On Thu, Sep 05, 2019 at 07:50:26PM +1000, Aleksa Sarai wrote:
> On 2019-09-05, Rasmus Villemoes <linux@xxxxxxxxxxxxxxxxxx> wrote:
> > On 04/09/2019 22.19, Aleksa Sarai wrote:
> > > A common pattern for syscall extensions is increasing the size of a
> > > struct passed from userspace, such that the zero-value of the new fields
> > > result in the old kernel behaviour (allowing for a mix of userspace and
> > > kernel vintages to operate on one another in most cases). This is done
> > > in both directions -- hence two helpers -- though it's more common to
> > > have to copy user space structs into kernel space.
> > >
> > > Previously there was no common lib/ function that implemented
> > > the necessary extension-checking semantics (and different syscalls
> > > implemented them slightly differently or incompletely[1]). A future
> > > patch replaces all of the common uses of this pattern to use the new
> > > copy_struct_{to,from}_user() helpers.
> > >
> > > [1]: For instance {sched_setattr,perf_event_open,clone3}(2) all do do
> > > similar checks to copy_struct_from_user() while rt_sigprocmask(2)
> > > always rejects differently-sized struct arguments.
> > >
> > > Suggested-by: Rasmus Villemoes <linux@xxxxxxxxxxxxxxxxxx>
> > > Signed-off-by: Aleksa Sarai <cyphar@xxxxxxxxxx>
> > > ---
> > > diff --git a/lib/struct_user.c b/lib/struct_user.c
> > > new file mode 100644
> > > index 000000000000..7301ab1bbe98
> > > --- /dev/null
> > > +++ b/lib/struct_user.c
> > > @@ -0,0 +1,182 @@
> > > +// SPDX-License-Identifier: GPL-2.0-or-later
> > > +/*
> > > + * Copyright (C) 2019 SUSE LLC
> > > + * Copyright (C) 2019 Aleksa Sarai <cyphar@xxxxxxxxxx>
> > > + */
> > > +
> > > +#include <linux/types.h>
> > > +#include <linux/export.h>
> > > +#include <linux/uaccess.h>
> > > +#include <linux/kernel.h>
> > > +#include <linux/string.h>
> > > +
> > > +#define BUFFER_SIZE 64
> > > +
> > > +/*
> > > + * "memset(p, 0, size)" but for user space buffers. Caller must have already
> > > + * checked access_ok(p, size).
> > > + */
> >
> > Isn't this __clear_user() exactly (perhaps except for the return value)?
> > Perhaps not every arch has that?
>
> I didn't know about clear_user() -- I will switch to it.
>
> > > +static int __memzero_user(void __user *p, size_t s)
> > > +{
> > > + const char zeros[BUFFER_SIZE] = {};
> > > + while (s > 0) {
> > > + size_t n = min(s, sizeof(zeros));
> > > +
> > > + if (__copy_to_user(p, zeros, n))
> > > + return -EFAULT;
> > > +
> > > + p += n;
> > > + s -= n;
> > > + }
> > > + return 0;
> > > +}
> > > +
> > > +/**
> > > + * copy_struct_to_user: copy a struct to user space
> > > + * @dst: Destination address, in user space.
> > > + * @usize: Size of @dst struct.
> > > + * @src: Source address, in kernel space.
> > > + * @ksize: Size of @src struct.
> > > + *
> > > + * Returns (in all cases, some data may have been copied):
> > > + * * -EFBIG: (@usize < @ksize) and there are non-zero trailing bytes in @src.
> > > + * * -EFAULT: access to user space failed.
> > > + */
> > > +int copy_struct_to_user(void __user *dst, size_t usize,
> > > + const void *src, size_t ksize)
> > > +{
> > > + size_t size = min(ksize, usize);
> > > + size_t rest = abs(ksize - usize);
> >
> > Eh, I'd avoid abs() here due to the funkiness of the implicit type
> > conversions - ksize-usize has type size_t, then that's coerced to an int
> > (or a long maybe?), the abs is applied which return an int/long (or
> > unsigned versions?). Something like "rest = max(ksize, usize) - size;"
> > is more obviously correct and doesn't fall into any
> > narrowing/widening/sign extending traps.
>
> Yeah, I originally used "max(ksize, usize) - size" for that reason but
> was worried it looked too funky (and some quick tests showed that abs()
> gives the right results in most cases -- though I just realised it would
> probably not give the right results around SIZE_MAX). I'll switch back.
>
> > > + if (unlikely(usize > PAGE_SIZE))
> > > + return -EFAULT;
> >
> > Please don't. That is a restriction on all future extensions - once a
> > kernel is shipped with a syscall using this helper with that arbitrary
> > restriction in place, that syscall is forever prevented from extending
> > its arg struct beyond PAGE_SIZE (which is arch-dependent anyway). Sure,
> > it's hard to imagine, but who'd have thought 32 O_* or CLONE_* bits
> > weren't enough for everybody?
> >
> > This is only for future compatibility, and if someone runs an app
> > compiled against 7.3 headers on a 5.4 kernel, they probably don't care
> > about performance, but they would like their app to run.
>
> I'm not sure I agree that the limit is in place *forever* -- it's
> generally not a break in compatibility to convert an error into a
> success (though, there are counterexamples such as mknod(2) -- but that
> was a very specific case).
>
> You're right that it would mean that some very new code won't run on
> very ancient kernels (assuming we ever pass around structs that
> massive), but there should be a reasonable trade-off here IMHO.
Passing a struct larger than a PAGE_SIZE right now (at least for all
those calls that would make use of this helper at the moment) is to be
considered a bug.
The PAGE_SIZE check is a reasonable heuristic. It's an assumption that
is pretty common in the kernel in other places as well. Plus the
possibility of DoS.
>
> If we allow very large sizes, a program could probably DoS the kernel by
> allocating a moderately-large block of memory and then spawning a bunch
> of threads that all cause the kernel to re-check that the same 1GB block
> of memory is zeroed. I haven't tried, but it seems like it's best to
> avoid the possibility altogether.
>
> > > + }
> > > + /* Copy the interoperable parts of the struct. */
> > > + if (__copy_to_user(dst, src, size))
> > > + return -EFAULT;
> >
> > I think I understand why you put this last instead of handling the
> > buffer in the "natural" order. However,
> > I'm wondering whether we should actually do this copy before checking
> > that the extra kernel bytes are 0 - the user will still be told that
> > there was some extra information via the -EFBIG/-E2BIG return, but maybe
> > in some cases the part he understands is good enough. But I also guess
> > we have to look to existing users to see whether that would prevent them
> > from being converted to using this helper.
> >
> > linux-api folks, WDYT?
>
> Regarding the order, I just copied what sched and perf already do. I
> wouldn't mind doing it the other way around -- though I am a little
> cautious about implicitly making guarantees like that. The syscall that
> uses copy_struct_to_user() might not want to make that guarantee (it
> might not make sense for them), and there are some -E2BIG returns that
> won't result in data being copied (usize > PAGE_SIZE).
>
> As for feedback, this is syscall-dependent at the moment. The sched and
> perf users explicitly return the size of the kernel structure (by
> overwriting uattr->size if -E2BIG is returned) for copies in either
> direction. So users arguably already have some kind of feedback about
> size issues. clone3() on the other hand doesn't do that (though it
> doesn't copy anything to user-space so this isn't relevant to this
> particular question).
>
> Effectively, I'd like to see someone argue that this is something that
> they would personally want (before we do it).
I think the order you have right now is fine. I don't see the point of
doing work first before we have verified that things are sane.
