While implementing rseq selftest for s390 a glibc problem with tls variables alignment has been discovered. It turned out to be a general problem affecting several architectures. The bug opened for this problem:
https://sourceware.org/bugzilla/show_bug.cgi?id=23403
There is no fix yet. On s390 __rseq_abi ends up aligned to 0x10 instead of 0x20 which makes rseq selftest fail every time.
The change proposed adds __rseq_abi misalignment check, produces user friendly message and skips the test.
Vasily Gorbik (1): rseq/selftests: add __rseq_abi misalignment check
tools/testing/selftests/rseq/rseq.c | 19 +++++++++++++++++++ .../testing/selftests/rseq/run_param_test.sh | 4 ++-- 2 files changed, 21 insertions(+), 2 deletions(-)
The kernel rseq syscall expects that struct rseq is 32 bytes aligned and returns EINVAL otherwise. Even though __rseq_abi is declared as static and proper aligned attribute is present __rseq_abi is a part of thread local storage. It turns out that on some platforms TLS itself is not properly aligned (at least for threads created), which is a glibc nptl bug and should be eventually fixed and backported. But in a meanwhile add __rseq_abi misalignment check, which would detect this situation and skip rseq test with some user friendly message.
glibc bug: https://sourceware.org/bugzilla/show_bug.cgi?id=23403
Signed-off-by: Vasily Gorbik gor@linux.ibm.com --- tools/testing/selftests/rseq/rseq.c | 19 +++++++++++++++++++ .../testing/selftests/rseq/run_param_test.sh | 4 ++-- 2 files changed, 21 insertions(+), 2 deletions(-)
diff --git a/tools/testing/selftests/rseq/rseq.c b/tools/testing/selftests/rseq/rseq.c index 4847e97ed049..50b13318395a 100644 --- a/tools/testing/selftests/rseq/rseq.c +++ b/tools/testing/selftests/rseq/rseq.c @@ -64,6 +64,23 @@ static int sys_rseq(volatile struct rseq *rseq_abi, uint32_t rseq_len, return syscall(__NR_rseq, rseq_abi, rseq_len, flags, sig); }
+/* + * rseq syscall might fail on some platforms due to wrong alignment of TLS + * variables: + * https://sourceware.org/bugzilla/show_bug.cgi?id=23403 + * + * check if glibc bug is present and skip the test in this case + */ +static void assert_rseq_abi_aligned() +{ + if ((unsigned long)&__rseq_abi & (__alignof__(__rseq_abi) - 1)) { + fputs("__rseq_abi is not properly aligned, which is a known \n" + "glibc nptl bug (https://sourceware.org/bugzilla/show_bug.cgi?id=23403).%5Cn" + "You need a fixed version of glibc to run this test.\n", stderr); + exit(4); /* skip this test */ + } +} + int rseq_register_current_thread(void) { int rc, ret = 0; @@ -72,6 +89,7 @@ int rseq_register_current_thread(void) signal_off_save(&oldset); if (refcount++) goto end; + assert_rseq_abi_aligned(); rc = sys_rseq(&__rseq_abi, sizeof(struct rseq), 0, RSEQ_SIG); if (!rc) { assert(rseq_current_cpu_raw() >= 0); @@ -94,6 +112,7 @@ int rseq_unregister_current_thread(void) signal_off_save(&oldset); if (--refcount) goto end; + assert_rseq_abi_aligned(); rc = sys_rseq(&__rseq_abi, sizeof(struct rseq), RSEQ_FLAG_UNREGISTER, RSEQ_SIG); if (!rc) diff --git a/tools/testing/selftests/rseq/run_param_test.sh b/tools/testing/selftests/rseq/run_param_test.sh index 3acd6d75ff9f..56caf5e3de3e 100755 --- a/tools/testing/selftests/rseq/run_param_test.sh +++ b/tools/testing/selftests/rseq/run_param_test.sh @@ -34,9 +34,9 @@ function do_tests() local i=0 while [ "$i" -lt "${#TEST_LIST[@]}" ]; do echo "Running test ${TEST_NAME[$i]}" - ./param_test ${TEST_LIST[$i]} -r ${REPS} ${@} ${EXTRA_ARGS} || exit 1 + ./param_test ${TEST_LIST[$i]} -r ${REPS} ${@} ${EXTRA_ARGS} || exit $? echo "Running compare-twice test ${TEST_NAME[$i]}" - ./param_test_compare_twice ${TEST_LIST[$i]} -r ${REPS} ${@} ${EXTRA_ARGS} || exit 1 + ./param_test_compare_twice ${TEST_LIST[$i]} -r ${REPS} ${@} ${EXTRA_ARGS} || exit $? let "i++" done }
The kernel rseq syscall expects that struct rseq is 32 bytes aligned and returns EINVAL otherwise. Even though __rseq_abi is declared as static and proper aligned attribute is present __rseq_abi is a part of thread local storage. It turns out that on some platforms TLS itself is not properly aligned (at least for threads created), which is a glibc nptl bug and should be eventually fixed and backported. But in a meanwhile add __rseq_abi misalignment check, which would detect this situation and skip rseq test with some user friendly message.
