[v2,2/9] nptl: Deallocate the thread stack on setup failure (BZ #19511)

  To setup either the thread scheduling parameters or affinity,
pthread_create enforce synchronization on created thread wait until
its parent either unlock the 'struct pthread' or send a cancellation
signal if a failure occurs.

However, cancelling the thread does not deallocate the newly created
stack since cancellation expects that a pthread_join to deallocate any
allocated thread resouces (threads stack or TLS).

I used a different strategy than the one proposed on BZ#19511
comment #4.  Making the parent responsible for the cleanup requires
additional synchronization similar to what pthread_join does.  Instead,
this patch reassigns an unused 'struct pthread' member,
parent_cancelhandling_unsed, to indicate whether the setup has failed
and set the thread itself to deallocate the allocated resouces (similar
to what detached mode does).

This strategy also simplifies by not using thread cancellation and
thus not running libgcc_so load in the signal handler (which is
avoided in thread cancellation since 'pthread_cancel' is the one
responsible to dlopen libgcc_s).

Checked on x86_64-linux-gnu and aarch64-linux-gnu.
---
 nptl/allocatestack.c  |  1 +
 nptl/descr.h          |  5 ++--
 nptl/pthread_create.c | 70 ++++++++++++++++++++-----------------------
 3 files changed, 36 insertions(+), 40 deletions(-)
  

* Adhemerval Zanella via Libc-alpha:

> @@ -424,17 +409,22 @@ start_thread (void *arg)
>  	 have ownership (see CONCURRENCY NOTES above).  */
>        if (__glibc_unlikely (pd->stopped_start))
>  	{
>  	  /* Get the lock the parent locked to force synchronization.  */
>  	  lll_lock (pd->lock, LLL_PRIVATE);
>  
>  	  /* We have ownership of PD now.  */
> +	  if (pd->setup_failed == 1)
> +	    {
> +	      /* In the case of a failed setup, the created thread will be
> +		 responsible to free up the allocated resources.  The
> +		 detached mode ensure it won't be joined and it will trigger
> +		 the required cleanup.  */
> +	      pd->joinid = pd;
> +	      __do_cancel ();
> +	    }
>  
>  	  /* And give it up right away.  */
>  	  lll_unlock (pd->lock, LLL_PRIVATE);
>  	}

I don't think this is quite right.  The failed thread should not linger
around after pthread_create has reported failure to the caller.  It's
better than what we had before, so maybe we should use this version and
clean it up further later.

At the point of the __do_cancel call, unwinding may not yet initialized,
so the implied dlopen happens on a potentially very small stack.  It
should be possible to replace it with a goto to the cleanup code.  Or
maybe put the cleanup code into a separate function and call it here and
on the regular cleanup path.

Apart from that, it looks fine to me.

Thanks,
Florian
  

On 01/06/2021 05:32, Florian Weimer wrote:
> * Adhemerval Zanella via Libc-alpha:
> 
>> @@ -424,17 +409,22 @@ start_thread (void *arg)
>>  	 have ownership (see CONCURRENCY NOTES above).  */
>>        if (__glibc_unlikely (pd->stopped_start))
>>  	{
>>  	  /* Get the lock the parent locked to force synchronization.  */
>>  	  lll_lock (pd->lock, LLL_PRIVATE);
>>  
>>  	  /* We have ownership of PD now.  */
>> +	  if (pd->setup_failed == 1)
>> +	    {
>> +	      /* In the case of a failed setup, the created thread will be
>> +		 responsible to free up the allocated resources.  The
>> +		 detached mode ensure it won't be joined and it will trigger
>> +		 the required cleanup.  */
>> +	      pd->joinid = pd;
>> +	      __do_cancel ();
>> +	    }
>>  
>>  	  /* And give it up right away.  */
>>  	  lll_unlock (pd->lock, LLL_PRIVATE);
>>  	}
> 
> I don't think this is quite right.  The failed thread should not linger
> around after pthread_create has reported failure to the caller.  It's
> better than what we had before, so maybe we should use this version and
> clean it up further later.

