* [PATCH] stdlib: Fix data race in __run_exit_handlers
@ 2020-09-20 9:57 Vitaly Buka via Libc-alpha
2020-09-20 12:09 ` Vitaly Buka via Libc-alpha
0 siblings, 1 reply; 24+ messages in thread
From: Vitaly Buka via Libc-alpha @ 2020-09-20 9:57 UTC (permalink / raw)
To: libc-alpha; +Cc: Vitaly Buka
Read f->func.cxa under the lock.
It's incredible rare so I failed to create a test. However we have
binary which according to logs add __cxa_atexit with unique arg
and then callback is called twice for the same arg.
There is a clear data race:
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
---
stdlib/exit.c | 20 ++++++++++++++------
1 file changed, 14 insertions(+), 6 deletions(-)
diff --git a/stdlib/exit.c b/stdlib/exit.c
index 7bca1cdc14..93cdec82f0 100644
--- a/stdlib/exit.c
+++ b/stdlib/exit.c
@@ -72,44 +72,52 @@ __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;
+ /* Unlock the list while we call a foreign function. */
+ __libc_lock_unlock (__exit_funcs_lock);
#ifdef PTR_DEMANGLE
PTR_DEMANGLE (onfct);
#endif
- onfct (status, f->func.on.arg);
+ onfct (status, arg);
+ __libc_lock_lock (__exit_funcs_lock);
break;
case ef_at:
atfct = f->func.at;
+ /* Unlock the list while we call a foreign function. */
+ __libc_lock_unlock (__exit_funcs_lock);
#ifdef PTR_DEMANGLE
PTR_DEMANGLE (atfct);
#endif
atfct ();
+ __libc_lock_lock (__exit_funcs_lock);
break;
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;
+ /* Unlock the list while we call a foreign function. */
+ __libc_lock_unlock (__exit_funcs_lock);
#ifdef PTR_DEMANGLE
PTR_DEMANGLE (cxafct);
#endif
- cxafct (f->func.cxa.arg, status);
+ 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
--
2.28.0.681.g6f77f65b4e-goog
^ permalink raw reply related [flat|nested] 24+ messages in thread
* [PATCH] stdlib: Fix data race in __run_exit_handlers
2020-09-20 9:57 [PATCH] stdlib: Fix data race in __run_exit_handlers Vitaly Buka via Libc-alpha
@ 2020-09-20 12:09 ` Vitaly Buka via Libc-alpha
2020-09-20 20:41 ` Paul Pluzhnikov via Libc-alpha
0 siblings, 1 reply; 24+ messages in thread
From: Vitaly Buka via Libc-alpha @ 2020-09-20 12:09 UTC (permalink / raw)
To: libc-alpha
Read f->func.cxa under the lock.
There is a clear data race:
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
It's incredibly rare so the test without patch fails only about 10%.
---
stdlib/Makefile | 4 +-
stdlib/exit.c | 20 +++++---
stdlib/test-cxa_atexit-race2.c | 94 ++++++++++++++++++++++++++++++++++
3 files changed, 111 insertions(+), 7 deletions(-)
create mode 100644 stdlib/test-cxa_atexit-race2.c
diff --git a/stdlib/Makefile b/stdlib/Makefile
index 4615f6dfe7..f622fa9e0b 100644
--- a/stdlib/Makefile
+++ b/stdlib/Makefile
@@ -82,7 +82,8 @@ 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 \
@@ -101,6 +102,7 @@ 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-test-dlclose-exit-race = $(shared-thread-library) $(libdl)
diff --git a/stdlib/exit.c b/stdlib/exit.c
index 7bca1cdc14..93cdec82f0 100644
--- a/stdlib/exit.c
+++ b/stdlib/exit.c
@@ -72,44 +72,52 @@ __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;
+ /* Unlock the list while we call a foreign function. */
+ __libc_lock_unlock (__exit_funcs_lock);
#ifdef PTR_DEMANGLE
PTR_DEMANGLE (onfct);
#endif
- onfct (status, f->func.on.arg);
+ onfct (status, arg);
+ __libc_lock_lock (__exit_funcs_lock);
break;
case ef_at:
atfct = f->func.at;
+ /* Unlock the list while we call a foreign function. */
+ __libc_lock_unlock (__exit_funcs_lock);
#ifdef PTR_DEMANGLE
PTR_DEMANGLE (atfct);
#endif
atfct ();
+ __libc_lock_lock (__exit_funcs_lock);
break;
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;
+ /* Unlock the list while we call a foreign function. */
+ __libc_lock_unlock (__exit_funcs_lock);
#ifdef PTR_DEMANGLE
PTR_DEMANGLE (cxafct);
#endif
- cxafct (f->func.cxa.arg, status);
+ 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
diff --git a/stdlib/test-cxa_atexit-race2.c b/stdlib/test-cxa_atexit-race2.c
new file mode 100644
index 0000000000..08acafe19f
--- /dev/null
+++ b/stdlib/test-cxa_atexit-race2.c
@@ -0,0 +1,94 @@
+/* Support file for atexit/exit, etc. race tests.
+ Copyright (C) 2017-2020 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". */
+
+/* The atexit/exit, at_quick_exit/quick_exit, __cxa_atexit/exit, etc.
+ exhibited data race while accessing destructor function list (Bug 14333).
+
+ This test spawns large number of threads, and check the same
+ desctructor is not called more then once. */
+
+#include <limits.h>
+#include <stdatomic.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <support/xthread.h>
+
+extern void *__dso_handle;
+
+const size_t kNumThreads = 50;
+const size_t kStacksize =
+#ifdef PTHREAD_STACK_MIN
+ 0x20000 < PTHREAD_STACK_MIN ? PTHREAD_STACK_MIN :
+#endif
+ 0x20000;
+
+static atomic_intptr_t added;
+static atomic_intptr_t done;
+static void *prev;
+
+static void
+atexitcb (void *arg)
+{
+ if (arg == prev)
+ abort ();
+ prev = arg;
+ ++done;
+}
+
+int __cxa_atexit (void (*func) (void *), void *arg, void *d);
+
+static void *
+threadfunc (void *unused)
+{
+ for (; done < 1e6;)
+ {
+ if (added < done + 100)
+ {
+ __cxa_atexit (&atexitcb, (void *)(++added), __dso_handle);
+ }
+ }
+ return 0;
+}
+
+static int
+do_test (void)
+{
+ size_t i;
+ pthread_attr_t attr;
+
+ xpthread_attr_init (&attr);
+ xpthread_attr_setdetachstate (&attr, 1);
+
+ /* With default 8MiB Linux stack size, creating 1024 threads can cause
+ VM exhausiton on 32-bit machines. Reduce stack size of each thread to
+ 128KiB for a maximum required VM size of 128MiB. */
+ xpthread_attr_setstacksize (&attr, kStacksize);
+
+ for (i = 0; i < kNumThreads; ++i)
+ {
+ xpthread_create (&attr, threadfunc, NULL);
+ }
+ xpthread_attr_destroy (&attr);
+
+ exit (0);
+}
+
+#define TEST_FUNCTION do_test
+#include <support/test-driver.c>
--
2.28.0.681.g6f77f65b4e-goog
^ permalink raw reply related [flat|nested] 24+ messages in thread
* Re: [PATCH] stdlib: Fix data race in __run_exit_handlers
2020-09-20 12:09 ` Vitaly Buka via Libc-alpha
@ 2020-09-20 20:41 ` Paul Pluzhnikov via Libc-alpha
2020-09-20 21:26 ` Vitaly Buka via Libc-alpha
0 siblings, 1 reply; 24+ messages in thread
From: Paul Pluzhnikov via Libc-alpha @ 2020-09-20 20:41 UTC (permalink / raw)
To: Vitaly Buka; +Cc: GLIBC Devel
On Sun, Sep 20, 2020 at 5:10 AM Vitaly Buka via Libc-alpha
<libc-alpha@sourceware.org> wrote:
> +static void *
> +threadfunc (void *unused)
> +{
> + for (; done < 1e6;)
> + {
> + if (added < done + 100)
> + {
> + __cxa_atexit (&atexitcb, (void *)(++added), __dso_handle);
Isn't there a data race on "added" here (in addition to a data race on "done")?
