[1/2] setjmp: Use BSD sematic as default for setjmp

  POSIX relaxed the relation of setjmp/longjmp and the signal mask
save/restore, meaning that setjmp does not require to be routed to
_setjmp to be standard compliant.

This is done to avoid breakage of SIGABRT handlers, since to fully
make abort AS-safe, it is required to remove the recurisve lock
used to unblock SIGABRT prior raised the signal.

Also, it allows caller to actually use setjmp, since from
7011c2622fe3e10a29dbe74f06aaebd07710127d the symbol is unconditionally
routed to _setjmp.

Checked on x86_64-linux-gnu.
---
 manual/setjmp.texi                  | 14 ++++----------
 nptl/pthread_create.c               |  3 ++-
 setjmp/setjmp.h                     |  5 -----
 sysdeps/nptl/libc_start_call_main.h |  3 ++-
 4 files changed, 8 insertions(+), 17 deletions(-)
  

* Adhemerval Zanella via Libc-alpha:

> POSIX relaxed the relation of setjmp/longjmp and the signal mask
> save/restore, meaning that setjmp does not require to be routed to
> _setjmp to be standard compliant.
>
> This is done to avoid breakage of SIGABRT handlers, since to fully
> make abort AS-safe, it is required to remove the recurisve lock
> used to unblock SIGABRT prior raised the signal.
>
> Also, it allows caller to actually use setjmp, since from
> 7011c2622fe3e10a29dbe74f06aaebd07710127d the symbol is unconditionally
> routed to _setjmp.

Doesn't this have non-trivial performance impact?

Thanks,
Florian
  

On 01/08/23 05:35, Florian Weimer wrote:
> * Adhemerval Zanella via Libc-alpha:
> 
>> POSIX relaxed the relation of setjmp/longjmp and the signal mask
>> save/restore, meaning that setjmp does not require to be routed to
>> _setjmp to be standard compliant.
>>
>> This is done to avoid breakage of SIGABRT handlers, since to fully
>> make abort AS-safe, it is required to remove the recurisve lock
>> used to unblock SIGABRT prior raised the signal.
>>
>> Also, it allows caller to actually use setjmp, since from
>> 7011c2622fe3e10a29dbe74f06aaebd07710127d the symbol is unconditionally
>> routed to _setjmp.
> 
> Doesn't this have non-trivial performance impact?

Yes, it is two extra sigprocmask to get/set the signal mask.  This is not 
*strictly* required, but the SIGABRT on abort generates racy conditions on
process creation and.  This patch can be dropped, but it would mean that
to get expected semantic for abort handlers will need to use sigsetjmp (..., 1)
instead of setjmp.
  

* Adhemerval Zanella Netto:

> On 01/08/23 05:35, Florian Weimer wrote:
>> * Adhemerval Zanella via Libc-alpha:
>> 
>>> POSIX relaxed the relation of setjmp/longjmp and the signal mask
>>> save/restore, meaning that setjmp does not require to be routed to
>>> _setjmp to be standard compliant.
>>>
>>> This is done to avoid breakage of SIGABRT handlers, since to fully
>>> make abort AS-safe, it is required to remove the recurisve lock
>>> used to unblock SIGABRT prior raised the signal.
>>>
>>> Also, it allows caller to actually use setjmp, since from
>>> 7011c2622fe3e10a29dbe74f06aaebd07710127d the symbol is unconditionally
>>> routed to _setjmp.
>> 
>> Doesn't this have non-trivial performance impact?
>
> Yes, it is two extra sigprocmask to get/set the signal mask.  This is
> not *strictly* required, but the SIGABRT on abort generates racy
> conditions on process creation and.  This patch can be dropped, but it
> would mean that to get expected semantic for abort handlers will need
> to use sigsetjmp (..., 1) instead of setjmp.

Sorry, I don't understand?  With the current locking, this change should
really not be required because the user SIGABRT handler does not run
with the signal mask changed.

