Re: [PATCH v2] [gdb] Fix heap-use-after-free in select_event_lwp

[Tom de Vries <tdevries@suse.de>]

On 2/21/24 18:42, Pedro Alves wrote:
> On 2024-02-19 15:04, Tom de Vries wrote:
>> On 2/9/24 16:46, Pedro Alves wrote:
>>> On 2024-01-23 11:48, Tom de Vries wrote:
>>> It sounds like we were about to report a stop for a thread that isn't marked as stopped?
>>> Now it looks to me that _that_ would be the bug to fix.
>>
>> This patch adds an assert to catch the bug you mention, and a fix in wait_lwp:
>> ...
>> diff --git a/gdb/linux-nat.c b/gdb/linux-nat.c
>> index e91c57ba239..5022da9abd2 100644
>> --- a/gdb/linux-nat.c
>> +++ b/gdb/linux-nat.c
>> @@ -2210,6 +2210,7 @@ wait_lwp (struct lwp_info *lp)
>>            core.  Store it in lp->waitstatus, because lp->status
>>            would be ambiguous (W_EXITCODE(0,0) == 0).  */
>>             lp->waitstatus = host_status_to_waitstatus (status);
>> +          lp->stopped = 1;
>>             return 0;
>>           }
>>
>> @@ -3368,6 +3369,7 @@ linux_nat_wait_1 (ptid_t ptid, struct target_waitstatus *ourstatus,
>>       }
>>
>>     gdb_assert (lp);
>> +  gdb_assert (lp->stopped);
>>
>>     status = lp->status;
>>     lp->status = 0;
>> ...
>>
>> This fixes the problem observed in the PR, and passes testing on x86_64-linux and aarch64-linux.
>>
>> WDYT?
> 
> Ah, thanks, so it was an exiting lwp that we're going to report the exit for.
> 
> Yes, this is the right thing to do.  We do the same in linux_nat_filter_event:
> 
>    /* This LWP is stopped now.  (And if dead, this prevents it from
>       ever being continued.)  */
>    lp->stopped = 1;
> 
> 
> gdbserver has a mark_lwp_dead function to centralize setting up various lwp flags such that the
> rest of the code doesn't mishandle them, and it seems like a good idea to do a similar thing in
> gdb as well.  Like below.  I ran the testsuite and saw no regressions.
> 

I agree that it's a good idea, and also ran the test-suite and saw no 
regressions.

> I tried to reproduce the gdb.base/vfork-follow-parent.exp issue described in the bug, but when I
> run that with a gdb built with -O0 -fsanitize=thread, the testcase just seemingly hangs starting
> up until I ctrl-c:
> 
>   Running /home/pedro/rocm/gdb/build-master-sanitize-thread/gdb/testsuite/../../../src/gdb/testsuite/gdb.base/vfork-follow-parent.exp ...
>   WARNING: Couldn't set the height to 0
>   WARNING: Couldn't set the width to 0.
>   ^Cgot a INT signal, interrupted by user
> 
> and when I look at gdb.log, I see it is full of the same report over and over:
> 
> ...
> WARNING: ThreadSanitizer: signal-unsafe call inside of a signal (pid=3086741)
>      #0 free ../../../../src/libsanitizer/tsan/tsan_interceptors_posix.cpp:711 (libtsan.so.0+0x37ab8)
>      #1 __gconv_close_transform iconv/gconv_db.c:800 (libc.so.6+0x2be3a)
>      #2 ext_lang_initialization() ../../src/gdb/extension.c:339 (gdb+0x5c6a3f)
>      #3 captured_main_1 ../../src/gdb/main.c:1058 (gdb+0x7c94da)
>      #4 captured_main ../../src/gdb/main.c:1329 (gdb+0x7ca7be)
>      #5 gdb_main(captured_main_args*) ../../src/gdb/main.c:1358 (gdb+0x7ca886)
>      #6 main ../../src/gdb/gdb.c:39 (gdb+0xf98b3)
> 
> SUMMARY: ThreadSanitizer: signal-unsafe call inside of a signal iconv/gconv_db.c:800 in __gconv_close_transform
> ...
> 
> 

Hmm, I can't reproduce that, not on aarch64-linux f39, nor on Leap 15.4 
x86_64-linux.