What prevents two threads from observing "added == 100" at the same
time and adding two calls with value of 101, which would later trigger
abort() in exitcb()?
> + /* With default 8MiB Linux stack size, creating 1024 threads can cause
> + VM exhausiton on 32-bit machines. Reduce stack size of each thread to
> + 128KiB for a maximum required VM size of 128MiB. */
This comment is far removed from the computation of kStacksize (and
the name violates the naming conventions used here).
I suggest:
size_t stack_size = 128 << 10; /* 128KiB */
if (stack_size < PTHREAD_STACK_MIN) stack_size = PTHREAD_STACK_MIN;
Also, I suspect that 32KiB would be more than enough for stack size here.
> + for (i = 0; i < kNumThreads; ++i)
Since kNumThreads isn't used anywhere else, I suggest making it a local:
const int num_threads = 50;
--
Paul Pluzhnikov
^ permalink raw reply [flat|nested] 24+ messages in thread
* Re: [PATCH] stdlib: Fix data race in __run_exit_handlers
2020-09-20 20:41 ` Paul Pluzhnikov via Libc-alpha
@ 2020-09-20 21:26 ` Vitaly Buka via Libc-alpha
2020-09-20 23:36 ` Vitaly Buka via Libc-alpha
0 siblings, 1 reply; 24+ messages in thread
From: Vitaly Buka via Libc-alpha @ 2020-09-20 21:26 UTC (permalink / raw)
To: Paul Pluzhnikov; +Cc: GLIBC Devel
On Sun, 20 Sep 2020 at 13:42, Paul Pluzhnikov <ppluzhnikov@google.com>
wrote:
> On Sun, Sep 20, 2020 at 5:10 AM Vitaly Buka via Libc-alpha
> <libc-alpha@sourceware.org> wrote:
>
> > +static void *
> > +threadfunc (void *unused)
> > +{
> > + for (; done < 1e6;)
> > + {
> > + if (added < done + 100)
> > + {
> > + __cxa_atexit (&atexitcb, (void *)(++added), __dso_handle);
>
> Isn't there a data race on "added" here (in addition to a data race on
> "done")?
> What prevents two threads from observing "added == 100" at the same
> time and adding two calls with value of 101, which would later trigger
> abort() in exitcb()?
>
They are atomic. Isn't (++added) guarantee to return different values in
all threads?
>
> > + /* With default 8MiB Linux stack size, creating 1024 threads can cause
> > + VM exhausiton on 32-bit machines. Reduce stack size of each
> thread to
> > + 128KiB for a maximum required VM size of 128MiB. */
>
> This comment is far removed from the computation of kStacksize (and
> the name violates the naming conventions used here).
>
> I suggest:
>
> size_t stack_size = 128 << 10; /* 128KiB */
> if (stack_size < PTHREAD_STACK_MIN) stack_size = PTHREAD_STACK_MIN;
>
> Also, I suspect that 32KiB would be more than enough for stack size here.
>
> > + for (i = 0; i < kNumThreads; ++i)
>
> Since kNumThreads isn't used anywhere else, I suggest making it a local:
>
> const int num_threads = 50;
>
> --
> Paul Pluzhnikov
>
^ permalink raw reply [flat|nested] 24+ messages in thread
* Re: [PATCH] stdlib: Fix data race in __run_exit_handlers
2020-09-20 21:26 ` Vitaly Buka via Libc-alpha
@ 2020-09-20 23:36 ` Vitaly Buka via Libc-alpha
2020-09-20 23:37 ` Vitaly Buka via Libc-alpha
0 siblings, 1 reply; 24+ messages in thread
From: Vitaly Buka via Libc-alpha @ 2020-09-20 23:36 UTC (permalink / raw)
To: Paul Pluzhnikov; +Cc: GLIBC Devel
Oh, this atomics are not what I thought. So the test reproduced bugs in the
test.
I tried different approaches and I failed to reproduce the data race.
On Sun, 20 Sep 2020 at 14:26, Vitaly Buka <vitalybuka@google.com> wrote:
>
>
> On Sun, 20 Sep 2020 at 13:42, Paul Pluzhnikov <ppluzhnikov@google.com>
> wrote:
>
>> On Sun, Sep 20, 2020 at 5:10 AM Vitaly Buka via Libc-alpha
>> <libc-alpha@sourceware.org> wrote:
>>
>> > +static void *
>> > +threadfunc (void *unused)
>> > +{
>> > + for (; done < 1e6;)
>> > + {
>> > + if (added < done + 100)
>> > + {
>> > + __cxa_atexit (&atexitcb, (void *)(++added), __dso_handle);
>>
>> Isn't there a data race on "added" here (in addition to a data race on
>> "done")?
>> What prevents two threads from observing "added == 100" at the same
>> time and adding two calls with value of 101, which would later trigger
>> abort() in exitcb()?
>>
>
> They are atomic. Isn't (++added) guarantee to return different values in
> all threads?
>
>
>>
>> > + /* With default 8MiB Linux stack size, creating 1024 threads can
>> cause
>> > + VM exhausiton on 32-bit machines. Reduce stack size of each
>> thread to
>> > + 128KiB for a maximum required VM size of 128MiB. */
>>
>> This comment is far removed from the computation of kStacksize (and
>> the name violates the naming conventions used here).
>>
>> I suggest:
>>
>> size_t stack_size = 128 << 10; /* 128KiB */
>> if (stack_size < PTHREAD_STACK_MIN) stack_size = PTHREAD_STACK_MIN;
>>
>> Also, I suspect that 32KiB would be more than enough for stack size here.
>>
>> > + for (i = 0; i < kNumThreads; ++i)
>>
>> Since kNumThreads isn't used anywhere else, I suggest making it a local:
>>
>> const int num_threads = 50;
>>
>> --
>> Paul Pluzhnikov
>>
>
^ permalink raw reply [flat|nested] 24+ messages in thread
* [PATCH] stdlib: Fix data race in __run_exit_handlers
2020-09-20 23:36 ` Vitaly Buka via Libc-alpha
@ 2020-09-20 23:37 ` Vitaly Buka via Libc-alpha
2020-09-21 8:31 ` Vitaly Buka via Libc-alpha
0 siblings, 1 reply; 24+ messages in thread
From: Vitaly Buka via Libc-alpha @ 2020-09-20 23:37 UTC (permalink / raw)
To: libc-alpha
Read f->func.cxa under the lock.
There is a clear data race:
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
It's incredibly rare so I can't create reasonable test.
