. You can also use a map for this purpose.

How to access perf_sample_data?
Typically sampling bpf program has the signature like
int do_sample(struct bpf_perf_event_data *ctx)
During program run, the program really got the following ctx
struct bpf_perf_event_data_kern {
bpf_user_pt_regs_t *regs;
struct perf_sample_data *data;
struct perf_event *event;
};
So you can use bpf_probe_read_kernel() to get 'data' and then other
information inside perf_sample_data.

You can experiment this way if you really like to see whether it can
solve your problem or not. If it works, I guess we can add a bpf helper
to do the copy as such copy length tends to be big.



Thanks,
Bradley



On 7/4/21, 9:04 pm, "Bradley Schatz" <bradley@...> wrote:

> What does '[unknown] [perf-18047.map]' mean? Does this mean
> perf-18047.map is not found? If the perf-<pid>.map file cannot be found,
> symbolization won't be possible. Maybe you want to double check this?

The file perf-18047.map is there and from other parts of the stack trace I can see it being used to successfully resolve symbols.

Thanks!



On 7/4/21, 4:44 am, "Y Song" <ys114321@...> wrote:

On Mon, Apr 5, 2021 at 10:08 PM Bradley Schatz
<bradley@...> wrote:
>
> Thanks for the suggestion. I found a tunable to keep the JNI shared library in memory after loading. As you can see below, it is no longer showing as deleted.
>
> 13238272 bytes in 404 allocations from stack
> [unknown] [jna2576903844543447777.tmp]
> [unknown] [perf-18047.map]

I have no experience with perf-map-agent, but the following is what I guess:
[perf-18047.map] is used to find the mapping between address and symbol.
What does '[unknown] [perf-18047.map]' mean? Does this mean
perf-18047.map is not found? If the perf-<pid>.map file cannot be found,
symbolization won't be possible. Maybe you want to double check this?