Thanks,
Florian
  

On 02/08/23 04:59, Florian Weimer wrote:
> * Adhemerval Zanella Netto:
> 
>> On 01/08/23 05:35, Florian Weimer wrote:
>>> * Adhemerval Zanella via Libc-alpha:
>>>
>>>> POSIX relaxed the relation of setjmp/longjmp and the signal mask
>>>> save/restore, meaning that setjmp does not require to be routed to
>>>> _setjmp to be standard compliant.
>>>>
>>>> This is done to avoid breakage of SIGABRT handlers, since to fully
>>>> make abort AS-safe, it is required to remove the recurisve lock
>>>> used to unblock SIGABRT prior raised the signal.
>>>>
>>>> Also, it allows caller to actually use setjmp, since from
>>>> 7011c2622fe3e10a29dbe74f06aaebd07710127d the symbol is unconditionally
>>>> routed to _setjmp.
>>>
>>> Doesn't this have non-trivial performance impact?
>>
>> Yes, it is two extra sigprocmask to get/set the signal mask.  This is
>> not *strictly* required, but the SIGABRT on abort generates racy
>> conditions on process creation and.  This patch can be dropped, but it
>> would mean that to get expected semantic for abort handlers will need
>> to use sigsetjmp (..., 1) instead of setjmp.
> 
> Sorry, I don't understand?  With the current locking, this change should
> really not be required because the user SIGABRT handler does not run
> with the signal mask changed.
>

This change is only required to keep the same semantic of setjmp/longjmp
regarding SIGABRT handler, where current code keeps subsequent SIGABRT 
installed with default flags to not keep the signal masked.  Otherwise,
users that callers that handle SIGABRT will need to either a different
sigaction mask that do no change the blocked signals while handling
the signal, call sigprocmask after SIGABRT returns from longjmp, or
use sigsetjmp.
  

* Adhemerval Zanella Netto:

> On 02/08/23 04:59, Florian Weimer wrote:
>> * Adhemerval Zanella Netto:
>> 
>>> On 01/08/23 05:35, Florian Weimer wrote:
>>>> * Adhemerval Zanella via Libc-alpha:
>>>>
>>>>> POSIX relaxed the relation of setjmp/longjmp and the signal mask
>>>>> save/restore, meaning that setjmp does not require to be routed to
>>>>> _setjmp to be standard compliant.
>>>>>
>>>>> This is done to avoid breakage of SIGABRT handlers, since to fully
>>>>> make abort AS-safe, it is required to remove the recurisve lock
>>>>> used to unblock SIGABRT prior raised the signal.
>>>>>
>>>>> Also, it allows caller to actually use setjmp, since from
>>>>> 7011c2622fe3e10a29dbe74f06aaebd07710127d the symbol is unconditionally
>>>>> routed to _setjmp.
>>>>
>>>> Doesn't this have non-trivial performance impact?
>>>
>>> Yes, it is two extra sigprocmask to get/set the signal mask.  This is
>>> not *strictly* required, but the SIGABRT on abort generates racy
>>> conditions on process creation and.  This patch can be dropped, but it
>>> would mean that to get expected semantic for abort handlers will need
>>> to use sigsetjmp (..., 1) instead of setjmp.
>> 
>> Sorry, I don't understand?  With the current locking, this change should
>> really not be required because the user SIGABRT handler does not run
>> with the signal mask changed.
>>
>
> This change is only required to keep the same semantic of setjmp/longjmp
> regarding SIGABRT handler, where current code keeps subsequent SIGABRT 
> installed with default flags to not keep the signal masked.  Otherwise,
> users that callers that handle SIGABRT will need to either a different
> sigaction mask that do no change the blocked signals while handling
> the signal, call sigprocmask after SIGABRT returns from longjmp, or
> use sigsetjmp.

Sorry, I still don't see it.  The new code switches the handler to
SIG_DFL atomically and blocks further sigaction calls.  This extends to
subprocesses because creating them is inhibited, too.  I think this
means that the difference in signal handler masking is not observable.

Thanks,
Florian
  

On 02/08/23 09:42, Florian Weimer wrote:
> * Adhemerval Zanella Netto:
> 
>> On 02/08/23 04:59, Florian Weimer wrote:
>>> * Adhemerval Zanella Netto:
>>>
>>>> On 01/08/23 05:35, Florian Weimer wrote:
>>>>> * Adhemerval Zanella via Libc-alpha:
>>>>>
>>>>>> POSIX relaxed the relation of setjmp/longjmp and the signal mask
>>>>>> save/restore, meaning that setjmp does not require to be routed to
>>>>>> _setjmp to be standard compliant.
>>>>>>
>>>>>> This is done to avoid breakage of SIGABRT handlers, since to fully
>>>>>> make abort AS-safe, it is required to remove the recurisve lock
>>>>>> used to unblock SIGABRT prior raised the signal.
>>>>>>
>>>>>> Also, it allows caller to actually use setjmp, since from
>>>>>> 7011c2622fe3e10a29dbe74f06aaebd07710127d the symbol is unconditionally
>>>>>> routed to _setjmp.
>>>>>
>>>>> Doesn't this have non-trivial performance impact?
>>>>
>>>> Yes, it is two extra sigprocmask to get/set the signal mask.  This is
>>>> not *strictly* required, but the SIGABRT on abort generates racy
>>>> conditions on process creation and.  This patch can be dropped, but it
>>>> would mean that to get expected semantic for abort handlers will need
>>>> to use sigsetjmp (..., 1) instead of setjmp.
>>>
>>> Sorry, I don't understand?  With the current locking, this change should
>>> really not be required because the user SIGABRT handler does not run
>>> with the signal mask changed.
>>>
>>
>> This change is only required to keep the same semantic of setjmp/longjmp
>> regarding SIGABRT handler, where current code keeps subsequent SIGABRT 
>> installed with default flags to not keep the signal masked.  Otherwise,
>> users that callers that handle SIGABRT will need to either a different
>> sigaction mask that do no change the blocked signals while handling
>> the signal, call sigprocmask after SIGABRT returns from longjmp, or
>> use sigsetjmp.
> 
> Sorry, I still don't see it.  The new code switches the handler to
> SIG_DFL atomically and blocks further sigaction calls.  This extends to
> subprocesses because creating them is inhibited, too.  I think this
> means that the difference in signal handler masking is not observable.

