set_bit() is broken on i386?
/*
* setbit.c -- test the Linux set_bit() function
*
* Compare the output of this program with and without the
* -finline-functions option to GCC.
*
* If they are not the same, set_bit is broken.
*
* Result on i386 with gcc 3.3.2 (Fedora Core 2):
*
* [me@d2 t]$ gcc -O2 -o setbit.ex setbit.c ; ./setbit.ex
* 00010001
* [me@d2 t]$ gcc -O2 -o setbit.ex -finline-functions setbit.c ; ./setbit.ex
* 00000001
*/
#define _GNU_SOURCE
#include <stdio.h>
#include <stdlib.h>
#define inline __attribute__((always_inline))
/*
* From 2.6.15/include/asm-i386/bitops.h -- needs a memory clobber? */
#define ADDR (*(volatile long *) addr)
static inline void set_bit(int nr, volatile unsigned long * addr)
{
__asm__ __volatile__( "lock ; "
"btsl %1,%0"
:"=m" (ADDR)
:"Ir" (nr));
}unsigned long b[2];
int main(int argc, char * const argv[])
{
b[1] = 0x1;
set_bit(12 * sizeof(unsigned long), b);
printf("%08x\n", b[1]);
return 0;
}Signature · 2 lines
Chuck Currently reading: _Sun Of Suns_ by Karl Schroeder
Quoted 8 lines · Chuck Ebbert
#define ADDR (*(volatile long *) addr)
static inline voidset_bit(int nr, volatile unsigned long * addr){ __asm__ __volatile__( "lock ; " "btsl %1,%0" :"=m" (ADDR) :"Ir" (nr)); }
The asm needs a memory clobber in order to avoid reordering with the assignment to b[1]:
static inline void set_bit(int nr, volatile unsigned long * addr)
{
__asm__ __volatile__( "lock ; "
"btsl %1,%0"
:"=m" (ADDR)
:"Ir" (nr)
: "memory");
}This works consistently correctly for me.
Cheers,
Trond
Quoted 6 lines · Chuck Ebbert
/*
*setbit.c-- test the Linuxset_bit()function *
* Compare the output of this program with and without the
* -finline-functions option to GCC.
*
Quoted 8 lines
* If they are not the same,
set_bitis broken.
*
* Result on i386 with gcc 3.3.2 (Fedora Core 2):
*
* [me@d2 t]$ gcc -O2 -o setbit.exsetbit.c; ./setbit.ex
* 00010001
* [me@d2 t]$ gcc -O2 -o setbit.ex -finline-functionssetbit.c; ./setbit.ex
* 00000001
It certainly seems to be gcc version dependent (and I'd not be surprised if the exact combo of -O2, -Os, and -mfoo and -fwhatever flags as well). Trond is probably right that a memory clobber will force gcc to DTIT (Do The Intended Thing, which may be different from a DTRT) regardless of what gcc's code generator decides to do....
% gcc -v
Using built-in specs.
Target: i386-redhat-linux
Configured with: ../configure --prefix=/usr --mandir=/usr/share/man --infodir=/usr/share/info --enable-shared --enable-threads=posix --enable-checking=release --with-system-zlib --enable-__cxa_atexit --disable-libunwind-exceptions --enable-libgcj-multifile --enable-languages=c,c++,objc,obj-c++,java,fortran,ada --enable-java-awt=gtk --disable-dssi --with-java-home=/usr/lib/jvm/java-1.4.2-gcj-1.4.2.0/jre --host=i386-redhat-linux Thread model: posix
gcc version 4.1.0 20060117 (Red Hat 4.1.0-0.15)
% gcc -O2 -o setbit.ex setbit.c ; ./setbit.ex
00000001
% gcc -O2 -o setbit.ex -finline-functions setbit.c ; ./setbit.ex
00000001
Fedora Core -devel tree as of this morning (so sort-of FC5 test2).
Quoted 36 lines · Chuck Ebbert
/*
*setbit.c-- test the Linuxset_bit()function
*
* Compare the output of this program with and without the
* -finline-functions option to GCC.
*
* If they are not the same,set_bitis broken.
