It was observed[1] on arm64 that __builtin_strlen led to an infinite loop in the get_size selftest. This is because __builtin_strlen (and other builtins) may sometimes result in a call to the C library function. The C library implementation of strlen uses an IFUNC resolver to load the most efficient strlen implementation for the underlying machine and hence has a PLT indirection even for static binaries. Because this binary avoids the C library startup routines, the PLT initialization never happens and hence the program gets stuck in an infinite loop.
On x86_64 the __builtin_strlen just happens to expand inline and avoid the call but that is not always guaranteed.
Further, while testing on x86_64 (Fedora 31), it was observed that the test also failed with a segfault inside write() because the generated code for the write function in glibc seems to access TLS before the syscall (probably due to the cancellation point check) and fails because TLS is not initialised.
To mitigate these problems, this patch reduces the interface with the C library to just the syscall function. The syscall function still sets errno on failure, which is undesirable but for now it only affects cases where syscalls fail.
[1] https://bugs.linaro.org/show_bug.cgi?id=5479
Signed-off-by: Siddhesh Poyarekar siddhesh@gotplt.org Reported-by: Masami Hiramatsu masami.hiramatsu@linaro.org --- tools/testing/selftests/size/get_size.c | 24 ++++++++++++++++++------ 1 file changed, 18 insertions(+), 6 deletions(-)
diff --git a/tools/testing/selftests/size/get_size.c b/tools/testing/selftests/size/get_size.c index d4b59ab979a0..f55943b6d1e2 100644 --- a/tools/testing/selftests/size/get_size.c +++ b/tools/testing/selftests/size/get_size.c @@ -12,23 +12,35 @@ * own execution. It also attempts to have as few dependencies * on kernel features as possible. * - * It should be statically linked, with startup libs avoided. - * It uses no library calls, and only the following 3 syscalls: + * It should be statically linked, with startup libs avoided. It uses + * no library calls except the syscall() function for the following 3 + * syscalls: * sysinfo(), write(), and _exit() * * For output, it avoids printf (which in some C libraries * has large external dependencies) by implementing it's own * number output and print routines, and using __builtin_strlen() + * + * The test may crash if any of the above syscalls fails because in some + * libc implementations (e.g. the GNU C Library) errno is saved in + * thread-local storage, which does not get initialized due to avoiding + * startup libs. */
#include <sys/sysinfo.h> #include <unistd.h> +#include <sys/syscall.h>
#define STDOUT_FILENO 1
static int print(const char *s) { - return write(STDOUT_FILENO, s, __builtin_strlen(s)); + size_t len = 0; + + while (s[len] != '\0') + len++; + + return syscall(SYS_write, STDOUT_FILENO, s, len); }
static inline char *num_to_str(unsigned long num, char *buf, int len) @@ -80,12 +92,12 @@ void _start(void) print("TAP version 13\n"); print("# Testing system size.\n");
- ccode = sysinfo(&info); + ccode = syscall(SYS_sysinfo, &info); if (ccode < 0) { print("not ok 1"); print(test_name); print(" ---\n reason: "could not get sysinfo"\n ...\n"); - _exit(ccode); + syscall(SYS_exit, ccode); } print("ok 1"); print(test_name); @@ -101,5 +113,5 @@ void _start(void) print(" ...\n"); print("1..1\n");
- _exit(0); + syscall(SYS_exit, 0); }
Hi,
[Cc: Tim Bird]
2020年1月14日(火) 1:43 Siddhesh Poyarekar siddhesh@gotplt.org:
It was observed[1] on arm64 that __builtin_strlen led to an infinite loop in the get_size selftest. This is because __builtin_strlen (and other builtins) may sometimes result in a call to the C library function. The C library implementation of strlen uses an IFUNC resolver to load the most efficient strlen implementation for the underlying machine and hence has a PLT indirection even for static binaries. Because this binary avoids the C library startup routines, the PLT initialization never happens and hence the program gets stuck in an infinite loop.
On x86_64 the __builtin_strlen just happens to expand inline and avoid the call but that is not always guaranteed.
