Patchwork Tests for minimal signal handler functionality in MINSIGSTKSZ space.

login
register
mail settings
Submitter Zack Weinberg
Date Jan. 16, 2019, 2:06 p.m.
Message ID <CAKCAbMiaUYOLkRD36MEVJRuzbAyjC7DEYNz68ZuJdWDrq=YPRQ@mail.gmail.com>
Download mbox | patch
Permalink /patch/31073/
State Committed
Headers show

Comments

Zack Weinberg - Jan. 16, 2019, 2:06 p.m.
The patch I just pushed is attached to this message.  Thanks for the
quick reviews.  A couple of notes:

> I'm not sure if it's explicitly spelled out that the default disposition
> for SIGABRT is to core

This is indeed specified in POSIX, see the table of default signal actions in
http://pubs.opengroup.org/onlinepubs/9699919799/basedefs/signal.h.html .

> +/* C2011 7.4.1.1p5 specifies that only the following operations are
> +   guaranteed to be well-defined inside an asynchronous signal handler:
> ...
> +     * calling abort, _Exit, quick_exit, or signal
> +       * signal may only be called with its first argument equal to the
> +         number of the signal that caused the handler to be called

It occurred to me this morning that signal is implemented using
sigaction, and a 'struct sigaction' is 152 bytes on x86-64 (because of
the sigset_t).  Normally that's nothing to worry about, but when
there's only 2048 bytes of stack in total and the kernel has filled
more than half of that with ucontext data, it _could_ be a problem.
I left the test as is for now, but I suggest that if it does crash on
any architecture we should drop that part of tst-minsigstksz-1.c
without hesitation.

It also occurs to me that on some architectures MINSIGSTKSZ is less
than a page; on those architectures, the rounding done in xsigstack.c
means we aren't _really_ testing this stuff in MINSIGSTKSZ space.
Since overflow is much more of a concern than underflow, what do
people think of adjusting the code in xsigstack.c so that the area
actually passed to sigaltstack will not be rounded and will be right
up against the guard in the direction of overflow?  This would mean
xsigstack.c has to know which direction the stack grows, but I think
we already have internal macros for that, so it's not a huge problem.
I would do this as a follow-up patch if it's agreed to be a good idea.

zw
Florian Weimer - Jan. 16, 2019, 2:16 p.m.
* Zack Weinberg:

> It also occurs to me that on some architectures MINSIGSTKSZ is less
> than a page; on those architectures, the rounding done in xsigstack.c
> means we aren't _really_ testing this stuff in MINSIGSTKSZ space.
> Since overflow is much more of a concern than underflow, what do
> people think of adjusting the code in xsigstack.c so that the area
> actually passed to sigaltstack will not be rounded and will be right
> up against the guard in the direction of overflow?  This would mean
> xsigstack.c has to know which direction the stack grows, but I think
> we already have internal macros for that,

_STACK_GROWS_DOWN, _STACK_GROWS_UP, and NEED_SEPARATE_REGISTER_STACK on
ia64.  But ia64 uses stack sizes which are a multiple of the page size,
so you are lucky. 8-)

Thanks,
Florian
Carlos O'Donell - Jan. 16, 2019, 5:06 p.m.
On 1/16/19 9:06 AM, Zack Weinberg wrote:
> The patch I just pushed is attached to this message.  Thanks for the
> quick reviews.  A couple of notes:
> 
>> I'm not sure if it's explicitly spelled out that the default disposition
>> for SIGABRT is to core
> 
> This is indeed specified in POSIX, see the table of default signal actions in
> http://pubs.opengroup.org/onlinepubs/9699919799/basedefs/signal.h.html .
> 
>> +/* C2011 7.4.1.1p5 specifies that only the following operations are
>> +   guaranteed to be well-defined inside an asynchronous signal handler:
>> ...
>> +     * calling abort, _Exit, quick_exit, or signal
>> +       * signal may only be called with its first argument equal to the
>> +         number of the signal that caused the handler to be called
> 
> It occurred to me this morning that signal is implemented using
> sigaction, and a 'struct sigaction' is 152 bytes on x86-64 (because of
> the sigset_t).  Normally that's nothing to worry about, but when
> there's only 2048 bytes of stack in total and the kernel has filled
> more than half of that with ucontext data, it _could_ be a problem.
> I left the test as is for now, but I suggest that if it does crash on
> any architecture we should drop that part of tst-minsigstksz-1.c
> without hesitation.
> 
> It also occurs to me that on some architectures MINSIGSTKSZ is less
> than a page; on those architectures, the rounding done in xsigstack.c
> means we aren't _really_ testing this stuff in MINSIGSTKSZ space.
> Since overflow is much more of a concern than underflow, what do
> people think of adjusting the code in xsigstack.c so that the area
> actually passed to sigaltstack will not be rounded and will be right
> up against the guard in the direction of overflow?  This would mean
> xsigstack.c has to know which direction the stack grows, but I think
> we already have internal macros for that, so it's not a huge problem.
> I would do this as a follow-up patch if it's agreed to be a good idea.