Agreed, but I think this is really an improvement over current strategy
as I commented earlier [1].  I think to proper fix would move the
lock prior any initialization so we don't need to run any unwinding
operations, skip directly to INTERNAL_SYSCALL_CALL (exit, 0), and let the
parent join the thread and do the resource deallocation.

> 
> At the point of the __do_cancel call, unwinding may not yet initialized,
> so the implied dlopen happens on a potentially very small stack.  It
> should be possible to replace it with a goto to the cleanup code.  Or
> maybe put the cleanup code into a separate function and call it here and
> on the regular cleanup path.
> 
> Apart from that, it looks fine to me.

The __do_cancel is not done really in asynchronous mode, so it will use
the allocated thread stack for the dlopen.  But I think with a better
strategy to move the 'struct thread' lock earlier than any initialization
will avoid such issue.

> 
> Thanks,
> Florian
> 

[1] https://sourceware.org/pipermail/libc-alpha/2021-May/126879.html
  

On 01/06/2021 10:08, Adhemerval Zanella wrote:
> 
> 
> On 01/06/2021 05:32, Florian Weimer wrote:
>> * Adhemerval Zanella via Libc-alpha:
>>
>>> @@ -424,17 +409,22 @@ start_thread (void *arg)
>>>  	 have ownership (see CONCURRENCY NOTES above).  */
>>>        if (__glibc_unlikely (pd->stopped_start))
>>>  	{
>>>  	  /* Get the lock the parent locked to force synchronization.  */
>>>  	  lll_lock (pd->lock, LLL_PRIVATE);
>>>  
>>>  	  /* We have ownership of PD now.  */
>>> +	  if (pd->setup_failed == 1)
>>> +	    {
>>> +	      /* In the case of a failed setup, the created thread will be
>>> +		 responsible to free up the allocated resources.  The
>>> +		 detached mode ensure it won't be joined and it will trigger
>>> +		 the required cleanup.  */
>>> +	      pd->joinid = pd;
>>> +	      __do_cancel ();
>>> +	    }
>>>  
>>>  	  /* And give it up right away.  */
>>>  	  lll_unlock (pd->lock, LLL_PRIVATE);
>>>  	}
>>
>> I don't think this is quite right.  The failed thread should not linger
>> around after pthread_create has reported failure to the caller.  It's
>> better than what we had before, so maybe we should use this version and
>> clean it up further later.
> 
> Agreed, but I think this is really an improvement over current strategy
> as I commented earlier [1].  I think to proper fix would move the
> lock prior any initialization so we don't need to run any unwinding
> operations, skip directly to INTERNAL_SYSCALL_CALL (exit, 0), and let the
> parent join the thread and do the resource deallocation.
> 
>>
>> At the point of the __do_cancel call, unwinding may not yet initialized,
>> so the implied dlopen happens on a potentially very small stack.  It
>> should be possible to replace it with a goto to the cleanup code.  Or
>> maybe put the cleanup code into a separate function and call it here and
>> on the regular cleanup path.
>>
>> Apart from that, it looks fine to me.
> 
> The __do_cancel is not done really in asynchronous mode, so it will use
> the allocated thread stack for the dlopen.  But I think with a better
> strategy to move the 'struct thread' lock earlier than any initialization
> will avoid such issue.
> 
>>
>> Thanks,
>> Florian
>>
> 
> [1] https://sourceware.org/pipermail/libc-alpha/2021-May/126879.html
> 

The patch below on top of this patch implements the synchronous thread
join in case of setup failure.  It shows no regressions on testcases,
and it is slight better because it avoid calling the __pthread_unwind for
such cases (and issuing the libgcc dlopen)

In fact since there is no registered user case that might required
calling either cancellation cleanup handler or thread destructors,
I think there is no need to actually issue a pthread_exit like
termination.

