| Message ID | 20230828213721.2957677-1-aurelien@aurel32.net |
|---|---|
| State | New |
| Headers | show |
| Series | io: Fix record locking contants for powerpc64 with __USE_FILE_OFFSET64 | expand |
* Aurelien Jarno: > diff --git a/sysdeps/unix/sysv/linux/powerpc/bits/fcntl.h b/sysdeps/unix/sysv/linux/powerpc/bits/fcntl.h > index f7615a447e..d8a291a331 100644 > --- a/sysdeps/unix/sysv/linux/powerpc/bits/fcntl.h > +++ b/sysdeps/unix/sysv/linux/powerpc/bits/fcntl.h > @@ -33,7 +33,7 @@ > # define __O_LARGEFILE 0200000 > #endif > > -#if __WORDSIZE == 64 > +#if __WORDSIZE == 64 && !defined __USE_FILE_OFFSET64 > # define F_GETLK 5 > # define F_SETLK 6 > # define F_SETLKW 7 I find this puzzling. Why would __USE_FILE_OFFSET64 have an effect if __WORDSIZE is 64? Thanks, Florian
On 29/08/23 04:59, Florian Weimer via Libc-alpha wrote: > * Aurelien Jarno: > >> diff --git a/sysdeps/unix/sysv/linux/powerpc/bits/fcntl.h b/sysdeps/unix/sysv/linux/powerpc/bits/fcntl.h >> index f7615a447e..d8a291a331 100644 >> --- a/sysdeps/unix/sysv/linux/powerpc/bits/fcntl.h >> +++ b/sysdeps/unix/sysv/linux/powerpc/bits/fcntl.h >> @@ -33,7 +33,7 @@ >> # define __O_LARGEFILE 0200000 >> #endif >> >> -#if __WORDSIZE == 64 >> +#if __WORDSIZE == 64 && !defined __USE_FILE_OFFSET64 >> # define F_GETLK 5 >> # define F_SETLK 6 >> # define F_SETLKW 7 > > I find this puzzling. Why would __USE_FILE_OFFSET64 have an effect if > __WORDSIZE is 64? This is a historical artifact from powerpc64. Instead of following other 64-bit architectures and define F_GETLK the same whether _FILE_OFFSET_BITS is defined, the port used powerpc definitiosn that required different values to support LFS. This patch is not wrong, but at same time not really required. The powercp64 fcntl will handle F_GETLK/F_SETLK/F_SETLKW with the historical values with the FCNTL_ADJUST_CMD macro, so old binaries will continue to work as expected.
On 2023-08-29 10:20, Adhemerval Zanella Netto wrote: > > > On 29/08/23 04:59, Florian Weimer via Libc-alpha wrote: > > * Aurelien Jarno: > > > >> diff --git a/sysdeps/unix/sysv/linux/powerpc/bits/fcntl.h b/sysdeps/unix/sysv/linux/powerpc/bits/fcntl.h > >> index f7615a447e..d8a291a331 100644 > >> --- a/sysdeps/unix/sysv/linux/powerpc/bits/fcntl.h > >> +++ b/sysdeps/unix/sysv/linux/powerpc/bits/fcntl.h > >> @@ -33,7 +33,7 @@ > >> # define __O_LARGEFILE 0200000 > >> #endif > >> > >> -#if __WORDSIZE == 64 > >> +#if __WORDSIZE == 64 && !defined __USE_FILE_OFFSET64 > >> # define F_GETLK 5 > >> # define F_SETLK 6 > >> # define F_SETLKW 7 > > > > I find this puzzling. Why would __USE_FILE_OFFSET64 have an effect if > > __WORDSIZE is 64? > > This is a historical artifact from powerpc64. Instead of following other 64-bit > architectures and define F_GETLK the same whether _FILE_OFFSET_BITS is defined, > the port used powerpc definitiosn that required different values to support LFS. > > This patch is not wrong, but at same time not really required. The powercp64 > fcntl will handle F_GETLK/F_SETLK/F_SETLKW with the historical values with > the FCNTL_ADJUST_CMD macro, so old binaries will continue to work as expected. It does break some binaries, for instance File-FcntlLock [1] with the following scenario: - perl is built against glibc 2.37 - system is upgraded to glibc 2.38 - File-FcntlLock is built against glibc 2.38: it does not work as the value of F_GETLK has changed from 12 to 5 [2]. In short these constants are not exclusively used by the glibc. They might be referenced in some libraries, and changing their values break things. [1] https://metacpan.org/dist/File-FcntlLock [2] https://metacpan.org/release/JTT/File-FcntlLock-0.22/source/FcntlLock.xs#L74
