| Message ID | 1471521815-4340-1-git-send-email-jlayton@redhat.com |
|---|---|
| State | New |
| Headers | show |
On 18 Aug 2016 08:03, Jeff Layton wrote: > +2016-08-18 Jeff Layton <jlayton@redhat.com> > + * sysdeps/unix/sysv/linux/bits/fcntl-linux.h: > + Add F_OFD_GETLK32, F_OFD_SETLK32, F_OFD_SETLKW32 Should be a blank line after the first one. look at all the other entries in this file as an example. > --- a/sysdeps/unix/sysv/linux/bits/fcntl-linux.h > +++ b/sysdeps/unix/sysv/linux/bits/fcntl-linux.h > @@ -127,11 +127,20 @@ > This means that they are inherited across fork or clone with CLONE_FILES > like BSD (flock) locks, and they are only released automatically when the > last reference to the the file description against which they were acquired > - is put. */ > + is put. */ > #ifdef __USE_GNU > -# define F_OFD_GETLK 36 > -# define F_OFD_SETLK 37 > -# define F_OFD_SETLKW 38 > +# if __OFF_T_MATCHES_OFF64_T || defined __USE_FILE_OFFSET64 > +# define F_OFD_GETLK 36 > +# define F_OFD_SETLK 37 > +# define F_OFD_SETLKW 38 > +# else > +# define F_OFD_GETLK32 39 > +# define F_OFD_SETLK32 40 > +# define F_OFD_SETLKW32 41 > +# define F_OFD_GETLK F_OFD_GETLK32 > +# define F_OFD_SETLK F_OFD_SETLK32 > +# define F_OFD_SETLKW F_OFD_SETLKW32 > +# endif > #endif i think we should define *64 and *32 variants all the time, and then route the F_OFD_GETLK/etc... to them based on compile mode. -mike
On Thu, 2016-08-18 at 06:04 -0700, Mike Frysinger wrote: > On 18 Aug 2016 08:03, Jeff Layton wrote: > > > > > > +2016-08-18 Jeff Layton <jlayton@redhat.com> > > > > + * sysdeps/unix/sysv/linux/bits/fcntl-linux.h: > > > > + Add F_OFD_GETLK32, F_OFD_SETLK32, F_OFD_SETLKW32 > > Should be a blank line after the first one. look at all the other > entries in this file as an example. > Ok, will fix... > > > > --- a/sysdeps/unix/sysv/linux/bits/fcntl-linux.h > > +++ b/sysdeps/unix/sysv/linux/bits/fcntl-linux.h > > @@ -127,11 +127,20 @@ > > This means that they are inherited across fork or clone with CLONE_FILES > > like BSD (flock) locks, and they are only released automatically when the > > last reference to the the file description against which they were acquired > > - is put. */ > > + is put. */ > > #ifdef __USE_GNU > > > > -# define F_OFD_GETLK 36 > > > > -# define F_OFD_SETLK 37 > > > > -# define F_OFD_SETLKW 38 > > +# if __OFF_T_MATCHES_OFF64_T || defined __USE_FILE_OFFSET64 > > > > +# define F_OFD_GETLK 36 > > > > +# define F_OFD_SETLK 37 > > > > +# define F_OFD_SETLKW 38 > > +# else > > > > +# define F_OFD_GETLK32 39 > > > > +# define F_OFD_SETLK32 40 > > > > +# define F_OFD_SETLKW32 41 > > > > +# define F_OFD_GETLK F_OFD_GETLK32 > > > > +# define F_OFD_SETLK F_OFD_SETLK32 > > > > +# define F_OFD_SETLKW F_OFD_SETLKW32 > > +# endif > > #endif > > i think we should define *64 and *32 variants all the time, and > then route the F_OFD_GETLK/etc... to them based on compile mode. > -mike Sorry, I don't quite understand here. The whole point is that the existing F_OFD_* constants are already implicitly 64-bit. Why do we need separate constants postfixed with "64" that no one will ever use?