---
stdlib/exit.c | 20 ++++++++++++++------
1 file changed, 14 insertions(+), 6 deletions(-)
diff --git a/stdlib/exit.c b/stdlib/exit.c
index 7bca1cdc14..9c79aa4a62 100644
--- a/stdlib/exit.c
+++ b/stdlib/exit.c
@@ -72,44 +72,52 @@ __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;
#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;
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
--
2.28.0.681.g6f77f65b4e-goog
^ permalink raw reply related [flat|nested] 24+ messages in thread
* [PATCH] stdlib: Fix data race in __run_exit_handlers
2020-09-20 23:37 ` Vitaly Buka via Libc-alpha
@ 2020-09-21 8:31 ` Vitaly Buka via Libc-alpha
2020-09-30 16:01 ` Joseph Myers
0 siblings, 1 reply; 24+ messages in thread
From: Vitaly Buka via Libc-alpha @ 2020-09-21 8:31 UTC (permalink / raw)
To: libc-alpha
Read f->func.cxa under the lock.
There is a clear data race:
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
Without patch abd with NDEBUG the test fails about 30% of time.
But it fails very rarely with asserts enabled.
---
stdlib/Makefile | 4 +-
stdlib/exit.c | 20 ++++++---
stdlib/test-cxa_atexit-race2.c | 76 ++++++++++++++++++++++++++++++++++
3 files changed, 93 insertions(+), 7 deletions(-)
create mode 100644 stdlib/test-cxa_atexit-race2.c
diff --git a/stdlib/Makefile b/stdlib/Makefile
index 4615f6dfe7..f622fa9e0b 100644
--- a/stdlib/Makefile
+++ b/stdlib/Makefile
@@ -82,7 +82,8 @@ 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 \
@@ -101,6 +102,7 @@ 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-test-dlclose-exit-race = $(shared-thread-library) $(libdl)
diff --git a/stdlib/exit.c b/stdlib/exit.c
index 7bca1cdc14..9c79aa4a62 100644
--- a/stdlib/exit.c
+++ b/stdlib/exit.c
@@ -72,44 +72,52 @@ __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;
#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;
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
diff --git a/stdlib/test-cxa_atexit-race2.c b/stdlib/test-cxa_atexit-race2.c
new file mode 100644
index 0000000000..9eb72471da
--- /dev/null
+++ b/stdlib/test-cxa_atexit-race2.c
@@ -0,0 +1,76 @@
+/* Support file for atexit/exit, etc. race tests.
+ Copyright (C) 2017-2020 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". */
+
+/* The atexit/exit, at_quick_exit/quick_exit, __cxa_atexit/exit, etc.
+ exhibited data race while calling destructors.
+
+ This test spawns multiple threads, and check the same desctructor is
+ not called more then once. */
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <support/xthread.h>
+
+static const int thread_num = 4;
+
+static void
+atexitcb (void *arg)
+{
+ static void *prev;
+ if (arg == prev)
+ {
+ printf ("%p\n", arg);
+ abort ();
+ }
+ prev = arg;
+}
+
+int __cxa_atexit (void (*func) (void *), void *arg, void *d);
+
+static void *
+threadfunc (void *arg)
+{
+ for (int i = 0; i < 100; ++i)
+ {
+ for (int j = 0; j < 4000; ++j)
+ __cxa_atexit (&atexitcb, arg += thread_num, 0);
+ pthread_yield ();
+ }
+ return 0;
+}
+
+static int
+do_test (void)
+{
+ size_t i;
+ pthread_attr_t attr;
+
+ xpthread_attr_init (&attr);
+ xpthread_attr_setdetachstate (&attr, 1);
+
+ for (i = 0; i < thread_num; ++i)
+ xpthread_create (&attr, threadfunc, (void *)i);
+ xpthread_attr_destroy (&attr);
+
+ exit (0);
+}
+
+#define TEST_FUNCTION do_test
+#include <support/test-driver.c>
--
2.28.0.681.g6f77f65b4e-goog
^ permalink raw reply related [flat|nested] 24+ messages in thread
* Re: [PATCH] stdlib: Fix data race in __run_exit_handlers
2020-09-21 8:31 ` Vitaly Buka via Libc-alpha
@ 2020-09-30 16:01 ` Joseph Myers
2021-04-17 16:16 ` Vitaly Buka via Libc-alpha
2021-04-17 17:36 ` Paul Pluzhnikov via Libc-alpha
0 siblings, 2 replies; 24+ messages in thread
From: Joseph Myers @ 2020-09-30 16:01 UTC (permalink / raw)
To: Vitaly Buka; +Cc: libc-alpha
Is there a bug report for this in Bugzilla? (Any bug being fixed that was
user-visible in a release should have a bug report filed. Then once the
fix is in, the bug report should be marked RESOLVED/FIXED with the target
milestone set to the first release that will have the fix; that's how we
generate the list of fixed bugs for the NEWS file.)
--
Joseph S. Myers
joseph@codesourcery.com
^ permalink raw reply [flat|nested] 24+ messages in thread
* [PATCH] stdlib: Fix data race in __run_exit_handlers
2020-09-30 16:01 ` Joseph Myers
@ 2021-04-17 16:16 ` Vitaly Buka via Libc-alpha
2021-04-17 17:11 ` Vitaly Buka via Libc-alpha
2021-04-17 17:36 ` Paul Pluzhnikov via Libc-alpha
1 sibling, 1 reply; 24+ messages in thread
From: Vitaly Buka via Libc-alpha @ 2021-04-17 16:16 UTC (permalink / raw)
To: libc-alpha; +Cc: Vitaly Buka
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
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).
The test needs multiple iterations to consistently fail without the fix.
---
stdlib/Makefile | 4 +-
stdlib/exit.c | 28 +++++---
stdlib/test-cxa_atexit-race2.c | 114 +++++++++++++++++++++++++++++++++
3 files changed, 135 insertions(+), 11 deletions(-)
create mode 100644 stdlib/test-cxa_atexit-race2.c
diff --git a/stdlib/Makefile b/stdlib/Makefile
index a9ad849531..c11a794d2a 100644
--- a/stdlib/Makefile
+++ b/stdlib/Makefile
@@ -82,7 +82,8 @@ 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
@@ -97,6 +98,7 @@ 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-test-dlclose-exit-race = $(shared-thread-library) $(libdl)
diff --git a/stdlib/exit.c b/stdlib/exit.c
index 0a4944fa05..9c669fd1a8 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
@@ -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;
@@ -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;
}
@@ -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;
#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;
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
@@ -123,9 +130,10 @@ __run_exit_handlers (int status, struct exit_function_list **listp,
allocate element. */
free (cur);
- __libc_lock_unlock (__exit_funcs_lock);
}
+ __libc_lock_unlock (__exit_funcs_lock);
+
if (run_list_atexit)
RUN_HOOK (__libc_atexit, ());
diff --git a/stdlib/test-cxa_atexit-race2.c b/stdlib/test-cxa_atexit-race2.c
new file mode 100644
index 0000000000..04656d261c
--- /dev/null
+++ b/stdlib/test-cxa_atexit-race2.c
@@ -0,0 +1,114 @@
+/* Support file for atexit/exit, etc. race tests.
+ 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". */
+
+/* The atexit/exit, at_quick_exit/quick_exit, __cxa_atexit/exit, etc.
+ exhibited data race while calling destructors.