glibc bug: https://sourceware.org/bugzilla/show_bug.cgi?id=23403
Signed-off-by: Vasily Gorbik gor@linux.ibm.com --- tools/testing/selftests/rseq/rseq.c | 20 +++++++++++++++++++ .../testing/selftests/rseq/run_param_test.sh | 4 ++-- 2 files changed, 22 insertions(+), 2 deletions(-)
diff --git a/tools/testing/selftests/rseq/rseq.c b/tools/testing/selftests/rseq/rseq.c index 4847e97ed049..3de7f0e7442e 100644 --- a/tools/testing/selftests/rseq/rseq.c +++ b/tools/testing/selftests/rseq/rseq.c @@ -64,6 +64,24 @@ static int sys_rseq(volatile struct rseq *rseq_abi, uint32_t rseq_len, return syscall(__NR_rseq, rseq_abi, rseq_len, flags, sig); }
+/* + * rseq syscall might fail on some platforms due to wrong alignment of TLS + * variables: + * https://sourceware.org/bugzilla/show_bug.cgi?id=23403 + * + * check if glibc bug is present and skip the test in this case + */ +static void assert_rseq_abi_aligned(void) +{ + if ((unsigned long)&__rseq_abi & (__alignof__(__rseq_abi) - 1)) { + fputs("__rseq_abi is not properly aligned, which is a known\n" + "glibc nptl bug (https://sourceware.org/bugzilla/show_bug.cgi?id=23403).%5Cn" + "You need a fixed version of glibc to run this test.\n", + stderr); + exit(4); /* skip this test */ + } +} + int rseq_register_current_thread(void) { int rc, ret = 0; @@ -72,6 +90,7 @@ int rseq_register_current_thread(void) signal_off_save(&oldset); if (refcount++) goto end; + assert_rseq_abi_aligned(); rc = sys_rseq(&__rseq_abi, sizeof(struct rseq), 0, RSEQ_SIG); if (!rc) { assert(rseq_current_cpu_raw() >= 0); @@ -94,6 +113,7 @@ int rseq_unregister_current_thread(void) signal_off_save(&oldset); if (--refcount) goto end; + assert_rseq_abi_aligned(); rc = sys_rseq(&__rseq_abi, sizeof(struct rseq), RSEQ_FLAG_UNREGISTER, RSEQ_SIG); if (!rc) diff --git a/tools/testing/selftests/rseq/run_param_test.sh b/tools/testing/selftests/rseq/run_param_test.sh index 3acd6d75ff9f..56caf5e3de3e 100755 --- a/tools/testing/selftests/rseq/run_param_test.sh +++ b/tools/testing/selftests/rseq/run_param_test.sh @@ -34,9 +34,9 @@ function do_tests() local i=0 while [ "$i" -lt "${#TEST_LIST[@]}" ]; do echo "Running test ${TEST_NAME[$i]}" - ./param_test ${TEST_LIST[$i]} -r ${REPS} ${@} ${EXTRA_ARGS} || exit 1 + ./param_test ${TEST_LIST[$i]} -r ${REPS} ${@} ${EXTRA_ARGS} || exit $? echo "Running compare-twice test ${TEST_NAME[$i]}" - ./param_test_compare_twice ${TEST_LIST[$i]} -r ${REPS} ${@} ${EXTRA_ARGS} || exit 1 + ./param_test_compare_twice ${TEST_LIST[$i]} -r ${REPS} ${@} ${EXTRA_ARGS} || exit $? let "i++" done }
On Mon, Aug 06, 2018 at 03:27:23PM +0200, Vasily Gorbik wrote:
The kernel rseq syscall expects that struct rseq is 32 bytes aligned and returns EINVAL otherwise. Even though __rseq_abi is declared as static and proper aligned attribute is present __rseq_abi is a part of thread local storage. It turns out that on some platforms TLS itself is not properly aligned (at least for threads created), which is a glibc nptl bug and should be eventually fixed and backported. But in a meanwhile add __rseq_abi misalignment check, which would detect this situation and skip rseq test with some user friendly message.