But this change it no to handle if raise returns, but rather if you have
a SIGABRT handler that does not (like the fortify tests) installed with
default flags.  In this case, the kernel will add SIGABRT on the masked
signal, longjmp will return to setjmp with the SIGBRT handler set mask,
and the next SIGABRT won't trigger the handler
  

* Adhemerval Zanella Netto:

> On 02/08/23 09:42, Florian Weimer wrote:
>> * Adhemerval Zanella Netto:
>> 
>>> On 02/08/23 04:59, Florian Weimer wrote:
>>>> * Adhemerval Zanella Netto:
>>>>
>>>>> On 01/08/23 05:35, Florian Weimer wrote:
>>>>>> * Adhemerval Zanella via Libc-alpha:
>>>>>>
>>>>>>> POSIX relaxed the relation of setjmp/longjmp and the signal mask
>>>>>>> save/restore, meaning that setjmp does not require to be routed to
>>>>>>> _setjmp to be standard compliant.
>>>>>>>
>>>>>>> This is done to avoid breakage of SIGABRT handlers, since to fully
>>>>>>> make abort AS-safe, it is required to remove the recurisve lock
>>>>>>> used to unblock SIGABRT prior raised the signal.
>>>>>>>
>>>>>>> Also, it allows caller to actually use setjmp, since from
>>>>>>> 7011c2622fe3e10a29dbe74f06aaebd07710127d the symbol is unconditionally
>>>>>>> routed to _setjmp.
>>>>>>
>>>>>> Doesn't this have non-trivial performance impact?
>>>>>
>>>>> Yes, it is two extra sigprocmask to get/set the signal mask.  This is
>>>>> not *strictly* required, but the SIGABRT on abort generates racy
>>>>> conditions on process creation and.  This patch can be dropped, but it
>>>>> would mean that to get expected semantic for abort handlers will need
>>>>> to use sigsetjmp (..., 1) instead of setjmp.
>>>>
>>>> Sorry, I don't understand?  With the current locking, this change should
>>>> really not be required because the user SIGABRT handler does not run
>>>> with the signal mask changed.
>>>>
>>>
>>> This change is only required to keep the same semantic of setjmp/longjmp
>>> regarding SIGABRT handler, where current code keeps subsequent SIGABRT 
>>> installed with default flags to not keep the signal masked.  Otherwise,
>>> users that callers that handle SIGABRT will need to either a different
>>> sigaction mask that do no change the blocked signals while handling
>>> the signal, call sigprocmask after SIGABRT returns from longjmp, or
>>> use sigsetjmp.
>> 
>> Sorry, I still don't see it.  The new code switches the handler to
>> SIG_DFL atomically and blocks further sigaction calls.  This extends to
>> subprocesses because creating them is inhibited, too.  I think this
>> means that the difference in signal handler masking is not observable.
>
> But this change it no to handle if raise returns, but rather if you have
> a SIGABRT handler that does not (like the fortify tests) installed with
> default flags.  In this case, the kernel will add SIGABRT on the masked
> signal, longjmp will return to setjmp with the SIGBRT handler set mask,
> and the next SIGABRT won't trigger the handler

Ahh, you mean because use removed the signal unblocking from abort?

If the signal is blocked, it is not delivered before it is unblocked.
This means that the handler will not observe it blocked.

But POSIX says this:

| The abort() function shall override blocking or ignoring the SIGABRT
| signal.

It also says:

| The SIGABRT signal shall be sent to the calling process as if by means
| of raise() with the argument SIGABRT.