Further, while testing on x86_64 (Fedora 31), it was observed that the test also failed with a segfault inside write() because the generated code for the write function in glibc seems to access TLS before the syscall (probably due to the cancellation point check) and fails because TLS is not initialised.
To mitigate these problems, this patch reduces the interface with the C library to just the syscall function. The syscall function still sets errno on failure, which is undesirable but for now it only affects cases where syscalls fail.
Thank you for the fix! I confirmed this fixes the issue.
---- root@devnote2:/opt/kselftest/size# ./get_size TAP version 13 # Testing system size. ok 1 get runtime memory use # System runtime memory report (units in Kilobytes): --- Total: 16085116 Free: 2042880 Buffer: 814052 In use: 13228184 ... 1..1 ----
Tested-by: Masami Hiramatsu masami.hiramatsu@linaro.org
Signed-off-by: Siddhesh Poyarekar siddhesh@gotplt.org Reported-by: Masami Hiramatsu masami.hiramatsu@linaro.org
tools/testing/selftests/size/get_size.c | 24 ++++++++++++++++++------ 1 file changed, 18 insertions(+), 6 deletions(-)
diff --git a/tools/testing/selftests/size/get_size.c b/tools/testing/selftests/size/get_size.c index d4b59ab979a0..f55943b6d1e2 100644 --- a/tools/testing/selftests/size/get_size.c +++ b/tools/testing/selftests/size/get_size.c @@ -12,23 +12,35 @@
- own execution. It also attempts to have as few dependencies
- on kernel features as possible.
- It should be statically linked, with startup libs avoided.
- It uses no library calls, and only the following 3 syscalls:
- It should be statically linked, with startup libs avoided. It uses
- no library calls except the syscall() function for the following 3
- syscalls:
- sysinfo(), write(), and _exit()
- For output, it avoids printf (which in some C libraries
- has large external dependencies) by implementing it's own
- number output and print routines, and using __builtin_strlen()
- The test may crash if any of the above syscalls fails because in some
- libc implementations (e.g. the GNU C Library) errno is saved in
- thread-local storage, which does not get initialized due to avoiding
*/
- startup libs.
#include <sys/sysinfo.h> #include <unistd.h> +#include <sys/syscall.h>
#define STDOUT_FILENO 1
static int print(const char *s) {
return write(STDOUT_FILENO, s, __builtin_strlen(s));
size_t len = 0;
while (s[len] != '\0')
len++;
return syscall(SYS_write, STDOUT_FILENO, s, len);
}
static inline char *num_to_str(unsigned long num, char *buf, int len) @@ -80,12 +92,12 @@ void _start(void) print("TAP version 13\n"); print("# Testing system size.\n");
ccode = sysinfo(&info);
ccode = syscall(SYS_sysinfo, &info); if (ccode < 0) { print("not ok 1"); print(test_name); print(" ---\n reason: \"could not get sysinfo\"\n ...\n");
_exit(ccode);
syscall(SYS_exit, ccode); } print("ok 1"); print(test_name);
@@ -101,5 +113,5 @@ void _start(void) print(" ...\n"); print("1..1\n");
_exit(0);
syscall(SYS_exit, 0);
}
2.24.1
-- Masami Hiramatsu
-----Original Message----- From: Masami Hiramatsu masami.hiramatsu@linaro.org Sent: Tuesday, January 14, 2020 8:14 PM To: Siddhesh Poyarekar siddhesh@gotplt.org Cc: linux-kselftest@vger.kernel.org; Shuah Khan shuah@kernel.org; Linux kernel mailing list linux-kernel@vger.kernel.org; Bird, Tim Tim.Bird@sony.com Subject: Re: [PATCH] kselftest: Minimise dependency of get_size on C library interfaces
Hi,
[Cc: Tim Bird]
2020年1月14日(火) 1:43 Siddhesh Poyarekar siddhesh@gotplt.org:
It was observed[1] on arm64 that __builtin_strlen led to an infinite loop in the get_size selftest. This is because __builtin_strlen (and other builtins) may sometimes result in a call to the C library function. The C library implementation of strlen uses an IFUNC resolver to load the most efficient strlen implementation for the underlying machine and hence has a PLT indirection even for static binaries. Because this binary avoids the C library startup routines, the PLT initialization never happens and hence the program gets stuck in an infinite loop.
