Re: [PATCH] stdlib: Fix data race in __run_exit_handlers

[Adhemerval Zanella <adhemerval.zanella@linaro.org>]

On 26/04/2021 16:27, Vitaly Buka via Libc-alpha wrote:
> Fixes https://sourceware.org/bugzilla/show_bug.cgi?id=27749
> 
> Keep __exit_funcs_lock almost all the time and unlock it only to execute
> callbacks. This fixed two issues.
> 
> 1. f->func.cxa was modified outside the lock with rare data race like:
> 	thread 0: __run_exit_handlers unlock __exit_funcs_lock
> 	thread 1: __internal_atexit locks __exit_funcs_lock
> 	thread 0: f->flavor = ef_free;
> 	thread 1: sees ef_free and use it as new
> 	thread 1: new->func.cxa.fn = (void (*) (void *, int)) func;
> 	thread 1: new->func.cxa.arg = arg;
> 	thread 1: new->flavor = ef_cxa;
> 	thread 0: cxafct = f->func.cxa.fn;  // it's wrong fn!
> 	thread 0: cxafct (f->func.cxa.arg, status);  // it's wrong arg!
> 	thread 0: goto restart;
> 	thread 0: call the same exit_function again as it's ef_cxa

Ok, the small window between fetching the function pointer and argument
from the list is triggering a race condition.

> 
> 2. Don't unlock in main while loop after *listp = cur->next. If *listp
>    is NULL and __exit_funcs_done is false another thread may fail in
>    __new_exitfn on assert (l != NULL):
> 	 thread 0: *listp = cur->next;  // It can be the last: *listp = NULL.
> 	 thread 0: __libc_lock_unlock
> 	 thread 1: __libc_lock_lock in __on_exit
> 	 thread 1: __new_exitfn
> 	 thread 1: if (__exit_funcs_done)  // false: thread 0 isn't there yet.
> 	 thread 1: l = *listp
> 	 thread 1: moves one and crashes on assert (l != NULL);

Yeah, this is tricky but it does look correct.  I guess the lock/unlock
during the loop was added to give a chance to concurrent 
__cxa_atexit / on_exit to have a chance to add a new callback, but it 
also only complicates things as you noted.  We might try to fix it on the
__new_exitfn (to avoid the assert), but I see the current approach of 
locking the list and only unlocking while running the callback is the
right approach.

The patch look ok in general, I added some comments below.  I have
adjusted the patch based on my comments [1], if you are ok with them
I can push it upstream.

[1] https://sourceware.org/git/?p=glibc.git;a=shortlog;h=refs/heads/azanella/bz27749-atexit-fix

> 
> The test needs multiple iterations to consistently fail without the fix.
> ---
>  stdlib/Makefile                |   4 +-
>  stdlib/exit.c                  |  28 ++++++---
>  stdlib/test-cxa_atexit-race2.c | 110 +++++++++++++++++++++++++++++++++
>  3 files changed, 131 insertions(+), 11 deletions(-)
>  create mode 100644 stdlib/test-cxa_atexit-race2.c
> 
> diff --git a/stdlib/Makefile b/stdlib/Makefile
> index b3b30ab73e..f5755a1654 100644
> --- a/stdlib/Makefile
> +++ b/stdlib/Makefile
> @@ -81,7 +81,8 @@ tests		:= tst-strtol tst-strtod testmb testrand testsort testdiv   \
>  		   tst-width-stdint tst-strfrom tst-strfrom-locale	    \
>  		   tst-getrandom tst-atexit tst-at_quick_exit 		    \
>  		   tst-cxa_atexit tst-on_exit test-atexit-race 		    \
> -		   test-at_quick_exit-race test-cxa_atexit-race             \
> +		   test-at_quick_exit-race test-cxa_atexit-race		    \
> +		   test-cxa_atexit-race2				    \
>  		   test-on_exit-race test-dlclose-exit-race 		    \
>  		   tst-makecontext-align test-bz22786 tst-strtod-nan-sign \
>  		   tst-swapcontext1 tst-setcontext4 tst-setcontext5 \
> @@ -100,6 +101,7 @@ endif
>  LDLIBS-test-atexit-race = $(shared-thread-library)
>  LDLIBS-test-at_quick_exit-race = $(shared-thread-library)
>  LDLIBS-test-cxa_atexit-race = $(shared-thread-library)
> +LDLIBS-test-cxa_atexit-race2 = $(shared-thread-library)
>  LDLIBS-test-on_exit-race = $(shared-thread-library)
>  LDLIBS-tst-canon-bz26341 = $(shared-thread-library)
>  
> diff --git a/stdlib/exit.c b/stdlib/exit.c
> index bed82733ad..f095b38ab3 100644
> --- a/stdlib/exit.c
> +++ b/stdlib/exit.c
> @@ -45,6 +45,8 @@ __run_exit_handlers (int status, struct exit_function_list **listp,
>      if (run_dtors)
>        __call_tls_dtors ();
>  
> +  __libc_lock_lock (__exit_funcs_lock);
> +
>    /* We do it this way to handle recursive calls to exit () made by
>       the functions registered with `atexit' and `on_exit'. We call
>       everyone on the list and use the status value in the last

Ok, it avoids the second race condition.

