Re: general protection fault in syscall_return_slowpath

[Dmitry Vyukov]

On Sun, Mar 8, 2020 at 7:35 PM 'Jann Horn' via syzkaller-bugs
<syzkaller-bugs@xxxxxxxxxxxxxxxx> wrote:
>
> On Sun, Mar 8, 2020 at 5:40 PM syzbot
> <syzbot+cd66e43794b178bb5cd6@xxxxxxxxxxxxxxxxxxxxxxxxx> wrote:
> > HEAD commit:    63623fd4 Merge tag 'for-linus' of git://git.kernel.org/pub..
> > git tree:       upstream
> > console output: https://syzkaller.appspot.com/x/log.txt?x=16cfeac3e00000
> > kernel config:  https://syzkaller.appspot.com/x/.config?x=5d2e033af114153f
> > dashboard link: https://syzkaller.appspot.com/bug?extid=cd66e43794b178bb5cd6
> > compiler:       clang version 10.0.0 (https://github.com/llvm/llvm-project/ c2443155a0fb245c8f17f2c1c72b6ea391e86e81)
> > syz repro:      https://syzkaller.appspot.com/x/repro.syz?x=12a42329e00000
> >
> > IMPORTANT: if you fix the bug, please add the following tag to the commit:
> > Reported-by: syzbot+cd66e43794b178bb5cd6@xxxxxxxxxxxxxxxxxxxxxxxxx
> >
> > general protection fault, probably for non-canonical address 0x1ffffffff1255a6b: 0000 [#1] PREEMPT SMP KASAN
> > CPU: 0 PID: 8742 Comm: syz-executor.2 Not tainted 5.6.0-rc3-syzkaller #0
> > Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011
> > RIP: 0010:arch_local_irq_disable arch/x86/include/asm/paravirt.h:757 [inline]
> > RIP: 0010:syscall_return_slowpath+0xeb/0x4a0 arch/x86/entry/common.c:277
> > Code: 00 10 0f 85 de 00 00 00 e8 b2 a3 76 00 48 c7 c0 58 d3 2a 89 48 c1 e8 03 80 3c 18 00 74 0c 48 c7 c7 58 d3 2a 89 e8 05 00 00 00 <00> 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
> > RSP: 0018:ffffc900020a7ed0 EFLAGS: 00010246
> > RAX: 1ffffffff1255a6b RBX: dffffc0000000000 RCX: ffff88808c512380
> > RDX: 0000000000000000 RSI: 0000000000000000 RDI: 0000000000000000
> > RBP: ffffc900020a7f10 R08: ffffffff810075bb R09: fffffbfff14d9182
> > R10: fffffbfff14d9182 R11: 0000000000000000 R12: 1ffff110118a2470
> > R13: 0000000000004000 R14: ffff88808c512380 R15: ffff88808c512380
> > FS:  000000000154f940(0000) GS:ffff8880aea00000(0000) knlGS:0000000000000000
> > CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> > CR2: 000000000076c000 CR3: 00000000a6b05000 CR4: 00000000001406f0
> > DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
> > DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
> > Call Trace:
> >  do_syscall_64+0x11f/0x1c0 arch/x86/entry/common.c:304
> >  entry_SYSCALL_64_after_hwframe+0x49/0xbe
> > BUG: kernel NULL pointer dereference, address: 0000000000000000
> > #PF: supervisor write access in kernel mode
> > #PF: error_code(0x0002) - not-present page
> > PGD 8fecc067 P4D 8fecc067 PUD 97953067 PMD 0
> > Oops: 0002 [#2] PREEMPT SMP KASAN
> > CPU: 0 PID: 8742 Comm: syz-executor.2 Not tainted 5.6.0-rc3-syzkaller #0
>
> Ugh, why does it build with -Werror...

Now I am realizing I don't know what's the proper way to turn off
warnings entirely...

We turn off this CONFIG_ERROR_ON_WARNING historically:
https://github.com/google/syzkaller/blob/2e9971bbbfb4df6ba0118353163a7703f3dbd6ec/dashboard/config/bits-syzbot.config#L17
and I thought that's enough. But now I realize it's not even a thing.
I see it referenced in some ChromeOS threads and there are some
discussions re upstreaming, but apparently it never existed upstream.

make has W=n, but it seems that it can only be used to produce more
warnings. We don't pass W=3 specifically and there is no W=0.

Should we always build with CFLAGS=-w? Is it guaranteed to work? Or is
there a better way?