At first glance this could be a tsan problem, given that the stack trace 
doesn't show entry into a signal handler.

> Does this patch work for you?
> 

It does.

As mentioned in the PR, on aarch64-linux this reproduces for me 5/10 
times, and with your patch 0/10 times.

Given that pretty much the entire patch is yours, do you want to proceed 
with this, or do you want me to integrate this into my patch with a 
Co-Authored-By tag?

Thanks,
- Tom

>  From 8cbc62be0dab8bb4b09fe05a52a348ddbfb2e33c Mon Sep 17 00:00:00 2001
> From: Pedro Alves <pedro@palves.net>
> Date: Wed, 21 Feb 2024 16:23:55 +0000
> Subject: [PATCH] mark_lwp_dead
> 
> Change-Id: I4a6169976f89bf714c478cbb2b7d4c32365e62a9
> ---
>   gdb/linux-nat.c | 34 +++++++++++++++++++++++++---------
>   1 file changed, 25 insertions(+), 9 deletions(-)
> 
> diff --git a/gdb/linux-nat.c b/gdb/linux-nat.c
> index e91c57ba239..b3d319e0a82 100644
> --- a/gdb/linux-nat.c
> +++ b/gdb/linux-nat.c
> @@ -2123,6 +2123,27 @@ wait_for_signal ()
>       }
>   }
>   
> +/* Mark LWP dead, with STATUS as exit status pending to report
> +   later.  */
> +
> +static void
> +mark_lwp_dead (lwp_info *lp, int status)
> +{
> +  /* Store the exit status lp->waitstatus, because lp->status would be
> +     ambiguous (W_EXITCODE(0,0) == 0).  */
> +  lp->waitstatus = host_status_to_waitstatus (status);
> +
> +  /* If we're processing LP's status, there should be no other event
> +     already recorded as pending.  */
> +  gdb_assert (lp->status == 0);
> +
> +  /* Dead LWP's aren't expected to reported a pending sigstop.  */
> +  lp->signalled = 0;
> +
> +  /* Prevent trying to stop it.  */
> +  lp->stopped = 1;
> +}
> +
>   /* Wait for LP to stop.  Returns the wait status, or 0 if the LWP has
>      exited.  */
>   
> @@ -2207,9 +2228,8 @@ wait_lwp (struct lwp_info *lp)
>   
>   	      /* If this is the leader exiting, it means the whole
>   		 process is gone.  Store the status to report to the
> -		 core.  Store it in lp->waitstatus, because lp->status
> -		 would be ambiguous (W_EXITCODE(0,0) == 0).  */
> -	      lp->waitstatus = host_status_to_waitstatus (status);
> +		 core.  */
> +	      mark_lwp_dead (lp, status);
>   	      return 0;
>   	    }
>   
> @@ -3013,12 +3033,7 @@ linux_nat_filter_event (int lwpid, int status)
>         linux_nat_debug_printf ("LWP %ld exited (resumed=%d)",
>   			      lp->ptid.lwp (), lp->resumed);
>   
> -      /* Dead LWP's aren't expected to reported a pending sigstop.  */
> -      lp->signalled = 0;
> -
> -      /* Store the pending event in the waitstatus, because
> -	 W_EXITCODE(0,0) == 0.  */
> -      lp->waitstatus = host_status_to_waitstatus (status);
> +      mark_lwp_dead (lp, status);
>         return;
>       }
>   
> @@ -3368,6 +3383,7 @@ linux_nat_wait_1 (ptid_t ptid, struct target_waitstatus *ourstatus,
>       }
>   
>     gdb_assert (lp);
> +  gdb_assert (lp->stopped);
>   
>     status = lp->status;
>     lp->status = 0;
> 
> base-commit: 99eeecc8d276e5af745e48825d66efff693a7678