I worried about this during the original review, but my conclusion was
that no sane user would do such an alignment to push their own stack
against the guard. At least I've never seen it done. This does mean that
the user has a variable amount of trailing stack space that depends on
the configured kernel page size, which is different from the userspace
configured static page size (but the same as the dynamically available
value from sysconf / AT_PAGESZ). If the kernel uses THP you also loose
because the page may be 2MiB and from userspace you don't know that.
So the test adjustment you propose is sensible, and I think would add
value to the test and catch out the currently silly values of 2048
(in most cases).

I expect to see some failures, so I think such an adjustment should
be made once 2.30 is open for development :-) Also keep in mind that
to be pedantic you need to know that the machine stack pointer
alignment requirement is not equal to the page alignment, maybe a note
there that MINSIGSTKSZ is such that it is a multiple of the stack
alignment?
Richard Henderson - Jan. 16, 2019, 10:44 p.m.
On 1/17/19 1:06 AM, Zack Weinberg wrote:
> It also occurs to me that on some architectures MINSIGSTKSZ is less
> than a page; on those architectures, the rounding done in xsigstack.c
> means we aren't _really_ testing this stuff in MINSIGSTKSZ space.
> Since overflow is much more of a concern than underflow, what do
> people think of adjusting the code in xsigstack.c so that the area
> actually passed to sigaltstack will not be rounded and will be right
> up against the guard in the direction of overflow?  This would mean
> xsigstack.c has to know which direction the stack grows, but I think
> we already have internal macros for that, so it's not a huge problem.
> I would do this as a follow-up patch if it's agreed to be a good idea.

Heh.  Should have read the whole thread first.
>From my other mail, you can clearly tell that I vote yes for this.  ;-)


r~
Richard Henderson - Jan. 16, 2019, 10:47 p.m.
On 1/17/19 4:06 AM, Carlos O'Donell wrote:
> I worried about this during the original review, but my conclusion was
> that no sane user would do such an alignment to push their own stack
> against the guard. At least I've never seen it done.

I have seen users use malloc to allocate stacks.  Which is basically the same
as putting the allocation next to the guard, except it corrupts the arena
instead.  :-P


r~
H.J. Lu - Jan. 17, 2019, 3:30 a.m.
On Wed, Jan 16, 2019 at 9:06 AM Carlos O'Donell <carlos@redhat.com> wrote:
>
> On 1/16/19 9:06 AM, Zack Weinberg wrote:
> > The patch I just pushed is attached to this message.  Thanks for the
> > quick reviews.  A couple of notes:
> >
> >> I'm not sure if it's explicitly spelled out that the default disposition
> >> for SIGABRT is to core
> >
> > This is indeed specified in POSIX, see the table of default signal actions in
> > http://pubs.opengroup.org/onlinepubs/9699919799/basedefs/signal.h.html .
> >
> >> +/* C2011 7.4.1.1p5 specifies that only the following operations are
> >> +   guaranteed to be well-defined inside an asynchronous signal handler:
> >> ...
> >> +     * calling abort, _Exit, quick_exit, or signal
> >> +       * signal may only be called with its first argument equal to the
> >> +         number of the signal that caused the handler to be called
> >
> > It occurred to me this morning that signal is implemented using
> > sigaction, and a 'struct sigaction' is 152 bytes on x86-64 (because of
> > the sigset_t).  Normally that's nothing to worry about, but when
> > there's only 2048 bytes of stack in total and the kernel has filled
> > more than half of that with ucontext data, it _could_ be a problem.
> > I left the test as is for now, but I suggest that if it does crash on
> > any architecture we should drop that part of tst-minsigstksz-1.c
> > without hesitation.
> >
> > It also occurs to me that on some architectures MINSIGSTKSZ is less
> > than a page; on those architectures, the rounding done in xsigstack.c
> > means we aren't _really_ testing this stuff in MINSIGSTKSZ space.
> > Since overflow is much more of a concern than underflow, what do
> > people think of adjusting the code in xsigstack.c so that the area
> > actually passed to sigaltstack will not be rounded and will be right
> > up against the guard in the direction of overflow?  This would mean
> > xsigstack.c has to know which direction the stack grows, but I think
> > we already have internal macros for that, so it's not a huge problem.
> > I would do this as a follow-up patch if it's agreed to be a good idea.
>
> I worried about this during the original review, but my conclusion was
> that no sane user would do such an alignment to push their own stack
> against the guard. At least I've never seen it done. This does mean that
> the user has a variable amount of trailing stack space that depends on
> the configured kernel page size, which is different from the userspace
> configured static page size (but the same as the dynamically available
> value from sysconf / AT_PAGESZ). If the kernel uses THP you also loose
> because the page may be 2MiB and from userspace you don't know that.
> So the test adjustment you propose is sensible, and I think would add
> value to the test and catch out the currently silly values of 2048
> (in most cases).
>
> I expect to see some failures, so I think such an adjustment should
> be made once 2.30 is open for development :-) Also keep in mind that
> to be pedantic you need to know that the machine stack pointer
> alignment requirement is not equal to the page alignment, maybe a note
> there that MINSIGSTKSZ is such that it is a multiple of the stack
> alignment?
>

