Re: Fix Linux fcntl OFD locks for non-LFS architectures (BZ#20251)

[Adhemerval Zanella <adhemerval dot zanella at linaro dot org>]

On 14/11/2018 17:36, Samuel Thibault wrote:
> Samuel Thibault, le jeu. 15 nov. 2018 02:29:39 +0100, a ecrit:
>> Samuel Thibault, le jeu. 15 nov. 2018 02:24:57 +0100, a ecrit:
>>> In login/utmp_file.c you have replaced calling __fcntl_nocancel by
>>> __fcntl64_nocancel, but shouldn't struct flock be replaced by struct
>>> flock64 too?
>>
>> Ah, no, applications *have* to use F_SETLK64 to use struct flock64, is
>> that it?
> 
> Mmm, no, as I read the Linux implementation, when calling fcntl(), one
> has to use F_SETLK64 to be able to use struct flock64, but when calling
> fcntl64(), struct flock64 is always used,a and thus login/utmp_file.c
> should really be useing struct flock64?
> 
> Samuel
> 

At least for Linux this specific usage is supported.  For 32-bit architectures,
Linux fcntl64 does:

 471 #if BITS_PER_LONG == 32                                                                
 472 SYSCALL_DEFINE3(fcntl64, unsigned int, fd, unsigned int, cmd,                          
 473                 unsigned long, arg)                                                    
 474 {   
 [...]
 492         switch (cmd) {  
 [...]
 511         default:
 512                 err = do_fcntl(fd, cmd, arg, f.file);
 513                 break;
 514         }

Which in turn will call fcntl_setlk. The issue that BZ#20251 is using
this same scenario but with Linux-specific OFD locks (which is not
supported).

However I agree that this is confusing and I think it would be an 
improvement if we explicit avoid use non-LFS interfaces within glibc.