*
* Result on i386 with gcc 3.3.2 (Fedora Core 2):
*
* [me@d2 t]$ gcc -O2 -o setbit.exsetbit.c; ./setbit.ex
* 00010001
* [me@d2 t]$ gcc -O2 -o setbit.ex -finline-functionssetbit.c; ./setbit.ex
* 00000001
*/
#define_GNU_SOURCE
#include <stdio.h>
#include <stdlib.h>#define inline __attribute__((always_inline))
/*
* From2.6.15/include/asm-i386/bitops.h-- needs a memory clobber? */
#define ADDR (*(volatile long *) addr)
static inline voidset_bit(int nr, volatile unsigned long * addr){ __asm__ __volatile__( "lock ; " "btsl %1,%0" :"=m" (ADDR) :"Ir" (nr)); }unsigned long b[2];
Quoted 7 lines
int
main(int argc, char * const argv[]){ b[1] = 0x1; set_bit(12 * sizeof(unsigned long), b); printf("%08x\n", b[1]); return 0; }
Yes, that's a bit of a trap. Seems that any place where the compiler "knows" the value of b[1], it'll cache it for the printf() push.
It's actually quite rare for the kernel to manipulate an unsigned long as both a target of the bitops functions and as a plain old unsigned long.
We have smp_mb__before_clear_bit() and smp_mb__after_clear_bit(), which need to be explicity used in the special case where we're using clear_bit() in still-rather-poorly-defined circumstances.
But what you have here is misbehaviour on UP, with set_bit(), due to compiler actions, which is quite unrelated.
So I guess we could add the rule "you need to use barrier() before treating a bitop-manipulated word as a word" or we can go stick all the barriers into bitops.h.
<does that>
OK, adding nine "memory"s to i386/bitops.h increases my allnoconfig vmlinux's text from 774636 bytes to 774928 (292 bytes more).
But the problem is that we've gone and altered the foo_bit() _semantics_ on x86. So people can now go and write code which will only work properly on x86 (even more than at present).
It'll be hard to find all the code which treats bitops-targets as scalars. It would be easier if we'd defined bitops as operating on a bitop_target_t or whatever, but we didn't do that.
Personally, I'd like to stick the barriers in there and sleep soundly. But we do need to make architectures all behave the same way (and define that way clearly), and the impact of the new semantics might be worse on some other architectures.
Quoted 11 lines · Trond Myklebust
Quoted 8 lines · Chuck Ebbert
#define ADDR (*(volatile long *) addr)
static inline voidset_bit(int nr, volatile unsigned long * addr){ __asm__ __volatile__( "lock ; " "btsl %1,%0" :"=m" (ADDR) :"Ir" (nr)); }
Quoted 2 lines
The asm needs a memory clobber in order to avoid reordering with the assignment to b[1]:
Check out 2.6.16-rc1, this has already been fixed.
Andreas.
Signature · 4 lines
Andreas Schwab, SuSE Labs, [email protected] SuSE Linux Products GmbH, Maxfeldstraße 5, 90409 Nürnberg, Germany PGP key fingerprint = 58CA 54C7 6D53 942B 1756 01D3 44D5 214B 8276 4ED5 "And now for something completely different."
Quoted 33 lines · On Fri, 20 Jan 2006, [email protected] wrote:
Quoted 6 lines · Chuck Ebbert
/*
*setbit.c-- test the Linuxset_bit()function *
* Compare the output of this program with and without the
* -finline-functions option to GCC.
*Quoted 8 lines
* If they are not the same,
set_bitis broken.
*
* Result on i386 with gcc 3.3.2 (Fedora Core 2):
*
* [me@d2 t]$ gcc -O2 -o setbit.exsetbit.c; ./setbit.ex
* 00010001
* [me@d2 t]$ gcc -O2 -o setbit.ex -finline-functionssetbit.c; ./setbit.ex
* 00000001It certainly seems to be gcc version dependent (and I'd not be surprised if the exact combo of -O2, -Os, and -mfoo and -fwhatever flags as well). Trond is probably right that a memory clobber will force gcc to DTIT (Do The Intended Thing, which may be different from a DTRT) regardless of what gcc's code generator decides to do....
% gcc -v
Using built-in specs.
Target: i386-redhat-linux
Configured with: ../configure --prefix=/usr --mandir=/usr/share/man --infodir=/usr/share/info --enable-shared --enable-threads=posix --enable-checking=release --with-system-zlib --enable-__cxa_atexit --disable-libunwind-exceptions --enable-libgcj-multifile --enable-languages=c,c++,objc,obj-c++,java,fortran,ada --enable-java-awt=gtk --disable-dssi --with-java-home=/usr/lib/jvm/java-1.4.2-gcj-1.4.2.0/jre --host=i386-redhat-linux Thread model: posix
gcc version 4.1.0 20060117 (Red Hat 4.1.0-0.15)
% gcc -O2 -o setbit.exsetbit.c; ./setbit.ex
00000001
% gcc -O2 -o setbit.ex -finline-functionssetbit.c; ./setbit.ex
00000001
Quoted 1 line
Fedora Core -devel tree as of this morning (so sort-of FC5 test2).