The new tests failed on AVX512 machines:

Program received signal SIGUSR1, User defined signal 1.
__GI_raise (sig=sig@entry=10) at ../sysdeps/unix/sysv/linux/raise.c:50
50   return ret;
(gdb) c
Continuing.

Program received signal SIGSEGV, Segmentation fault.
_dl_runtime_resolve_xsavec () at ../sysdeps/x86_64/dl-trampoline.h:93
93 movq %rax, REGISTER_SAVE_RAX(%rsp)
(gdb) bt
#0  _dl_runtime_resolve_xsavec () at ../sysdeps/x86_64/dl-trampoline.h:93
#1  0x0040248d in handler (unused=<optimized out>) at tst-minsigstksz-4.c:44
#2  <signal handler called>
#3  __GI_raise (sig=sig@entry=10) at ../sysdeps/unix/sysv/linux/raise.c:50
#4  0x004024da in do_test () at tst-minsigstksz-4.c:59
#5  0x00402cd6 in support_test_main (argc=1, argv=0xffffcef8,
    config=config@entry=0xffffcdf0) at support_test_main.c:350
#6  0x00402348 in main (argc=<optimized out>, argv=<optimized out>)
    at ../support/test-driver.c:168
(gdb)

AVX512 needs 2560 bytes to save processor state.
Zack Weinberg - Jan. 18, 2019, 2:52 a.m.
On Wed, Jan 16, 2019 at 10:31 PM H.J. Lu <hjl.tools@gmail.com> wrote:
>
> The new tests failed on AVX512 machines:
>
> Program received signal SIGUSR1, User defined signal 1.
> __GI_raise (sig=sig@entry=10) at ../sysdeps/unix/sysv/linux/raise.c:50
> 50   return ret;
> (gdb) c
> Continuing.
>
> Program received signal SIGSEGV, Segmentation fault.
> _dl_runtime_resolve_xsavec () at ../sysdeps/x86_64/dl-trampoline.h:93
> 93 movq %rax, REGISTER_SAVE_RAX(%rsp)
> (gdb) bt
> #0  _dl_runtime_resolve_xsavec () at ../sysdeps/x86_64/dl-trampoline.h:93
> #1  0x0040248d in handler (unused=<optimized out>) at tst-minsigstksz-4.c:44
> #2  <signal handler called>
> #3  __GI_raise (sig=sig@entry=10) at ../sysdeps/unix/sysv/linux/raise.c:50
> #4  0x004024da in do_test () at tst-minsigstksz-4.c:59
> #5  0x00402cd6 in support_test_main (argc=1, argv=0xffffcef8,
>     config=config@entry=0xffffcdf0) at support_test_main.c:350
> #6  0x00402348 in main (argc=<optimized out>, argv=<optimized out>)
>     at ../support/test-driver.c:168
> (gdb)
>
> AVX512 needs 2560 bytes to save processor state.

Well, this is the problem that we knew existed and can't fix quickly.
If I'm reading http://pubs.opengroup.org/onlinepubs/9699919799/functions/V2_chap02.html
correctly, we can't introduce new sysconf constants unilaterally
(there is no license to define system-specific _SC_* symbols).

I wonder if this test passes if you run it with LD_BIND_NOW=t in the
environment.  Forcing -z now for these tests might be the best we can
do for 2.29.

zw

Patch

From fbbc9a4e347dabb2d1662744e6a2e83b569ea3a4 Mon Sep 17 00:00:00 2001
From: Zack Weinberg <zackw@panix.com>
Date: Tue, 15 Jan 2019 14:58:15 -0500
Subject: [PATCH] Tests for minimal signal handler functionality in MINSIGSTKSZ
 space.

There is general agreement that the very short list of things that ISO
C says you can do in an async signal handler should all work when the
handler is running on an alternate signal stack with only MINSIGSTKSZ
space.  This patch adds tests to make sure those things do work.

To facilitate this, there is a new set of test support routines for
setting up alternate signal stacks; see support/xsignal.h for the API.

         * support/xsignal.h (xalloc_sigstack, xfree_sigstack)
         (xget_sigstack_location): New test support functions.
         * support/xsigstack.c: New file, implementing them.
         * support/tst-xsigstack.c: New test for them.
         * support/Makefile: Update.

