Bug 1295819

Description of the problem:

I have developed a prototype jit reader shared library for OpenJDK. It's purpose is to enable symbolic names for compiled Java methods be registered with gdb.

gdb recognises and invokes the hook function gdb_init_reader provided by my shared library. It also correctly invokes the read and destroy callbacks registered by gdb_init_reader. However, it never calls the get_frame_id and unwind_frame callbacks.

I debugged gdb (using gdb :-) and it turns out that the inferior process fails to install the required unwinder (jit_unwinder). This is supposed to be registered by function jit_inferior_created_hook.The only call to that function occurs in follow_exec (infrun.c) but follow_exec does not appear to be called when the JVM process is started.

I modified function post_create_inferior (infcmd.c) so that it calls jit_inferior_created_hook (I inserted the call at the end after the call to observer_notify_inferior_created) and this causes the relevant unwinder to be registered. With this patch the jit reader get_frame_id and unwind_frame callbacks are subsequently invoked and my prototype jit reader is able to list Java method and stub function names in backtraces.


Thgis was tested (and patched) on Fedora 21 with gdb 7.9.1. However, the same problem occurs with gdb 7.10.1 on Fedora 23.

How reproducible:

My JIT reader is fairly small and simple to build/use. However, it is not easy to reproduce the problem using this code because exercising the JIT reader requires building an OpenJDK release tweaked to generate info for JITted methods. I won't burden anyone with the details of how to achieve *that* unless it becomes absolutely necessary.

However, the bug does not need a working JIT reader (never mind the accompanying JIT compiler) to manifest the problem. A simple dummy JIT reader can be used to identify that the get_frame_id and unwind functions are not being called. You just need to implement the former to return an invalid frame (ip = 0x0, sp = 0x0) and the latter to return GDB_FAIL. You might add also add print statements to show whether they are being called or not. Alternatively, you can use gdb to debug the jit reader runnign in an inferior gdb and place breakpoints on the jit reader callbacks. I recommend placing the breakpoints on the unwinder function, jit_frame_sniffer.

Either method will also serve to show that my proposed patch does ensure the callbacks are invoked.

Steps to Reproduce:
1. implement adummy JIT reader
2. run gdb with gdb as target
3. from the outer gdb break jit_frame_sniffer
4. start the inner gdb from the outer gdb
5. register the dummy jit reader in the inferior gdb
4. start /any/ app from the inner gdb
5. hit Ctrl-C to break the app
6. type bt to exercise stack unwinding

Actual results:

jit_frame_sniffer will not get called.

Expected results:

jit_frame_sniffer should get called and should itself call the unwind callback.

Additional info:

The patch to fix this 7.9.1 is

$ diff -r gdb-7.9.1/gdb/infcmd.c tmp/gdb-7.9.1/gdb/infcmd.c
476,479d475
< 
<   // make sure we set up the jit reader
< 
<   jit_inferior_created_hook ();

The patch to fix this 7.10.1 is


$ diff -r gdb-7.10.1/gdb/infcmd.c tmp/gdb-7.10.1/gdb/infcmd.c
473,476d472
< 
<   // make sure we set up the jit reader
< 
<   jit_inferior_created_hook ();


I have several other issues with the way the jit reader works which I will report in separate bugzillas.