This is what I am getting under Gentoo with different gcc's:
$ export CC=gcc-3.4.5; $CC --version; rm ./setbit.ex; $CC -O2 -o setbit.ex setbit.c ; ./setbit.ex; rm ./setbit.ex; $CC -O2 -o setbit.ex -finline-functions setbit.c ; ./setbit.ex gcc-3.4.5 (GCC) 3.4.5 (Gentoo 3.4.5, ssp-3.4.5-1.0, pie-8.7.9) Copyright (C) 2004 Free Software Foundation, Inc. This is free software; see the source for copying conditions. There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
00000001
00000001
$ export CC=gcc-4.0.2; $CC --version; rm ./setbit.ex; $CC -O2 -o setbit.ex setbit.c ; ./setbit.ex; rm ./setbit.ex; $CC -O2 -o setbit.ex -finline-functions setbit.c ; ./setbit.ex gcc-4.0.2 (GCC) 4.0.2 (Gentoo 4.0.2-r3, pie-8.7.8) Copyright (C) 2005 Free Software Foundation, Inc. This is free software; see the source for copying conditions. There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
00000001
00000001
$ export CC=gcc-4.1.0-beta20060113; $CC --version; rm ./setbit.ex; $CC -O2 -o setbit.ex setbit.c ; ./setbit.ex; rm ./setbit.ex; $CC -O2 -o setbit.ex -finline-functions setbit.c ; ./setbit.ex gcc-4.1.0-beta20060113 (GCC) 4.1.0-beta20060113 (prerelease) Copyright (C) 2005 Free Software Foundation, Inc. This is free software; see the source for copying conditions. There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
00000001
00000001
And on some really old and not upgraded server (with Gentoo too of course):
$ export CC=gcc32; $CC --version; rm ./setbit.ex; $CC -O2 -o setbit.ex setbit.c ; ./setbit.ex; rm ./setbit.ex; $CC -O2 -o setbit.ex -finline-functions setbit.c ; ./setbit.ex gcc (GCC) 3.3.5-20050130 (Gentoo 3.3.5.20050130-r1, ssp-3.3.5.20050130-1, pie-8.7.7.1) Copyright (C) 2003 Free Software Foundation, Inc. This is free software; see the source for copying conditions. There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
00010001
00000001
So either this problem is present only on gcc < 3.4 or Gentoo patched it somehow in newer versions...
Thanks,
Grzegorz Kulewski
Quoted 14 lines · Andreas Schwab
Quoted 11 lines · Trond Myklebust
Quoted 8 lines · Chuck Ebbert
#define ADDR (*(volatile long *) addr)
static inline voidset_bit(int nr, volatile unsigned long * addr){ __asm__ __volatile__( "lock ; " "btsl %1,%0" :"=m" (ADDR) :"Ir" (nr)); }
Quoted 5 lines
Quoted 2 lines
The asm needs a memory clobber in order to avoid reordering with the assignment to b[1]:
Check out 2.6.16-rc1, this has already been fixed.
No, that doesn't fix this testcase.
We need to somehow tell the compiler "this assembly statement altered memory and you can't cache memory contents across it". That's what "memory" (ie: barrier()) does. I don't think there's a way of telling gcc _what_ memory was clobbered - just "all of memory".
Quoted 4 lines · Andrew Morton
We need to somehow tell the compiler "this assembly statement altered memory and you can't cache memory contents across it". That's what "memory" (ie:
barrier()) does. I don't think there's a way of telling gcc _what_ memory was clobbered - just "all of memory".
We _can_ (and do) tell gcc that the unsigned long at address "addr" will be clobbered. The problem here is that we're actually applying set_bit() to a bit array that is larger than the single long, so we are not necessarily clobbering "addr", but rather the long at addr + X.
Most non-386 architectures don't actually have this compiler reordering problem since they tend to convert the index into the bit array into an offset for addr + a remainder:
unsigned long *offset = addr + (nr / 8*sizeof(unsigned long)); unsigned long bit = (nr % 8*sizeof(unsigned long));
and then tell the compiler that the long at "offset" will be clobbered. The remaining architectures appear to already have the general memory clobber set in their asms (as far as I can see).
IOW: the 386 and x86_64 appear to be the problem cases here, and then only when applied to large bit arrays.
Cheers
Trond