>
> No improvement in granularity though.
>
> In the VM I'm using -XX:+PreserveFramePointer -XX:+UnlockDiagnosticVMOptions -XX:+DebugNonSafepoints. In perf_maps_agent, I'm using "unfoldall"
>
> Any other suggestions?
>
> Thanks!
>
>
>
>
> On 3/4/21, 2:42 am, "Y Song" <ys114321@...> wrote:
>
> On Wed, Mar 31, 2021 at 11:25 PM Bradley Schatz
> <bradley@...> wrote:
> >
> > Hi,
> >
> >
> >
> > I’m just starting to come to grips with bcc & perf-map-agent for introspecting java on linux, with the goal of identifying what appears to be an off-heap memory leak (using memleak).
> >
> >
> >
> > I appear to be getting reliable stack decoding for jvm library code and for jit’ed java methods (see below for an example of the former). However I am seeing some very short stack traces which don’t seem to decode (the latter three stacks of below).
> >
> >
> >
> > It’s looking to me like the frame starting with “jna…” is likely the native JNI shared library for the FFI library “JNA”.
> >
> >
> >
> > Any suggestions as to why these latter three are so brief and/or how I can increase the resolution?
>
> I can see the file has been marked as deleted.
>
> 34603008 bytes in 33 allocations from stack
>
> [unknown] [jna9005484735610534564.tmp (deleted)]
>
> [unknown] [perf-31566.map]
>
> 96468992 bytes in 92 allocations from stack
>
> [unknown] [jna9005484735610534564.tmp (deleted)]
>
> [unknown] [perf-31566.map]
>
> So the file has been removed in userspace and current bcc won't be
> pass to parse it since it takes the file name as
> "jna9005484735610534564.tmp (deleted)"
> The file name is actually taken from /proc/<pid>/maps.
>
> I am not sure whether you can hack to parse "jna9005484735610534564.tmp" or not.
> But I would consider it is unsafe to do that as the original file
> related info may just
> exist in kernel and there is a reference to it. For user space, it is
> either gone or
> could be replaced by something else. So the safest way is to find a place to
> do symbolization before file is gone or keep tmp file a little bit longer.
> >
> >
> >
> > Apologies if this is the wrong place for such a question. Thank you for your help.
> >
> >
> >
> > Kind regards,
> >
> > Bradley
> >
> >
> >
> >
> >
> >
> >
> >
> >
> > 119408 bytes in 71 allocations from stack
> >
> > os::malloc(unsigned long, MemoryType, NativeCallStack const&)+0xb5 [libjvm.so]
> >
> > CodeBlob::set_oop_maps(OopMapSet*) [clone .part.5]+0x75 [libjvm.so]
> >
> > CodeBlob::CodeBlob(char const*, CodeBuffer*, int, int, int, int, OopMapSet*)+0xe3 [libjvm.so]
> >
> > nmethod::nmethod(Method*, int, int, int, CodeOffsets*, int, DebugInformationRecorder*, Dependencies*, CodeBuffer*, int, OopMapSet*, ExceptionHandlerTable*, ImplicitExceptionTable*, AbstractCompiler*, int)+0x4d [libjvm.so]
> >
> > nmethod::new_nmethod(methodHandle, int, int, CodeOffsets*, int, DebugInformationRecorder*, Dependencies*, CodeBuffer*, int, OopMapSet*, ExceptionHandlerTable*, ImplicitExceptionTable*, AbstractCompiler*, int)+0x219 [libjvm.so]
> >
> > ciEnv::register_method(ciMethod*, int, CodeOffsets*, int, CodeBuffer*, int, OopMapSet*, ExceptionHandlerTable*, ImplicitExceptionTable*, AbstractCompiler*, int, bool, bool, RTMState)+0x1b1 [libjvm.so]
> >
> > Compile::Compile(ciEnv*, C2Compiler*, ciMethod*, int, bool, bool, bool)+0xe60 [libjvm.so]
> >
> > C2Compiler::compile_method(ciEnv*, ciMethod*, int)+0xa3 [libjvm.so]
> >
> > CompileBroker::invoke_compiler_on_method(CompileTask*)+0x808 [libjvm.so]
> >
> > CompileBroker::compiler_thread_loop()+0x6d8 [libjvm.so]
> >
> > JavaThread::thread_main_inner()+0x1c7 [libjvm.so]
> >
> > JavaThread::run()+0x2fa [libjvm.so]
> >
> > java_start(Thread*)+0x102 [libjvm.so]
> >
> > start_thread+0xf3 [libpthread-2.28.so]
> >
> > 34603008 bytes in 33 allocations from stack
> >
> > [unknown] [jna9005484735610534564.tmp (deleted)]
> >
> > [unknown] [perf-31566.map]
> >
> > 96468992 bytes in 92 allocations from stack
> >
> > [unknown] [jna9005484735610534564.tmp (deleted)]
> >
> > [unknown] [perf-31566.map]
> >
> > 295698432 bytes in 282 allocations from stack
> >
> > [unknown] [jna9005484735610534564.tmp (deleted)]
> >
> > [unknown] [perf-31566.map]
> >
> >
> >
> >
>





Yonghong Song
 

On Thu, Apr 15, 2021 at 11:09 PM Bradley Schatz
<bradley@...> wrote:

I still digging into this issue, and have hacked memleak/bcc to show addresses when they cant be resolved. In other places, I'm seeing jit complied java stack frames showing up alongside C ones, which is the expected behaviour when using Java with -XX:+PreserveFramePointer.

I'm still seeing these allocation traces with only two frames per the below.

11534336 bytes in 11 allocations from stack
[7efeebb115d4 unknown] [jna7632521838566054573.tmp ]
[7efec27f7440 unknown] [perf-15445.map]
14680064 bytes in 14 allocations from stack
[7efeebb115d4 unknown] [jna7632521838566054573.tmp ]
[7efec27f7380 unknown] [perf-15445.map]

I suspect the lower frame above is a libffi generated trampoline. Are there any known issues with the eBPF stack tracing infrastructure and such trampolines? Workarounds?
The bpf stack unwinder is using the kernel one which is the frame
pointer based and it may have issues with generated trampoline code
which may mess up frame pointer based stack chain. One possibility is
to use perf call-graph "dwarf" mode to get the raw data to user space
and use more powerful library like libunwind etc. to unwind the stack.
But I haven't do this together with bpf program yet, and cannot
describe whether and how to do "dwarf" call-graph with bpf program
together.


Thanks,
Bradley



On 7/4/21, 9:04 pm, "Bradley Schatz" <bradley@...> wrote:

> What does '[unknown] [perf-18047.map]' mean? Does this mean
> perf-18047.map is not found? If the perf-<pid>.map file cannot be found,
> symbolization won't be possible. Maybe you want to double check this?

The file perf-18047.map is there and from other parts of the stack trace I can see it being used to successfully resolve symbols.

Thanks!



