On Tue, Dec 17, 2019 at 10:38:09PM +0000, Catalin Marinas wrote:

On Tue, Dec 17, 2019 at 04:57:34PM +0100, Magnus Karlsson wrote:

quoted

On Tue, Dec 17, 2019 at 4:40 PM Catalin Marinas [off-list ref] wrote:

quoted

On Tue, Dec 17, 2019 at 02:27:22PM +0100, Magnus Karlsson wrote:

quoted

On Mon, Dec 16, 2019 at 4:10 PM Magnus Karlsson
[off-list ref] wrote:

quoted

On Mon, Dec 16, 2019 at 4:00 PM Daniel Borkmann [off-list ref] wrote:

quoted

On Sat, Dec 14, 2019 at 08:20:07AM -0800, syzbot wrote:

quoted

syzbot has found a reproducer for the following crash on:

HEAD commit:    1d1997db Revert "nfp: abm: fix memory leak in nfp_abm_u32_..
git tree:       net-next
console output: https://syzkaller.appspot.com/x/log.txt?x=1029f851e00000
kernel config:  https://syzkaller.appspot.com/x/.config?x=cef1fd5032faee91
dashboard link: https://syzkaller.appspot.com/bug?extid=9301f2f33873407d5b33
compiler:       gcc (GCC) 9.0.0 20181231 (experimental)
syz repro:      https://syzkaller.appspot.com/x/repro.syz?x=119d9fb1e00000

IMPORTANT: if you fix the bug, please add the following tag to the commit:
Reported-by: syzbot+9301f2f33873407d5b33@syzkaller.appspotmail.com

Bjorn / Magnus, given xsk below, PTAL, thanks!

Thanks. I will take a look at it right away.

/Magnus

After looking through the syzcaller report, I have the following
hypothesis that would dearly need some comments from MM-savy people
out there. Syzcaller creates, using mmap, a memory area that is

I guess that's not an anonymous mmap() since we don't seem to have a
struct page for src in cow_user_page() (the WARN_ON_ONCE path). Do you
have more information on the mmap() call?

I have this from the syzcaller logs:

mmap(&(0x7f0000001000/0x2000)=nil, 0x2000, 0xfffffe, 0x12, r8, 0x0)
getsockopt$XDP_MMAP_OFFSETS(r8, 0x11b, 0x7, &(0x7f0000001300),
&(0x7f0000000100)=0x60)

The full log can be found at:
https://syzkaller.appspot.com/x/repro.syz?x=119d9fb1e00000

Thanks. Prior to mmap, we have:

r8 = socket$xdp(0x2c, 0x3, 0x0)

So basically we have an mmap() on a socket descriptor with a subsequent
copy_to_user() writing this range. We do we even end up doing CoW on
such mapping?

It's a non-readable private mapping of a socket. Any write to it would
cause CoW.

BTW, how useful memory mapped sockets are? I don't know much about
networking, but it looks like a rarely used feature that substantially
increase attack surface. CAP_NET_RAW is easy to come by nowadays with
user-ns.

Few years back I was able to modify zero page via memory mapped socket...

Maybe the socket code should also implement the .fault()
file op. It needs more digging.

Caller definitely does a weird thing here that doesn't suppose to produce
a meaningful result. I think we can keep the warning for now just to make
sure we don't have any more-or-less legitimate obscure use-case.

But ultimately this WARN_ON_ONCE() has to be upgraded to SIGSEGV or
SIGBUS. Pretending that we do anything meaningful here by clearing the
page unlikely does anything useful to a user.

-- 
 Kirill A. Shutemov