Re: [PATCH v4 2/3] 32-bit ABIs: support stat syscall family

[Yury Norov <ynorov at caviumnetworks dot com>]

On Mon, Aug 29, 2016 at 10:34:38PM +0200, Arnd Bergmann wrote:
> On Monday 29 August 2016, Yury Norov wrote:
> > > 
> > > We've gone back and forth on the arm64+ilp32 definition of the
> > > existing stat64 , because that one cannot easily use the same
> > > generic structure. I originally thought that using the 64-bit
> > > layout of 'struct stat' was a good idea for arm64, but now
> > > I don't see much of an advantage either way (the existing arm32
> > > 'struct stat64' layout or the existing  arm64 'struct stat' layout).
> > > 
> > > Both of them are incompatible with the default layout for all
> > > other new 32-bit architectures in user space, so pick one of the
> > > two for the kernel interface, and convert it to the normal layout
> > > in glibc using __xstat_conv():
> > > 
> > > - if you use the arm64 'struct stat' layout (as in the current
> > >   kernel patches), you need a temporary buffer with the larger
> > >   size and then copy over the time fields one by one
> > > - if you instead use the arm32 'struct stat64' layout in the kernel,
> > >   the size is the same, and the conversion is a one-line
> > >   assignment of st_ino.
> > > 
> > >       Arnd
> > 
> > Ok. I'll do like this. I'd prefer 1st option, because 2nd implies
> > allocating big buffer on kernel side.
> 
> On the kernel side, both require allocating a buffer, the second
> one is just slightly bigger. The kernel uses its own internal
> 'struct kstat', which gets copied into one of the other structures.
> 
> 	Arnd

Ah, and in kernel too...

Anyway, using the arm32 'struct stat64' cannot help because
__xstat{,32,64}_conv() assumes copying of the kernel buffer to user
buffer.

We can, of course, write custom __xstat_conv() to avoid that copying. 
But then we'd write custom syscalls for stat family for aarch64/ilp32.
More simple and straightforward way then is to introduce custom layout
for struct stat, and avoid all conversions and memory wasting at all.
And this is what I suggested in the very first series of ilp32 support.
But Joseph declined it as non-generic solution.

So for me there are 3 options:
 - __xstat_conv() - hacky, complex and ineffective. I work on it right
   now, and I'm not happy.
 - current version with conversion of time fields only - adds new
   fresh portion of hacks, but little more effective.
 - initial version. Effective, simple, free of hacks but (because of)
   non-generic.

Stat-related code in glibc is probably the worst code in open source
I ever seen. That's why I don't want ilp32 to use it. But if it's
impossible, I'd end up with aarch64. At least it doesn't have broken
fields.