Re: [PATCH] kasan: fix the missing underflow in memmove and memcpy with CONFIG_KASAN_GENERIC=y

From: Walter Wu
Date: Wed Oct 02 2019 - 22:19:40 EST


On Wed, 2019-10-02 at 15:57 +0200, Dmitry Vyukov wrote:
> On Wed, Oct 2, 2019 at 2:15 PM Walter Wu <walter-zh.wu@xxxxxxxxxxxx> wrote:
> >
> > On Mon, 2019-09-30 at 12:36 +0800, Walter Wu wrote:
> > > On Fri, 2019-09-27 at 21:41 +0200, Dmitry Vyukov wrote:
> > > > On Fri, Sep 27, 2019 at 4:22 PM Walter Wu <walter-zh.wu@xxxxxxxxxxxx> wrote:
> > > > >
> > > > > On Fri, 2019-09-27 at 15:07 +0200, Dmitry Vyukov wrote:
> > > > > > On Fri, Sep 27, 2019 at 5:43 AM Walter Wu <walter-zh.wu@xxxxxxxxxxxx> wrote:
> > > > > > >
> > > > > > > memmove() and memcpy() have missing underflow issues.
> > > > > > > When -7 <= size < 0, then KASAN will miss to catch the underflow issue.
> > > > > > > It looks like shadow start address and shadow end address is the same,
> > > > > > > so it does not actually check anything.
> > > > > > >
> > > > > > > The following test is indeed not caught by KASAN:
> > > > > > >
> > > > > > > char *p = kmalloc(64, GFP_KERNEL);
> > > > > > > memset((char *)p, 0, 64);
> > > > > > > memmove((char *)p, (char *)p + 4, -2);
> > > > > > > kfree((char*)p);
> > > > > > >
> > > > > > > It should be checked here:
> > > > > > >
> > > > > > > void *memmove(void *dest, const void *src, size_t len)
> > > > > > > {
> > > > > > > check_memory_region((unsigned long)src, len, false, _RET_IP_);
> > > > > > > check_memory_region((unsigned long)dest, len, true, _RET_IP_);
> > > > > > >
> > > > > > > return __memmove(dest, src, len);
> > > > > > > }
> > > > > > >
> > > > > > > We fix the shadow end address which is calculated, then generic KASAN
> > > > > > > get the right shadow end address and detect this underflow issue.
> > > > > > >
> > > > > > > [1] https://bugzilla.kernel.org/show_bug.cgi?id=199341
> > > > > > >
> > > > > > > Signed-off-by: Walter Wu <walter-zh.wu@xxxxxxxxxxxx>
> > > > > > > Reported-by: Dmitry Vyukov <dvyukov@xxxxxxxxxx>
> > > > > > > ---
> > > > > > > lib/test_kasan.c | 36 ++++++++++++++++++++++++++++++++++++
> > > > > > > mm/kasan/generic.c | 8 ++++++--
> > > > > > > 2 files changed, 42 insertions(+), 2 deletions(-)
> > > > > > >
> > > > > > > diff --git a/lib/test_kasan.c b/lib/test_kasan.c
> > > > > > > index b63b367a94e8..8bd014852556 100644
> > > > > > > --- a/lib/test_kasan.c
> > > > > > > +++ b/lib/test_kasan.c
> > > > > > > @@ -280,6 +280,40 @@ static noinline void __init kmalloc_oob_in_memset(void)
> > > > > > > kfree(ptr);
> > > > > > > }
> > > > > > >
> > > > > > > +static noinline void __init kmalloc_oob_in_memmove_underflow(void)
> > > > > > > +{
> > > > > > > + char *ptr;
> > > > > > > + size_t size = 64;
> > > > > > > +
> > > > > > > + pr_info("underflow out-of-bounds in memmove\n");
> > > > > > > + ptr = kmalloc(size, GFP_KERNEL);
> > > > > > > + if (!ptr) {
> > > > > > > + pr_err("Allocation failed\n");
> > > > > > > + return;
> > > > > > > + }
> > > > > > > +
> > > > > > > + memset((char *)ptr, 0, 64);
> > > > > > > + memmove((char *)ptr, (char *)ptr + 4, -2);
> > > > > > > + kfree(ptr);
> > > > > > > +}
> > > > > > > +
> > > > > > > +static noinline void __init kmalloc_oob_in_memmove_overflow(void)
> > > > > > > +{
> > > > > > > + char *ptr;
> > > > > > > + size_t size = 64;
> > > > > > > +
> > > > > > > + pr_info("overflow out-of-bounds in memmove\n");
> > > > > > > + ptr = kmalloc(size, GFP_KERNEL);
> > > > > > > + if (!ptr) {
> > > > > > > + pr_err("Allocation failed\n");
> > > > > > > + return;
> > > > > > > + }
> > > > > > > +
> > > > > > > + memset((char *)ptr, 0, 64);
> > > > > > > + memmove((char *)ptr + size, (char *)ptr, 2);
> > > > > > > + kfree(ptr);
> > > > > > > +}
> > > > > > > +
> > > > > > > static noinline void __init kmalloc_uaf(void)
> > > > > > > {
> > > > > > > char *ptr;
> > > > > > > @@ -734,6 +768,8 @@ static int __init kmalloc_tests_init(void)
> > > > > > > kmalloc_oob_memset_4();
> > > > > > > kmalloc_oob_memset_8();
> > > > > > > kmalloc_oob_memset_16();
> > > > > > > + kmalloc_oob_in_memmove_underflow();
> > > > > > > + kmalloc_oob_in_memmove_overflow();
> > > > > > > kmalloc_uaf();
> > > > > > > kmalloc_uaf_memset();
> > > > > > > kmalloc_uaf2();
> > > > > > > diff --git a/mm/kasan/generic.c b/mm/kasan/generic.c
> > > > > > > index 616f9dd82d12..34ca23d59e67 100644
> > > > > > > --- a/mm/kasan/generic.c
> > > > > > > +++ b/mm/kasan/generic.c
> > > > > > > @@ -131,9 +131,13 @@ static __always_inline bool memory_is_poisoned_n(unsigned long addr,
> > > > > > > size_t size)
> > > > > > > {