Strictly speaking, it is impossible to comply with both requirements,
but I think the handler is expected to run even if SIGABRT is blocked.
As far as I understand it, the new code terminates the process in this
case, without ever running the handler.

Thanks,
Florian
  

On 02/08/23 10:17, Florian Weimer wrote:
> * Adhemerval Zanella Netto:
> 
>> On 02/08/23 09:42, Florian Weimer wrote:
>>> * Adhemerval Zanella Netto:
>>>
>>>> On 02/08/23 04:59, Florian Weimer wrote:
>>>>> * Adhemerval Zanella Netto:
>>>>>
>>>>>> On 01/08/23 05:35, Florian Weimer wrote:
>>>>>>> * Adhemerval Zanella via Libc-alpha:
>>>>>>>
>>>>>>>> POSIX relaxed the relation of setjmp/longjmp and the signal mask
>>>>>>>> save/restore, meaning that setjmp does not require to be routed to
>>>>>>>> _setjmp to be standard compliant.
>>>>>>>>
>>>>>>>> This is done to avoid breakage of SIGABRT handlers, since to fully
>>>>>>>> make abort AS-safe, it is required to remove the recurisve lock
>>>>>>>> used to unblock SIGABRT prior raised the signal.
>>>>>>>>
>>>>>>>> Also, it allows caller to actually use setjmp, since from
>>>>>>>> 7011c2622fe3e10a29dbe74f06aaebd07710127d the symbol is unconditionally
>>>>>>>> routed to _setjmp.
>>>>>>>
>>>>>>> Doesn't this have non-trivial performance impact?
>>>>>>
>>>>>> Yes, it is two extra sigprocmask to get/set the signal mask.  This is
>>>>>> not *strictly* required, but the SIGABRT on abort generates racy
>>>>>> conditions on process creation and.  This patch can be dropped, but it
>>>>>> would mean that to get expected semantic for abort handlers will need
>>>>>> to use sigsetjmp (..., 1) instead of setjmp.
>>>>>
>>>>> Sorry, I don't understand?  With the current locking, this change should
>>>>> really not be required because the user SIGABRT handler does not run
>>>>> with the signal mask changed.
>>>>>
>>>>
>>>> This change is only required to keep the same semantic of setjmp/longjmp
>>>> regarding SIGABRT handler, where current code keeps subsequent SIGABRT 
>>>> installed with default flags to not keep the signal masked.  Otherwise,
>>>> users that callers that handle SIGABRT will need to either a different
>>>> sigaction mask that do no change the blocked signals while handling
>>>> the signal, call sigprocmask after SIGABRT returns from longjmp, or
>>>> use sigsetjmp.
>>>
>>> Sorry, I still don't see it.  The new code switches the handler to
>>> SIG_DFL atomically and blocks further sigaction calls.  This extends to
>>> subprocesses because creating them is inhibited, too.  I think this
>>> means that the difference in signal handler masking is not observable.
>>
>> But this change it no to handle if raise returns, but rather if you have
>> a SIGABRT handler that does not (like the fortify tests) installed with
>> default flags.  In this case, the kernel will add SIGABRT on the masked
>> signal, longjmp will return to setjmp with the SIGBRT handler set mask,
>> and the next SIGABRT won't trigger the handler
> 
> Ahh, you mean because use removed the signal unblocking from abort?
> 
> If the signal is blocked, it is not delivered before it is unblocked.
> This means that the handler will not observe it blocked.
> 
> But POSIX says this:
> 
> | The abort() function shall override blocking or ignoring the SIGABRT
> | signal.
> 
> It also says:
> 
> | The SIGABRT signal shall be sent to the calling process as if by means
> | of raise() with the argument SIGABRT.
> 
> Strictly speaking, it is impossible to comply with both requirements,
> but I think the handler is expected to run even if SIGABRT is blocked.
> As far as I understand it, the new code terminates the process in this
> case, without ever running the handler.

The later has been changed with a new clarification [1]:

  The SIGABRT signal shall be sent to the calling [CX]thread[/CX] as if by
  means of raise() with the argument SIGABRT. [CX]If this signal does not 
  terminate the process (for example, if the signal is caught and the handler 
  returns), abort() may change the disposition of SIGABRT to SIG_DFL and send 
  the signal (in the same way) again. If a second signal is sent and it does 
  not terminate the process, the behavior is unspecified, except that the 
  abort() call shall not return.[/CX]

[1] https://austingroupbugs.net/view.php?id=906#c5851
  

* Adhemerval Zanella Netto:

