[PATCH v12 11/12] open: openat2(2) syscall

[Andy Lutomirski]

> On Sep 7, 2019, at 9:58 AM, Linus Torvalds <torvalds at linux-foundation.org> wrote:
> 
>> On Sat, Sep 7, 2019 at 5:40 AM Jeff Layton <jlayton at kernel.org> wrote:
>> 
>> After thinking about this a bit, I wonder if we might be better served
>> with a new set of OA2_* flags instead of repurposing the O_* flags?
> 
> I'd hate to have yet _another_ set of translation functions, and
> another chance of people just getting it wrong either in user space or
> the kernel.
> 
> So no. Let's not make another set of flags that has no sane way to
> have type-safety to avoid more confusion.
> 
> The new flags that _only_ work with openat2() might be named with a
> prefix/suffix to mark that, but I'm not sure it's a huge deal.
> 
>            

I agree with the philosophy, but I think it doesn’t apply in this case.  Here are the flags:

O_RDONLY, O_WRONLY, O_RDWR: not even a proper bitmask. The kernel already has the FMODE_ bits to make this make sense. How about we make the openat2 permission bits consistent with the internal representation and let the O_ permission bits remain as an awful translation.  The kernel already translates like this, and it already sucks.

O_CREAT, O_TMPFILE, O_NOCTTY, O_TRUNC: not modes on the fd at all.  These affect the meaning of open().  Heck, for openat2, NOCTTY should be this default.

O_EXCL: hopelessly overloaded.

O_APPEND, O_DIRECT, O_SYNC, O_DSYNC, O_LARGEFILE, O_NOATIME, O_PATH, O_NONBLOCK: genuine mode bits

O_CLOEXEC: special because it affects the fd, not the struct file.

Linus, you rejected resolveat() because you wanted a *nice* API that people would use and that might even be adopted by other OSes. Let’s please not make openat2() be a giant pile of crap in the name of consistency with open().  open(), frankly, is horrible.