Skip to contentSkip to conversation

a question regarding sys_poll() on x86_64 via tha ia32 layer

DISCUSSION Thomas Meyer 3 participants 5 messages Sep 15, 2011 VFS

lore
Share & tools
lore
#0 lore raw

cc'ed some people on this assumed bug.

Am Dienstag, den 13.09.2011, 14:04 +0200 schrieb Thomas Meyer:

Quoted 46 lines

Hello,

the ia32 poll system call is routed through the "standard" function sys_poll().

This function is defined as:

SYSCALL_DEFINE3(poll, struct pollfd __user *, ufds, unsigned int, nfds,
long, timeout_msecs)

in fs/select.c

timeout_msecs is of type long which is AFAIK is 4 bytes on x86 and 8 bytes on x86_64.

the test for sign (i.e. < 0) in the objdump is done against the 64 bit register (here %rbx):

ffffffff811313e0 <sys_poll>:

ffffffff811313e0:       55                      push   %rbp
ffffffff811313e1:       48 89 e5                mov    %rsp,%rbp
ffffffff811313e4:       48 83 ec 30             sub    $0x30,%rsp
ffffffff811313e8:       48 89 5d e8             mov    %rbx,-0x18(%rbp)
ffffffff811313ec:       48 89 d3                mov    %rdx,%rbx
ffffffff811313ef:       31 d2                   xor    %edx,%edx
ffffffff811313f1:       48 85 db                test   %rbx,%rbx
ffffffff811313f4:       4c 89 65 f0             mov    %r12,-0x10(%rbp)
ffffffff811313f8:       4c 89 6d f8             mov    %r13,-0x8(%rbp)
ffffffff811313fc:       41 89 f4                mov    %esi,%r12d
ffffffff811313ff:       49 89 fd                mov    %rdi,%r13
ffffffff81131402:       78 42                   js     ffffffff81131446 <sys_poll+0x66>

on an x86 kernel the test is done against %ebx

so when the system call is called with %rbx = 00000000ffffffff (i.e. -1 from %ebx) on an x86_64 kernel via the ia32 layer the test for sign will fail and the timer will be set.

btw. <sys/poll.h> seems to define the function as

extern int poll (struct pollfd *__fds, nfds_t __nfds, int __timeout);

what am I overloking?

mfg
thomas

#1 lore raw
Quoted 50 lines · Thomas Meyer

cc'ed some people on this assumed bug.

Am Dienstag, den 13.09.2011, 14:04 +0200 schrieb Thomas Meyer:

Quoted 46 lines

Hello,

the ia32 poll system call is routed through the "standard" function sys_poll().

This function is defined as:

SYSCALL_DEFINE3(poll, struct pollfd __user *, ufds, unsigned int, nfds,
long, timeout_msecs)

in fs/select.c

timeout_msecs is of type long which is AFAIK is 4 bytes on x86 and 8 bytes on x86_64.

the test for sign (i.e. < 0) in the objdump is done against the 64 bit register (here %rbx):

ffffffff811313e0 <sys_poll>:

ffffffff811313e0:       55                      push   %rbp
ffffffff811313e1:       48 89 e5                mov    %rsp,%rbp
ffffffff811313e4:       48 83 ec 30             sub    $0x30,%rsp
ffffffff811313e8:       48 89 5d e8             mov    %rbx,-0x18(%rbp)
ffffffff811313ec:       48 89 d3                mov    %rdx,%rbx
ffffffff811313ef:       31 d2                   xor    %edx,%edx
ffffffff811313f1:       48 85 db                test   %rbx,%rbx
ffffffff811313f4:       4c 89 65 f0             mov    %r12,-0x10(%rbp)
ffffffff811313f8:       4c 89 6d f8             mov    %r13,-0x8(%rbp)
ffffffff811313fc:       41 89 f4                mov    %esi,%r12d
ffffffff811313ff:       49 89 fd                mov    %rdi,%r13
ffffffff81131402:       78 42                   js     ffffffff81131446 <sys_poll+0x66>

on an x86 kernel the test is done against %ebx

so when the system call is called with %rbx = 00000000ffffffff (i.e. -1 from %ebx) on an x86_64 kernel via the ia32 layer the test for sign will fail and the timer will be set.

btw. <sys/poll.h> seems to define the function as

extern int poll (struct pollfd *__fds, nfds_t __nfds, int __timeout);

what am I overloking?

mfg
thomas

Its a plain bug, please submit a formal patch.

Probably not noticed because timer is set to more than 24 days.

fs/select.c +1 -1
index d33418f..e782258 100644
@@ -912,7 +912,7 @@ static long do_restart_poll(struct restart_block *restart_block)
 }
 
 SYSCALL_DEFINE3(poll, struct pollfd __user *, ufds, unsigned int, nfds,
-		long, timeout_msecs)
+		int, timeout_msecs)
 {
 	struct timespec end_time, *to = NULL;
 	int ret;
#2 lore raw

Am Donnerstag, den 15.09.2011, 14:23 +0200 schrieb Eric Dumazet:

Quoted 4 lines

Its a plain bug, please submit a formal patch.

Probably not noticed because timer is set to more than 24 days.

okay. I think a similar problem exists for sys_truncate() in ia32 layer on x86_64:

the ia32 syscall is directly wired to sys_truncate(char *, long).

signed long will use %rbx on x86_64 and %ebx on x86.

call sys_truncate on x86_64 from x86 userspace will handle length 0xffffffff not as -1.

maybe an easy fix would be to change long to unsigned long for sys_truncate(), like sys_ftruncate(). what do you think?

with kind regards
thomas

#3 lore raw
Quoted 11 lines · Thomas Meyer

Am Donnerstag, den 15.09.2011, 14:23 +0200 schrieb Eric Dumazet:

Quoted 4 lines

Its a plain bug, please submit a formal patch.

Probably not noticed because timer is set to more than 24 days.

okay. I think a similar problem exists for sys_truncate() in ia32 layer on x86_64:

the ia32 syscall is directly wired to sys_truncate(char *, long).

Quoted 7 lines

signed long will use %rbx on x86_64 and %ebx on x86.

call sys_truncate on x86_64 from x86 userspace will handle length 0xffffffff not as -1.

maybe an easy fix would be to change long to unsigned long for sys_truncate(), like sys_ftruncate(). what do you think?

I think you should take a look at fs/compat.c

#4 lore raw
In reply to Eric Dumazet

fs/select.c static long do_restart_poll(struct restart_block *restart_block) · line 915

Quoted 4 lines · Eric Dumazet

SYSCALL_DEFINE3(poll, struct pollfd __user *, ufds, unsigned int, nfds,

fs/select.c +1 -1
-		long, timeout_msecs)
+		int, timeout_msecs)

The preferred fix is normally to write a compat wrapper that sign extends.

Maybe someone on 64bit really wants to wait for a long time, but not forever.

-Andi

Signature · 1 line
[email protected] -- Speaking for myself only