Re: kernel BUG at mm/vmalloc.c:LINE! (2)

[Vegard Nossum]

On Mon, 11 Jan 2021 at 10:16, Dmitry Vyukov <dvyukov@xxxxxxxxxx> wrote:
>
> On Sun, Jan 10, 2021 at 10:34 PM syzbot
> <syzbot+5f326d255ca648131f87@xxxxxxxxxxxxxxxxxxxxxxxxx> wrote:
> >
> > syzbot suspects this issue was fixed by commit:
> >
> > commit 537cf4e3cc2f6cc9088dcd6162de573f603adc29
> > Author: Magnus Karlsson <magnus.karlsson@xxxxxxxxx>
> > Date:   Fri Nov 20 11:53:39 2020 +0000
> >
> >     xsk: Fix umem cleanup bug at socket destruct
> >
> > bisection log:  https://syzkaller.appspot.com/x/bisect.txt?x=139f3dfb500000
> > start commit:   e87d24fc Merge branch 'net-iucv-fixes-2020-11-09'
> > git tree:       net
> > kernel config:  https://syzkaller.appspot.com/x/.config?x=61033507391c77ff
> > dashboard link: https://syzkaller.appspot.com/bug?extid=5f326d255ca648131f87
> > syz repro:      https://syzkaller.appspot.com/x/repro.syz?x=10d10006500000
> > C reproducer:   https://syzkaller.appspot.com/x/repro.c?x=126c9eaa500000
> >
> > If the result looks correct, please mark the issue as fixed by replying with:
> >
> > #syz fix: xsk: Fix umem cleanup bug at socket destruct
> >
> > For information about bisection process see: https://goo.gl/tpsmEJ#bisection
>
> FTR, the bisection log looks clean, but this does not look like the
> fix for this. The reproducer does not destroy sockets.

I think it's the correct fix.

The crash report also has this, which shows the reproducer does
actually destroy sockets:

 xdp_umem_addr_unmap net/xdp/xdp_umem.c:44 [inline]
 xdp_umem_release net/xdp/xdp_umem.c:62 [inline]
 xdp_put_umem+0x113/0x330 net/xdp/xdp_umem.c:80
 xsk_destruct net/xdp/xsk.c:1150 [inline]
 xsk_destruct+0xc0/0xf0 net/xdp/xsk.c:1142
 __sk_destruct+0x4b/0x8f0 net/core/sock.c:1759
 rcu_do_batch kernel/rcu/tree.c:2476 [inline]

I've tested the reproducer on both 537cf4e3cc2f and 537cf4e3cc2f^ and
it only reproduces on 537cf4e3cc2f^ here (with the same stack trace as
the syzbot report).

The repro I used was
https://syzkaller.appspot.com/text?tag=ReproSyz&x=10d10006500000 which
is just:

r0 = socket$xdp(0x2c, 0x3, 0x0)
setsockopt$XDP_UMEM_REG(r0, 0x11b, 0x4,
&(0x7f0000000040)={&(0x7f0000000000)=""/2, 0x1000000, 0x1000}, 0x20)

so the socket definitely gets created/destroyed.

Feel free to undo if you disagree:

#syz fix: xsk: Fix umem cleanup bug at socket destruct


Vegard