+
+ This test spawns multiple threads, and check the same desctructor is
+ not called more then 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)
+{
+ --registered;
+ static void *prev;
+ if (arg == prev)
+ {
+ printf ("%p\n", arg);
+ abort ();
+ }
+ prev = arg;
+
+ while (todo > 0 && registered < 100)
+ ;
+}
+
+int __cxa_atexit (void (*func) (void *), void *arg, void *d);
+
+static void *cb_arg = NULL;
+static void
+add_handlers (void)
+{
+ int n = 10;
+ for (int i = 0; i < n; ++i)
+ __cxa_atexit (&atexit_cb, ++cb_arg, 0);
+ registered += n;
+ todo -= n;
+}
+
+static void *
+thread_func (void *arg)
+{
+ while (todo > 0)
+ if (registered < 10000)
+ add_handlers ();
+ return 0;
+}
+
+static void
+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 (!registered)
+ ;
+ exit (0);
+}
+
+static int
+do_test (void)
+{
+ for (int i = 0; i < 20; ++i)
+ {
+ for (int i = 0; i < 10; ++i)
+ if (fork () == 0)
+ test_and_exit ();
+
+ int status;
+ while (wait (&status) > 0)
+ {
+ if (!WIFEXITED (status))
+ {
+ printf ("Failed interation %d\n", i);
+ abort ();
+ }
+ }
+ }
+
+ exit (0);
+}
+
+#define TEST_FUNCTION do_test
+#include <support/test-driver.c>
--
2.31.1.368.gbe11c130af-goog
^ permalink raw reply related [flat|nested] 24+ messages in thread
* [PATCH] stdlib: Fix data race in __run_exit_handlers
2021-04-17 16:16 ` Vitaly Buka via Libc-alpha
@ 2021-04-17 17:11 ` Vitaly Buka via Libc-alpha
2021-04-17 17:13 ` Vitaly Buka via Libc-alpha
0 siblings, 1 reply; 24+ messages in thread
From: Vitaly Buka via Libc-alpha @ 2021-04-17 17:11 UTC (permalink / raw)
To: libc-alpha; +Cc: Vitaly Buka
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
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).
The test needs multiple iterations to consistently fail without the fix.
---
stdlib/Makefile | 4 +-
stdlib/exit.c | 28 +++++---
stdlib/test-cxa_atexit-race2.c | 114 +++++++++++++++++++++++++++++++++
3 files changed, 135 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
@@ -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;
@@ -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;
}
@@ -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;
#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;
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
@@ -123,9 +130,10 @@ __run_exit_handlers (int status, struct exit_function_list **listp,
allocate element. */
free (cur);
- __libc_lock_unlock (__exit_funcs_lock);
}
+ __libc_lock_unlock (__exit_funcs_lock);
+
if (run_list_atexit)
RUN_HOOK (__libc_atexit, ());
diff --git a/stdlib/test-cxa_atexit-race2.c b/stdlib/test-cxa_atexit-race2.c
new file mode 100644
index 0000000000..04656d261c
--- /dev/null
+++ b/stdlib/test-cxa_atexit-race2.c
@@ -0,0 +1,114 @@
+/* Support file for atexit/exit, etc. race tests.
+ 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". */
+
+/* The atexit/exit, at_quick_exit/quick_exit, __cxa_atexit/exit, etc.
+ exhibited data race while calling destructors.
+
+ This test spawns multiple threads, and check the same desctructor is
+ not called more then 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)
+{
+ --registered;
+ static void *prev;
+ if (arg == prev)
+ {
+ printf ("%p\n", arg);
+ abort ();
+ }
+ prev = arg;
+
+ while (todo > 0 && registered < 100)
+ ;
+}
+
+int __cxa_atexit (void (*func) (void *), void *arg, void *d);
+
+static void *cb_arg = NULL;
+static void
+add_handlers (void)
+{
+ int n = 10;
+ for (int i = 0; i < n; ++i)
+ __cxa_atexit (&atexit_cb, ++cb_arg, 0);
+ registered += n;
+ todo -= n;
+}
+
+static void *
+thread_func (void *arg)
+{
+ while (todo > 0)
+ if (registered < 10000)
+ add_handlers ();
+ return 0;
+}
+
+static void
+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 (!registered)
+ ;
+ exit (0);
+}
+
+static int
+do_test (void)
+{
+ for (int i = 0; i < 20; ++i)
+ {
+ for (int i = 0; i < 10; ++i)
+ if (fork () == 0)
+ test_and_exit ();
+
+ int status;
+ while (wait (&status) > 0)
+ {
+ if (!WIFEXITED (status))
+ {
+ printf ("Failed interation %d\n", i);
+ abort ();
+ }
+ }
+ }
+
+ exit (0);
+}
+
+#define TEST_FUNCTION do_test
+#include <support/test-driver.c>
--
2.31.1.368.gbe11c130af-goog
^ permalink raw reply related [flat|nested] 24+ messages in thread
* Re: [PATCH] stdlib: Fix data race in __run_exit_handlers
2021-04-17 17:11 ` Vitaly Buka via Libc-alpha
@ 2021-04-17 17:13 ` Vitaly Buka via Libc-alpha
2021-04-17 17:22 ` Vitaly Buka via Libc-alpha
0 siblings, 1 reply; 24+ messages in thread
From: Vitaly Buka via Libc-alpha @ 2021-04-17 17:13 UTC (permalink / raw)
To: GLIBC Devel
In the last patch I've fixed atomic usage in the test.
Here is the bug https://sourceware.org/bugzilla/show_bug.cgi?id=27749
On Sat, 17 Apr 2021 at 10:11, Vitaly Buka <vitalybuka@google.com> wrote:
> 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
>
> 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).
>
> The test needs multiple iterations to consistently fail without the fix.
> ---
> stdlib/Makefile | 4 +-
> stdlib/exit.c | 28 +++++---
> stdlib/test-cxa_atexit-race2.c | 114 +++++++++++++++++++++++++++++++++
> 3 files changed, 135 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
> @@ -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;
>
> @@ -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;
> }
>
> @@ -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;
> #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;
> 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
> @@ -123,9 +130,10 @@ __run_exit_handlers (int status, struct
> exit_function_list **listp,
> allocate element. */
> free (cur);
>
> - __libc_lock_unlock (__exit_funcs_lock);
> }
>
> + __libc_lock_unlock (__exit_funcs_lock);
> +
> if (run_list_atexit)
> RUN_HOOK (__libc_atexit, ());
>
> diff --git a/stdlib/test-cxa_atexit-race2.c
> b/stdlib/test-cxa_atexit-race2.c
> new file mode 100644
> index 0000000000..04656d261c
> --- /dev/null
> +++ b/stdlib/test-cxa_atexit-race2.c
> @@ -0,0 +1,114 @@
> +/* Support file for atexit/exit, etc. race tests.
> + 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". */
> +
> +/* The atexit/exit, at_quick_exit/quick_exit, __cxa_atexit/exit, etc.
> + exhibited data race while calling destructors.
