On 31 May 2016 17:00, Chris Metcalf wrote:
> On 5/31/2016 4:04 PM, Yury Norov wrote:
> > In path a63c7fa18a (Add sysdeps/unix/sysv/linux/generic/.) you add
> > this:
> > +++ b/sysdeps/unix/sysv/linux/generic/wordsize-32/preadv.c
> >
> > [...]
> >
> > +static ssize_t
> > +do_preadv (int fd, const struct iovec *vector, int count, off_t
> > offset)
> > +{
> > +  assert (sizeof (offset) == 4);
> > +  return INLINE_SYSCALL (preadv, __ALIGNMENT_COUNT (5, 6), fd,
> > +                         vector, count, __ALIGNMENT_ARG
> > +                         __LONG_LONG_PAIR (offset >> 31, offset));
> > +}
> > +
> >
> > And this is the code that is picked up if I choose wordsize-32 for my
> > AARCH64/ILP32. So I have questions.
> >
> > 1. What is the assert for? We agreed that all new ABIs will be 64-bit
> > off_t only.
> >
> > I fixed it internally like this:
> > +#ifndef __OFF_T_MATCHES_OFF64_T
> >          assert (sizeof (offset) == 4);
> > +#endif
> >
> > There is a bunch of similar assertions in glibc.
> >
> > 2. This one looks weird:
> > __LONG_LONG_PAIR (offset >> 31, offset))
> > Why 31-bit offset? And why you don't mask 2nd argument?
> > Later in your patch I see this:
> > +static ssize_t
> > +do_preadv64 (int fd, const struct iovec *vector, int count, off64_t
> > offset)
> >
> > +{
> > +  return INLINE_SYSCALL (preadv, __ALIGNMENT_COUNT (5, 6), fd,
> > +                         vector, count, __ALIGNMENT_ARG
> > +                         __LONG_LONG_PAIR ((off_t) (offset >> 32),
> > +                                           (off_t) (offset & 0xffffffff)));
> > +}
> >
> > And it looks correct to me. If 1st version is correct as well, I think
> > it should be commented.
> 
> I did this work before x32 came out, so I tried to model it more closely on
> the existing x86 compat API.  I agree that a 64-bit off_t model seems reasonable;
> however, the code does exactly what it does to match x86, namely preadv() takes
> a 32-bit offset, and preadv64() take a 64-bit offset.  The assert() in preadv to force
> sizeof to be 4 is exactly why in that routine we use (offset >> 31, offset).  Since
> we know offset fits in 32 bits, all we need to do is properly sign-extend it into
> 64 bits in the high register of the pair, which is what (offset >> 31) does - you end
> up with only 0 or -1, thus sign-extending the 32-bit signed off_t. Then in
> preadv64() we actually need to break apart the 64-bit offset into a high 32 bits
> and a low 32 bits, which is what (offset >> 32, offset & 0xffffffff) does.
> 
> For a 64-bit off_t you will want to not compile preadv.c at all, and instead make
> __libc_preadv() and friends be aliases of __libc_preadv64().

sounds like Adhemerval's pread/pwrite unify work should be extended to
the preadv/pwritev funcs.  it deals with the ilp32 case and uses the new
SYSCALL_LL macro to deal with the ugly shifting/masking.

check out these commits:
https://sourceware.org/git/?p=glibc.git;a=commit;h=071af4769fcdfe2cd349157b01f27c9571478ace
https://sourceware.org/git/?p=glibc.git;a=commit;h=77a4fbd53611720cd6ae845de560df5dd332b28e
https://sourceware.org/git/?p=glibc.git;a=commit;h=eeddfa91cbb1a619af135c7a9ac14251ec094b7a
-mike