[RFA] Fix regcache leak, and avoid possible regcache access after detach.

Message ID 20190216205810.1253-1-philippe.waroquiers@skynet.be
State New
Headers show
Series
  • [RFA] Fix regcache leak, and avoid possible regcache access after detach.
Related show

Commit Message

Philippe Waroquiers Feb. 16, 2019, 8:58 p.m.
Valgrind reports leaks like the below in various tests,
e.g. gdb.threads/attach-slow-waitpid.exp, gdb.ada/task_switch_in_core.exp, ...

Fix the leak by clearing the regcache when detaching from an inferior.
Note that these leaks are 'created' when GDB exits,
when the regcache::current_regcache is destroyed : the elements
of the forward_list are pointers, and the 'pointed to' memory is not
deleted by the forward_list destructor.

Nevertheless, fixing this leak is good as it makes a bunch of
tests 'leak clean'.

Also, it seems strange to keep a register cache for a process from
which GDB detached : it is not clear if this cache is still valid
after detach.  And effectively, when clearing only the regcache,
(and not the frame cache), then the frame cache was still 'pointing'
at this regcache and was used when switching to the child process
in the test gdb.threads/watchpoint-fork.exp, which seems strange.

So, we solve the leak and avoid possible accesses to the regcache
and frame cache of the detached inferior, by clearing both the
regcache and the frame cache.

Tested on debian/amd64, natively and under Valgrind.

==27679== VALGRIND_GDB_ERROR_BEGIN
==27679== 1,123 (72 direct, 1,051 indirect) bytes in 1 blocks are definitely lost in loss record 2,942 of 3,400
==27679==    at 0x4C2C4CC: operator new(unsigned long) (vg_replace_malloc.c:344)
==27679==    by 0x5CDF71: get_thread_arch_aspace_regcache(ptid_t, gdbarch*, address_space*) (regcache.c:330)
==27679==    by 0x5CE12A: get_thread_regcache (regcache.c:366)
==27679==    by 0x5CE12A: get_current_regcache() (regcache.c:372)
==27679==    by 0x4FF63D: post_create_inferior(target_ops*, int) (infcmd.c:452)
==27679==    by 0x43AF62: core_target_open(char const*, int) (corelow.c:458)
==27679==    by 0x408B68: cmd_func(cmd_list_element*, char const*, int) (cli-decode.c:1892)
==27679==    by 0x66B399: execute_command(char const*, int) (top.c:630)
==27679==    by 0x4B4D2B: command_handler(char const*) (event-top.c:583)
==27679==    by 0x4B505C: command_line_handler(std::unique_ptr<char, gdb::xfree_deleter<char> >&&) (event-top.c:770)
==27679==    by 0x4B3EA3: gdb_rl_callback_handler(char*) (event-top.c:213)
==27679==    by 0x4E68AC2: rl_callback_read_char (callback.c:283)
==27679==    by 0x4B3DDD: gdb_rl_callback_read_char_wrapper_noexcept() (event-top.c:175)
==27679==    by 0x4B3E48: gdb_rl_callback_read_char_wrapper(void*) (event-top.c:192)
==27679==    by 0x4B43EF: stdin_event_handler(int, void*) (event-top.c:511)
==27679==    by 0x4B31BC: gdb_wait_for_event(int) (event-loop.c:859)
==27679==    by 0x4B32F3: gdb_do_one_event() [clone .part.4] (event-loop.c:347)
==27679==    by 0x4B3484: gdb_do_one_event (common-exceptions.h:219)
==27679==    by 0x4B3484: start_event_loop() (event-loop.c:371)
==27679==    by 0x549967: captured_command_loop() (main.c:330)
==27679==    by 0x54A95C: captured_main (main.c:1176)
==27679==    by 0x54A95C: gdb_main(captured_main_args*) (main.c:1192)
==27679==    by 0x297517: main (gdb.c:32)

gdb/ChangeLog
2019-02-16  Philippe Waroquiers  <philippe.waroquiers@skynet.be>

	* target.c (target_detach): Clear the regcache and the
	frame cache.
---
 gdb/target.c | 11 +++++++++++
 1 file changed, 11 insertions(+)

-- 
2.20.1

Comments

Kevin Buettner Feb. 19, 2019, 11:44 p.m. | #1
On Sat, 16 Feb 2019 21:58:10 +0100
Philippe Waroquiers <philippe.waroquiers@skynet.be> wrote:

> Valgrind reports leaks like the below in various tests,

> e.g. gdb.threads/attach-slow-waitpid.exp, gdb.ada/task_switch_in_core.exp, ...