> +
> + This test spawns multiple threads, and check the same desctructor is
> + not called more then 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)
> +{
> + --registered;
> + static void *prev;
> + if (arg == prev)
> + {
> + printf ("%p\n", arg);
> + abort ();
> + }
> + prev = arg;
> +
> + while (todo > 0 && registered < 100)
> + ;
> +}
> +
> +int __cxa_atexit (void (*func) (void *), void *arg, void *d);
> +
> +static void *cb_arg = NULL;
> +static void
> +add_handlers (void)
> +{
> + int n = 10;
> + for (int i = 0; i < n; ++i)
> + __cxa_atexit (&atexit_cb, ++cb_arg, 0);
> + registered += n;
> + todo -= n;
> +}
> +
> +static void *
> +thread_func (void *arg)
> +{
> + while (todo > 0)
> + if (registered < 10000)
> + add_handlers ();
> + return 0;
> +}
> +
> +static void
> +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 (!registered)
> + ;
> + exit (0);
> +}
> +
> +static int
> +do_test (void)
> +{
> + for (int i = 0; i < 20; ++i)
> + {
> + for (int i = 0; i < 10; ++i)
> + if (fork () == 0)
> + test_and_exit ();
> +
> + int status;
> + while (wait (&status) > 0)
> + {
> + if (!WIFEXITED (status))
> + {
> + printf ("Failed interation %d\n", i);
> + abort ();
> + }
> + }
> + }
> +
> + exit (0);
> +}
> +
> +#define TEST_FUNCTION do_test
> +#include <support/test-driver.c>
> --
> 2.31.1.368.gbe11c130af-goog
>
>
^ permalink raw reply [flat|nested] 24+ messages in thread
* [PATCH] stdlib: Fix data race in __run_exit_handlers
2021-04-17 17:13 ` Vitaly Buka via Libc-alpha
@ 2021-04-17 17:22 ` Vitaly Buka via Libc-alpha
2021-04-17 18:01 ` Paul Pluzhnikov via Libc-alpha
0 siblings, 1 reply; 24+ messages in thread
From: Vitaly Buka via Libc-alpha @ 2021-04-17 17:22 UTC (permalink / raw)
To: libc-alpha; +Cc: Vitaly Buka
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
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).
The test needs multiple iterations to consistently fail without the fix.
---
stdlib/Makefile | 4 +-
stdlib/exit.c | 28 +++++---
stdlib/test-cxa_atexit-race2.c | 114 +++++++++++++++++++++++++++++++++
3 files changed, 135 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
@@ -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;
@@ -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;
}
@@ -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;
#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;
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
@@ -123,9 +130,10 @@ __run_exit_handlers (int status, struct exit_function_list **listp,
allocate element. */
free (cur);
- __libc_lock_unlock (__exit_funcs_lock);
}
+ __libc_lock_unlock (__exit_funcs_lock);
+
if (run_list_atexit)
RUN_HOOK (__libc_atexit, ());
diff --git a/stdlib/test-cxa_atexit-race2.c b/stdlib/test-cxa_atexit-race2.c
new file mode 100644
index 0000000000..0aed0e7eae
--- /dev/null
+++ b/stdlib/test-cxa_atexit-race2.c
@@ -0,0 +1,114 @@
+/* Support file for atexit/exit, etc. race tests.
+ 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". */
+
+/* The atexit/exit, at_quick_exit/quick_exit, __cxa_atexit/exit, etc.
+ exhibited data race while calling destructors.
+
+ This test spawns multiple threads, and check the same desctructor is
+ not called more then 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 (®istered, 1);
+ static void *prev;
+ if (arg == prev)
+ {
+ printf ("%p\n", arg);
+ abort ();
+ }
+ prev = arg;
+
+ while (atomic_load (&todo) > 0 && atomic_load (®istered) < 100)
+ ;
+}
+
+int __cxa_atexit (void (*func) (void *), void *arg, void *d);
+
+static void *cb_arg = NULL;
+static void
+add_handlers (void)
+{
+ int n = 10;
+ for (int i = 0; i < n; ++i)
+ __cxa_atexit (&atexit_cb, ++cb_arg, 0);
+ atomic_fetch_add (®istered, n);
+ atomic_fetch_sub (&todo, n);
+}
+
+static void *
+thread_func (void *arg)
+{
+ while (atomic_load (&todo) > 0)
+ if (atomic_load (®istered) < 10000)
+ add_handlers ();
+ return 0;
+}
+
+static void
+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 (®istered))
+ ;
+ exit (0);
+}
+
+static int
+do_test (void)
+{
+ for (int i = 0; i < 20; ++i)
+ {
+ for (int i = 0; i < 10; ++i)
+ if (fork () == 0)
+ test_and_exit ();
+
+ int status;
+ while (wait (&status) > 0)
+ {
+ if (!WIFEXITED (status))
+ {
+ printf ("Failed interation %d\n", i);
+ abort ();
+ }
+ }
+ }
+
+ exit (0);
+}
+
+#define TEST_FUNCTION do_test
+#include <support/test-driver.c>
--
2.31.1.368.gbe11c130af-goog
^ permalink raw reply related [flat|nested] 24+ messages in thread
* Re: [PATCH] stdlib: Fix data race in __run_exit_handlers
2020-09-30 16:01 ` Joseph Myers
2021-04-17 16:16 ` Vitaly Buka via Libc-alpha
@ 2021-04-17 17:36 ` Paul Pluzhnikov via Libc-alpha
2021-04-17 20:19 ` Florian Weimer
1 sibling, 1 reply; 24+ messages in thread
From: Paul Pluzhnikov via Libc-alpha @ 2021-04-17 17:36 UTC (permalink / raw)
To: Joseph Myers; +Cc: Vitaly Buka, GLIBC Devel
On Wed, Sep 30, 2020 at 9:01 AM Joseph Myers <joseph@codesourcery.com> wrote:
>
> Is there a bug report for this in Bugzilla?
This is really a continuation of BZ#14333, a fix for which was incomplete.
Do we want a new bug for this?
--
Paul Pluzhnikov
^ permalink raw reply [flat|nested] 24+ messages in thread
* Re: [PATCH] stdlib: Fix data race in __run_exit_handlers
2021-04-17 17:22 ` Vitaly Buka via Libc-alpha
@ 2021-04-17 18:01 ` Paul Pluzhnikov via Libc-alpha
2021-04-20 22:51 ` Vitaly Buka via Libc-alpha
0 siblings, 1 reply; 24+ messages in thread
From: Paul Pluzhnikov via Libc-alpha @ 2021-04-17 18:01 UTC (permalink / raw)
To: Vitaly Buka; +Cc: GLIBC Devel
On Sat, Apr 17, 2021 at 10:23 AM Vitaly Buka via Libc-alpha
<libc-alpha@sourceware.org> wrote:
> diff --git a/stdlib/test-cxa_atexit-race2.c b/stdlib/test-cxa_atexit-race2.c
> + This test spawns multiple threads, and check the same desctructor is
Typo. Suggest:
This test spawns multiple threads, and checks that the same destructor is
> + not called more then once. */
s/then/than/
> +static void *cb_arg = NULL;
> +static void
> +add_handlers (void)
> +{
> + int n = 10;
> + for (int i = 0; i < n; ++i)
> + __cxa_atexit (&atexit_cb, ++cb_arg, 0);
add_handlers() is called from many threads. This code appears to race on cb_arg.
--
Paul Pluzhnikov
^ permalink raw reply [flat|nested] 24+ messages in thread
* Re: [PATCH] stdlib: Fix data race in __run_exit_handlers
2021-04-17 17:36 ` Paul Pluzhnikov via Libc-alpha
@ 2021-04-17 20:19 ` Florian Weimer
2021-04-19 2:48 ` Vitaly Buka via Libc-alpha
0 siblings, 1 reply; 24+ messages in thread
From: Florian Weimer @ 2021-04-17 20:19 UTC (permalink / raw)
To: Paul Pluzhnikov via Libc-alpha; +Cc: Vitaly Buka, Joseph Myers
* Paul Pluzhnikov via Libc-alpha:
> On Wed, Sep 30, 2020 at 9:01 AM Joseph Myers <joseph@codesourcery.com> wrote:
>>
>> Is there a bug report for this in Bugzilla?
>
> This is really a continuation of BZ#14333, a fix for which was incomplete.
>
> Do we want a new bug for this?
Yes, please.