         * signal/tst-minsigstksz-1.c
         * signal/tst-minsigstksz-2.c
         * signal/tst-minsigstksz-3.c
         * signal/tst-minsigstksz-3a.c
         * signal/tst-minsigstksz-4.c: New tests.
         * signal/Makefile: Run them.
---
 ChangeLog                   |  15 +++++
 signal/Makefile             |   2 +
 signal/tst-minsigstksz-1.c  | 131 ++++++++++++++++++++++++++++++++++++
 signal/tst-minsigstksz-2.c  |  66 ++++++++++++++++++
 signal/tst-minsigstksz-3.c  |  64 ++++++++++++++++++
 signal/tst-minsigstksz-3a.c |  69 +++++++++++++++++++
 signal/tst-minsigstksz-4.c  |  65 ++++++++++++++++++
 support/Makefile            |   2 +
 support/tst-xsigstack.c     |  64 ++++++++++++++++++
 support/xsignal.h           |  17 +++++
 support/xsigstack.c         | 107 +++++++++++++++++++++++++++++
 11 files changed, 602 insertions(+)
 create mode 100644 signal/tst-minsigstksz-1.c
 create mode 100644 signal/tst-minsigstksz-2.c
 create mode 100644 signal/tst-minsigstksz-3.c
 create mode 100644 signal/tst-minsigstksz-3a.c
 create mode 100644 signal/tst-minsigstksz-4.c
 create mode 100644 support/tst-xsigstack.c
 create mode 100644 support/xsigstack.c

diff --git a/ChangeLog b/ChangeLog
index 36bb3ed885..7cc7916a17 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,3 +1,18 @@ 
+2019-01-16  Zack Weinberg  <zackw@panix.com>
+
+	 * support/xsignal.h (xalloc_sigstack, xfree_sigstack)
+	 (xget_sigstack_location): New test support functions.
+	 * support/xsigstack.c: New file, implementing them.
+	 * support/tst-xsigstack.c: New test for them.
+	 * support/Makefile: Update.
+
+	 * signal/tst-minsigstksz-1.c
+	 * signal/tst-minsigstksz-2.c
+	 * signal/tst-minsigstksz-3.c
+	 * signal/tst-minsigstksz-3a.c
+	 * signal/tst-minsigstksz-4.c: New tests.
+	 * signal/Makefile: Run them.
+
 2019-01-16  Siddhesh Poyarekar  <siddhesh@sourceware.org>
 
 	* po/libc.pot: Regenerate.
diff --git a/signal/Makefile b/signal/Makefile
index 647ce242e6..9597287bca 100644
--- a/signal/Makefile
+++ b/signal/Makefile
@@ -47,6 +47,8 @@  routines	:= signal raise killpg \
 
 tests		:= tst-signal tst-sigset tst-sigsimple tst-raise tst-sigset2 \
 		   tst-sigwait-eintr tst-sigaction \
+		   tst-minsigstksz-1 tst-minsigstksz-2 tst-minsigstksz-3 \
+		   tst-minsigstksz-3a tst-minsigstksz-4 \
 