On x86_64 the __builtin_strlen just happens to expand inline and avoid the call but that is not always guaranteed.
Further, while testing on x86_64 (Fedora 31), it was observed that the test also failed with a segfault inside write() because the generated code for the write function in glibc seems to access TLS before the syscall (probably due to the cancellation point check) and fails because TLS is not initialised.
To mitigate these problems, this patch reduces the interface with the C library to just the syscall function. The syscall function still sets errno on failure, which is undesirable but for now it only affects cases where syscalls fail.
Thank you for the fix! I confirmed this fixes the issue.
root@devnote2:/opt/kselftest/size# ./get_size TAP version 13 # Testing system size. ok 1 get runtime memory use
# System runtime memory report (units in Kilobytes):
Total: 16085116 Free: 2042880 Buffer: 814052 In use: 13228184 ...
1..1
Tested-by: Masami Hiramatsu masami.hiramatsu@linaro.org
Signed-off-by: Siddhesh Poyarekar siddhesh@gotplt.org Reported-by: Masami Hiramatsu masami.hiramatsu@linaro.org
tools/testing/selftests/size/get_size.c | 24 ++++++++++++++++++------ 1 file changed, 18 insertions(+), 6 deletions(-)
diff --git a/tools/testing/selftests/size/get_size.c b/tools/testing/selftests/size/get_size.c index d4b59ab979a0..f55943b6d1e2 100644 --- a/tools/testing/selftests/size/get_size.c +++ b/tools/testing/selftests/size/get_size.c @@ -12,23 +12,35 @@
- own execution. It also attempts to have as few dependencies
- on kernel features as possible.
- It should be statically linked, with startup libs avoided.
- It uses no library calls, and only the following 3 syscalls:
- It should be statically linked, with startup libs avoided. It uses
- no library calls except the syscall() function for the following 3
- syscalls:
- sysinfo(), write(), and _exit()
- For output, it avoids printf (which in some C libraries
- has large external dependencies) by implementing it's own
- number output and print routines, and using __builtin_strlen()
Since the code no longer uses __builtin_strlen(), this comment should change also, to say something like "and string length function.
- The test may crash if any of the above syscalls fails because in some
- libc implementations (e.g. the GNU C Library) errno is saved in
- thread-local storage, which does not get initialized due to avoiding
*/
- startup libs.
#include <sys/sysinfo.h> #include <unistd.h> +#include <sys/syscall.h>
#define STDOUT_FILENO 1
static int print(const char *s) {
return write(STDOUT_FILENO, s, __builtin_strlen(s));
size_t len = 0;
while (s[len] != '\0')
len++;
return syscall(SYS_write, STDOUT_FILENO, s, len);
}
static inline char *num_to_str(unsigned long num, char *buf, int len) @@ -80,12 +92,12 @@ void _start(void) print("TAP version 13\n"); print("# Testing system size.\n");
ccode = sysinfo(&info);
ccode = syscall(SYS_sysinfo, &info); if (ccode < 0) { print("not ok 1"); print(test_name); print(" ---\n reason: \"could not get sysinfo\"\n ...\n");
_exit(ccode);
syscall(SYS_exit, ccode); } print("ok 1"); print(test_name);
@@ -101,5 +113,5 @@ void _start(void) print(" ...\n"); print("1..1\n");
_exit(0);
syscall(SYS_exit, 0);
}
2.24.1
Thanks very much for this bug report and fix!!
Reviewed-by: Tim Bird tim.bird@sony.com
-- Tim
On 16/01/20 11:02 pm, Tim.Bird@sony.com wrote:
Since the code no longer uses __builtin_strlen(), this comment should change also, to say something like "and string length function.
Oops, I had made the change (to drop the __builtin_strlen mention since the length is just computed inline now) and somehow dropped it in some copy-paste confusion. Looks like a patchwork-bot applied my patch, so should I send an updated patch or just a comment fix follow-up?
Thanks, Siddhesh
linux-kselftest-mirror@lists.linaro.org