> @@ -53,8 +55,6 @@ __run_exit_handlers (int status, struct exit_function_list **listp,
>      {
>        struct exit_function_list *cur;
>  
> -      __libc_lock_lock (__exit_funcs_lock);
> -
>      restart:
>        cur = *listp;
>  

I think there is no need use the goto anymore, since there is no need
to unlock the lock within the loop (the goto can be just a continue).

> @@ -63,7 +63,6 @@ __run_exit_handlers (int status, struct exit_function_list **listp,
>  	  /* Exit processing complete.  We will not allow any more
>  	     atexit/on_exit registrations.  */
>  	  __exit_funcs_done = true;
> -	  __libc_lock_unlock (__exit_funcs_lock);
>  	  break;
>  	}
>  

Ok, there is no need to unlock on break anymore.

> @@ -72,44 +71,52 @@ __run_exit_handlers (int status, struct exit_function_list **listp,
>  	  struct exit_function *const f = &cur->fns[--cur->idx];
>  	  const uint64_t new_exitfn_called = __new_exitfn_called;
>  
> -	  /* Unlock the list while we call a foreign function.  */
> -	  __libc_lock_unlock (__exit_funcs_lock);
>  	  switch (f->flavor)
>  	    {
>  	      void (*atfct) (void);
>  	      void (*onfct) (int status, void *arg);
>  	      void (*cxafct) (void *arg, int status);
> +	      void *arg;
>  
>  	    case ef_free:
>  	    case ef_us:
>  	      break;
>  	    case ef_on:
>  	      onfct = f->func.on.fn;
> +	      arg = f->func.on.arg;
>  #ifdef PTR_DEMANGLE
>  	      PTR_DEMANGLE (onfct);
>  #endif
> -	      onfct (status, f->func.on.arg);
> +	      /* Unlock the list while we call a foreign function.  */
> +	      __libc_lock_unlock (__exit_funcs_lock);
> +	      onfct (status, arg);
> +	      __libc_lock_lock (__exit_funcs_lock);
>  	      break;
>  	    case ef_at:
>  	      atfct = f->func.at;

Ok.

>  #ifdef PTR_DEMANGLE
>  	      PTR_DEMANGLE (atfct);
>  #endif
> +	      /* Unlock the list while we call a foreign function.  */
> +	      __libc_lock_unlock (__exit_funcs_lock);
>  	      atfct ();
> +	      __libc_lock_lock (__exit_funcs_lock);
>  	      break;

Ok.

>  	    case ef_cxa:
>  	      /* To avoid dlclose/exit race calling cxafct twice (BZ 22180),
>  		 we must mark this function as ef_free.  */
>  	      f->flavor = ef_free;
>  	      cxafct = f->func.cxa.fn;
> +	      arg = f->func.cxa.arg;
>  #ifdef PTR_DEMANGLE
>  	      PTR_DEMANGLE (cxafct);
>  #endif
> -	      cxafct (f->func.cxa.arg, status);
> +	      /* Unlock the list while we call a foreign function.  */
> +	      __libc_lock_unlock (__exit_funcs_lock);
> +	      cxafct (arg, status);
> +	      __libc_lock_lock (__exit_funcs_lock);
>  	      break;
>  	    }
> -	  /* Re-lock again before looking at global state.  */
> -	  __libc_lock_lock (__exit_funcs_lock);
>  
>  	  if (__glibc_unlikely (new_exitfn_called != __new_exitfn_called))
>  	    /* The last exit function, or another thread, has registered

Ok.

> @@ -123,9 +130,10 @@ __run_exit_handlers (int status, struct exit_function_list **listp,
>  	   allocate element.  */
>  	free (cur);
>  
> -      __libc_lock_unlock (__exit_funcs_lock);

Just remove the extra newline below as well.

>      }
>  
> +  __libc_lock_unlock (__exit_funcs_lock);
> +
>    if (run_list_atexit)
>      RUN_HOOK (__libc_atexit, ());
>  

Ok.