 include ../Rules
 
diff --git a/signal/tst-minsigstksz-1.c b/signal/tst-minsigstksz-1.c
new file mode 100644
index 0000000000..00344d5fbf
--- /dev/null
+++ b/signal/tst-minsigstksz-1.c
@@ -0,0 +1,131 @@ 
+/* Tests of signal delivery on an alternate stack (nonlethal).
+   Copyright (C) 2019 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
+   <http://www.gnu.org/licenses/>.  */
+
+#include <support/xsignal.h>
+#include <support/support.h>
+#include <support/check.h>
+
+/* C2011 7.4.1.1p5 specifies that only the following operations are
+   guaranteed to be well-defined inside an asynchronous signal handler:
+     * any operation on a lock-free atomic object
+     * assigning a value to an object declared as volatile sig_atomic_t
+     * calling abort, _Exit, quick_exit, or signal
+       * signal may only be called with its first argument equal to the
+         number of the signal that caused the handler to be called
+
+   We use this list as a guideline for the set of operations that ought
+   also to be safe in a _synchronous_ signal delivered on an alternate
+   signal stack with only MINSIGSTKSZ bytes of space.
+
+   This test program tests all of the above operations that do not,
+   one way or another, cause the program to be terminated.  */
+
+/* We do not try to test atomic operations exhaustively, only a simple
+   atomic counter increment.  This is only safe if atomic_[u]int is
+   unconditionally lock-free.  */
+#ifdef __STDC_NO_ATOMICS__
+# define TEST_ATOMIC_OPS 0
+#else
+# include <stdatomic.h>
+# if ATOMIC_INT_LOCK_FREE != 2
+#  define TEST_ATOMIC_OPS 0
+# else
+#  define TEST_ATOMIC_OPS 1
+# endif
+#endif
+
+static volatile sig_atomic_t signal_flag = 0;
+static volatile sig_atomic_t signal_err = 0;
+static void
+handler_set_flag (int unused)
+{
+  signal_flag = 1;
+}
+
+static void
+handler_set_flag_once (int sig)
+{
+  signal_flag = 1;
+  if (signal (sig, SIG_IGN) == SIG_ERR)
+    /* It is not safe to call FAIL_EXIT1 here.  Set another flag instead.  */
+    signal_err = 1;
+}
+
+#if TEST_ATOMIC_OPS
+static atomic_uint signal_count = 0;
+static void
+handler_count_up_1 (int unused)
+{
+  atomic_fetch_add (&signal_count, 1);
+}
+#endif
+
+int
+do_test (void)
+{
+  void *sstk = xalloc_sigstack (0);
+  struct sigaction sa;
+
+  /* Test 1: setting a volatile sig_atomic_t flag.  */
+  sa.sa_handler = handler_set_flag;
+  sa.sa_flags   = SA_RESTART | SA_ONSTACK;
+  sigfillset (&sa.sa_mask);
+  if (sigaction (SIGUSR1, &sa, 0))
+    FAIL_EXIT1 ("sigaction (SIGUSR1, handler_set_flag): %m\n");
+
+  TEST_VERIFY_EXIT (signal_flag == 0);
+  raise (SIGUSR1);
+  TEST_VERIFY_EXIT (signal_flag == 1);
+  signal_flag = 0;
+  raise (SIGUSR1);
+  TEST_VERIFY_EXIT (signal_flag == 1);
+  signal_flag = 0;
+
+  /* Test 1: setting a volatile sig_atomic_t flag and then ignoring
+     further delivery of the signal. */
+  sa.sa_handler = handler_set_flag_once;
+  if (sigaction (SIGUSR1, &sa, 0))
+    FAIL_EXIT1 ("sigaction (SIGUSR1, handler_set_flag_once): %m\n");
+
+  raise (SIGUSR1);
+  TEST_VERIFY_EXIT (signal_flag == 1);
+  /* Note: if signal_err is 1, a system call failed, but we can't
+     report the error code because errno is indeterminate.  */
+  TEST_VERIFY_EXIT (signal_err == 0);
+
+  signal_flag = 0;
+  raise (SIGUSR1);
+  TEST_VERIFY_EXIT (signal_flag == 0);
+  TEST_VERIFY_EXIT (signal_err == 0);
+
+#if TEST_ATOMIC_OPS
+  sa.sa_handler = handler_count_up_1;
+  if (sigaction (SIGUSR1, &sa, 0))
+    FAIL_EXIT1 ("sigaction (SIGUSR1, handler_count_up_1): %m\n");
+
+  raise (SIGUSR1);
+  TEST_VERIFY_EXIT (atomic_load (&signal_count) == 1);
+  raise (SIGUSR1);
+  TEST_VERIFY_EXIT (atomic_load (&signal_count) == 2);
+#endif
+
+  xfree_sigstack (sstk);
+  return 0;
+}
+
+#include <support/test-driver.c>
diff --git a/signal/tst-minsigstksz-2.c b/signal/tst-minsigstksz-2.c
new file mode 100644
index 0000000000..3368dde6b8
--- /dev/null
+++ b/signal/tst-minsigstksz-2.c
@@ -0,0 +1,66 @@ 
+/* Tests of signal delivery on an alternate stack (abort).
+   Copyright (C) 2019 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
+   <http://www.gnu.org/licenses/>.  */
+
+#include <support/xsignal.h>
+#include <support/support.h>
+#include <support/check.h>
+#include <stdlib.h>
+
+/* C2011 7.4.1.1p5 specifies that only the following operations are
+   guaranteed to be well-defined inside an asynchronous signal handler:
+     * any operation on a lock-free atomic object
+     * assigning a value to an object declared as volatile sig_atomic_t
+     * calling abort, _Exit, quick_exit, or signal
+       * signal may only be called with its first argument equal to the
+         number of the signal that caused the handler to be called
+
+   We use this list as a guideline for the set of operations that ought
+   also to be safe in a _synchronous_ signal delivered on an alternate
+   signal stack with only MINSIGSTKSZ bytes of space.
+
+   This test program tests calls to abort.  Note that it does _not_
+   install a handler for SIGABRT, because that signal would also be
+   delivered on the alternate stack and MINSIGSTKSZ does not provide
+   enough space for delivery of nested signals.  */
+
+static void
+handler (int unused)
+{
+  abort ();
+}
+
+int
+do_test (void)
+{
+  void *sstk = xalloc_sigstack (0);
+  struct sigaction sa;
+
+  sa.sa_handler = handler;
+  sa.sa_flags   = SA_RESTART | SA_ONSTACK;
+  sigfillset (&sa.sa_mask);
+  if (sigaction (SIGUSR1, &sa, 0))
+    FAIL_RET ("sigaction (SIGUSR1, handler): %m\n");
+
+  raise (SIGUSR1);
+
+  xfree_sigstack (sstk);
+  FAIL_RET ("test process was not terminated by abort in signal handler");
+}
+
+#define EXPECTED_SIGNAL SIGABRT
+#include <support/test-driver.c>
diff --git a/signal/tst-minsigstksz-3.c b/signal/tst-minsigstksz-3.c
new file mode 100644
index 0000000000..a8d9a6369c
--- /dev/null
+++ b/signal/tst-minsigstksz-3.c