> > > > > > > unsigned long ret;
> > > > > > > + void *shadow_start = kasan_mem_to_shadow((void *)addr);
> > > > > > > + void *shadow_end = kasan_mem_to_shadow((void *)addr + size - 1) + 1;
> > > > > > >
> > > > > > > - ret = memory_is_nonzero(kasan_mem_to_shadow((void *)addr),
> > > > > > > - kasan_mem_to_shadow((void *)addr + size - 1) + 1);
> > > > > > > + if ((long)size < 0)
> > > > > > > + shadow_end = kasan_mem_to_shadow((void *)addr + size);
> > > > > >
> > > > > > Hi Walter,
> > > > > >
> > > > > > Thanks for working on this.
> > > > > >
> > > > > > If size<0, does it make sense to continue at all? We will still check
> > > > > > 1PB of shadow memory? What happens when we pass such huge range to
> > > > > > memory_is_nonzero?
> > > > > > Perhaps it's better to produce an error and bail out immediately if size<0?
> > > > >
> > > > > I agree with what you said. when size<0, it is indeed an unreasonable
> > > > > behavior, it should be blocked from continuing to do.
> > > > >
> > > > >
> > > > > > Also, what's the failure mode of the tests? Didn't they badly corrupt
> > > > > > memory? We tried to keep tests such that they produce the KASAN
> > > > > > reports, but don't badly corrupt memory b/c/ we need to run all of
> > > > > > them.
> > > > >
> > > > > Maybe we should first produce KASAN reports and then go to execute
> > > > > memmove() or do nothing? It looks like itâs doing the following.or?
> > > > >
> > > > > void *memmove(void *dest, const void *src, size_t len)
> > > > > {
> > > > > + if (long(len) <= 0)
> > > >
> > > > /\/\/\/\/\/\
> > > >
> > > > This check needs to be inside of check_memory_region, otherwise we
> > > > will have similar problems in all other places that use
> > > > check_memory_region.
> > > Thanks for your reminder.
> > >
> > > bool check_memory_region(unsigned long addr, size_t size, bool write,
> > > unsigned long ret_ip)
> > > {
> > > + if (long(size) < 0) {
> > > + kasan_report_invalid_size(src, dest, len, _RET_IP_);
> > > + return false;
> > > + }
> > > +
> > > return check_memory_region_inline(addr, size, write, ret_ip);
> > > }
> > >
> > > > But check_memory_region already returns a bool, so we could check that
> > > > bool and return early.
> > >
> > > When size<0, we should only show one KASAN report, and should we only
> > > limit to return when size<0 is true? If yse, then __memmove() will do
> > > nothing.
> > >
> > >
> > > void *memmove(void *dest, const void *src, size_t len)
> > > {
> > > - check_memory_region((unsigned long)src, len, false, _RET_IP_);
> > > + if(!check_memory_region((unsigned long)src, len, false,
> > > _RET_IP_)
> > > + && long(size) < 0)
> > > + return;
> > > +
> > > check_memory_region((unsigned long)dest, len, true, _RET_IP_);
> > >
> > > return __memmove(dest, src, len);
> > >
> > > >
> > Hi Dmitry,
> >
> > What do you think the following code is better than the above one.
> > In memmmove/memset/memcpy, they need to determine whether size < 0 is
> > true. we directly determine whether size is negative in memmove and
> > return early. it avoid to generate repeated KASAN report. Is it better?
> >
> > void *memmove(void *dest, const void *src, size_t len)
> > {
> > + if (long(size) < 0) {
> > + kasan_report_invalid_size(src, dest, len, _RET_IP_);
> > + return;
> > + }
> > +
> > check_memory_region((unsigned long)src, len, false, _RET_IP_);
> > check_memory_region((unsigned long)dest, len, true, _RET_IP_);
> >
> >
> > check_memory_region() still has to check whether the size is negative.
> > but memmove/memset/memcpy generate invalid size KASAN report will not be
> > there.
>
>
> If check_memory_region() will do the check, why do we need to
> duplicate it inside of memmove and all other range functions?
>
Yes, I know it has duplication, but if we don't have to determine size<0
in memmove, then all check_memory_region return false will do nothing,
it includes other memory corruption behaviors, this is my original
concern.

> I would do:
>
> void *memmove(void *dest, const void *src, size_t len)
> {
> if (check_memory_region((unsigned long)src, len, false, _RET_IP_))
> return;
if check_memory_region return TRUE is to do nothing, but it is no memory
corruption? Should it return early when check_memory_region return a
FALSE?

>
> This avoids duplicating the check, adds minimal amount of code to
> range functions and avoids adding kasan_report_invalid_size.
Thanks for your suggestion.
We originally want to show complete information(destination address,
source address, and its length), but add minimal amount of code into
kasan_report(), it should be good.