> 

> Fix the leak by clearing the regcache when detaching from an inferior.

> Note that these leaks are 'created' when GDB exits,

> when the regcache::current_regcache is destroyed : the elements

> of the forward_list are pointers, and the 'pointed to' memory is not

> deleted by the forward_list destructor.

> 

> Nevertheless, fixing this leak is good as it makes a bunch of

> tests 'leak clean'.

> 

> Also, it seems strange to keep a register cache for a process from

> which GDB detached : it is not clear if this cache is still valid

> after detach.  And effectively, when clearing only the regcache,

> (and not the frame cache), then the frame cache was still 'pointing'

> at this regcache and was used when switching to the child process

> in the test gdb.threads/watchpoint-fork.exp, which seems strange.

> 

> So, we solve the leak and avoid possible accesses to the regcache

> and frame cache of the detached inferior, by clearing both the

> regcache and the frame cache.

> 

> Tested on debian/amd64, natively and under Valgrind.


Before committing, could you also test against native-gdbserver?
I.e...

make check RUNTESTFLAGS="--target_board=native-gdbserver"

(I'd like the detach code in remote.c to be exercised as well.  I've
run into some subtle bugs in this area of remote.c in the past.)

My remaining comments consist of a few nits...

> diff --git a/gdb/target.c b/gdb/target.c

> index 116510e8cb..0f10628590 100644

> --- a/gdb/target.c

> +++ b/gdb/target.c

> @@ -2010,6 +2010,8 @@ target_preopen (int from_tty)

>  void

>  target_detach (inferior *inf, int from_tty)

>  {

> +  ptid_t pid_ptid = ptid_t (inf->pid);

> +


Please move this declaration / initialization to the line just before
where it's used.  Since it's only used once, you could also just pass
ptid_t (inf->pid) directly to registers_changed_ptid() and get rid
of the declaration entirely.

>    /* As long as some to_detach implementations rely on the current_inferior

>       (either directly, or indirectly, like through target_gdbarch or by

>       reading memory), INF needs to be the current inferior.  When that

> @@ -2029,6 +2031,15 @@ target_detach (inferior *inf, int from_tty)

>    prepare_for_detach ();

>  

>    current_top_target ()->detach (inf, from_tty);

> +

> +  /* After we have detached, clear the register cache for this inferior.  */

> +  registers_changed_ptid (pid_ptid);

> +

> +  /* We have to ensure we have no frame cached left.  Normally,


s/cached/cache/

> +     registers_changed_ptid (pid_ptid) calls reinit_frame_cache when

> +     inferior_ptid matches pid_ptid, but in our case, it does not

> +     call it, as inferior_pid has been reset.  */


s/inferior_pid/inferior_ptid/

Thanks for that comment though; I was wondering why reinit_frame_cache
needed to be called.

> +  reinit_frame_cache ();

>  }


Okay with those changes, assuming that all goes well with the
native-gdbserver testing.
Philippe Waroquiers Feb. 27, 2019, 9:44 p.m. | #2
On Tue, 2019-02-19 at 16:44 -0700, Kevin Buettner wrote:
> Okay with those changes, assuming that all goes well with the

> native-gdbserver testing.

Thanks for the review.

I have just pushed the patch, after doing the suggested fixes,
successfully tested with native gdbserver, and waited for the
branching just to be sure.

Philippe

Patch

diff --git a/gdb/target.c b/gdb/target.c
index 116510e8cb..0f10628590 100644
--- a/gdb/target.c
+++ b/gdb/target.c
@@ -2010,6 +2010,8 @@  target_preopen (int from_tty)
 void
 target_detach (inferior *inf, int from_tty)
 {
+  ptid_t pid_ptid = ptid_t (inf->pid);
+
   /* As long as some to_detach implementations rely on the current_inferior
      (either directly, or indirectly, like through target_gdbarch or by
      reading memory), INF needs to be the current inferior.  When that
@@ -2029,6 +2031,15 @@  target_detach (inferior *inf, int from_tty)
   prepare_for_detach ();
 
   current_top_target ()->detach (inf, from_tty);
+
+  /* After we have detached, clear the register cache for this inferior.  */
+  registers_changed_ptid (pid_ptid);
+
+  /* We have to ensure we have no frame cached left.  Normally,
+     registers_changed_ptid (pid_ptid) calls reinit_frame_cache when
+     inferior_ptid matches pid_ptid, but in our case, it does not
+     call it, as inferior_pid has been reset.  */
+  reinit_frame_cache ();
 }
 
 void