@@ -0,0 +1,64 @@ 
+/* Tests of signal delivery on an alternate stack (_Exit).
+   Copyright (C) 2019 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
+   <http://www.gnu.org/licenses/>.  */
+
+#include <support/xsignal.h>
+#include <support/support.h>
+#include <support/check.h>
+#include <stdlib.h>
+
+/* C2011 7.4.1.1p5 specifies that only the following operations are
+   guaranteed to be well-defined inside an asynchronous signal handler:
+     * any operation on a lock-free atomic object
+     * assigning a value to an object declared as volatile sig_atomic_t
+     * calling abort, _Exit, quick_exit, or signal
+       * signal may only be called with its first argument equal to the
+         number of the signal that caused the handler to be called
+
+   We use this list as a guideline for the set of operations that ought
+   also to be safe in a _synchronous_ signal delivered on an alternate
+   signal stack with only MINSIGSTKSZ bytes of space.
+
+   This test program tests calls to _Exit.  */
+
+#define EXPECTED_STATUS 3
+
+static void
+handler (int unused)
+{
+  _Exit (EXPECTED_STATUS);
+}
+
+int
+do_test (void)
+{
+  void *sstk = xalloc_sigstack (0);
+  struct sigaction sa;
+
+  sa.sa_handler = handler;
+  sa.sa_flags   = SA_RESTART | SA_ONSTACK;
+  sigfillset (&sa.sa_mask);
+  if (sigaction (SIGUSR1, &sa, 0))
+    FAIL_RET ("sigaction (SIGUSR1, handler): %m\n");
+
+  raise (SIGUSR1);
+
+  xfree_sigstack (sstk);
+  FAIL_RET ("test process was not terminated by _Exit in signal handler");
+}
+
+#include <support/test-driver.c>
diff --git a/signal/tst-minsigstksz-3a.c b/signal/tst-minsigstksz-3a.c
new file mode 100644
index 0000000000..b58b8d01ba
--- /dev/null
+++ b/signal/tst-minsigstksz-3a.c
@@ -0,0 +1,69 @@ 
+/* Tests of signal delivery on an alternate stack (_exit).
+   Copyright (C) 2019 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
+   <http://www.gnu.org/licenses/>.  */
+
+#include <support/xsignal.h>
+#include <support/support.h>
+#include <support/check.h>
+#include <unistd.h>
+
+/* C2011 7.4.1.1p5 specifies that only the following operations are
+   guaranteed to be well-defined inside an asynchronous signal handler:
+     * any operation on a lock-free atomic object
+     * assigning a value to an object declared as volatile sig_atomic_t
+     * calling abort, _Exit, quick_exit, or signal
+       * signal may only be called with its first argument equal to the
+         number of the signal that caused the handler to be called
+
+   We use this list as a guideline for the set of operations that ought
+   also to be safe in a _synchronous_ signal delivered on an alternate
+   signal stack with only MINSIGSTKSZ bytes of space.
+
+   This test program tests calls to _exit, which is the same function
+   as _Exit, but specified by POSIX rather than ISO C.  For reasons
+   unknown to the author of this program, the C committee did not
+   think it could standardize _exit under that name; regardless, in a
+   POSIX-conformant environment, they should be completely
+   interchangeable.  */
+
+#define EXPECTED_STATUS 3
+
+static void
+handler (int unused)
+{
+  _exit (EXPECTED_STATUS);
+}
+
+int
+do_test (void)
+{
+  void *sstk = xalloc_sigstack (0);
+  struct sigaction sa;
+
+  sa.sa_handler = handler;
+  sa.sa_flags   = SA_RESTART | SA_ONSTACK;
+  sigfillset (&sa.sa_mask);
+  if (sigaction (SIGUSR1, &sa, 0))
+    FAIL_RET ("sigaction (SIGUSR1, handler): %m\n");
+
+  raise (SIGUSR1);
+
+  xfree_sigstack (sstk);
+  FAIL_RET ("test process was not terminated by _exit in signal handler");
+}
+
+#include <support/test-driver.c>
diff --git a/signal/tst-minsigstksz-4.c b/signal/tst-minsigstksz-4.c
new file mode 100644
index 0000000000..0dc63b4dd4
--- /dev/null
+++ b/signal/tst-minsigstksz-4.c
@@ -0,0 +1,65 @@ 
+/* Tests of signal delivery on an alternate stack (quick_exit).
+   Copyright (C) 2019 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
+   <http://www.gnu.org/licenses/>.  */
+
+#include <support/xsignal.h>
+#include <support/support.h>
+#include <support/check.h>
+#include <stdlib.h>
+
+/* C2011 7.4.1.1p5 specifies that only the following operations are
+   guaranteed to be well-defined inside an asynchronous signal handler:
+     * any operation on a lock-free atomic object
+     * assigning a value to an object declared as volatile sig_atomic_t
+     * calling abort, _Exit, quick_exit, or signal
+       * signal may only be called with its first argument equal to the
+         number of the signal that caused the handler to be called
+
+   We use this list as a guideline for the set of operations that ought
+   also to be safe in a _synchronous_ signal delivered on an alternate
+   signal stack with only MINSIGSTKSZ bytes of space.
+
+   This test program tests calls to quick_exit.  Note that this is only
+   safe when there are no at_quick_exit callbacks.  */
+
+#define EXPECTED_STATUS 3
+
+static void
+handler (int unused)
+{
+  quick_exit (EXPECTED_STATUS);
+}
+
+int
+do_test (void)
+{
+  void *sstk = xalloc_sigstack (0);
+  struct sigaction sa;
+
+  sa.sa_handler = handler;
+  sa.sa_flags   = SA_RESTART | SA_ONSTACK;
+  sigfillset (&sa.sa_mask);
+  if (sigaction (SIGUSR1, &sa, 0))
+    FAIL_RET ("sigaction (SIGUSR1, handler): %m\n");
+
+  raise (SIGUSR1);
+
+  xfree_sigstack (sstk);
+  FAIL_RET ("test process was not terminated by quick_exit in signal handler");
+}
+
+#include <support/test-driver.c>
diff --git a/support/Makefile b/support/Makefile
index 6ac4447def..432cf2fe6c 100644
--- a/support/Makefile
+++ b/support/Makefile
@@ -145,6 +145,7 @@  libsupport-routines = \
   xsetsockopt \
   xsigaction \
   xsignal \
+  xsigstack \
   xsocket \
   xstrdup \
   xstrndup \
@@ -205,6 +206,7 @@  tests = \
   tst-test_compare_blob \
   tst-test_compare_string \
   tst-xreadlink \
+  tst-xsigstack \
 
