Re: tty, fbcon: use-after-free in fbcon_invert_region

From: Dmitry Vyukov
Date: Thu Oct 13 2016 - 07:09:23 EST


On Wed, Oct 12, 2016 at 12:48 AM, Scot Doyle <lkml14@xxxxxxxxxxxxx> wrote:
>> On Tue, Oct 11, 2016 at 3:43 AM, Scot Doyle <lkml14@xxxxxxxxxxxxx> wrote:
>> > I wonder if the text selection is outside the newly resized vc?
>> > Does this patch help?
>> >
>> > --- vt.c 2016-10-11 00:32:43.079605599 -0000
>> > +++ vt.c.new 2016-10-11 00:36:12.744650759 -0000
>> > @@ -874,6 +874,9 @@
>> > if (!newscreen)
>> > return -ENOMEM;
>> >
>> > + if (vc == sel_cons)
>> > + clear_selection();
>> > +
>> > old_rows = vc->vc_rows;
>> > old_row_size = vc->vc_size_row;
>>
>> This helped with the use-after-frees and out-of-bounds.
>> Tested-by: Dmitry Vyukov <dvyukov@xxxxxxxxxx>
>>
>> However, now the test program hanged in D unkillable stack on some
>> kind of kernel deadlock. Don't know if it's induced by your patch, or
>> just another bug. At least there are no vc_do_resize in stacks.
>>
>> # ps afxu | grep a.out
>> root 6163 6.5 0.0 0 0 pts/0 Zl 13:25 0:00 |
>> \_ [a.out] <defunct>
>>
>> # ls /proc/6163/task/
>> 6163 6191 6193 6194 6201
>>
>> # cat /proc/6163/task/*/stack
>> [< inline >] down_read_failed drivers/tty/tty_ldsem.c:241
>> [<ffffffff831b8da6>] __ldsem_down_read_nested+0x2a6/0x5b0 drivers/tty/tty_ldsem.c:332
>> [<ffffffff831b23f5>] tty_ldisc_ref_wait+0x35/0xb0 drivers/tty/tty_ldisc.c:274
>> [<ffffffff831962b7>] tty_write+0x177/0x840 drivers/tty/tty_io.c:1250
>> [<ffffffff8182c700>] __vfs_write+0x110/0x620 fs/read_write.c:510
>> [<ffffffff8182dc05>] vfs_write+0x175/0x4e0 fs/read_write.c:560
>> [< inline >] SYSC_write fs/read_write.c:607
>> [<ffffffff818314c9>] SyS_write+0xd9/0x1b0 fs/read_write.c:599
>> [<ffffffff86daf545>] entry_SYSCALL_64_fastpath+0x23/0xc6
>> arch/x86/entry/entry_64.S:208
>
> The patch below removes the resize ioctl's from the first test program.
> Are there any use-after-free/out-of-bounds errors when running the patched
> test program on the unpatched kernel? If not, but there are still
> deadlocks, then perhaps they aren't caused by the proposed kernel patch?

Yes, I've removed these:

ioctl(0, 0x5609ul, 0); // VT_RESIZE
ioctl(0, 0x5414ul, 0); // TIOCSWINSZ

It does not crash, but still deadlocks. So I guess your patch fixes
the crashes. Please mail a patch with the fix.


> --- test.c
> +++ test.c.new
> @@ -141,8 +141,6 @@
> NONFAILING(*(uint16_t*)0x20f77ff9 = (uint16_t)0x6);
> NONFAILING(*(uint16_t*)0x20f77ffb = (uint16_t)0x3f);
> NONFAILING(*(uint16_t*)0x20f77ffd = (uint16_t)0x0);
> - r[17] = execute_syscall(__NR_ioctl, r[8], 0x541cul, 0x20f77ff4ul, 0,
> - 0, 0, 0, 0, 0);
> break;
> case 8:
> NONFAILING(*(uint32_t*)0x20f6dffc = (uint32_t)0x5);
> @@ -212,8 +210,6 @@
> NONFAILING(*(uint16_t*)0x20f78ffa = (uint16_t)0xeb8);
> NONFAILING(*(uint16_t*)0x20f78ffc = (uint16_t)0x9);
> NONFAILING(*(uint16_t*)0x20f78ffe = (uint16_t)0x7);
> - r2[17] = execute_syscall(__NR_ioctl, r2[5], 0x5609ul, 0x20f78ffaul, 0,
> - 0, 0, 0, 0, 0);
> break;
> case 8:
> r2[18] =
> @@ -273,8 +269,6 @@
> NONFAILING(*(uint16_t*)0x20f70002 = (uint16_t)0x2);
> NONFAILING(*(uint16_t*)0x20f70004 = (uint16_t)0xd1e);
> NONFAILING(*(uint16_t*)0x20f70006 = (uint16_t)0x7);
> - r2[34] = execute_syscall(__NR_ioctl, r2[5], 0x5414ul, 0x20f70000ul, 0,
> - 0, 0, 0, 0, 0);
> break;
> }
> return 0;
>