On 2019-07-18, Rasmus Villemoes <linux@rasmusvillemoes.dk> wrote:
> On 06/07/2019 16.57, Aleksa Sarai wrote:
> > --- a/fs/open.c
> > +++ b/fs/open.c
> > @@ -928,24 +928,32 @@ struct file *open_with_fake_path(const struct path *path, int flags,
> >  }
> >  EXPORT_SYMBOL(open_with_fake_path);
> >  
> > -static inline int build_open_flags(int flags, umode_t mode, struct open_flags *op)
> > +static inline int build_open_flags(struct open_how how, struct open_flags *op)
> >  {
> 
> How does passing such a huge struct by value affect code generation?
> Does gcc actually inline the function (and does it even inline the old
> one given that it's already non-trivial and has more than one caller).

I'm not sure, but I'll just do what you suggested with passing a const
reference and just copying the few fields that actually are touched by
this function.

> >  
> > diff --git a/include/linux/fcntl.h b/include/linux/fcntl.h
> > index 2868ae6c8fc1..e59917292213 100644
> > --- a/include/linux/fcntl.h
> > +++ b/include/linux/fcntl.h
> > @@ -4,13 +4,26 @@
> >  
> >  #include <uapi/linux/fcntl.h>
> >  
> > -/* list of all valid flags for the open/openat flags argument: */
> > +/* Should open_how.mode be set for older syscalls wrappers? */
> > +#define OPENHOW_MODE(flags, mode) \
> > +	(((flags) | (O_CREAT | __O_TMPFILE)) ? (mode) : 0)
> > +
> 
> Typo: (((flags) & (O_CREAT | __O_TMPFILE)) ? (mode) : 0)

Yup, thanks. I'm not sure why my tests passed on v9 with this bug (they
didn't pass in my v10-draft until I fixed this bug earlier today).

> 
> > +/**
> > + * Arguments for how openat2(2) should open the target path. If @extra is zero,
> > + * then openat2(2) is identical to openat(2).
> > + *
> > + * @flags: O_* flags (unknown flags ignored).
> > + * @mode: O_CREAT file mode (ignored otherwise).
> 
> should probably say "O_CREAT/O_TMPFILE file mode".

:+1:

> > + * @upgrade_mask: restrict how the O_PATH may be re-opened (ignored otherwise).
> > + * @resolve: RESOLVE_* flags (-EINVAL on unknown flags).
> > + * @reserved: reserved for future extensions, must be zeroed.
> > + */
> > +struct open_how {
> > +	__u32 flags;
> > +	union {
> > +		__u16 mode;
> > +		__u16 upgrade_mask;
> > +	};
> > +	__u16 resolve;
> 
> So mode and upgrade_mask are naturally u16 aka mode_t. And yes, they
> probably never need to be used together, so the union works. That then
> makes the next member 2-byte aligned, so using a u16 for the resolve
> flags brings us to an 8-byte boundary, and 11 unused flag bits should be
> enough for a while. But it seems a bit artificial to cram all this
> together and then add 56 bytes of reserved space.

I will happily admit that padding to 64 bytes is probably _very_ extreme
(I picked it purely because it's the size of a cache-line so anything
bigger makes even less sense). I was hoping someone would suggest a
better size once I posted the patchset, since I couldn't think of a good
answer myself.

Do you have any suggestions for a better layout or padding size?

-- 
Aleksa Sarai
Senior Software Engineer (Containers)
SUSE Linux GmbH
<https://www.cyphar.com/>