On Tue, Aug 25, 2020 at 02:27:11PM +1000, Dave Chinner wrote:
> On Mon, Aug 24, 2020 at 09:35:59PM -0600, Andreas Dilger wrote:
> > On Aug 24, 2020, at 9:26 PM, Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
> > >
> > > On Tue, Aug 25, 2020 at 10:27:35AM +1000, Dave Chinner wrote:
> > >>> do {
> > >>> - unsigned offset, bytes;
> > >>> -
> > >>> - offset = offset_in_page(pos);
> > >>> - bytes = min_t(loff_t, PAGE_SIZE - offset, count);
> > >>> + loff_t bytes;
> > >>>
> > >>> if (IS_DAX(inode))
> > >>> - status = dax_iomap_zero(pos, offset, bytes, iomap);
> > >>> + bytes = dax_iomap_zero(pos, length, iomap);
> > >>
> > >> Hmmm. everything is loff_t here, but the callers are defining length
> > >> as u64, not loff_t. Is there a potential sign conversion problem
> > >> here? (sure 64 bit is way beyond anything we'll pass here, but...)
> > >
> > > I've gone back and forth on the correct type for 'length' a few times.
> > > size_t is too small (not for zeroing, but for seek()). An unsigned type
> > > seems right -- a length can't be negative, and we don't want to give
> > > the impression that it can. But the return value from these functions
> > > definitely needs to be signed so we can represent an error. So a u64
> > > length with an loff_t return type feels like the best solution. And
> > > the upper layers have to promise not to pass in a length that's more
> > > than 2^63-1.
> >
> > The problem with allowing a u64 as the length is that it leads to the
> > possibility of an argument value that cannot be returned. Checking
> > length < 0 is not worse than checking length > 0x7ffffffffffffff,
> > and has the benefit of consistency with the other argument types and
> > signs...
The callee should just trust that the caller isn't going to do that.
File sizes can't be more than 2^63-1 bytes, so an extent of a file also
can't be more than 2^63-1 bytes.
> I think the problem here is that we have no guaranteed 64 bit size
> type. when that was the case with off_t, we created loff_t to always
> represent a 64 bit offset value. However, we never created one for
> the count/size that is passed alongside loff_t in many places - it
> was said that "syscalls are limited to 32 bit sizes" and
> "size_t is 64 bit on 64 bit platforms" and so on and so we still
> don't have a clean way to pass 64 bit sizes through the IO path.
>
> We've been living with this shitty situation for a long time now, so
> perhaps it's time for us to define lsize_t for 64 bit lengths and
> start using that everywhere that needs a 64 bit clean path
> through the code, regardless of whether the arch is 32 or 64 bit...
>
> Thoughts?
I don't think the THP patches should be blocked on this expedition.
We have a guaranteed 64 bit type -- it's u64. I don't think defining
lsize_t is going to fix anything. The next big problem to fix will be
supporting storage >16EiB, but I think that's a project that can start
in ten years and still be here before anyone but the TLAs have that much
storage in a single device.
Any objection to leaving this patch as-is with a u64 length?