On 7/4/21, 4:44 am, "Y Song" <ys114321@...> wrote:

On Mon, Apr 5, 2021 at 10:08 PM Bradley Schatz
<bradley@...> wrote:
>
> Thanks for the suggestion. I found a tunable to keep the JNI shared library in memory after loading. As you can see below, it is no longer showing as deleted.
>
> 13238272 bytes in 404 allocations from stack
> [unknown] [jna2576903844543447777.tmp]
> [unknown] [perf-18047.map]

I have no experience with perf-map-agent, but the following is what I guess:
[perf-18047.map] is used to find the mapping between address and symbol.
What does '[unknown] [perf-18047.map]' mean? Does this mean
perf-18047.map is not found? If the perf-<pid>.map file cannot be found,
symbolization won't be possible. Maybe you want to double check this?

>
> No improvement in granularity though.
>
> In the VM I'm using -XX:+PreserveFramePointer -XX:+UnlockDiagnosticVMOptions -XX:+DebugNonSafepoints. In perf_maps_agent, I'm using "unfoldall"
>
> Any other suggestions?
>
> Thanks!
>
>
>
>
> On 3/4/21, 2:42 am, "Y Song" <ys114321@...> wrote:
>
> On Wed, Mar 31, 2021 at 11:25 PM Bradley Schatz
> <bradley@...> wrote:
> >
> > Hi,
> >
> >
> >
> > I’m just starting to come to grips with bcc & perf-map-agent for introspecting java on linux, with the goal of identifying what appears to be an off-heap memory leak (using memleak).
> >
> >
> >
> > I appear to be getting reliable stack decoding for jvm library code and for jit’ed java methods (see below for an example of the former). However I am seeing some very short stack traces which don’t seem to decode (the latter three stacks of below).
> >
> >
> >
> > It’s looking to me like the frame starting with “jna…” is likely the native JNI shared library for the FFI library “JNA”.
> >
> >
> >
> > Any suggestions as to why these latter three are so brief and/or how I can increase the resolution?
>
> I can see the file has been marked as deleted.
>
> 34603008 bytes in 33 allocations from stack
>
> [unknown] [jna9005484735610534564.tmp (deleted)]
>
> [unknown] [perf-31566.map]
>
> 96468992 bytes in 92 allocations from stack
>
> [unknown] [jna9005484735610534564.tmp (deleted)]
>
> [unknown] [perf-31566.map]
>
> So the file has been removed in userspace and current bcc won't be
> pass to parse it since it takes the file name as
> "jna9005484735610534564.tmp (deleted)"
> The file name is actually taken from /proc/<pid>/maps.
>
> I am not sure whether you can hack to parse "jna9005484735610534564.tmp" or not.
> But I would consider it is unsafe to do that as the original file
> related info may just
> exist in kernel and there is a reference to it. For user space, it is
> either gone or
> could be replaced by something else. So the safest way is to find a place to
> do symbolization before file is gone or keep tmp file a little bit longer.
> >
> >
> >
> > Apologies if this is the wrong place for such a question. Thank you for your help.
> >
> >
> >
> > Kind regards,
> >
> > Bradley
> >
> >
> >
> >
> >
> >
> >
> >
> >
> > 119408 bytes in 71 allocations from stack
> >
> > os::malloc(unsigned long, MemoryType, NativeCallStack const&)+0xb5 [libjvm.so]
> >
> > CodeBlob::set_oop_maps(OopMapSet*) [clone .part.5]+0x75 [libjvm.so]
> >
> > CodeBlob::CodeBlob(char const*, CodeBuffer*, int, int, int, int, OopMapSet*)+0xe3 [libjvm.so]
> >
> > nmethod::nmethod(Method*, int, int, int, CodeOffsets*, int, DebugInformationRecorder*, Dependencies*, CodeBuffer*, int, OopMapSet*, ExceptionHandlerTable*, ImplicitExceptionTable*, AbstractCompiler*, int)+0x4d [libjvm.so]
> >
> > nmethod::new_nmethod(methodHandle, int, int, CodeOffsets*, int, DebugInformationRecorder*, Dependencies*, CodeBuffer*, int, OopMapSet*, ExceptionHandlerTable*, ImplicitExceptionTable*, AbstractCompiler*, int)+0x219 [libjvm.so]
> >
> > ciEnv::register_method(ciMethod*, int, CodeOffsets*, int, CodeBuffer*, int, OopMapSet*, ExceptionHandlerTable*, ImplicitExceptionTable*, AbstractCompiler*, int, bool, bool, RTMState)+0x1b1 [libjvm.so]
> >
> > Compile::Compile(ciEnv*, C2Compiler*, ciMethod*, int, bool, bool, bool)+0xe60 [libjvm.so]
> >
> > C2Compiler::compile_method(ciEnv*, ciMethod*, int)+0xa3 [libjvm.so]
> >
> > CompileBroker::invoke_compiler_on_method(CompileTask*)+0x808 [libjvm.so]
> >
> > CompileBroker::compiler_thread_loop()+0x6d8 [libjvm.so]
> >
> > JavaThread::thread_main_inner()+0x1c7 [libjvm.so]
> >
> > JavaThread::run()+0x2fa [libjvm.so]
> >
> > java_start(Thread*)+0x102 [libjvm.so]
> >
> > start_thread+0xf3 [libpthread-2.28.so]
> >
> > 34603008 bytes in 33 allocations from stack
> >
> > [unknown] [jna9005484735610534564.tmp (deleted)]
> >
> > [unknown] [perf-31566.map]
> >
> > 96468992 bytes in 92 allocations from stack
> >
> > [unknown] [jna9005484735610534564.tmp (deleted)]
> >
> > [unknown] [perf-31566.map]
> >
> > 295698432 bytes in 282 allocations from stack
> >
> > [unknown] [jna9005484735610534564.tmp (deleted)]
> >
> > [unknown] [perf-31566.map]
> >
> >
> >
> >
>


