Date:   Wed, 16 Jan 2019 08:15:14 -0800
From:   "Paul E. McKenney" <paulmck@...ux.ibm.com>
To:     Dmitry Vyukov <dvyukov@...gle.com>
Cc:     Jan Kara <jack@...e.cz>,
        Tetsuo Handa <penguin-kernel@...ove.sakura.ne.jp>,
        Al Viro <viro@...iv.linux.org.uk>,
        linux-fsdevel <linux-fsdevel@...r.kernel.org>,
        LKML <linux-kernel@...r.kernel.org>,
        Alan Stern <stern@...land.harvard.edu>,
        Andrea Parri <andrea.parri@...rulasolutions.com>
Subject: Re: [PATCH] fs: ratelimit __find_get_block_slow() failure message.

On Wed, Jan 16, 2019 at 04:33:05PM +0100, Dmitry Vyukov wrote:
> On Wed, Jan 16, 2019 at 3:51 PM Jan Kara <jack@...e.cz> wrote:
> >
> > On Wed 16-01-19 13:37:22, Dmitry Vyukov wrote:
> > > On Wed, Jan 16, 2019 at 12:56 PM Jan Kara <jack@...e.cz> wrote:
> > > >
> > > > On Wed 16-01-19 12:03:27, Dmitry Vyukov wrote:
> > > > > On Wed, Jan 16, 2019 at 11:43 AM Jan Kara <jack@...e.cz> wrote:
> > > > > >
> > > > > > On Wed 16-01-19 10:47:56, Dmitry Vyukov wrote:
> > > > > > > On Fri, Jan 11, 2019 at 1:46 PM Tetsuo Handa
> > > > > > > <penguin-kernel@...ove.sakura.ne.jp> wrote:
> > > > > > > >
> > > > > > > > On 2019/01/11 19:48, Dmitry Vyukov wrote:
> > > > > > > > >> How did you arrive to the conclusion that it is harmless?
> > > > > > > > >> There is only one relevant standard covering this, which is the C
> > > > > > > > >> language standard, and it is very clear on this -- this has Undefined
> > > > > > > > >> Behavior, that is the same as, for example, reading/writing random
> > > > > > > > >> pointers.
> > > > > > > > >>
> > > > > > > > >> Check out this on how any race that you might think is benign can be
> > > > > > > > >> badly miscompiled and lead to arbitrary program behavior:
> > > > > > > > >> https://software.intel.com/en-us/blogs/2013/01/06/benign-data-races-what-could-possibly-go-wrong
> > > > > > > > >
> > > > > > > > > Also there is no other practical definition of data race for automatic
> > > > > > > > > data race detectors than: two conflicting non-atomic concurrent
> > > > > > > > > accesses. Which this code is. Which means that if we continue writing
> > > > > > > > > such code we are not getting data race detection and don't detect
> > > > > > > > > thousands of races in kernel code that one may consider more harmful
> > > > > > > > > than this one the easy way. And instead will spent large amounts of
> > > > > > > > > time to fix some of then the hard way, and leave the rest as just too
> > > > > > > > > hard to debug so let the kernel continue crashing from time to time (I
> > > > > > > > > believe a portion of currently open syzbot bugs that developers just
> > > > > > > > > left as "I don't see how this can happen" are due to such races).
> > > > > > > > >
> > > > > > > >
> > > > > > > > I still cannot catch. Read/write of sizeof(long) bytes at naturally
> > > > > > > > aligned address is atomic, isn't it?
> > > > > > >
> > > > > > > Nobody guarantees this. According to C non-atomic conflicting
> > > > > > > reads/writes of sizeof(long) cause undefined behavior of the whole
> > > > > > > program.
> > > > > >
> > > > > > Yes, but to be fair the kernel has always relied on long accesses to be
> > > > > > atomic pretty heavily so that it is now de-facto standard for the kernel
> > > > > > AFAICT. I understand this makes life for static checkers hard but such is
> > > > > > reality.
> > > > >
> > > > > Yes, but nobody never defined what "a long access" means. And if you
> > > > > see a function that accepts a long argument and stores it into a long
> > > > > field, no, it does not qualify. I bet this will come at surprise to
> > > > > lots of developers.
> > > >
> > > > Yes, inlining and other optimizations can screw you.
> > > >
> > > > > Check out this fix and try to extrapolate how this "function stores
> > > > > long into a long leads to a serious security bug" can actually be
> > > > > applied to whole lot of places after inlining (or when somebody just
> > > > > slightly shuffles code in a way that looks totally safe) that also
> > > > > kinda look safe and atomic:
> > > > > https://lore.kernel.org/patchwork/patch/599779/
> > > > > So where is the boundary between "a long access" that is atomic and
> > > > > the one that is not necessary atomic?
> > > >
> > > > So I tend to rely on "long access being atomic" for opaque values (no
> > > > flags, no counters, ...). Just value that gets fetched from some global
> > > > variable / other data structure, stored, read, and possibly compared for
> > > > equality. I agree the compiler could still screw you if it could infer how
> > > > that value got initially created and try to be clever about it...
> > >
> > > So can you, or somebody else, define a set of rules that we can use to
> > > discriminate each particular case? How can we avoid that "the compiler
> > > could still screw you"?
> > >
> > > Inlining is always enabled, right, so one needs to take into account
> > > everything that's possibly can be inlined. Now or in future. And also
> > > link-time-code generation, if we don't use it we are dropping 10% of
> > > performance on the floor.
> > > Also, ensuring that the code works when it's first submitted is the
> > > smaller part of the problem. It's ensuring that it continues to work
> > > in future what's more important. Try to imagine what amount of burden
> > > this puts onto all developers who touch any kernel code in future.
> > > Basically if you slightly alter local logic in a function that does
> > > not do any loads/stores, you can screw multiple "proofs" that long
> > > accesses are atomic. Or, you just move a function from .c file to .h.
> > > I can bet nobody re-proofs all "long accesses are atomic" around the
> > > changed code during code reviews, so these things break over time.
> > > Or, even if only comparisons are involved (that you mentioned as
> > > "safe") I see how that can actually affect compilation process. Say,
> > > we are in the branch where 2 variables compare equal, now since no
> > > concurrency is involved from compiler point of view, it can, say,
> > > discard one variable and then re-load it from the other variable's
> > > location, and say not the other variable has value that the other one
> > > must never have. I don't have a full scenario, but that's exactly the
> > > point. One will never see all possibilities.
> > >
> > > It all becomes super slippery slope very quickly. And we do want
> > > compiler to generate as fast code as possible and do all these
> > > optimizations. And it's not that there are big objective reasons to
> > > not just mark all concurrent accesses and stop spending large amounts
> > > of time on these "proofs".
> >
> > I guess you've convinced me that somehow marking such accesses is
> > desirable. So is using atomic_long_t and atomic_long_set()
> > / atomic_long_read() for manipulation instead what you suggest?
> 
> +Paul, Alan, Andrea, who may better know the long-term direction.
> I think the closer analog of "plain long read/write that is intended
> to be atomic" is using READ_ONCE/WRITE_ONCE with a normal non-atomic
> type. It should achieve the goal of just read/write it once without
> any tricks, but also makes them visible/searchable for humans and
> tools and potential future refactoring to something else.

Agreed, use of READ_ONCE() and WRITE_ONCE() should suffice.

The atomic_long_set() and atomic_long_read() would also work, but are
a bit more typing.  ;-)

							Thanx, Paul

Powered by blists - more mailing lists