On 18 Aug 2016 09:49, Jeff Layton wrote: > On Thu, 2016-08-18 at 06:04 -0700, Mike Frysinger wrote: > > On 18 Aug 2016 08:03, Jeff Layton wrote: > > > --- a/sysdeps/unix/sysv/linux/bits/fcntl-linux.h > > > +++ b/sysdeps/unix/sysv/linux/bits/fcntl-linux.h > > > @@ -127,11 +127,20 @@ > > > This means that they are inherited across fork or clone with CLONE_FILES > > > like BSD (flock) locks, and they are only released automatically when the > > > last reference to the the file description against which they were acquired > > > - is put. */ > > > + is put. */ > > > #ifdef __USE_GNU > > > > > -# define F_OFD_GETLK 36 > > > > > -# define F_OFD_SETLK 37 > > > > > -# define F_OFD_SETLKW 38 > > > +# if __OFF_T_MATCHES_OFF64_T || defined __USE_FILE_OFFSET64 > > > > > +# define F_OFD_GETLK 36 > > > > > +# define F_OFD_SETLK 37 > > > > > +# define F_OFD_SETLKW 38 > > > +# else > > > > > +# define F_OFD_GETLK32 39 > > > > > +# define F_OFD_SETLK32 40 > > > > > +# define F_OFD_SETLKW32 41 > > > > > +# define F_OFD_GETLK F_OFD_GETLK32 > > > > > +# define F_OFD_SETLK F_OFD_SETLK32 > > > > > +# define F_OFD_SETLKW F_OFD_SETLKW32 > > > +# endif > > > #endif > > > > i think we should define *64 and *32 variants all the time, and > > then route the F_OFD_GETLK/etc... to them based on compile mode. > > Sorry, I don't quite understand here. The whole point is that the > existing F_OFD_* constants are already implicitly 64-bit. Why do we > need separate constants postfixed with "64" that no one will ever use? you're making them not explicitly 64-bit when off_t!=off64_t. all the other commands in this file have used the convention: <cmd>: automatically 32-bit or 64-bit <cmd>64: always 64-bit so the header would do: #define F_OFD_GETLK64 36 #define F_OFD_SETLK64 37 #define F_OFD_SETLKW64 38 #if __OFF_T_MATCHES_OFF64_T || defined __USE_FILE_OFFSET64 # define F_OFD_GETLK F_OFD_GETLK64 # define F_OFD_SETLK F_OFD_SETLK64 # define F_OFD_SETLKW F_OFD_SETLKW64 #else # define F_OFD_GETLK 39 # define F_OFD_SETLK 40 # define F_OFD_SETLKW 41 #endif the defines deviate a bit from the names used on the kernel side, but we've already done that, and you're proposed patch does too. -mike
diff --git a/ChangeLog b/ChangeLog index 59c68d8b9d0d..4a2488b50054 100644 --- a/ChangeLog +++ b/ChangeLog @@ -1,3 +1,7 @@ +2016-08-18 Jeff Layton <jlayton@redhat.com> + * sysdeps/unix/sysv/linux/bits/fcntl-linux.h: + Add F_OFD_GETLK32, F_OFD_SETLK32, F_OFD_SETLKW32 + 2016-08-18 Stefan Liebler <stli@linux.vnet.ibm.com> * sysdeps/ieee754/dbl-64/k_rem_pio2.c (__kernel_rem_pio2): diff --git a/sysdeps/unix/sysv/linux/bits/fcntl-linux.h b/sysdeps/unix/sysv/linux/bits/fcntl-linux.h index 7e5b0aecdcb4..9cb1b5f6f3ec 100644 --- a/sysdeps/unix/sysv/linux/bits/fcntl-linux.h +++ b/sysdeps/unix/sysv/linux/bits/fcntl-linux.h @@ -127,11 +127,20 @@ This means that they are inherited across fork or clone with CLONE_FILES like BSD (flock) locks, and they are only released automatically when the last reference to the the file description against which they were acquired - is put. */ + is put. */ #ifdef __USE_GNU -# define F_OFD_GETLK 36 -# define F_OFD_SETLK 37 -# define F_OFD_SETLKW 38 +# if __OFF_T_MATCHES_OFF64_T || defined __USE_FILE_OFFSET64 +# define F_OFD_GETLK 36 +# define F_OFD_SETLK 37 +# define F_OFD_SETLKW 38 +# else +# define F_OFD_GETLK32 39 +# define F_OFD_SETLK32 40 +# define F_OFD_SETLKW32 41 +# define F_OFD_GETLK F_OFD_GETLK32 +# define F_OFD_SETLK F_OFD_SETLK32 +# define F_OFD_SETLKW F_OFD_SETLKW32 +# endif #endif #ifdef __USE_LARGEFILE64
The original assumption with OFD lock support was that userland would always pass in a struct flock that has been set up for large file support (LFS). It's possible however for someone to build a 32-bit program without large file support, and still try to use OFD locks. If that happens then it will send a non-LFS struct flock to the kernel, which will then try to interpret it as a LFS one, which can cause unexpected results. My original idea was to just ensure that the build would fail in this case by not defining the F_OFD_* constants when LFS support was not enabled. Instead though, it's probably better long-term to just go ahead and support OFD locks with non-LFS struct flock. To do that without breaking programs that are already working, we add a set of F_OFD_*32 constants, and add the plumbing to the kernel to make those do the right thing. This patch adds the same constants to glibc, and has it redefine the "normal" F_OFD_* constants to their *32 equvalents when LFS support is not in use. Signed-off-by: Jeff Layton <jlayton@redhat.com> --- ChangeLog | 4 ++++ sysdeps/unix/sysv/linux/bits/fcntl-linux.h | 17 +++++++++++++---- 2 files changed, 17 insertions(+), 4 deletions(-)