^ permalink raw reply [flat|nested] 24+ messages in thread
* [PATCH] stdlib: Fix data race in __run_exit_handlers
2021-04-17 20:19 ` Florian Weimer
@ 2021-04-19 2:48 ` Vitaly Buka via Libc-alpha
2021-04-19 2:57 ` Vitaly Buka via Libc-alpha
0 siblings, 1 reply; 24+ messages in thread
From: Vitaly Buka via Libc-alpha @ 2021-04-19 2:48 UTC (permalink / raw)
To: libc-alpha; +Cc: Vitaly Buka
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
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).
The test needs multiple iterations to consistently fail without the fix.
---
stdlib/Makefile | 4 +-
stdlib/exit.c | 28 +++++---
stdlib/test-cxa_atexit-race2.c | 114 +++++++++++++++++++++++++++++++++
3 files changed, 135 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
@@ -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;
@@ -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;
}
@@ -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;
#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;
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
@@ -123,9 +130,10 @@ __run_exit_handlers (int status, struct exit_function_list **listp,
allocate element. */
free (cur);
- __libc_lock_unlock (__exit_funcs_lock);
}
+ __libc_lock_unlock (__exit_funcs_lock);
+
if (run_list_atexit)
RUN_HOOK (__libc_atexit, ());
diff --git a/stdlib/test-cxa_atexit-race2.c b/stdlib/test-cxa_atexit-race2.c
new file mode 100644
index 0000000000..8f97ad2031
--- /dev/null
+++ b/stdlib/test-cxa_atexit-race2.c
@@ -0,0 +1,114 @@
+/* Support file for atexit/exit, etc. race tests.
+ 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". */
+
+/* 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 then 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 (®istered, 1);
+ static void *prev;
+ if (arg == prev)
+ {
+ printf ("%p\n", arg);
+ abort ();
+ }
+ prev = arg;
+
+ while (atomic_load (&todo) > 0 && atomic_load (®istered) < 100)
+ ;
+}
+
+int __cxa_atexit (void (*func) (void *), void *arg, void *d);
+
+static void *cb_arg = NULL;
+static void
+add_handlers (void)
+{
+ int n = 10;
+ for (int i = 0; i < n; ++i)
+ __cxa_atexit (&atexit_cb, ++cb_arg, 0);
+ atomic_fetch_add (®istered, n);
+ atomic_fetch_sub (&todo, n);
+}
+
+static void *
+thread_func (void *arg)
+{
+ while (atomic_load (&todo) > 0)
+ if (atomic_load (®istered) < 10000)
+ add_handlers ();
+ return 0;
+}
+
+static void
+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 (®istered))
+ ;
+ exit (0);
+}
+
+static int
+do_test (void)
+{
+ for (int i = 0; i < 20; ++i)
+ {
+ for (int i = 0; i < 10; ++i)
+ if (fork () == 0)
+ test_and_exit ();
+
+ int status;
+ while (wait (&status) > 0)
+ {
+ if (!WIFEXITED (status))
+ {
+ printf ("Failed interation %d\n", i);
+ abort ();
+ }
+ }
+ }
+
+ exit (0);
+}
+
+#define TEST_FUNCTION do_test
+#include <support/test-driver.c>
--
2.31.1.368.gbe11c130af-goog
^ permalink raw reply related [flat|nested] 24+ messages in thread
* [PATCH] stdlib: Fix data race in __run_exit_handlers
2021-04-19 2:48 ` Vitaly Buka via Libc-alpha
@ 2021-04-19 2:57 ` Vitaly Buka via Libc-alpha
0 siblings, 0 replies; 24+ messages in thread
From: Vitaly Buka via Libc-alpha @ 2021-04-19 2:57 UTC (permalink / raw)
To: libc-alpha; +Cc: Vitaly Buka
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
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).
The test needs multiple iterations to consistently fail without the fix.
---
stdlib/Makefile | 4 +-
stdlib/exit.c | 28 +++++---
stdlib/test-cxa_atexit-race2.c | 114 +++++++++++++++++++++++++++++++++
3 files changed, 135 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
@@ -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;
@@ -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;
}
@@ -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;
#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;
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
@@ -123,9 +130,10 @@ __run_exit_handlers (int status, struct exit_function_list **listp,
allocate element. */
free (cur);
- __libc_lock_unlock (__exit_funcs_lock);
}
+ __libc_lock_unlock (__exit_funcs_lock);
+
if (run_list_atexit)
RUN_HOOK (__libc_atexit, ());
diff --git a/stdlib/test-cxa_atexit-race2.c b/stdlib/test-cxa_atexit-race2.c
new file mode 100644
index 0000000000..27f142c6a3
--- /dev/null
+++ b/stdlib/test-cxa_atexit-race2.c
@@ -0,0 +1,114 @@
+/* Support file for atexit/exit, etc. race tests.
+ 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". */
+
+/* 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 (®istered, 1);
+ static void *prev;
+ if (arg == prev)
+ {
+ printf ("%p\n", arg);
+ abort ();
+ }
+ prev = arg;
+
+ while (atomic_load (&todo) > 0 && atomic_load (®istered) < 100)
+ ;
+}
+
+int __cxa_atexit (void (*func) (void *), void *arg, void *d);
+
+static void *cb_arg = NULL;
+static void
+add_handlers (void)
+{
+ int n = 10;
+ for (int i = 0; i < n; ++i)
+ __cxa_atexit (&atexit_cb, ++cb_arg, 0);
+ atomic_fetch_add (®istered, n);
+ atomic_fetch_sub (&todo, n);
+}
+
+static void *
+thread_func (void *arg)
+{
+ while (atomic_load (&todo) > 0)
+ if (atomic_load (®istered) < 10000)
+ add_handlers ();
+ return 0;
+}
+
+static void
+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 (®istered))
+ ;
+ exit (0);
+}
+
+static int
+do_test (void)
+{
+ for (int i = 0; i < 20; ++i)
+ {
+ for (int i = 0; i < 10; ++i)
+ if (fork () == 0)
+ test_and_exit ();
+
+ int status;
+ while (wait (&status) > 0)
+ {
+ if (!WIFEXITED (status))
+ {
+ printf ("Failed interation %d\n", i);
+ abort ();
+ }
+ }
+ }
+
+ exit (0);
+}
+
+#define TEST_FUNCTION do_test
+#include <support/test-driver.c>
--
2.31.1.368.gbe11c130af-goog
^ permalink raw reply related [flat|nested] 24+ messages in thread
* Re: [PATCH] stdlib: Fix data race in __run_exit_handlers
2021-04-17 18:01 ` Paul Pluzhnikov via Libc-alpha
@ 2021-04-20 22:51 ` Vitaly Buka via Libc-alpha
2021-04-20 23:40 ` Paul Pluzhnikov via Libc-alpha
0 siblings, 1 reply; 24+ messages in thread
From: Vitaly Buka via Libc-alpha @ 2021-04-20 22:51 UTC (permalink / raw)
To: Paul Pluzhnikov; +Cc: GLIBC Devel
I guess I forgot to reply to these comments.
On Sat, 17 Apr 2021 at 11:02, Paul Pluzhnikov <ppluzhnikov@google.com>
wrote:
> On Sat, Apr 17, 2021 at 10:23 AM Vitaly Buka via Libc-alpha
> <libc-alpha@sourceware.org> wrote:
>
> > diff --git a/stdlib/test-cxa_atexit-race2.c
> b/stdlib/test-cxa_atexit-race2.c
>
> > + This test spawns multiple threads, and check the same desctructor is
>
> Typo. Suggest:
>
> This test spawns multiple threads, and checks that the same destructor is
>
Done in the last patch.