glibc bug: https://sourceware.org/bugzilla/show_bug.cgi?id=23403
Signed-off-by: Vasily Gorbik gor@linux.ibm.com
Acked-by: Paul E. McKenney paulmck@linux.vnet.ibm.com
tools/testing/selftests/rseq/rseq.c | 20 +++++++++++++++++++ .../testing/selftests/rseq/run_param_test.sh | 4 ++-- 2 files changed, 22 insertions(+), 2 deletions(-)
diff --git a/tools/testing/selftests/rseq/rseq.c b/tools/testing/selftests/rseq/rseq.c index 4847e97ed049..3de7f0e7442e 100644 --- a/tools/testing/selftests/rseq/rseq.c +++ b/tools/testing/selftests/rseq/rseq.c @@ -64,6 +64,24 @@ static int sys_rseq(volatile struct rseq *rseq_abi, uint32_t rseq_len, return syscall(__NR_rseq, rseq_abi, rseq_len, flags, sig); }
+/*
- rseq syscall might fail on some platforms due to wrong alignment of TLS
- variables:
- check if glibc bug is present and skip the test in this case
- */
+static void assert_rseq_abi_aligned(void) +{
- if ((unsigned long)&__rseq_abi & (__alignof__(__rseq_abi) - 1)) {
fputs("__rseq_abi is not properly aligned, which is a known\n""glibc nptl bug (https://sourceware.org/bugzilla/show_bug.cgi?id=23403).\n""You need a fixed version of glibc to run this test.\n",stderr);exit(4); /* skip this test */- }
+}
int rseq_register_current_thread(void) { int rc, ret = 0; @@ -72,6 +90,7 @@ int rseq_register_current_thread(void) signal_off_save(&oldset); if (refcount++) goto end;
- assert_rseq_abi_aligned(); rc = sys_rseq(&__rseq_abi, sizeof(struct rseq), 0, RSEQ_SIG); if (!rc) { assert(rseq_current_cpu_raw() >= 0);
@@ -94,6 +113,7 @@ int rseq_unregister_current_thread(void) signal_off_save(&oldset); if (--refcount) goto end;
- assert_rseq_abi_aligned(); rc = sys_rseq(&__rseq_abi, sizeof(struct rseq), RSEQ_FLAG_UNREGISTER, RSEQ_SIG); if (!rc)
diff --git a/tools/testing/selftests/rseq/run_param_test.sh b/tools/testing/selftests/rseq/run_param_test.sh index 3acd6d75ff9f..56caf5e3de3e 100755 --- a/tools/testing/selftests/rseq/run_param_test.sh +++ b/tools/testing/selftests/rseq/run_param_test.sh @@ -34,9 +34,9 @@ function do_tests() local i=0 while [ "$i" -lt "${#TEST_LIST[@]}" ]; do echo "Running test ${TEST_NAME[$i]}"
./param_test ${TEST_LIST[$i]} -r ${REPS} ${@} ${EXTRA_ARGS} || exit 1
./param_test ${TEST_LIST[$i]} -r ${REPS} ${@} ${EXTRA_ARGS} || exit $?
./param_test_compare_twice ${TEST_LIST[$i]} -r ${REPS} ${@} ${EXTRA_ARGS} || exit 1
./param_test_compare_twice ${TEST_LIST[$i]} -r ${REPS} ${@} ${EXTRA_ARGS} || exit $?}
2.18.0.13.gd42ae10
-- To unsubscribe from this list: send the line "unsubscribe linux-kselftest" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
----- On Aug 6, 2018, at 7:47 AM, gor gor@linux.ibm.com wrote:
While implementing rseq selftest for s390 a glibc problem with tls variables alignment has been discovered. It turned out to be a general problem affecting several architectures. The bug opened for this problem:
https://sourceware.org/bugzilla/show_bug.cgi?id=23403
There is no fix yet. On s390 __rseq_abi ends up aligned to 0x10 instead of 0x20 which makes rseq selftest fail every time.
The change proposed adds __rseq_abi misalignment check, produces user friendly message and skips the test.
That's a very unfortunate situation. I'm concerned about adding glibc-specific error messages in rseq selftests though. I'm curious to hear what others think about this.
I would have thought simply improving rseq registration error handling from having the test program return nonzero to add a perror() in there would be a more generic way to handle this.
Regarding the message printed by your check: "you need a fixed version of glibc to run this test". I disagree with it. Someone can effectively run the test on a bogus glibc and it serves its purpose: it reports that glibc is buggy.
I would understand adding this kind of test in an user-facing application or library to detect bogus glibc (in fact I've used similar approaches in lttng-ust to detect bogus compilers), but why add this to skip a selftest program, which sole purpose is to test the stack underneath it ?
Thanks,
Mathieu
Vasily Gorbik (1): rseq/selftests: add __rseq_abi misalignment check
tools/testing/selftests/rseq/rseq.c | 19 +++++++++++++++++++ .../testing/selftests/rseq/run_param_test.sh | 4 ++-- 2 files changed, 21 insertions(+), 2 deletions(-)
-- 2.18.0.13.gd42ae10
linux-kselftest-mirror@lists.linaro.org
-
Mathieu Desnoyers -
Paul E. McKenney -
Vasily Gorbik