This is the mail archive of the gdb-testers@sourceware.org mailing list for the GDB project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

[binutils-gdb] Extend JIT-reader test and fix GDB problems that exposes


*** TEST RESULTS FOR COMMIT 20aa2c606ef682889722b03b1d874befa84fbf53 ***

Author: Pedro Alves <palves@redhat.com>
Branch: master
Commit: 20aa2c606ef682889722b03b1d874befa84fbf53

Extend JIT-reader test and fix GDB problems that exposes

The jit-reader.exp test isn't really exercising the jit-reader's
unwinder API at all.  This commit address that, and then fixes GDB
problems exposed.

- The custom JIT reader provided for the jit-reader.exp testcase
  always rejects the jitted function's frame...

  This is because the custom JIT reader in the testcase never ever
  sets state->code_begin/end, so the bounds check in
  gdb.base/jitreader.c:unwind_frame:

   if (this_ip >= state->code_end || this_ip < state->code_begin)
     return GDB_FAIL;

  tends to fail, unless you're "lucky" (because it references
  uninitialized data).

  The result is that GDB is always actually using a built-in unwinder
  for the jitted function.

- The provided unwinder doesn't do anything that GDB's built-in
  unwinder can't do.

  IOW, we can't really tell whether the JIT reader's unwinder is
  working or not.

  I fixed that by making the jitted function mangle its own stack
  pointer with a xor, and then teaching the jit unwinder to demangle
  it back (another xor).  So now "backtrace" with GDB's built-in
  unwinder fails while with the jit unwinder, it succeeds.

- GDB crashes after unloading the JIT reader, and flushing frames...

  I made the testcase use the "flushregs" command after unloading the
  JIT reader, to force the JIT frames to be flushed.  However, that
  crashes GDB...

  When reinit_frame_cache tears down a frame's cache, it calls its
  unwinder's dealloc_cache method, which for JIT frames ends up in
  jit.c:jit_dealloc_cache.  This function calls each of the frame's
  gdb_reg_value's "free" pointer:

   for (i = 0; i < gdbarch_num_regs (frame_arch); i++)
     if (priv_data->registers[i] && priv_data->registers[i]->free)
       priv_data->registers[i]->free (priv_data->registers[i]);

  and the problem is these gdb_reg_value instances have been returned
  by the JIT reader that has been already unloaded, and their "free"
  function pointers likely point to functions in the DSO that has
  already been unloaded...

  A fix for that could be to call reinit_frame_cache in
  jit_reader_unload_command _before_ unloading the jit reader DSO so
  that the jit reader is given a chance to clean up the gdb_reg_values
  before it is unloaded.  However, the fix for the point below makes
  this unnecessary, because it stops jit.c from keeping around
  gdb_reg_values in the first place.

- However, it still makes sense to clear the frame cache when loading
  or unloading a JIT unwinder.

  This makes testing a JIT unwinder a bit simpler.

- Not only the frame cache actually -- gdb is not unloading the
  jit-registered objfiles when the JIT reader is unloaded, and not
  loading the already-registered descriptors when a JIT reader is
  loaded.

  The new test exercises unloading the jit reader, loading it back
  again, and then making sure the JIT reader's unwinder works again.
  Without the unload/re-load of already-read descriptors, the newly
  loaded JIT would have no idea where the new function is, because
  it's stored at symbol read time.

- I added a couple "info frame" calls to the test, and that
  crashes GDB...

  The problem is that jit_frame_prev_register assumes it'll only be
  called for raw registers, so when it gets a pseudo register number,
  the "priv->registers[reg]" access is really an out-of-bounds access.

  To fix that, I made jit_frame_prev_register use
  gdbarch_pseudo_register_read_value for reading the pseudo-registers.
  However, that works with a regcache and we don't have one.  To fix
  that, I made the JIT unwinder store a regcache in its cache instead
  of an array of gdb_reg_value pointers.

gdb/ChangeLog:
2016-07-01  Pedro Alves  <palves@redhat.com>
	    Tom Tromey  <tom@tromey.com>

	* jit.c (jit_reader_load_command): Call reinit_frame_cache and
	jit_inferior_created_hook.
	(jit_reader_unload_command): Call reinit_frame_cache and
	jit_inferior_exit_hook.
	* jit.c (struct jit_unwind_private) <registers>: Delete field.
	<regcache>: New field.
	(jit_unwind_reg_set_impl): Set the register's value in the
	regcache.  Free the passed-in gdb_reg_value.
	(jit_dealloc_cache): Adjust to free the regcache.
	(jit_frame_sniffer): Allocate a regcache instead of an array of
	gdb_reg_value pointers.
	(jit_frame_this_id): Adjust.
	(jit_frame_prev_register): Read raw registers off of the regcache
	instead of from the gdb_reg_value pointer array.  Use
	gdbarch_pseudo_register_read_value to read pseudo registers.
	* regcache.c (regcache_raw_set_cached_value): New function,
	factored out from ...
	(regcache_raw_write): ... here.
	* regcache.h (regcache_raw_set_cached_value): Declare.

gdb/testsuite/ChangeLog:
2016-07-01  Pedro Alves  <palves@redhat.com>

	* gdb.base/jit-reader.exp (info_registers_current_frame): New
	procedure.
	(jit_reader_test): Test the jit reader's unwinder.
	* gdb.base/jithost.c (jit_function_00_code): New global.
	(main): Use memcpy to fill in the mmapped code, instead of poking
	bytes manually here.
	* gdb.base/jitreader.c (enum register_mapping) <AMD64_RBP>: New
	value.
	(read_debug_info): Save the function's range.
	(read_sp): New function.
	(unwind_frame): Use it.  Also unwind RBP.
	(get_frame_id): Use read_sp.
	(gdb_init_reader): Use calloc instead of malloc.
	* lib/gdb.exp (get_hexadecimal_valueof): Add optional 'test'
	parameter.  Use gdb_test_multiple.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]