 ifeq ($(run-built-tests),yes)
 tests-special = \
diff --git a/support/tst-xsigstack.c b/support/tst-xsigstack.c
new file mode 100644
index 0000000000..42859c79e9
--- /dev/null
+++ b/support/tst-xsigstack.c
@@ -0,0 +1,64 @@ 
+/* Test of sigaltstack wrappers.
+   Copyright (C) 2019 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
+   <http://www.gnu.org/licenses/>.  */
+
+#include <support/xsignal.h>
+#include <support/support.h>
+#include <support/xunistd.h>
+#include <support/check.h>
+
+#include <stdint.h>
+#include <stdio.h>
+
+static volatile uintptr_t handler_stackaddr;
+
+static void
+handler (int unused)
+{
+  int var;
+  handler_stackaddr = (uintptr_t) &var;
+}
+
+int
+do_test (void)
+{
+  void *sstk = xalloc_sigstack (0);
+
+  unsigned char *sp;
+  size_t size;
+  xget_sigstack_location (sstk, &sp, &size);
+  printf ("signal stack installed: sp=%p size=%zu\n", sp, size);
+
+  struct sigaction sa;
+  sa.sa_handler = handler;
+  sa.sa_flags   = SA_RESTART | SA_ONSTACK;
+  sigfillset (&sa.sa_mask);
+  if (sigaction (SIGUSR1, &sa, 0))
+    FAIL_RET ("sigaction (SIGUSR1, handler): %m\n");
+
+  raise (SIGUSR1);
+
+  uintptr_t haddr = handler_stackaddr;
+  printf ("address of handler local variable: %p\n", (void *)haddr);
+  TEST_VERIFY ((uintptr_t)sp < haddr);
+  TEST_VERIFY (haddr < (uintptr_t)sp + size);
+
+  xfree_sigstack (sstk);
+  return 0;
+}
+
+#include <support/test-driver.c>
diff --git a/support/xsignal.h b/support/xsignal.h
index f3308883a4..d89e2bc575 100644
--- a/support/xsignal.h
+++ b/support/xsignal.h
@@ -37,6 +37,23 @@  void xsigaction (int sig, const struct sigaction *newact,
 
 void xpthread_sigmask (int how, const sigset_t *set, sigset_t *oldset);
 