Bradley Schatz
 

I still digging into this issue, and have hacked memleak/bcc to show addresses when they cant be resolved. In other places, I'm seeing jit complied java stack frames showing up alongside C ones, which is the expected behaviour when using Java with -XX:+PreserveFramePointer.

I'm still seeing these allocation traces with only two frames per the below.

11534336 bytes in 11 allocations from stack
[7efeebb115d4 unknown] [jna7632521838566054573.tmp ]
[7efec27f7440 unknown] [perf-15445.map]
14680064 bytes in 14 allocations from stack
[7efeebb115d4 unknown] [jna7632521838566054573.tmp ]
[7efec27f7380 unknown] [perf-15445.map]

I suspect the lower frame above is a libffi generated trampoline. Are there any known issues with the eBPF stack tracing infrastructure and such trampolines? Workarounds?

Thanks,
Bradley



On 7/4/21, 9:04 pm, "Bradley Schatz" <bradley@...> wrote:

> What does '[unknown] [perf-18047.map]' mean? Does this mean
> perf-18047.map is not found? If the perf-<pid>.map file cannot be found,
> symbolization won't be possible. Maybe you want to double check this?

The file perf-18047.map is there and from other parts of the stack trace I can see it being used to successfully resolve symbols.

Thanks!



On 7/4/21, 4:44 am, "Y Song" <ys114321@...> wrote:

On Mon, Apr 5, 2021 at 10:08 PM Bradley Schatz
<bradley@...> wrote:
>
> Thanks for the suggestion. I found a tunable to keep the JNI shared library in memory after loading. As you can see below, it is no longer showing as deleted.
>
> 13238272 bytes in 404 allocations from stack
> [unknown] [jna2576903844543447777.tmp]
> [unknown] [perf-18047.map]

I have no experience with perf-map-agent, but the following is what I guess:
[perf-18047.map] is used to find the mapping between address and symbol.
What does '[unknown] [perf-18047.map]' mean? Does this mean
perf-18047.map is not found? If the perf-<pid>.map file cannot be found,
symbolization won't be possible. Maybe you want to double check this?

