[lkml]   [2014]   [Jul]   [7]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectAw: Re: [PATCH] fix fanotify_mark() breakage on big endian 32bit kernel
Hi Heiko,

> On Fri, Jul 04, 2014 at 05:12:35PM +0200, Helge Deller wrote:
> > This patch affects big endian architectures only.
> >
> > On those with 32bit userspace and 64bit kernel (CONFIG_COMPAT=y) the
> > 64bit mask parameter is correctly constructed out of two 32bit values in
> > the compat_fanotify_mark() function and then passed as 64bit parameter
> > to the fanotify_mark() syscall.
> >
> > But for the CONFIG_COMPAT=n case (32bit kernel & userspace),
> > compat_fanotify_mark() isn't used and the fanotify_mark syscall implementation
> > is used directly. In that case the upper and lower 32 bits of the 64bit mask
> > parameter is still swapped on big endian machines and thus leads to
> > fanotify_mark failing with -EINVAL.
> Why do you think upper and lower 32 bits are swapped on big endian machines?

I assumed it, because I see this behaviour on parisc, and because of this commit
from you regarding the compat-case. I do recognize, that in this patch the u64 value
is constructed out of the two 32bit values to hand it over. So, this patch is OK.

commit 592f6b842f64e416c7598a1b97c649b34241e22d
Author: Heiko Carstens <>
Date: Mon Jan 27 17:07:19 2014 -0800

compat: fix sys_fanotify_mark

Commit 91c2e0bcae72 ("unify compat fanotify_mark(2), switch to
COMPAT_SYSCALL_DEFINE") added a new unified compat fanotify_mark syscall
to be used by all architectures.

Unfortunately the unified version merges the split mask parameter in a
wrong way: the lower and higher word got swapped.

This was discovered with glibc's tst-fanotify test case.

> > Here is a strace of the same 32bit executable (fanotify01 testcase from LTP):
> >
> > On a 64bit kernel it suceeds:
> > syscall_322(0, 0, 0x3, 0x3, 0x266c8, 0x1) = 0x3
> > syscall_323(0x3, 0x1, 0, 0x3b, 0xffffff9c, 0x266c8) = 0
> >
> > On a 32bit kernel it fails:
> > syscall_322(0, 0, 0x3, 0x3, 0x266c8, 0x1) = 0x3
> > syscall_323(0x3, 0x1, 0, 0x3b, 0xffffff9c, 0x266c8) = -1 (errno 22)
> So "0" and "0x3b" together should be the 64 bit "0x3b" mask, this looks just
> fine.
> > diff --git a/fs/notify/fanotify/fanotify_user.c b/fs/notify/fanotify/fanotify_user.c
> > index 3fdc8a3..374261c 100644
> > --- a/fs/notify/fanotify/fanotify_user.c
> > +++ b/fs/notify/fanotify/fanotify_user.c
> > @@ -787,6 +787,10 @@ SYSCALL_DEFINE5(fanotify_mark, int, fanotify_fd, unsigned int, flags,
> > struct path path;
> > int ret;
> >
> > +#if defined(__BIG_ENDIAN) && !defined(CONFIG_64BIT)
> > + mask = (mask << 32) | (mask >> 32);
> > +#endif
> > +
> > pr_debug("%s: fanotify_fd=%d flags=%x dfd=%d pathname=%p mask=%llx\n",
> > __func__, fanotify_fd, flags, dfd, pathname, mask);
> Did you activate this pr_debug()? I'm really wondering what the output looks
> like on your machine.

Just tested it.
On 3.16.0-rc4-32bit (without my patch)
syscall_323(0x3, 0x1, 0, 0x3b, 0xffffff9c, 0x266c8) = -1 (errno 22)
SYSC_fanotify_mark: fanotify_fd=3 flags=1 dfd=-100 pathname=000266c8 mask=3b00000000

and on 3.16.0-rc4-32bit+ (*with* my patch, same executable file):
syscall_323(0x3, 0x1, 0, 0x3b, 0xffffff9c, 0x266c8) = 0
SYSC_fanotify_mark: fanotify_fd=3 flags=1 dfd=-100 pathname=000266c8 mask=3b

So, my patch works as expected.

The Linux Test Project (LTP) uses in testcases/kernel/syscalls/fanotify/fanotify.h this coding, which is IMHO
correct as it would break your commit 592f6b842f64e416c7598a1b97c649b34241e22d otherwise:
long myfanotify_mark(int fd, unsigned int flags, uint64_t mask,
int dfd, const char *pathname)
#if LTP_USE_64_ABI
return ltp_syscall(__NR_fanotify_mark, fd, flags, mask, dfd, pathname);
return ltp_syscall(__NR_fanotify_mark, fd, flags,
__LONG_LONG_PAIR((unsigned long) (mask >> 32),
(unsigned long) mask),
dfd, (unsigned long) pathname);

with __LONG_LONG_PAIR() defined in /usr/include/endian.h:
# define __LONG_LONG_PAIR(HI, LO) LO, HI
# define __LONG_LONG_PAIR(HI, LO) HI, LO

and in glibc sysdeps/unix/sysv/linux/sys/fanotify.h I see:
extern int fanotify_mark (int __fanotify_fd, unsigned int __flags, uint64_t __mask, int __dfd, const char *__pathname);
sysdeps/unix/sysv/linux/s390/s390-32/syscalls.list:fanotify_mark EXTRA fanotify_mark i:iiiiis fanotify_mark

> At least an s390 the C ABI defines that 64 bit values are split into an
> even odd register pair, where the most significant bits are in the even numbered
> register.

and Dave wrote for hppa:
> In GCC, we typically have an odd even register pair to hold 64-bit
> values as register r0 is not usable.

This seems different.

> So for sys_fanotify_mark everything is fine on s390, and probably most other
> architectures as well. Having a 64 bit syscall parameter indeed does work,
> if all the architecture specific details have been correctly considered.

I think this is the problem!
For parisc the architecture specifc details have not been considered correctly.
I tried this test:

static int low32, high32;
SYSCALL_DEFINE5(fanotify_mark_test, int, fanotify_fd, unsigned int, flags,
__u64, mask, int, dfd, const char __user *, pathname)
low32 = (int) mask;
high32 = (int) (mask >> 32);

and got:

.section .text.SyS_fanotify_mark_test,"ax",@progbits
.align 4
.globl SyS_fanotify_mark_test
.type SyS_fanotify_mark_test, @function
copy %r3,%r1
copy %r30,%r3
stwm %r1,64(%r30)
addil LR'low32-$global$,%r27
ldi 0,%r28
stw %r24,RR'low32-$global$(%r1)
addil LR'high32-$global$,%r27
stw %r23,RR'high32-$global$(%r1)
ldo 64(%r3),%r30
bv %r0(%r2)
ldwm -64(%r30),%r3

So on hppa r26 is fanotify_fd, %r25 is flags, %r24/%r23 is lower/higher 32bits of mask.
For the mask parameter this is different to what the __LONG_LONG_PAIR() marcro
would hand over to the syscall (which would be %r24/%r23 as higher/lower 32bits).

So, the problem is the usage of __u64 in the 32bit API. It has to be handled architecture-specific.
It seems to work for little-endian machines, and probably (by luck?!?) for s390, but I'm not sure if
it maybe breaks (like on parisc) on other arches, e.g. what about sparc?

For parisc I can work around that problem in the architecture-specifc coding, but I still
think using __64 here is wrong and just may lead to such bugs.


 \ /
  Last update: 2014-07-07 16:41    [W:0.087 / U:1.100 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site