+/* Allocate and activate an alternate signal stack.  This stack will
+   have SIZE + MINSIGSTKSZ bytes of space, rounded up to a whole
+   number of pages.  There will be large (at least 1 MiB) inaccessible
+   guard bands on either side of it.  The return value is a cookie
+   that can be passed to xfree_sigstack to deactivate and deallocate
+   the stack again.  It is not necessary to call sigaltstack after
+   calling this function.  Terminates the process on error.  */
+void *xalloc_sigstack (size_t size);
+
+/* Deactivate and deallocate a signal stack created by xalloc_sigstack.  */
+void xfree_sigstack (void *stack);
+
+/* Extract the actual address and size of the alternate signal stack from
+   the cookie returned by xalloc_sigstack.  */
+void xget_sigstack_location (const void *stack, unsigned char **addrp,
+                             size_t *sizep);
+
 __END_DECLS
 
 #endif /* SUPPORT_SIGNAL_H */
diff --git a/support/xsigstack.c b/support/xsigstack.c
new file mode 100644
index 0000000000..cebfa19aa5
--- /dev/null
+++ b/support/xsigstack.c
@@ -0,0 +1,107 @@ 
+/* sigaltstack wrappers.
+   Copyright (C) 2019 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
+   <http://www.gnu.org/licenses/>.  */
+
+#include <support/xsignal.h>
+#include <support/support.h>
+#include <support/xunistd.h>
+#include <support/check.h>
+
+#include <stdlib.h>
+#include <string.h>
+#include <sys/mman.h>
+#include <sys/param.h> /* roundup, MAX */
+
+/* The "cookie" returned by xalloc_sigstack points to one of these
+   structures.  */
+struct sigstack_desc
+{
+  void *alloc_base;  /* Base address of the complete allocation.  */
+  size_t alloc_size; /* Size of the complete allocation.  */
+  stack_t alt_stack; /* The address and size of the stack itself.  */
+  stack_t old_stack; /* The previous signal stack.  */
+};
+
+void *
+xalloc_sigstack (size_t size)
+{
+  size_t pagesize = sysconf (_SC_PAGESIZE);
+  if (pagesize == -1)
+    FAIL_EXIT1 ("sysconf (_SC_PAGESIZE): %m\n");
+
+  /* Always supply at least MINSIGSTKSZ space; passing 0 as size means
+     only that much space.  No matter what the number is, round it up
+     to a whole number of pages.  */
+  size_t stacksize = roundup (size + MINSIGSTKSZ, pagesize);
+
+  /* The guard bands need to be large enough to intercept offset
+     accesses from a stack address that might otherwise hit another
+     mapping.  Make them at least twice as big as the stack itself, to
+     defend against an offset by the entire size of a large
+     stack-allocated array.  The minimum is 1MiB, which is arbitrarily
+     chosen to be larger than any "typical" wild pointer offset.
+     Again, no matter what the number is, round it up to a whole
+     number of pages.  */
+  size_t guardsize = roundup (MAX (2 * stacksize, 1024 * 1024), pagesize);
+
+  struct sigstack_desc *desc = xmalloc (sizeof (struct sigstack_desc));
+  desc->alloc_size = guardsize + stacksize + guardsize;
+  /* Use MAP_NORESERVE so that RAM will not be wasted on the guard
+     bands; touch all the pages of the actual stack before returning,
+     so we know they are allocated.  */
+  desc->alloc_base = xmmap (0,
+                            desc->alloc_size,
+                            PROT_READ|PROT_WRITE,
+                            MAP_PRIVATE|MAP_ANONYMOUS|MAP_NORESERVE|MAP_STACK,
+                            -1);
+
+  xmprotect (desc->alloc_base, guardsize, PROT_NONE);
+  xmprotect (desc->alloc_base + guardsize + stacksize, guardsize, PROT_NONE);
+  memset (desc->alloc_base + guardsize, 0xA5, stacksize);
+
+  desc->alt_stack.ss_sp    = desc->alloc_base + guardsize;
+  desc->alt_stack.ss_flags = 0;
+  desc->alt_stack.ss_size  = stacksize;
+
+  if (sigaltstack (&desc->alt_stack, &desc->old_stack))
+    FAIL_EXIT1 ("sigaltstack (new stack: sp=%p, size=%zu, flags=%u): %m\n",
+                desc->alt_stack.ss_sp, desc->alt_stack.ss_size,
+                desc->alt_stack.ss_flags);
+
+  return desc;
+}
+
+void
+xfree_sigstack (void *stack)
+{
+  struct sigstack_desc *desc = stack;
+
+  if (sigaltstack (&desc->old_stack, 0))
+    FAIL_EXIT1 ("sigaltstack (restore old stack: sp=%p, size=%zu, flags=%u): "
+                "%m\n", desc->old_stack.ss_sp, desc->old_stack.ss_size,
+                desc->old_stack.ss_flags);
+  xmunmap (desc->alloc_base, desc->alloc_size);
+  free (desc);
+}
+
+void
+xget_sigstack_location (const void *stack, unsigned char **addrp, size_t *sizep)
+{
+  const struct sigstack_desc *desc = stack;
+  *addrp = desc->alt_stack.ss_sp;
+  *sizep = desc->alt_stack.ss_size;
+}
-- 
2.20.1