>
> No improvement in granularity though.
>
> In the VM I'm using -XX:+PreserveFramePointer -XX:+UnlockDiagnosticVMOptions -XX:+DebugNonSafepoints. In perf_maps_agent, I'm using "unfoldall"
>
> Any other suggestions?
>
> Thanks!
>
>
>
>
> On 3/4/21, 2:42 am, "Y Song" <ys114321@...> wrote:
>
> On Wed, Mar 31, 2021 at 11:25 PM Bradley Schatz
> <bradley@...> wrote:
> >
> > Hi,
> >
> >
> >
> > I’m just starting to come to grips with bcc & perf-map-agent for introspecting java on linux, with the goal of identifying what appears to be an off-heap memory leak (using memleak).
> >
> >
> >
> > I appear to be getting reliable stack decoding for jvm library code and for jit’ed java methods (see below for an example of the former). However I am seeing some very short stack traces which don’t seem to decode (the latter three stacks of below).
> >
> >
> >
> > It’s looking to me like the frame starting with “jna…” is likely the native JNI shared library for the FFI library “JNA”.
> >
> >
> >
> > Any suggestions as to why these latter three are so brief and/or how I can increase the resolution?
>
> I can see the file has been marked as deleted.
>
> 34603008 bytes in 33 allocations from stack
>
> [unknown] [jna9005484735610534564.tmp (deleted)]
>
> [unknown] [perf-31566.map]
>
> 96468992 bytes in 92 allocations from stack
>
> [unknown] [jna9005484735610534564.tmp (deleted)]
>
> [unknown] [perf-31566.map]
>
> So the file has been removed in userspace and current bcc won't be
> pass to parse it since it takes the file name as
> "jna9005484735610534564.tmp (deleted)"
> The file name is actually taken from /proc/<pid>/maps.
>
> I am not sure whether you can hack to parse "jna9005484735610534564.tmp" or not.
> But I would consider it is unsafe to do that as the original file
> related info may just
> exist in kernel and there is a reference to it. For user space, it is
> either gone or
> could be replaced by something else. So the safest way is to find a place to
> do symbolization before file is gone or keep tmp file a little bit longer.
> >
> >
> >
> > Apologies if this is the wrong place for such a question. Thank you for your help.
> >
> >
> >
> > Kind regards,
> >
> > Bradley
> >
> >
> >
> >
> >
> >
> >
> >
> >
> > 119408 bytes in 71 allocations from stack
> >
> > os::malloc(unsigned long, MemoryType, NativeCallStack const&)+0xb5 [libjvm.so]
> >
> > CodeBlob::set_oop_maps(OopMapSet*) [clone .part.5]+0x75 [libjvm.so]
> >
> > CodeBlob::CodeBlob(char const*, CodeBuffer*, int, int, int, int, OopMapSet*)+0xe3 [libjvm.so]
> >
> > nmethod::nmethod(Method*, int, int, int, CodeOffsets*, int, DebugInformationRecorder*, Dependencies*, CodeBuffer*, int, OopMapSet*, ExceptionHandlerTable*, ImplicitExceptionTable*, AbstractCompiler*, int)+0x4d [libjvm.so]
> >
> > nmethod::new_nmethod(methodHandle, int, int, CodeOffsets*, int, DebugInformationRecorder*, Dependencies*, CodeBuffer*, int, OopMapSet*, ExceptionHandlerTable*, ImplicitExceptionTable*, AbstractCompiler*, int)+0x219 [libjvm.so]
> >
> > ciEnv::register_method(ciMethod*, int, CodeOffsets*, int, CodeBuffer*, int, OopMapSet*, ExceptionHandlerTable*, ImplicitExceptionTable*, AbstractCompiler*, int, bool, bool, RTMState)+0x1b1 [libjvm.so]
> >
> > Compile::Compile(ciEnv*, C2Compiler*, ciMethod*, int, bool, bool, bool)+0xe60 [libjvm.so]
> >
> > C2Compiler::compile_method(ciEnv*, ciMethod*, int)+0xa3 [libjvm.so]
> >
> > CompileBroker::invoke_compiler_on_method(CompileTask*)+0x808 [libjvm.so]
> >
> > CompileBroker::compiler_thread_loop()+0x6d8 [libjvm.so]
> >
> > JavaThread::thread_main_inner()+0x1c7 [libjvm.so]
> >
> > JavaThread::run()+0x2fa [libjvm.so]
> >
> > java_start(Thread*)+0x102 [libjvm.so]
> >
> > start_thread+0xf3 [libpthread-2.28.so]
> >
> > 34603008 bytes in 33 allocations from stack
> >
> > [unknown] [jna9005484735610534564.tmp (deleted)]
> >
> > [unknown] [perf-31566.map]
> >
> > 96468992 bytes in 92 allocations from stack
> >
> > [unknown] [jna9005484735610534564.tmp (deleted)]
> >
> > [unknown] [perf-31566.map]
> >
> > 295698432 bytes in 282 allocations from stack
> >
> > [unknown] [jna9005484735610534564.tmp (deleted)]
> >
> > [unknown] [perf-31566.map]
> >
> >
> >
> >
>