On 29/08/23 13:30, Aurelien Jarno wrote: > On 2023-08-29 10:20, Adhemerval Zanella Netto wrote: >> >> >> On 29/08/23 04:59, Florian Weimer via Libc-alpha wrote: >>> * Aurelien Jarno: >>> >>>> diff --git a/sysdeps/unix/sysv/linux/powerpc/bits/fcntl.h b/sysdeps/unix/sysv/linux/powerpc/bits/fcntl.h >>>> index f7615a447e..d8a291a331 100644 >>>> --- a/sysdeps/unix/sysv/linux/powerpc/bits/fcntl.h >>>> +++ b/sysdeps/unix/sysv/linux/powerpc/bits/fcntl.h >>>> @@ -33,7 +33,7 @@ >>>> # define __O_LARGEFILE 0200000 >>>> #endif >>>> >>>> -#if __WORDSIZE == 64 >>>> +#if __WORDSIZE == 64 && !defined __USE_FILE_OFFSET64 >>>> # define F_GETLK 5 >>>> # define F_SETLK 6 >>>> # define F_SETLKW 7 >>> >>> I find this puzzling. Why would __USE_FILE_OFFSET64 have an effect if >>> __WORDSIZE is 64? >> >> This is a historical artifact from powerpc64. Instead of following other 64-bit >> architectures and define F_GETLK the same whether _FILE_OFFSET_BITS is defined, >> the port used powerpc definitiosn that required different values to support LFS. >> >> This patch is not wrong, but at same time not really required. The powercp64 >> fcntl will handle F_GETLK/F_SETLK/F_SETLKW with the historical values with >> the FCNTL_ADJUST_CMD macro, so old binaries will continue to work as expected. > > It does break some binaries, for instance File-FcntlLock [1] with the > following scenario: > - perl is built against glibc 2.37 > - system is upgraded to glibc 2.38 > - File-FcntlLock is built against glibc 2.38: it does not work as the > value of F_GETLK has changed from 12 to 5 [2]. > > In short these constants are not exclusively used by the glibc. They > might be referenced in some libraries, and changing their values break > things. > > [1] https://metacpan.org/dist/File-FcntlLock > [2] https://metacpan.org/release/JTT/File-FcntlLock-0.22/source/FcntlLock.xs#L74 > Sigh, you are correct. And it is also means that default and LFS objects are essentially incompatible on powerpc64 in this regard for this very reason, but I agree that we should keep the compatibility. Could add a LFS test, just to make sure the lock constants are correct: diff --git a/io/Makefile b/io/Makefile index 6ccc0e8691..8a3c83a3bb 100644 --- a/io/Makefile +++ b/io/Makefile @@ -192,6 +192,7 @@ tests := \ tst-fchownat \ tst-fcntl \ tst-fcntl-lock \ + tst-fcntl-lock-lfs \ tst-fstatat \ tst-fts \ tst-fts-lfs \ diff --git a/io/tst-fcntl-lock-lfs.c b/io/tst-fcntl-lock-lfs.c new file mode 100644 index 0000000000..f2a909fb02 --- /dev/null +++ b/io/tst-fcntl-lock-lfs.c @@ -0,0 +1,2 @@ +#define _FILE_OFFSET_BITS 64 +#include <io/tst-fcntl-lock.c> It won't really trigger this issue, but it a extra sanity test.
On Aug 29 2023, Aurelien Jarno wrote: > It does break some binaries, for instance File-FcntlLock [1] with the > following scenario: > - perl is built against glibc 2.37 > - system is upgraded to glibc 2.38 > - File-FcntlLock is built against glibc 2.38: it does not work as the > value of F_GETLK has changed from 12 to 5 [2]. It happens all the time that an API change breaks some combination of recompiled and not recompiled objects. But it is not an ABI break, only a QoI issue.
diff --git a/sysdeps/unix/sysv/linux/powerpc/bits/fcntl.h b/sysdeps/unix/sysv/linux/powerpc/bits/fcntl.h index f7615a447e..d8a291a331 100644 --- a/sysdeps/unix/sysv/linux/powerpc/bits/fcntl.h +++ b/sysdeps/unix/sysv/linux/powerpc/bits/fcntl.h @@ -33,7 +33,7 @@ # define __O_LARGEFILE 0200000 #endif -#if __WORDSIZE == 64 +#if __WORDSIZE == 64 && !defined __USE_FILE_OFFSET64 # define F_GETLK 5 # define F_SETLK 6 # define F_SETLKW 7
Commit 5f828ff824e3b7cd1 ("io: Fix F_GETLK, F_SETLK, and F_SETLKW for powerpc64") fixed an issue with the value of the lock constants on powerpc64 when not using __USE_FILE_OFFSET64, but it ended-up also changing the value when using __USE_FILE_OFFSET64 causing an ABI breakage. Fix that by also checking that define, restoring the pre 4d0fe291aed3a476a commit values: Default values: - F_GETLK: 5 - F_SETLK: 6 - F_SETLKW: 7 With -D_FILE_OFFSET_BITS=64: - F_GETLK: 12 - F_SETLK: 13 - F_SETLKW: 14 Resolves: BZ #30804. Signed-off-by: Aurelien Jarno <aurelien@aurel32.net> --- sysdeps/unix/sysv/linux/powerpc/bits/fcntl.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)