Re: [musl] [PATCH v1] add renameat2 linux syscall wrapper

[Rich Felker <dalias@libc.org>]

On Mon, Apr 22, 2024 at 09:39:17PM -0700, Tony Ambardar wrote:
> On Mon, Apr 22, 2024 at 06:47:26PM -0400, Rich Felker wrote:
> > On Sun, Apr 21, 2024 at 08:36:40AM -0700, Tony Ambardar wrote:
> > > This syscall is available since Linux 3.15 and also implemented in glibc
> > > from version 2.28. It is commonly used in filesystem or security contexts.
> > > 
> > > Defines RENAME_NOREPLACE, RENAME_EXCHANGE, RENAME_WHITEOUT are guarded by
> > > _GNU_SOURCE as with glibc.
> > > 
> > > Signed-off-by: Tony Ambardar <Tony.Ambardar@gmail.com>
> > > ---
> > >  include/stdio.h       | 7 +++++++
> > >  src/linux/renameat2.c | 8 ++++++++
> > >  2 files changed, 15 insertions(+)
> > >  create mode 100644 src/linux/renameat2.c
> > > 
> > > diff --git a/include/stdio.h b/include/stdio.h
> > > index cb858618..8312c3bf 100644
> > > --- a/include/stdio.h
> > > +++ b/include/stdio.h
> > > @@ -158,6 +158,13 @@ char *ctermid(char *);
> > >  #define L_ctermid 20
> > >  #endif
> > >  
> > > +#if defined(_GNU_SOURCE)
> > > +#define RENAME_NOREPLACE (1 << 0)
> > > +#define RENAME_EXCHANGE (1 << 1)
> > > +#define RENAME_WHITEOUT (1 << 2)
> > > +
> > > +int renameat2(int, const char *, int, const char *, unsigned int);
> > > +#endif
> > 
> > s/unsigned int/unsigned/ and maybe just write out the constants? I
> > think that's the style musl uses most places.
> 
> Right, the constants-by-shift mirrored glibc and I saw these also in musl
> e.g. hwcap.h, elf.h, mount.h, syslog.h, etc. I'm not wedded to either but
> just wanted to be consistent. Please confirm a preference and I'll update.
> 
> > 
> > >  #if defined(_XOPEN_SOURCE) || defined(_GNU_SOURCE) \
> > >   || defined(_BSD_SOURCE)
> > > diff --git a/src/linux/renameat2.c b/src/linux/renameat2.c
> > > new file mode 100644
> > > index 00000000..3062aa15
> > > --- /dev/null
> > > +++ b/src/linux/renameat2.c
> > > @@ -0,0 +1,8 @@
> > > +#define _GNU_SOURCE
> > > +#include <stdio.h>
> > > +#include "syscall.h"
> > > +
> > > +int renameat2(int oldfd, const char *old, int newfd, const char *new, unsigned int flags)
> > > +{
> > > +	return syscall(SYS_renameat2, oldfd, old, newfd, new, flags);
> > > +}
> > > -- 
> > > 2.34.1
> > 
> > This probably at least needs to support flags==0 on kernels without
> > SYS_renameat2 by calling renameat in that case. Then I'm not sure if
> > ENOSYS should be kept if the new syscall is missing, or just EINVAL or
> > whatever is used to report unsupported flags.
> 
> I had wondered about this too and looked at all 'arch/.../syscall.h.in'
> files. All arches support SYS_renameat2, but riscv, aarch64, loongarch64
> and or1k are missing SYS_rename and/or SYS_renameat. Also, current code in
> stdio/rename.c and unistd/renameat.c may fall back _unconditionally_ to
> use SYS_renameat2 in these cases, with the assumption it's always present.
> 
> I think this means we're OK? Or do we allow for old kernels which lack
> SYS_renameat2, in which case some existing code would need updating?

No, this just means it's unconditionally present on some archs
(because they were added after SYS_renameat2 was added), not that it
can be assumed to be present in general.

Rich