> diff --git a/stdlib/test-cxa_atexit-race2.c b/stdlib/test-cxa_atexit-race2.c
> new file mode 100644
> index 0000000000..d8c3d418e7
> --- /dev/null
> +++ b/stdlib/test-cxa_atexit-race2.c
> @@ -0,0 +1,110 @@
> +/* Support file for atexit/exit, etc. race tests.

I think it would be good to add a reference to the bug report.

> +   Copyright (C) 2017-2021 Free Software Foundation, Inc.
> +   This file is part of the GNU C Library.
> +
> +   The GNU C Library is free software; you can redistribute it and/or
> +   modify it under the terms of the GNU Lesser General Public
> +   License as published by the Free Software Foundation; either
> +   version 2.1 of the License, or (at your option) any later version.
> +
> +   The GNU C Library is distributed in the hope that it will be useful,
> +   but WITHOUT ANY WARRANTY; without even the implied warranty of
> +   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
> +   Lesser General Public License for more details.
> +
> +   You should have received a copy of the GNU Lesser General Public
> +   License along with the GNU C Library; if not, see
> +   <https://www.gnu.org/licenses/>.  */
> +
> +/* This file must be run from within a directory called "stdlib".  */

I don't think this true.

> +
> +/* The atexit/exit, at_quick_exit/quick_exit, __cxa_atexit/exit, etc. exhibited
> +   data race while calling destructors.
> +
> +   This test registers destructors from the background thread, and checks that
> +   the same destructor is not called more than once.  */
> +
> +#include <stdatomic.h>
> +#include <stdio.h>
> +#include <stdlib.h>
> +#include <support/xthread.h>
> +#include <sys/wait.h>
> +#include <unistd.h>
> +
> +static atomic_int registered;
> +static atomic_int todo = 100000;
> +
> +static void
> +atexit_cb (void *arg)
> +{
> +  atomic_fetch_sub (&registered, 1);
> +  static void *prev;
> +  if (arg == prev)
> +    {
> +      printf ("%p\n", arg);
> +      abort ();

Use FAIL_EXIT1 here.

> +    }
> +  prev = arg;
> +
> +  while (atomic_load (&todo) > 0 && atomic_load (&registered) < 100)
> +    ;
> +}
> +
> +int __cxa_atexit (void (*func) (void *), void *arg, void *d);
> +
> +static void *
> +thread_func (void *arg)
> +{
> +  void *cb_arg = NULL;
> +  while (atomic_load (&todo) > 0)

Add a open bracket here.

> +    if (atomic_load (&registered) < 10000)
> +      {
> +        int n = 10;
> +        for (int i = 0; i < n; ++i)
> +          __cxa_atexit (&atexit_cb, ++cb_arg, 0);
> +        atomic_fetch_add (&registered, n);
> +        atomic_fetch_sub (&todo, n);
> +      }
> +  return 0;

Use NULL here.

> +}
> +
> +static void

I would add a _Noreturn here.

> +test_and_exit (void)
> +{
> +  pthread_attr_t attr;
> +
> +  xpthread_attr_init (&attr);
> +  xpthread_attr_setdetachstate (&attr, 1);
> +
> +  xpthread_create (&attr, thread_func, NULL);
> +  xpthread_attr_destroy (&attr);
> +  while (!atomic_load (&registered))

Check for 0 here (unless the return value is a bool the type check
should be explicit).

> +    ;
> +  exit (0);
> +}
> +
> +static int
> +do_test (void)
> +{
> +  for (int i = 0; i < 20; ++i)
> +    {
> +      for (int i = 0; i < 10; ++i)
> +        if (fork () == 0)

Use xfork.

> +          test_and_exit ();
> +
> +      int status;
> +      while (wait (&status) > 0)
> +        {
> +          if (!WIFEXITED (status))

I prefer if we limit the number of wait call to check for invalid
return codes:

  for (int i = 0; i < 10; ++i)
    {
      int status;
      xwaitpid (0, &status, 0);
      if (!WIFEXITED (status))
        FAIL_EXIT1 ("Failed iterations %d", i);
      TEST_COMPARE (WEXITSTATUS (status), 0);
    }

> +            {
> +              printf ("Failed interation %d\n", i);
> +              abort ();

Use FAIL_EXIT1 here.

> +            }
> +        }
> +    }
> +
> +  exit (0);

There is no need to add an exit here.

> +}
> +
> +#define TEST_FUNCTION do_test
> +#include <support/test-driver.c>
>