fcntl36: fix 32-bit sporadic failures

No kernel bug, problem was found in my reproducer. Can anyone confirm this fixes 32-bit arm too?

Jan,

Running in a container, armhf userland, with arm64 kernel (real HW: hikey):

$ uname -a
Linux ltpfcntl36 4.16.0-2-arm64 #1 SMP Debian 4.16.12-1 (2018-05-27)
armv8l GNU/Linux

$ file ./fcntl36
./fcntl36: ELF 32-bit LSB pie executable ARM, EABI5 version 1 (SYSV),
dynamically linked, interpreter /lib/ld-linux-armhf.so.3, for
GNU/Linux 3.2.0,
BuildID[sha1]=074ba3622d50f62108260d2912ae5dc0e8d52d46, with
debug_info, not stripped

My un-patched binary shows me:

[pid 3501] fcntl64(11, F_OFD_SETLK, {l_type=F_UNLCK, l_whence=SEEK_SET, l_start=16384, l_len=4096} <unfinished ...>

no F_SETLK for this binary.

Running in a container, arm64 userland, with arm64 kernel (real HW: hikey):

$ file ./fcntl36
./fcntl36: ELF 64-bit LSB pie executable ARM aarch64, version 1
(SYSV), dynamically linked, interpreter /lib/ld-linux-aarch64.so.1,
for GNU/Linux 3.7.0,
BuildID[sha1]=8fb1d0e0e0d027fce5c8d11742ce5bdb7678749b, with
debug_info, not stripped

Same un-patched binary shows me:

[pid 483] fcntl(10, F_OFD_SETLKW, {l_type=F_UNLCK,
l_whence=SEEK_SET, l_start=13312, l_len=4096} <unfinished ...>

Do you want me to test patched LTP in arm32/64 to make sure they work

and you check if it fixes the issue for i686 on your side ?

Or should
I compile it some other way ?

Let me know what you need.

Rafael

Quoted 2 lines · Jan Stancek

No kernel bug, problem was found in my reproducer. Can anyone confirm this fixes 32-bit arm too?

fcntl36 testcase has been observed to sporadically fail, when running 32-bit user-space on 64-bit kernel.

Strace shows that region length is 0 for some commands:
fcntl64(6, F_SETLK, {l_type=F_UNLCK, l_whence=SEEK_SET, l_start=5120, l_len=0}

This is because testcase is passing struct flock64, but command is F_SETLK, not F_SETLK64.

So, kernel treats argument for POSIX command as 32-bit "struct flock":

static long do_compat_fcntl64(unsigned int fd, unsigned int cmd,
...

        case F_SETLK:
        case F_SETLKW:
                err = get_compat_flock(&flock, compat_ptr(arg));
                ...

in contrast to F_OFD_* commands, where argument is treated as "struct flock64":

        case F_SETLK64:
        case F_SETLKW64:
        case F_OFD_SETLK:
        case F_OFD_SETLKW:
                err = get_compat_flock64(&flock, compat_ptr(arg));
                ...

Switch argument of POSIX commands to 'struct flock' and leave it to glibc to pick correct syscall and command.

I tested fcntl36 and fcntl36_64 on x86_64 kernel (as 64bit + 32bit binary) and i386 kernel by running 100 iterations.

fcntl36 testcase has been observed to sporadically fail, when running 32-bit user-space on 64-bit kernel.

Strace shows that region length is 0 for some commands:
fcntl64(6, F_SETLK, {l_type=F_UNLCK, l_whence=SEEK_SET, l_start=5120, l_len=0}

This is because testcase is passing struct flock64, but command is F_SETLK, not F_SETLK64.

So, kernel treats argument for POSIX command as 32-bit "struct flock":

static long do_compat_fcntl64(unsigned int fd, unsigned int cmd,
...

        case F_SETLK:
        case F_SETLKW:
                err = get_compat_flock(&flock, compat_ptr(arg));
                ...

in contrast to F_OFD_* commands, where argument is treated as "struct flock64":

        case F_SETLK64:
        case F_SETLKW64:
        case F_OFD_SETLK:
        case F_OFD_SETLKW:
                err = get_compat_flock64(&flock, compat_ptr(arg));
                ...

Switch argument of POSIX commands to 'struct flock' and leave it to glibc to pick correct syscall and command.

I tested fcntl36 and fcntl36_64 on x86_64 kernel (as 64bit + 32bit binary) and i386 kernel by running 100 iterations.

Signed-off-by Jan Stancek <[email protected]>

---
No kernel bug, problem was found in my reproducer. Can anyone confirm this fixes 32-bit arm too?

 testcases/kernel/syscalls/fcntl/fcntl36.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

testcases/kernel/syscalls/fcntl/fcntl36.c +2 -3

index 3246d13892cd..81bd5a647e4c 100644
@@ -127,7 +127,7 @@ static void *fn_posix_w(void *arg)
        int fd = SAFE_OPEN(fname, O_RDWR);
        long wt = pa->cnt;
 
-       struct flock64 lck = {
+       struct flock lck = {
                .l_whence = SEEK_SET,
                .l_start  = pa->offset,
                .l_len    = pa->length,

@@ -227,7 +227,7 @@ static void *fn_posix_r(void *arg)
        int i;
        int fd = SAFE_OPEN(fname, O_RDWR);
 
-       struct flock64 lck = {
+       struct flock lck = {
                .l_whence = SEEK_SET,
                .l_start  = pa->offset,
                .l_len    = pa->length,
--

1.8.3.1

Quoted 31 lines

fcntl36 testcase has been observed to sporadically fail, when running 32-bit user-space on 64-bit kernel.

Strace shows that region length is 0 for some commands:
fcntl64(6, F_SETLK, {l_type=F_UNLCK, l_whence=SEEK_SET, l_start=5120, l_len=0}

This is because testcase is passing struct flock64, but command is F_SETLK, not F_SETLK64.

So, kernel treats argument for POSIX command as 32-bit "struct flock":

static long do_compat_fcntl64(unsigned int fd, unsigned int cmd,
...

        case F_SETLK:
        case F_SETLKW:
                err = get_compat_flock(&flock, compat_ptr(arg));
                ...

in contrast to F_OFD_* commands, where argument is treated as "struct flock64":

        case F_SETLK64:
        case F_SETLKW64:
        case F_OFD_SETLK:
        case F_OFD_SETLKW:
                err = get_compat_flock64(&flock, compat_ptr(arg));
                ...

Switch argument of POSIX commands to 'struct flock' and leave it to glibc to pick correct syscall and command.

Quoted 4 lines

I tested fcntl36 and fcntl36_64 on x86_64 kernel (as 64bit + 32bit binary) and i386 kernel by running 100 iterations.

Signed-off-by Jan Stancek <[email protected]>

Tested-by Naresh Kamboju <[email protected]>