>> Ahh, you mean because use removed the signal unblocking from abort?
>> 
>> If the signal is blocked, it is not delivered before it is unblocked.
>> This means that the handler will not observe it blocked.
>> 
>> But POSIX says this:
>> 
>> | The abort() function shall override blocking or ignoring the SIGABRT
>> | signal.
>> 
>> It also says:
>> 
>> | The SIGABRT signal shall be sent to the calling process as if by means
>> | of raise() with the argument SIGABRT.
>> 
>> Strictly speaking, it is impossible to comply with both requirements,
>> but I think the handler is expected to run even if SIGABRT is blocked.
>> As far as I understand it, the new code terminates the process in this
>> case, without ever running the handler.
>
> The later has been changed with a new clarification [1]:
>
>   The SIGABRT signal shall be sent to the calling [CX]thread[/CX] as if by
>   means of raise() with the argument SIGABRT. [CX]If this signal does not 
>   terminate the process (for example, if the signal is caught and the handler 
>   returns), abort() may change the disposition of SIGABRT to SIG_DFL and send 
>   the signal (in the same way) again. If a second signal is sent and it does 
>   not terminate the process, the behavior is unspecified, except that the 
>   abort() call shall not return.[/CX]
>
> [1] https://austingroupbugs.net/view.php?id=906#c5851

Okay, I missed that change.  So removing the unblocking should be okay
after this specification change.  I still don't see how the removal of
unblocking changes the signal mask observed by the signal handler,
though.

Thanks,
Florian
  

On 02/08/23 11:43, Florian Weimer wrote:
> * Adhemerval Zanella Netto:
> 
>>> Ahh, you mean because use removed the signal unblocking from abort?
>>>
>>> If the signal is blocked, it is not delivered before it is unblocked.
>>> This means that the handler will not observe it blocked.
>>>
>>> But POSIX says this:
>>>
>>> | The abort() function shall override blocking or ignoring the SIGABRT
>>> | signal.
>>>
>>> It also says:
>>>
>>> | The SIGABRT signal shall be sent to the calling process as if by means
>>> | of raise() with the argument SIGABRT.
>>>
>>> Strictly speaking, it is impossible to comply with both requirements,
>>> but I think the handler is expected to run even if SIGABRT is blocked.
>>> As far as I understand it, the new code terminates the process in this
>>> case, without ever running the handler.
>>
>> The later has been changed with a new clarification [1]:
>>
>>   The SIGABRT signal shall be sent to the calling [CX]thread[/CX] as if by
>>   means of raise() with the argument SIGABRT. [CX]If this signal does not 
>>   terminate the process (for example, if the signal is caught and the handler 
>>   returns), abort() may change the disposition of SIGABRT to SIG_DFL and send 
>>   the signal (in the same way) again. If a second signal is sent and it does 
>>   not terminate the process, the behavior is unspecified, except that the 
>>   abort() call shall not return.[/CX]
>>
>> [1] https://austingroupbugs.net/view.php?id=906#c5851
> 
> Okay, I missed that change.  So removing the unblocking should be okay
> after this specification change.  I still don't see how the removal of
> unblocking changes the signal mask observed by the signal handler,
> though.

It is not by the signal handler, but rather after a SIGABRT handler issues
longjmp.  With default flags (0), SIGABRT will continue to be blocked:

$ cat test.c
#include <assert.h>
#include <setjmp.h>
#include <signal.h>
#include <stdio.h>
#include <stdlib.h>

static jmp_buf jb;

static void
sigabrt_handler (int sig)
{
  longjmp (jb, 1);
}

int main (int argc, char *argv[])
{
  sigset_t set;

  assert (sigprocmask (SIG_BLOCK, NULL, &set) == 0);
  printf ("SIGABRT blocked: %d\n", sigismember (&set, SIGABRT));

  struct sigaction sa = { .sa_handler = sigabrt_handler, .sa_flags = 0 };
  sigemptyset (&sa.sa_mask);
  assert (sigaction (SIGABRT, &sa, 0) == 0);

  if (setjmp (jb) == 0)
    abort ();

  printf ("first abort did not terminated\n");

  assert (sigprocmask (SIG_BLOCK, NULL, &set) == 0);
  printf ("SIGABRT blocked: %d\n", sigismember (&set, SIGABRT));

  if (setjmp (jb) == 0)
    abort ();

  printf ("second abort did not terminated\n");

  return 0;
}
$ gcc test.c -o test && ./test
SIGABRT blocked: 0
first abort did not terminated
SIGABRT blocked: 1
second abort did not terminated

By continuing to use _setjmp as 'setjmp' and not unblocking SIGABRT on abort
call, the process will be aborted regardless.  This is due the difference
historically between BSD and SysV regarding whether setjmp/longjmp should
save/restore the signal mask (POSIX current allows both semantics).