>
> > + not called more then once. */
>
> s/then/than/
>
Done in the last patch.
>
> > +static void *cb_arg = NULL;
> > +static void
> > +add_handlers (void)
> > +{
> > + int n = 10;
> > + for (int i = 0; i < n; ++i)
> > + __cxa_atexit (&atexit_cb, ++cb_arg, 0);
>
> add_handlers() is called from many threads. This code appears to race on
> cb_arg.
>
We don't have a data race as add_handlers is called from a single
background thread.
Previous patches had more threads but they didn't help to reproduce the
issues.
>
>
> --
> Paul Pluzhnikov
>
^ permalink raw reply [flat|nested] 24+ messages in thread
* Re: [PATCH] stdlib: Fix data race in __run_exit_handlers
2021-04-20 22:51 ` Vitaly Buka via Libc-alpha
@ 2021-04-20 23:40 ` Paul Pluzhnikov via Libc-alpha
2021-04-26 19:20 ` Vitaly Buka via Libc-alpha
0 siblings, 1 reply; 24+ messages in thread
From: Paul Pluzhnikov via Libc-alpha @ 2021-04-20 23:40 UTC (permalink / raw)
To: Vitaly Buka; +Cc: GLIBC Devel
On Tue, Apr 20, 2021 at 3:51 PM Vitaly Buka <vitalybuka@google.com> wrote:
>> > +static void *cb_arg = NULL;
>> > +static void
>> > +add_handlers (void)
>> > +{
>> > + int n = 10;
>> > + for (int i = 0; i < n; ++i)
>> > + __cxa_atexit (&atexit_cb, ++cb_arg, 0);
>>
>> add_handlers() is called from many threads. This code appears to race on cb_arg.
>
>
> We don't have a data race as add_handlers is called from a single background thread.
> Previous patches had more threads but they didn't help to reproduce the issues.
Thanks.
I think cb_arg can be moved into add_handlers() and doesn't have to be
static anymore (if it were a local, there wouldn't be a question of a
race in the first place).
I've looked at the code and the first data race description at the
start of this thread.
I agree that this is the right fix for it.
I have not yet understood the second interaction (between
__run_exit_handlers and __new_exitfn), but I am not sure I really need
to: the patch seems correct.
One other change I would make is to move the unlock before
PTR_DEMANGLE (since it doesn't use any of the data guarded by this
lock).
--
Paul Pluzhnikov
^ permalink raw reply [flat|nested] 24+ messages in thread
* [PATCH] stdlib: Fix data race in __run_exit_handlers
2021-04-20 23:40 ` Paul Pluzhnikov via Libc-alpha
@ 2021-04-26 19:20 ` Vitaly Buka via Libc-alpha
2021-04-26 19:23 ` Vitaly Buka via Libc-alpha
0 siblings, 1 reply; 24+ messages in thread
From: Vitaly Buka via Libc-alpha @ 2021-04-26 19:20 UTC (permalink / raw)
To: libc-alpha; +Cc: Vitaly Buka
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
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);
The test needs multiple iterations to consistently fail without the fix.
---
stdlib/Makefile | 4 +-
stdlib/exit.c | 28 ++++++---
stdlib/test-cxa_atexit-race2.c | 109 +++++++++++++++++++++++++++++++++
3 files changed, 130 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
@@ -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;
@@ -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;
}
@@ -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;
#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;
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
@@ -123,9 +130,10 @@ __run_exit_handlers (int status, struct exit_function_list **listp,
allocate element. */
free (cur);
- __libc_lock_unlock (__exit_funcs_lock);
}
+ __libc_lock_unlock (__exit_funcs_lock);
+
if (run_list_atexit)
RUN_HOOK (__libc_atexit, ());
diff --git a/stdlib/test-cxa_atexit-race2.c b/stdlib/test-cxa_atexit-race2.c
new file mode 100644
index 0000000000..9ab110ebaf
--- /dev/null
+++ b/stdlib/test-cxa_atexit-race2.c
@@ -0,0 +1,109 @@
+/* Support file for atexit/exit, etc. race tests.
+ 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". */
+
+/* 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 (®istered, 1);
+ static void *prev;
+ if (arg == prev)
+ {
+ printf ("%p\n", arg);
+ abort ();
+ }
+ prev = arg;
+
+ while (atomic_load (&todo) > 0 && atomic_load (®istered) < 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)
+ if (atomic_load(®istered) < 10000) {
+ int n = 10;
+ for (int i = 0; i < n; ++i) __cxa_atexit(&atexit_cb, ++cb_arg, 0);
+ atomic_fetch_add(®istered, n);
+ atomic_fetch_sub(&todo, n);
+ }
+ return 0;
+}
+
+static void
+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 (®istered))
+ ;
+ exit (0);
+}
+
+static int
+do_test (void)
+{
+ for (int i = 0; i < 20; ++i)
+ {
+ for (int i = 0; i < 10; ++i)
+ if (fork () == 0)
+ test_and_exit ();
+
+ int status;
+ while (wait (&status) > 0)
+ {
+ if (!WIFEXITED (status))
+ {
+ printf ("Failed interation %d\n", i);
+ abort ();
+ }
+ }
+ }
+
+ exit (0);
+}
+
+#define TEST_FUNCTION do_test
+#include <support/test-driver.c>
--
2.31.1.498.g6c1eba8ee3d-goog
^ permalink raw reply related [flat|nested] 24+ messages in thread
* Re: [PATCH] stdlib: Fix data race in __run_exit_handlers
2021-04-26 19:20 ` Vitaly Buka via Libc-alpha
@ 2021-04-26 19:23 ` Vitaly Buka via Libc-alpha
2021-04-26 19:27 ` Vitaly Buka via Libc-alpha
0 siblings, 1 reply; 24+ messages in thread
From: Vitaly Buka via Libc-alpha @ 2021-04-26 19:23 UTC (permalink / raw)
To: GLIBC Devel
The last patch removes static cb_arg and improves description of race case
2.
On Mon, 26 Apr 2021 at 12:20, Vitaly Buka <vitalybuka@google.com> 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
>
> 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);
>
> The test needs multiple iterations to consistently fail without the fix.
> ---
> stdlib/Makefile | 4 +-
> stdlib/exit.c | 28 ++++++---
> stdlib/test-cxa_atexit-race2.c | 109 +++++++++++++++++++++++++++++++++
> 3 files changed, 130 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
> @@ -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;
>
> @@ -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;
> }
>
> @@ -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;
> #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;
> 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
> @@ -123,9 +130,10 @@ __run_exit_handlers (int status, struct
> exit_function_list **listp,
> allocate element. */
> free (cur);
>
> - __libc_lock_unlock (__exit_funcs_lock);
> }
>
> + __libc_lock_unlock (__exit_funcs_lock);
> +
> if (run_list_atexit)
> RUN_HOOK (__libc_atexit, ());
>
> diff --git a/stdlib/test-cxa_atexit-race2.c
> b/stdlib/test-cxa_atexit-race2.c
> new file mode 100644
> index 0000000000..9ab110ebaf
> --- /dev/null
> +++ b/stdlib/test-cxa_atexit-race2.c
> @@ -0,0 +1,109 @@
> +/* Support file for atexit/exit, etc. race tests.