--

diff --git a/nptl/pthread_create.c b/nptl/pthread_create.c
index 018af30c85..7a4c742416 100644
--- a/nptl/pthread_create.c
+++ b/nptl/pthread_create.c
@@ -346,6 +346,23 @@ start_thread (void *arg)
 {
   struct pthread *pd = arg;
 
+  /* We are either in (a) or (b), and in either case we either own PD already
+     (2) or are about to own PD (1), and so our only restriction would be that
+     we can't free PD until we know we have ownership (see CONCURRENCY NOTES
+     above).  */
+  if (__glibc_unlikely (pd->stopped_start))
+    {
+      /* Get the lock the parent locked to force synchronization.  */
+      lll_lock (pd->lock, LLL_PRIVATE);
+
+      /* We have ownership of PD now.  */
+      if (pd->setup_failed == 1)
+	goto out;
+
+      /* And give it up right away.  */
+      lll_unlock (pd->lock, LLL_PRIVATE);
+     }
+
   /* Initialize resolver state pointer.  */
   __resp = &pd->res;
 
@@ -403,30 +420,6 @@ start_thread (void *arg)
       /* Store the new cleanup handler info.  */
       THREAD_SETMEM (pd, cleanup_jmp_buf, &unwind_buf);
 
-      /* We are either in (a) or (b), and in either case we either own
-         PD already (2) or are about to own PD (1), and so our only
-	 restriction would be that we can't free PD until we know we
-	 have ownership (see CONCURRENCY NOTES above).  */
-      if (__glibc_unlikely (pd->stopped_start))
-	{
-	  /* Get the lock the parent locked to force synchronization.  */
-	  lll_lock (pd->lock, LLL_PRIVATE);
-
-	  /* We have ownership of PD now.  */
-	  if (pd->setup_failed == 1)
-	    {
-	      /* In the case of a failed setup, the created thread will be
-		 responsible to free up the allocated resources.  The
-		 detached mode ensure it won't be joined and it will trigger
-		 the required cleanup.  */
-	      pd->joinid = pd;
-	      __do_cancel ();
-	    }
-
-	  /* And give it up right away.  */
-	  lll_unlock (pd->lock, LLL_PRIVATE);
-	}
-
       LIBC_PROBE (pthread_start, 3, (pthread_t) pd, pd->start_routine, pd->arg);
 
       /* Run the code the user provided.  */
@@ -556,6 +549,7 @@ start_thread (void *arg)
     /* Free the TCB.  */
     __nptl_free_tcb (pd);
 
+out:
   /* We cannot call '_exit' here.  '_exit' will terminate the process.
 
      The 'exit' implementation in the kernel will signal when the
@@ -802,6 +796,8 @@ __pthread_create_2_1 (pthread_t *newthread, const pthread_attr_t *attr,
      thread.  */
   __libc_signal_restore_set (&original_sigmask);
 
+  int setup_failed = 0;
+
   if (__glibc_unlikely (retval != 0))
     {
       if (thread_ran)
@@ -814,7 +810,7 @@ __pthread_create_2_1 (pthread_t *newthread, const pthread_attr_t *attr,
   	  /* State (a), we own PD. The thread blocked on this lock will
 	     finish due setup_failed being set.  The thread will be
 	     responsible to cleanup any allocated resource.  */
-	  pd->setup_failed = 1;
+	  setup_failed = pd->setup_failed = 1;
 	  lll_unlock (pd->lock, LLL_PRIVATE);
         }
       else
@@ -857,6 +853,9 @@ __pthread_create_2_1 (pthread_t *newthread, const pthread_attr_t *attr,
       THREAD_SETMEM (THREAD_SELF, header.multiple_threads, 1);
     }
 
+  if (setup_failed == 1)
+    __pthread_join (*newthread, NULL);
+
  out:
   if (destroy_default_attr)
     __pthread_attr_destroy (&default_attr.external);