Buggy commit tracked to: "Re: [PATCH 2/9] iov_iter: move rw_copy_check_uvector() into lib/iov_iter.c"

[David Laight]

From: David Hildenbrand
> Sent: 22 October 2020 10:19
> 
> On 22.10.20 11:01, Greg KH wrote:
> > On Thu, Oct 22, 2020 at 10:48:59AM +0200, David Hildenbrand wrote:
> >> On 22.10.20 10:40, David Laight wrote:
> >>> From: David Hildenbrand
> >>>> Sent: 22 October 2020 09:35
> >>>>
> >>>> On 22.10.20 10:26, Greg KH wrote:
> >>>>> On Thu, Oct 22, 2020 at 12:39:14AM +0100, Al Viro wrote:
> >>>>>> On Wed, Oct 21, 2020 at 06:13:01PM +0200, Greg KH wrote:
> >>>>>>> On Fri, Sep 25, 2020 at 06:51:39AM +0200, Christoph Hellwig wrote:
> >>>>>>>> From: David Laight <David.Laight at ACULAB.COM>
> >>>>>>>>
> >>>>>>>> This lets the compiler inline it into import_iovec() generating
> >>>>>>>> much better code.
> >>>>>>>>
> >>>>>>>> Signed-off-by: David Laight <david.laight at aculab.com>
> >>>>>>>> Signed-off-by: Christoph Hellwig <hch at lst.de>
> >>>>>>>> ---
> >>>>>>>>  fs/read_write.c | 179 ------------------------------------------------
> >>>>>>>>  lib/iov_iter.c  | 176 +++++++++++++++++++++++++++++++++++++++++++++++
> >>>>>>>>  2 files changed, 176 insertions(+), 179 deletions(-)
> >>>>>>>
> >>>>>>> Strangely, this commit causes a regression in Linus's tree right now.
> >>>>>>>
> >>>>>>> I can't really figure out what the regression is, only that this commit
> >>>>>>> triggers a "large Android system binary" from working properly.  There's
> >>>>>>> no kernel log messages anywhere, and I don't have any way to strace the
> >>>>>>> thing in the testing framework, so any hints that people can provide
> >>>>>>> would be most appreciated.
> >>>>>>
> >>>>>> It's a pure move - modulo changed line breaks in the argument lists
> >>>>>> the functions involved are identical before and after that (just checked
> >>>>>> that directly, by checking out the trees before and after, extracting two
> >>>>>> functions in question from fs/read_write.c and lib/iov_iter.c (before and
> >>>>>> after, resp.) and checking the diff between those.
> >>>>>>
> >>>>>> How certain is your bisection?
> >>>>>
> >>>>> The bisection is very reproducable.
> >>>>>
> >>>>> But, this looks now to be a compiler bug.  I'm using the latest version
> >>>>> of clang and if I put "noinline" at the front of the function,
> >>>>> everything works.
> >>>>
> >>>> Well, the compiler can do more invasive optimizations when inlining. If
> >>>> you have buggy code that relies on some unspecified behavior, inlining
> >>>> can change the behavior ... but going over that code, there isn't too
> >>>> much action going on. At least nothing screamed at me.
> >>>
> >>> Apart from all the optimisations that get rid off the 'pass be reference'
> >>> parameters and strange conditional tests.
> >>> Plenty of scope for the compiler getting it wrong.
> >>> But nothing even vaguely illegal.
> >>
> >> Not the first time that people blame the compiler to then figure out
> >> that something else is wrong ... but maybe this time is different :)
> >
> > I agree, I hate to blame the compiler, that's almost never the real
> > problem, but this one sure "feels" like it.
> >
> > I'm running some more tests, trying to narrow things down as just adding
> > a "noinline" to the function that got moved here doesn't work on Linus's
> > tree at the moment because the function was split into multiple
> > functions.
> >
> > Give me a few hours...
> 
> I might be wrong but
> 
> a) import_iovec() uses:
> - unsigned nr_segs -> int
> - unsigned fast_segs -> int
> b) rw_copy_check_uvector() uses:
> - unsigned long nr_segs -> long
> - unsigned long fast_seg -> long
> 
> So when calling rw_copy_check_uvector(), we have to zero-extend the
> registers used for passing the arguments. That's definitely done when
> calling the function explicitly. Maybe when inlining something is messed up?

That's also not needed on x86-64 - the high bits get cleared by 32bit writes.
But, IIRC, arm64 leaves them unchanged or undefined.

I guessing that every array access uses a *(Rx + Ry) addressing
mode. So indexing an array even with 'unsigned int' requires
an explicit zero-extend on arm64?
(x86-64 ends up with an explicit sign extend when indexing an
array with 'signed int'.)

	David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)