> + 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". */
> +
> +/* 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 (®istered, 1);
> + static void *prev;
> + if (arg == prev)
> + {
> + printf ("%p\n", arg);
> + abort ();
> + }
> + prev = arg;
> +
> + while (atomic_load (&todo) > 0 && atomic_load (®istered) < 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)
> + if (atomic_load(®istered) < 10000) {
> + int n = 10;
> + for (int i = 0; i < n; ++i) __cxa_atexit(&atexit_cb, ++cb_arg, 0);
> + atomic_fetch_add(®istered, n);
> + atomic_fetch_sub(&todo, n);
> + }
> + return 0;
> +}
> +
> +static void
> +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 (®istered))
> + ;
> + exit (0);
> +}
> +
> +static int
> +do_test (void)
> +{
> + for (int i = 0; i < 20; ++i)
> + {
> + for (int i = 0; i < 10; ++i)
> + if (fork () == 0)
> + test_and_exit ();
> +
> + int status;
> + while (wait (&status) > 0)
> + {
> + if (!WIFEXITED (status))
> + {
> + printf ("Failed interation %d\n", i);
> + abort ();
> + }
> + }
> + }
> +
> + exit (0);
> +}
> +
> +#define TEST_FUNCTION do_test
> +#include <support/test-driver.c>
> --
> 2.31.1.498.g6c1eba8ee3d-goog
>
>
^ permalink raw reply [flat|nested] 24+ messages in thread
* [PATCH] stdlib: Fix data race in __run_exit_handlers
2021-04-26 19:23 ` Vitaly Buka via Libc-alpha
@ 2021-04-26 19:27 ` Vitaly Buka via Libc-alpha
2021-05-13 13:15 ` Adhemerval Zanella via Libc-alpha
0 siblings, 1 reply; 24+ messages in thread
From: Vitaly Buka via Libc-alpha @ 2021-04-26 19:27 UTC (permalink / raw)
To: libc-alpha; +Cc: Vitaly Buka
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
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);
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
@@ -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;
@@ -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;
}
@@ -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;
#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;
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
@@ -123,9 +130,10 @@ __run_exit_handlers (int status, struct exit_function_list **listp,
allocate element. */
free (cur);
- __libc_lock_unlock (__exit_funcs_lock);
}
+ __libc_lock_unlock (__exit_funcs_lock);
+
if (run_list_atexit)
RUN_HOOK (__libc_atexit, ());
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.
+ 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". */
+
+/* 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 (®istered, 1);
+ static void *prev;
+ if (arg == prev)
+ {
+ printf ("%p\n", arg);
+ abort ();
+ }
+ prev = arg;
+
+ while (atomic_load (&todo) > 0 && atomic_load (®istered) < 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)
+ if (atomic_load (®istered) < 10000)
+ {
+ int n = 10;
+ for (int i = 0; i < n; ++i)
+ __cxa_atexit (&atexit_cb, ++cb_arg, 0);
+ atomic_fetch_add (®istered, n);
+ atomic_fetch_sub (&todo, n);
+ }
+ return 0;
+}
+
+static void
+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 (®istered))
+ ;
+ exit (0);
+}
+
+static int
+do_test (void)
+{
+ for (int i = 0; i < 20; ++i)
+ {
+ for (int i = 0; i < 10; ++i)
+ if (fork () == 0)
+ test_and_exit ();
+
+ int status;
+ while (wait (&status) > 0)
+ {
+ if (!WIFEXITED (status))
+ {
+ printf ("Failed interation %d\n", i);
+ abort ();
+ }
+ }
+ }
+
+ exit (0);
+}
+
+#define TEST_FUNCTION do_test
+#include <support/test-driver.c>
--
2.31.1.498.g6c1eba8ee3d-goog
^ permalink raw reply related [flat|nested] 24+ messages in thread
* Re: [PATCH] stdlib: Fix data race in __run_exit_handlers
2021-04-26 19:27 ` Vitaly Buka via Libc-alpha
@ 2021-05-13 13:15 ` Adhemerval Zanella via Libc-alpha
2021-05-14 6:50 ` Vitaly Buka via Libc-alpha
0 siblings, 1 reply; 24+ messages in thread
From: Adhemerval Zanella via Libc-alpha @ 2021-05-13 13:15 UTC (permalink / raw)
To: Vitaly Buka, libc-alpha
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 (®istered, 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 (®istered) < 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 (®istered) < 10000)
> + {
> + int n = 10;
> + for (int i = 0; i < n; ++i)
> + __cxa_atexit (&atexit_cb, ++cb_arg, 0);
> + atomic_fetch_add (®istered, 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 (®istered))
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>
>
^ permalink raw reply [flat|nested] 24+ messages in thread
* Re: [PATCH] stdlib: Fix data race in __run_exit_handlers
2021-05-13 13:15 ` Adhemerval Zanella via Libc-alpha
@ 2021-05-14 6:50 ` Vitaly Buka via Libc-alpha
0 siblings, 0 replies; 24+ messages in thread
From: Vitaly Buka via Libc-alpha @ 2021-05-14 6:50 UTC (permalink / raw)
To: Adhemerval Zanella; +Cc: GLIBC Devel
Thank you. These improvements look good to me. Please push it.
On Thu, 13 May 2021 at 06:15, Adhemerval Zanella <
adhemerval.zanella@linaro.org> wrote:
>
>
> 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 (®istered, 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 (®istered) < 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 (®istered) < 10000)
> > + {
> > + int n = 10;
> > + for (int i = 0; i < n; ++i)
> > + __cxa_atexit (&atexit_cb, ++cb_arg, 0);
> > + atomic_fetch_add (®istered, 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 (®istered))
>
> 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>
> >
>
^ permalink raw reply [flat|nested] 24+ messages in thread
end of thread, other threads:[~2021-05-14 6:51 UTC | newest]
Thread overview: 24+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2020-09-20 9:57 [PATCH] stdlib: Fix data race in __run_exit_handlers Vitaly Buka via Libc-alpha
2020-09-20 12:09 ` Vitaly Buka via Libc-alpha
2020-09-20 20:41 ` Paul Pluzhnikov via Libc-alpha
2020-09-20 21:26 ` Vitaly Buka via Libc-alpha
2020-09-20 23:36 ` Vitaly Buka via Libc-alpha
2020-09-20 23:37 ` Vitaly Buka via Libc-alpha
2020-09-21 8:31 ` Vitaly Buka via Libc-alpha
2020-09-30 16:01 ` Joseph Myers
2021-04-17 16:16 ` Vitaly Buka via Libc-alpha
2021-04-17 17:11 ` Vitaly Buka via Libc-alpha
2021-04-17 17:13 ` Vitaly Buka via Libc-alpha
2021-04-17 17:22 ` Vitaly Buka via Libc-alpha
2021-04-17 18:01 ` Paul Pluzhnikov via Libc-alpha
2021-04-20 22:51 ` Vitaly Buka via Libc-alpha
2021-04-20 23:40 ` Paul Pluzhnikov via Libc-alpha
2021-04-26 19:20 ` Vitaly Buka via Libc-alpha
2021-04-26 19:23 ` Vitaly Buka via Libc-alpha
2021-04-26 19:27 ` Vitaly Buka via Libc-alpha
2021-05-13 13:15 ` Adhemerval Zanella via Libc-alpha
2021-05-14 6:50 ` Vitaly Buka via Libc-alpha
2021-04-17 17:36 ` Paul Pluzhnikov via Libc-alpha
2021-04-17 20:19 ` Florian Weimer
2021-04-19 2:48 ` Vitaly Buka via Libc-alpha
2021-04-19 2:57 ` Vitaly Buka via Libc-alpha
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).